Skip to content

FreeToken brings datacenter-scale model serving to your desktop. Run massive models locally, fast and efficiently.

feat(cpu-moe): serve GGUF experts on the cpu and hybrid backends through ggml-cpu (#704)

Xiaoze Fanc8559fdff68c
diff --git a/docs/cli.md b/docs/cli.md
index 4918a317..69848f2e 100644
--- a/docs/cli.md
+++ b/docs/cli.md
@@ -89,7 +89,7 @@ See [models.md](models.md#moe-strategies) for what each strategy does.
| `--moe-cache-split` | equal | How the pools share the slots: `equal` (in proportion to their layer counts), or `profile:<path>`: every pool keeps `top_k` slots and the rest go to the most used experts in the profile's per-layer expert counts. With `offload`, a pool smaller than the expert count lowers `--max-running-requests` to its slots / top_k |
| `--kv-reserve-tokens` | 8192 | KV token floor reserved before `--moe-cache-auto` fills experts |
| `--moe-cpu-threads` | physical cores | CPU worker threads for the cpu/hybrid executor |
-| `--moe-cpu-layers` | all on GPU | With `offload`: which MoE layers decode on CPU (`3,7,11`, a count, a fraction, or `auto`). `auto` is for Windows/WSL only, where CUDA pinned memory is capped; every value needs an expert format the CPU executor serves (bf16, fp8_block, nvfp4, mxfp4) |
+| `--moe-cpu-layers` | all on GPU | With `offload`: which MoE layers decode on CPU (`3,7,11`, a count, a fraction, or `auto`). `auto` is for Windows/WSL only, where CUDA pinned memory is capped; every value needs an expert format the CPU executor serves (bf16, fp8_block, nvfp4, mxfp4, GGUF on x86 CPUs) |
| `--moe-hybrid-max-fetch` | auto | With `hybrid`: max experts fetched over PCIe per layer per step; rest computed on CPU |
| `--moe-prefill-hit-d2d` | off | Prefill: copy cache-hit experts device-side, stream only misses (CUDA >= 13) |
| `--disable-moe-prefill-overlap` | overlap on | Disable the two-buffer prefill copy overlap |
diff --git a/python/freetoken/engine/engine.py b/python/freetoken/engine/engine.py
index 980bb043..326c467d 100644
--- a/python/freetoken/engine/engine.py
+++ b/python/freetoken/engine/engine.py
@@ -1479,6 +1479,10 @@ def _cpu_moe_executor_viable(model_config) -> bool:
return False
expert_quant = getattr(model_config, "expert_quant", "none")
fmt = expert_quant if expert_quant != "none" else (moe_wfmt or "bf16")
+ if fmt == "gguf":
+ from freetoken.kernel import ggml_cpu
+
+ return fmt in _cpu_moe.weight_formats() and ggml_cpu.select() is not None
return fmt == "mxfp4" or fmt in _cpu_moe.weight_formats()
diff --git a/python/freetoken/kernel/csrc/cpu_moe/cpu_moe_ext.cpp b/python/freetoken/kernel/csrc/cpu_moe/cpu_moe_ext.cpp
index 6236816e..c4887efb 100644
--- a/python/freetoken/kernel/csrc/cpu_moe/cpu_moe_ext.cpp
+++ b/python/freetoken/kernel/csrc/cpu_moe/cpu_moe_ext.cpp
@@ -46,8 +46,9 @@ std::unique_ptr<MoeFormat> make_bf16_format(const MoeParams& p, const FormatArgs
std::unique_ptr<MoeFormat> make_nvfp4_format(const MoeParams& p, const FormatArgs& a);
std::unique_ptr<MoeFormat> make_mxfp4_format(const MoeParams& p, const FormatArgs& a);
std::unique_ptr<MoeFormat> make_ds_fp4_format(const MoeParams& p, const FormatArgs& a);
-std::unique_ptr<MoeFormat> make_q4_0_format(const MoeParams& p, const FormatArgs& a);
std::unique_ptr<MoeFormat> make_fp8_block_format(const MoeParams& p, const FormatArgs& a);
+std::unique_ptr<MoeFormat> make_gguf_format(const MoeParams& p, const FormatArgs& a);
+void set_ggml_cpu(uintptr_t addr);
namespace {
@@ -94,8 +95,8 @@ inline MoeTask from_word(uint64_t w) {
// Every weight format this module serves, keyed by the name cpu_executor.py passes; weight_formats() exports the names.
const std::pair<const char*, std::unique_ptr<MoeFormat> (*)(const MoeParams&, const FormatArgs&)> kFormats[] = {
- {"bf16", make_bf16_format}, {"nvfp4", make_nvfp4_format}, {"mxfp4_triton", make_mxfp4_format},
- {"ds_fp4", make_ds_fp4_format}, {"q4_0", make_q4_0_format}, {"fp8_block", make_fp8_block_format},
+ {"bf16", make_bf16_format}, {"nvfp4", make_nvfp4_format}, {"mxfp4_triton", make_mxfp4_format},
+ {"ds_fp4", make_ds_fp4_format}, {"fp8_block", make_fp8_block_format}, {"gguf", make_gguf_format},
};
std::unique_ptr<MoeFormat> make_format(const std::string& name, const MoeParams& p, const FormatArgs& a) {
@@ -573,6 +574,8 @@ PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
// (act_apply falls through to gelu_tanh); the probe turns a stale extension
// into a loud rebuild instruction instead of wrong model outputs.
m.def("max_generic_act_id", []() { return static_cast<int>(freetoken::cpu_moe::ACT_SWIGLU_CLAMP); });
+ // The ggml-cpu variant the gguf format calls: the kernel table address that freetoken.kernel.ggml_cpu.select() returns.
+ m.def("set_ggml_cpu", &freetoken::cpu_moe::set_ggml_cpu, py::arg("api"));
m.def("weight_formats", []() {
py::list names;
for (const auto& f : freetoken::cpu_moe::kFormats) names.append(f.first);
diff --git a/python/freetoken/kernel/csrc/cpu_moe/kernels/gguf.cpp b/python/freetoken/kernel/csrc/cpu_moe/kernels/gguf.cpp
new file mode 100644
index 00000000..c1641af5
--- /dev/null
+++ b/python/freetoken/kernel/csrc/cpu_moe/kernels/gguf.cpp
@@ -0,0 +1,107 @@
+#include <atomic>
+#include <mutex>
+#include <stdexcept>
+#include <string>
+#include <vector>
+
+#include "../format.h"
+#include "cpu_kernels.h"
+
+namespace freetoken::cpu_moe {
+namespace {
+
+std::atomic<const ft_ggml_cpu_api*> g_ggml{nullptr};
+
+const ft_ggml_cpu_api& ggml_api() {
+ const ft_ggml_cpu_api* api = g_ggml.load();
+ if (api == nullptr) throw std::logic_error("CPU MoE gguf experts need a ggml-cpu variant; call _cpu_moe.set_ggml_cpu first");
+ return *api;
+}
+
+} // namespace
+
+void set_ggml_cpu(uintptr_t addr) {
+ const auto* api = reinterpret_cast<const ft_ggml_cpu_api*>(addr);
+ if (api == nullptr || api->abi != FT_GGML_CPU_ABI) throw std::invalid_argument("set_ggml_cpu: not a ggml-cpu kernel table of this build");
+ const ft_ggml_cpu_api* prev = nullptr;
+ if (!g_ggml.compare_exchange_strong(prev, api) && prev != api)
+ throw std::invalid_argument(std::string("set_ggml_cpu: this process already uses the ") + prev->name + " ggml-cpu variant");
+}
+
+namespace {
+
+// Packed ggml rows of the layers whose experts have this gate_up type and this down type, dotted by ggml-cpu against the activation converted to each type's vec_dot input (q8_0, q8_1 or q8_K, or the float type itself).
+class GgufFormat final : public MoeFormat {
+ public:
+ GgufFormat(const MoeParams& p, const FormatArgs& a)
+ : MoeFormat(p), ggml_(ggml_api()), gate_up_(a.bank("gate_up")), down_(a.bank("down")) {
+ static std::once_flag tables;
+ std::call_once(tables, ggml_.init);
+ gu_ = kernel(static_cast<int>(a.option("gate_up_type")), p.H, "gate_up");
+ dn_ = kernel(static_cast<int>(a.option("down_type")), p.I, "down");
+ xf_.resize(p.H);
+ xq_.resize(static_cast<size_t>(p.max_tokens) * gu_.act_row_bytes);
+ gq_.resize(static_cast<size_t>(p.max_tokens) * p.top_k * dn_.act_row_bytes);
+ }
+
+ std::string isa_name() const override { return std::string(ggml_.name) + "(ggml-cpu)"; }
+
+ void prep_input(int tok, const bf16_t* x) override {
+ for (int i = 0; i < p.H; ++i) xf_[i] = bf16_to_f32(x[i]);
+ gu_.from_float(xf_.data(), xq_.data() + static_cast<size_t>(tok) * gu_.act_row_bytes, p.H);
+ }
+
+ bool preps_inter() const override { return true; }
+
+ void prep_inter(int64_t r, bf16_t* g) override {
+ thread_local std::vector<float> gf;
+ gf.resize(p.I);
+ for (int i = 0; i < p.I; ++i) gf[i] = bf16_to_f32(g[i]);
+ dn_.from_float(gf.data(), gq_.data() + static_cast<size_t>(r) * dn_.act_row_bytes, p.I);
+ }
+
+ void gate_up(int layer, int e, int tok, int i0, int i1, const bf16_t* /*x*/, float w_in, bf16_t* g) const override {
+ const uint8_t* w = static_cast<const uint8_t*>(tbl_at(gate_up_, layer)) + static_cast<size_t>(e) * 2 * p.I * gu_.row_bytes;
+ const uint8_t* xq = xq_.data() + static_cast<size_t>(tok) * gu_.act_row_bytes;
+ gated_rows(p, i0, i1, w_in, g, [&](int row) {
+ float s;
+ gu_.vec_dot(p.H, &s, 0, w + static_cast<size_t>(row) * gu_.row_bytes, 0, xq, 0, 1);
+ return s;
+ });
+ }
+
+ void down(int layer, int h0, int h1, const DownRoute* routes, int n, float* out) const override {
+ const uint8_t* w = static_cast<const uint8_t*>(tbl_at(down_, layer));
+ for (int h = h0; h < h1; ++h)
+ for (int j = 0; j < n; ++j)
+ dn_.vec_dot(p.I, &out[j * kRowBlock + h - h0], 0, w + (static_cast<size_t>(routes[j].e) * p.H + h) * dn_.row_bytes, 0,
+ gq_.data() + static_cast<size_t>(routes[j].r) * dn_.act_row_bytes, 0, 1);
+ }
+
+ private:
+ ft_ggml_row_kernel kernel(int type, int k, const char* role) const {
+ ft_ggml_row_kernel out{};
+ const int rc = ggml_.row_kernel(type, k, &out);
+ const std::string name = ggml_.type_name(type);
+ if (rc == FT_GGML_NO_KERNEL) throw std::invalid_argument("ggml-cpu has no dot product for " + name + " experts");
+ if (rc != FT_GGML_OK)
+ throw std::invalid_argument("CPU MoE " + name + " " + role + " rows of " + std::to_string(k) + " values are not whole blocks");
+ return out;
+ }
+
+ const ft_ggml_cpu_api& ggml_;
+ const uint64_t* gate_up_;
+ const uint64_t* down_;
+ ft_ggml_row_kernel gu_{}, dn_{};
+ std::vector<float> xf_;
+ // the activation in the vec_dot input type: x per token, g per route
+ AlignedVector<uint8_t> xq_, gq_;
+};
+
+} // namespace
+
+std::unique_ptr<MoeFormat> make_gguf_format(const MoeParams& p, const FormatArgs& a) {
+ return std::make_unique<GgufFormat>(p, a);
+}
+
+} // namespace freetoken::cpu_moe
diff --git a/python/freetoken/kernel/csrc/cpu_moe/kernels/q4_0.cpp b/python/freetoken/kernel/csrc/cpu_moe/kernels/q4_0.cpp
deleted file mode 100644
index 600a194e..00000000
--- a/python/freetoken/kernel/csrc/cpu_moe/kernels/q4_0.cpp
+++ /dev/null
@@ -1,216 +0,0 @@
-#include <stdexcept>
-#include <vector>
-
-#include "../format.h"
-
-namespace freetoken::cpu_moe {
-namespace {
-
-// -------------------------------- Q4_0 (W4A8) --------------------------------
-// Native GGUF Q4_0 experts (gemma4 GGUF): per-32 block = fp16 scale d + 16 packed
-// bytes; byte j holds element j in its low nibble and j+16 in its high nibble, so a
-// block's storage order is [lo0..lo15, hi0..hi15] and w = (nibble - 8) * d. Matches
-// the reference dequant (models/gguf/dequant.py) and the packed banks the GPU offload
-// path streams.
-//
-// llama.cpp ggml_vec_dot_q4_0_q8_0: W4A8. The activation is pre-quantized to Q8_0
-// (per-32-block int8 ``aq`` + fp32 scale ``asb``); each block unpacks its 16 bytes to
-// 32 int8 weights in [-8,7] (bytes_from_nibbles_32: low nibbles -> elems 0..15, high
-// -> 16..31) and runs an integer block dot -- VPDPBUSD (AVX-VNNI) or VPMADDUBSW+VPMADDWD
-// (AVX2) with the ggml sign trick |w|*(sign(w)*a)=w*a, or a scalar int loop -- then
-// scales the block sum by wd*xd in fp32. No fp weight dequant / shuffle chain. The GPU
-// offload path (ggml_moe_a8_vec / MMVQ) is also W4A8, so cpu and hybrid stay close.
-using q4dot_fn = float (*)(const uint8_t*, const int8_t*, const float*, int);
-
-float q4_0_dot_i8_scalar(const uint8_t* w, const int8_t* aq, const float* asb, int K) {
- float acc = 0.0f;
- const int nb = K / 32;
- for (int b = 0; b < nb; ++b) {
- const uint8_t* blk = w + (size_t)b * 18;
- uint16_t dh;
- std::memcpy(&dh, blk, sizeof(dh));
- const uint8_t* q = blk + 2; // 16 nibble bytes
- const int8_t* a = aq + (size_t)b * 32;
- int isum = 0;
- for (int j = 0; j < 16; ++j) {
- isum += ((int)(q[j] & 0x0F) - 8) * (int)a[j]; // elem j
- isum += ((int)(q[j] >> 4) - 8) * (int)a[16 + j]; // elem 16+j
- }
- acc += fp16_to_f32(dh) * asb[b] * (float)isum;
- }
- return acc;
-}
-
-#if CPU_MOE_X86
-// fp16 block scale -> fp32 via HW F16C (single value in lane 0).
-CPU_MOE_TARGET("f16c")
-static inline float q4_scale(uint16_t h) {
- return _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128((int)h)));
-}
-
-// Unpack one Q4_0 block's 16 bytes -> 32 int8 weights in [-8,7] (elems 0..15 = low
-// nibbles, 16..31 = high nibbles). ``eight`` = _mm256_set1_epi8(8).
-CPU_MOE_TARGET("avx2")
-static inline __m256i q4_unpack32(const uint8_t* blk, __m128i mask, __m256i eight) {
- const __m128i qb = _mm_loadu_si128(reinterpret_cast<const __m128i*>(blk + 2));
- const __m128i lo = _mm_and_si128(qb, mask);
- const __m128i hi = _mm_and_si128(_mm_srli_epi16(qb, 4), mask);
- return _mm256_sub_epi8(_mm256_set_m128i(hi, lo), eight);
-}
-
-// AVX2 W4A8 (llama.cpp non-VNNI mul_sum_i8_pairs): integer block dot via VPMADDUBSW +
-// VPMADDWD (sign trick), scaled by wd*xd. |aw*sa| pair sums <= 8*127*2 < 32767 -> no
-// int16 saturation. This is the fast path on AVX2 CPUs without AVX-VNNI (and the
-// avx512-tier fallback, since the block dot is 256-bit either way).
-CPU_MOE_TARGET("avx2,fma,f16c")
-float q4_0_dot_i8_avx2(const uint8_t* w, const int8_t* aq, const float* asb, int K) {
- const __m128i mask = _mm_set1_epi8(0x0F);
- const __m256i eight = _mm256_set1_epi8(8);
- const __m256i ones16 = _mm256_set1_epi16(1);
- __m256 accF = _mm256_setzero_ps();
- const int nb = K / 32;
- for (int b = 0; b < nb; ++b) {
- const uint8_t* blk = w + (size_t)b * 18;
- _mm_prefetch(reinterpret_cast<const char*>(blk) + 512, _MM_HINT_T0);
- uint16_t dh;
- std::memcpy(&dh, blk, sizeof(dh));
- __m256i wq = q4_unpack32(blk, mask, eight);
- __m256i a = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(aq + (size_t)b * 32));
- __m256i aw = _mm256_sign_epi8(wq, wq); // |wq| (unsigned operand)
- __m256i sa = _mm256_sign_epi8(a, wq); // sign(wq) * a (signed operand)
- __m256i d32 = _mm256_madd_epi16(_mm256_maddubs_epi16(aw, sa), ones16); // 8 int32
- accF = _mm256_fmadd_ps(_mm256_cvtepi32_ps(d32), _mm256_set1_ps(q4_scale(dh) * asb[b]), accF);
- }
- return hsum256(accF);
-}
-
-// AVX-VNNI W4A8: one VPDPBUSD per block (the fast path on modern CPUs).
-CPU_MOE_TARGET("avx2,avxvnni,fma,f16c")
-float q4_0_dot_i8_vnni(const uint8_t* w, const int8_t* aq, const float* asb, int K) {
- const __m128i mask = _mm_set1_epi8(0x0F);
- const __m256i eight = _mm256_set1_epi8(8);
- __m256 accF = _mm256_setzero_ps();
- const int nb = K / 32;
- for (int b = 0; b < nb; ++b) {
- const uint8_t* blk = w + (size_t)b * 18;
- _mm_prefetch(reinterpret_cast<const char*>(blk) + 512, _MM_HINT_T0);
- uint16_t dh;
- std::memcpy(&dh, blk, sizeof(dh));
- __m256i wq = q4_unpack32(blk, mask, eight);
- __m256i a = _mm256_loadu_si256(reinterpret_cast<const __m256i*>(aq + (size_t)b * 32));
- __m256i aw = _mm256_sign_epi8(wq, wq); // |wq| (unsigned operand)
- __m256i sa = _mm256_sign_epi8(a, wq); // sign(wq) * a (signed operand)
- __m256i di = _mm256_dpbusd_avx_epi32(_mm256_setzero_si256(), aw, sa);
- // All 32 elems of the block share wd*xd; distribute over di's 8 partial sums and
- // reduce at the end (equivalent to scale * block_total).
- accF = _mm256_fmadd_ps(_mm256_cvtepi32_ps(di), _mm256_set1_ps(q4_scale(dh) * asb[b]), accF);
- }
- return hsum256(accF);
-}
-#endif // CPU_MOE_X86
-
-// All tiers are W4A8 (int8 activations pre-quantized to Q8_0). AVX-VNNI is orthogonal to
-// the ISA tier (gated by cpu_has_avxvnni() / FREETOKEN_CPU_MOE_NO_VNNI), so it wins when
-// present; otherwise the 256-bit VPMADDUBSW kernel covers both the avx2 and avx512 tiers.
-q4dot_fn select_q4dot() {
- const IsaTier t = pick_isa();
-#if CPU_MOE_X86
- if (cpu_has_avxvnni()) return q4_0_dot_i8_vnni;
- if (t >= ISA_AVX2) return q4_0_dot_i8_avx2;
-#endif
- (void)t;
- return q4_0_dot_i8_scalar;
-}
-
-// Quantize a bf16 activation row to Q8_0 (llama.cpp): per-32-block symmetric int8 in
-// natural order + one fp32 scale (amax/127) per block. Done once per token/route,
-// amortized over every output row that the q4_0 W4A8 GEMV reads. K % 32 == 0.
-void quant_q8_0(const bf16_t* x, int K, int8_t* aq, float* asb) {
- const int nb = K / 32;
- for (int b = 0; b < nb; ++b) {
- const bf16_t* xb = x + (size_t)b * 32;
- float xf[32], amax = 0.0f;
- for (int j = 0; j < 32; ++j) {
- xf[j] = bf16_to_f32(xb[j]);
- amax = std::max(amax, std::fabs(xf[j]));
- }
- const float d = amax > 0.0f ? amax / 127.0f : 1.0f;
- asb[b] = d;
- const float inv = amax > 0.0f ? 1.0f / d : 0.0f;
- int8_t* o = aq + (size_t)b * 32;
- for (int j = 0; j < 32; ++j)
- o[j] = (int8_t)std::max(-127, std::min(127, (int)std::lround(xf[j] * inv)));
- }
-}
-
-class Q4_0Format final : public MoeFormat {
- public:
- Q4_0Format(const MoeParams& p, const FormatArgs& a) : MoeFormat(p), dot_(select_q4dot()) {
- gate_up_ = a.bank("gate_up");
- down_ = a.bank("down");
- const size_t tokens = p.max_tokens, routes = (size_t)p.max_tokens * p.top_k;
- xi8_.resize(tokens * p.H);
- xas_.resize(tokens * (p.H / 32));
- gi8_.resize(routes * p.I);
- gas_.resize(routes * (p.I / 32));
- if (p.H % 32 != 0 || p.I % 32 != 0)
- throw std::runtime_error("Q4_0 CPU MoE requires H and I to be multiples of 32");
- gu_row_bytes_ = (p.H / 32) * 18; // K = H (gate_up rows)
- dn_row_bytes_ = (p.I / 32) * 18; // K = I (down rows)
- // select_q4dot picks VPDPBUSD / VPMADDUBSW / scalar for the tier, so the tag reflects which of those it resolved to.
- isa_ = std::string(tier_name(pick_isa())) + (cpu_has_avxvnni() ? "+vnni(q4_0-w4a8)" : "+q4_0-w4a8");
- }
-
- std::string isa_name() const override { return isa_; }
-
-
- void prep_input(int tok, const bf16_t* x) override {
- quant_q8_0(x, p.H, xi8_.data() + (size_t)tok * p.H, xas_.data() + (size_t)tok * (p.H / 32));
- }
-
- bool preps_inter() const override { return true; }
-
- void prep_inter(int64_t r, bf16_t* g) override {
- quant_q8_0(g, p.I, gi8_.data() + (size_t)r * p.I, gas_.data() + (size_t)r * (p.I / 32));
- }
-
- void gate_up(int layer, int e, int tok, int i0, int i1, const bf16_t* /*x*/, float w_in,
- bf16_t* g) const override {
- const uint8_t* packed = static_cast<const uint8_t*>(tbl_at(gate_up_, layer));
- const int8_t* xi8 = xi8_.data() + (size_t)tok * p.H;
- const float* xas = xas_.data() + (size_t)tok * (p.H / 32);
- gated_rows(p, i0, i1, w_in, g, [&](int row) {
- return dot_(packed + ((size_t)e * (2 * p.I) + row) * (size_t)gu_row_bytes_, xi8, xas, p.H);
- });
- }
-
- void down(int layer, int h0, int h1, const DownRoute* routes, int n, float* out) const override {
- const uint8_t* packed = static_cast<const uint8_t*>(tbl_at(down_, layer));
- for (int h = h0; h < h1; ++h) {
- for (int j = 0; j < n; ++j) {
- const size_t r = (size_t)routes[j].r;
- out[j * kRowBlock + h - h0] = dot_(packed + ((size_t)routes[j].e * p.H + h) * (size_t)dn_row_bytes_,
- gi8_.data() + r * p.I, gas_.data() + r * (p.I / 32), p.I);
- }
- }
- }
-
- private:
- q4dot_fn dot_;
- // GGUF Q4_0 rows of 32-element blocks (fp16 scale + 16 nibble bytes): gate_up [E,2I,H/32*18], down [E,H,I/32*18].
- const uint64_t* gate_up_ = nullptr;
- const uint64_t* down_ = nullptr;
- int gu_row_bytes_ = 0, dn_row_bytes_ = 0;
- std::string isa_;
- // Q8_0 activations: per-32-block int8 + fp32 scale, for the input and the intermediate.
- AlignedVector<int8_t> xi8_, gi8_;
- AlignedVector<float> xas_, gas_;
-};
-
-} // namespace
-
-std::unique_ptr<MoeFormat> make_q4_0_format(const MoeParams& p, const FormatArgs& a) {
- return std::make_unique<Q4_0Format>(p, a);
-}
-
-} // namespace freetoken::cpu_moe
diff --git a/python/freetoken/kernel/csrc/ggml/VERSION b/python/freetoken/kernel/csrc/ggml/VERSION
index 6eabb0c1..3c66f9b7 100644
--- a/python/freetoken/kernel/csrc/ggml/VERSION
+++ b/python/freetoken/kernel/csrc/ggml/VERSION
@@ -1,7 +1,8 @@
llama.cpp 8c2134a8755d8a9792af513dab20282affe903e4 (2026-10-14)
-Only the ggml files that the GGUF ops (freetoken/ggml_ops.cu, built by kernel/ggml.py) compile or include:
-ggml/include -> include, ggml/src -> src, ggml/src/ggml-cuda -> ggml-cuda.
+Only the ggml files that FreeToken compiles or includes:
+- the GGUF ops (freetoken/ggml_ops.cu, built by kernel/ggml.py): ggml/include -> include, ggml/src -> src, ggml/src/ggml-cuda -> ggml-cuda.
+- the CPU MoE row kernels (freetoken/cpu_kernels.c and freetoken/variant_module.cpp, one freetoken.kernel.ggml_cpu module per x86 variant, built by setup.py): ggml/src/{ggml-quants.c,ggml-quants.h,ggml-backend-impl.h} -> src, ggml/include/ggml-cpu.h -> include, ggml/src/ggml-cpu/{quants.c,quants.h,vec.cpp,vec.h,ggml-cpu-impl.h,simd-mappings.h,arch-fallback.h,arch/x86/quants.c,arch/x86/cpu-feats.cpp} -> ggml-cpu. These have no local changes.
Local changes, kept in freetoken.patch (already applied):
- quantize.cu/.cuh: the q8_1 quantize kernels read fp32, bf16 or fp16 activations.
diff --git a/python/freetoken/kernel/csrc/ggml/freetoken/cpu_kernels.c b/python/freetoken/kernel/csrc/ggml/freetoken/cpu_kernels.c
new file mode 100644
index 00000000..79a03638
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/freetoken/cpu_kernels.c
@@ -0,0 +1,172 @@
+// The ggml-base symbols and lookup tables that the vendored ggml-cpu quant kernels reference, and the row kernel of each weight type.
+
+#include "cpu_kernels.h"
+
+#include "ggml-impl.h"
+#include "ggml-cpu/ggml-cpu-impl.h"
+#include "ggml-cpu/quants.h"
+#include "ggml-cpu/simd-mappings.h"
+#include "ggml-cpu/vec.h"
+
+#include <stdarg.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+float ggml_table_f32_f16[1 << 16];
+float ggml_table_f32_e8m0_half[1 << 8];
+float ggml_table_f32_ue4m3[1 << 8];
+
+void ggml_abort(const char * file, int line, const char * fmt, ...) {
+ va_list args;
+ va_start(args, fmt);
+ fprintf(stderr, "ggml: %s:%d: ", file, line);
+ vfprintf(stderr, fmt, args);
+ va_end(args);
+ fputc('\n', stderr);
+ fflush(stderr);
+ abort();
+}
+
+struct type_info {
+ const char * name;
+ int64_t blck_size;
+ size_t type_size;
+};
+
+static const struct type_info k_types[GGML_TYPE_COUNT] = {
+ [GGML_TYPE_F32] = {"f32", 1, sizeof(float)},
+ [GGML_TYPE_F16] = {"f16", 1, sizeof(ggml_fp16_t)},
+ [GGML_TYPE_BF16] = {"bf16", 1, sizeof(ggml_bf16_t)},
+ [GGML_TYPE_Q1_0] = {"q1_0", QK1_0, sizeof(block_q1_0)},
+ [GGML_TYPE_Q2_0] = {"q2_0", QK2_0, sizeof(block_q2_0)},
+ [GGML_TYPE_Q4_0] = {"q4_0", QK4_0, sizeof(block_q4_0)},
+ [GGML_TYPE_Q4_1] = {"q4_1", QK4_1, sizeof(block_q4_1)},
+ [GGML_TYPE_Q5_0] = {"q5_0", QK5_0, sizeof(block_q5_0)},
+ [GGML_TYPE_Q5_1] = {"q5_1", QK5_1, sizeof(block_q5_1)},
+ [GGML_TYPE_Q8_0] = {"q8_0", QK8_0, sizeof(block_q8_0)},
+ [GGML_TYPE_Q8_1] = {"q8_1", QK8_1, sizeof(block_q8_1)},
+ [GGML_TYPE_MXFP4] = {"mxfp4", QK_MXFP4, sizeof(block_mxfp4)},
+ [GGML_TYPE_NVFP4] = {"nvfp4", QK_NVFP4, sizeof(block_nvfp4)},
+ [GGML_TYPE_Q2_K] = {"q2_K", QK_K, sizeof(block_q2_K)},
+ [GGML_TYPE_Q3_K] = {"q3_K", QK_K, sizeof(block_q3_K)},
+ [GGML_TYPE_Q4_K] = {"q4_K", QK_K, sizeof(block_q4_K)},
+ [GGML_TYPE_Q5_K] = {"q5_K", QK_K, sizeof(block_q5_K)},
+ [GGML_TYPE_Q6_K] = {"q6_K", QK_K, sizeof(block_q6_K)},
+ [GGML_TYPE_Q8_K] = {"q8_K", QK_K, sizeof(block_q8_K)},
+ [GGML_TYPE_TQ1_0] = {"tq1_0", QK_K, sizeof(block_tq1_0)},
+ [GGML_TYPE_TQ2_0] = {"tq2_0", QK_K, sizeof(block_tq2_0)},
+ [GGML_TYPE_IQ2_XXS] = {"iq2_xxs", QK_K, sizeof(block_iq2_xxs)},
+ [GGML_TYPE_IQ2_XS] = {"iq2_xs", QK_K, sizeof(block_iq2_xs)},
+ [GGML_TYPE_IQ2_S] = {"iq2_s", QK_K, sizeof(block_iq2_s)},
+ [GGML_TYPE_IQ3_XXS] = {"iq3_xxs", QK_K, sizeof(block_iq3_xxs)},
+ [GGML_TYPE_IQ3_S] = {"iq3_s", QK_K, sizeof(block_iq3_s)},
+ [GGML_TYPE_IQ1_S] = {"iq1_s", QK_K, sizeof(block_iq1_s)},
+ [GGML_TYPE_IQ1_M] = {"iq1_m", QK_K, sizeof(block_iq1_m)},
+ [GGML_TYPE_IQ4_NL] = {"iq4_nl", QK4_NL, sizeof(block_iq4_nl)},
+ [GGML_TYPE_IQ4_XS] = {"iq4_xs", QK_K, sizeof(block_iq4_xs)},
+};
+
+static int known_type(int type) { return type >= 0 && type < GGML_TYPE_COUNT && k_types[type].name != NULL; }
+
+int64_t ggml_blck_size(enum ggml_type type) { return k_types[type].blck_size; }
+
+size_t ggml_type_size(enum ggml_type type) { return k_types[type].type_size; }
+
+size_t ggml_row_size(enum ggml_type type, int64_t ne) { return k_types[type].type_size * ne / k_types[type].blck_size; }
+
+const char * ggml_type_name(enum ggml_type type) { return known_type(type) ? k_types[type].name : "unknown"; }
+
+const char * ft_ggml_type_name(int type) { return ggml_type_name((enum ggml_type) type); }
+
+struct dot_info {
+ ft_ggml_vec_dot_t vec_dot;
+ int act_type;
+};
+
+// The weight types of type_traits_cpu in ggml-cpu.c, with their vec_dot and vec_dot_type.
+static const struct dot_info k_dots[GGML_TYPE_COUNT] = {
+ [GGML_TYPE_F32] = {(ft_ggml_vec_dot_t) ggml_vec_dot_f32, GGML_TYPE_F32},
+ [GGML_TYPE_F16] = {(ft_ggml_vec_dot_t) ggml_vec_dot_f16, GGML_TYPE_F16},
+ [GGML_TYPE_BF16] = {(ft_ggml_vec_dot_t) ggml_vec_dot_bf16, GGML_TYPE_BF16},
+ [GGML_TYPE_Q1_0] = {ggml_vec_dot_q1_0_q8_0, GGML_TYPE_Q8_0},
+ [GGML_TYPE_Q2_0] = {ggml_vec_dot_q2_0_q8_0, GGML_TYPE_Q8_0},
+ [GGML_TYPE_Q4_0] = {ggml_vec_dot_q4_0_q8_0, GGML_TYPE_Q8_0},
+ [GGML_TYPE_Q4_1] = {ggml_vec_dot_q4_1_q8_1, GGML_TYPE_Q8_1},
+ [GGML_TYPE_Q5_0] = {ggml_vec_dot_q5_0_q8_0, GGML_TYPE_Q8_0},
+ [GGML_TYPE_Q5_1] = {ggml_vec_dot_q5_1_q8_1, GGML_TYPE_Q8_1},
+ [GGML_TYPE_Q8_0] = {ggml_vec_dot_q8_0_q8_0, GGML_TYPE_Q8_0},
+ [GGML_TYPE_MXFP4] = {ggml_vec_dot_mxfp4_q8_0, GGML_TYPE_Q8_0},
+ [GGML_TYPE_NVFP4] = {ggml_vec_dot_nvfp4_q8_0, GGML_TYPE_Q8_0},
+ [GGML_TYPE_Q2_K] = {ggml_vec_dot_q2_K_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_Q3_K] = {ggml_vec_dot_q3_K_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_Q4_K] = {ggml_vec_dot_q4_K_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_Q5_K] = {ggml_vec_dot_q5_K_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_Q6_K] = {ggml_vec_dot_q6_K_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_TQ1_0] = {ggml_vec_dot_tq1_0_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_TQ2_0] = {ggml_vec_dot_tq2_0_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_IQ2_XXS] = {ggml_vec_dot_iq2_xxs_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_IQ2_XS] = {ggml_vec_dot_iq2_xs_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_IQ2_S] = {ggml_vec_dot_iq2_s_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_IQ3_XXS] = {ggml_vec_dot_iq3_xxs_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_IQ3_S] = {ggml_vec_dot_iq3_s_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_IQ1_S] = {ggml_vec_dot_iq1_s_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_IQ1_M] = {ggml_vec_dot_iq1_m_q8_K, GGML_TYPE_Q8_K},
+ [GGML_TYPE_IQ4_NL] = {ggml_vec_dot_iq4_nl_q8_0, GGML_TYPE_Q8_0},
+ [GGML_TYPE_IQ4_XS] = {ggml_vec_dot_iq4_xs_q8_K, GGML_TYPE_Q8_K},
+};
+
+static void fp32_to_fp32_row(const float * x, void * y, int64_t k) { memcpy(y, x, (size_t) k * sizeof(float)); }
+
+static void fp32_to_fp16_row(const float * x, void * y, int64_t k) {
+ ggml_fp16_t * out = y;
+ for (int64_t i = 0; i < k; ++i) {
+ out[i] = GGML_CPU_FP32_TO_FP16(x[i]);
+ }
+}
+
+static void fp32_to_bf16_row(const float * x, void * y, int64_t k) {
+ ggml_bf16_t * out = y;
+ for (int64_t i = 0; i < k; ++i) {
+ out[i] = GGML_FP32_TO_BF16(x[i]);
+ }
+}
+
+static ft_ggml_from_float_t act_quantizer(int act_type) {
+ switch (act_type) {
+ case GGML_TYPE_F32: return fp32_to_fp32_row;
+ case GGML_TYPE_F16: return fp32_to_fp16_row;
+ case GGML_TYPE_BF16: return fp32_to_bf16_row;
+ case GGML_TYPE_Q8_0: return quantize_row_q8_0;
+ case GGML_TYPE_Q8_1: return quantize_row_q8_1;
+ case GGML_TYPE_Q8_K: return quantize_row_q8_K;
+ default: return NULL;
+ }
+}
+
+// The tables ggml_cpu_init fills for the quant kernels.
+void ft_ggml_cpu_init(void) {
+ for (int i = 0; i < (1 << 16); ++i) {
+ ggml_table_f32_f16[i] = GGML_COMPUTE_FP16_TO_FP32((ggml_fp16_t) i);
+ }
+ for (int i = 0; i < (1 << 8); ++i) {
+ ggml_table_f32_e8m0_half[i] = GGML_E8M0_TO_FP32_HALF(i);
+ ggml_table_f32_ue4m3[i] = ggml_ue4m3_to_fp32((uint8_t) i);
+ }
+}
+
+int ft_ggml_cpu_row_kernel(int type, int64_t k, struct ft_ggml_row_kernel * out) {
+ if (!known_type(type) || k_dots[type].vec_dot == NULL) {
+ return FT_GGML_NO_KERNEL;
+ }
+ const int act = k_dots[type].act_type;
+ if (k % k_types[type].blck_size != 0 || k % k_types[act].blck_size != 0) {
+ return FT_GGML_PARTIAL_BLOCK;
+ }
+ out->vec_dot = k_dots[type].vec_dot;
+ out->from_float = act_quantizer(act);
+ out->act_type = act;
+ out->row_bytes = ggml_row_size((enum ggml_type) type, k);
+ out->act_row_bytes = ggml_row_size((enum ggml_type) act, k);
+ return FT_GGML_OK;
+}
diff --git a/python/freetoken/kernel/csrc/ggml/freetoken/cpu_kernels.h b/python/freetoken/kernel/csrc/ggml/freetoken/cpu_kernels.h
new file mode 100644
index 00000000..d58c1748
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/freetoken/cpu_kernels.h
@@ -0,0 +1,43 @@
+// ggml-cpu row kernels for the CPU MoE executor. setup.py builds them once per x86 variant into freetoken.kernel.ggml_cpu._<variant>; call them only through the table of a variant whose score() is not 0.
+#pragma once
+
+#include <stddef.h>
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+typedef void (*ft_ggml_vec_dot_t)(int n, float * s, size_t bs, const void * x, size_t bx, const void * y, size_t by, int nrc);
+typedef void (*ft_ggml_from_float_t)(const float * x, void * y, int64_t k);
+
+struct ft_ggml_row_kernel {
+ ft_ggml_vec_dot_t vec_dot; // one weight row . one activation row in act_type
+ ft_ggml_from_float_t from_float; // fp32 activation row -> act_type
+ int act_type;
+ size_t row_bytes; // one weight row of k values
+ size_t act_row_bytes; // one activation row of k values in act_type
+};
+
+enum { FT_GGML_OK = 0, FT_GGML_NO_KERNEL = 1, FT_GGML_PARTIAL_BLOCK = 2 };
+
+void ft_ggml_cpu_init(void);
+
+// The kernels for weight rows of ggml type `type` over k values.
+int ft_ggml_cpu_row_kernel(int type, int64_t k, struct ft_ggml_row_kernel * out);
+
+const char * ft_ggml_type_name(int type);
+
+// One variant's functions, as its module's api() hands them to _cpu_moe.set_ggml_cpu.
+#define FT_GGML_CPU_ABI 1
+struct ft_ggml_cpu_api {
+ int abi; // FT_GGML_CPU_ABI of the module that built it
+ const char * name; // the variant, e.g. "haswell"
+ void (*init)(void);
+ int (*row_kernel)(int type, int64_t k, struct ft_ggml_row_kernel * out);
+ const char * (*type_name)(int type);
+};
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/python/freetoken/kernel/csrc/ggml/freetoken/variant_module.cpp b/python/freetoken/kernel/csrc/ggml/freetoken/variant_module.cpp
new file mode 100644
index 00000000..8f3d3e52
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/freetoken/variant_module.cpp
@@ -0,0 +1,32 @@
+// One ggml-cpu variant as a Python module: score() is 0 unless this CPU has every ISA extension the variant is built for, and api() is the address of its kernel table.
+// setup.py builds this file without the variant's ISA flags, like upstream's <variant>-feats target, so importing the module and calling score() is safe on any x86 CPU.
+#define PY_SSIZE_T_CLEAN
+#include <Python.h>
+
+#include "cpu_kernels.h"
+#include "ggml-cpu/arch/x86/cpu-feats.cpp"
+
+#define FT_CAT_(a, b) a##b
+#define FT_CAT(a, b) FT_CAT_(a, b)
+#define FT_STR_(a) #a
+#define FT_STR(a) FT_STR_(a)
+
+namespace {
+
+const ft_ggml_cpu_api kApi = {FT_GGML_CPU_ABI, FT_STR(FT_GGML_VARIANT), ft_ggml_cpu_init, ft_ggml_cpu_row_kernel, ft_ggml_type_name};
+
+PyObject* score(PyObject*, PyObject*) { return PyLong_FromLong(ggml_backend_cpu_x86_score()); }
+
+PyObject* api(PyObject*, PyObject*) { return PyLong_FromVoidPtr(const_cast<ft_ggml_cpu_api*>(&kApi)); }
+
+PyMethodDef kMethods[] = {
+ {"score", score, METH_NOARGS, "0 if this CPU lacks an ISA extension the variant is built for, else upstream's variant score."},
+ {"api", api, METH_NOARGS, "The address of the variant's kernel table, for _cpu_moe.set_ggml_cpu."},
+ {nullptr, nullptr, 0, nullptr},
+};
+
+PyModuleDef kModule = {PyModuleDef_HEAD_INIT, "_" FT_STR(FT_GGML_VARIANT), nullptr, -1, kMethods};
+
+} // namespace
+
+PyMODINIT_FUNC FT_CAT(PyInit__, FT_GGML_VARIANT)(void) { return PyModule_Create(&kModule); }
diff --git a/python/freetoken/kernel/csrc/ggml/ggml-cpu/arch-fallback.h b/python/freetoken/kernel/csrc/ggml/ggml-cpu/arch-fallback.h
new file mode 100644
index 00000000..4e59d0ae
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/ggml-cpu/arch-fallback.h
@@ -0,0 +1,375 @@
+
+#pragma once
+
+// Rename `_generic` functions if no native implementation is available.
+// This effectively selects the generic implementation.
+
+#if defined(GGML_CPU_GENERIC)
+// quants.c
+#define quantize_row_q8_0_generic quantize_row_q8_0
+#define quantize_row_q8_1_generic quantize_row_q8_1
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_q4_0_q8_0_generic ggml_vec_dot_q4_0_q8_0
+#define ggml_vec_dot_q4_1_q8_1_generic ggml_vec_dot_q4_1_q8_1
+#define ggml_vec_dot_q5_0_q8_0_generic ggml_vec_dot_q5_0_q8_0
+#define ggml_vec_dot_q5_1_q8_1_generic ggml_vec_dot_q5_1_q8_1
+#define ggml_vec_dot_q8_0_q8_0_generic ggml_vec_dot_q8_0_q8_0
+#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
+#define ggml_vec_dot_q3_K_q8_K_generic ggml_vec_dot_q3_K_q8_K
+#define ggml_vec_dot_q4_K_q8_K_generic ggml_vec_dot_q4_K_q8_K
+#define ggml_vec_dot_q5_K_q8_K_generic ggml_vec_dot_q5_K_q8_K
+#define ggml_vec_dot_q6_K_q8_K_generic ggml_vec_dot_q6_K_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q1_0_4x4_q8_0_generic ggml_gemv_q1_0_4x4_q8_0
+#define ggml_gemv_q1_0_4x8_q8_0_generic ggml_gemv_q1_0_4x8_q8_0
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
+#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
+#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
+#define ggml_gemm_q1_0_4x4_q8_0_generic ggml_gemm_q1_0_4x4_q8_0
+#define ggml_gemm_q1_0_4x8_q8_0_generic ggml_gemm_q1_0_4x8_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
+#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
+#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
+#elif defined(__aarch64__) || defined(__arm__) || defined(_M_ARM) || defined(_M_ARM64) || defined(_M_ARM64EC)
+// repack.cpp
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#elif defined(__x86_64__) || defined(__i386__) || defined(_M_IX86) || defined(_M_X64)
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
+#define ggml_gemv_q1_0_4x4_q8_0_generic ggml_gemv_q1_0_4x4_q8_0
+#define ggml_gemv_q1_0_4x8_q8_0_generic ggml_gemv_q1_0_4x8_q8_0
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
+#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
+#define ggml_gemm_q1_0_4x4_q8_0_generic ggml_gemm_q1_0_4x4_q8_0
+#define ggml_gemm_q1_0_4x8_q8_0_generic ggml_gemm_q1_0_4x8_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
+#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
+#elif defined(__POWERPC__) || defined(__powerpc__)
+// ref: https://github.com/ggml-org/llama.cpp/pull/14146#issuecomment-2972561679
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q1_0_4x4_q8_0_generic ggml_gemv_q1_0_4x4_q8_0
+#define ggml_gemv_q1_0_4x8_q8_0_generic ggml_gemv_q1_0_4x8_q8_0
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
+#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
+#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
+#define ggml_gemm_q1_0_4x4_q8_0_generic ggml_gemm_q1_0_4x4_q8_0
+#define ggml_gemm_q1_0_4x8_q8_0_generic ggml_gemm_q1_0_4x8_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
+#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
+#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
+#elif defined(__loongarch64)
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q1_0_4x4_q8_0_generic ggml_gemv_q1_0_4x4_q8_0
+#define ggml_gemv_q1_0_4x8_q8_0_generic ggml_gemv_q1_0_4x8_q8_0
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
+#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
+#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
+#define ggml_gemm_q1_0_4x4_q8_0_generic ggml_gemm_q1_0_4x4_q8_0
+#define ggml_gemm_q1_0_4x8_q8_0_generic ggml_gemm_q1_0_4x8_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
+#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
+#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
+#elif defined(__riscv)
+// quants.c
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x1_generic ggml_quantize_mat_q8_0_4x1
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_K_4x1_generic ggml_quantize_mat_q8_K_4x1
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q1_0_4x4_q8_0_generic ggml_gemv_q1_0_4x4_q8_0
+#define ggml_gemv_q1_0_4x8_q8_0_generic ggml_gemv_q1_0_4x8_q8_0
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
+#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
+#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
+#define ggml_gemm_q1_0_4x4_q8_0_generic ggml_gemm_q1_0_4x4_q8_0
+#define ggml_gemm_q1_0_4x8_q8_0_generic ggml_gemm_q1_0_4x8_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
+#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
+#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
+#elif defined(__s390x__)
+// quants.c
+#define quantize_row_q8_K_generic quantize_row_q8_K
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_q2_K_q8_K_generic ggml_vec_dot_q2_K_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q1_0_4x4_q8_0_generic ggml_gemv_q1_0_4x4_q8_0
+#define ggml_gemv_q1_0_4x8_q8_0_generic ggml_gemv_q1_0_4x8_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
+#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
+#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
+#define ggml_gemm_q1_0_4x4_q8_0_generic ggml_gemm_q1_0_4x4_q8_0
+#define ggml_gemm_q1_0_4x8_q8_0_generic ggml_gemm_q1_0_4x8_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
+#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
+#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
+#elif defined(__wasm__)
+// quants.c
+#define ggml_vec_dot_tq1_0_q8_K_generic ggml_vec_dot_tq1_0_q8_K
+#define ggml_vec_dot_tq2_0_q8_K_generic ggml_vec_dot_tq2_0_q8_K
+#define ggml_vec_dot_iq2_xxs_q8_K_generic ggml_vec_dot_iq2_xxs_q8_K
+#define ggml_vec_dot_iq2_xs_q8_K_generic ggml_vec_dot_iq2_xs_q8_K
+#define ggml_vec_dot_iq2_s_q8_K_generic ggml_vec_dot_iq2_s_q8_K
+#define ggml_vec_dot_iq3_xxs_q8_K_generic ggml_vec_dot_iq3_xxs_q8_K
+#define ggml_vec_dot_iq3_s_q8_K_generic ggml_vec_dot_iq3_s_q8_K
+#define ggml_vec_dot_iq1_s_q8_K_generic ggml_vec_dot_iq1_s_q8_K
+#define ggml_vec_dot_iq1_m_q8_K_generic ggml_vec_dot_iq1_m_q8_K
+#define ggml_vec_dot_iq4_nl_q8_0_generic ggml_vec_dot_iq4_nl_q8_0
+#define ggml_vec_dot_iq4_xs_q8_K_generic ggml_vec_dot_iq4_xs_q8_K
+#define ggml_vec_dot_mxfp4_q8_0_generic ggml_vec_dot_mxfp4_q8_0
+#define ggml_vec_dot_nvfp4_q8_0_generic ggml_vec_dot_nvfp4_q8_0
+#define ggml_vec_dot_q1_0_q8_0_generic ggml_vec_dot_q1_0_q8_0
+#define ggml_vec_dot_q2_0_q8_0_generic ggml_vec_dot_q2_0_q8_0
+// repack.cpp
+#define ggml_quantize_mat_q8_0_4x4_generic ggml_quantize_mat_q8_0_4x4
+#define ggml_quantize_mat_q8_0_4x8_generic ggml_quantize_mat_q8_0_4x8
+#define ggml_quantize_mat_q8_K_4x4_generic ggml_quantize_mat_q8_K_4x4
+#define ggml_quantize_mat_q8_K_4x8_generic ggml_quantize_mat_q8_K_4x8
+#define ggml_gemv_q1_0_4x4_q8_0_generic ggml_gemv_q1_0_4x4_q8_0
+#define ggml_gemv_q1_0_4x8_q8_0_generic ggml_gemv_q1_0_4x8_q8_0
+#define ggml_gemv_q4_0_4x4_q8_0_generic ggml_gemv_q4_0_4x4_q8_0
+#define ggml_gemv_q4_0_4x8_q8_0_generic ggml_gemv_q4_0_4x8_q8_0
+#define ggml_gemv_q4_0_8x8_q8_0_generic ggml_gemv_q4_0_8x8_q8_0
+#define ggml_gemv_q2_K_8x8_q8_K_generic ggml_gemv_q2_K_8x8_q8_K
+#define ggml_gemv_q4_K_8x4_q8_K_generic ggml_gemv_q4_K_8x4_q8_K
+#define ggml_gemv_q4_K_8x8_q8_K_generic ggml_gemv_q4_K_8x8_q8_K
+#define ggml_gemv_q5_K_8x4_q8_K_generic ggml_gemv_q5_K_8x4_q8_K
+#define ggml_gemv_q5_K_8x8_q8_K_generic ggml_gemv_q5_K_8x8_q8_K
+#define ggml_gemv_q6_K_8x4_q8_K_generic ggml_gemv_q6_K_8x4_q8_K
+#define ggml_gemv_q6_K_8x8_q8_K_generic ggml_gemv_q6_K_8x8_q8_K
+#define ggml_gemv_iq4_nl_4x4_q8_0_generic ggml_gemv_iq4_nl_4x4_q8_0
+#define ggml_gemv_iq4_nl_8x8_q8_0_generic ggml_gemv_iq4_nl_8x8_q8_0
+#define ggml_gemv_mxfp4_4x4_q8_0_generic ggml_gemv_mxfp4_4x4_q8_0
+#define ggml_gemv_mxfp4_8x8_q8_0_generic ggml_gemv_mxfp4_8x8_q8_0
+#define ggml_gemv_q8_0_4x4_q8_0_generic ggml_gemv_q8_0_4x4_q8_0
+#define ggml_gemv_q8_0_4x8_q8_0_generic ggml_gemv_q8_0_4x8_q8_0
+#define ggml_gemm_q1_0_4x4_q8_0_generic ggml_gemm_q1_0_4x4_q8_0
+#define ggml_gemm_q1_0_4x8_q8_0_generic ggml_gemm_q1_0_4x8_q8_0
+#define ggml_gemm_q4_0_4x4_q8_0_generic ggml_gemm_q4_0_4x4_q8_0
+#define ggml_gemm_q4_0_4x8_q8_0_generic ggml_gemm_q4_0_4x8_q8_0
+#define ggml_gemm_q4_0_8x8_q8_0_generic ggml_gemm_q4_0_8x8_q8_0
+#define ggml_gemm_q2_K_8x8_q8_K_generic ggml_gemm_q2_K_8x8_q8_K
+#define ggml_gemm_q4_K_8x4_q8_K_generic ggml_gemm_q4_K_8x4_q8_K
+#define ggml_gemm_q4_K_8x8_q8_K_generic ggml_gemm_q4_K_8x8_q8_K
+#define ggml_gemm_q5_K_8x4_q8_K_generic ggml_gemm_q5_K_8x4_q8_K
+#define ggml_gemm_q5_K_8x8_q8_K_generic ggml_gemm_q5_K_8x8_q8_K
+#define ggml_gemm_q6_K_8x4_q8_K_generic ggml_gemm_q6_K_8x4_q8_K
+#define ggml_gemm_q6_K_8x8_q8_K_generic ggml_gemm_q6_K_8x8_q8_K
+#define ggml_gemm_iq4_nl_4x4_q8_0_generic ggml_gemm_iq4_nl_4x4_q8_0
+#define ggml_gemm_iq4_nl_8x8_q8_0_generic ggml_gemm_iq4_nl_8x8_q8_0
+#define ggml_gemm_mxfp4_4x4_q8_0_generic ggml_gemm_mxfp4_4x4_q8_0
+#define ggml_gemm_mxfp4_8x8_q8_0_generic ggml_gemm_mxfp4_8x8_q8_0
+#define ggml_gemm_q8_0_4x4_q8_0_generic ggml_gemm_q8_0_4x4_q8_0
+#define ggml_gemm_q8_0_4x8_q8_0_generic ggml_gemm_q8_0_4x8_q8_0
+#endif
diff --git a/python/freetoken/kernel/csrc/ggml/ggml-cpu/arch/x86/cpu-feats.cpp b/python/freetoken/kernel/csrc/ggml/ggml-cpu/arch/x86/cpu-feats.cpp
new file mode 100644
index 00000000..d775a036
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/ggml-cpu/arch/x86/cpu-feats.cpp
@@ -0,0 +1,327 @@
+#include "ggml-backend-impl.h"
+
+#if defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
+
+#ifdef _MSC_VER
+#include <intrin.h>
+#endif
+
+#include <cstring>
+#include <vector>
+#include <bitset>
+#include <array>
+#include <string>
+
+// ref: https://cdrdv2-public.intel.com/782156/325383-sdm-vol-2abcd.pdf
+struct cpuid_x86 {
+ bool SSE3(void) { return f_1_ecx[0]; }
+ bool PCLMULQDQ(void) { return f_1_ecx[1]; }
+ bool MONITOR(void) { return f_1_ecx[3]; }
+ bool SSSE3(void) { return f_1_ecx[9]; }
+ bool FMA(void) { return f_1_ecx[12]; }
+ bool CMPXCHG16B(void) { return f_1_ecx[13]; }
+ bool SSE41(void) { return f_1_ecx[19]; }
+ bool SSE42(void) { return f_1_ecx[20]; }
+ bool MOVBE(void) { return f_1_ecx[22]; }
+ bool POPCNT(void) { return f_1_ecx[23]; }
+ bool AES(void) { return f_1_ecx[25]; }
+ bool XSAVE(void) { return f_1_ecx[26]; }
+ bool OSXSAVE(void) { return f_1_ecx[27]; }
+ bool AVX(void) { return f_1_ecx[28]; }
+ bool F16C(void) { return f_1_ecx[29]; }
+ bool RDRAND(void) { return f_1_ecx[30]; }
+
+ bool MSR(void) { return f_1_edx[5]; }
+ bool CX8(void) { return f_1_edx[8]; }
+ bool SEP(void) { return f_1_edx[11]; }
+ bool CMOV(void) { return f_1_edx[15]; }
+ bool CLFSH(void) { return f_1_edx[19]; }
+ bool MMX(void) { return f_1_edx[23]; }
+ bool FXSR(void) { return f_1_edx[24]; }
+ bool SSE(void) { return f_1_edx[25]; }
+ bool SSE2(void) { return f_1_edx[26]; }
+
+ bool FSGSBASE(void) { return f_7_ebx[0]; }
+ bool BMI1(void) { return f_7_ebx[3]; }
+ bool HLE(void) { return is_intel && f_7_ebx[4]; }
+ bool AVX2(void) { return f_7_ebx[5]; }
+ bool BMI2(void) { return f_7_ebx[8]; }
+ bool ERMS(void) { return f_7_ebx[9]; }
+ bool INVPCID(void) { return f_7_ebx[10]; }
+ bool RTM(void) { return is_intel && f_7_ebx[11]; }
+ bool AVX512F(void) { return f_7_ebx[16]; }
+ bool AVX512DQ(void) { return f_7_ebx[17]; }
+ bool RDSEED(void) { return f_7_ebx[18]; }
+ bool ADX(void) { return f_7_ebx[19]; }
+ bool AVX512PF(void) { return f_7_ebx[26]; }
+ bool AVX512ER(void) { return f_7_ebx[27]; }
+ bool AVX512CD(void) { return f_7_ebx[28]; }
+ bool AVX512BW(void) { return f_7_ebx[30]; }
+ bool AVX512VL(void) { return f_7_ebx[31]; }
+
+ bool SHA(void) { return f_7_ebx[29]; }
+
+ bool PREFETCHWT1(void) { return f_7_ecx[0]; }
+
+ bool LAHF(void) { return f_81_ecx[0]; }
+ bool LZCNT(void) { return is_intel && f_81_ecx[5]; }
+ bool ABM(void) { return is_amd && f_81_ecx[5]; }
+ bool SSE4a(void) { return is_amd && f_81_ecx[6]; }
+ bool XOP(void) { return is_amd && f_81_ecx[11]; }
+ bool TBM(void) { return is_amd && f_81_ecx[21]; }
+
+ bool SYSCALL(void) { return is_intel && f_81_edx[11]; }
+ bool MMXEXT(void) { return is_amd && f_81_edx[22]; }
+ bool RDTSCP(void) { return is_intel && f_81_edx[27]; }
+ bool _3DNOWEXT(void) { return is_amd && f_81_edx[30]; }
+ bool _3DNOW(void) { return is_amd && f_81_edx[31]; }
+
+ bool AVX512_VBMI(void) { return f_7_ecx[1]; }
+ bool AVX512_VNNI(void) { return f_7_ecx[11]; }
+ bool AVX512_FP16(void) { return f_7_edx[23]; }
+ bool AVX512_BF16(void) { return f_7_1_eax[5]; }
+ bool AVX_VNNI(void) { return f_7_1_eax[4]; }
+
+ bool AMX_TILE(void) { return f_7_edx[24]; }
+ bool AMX_INT8(void) { return f_7_edx[25]; }
+ bool AMX_FP16(void) { return f_7_1_eax[21]; }
+ bool AMX_BF16(void) { return f_7_edx[22]; }
+
+#ifdef _MSC_VER
+ static void cpuid(int cpu_info[4], int eax) {
+ __cpuid(cpu_info, eax);
+ }
+ static void cpuidex(int cpu_info[4], int eax, int ecx) {
+ __cpuidex(cpu_info, eax, ecx);
+ }
+#else
+ static void cpuid(int cpu_info[4], int eax) {
+ __asm__ __volatile__(
+ "cpuid"
+ : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
+ : "a"(eax), "c"(0));
+ }
+ static void cpuidex(int cpu_info[4], int eax, int ecx) {
+ __asm__ __volatile__(
+ "cpuid"
+ : "=a"(cpu_info[0]), "=b"(cpu_info[1]), "=c"(cpu_info[2]), "=d"(cpu_info[3])
+ : "a"(eax), "c"(ecx));
+ }
+#endif
+
+ cpuid_x86() {
+ std::array<int, 4> cpui;
+ std::vector<std::array<int, 4>> data;
+
+ // calling __cpuid with 0x0 as the function_id argument
+ // gets the number of the highest valid function ID.
+ cpuid(cpui.data(), 0);
+ int n_ids = cpui[0];
+
+ for (int i = 0; i <= n_ids; ++i) {
+ cpuidex(cpui.data(), i, 0);
+ data.push_back(cpui);
+ }
+
+ // capture vendor string
+ char vendor[0x20] = {};
+ *reinterpret_cast<int *>(vendor) = data[0][1];
+ *reinterpret_cast<int *>(vendor + 4) = data[0][3];
+ *reinterpret_cast<int *>(vendor + 8) = data[0][2];
+ this->vendor = vendor;
+ if (this->vendor == "GenuineIntel") {
+ is_intel = true;
+ } else if (this->vendor == "AuthenticAMD") {
+ is_amd = true;
+ }
+
+ // load bitset with flags for function 0x00000001
+ if (n_ids >= 1) {
+ f_1_ecx = data[1][2];
+ f_1_edx = data[1][3];
+ }
+
+ // load bitset with flags for function 0x00000007
+ if (n_ids >= 7) {
+ f_7_ebx = data[7][1];
+ f_7_ecx = data[7][2];
+ f_7_edx = data[7][3];
+ cpuidex(cpui.data(), 7, 1);
+ f_7_1_eax = cpui[0];
+ }
+
+ // calling __cpuid with 0x80000000 as the function_id argument
+ // gets the number of the highest valid extended ID.
+ cpuid(cpui.data(), 0x80000000);
+ unsigned int n_ex_ids = cpui[0];
+
+ std::vector<std::array<int, 4>> ext_data;
+ for (unsigned int i = 0x80000000; i <= n_ex_ids; ++i) {
+ cpuidex(cpui.data(), i, 0);
+ ext_data.push_back(cpui);
+ }
+
+ // load bitset with flags for function 0x80000001
+ if (n_ex_ids >= 0x80000001) {
+ f_81_ecx = ext_data[1][2];
+ f_81_edx = ext_data[1][3];
+ }
+
+ // interpret CPU brand string if reported
+ char brand[0x40] = {};
+ if (n_ex_ids >= 0x80000004) {
+ std::memcpy(brand, ext_data[2].data(), sizeof(cpui));
+ std::memcpy(brand + 16, ext_data[3].data(), sizeof(cpui));
+ std::memcpy(brand + 32, ext_data[4].data(), sizeof(cpui));
+ this->brand = brand;
+ }
+ }
+
+ bool is_intel = false;
+ bool is_amd = false;
+ std::string vendor;
+ std::string brand;
+ std::bitset<32> f_1_ecx;
+ std::bitset<32> f_1_edx;
+ std::bitset<32> f_7_ebx;
+ std::bitset<32> f_7_ecx;
+ std::bitset<32> f_7_edx;
+ std::bitset<32> f_7_1_eax;
+ std::bitset<32> f_81_ecx;
+ std::bitset<32> f_81_edx;
+};
+
+#if 0
+void test_x86_is() {
+ cpuid_x86 is;
+ printf("CPU Vendor: %s\n", is.vendor.c_str());
+ printf("Brand: %s\n", is.brand.c_str());
+ printf("is_intel: %d\n", is.is_intel);
+ printf("is_amd: %d\n", is.is_amd);
+ printf("sse3: %d\n", is.SSE3());
+ printf("pclmulqdq: %d\n", is.PCLMULQDQ());
+ printf("ssse3: %d\n", is.SSSE3());
+ printf("fma: %d\n", is.FMA());
+ printf("cmpxchg16b: %d\n", is.CMPXCHG16B());
+ printf("sse41: %d\n", is.SSE41());
+ printf("sse42: %d\n", is.SSE42());
+ printf("movbe: %d\n", is.MOVBE());
+ printf("popcnt: %d\n", is.POPCNT());
+ printf("aes: %d\n", is.AES());
+ printf("xsave: %d\n", is.XSAVE());
+ printf("osxsave: %d\n", is.OSXSAVE());
+ printf("avx: %d\n", is.AVX());
+ printf("f16c: %d\n", is.F16C());
+ printf("rdrand: %d\n", is.RDRAND());
+ printf("msr: %d\n", is.MSR());
+ printf("cx8: %d\n", is.CX8());
+ printf("sep: %d\n", is.SEP());
+ printf("cmov: %d\n", is.CMOV());
+ printf("clflush: %d\n", is.CLFSH());
+ printf("mmx: %d\n", is.MMX());
+ printf("fxsr: %d\n", is.FXSR());
+ printf("sse: %d\n", is.SSE());
+ printf("sse2: %d\n", is.SSE2());
+ printf("fsgsbase: %d\n", is.FSGSBASE());
+ printf("bmi1: %d\n", is.BMI1());
+ printf("hle: %d\n", is.HLE());
+ printf("avx2: %d\n", is.AVX2());
+ printf("bmi2: %d\n", is.BMI2());
+ printf("erms: %d\n", is.ERMS());
+ printf("invpcid: %d\n", is.INVPCID());
+ printf("rtm: %d\n", is.RTM());
+ printf("avx512f: %d\n", is.AVX512F());
+ printf("rdseed: %d\n", is.RDSEED());
+ printf("adx: %d\n", is.ADX());
+ printf("avx512pf: %d\n", is.AVX512PF());
+ printf("avx512er: %d\n", is.AVX512ER());
+ printf("avx512cd: %d\n", is.AVX512CD());
+ printf("sha: %d\n", is.SHA());
+ printf("prefetchwt1: %d\n", is.PREFETCHWT1());
+ printf("lahf: %d\n", is.LAHF());
+ printf("lzcnt: %d\n", is.LZCNT());
+ printf("abm: %d\n", is.ABM());
+ printf("sse4a: %d\n", is.SSE4a());
+ printf("xop: %d\n", is.XOP());
+ printf("tbm: %d\n", is.TBM());
+ printf("syscall: %d\n", is.SYSCALL());
+ printf("mmxext: %d\n", is.MMXEXT());
+ printf("rdtscp: %d\n", is.RDTSCP());
+ printf("3dnowext: %d\n", is._3DNOWEXT());
+ printf("3dnow: %d\n", is._3DNOW());
+ printf("avx512_vbmi: %d\n", is.AVX512_VBMI());
+ printf("avx512_vnni: %d\n", is.AVX512_VNNI());
+ printf("avx512_fp16: %d\n", is.AVX512_FP16());
+ printf("avx512_bf16: %d\n", is.AVX512_BF16());
+ printf("amx_tile: %d\n", is.AMX_TILE());
+ printf("amx_int8: %d\n", is.AMX_INT8());
+ printf("amx_fp16: %d\n", is.AMX_FP16());
+ printf("amx_bf16: %d\n", is.AMX_BF16());
+}
+#endif
+
+static int ggml_backend_cpu_x86_score() {
+ // FIXME: this does not check for OS support
+
+ int score = 1;
+ cpuid_x86 is;
+
+#ifdef GGML_FMA
+ if (!is.FMA()) { return 0; }
+ score += 1;
+#endif
+#ifdef GGML_F16C
+ if (!is.F16C()) { return 0; }
+ score += 1<<1;
+#endif
+#ifdef GGML_SSE42
+ if (!is.SSE42()) { return 0; }
+ score += 1<<2;
+#endif
+#ifdef GGML_BMI2
+ if (!is.BMI2()) { return 0; }
+ score += 1<<3;
+#endif
+#ifdef GGML_AVX
+ if (!is.AVX()) { return 0; }
+ score += 1<<4;
+#endif
+#ifdef GGML_AVX2
+ if (!is.AVX2()) { return 0; }
+ score += 1<<5;
+#endif
+#ifdef GGML_AVX_VNNI
+ if (!is.AVX_VNNI()) { return 0; }
+ score += 1<<6;
+#endif
+#ifdef GGML_AVX512
+ if (!is.AVX512F()) { return 0; }
+ if (!is.AVX512CD()) { return 0; }
+ if (!is.AVX512VL()) { return 0; }
+ if (!is.AVX512DQ()) { return 0; }
+ if (!is.AVX512BW()) { return 0; }
+ score += 1<<7;
+#endif
+#ifdef GGML_AVX512_VBMI
+ if (!is.AVX512_VBMI()) { return 0; }
+ score += 1<<8;
+#endif
+#ifdef GGML_AVX512_BF16
+ if (!is.AVX512_BF16()) { return 0; }
+ score += 1<<9;
+#endif
+#ifdef GGML_AVX512_VNNI
+ if (!is.AVX512_VNNI()) { return 0; }
+ score += 1<<10;
+#endif
+#ifdef GGML_AMX_INT8
+ if (!is.AMX_INT8()) { return 0; }
+ score += 1<<11;
+#endif
+
+ return score;
+}
+
+GGML_BACKEND_DL_SCORE_IMPL(ggml_backend_cpu_x86_score)
+
+#endif // defined(__x86_64__) || (defined(_MSC_VER) && defined(_M_AMD64))
diff --git a/python/freetoken/kernel/csrc/ggml/ggml-cpu/arch/x86/quants.c b/python/freetoken/kernel/csrc/ggml/ggml-cpu/arch/x86/quants.c
new file mode 100644
index 00000000..21cd6312
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/ggml-cpu/arch/x86/quants.c
@@ -0,0 +1,4172 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu.h"
+#include "simd-mappings.h"
+
+#include "../../quants.h"
+#include "../../ggml-cpu-impl.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h> // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+// some compilers don't provide _mm256_set_m128i, e.g. gcc 7
+#define MM256_SET_M128I(a, b) _mm256_insertf128_si256(_mm256_castsi128_si256(b), (a), 1)
+
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
+// multiply int8_t, add results pairwise twice
+static inline __m128i mul_sum_i8_pairs(const __m128i x, const __m128i y) {
+ // Get absolute values of x vectors
+ const __m128i ax = _mm_sign_epi8(x, x);
+ // Sign the values of the y vectors
+ const __m128i sy = _mm_sign_epi8(y, x);
+ // Perform multiplication and create 16-bit values
+ const __m128i dot = _mm_maddubs_epi16(ax, sy);
+ const __m128i ones = _mm_set1_epi16(1);
+ return _mm_madd_epi16(ones, dot);
+}
+
+#if __AVX__ || __AVX2__ || __AVX512F__
+// horizontally add 8 floats
+static inline float hsum_float_8(const __m256 x) {
+ __m128 res = _mm256_extractf128_ps(x, 1);
+ res = _mm_add_ps(res, _mm256_castps256_ps128(x));
+ res = _mm_add_ps(res, _mm_movehl_ps(res, res));
+ res = _mm_add_ss(res, _mm_movehdup_ps(res));
+ return _mm_cvtss_f32(res);
+}
+
+// horizontally add 8 int32_t
+static inline int hsum_i32_8(const __m256i a) {
+ const __m128i sum128 = _mm_add_epi32(_mm256_castsi256_si128(a), _mm256_extractf128_si256(a, 1));
+ const __m128i hi64 = _mm_unpackhi_epi64(sum128, sum128);
+ const __m128i sum64 = _mm_add_epi32(hi64, sum128);
+ const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
+ return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
+}
+
+// horizontally add 4 int32_t
+static inline int hsum_i32_4(const __m128i a) {
+ const __m128i hi64 = _mm_unpackhi_epi64(a, a);
+ const __m128i sum64 = _mm_add_epi32(hi64, a);
+ const __m128i hi32 = _mm_shuffle_epi32(sum64, _MM_SHUFFLE(2, 3, 0, 1));
+ return _mm_cvtsi128_si32(_mm_add_epi32(sum64, hi32));
+}
+
+#if defined(__AVX2__) || defined(__AVX512F__)
+static inline __m256i mul_add_epi8(const __m256i x, const __m256i y) {
+ const __m256i ax = _mm256_sign_epi8(x, x);
+ const __m256i sy = _mm256_sign_epi8(y, x);
+ return _mm256_maddubs_epi16(ax, sy);
+}
+
+// spread 32 bits to 32 bytes { 0x00, 0xFF }
+static inline __m256i bytes_from_bits_32(const uint8_t * x) {
+ uint32_t x32;
+ memcpy(&x32, x, sizeof(uint32_t));
+ const __m256i shuf_mask = _mm256_set_epi64x(
+ 0x0303030303030303, 0x0202020202020202,
+ 0x0101010101010101, 0x0000000000000000);
+ __m256i bytes = _mm256_shuffle_epi8(_mm256_set1_epi32(x32), shuf_mask);
+ const __m256i bit_mask = _mm256_set1_epi64x(0x7fbfdfeff7fbfdfe);
+ bytes = _mm256_or_si256(bytes, bit_mask);
+ return _mm256_cmpeq_epi8(bytes, _mm256_set1_epi64x(-1));
+}
+
+// Unpack 32 4-bit fields into 32 bytes
+// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
+static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
+{
+ const __m128i tmp = _mm_loadu_si128((const __m128i *)rsi);
+ const __m256i bytes = MM256_SET_M128I(_mm_srli_epi16(tmp, 4), tmp);
+ const __m256i lowMask = _mm256_set1_epi8( 0xF );
+ return _mm256_and_si256(lowMask, bytes);
+}
+
+// add int16_t pairwise and return as float vector
+static inline __m256 sum_i16_pairs_float(const __m256i x) {
+ const __m256i ones = _mm256_set1_epi16(1);
+ const __m256i summed_pairs = _mm256_madd_epi16(ones, x);
+ return _mm256_cvtepi32_ps(summed_pairs);
+}
+
+static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
+#if defined(__AVX512VNNI__) && defined(__AVX512VL__)
+ const __m256i zero = _mm256_setzero_si256();
+ const __m256i summed_pairs = _mm256_dpbusd_epi32(zero, ax, sy);
+ return _mm256_cvtepi32_ps(summed_pairs);
+#elif defined(__AVXVNNI__)
+ const __m256i zero = _mm256_setzero_si256();
+ const __m256i summed_pairs = _mm256_dpbusd_avx_epi32(zero, ax, sy);
+ return _mm256_cvtepi32_ps(summed_pairs);
+#else
+ // Perform multiplication and create 16-bit values
+ const __m256i dot = _mm256_maddubs_epi16(ax, sy);
+ return sum_i16_pairs_float(dot);
+#endif
+}
+
+// multiply int8_t, add results pairwise twice and return as float vector
+static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
+#if __AVXVNNIINT8__
+ const __m256i zero = _mm256_setzero_si256();
+ const __m256i summed_pairs = _mm256_dpbssd_epi32(zero, x, y);
+ return _mm256_cvtepi32_ps(summed_pairs);
+#else
+ // Get absolute values of x vectors
+ const __m256i ax = _mm256_sign_epi8(x, x);
+ // Sign the values of the y vectors
+ const __m256i sy = _mm256_sign_epi8(y, x);
+ return mul_sum_us8_pairs_float(ax, sy);
+#endif
+}
+
+static inline __m128i packNibbles( __m256i bytes )
+{
+ // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
+#if __AVX512F__
+ const __m256i bytes_srli_4 = _mm256_srli_epi16(bytes, 4); // 0000_0000_abcd_0000
+ bytes = _mm256_or_si256(bytes, bytes_srli_4); // 0000_abcd_abcd_efgh
+ return _mm256_cvtepi16_epi8(bytes); // abcd_efgh
+#else
+ const __m256i lowByte = _mm256_set1_epi16( 0xFF );
+ __m256i high = _mm256_andnot_si256( lowByte, bytes );
+ __m256i low = _mm256_and_si256( lowByte, bytes );
+ high = _mm256_srli_epi16( high, 4 );
+ bytes = _mm256_or_si256( low, high );
+
+ // Compress uint16_t lanes into bytes
+ __m128i r0 = _mm256_castsi256_si128( bytes );
+ __m128i r1 = _mm256_extracti128_si256( bytes, 1 );
+ return _mm_packus_epi16( r0, r1 );
+#endif
+}
+#elif defined(__AVX__)
+static inline __m128i packNibbles( __m128i bytes1, __m128i bytes2 )
+{
+ // Move bits within 16-bit lanes from 0000_abcd_0000_efgh into 0000_0000_abcd_efgh
+ const __m128i lowByte = _mm_set1_epi16( 0xFF );
+ __m128i high = _mm_andnot_si128( lowByte, bytes1 );
+ __m128i low = _mm_and_si128( lowByte, bytes1 );
+ high = _mm_srli_epi16( high, 4 );
+ bytes1 = _mm_or_si128( low, high );
+ high = _mm_andnot_si128( lowByte, bytes2 );
+ low = _mm_and_si128( lowByte, bytes2 );
+ high = _mm_srli_epi16( high, 4 );
+ bytes2 = _mm_or_si128( low, high );
+
+ return _mm_packus_epi16( bytes1, bytes2);
+}
+
+static inline __m128i mul_add_epi8_sse(const __m128i x, const __m128i y) {
+ const __m128i ax = _mm_sign_epi8(x, x);
+ const __m128i sy = _mm_sign_epi8(y, x);
+ return _mm_maddubs_epi16(ax, sy);
+}
+
+// spread 32 bits to 32 bytes { 0x00, 0xFF }
+static inline __m256i bytes_from_bits_32(const uint8_t * x) {
+ uint32_t x32;
+ memcpy(&x32, x, sizeof(uint32_t));
+ const __m128i shuf_maskl = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
+ const __m128i shuf_maskh = _mm_set_epi64x(0x0303030303030303, 0x0202020202020202);
+ __m128i bytesl = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskl);
+ __m128i bytesh = _mm_shuffle_epi8(_mm_set1_epi32(x32), shuf_maskh);
+ const __m128i bit_mask = _mm_set1_epi64x(0x7fbfdfeff7fbfdfe);
+ bytesl = _mm_or_si128(bytesl, bit_mask);
+ bytesh = _mm_or_si128(bytesh, bit_mask);
+ bytesl = _mm_cmpeq_epi8(bytesl, _mm_set1_epi64x(-1));
+ bytesh = _mm_cmpeq_epi8(bytesh, _mm_set1_epi64x(-1));
+ return MM256_SET_M128I(bytesh, bytesl);
+}
+
+// Unpack 32 4-bit fields into 32 bytes
+// The output vector contains 32 bytes, each one in [ 0 .. 15 ] interval
+static inline __m256i bytes_from_nibbles_32(const uint8_t * rsi)
+{
+ // Load 16 bytes from memory
+ __m128i tmpl = _mm_loadu_si128((const __m128i *)rsi);
+ __m128i tmph = _mm_srli_epi16(tmpl, 4);
+ const __m128i lowMask = _mm_set1_epi8(0xF);
+ tmpl = _mm_and_si128(lowMask, tmpl);
+ tmph = _mm_and_si128(lowMask, tmph);
+ return MM256_SET_M128I(tmph, tmpl);
+}
+
+// add int16_t pairwise and return as float vector
+static inline __m256 sum_i16_pairs_float(const __m128i xh, const __m128i xl) {
+ const __m128i ones = _mm_set1_epi16(1);
+ const __m128i summed_pairsl = _mm_madd_epi16(ones, xl);
+ const __m128i summed_pairsh = _mm_madd_epi16(ones, xh);
+ const __m256i summed_pairs = MM256_SET_M128I(summed_pairsh, summed_pairsl);
+ return _mm256_cvtepi32_ps(summed_pairs);
+}
+
+static inline __m256 mul_sum_us8_pairs_float(const __m256i ax, const __m256i sy) {
+ const __m128i axl = _mm256_castsi256_si128(ax);
+ const __m128i axh = _mm256_extractf128_si256(ax, 1);
+ const __m128i syl = _mm256_castsi256_si128(sy);
+ const __m128i syh = _mm256_extractf128_si256(sy, 1);
+ // Perform multiplication and create 16-bit values
+ const __m128i dotl = _mm_maddubs_epi16(axl, syl);
+ const __m128i doth = _mm_maddubs_epi16(axh, syh);
+ return sum_i16_pairs_float(doth, dotl);
+}
+
+// multiply int8_t, add results pairwise twice and return as float vector
+static inline __m256 mul_sum_i8_pairs_float(const __m256i x, const __m256i y) {
+ const __m128i xl = _mm256_castsi256_si128(x);
+ const __m128i xh = _mm256_extractf128_si256(x, 1);
+ const __m128i yl = _mm256_castsi256_si128(y);
+ const __m128i yh = _mm256_extractf128_si256(y, 1);
+ // Get absolute values of x vectors
+ const __m128i axl = _mm_sign_epi8(xl, xl);
+ const __m128i axh = _mm_sign_epi8(xh, xh);
+ // Sign the values of the y vectors
+ const __m128i syl = _mm_sign_epi8(yl, xl);
+ const __m128i syh = _mm_sign_epi8(yh, xh);
+ // Perform multiplication and create 16-bit values
+ const __m128i dotl = _mm_maddubs_epi16(axl, syl);
+ const __m128i doth = _mm_maddubs_epi16(axh, syh);
+ return sum_i16_pairs_float(doth, dotl);
+}
+
+// larger version of mul_sum_i8_pairs_float where x and y are each represented by four 128-bit vectors
+static inline __m256 mul_sum_i8_quad_float(const __m128i x_1_0, const __m128i x_1_1, const __m128i x_2_0, const __m128i x_2_1,
+ const __m128i y_1_0, const __m128i y_1_1, const __m128i y_2_0, const __m128i y_2_1) {
+ const __m128i mone = _mm_set1_epi16(1);
+
+ const __m128i p16_1_0 = mul_add_epi8_sse(x_1_0, y_1_0);
+ const __m128i p16_1_1 = mul_add_epi8_sse(x_1_1, y_1_1);
+ const __m128i p16_2_0 = mul_add_epi8_sse(x_2_0, y_2_0);
+ const __m128i p16_2_1 = mul_add_epi8_sse(x_2_1, y_2_1);
+ const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, mone);
+ const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, mone);
+ const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, mone);
+ const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, mone);
+ const __m128i p_1 = _mm_add_epi32(p_1_0, p_1_1);
+ const __m128i p_2 = _mm_add_epi32(p_2_0, p_2_1);
+ return _mm256_cvtepi32_ps(MM256_SET_M128I(p_2, p_1));
+}
+
+// quad fp16 delta calculation
+static inline __m256 quad_fp16_delta_float(const float x0, const float y0, const float x1, const float y1) {
+ // GGML_CPU_FP16_TO_FP32 is faster than Intel F16C
+ return _mm256_set_m128(_mm_set1_ps(GGML_CPU_FP16_TO_FP32(x1) * GGML_CPU_FP16_TO_FP32(y1)),
+ _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x0) * GGML_CPU_FP16_TO_FP32(y0)));
+}
+
+static inline __m256 quad_mx_delta_float(const uint8_t x0, const float y0, const uint8_t x1, const float y1) {
+ return _mm256_set_m128(_mm_set1_ps(GGML_CPU_E8M0_TO_FP32_HALF(x1) * GGML_CPU_FP16_TO_FP32(y1)),
+ _mm_set1_ps(GGML_CPU_E8M0_TO_FP32_HALF(x0) * GGML_CPU_FP16_TO_FP32(y0)));
+}
+#endif
+#elif defined(__SSSE3__)
+static inline __m128i bytes_from_bits_16(const uint8_t * x) {
+ uint16_t x16;
+ memcpy(&x16, x, sizeof(uint16_t));
+
+ const __m128i shuf_mask = _mm_set_epi64x(0x0101010101010101, 0x0000000000000000);
+ __m128i bytes = _mm_shuffle_epi8(_mm_set1_epi16((short) x16), shuf_mask);
+ const __m128i bit_mask = _mm_set_epi64x(0x7fbfdfeff7fbfdfe, 0x7fbfdfeff7fbfdfe);
+ bytes = _mm_or_si128(bytes, bit_mask);
+
+ return _mm_cmpeq_epi8(bytes, _mm_set1_epi64x(-1));
+}
+
+// horizontally add 4x4 floats
+static inline float hsum_float_4x4(const __m128 a, const __m128 b, const __m128 c, const __m128 d) {
+ __m128 res_0 =_mm_hadd_ps(a, b);
+ __m128 res_1 =_mm_hadd_ps(c, d);
+ __m128 res =_mm_hadd_ps(res_0, res_1);
+ res =_mm_hadd_ps(res, res);
+ res =_mm_hadd_ps(res, res);
+
+ return _mm_cvtss_f32(res);
+}
+#endif // __AVX__ || __AVX2__ || __AVX512F__
+#endif // defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__SSSE3__)
+
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+ assert(QK8_0 == 32);
+ assert(k % QK8_0 == 0);
+ const int nb = k / QK8_0;
+
+ block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__AVX2__) || defined(__AVX__)
+ for (int i = 0; i < nb; i++) {
+ // Load elements into 4 AVX vectors
+ __m256 v0 = _mm256_loadu_ps( x );
+ __m256 v1 = _mm256_loadu_ps( x + 8 );
+ __m256 v2 = _mm256_loadu_ps( x + 16 );
+ __m256 v3 = _mm256_loadu_ps( x + 24 );
+ x += 32;
+
+ // Compute max(abs(e)) for the block
+ const __m256 signBit = _mm256_set1_ps( -0.0f );
+ __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
+ maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
+ maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
+ maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
+
+ __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
+ max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
+ max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
+ const float maxScalar = _mm_cvtss_f32( max4 );
+
+ // Quantize these floats
+ const float d = maxScalar / 127.f;
+ y[i].d = GGML_CPU_FP32_TO_FP16(d);
+ const float id = ( maxScalar != 0.0f ) ? 127.f / maxScalar : 0.0f;
+ const __m256 mul = _mm256_set1_ps( id );
+
+ // Apply the multiplier
+ v0 = _mm256_mul_ps( v0, mul );
+ v1 = _mm256_mul_ps( v1, mul );
+ v2 = _mm256_mul_ps( v2, mul );
+ v3 = _mm256_mul_ps( v3, mul );
+
+ // Round to nearest integer
+ v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
+ v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
+ v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
+ v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
+
+ // Convert floats to integers
+ __m256i i0 = _mm256_cvtps_epi32( v0 );
+ __m256i i1 = _mm256_cvtps_epi32( v1 );
+ __m256i i2 = _mm256_cvtps_epi32( v2 );
+ __m256i i3 = _mm256_cvtps_epi32( v3 );
+
+#if defined(__AVX2__)
+ // Convert int32 to int16
+ i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15
+ i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31
+ // Convert int16 to int8
+ i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
+
+ // We got our precious signed bytes, but the order is now wrong
+ // These AVX2 pack instructions process 16-byte pieces independently
+ // The following instruction is fixing the order
+ const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
+ i0 = _mm256_permutevar8x32_epi32( i0, perm );
+
+ _mm256_storeu_si256((__m256i *)y[i].qs, i0);
+#else
+ // Since we don't have in AVX some necessary functions,
+ // we split the registers in half and call AVX2 analogs from SSE
+ __m128i ni0 = _mm256_castsi256_si128( i0 );
+ __m128i ni1 = _mm256_extractf128_si256( i0, 1);
+ __m128i ni2 = _mm256_castsi256_si128( i1 );
+ __m128i ni3 = _mm256_extractf128_si256( i1, 1);
+ __m128i ni4 = _mm256_castsi256_si128( i2 );
+ __m128i ni5 = _mm256_extractf128_si256( i2, 1);
+ __m128i ni6 = _mm256_castsi256_si128( i3 );
+ __m128i ni7 = _mm256_extractf128_si256( i3, 1);
+
+ // Convert int32 to int16
+ ni0 = _mm_packs_epi32( ni0, ni1 );
+ ni2 = _mm_packs_epi32( ni2, ni3 );
+ ni4 = _mm_packs_epi32( ni4, ni5 );
+ ni6 = _mm_packs_epi32( ni6, ni7 );
+ // Convert int16 to int8
+ ni0 = _mm_packs_epi16( ni0, ni2 );
+ ni4 = _mm_packs_epi16( ni4, ni6 );
+
+ _mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0);
+ _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
+#endif
+ }
+#else
+ GGML_UNUSED(nb);
+ // scalar
+ quantize_row_q8_0_ref(x, y, k);
+#endif
+}
+
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+ assert(k % QK8_1 == 0);
+ const int nb = k / QK8_1;
+
+ block_q8_1 * GGML_RESTRICT y = vy;
+#if defined(__AVX2__) || defined(__AVX__)
+ for (int i = 0; i < nb; i++) {
+ // Load elements into 4 AVX vectors
+ __m256 v0 = _mm256_loadu_ps( x );
+ __m256 v1 = _mm256_loadu_ps( x + 8 );
+ __m256 v2 = _mm256_loadu_ps( x + 16 );
+ __m256 v3 = _mm256_loadu_ps( x + 24 );
+ x += 32;
+
+ // Compute max(abs(e)) for the block
+ const __m256 signBit = _mm256_set1_ps( -0.0f );
+ __m256 maxAbs = _mm256_andnot_ps( signBit, v0 );
+ maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v1 ) );
+ maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v2 ) );
+ maxAbs = _mm256_max_ps( maxAbs, _mm256_andnot_ps( signBit, v3 ) );
+
+ __m128 max4 = _mm_max_ps( _mm256_extractf128_ps( maxAbs, 1 ), _mm256_castps256_ps128( maxAbs ) );
+ max4 = _mm_max_ps( max4, _mm_movehl_ps( max4, max4 ) );
+ max4 = _mm_max_ss( max4, _mm_movehdup_ps( max4 ) );
+ const float max_scalar = _mm_cvtss_f32( max4 );
+
+ // Quantize these floats
+ const float d = max_scalar / 127.f;
+ y[i].d = GGML_CPU_FP32_TO_FP16(d);
+ const float id = ( max_scalar != 0.0f ) ? 127.f / max_scalar : 0.0f;
+ const __m256 mul = _mm256_set1_ps( id );
+
+ // Apply the multiplier
+ v0 = _mm256_mul_ps( v0, mul );
+ v1 = _mm256_mul_ps( v1, mul );
+ v2 = _mm256_mul_ps( v2, mul );
+ v3 = _mm256_mul_ps( v3, mul );
+
+ // Round to nearest integer
+ v0 = _mm256_round_ps( v0, _MM_ROUND_NEAREST );
+ v1 = _mm256_round_ps( v1, _MM_ROUND_NEAREST );
+ v2 = _mm256_round_ps( v2, _MM_ROUND_NEAREST );
+ v3 = _mm256_round_ps( v3, _MM_ROUND_NEAREST );
+
+ // Convert floats to integers
+ __m256i i0 = _mm256_cvtps_epi32( v0 );
+ __m256i i1 = _mm256_cvtps_epi32( v1 );
+ __m256i i2 = _mm256_cvtps_epi32( v2 );
+ __m256i i3 = _mm256_cvtps_epi32( v3 );
+
+#if defined(__AVX2__)
+ // Compute the sum of the quants and set y[i].s
+ y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_8(_mm256_add_epi32(_mm256_add_epi32(i0, i1), _mm256_add_epi32(i2, i3))));
+
+ // Convert int32 to int16
+ i0 = _mm256_packs_epi32( i0, i1 ); // 0, 1, 2, 3, 8, 9, 10, 11, 4, 5, 6, 7, 12, 13, 14, 15
+ i2 = _mm256_packs_epi32( i2, i3 ); // 16, 17, 18, 19, 24, 25, 26, 27, 20, 21, 22, 23, 28, 29, 30, 31
+ // Convert int16 to int8
+ i0 = _mm256_packs_epi16( i0, i2 ); // 0, 1, 2, 3, 8, 9, 10, 11, 16, 17, 18, 19, 24, 25, 26, 27, 4, 5, 6, 7, 12, 13, 14, 15, 20, 21, 22, 23, 28, 29, 30, 31
+
+ // We got our precious signed bytes, but the order is now wrong
+ // These AVX2 pack instructions process 16-byte pieces independently
+ // The following instruction is fixing the order
+ const __m256i perm = _mm256_setr_epi32( 0, 4, 1, 5, 2, 6, 3, 7 );
+ i0 = _mm256_permutevar8x32_epi32( i0, perm );
+
+ _mm256_storeu_si256((__m256i *)y[i].qs, i0);
+#else
+ // Since we don't have in AVX some necessary functions,
+ // we split the registers in half and call AVX2 analogs from SSE
+ __m128i ni0 = _mm256_castsi256_si128( i0 );
+ __m128i ni1 = _mm256_extractf128_si256( i0, 1);
+ __m128i ni2 = _mm256_castsi256_si128( i1 );
+ __m128i ni3 = _mm256_extractf128_si256( i1, 1);
+ __m128i ni4 = _mm256_castsi256_si128( i2 );
+ __m128i ni5 = _mm256_extractf128_si256( i2, 1);
+ __m128i ni6 = _mm256_castsi256_si128( i3 );
+ __m128i ni7 = _mm256_extractf128_si256( i3, 1);
+
+ // Compute the sum of the quants and set y[i].s
+ const __m128i s0 = _mm_add_epi32(_mm_add_epi32(ni0, ni1), _mm_add_epi32(ni2, ni3));
+ const __m128i s1 = _mm_add_epi32(_mm_add_epi32(ni4, ni5), _mm_add_epi32(ni6, ni7));
+ y[i].s = GGML_CPU_FP32_TO_FP16(d * hsum_i32_4(_mm_add_epi32(s0, s1)));
+
+ // Convert int32 to int16
+ ni0 = _mm_packs_epi32( ni0, ni1 );
+ ni2 = _mm_packs_epi32( ni2, ni3 );
+ ni4 = _mm_packs_epi32( ni4, ni5 );
+ ni6 = _mm_packs_epi32( ni6, ni7 );
+ // Convert int16 to int8
+ ni0 = _mm_packs_epi16( ni0, ni2 );
+ ni4 = _mm_packs_epi16( ni4, ni6 );
+
+ _mm_storeu_si128((__m128i *)(y[i].qs + 0), ni0);
+ _mm_storeu_si128((__m128i *)(y[i].qs + 16), ni4);
+#endif
+ }
+#else
+ GGML_UNUSED(nb);
+ // scalar
+ quantize_row_q8_1_ref(x, y, k);
+#endif
+}
+
+// placeholder implementation for Apple targets
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ quantize_row_q8_K_ref(x, y, k);
+}
+
+//===================================== Dot products =================================
+
+//
+// Helper functions
+//
+
+#if __AVX__ || __AVX2__ || __AVX512F__
+
+// shuffles to pick the required scales in dot products
+static inline __m256i get_scale_shuffle_q3k(int i) {
+ static const uint8_t k_shuffle[128] = {
+ 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+ 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+ 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+ 12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13, 14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,
+ };
+ return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
+}
+static inline __m256i get_scale_shuffle_k4(int i) {
+ static const uint8_t k_shuffle[256] = {
+ 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1,
+ 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3, 2, 3,
+ 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5, 4, 5,
+ 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7, 6, 7,
+ 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9, 8, 9,
+ 10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,10,11,
+ 12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,12,13,
+ 14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15,14,15
+ };
+ return _mm256_loadu_si256((const __m256i*)k_shuffle + i);
+}
+static inline __m128i get_scale_shuffle(int i) {
+ static const uint8_t k_shuffle[128] = {
+ 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+ 2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3,
+ 4, 4, 4, 4, 4, 4, 4, 4, 5, 5, 5, 5, 5, 5, 5, 5,
+ 6, 6, 6, 6, 6, 6, 6, 6, 7, 7, 7, 7, 7, 7, 7, 7,
+ 8, 8, 8, 8, 8, 8, 8, 8, 9, 9, 9, 9, 9, 9, 9, 9,
+ 10,10,10,10,10,10,10,10, 11,11,11,11,11,11,11,11,
+ 12,12,12,12,12,12,12,12, 13,13,13,13,13,13,13,13,
+ 14,14,14,14,14,14,14,14, 15,15,15,15,15,15,15,15
+ };
+ return _mm_loadu_si128((const __m128i*)k_shuffle + i);
+}
+#endif
+
+void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK1_0;
+ const int nb = n / qk;
+
+ assert(n % qk == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q1_0 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__AVX2__)
+ const __m256i ones_8 = _mm256_set1_epi8(1);
+ const __m256i ones_16 = _mm256_set1_epi16(1);
+ const __m256i byte_shuf = _mm256_setr_epi8(
+ 0, 0, 0, 0, 0, 0, 0, 0, 1, 1, 1, 1, 1, 1, 1, 1,
+ 2, 2, 2, 2, 2, 2, 2, 2, 3, 3, 3, 3, 3, 3, 3, 3);
+ const __m256i bit_masks = _mm256_setr_epi8(
+ 1, 2, 4, 8, 16, 32, 64, (char) -128, 1, 2, 4, 8, 16, 32, 64, (char) -128,
+ 1, 2, 4, 8, 16, 32, 64, (char) -128, 1, 2, 4, 8, 16, 32, 64, (char) -128);
+ const __m256i zero = _mm256_setzero_si256();
+ __m256 acc = _mm256_setzero_ps();
+
+ for (int ib = 0; ib < nb; ++ib) {
+ const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+ const uint32_t * GGML_RESTRICT qs32 = (const uint32_t *) x[ib].qs;
+ const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+
+ __m256 acc_block;
+ {
+ const __m256i qy = _mm256_loadu_si256((const __m256i *) y_ptr[0].qs);
+ const __m256i sm = _mm256_cmpeq_epi8(
+ _mm256_and_si256(_mm256_shuffle_epi8(_mm256_set1_epi32((int) qs32[0]), byte_shuf), bit_masks), zero);
+ const __m256i sy = _mm256_sub_epi8(_mm256_xor_si256(qy, sm), sm);
+ const __m256i s32 = _mm256_madd_epi16(_mm256_maddubs_epi16(ones_8, sy), ones_16);
+ acc_block = _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[0].d)), _mm256_cvtepi32_ps(s32));
+ }
+ for (int K = 1; K < 4; ++K) {
+ const __m256i qy = _mm256_loadu_si256((const __m256i *) y_ptr[K].qs);
+ const __m256i sm = _mm256_cmpeq_epi8(
+ _mm256_and_si256(_mm256_shuffle_epi8(_mm256_set1_epi32((int) qs32[K]), byte_shuf), bit_masks), zero);
+ const __m256i sy = _mm256_sub_epi8(_mm256_xor_si256(qy, sm), sm);
+ const __m256i s32 = _mm256_madd_epi16(_mm256_maddubs_epi16(ones_8, sy), ones_16);
+ acc_block = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[K].d)), _mm256_cvtepi32_ps(s32), acc_block);
+ }
+ acc = _mm256_fmadd_ps(_mm256_set1_ps(d0), acc_block, acc);
+ }
+
+ *s = hsum_float_8(acc);
+#elif defined(__AVX__)
+ const __m128i ones_8 = _mm_set1_epi8(1);
+ const __m128i ones_16 = _mm_set1_epi16(1);
+ const __m128i zero = _mm_setzero_si128();
+ __m256 acc = _mm256_setzero_ps();
+
+ for (int ib = 0; ib < nb; ++ib) {
+ const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+ const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+ __m256 acc_block;
+ {
+ const __m256i bit_mask = bytes_from_bits_32(&x[ib].qs[0]);
+ const __m128i bit_mask_0 = _mm256_castsi256_si128(bit_mask);
+ const __m128i bit_mask_1 = _mm256_extractf128_si256(bit_mask, 1);
+ const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[0].qs[0]);
+ const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[0].qs[16]);
+ const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero);
+ const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero);
+ const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0);
+ const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1);
+ const __m128i sum16_0 = _mm_maddubs_epi16(ones_8, sy_0);
+ const __m128i sum16_1 = _mm_maddubs_epi16(ones_8, sy_1);
+ const __m128i sum32_0 = _mm_madd_epi16(sum16_0, ones_16);
+ const __m128i sum32_1 = _mm_madd_epi16(sum16_1, ones_16);
+ const __m256 q = _mm256_cvtepi32_ps(MM256_SET_M128I(sum32_1, sum32_0));
+ acc_block = _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[0].d)), q);
+ }
+ for(int K = 1; K < 4; ++K) {
+ const __m256i bit_mask = bytes_from_bits_32(&x[ib].qs[(K) * 4]);
+ const __m128i bit_mask_0 = _mm256_castsi256_si128(bit_mask);
+ const __m128i bit_mask_1 = _mm256_extractf128_si256(bit_mask, 1);
+ const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[(K)].qs[0]);
+ const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[(K)].qs[16]);
+ const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero);
+ const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero);
+ const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0);
+ const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1);
+ const __m128i sum16_0 = _mm_maddubs_epi16(ones_8, sy_0);
+ const __m128i sum16_1 = _mm_maddubs_epi16(ones_8, sy_1);
+ const __m128i sum32_0 = _mm_madd_epi16(sum16_0, ones_16);
+ const __m128i sum32_1 = _mm_madd_epi16(sum16_1, ones_16);
+ const __m256 q = _mm256_cvtepi32_ps(MM256_SET_M128I(sum32_1, sum32_0));
+ acc_block = _mm256_add_ps(acc_block, _mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[(K)].d)), q));
+ }
+#undef Q1_AVX_BLOCK
+
+ acc = _mm256_add_ps(acc, _mm256_mul_ps(_mm256_set1_ps(d0), acc_block));
+ }
+
+ *s = hsum_float_8(acc);
+#elif defined(__SSSE3__)
+ const __m128i ones_8 = _mm_set1_epi8(1);
+ const __m128i ones_16 = _mm_set1_epi16(1);
+ const __m128i zero = _mm_setzero_si128();
+ __m128 acc_0 = _mm_setzero_ps();
+ __m128 acc_1 = _mm_setzero_ps();
+ __m128 acc_2 = _mm_setzero_ps();
+ __m128 acc_3 = _mm_setzero_ps();
+
+ for (int ib = 0; ib < nb; ++ib) {
+ const __m128 d0 = _mm_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
+ const block_q8_0 * GGML_RESTRICT y_ptr = &y[ib * 4];
+
+#define Q1_SSSE3_BLOCK(QS_OFF, Y_IDX, ACC) \
+ { \
+ const __m128i bit_mask_0 = bytes_from_bits_16(&x[ib].qs[(QS_OFF) + 0]); \
+ const __m128i bit_mask_1 = bytes_from_bits_16(&x[ib].qs[(QS_OFF) + 2]); \
+ const __m128i qy_0 = _mm_loadu_si128((const __m128i *) &y_ptr[(Y_IDX)].qs[0]); \
+ const __m128i qy_1 = _mm_loadu_si128((const __m128i *) &y_ptr[(Y_IDX)].qs[16]); \
+ const __m128i sign_mask_0 = _mm_cmpeq_epi8(bit_mask_0, zero); \
+ const __m128i sign_mask_1 = _mm_cmpeq_epi8(bit_mask_1, zero); \
+ const __m128i sy_0 = _mm_sub_epi8(_mm_xor_si128(qy_0, sign_mask_0), sign_mask_0); \
+ const __m128i sy_1 = _mm_sub_epi8(_mm_xor_si128(qy_1, sign_mask_1), sign_mask_1); \
+ const __m128i sum_0 = _mm_madd_epi16(_mm_maddubs_epi16(ones_8, sy_0), ones_16); \
+ const __m128i sum_1 = _mm_madd_epi16(_mm_maddubs_epi16(ones_8, sy_1), ones_16); \
+ const __m128 q = _mm_cvtepi32_ps(_mm_add_epi32(sum_0, sum_1)); \
+ (ACC) = _mm_add_ps((ACC), _mm_mul_ps(_mm_mul_ps(d0, _mm_set1_ps(GGML_CPU_FP16_TO_FP32(y_ptr[(Y_IDX)].d))), q)); \
+ }
+ Q1_SSSE3_BLOCK(0, 0, acc_0)
+ Q1_SSSE3_BLOCK(4, 1, acc_1)
+ Q1_SSSE3_BLOCK(8, 2, acc_2)
+ Q1_SSSE3_BLOCK(12, 3, acc_3)
+#undef Q1_SSSE3_BLOCK
+ }
+
+ *s = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+#else
+ UNUSED(nb);
+ UNUSED(x);
+ UNUSED(y);
+ ggml_vec_dot_q1_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+#if defined(__AVX2__)
+// four consecutive Q2_0 weights per byte: byte b becomes b & 3, (b >> 2) & 3, (b >> 4) & 3, (b >> 6) & 3 (OR, never add: the overlapping copies only touch masked-off bits)
+static inline __m256i q2_0_unpack_32_avx2(const uint8_t * qs) {
+ __m256i v = _mm256_cvtepu8_epi32(_mm_loadl_epi64((const __m128i *) qs));
+ v = _mm256_or_si256(v, _mm256_slli_epi32(v, 12));
+ v = _mm256_or_si256(v, _mm256_slli_epi32(v, 6));
+ return _mm256_and_si256(v, _mm256_set1_epi32(0x03030303));
+}
+
+// 32 weights against one Q8_0 block: sum((v - 1) * y) = sum(v * y) - sum(y), exact for every int8 y
+static inline __m256 q2_0_dot_32_avx2(const uint8_t * qs, const int8_t * qy) {
+ const __m256i v = q2_0_unpack_32_avx2(qs);
+ const __m256i y = _mm256_loadu_si256((const __m256i *) qy);
+ const __m256i p = _mm256_sub_epi16(_mm256_maddubs_epi16(v, y), _mm256_maddubs_epi16(_mm256_set1_epi8(1), y));
+ return _mm256_cvtepi32_ps(_mm256_madd_epi16(p, _mm256_set1_epi16(1)));
+}
+
+static inline __m256 q2_0_acc_block_avx2(const block_q2_0 * x, const block_q8_0 * y, __m256 acc) {
+ const float dx = GGML_CPU_FP16_TO_FP32(x->d);
+ const __m256 d0 = _mm256_set1_ps(dx * GGML_CPU_FP16_TO_FP32(y[0].d));
+ const __m256 d1 = _mm256_set1_ps(dx * GGML_CPU_FP16_TO_FP32(y[1].d));
+ const __m256 p0 = q2_0_dot_32_avx2(x->qs + 0, y[0].qs);
+ const __m256 p1 = q2_0_dot_32_avx2(x->qs + 8, y[1].qs);
+#if defined(__FMA__)
+ acc = _mm256_fmadd_ps(d0, p0, acc);
+ return _mm256_fmadd_ps(d1, p1, acc);
+#else
+ acc = _mm256_add_ps(acc, _mm256_mul_ps(d0, p0));
+ return _mm256_add_ps(acc, _mm256_mul_ps(d1, p1));
+#endif
+}
+#endif
+
+void ggml_vec_dot_q2_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+#if defined(__AVX2__)
+ assert(n % QK2_0 == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q2_0 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+ const int nb = n / QK2_0;
+
+ // two Q2_0 blocks per step, each into its own accumulator (two FMA chains)
+ __m256 acc0 = _mm256_setzero_ps();
+ __m256 acc1 = _mm256_setzero_ps();
+ int i = 0;
+ for (; i + 1 < nb; i += 2) {
+ acc0 = q2_0_acc_block_avx2(x + i + 0, y + 2*i + 0, acc0);
+ acc1 = q2_0_acc_block_avx2(x + i + 1, y + 2*i + 2, acc1);
+ }
+ if (i < nb) {
+ acc0 = q2_0_acc_block_avx2(x + i, y + 2*i, acc0);
+ }
+
+ *s = hsum_float_8(_mm256_add_ps(acc0, acc1));
+#else
+ ggml_vec_dot_q2_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK8_0;
+ const int nb = n / qk;
+
+ assert(n % qk == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q4_0 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ int ib = 0;
+ float sumf = 0;
+
+#if defined(__AVX2__)
+ // Initialize accumulator with zeros
+ __m256 acc = _mm256_setzero_ps();
+
+ // Main loop
+ for (; ib < nb; ++ib) {
+ /* Compute combined scale for the block */
+ const __m256 d = _mm256_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
+
+ __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+
+ // Now we have a vector with bytes in [ 0 .. 15 ] interval. Offset them into [ -8 .. +7 ] interval.
+ const __m256i off = _mm256_set1_epi8( 8 );
+ qx = _mm256_sub_epi8( qx, off );
+
+ __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+ const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+ /* Multiply q with scale and accumulate */
+ acc = _mm256_fmadd_ps( d, q, acc );
+ }
+
+ sumf = hsum_float_8(acc);
+#elif defined(__AVX__)
+ __m256 accum = _mm256_setzero_ps();
+ for (; ib + 1 < nb; ib += 2) {
+ const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+ const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+ const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+ const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+ const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+ const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+ const __m128i q4b_1_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_1), _mm_set1_epi8(8));
+ const __m128i q4b_1_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_1, 4)), _mm_set1_epi8(8));
+ const __m128i q4b_2_0 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), q4bits_2), _mm_set1_epi8(8));
+ const __m128i q4b_2_1 = _mm_sub_epi8(_mm_and_si128(_mm_set1_epi8(15), _mm_srli_epi16(q4bits_2, 4)), _mm_set1_epi8(8));
+
+ const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+ const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+ const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+ const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+ const __m128i p_1 = _mm_add_epi16(p16_1_0, p16_1_1);
+ const __m128i p_2 = _mm_add_epi16(p16_2_0, p16_2_1);
+ const __m256 p = sum_i16_pairs_float(p_2, p_1);
+
+ const __m256 deltas = quad_fp16_delta_float(x[ib].d, y[ib].d, x[ib + 1].d, y[ib + 1].d);
+ accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
+ }
+
+ sumf = hsum_float_8(accum);
+#elif defined(__SSSE3__)
+ // set constants
+ const __m128i lowMask = _mm_set1_epi8(0xF);
+ const __m128i off = _mm_set1_epi8(8);
+
+ // Initialize accumulator with zeros
+ __m128 acc_0 = _mm_setzero_ps();
+ __m128 acc_1 = _mm_setzero_ps();
+ __m128 acc_2 = _mm_setzero_ps();
+ __m128 acc_3 = _mm_setzero_ps();
+
+ for (; ib + 1 < nb; ib += 2) {
+ _mm_prefetch(&x[ib] + sizeof(block_q4_0), _MM_HINT_T0);
+ _mm_prefetch(&y[ib] + sizeof(block_q8_0), _MM_HINT_T0);
+
+ // Compute combined scale for the block 0 and 1
+ const __m128 d_0_1 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d) );
+
+ const __m128i tmp_0_1 = _mm_loadu_si128((const __m128i *)x[ib].qs);
+
+ __m128i bx_0 = _mm_and_si128(lowMask, tmp_0_1);
+ __m128i by_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
+ bx_0 = _mm_sub_epi8(bx_0, off);
+ const __m128i i32_0 = mul_sum_i8_pairs(bx_0, by_0);
+
+ __m128i bx_1 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_0_1, 4));
+ __m128i by_1 = _mm_loadu_si128((const __m128i *)(y[ib].qs + 16));
+ bx_1 = _mm_sub_epi8(bx_1, off);
+ const __m128i i32_1 = mul_sum_i8_pairs(bx_1, by_1);
+
+ _mm_prefetch(&x[ib] + 2 * sizeof(block_q4_0), _MM_HINT_T0);
+ _mm_prefetch(&y[ib] + 2 * sizeof(block_q8_0), _MM_HINT_T0);
+
+ // Compute combined scale for the block 2 and 3
+ const __m128 d_2_3 = _mm_set1_ps( GGML_CPU_FP16_TO_FP32(x[ib + 1].d) * GGML_CPU_FP16_TO_FP32(y[ib + 1].d) );
+
+ const __m128i tmp_2_3 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+
+ __m128i bx_2 = _mm_and_si128(lowMask, tmp_2_3);
+ __m128i by_2 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+ bx_2 = _mm_sub_epi8(bx_2, off);
+ const __m128i i32_2 = mul_sum_i8_pairs(bx_2, by_2);
+
+ __m128i bx_3 = _mm_and_si128(lowMask, _mm_srli_epi64(tmp_2_3, 4));
+ __m128i by_3 = _mm_loadu_si128((const __m128i *)(y[ib + 1].qs + 16));
+ bx_3 = _mm_sub_epi8(bx_3, off);
+ const __m128i i32_3 = mul_sum_i8_pairs(bx_3, by_3);
+
+ // Convert int32_t to float
+ __m128 p0 = _mm_cvtepi32_ps(i32_0);
+ __m128 p1 = _mm_cvtepi32_ps(i32_1);
+ __m128 p2 = _mm_cvtepi32_ps(i32_2);
+ __m128 p3 = _mm_cvtepi32_ps(i32_3);
+
+ // Apply the scale
+ __m128 p0_d = _mm_mul_ps( d_0_1, p0 );
+ __m128 p1_d = _mm_mul_ps( d_0_1, p1 );
+ __m128 p2_d = _mm_mul_ps( d_2_3, p2 );
+ __m128 p3_d = _mm_mul_ps( d_2_3, p3 );
+
+ // Acummulate
+ acc_0 = _mm_add_ps(p0_d, acc_0);
+ acc_1 = _mm_add_ps(p1_d, acc_1);
+ acc_2 = _mm_add_ps(p2_d, acc_2);
+ acc_3 = _mm_add_ps(p3_d, acc_3);
+ }
+
+ sumf = hsum_float_4x4(acc_0, acc_1, acc_2, acc_3);
+
+#endif
+ for (; ib < nb; ++ib) {
+ int sumi0 = 0;
+ int sumi1 = 0;
+
+ for (int j = 0; j < qk/2; ++j) {
+ const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+ const int v1 = (x[ib].qs[j] >> 4) - 8;
+
+ sumi0 += (v0 * y[ib].qs[j]);
+ sumi1 += (v1 * y[ib].qs[j + qk/2]);
+ }
+
+ int sumi = sumi0 + sumi1;
+ sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
+ }
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK8_1;
+ const int nb = n / qk;
+
+ assert(n % qk == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q4_1 * GGML_RESTRICT x = vx;
+ const block_q8_1 * GGML_RESTRICT y = vy;
+
+ int ib = 0;
+
+#if defined(__AVX2__) || defined(__AVX__)
+ // Initialize accumulator with zeros
+ __m256 acc = _mm256_setzero_ps();
+
+ float summs = 0;
+
+ // Main loop
+ for (; ib < nb; ++ib) {
+ const float d0 = GGML_CPU_FP16_TO_FP32(x[ib].d);
+ const float d1 = GGML_CPU_FP16_TO_FP32(y[ib].d);
+
+ summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
+
+ const __m256 d0v = _mm256_set1_ps( d0 );
+ const __m256 d1v = _mm256_set1_ps( d1 );
+
+ // Compute combined scales
+ const __m256 d0d1 = _mm256_mul_ps( d0v, d1v );
+
+ // Load 16 bytes, and unpack 4 bit fields into bytes, making 32 bytes
+ const __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+ const __m256i qy = _mm256_loadu_si256( (const __m256i *)y[ib].qs );
+
+ const __m256 xy = mul_sum_us8_pairs_float(qx, qy);
+
+ // Accumulate d0*d1*x*y
+#if defined(__AVX2__)
+ acc = _mm256_fmadd_ps( d0d1, xy, acc );
+#else
+ acc = _mm256_add_ps( _mm256_mul_ps( d0d1, xy ), acc );
+#endif
+ }
+
+ *s = hsum_float_8(acc) + summs;
+#else
+ UNUSED(nb);
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(ib);
+ ggml_vec_dot_q4_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+ assert(n % QK_MXFP4 == 0);
+ static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
+
+ const block_mxfp4 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_MXFP4;
+
+ int ib = 0;
+ float sumf = 0;
+
+#if defined __AVX2__
+
+ const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_fp4);
+ const __m128i m4b = _mm_set1_epi8(0x0f);
+ const __m256i mone = _mm256_set1_epi16(1);
+
+ __m256 accum1 = _mm256_setzero_ps();
+ __m256 accum2 = _mm256_setzero_ps();
+
+ for (; ib + 1 < nb; ib += 2) {
+ const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
+ const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
+ const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[ib + 0].qs);
+ const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[ib + 1].qs);
+ const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
+ _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
+ const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
+ _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
+ const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+ const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+ const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
+ const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
+ const __m256 scale0 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib + 0].e));
+ const __m256 scale1 = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib + 1].e));
+ accum1 = _mm256_fmadd_ps(scale0, _mm256_cvtepi32_ps(p_1), accum1);
+ accum2 = _mm256_fmadd_ps(scale1, _mm256_cvtepi32_ps(p_2), accum2);
+ }
+
+ sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
+
+#elif defined __AVX__
+ const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_fp4);
+ const __m128i m4b = _mm_set1_epi8(0x0f);
+
+ __m256 accum = _mm256_setzero_ps();
+ for (; ib + 1 < nb; ib += 2) {
+ const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+ const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+ const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+ const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+ const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+ const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+ const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+ const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+ const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+ const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+
+ const __m256 p = mul_sum_i8_quad_float(q4b_1_0, q4b_1_1, q4b_2_0, q4b_2_1, q8b_1_0, q8b_1_1, q8b_2_0, q8b_2_1);
+ const __m256 deltas = quad_mx_delta_float(x[ib].e, y[ib].d, x[ib + 1].e, y[ib + 1].d);
+ accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
+ }
+
+ sumf = hsum_float_8(accum);
+
+#endif
+ for (; ib < nb; ++ib) {
+ const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_E8M0_TO_FP32_HALF(x[ib].e);
+ int sumi1 = 0;
+ int sumi2 = 0;
+ for (int j = 0; j < QK_MXFP4/2; ++j) {
+ sumi1 += y[ib].qs[j + 0] * kvalues_fp4[x[ib].qs[j] & 0xf];
+ sumi2 += y[ib].qs[j + QK_MXFP4/2] * kvalues_fp4[x[ib].qs[j] >> 4];
+ }
+ sumf += d * (sumi1 + sumi2);
+ }
+ *s = sumf;
+}
+
+void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+ assert(n % QK_NVFP4 == 0);
+
+ const block_nvfp4 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_NVFP4;
+ int ib = 0;
+ float sumf = 0;
+
+#if defined(__AVX2__)
+
+ const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_fp4);
+ const __m128i m4b = _mm_set1_epi8(0x0f);
+ const __m256i mone = _mm256_set1_epi16(1);
+
+ __m256 accum = _mm256_setzero_ps();
+ for(; ib < nb; ib++){
+
+ const __m128i q4bits_01 = _mm_loadu_si128((const __m128i *)(x[ib].qs + 0));
+ const __m128i q4bits_23 = _mm_loadu_si128((const __m128i *)(x[ib].qs + 16));
+
+ const __m256i q8_01 = _mm256_loadu_si256((const __m256i *)y[2*ib + 0].qs);
+ const __m256i q8_23 = _mm256_loadu_si256((const __m256i *)y[2*ib + 1].qs);
+
+ const __m128i q4_01_lo = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_01, m4b));
+ const __m128i q4_01_hi = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_01, 4), m4b));
+ const __m128i q4_23_lo = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_23, m4b));
+ const __m128i q4_23_hi = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_23, 4), m4b));
+
+ //reordering
+ const __m256i q4_01 = MM256_SET_M128I(_mm_unpackhi_epi64(q4_01_lo,q4_01_hi), _mm_unpacklo_epi64(q4_01_lo,q4_01_hi));
+ const __m256i q4_23 = MM256_SET_M128I(_mm_unpackhi_epi64(q4_23_lo,q4_23_hi),_mm_unpacklo_epi64(q4_23_lo,q4_23_hi));
+
+ const __m256i p01 = mul_add_epi8(q4_01,q8_01);
+ const __m256i p_1 = _mm256_madd_epi16(p01, mone);
+
+ const __m256i p23 = mul_add_epi8(q4_23,q8_23);
+ const __m256i p_2 = _mm256_madd_epi16(p23, mone);
+
+ const float dy0 = GGML_CPU_FP16_TO_FP32(y[2*ib].d);
+ const float dy1 = GGML_CPU_FP16_TO_FP32(y[2*ib+1].d);
+
+ const float s0 = GGML_CPU_UE4M3_TO_FP32(x[ib].d[0]) * dy0;
+ const float s1 = GGML_CPU_UE4M3_TO_FP32(x[ib].d[1]) * dy0;
+ const float s2 = GGML_CPU_UE4M3_TO_FP32(x[ib].d[2]) * dy1;
+ const float s3 = GGML_CPU_UE4M3_TO_FP32(x[ib].d[3]) * dy1;
+
+ const __m256 scales01 = _mm256_set_m128(_mm_set1_ps(s1), _mm_set1_ps(s0));
+ const __m256 scales23 = _mm256_set_m128(_mm_set1_ps(s3), _mm_set1_ps(s2));
+
+ accum = _mm256_fmadd_ps(scales01, _mm256_cvtepi32_ps(p_1), accum);
+ accum = _mm256_fmadd_ps(scales23, _mm256_cvtepi32_ps(p_2), accum);
+ }
+ sumf = hsum_float_8(accum);
+
+#elif defined(__AVX__)
+
+ const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_fp4);
+ const __m128i m4b = _mm_set1_epi8(0x0f);
+
+ __m256 accum = _mm256_setzero_ps();
+ for(; ib < nb; ib++){
+
+ const __m128i q4bits_01 = _mm_loadu_si128((const __m128i *)(x[ib].qs + 0));
+ const __m128i q4bits_23 = _mm_loadu_si128((const __m128i *)(x[ib].qs + 16));
+
+ const __m128i q8_0 = _mm_loadu_si128((const __m128i *)(y[2*ib + 0].qs + 0));
+ const __m128i q8_1 = _mm_loadu_si128((const __m128i *)(y[2*ib + 0].qs + 16));
+ const __m128i q8_2 = _mm_loadu_si128((const __m128i *)(y[2*ib + 1].qs + 0));
+ const __m128i q8_3 = _mm_loadu_si128((const __m128i *)(y[2*ib + 1].qs + 16));
+
+ const __m128i q4_01_lo = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_01, m4b));
+ const __m128i q4_01_hi = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_01, 4), m4b));
+ const __m128i q4_23_lo = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_23, m4b));
+ const __m128i q4_23_hi = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_23, 4), m4b));
+
+ const __m128i q4_0 = _mm_unpacklo_epi64(q4_01_lo, q4_01_hi);
+ const __m128i q4_1 = _mm_unpackhi_epi64(q4_01_lo, q4_01_hi);
+ const __m128i q4_2 = _mm_unpacklo_epi64(q4_23_lo, q4_23_hi);
+ const __m128i q4_3 = _mm_unpackhi_epi64(q4_23_lo, q4_23_hi);
+
+ const __m128i p0_i32 = mul_sum_i8_pairs(q4_0, q8_0);
+ const __m128i p1_i32 = mul_sum_i8_pairs(q4_1, q8_1);
+ const __m128i p2_i32 = mul_sum_i8_pairs(q4_2, q8_2);
+ const __m128i p3_i32 = mul_sum_i8_pairs(q4_3, q8_3);
+
+ const __m128 p0 = _mm_cvtepi32_ps(p0_i32);
+ const __m128 p1 = _mm_cvtepi32_ps(p1_i32);
+ const __m128 p2 = _mm_cvtepi32_ps(p2_i32);
+ const __m128 p3 = _mm_cvtepi32_ps(p3_i32);
+
+ const __m256 p01 = _mm256_set_m128(p1, p0);
+ const __m256 p23 = _mm256_set_m128(p3, p2);
+
+ const float dy0 = GGML_CPU_FP16_TO_FP32(y[2*ib].d);
+ const float dy1 = GGML_CPU_FP16_TO_FP32(y[2*ib+1].d);
+
+ const float s0 = GGML_CPU_UE4M3_TO_FP32(x[ib].d[0]) * dy0;
+ const float s1 = GGML_CPU_UE4M3_TO_FP32(x[ib].d[1]) * dy0;
+ const float s2 = GGML_CPU_UE4M3_TO_FP32(x[ib].d[2]) * dy1;
+ const float s3 = GGML_CPU_UE4M3_TO_FP32(x[ib].d[3]) * dy1;
+
+ const __m256 scales01 = _mm256_set_m128(_mm_set1_ps(s1), _mm_set1_ps(s0));
+ const __m256 scales23 = _mm256_set_m128(_mm_set1_ps(s3), _mm_set1_ps(s2));
+
+ accum = _mm256_add_ps(accum, _mm256_mul_ps(p01, scales01));
+ accum = _mm256_add_ps(accum, _mm256_mul_ps(p23, scales23));
+ }
+ sumf = hsum_float_8(accum);
+
+#endif
+
+ for (;ib < nb; ++ib) {
+ for (int s_idx = 0; s_idx < 4; ++s_idx) {
+ const float d = GGML_CPU_UE4M3_TO_FP32(x[ib].d[s_idx]);
+ const int q8_block = s_idx / 2;
+ const int q8_off = (s_idx % 2) * QK_NVFP4_SUB;
+ const float dy = GGML_CPU_FP16_TO_FP32(y[2*ib + q8_block].d);
+
+ int sumi_lo = 0, sumi_hi = 0;
+ for (int j = 0; j < QK_NVFP4_SUB/2; ++j) {
+ const uint8_t qv = x[ib].qs[s_idx*(QK_NVFP4_SUB/2) + j];
+ sumi_lo += y[2*ib + q8_block].qs[q8_off + j + 0] * kvalues_fp4[qv & 0xf];
+ sumi_hi += y[2*ib + q8_block].qs[q8_off + j + QK_NVFP4_SUB/2] * kvalues_fp4[qv >> 4];
+ }
+
+ sumf += dy * d * (sumi_lo + sumi_hi);
+ }
+ }
+ *s = sumf;
+}
+
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK8_0;
+ const int nb = n / qk;
+
+ int ib = 0;
+
+ assert(n % qk == 0);
+ assert(qk == QK5_0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q5_0 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+#if defined(__AVX2__)
+ // Initialize accumulator with zeros
+ __m256 acc = _mm256_setzero_ps();
+
+ // Main loop
+ for (; ib < nb; ++ib) {
+ /* Compute combined scale for the block */
+ const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
+
+ __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+ __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+ bxhi = _mm256_andnot_si256(bxhi, _mm256_set1_epi8((char)0xF0));
+ qx = _mm256_or_si256(qx, bxhi);
+
+ __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+ const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+ /* Multiply q with scale and accumulate */
+ acc = _mm256_fmadd_ps(d, q, acc);
+ }
+
+ *s = hsum_float_8(acc);
+#elif defined(__AVX__)
+ // Initialize accumulator with zeros
+ __m256 acc = _mm256_setzero_ps();
+ __m128i mask = _mm_set1_epi8((char)0xF0);
+
+ // Main loop
+ for (; ib < nb; ++ib) {
+ /* Compute combined scale for the block */
+ const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
+
+ __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
+ const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+ __m128i bxhil = _mm256_castsi256_si128(bxhi);
+ __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
+ bxhil = _mm_andnot_si128(bxhil, mask);
+ bxhih = _mm_andnot_si128(bxhih, mask);
+ __m128i bxl = _mm256_castsi256_si128(bx_0);
+ __m128i bxh = _mm256_extractf128_si256(bx_0, 1);
+ bxl = _mm_or_si128(bxl, bxhil);
+ bxh = _mm_or_si128(bxh, bxhih);
+ bx_0 = MM256_SET_M128I(bxh, bxl);
+
+ const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+ const __m256 q = mul_sum_i8_pairs_float(bx_0, by_0);
+
+ /* Multiply q with scale and accumulate */
+ acc = _mm256_add_ps(_mm256_mul_ps(d, q), acc);
+ }
+
+ *s = hsum_float_8(acc);
+#else
+ UNUSED(nb);
+ UNUSED(ib);
+ UNUSED(x);
+ UNUSED(y);
+ ggml_vec_dot_q5_0_q8_0_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK8_1;
+ const int nb = n / qk;
+
+ int ib = 0;
+
+ assert(n % qk == 0);
+ assert(qk == QK5_1);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q5_1 * GGML_RESTRICT x = vx;
+ const block_q8_1 * GGML_RESTRICT y = vy;
+
+#if defined(__AVX2__)
+ // Initialize accumulator with zeros
+ __m256 acc = _mm256_setzero_ps();
+
+ float summs = 0.0f;
+
+ // Main loop
+ for (; ib < nb; ++ib) {
+ const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
+
+ summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
+
+ __m256i qx = bytes_from_nibbles_32(x[ib].qs);
+ __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+ bxhi = _mm256_and_si256(bxhi, _mm256_set1_epi8(0x10));
+ qx = _mm256_or_si256(qx, bxhi);
+
+ const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d));
+ const __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+ const __m256 q = mul_sum_us8_pairs_float(qx, qy);
+
+ acc = _mm256_fmadd_ps(q, _mm256_mul_ps(dx, dy), acc);
+ }
+
+ *s = hsum_float_8(acc) + summs;
+#elif defined(__AVX__)
+ // Initialize accumulator with zeros
+ __m256 acc = _mm256_setzero_ps();
+ __m128i mask = _mm_set1_epi8(0x10);
+
+ float summs = 0.0f;
+
+ // Main loop
+ for (; ib < nb; ++ib) {
+ const __m256 dx = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d));
+
+ summs += GGML_CPU_FP16_TO_FP32(x[ib].m) * GGML_CPU_FP16_TO_FP32(y[ib].s);
+
+ __m256i bx_0 = bytes_from_nibbles_32(x[ib].qs);
+ const __m256i bxhi = bytes_from_bits_32(x[ib].qh);
+ __m128i bxhil = _mm256_castsi256_si128(bxhi);
+ __m128i bxhih = _mm256_extractf128_si256(bxhi, 1);
+ bxhil = _mm_and_si128(bxhil, mask);
+ bxhih = _mm_and_si128(bxhih, mask);
+ __m128i bxl = _mm256_castsi256_si128(bx_0);
+ __m128i bxh = _mm256_extractf128_si256(bx_0, 1);
+ bxl = _mm_or_si128(bxl, bxhil);
+ bxh = _mm_or_si128(bxh, bxhih);
+ bx_0 = MM256_SET_M128I(bxh, bxl);
+
+ const __m256 dy = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib].d));
+ const __m256i by_0 = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+ const __m256 q = mul_sum_us8_pairs_float(bx_0, by_0);
+
+ acc = _mm256_add_ps(_mm256_mul_ps(q, _mm256_mul_ps(dx, dy)), acc);
+ }
+
+ *s = hsum_float_8(acc) + summs;
+#else
+ UNUSED(nb);
+ UNUSED(ib);
+ UNUSED(x);
+ UNUSED(y);
+ ggml_vec_dot_q5_1_q8_1_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK8_0;
+ const int nb = n / qk;
+
+ assert(n % qk == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q8_0 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ int ib = 0;
+ float sumf = 0;
+
+#if defined(__AVX2__)
+ // Initialize accumulator with zeros
+ __m256 acc = _mm256_setzero_ps();
+
+ // Main loop
+ for (; ib < nb; ++ib) {
+ // Compute combined scale for the block
+ const __m256 d = _mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ib].d) * GGML_CPU_FP16_TO_FP32(y[ib].d));
+ __m256i qx = _mm256_loadu_si256((const __m256i *)x[ib].qs);
+ __m256i qy = _mm256_loadu_si256((const __m256i *)y[ib].qs);
+
+ const __m256 q = mul_sum_i8_pairs_float(qx, qy);
+
+ // Multiply q with scale and accumulate
+ acc = _mm256_fmadd_ps( d, q, acc );
+ }
+
+ sumf = hsum_float_8(acc);
+#elif defined(__AVX__)
+ __m256 accum = _mm256_setzero_ps();
+
+ for (; ib + 1 < nb; ib += 2) {
+ const __m128i qx_1_0 = _mm_loadu_si128((const __m128i *)x[ib].qs);
+ const __m128i qx_1_1 = _mm_loadu_si128((const __m128i *)x[ib].qs + 1);
+ const __m128i qx_2_0 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+ const __m128i qx_2_1 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs + 1);
+ const __m128i qy_1_0 = _mm_loadu_si128((const __m128i *)y[ib].qs);
+ const __m128i qy_1_1 = _mm_loadu_si128((const __m128i *)y[ib].qs + 1);
+ const __m128i qy_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+ const __m128i qy_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+ const __m256 p = mul_sum_i8_quad_float(qx_1_0, qx_1_1, qx_2_0, qx_2_1, qy_1_0, qy_1_1, qy_2_0, qy_2_1);
+ const __m256 deltas = quad_fp16_delta_float(x[ib].d, y[ib].d, x[ib + 1].d, y[ib + 1].d);
+ accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
+ }
+
+ sumf = hsum_float_8(accum);
+#endif
+ for (; ib < nb; ++ib) {
+ int sumi = 0;
+
+ for (int j = 0; j < qk; j++) {
+ sumi += x[ib].qs[j]*y[ib].qs[j];
+ }
+
+ sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
+ }
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_tq1_0 * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+ __m256 sumf = _mm256_setzero_ps();
+
+ for (int i = 0; i < nb; ++i) {
+ // 16-bit sums
+ __m256i sumi0 = _mm256_setzero_si256();
+ __m256i sumi1 = _mm256_setzero_si256();
+ __m256i sumi2 = _mm256_setzero_si256();
+
+ // first 32 bytes of 5 elements
+ {
+ __m256i qx0 = _mm256_loadu_si256((const __m256i *) (x[i].qs));
+ // 8-bit multiplies with shifts, masks and adds
+ __m256i qx1 = _mm256_add_epi8(qx0, _mm256_add_epi8(qx0, qx0)); // 1 * 3
+ __m256i qx2 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx0, 3), _mm256_set1_epi8(-8)), qx0); // 1 * 9
+ __m256i qx3 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx1, 3), _mm256_set1_epi8(-8)), qx1); // 3 * 9
+ __m256i qx4 = _mm256_add_epi8(_mm256_and_si256(_mm256_slli_epi16(qx2, 3), _mm256_set1_epi8(-8)), qx2); // 9 * 9
+
+ // TODO: can _mm256_mulhi_epu16 be faster even if 16-bits?
+
+ // Cancel the +1 from avg so that it behaves like a halving add
+ qx0 = _mm256_subs_epu8(qx0, _mm256_set1_epi8(1));
+ qx1 = _mm256_subs_epu8(qx1, _mm256_set1_epi8(1));
+ qx2 = _mm256_subs_epu8(qx2, _mm256_set1_epi8(1));
+ qx3 = _mm256_subs_epu8(qx3, _mm256_set1_epi8(1));
+ qx4 = _mm256_subs_epu8(qx4, _mm256_set1_epi8(1));
+ // Multiply by 3 and get the top 2 bits
+ qx0 = _mm256_avg_epu8(qx0, _mm256_avg_epu8(qx0, _mm256_setzero_si256()));
+ qx1 = _mm256_avg_epu8(qx1, _mm256_avg_epu8(qx1, _mm256_setzero_si256()));
+ qx2 = _mm256_avg_epu8(qx2, _mm256_avg_epu8(qx2, _mm256_setzero_si256()));
+ qx3 = _mm256_avg_epu8(qx3, _mm256_avg_epu8(qx3, _mm256_setzero_si256()));
+ qx4 = _mm256_avg_epu8(qx4, _mm256_avg_epu8(qx4, _mm256_setzero_si256()));
+ qx0 = _mm256_and_si256(_mm256_srli_epi16(qx0, 6), _mm256_set1_epi8(3));
+ qx1 = _mm256_and_si256(_mm256_srli_epi16(qx1, 6), _mm256_set1_epi8(3));
+ qx2 = _mm256_and_si256(_mm256_srli_epi16(qx2, 6), _mm256_set1_epi8(3));
+ qx3 = _mm256_and_si256(_mm256_srli_epi16(qx3, 6), _mm256_set1_epi8(3));
+ qx4 = _mm256_and_si256(_mm256_srli_epi16(qx4, 6), _mm256_set1_epi8(3));
+
+ const __m256i qy0 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 0));
+ const __m256i qy1 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 32));
+ const __m256i qy2 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 64));
+ const __m256i qy3 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 96));
+ const __m256i qy4 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 128));
+
+ qx0 = _mm256_maddubs_epi16(qx0, qy0);
+ qx1 = _mm256_maddubs_epi16(qx1, qy1);
+ qx2 = _mm256_maddubs_epi16(qx2, qy2);
+ qx3 = _mm256_maddubs_epi16(qx3, qy3);
+ qx4 = _mm256_maddubs_epi16(qx4, qy4);
+
+ sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(qx0, qx1));
+ sumi1 = _mm256_add_epi16(sumi1, _mm256_add_epi16(qx2, qx3));
+ sumi2 = _mm256_add_epi16(sumi2, qx4);
+ }
+
+ // last 16 bytes of 5-element, along with the 4 bytes of 4 elements
+ {
+ __m128i qx0 = _mm_loadu_si128((const __m128i *) (x[i].qs + 32));
+ uint32_t qh;
+ memcpy(&qh, x[i].qh, sizeof(qh)); // potentially unaligned
+ __m256i qx5_l = _mm256_cvtepu8_epi16(_mm_set1_epi32(qh));
+ __m128i qx1 = _mm_add_epi8(qx0, _mm_add_epi8(qx0, qx0)); // 1 * 3
+ __m128i qx2 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx0, 3), _mm_set1_epi8(-8)), qx0); // 1 * 9
+ __m128i qx3 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx1, 3), _mm_set1_epi8(-8)), qx1); // 3 * 9
+ __m128i qx4 = _mm_add_epi8(_mm_and_si128(_mm_slli_epi16(qx2, 3), _mm_set1_epi8(-8)), qx2); // 9 * 9
+ __m256i qx01 = MM256_SET_M128I(qx1, qx0);
+ __m256i qx23 = MM256_SET_M128I(qx3, qx2);
+
+ // avx2 does not have 8-bit multiplies, so 16-bit it is.
+ qx5_l = _mm256_mullo_epi16(qx5_l, _mm256_set_epi16(27, 27, 27, 27, 9, 9, 9, 9, 3, 3, 3, 3, 1, 1, 1, 1));
+ qx5_l = _mm256_and_si256(qx5_l, _mm256_set1_epi16(0xFF));
+ __m128i qx5 = _mm_packus_epi16(_mm256_castsi256_si128(qx5_l), _mm256_extracti128_si256(qx5_l, 1));
+
+ __m256i qx45 = MM256_SET_M128I(qx5, qx4);
+
+ // Cancel the +1 from avg so that it behaves like a halving add
+ qx01 = _mm256_subs_epu8(qx01, _mm256_set1_epi8(1));
+ qx23 = _mm256_subs_epu8(qx23, _mm256_set1_epi8(1));
+ qx45 = _mm256_subs_epu8(qx45, _mm256_set1_epi8(1));
+ // Multiply by 3 and get the top 2 bits
+ qx01 = _mm256_avg_epu8(qx01, _mm256_avg_epu8(qx01, _mm256_setzero_si256()));
+ qx23 = _mm256_avg_epu8(qx23, _mm256_avg_epu8(qx23, _mm256_setzero_si256()));
+ qx45 = _mm256_avg_epu8(qx45, _mm256_avg_epu8(qx45, _mm256_setzero_si256()));
+ qx01 = _mm256_and_si256(_mm256_srli_epi16(qx01, 6), _mm256_set1_epi8(3));
+ qx23 = _mm256_and_si256(_mm256_srli_epi16(qx23, 6), _mm256_set1_epi8(3));
+ qx45 = _mm256_and_si256(_mm256_srli_epi16(qx45, 6), _mm256_set1_epi8(3));
+
+ const __m256i qy01 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 160));
+ const __m256i qy23 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 192));
+ const __m256i qy45 = _mm256_loadu_si256((const __m256i *) (y[i].qs + 224));
+
+ qx01 = _mm256_maddubs_epi16(qx01, qy01);
+ qx23 = _mm256_maddubs_epi16(qx23, qy23);
+ qx45 = _mm256_maddubs_epi16(qx45, qy45);
+
+ sumi0 = _mm256_add_epi16(sumi0, qx01);
+ sumi1 = _mm256_add_epi16(sumi1, qx23);
+ sumi2 = _mm256_add_epi16(sumi2, qx45);
+ }
+
+ const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
+ const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d));
+
+ sumi0 = _mm256_sub_epi16(sumi0, ysum);
+ sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(sumi1, sumi2));
+ sumi0 = _mm256_madd_epi16(sumi0, _mm256_set1_epi16(1));
+
+ sumf = _mm256_add_ps(_mm256_mul_ps(_mm256_cvtepi32_ps(sumi0), d), sumf);
+ }
+
+ *s = hsum_float_8(sumf);
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ ggml_vec_dot_tq1_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_tq2_0 * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+ __m256 sumf = _mm256_setzero_ps();
+
+ for (int i = 0; i < nb; ++i) {
+ // 16-bit sums, because 256*127 still fits
+ __m256i sumi0 = _mm256_setzero_si256();
+ __m256i sumi1 = _mm256_setzero_si256();
+
+ for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+ __m256i qx0 = _mm256_loadu_si256((const __m256i *) (x[i].qs + j));
+ __m256i qx1 = _mm256_srli_epi16(qx0, 2);
+ __m256i qx2 = _mm256_srli_epi16(qx0, 4);
+ __m256i qx3 = _mm256_srli_epi16(qx0, 6);
+
+ // 0, 1, 2 (should not be 3)
+ qx0 = _mm256_and_si256(qx0, _mm256_set1_epi8(3));
+ qx1 = _mm256_and_si256(qx1, _mm256_set1_epi8(3));
+ qx2 = _mm256_and_si256(qx2, _mm256_set1_epi8(3));
+ qx3 = _mm256_and_si256(qx3, _mm256_set1_epi8(3));
+
+ const __m256i qy0 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 0));
+ const __m256i qy1 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 32));
+ const __m256i qy2 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 64));
+ const __m256i qy3 = _mm256_loadu_si256((const __m256i *) (y[i].qs + j*4 + 96));
+
+ qx0 = _mm256_maddubs_epi16(qx0, qy0);
+ qx1 = _mm256_maddubs_epi16(qx1, qy1);
+ qx2 = _mm256_maddubs_epi16(qx2, qy2);
+ qx3 = _mm256_maddubs_epi16(qx3, qy3);
+
+ sumi0 = _mm256_add_epi16(sumi0, _mm256_add_epi16(qx0, qx1));
+ sumi1 = _mm256_add_epi16(sumi1, _mm256_add_epi16(qx2, qx3));
+ }
+
+ const __m256i ysum = _mm256_loadu_si256((const __m256i *) y[i].bsums);
+ const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d));
+
+ sumi0 = _mm256_add_epi16(sumi0, sumi1);
+ sumi0 = _mm256_sub_epi16(sumi0, ysum);
+ sumi0 = _mm256_madd_epi16(sumi0, _mm256_set1_epi16(1));
+
+ sumf = _mm256_add_ps(_mm256_mul_ps(_mm256_cvtepi32_ps(sumi0), d), sumf);
+ }
+
+ *s = hsum_float_8(sumf);
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ ggml_vec_dot_tq2_0_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q2_K * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+#if defined __AVX2__
+
+ const __m256i m3 = _mm256_set1_epi8(3);
+ const __m128i m4 = _mm_set1_epi8(0xF);
+
+ __m256 acc = _mm256_setzero_ps();
+
+ for (int i = 0; i < nb; ++i) {
+
+ const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+ const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+ const uint8_t * GGML_RESTRICT q2 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+ const __m128i scales8 = _mm_and_si128(mins_and_scales, m4);
+ const __m128i mins8 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
+ const __m256i mins = _mm256_cvtepi8_epi16(mins8);
+ const __m256i prod = _mm256_madd_epi16(mins, _mm256_loadu_si256((const __m256i*)y[i].bsums));
+
+ acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(prod), acc);
+
+ const __m256i all_scales = _mm256_cvtepi8_epi16(scales8);
+ const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
+ const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
+ const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
+
+ __m256i sumi = _mm256_setzero_si256();
+
+ for (int j = 0; j < QK_K/128; ++j) {
+
+ const __m256i q2bits = _mm256_loadu_si256((const __m256i*)q2); q2 += 32;
+
+ const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+ const __m256i q2_0 = _mm256_and_si256(q2bits, m3);
+ const __m256i q2_1 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 2), m3);
+ const __m256i q2_2 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 4), m3);
+ const __m256i q2_3 = _mm256_and_si256(_mm256_srli_epi16(q2bits, 6), m3);
+
+ __m256i p0 = _mm256_maddubs_epi16(q2_0, q8_0);
+ __m256i p1 = _mm256_maddubs_epi16(q2_1, q8_1);
+ __m256i p2 = _mm256_maddubs_epi16(q2_2, q8_2);
+ __m256i p3 = _mm256_maddubs_epi16(q2_3, q8_3);
+
+ p0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(0)), p0);
+ p1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(1)), p1);
+ p2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(2)), p2);
+ p3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(3)), p3);
+
+ p0 = _mm256_add_epi32(p0, p1);
+ p2 = _mm256_add_epi32(p2, p3);
+
+ sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p0, p2));
+ }
+
+ acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+
+ }
+
+ *s = hsum_float_8(acc);
+
+#elif defined __AVX__
+
+ const __m128i m3 = _mm_set1_epi8(0x3);
+ const __m128i m4 = _mm_set1_epi8(0xF);
+ const __m128i m2 = _mm_set1_epi8(0x2);
+
+ __m256 acc = _mm256_setzero_ps();
+
+ for (int i = 0; i < nb; ++i) {
+
+ const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+ const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+ const uint8_t * GGML_RESTRICT q2 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ // load mins and scales from block_q2_K.scales[QK_K/16]
+ const __m128i mins_and_scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+ const __m128i scales16 = _mm_and_si128(mins_and_scales, m4);
+ const __m128i mins16 = _mm_and_si128(_mm_srli_epi16(mins_and_scales, 4), m4);
+ const __m128i mins_0 = _mm_cvtepi8_epi16(mins16);
+ const __m128i mins_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(mins16, mins16));
+
+ // summs = y[i].bsums * (x[i].scales >> 4) in 16bits*8*2 to 32bits*4*2
+ const __m128i summs_0 = _mm_madd_epi16(mins_0, _mm_loadu_si128((const __m128i*)&y[i].bsums[0]));
+ const __m128i summs_1 = _mm_madd_epi16(mins_1, _mm_loadu_si128((const __m128i*)&y[i].bsums[8]));
+
+ // sumf += -dmin * summs in 32bits*8
+ acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dmin), _mm256_cvtepi32_ps(MM256_SET_M128I(summs_1, summs_0))), acc);
+
+ const __m128i scales_0 = _mm_cvtepi8_epi16(scales16);
+ const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales16, scales16));
+ const __m128i scales[2] = { scales_0, scales_1 };
+
+ __m128i sumi_0 = _mm_setzero_si128();
+ __m128i sumi_1 = _mm_setzero_si128();
+
+ for (int j = 0; j < QK_K/128; ++j) {
+
+ // load Q8 quants int8*16*8 from block_q8_K.qs[QK_K]
+ const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+ // load 2bits*16*8 from block_q2_K.qs[QK_K/4]
+ __m128i q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
+ const __m128i q2_0 = _mm_and_si128(q2bits, m3);
+ const __m128i q2_2 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+ const __m128i q2_4 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+ const __m128i q2_6 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+ q2bits = _mm_loadu_si128((const __m128i*)q2); q2 += 16;
+ const __m128i q2_1 = _mm_and_si128(q2bits, m3);
+ const __m128i q2_3 = _mm_and_si128(_mm_srli_epi16(q2bits, 2), m3);
+ const __m128i q2_5 = _mm_and_si128(_mm_srli_epi16(q2bits, 4), m3);
+ const __m128i q2_7 = _mm_and_si128(_mm_srli_epi16(q2bits, 6), m3);
+
+ // isuml = q8[l] * ((q2[l] >> shift) & 3) in 8bits*16*8 to 16bits*8*8
+ __m128i p0 = _mm_maddubs_epi16(q2_0, q8_0);
+ __m128i p1 = _mm_maddubs_epi16(q2_1, q8_1);
+ __m128i p2 = _mm_maddubs_epi16(q2_2, q8_2);
+ __m128i p3 = _mm_maddubs_epi16(q2_3, q8_3);
+ __m128i p4 = _mm_maddubs_epi16(q2_4, q8_4);
+ __m128i p5 = _mm_maddubs_epi16(q2_5, q8_5);
+ __m128i p6 = _mm_maddubs_epi16(q2_6, q8_6);
+ __m128i p7 = _mm_maddubs_epi16(q2_7, q8_7);
+
+ // isum += (x[i].scales[is++] & 0xF) * isuml in 16bits*8*8 to 32bits*4*8
+ __m128i shuffle = _mm_set1_epi16(0x0100);
+ p0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p0);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p1);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p2);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p3);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p4);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p5);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p6);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p7);
+
+ p0 = _mm_add_epi32(p0, p1);
+ p2 = _mm_add_epi32(p2, p3);
+ p4 = _mm_add_epi32(p4, p5);
+ p6 = _mm_add_epi32(p6, p7);
+
+ // isum in 32bits*4*2
+ sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p0, p2));
+ sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p4, p6));
+ }
+
+ // sumf += dall * isum - dmin * summs in 32bits
+ __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+ acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&dall), _mm256_cvtepi32_ps(sumi)), acc);
+ }
+
+ *s = hsum_float_8(acc);
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ ggml_vec_dot_q2_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const uint32_t kmask1 = 0x03030303;
+ const uint32_t kmask2 = 0x0f0f0f0f;
+
+ const block_q3_K * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+#if defined __AVX2__
+
+ const __m256i m3 = _mm256_set1_epi8(3);
+ const __m256i mone = _mm256_set1_epi8(1);
+ const __m128i m32 = _mm_set1_epi8(32);
+
+ __m256 acc = _mm256_setzero_ps();
+
+ uint32_t aux[3];
+
+ for (int i = 0; i < nb; ++i) {
+
+ const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+ const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ // Set up scales
+ memcpy(aux, x[i].scales, 12);
+ __m128i scales128 = _mm_set_epi32(
+ ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
+ ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
+ (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
+ (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
+ scales128 = _mm_sub_epi8(scales128, m32);
+ const __m256i all_scales = _mm256_cvtepi8_epi16(scales128);
+ const __m128i l_scales = _mm256_extracti128_si256(all_scales, 0);
+ const __m128i h_scales = _mm256_extracti128_si256(all_scales, 1);
+ const __m256i scales[2] = {MM256_SET_M128I(l_scales, l_scales), MM256_SET_M128I(h_scales, h_scales)};
+
+ // high bit
+ const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].hmask);
+
+ // integer accumulator
+ __m256i sumi = _mm256_setzero_si256();
+
+ int bit = 0;
+ int is = 0;
+
+ for (int j = 0; j < QK_K/128; ++j) {
+ // load low 2 bits
+ const __m256i q3bits = _mm256_loadu_si256((const __m256i*)q3); q3 += 32;
+
+ // prepare low and high bits
+ const __m256i q3l_0 = _mm256_and_si256(q3bits, m3);
+ const __m256i q3h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+ ++bit;
+
+ const __m256i q3l_1 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 2), m3);
+ const __m256i q3h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+ ++bit;
+
+ const __m256i q3l_2 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 4), m3);
+ const __m256i q3h_2 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+ ++bit;
+
+ const __m256i q3l_3 = _mm256_and_si256(_mm256_srli_epi16(q3bits, 6), m3);
+ const __m256i q3h_3 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_andnot_si256(hbits, _mm256_slli_epi16(mone, bit)), bit), 2);
+ ++bit;
+
+ // load Q8 quants
+ const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+ // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
+ // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+ // and 2 if the high bit was set)
+ __m256i q8s_0 = _mm256_maddubs_epi16(q3h_0, q8_0);
+ __m256i q8s_1 = _mm256_maddubs_epi16(q3h_1, q8_1);
+ __m256i q8s_2 = _mm256_maddubs_epi16(q3h_2, q8_2);
+ __m256i q8s_3 = _mm256_maddubs_epi16(q3h_3, q8_3);
+
+ __m256i p16_0 = _mm256_maddubs_epi16(q3l_0, q8_0);
+ __m256i p16_1 = _mm256_maddubs_epi16(q3l_1, q8_1);
+ __m256i p16_2 = _mm256_maddubs_epi16(q3l_2, q8_2);
+ __m256i p16_3 = _mm256_maddubs_epi16(q3l_3, q8_3);
+
+ p16_0 = _mm256_sub_epi16(p16_0, q8s_0);
+ p16_1 = _mm256_sub_epi16(p16_1, q8s_1);
+ p16_2 = _mm256_sub_epi16(p16_2, q8s_2);
+ p16_3 = _mm256_sub_epi16(p16_3, q8s_3);
+
+ // multiply with scales
+ p16_0 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 0)), p16_0);
+ p16_1 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 1)), p16_1);
+ p16_2 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 2)), p16_2);
+ p16_3 = _mm256_madd_epi16(_mm256_shuffle_epi8(scales[j], get_scale_shuffle_q3k(is + 3)), p16_3);
+
+ // accumulate
+ p16_0 = _mm256_add_epi32(p16_0, p16_1);
+ p16_2 = _mm256_add_epi32(p16_2, p16_3);
+ sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_2));
+
+ }
+
+ // multiply with block scale and accumulate
+ acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+
+ }
+
+ *s = hsum_float_8(acc);
+
+#elif defined __AVX__
+
+ const __m128i m3 = _mm_set1_epi8(3);
+ const __m128i mone = _mm_set1_epi8(1);
+ const __m128i m32 = _mm_set1_epi8(32);
+ const __m128i m2 = _mm_set1_epi8(2);
+
+ __m256 acc = _mm256_setzero_ps();
+
+ const uint32_t *aux;
+
+ for (int i = 0; i < nb; ++i) {
+
+ const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+ const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ // Set up scales
+ aux = (const uint32_t *)x[i].scales;
+ __m128i scales128 = _mm_set_epi32(
+ ((aux[1] >> 4) & kmask2) | (((aux[2] >> 6) & kmask1) << 4),
+ ((aux[0] >> 4) & kmask2) | (((aux[2] >> 4) & kmask1) << 4),
+ (aux[1] & kmask2) | (((aux[2] >> 2) & kmask1) << 4),
+ (aux[0] & kmask2) | (((aux[2] >> 0) & kmask1) << 4));
+ scales128 = _mm_sub_epi8(scales128, m32);
+ const __m128i scales_0 = _mm_cvtepi8_epi16(scales128);
+ const __m128i scales_1 = _mm_cvtepi8_epi16(_mm_unpackhi_epi64(scales128, scales128));
+ const __m128i scales[2] = { scales_0, scales_1 };
+
+ // high bit *128*2 from block_q3_K.hmask[QK_K/8]
+ const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].hmask[0]);
+ const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].hmask[16]);
+
+ // integer accumulator
+ __m128i sumi_0 = _mm_setzero_si128();
+ __m128i sumi_1 = _mm_setzero_si128();
+
+ for (int j = 0; j < QK_K/128; ++j) {
+ // load low 2 bits *64*2 from block_q3_K.qs[QK_K/4]
+ const __m128i q3bits_0 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
+ const __m128i q3bits_1 = _mm_loadu_si128((const __m128i*)q3); q3 += 16;
+
+ // prepare low and high bits
+ const int bit = j << 2;
+
+ const __m128i q3l_0 = _mm_and_si128(q3bits_0, m3);
+ const __m128i q3l_1 = _mm_and_si128(q3bits_1, m3);
+ const __m128i q3h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit)), bit), 2);
+ const __m128i q3h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit)), bit), 2);
+
+ const __m128i q3l_2 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 2), m3);
+ const __m128i q3l_3 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 2), m3);
+ const __m128i q3h_2 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
+ const __m128i q3h_3 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+1)), bit+1), 2);
+
+ const __m128i q3l_4 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 4), m3);
+ const __m128i q3l_5 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 4), m3);
+ const __m128i q3h_4 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
+ const __m128i q3h_5 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+2)), bit+2), 2);
+
+ const __m128i q3l_6 = _mm_and_si128(_mm_srli_epi16(q3bits_0, 6), m3);
+ const __m128i q3l_7 = _mm_and_si128(_mm_srli_epi16(q3bits_1, 6), m3);
+ const __m128i q3h_6 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_0, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
+ const __m128i q3h_7 = _mm_slli_epi16(_mm_srli_epi16(_mm_andnot_si128(hbits_1, _mm_slli_epi16(mone, bit+3)), bit+3), 2);
+
+ // load Q8 quants from block_q8_K.qs[QK_K]
+ const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+ // Dot product: we multiply the 2 low bits and 1 high bit part separately, so we can use _mm256_maddubs_epi16,
+ // and then subtract. The high bit part has the 2 already subtracted (and so, it is zero if the high bit was not set,
+ // and 2 if the high bit was set)
+ __m128i q8s_0 = _mm_maddubs_epi16(q3h_0, q8_0);
+ __m128i q8s_1 = _mm_maddubs_epi16(q3h_1, q8_1);
+ __m128i q8s_2 = _mm_maddubs_epi16(q3h_2, q8_2);
+ __m128i q8s_3 = _mm_maddubs_epi16(q3h_3, q8_3);
+ __m128i q8s_4 = _mm_maddubs_epi16(q3h_4, q8_4);
+ __m128i q8s_5 = _mm_maddubs_epi16(q3h_5, q8_5);
+ __m128i q8s_6 = _mm_maddubs_epi16(q3h_6, q8_6);
+ __m128i q8s_7 = _mm_maddubs_epi16(q3h_7, q8_7);
+
+ __m128i p16_0 = _mm_maddubs_epi16(q3l_0, q8_0);
+ __m128i p16_1 = _mm_maddubs_epi16(q3l_1, q8_1);
+ __m128i p16_2 = _mm_maddubs_epi16(q3l_2, q8_2);
+ __m128i p16_3 = _mm_maddubs_epi16(q3l_3, q8_3);
+ __m128i p16_4 = _mm_maddubs_epi16(q3l_4, q8_4);
+ __m128i p16_5 = _mm_maddubs_epi16(q3l_5, q8_5);
+ __m128i p16_6 = _mm_maddubs_epi16(q3l_6, q8_6);
+ __m128i p16_7 = _mm_maddubs_epi16(q3l_7, q8_7);
+
+ p16_0 = _mm_sub_epi16(p16_0, q8s_0);
+ p16_1 = _mm_sub_epi16(p16_1, q8s_1);
+ p16_2 = _mm_sub_epi16(p16_2, q8s_2);
+ p16_3 = _mm_sub_epi16(p16_3, q8s_3);
+ p16_4 = _mm_sub_epi16(p16_4, q8s_4);
+ p16_5 = _mm_sub_epi16(p16_5, q8s_5);
+ p16_6 = _mm_sub_epi16(p16_6, q8s_6);
+ p16_7 = _mm_sub_epi16(p16_7, q8s_7);
+
+ // multiply with scales
+ __m128i shuffle = _mm_set1_epi16(0x0100);
+ p16_0 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_0);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p16_1 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_1);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p16_2 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_2);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p16_3 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_3);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p16_4 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_4);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p16_5 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_5);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p16_6 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_6);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ p16_7 = _mm_madd_epi16(_mm_shuffle_epi8(scales[j], shuffle), p16_7);
+
+ // accumulate
+ p16_0 = _mm_add_epi32(p16_0, p16_1);
+ p16_2 = _mm_add_epi32(p16_2, p16_3);
+ p16_4 = _mm_add_epi32(p16_4, p16_5);
+ p16_6 = _mm_add_epi32(p16_6, p16_7);
+ sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+ sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_4, p16_6));
+
+ }
+
+ // multiply with block scale and accumulate
+ __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+ acc = _mm256_add_ps(_mm256_mul_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi)), acc);
+
+ }
+
+ *s = hsum_float_8(acc);
+
+#else
+ UNUSED(kmask1);
+ UNUSED(kmask2);
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ ggml_vec_dot_q3_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q4_K * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ static const uint32_t kmask1 = 0x3f3f3f3f;
+ static const uint32_t kmask2 = 0x0f0f0f0f;
+ static const uint32_t kmask3 = 0x03030303;
+
+ uint32_t utmp[4];
+
+#if defined __AVX2__
+
+ const __m256i m4 = _mm256_set1_epi8(0xF);
+
+ __m256 acc = _mm256_setzero_ps();
+ __m128 acc_m = _mm_setzero_ps();
+
+ for (int i = 0; i < nb; ++i) {
+
+ const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+ const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+ memcpy(utmp, x[i].scales, 12);
+ utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+ const uint32_t uaux = utmp[1] & kmask1;
+ utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+ utmp[2] = uaux;
+ utmp[0] &= kmask1;
+
+ const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
+
+ const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
+ const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
+ const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
+ acc_m = _mm_fmadd_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod), acc_m);
+
+ const __m128i sc128 = _mm256_extracti128_si256(mins_and_scales, 0);
+ const __m256i scales = MM256_SET_M128I(sc128, sc128);
+
+ __m256i sumi = _mm256_setzero_si256();
+
+ for (int j = 0; j < QK_K/64; ++j) {
+
+ const __m256i scale_l = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
+ const __m256i scale_h = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
+
+ const __m256i q4bits = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
+ const __m256i q4l = _mm256_and_si256(q4bits, m4);
+ const __m256i q4h = _mm256_and_si256(_mm256_srli_epi16(q4bits, 4), m4);
+
+ const __m256i q8l = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ __m256i p16l = _mm256_maddubs_epi16(q4l, q8l);
+ p16l = _mm256_madd_epi16(scale_l, p16l);
+
+ const __m256i q8h = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ __m256i p16h = _mm256_maddubs_epi16(q4h, q8h);
+ p16h = _mm256_madd_epi16(scale_h, p16h);
+ const __m256i sumj = _mm256_add_epi32(p16l, p16h);
+
+ sumi = _mm256_add_epi32(sumi, sumj);
+ }
+
+ __m256 vd = _mm256_set1_ps(d);
+ acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
+
+ }
+
+ acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
+ acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
+
+ *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
+
+#elif defined __AVX__
+
+ const __m128i m4 = _mm_set1_epi8(0xF);
+ const __m128i m2 = _mm_set1_epi8(0x2);
+
+ __m256 acc = _mm256_setzero_ps();
+ __m128 acc_m = _mm_setzero_ps();
+
+ for (int i = 0; i < nb; ++i) {
+
+ const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+ const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+ const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ memcpy(utmp, x[i].scales, 12);
+ utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+ const uint32_t uaux = utmp[1] & kmask1;
+ utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+ utmp[2] = uaux;
+ utmp[0] &= kmask1;
+
+ const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
+ const __m128i scales = _mm_cvtepu8_epi16(utmps);
+ const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
+
+ const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
+ const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
+ const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
+ const __m128i prod = _mm_madd_epi16(mins, q8s);
+ acc_m = _mm_add_ps(_mm_mul_ps(_mm_set1_ps(dmin), _mm_cvtepi32_ps(prod)), acc_m);
+
+ __m128i sumi_0 = _mm_setzero_si128();
+ __m128i sumi_1 = _mm_setzero_si128();
+
+ __m128i shuffle = _mm_set1_epi16(0x0100);
+ for (int j = 0; j < QK_K/64; ++j) {
+
+ const __m128i scale_l = _mm_shuffle_epi8(scales, shuffle);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ const __m128i scale_h = _mm_shuffle_epi8(scales, shuffle);
+ shuffle = _mm_add_epi16(shuffle, m2);
+
+ __m128i q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+ const __m128i q4l_0 = _mm_and_si128(q4bits, m4);
+ const __m128i q4h_0 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
+ q4bits = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+ const __m128i q4l_1 = _mm_and_si128(q4bits, m4);
+ const __m128i q4h_1 = _mm_and_si128(_mm_srli_epi16(q4bits, 4), m4);
+
+ const __m128i q8l_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ __m128i p16l = _mm_maddubs_epi16(q4l_0, q8l_0);
+ p16l = _mm_madd_epi16(scale_l, p16l);
+ sumi_0 = _mm_add_epi32(sumi_0, p16l);
+ const __m128i q8l_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ p16l = _mm_maddubs_epi16(q4l_1, q8l_1);
+ p16l = _mm_madd_epi16(scale_l, p16l);
+ sumi_1 = _mm_add_epi32(sumi_1, p16l);
+
+ const __m128i q8h_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ __m128i p16h = _mm_maddubs_epi16(q4h_0, q8h_0);
+ p16h = _mm_madd_epi16(scale_h, p16h);
+ sumi_0 = _mm_add_epi32(sumi_0, p16h);
+ const __m128i q8h_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ p16h = _mm_maddubs_epi16(q4h_1, q8h_1);
+ p16h = _mm_madd_epi16(scale_h, p16h);
+ sumi_1 = _mm_add_epi32(sumi_1, p16h);
+
+ }
+
+ __m256 vd = _mm256_set1_ps(d);
+ __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+ acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
+
+ }
+
+ acc_m = _mm_add_ps(acc_m, _mm_movehl_ps(acc_m, acc_m));
+ acc_m = _mm_add_ss(acc_m, _mm_movehdup_ps(acc_m));
+
+ *s = hsum_float_8(acc) + _mm_cvtss_f32(acc_m);
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ UNUSED(kmask1);
+ UNUSED(kmask2);
+ UNUSED(kmask3);
+ UNUSED(utmp);
+ ggml_vec_dot_q4_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q5_K * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ static const uint32_t kmask1 = 0x3f3f3f3f;
+ static const uint32_t kmask2 = 0x0f0f0f0f;
+ static const uint32_t kmask3 = 0x03030303;
+
+ uint32_t utmp[4];
+
+#if defined __AVX2__
+
+ const __m256i m4 = _mm256_set1_epi8(0xF);
+ const __m128i mzero = _mm_setzero_si128();
+ const __m256i mone = _mm256_set1_epi8(1);
+
+ __m256 acc = _mm256_setzero_ps();
+
+ float summs = 0.f;
+
+ for (int i = 0; i < nb; ++i) {
+ const uint8_t * GGML_RESTRICT q5 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+ const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+ memcpy(utmp, x[i].scales, 12);
+ utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+ const uint32_t uaux = utmp[1] & kmask1;
+ utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+ utmp[2] = uaux;
+ utmp[0] &= kmask1;
+
+ const __m256i mins_and_scales = _mm256_cvtepu8_epi16(_mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]));
+
+ const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
+ const __m128i q8s = _mm_hadd_epi16(_mm256_extracti128_si256(q8sums, 0), _mm256_extracti128_si256(q8sums, 1));
+ const __m128i prod = _mm_madd_epi16(_mm256_extracti128_si256(mins_and_scales, 1), q8s);
+ const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
+ summs += dmin * _mm_extract_epi32(hsum, 0);
+
+ const __m128i sc128 = _mm256_extracti128_si256(mins_and_scales, 0);
+ const __m256i scales = MM256_SET_M128I(sc128, sc128);
+
+ const __m256i hbits = _mm256_loadu_si256((const __m256i*)x[i].qh);
+ __m256i hmask = mone;
+
+ __m256i sumi = _mm256_setzero_si256();
+
+ int bit = 0;
+
+ for (int j = 0; j < QK_K/64; ++j) {
+
+ const __m256i scale_0 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+0));
+ const __m256i scale_1 = _mm256_shuffle_epi8(scales, get_scale_shuffle_k4(2*j+1));
+
+ const __m256i q5bits = _mm256_loadu_si256((const __m256i*)q5); q5 += 32;
+
+ const __m256i q5l_0 = _mm256_and_si256(q5bits, m4);
+ const __m256i q5h_0 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
+ const __m256i q5_0 = _mm256_add_epi8(q5l_0, q5h_0);
+ hmask = _mm256_slli_epi16(hmask, 1);
+
+ const __m256i q5l_1 = _mm256_and_si256(_mm256_srli_epi16(q5bits, 4), m4);
+ const __m256i q5h_1 = _mm256_slli_epi16(_mm256_srli_epi16(_mm256_and_si256(hbits, hmask), bit++), 4);
+ const __m256i q5_1 = _mm256_add_epi8(q5l_1, q5h_1);
+ hmask = _mm256_slli_epi16(hmask, 1);
+
+ const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+ __m256i p16_0 = _mm256_maddubs_epi16(q5_0, q8_0);
+ __m256i p16_1 = _mm256_maddubs_epi16(q5_1, q8_1);
+
+ p16_0 = _mm256_madd_epi16(scale_0, p16_0);
+ p16_1 = _mm256_madd_epi16(scale_1, p16_1);
+
+ sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+
+ }
+
+ __m256 vd = _mm256_set1_ps(d);
+ acc = _mm256_fmadd_ps(vd, _mm256_cvtepi32_ps(sumi), acc);
+
+ }
+
+ *s = hsum_float_8(acc) + summs;
+
+#elif defined __AVX__
+
+ const __m128i m4 = _mm_set1_epi8(0xF);
+ const __m128i mzero = _mm_setzero_si128();
+ const __m128i mone = _mm_set1_epi8(1);
+ const __m128i m2 = _mm_set1_epi8(2);
+
+ __m256 acc = _mm256_setzero_ps();
+
+ float summs = 0.f;
+
+ for (int i = 0; i < nb; ++i) {
+
+ const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+ const float dmin = -y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+ const uint8_t * GGML_RESTRICT q5 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ memcpy(utmp, x[i].scales, 12);
+ utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+ const uint32_t uaux = utmp[1] & kmask1;
+ utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+ utmp[2] = uaux;
+ utmp[0] &= kmask1;
+
+ const __m128i utmps = _mm_set_epi32(utmp[3], utmp[2], utmp[1], utmp[0]);
+ const __m128i scales = _mm_cvtepu8_epi16(utmps);
+ const __m128i mins = _mm_cvtepu8_epi16(_mm_unpackhi_epi64(utmps, utmps));
+
+ const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)&y[i].bsums[0]);
+ const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)&y[i].bsums[8]);
+ const __m128i q8s = _mm_hadd_epi16(q8sums_0, q8sums_1);
+ const __m128i prod = _mm_madd_epi16(mins, q8s);
+ const __m128i hsum = _mm_hadd_epi32(_mm_hadd_epi32(prod, mzero), mzero);
+ summs += dmin * _mm_extract_epi32(hsum, 0);
+
+ const __m128i hbits_0 = _mm_loadu_si128((const __m128i*)&x[i].qh[0]);
+ const __m128i hbits_1 = _mm_loadu_si128((const __m128i*)&x[i].qh[16]);
+ __m128i hmask = mone;
+
+ __m128i sumi_0 = _mm_setzero_si128();
+ __m128i sumi_1 = _mm_setzero_si128();
+
+ int bit = 0;
+
+ __m128i shuffle = _mm_set1_epi16(0x0100);
+ for (int j = 0; j < QK_K/64; ++j) {
+
+ const __m128i scale_0 = _mm_shuffle_epi8(scales, shuffle);
+ shuffle = _mm_add_epi16(shuffle, m2);
+ const __m128i scale_1 = _mm_shuffle_epi8(scales, shuffle);
+ shuffle = _mm_add_epi16(shuffle, m2);
+
+ const __m128i q5bits_0 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
+ const __m128i q5bits_1 = _mm_loadu_si128((const __m128i*)q5); q5 += 16;
+
+ __m128i q5l_0 = _mm_and_si128(q5bits_0, m4);
+ __m128i q5l_1 = _mm_and_si128(q5bits_1, m4);
+ __m128i q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
+ __m128i q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
+ __m128i q5_0 = _mm_add_epi8(q5l_0, q5h_0);
+ __m128i q5_1 = _mm_add_epi8(q5l_1, q5h_1);
+ hmask = _mm_slli_epi16(hmask, 1);
+
+ __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ __m128i p16_0 = _mm_maddubs_epi16(q5_0, q8_0);
+ __m128i p16_1 = _mm_maddubs_epi16(q5_1, q8_1);
+ p16_0 = _mm_madd_epi16(scale_0, p16_0);
+ p16_1 = _mm_madd_epi16(scale_0, p16_1);
+
+ q5l_0 = _mm_and_si128(_mm_srli_epi16(q5bits_0, 4), m4);
+ q5l_1 = _mm_and_si128(_mm_srli_epi16(q5bits_1, 4), m4);
+ q5h_0 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_0, hmask), bit), 4);
+ q5h_1 = _mm_slli_epi16(_mm_srli_epi16(_mm_and_si128(hbits_1, hmask), bit++), 4);
+ q5_0 = _mm_add_epi8(q5l_0, q5h_0);
+ q5_1 = _mm_add_epi8(q5l_1, q5h_1);
+ hmask = _mm_slli_epi16(hmask, 1);
+
+ q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ __m128i p16_2 = _mm_maddubs_epi16(q5_0, q8_0);
+ __m128i p16_3 = _mm_maddubs_epi16(q5_1, q8_1);
+ p16_2 = _mm_madd_epi16(scale_1, p16_2);
+ p16_3 = _mm_madd_epi16(scale_1, p16_3);
+
+ sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+ sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
+
+ }
+
+ __m256 vd = _mm256_set1_ps(d);
+ __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+ acc = _mm256_add_ps(_mm256_mul_ps(vd, _mm256_cvtepi32_ps(sumi)), acc);
+
+ }
+
+ *s = hsum_float_8(acc) + summs;
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ UNUSED(kmask1);
+ UNUSED(kmask2);
+ UNUSED(kmask3);
+ UNUSED(utmp);
+ ggml_vec_dot_q5_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q6_K * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+#if defined __AVX2__
+
+ const __m256i m3 = _mm256_set1_epi8(3);
+ const __m256i m15 = _mm256_set1_epi8(15);
+
+ __m256 acc = _mm256_setzero_ps();
+
+ for (int i = 0; i < nb; ++i) {
+
+ const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+ const uint8_t * GGML_RESTRICT q4 = x[i].ql;
+ const uint8_t * GGML_RESTRICT qh = x[i].qh;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ const __m256i q8sums = _mm256_loadu_si256((const __m256i*)y[i].bsums);
+ const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+ const __m256i scales_16 = _mm256_cvtepi8_epi16(scales);
+ const __m256i q8sclsub = _mm256_slli_epi32(_mm256_madd_epi16(q8sums, scales_16), 5);
+
+ __m256i sumi = _mm256_setzero_si256();
+
+ int is = 0;
+
+ for (int j = 0; j < QK_K/128; ++j) {
+ const __m256i q4bits1 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
+ const __m256i q4bits2 = _mm256_loadu_si256((const __m256i*)q4); q4 += 32;
+ const __m256i q4bitsH = _mm256_loadu_si256((const __m256i*)qh); qh += 32;
+
+ const __m256i q4h_0 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, m3), 4);
+ const __m256i q4h_1 = _mm256_slli_epi16(_mm256_and_si256(q4bitsH, _mm256_set1_epi8(12)), 2);
+ const __m256i q4h_2 = _mm256_and_si256(q4bitsH, _mm256_set1_epi8(48));
+ const __m256i q4h_3 = _mm256_srli_epi16(_mm256_and_si256(q4bitsH, _mm256_set1_epi8(-64)), 2);
+
+ const __m256i q4_0 = _mm256_or_si256(_mm256_and_si256(q4bits1, m15), q4h_0);
+ const __m256i q4_1 = _mm256_or_si256(_mm256_and_si256(q4bits2, m15), q4h_1);
+ const __m256i q4_2 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits1, 4), m15), q4h_2);
+ const __m256i q4_3 = _mm256_or_si256(_mm256_and_si256(_mm256_srli_epi16(q4bits2, 4), m15), q4h_3);
+
+ const __m256i q8_0 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ const __m256i q8_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ const __m256i q8_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ const __m256i q8_3 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+ __m256i p16_0 = _mm256_maddubs_epi16(q4_0, q8_0);
+ __m256i p16_1 = _mm256_maddubs_epi16(q4_1, q8_1);
+ __m256i p16_2 = _mm256_maddubs_epi16(q4_2, q8_2);
+ __m256i p16_3 = _mm256_maddubs_epi16(q4_3, q8_3);
+
+ const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
+ const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
+ const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
+ const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
+ is += 4;
+
+ p16_0 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_0), p16_0);
+ p16_1 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_1), p16_1);
+ p16_2 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_2), p16_2);
+ p16_3 = _mm256_madd_epi16(_mm256_cvtepi8_epi16(scale_3), p16_3);
+
+ sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_0, p16_1));
+ sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p16_2, p16_3));
+
+ }
+
+ sumi = _mm256_sub_epi32(sumi, q8sclsub);
+ acc = _mm256_fmadd_ps(_mm256_broadcast_ss(&d), _mm256_cvtepi32_ps(sumi), acc);
+ }
+
+ *s = hsum_float_8(acc);
+
+#elif defined __AVX__
+
+ const __m128i m3 = _mm_set1_epi8(3);
+ const __m128i m15 = _mm_set1_epi8(15);
+
+ __m256 acc = _mm256_setzero_ps();
+
+ for (int i = 0; i < nb; ++i) {
+
+ const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+ const uint8_t * GGML_RESTRICT q4 = x[i].ql;
+ const uint8_t * GGML_RESTRICT qh = x[i].qh;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ // handle the q6_k -32 offset separately using bsums
+ const __m128i q8sums_0 = _mm_loadu_si128((const __m128i*)y[i].bsums);
+ const __m128i q8sums_1 = _mm_loadu_si128((const __m128i*)y[i].bsums + 1);
+ const __m128i scales = _mm_loadu_si128((const __m128i*)x[i].scales);
+ const __m128i scales_16_0 = _mm_cvtepi8_epi16(scales);
+ const __m128i scales_16_1 = _mm_cvtepi8_epi16(_mm_bsrli_si128(scales, 8));
+ const __m128i q8sclsub_0 = _mm_slli_epi32(_mm_madd_epi16(q8sums_0, scales_16_0), 5);
+ const __m128i q8sclsub_1 = _mm_slli_epi32(_mm_madd_epi16(q8sums_1, scales_16_1), 5);
+
+ __m128i sumi_0 = _mm_setzero_si128();
+ __m128i sumi_1 = _mm_setzero_si128();
+
+ int is = 0;
+
+ for (int j = 0; j < QK_K/128; ++j) {
+
+ const __m128i q4bitsH_0 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
+ const __m128i q4bitsH_1 = _mm_loadu_si128((const __m128i*)qh); qh += 16;
+
+ const __m128i q4h_0 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, m3), 4);
+ const __m128i q4h_1 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, m3), 4);
+ const __m128i q4h_2 = _mm_slli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(12)), 2);
+ const __m128i q4h_3 = _mm_slli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(12)), 2);
+ const __m128i q4h_4 = _mm_and_si128(q4bitsH_0, _mm_set1_epi8(48));
+ const __m128i q4h_5 = _mm_and_si128(q4bitsH_1, _mm_set1_epi8(48));
+ const __m128i q4h_6 = _mm_srli_epi16(_mm_and_si128(q4bitsH_0, _mm_set1_epi8(-64)), 2);
+ const __m128i q4h_7 = _mm_srli_epi16(_mm_and_si128(q4bitsH_1, _mm_set1_epi8(-64)), 2);
+
+ const __m128i q4bits1_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+ const __m128i q4bits1_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+ const __m128i q4bits2_0 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+ const __m128i q4bits2_1 = _mm_loadu_si128((const __m128i*)q4); q4 += 16;
+
+ const __m128i q4_0 = _mm_or_si128(_mm_and_si128(q4bits1_0, m15), q4h_0);
+ const __m128i q4_1 = _mm_or_si128(_mm_and_si128(q4bits1_1, m15), q4h_1);
+ const __m128i q4_2 = _mm_or_si128(_mm_and_si128(q4bits2_0, m15), q4h_2);
+ const __m128i q4_3 = _mm_or_si128(_mm_and_si128(q4bits2_1, m15), q4h_3);
+ const __m128i q4_4 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_0, 4), m15), q4h_4);
+ const __m128i q4_5 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits1_1, 4), m15), q4h_5);
+ const __m128i q4_6 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_0, 4), m15), q4h_6);
+ const __m128i q4_7 = _mm_or_si128(_mm_and_si128(_mm_srli_epi16(q4bits2_1, 4), m15), q4h_7);
+
+ const __m128i q8_0 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_1 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_2 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_3 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_4 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_5 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_6 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+ const __m128i q8_7 = _mm_loadu_si128((const __m128i*)q8); q8 += 16;
+
+ __m128i p16_0 = _mm_maddubs_epi16(q4_0, q8_0);
+ __m128i p16_1 = _mm_maddubs_epi16(q4_1, q8_1);
+ __m128i p16_2 = _mm_maddubs_epi16(q4_2, q8_2);
+ __m128i p16_3 = _mm_maddubs_epi16(q4_3, q8_3);
+ __m128i p16_4 = _mm_maddubs_epi16(q4_4, q8_4);
+ __m128i p16_5 = _mm_maddubs_epi16(q4_5, q8_5);
+ __m128i p16_6 = _mm_maddubs_epi16(q4_6, q8_6);
+ __m128i p16_7 = _mm_maddubs_epi16(q4_7, q8_7);
+
+ const __m128i scale_0 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 0));
+ const __m128i scale_1 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 1));
+ const __m128i scale_2 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 2));
+ const __m128i scale_3 = _mm_shuffle_epi8(scales, get_scale_shuffle(is + 3));
+ is += 4;
+
+ p16_0 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_0), p16_0);
+ p16_1 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_0, 8)), p16_1);
+ p16_2 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_1), p16_2);
+ p16_3 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_1, 8)), p16_3);
+ p16_4 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_2), p16_4);
+ p16_5 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_2, 8)), p16_5);
+ p16_6 = _mm_madd_epi16(_mm_cvtepi8_epi16(scale_3), p16_6);
+ p16_7 = _mm_madd_epi16(_mm_cvtepi8_epi16(_mm_bsrli_si128(scale_3, 8)), p16_7);
+
+ sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_0, p16_2));
+ sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_1, p16_3));
+ sumi_0 = _mm_add_epi32(sumi_0, _mm_add_epi32(p16_4, p16_6));
+ sumi_1 = _mm_add_epi32(sumi_1, _mm_add_epi32(p16_5, p16_7));
+
+ }
+
+ sumi_0 = _mm_sub_epi32(sumi_0, q8sclsub_0);
+ sumi_1 = _mm_sub_epi32(sumi_1, q8sclsub_1);
+ const __m256i sumi = MM256_SET_M128I(sumi_1, sumi_0);
+ acc = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi)), acc);
+ }
+
+ *s = hsum_float_8(acc);
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ ggml_vec_dot_q6_K_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+#if defined (__AVX__) || defined (__AVX2__)
+static const int8_t keven_signs_q2xs[1024] = {
+ 1, 1, 1, 1, 1, 1, 1, 1, -1, 1, 1, 1, 1, 1, 1, -1, 1, -1, 1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, 1, 1, 1,
+ 1, 1, -1, 1, 1, 1, 1, -1, -1, 1, -1, 1, 1, 1, 1, 1, 1, -1, -1, 1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, 1, -1,
+ 1, 1, 1, -1, 1, 1, 1, -1, -1, 1, 1, -1, 1, 1, 1, 1, 1, -1, 1, -1, 1, 1, 1, 1, -1, -1, 1, -1, 1, 1, 1, -1,
+ 1, 1, -1, -1, 1, 1, 1, 1, -1, 1, -1, -1, 1, 1, 1, -1, 1, -1, -1, -1, 1, 1, 1, -1, -1, -1, -1, -1, 1, 1, 1, 1,
+ 1, 1, 1, 1, -1, 1, 1, -1, -1, 1, 1, 1, -1, 1, 1, 1, 1, -1, 1, 1, -1, 1, 1, 1, -1, -1, 1, 1, -1, 1, 1, -1,
+ 1, 1, -1, 1, -1, 1, 1, 1, -1, 1, -1, 1, -1, 1, 1, -1, 1, -1, -1, 1, -1, 1, 1, -1, -1, -1, -1, 1, -1, 1, 1, 1,
+ 1, 1, 1, -1, -1, 1, 1, 1, -1, 1, 1, -1, -1, 1, 1, -1, 1, -1, 1, -1, -1, 1, 1, -1, -1, -1, 1, -1, -1, 1, 1, 1,
+ 1, 1, -1, -1, -1, 1, 1, -1, -1, 1, -1, -1, -1, 1, 1, 1, 1, -1, -1, -1, -1, 1, 1, 1, -1, -1, -1, -1, -1, 1, 1, -1,
+ 1, 1, 1, 1, 1, -1, 1, -1, -1, 1, 1, 1, 1, -1, 1, 1, 1, -1, 1, 1, 1, -1, 1, 1, -1, -1, 1, 1, 1, -1, 1, -1,
+ 1, 1, -1, 1, 1, -1, 1, 1, -1, 1, -1, 1, 1, -1, 1, -1, 1, -1, -1, 1, 1, -1, 1, -1, -1, -1, -1, 1, 1, -1, 1, 1,
+ 1, 1, 1, -1, 1, -1, 1, 1, -1, 1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, 1, -1, -1, -1, 1, -1, 1, -1, 1, 1,
+ 1, 1, -1, -1, 1, -1, 1, -1, -1, 1, -1, -1, 1, -1, 1, 1, 1, -1, -1, -1, 1, -1, 1, 1, -1, -1, -1, -1, 1, -1, 1, -1,
+ 1, 1, 1, 1, -1, -1, 1, 1, -1, 1, 1, 1, -1, -1, 1, -1, 1, -1, 1, 1, -1, -1, 1, -1, -1, -1, 1, 1, -1, -1, 1, 1,
+ 1, 1, -1, 1, -1, -1, 1, -1, -1, 1, -1, 1, -1, -1, 1, 1, 1, -1, -1, 1, -1, -1, 1, 1, -1, -1, -1, 1, -1, -1, 1, -1,
+ 1, 1, 1, -1, -1, -1, 1, -1, -1, 1, 1, -1, -1, -1, 1, 1, 1, -1, 1, -1, -1, -1, 1, 1, -1, -1, 1, -1, -1, -1, 1, -1,
+ 1, 1, -1, -1, -1, -1, 1, 1, -1, 1, -1, -1, -1, -1, 1, -1, 1, -1, -1, -1, -1, -1, 1, -1, -1, -1, -1, -1, -1, -1, 1, 1,
+ 1, 1, 1, 1, 1, 1, -1, -1, -1, 1, 1, 1, 1, 1, -1, 1, 1, -1, 1, 1, 1, 1, -1, 1, -1, -1, 1, 1, 1, 1, -1, -1,
+ 1, 1, -1, 1, 1, 1, -1, 1, -1, 1, -1, 1, 1, 1, -1, -1, 1, -1, -1, 1, 1, 1, -1, -1, -1, -1, -1, 1, 1, 1, -1, 1,
+ 1, 1, 1, -1, 1, 1, -1, 1, -1, 1, 1, -1, 1, 1, -1, -1, 1, -1, 1, -1, 1, 1, -1, -1, -1, -1, 1, -1, 1, 1, -1, 1,
+ 1, 1, -1, -1, 1, 1, -1, -1, -1, 1, -1, -1, 1, 1, -1, 1, 1, -1, -1, -1, 1, 1, -1, 1, -1, -1, -1, -1, 1, 1, -1, -1,
+ 1, 1, 1, 1, -1, 1, -1, 1, -1, 1, 1, 1, -1, 1, -1, -1, 1, -1, 1, 1, -1, 1, -1, -1, -1, -1, 1, 1, -1, 1, -1, 1,
+ 1, 1, -1, 1, -1, 1, -1, -1, -1, 1, -1, 1, -1, 1, -1, 1, 1, -1, -1, 1, -1, 1, -1, 1, -1, -1, -1, 1, -1, 1, -1, -1,
+ 1, 1, 1, -1, -1, 1, -1, -1, -1, 1, 1, -1, -1, 1, -1, 1, 1, -1, 1, -1, -1, 1, -1, 1, -1, -1, 1, -1, -1, 1, -1, -1,
+ 1, 1, -1, -1, -1, 1, -1, 1, -1, 1, -1, -1, -1, 1, -1, -1, 1, -1, -1, -1, -1, 1, -1, -1, -1, -1, -1, -1, -1, 1, -1, 1,
+ 1, 1, 1, 1, 1, -1, -1, 1, -1, 1, 1, 1, 1, -1, -1, -1, 1, -1, 1, 1, 1, -1, -1, -1, -1, -1, 1, 1, 1, -1, -1, 1,
+ 1, 1, -1, 1, 1, -1, -1, -1, -1, 1, -1, 1, 1, -1, -1, 1, 1, -1, -1, 1, 1, -1, -1, 1, -1, -1, -1, 1, 1, -1, -1, -1,
+ 1, 1, 1, -1, 1, -1, -1, -1, -1, 1, 1, -1, 1, -1, -1, 1, 1, -1, 1, -1, 1, -1, -1, 1, -1, -1, 1, -1, 1, -1, -1, -1,
+ 1, 1, -1, -1, 1, -1, -1, 1, -1, 1, -1, -1, 1, -1, -1, -1, 1, -1, -1, -1, 1, -1, -1, -1, -1, -1, -1, -1, 1, -1, -1, 1,
+ 1, 1, 1, 1, -1, -1, -1, -1, -1, 1, 1, 1, -1, -1, -1, 1, 1, -1, 1, 1, -1, -1, -1, 1, -1, -1, 1, 1, -1, -1, -1, -1,
+ 1, 1, -1, 1, -1, -1, -1, 1, -1, 1, -1, 1, -1, -1, -1, -1, 1, -1, -1, 1, -1, -1, -1, -1, -1, -1, -1, 1, -1, -1, -1, 1,
+ 1, 1, 1, -1, -1, -1, -1, 1, -1, 1, 1, -1, -1, -1, -1, -1, 1, -1, 1, -1, -1, -1, -1, -1, -1, -1, 1, -1, -1, -1, -1, 1,
+ 1, 1, -1, -1, -1, -1, -1, -1, -1, 1, -1, -1, -1, -1, -1, 1, 1, -1, -1, -1, -1, -1, -1, 1, -1, -1, -1, -1, -1, -1, -1, -1,
+};
+#endif
+
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq2_xxs * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+
+ const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+ uint32_t aux32[4];
+ const uint8_t * aux8 = (const uint8_t *)aux32;
+
+ __m256 accumf = _mm256_setzero_ps();
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ __m256i sumi1 = _mm256_setzero_si256();
+ __m256i sumi2 = _mm256_setzero_si256();
+ for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+ const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+ const __m256i q2_1 = _mm256_set_epi64x(iq2xxs_grid[aux8[ 3]], iq2xxs_grid[aux8[ 2]], iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+ const __m256i q2_2 = _mm256_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]], iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+ const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+ signs64[(aux32[1] >> 7) & 127], signs64[(aux32[1] >> 0) & 127]);
+ const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127],
+ signs64[(aux32[3] >> 7) & 127], signs64[(aux32[3] >> 0) & 127]);
+ const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
+ const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
+ const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1);
+ const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2);
+ const uint16_t ls1 = aux32[1] >> 28;
+ const uint16_t ls2 = aux32[3] >> 28;
+ const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+ const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+ sumi1 = _mm256_add_epi32(sumi1, p1);
+ sumi2 = _mm256_add_epi32(sumi2, p2);
+ }
+
+ accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+ }
+
+ *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+ const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+ uint32_t aux32[4];
+ const uint8_t * aux8 = (const uint8_t *)aux32;
+
+ __m256 accumf = _mm256_setzero_ps();
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ __m128i sumi1_0 = _mm_setzero_si128();
+ __m128i sumi1_1 = _mm_setzero_si128();
+ __m128i sumi2_0 = _mm_setzero_si128();
+ __m128i sumi2_1 = _mm_setzero_si128();
+ for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+ const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ memcpy(aux32, q2, 4*sizeof(uint32_t)); q2 += 8;
+ const __m128i q2_1_0 = _mm_set_epi64x(iq2xxs_grid[aux8[1]], iq2xxs_grid[aux8[0]]);
+ const __m128i q2_1_1 = _mm_set_epi64x(iq2xxs_grid[aux8[3]], iq2xxs_grid[aux8[2]]);
+ const __m128i q2_2_0 = _mm_set_epi64x(iq2xxs_grid[aux8[9]], iq2xxs_grid[aux8[8]]);
+ const __m128i q2_2_1 = _mm_set_epi64x(iq2xxs_grid[aux8[11]], iq2xxs_grid[aux8[10]]);
+ const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[1] >> 7) & 127], signs64[(aux32[1] >> 0) & 127]);
+ const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
+ const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[3] >> 7) & 127], signs64[(aux32[3] >> 0) & 127]);
+ const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[3] >> 21) & 127], signs64[(aux32[3] >> 14) & 127]);
+ const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
+ const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
+ const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
+ const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
+ const __m128i dot1_0 = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+ const __m128i dot1_1 = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+ const __m128i dot2_0 = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+ const __m128i dot2_1 = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+ const uint16_t ls1 = aux32[1] >> 28;
+ const uint16_t ls2 = aux32[3] >> 28;
+ const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+ const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+ const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+ const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+ sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+ sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+ sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+ sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+ }
+
+ accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+ }
+
+ *s = 0.125f * hsum_float_8(accumf);
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ ggml_vec_dot_iq2_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_iq2_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq2_xs * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+
+ const __m256i mone = _mm256_set1_epi8(1);
+ static const char block_sign_shuffle_mask_1[32] = {
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+ 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+ };
+ static const char block_sign_shuffle_mask_2[32] = {
+ 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+ 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+ };
+ static const uint8_t bit_selector_mask_bytes[32] = {
+ 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+ 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+ };
+
+ const __m256i bit_selector_mask = _mm256_loadu_si256((const __m256i*)bit_selector_mask_bytes);
+ const __m256i block_sign_shuffle_1 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_1);
+ const __m256i block_sign_shuffle_2 = _mm256_loadu_si256((const __m256i*)block_sign_shuffle_mask_2);
+
+ static const uint8_t k_bit_helper[32] = {
+ 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+ 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+ };
+ const __m256i bit_helper = _mm256_loadu_si256((const __m256i*)k_bit_helper);
+ const __m256i m511 = _mm256_set1_epi16(511);
+ const __m128i m4 = _mm_set1_epi8(0xf);
+ const __m128i m1 = _mm_set1_epi8(1);
+
+ uint64_t aux64;
+
+ // somewhat hacky, but gives a significant boost in performance
+ __m256i aux_gindex;
+ const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+ __m256 accumf = _mm256_setzero_ps();
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ memcpy(&aux64, x[i].scales, 8);
+ __m128i stmp = _mm_set1_epi64x(aux64);
+ stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
+ const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
+
+ __m256i sumi1 = _mm256_setzero_si256();
+ __m256i sumi2 = _mm256_setzero_si256();
+ for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+ const __m256i q2_data = _mm256_loadu_si256((const __m256i*)q2); q2 += 16;
+ aux_gindex = _mm256_and_si256(q2_data, m511);
+
+ const __m256i partial_sign_bits = _mm256_srli_epi16(q2_data, 9);
+ const __m256i partial_sign_bits_upper = _mm256_srli_epi16(q2_data, 13);
+ const __m256i partial_sign_bits_for_counting = _mm256_xor_si256(partial_sign_bits, partial_sign_bits_upper);
+
+ const __m256i odd_bits = _mm256_shuffle_epi8(bit_helper, partial_sign_bits_for_counting);
+ const __m256i full_sign_bits = _mm256_or_si256(partial_sign_bits, odd_bits);
+
+ const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ const __m256i q8_3 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ const __m256i q8_4 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+
+ const __m256i q2_1 = _mm256_set_epi64x(iq2xs_grid[gindex[ 3]], iq2xs_grid[gindex[ 2]],
+ iq2xs_grid[gindex[ 1]], iq2xs_grid[gindex[ 0]]);
+ const __m256i q2_2 = _mm256_set_epi64x(iq2xs_grid[gindex[ 7]], iq2xs_grid[gindex[ 6]],
+ iq2xs_grid[gindex[ 5]], iq2xs_grid[gindex[ 4]]);
+ const __m256i q2_3 = _mm256_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]],
+ iq2xs_grid[gindex[ 9]], iq2xs_grid[gindex[ 8]]);
+ const __m256i q2_4 = _mm256_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]],
+ iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+
+ const __m128i full_signs_l = _mm256_castsi256_si128(full_sign_bits);
+ const __m128i full_signs_h = _mm256_extractf128_si256(full_sign_bits, 1);
+ const __m256i full_signs_1 = MM256_SET_M128I(full_signs_l, full_signs_l);
+ const __m256i full_signs_2 = MM256_SET_M128I(full_signs_h, full_signs_h);
+
+ __m256i signs;
+ signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_1);
+ signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+ const __m256i q8s_1 = _mm256_sign_epi8(q8_1, _mm256_or_si256(signs, mone));
+
+ signs = _mm256_shuffle_epi8(full_signs_1, block_sign_shuffle_2);
+ signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+ const __m256i q8s_2 = _mm256_sign_epi8(q8_2, _mm256_or_si256(signs, mone));
+
+ signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_1);
+ signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+ const __m256i q8s_3 = _mm256_sign_epi8(q8_3, _mm256_or_si256(signs, mone));
+
+ signs = _mm256_shuffle_epi8(full_signs_2, block_sign_shuffle_2);
+ signs = _mm256_cmpeq_epi8(_mm256_and_si256(signs, bit_selector_mask), bit_selector_mask);
+ const __m256i q8s_4 = _mm256_sign_epi8(q8_4, _mm256_or_si256(signs, mone));
+
+ const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1);
+ const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2);
+ const __m256i dot3 = _mm256_maddubs_epi16(q2_3, q8s_3);
+ const __m256i dot4 = _mm256_maddubs_epi16(q2_4, q8s_4);
+
+ const __m256i sc1 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0)));
+ const __m256i sc2 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1)));
+ const __m256i sc3 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2)));
+ const __m256i sc4 = _mm256_cvtepi8_epi16(_mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3)));
+
+ sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot1, sc1));
+ sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot2, sc2));
+ sumi1 = _mm256_add_epi32(sumi1, _mm256_madd_epi16(dot3, sc3));
+ sumi2 = _mm256_add_epi32(sumi2, _mm256_madd_epi16(dot4, sc4));
+ }
+
+ accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+ }
+
+ *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+ const __m128i mone = _mm_set1_epi8(1);
+ static const char block_sign_shuffle_mask_1[32] = {
+ 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02,
+ 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06,
+ };
+ static const char block_sign_shuffle_mask_2[32] = {
+ 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a, 0x0a,
+ 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0c, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e, 0x0e,
+ };
+ static const uint8_t bit_selector_mask_bytes[32] = {
+ 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+ 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+ };
+
+ const __m128i bit_selector_mask_0 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes);
+ const __m128i bit_selector_mask_1 = _mm_loadu_si128((const __m128i*)bit_selector_mask_bytes + 1);
+ const __m128i block_sign_shuffle_1_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1);
+ const __m128i block_sign_shuffle_1_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_1 + 1);
+ const __m128i block_sign_shuffle_2_0 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2);
+ const __m128i block_sign_shuffle_2_1 = _mm_loadu_si128((const __m128i*)block_sign_shuffle_mask_2 + 1);
+
+ static const uint8_t k_bit_helper[32] = {
+ 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+ 0x00, 0x80, 0x80, 0x00, 0x80, 0x00, 0x00, 0x80, 0x80, 0x00, 0x00, 0x80, 0x00, 0x80, 0x80, 0x00,
+ };
+ const __m128i bit_helper_0 = _mm_loadu_si128((const __m128i*)k_bit_helper);
+ const __m128i bit_helper_1 = _mm_loadu_si128((const __m128i*)k_bit_helper + 1);
+ const __m128i m511 = _mm_set1_epi16(511);
+ const __m128i m4 = _mm_set1_epi8(0xf);
+ const __m128i m1 = _mm_set1_epi8(1);
+
+ uint64_t aux64;
+
+ // somewhat hacky, but gives a significant boost in performance
+ __m256i aux_gindex;
+ const uint16_t * gindex = (const uint16_t *)&aux_gindex;
+
+ __m256 accumf = _mm256_setzero_ps();
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ memcpy(&aux64, x[i].scales, 8);
+ __m128i stmp = _mm_set1_epi64x(aux64);
+ stmp = _mm_unpacklo_epi8(_mm_and_si128(stmp, m4), _mm_and_si128(_mm_srli_epi16(stmp, 4), m4));
+ const __m128i scales = _mm_add_epi8(_mm_slli_epi16(stmp, 1), m1);
+
+ __m128i sumi1_0 = _mm_setzero_si128();
+ __m128i sumi1_1 = _mm_setzero_si128();
+ __m128i sumi2_0 = _mm_setzero_si128();
+ __m128i sumi2_1 = _mm_setzero_si128();
+ for (int ib32 = 0; ib32 < QK_K/32; ib32 += 4) {
+
+ const __m128i q2_data_0 = _mm_loadu_si128((const __m128i*)q2);
+ const __m128i q2_data_1 = _mm_loadu_si128((const __m128i*)q2 + 1); q2 += 16;
+ aux_gindex = MM256_SET_M128I(_mm_and_si128(q2_data_1, m511), _mm_and_si128(q2_data_0, m511));
+
+ const __m128i partial_sign_bits_0 = _mm_srli_epi16(q2_data_0, 9);
+ const __m128i partial_sign_bits_1 = _mm_srli_epi16(q2_data_1, 9);
+ const __m128i partial_sign_bits_upper_0 = _mm_srli_epi16(q2_data_0, 13);
+ const __m128i partial_sign_bits_upper_1 = _mm_srli_epi16(q2_data_1, 13);
+ const __m128i partial_sign_bits_for_counting_0 = _mm_xor_si128(partial_sign_bits_0, partial_sign_bits_upper_0);
+ const __m128i partial_sign_bits_for_counting_1 = _mm_xor_si128(partial_sign_bits_1, partial_sign_bits_upper_1);
+
+ const __m128i odd_bits_0 = _mm_shuffle_epi8(bit_helper_0, partial_sign_bits_for_counting_0);
+ const __m128i odd_bits_1 = _mm_shuffle_epi8(bit_helper_1, partial_sign_bits_for_counting_1);
+ const __m128i full_sign_bits_0 = _mm_or_si128(partial_sign_bits_0, odd_bits_0);
+ const __m128i full_sign_bits_1 = _mm_or_si128(partial_sign_bits_1, odd_bits_1);
+
+ const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_3_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_3_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_4_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_4_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+ const __m128i q2_1_0 = _mm_set_epi64x(iq2xs_grid[gindex[1]], iq2xs_grid[gindex[0]]);
+ const __m128i q2_1_1 = _mm_set_epi64x(iq2xs_grid[gindex[3]], iq2xs_grid[gindex[2]]);
+ const __m128i q2_2_0 = _mm_set_epi64x(iq2xs_grid[gindex[5]], iq2xs_grid[gindex[4]]);
+ const __m128i q2_2_1 = _mm_set_epi64x(iq2xs_grid[gindex[7]], iq2xs_grid[gindex[6]]);
+ const __m128i q2_3_0 = _mm_set_epi64x(iq2xs_grid[gindex[9]], iq2xs_grid[gindex[8]]);
+ const __m128i q2_3_1 = _mm_set_epi64x(iq2xs_grid[gindex[11]], iq2xs_grid[gindex[10]]);
+ const __m128i q2_4_0 = _mm_set_epi64x(iq2xs_grid[gindex[13]], iq2xs_grid[gindex[12]]);
+ const __m128i q2_4_1 = _mm_set_epi64x(iq2xs_grid[gindex[15]], iq2xs_grid[gindex[14]]);
+
+ // AVX2 full_signs_1 is full_sign_bits_0 here
+ // AVX2 full_signs_2 is full_sign_bits_1 here
+ __m128i signs_0, signs_1;
+ signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_0);
+ signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_1_1);
+ signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+ signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+ const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, _mm_or_si128(signs_0, mone));
+ const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, _mm_or_si128(signs_1, mone));
+
+ signs_0 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_0);
+ signs_1 = _mm_shuffle_epi8(full_sign_bits_0, block_sign_shuffle_2_1);
+ signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+ signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+ const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, _mm_or_si128(signs_0, mone));
+ const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, _mm_or_si128(signs_1, mone));
+
+ signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_0);
+ signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_1_1);
+ signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+ signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+ const __m128i q8s_3_0 = _mm_sign_epi8(q8_3_0, _mm_or_si128(signs_0, mone));
+ const __m128i q8s_3_1 = _mm_sign_epi8(q8_3_1, _mm_or_si128(signs_1, mone));
+
+ signs_0 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_0);
+ signs_1 = _mm_shuffle_epi8(full_sign_bits_1, block_sign_shuffle_2_1);
+ signs_0 = _mm_cmpeq_epi8(_mm_and_si128(signs_0, bit_selector_mask_0), bit_selector_mask_0);
+ signs_1 = _mm_cmpeq_epi8(_mm_and_si128(signs_1, bit_selector_mask_1), bit_selector_mask_1);
+ const __m128i q8s_4_0 = _mm_sign_epi8(q8_4_0, _mm_or_si128(signs_0, mone));
+ const __m128i q8s_4_1 = _mm_sign_epi8(q8_4_1, _mm_or_si128(signs_1, mone));
+
+ const __m128i dot1_0 = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+ const __m128i dot1_1 = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+ const __m128i dot2_0 = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+ const __m128i dot2_1 = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+ const __m128i dot3_0 = _mm_maddubs_epi16(q2_3_0, q8s_3_0);
+ const __m128i dot3_1 = _mm_maddubs_epi16(q2_3_1, q8s_3_1);
+ const __m128i dot4_0 = _mm_maddubs_epi16(q2_4_0, q8s_4_0);
+ const __m128i dot4_1 = _mm_maddubs_epi16(q2_4_1, q8s_4_1);
+
+ __m128i sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+0));
+ const __m128i sc1_0 = _mm_cvtepi8_epi16(sc_tmp);
+ const __m128i sc1_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+ sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+1));
+ const __m128i sc2_0 = _mm_cvtepi8_epi16(sc_tmp);
+ const __m128i sc2_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+ sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+2));
+ const __m128i sc3_0 = _mm_cvtepi8_epi16(sc_tmp);
+ const __m128i sc3_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+ sc_tmp = _mm_shuffle_epi8(scales, get_scale_shuffle(ib32+3));
+ const __m128i sc4_0 = _mm_cvtepi8_epi16(sc_tmp);
+ const __m128i sc4_1 = _mm_cvtepi8_epi16(_mm_srli_si128(sc_tmp, 8));
+
+ sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot1_0, sc1_0));
+ sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot1_1, sc1_1));
+ sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot2_0, sc2_0));
+ sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot2_1, sc2_1));
+ sumi1_0 = _mm_add_epi32(sumi1_0, _mm_madd_epi16(dot3_0, sc3_0));
+ sumi1_1 = _mm_add_epi32(sumi1_1, _mm_madd_epi16(dot3_1, sc3_1));
+ sumi2_0 = _mm_add_epi32(sumi2_0, _mm_madd_epi16(dot4_0, sc4_0));
+ sumi2_1 = _mm_add_epi32(sumi2_1, _mm_madd_epi16(dot4_1, sc4_1));
+ }
+
+ accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+ }
+
+ *s = 0.125f * hsum_float_8(accumf);
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ ggml_vec_dot_iq2_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_iq2_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq2_s * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+
+ static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+ 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+ };
+
+ static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+ 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+ };
+
+ const __m128i m4 = _mm_set1_epi8(0xf);
+ const __m128i m1 = _mm_set1_epi8(1);
+
+ const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
+ const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
+
+ uint64_t aux64;
+
+ __m256 accumf = _mm256_setzero_ps();
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint8_t * GGML_RESTRICT qs = x[i].qs;
+ const uint8_t * GGML_RESTRICT qh = x[i].qh;
+ const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ memcpy(&aux64, x[i].scales, 8);
+ const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1);
+ const __m256i scales16 = _mm256_cvtepi8_epi16(scales8); // 0 2 4 6 8 10 12 14 1 3 5 7 9 11 13 15
+
+ __m256i sumi1 = _mm256_setzero_si256();
+ __m256i sumi2 = _mm256_setzero_si256();
+ for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+ const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ const __m256i q2_1 = _mm256_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+ iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)],
+ iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+ iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+ const __m256i q2_2 = _mm256_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+ iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)],
+ iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+ iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+ qs += 8;
+
+ __m256i aux256 = _mm256_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
+ aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+ const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
+ const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
+
+ aux256 = _mm256_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
+ aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+ const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
+ const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
+
+ signs += 4;
+
+ const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1); // blocks 2*ib32+0, 2*ib32+1
+ const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2); // blocks 2*ib32+2, 2*ib32+3
+
+ const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+0)));
+ const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_shuffle_epi8(scales16, get_scale_shuffle_k4(ib32+1)));
+ sumi1 = _mm256_add_epi32(sumi1, p1);
+ sumi2 = _mm256_add_epi32(sumi2, p2);
+ }
+
+ accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+ }
+
+ *s = 0.125f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+ static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+ 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+ };
+
+ static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+ 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+ };
+
+ const __m128i m4 = _mm_set1_epi8(0xf);
+ const __m128i m1 = _mm_set1_epi8(1);
+
+ const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
+ const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
+ const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
+ const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
+
+ uint64_t aux64;
+
+ __m256 accumf = _mm256_setzero_ps();
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint8_t * GGML_RESTRICT qs = x[i].qs;
+ const uint8_t * GGML_RESTRICT qh = x[i].qh;
+ const uint16_t * GGML_RESTRICT signs = (const uint16_t *)(x[i].qs + QK_K/8);
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+
+ memcpy(&aux64, x[i].scales, 8);
+ const __m128i scales8 = _mm_add_epi8(_mm_slli_epi16(_mm_and_si128(_mm_set_epi64x(aux64 >> 4, aux64), m4), 1), m1);
+ const __m128i scales16_0 = _mm_cvtepi8_epi16(scales8);
+ const __m128i scales16_1 = _mm_cvtepi8_epi16(_mm_srli_si128(scales8, 8));
+
+ __m128i sumi1_0 = _mm_setzero_si128();
+ __m128i sumi1_1 = _mm_setzero_si128();
+ __m128i sumi2_0 = _mm_setzero_si128();
+ __m128i sumi2_1 = _mm_setzero_si128();
+ for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+ const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q2_1_0 = _mm_set_epi64x(iq2s_grid[qs[1] | ((qh[ib32+0] << 6) & 0x300)],
+ iq2s_grid[qs[0] | ((qh[ib32+0] << 8) & 0x300)]);
+ const __m128i q2_1_1 = _mm_set_epi64x(iq2s_grid[qs[3] | ((qh[ib32+0] << 2) & 0x300)],
+ iq2s_grid[qs[2] | ((qh[ib32+0] << 4) & 0x300)]);
+ const __m128i q2_2_0 = _mm_set_epi64x(iq2s_grid[qs[5] | ((qh[ib32+1] << 6) & 0x300)],
+ iq2s_grid[qs[4] | ((qh[ib32+1] << 8) & 0x300)]);
+ const __m128i q2_2_1 = _mm_set_epi64x(iq2s_grid[qs[7] | ((qh[ib32+1] << 2) & 0x300)],
+ iq2s_grid[qs[6] | ((qh[ib32+1] << 4) & 0x300)]);
+ qs += 8;
+
+ __m128i aux128_0 = _mm_set1_epi32(signs[0] | ((uint32_t) signs[1] << 16));
+ __m128i aux128_1 = aux128_0;
+ aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+ aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+ const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+ const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+ const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
+ const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
+
+ aux128_0 = _mm_set1_epi32(signs[2] | ((uint32_t) signs[3] << 16));
+ aux128_1 = aux128_0;
+ aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+ aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+ const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+ const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+ const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
+ const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
+
+ signs += 4;
+
+ const __m128i dot1_0 = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+ const __m128i dot1_1 = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+ const __m128i dot2_0 = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+ const __m128i dot2_1 = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+
+ const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 0)));
+ const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+0), 1)));
+ const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_shuffle_epi8(scales16_0, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 0)));
+ const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_shuffle_epi8(scales16_1, _mm256_extractf128_si256(get_scale_shuffle_k4(ib32+1), 1)));
+ sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+ sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+ sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+ sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+ }
+
+ accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+ }
+
+ *s = 0.125f * hsum_float_8(accumf);
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ ggml_vec_dot_iq2_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq3_xxs * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+
+ const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+ uint32_t aux32[2];
+
+ __m256 accumf = _mm256_setzero_ps();
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+ const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ __m256i sumi1 = _mm256_setzero_si256();
+ __m256i sumi2 = _mm256_setzero_si256();
+ for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+ const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ const __m256i q2_1 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+ iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+ q3 += 8;
+ const __m256i q2_2 = _mm256_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]],
+ iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+ q3 += 8;
+ memcpy(aux32, gas, 8); gas += 8;
+ const __m256i s2_1 = _mm256_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127],
+ signs64[(aux32[0] >> 7) & 127], signs64[(aux32[0] >> 0) & 127]);
+ const __m256i s2_2 = _mm256_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127],
+ signs64[(aux32[1] >> 7) & 127], signs64[(aux32[1] >> 0) & 127]);
+ const __m256i q8s_1 = _mm256_sign_epi8(q8_1, s2_1);
+ const __m256i q8s_2 = _mm256_sign_epi8(q8_2, s2_2);
+ const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1);
+ const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2);
+ const uint16_t ls1 = aux32[0] >> 28;
+ const uint16_t ls2 = aux32[1] >> 28;
+ const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+ const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+ sumi1 = _mm256_add_epi32(sumi1, p1);
+ sumi2 = _mm256_add_epi32(sumi2, p2);
+ }
+
+ accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+ }
+
+ *s = 0.25f * hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+ const uint64_t * signs64 = (const uint64_t *)keven_signs_q2xs;
+
+ uint32_t aux32[2];
+
+ __m256 accumf = _mm256_setzero_ps();
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+ const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ __m128i sumi1_0 = _mm_setzero_si128();
+ __m128i sumi1_1 = _mm_setzero_si128();
+ __m128i sumi2_0 = _mm_setzero_si128();
+ __m128i sumi2_1 = _mm_setzero_si128();
+ for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+ const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q2_1_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+ const __m128i q2_1_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
+ q3 += 8;
+ const __m128i q2_2_0 = _mm_set_epi32(iq3xxs_grid[q3[3]], iq3xxs_grid[q3[2]], iq3xxs_grid[q3[1]], iq3xxs_grid[q3[0]]);
+ const __m128i q2_2_1 = _mm_set_epi32(iq3xxs_grid[q3[7]], iq3xxs_grid[q3[6]], iq3xxs_grid[q3[5]], iq3xxs_grid[q3[4]]);
+ q3 += 8;
+ memcpy(aux32, gas, 8); gas += 8;
+ const __m128i s2_1_0 = _mm_set_epi64x(signs64[(aux32[0] >> 7) & 127], signs64[(aux32[0] >> 0) & 127]);
+ const __m128i s2_1_1 = _mm_set_epi64x(signs64[(aux32[0] >> 21) & 127], signs64[(aux32[0] >> 14) & 127]);
+ const __m128i s2_2_0 = _mm_set_epi64x(signs64[(aux32[1] >> 7) & 127], signs64[(aux32[1] >> 0) & 127]);
+ const __m128i s2_2_1 = _mm_set_epi64x(signs64[(aux32[1] >> 21) & 127], signs64[(aux32[1] >> 14) & 127]);
+ const __m128i q8s_1_0 = _mm_sign_epi8(q8_1_0, s2_1_0);
+ const __m128i q8s_1_1 = _mm_sign_epi8(q8_1_1, s2_1_1);
+ const __m128i q8s_2_0 = _mm_sign_epi8(q8_2_0, s2_2_0);
+ const __m128i q8s_2_1 = _mm_sign_epi8(q8_2_1, s2_2_1);
+ const __m128i dot1_0 = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+ const __m128i dot1_1 = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+ const __m128i dot2_0 = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+ const __m128i dot2_1 = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+ const uint16_t ls1 = aux32[0] >> 28;
+ const uint16_t ls2 = aux32[1] >> 28;
+ const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+ const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+ const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+ const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+ sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+ sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+ sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+ sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+ }
+
+ accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+ }
+
+ *s = 0.25f * hsum_float_8(accumf);
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ ggml_vec_dot_iq3_xxs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_iq3_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq3_s * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+#if defined(__AVX2__)
+
+ static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+ 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+ };
+
+ static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+ 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+ };
+
+ const __m256i mask1 = _mm256_loadu_si256((const __m256i*)k_mask1);
+ const __m256i mask2 = _mm256_loadu_si256((const __m256i*)k_mask2);
+
+ const __m256i idx_shift = _mm256_set_epi32(1, 2, 3, 4, 5, 6, 7, 8);
+ const __m256i idx_mask = _mm256_set1_epi32(256);
+
+ typedef union {
+ __m256i vec[2];
+ uint32_t index[16];
+ } index_t;
+
+ index_t idx;
+
+ __m256 accumf = _mm256_setzero_ps();
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint8_t * GGML_RESTRICT qs = x[i].qs;
+ const uint8_t * GGML_RESTRICT qh = x[i].qh;
+ const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ __m256i sumi1 = _mm256_setzero_si256();
+ __m256i sumi2 = _mm256_setzero_si256();
+ for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+ const __m256i q8_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ const __m256i q8_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ const __m256i idx_l = _mm256_cvtepu8_epi16(_mm_loadu_si128((const __m128i *)qs)); qs += 16;
+ idx.vec[0] = _mm256_set1_epi32(qh[ib32+0]);
+ idx.vec[1] = _mm256_set1_epi32(qh[ib32+1]);
+ idx.vec[0] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[0], idx_shift), idx_mask);
+ idx.vec[1] = _mm256_and_si256(_mm256_sllv_epi32(idx.vec[1], idx_shift), idx_mask);
+ idx.vec[0] = _mm256_or_si256(idx.vec[0], _mm256_cvtepi16_epi32(_mm256_castsi256_si128(idx_l)));
+ idx.vec[1] = _mm256_or_si256(idx.vec[1], _mm256_cvtepi16_epi32(_mm256_extractf128_si256(idx_l, 1)));
+
+ // At leat on my CPU (Ryzen 7950X), using _mm256_i32gather_epi32 is slower than _mm256_set_epi32. Strange.
+ //const __m256i q2_1 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[0], 4);
+ //const __m256i q2_2 = _mm256_i32gather_epi32((const int *)iq3s_grid, idx.vec[1], 4);
+ const __m256i q2_1 = _mm256_set_epi32(
+ iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]],
+ iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]
+ );
+ const __m256i q2_2 = _mm256_set_epi32(
+ iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]],
+ iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[ 9]], iq3s_grid[idx.index[ 8]]
+ );
+
+ __m256i aux256 = _mm256_set1_epi32(signs[0] | (signs[1] << 16));
+ aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+ const __m256i s2_1 = _mm256_cmpeq_epi8(aux256, mask2);
+ const __m256i q8s_1 = _mm256_sub_epi8(_mm256_xor_si256(s2_1, q8_1), s2_1);
+
+ aux256 = _mm256_set1_epi32(signs[2] | (signs[3] << 16));
+ aux256 = _mm256_and_si256(_mm256_shuffle_epi8(aux256,mask1), mask2);
+ const __m256i s2_2 = _mm256_cmpeq_epi8(aux256, mask2);
+ const __m256i q8s_2 = _mm256_sub_epi8(_mm256_xor_si256(s2_2, q8_2), s2_2);
+
+ signs += 4;
+
+ const __m256i dot1 = _mm256_maddubs_epi16(q2_1, q8s_1);
+ const __m256i dot2 = _mm256_maddubs_epi16(q2_2, q8s_2);
+ const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+ const uint16_t ls2 = x[i].scales[ib32/2] >> 4;
+ const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(2*ls1+1));
+ const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(2*ls2+1));
+ sumi1 = _mm256_add_epi32(sumi1, p1);
+ sumi2 = _mm256_add_epi32(sumi2, p2);
+ }
+
+ accumf = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accumf);
+
+ }
+
+ *s = hsum_float_8(accumf);
+
+#elif defined(__AVX__)
+ static const uint8_t k_mask1[32] = {0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01,
+ 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03
+ };
+
+ static const uint8_t k_mask2[32] = {0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+ 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80, 0x01, 0x02, 0x04, 0x08, 0x10, 0x20, 0x40, 0x80,
+ };
+
+ const __m128i mask1_0 = _mm_loadu_si128((const __m128i*)k_mask1);
+ const __m128i mask1_1 = _mm_loadu_si128((const __m128i*)k_mask1 + 1);
+ const __m128i mask2_0 = _mm_loadu_si128((const __m128i*)k_mask2);
+ const __m128i mask2_1 = _mm_loadu_si128((const __m128i*)k_mask2 + 1);
+
+ const __m128i idx_mul_0 = _mm_set_epi32(32, 64, 128, 256);
+ const __m128i idx_mul_1 = _mm_set_epi32(2, 4, 8, 16);
+ const __m128i idx_mask = _mm_set1_epi32(256);
+
+ typedef union {
+ __m128i vec[4];
+ uint32_t index[16];
+ } index_t;
+
+ index_t idx;
+
+ __m256 accumf = _mm256_setzero_ps();
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint8_t * GGML_RESTRICT qs = x[i].qs;
+ const uint8_t * GGML_RESTRICT qh = x[i].qh;
+ const uint16_t * GGML_RESTRICT signs = (const uint16_t *)x[i].signs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ __m128i sumi1_0 = _mm_setzero_si128();
+ __m128i sumi1_1 = _mm_setzero_si128();
+ __m128i sumi2_0 = _mm_setzero_si128();
+ __m128i sumi2_1 = _mm_setzero_si128();
+ for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+ const __m128i q8_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i qs_tmp = _mm_loadu_si128((const __m128i *)qs);
+ const __m128i idx_l_0 = _mm_cvtepu8_epi16(qs_tmp);
+ const __m128i idx_l_1 = _mm_cvtepu8_epi16(_mm_srli_si128(qs_tmp, 8)); qs += 16;
+ idx.vec[0] = _mm_set1_epi32(qh[ib32+0]);
+ idx.vec[1] = idx.vec[0];
+ idx.vec[2] = _mm_set1_epi32(qh[ib32+1]);
+ idx.vec[3] = idx.vec[2];
+
+ idx.vec[0] = _mm_and_si128(_mm_mullo_epi32(idx.vec[0], idx_mul_0), idx_mask);
+ idx.vec[1] = _mm_and_si128(_mm_mullo_epi32(idx.vec[1], idx_mul_1), idx_mask);
+ idx.vec[2] = _mm_and_si128(_mm_mullo_epi32(idx.vec[2], idx_mul_0), idx_mask);
+ idx.vec[3] = _mm_and_si128(_mm_mullo_epi32(idx.vec[3], idx_mul_1), idx_mask);
+
+ idx.vec[0] = _mm_or_si128(idx.vec[0], _mm_cvtepi16_epi32(idx_l_0));
+ idx.vec[1] = _mm_or_si128(idx.vec[1], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_0, 8)));
+ idx.vec[2] = _mm_or_si128(idx.vec[2], _mm_cvtepi16_epi32(idx_l_1));
+ idx.vec[3] = _mm_or_si128(idx.vec[3], _mm_cvtepi16_epi32(_mm_srli_si128(idx_l_1, 8)));
+
+ const __m128i q2_1_0 = _mm_set_epi32(iq3s_grid[idx.index[3]], iq3s_grid[idx.index[2]], iq3s_grid[idx.index[1]], iq3s_grid[idx.index[0]]);
+ const __m128i q2_1_1 = _mm_set_epi32(iq3s_grid[idx.index[7]], iq3s_grid[idx.index[6]], iq3s_grid[idx.index[5]], iq3s_grid[idx.index[4]]);
+ const __m128i q2_2_0 = _mm_set_epi32(iq3s_grid[idx.index[11]], iq3s_grid[idx.index[10]], iq3s_grid[idx.index[9]], iq3s_grid[idx.index[8]]);
+ const __m128i q2_2_1 = _mm_set_epi32(iq3s_grid[idx.index[15]], iq3s_grid[idx.index[14]], iq3s_grid[idx.index[13]], iq3s_grid[idx.index[12]]);
+
+ __m128i aux128_0 = _mm_set1_epi32(signs[0] | (signs[1] << 16));
+ __m128i aux128_1 = aux128_0;
+ aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+ aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+ const __m128i s2_1_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+ const __m128i s2_1_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+ const __m128i q8s_1_0 = _mm_sub_epi8(_mm_xor_si128(s2_1_0, q8_1_0), s2_1_0);
+ const __m128i q8s_1_1 = _mm_sub_epi8(_mm_xor_si128(s2_1_1, q8_1_1), s2_1_1);
+
+ aux128_0 = _mm_set1_epi32(signs[2] | (signs[3] << 16));
+ aux128_1 = aux128_0;
+ aux128_0 = _mm_and_si128(_mm_shuffle_epi8(aux128_0,mask1_0), mask2_0);
+ aux128_1 = _mm_and_si128(_mm_shuffle_epi8(aux128_1,mask1_1), mask2_1);
+ const __m128i s2_2_0 = _mm_cmpeq_epi8(aux128_0, mask2_0);
+ const __m128i s2_2_1 = _mm_cmpeq_epi8(aux128_1, mask2_1);
+ const __m128i q8s_2_0 = _mm_sub_epi8(_mm_xor_si128(s2_2_0, q8_2_0), s2_2_0);
+ const __m128i q8s_2_1 = _mm_sub_epi8(_mm_xor_si128(s2_2_1, q8_2_1), s2_2_1);
+
+ signs += 4;
+
+ const __m128i dot1_0 = _mm_maddubs_epi16(q2_1_0, q8s_1_0);
+ const __m128i dot1_1 = _mm_maddubs_epi16(q2_1_1, q8s_1_1);
+ const __m128i dot2_0 = _mm_maddubs_epi16(q2_2_0, q8s_2_0);
+ const __m128i dot2_1 = _mm_maddubs_epi16(q2_2_1, q8s_2_1);
+ const uint16_t ls1 = x[i].scales[ib32/2] & 0xf;
+ const uint16_t ls2 = x[i].scales[ib32/2] >> 4;
+ const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(2*ls1+1));
+ const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(2*ls1+1));
+ const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(2*ls2+1));
+ const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(2*ls2+1));
+ sumi1_0 = _mm_add_epi32(sumi1_0, p1_0);
+ sumi1_1 = _mm_add_epi32(sumi1_1, p1_1);
+ sumi2_0 = _mm_add_epi32(sumi2_0, p2_0);
+ sumi2_1 = _mm_add_epi32(sumi2_1, p2_1);
+ }
+
+ accumf = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(_mm_add_epi32(sumi1_1, sumi2_1), _mm_add_epi32(sumi1_0, sumi2_0)))), accumf);
+
+ }
+
+ *s = hsum_float_8(accumf);
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ ggml_vec_dot_iq3_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_iq1_s_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq1_s * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+#if defined __AVX2__
+
+ __m256 accum = _mm256_setzero_ps();
+ float accum1 = 0;
+ for (int i = 0; i < nb; ++i) {
+
+ const int8_t * q8 = y[i].qs;
+ const uint8_t * qs = x[i].qs;
+ const uint16_t * qh = x[i].qh;
+
+ __m256i sumi = _mm256_setzero_si256();
+ int sumi1 = 0;
+ for (int ib = 0; ib < QK_K/32; ib += 2) {
+#ifdef __BMI2__
+ const uint64_t packed_idx1 = _pdep_u64(*(const uint32_t *)qs, 0x00ff00ff00ff00ffULL) | _pdep_u64(qh[ib], 0x700070007000700ULL);
+ const uint64_t packed_idx2 = _pdep_u64(*(const uint32_t *)(qs + 4), 0x00ff00ff00ff00ffULL) | _pdep_u64(qh[ib + 1], 0x700070007000700ULL);
+ const uint16_t *idx1 = (const uint16_t *)(&packed_idx1);
+ const uint16_t *idx2 = (const uint16_t *)(&packed_idx2);
+ const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[idx1[3]], iq1s_grid[idx1[2]], iq1s_grid[idx1[1]], iq1s_grid[idx1[0]]);
+ const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[idx2[3]], iq1s_grid[idx2[2]], iq1s_grid[idx2[1]], iq1s_grid[idx2[0]]);
+#else
+ const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)],
+ iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
+ const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)],
+ iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
+#endif
+ qs += 8;
+ const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+ const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
+ const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
+ const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+ const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+ const __m256i p1 = _mm256_madd_epi16(dot1, _mm256_set1_epi16(ls1));
+ const __m256i p2 = _mm256_madd_epi16(dot2, _mm256_set1_epi16(ls2));
+
+ sumi = _mm256_add_epi32(sumi, _mm256_add_epi32(p1, p2));
+ sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+ + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+ }
+
+ const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+ accum = _mm256_fmadd_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(sumi), accum);
+ accum1 += d * sumi1;
+
+ }
+
+ *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
+#elif defined __AVX__
+ __m256 accum = _mm256_setzero_ps();
+ float accum1 = 0;
+ for (int i = 0; i < nb; ++i) {
+
+ const int8_t * q8 = y[i].qs;
+ const uint8_t * qs = x[i].qs;
+ const uint16_t * qh = x[i].qh;
+
+ __m128i sumi1_0 = _mm_setzero_si128();
+ __m128i sumi1_1 = _mm_setzero_si128();
+ int sumi1 = 0;
+ for (int ib = 0; ib < QK_K/32; ib += 2) {
+ const __m128i q1b_1_0 = _mm_set_epi64x(iq1s_grid[qs[1] | ((qh[ib+0] << 5) & 0x700)], iq1s_grid[qs[0] | ((qh[ib+0] << 8) & 0x700)]);
+ const __m128i q1b_1_1 = _mm_set_epi64x(iq1s_grid[qs[3] | ((qh[ib+0] >> 1) & 0x700)], iq1s_grid[qs[2] | ((qh[ib+0] << 2) & 0x700)]);
+ const __m128i q1b_2_0 = _mm_set_epi64x(iq1s_grid[qs[5] | ((qh[ib+1] << 5) & 0x700)], iq1s_grid[qs[4] | ((qh[ib+1] << 8) & 0x700)]);
+ const __m128i q1b_2_1 = _mm_set_epi64x(iq1s_grid[qs[7] | ((qh[ib+1] >> 1) & 0x700)], iq1s_grid[qs[6] | ((qh[ib+1] << 2) & 0x700)]);
+ qs += 8;
+ const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+ const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
+ const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
+ const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
+ const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
+ const int16_t ls1 = 2*((qh[ib+0] >> 12) & 7) + 1;
+ const int16_t ls2 = 2*((qh[ib+1] >> 12) & 7) + 1;
+ const __m128i p1_0 = _mm_madd_epi16(dot1_0, _mm_set1_epi16(ls1));
+ const __m128i p1_1 = _mm_madd_epi16(dot1_1, _mm_set1_epi16(ls1));
+ const __m128i p2_0 = _mm_madd_epi16(dot2_0, _mm_set1_epi16(ls2));
+ const __m128i p2_1 = _mm_madd_epi16(dot2_1, _mm_set1_epi16(ls2));
+
+ sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
+ sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
+ sumi1 += (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]) * (qh[ib+0] & 0x8000 ? -1 : 1) * ls1
+ + (y[i].bsums[2*ib+2] + y[i].bsums[2*ib+3]) * (qh[ib+1] & 0x8000 ? -1 : 1) * ls2;
+ }
+
+ const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+ accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(d), _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum);
+ accum1 += d * sumi1;
+
+ }
+
+ *s = hsum_float_8(accum) + IQ1S_DELTA * accum1;
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ ggml_vec_dot_iq1_s_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_iq1_m_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq1_m * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ iq1m_scale_t scale;
+
+#if defined __AVX2__
+
+ const __m256i mask = _mm256_set1_epi16(0x7);
+ const __m256i mone = _mm256_set1_epi16(1);
+ const __m256i mone8 = _mm256_set1_epi8(1);
+ const __m256i mtwo8 = _mm256_set1_epi8(2);
+ // VPSHUFB cannot cross 128-bit lanes so odd shifts go to upper half.
+ const __m256i scales_shift = _mm256_set_epi64x(9, 3, 6, 0);
+
+ __m256 accum1 = _mm256_setzero_ps();
+ __m256 accum2 = _mm256_setzero_ps();
+ for (int i = 0; i < nb; ++i) {
+
+ const int8_t * q8 = y[i].qs;
+ const uint8_t * qs = x[i].qs;
+ const uint8_t * qh = x[i].qh;
+ const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+ scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+ // Extract 3-bit scales (16 values)
+ __m256i scales = _mm256_set1_epi64x(*(const uint64_t*)sc);
+ scales = _mm256_srlv_epi64(scales, scales_shift);
+ scales = _mm256_add_epi16(_mm256_slli_epi16(_mm256_and_si256(scales, mask), 1), mone);
+
+ // Indices to repeat each scale 8 times.
+ __m256i scales_idx1 = _mm256_set1_epi16(0x0100);
+ __m256i scales_idx2 = _mm256_add_epi8(scales_idx1, _mm256_set1_epi8(8));
+
+ __m256i sumi1 = _mm256_setzero_si256();
+ __m256i sumi2 = _mm256_setzero_si256();
+ for (int ib = 0; ib < QK_K/32; ib += 2) {
+#ifdef __BMI2__
+ const uint64_t packed_idx1 = _pdep_u64(*(const uint32_t *)qs, 0x00ff00ff00ff00ffULL)
+ | _pdep_u64(*(const uint16_t*)(qh) & 0x7777, 0xf000f000f000f00ULL);
+ const uint64_t packed_idx2 = _pdep_u64(*(const uint32_t *)(qs + 4), 0x00ff00ff00ff00ffULL)
+ | _pdep_u64(*(const uint16_t*)(qh + 2) & 0x7777, 0xf000f000f000f00ULL);
+ const uint16_t *idx1 = (const uint16_t *)(&packed_idx1);
+ const uint16_t *idx2 = (const uint16_t *)(&packed_idx2);
+ const __m256i q1b_1 = _mm256_set_epi64x(iq1s_grid[idx1[3]], iq1s_grid[idx1[2]], iq1s_grid[idx1[1]], iq1s_grid[idx1[0]]);
+ const __m256i q1b_2 = _mm256_set_epi64x(iq1s_grid[idx2[3]], iq1s_grid[idx2[2]], iq1s_grid[idx2[1]], iq1s_grid[idx2[0]]);
+
+ // Convert signs to bytes 0x81 (negative) or 0x01 (positive)
+ const uint64_t delta_sign = _pdep_u64(*(const uint32_t*)(qh) & 0x88888888, 0xf0f0f0f0f0f0f0f0ULL);
+ const __m256i delta1 = _mm256_or_si256(mone8, _mm256_cvtepi8_epi64(_mm_set1_epi32(delta_sign)));
+ const __m256i delta2 = _mm256_or_si256(mone8, _mm256_cvtepi8_epi64(_mm_set1_epi32(delta_sign >> 32)));
+#else
+ const __m256i q1b_1 = _mm256_set_epi64x(
+ iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)],
+ iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]
+ );
+ const __m256i q1b_2 = _mm256_set_epi64x(
+ iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)],
+ iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]
+ );
+
+ const __m256i delta1 = _mm256_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+ qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
+ qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+ qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+ const __m256i delta2 = _mm256_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+ qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101,
+ qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+ qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+#endif
+ const __m256i q8b_1 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+ const __m256i q8b_2 = _mm256_loadu_si256((const __m256i*)q8); q8 += 32;
+
+ const __m256i dot1 = mul_add_epi8(q1b_1, q8b_1);
+ const __m256i dot2 = mul_add_epi8(q1b_2, q8b_2);
+ const __m256i dot3 = _mm256_maddubs_epi16(mone8, _mm256_sign_epi8(q8b_1, delta1));
+ const __m256i dot4 = _mm256_maddubs_epi16(mone8, _mm256_sign_epi8(q8b_2, delta2));
+
+ __m256i scale1 = _mm256_shuffle_epi8(scales, scales_idx1);
+ __m256i scale2 = _mm256_shuffle_epi8(scales, scales_idx2);
+
+ scales_idx1 = _mm256_add_epi8(scales_idx1, mtwo8);
+ scales_idx2 = _mm256_add_epi8(scales_idx2, mtwo8);
+
+ const __m256i p1 = _mm256_madd_epi16(dot1, scale1);
+ const __m256i p2 = _mm256_madd_epi16(dot2, scale2);
+ const __m256i p3 = _mm256_madd_epi16(dot3, scale1);
+ const __m256i p4 = _mm256_madd_epi16(dot4, scale2);
+
+ sumi1 = _mm256_add_epi32(sumi1, _mm256_add_epi32(p1, p2));
+ sumi2 = _mm256_add_epi32(sumi2, _mm256_add_epi32(p3, p4));
+
+ qs += 8; qh += 4;
+ }
+
+ const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16));
+
+ accum1 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi1), accum1);
+ accum2 = _mm256_fmadd_ps(d, _mm256_cvtepi32_ps(sumi2), accum2);
+ }
+
+ *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
+
+#elif defined __AVX__
+ const __m128i mask = _mm_set1_epi16(0x7);
+ const __m128i mone = _mm_set1_epi16(1);
+
+ __m256 accum1 = _mm256_setzero_ps();
+ __m256 accum2 = _mm256_setzero_ps();
+ for (int i = 0; i < nb; ++i) {
+
+ const int8_t * q8 = y[i].qs;
+ const uint8_t * qs = x[i].qs;
+ const uint8_t * qh = x[i].qh;
+ const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+ scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+ __m128i sumi1_0 = _mm_setzero_si128();
+ __m128i sumi1_1 = _mm_setzero_si128();
+ __m128i sumi2_0 = _mm_setzero_si128();
+ __m128i sumi2_1 = _mm_setzero_si128();
+ for (int ib = 0; ib < QK_K/32; ib += 2) {
+ const __m128i q1b_1_0 = _mm_set_epi64x(
+ iq1s_grid[qs[1] | (((uint16_t)qh[0] << 4) & 0x700)], iq1s_grid[qs[0] | (((uint16_t)qh[0] << 8) & 0x700)]);
+ const __m128i q1b_1_1 = _mm_set_epi64x(
+ iq1s_grid[qs[3] | (((uint16_t)qh[1] << 4) & 0x700)], iq1s_grid[qs[2] | (((uint16_t)qh[1] << 8) & 0x700)]);
+ const __m128i q1b_2_0 = _mm_set_epi64x(
+ iq1s_grid[qs[5] | (((uint16_t)qh[2] << 4) & 0x700)], iq1s_grid[qs[4] | (((uint16_t)qh[2] << 8) & 0x700)]);
+ const __m128i q1b_2_1 = _mm_set_epi64x(
+ iq1s_grid[qs[7] | (((uint16_t)qh[3] << 4) & 0x700)], iq1s_grid[qs[6] | (((uint16_t)qh[3] << 8) & 0x700)]);
+ const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+
+ const __m128i dot1_0 = mul_add_epi8_sse(q1b_1_0, q8b_1_0);
+ const __m128i dot1_1 = mul_add_epi8_sse(q1b_1_1, q8b_1_1);
+ const __m128i dot2_0 = mul_add_epi8_sse(q1b_2_0, q8b_2_0);
+ const __m128i dot2_1 = mul_add_epi8_sse(q1b_2_1, q8b_2_1);
+
+ const __m128i delta1_0 = _mm_set_epi64x(qh[0] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+ qh[0] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+ const __m128i delta1_1 = _mm_set_epi64x(qh[1] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+ qh[1] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+ const __m128i delta2_0 = _mm_set_epi64x(qh[2] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+ qh[2] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+ const __m128i delta2_1 = _mm_set_epi64x(qh[3] & 0x80 ? 0xffffffffffffffff : 0x0101010101010101,
+ qh[3] & 0x08 ? 0xffffffffffffffff : 0x0101010101010101);
+
+ const __m128i dot3_0 = mul_add_epi8_sse(delta1_0, q8b_1_0);
+ const __m128i dot3_1 = mul_add_epi8_sse(delta1_1, q8b_1_1);
+ const __m128i dot4_0 = mul_add_epi8_sse(delta2_0, q8b_2_0);
+ const __m128i dot4_1 = mul_add_epi8_sse(delta2_1, q8b_2_1);
+
+ __m128i scale1_0 = _mm_set1_epi16(sc[ib/2] >> 0);
+ __m128i scale1_1 = _mm_set1_epi16(sc[ib/2] >> 3);
+ __m128i scale2_0 = _mm_set1_epi16(sc[ib/2] >> 6);
+ __m128i scale2_1 = _mm_set1_epi16(sc[ib/2] >> 9);
+
+ scale1_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_0, mask), 1), mone);
+ scale1_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale1_1, mask), 1), mone);
+ scale2_0 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_0, mask), 1), mone);
+ scale2_1 = _mm_add_epi16(_mm_slli_epi16(_mm_and_si128(scale2_1, mask), 1), mone);
+ const __m128i p1_0 = _mm_madd_epi16(dot1_0, scale1_0);
+ const __m128i p1_1 = _mm_madd_epi16(dot1_1, scale1_1);
+ const __m128i p2_0 = _mm_madd_epi16(dot2_0, scale2_0);
+ const __m128i p2_1 = _mm_madd_epi16(dot2_1, scale2_1);
+ const __m128i p3_0 = _mm_madd_epi16(dot3_0, scale1_0);
+ const __m128i p3_1 = _mm_madd_epi16(dot3_1, scale1_1);
+ const __m128i p4_0 = _mm_madd_epi16(dot4_0, scale2_0);
+ const __m128i p4_1 = _mm_madd_epi16(dot4_1, scale2_1);
+
+ sumi1_0 = _mm_add_epi32(sumi1_0, _mm_add_epi32(p1_0, p2_0));
+ sumi1_1 = _mm_add_epi32(sumi1_1, _mm_add_epi32(p1_1, p2_1));
+ sumi2_0 = _mm_add_epi32(sumi2_0, _mm_add_epi32(p3_0, p4_0));
+ sumi2_1 = _mm_add_epi32(sumi2_1, _mm_add_epi32(p3_1, p4_1));
+
+ qs += 8; qh += 4;
+ }
+
+ const __m256 d = _mm256_set1_ps(y[i].d * GGML_CPU_FP16_TO_FP32(scale.f16));
+
+ accum1 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi1_1, sumi1_0))), accum1);
+ accum2 = _mm256_add_ps(_mm256_mul_ps(d, _mm256_cvtepi32_ps(MM256_SET_M128I(sumi2_1, sumi2_0))), accum2);
+ }
+
+ *s = hsum_float_8(accum1) + IQ1M_DELTA * hsum_float_8(accum2);
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ UNUSED(scale);
+ ggml_vec_dot_iq1_m_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
+
+void ggml_vec_dot_iq4_nl_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+ assert(n % QK4_NL == 0);
+ static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
+
+ const block_iq4_nl * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK4_NL;
+
+ int ib = 0;
+ float sumf = 0;
+
+#if defined __AVX2__
+
+ const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+ const __m128i m4b = _mm_set1_epi8(0x0f);
+ const __m256i mone = _mm256_set1_epi16(1);
+
+ __m256 accum1 = _mm256_setzero_ps();
+ __m256 accum2 = _mm256_setzero_ps();
+ for (; ib + 1 < nb; ib += 2) {
+ const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)x[ib + 0].qs);
+ const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)x[ib + 1].qs);
+ const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)y[ib + 0].qs);
+ const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)y[ib + 1].qs);
+ const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
+ _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
+ const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
+ _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
+ const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+ const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+ const __m256i p_1 = _mm256_madd_epi16(p16_1, mone);
+ const __m256i p_2 = _mm256_madd_epi16(p16_2, mone);
+ accum1 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 0].d)*GGML_CPU_FP16_TO_FP32(x[ib + 0].d)),
+ _mm256_cvtepi32_ps(p_1), accum1);
+ accum2 = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(y[ib + 1].d)*GGML_CPU_FP16_TO_FP32(x[ib + 1].d)),
+ _mm256_cvtepi32_ps(p_2), accum2);
+ }
+
+ sumf = hsum_float_8(_mm256_add_ps(accum1, accum2));
+
+#elif defined __AVX__
+ const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+ const __m128i m4b = _mm_set1_epi8(0x0f);
+
+ __m256 accum = _mm256_setzero_ps();
+ for (; ib + 1 < nb; ib += 2) {
+ const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)x[ib + 0].qs);
+ const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)x[ib + 1].qs);
+ const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs);
+ const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)y[ib + 0].qs + 1);
+ const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs);
+ const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)y[ib + 1].qs + 1);
+
+ const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+ const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+ const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+ const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+
+ const __m256 p = mul_sum_i8_quad_float(q4b_1_0, q4b_1_1, q4b_2_0, q4b_2_1, q8b_1_0, q8b_1_1, q8b_2_0, q8b_2_1);
+ const __m256 deltas = quad_fp16_delta_float(x[ib].d, y[ib].d, x[ib + 1].d, y[ib + 1].d);
+ accum = _mm256_add_ps(_mm256_mul_ps(deltas, p), accum);
+ }
+
+ sumf = hsum_float_8(accum);
+
+#endif
+ for (; ib < nb; ++ib) {
+ const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
+ int sumi1 = 0, sumi2 = 0;
+ for (int j = 0; j < QK4_NL/2; ++j) {
+ sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
+ sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >> 4];
+ }
+ sumf += d * (sumi1 + sumi2);
+ }
+ *s = sumf;
+}
+
+void ggml_vec_dot_iq4_xs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+ assert(n % QK_K == 0);
+
+ const block_iq4_xs * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+#if defined __AVX2__
+
+ const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+ const __m128i m4b = _mm_set1_epi8(0x0f);
+
+ __m256 accum = _mm256_setzero_ps();
+ for (int ibl = 0; ibl < nb; ++ibl) {
+ const uint8_t * qs = x[ibl].qs;
+ const int8_t * q8 = y[ibl].qs;
+ uint16_t sh = x[ibl].scales_h;
+ __m256i sumi1 = _mm256_setzero_si256();
+ __m256i sumi2 = _mm256_setzero_si256();
+ for (int ib = 0; ib < QK_K/32; ib += 2) {
+ const __m128i q4bits_1 = _mm_loadu_si128((const __m128i*)qs); qs += 16;
+ const __m128i q4bits_2 = _mm_loadu_si128((const __m128i*)qs); qs += 16;
+ const __m256i q8b_1 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ const __m256i q8b_2 = _mm256_loadu_si256((const __m256i *)q8); q8 += 32;
+ const __m256i q4b_1 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b)),
+ _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b)));
+ const __m256i q4b_2 = MM256_SET_M128I(_mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b)),
+ _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b)));
+ const __m256i p16_1 = mul_add_epi8(q4b_1, q8b_1);
+ const __m256i p16_2 = mul_add_epi8(q4b_2, q8b_2);
+ const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
+ const int16_t ls2 = ((x[ibl].scales_l[ib/2] >> 4) | ((sh << 2) & 0x30)) - 32;
+ sh >>= 4;
+ const __m256i p_1 = _mm256_madd_epi16(p16_1, _mm256_set1_epi16(ls1));
+ const __m256i p_2 = _mm256_madd_epi16(p16_2, _mm256_set1_epi16(ls2));
+ sumi1 = _mm256_add_epi32(p_1, sumi1);
+ sumi2 = _mm256_add_epi32(p_2, sumi2);
+ }
+ accum = _mm256_fmadd_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+ _mm256_cvtepi32_ps(_mm256_add_epi32(sumi1, sumi2)), accum);
+ }
+
+ *s = hsum_float_8(accum);
+
+#elif defined __AVX__
+ const __m128i values128 = _mm_loadu_si128((const __m128i*)kvalues_iq4nl);
+ const __m128i m4b = _mm_set1_epi8(0x0f);
+
+ __m256 accum = _mm256_setzero_ps();
+ for (int ibl = 0; ibl < nb; ++ibl) {
+ const uint8_t * qs = x[ibl].qs;
+ const int8_t * q8 = y[ibl].qs;
+ uint16_t sh = x[ibl].scales_h;
+ __m128i sumi1_0 = _mm_setzero_si128();
+ __m128i sumi1_1 = _mm_setzero_si128();
+ __m128i sumi2_0 = _mm_setzero_si128();
+ __m128i sumi2_1 = _mm_setzero_si128();
+ for (int ib = 0; ib < QK_K/32; ib += 2) {
+ const __m128i q4bits_1 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
+ const __m128i q4bits_2 = _mm_loadu_si128((const __m128i *)qs); qs += 16;
+ const __m128i q8b_1_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8b_1_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8b_2_0 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q8b_2_1 = _mm_loadu_si128((const __m128i *)q8); q8 += 16;
+ const __m128i q4b_1_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_1, m4b));
+ const __m128i q4b_1_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_1, 4), m4b));
+ const __m128i q4b_2_0 = _mm_shuffle_epi8(values128, _mm_and_si128(q4bits_2, m4b));
+ const __m128i q4b_2_1 = _mm_shuffle_epi8(values128, _mm_and_si128(_mm_srli_epi16(q4bits_2, 4), m4b));
+ const __m128i p16_1_0 = mul_add_epi8_sse(q4b_1_0, q8b_1_0);
+ const __m128i p16_1_1 = mul_add_epi8_sse(q4b_1_1, q8b_1_1);
+ const __m128i p16_2_0 = mul_add_epi8_sse(q4b_2_0, q8b_2_0);
+ const __m128i p16_2_1 = mul_add_epi8_sse(q4b_2_1, q8b_2_1);
+ const int16_t ls1 = ((x[ibl].scales_l[ib/2] & 0xf) | ((sh << 4) & 0x30)) - 32;
+ const int16_t ls2 = ((x[ibl].scales_l[ib/2] >> 4) | ((sh << 2) & 0x30)) - 32;
+ sh >>= 4;
+ const __m128i p_1_0 = _mm_madd_epi16(p16_1_0, _mm_set1_epi16(ls1));
+ const __m128i p_1_1 = _mm_madd_epi16(p16_1_1, _mm_set1_epi16(ls1));
+ const __m128i p_2_0 = _mm_madd_epi16(p16_2_0, _mm_set1_epi16(ls2));
+ const __m128i p_2_1 = _mm_madd_epi16(p16_2_1, _mm_set1_epi16(ls2));
+ sumi1_0 = _mm_add_epi32(p_1_0, sumi1_0);
+ sumi1_1 = _mm_add_epi32(p_1_1, sumi1_1);
+ sumi2_0 = _mm_add_epi32(p_2_0, sumi2_0);
+ sumi2_1 = _mm_add_epi32(p_2_1, sumi2_1);
+ }
+ __m128i sumi12_0 = _mm_add_epi32(sumi1_0, sumi2_0);
+ __m128i sumi12_1 = _mm_add_epi32(sumi1_1, sumi2_1);
+ accum = _mm256_add_ps(_mm256_mul_ps(_mm256_set1_ps(GGML_CPU_FP16_TO_FP32(x[ibl].d)*y[ibl].d),
+ _mm256_cvtepi32_ps(MM256_SET_M128I(sumi12_1, sumi12_0))), accum);
+ }
+
+ *s = hsum_float_8(accum);
+
+#else
+ UNUSED(x);
+ UNUSED(y);
+ UNUSED(nb);
+ ggml_vec_dot_iq4_xs_q8_K_generic(n, s, bs, vx, bx, vy, by, nrc);
+#endif
+}
diff --git a/python/freetoken/kernel/csrc/ggml/ggml-cpu/ggml-cpu-impl.h b/python/freetoken/kernel/csrc/ggml/ggml-cpu/ggml-cpu-impl.h
new file mode 100644
index 00000000..82345b20
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/ggml-cpu/ggml-cpu-impl.h
@@ -0,0 +1,538 @@
+#pragma once
+
+// GGML CPU internal header
+
+#include "ggml.h"
+#include "ggml-impl.h"
+
+#include <stdlib.h> // load `stdlib.h` before other headers to work around MinGW bug: https://sourceforge.net/p/mingw-w64/bugs/192/
+//#include <stddef.h>
+#include <stdbool.h>
+#include <string.h> // memcpy
+#include <math.h> // fabsf
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+struct ggml_compute_params {
+ // ith = thread index, nth = number of threads
+ int ith, nth;
+
+ // work buffer for all threads
+ size_t wsize;
+ void * wdata;
+
+ struct ggml_threadpool * threadpool;
+
+ // use reference implementation
+ bool use_ref;
+};
+
+
+#if defined(_MSC_VER)
+
+#define m512bh(p) p
+#define m512i(p) p
+
+#else
+
+#define m512bh(p) (__m512bh)(p)
+#define m512i(p) (__m512i)(p)
+
+#endif
+
+// __FMA__ and __F16C__ are not defined in MSVC, however they are implied with AVX2/AVX512
+#if defined(_MSC_VER) && (defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __FMA__
+#define __FMA__
+#endif
+#ifndef __F16C__
+#define __F16C__
+#endif
+#endif
+
+// __SSE3__ and __SSSE3__ are not defined in MSVC, but SSE3/SSSE3 are present when AVX/AVX2/AVX512 are available
+#if defined(_MSC_VER) && (defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__))
+#ifndef __SSE3__
+#define __SSE3__
+#endif
+#ifndef __SSSE3__
+#define __SSSE3__
+#endif
+#endif
+
+#if defined(__s390x__) && defined(__VEC__)
+#ifndef __VXE__
+#define __VXE__
+#endif // __VXE__
+#ifndef __VXE2__
+#define __VXE2__
+#endif // __VXE2__
+#endif // __s390x__ && __VEC__
+
+#if defined(__ARM_FEATURE_SVE) && defined(__linux__)
+#include <sys/prctl.h>
+#endif
+
+#if defined(__ARM_NEON)
+
+// ref: https://github.com/ggml-org/llama.cpp/pull/5404
+#if defined(_MSC_VER) && !defined(__clang__)
+#define ggml_vld1q_u32(w,x,y,z) { ((w) + ((uint64_t)(x) << 32)), ((y) + ((uint64_t)(z) << 32)) }
+#else
+#define ggml_vld1q_u32(w,x,y,z) { (w), (x), (y), (z) }
+#endif // _MSC_VER
+
+#if !defined(__aarch64__)
+
+// 32-bit ARM compatibility
+
+// vaddlvq_s16
+// vpaddq_s16
+// vpaddq_s32
+// vaddvq_s32
+// vaddvq_f32
+// vmaxvq_f32
+// vcvtnq_s32_f32
+// vzip1_u8
+// vzip2_u8
+
+inline static int32_t vaddlvq_s16(int16x8_t v) {
+ int32x4_t v0 = vreinterpretq_s32_s64(vpaddlq_s32(vpaddlq_s16(v)));
+ return vgetq_lane_s32(v0, 0) + vgetq_lane_s32(v0, 2);
+}
+
+inline static int16x8_t vpaddq_s16(int16x8_t a, int16x8_t b) {
+ int16x4_t a0 = vpadd_s16(vget_low_s16(a), vget_high_s16(a));
+ int16x4_t b0 = vpadd_s16(vget_low_s16(b), vget_high_s16(b));
+ return vcombine_s16(a0, b0);
+}
+
+inline static int32x4_t vpaddq_s32(int32x4_t a, int32x4_t b) {
+ int32x2_t a0 = vpadd_s32(vget_low_s32(a), vget_high_s32(a));
+ int32x2_t b0 = vpadd_s32(vget_low_s32(b), vget_high_s32(b));
+ return vcombine_s32(a0, b0);
+}
+
+inline static int32_t vaddvq_s32(int32x4_t v) {
+ return vgetq_lane_s32(v, 0) + vgetq_lane_s32(v, 1) + vgetq_lane_s32(v, 2) + vgetq_lane_s32(v, 3);
+}
+
+inline static float vaddvq_f32(float32x4_t v) {
+ return vgetq_lane_f32(v, 0) + vgetq_lane_f32(v, 1) + vgetq_lane_f32(v, 2) + vgetq_lane_f32(v, 3);
+}
+
+inline static float vmaxvq_f32(float32x4_t v) {
+ return
+ MAX(MAX(vgetq_lane_f32(v, 0), vgetq_lane_f32(v, 1)),
+ MAX(vgetq_lane_f32(v, 2), vgetq_lane_f32(v, 3)));
+}
+
+inline static int32x4_t vcvtnq_s32_f32(float32x4_t v) {
+ int32x4_t res;
+
+ res[0] = roundf(vgetq_lane_f32(v, 0));
+ res[1] = roundf(vgetq_lane_f32(v, 1));
+ res[2] = roundf(vgetq_lane_f32(v, 2));
+ res[3] = roundf(vgetq_lane_f32(v, 3));
+
+ return res;
+}
+
+inline static uint8x8_t vzip1_u8(uint8x8_t a, uint8x8_t b) {
+ uint8x8_t res;
+
+ res[0] = a[0]; res[1] = b[0];
+ res[2] = a[1]; res[3] = b[1];
+ res[4] = a[2]; res[5] = b[2];
+ res[6] = a[3]; res[7] = b[3];
+
+ return res;
+}
+
+inline static uint8x8_t vzip2_u8(uint8x8_t a, uint8x8_t b) {
+ uint8x8_t res;
+
+ res[0] = a[4]; res[1] = b[4];
+ res[2] = a[5]; res[3] = b[5];
+ res[4] = a[6]; res[5] = b[6];
+ res[6] = a[7]; res[7] = b[7];
+
+ return res;
+}
+
+// vld1q_s16_x2
+// vld1q_u8_x2
+// vld1q_u8_x4
+// vld1q_s8_x2
+// vld1q_s8_x4
+// TODO: double-check these work correctly
+
+typedef struct ggml_int16x8x2_t {
+ int16x8_t val[2];
+} ggml_int16x8x2_t;
+
+inline static ggml_int16x8x2_t ggml_vld1q_s16_x2(const int16_t * ptr) {
+ ggml_int16x8x2_t res;
+
+ res.val[0] = vld1q_s16(ptr + 0);
+ res.val[1] = vld1q_s16(ptr + 8);
+
+ return res;
+}
+
+typedef struct ggml_uint8x16x2_t {
+ uint8x16_t val[2];
+} ggml_uint8x16x2_t;
+
+inline static ggml_uint8x16x2_t ggml_vld1q_u8_x2(const uint8_t * ptr) {
+ ggml_uint8x16x2_t res;
+
+ res.val[0] = vld1q_u8(ptr + 0);
+ res.val[1] = vld1q_u8(ptr + 16);
+
+ return res;
+}
+
+typedef struct ggml_uint8x16x4_t {
+ uint8x16_t val[4];
+} ggml_uint8x16x4_t;
+
+inline static ggml_uint8x16x4_t ggml_vld1q_u8_x4(const uint8_t * ptr) {
+ ggml_uint8x16x4_t res;
+
+ res.val[0] = vld1q_u8(ptr + 0);
+ res.val[1] = vld1q_u8(ptr + 16);
+ res.val[2] = vld1q_u8(ptr + 32);
+ res.val[3] = vld1q_u8(ptr + 48);
+
+ return res;
+}
+
+typedef struct ggml_int8x16x2_t {
+ int8x16_t val[2];
+} ggml_int8x16x2_t;
+
+inline static ggml_int8x16x2_t ggml_vld1q_s8_x2(const int8_t * ptr) {
+ ggml_int8x16x2_t res;
+
+ res.val[0] = vld1q_s8(ptr + 0);
+ res.val[1] = vld1q_s8(ptr + 16);
+
+ return res;
+}
+
+typedef struct ggml_int8x16x4_t {
+ int8x16_t val[4];
+} ggml_int8x16x4_t;
+
+inline static ggml_int8x16x4_t ggml_vld1q_s8_x4(const int8_t * ptr) {
+ ggml_int8x16x4_t res;
+
+ res.val[0] = vld1q_s8(ptr + 0);
+ res.val[1] = vld1q_s8(ptr + 16);
+ res.val[2] = vld1q_s8(ptr + 32);
+ res.val[3] = vld1q_s8(ptr + 48);
+
+ return res;
+}
+
+// NOTE: not tested
+inline static int8x16_t ggml_vqtbl1q_s8(int8x16_t a, uint8x16_t b) {
+ int8x16_t res;
+
+ res[ 0] = a[b[ 0]];
+ res[ 1] = a[b[ 1]];
+ res[ 2] = a[b[ 2]];
+ res[ 3] = a[b[ 3]];
+ res[ 4] = a[b[ 4]];
+ res[ 5] = a[b[ 5]];
+ res[ 6] = a[b[ 6]];
+ res[ 7] = a[b[ 7]];
+ res[ 8] = a[b[ 8]];
+ res[ 9] = a[b[ 9]];
+ res[10] = a[b[10]];
+ res[11] = a[b[11]];
+ res[12] = a[b[12]];
+ res[13] = a[b[13]];
+ res[14] = a[b[14]];
+ res[15] = a[b[15]];
+
+ return res;
+}
+
+// NOTE: not tested
+inline static uint8x16_t ggml_vqtbl1q_u8(uint8x16_t a, uint8x16_t b) {
+ uint8x16_t res;
+
+ res[ 0] = a[b[ 0]];
+ res[ 1] = a[b[ 1]];
+ res[ 2] = a[b[ 2]];
+ res[ 3] = a[b[ 3]];
+ res[ 4] = a[b[ 4]];
+ res[ 5] = a[b[ 5]];
+ res[ 6] = a[b[ 6]];
+ res[ 7] = a[b[ 7]];
+ res[ 8] = a[b[ 8]];
+ res[ 9] = a[b[ 9]];
+ res[10] = a[b[10]];
+ res[11] = a[b[11]];
+ res[12] = a[b[12]];
+ res[13] = a[b[13]];
+ res[14] = a[b[14]];
+ res[15] = a[b[15]];
+
+ return res;
+}
+
+#else
+
+#define ggml_int16x8x2_t int16x8x2_t
+#define ggml_uint8x16x2_t uint8x16x2_t
+#define ggml_uint8x16x4_t uint8x16x4_t
+#define ggml_int8x16x2_t int8x16x2_t
+#define ggml_int8x16x4_t int8x16x4_t
+
+#define ggml_vld1q_s16_x2 vld1q_s16_x2
+#define ggml_vld1q_u8_x2 vld1q_u8_x2
+#define ggml_vld1q_u8_x4 vld1q_u8_x4
+#define ggml_vld1q_s8_x2 vld1q_s8_x2
+#define ggml_vld1q_s8_x4 vld1q_s8_x4
+#define ggml_vqtbl1q_s8 vqtbl1q_s8
+#define ggml_vqtbl1q_u8 vqtbl1q_u8
+
+#endif // !defined(__aarch64__)
+
+#if !defined(__ARM_FEATURE_DOTPROD)
+
+// NOTE: this fallback produces the same total sum as native vdotq_s32 but with different per-lane grouping — do not use when individual lane values matter.
+inline static int32x4_t ggml_vdotq_s32(int32x4_t acc, int8x16_t a, int8x16_t b) {
+ const int16x8_t p0 = vmull_s8(vget_low_s8 (a), vget_low_s8 (b));
+ const int16x8_t p1 = vmull_s8(vget_high_s8(a), vget_high_s8(b));
+
+ return vaddq_s32(acc, vaddq_s32(vpaddlq_s16(p0), vpaddlq_s16(p1)));
+}
+
+#else
+
+#define ggml_vdotq_s32(a, b, c) vdotq_s32(a, b, c)
+
+#endif // !defined(__ARM_FEATURE_DOTPROD)
+
+static inline int32x4_t ggml_nvfp4_dot8(const int8x8_t q4_lo, const int8x8_t q8_lo,
+ const int8x8_t q4_hi, const int8x8_t q8_hi) {
+ const int16x8_t p_lo = vmull_s8(q4_lo, q8_lo);
+ const int16x8_t p_hi = vmull_s8(q4_hi, q8_hi);
+ const int32x4_t sum_lo = vpaddlq_s16(p_lo);
+ const int32x4_t sum_hi = vpaddlq_s16(p_hi);
+ return vaddq_s32(sum_lo, sum_hi);
+}
+
+#endif // defined(__ARM_NEON)
+
+#ifdef __wasm_simd128__
+#include <wasm_simd128.h>
+#endif
+
+#ifdef __POWER9_VECTOR__
+#include <altivec.h>
+#endif
+
+#if defined(_MSC_VER) || defined(__MINGW32__)
+#include <intrin.h>
+#elif defined(__SSE__) || defined(__SSE3__) || defined(__SSSE3__) || defined(__AVX__) || defined(__F16C__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__AVX512BF16__)
+#include <immintrin.h>
+#endif
+
+#ifdef __riscv_v_intrinsic
+#include <riscv_vector.h>
+#endif
+
+#if defined(__loongarch64)
+#if defined(__loongarch_asx)
+#include <lasxintrin.h>
+#endif
+#if defined(__loongarch_sx)
+#include <lsxintrin.h>
+#endif
+#endif
+
+#if defined(__VXE__) || defined(__VXE2__)
+#include <vecintrin.h>
+
+#define vec_neg(a) (-(a)) // Vector Negate
+#define vec_add(a, b) ((a) + (b)) // Vector Add
+#define vec_sub(a, b) ((a) - (b)) // Vector Subtract
+#define vec_mul(a, b) ((a) * (b)) // Vector Multiply
+#define vec_div(a, b) ((a) / (b)) // Vector Divide
+#define vec_sl(a, b) ((a) << (b)) // Vector Shift Left
+#define vec_sra(a, b) ((a) >> (b)) // Vector Shift Right
+#define vec_sr(a, b) ((a) >> (b)) // Vector Shift Right Algebraic
+#define vec_slo(a, b) vec_slb(a, (b) << 64) // Vector Shift Left by Octet
+#define vec_sro(a, b) vec_srb(a, (b) << 64) // Vector Shift Right by Octet
+
+#ifndef vec_and
+#define vec_and(a, b) ((a) & (b)) // Vector AND
+#endif
+
+#ifndef vec_or
+#define vec_or(a, b) ((a) | (b)) // Vector OR
+#endif
+
+#ifndef vec_xor
+#define vec_xor(a, b) ((a) ^ (b)) // Vector XOR
+#endif
+
+typedef signed char char8x16_t __attribute__((vector_size(16)));
+typedef unsigned char uchar8x16_t __attribute__((vector_size(16)));
+
+typedef int8_t int8x16_t __attribute__((vector_size(16)));
+typedef int16_t int16x8_t __attribute__((vector_size(16)));
+typedef int32_t int32x4_t __attribute__((vector_size(16)));
+typedef int64_t int64x2_t __attribute__((vector_size(16)));
+
+typedef uint8_t uint8x16_t __attribute__((vector_size(16)));
+typedef uint16_t uint16x8_t __attribute__((vector_size(16)));
+typedef uint32_t uint32x4_t __attribute__((vector_size(16)));
+typedef uint64_t uint64x2_t __attribute__((vector_size(16)));
+
+typedef float float32x4_t __attribute__((vector_size(16)));
+typedef double double64x2_t __attribute__((vector_size(16)));
+
+typedef struct ggml_uint8x16x2_t {
+ uint8x16_t val[2];
+} ggml_uint8x16x2_t;
+
+inline static ggml_uint8x16x2_t ggml_vec_xl_u8x2(const uint8_t * ptr) {
+ ggml_uint8x16x2_t res;
+
+ res.val[0] = vec_xl( 0, ptr);
+ res.val[1] = vec_xl(16, ptr);
+
+ return res;
+}
+
+typedef struct ggml_uint8x16x4_t {
+ uint8x16_t val[4];
+} ggml_uint8x16x4_t;
+
+inline static ggml_uint8x16x4_t ggml_vec_xl_u8x4(const uint8_t * ptr) {
+ ggml_uint8x16x4_t res;
+
+ res.val[0] = vec_xl( 0, ptr);
+ res.val[1] = vec_xl(16, ptr);
+ res.val[2] = vec_xl(32, ptr);
+ res.val[3] = vec_xl(48, ptr);
+
+ return res;
+}
+
+typedef struct ggml_int8x16x4_t {
+ int8x16_t val[4];
+} ggml_int8x16x4_t;
+
+inline static ggml_int8x16x4_t ggml_vec_xl_s8x4(const int8_t * ptr) {
+ ggml_int8x16x4_t res;
+
+ res.val[0] = vec_xl( 0, ptr);
+ res.val[1] = vec_xl(16, ptr);
+ res.val[2] = vec_xl(32, ptr);
+ res.val[3] = vec_xl(48, ptr);
+
+ return res;
+}
+
+typedef struct ggml_int16x8x2_t {
+ int16x8_t val[2];
+} ggml_int16x8x2_t;
+
+inline static ggml_int16x8x2_t ggml_vec_xl_s16x2(const int16_t * ptr) {
+ ggml_int16x8x2_t res;
+
+ res.val[0] = vec_xl( 0, ptr);
+ res.val[1] = vec_xl(16, ptr);
+
+ return res;
+}
+
+/*
+ ! WARNING: Very slow. Use vec_perm if possible. Refer to iq4_xs
+ ! or iq4_nl for example implementation.
+*/
+inline static int8x16_t ggml_vec_tbl(int8x16_t a, uint8x16_t b) {
+ int8x16_t res;
+
+ res[ 0] = a[b[ 0]];
+ res[ 1] = a[b[ 1]];
+ res[ 2] = a[b[ 2]];
+ res[ 3] = a[b[ 3]];
+ res[ 4] = a[b[ 4]];
+ res[ 5] = a[b[ 5]];
+ res[ 6] = a[b[ 6]];
+ res[ 7] = a[b[ 7]];
+ res[ 8] = a[b[ 8]];
+ res[ 9] = a[b[ 9]];
+ res[10] = a[b[10]];
+ res[11] = a[b[11]];
+ res[12] = a[b[12]];
+ res[13] = a[b[13]];
+ res[14] = a[b[14]];
+ res[15] = a[b[15]];
+
+ return res;
+}
+
+inline static int16x8_t vec_padd_s16(int16x8_t a, int16x8_t b) {
+ const uchar8x16_t v_maske = { 0, 1, 4, 5, 8, 9, 12, 13,
+ 16, 17, 20, 21, 24, 25, 28, 29 };
+
+ const int16x8_t v_abo = vec_pack((int32x4_t)a, (int32x4_t)b);
+ const int16x8_t v_abe = vec_perm(a, b, v_maske);
+ return v_abo + v_abe;
+}
+
+/**
+ * @see https://github.com/ggml-org/llama.cpp/pull/14037
+ */
+inline static float vec_hsum_f32x4(float32x4_t v) {
+ float32x4_t v_temp = v + vec_reve(v);
+ return v_temp[0] + v_temp[1];
+}
+
+inline static int32_t vec_hsum_i32x4(int32x4_t v) {
+ int32x4_t v_temp = v + vec_reve(v);
+ return v_temp[0] + v_temp[1];
+}
+
+inline static int32x4_t ggml_vec_dot(int32x4_t acc, int8x16_t a, int8x16_t b) {
+ const int16x8_t p = vec_mule(a, b) + vec_mulo(a, b);
+ return acc + (vec_unpackh(p) + vec_unpackl(p));
+}
+
+#endif
+
+#if defined(__loongarch_sx)
+/* float type data load instructions */
+static __m128 __lsx_vreplfr2vr_s(const float val) {
+ v4f32 res = {val, val, val, val};
+ return (__m128)res;
+}
+#endif
+
+#if defined(__loongarch_asx)
+static __m256 __lasx_xvreplfr2vr_s(const float val) {
+ v8f32 res = {val, val, val, val, val, val, val, val};
+ return (__m256)res;
+}
+#endif
+
+// TODO: move to ggml-threading
+void ggml_barrier(struct ggml_threadpool * tp);
+
+void ggml_threadpool_chunk_set(struct ggml_threadpool * tp, int value);
+int ggml_threadpool_chunk_add(struct ggml_threadpool * tp, int value);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/python/freetoken/kernel/csrc/ggml/ggml-cpu/quants.c b/python/freetoken/kernel/csrc/ggml/ggml-cpu/quants.c
new file mode 100644
index 00000000..5e36459f
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/ggml-cpu/quants.c
@@ -0,0 +1,1339 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+
+#include "ggml-cpu-impl.h"
+#include "simd-mappings.h"
+#include "ggml-quants.h"
+#include "quants.h"
+
+#include "arch-fallback.h"
+
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h> // for GGML_ASSERT
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+void quantize_row_q1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ quantize_row_q1_0_ref(x, y, k);
+}
+
+void quantize_row_q2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ quantize_row_q2_0_ref(x, y, k);
+}
+
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ quantize_row_q4_0_ref(x, y, k);
+}
+
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ quantize_row_q4_1_ref(x, y, k);
+}
+
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ quantize_row_q5_0_ref(x, y, k);
+}
+
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ quantize_row_q5_1_ref(x, y, k);
+}
+
+void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ quantize_row_q8_0_ref(x, y, k);
+}
+
+void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ quantize_row_q8_1_ref(x, y, k);
+}
+
+void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ quantize_row_mxfp4_ref(x, y, k);
+}
+
+void quantize_row_nvfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ quantize_row_nvfp4_ref(x, y, k);
+}
+
+//
+// 2-6 bit quantization in super-blocks
+//
+
+//========================- 2-bit (de)-quantization
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+ quantize_row_q2_K_ref(x, vy, k);
+}
+
+//========================= 3-bit (de)-quantization
+
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+ quantize_row_q3_K_ref(x, vy, k);
+}
+
+// ====================== 4-bit (de)-quantization
+
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+ assert(k % QK_K == 0);
+ block_q4_K * GGML_RESTRICT y = vy;
+ quantize_row_q4_K_ref(x, y, k);
+}
+
+// ====================== 5-bit (de)-quantization
+
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+ assert(k % QK_K == 0);
+ block_q5_K * GGML_RESTRICT y = vy;
+ quantize_row_q5_K_ref(x, y, k);
+}
+
+// ====================== 6-bit (de)-quantization
+
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+ assert(k % QK_K == 0);
+ block_q6_K * GGML_RESTRICT y = vy;
+ quantize_row_q6_K_ref(x, y, k);
+}
+
+// ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs)
+
+void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+ assert(k % QK_K == 0);
+ block_tq1_0 * GGML_RESTRICT y = vy;
+ quantize_row_tq1_0_ref(x, y, k);
+}
+
+void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+ assert(k % QK_K == 0);
+ block_tq2_0 * GGML_RESTRICT y = vy;
+ quantize_row_tq2_0_ref(x, y, k);
+}
+
+//===================================== Q8_K ==============================================
+
+void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ quantize_row_q8_K_ref(x, y, k);
+}
+
+//===================================== Dot products =================================
+
+void ggml_vec_dot_q1_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK1_0;
+ const int nb = n / qk;
+
+ assert(n % qk == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q1_0 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ float sumf = 0.0;
+
+ for (int i = 0; i < nb; i++) {
+ const float d0 = GGML_CPU_FP16_TO_FP32(x[i].d);
+
+ float sumi = 0.0f;
+
+ for (int k = 0; k < 4; k++) {
+ const block_q8_0 * GGML_RESTRICT yb = &y[i * 4 + k];
+ const float d1 = GGML_CPU_FP16_TO_FP32(yb->d);
+ int sumi_block = 0;
+
+ const uint8_t * GGML_RESTRICT bits = &x[i].qs[k * 4];
+ const int8_t * GGML_RESTRICT qy = yb->qs;
+
+ for (int b = 0; b < 4; ++b, qy += 8) {
+ const unsigned mask = bits[b];
+ sumi_block += ((mask & 0x01) ? qy[0] : -qy[0])
+ + ((mask & 0x02) ? qy[1] : -qy[1])
+ + ((mask & 0x04) ? qy[2] : -qy[2])
+ + ((mask & 0x08) ? qy[3] : -qy[3])
+ + ((mask & 0x10) ? qy[4] : -qy[4])
+ + ((mask & 0x20) ? qy[5] : -qy[5])
+ + ((mask & 0x40) ? qy[6] : -qy[6])
+ + ((mask & 0x80) ? qy[7] : -qy[7]);
+ }
+
+ sumi += d1 * sumi_block;
+ }
+
+ sumf += d0 * sumi;
+ }
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_q2_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK2_0;
+ const int nb = n / qk;
+
+ assert(n % qk == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q2_0 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ float sumf = 0.0f;
+
+ for (int i = 0; i < nb; i++) {
+ const float d0 = GGML_CPU_FP16_TO_FP32(x[i].d);
+
+ float sumi = 0.0f;
+
+ // group 64: one Q2_0 block (64 weights) maps to two Q8_0 blocks (2 * 32 = 64)
+ for (int k = 0; k < 2; k++) {
+ const block_q8_0 * GGML_RESTRICT yb = &y[i * 2 + k];
+ const float d1 = GGML_CPU_FP16_TO_FP32(yb->d);
+ int sumi_block = 0;
+
+ const uint8_t * GGML_RESTRICT qs = &x[i].qs[k * 8];
+ const int8_t * GGML_RESTRICT qy = yb->qs;
+
+ for (int b = 0; b < 8; ++b) {
+ const uint8_t byte = qs[b];
+ // Extract 4 two-bit values, map {0,1,2,3} -> {-1,0,1,2}
+ sumi_block += ((int)((byte >> 0) & 3) - 1) * qy[b*4 + 0];
+ sumi_block += ((int)((byte >> 2) & 3) - 1) * qy[b*4 + 1];
+ sumi_block += ((int)((byte >> 4) & 3) - 1) * qy[b*4 + 2];
+ sumi_block += ((int)((byte >> 6) & 3) - 1) * qy[b*4 + 3];
+ }
+
+ sumi += d1 * sumi_block;
+ }
+
+ sumf += d0 * sumi;
+ }
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK8_0;
+ const int nb = n / qk;
+
+ assert(n % qk == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q4_0 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ int ib = 0;
+ float sumf = 0;
+
+ for (; ib < nb; ++ib) {
+ int sumi0 = 0;
+ int sumi1 = 0;
+
+ for (int j = 0; j < qk/2; ++j) {
+ const int v0 = (x[ib].qs[j] & 0x0F) - 8;
+ const int v1 = (x[ib].qs[j] >> 4) - 8;
+
+ sumi0 += (v0 * y[ib].qs[j]);
+ sumi1 += (v1 * y[ib].qs[j + qk/2]);
+ }
+
+ int sumi = sumi0 + sumi1;
+ sumf += sumi*GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d);
+ }
+
+ *s = sumf;
+}
+
+// TODO: add WASM SIMD
+void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK8_1;
+ const int nb = n / qk;
+
+ assert(n % qk == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q4_1 * GGML_RESTRICT x = vx;
+ const block_q8_1 * GGML_RESTRICT y = vy;
+
+ int ib = 0;
+ float sumf = 0;
+
+ for (; ib < nb; ++ib) {
+ int sumi0 = 0;
+ int sumi1 = 0;
+
+ for (int j = 0; j < qk/2; ++j) {
+ const int v0 = (x[ib].qs[j] & 0x0F);
+ const int v1 = (x[ib].qs[j] >> 4);
+
+ sumi0 += (v0 * y[ib].qs[j]);
+ sumi1 += (v1 * y[ib].qs[j + qk/2]);
+ }
+
+ int sumi = sumi0 + sumi1;
+ sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+ }
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+ assert(n % QK_MXFP4 == 0);
+ static_assert(QK_MXFP4 == QK8_0, "QK_MXFP4 and QK8_0 must be the same");
+
+ const block_mxfp4 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_MXFP4;
+
+ int ib = 0;
+ float sumf = 0;
+
+ for (; ib < nb; ++ib) {
+ const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_E8M0_TO_FP32_HALF(x[ib].e);
+
+ int sumi1 = 0;
+ int sumi2 = 0;
+ for (int j = 0; j < QK_MXFP4/2; ++j) {
+ sumi1 += y[ib].qs[j + 0] * kvalues_mxfp4[x[ib].qs[j] & 0xf];
+ sumi2 += y[ib].qs[j + QK_MXFP4/2] * kvalues_mxfp4[x[ib].qs[j] >> 4];
+ }
+ sumf += d * (sumi1 + sumi2);
+ }
+ *s = sumf;
+}
+
+// NVFP4: super-block of 64 elements = 4 sub-blocks of 16 = 2 q8_0 blocks
+void ggml_vec_dot_nvfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+ assert(n % QK_NVFP4 == 0);
+
+ const block_nvfp4 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_NVFP4;
+
+ float sumf = 0;
+
+ for (int ib = 0; ib < nb; ++ib) {
+ for (int s_idx = 0; s_idx < 4; ++s_idx) {
+ const float d = ggml_ue4m3_to_fp32(x[ib].d[s_idx]);
+ const int q8_block = s_idx / 2;
+ const int q8_off = (s_idx % 2) * QK_NVFP4_SUB;
+ const float dy = GGML_CPU_FP16_TO_FP32(y[2*ib + q8_block].d);
+
+ int sumi_lo = 0, sumi_hi = 0;
+ for (int j = 0; j < QK_NVFP4_SUB/2; ++j) {
+ const uint8_t qv = x[ib].qs[s_idx*(QK_NVFP4_SUB/2) + j];
+ sumi_lo += y[2*ib + q8_block].qs[q8_off + j + 0] * kvalues_mxfp4[qv & 0xf];
+ sumi_hi += y[2*ib + q8_block].qs[q8_off + j + QK_NVFP4_SUB/2] * kvalues_mxfp4[qv >> 4];
+ }
+
+ sumf += dy * d * (sumi_lo + sumi_hi);
+ }
+ }
+ *s = sumf;
+}
+
+void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK8_0;
+ const int nb = n / qk;
+
+ int ib = 0;
+ float sumf = 0;
+
+ assert(n % qk == 0);
+ assert(qk == QK5_0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q5_0 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ for (; ib < nb; ++ib) {
+ uint32_t qh;
+ memcpy(&qh, x[ib].qh, sizeof(qh));
+
+ int sumi0 = 0;
+ int sumi1 = 0;
+
+ for (int j = 0; j < qk/2; ++j) {
+ const uint8_t xh_0 = ((qh & (1u << (j + 0 ))) >> (j + 0 )) << 4;
+ const uint8_t xh_1 = ((qh & (1u << (j + 16))) >> (j + 12));
+
+ const int32_t x0 = (int8_t)(((x[ib].qs[j] & 0x0F) | xh_0) - 16);
+ const int32_t x1 = (int8_t)(((x[ib].qs[j] >> 4) | xh_1) - 16);
+
+ sumi0 += (x0 * y[ib].qs[j]);
+ sumi1 += (x1 * y[ib].qs[j + qk/2]);
+ }
+
+ int sumi = sumi0 + sumi1;
+ sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d)) * sumi;
+ }
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK8_1;
+ const int nb = n / qk;
+
+ int ib = 0;
+ float sumf = 0;
+
+ assert(n % qk == 0);
+ assert(qk == QK5_1);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q5_1 * GGML_RESTRICT x = vx;
+ const block_q8_1 * GGML_RESTRICT y = vy;
+
+ for (; ib < nb; ++ib) {
+ uint32_t qh;
+ memcpy(&qh, x[ib].qh, sizeof(qh));
+
+ int sumi0 = 0;
+ int sumi1 = 0;
+
+ for (int j = 0; j < qk/2; ++j) {
+ const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10;
+ const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10;
+
+ const int32_t x0 = (x[ib].qs[j] & 0xF) | xh_0;
+ const int32_t x1 = (x[ib].qs[j] >> 4) | xh_1;
+
+ sumi0 += (x0 * y[ib].qs[j]);
+ sumi1 += (x1 * y[ib].qs[j + qk/2]);
+ }
+
+ int sumi = sumi0 + sumi1;
+ sumf += (GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d))*sumi + GGML_CPU_FP16_TO_FP32(x[ib].m)*GGML_CPU_FP16_TO_FP32(y[ib].s);
+ }
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ const int qk = QK8_0;
+ const int nb = n / qk;
+
+ assert(n % qk == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q8_0 * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ int ib = 0;
+ float sumf = 0;
+
+ for (; ib < nb; ++ib) {
+ int sumi = 0;
+
+ for (int j = 0; j < qk; j++) {
+ sumi += x[ib].qs[j]*y[ib].qs[j];
+ }
+
+ sumf += sumi*(GGML_CPU_FP16_TO_FP32(x[ib].d)*GGML_CPU_FP16_TO_FP32(y[ib].d));
+ }
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_tq1_0 * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
+
+ float sumf = 0.0f;
+
+ for (int i = 0; i < nb; ++i) {
+ int sum = 0;
+
+ for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
+ for (size_t l = 0; l < 5; ++l) {
+ for (size_t m = 0; m < 32; ++m) {
+ uint8_t q = x[i].qs[j + m] * pow3[l];
+ uint16_t xi = ((uint16_t) q * 3) >> 8;
+ sum += (xi - 1) * y[i].qs[j*5 + l*32 + m];
+ }
+ }
+ }
+ for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
+ for (size_t l = 0; l < 5; ++l) {
+ for (size_t m = 0; m < 16; ++m) {
+ uint8_t q = x[i].qs[j + m] * pow3[l];
+ uint16_t xi = ((uint16_t) q * 3) >> 8;
+ sum += (xi - 1) * y[i].qs[j*5 + l*16 + m];
+ }
+ }
+ }
+
+ for (size_t l = 0; l < 4; ++l) {
+ for (size_t j = 0; j < sizeof(x->qh); ++j) {
+ uint8_t q = x[i].qh[j] * pow3[l];
+ uint16_t xi = ((uint16_t) q * 3) >> 8;
+ sum += (xi - 1) * y[i].qs[sizeof(x->qs)*5 + l*sizeof(x->qh) + j];
+ }
+ }
+
+ sumf += (float) sum * (GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d);
+ }
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_tq2_0 * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+ float sumf = 0.0f;
+
+ for (int i = 0; i < nb; ++i) {
+ int32_t sumi = 0;
+
+ for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+ for (size_t l = 0; l < 4; ++l) {
+ for (size_t k = 0; k < 32; ++k) {
+ sumi += y[i].qs[j*4 + l*32 + k] * (((x[i].qs[j + k] >> (l*2)) & 3) - 1);
+ }
+ }
+ }
+
+ const float d = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+
+ sumf += (float) sumi * d;
+ }
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q2_K * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ float sumf = 0;
+
+ for (int i = 0; i < nb; ++i) {
+
+ const uint8_t * q2 = x[i].qs;
+ const int8_t * q8 = y[i].qs;
+ const uint8_t * sc = x[i].scales;
+
+ int summs = 0;
+ for (int j = 0; j < 16; ++j) {
+ summs += y[i].bsums[j] * (sc[j] >> 4);
+ }
+
+ const float dall = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].d);
+ const float dmin = y[i].d * GGML_CPU_FP16_TO_FP32(x[i].dmin);
+
+ int isum = 0;
+ int is = 0;
+ int d;
+ for (int k = 0; k < QK_K/128; ++k) {
+ int shift = 0;
+ for (int j = 0; j < 4; ++j) {
+ d = sc[is++] & 0xF;
+ int isuml = 0;
+ for (int l = 0; l < 16; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+ isum += d * isuml;
+ d = sc[is++] & 0xF;
+ isuml = 0;
+ for (int l = 16; l < 32; ++l) isuml += q8[l] * ((q2[l] >> shift) & 3);
+ isum += d * isuml;
+ shift += 2;
+ q8 += 32;
+ }
+ q2 += 32;
+ }
+ sumf += dall * isum - dmin * summs;
+ }
+ *s = sumf;
+}
+
+void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const uint32_t kmask1 = 0x03030303;
+ const uint32_t kmask2 = 0x0f0f0f0f;
+
+ const block_q3_K * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ // scalar version
+ // This function is written like this so the compiler can manage to vectorize most of it
+ // Using -Ofast, GCC and clang manage to produce code that is within a factor of 2 or so from the
+ // manually vectorized version above. Every other version I tried would run at least 4 times slower.
+ // The ideal situation would be if we could just write the code once, and the compiler would
+ // automatically produce the best possible set of machine instructions, instead of us having to manually
+ // write vectorized versions for AVX, ARM_NEON, etc.
+
+ int8_t aux8[QK_K];
+ int16_t aux16[8];
+ float sums [8];
+ int32_t aux32[8];
+ memset(sums, 0, 8*sizeof(float));
+
+ uint32_t auxs[4];
+ const int8_t * scales = (const int8_t*)auxs;
+
+ float sumf = 0;
+ for (int i = 0; i < nb; ++i) {
+ const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+ const uint8_t * GGML_RESTRICT hm = x[i].hmask;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ memset(aux32, 0, 8*sizeof(int32_t));
+ int8_t * GGML_RESTRICT a = aux8;
+ uint8_t m = 1;
+ for (int j = 0; j < QK_K; j += 128) {
+ for (int l = 0; l < 32; ++l) a[l] = q3[l] & 3;
+ for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+ a += 32; m <<= 1;
+ for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 2) & 3;
+ for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+ a += 32; m <<= 1;
+ for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 4) & 3;
+ for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+ a += 32; m <<= 1;
+ for (int l = 0; l < 32; ++l) a[l] = (q3[l] >> 6) & 3;
+ for (int l = 0; l < 32; ++l) a[l] -= (hm[l] & m ? 0 : 4);
+ a += 32; m <<= 1;
+ q3 += 32;
+ }
+ a = aux8;
+
+ memcpy(auxs, x[i].scales, 12);
+ uint32_t tmp = auxs[2];
+ auxs[2] = ((auxs[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
+ auxs[3] = ((auxs[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
+ auxs[0] = (auxs[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
+ auxs[1] = (auxs[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
+ for (int j = 0; j < QK_K/16; ++j) {
+ for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+ for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+ q8 += 8; a += 8;
+ for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+ for (int l = 0; l < 8; ++l) aux32[l] += (scales[j] - 32) * aux16[l];
+ q8 += 8; a += 8;
+ }
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+ }
+ for (int l = 0; l < 8; ++l) sumf += sums[l];
+ *s = sumf;
+}
+
+void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q4_K * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ static const uint32_t kmask1 = 0x3f3f3f3f;
+ static const uint32_t kmask2 = 0x0f0f0f0f;
+ static const uint32_t kmask3 = 0x03030303;
+
+ uint32_t utmp[4];
+
+ const uint8_t * scales = (const uint8_t*)&utmp[0];
+ const uint8_t * mins = (const uint8_t*)&utmp[2];
+
+ int8_t aux8[QK_K];
+ int16_t aux16[8];
+ float sums [8];
+ int32_t aux32[8];
+ memset(sums, 0, 8*sizeof(float));
+
+ float sumf = 0;
+ for (int i = 0; i < nb; ++i) {
+ const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ memset(aux32, 0, 8*sizeof(int32_t));
+ int8_t * GGML_RESTRICT a = aux8;
+ for (int j = 0; j < QK_K/64; ++j) {
+ for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+ a += 32;
+ for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
+ a += 32; q4 += 32;
+ }
+ memcpy(utmp, x[i].scales, 12);
+ utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+ const uint32_t uaux = utmp[1] & kmask1;
+ utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+ utmp[2] = uaux;
+ utmp[0] &= kmask1;
+
+ int sumi = 0;
+ for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+ a = aux8;
+ int is = 0;
+ for (int j = 0; j < QK_K/32; ++j) {
+ int32_t scale = scales[is++];
+ for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+ for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+ q8 += 8; a += 8;
+ for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+ for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+ q8 += 8; a += 8;
+ for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+ for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+ q8 += 8; a += 8;
+ for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+ for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+ q8 += 8; a += 8;
+ }
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+ const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+ sumf -= dmin * sumi;
+ }
+ for (int l = 0; l < 8; ++l) sumf += sums[l];
+ *s = sumf;
+}
+
+void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q5_K * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ static const uint32_t kmask1 = 0x3f3f3f3f;
+ static const uint32_t kmask2 = 0x0f0f0f0f;
+ static const uint32_t kmask3 = 0x03030303;
+
+ uint32_t utmp[4];
+
+ const uint8_t * scales = (const uint8_t*)&utmp[0];
+ const uint8_t * mins = (const uint8_t*)&utmp[2];
+
+ int8_t aux8[QK_K];
+ int16_t aux16[8];
+ float sums [8];
+ int32_t aux32[8];
+ memset(sums, 0, 8*sizeof(float));
+
+ float sumf = 0;
+ for (int i = 0; i < nb; ++i) {
+ const uint8_t * GGML_RESTRICT q4 = x[i].qs;
+ const uint8_t * GGML_RESTRICT hm = x[i].qh;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ memset(aux32, 0, 8*sizeof(int32_t));
+ int8_t * GGML_RESTRICT a = aux8;
+ uint8_t m = 1;
+ for (int j = 0; j < QK_K/64; ++j) {
+ for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] & 0xF);
+ for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+ a += 32; m <<= 1;
+ for (int l = 0; l < 32; ++l) a[l] = (int8_t)(q4[l] >> 4);
+ for (int l = 0; l < 32; ++l) a[l] += (hm[l] & m ? 16 : 0);
+ a += 32; m <<= 1;
+ q4 += 32;
+ }
+ memcpy(utmp, x[i].scales, 12);
+ utmp[3] = ((utmp[2] >> 4) & kmask2) | (((utmp[1] >> 6) & kmask3) << 4);
+ const uint32_t uaux = utmp[1] & kmask1;
+ utmp[1] = (utmp[2] & kmask2) | (((utmp[0] >> 6) & kmask3) << 4);
+ utmp[2] = uaux;
+ utmp[0] &= kmask1;
+
+ int sumi = 0;
+ for (int j = 0; j < QK_K/16; ++j) sumi += y[i].bsums[j] * mins[j/2];
+ a = aux8;
+ int is = 0;
+ for (int j = 0; j < QK_K/32; ++j) {
+ int32_t scale = scales[is++];
+ for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+ for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+ q8 += 8; a += 8;
+ for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+ for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+ q8 += 8; a += 8;
+ for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+ for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+ q8 += 8; a += 8;
+ for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+ for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+ q8 += 8; a += 8;
+ }
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+ const float dmin = GGML_CPU_FP16_TO_FP32(x[i].dmin) * y[i].d;
+ sumf -= dmin * sumi;
+ }
+ for (int l = 0; l < 8; ++l) sumf += sums[l];
+ *s = sumf;
+}
+
+void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_q6_K * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ int8_t aux8[QK_K];
+ int16_t aux16[8];
+ float sums [8];
+ int32_t aux32[8];
+ memset(sums, 0, 8*sizeof(float));
+
+ float sumf = 0;
+ for (int i = 0; i < nb; ++i) {
+ const uint8_t * GGML_RESTRICT q4 = x[i].ql;
+ const uint8_t * GGML_RESTRICT qh = x[i].qh;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ memset(aux32, 0, 8*sizeof(int32_t));
+ int8_t * GGML_RESTRICT a = aux8;
+ for (int j = 0; j < QK_K; j += 128) {
+ for (int l = 0; l < 32; ++l) {
+ a[l + 0] = (int8_t)((q4[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+ a[l + 32] = (int8_t)((q4[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+ a[l + 64] = (int8_t)((q4[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+ a[l + 96] = (int8_t)((q4[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+ }
+ a += 128;
+ q4 += 64;
+ qh += 32;
+ }
+ a = aux8;
+ int is = 0;
+ for (int j = 0; j < QK_K/16; ++j) {
+ int scale = x[i].scales[is++];
+ for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+ for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+ q8 += 8; a += 8;
+ for (int l = 0; l < 8; ++l) aux16[l] = q8[l] * a[l];
+ for (int l = 0; l < 8; ++l) aux32[l] += scale * aux16[l];
+ q8 += 8; a += 8;
+ }
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ for (int l = 0; l < 8; ++l) sums[l] += d * aux32[l];
+ }
+ for (int l = 0; l < 8; ++l) sumf += sums[l];
+ *s = sumf;
+}
+
+void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq2_xxs * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ uint32_t aux32[2];
+ const uint8_t * aux8 = (const uint8_t *)aux32;
+
+ float sumf = 0.f;
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ int32_t bsum = 0;
+ for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+ memcpy(aux32, q2, 2*sizeof(uint32_t));
+ q2 += 4;
+ const uint32_t ls = 2*(aux32[1] >> 28) + 1;
+ int32_t sumi = 0;
+ for (int l = 0; l < 4; ++l) {
+ const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+ const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+ for (int j = 0; j < 8; ++j) {
+ sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+ }
+ q8 += 8;
+ }
+ bsum += sumi * ls;
+ }
+ sumf += d * bsum;
+ }
+ *s = 0.125f * sumf;
+}
+
+void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq2_xs * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ float sumf = 0.f;
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint16_t * GGML_RESTRICT q2 = x[i].qs;
+ const uint8_t * GGML_RESTRICT sc = x[i].scales;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ int32_t bsum = 0;
+ for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+ const uint16_t ls1 = 2*(sc[ib32] & 0xf) + 1;
+ const uint16_t ls2 = 2*(sc[ib32] >> 4) + 1;
+ int32_t sumi = 0;
+ for (int l = 0; l < 2; ++l) {
+ const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+ const uint8_t signs = ksigns_iq2xs[q2[l] >> 9];
+ for (int j = 0; j < 8; ++j) {
+ sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+ }
+ q8 += 8;
+ }
+ bsum += sumi * ls1;
+ sumi = 0;
+ for (int l = 2; l < 4; ++l) {
+ const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (q2[l] & 511));
+ const uint8_t signs = ksigns_iq2xs[q2[l] >> 9];
+ for (int j = 0; j < 8; ++j) {
+ sumi += grid[j] * q8[j] * (signs & kmask_iq2xs[j] ? -1 : 1);
+ }
+ q8 += 8;
+ }
+ bsum += sumi * ls2;
+ q2 += 4;
+ }
+ sumf += d * bsum;
+ }
+ *s = 0.125f * sumf;
+}
+
+void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq2_s * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ float sumf = 0;
+ for (int i = 0; i < nb; i++) {
+
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const int8_t * q8 = y[i].qs;
+ const uint8_t * qs = x[i].qs;
+ const uint8_t * qh = x[i].qh;
+ const uint8_t * signs = qs + QK_K/8;
+
+ int bsum = 0;
+ for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+ int ls1 = 1 + 2*(x[i].scales[ib32] & 0xf);
+ int ls2 = 1 + 2*(x[i].scales[ib32] >> 4);
+ int sumi1 = 0, sumi2 = 0;
+ for (int l = 0; l < 2; ++l) {
+ const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+ for (int j = 0; j < 8; ++j) {
+ sumi1 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+ }
+ q8 += 8;
+ }
+ for (int l = 2; l < 4; ++l) {
+ const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+ for (int j = 0; j < 8; ++j) {
+ sumi2 += q8[j] * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1 : 1);
+ }
+ q8 += 8;
+ }
+ bsum += ls1 * sumi1 + ls2 * sumi2;
+ qs += 4;
+ signs += 4;
+ }
+
+ sumf += d * bsum;
+ }
+
+ *s = 0.125f * sumf;
+}
+
+void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq3_xxs * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ uint32_t aux32;
+
+ float sumf = 0.f;
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint8_t * GGML_RESTRICT q3 = x[i].qs;
+ const uint8_t * GGML_RESTRICT gas = x[i].qs + QK_K/4;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ int32_t bsum = 0;
+ for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+ memcpy(&aux32, gas, sizeof(uint32_t)); gas += sizeof(uint32_t);
+ const uint32_t ls = 2*(aux32 >> 28) + 1;
+ int32_t sumi = 0;
+ for (int l = 0; l < 4; ++l) {
+ const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + q3[2*l+0]);
+ const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + q3[2*l+1]);
+ const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+ for (int j = 0; j < 4; ++j) {
+ sumi += grid1[j] * q8[j+0] * (signs & kmask_iq2xs[j+0] ? -1 : 1);
+ sumi += grid2[j] * q8[j+4] * (signs & kmask_iq2xs[j+4] ? -1 : 1);
+ }
+ q8 += 8;
+ }
+ q3 += 8;
+ bsum += sumi * ls;
+ }
+ sumf += d * bsum;
+ }
+ *s = 0.25f * sumf;
+}
+
+void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq3_s * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ float sumf = 0.f;
+ for (int i = 0; i < nb; ++i) {
+ const float d = GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d;
+ const uint8_t * GGML_RESTRICT qs = x[i].qs;
+ const uint8_t * GGML_RESTRICT qh = x[i].qh;
+ const uint8_t * GGML_RESTRICT signs = x[i].signs;
+ const int8_t * GGML_RESTRICT q8 = y[i].qs;
+ int32_t bsum = 0;
+ for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+ const uint32_t ls1 = 2*(x[i].scales[ib32/2] & 0xf) + 1;
+ const uint32_t ls2 = 2*(x[i].scales[ib32/2] >> 4) + 1;
+ int32_t sumi = 0;
+ for (int l = 0; l < 4; ++l) {
+ const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+0] << (8-2*l)) & 256)));
+ const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+0] << (7-2*l)) & 256)));
+ for (int j = 0; j < 4; ++j) {
+ sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+ sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+ }
+ q8 += 8;
+ }
+ qs += 8;
+ signs += 4;
+ bsum += sumi * ls1;
+ sumi = 0;
+ for (int l = 0; l < 4; ++l) {
+ const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[ib32+1] << (8-2*l)) & 256)));
+ const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[ib32+1] << (7-2*l)) & 256)));
+ for (int j = 0; j < 4; ++j) {
+ sumi += grid1[j] * q8[j+0] * (signs[l] & kmask_iq2xs[j+0] ? -1 : 1);
+ sumi += grid2[j] * q8[j+4] * (signs[l] & kmask_iq2xs[j+4] ? -1 : 1);
+ }
+ q8 += 8;
+ }
+ qs += 8;
+ signs += 4;
+ bsum += sumi * ls2;
+ }
+ sumf += d * bsum;
+ }
+ *s = sumf;
+}
+
+void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq1_s * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ float sumf = 0;
+ for (int i = 0; i < nb; i++) {
+
+ const int8_t * q8 = y[i].qs;
+ const uint8_t * qs = x[i].qs;
+ const uint16_t * qh = x[i].qh;
+
+ int sumi = 0, sumi1 = 0;
+ for (int ib = 0; ib < QK_K/32; ++ib) {
+ const int ls = 2*((qh[ib] >> 12) & 7) + 1;
+ const int delta = qh[ib] & 0x8000 ? -1 : 1;
+ int lsum = 0;
+ for (int l = 0; l < 4; ++l) {
+ const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
+ for (int j = 0; j < 8; ++j) {
+ lsum += q8[j] * grid[j];
+ }
+ q8 += 8;
+ }
+ sumi += ls * lsum;
+ sumi1 += ls * delta * (y[i].bsums[2*ib+0] + y[i].bsums[2*ib+1]);
+ qs += 4;
+ }
+
+ sumf += GGML_CPU_FP16_TO_FP32(x[i].d) * y[i].d * (sumi + IQ1S_DELTA * sumi1);
+ }
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(n % QK_K == 0);
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+
+ const block_iq1_m * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ iq1m_scale_t scale;
+
+ int sum1[2], sum2[2], delta[4];
+
+ float sumf = 0;
+ for (int i = 0; i < nb; i++) {
+
+ const int8_t * q8 = y[i].qs;
+ const uint8_t * qs = x[i].qs;
+ const uint8_t * qh = x[i].qh;
+ const uint16_t * sc = (const uint16_t *)x[i].scales;
+
+ scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+
+ int sumi1 = 0, sumi2 = 0;
+ for (int ib = 0; ib < QK_K/32; ++ib) {
+ delta[0] = qh[0] & 0x08 ? -1 : 1;
+ delta[1] = qh[0] & 0x80 ? -1 : 1;
+ delta[2] = qh[1] & 0x08 ? -1 : 1;
+ delta[3] = qh[1] & 0x80 ? -1 : 1;
+ sum1[0] = sum1[1] = sum2[0] = sum2[1] = 0;
+ for (int l = 0; l < 4; ++l) {
+ const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((uint16_t)qh[l/2] << (8 - 4*(l%2))) & 0x700)));
+ int lsum1 = 0, lsum2 = 0;
+ for (int j = 0; j < 8; ++j) {
+ lsum1 += q8[j] * grid[j];
+ lsum2 += q8[j];
+ }
+ q8 += 8;
+ sum1[l/2] += lsum1;
+ sum2[l/2] += lsum2*delta[l];
+ }
+
+ const int ls1 = 2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1;
+ const int ls2 = 2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1;
+
+ sumi1 += sum1[0] * ls1 + sum1[1] * ls2;
+ sumi2 += sum2[0] * ls1 + sum2[1] * ls2;
+ qs += 4;
+ qh += 2;
+ }
+
+ sumf += GGML_CPU_FP16_TO_FP32(scale.f16) * y[i].d * (sumi1 + IQ1M_DELTA * sumi2);
+ }
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+ assert(n % QK4_NL == 0);
+ static_assert(QK4_NL == QK8_0, "QK4_NL and QK8_0 must be the same");
+
+ const block_iq4_nl * GGML_RESTRICT x = vx;
+ const block_q8_0 * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK4_NL;
+
+ int ib = 0;
+ float sumf = 0;
+
+ for (; ib < nb; ++ib) {
+ const float d = GGML_CPU_FP16_TO_FP32(y[ib].d)*GGML_CPU_FP16_TO_FP32(x[ib].d);
+ int sumi1 = 0, sumi2 = 0;
+ for (int j = 0; j < QK4_NL/2; ++j) {
+ sumi1 += y[ib].qs[j+ 0] * kvalues_iq4nl[x[ib].qs[j] & 0xf];
+ sumi2 += y[ib].qs[j+QK4_NL/2] * kvalues_iq4nl[x[ib].qs[j] >> 4];
+ }
+ sumf += d * (sumi1 + sumi2);
+ }
+ *s = sumf;
+}
+
+void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc) {
+ assert(nrc == 1);
+ UNUSED(nrc);
+ UNUSED(bx);
+ UNUSED(by);
+ UNUSED(bs);
+ assert(n % QK_K == 0);
+
+ const block_iq4_xs * GGML_RESTRICT x = vx;
+ const block_q8_K * GGML_RESTRICT y = vy;
+
+ const int nb = n / QK_K;
+
+ float sumf = 0;
+ for (int ibl = 0; ibl < nb; ++ibl) {
+ const float d4d8 = GGML_CPU_FP16_TO_FP32(x[ibl].d) * y[ibl].d;
+ uint16_t h = x[ibl].scales_h;
+ const uint8_t * qs = x[ibl].qs;
+ const int8_t * q8 = y[ibl].qs;
+ for (int ib = 0; ib < QK_K/32; ib += 2) {
+ const uint8_t ls1 = (x[ibl].scales_l[ib/2] & 0xf) | ((h << 4) & 0x30);
+ const uint8_t ls2 = (x[ibl].scales_l[ib/2] >> 4) | ((h << 2) & 0x30);
+ h >>= 4;
+ const float d1 = d4d8*(ls1 - 32);
+ const float d2 = d4d8*(ls2 - 32);
+ int sumi1 = 0, sumi2 = 0;
+ for (int j = 0; j < 16; ++j) {
+ sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+ sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4];
+ }
+ sumf += d1 * (sumi1 + sumi2);
+ qs += 16;
+ q8 += 32;
+ sumi1 = sumi2 = 0;
+ for (int j = 0; j < 16; ++j) {
+ sumi1 += q8[j+ 0] * kvalues_iq4nl[qs[j] & 0xf];
+ sumi2 += q8[j+16] * kvalues_iq4nl[qs[j] >> 4];
+ }
+ sumf += d2 * (sumi1 + sumi2);
+ qs += 16;
+ q8 += 32;
+ }
+ }
+ *s = sumf;
+}
+
+// ============================ 4-bit non-linear quants
+
+void quantize_row_iq4_nl(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK4_NL == 0);
+ quantize_row_iq4_nl_ref(x, y, k);
+}
+
+void quantize_row_iq4_xs(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ quantize_iq4_xs(x, y, 1, k, NULL);
+}
diff --git a/python/freetoken/kernel/csrc/ggml/ggml-cpu/quants.h b/python/freetoken/kernel/csrc/ggml/ggml-cpu/quants.h
new file mode 100644
index 00000000..93ea7eef
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/ggml-cpu/quants.h
@@ -0,0 +1,106 @@
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML CPU internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// Quantization
+void quantize_row_q1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_1(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_mxfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_nvfp4(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_q2_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q3_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q4_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q5_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q6_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_q8_K(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_tq1_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_tq2_0(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+void quantize_row_iq4_nl (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void quantize_row_iq4_xs (const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+
+// Dot product
+void ggml_vec_dot_q1_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q2_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_mxfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_nvfp4_q8_0(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_q2_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_tq1_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq2_0_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_iq2_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0 (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+// Generic implementation
+void quantize_row_q8_0_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void quantize_row_q8_1_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k);
+void quantize_row_q8_K_generic(const float * GGML_RESTRICT x, void * GGML_RESTRICT y, int64_t k);
+void ggml_vec_dot_q1_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q2_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_1_q8_1_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q8_0_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_mxfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_nvfp4_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_tq1_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_tq2_0_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+void ggml_vec_dot_q2_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q3_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q4_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q5_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_q6_K_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq2_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_xxs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq3_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_s_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq1_m_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_nl_q8_0_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+void ggml_vec_dot_iq4_xs_q8_K_generic(int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT vx, size_t bx, const void * GGML_RESTRICT vy, size_t by, int nrc);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/python/freetoken/kernel/csrc/ggml/ggml-cpu/simd-mappings.h b/python/freetoken/kernel/csrc/ggml/ggml-cpu/simd-mappings.h
new file mode 100644
index 00000000..fb53ecdf
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/ggml-cpu/simd-mappings.h
@@ -0,0 +1,1366 @@
+#pragma once
+
+#include "ggml-cpu-impl.h"
+
+#ifdef __ARM_FEATURE_SVE
+#include <arm_sve.h>
+#endif // __ARM_FEATURE_SVE
+
+#if defined(__ARM_NEON) && !defined(__CUDACC__) && !defined(__MUSACC__)
+// if YCM cannot find <arm_neon.h>, make a symbolic link to it, for example:
+//
+// $ ln -sfn /Library/Developer/CommandLineTools/usr/lib/clang/13.1.6/include/arm_neon.h ./src/
+//
+#include <arm_neon.h>
+#endif
+
+#if defined(__riscv_v_intrinsic)
+#include <riscv_vector.h>
+#endif
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//
+// simd mappings
+//
+
+// FP16 to FP32 conversion
+
+// 16-bit float
+// on Arm, we use __fp16, which requires the IEEE fp16 format: implied on
+// AArch64, selected by -mfp16-format=ieee on 32 bit Arm, where the compiler
+// may otherwise reject the type
+// on x86, we use uint16_t
+//
+// for old CUDA compilers (<= 11), we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/10616
+// for MUSA compilers , we use uint16_t: ref https://github.com/ggml-org/llama.cpp/pull/11843
+//
+#if defined(__ARM_NEON) && defined(__ARM_FP16_FORMAT_IEEE) && !(defined(__CUDACC__) && __CUDACC_VER_MAJOR__ <= 11) && !defined(__MUSACC__)
+ #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) neon_compute_fp16_to_fp32(x)
+ #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) neon_compute_fp32_to_fp16(x)
+
+ #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+
+ static inline float neon_compute_fp16_to_fp32(ggml_fp16_t h) {
+ __fp16 tmp;
+ memcpy(&tmp, &h, sizeof(ggml_fp16_t));
+ return (float)tmp;
+ }
+
+ static inline ggml_fp16_t neon_compute_fp32_to_fp16(float f) {
+ ggml_fp16_t res;
+ __fp16 tmp = f;
+ memcpy(&res, &tmp, sizeof(ggml_fp16_t));
+ return res;
+ }
+#elif defined(__F16C__)
+ #ifdef _MSC_VER
+ #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) _mm_cvtss_f32(_mm_cvtph_ps(_mm_cvtsi32_si128(x)))
+ #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) _mm_extract_epi16(_mm_cvtps_ph(_mm_set_ss(x), 0), 0)
+ #else
+ #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) _cvtsh_ss(x)
+ #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) _cvtss_sh(x, 0)
+ #endif
+#elif defined(__POWER9_VECTOR__)
+ #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) power_compute_fp16_to_fp32(x)
+ #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) power_compute_fp32_to_fp16(x)
+ /* the inline asm below is about 12% faster than the lookup method */
+ #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+ #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
+
+ static inline float power_compute_fp16_to_fp32(ggml_fp16_t h) {
+ float f;
+ double d;
+ __asm__(
+ "mtfprd %0,%2\n"
+ "xscvhpdp %0,%0\n"
+ "frsp %1,%0\n" :
+ /* temp */ "=d"(d),
+ /* out */ "=f"(f):
+ /* in */ "r"(h));
+ return f;
+ }
+
+ static inline ggml_fp16_t power_compute_fp32_to_fp16(float f) {
+ double d;
+ ggml_fp16_t r;
+ __asm__( /* xscvdphp can work on double or single precision */
+ "xscvdphp %0,%2\n"
+ "mffprd %1,%0\n" :
+ /* temp */ "=d"(d),
+ /* out */ "=r"(r):
+ /* in */ "f"(f));
+ return r;
+ }
+#elif defined(__riscv) && defined(__riscv_zfhmin)
+ static inline float riscv_compute_fp16_to_fp32(ggml_fp16_t h) {
+ _Float16 hf;
+ memcpy(&hf, &h, sizeof(ggml_fp16_t));
+ return hf;
+ }
+
+ static inline ggml_fp16_t riscv_compute_fp32_to_fp16(float f) {
+ ggml_fp16_t res;
+ _Float16 hf = (_Float16)f;
+ memcpy(&res, &hf, sizeof(ggml_fp16_t));
+ return res;
+ }
+
+ #define GGML_CPU_COMPUTE_FP16_TO_FP32(x) riscv_compute_fp16_to_fp32(x)
+ #define GGML_CPU_COMPUTE_FP32_TO_FP16(x) riscv_compute_fp32_to_fp16(x)
+ #define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+ #define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+// precomputed f32 table for f16 (256 KB)
+// defined in ggml-cpu.c, initialized in ggml_cpu_init()
+extern float ggml_table_f32_f16[1 << 16];
+
+// precomputed f32 table for e8m0 half (1 KB)
+// defined in ggml-cpu.c, initialized in ggml_cpu_init()
+extern float ggml_table_f32_e8m0_half[1 << 8];
+
+// precomputed f32 table for ue4m3 (1 KB)
+// defined in ggml-cpu.c, initialized in ggml_cpu_init()
+extern float ggml_table_f32_ue4m3[1 << 8];
+
+// Use lookup table for E8M0 on x86 (faster than bit manipulation)
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+#define GGML_CPU_E8M0_TO_FP32_HALF(x) ggml_table_f32_e8m0_half[(uint8_t)(x)]
+#else
+#define GGML_CPU_E8M0_TO_FP32_HALF(x) GGML_E8M0_TO_FP32_HALF(x)
+#endif
+
+// Use lookup table for UE4M3 on x86 and ARM (faster than bit manipulation)
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__) || defined(__ARM_NEON)
+#define GGML_CPU_UE4M3_TO_FP32(x) ggml_table_f32_ue4m3[(uint8_t)(x)]
+#else
+#define GGML_CPU_UE4M3_TO_FP32(x) ggml_ue4m3_to_fp32(x)
+#endif
+
+// Pre AVX512 CPUs (AVX2 with F16C, such as i5-8th gen or i7-4790)
+// do not produce codegen as checks below don't test for __F16C__
+// but this check does.
+#if defined(__F16C__)
+#define GGML_CPU_FP16_TO_FP32(x) GGML_CPU_COMPUTE_FP16_TO_FP32(x)
+#define GGML_CPU_FP32_TO_FP16(x) GGML_CPU_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+// On ARM NEON, it's quicker to directly convert x -> x instead of calling into ggml_lookup_fp16_to_fp32,
+// so we define GGML_CPU_FP16_TO_FP32 and GGML_CPU_FP32_TO_FP16 elsewhere for NEON.
+// This is also true for POWER9.
+#if !defined(GGML_CPU_FP16_TO_FP32)
+inline static float ggml_lookup_fp16_to_fp32(ggml_fp16_t f) {
+ uint16_t s;
+ memcpy(&s, &f, sizeof(uint16_t));
+ return ggml_table_f32_f16[s];
+}
+
+#define GGML_CPU_FP16_TO_FP32(x) ggml_lookup_fp16_to_fp32(x)
+#endif
+
+#if !defined(GGML_CPU_FP32_TO_FP16)
+#define GGML_CPU_FP32_TO_FP16(x) GGML_COMPUTE_FP32_TO_FP16(x)
+#endif
+
+
+// we define a common set of C macros which map to specific intrinsics based on the current architecture
+// we then implement the fundamental computation operations below using only these macros
+// adding support for new architectures requires to define the corresponding SIMD macros
+//
+// GGML_F32_STEP / GGML_F16_STEP
+// number of elements to process in a single step
+//
+// GGML_F32_EPR / GGML_F16_EPR
+// number of elements to fit in a single register
+//
+
+#if defined(__ARM_FEATURE_SVE) && defined(__ARM_FEATURE_FMA)
+
+#define GGML_SIMD
+
+// F32 SVE
+#define GGML_F32_EPR 8
+#define DEFAULT_PG svptrue_b32()
+
+#define GGML_F32xt svfloat32_t
+#define GGML_F32xt_ZERO svdup_n_f32(0.0f)
+#define GGML_F32xt_SET1(x) svdup_n_f32(x)
+#define GGML_F32xt_LOAD_IMPL(pg, a) svld1_f32(pg, a)
+#define GGML_F32xt_LOAD(a) GGML_F32xt_LOAD_IMPL(DEFAULT_PG, a)
+#define GGML_F32xt_STORE_IMPL(pg, a, b) svst1_f32(pg, a, b)
+#define GGML_F32xt_STORE(a, b) GGML_F32xt_STORE_IMPL(DEFAULT_PG, a, b)
+#define GGML_F32xt_FMA_IMPL(pg, a, b, c) svmad_f32_m(pg, b, c, a)
+#define GGML_F32xt_FMA(a, b, c) GGML_F32xt_FMA_IMPL(DEFAULT_PG, a, b, c)
+#define GGML_F32xt_ADD_IMPL(pg, a, b) svadd_f32_m(pg, a, b)
+#define GGML_F32xt_ADD(a, b) GGML_F32xt_ADD_IMPL(DEFAULT_PG, a, b)
+#define GGML_F32xt_MUL_IMPL(pg, a, b) svmul_f32_m(pg, a, b)
+#define GGML_F32xt_MUL(a, b) GGML_F32xt_MUL_IMPL(DEFAULT_PG, a, b)
+#define GGML_F32xt_REDUCE_ONE_IMPL(pg, a) svaddv(pg, a)
+#define GGML_F32xt_REDUCE_ONE(a) GGML_F32xt_REDUCE_ONE_IMPL(DEFAULT_PG, a)
+#define GGML_F32xt_REDUCE_IMPL(pg, res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8) \
+{ \
+ sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum2); \
+ sum3 = svadd_f32_m(DEFAULT_PG, sum3, sum4); \
+ sum5 = svadd_f32_m(DEFAULT_PG, sum5, sum6); \
+ sum7 = svadd_f32_m(DEFAULT_PG, sum7, sum8); \
+ sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum3); \
+ sum5 = svadd_f32_m(DEFAULT_PG, sum5, sum7); \
+ sum1 = svadd_f32_m(DEFAULT_PG, sum1, sum5); \
+ (res) = (ggml_float) GGML_F32xt_REDUCE_ONE(sum1); \
+}
+#define GGML_F32xt_REDUCE(res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8) \
+ GGML_F32xt_REDUCE_IMPL(DEFAULT_PG, res, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8)
+
+#define GGML_F32_VEC GGML_F32xt
+#define GGML_F32_VEC_ZERO GGML_F32xt_ZERO
+#define GGML_F32_VEC_SET1 GGML_F32xt_SET1
+#define GGML_F32_VEC_LOAD GGML_F32xt_LOAD
+#define GGML_F32_VEC_STORE GGML_F32xt_STORE
+#define GGML_F32_VEC_FMA GGML_F32xt_FMA
+#define GGML_F32_VEC_ADD GGML_F32xt_ADD
+#define GGML_F32_VEC_MUL GGML_F32xt_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32xt_REDUCE
+
+// F16 SVE
+#define DEFAULT_PG32 svptrue_b32()
+#define DEFAULT_PG16 svptrue_b16()
+
+#define GGML_F32Cxt svfloat16_t
+#define GGML_F32Cxt_ZERO svdup_n_f16(0.0f)
+#define GGML_F32Cxt_SET1(x) svdup_n_f16(x)
+#define GGML_F32Cxt_LOAD(p) svld1_f16(DEFAULT_PG16, (const __fp16 *)(p))
+#define GGML_F32Cxt_STORE(dst_ptr, src_vec) svst1_f16(DEFAULT_PG16, (__fp16 *)(dst_ptr), (src_vec))
+
+#define GGML_F32Cxt_FMA_IMPL(pg, a, b, c) svmad_f16_x(pg, b, c, a)
+#define GGML_F32Cxt_FMA(a, b, c) GGML_F32Cxt_FMA_IMPL(DEFAULT_PG16, a, b, c)
+#define GGML_F32Cxt_ADD_IMPL(pg, a, b) svadd_f16_x(pg, a, b)
+#define GGML_F32Cxt_ADD(a, b) GGML_F32Cxt_ADD_IMPL(DEFAULT_PG16, a, b)
+#define GGML_F32Cxt_MUL_IMPL(pg, a, b) svmul_f16_x(pg, a, b)
+#define GGML_F32Cxt_MUL(a, b) GGML_F32Cxt_MUL_IMPL(DEFAULT_PG16, a, b)
+#define GGML_F32Cxt_REDUCE GGML_F16xt_REDUCE_MIXED
+
+#define GGML_F16x_VEC GGML_F32Cxt
+#define GGML_F16x_VEC_ZERO GGML_F32Cxt_ZERO
+#define GGML_F16x_VEC_SET1 GGML_F32Cxt_SET1
+#define GGML_F16x_VEC_LOAD(p, i) GGML_F32Cxt_LOAD(p)
+#define GGML_F16x_VEC_STORE(p, r, i) GGML_F32Cxt_STORE((__fp16 *)(p), r)
+#define GGML_F16x_VEC_FMA GGML_F32Cxt_FMA
+#define GGML_F16x_VEC_ADD GGML_F32Cxt_ADD
+#define GGML_F16x_VEC_MUL GGML_F32Cxt_MUL
+#define GGML_F16x_VEC_REDUCE GGML_F32Cxt_REDUCE
+
+#define GGML_F16xt_REDUCE_ONE_IMPL(pg, a) svaddv_f16(pg, a)
+#define GGML_F16xt_REDUCE_ONE(a) GGML_F16xt_REDUCE_ONE_IMPL(DEFAULT_PG16, a)
+
+#define GGML_F16xt_REDUCE_MIXED_IMPL(pg16, res, sum1, sum2, sum3, sum4) \
+{ \
+ sum1 = svadd_f16_x(pg16, sum1, sum2); \
+ sum3 = svadd_f16_x(pg16, sum3, sum4); \
+ sum1 = svadd_f16_x(pg16, sum1, sum3); \
+ __fp16 sum_f16 = svaddv_f16(pg16, sum1); \
+ (res) = (ggml_float) sum_f16; \
+}
+#define GGML_F16xt_REDUCE_MIXED(res, sum1, sum2, sum3, sum4) \
+ GGML_F16xt_REDUCE_MIXED_IMPL(DEFAULT_PG16, res, sum1, sum2, sum3, sum4)
+
+// F16 NEON
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+ #define GGML_F16_STEP 32
+ #define GGML_F16_EPR 8
+
+ #define GGML_F16x8 float16x8_t
+ #define GGML_F16x8_ZERO vdupq_n_f16(0.0f)
+ #define GGML_F16x8_SET1(x) vdupq_n_f16(x)
+ #define GGML_F16x8_LOAD(x) vld1q_f16((const __fp16 *)(x))
+ #define GGML_F16x8_STORE vst1q_f16
+ #define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c)
+ #define GGML_F16x8_ADD vaddq_f16
+ #define GGML_F16x8_MUL vmulq_f16
+ #define GGML_F16x8_REDUCE(res, x) \
+ do { \
+ int offset = GGML_F16_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ (x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ (x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ (x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
+ } \
+ const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 ((x)[0])); \
+ const float32x4_t t1 = vcvt_f32_f16(vget_high_f16((x)[0])); \
+ (res) = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1)); \
+ } while (0)
+
+ #define GGML_F16_VEC GGML_F16x8
+ #define GGML_F16_VEC_ZERO GGML_F16x8_ZERO
+ #define GGML_F16_VEC_SET1 GGML_F16x8_SET1
+ #define GGML_F16_VEC_LOAD(p, i) GGML_F16x8_LOAD(p)
+ #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((__fp16 *)(p), (r)[i])
+ #define GGML_F16_VEC_FMA GGML_F16x8_FMA
+ #define GGML_F16_VEC_ADD GGML_F16x8_ADD
+ #define GGML_F16_VEC_MUL GGML_F16x8_MUL
+ #define GGML_F16_VEC_REDUCE GGML_F16x8_REDUCE
+#else
+ // if FP16 vector arithmetic is not supported, we use FP32 instead
+ // and take advantage of the vcvt_ functions to convert to/from FP16
+
+ #define GGML_F16_STEP 16
+ #define GGML_F16_EPR 4
+
+ #define GGML_F32Cx4 float32x4_t
+ #define GGML_F32Cx4_ZERO vdupq_n_f32(0.0f)
+ #define GGML_F32Cx4_SET1(x) vdupq_n_f32(x)
+ #define GGML_F32Cx4_LOAD(x) vcvt_f32_f16(vld1_f16((const __fp16 *)(x)))
+ #define GGML_F32Cx4_STORE(x, y) vst1_f16(x, vcvt_f16_f32(y))
+ #define GGML_F32Cx4_FMA(a, b, c) vfmaq_f32(a, b, c)
+ #define GGML_F32Cx4_ADD vaddq_f32
+ #define GGML_F32Cx4_MUL vmulq_f32
+ #define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE
+
+ #define GGML_F16_VEC GGML_F32Cx4
+ #define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO
+ #define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1
+ #define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p)
+ #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE((__fp16 *)(p), r[i])
+ #define GGML_F16_VEC_FMA GGML_F32Cx4_FMA
+ #define GGML_F16_VEC_ADD GGML_F32Cx4_ADD
+ #define GGML_F16_VEC_MUL GGML_F32Cx4_MUL
+ #define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE
+#endif
+
+#elif defined(__ARM_NEON) && defined(__ARM_FEATURE_FMA) && defined(__ARM_FP16_FORMAT_IEEE)
+
+#define GGML_SIMD
+
+// F32 NEON
+
+#define GGML_F32_STEP 16
+#define GGML_F32_EPR 4
+
+#define GGML_F32x4 float32x4_t
+#define GGML_F32x4_ZERO vdupq_n_f32(0.0f)
+#define GGML_F32x4_SET1(x) vdupq_n_f32(x)
+#define GGML_F32x4_LOAD vld1q_f32
+#define GGML_F32x4_STORE vst1q_f32
+#define GGML_F32x4_FMA(a, b, c) vfmaq_f32(a, b, c)
+#define GGML_F32x4_ADD vaddq_f32
+#define GGML_F32x4_MUL vmulq_f32
+#define GGML_F32x4_REDUCE_ONE(x) vaddvq_f32(x)
+#define GGML_F32x4_REDUCE(res, x) \
+{ \
+ int offset = GGML_F32_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ (x)[i] = vaddq_f32((x)[i], (x)[offset+i]); \
+ } \
+ (res) = (ggml_float) GGML_F32x4_REDUCE_ONE((x)[0]); \
+}
+
+#define GGML_F32_VEC GGML_F32x4
+#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
+#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
+#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
+#define GGML_F32_VEC_STORE GGML_F32x4_STORE
+#define GGML_F32_VEC_FMA GGML_F32x4_FMA
+#define GGML_F32_VEC_ADD GGML_F32x4_ADD
+#define GGML_F32_VEC_MUL GGML_F32x4_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
+
+// F16 NEON
+
+#if defined(__ARM_FEATURE_FP16_VECTOR_ARITHMETIC)
+ #define GGML_F16_STEP 32
+ #define GGML_F16_EPR 8
+
+ #define GGML_F16x8 float16x8_t
+ #define GGML_F16x8_ZERO vdupq_n_f16(0.0f)
+ #define GGML_F16x8_SET1(x) vdupq_n_f16(x)
+ #define GGML_F16x8_LOAD(x) vld1q_f16((const __fp16 *)(x))
+ #define GGML_F16x8_STORE vst1q_f16
+ #define GGML_F16x8_FMA(a, b, c) vfmaq_f16(a, b, c)
+ #define GGML_F16x8_ADD vaddq_f16
+ #define GGML_F16x8_MUL vmulq_f16
+ #define GGML_F16x8_REDUCE(res, x) \
+ do { \
+ int offset = GGML_F16_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ (x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ (x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ (x)[i] = vaddq_f16((x)[i], (x)[offset+i]); \
+ } \
+ const float32x4_t t0 = vcvt_f32_f16(vget_low_f16 ((x)[0])); \
+ const float32x4_t t1 = vcvt_f32_f16(vget_high_f16((x)[0])); \
+ (res) = (ggml_float) vaddvq_f32(vaddq_f32(t0, t1)); \
+ } while (0)
+
+ #define GGML_F16_VEC GGML_F16x8
+ #define GGML_F16_VEC_ZERO GGML_F16x8_ZERO
+ #define GGML_F16_VEC_SET1 GGML_F16x8_SET1
+ #define GGML_F16_VEC_LOAD(p, i) GGML_F16x8_LOAD(p)
+ #define GGML_F16_VEC_STORE(p, r, i) GGML_F16x8_STORE((__fp16 *)(p), (r)[i])
+ #define GGML_F16_VEC_FMA GGML_F16x8_FMA
+ #define GGML_F16_VEC_ADD GGML_F16x8_ADD
+ #define GGML_F16_VEC_MUL GGML_F16x8_MUL
+ #define GGML_F16_VEC_REDUCE GGML_F16x8_REDUCE
+#else
+ // if FP16 vector arithmetic is not supported, we use FP32 instead
+ // and take advantage of the vcvt_ functions to convert to/from FP16
+
+ #define GGML_F16_STEP 16
+ #define GGML_F16_EPR 4
+
+ #define GGML_F32Cx4 float32x4_t
+ #define GGML_F32Cx4_ZERO vdupq_n_f32(0.0f)
+ #define GGML_F32Cx4_SET1(x) vdupq_n_f32(x)
+ #define GGML_F32Cx4_LOAD(x) vcvt_f32_f16(vld1_f16((const __fp16 *)(x)))
+ #define GGML_F32Cx4_STORE(x, y) vst1_f16(x, vcvt_f16_f32(y))
+ #define GGML_F32Cx4_FMA(a, b, c) vfmaq_f32(a, b, c)
+ #define GGML_F32Cx4_ADD vaddq_f32
+ #define GGML_F32Cx4_MUL vmulq_f32
+ #define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE
+
+ #define GGML_F16_VEC GGML_F32Cx4
+ #define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO
+ #define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1
+ #define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p)
+ #define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE((__fp16 *)(p), r[i])
+ #define GGML_F16_VEC_FMA GGML_F32Cx4_FMA
+ #define GGML_F16_VEC_ADD GGML_F32Cx4_ADD
+ #define GGML_F16_VEC_MUL GGML_F32Cx4_MUL
+ #define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE
+#endif
+
+#elif defined(__AVX512F__)
+
+#define GGML_SIMD
+
+// F32 AVX512
+
+#define GGML_F32_STEP 64
+#define GGML_F32_EPR 16
+
+#define GGML_F32x16 __m512
+#define GGML_F32x16_ZERO _mm512_setzero_ps()
+#define GGML_F32x16_SET1(x) _mm512_set1_ps(x)
+#define GGML_F32x16_LOAD _mm512_loadu_ps
+#define GGML_F32x16_STORE _mm512_storeu_ps
+// _mm512_fmadd_ps is defined in AVX512F so no guard is required
+#define GGML_F32x16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a)
+#define GGML_F32x16_ADD _mm512_add_ps
+#define GGML_F32x16_MUL _mm512_mul_ps
+#define GGML_F32x16_REDUCE(res, x) \
+do { \
+ int offset = GGML_F32_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm512_add_ps(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm512_add_ps(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm512_add_ps(x[i], x[offset+i]); \
+ } \
+ res = (ggml_float) _mm512_reduce_add_ps(x[0]); \
+} while (0)
+
+// TODO: is this optimal ?
+
+#define GGML_F32_VEC GGML_F32x16
+#define GGML_F32_VEC_ZERO GGML_F32x16_ZERO
+#define GGML_F32_VEC_SET1 GGML_F32x16_SET1
+#define GGML_F32_VEC_LOAD GGML_F32x16_LOAD
+#define GGML_F32_VEC_STORE GGML_F32x16_STORE
+#define GGML_F32_VEC_FMA GGML_F32x16_FMA
+#define GGML_F32_VEC_ADD GGML_F32x16_ADD
+#define GGML_F32_VEC_MUL GGML_F32x16_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x16_REDUCE
+
+// F16 AVX512
+
+#if defined(__AVX512FP16__)
+
+#define GGML_F16_STEP 128
+#define GGML_F16_EPR 32
+
+#define GGML_F16x32 __m512h
+#define GGML_F16x32_ZERO _mm512_setzero_ph()
+#define GGML_F16x32_SET1(x) _mm512_set1_ph(__extension__(_Float16)(x))
+#define GGML_F16x32_LOAD(x) _mm512_loadu_ph(x)
+#define GGML_F16x32_STORE(x, y) _mm512_storeu_ph(x, y)
+#define GGML_F16x32_FMA(a, b, c) _mm512_fmadd_ph(b, c, a)
+#define GGML_F16x32_ADD _mm512_add_ph
+#define GGML_F16x32_MUL _mm512_mul_ph
+#define GGML_F16x32_REDUCE(res, x) \
+do { \
+ int offset = GGML_F16_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm512_add_ph(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm512_add_ph(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm512_add_ph(x[i], x[offset+i]); \
+ } \
+ res = (ggml_float) _mm512_reduce_add_ph(x[0]); \
+} while (0)
+
+#define GGML_F16_VEC GGML_F16x32
+#define GGML_F16_VEC_ZERO GGML_F16x32_ZERO
+#define GGML_F16_VEC_SET1 GGML_F16x32_SET1
+#define GGML_F16_VEC_LOAD(p, i) GGML_F16x32_LOAD(p)
+#define GGML_F16_VEC_STORE(p, r, i) GGML_F16x32_STORE(p, r[i])
+#define GGML_F16_VEC_FMA GGML_F16x32_FMA
+#define GGML_F16_VEC_ADD GGML_F16x32_ADD
+#define GGML_F16_VEC_MUL GGML_F16x32_MUL
+#define GGML_F16_VEC_REDUCE GGML_F16x32_REDUCE
+
+#else // Fallback FP16 <-> FP32
+
+#define GGML_F16_STEP 64
+#define GGML_F16_EPR 16
+
+#define GGML_F32Cx16 __m512
+#define GGML_F32Cx16_ZERO _mm512_setzero_ps()
+#define GGML_F32Cx16_SET1(x) _mm512_set1_ps(x)
+
+// unlike _mm256_cvt intrinsics that require F16C, _mm512_cvt is defined in AVX512F
+// so F16C guard isn't required
+#define GGML_F32Cx16_LOAD(x) _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(x)))
+#define GGML_F32Cx16_STORE(x, y) _mm256_storeu_si256((__m256i *)(x), _mm512_cvtps_ph(y, 0))
+
+#define GGML_F32Cx16_FMA(a, b, c) _mm512_fmadd_ps(b, c, a)
+#define GGML_F32Cx16_ADD _mm512_add_ps
+#define GGML_F32Cx16_MUL _mm512_mul_ps
+#define GGML_F32Cx16_REDUCE(res, x) \
+do { \
+ int offset = GGML_F32_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm512_add_ps(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm512_add_ps(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm512_add_ps(x[i], x[offset+i]); \
+ } \
+ res = (ggml_float) _mm512_reduce_add_ps(x[0]); \
+} while (0)
+
+#define GGML_F16_VEC GGML_F32Cx16
+#define GGML_F16_VEC_ZERO GGML_F32Cx16_ZERO
+#define GGML_F16_VEC_SET1 GGML_F32Cx16_SET1
+#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx16_LOAD(p)
+#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx16_STORE(p, r[i])
+#define GGML_F16_VEC_FMA GGML_F32Cx16_FMA
+#define GGML_F16_VEC_ADD GGML_F32Cx16_ADD
+#define GGML_F16_VEC_MUL GGML_F32Cx16_MUL
+
+#define GGML_F16_VEC_REDUCE GGML_F32Cx16_REDUCE
+
+#endif // __AVX512FP16__
+#elif defined(__AVX__)
+
+#define GGML_SIMD
+
+// F32 AVX
+
+#define GGML_F32_STEP 32
+#define GGML_F32_EPR 8
+
+#define GGML_F32x8 __m256
+#define GGML_F32x8_ZERO _mm256_setzero_ps()
+#define GGML_F32x8_SET1(x) _mm256_set1_ps(x)
+#define GGML_F32x8_LOAD _mm256_loadu_ps
+#define GGML_F32x8_STORE _mm256_storeu_ps
+#if defined(__FMA__)
+ #define GGML_F32x8_FMA(a, b, c) _mm256_fmadd_ps(b, c, a)
+#else
+ #define GGML_F32x8_FMA(a, b, c) _mm256_add_ps(_mm256_mul_ps(b, c), a)
+#endif
+#define GGML_F32x8_ADD _mm256_add_ps
+#define GGML_F32x8_MUL _mm256_mul_ps
+#define GGML_F32x8_REDUCE(res, x) \
+do { \
+ int offset = GGML_F32_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm256_add_ps(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm256_add_ps(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm256_add_ps(x[i], x[offset+i]); \
+ } \
+ const __m128 t0 = _mm_add_ps(_mm256_castps256_ps128(x[0]), \
+ _mm256_extractf128_ps(x[0], 1)); \
+ const __m128 t1 = _mm_hadd_ps(t0, t0); \
+ res = (ggml_float) _mm_cvtss_f32(_mm_hadd_ps(t1, t1)); \
+} while (0)
+// TODO: is this optimal ?
+
+#define GGML_F32_VEC GGML_F32x8
+#define GGML_F32_VEC_ZERO GGML_F32x8_ZERO
+#define GGML_F32_VEC_SET1 GGML_F32x8_SET1
+#define GGML_F32_VEC_LOAD GGML_F32x8_LOAD
+#define GGML_F32_VEC_STORE GGML_F32x8_STORE
+#define GGML_F32_VEC_FMA GGML_F32x8_FMA
+#define GGML_F32_VEC_ADD GGML_F32x8_ADD
+#define GGML_F32_VEC_MUL GGML_F32x8_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x8_REDUCE
+
+// F16 AVX
+
+#define GGML_F16_STEP 32
+#define GGML_F16_EPR 8
+
+// F16 arithmetic is not supported by AVX, so we use F32 instead
+
+#define GGML_F32Cx8 __m256
+#define GGML_F32Cx8_ZERO _mm256_setzero_ps()
+#define GGML_F32Cx8_SET1(x) _mm256_set1_ps(x)
+
+#if defined(__F16C__)
+// the _mm256_cvt intrinsics require F16C
+#define GGML_F32Cx8_LOAD(x) _mm256_cvtph_ps(_mm_loadu_si128((const __m128i *)(x)))
+#define GGML_F32Cx8_STORE(x, y) _mm_storeu_si128((__m128i *)(x), _mm256_cvtps_ph(y, 0))
+#else
+static inline __m256 __avx_f32cx8_load(const ggml_fp16_t * x) {
+ float tmp[8];
+
+ for (int i = 0; i < 8; i++) {
+ tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
+ }
+
+ return _mm256_loadu_ps(tmp);
+}
+static inline void __avx_f32cx8_store(ggml_fp16_t *x, __m256 y) {
+ float arr[8];
+
+ _mm256_storeu_ps(arr, y);
+
+ for (int i = 0; i < 8; i++)
+ x[i] = GGML_CPU_FP32_TO_FP16(arr[i]);
+}
+#define GGML_F32Cx8_LOAD(x) __avx_f32cx8_load(x)
+#define GGML_F32Cx8_STORE(x, y) __avx_f32cx8_store(x, y)
+#endif
+
+#define GGML_F32Cx8_FMA GGML_F32x8_FMA
+#define GGML_F32Cx8_ADD _mm256_add_ps
+#define GGML_F32Cx8_MUL _mm256_mul_ps
+#define GGML_F32Cx8_REDUCE GGML_F32x8_REDUCE
+
+#define GGML_F16_VEC GGML_F32Cx8
+#define GGML_F16_VEC_ZERO GGML_F32Cx8_ZERO
+#define GGML_F16_VEC_SET1 GGML_F32Cx8_SET1
+#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx8_LOAD(p)
+#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx8_STORE(p, r[i])
+#define GGML_F16_VEC_FMA GGML_F32Cx8_FMA
+#define GGML_F16_VEC_ADD GGML_F32Cx8_ADD
+#define GGML_F16_VEC_MUL GGML_F32Cx8_MUL
+#define GGML_F16_VEC_REDUCE GGML_F32Cx8_REDUCE
+
+#elif defined(__POWER9_VECTOR__)
+
+#define GGML_SIMD
+
+// F32 POWER9
+
+#define GGML_F32_STEP 32
+#define GGML_F32_EPR 4
+
+#define GGML_F32x4 vector float
+#define GGML_F32x4_ZERO {0.0f}
+#define GGML_F32x4_SET1 vec_splats
+#define GGML_F32x4_LOAD(p) vec_xl(0, p)
+#define GGML_F32x4_STORE(p, r) vec_xst(r, 0, p)
+#define GGML_F32x4_FMA(a, b, c) vec_madd(b, c, a)
+#define GGML_F32x4_ADD vec_add
+#define GGML_F32x4_MUL vec_mul
+#define GGML_F32x4_REDUCE(res, x) \
+{ \
+ int offset = GGML_F32_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = vec_add(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = vec_add(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = vec_add(x[i], x[offset+i]); \
+ } \
+ res = vec_extract(x[0], 0) + \
+ vec_extract(x[0], 1) + \
+ vec_extract(x[0], 2) + \
+ vec_extract(x[0], 3); \
+}
+#define GGML_F32x4_REDUCE_4(res, s0, s1, s2, s3) \
+{ \
+ vector float v = vec_add(vec_add(s0, s1), \
+ vec_add(s2, s3)); \
+ v = vec_add(v, vec_sld(v, v, 8)); \
+ v = vec_add(v, vec_sld(v, v, 4)); \
+ res += (ggml_float) vec_extract(v, 0); \
+}
+
+#define GGML_F32_VEC GGML_F32x4
+#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
+#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
+#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
+#define GGML_F32_VEC_STORE GGML_F32x4_STORE
+#define GGML_F32_VEC_FMA GGML_F32x4_FMA
+#define GGML_F32_VEC_ADD GGML_F32x4_ADD
+#define GGML_F32_VEC_MUL GGML_F32x4_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
+
+// F16 POWER9
+#define GGML_F16_STEP GGML_F32_STEP
+#define GGML_F16_EPR GGML_F32_EPR
+#define GGML_F16_VEC GGML_F32x4
+#define GGML_F16_VEC_ZERO GGML_F32x4_ZERO
+#define GGML_F16_VEC_SET1 GGML_F32x4_SET1
+#define GGML_F16_VEC_FMA GGML_F32x4_FMA
+#define GGML_F16_VEC_ADD GGML_F32x4_ADD
+#define GGML_F16_VEC_MUL GGML_F32x4_MUL
+#define GGML_F16_VEC_REDUCE GGML_F32x4_REDUCE
+// Use vec_xl, not vec_ld, in case the load address is not aligned.
+#define GGML_F16_VEC_LOAD(p, i) (i & 0x1) ? \
+ vec_extract_fp32_from_shorth(vec_xl(0, p - GGML_F16_EPR)) : \
+ vec_extract_fp32_from_shortl(vec_xl(0, p))
+static inline unsigned char ggml_endian_byte(int i) {
+ uint16_t tmp_val = 1;
+ return ((unsigned char *)&tmp_val)[i];
+}
+#define GGML_ENDIAN_BYTE(i) ggml_endian_byte(i)
+#define GGML_F16_VEC_STORE(p, r, i) \
+ if (i & 0x1) \
+ vec_xst(vec_pack_to_short_fp32(r[i - GGML_ENDIAN_BYTE(1)], \
+ r[i - GGML_ENDIAN_BYTE(0)]), \
+ 0, p - GGML_F16_EPR)
+
+//BF16 POWER9
+#define GGML_BF16_STEP 16
+#define GGML_BF16_EPR 8
+
+#define GGML_BF16x8 vector unsigned short
+#define GGML_BF16x8_ZERO vec_splats((unsigned short)0)
+#define GGML_BF16x8_LOAD(p) vec_xl(0, (const unsigned short *)(p))
+
+#define GGML_BF16_VEC GGML_BF16x8
+#define GGML_BF16_VEC_ZERO GGML_BF16x8_ZERO
+#define GGML_BF16_VEC_LOAD GGML_BF16x8_LOAD
+#if defined(__LITTLE_ENDIAN__)
+#define GGML_BF16_TO_F32_LO(v) ((vector float) vec_mergel(GGML_BF16_VEC_ZERO, (v)))
+#define GGML_BF16_TO_F32_HI(v) ((vector float) vec_mergeh(GGML_BF16_VEC_ZERO, (v)))
+#else
+#define GGML_BF16_TO_F32_LO(v) ((vector float) vec_mergel((v), GGML_BF16_VEC_ZERO))
+#define GGML_BF16_TO_F32_HI(v) ((vector float) vec_mergeh((v), GGML_BF16_VEC_ZERO))
+#endif
+#define GGML_BF16_FMA_LO(acc, x, y) \
+ (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_LO(x), GGML_BF16_TO_F32_LO(y))
+#define GGML_BF16_FMA_HI(acc, x, y) \
+ (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_HI(x), GGML_BF16_TO_F32_HI(y))
+
+#elif defined(__wasm_simd128__)
+
+#define GGML_SIMD
+
+// F32 WASM
+
+#define GGML_F32_STEP 16
+#define GGML_F32_EPR 4
+
+#define GGML_F32x4 v128_t
+#define GGML_F32x4_ZERO wasm_f32x4_splat(0.0f)
+#define GGML_F32x4_SET1(x) wasm_f32x4_splat(x)
+#define GGML_F32x4_LOAD wasm_v128_load
+#define GGML_F32x4_STORE wasm_v128_store
+#define GGML_F32x4_FMA(a, b, c) wasm_f32x4_add(wasm_f32x4_mul(b, c), a)
+#define GGML_F32x4_ADD wasm_f32x4_add
+#define GGML_F32x4_MUL wasm_f32x4_mul
+#define GGML_F32x4_REDUCE(res, x) \
+{ \
+ int offset = GGML_F32_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = wasm_f32x4_add(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = wasm_f32x4_add(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = wasm_f32x4_add(x[i], x[offset+i]); \
+ } \
+ res = wasm_f32x4_extract_lane(x[0], 0) + \
+ wasm_f32x4_extract_lane(x[0], 1) + \
+ wasm_f32x4_extract_lane(x[0], 2) + \
+ wasm_f32x4_extract_lane(x[0], 3); \
+}
+
+#define GGML_F32_VEC GGML_F32x4
+#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
+#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
+#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
+#define GGML_F32_VEC_STORE GGML_F32x4_STORE
+#define GGML_F32_VEC_FMA GGML_F32x4_FMA
+#define GGML_F32_VEC_ADD GGML_F32x4_ADD
+#define GGML_F32_VEC_MUL GGML_F32x4_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
+
+// F16 WASM
+
+#define GGML_F16_STEP 16
+#define GGML_F16_EPR 4
+
+inline static v128_t __wasm_f16x4_load(const ggml_fp16_t * p) {
+ float tmp[4];
+
+ tmp[0] = GGML_CPU_FP16_TO_FP32(p[0]);
+ tmp[1] = GGML_CPU_FP16_TO_FP32(p[1]);
+ tmp[2] = GGML_CPU_FP16_TO_FP32(p[2]);
+ tmp[3] = GGML_CPU_FP16_TO_FP32(p[3]);
+
+ return wasm_v128_load(tmp);
+}
+
+inline static void __wasm_f16x4_store(ggml_fp16_t * p, v128_t x) {
+ float tmp[4];
+
+ wasm_v128_store(tmp, x);
+
+ p[0] = GGML_CPU_FP32_TO_FP16(tmp[0]);
+ p[1] = GGML_CPU_FP32_TO_FP16(tmp[1]);
+ p[2] = GGML_CPU_FP32_TO_FP16(tmp[2]);
+ p[3] = GGML_CPU_FP32_TO_FP16(tmp[3]);
+}
+
+#define GGML_F16x4 v128_t
+#define GGML_F16x4_ZERO wasm_f32x4_splat(0.0f)
+#define GGML_F16x4_SET1(x) wasm_f32x4_splat(x)
+#define GGML_F16x4_LOAD(x) __wasm_f16x4_load(x)
+#define GGML_F16x4_STORE(x, y) __wasm_f16x4_store(x, y)
+#define GGML_F16x4_FMA GGML_F32x4_FMA
+#define GGML_F16x4_ADD wasm_f32x4_add
+#define GGML_F16x4_MUL wasm_f32x4_mul
+#define GGML_F16x4_REDUCE(res, x) \
+{ \
+ int offset = GGML_F16_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = wasm_f32x4_add(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = wasm_f32x4_add(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = wasm_f32x4_add(x[i], x[offset+i]); \
+ } \
+ res = (ggml_float) (wasm_f32x4_extract_lane(x[0], 0) + \
+ wasm_f32x4_extract_lane(x[0], 1) + \
+ wasm_f32x4_extract_lane(x[0], 2) + \
+ wasm_f32x4_extract_lane(x[0], 3)); \
+}
+
+#define GGML_F16_VEC GGML_F16x4
+#define GGML_F16_VEC_ZERO GGML_F16x4_ZERO
+#define GGML_F16_VEC_SET1 GGML_F16x4_SET1
+#define GGML_F16_VEC_LOAD(p, i) GGML_F16x4_LOAD(p)
+#define GGML_F16_VEC_STORE(p, r, i) GGML_F16x4_STORE(p, r[i])
+#define GGML_F16_VEC_FMA GGML_F16x4_FMA
+#define GGML_F16_VEC_ADD GGML_F16x4_ADD
+#define GGML_F16_VEC_MUL GGML_F16x4_MUL
+#define GGML_F16_VEC_REDUCE GGML_F16x4_REDUCE
+
+#elif defined(__SSE3__)
+
+#define GGML_SIMD
+
+// F32 SSE
+
+#define GGML_F32_STEP 32
+#define GGML_F32_EPR 4
+
+#define GGML_F32x4 __m128
+#define GGML_F32x4_ZERO _mm_setzero_ps()
+#define GGML_F32x4_SET1(x) _mm_set1_ps(x)
+#define GGML_F32x4_LOAD _mm_loadu_ps
+#define GGML_F32x4_STORE _mm_storeu_ps
+#if defined(__FMA__)
+ // TODO: Does this work?
+ #define GGML_F32x4_FMA(a, b, c) _mm_fmadd_ps(b, c, a)
+#else
+ #define GGML_F32x4_FMA(a, b, c) _mm_add_ps(_mm_mul_ps(b, c), a)
+#endif
+#define GGML_F32x4_ADD _mm_add_ps
+#define GGML_F32x4_MUL _mm_mul_ps
+#define GGML_F32x4_REDUCE(res, x) \
+{ \
+ int offset = GGML_F32_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm_add_ps(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm_add_ps(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = _mm_add_ps(x[i], x[offset+i]); \
+ } \
+ const __m128 t0 = _mm_hadd_ps(x[0], x[0]); \
+ res = (ggml_float) _mm_cvtss_f32(_mm_hadd_ps(t0, t0)); \
+}
+// TODO: is this optimal ?
+
+#define GGML_F32_VEC GGML_F32x4
+#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
+#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
+#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
+#define GGML_F32_VEC_STORE GGML_F32x4_STORE
+#define GGML_F32_VEC_FMA GGML_F32x4_FMA
+#define GGML_F32_VEC_ADD GGML_F32x4_ADD
+#define GGML_F32_VEC_MUL GGML_F32x4_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
+
+// F16 SSE
+
+#define GGML_F16_STEP 32
+#define GGML_F16_EPR 4
+
+static inline __m128 __sse_f16x4_load(const ggml_fp16_t * x) {
+ float tmp[4];
+
+ tmp[0] = GGML_CPU_FP16_TO_FP32(x[0]);
+ tmp[1] = GGML_CPU_FP16_TO_FP32(x[1]);
+ tmp[2] = GGML_CPU_FP16_TO_FP32(x[2]);
+ tmp[3] = GGML_CPU_FP16_TO_FP32(x[3]);
+
+ return _mm_loadu_ps(tmp);
+}
+
+static inline void __sse_f16x4_store(ggml_fp16_t * x, __m128 y) {
+ float arr[4];
+
+ _mm_storeu_ps(arr, y);
+
+ x[0] = GGML_CPU_FP32_TO_FP16(arr[0]);
+ x[1] = GGML_CPU_FP32_TO_FP16(arr[1]);
+ x[2] = GGML_CPU_FP32_TO_FP16(arr[2]);
+ x[3] = GGML_CPU_FP32_TO_FP16(arr[3]);
+}
+
+#define GGML_F32Cx4 __m128
+#define GGML_F32Cx4_ZERO _mm_setzero_ps()
+#define GGML_F32Cx4_SET1(x) _mm_set1_ps(x)
+#define GGML_F32Cx4_LOAD(x) __sse_f16x4_load(x)
+#define GGML_F32Cx4_STORE(x, y) __sse_f16x4_store(x, y)
+#define GGML_F32Cx4_FMA GGML_F32x4_FMA
+#define GGML_F32Cx4_ADD _mm_add_ps
+#define GGML_F32Cx4_MUL _mm_mul_ps
+#define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE
+
+#define GGML_F16_VEC GGML_F32Cx4
+#define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO
+#define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1
+#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p)
+#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE(p, r[i])
+#define GGML_F16_VEC_FMA GGML_F32Cx4_FMA
+#define GGML_F16_VEC_ADD GGML_F32Cx4_ADD
+#define GGML_F16_VEC_MUL GGML_F32Cx4_MUL
+#define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE
+
+#elif defined(__loongarch_asx)
+
+#define GGML_SIMD
+
+// F32 LASX
+#define GGML_F32_STEP 32
+#define GGML_F32_EPR 8
+
+#define GGML_F32x8 __m256
+#define GGML_F32x8_ZERO (__m256)__lasx_xvldi(0)
+#define GGML_F32x8_SET1(x) (__m256)__lasx_xvreplfr2vr_s((x))
+#define GGML_F32x8_LOAD(x) (__m256)__lasx_xvld((x), 0)
+#define GGML_F32x8_STORE(x,y) __lasx_xvst((y), (x), 0)
+#define GGML_F32x8_FMA(a, b, c) __lasx_xvfmadd_s(b, c, a)
+#define GGML_F32x8_ADD __lasx_xvfadd_s
+#define GGML_F32x8_MUL __lasx_xvfmul_s
+#define GGML_F32x8_REDUCE(res, x) \
+do { \
+ int offset = GGML_F32_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = __lasx_xvfadd_s(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = __lasx_xvfadd_s(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = __lasx_xvfadd_s(x[i], x[offset+i]); \
+ } \
+ float *tmp_p = (float *)&x[0]; \
+ res = tmp_p[0] + tmp_p[1] + tmp_p[2] + tmp_p[3] + tmp_p[4] + tmp_p[5] + tmp_p[6] + tmp_p[7]; \
+} while (0)
+// TODO: is this optimal ?
+
+#define GGML_F32_VEC GGML_F32x8
+#define GGML_F32_VEC_ZERO GGML_F32x8_ZERO
+#define GGML_F32_VEC_SET1 GGML_F32x8_SET1
+#define GGML_F32_VEC_LOAD GGML_F32x8_LOAD
+#define GGML_F32_VEC_STORE GGML_F32x8_STORE
+#define GGML_F32_VEC_FMA GGML_F32x8_FMA
+#define GGML_F32_VEC_ADD GGML_F32x8_ADD
+#define GGML_F32_VEC_MUL GGML_F32x8_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x8_REDUCE
+
+// F16 LASX
+
+#define GGML_F16_STEP 32
+#define GGML_F16_EPR 8
+
+// F16 arithmetic is not supported by LASX, so we use F32 instead
+
+#define GGML_F32Cx8 __m256
+#define GGML_F32Cx8_ZERO (__m256)__lasx_xvldi(0)
+#define GGML_F32Cx8_SET1(x) (__m256)__lasx_xvreplfr2vr_s((x))
+
+static inline __m256 __lasx_f32cx8_load(const ggml_fp16_t * x) {
+ __m256i a;
+ memcpy(&a, x, sizeof(ggml_fp16_t) * 8);
+ a = __lasx_xvpermi_d(a, 0 | (1 << 4));
+ return __lasx_xvfcvtl_s_h(a);
+}
+
+static inline void __lasx_f32cx8_store(ggml_fp16_t * x, __m256 y) {
+ __m256i a = __lasx_xvfcvt_h_s(y, y);
+ a = __lasx_xvpermi_d(a, 0 | (2 << 2));
+ memcpy(x, &a, sizeof(ggml_fp16_t) * 8);
+}
+#define GGML_F32Cx8_LOAD(x) __lasx_f32cx8_load(x)
+#define GGML_F32Cx8_STORE(x, y) __lasx_f32cx8_store(x, y)
+
+#define GGML_F32Cx8_FMA GGML_F32x8_FMA
+#define GGML_F32Cx8_ADD __lasx_xvfadd_s
+#define GGML_F32Cx8_MUL __lasx_xvfmul_s
+#define GGML_F32Cx8_REDUCE GGML_F32x8_REDUCE
+
+#define GGML_F16_VEC GGML_F32Cx8
+#define GGML_F16_VEC_ZERO GGML_F32Cx8_ZERO
+#define GGML_F16_VEC_SET1 GGML_F32Cx8_SET1
+#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx8_LOAD(p)
+#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx8_STORE(p, r[i])
+#define GGML_F16_VEC_FMA GGML_F32Cx8_FMA
+#define GGML_F16_VEC_ADD GGML_F32Cx8_ADD
+#define GGML_F16_VEC_MUL GGML_F32Cx8_MUL
+#define GGML_F16_VEC_REDUCE GGML_F32Cx8_REDUCE
+
+#elif defined(__loongarch_sx)
+
+#define GGML_SIMD
+
+// F32 LSX
+
+#define GGML_F32_STEP 32
+#define GGML_F32_EPR 4
+
+#define GGML_F32x4 __m128
+#define GGML_F32x4_ZERO (__m128)__lsx_vldi(0)
+#define GGML_F32x4_SET1(x) (__m128)__lsx_vreplfr2vr_s((x))
+#define GGML_F32x4_LOAD(x) (__m128)__lsx_vld((x), 0)
+#define GGML_F32x4_STORE(x, y) __lsx_vst(y, x, 0)
+#define GGML_F32x4_FMA(a, b, c) __lsx_vfmadd_s(b, c, a)
+#define GGML_F32x4_ADD __lsx_vfadd_s
+#define GGML_F32x4_MUL __lsx_vfmul_s
+
+#define GGML_F32x4_REDUCE(res, x) \
+{ \
+ int offset = GGML_F32_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = __lsx_vfadd_s(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = __lsx_vfadd_s(x[i], x[offset+i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = __lsx_vfadd_s(x[i], x[offset+i]); \
+ } \
+ __m128i t0 = __lsx_vpickev_w((__m128i)x[0], (__m128i)x[0]); \
+ __m128i t1 = __lsx_vpickod_w((__m128i)x[0], (__m128i)x[0]); \
+ __m128 t2 = __lsx_vfadd_s((__m128)t0, (__m128)t1); \
+ __m128i t3 = __lsx_vpickev_w((__m128i)t2, (__m128i)t2); \
+ __m128i t4 = __lsx_vpickod_w((__m128i)t2, (__m128i)t2); \
+ __m128 t5 = __lsx_vfadd_s((__m128)t3, (__m128)t4); \
+ res = (ggml_float) ((v4f32)t5)[0]; \
+}
+
+#define GGML_F32_VEC GGML_F32x4
+#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
+#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
+#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
+#define GGML_F32_VEC_STORE GGML_F32x4_STORE
+#define GGML_F32_VEC_FMA GGML_F32x4_FMA
+#define GGML_F32_VEC_ADD GGML_F32x4_ADD
+#define GGML_F32_VEC_MUL GGML_F32x4_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
+
+// F16 LSX
+
+#define GGML_F16_STEP 32
+#define GGML_F16_EPR 4
+
+static inline __m128 __lsx_f16x4_load(const ggml_fp16_t * x) {
+ return __lsx_vfcvtl_s_h(__lsx_vld((const void *)x, 0));
+}
+
+static inline void __lsx_f16x4_store(ggml_fp16_t * x, __m128 y) {
+ __m128i a = __lsx_vfcvt_h_s(y, y);
+ memcpy(x, &a, sizeof(ggml_fp16_t) * 4);
+}
+
+#define GGML_F32Cx4 __m128
+#define GGML_F32Cx4_ZERO (__m128)__lsx_vldi(0)
+#define GGML_F32Cx4_SET1(x) (__m128)__lsx_vreplfr2vr_s((x))
+#define GGML_F32Cx4_LOAD(x) (__m128)__lsx_f16x4_load(x)
+#define GGML_F32Cx4_STORE(x, y) __lsx_f16x4_store(x, y)
+#define GGML_F32Cx4_FMA GGML_F32x4_FMA
+#define GGML_F32Cx4_ADD __lsx_vfadd_s
+#define GGML_F32Cx4_MUL __lsx_vfmul_s
+#define GGML_F32Cx4_REDUCE GGML_F32x4_REDUCE
+
+#define GGML_F16_VEC GGML_F32Cx4
+#define GGML_F16_VEC_ZERO GGML_F32Cx4_ZERO
+#define GGML_F16_VEC_SET1 GGML_F32Cx4_SET1
+#define GGML_F16_VEC_LOAD(p, i) GGML_F32Cx4_LOAD(p)
+#define GGML_F16_VEC_STORE(p, r, i) GGML_F32Cx4_STORE(p, r[i])
+#define GGML_F16_VEC_FMA GGML_F32Cx4_FMA
+#define GGML_F16_VEC_ADD GGML_F32Cx4_ADD
+#define GGML_F16_VEC_MUL GGML_F32Cx4_MUL
+#define GGML_F16_VEC_REDUCE GGML_F32Cx4_REDUCE
+
+#elif defined(__VXE__) || defined(__VXE2__)
+
+#define GGML_SIMD
+
+// F32 s390x
+
+#define GGML_F32_STEP 32
+#define GGML_F32_EPR 4
+
+#define GGML_F32x4 float32x4_t
+#define GGML_F32x4_ZERO vec_splats(0.0f)
+#define GGML_F32x4_SET1 vec_splats
+#define GGML_F32x4_LOAD(p) vec_xl(0, p)
+#define GGML_F32x4_STORE(p, r) vec_xst(r, 0, p)
+#define GGML_F32x4_FMA(a, b, c) vec_madd(b, c, a)
+#define GGML_F32x4_ADD vec_add
+#define GGML_F32x4_MUL vec_mul
+#define GGML_F32x4_REDUCE(res, x) \
+{ \
+ int offset = GGML_F32_ARR >> 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = vec_add(x[i], x[offset + i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = vec_add(x[i], x[offset + i]); \
+ } \
+ offset >>= 1; \
+ for (int i = 0; i < offset; ++i) { \
+ x[i] = vec_add(x[i], x[offset + i]); \
+ } \
+ float32x4_t tmp = x[0] + vec_reve(x[0]); \
+ res = tmp[0] + tmp[1]; \
+}
+#define GGML_F32x4_REDUCE_4(res, s0, s1, s2, s3) \
+{ \
+ float32x4_t v = vec_add(vec_add(s0, s1), \
+ vec_add(s2, s3)); \
+ v = vec_add(v, vec_sld(v, v, 8)); \
+ v = vec_add(v, vec_sld(v, v, 4)); \
+ res += (ggml_float)vec_extract(v, 0); \
+}
+
+#define GGML_F32_VEC GGML_F32x4
+#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
+#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
+#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
+#define GGML_F32_VEC_STORE GGML_F32x4_STORE
+#define GGML_F32_VEC_FMA GGML_F32x4_FMA
+#define GGML_F32_VEC_ADD GGML_F32x4_ADD
+#define GGML_F32_VEC_MUL GGML_F32x4_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
+
+// F16 s390x
+#define GGML_F16_STEP GGML_F32_STEP
+#define GGML_F16_EPR GGML_F32_EPR
+
+static inline uint32x4_t __lzs_f32cx4_to_f16(float32x4_t v_f) {
+ float32x4_t v_base = vec_mul(vec_mul(vec_abs(v_f), vec_splats(0x1.0p+112f)), vec_splats(0x1.0p-110f));
+
+ const uint32x4_t v_w = (uint32x4_t)v_f;
+ const uint32x4_t v_shl1_w = vec_add(v_w, v_w);
+ const uint32x4_t v_sign = vec_and(v_w, vec_splats(UINT32_C(0x80000000)));
+ const uint32x4_t v_bias = vec_max(vec_and(v_shl1_w, vec_splats(UINT32_C(0xFF000000))), vec_splats(UINT32_C(0x71000000)));
+
+ v_base = vec_add((float32x4_t)vec_add(vec_sr(v_bias, 1), vec_splats(UINT32_C(0x07800000))), v_base);
+
+ const uint32x4_t v_bits = (uint32x4_t)v_base;
+ const uint32x4_t v_nonsign = vec_add(vec_and(vec_sr(v_bits, 13), vec_splats(UINT32_C(0x00007C00))),
+ vec_and(v_bits, vec_splats(UINT32_C(0x00000FFF))));
+ const uint32x4_t v_is_nan = (uint32x4_t)vec_cmpgt(v_shl1_w, vec_splats(UINT32_C(0xFF000000)));
+
+ return vec_or(vec_sr(v_sign, 16), vec_sel(v_nonsign, vec_splats(UINT32_C(0x7E00)), v_is_nan));
+}
+
+static inline float32x4_t __lzs_f16cx4_load(const ggml_fp16_t * x) {
+ float tmp[4];
+
+ for (int i = 0; i < 4; i++) {
+ tmp[i] = GGML_CPU_FP16_TO_FP32(x[i]);
+ }
+
+ // note: keep type-cast here to prevent compiler bugs
+ // see: https://github.com/ggml-org/llama.cpp/issues/12846
+ return vec_xl(0, (const float *)(tmp));
+}
+
+static inline void __lzs_f16cx4_store(ggml_fp16_t * x, float32x4_t v_y) {
+ const uint32x4_t v_h = __lzs_f32cx4_to_f16(v_y);
+ const uint64_t tmp = ((uint64x2_t)vec_pack(v_h, v_h))[0];
+ memcpy(x, &tmp, sizeof(tmp));
+}
+
+#define GGML_F16_VEC GGML_F32x4
+#define GGML_F16_VEC_ZERO GGML_F32x4_ZERO
+#define GGML_F16_VEC_SET1 GGML_F32x4_SET1
+#define GGML_F16_VEC_LOAD(p, i) __lzs_f16cx4_load(p)
+#define GGML_F16_VEC_STORE(p, r, i) __lzs_f16cx4_store(p, r[i])
+#define GGML_F16_VEC_FMA GGML_F32x4_FMA
+#define GGML_F16_VEC_ADD GGML_F32x4_ADD
+#define GGML_F16_VEC_MUL GGML_F32x4_MUL
+#define GGML_F16_VEC_REDUCE GGML_F32x4_REDUCE
+
+// BF16 s390x
+#define GGML_BF16_STEP 16
+#define GGML_BF16_EPR 8
+
+#define GGML_BF16x8 __vector unsigned short
+#define GGML_BF16x8_ZERO vec_splats((unsigned short)0)
+#define GGML_BF16x8_LOAD(p) vec_xl(0, (const unsigned short *)(p))
+
+#define GGML_BF16_VEC GGML_BF16x8
+#define GGML_BF16_VEC_ZERO GGML_BF16x8_ZERO
+#define GGML_BF16_VEC_LOAD GGML_BF16x8_LOAD
+#define GGML_BF16_TO_F32_LO(v) ((float32x4_t) vec_mergel((v), GGML_BF16_VEC_ZERO))
+#define GGML_BF16_TO_F32_HI(v) ((float32x4_t) vec_mergeh((v), GGML_BF16_VEC_ZERO))
+#define GGML_BF16_FMA_LO(acc, x, y) \
+ (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_LO(x), GGML_BF16_TO_F32_LO(y))
+#define GGML_BF16_FMA_HI(acc, x, y) \
+ (acc) = GGML_F32x4_FMA((acc), GGML_BF16_TO_F32_HI(x), GGML_BF16_TO_F32_HI(y))
+
+#elif defined(__riscv_v_intrinsic)
+
+// compatible with vlen >= 128
+
+#define GGML_SIMD
+
+// F32
+
+#define GGML_F32_STEP 16
+#define GGML_F32_EPR 4
+
+#define GGML_F32x4 vfloat32m1_t
+#define GGML_F32x4_ZERO __riscv_vfmv_v_f_f32m1(0.0f, GGML_F32_EPR)
+#define GGML_F32x4_SET1(x) __riscv_vfmv_v_f_f32m1(x, GGML_F32_EPR)
+#define GGML_F32x4_LOAD(x) __riscv_vle32_v_f32m1(x, GGML_F32_EPR)
+#define GGML_F32x4_STORE(b, v) __riscv_vse32_v_f32m1(b, v, GGML_F32_EPR)
+#define GGML_F32x4_FMA(a, b, c) __riscv_vfmacc_vv_f32m1(a, b, c, GGML_F32_EPR)
+#define GGML_F32x4_ADD(a, b) __riscv_vfadd_vv_f32m1(a, b, GGML_F32_EPR)
+#define GGML_F32x4_MUL(a, b) __riscv_vfmul_vv_f32m1(a, b, GGML_F32_EPR)
+
+#define GGML_F32_VEC GGML_F32x4
+#define GGML_F32_VEC_ZERO GGML_F32x4_ZERO
+#define GGML_F32_VEC_SET1 GGML_F32x4_SET1
+#define GGML_F32_VEC_LOAD GGML_F32x4_LOAD
+#define GGML_F32_VEC_STORE GGML_F32x4_STORE
+#define GGML_F32_VEC_FMA GGML_F32x4_FMA
+#define GGML_F32_VEC_ADD GGML_F32x4_ADD
+#define GGML_F32_VEC_MUL GGML_F32x4_MUL
+#define GGML_F32_VEC_REDUCE GGML_F32x4_REDUCE
+
+#endif
+
+// GGML_F32_ARR / GGML_F16_ARR
+// number of registers to use per step
+#ifdef GGML_SIMD
+#define GGML_F32_ARR (GGML_F32_STEP/GGML_F32_EPR)
+#define GGML_F16_ARR (GGML_F16_STEP/GGML_F16_EPR)
+#endif
+
+// GGML_F16_DOT_*
+// like GGML_F16_* but for dot products which need F32 accumulation on AVX512-FP16
+
+#if defined(__AVX512FP16__)
+
+#define GGML_F16_DOT_STEP GGML_F32_STEP
+#define GGML_F16_DOT_EPR GGML_F32_EPR
+#define GGML_F16_DOT_ARR GGML_F32_ARR
+#define GGML_F16_DOT_VEC GGML_F32x16
+#define GGML_F16_DOT_VEC_ZERO GGML_F32x16_ZERO
+#define GGML_F16_DOT_VEC_LOAD(p, i) _mm512_cvtph_ps(_mm256_loadu_si256((const __m256i *)(p)))
+#define GGML_F16_DOT_VEC_FMA GGML_F32x16_FMA
+#define GGML_F16_DOT_VEC_REDUCE GGML_F32x16_REDUCE
+
+#else
+
+#define GGML_F16_DOT_STEP GGML_F16_STEP
+#define GGML_F16_DOT_EPR GGML_F16_EPR
+#define GGML_F16_DOT_ARR GGML_F16_ARR
+#define GGML_F16_DOT_VEC GGML_F16_VEC
+#define GGML_F16_DOT_VEC_ZERO GGML_F16_VEC_ZERO
+#define GGML_F16_DOT_VEC_LOAD GGML_F16_VEC_LOAD
+#define GGML_F16_DOT_VEC_FMA GGML_F16_VEC_FMA
+#define GGML_F16_DOT_VEC_REDUCE GGML_F16_VEC_REDUCE
+
+#endif // defined(__AVX512FP16__)
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/python/freetoken/kernel/csrc/ggml/ggml-cpu/vec.cpp b/python/freetoken/kernel/csrc/ggml/ggml-cpu/vec.cpp
new file mode 100644
index 00000000..3918a6f4
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/ggml-cpu/vec.cpp
@@ -0,0 +1,613 @@
+#include "vec.h"
+
+#include <cassert>
+
+// precomputed gelu table for f16 (128 KB)
+ggml_fp16_t ggml_table_gelu_f16[1 << 16];
+
+// precomputed quick gelu table for f16 (128 KB)
+ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
+
+void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc) {
+ assert(nrc == 1);
+ GGML_UNUSED(nrc);
+ GGML_UNUSED(bx);
+ GGML_UNUSED(by);
+ GGML_UNUSED(bs);
+
+#if defined(GGML_SIMD)
+ float sumf = 0.0f;
+
+ #if defined(__ARM_FEATURE_SVE)
+ const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
+ const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
+ const int ggml_f32_step = 8 * ggml_f32_epr; // choose 8 SVE registers
+
+ const int np = (n & ~(ggml_f32_step - 1));
+ svfloat32_t sum1 = svdup_n_f32(0.0f);
+ svfloat32_t sum2 = svdup_n_f32(0.0f);
+ svfloat32_t sum3 = svdup_n_f32(0.0f);
+ svfloat32_t sum4 = svdup_n_f32(0.0f);
+ svfloat32_t sum5 = svdup_n_f32(0.0f);
+ svfloat32_t sum6 = svdup_n_f32(0.0f);
+ svfloat32_t sum7 = svdup_n_f32(0.0f);
+ svfloat32_t sum8 = svdup_n_f32(0.0f);
+ svfloat32_t ax1,ax2,ax3,ax4,ax5,ax6,ax7,ax8;
+ svfloat32_t ay1,ay2,ay3,ay4,ay5,ay6,ay7,ay8;
+ for (int i = 0; i < np; i += ggml_f32_step) {
+ ax1 = GGML_F32_VEC_LOAD(x + i);
+ ay1 = GGML_F32_VEC_LOAD(y + i);
+ sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
+
+ ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
+ ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
+ sum2 = GGML_F32_VEC_FMA(sum2, ax2, ay2);
+
+ ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
+ ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
+ sum3 = GGML_F32_VEC_FMA(sum3, ax3, ay3);
+
+ ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
+ ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
+ sum4 = GGML_F32_VEC_FMA(sum4, ax4, ay4);
+
+ ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
+ ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
+ sum5 = GGML_F32_VEC_FMA(sum5, ax5, ay5);
+
+ ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
+ ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
+ sum6 = GGML_F32_VEC_FMA(sum6, ax6, ay6);
+
+ ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
+ ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
+ sum7 = GGML_F32_VEC_FMA(sum7, ax7, ay7);
+
+ ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
+ ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
+ sum8 = GGML_F32_VEC_FMA(sum8, ax8, ay8);
+ }
+ // leftovers
+ // Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
+ const int np2 = (n & ~(ggml_f32_epr - 1));
+ for (int i = np; i < np2; i += ggml_f32_epr) {
+ ax1 = GGML_F32_VEC_LOAD(x + i);
+ ay1 = GGML_F32_VEC_LOAD(y + i);
+ sum1 = GGML_F32_VEC_FMA(sum1, ax1, ay1);
+ }
+ // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmla on available elements only
+ if (np2 < n) {
+ svbool_t pg = svwhilelt_b32(np2, n);
+ ax1 = svld1_f32(pg, x + np2);
+ ay1 = svld1_f32(pg, y + np2);
+ sum1 = svmla_f32_m(pg, sum1, ax1, ay1);
+ }
+ // reduce sum1,sum2 to sum1
+ GGML_F32_VEC_REDUCE(sumf, sum1, sum2, sum3, sum4, sum5, sum6, sum7, sum8);
+ #elif defined(__riscv_v_intrinsic)
+ int vl = __riscv_vsetvlmax_e32m8();
+ vfloat32m1_t vs = __riscv_vfmv_v_f_f32m1(0.0f, 1);
+ vfloat32m8_t vsum;
+ vfloat32m8_t ax;
+ vfloat32m8_t ay;
+ vsum = __riscv_vfmv_v_f_f32m8_tu(vsum, 0.0f, vl);
+ for (int i = 0; i < n; i += vl) {
+ vl = __riscv_vsetvl_e32m8(n - i);
+ ax = __riscv_vle32_v_f32m8_tu(ax, &x[i], vl);
+ ay = __riscv_vle32_v_f32m8_tu(ay, &y[i], vl);
+ vsum = __riscv_vfmacc_vv_f32m8_tu(vsum, ax, ay, vl);
+ }
+ vl = __riscv_vsetvlmax_e32m8();
+ vs = __riscv_vfredusum_vs_f32m8_f32m1(vsum, vs, vl);
+ sumf += __riscv_vfmv_f_s_f32m1_f32(vs);
+ #else
+ const int np = (n & ~(GGML_F32_STEP - 1));
+
+ GGML_F32_VEC sum[GGML_F32_ARR] = { GGML_F32_VEC_ZERO };
+
+ GGML_F32_VEC ax[GGML_F32_ARR];
+ GGML_F32_VEC ay[GGML_F32_ARR];
+
+ for (int i = 0; i < np; i += GGML_F32_STEP) {
+ for (int j = 0; j < GGML_F32_ARR; j++) {
+ ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
+ ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+
+ sum[j] = GGML_F32_VEC_FMA(sum[j], ax[j], ay[j]);
+ }
+ }
+
+ // reduce sum0..sum3 to sum0
+ GGML_F32_VEC_REDUCE(sumf, sum);
+
+ // leftovers
+ for (int i = np; i < n; ++i) {
+ sumf += x[i]*y[i];
+ }
+ #endif
+#else
+ // scalar
+ ggml_float sumf = 0.0;
+ for (int i = 0; i < n; ++i) {
+ sumf += (ggml_float)(x[i]*y[i]);
+ }
+#endif
+
+ *s = sumf;
+}
+
+void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc) {
+ assert(nrc == 1);
+ GGML_UNUSED(nrc);
+ GGML_UNUSED(bx);
+ GGML_UNUSED(by);
+ GGML_UNUSED(bs);
+ int i = 0;
+ ggml_float sumf = 0;
+
+#if defined(__AVX512BF16__)
+ __m512 c1 = _mm512_setzero_ps();
+ __m512 c2 = _mm512_setzero_ps();
+ for (; i + 64 <= n; i += 64) {
+ c1 = _mm512_dpbf16_ps(c1, m512bh(_mm512_loadu_si512((x + i))),
+ m512bh(_mm512_loadu_si512((y + i))));
+ c2 = _mm512_dpbf16_ps(c2, m512bh(_mm512_loadu_si512((x + i + 32))),
+ m512bh(_mm512_loadu_si512((y + i + 32))));
+ }
+ sumf += (ggml_float)_mm512_reduce_add_ps(c1);
+ sumf += (ggml_float)_mm512_reduce_add_ps(c2);
+
+#elif defined(__AVX512F__)
+#define LOAD(p) _mm512_castsi512_ps(_mm512_slli_epi32(_mm512_cvtepu16_epi32(_mm256_loadu_si256((const __m256i *)(p))), 16))
+ __m512 c1 = _mm512_setzero_ps();
+ __m512 c2 = _mm512_setzero_ps();
+ for (; i + 32 <= n; i += 32) {
+ c1 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i), LOAD(y + i)), c1);
+ c2 = _mm512_add_ps(_mm512_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c2);
+ }
+ sumf += (ggml_float)_mm512_reduce_add_ps(c1);
+ sumf += (ggml_float)_mm512_reduce_add_ps(c2);
+
+#undef LOAD
+#elif defined(__AVX2__) || defined(__AVX__)
+#if defined(__AVX2__)
+#define LOAD(p) _mm256_castsi256_ps(_mm256_slli_epi32(_mm256_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16))
+#else
+#define LOAD(p) _mm256_castsi256_ps(_mm256_insertf128_si256(_mm256_castsi128_si256(_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_loadu_si128((const __m128i *)(p))), 16)), (_mm_slli_epi32(_mm_cvtepu16_epi32(_mm_bsrli_si128(_mm_loadu_si128((const __m128i *)(p)), 8)), 16)), 1))
+#endif
+ __m256 c1 = _mm256_setzero_ps();
+ __m256 c2 = _mm256_setzero_ps();
+ __m256 c3 = _mm256_setzero_ps();
+ __m256 c4 = _mm256_setzero_ps();
+ for (; i + 32 <= n; i += 32) {
+ c1 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i), LOAD(y + i)), c1);
+ c2 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 8), LOAD(y + i + 8)), c2);
+ c3 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 16), LOAD(y + i + 16)), c3);
+ c4 = _mm256_add_ps(_mm256_mul_ps(LOAD(x + i + 24), LOAD(y + i + 24)), c4);
+ }
+ __m128 g;
+ c1 = _mm256_add_ps(_mm256_add_ps(c1, c3),
+ _mm256_add_ps(c2, c4));
+ g = _mm_add_ps(_mm256_extractf128_ps(c1, 1),
+ _mm256_castps256_ps128(c1));
+ g = _mm_add_ps(g, _mm_movehl_ps(g, g));
+ g = _mm_add_ss(g, _mm_movehdup_ps(g));
+ sumf += (ggml_float)_mm_cvtss_f32(g);
+
+#undef LOAD
+#elif defined(__riscv_v_intrinsic) && defined(__riscv_zvfbfwma)
+ size_t vl = __riscv_vsetvlmax_e32m4();
+
+ // initialize accumulators to all zeroes
+ vfloat32m4_t vsum0 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+ vfloat32m4_t vsum1 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+
+ // calculate step size
+ const size_t epr = __riscv_vsetvlmax_e16m2();
+ const size_t step = epr * 2;
+ const int np = (n & ~(step - 1));
+
+ // unroll by 2
+ for (; i < np; i += step) {
+ vbfloat16m2_t ax0 = __riscv_vle16_v_bf16m2((const __bf16 *)&x[i], epr);
+ vbfloat16m2_t ay0 = __riscv_vle16_v_bf16m2((const __bf16 *)&y[i], epr);
+ vsum0 = __riscv_vfwmaccbf16_vv_f32m4(vsum0, ax0, ay0, epr);
+ __asm__ __volatile__ ("" ::: "memory");
+
+ vbfloat16m2_t ax1 = __riscv_vle16_v_bf16m2((const __bf16 *)&x[i + epr], epr);
+ vbfloat16m2_t ay1 = __riscv_vle16_v_bf16m2((const __bf16 *)&y[i + epr], epr);
+ vsum1 = __riscv_vfwmaccbf16_vv_f32m4(vsum1, ax1, ay1, epr);
+ __asm__ __volatile__ ("" ::: "memory");
+ }
+
+ // accumulate in 1 register
+ vsum0 = __riscv_vfadd_vv_f32m4(vsum0, vsum1, vl);
+
+ // leftovers
+ for (i = np; i < n; i += vl) {
+ vl = __riscv_vsetvl_e16m2(n - i);
+ vbfloat16m2_t ax0 = __riscv_vle16_v_bf16m2((const __bf16 *)&x[i], vl);
+ vbfloat16m2_t ay0 = __riscv_vle16_v_bf16m2((const __bf16 *)&y[i], vl);
+ vsum0 = __riscv_vfwmaccbf16_vv_f32m4(vsum0, ax0, ay0, vl);
+ }
+
+ // reduce
+ vl = __riscv_vsetvlmax_e32m4();
+ vfloat32m1_t redsum = __riscv_vfredusum_vs_f32m4_f32m1(vsum0, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
+ sumf += __riscv_vfmv_f_s_f32m1_f32(redsum);
+
+#elif defined(__POWER9_VECTOR__) || defined(__VXE__) || defined(__VXE2__)
+ const int np = (n & ~(GGML_BF16_STEP - 1));
+ if (np > 0) {
+ GGML_F32_VEC sum[4] = {GGML_F32_VEC_ZERO};
+ for (; i < np; i += GGML_BF16_STEP) {
+ GGML_BF16_VEC vx0 = GGML_BF16_VEC_LOAD(x + i);
+ GGML_BF16_VEC vx1 = GGML_BF16_VEC_LOAD(x + i + 8);
+ GGML_BF16_VEC vy0 = GGML_BF16_VEC_LOAD(y + i);
+ GGML_BF16_VEC vy1 = GGML_BF16_VEC_LOAD(y + i + 8);
+ GGML_BF16_FMA_LO(sum[0], vx0, vy0);
+ GGML_BF16_FMA_HI(sum[1], vx0, vy0);
+ GGML_BF16_FMA_LO(sum[2], vx1, vy1);
+ GGML_BF16_FMA_HI(sum[3], vx1, vy1);
+ }
+ GGML_F32x4_REDUCE_4(sumf, sum[0], sum[1], sum[2], sum[3]);
+ }
+#endif
+
+ for (; i < n; ++i) {
+ sumf += (ggml_float)(GGML_BF16_TO_FP32(x[i]) *
+ GGML_BF16_TO_FP32(y[i]));
+ }
+ *s = sumf;
+}
+
+void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc) {
+ assert(nrc == 1);
+ GGML_UNUSED(nrc);
+ GGML_UNUSED(bx);
+ GGML_UNUSED(by);
+ GGML_UNUSED(bs);
+
+ ggml_float sumf = 0.0;
+
+
+#if defined(GGML_SIMD)
+ #if defined(__ARM_FEATURE_SVE)
+ const int ggml_f16_epr = svcnth();
+ const int ggml_f16_step = 8 * ggml_f16_epr;
+ const int np = n - (n % ggml_f16_step);
+ const int np2 = n - (n % ggml_f16_epr);
+
+ svfloat32_t sum1_lo = svdup_n_f32(0.0f);
+ svfloat32_t sum1_hi = svdup_n_f32(0.0f);
+ svfloat32_t sum2_lo = svdup_n_f32(0.0f);
+ svfloat32_t sum2_hi = svdup_n_f32(0.0f);
+ svfloat32_t sum3_lo = svdup_n_f32(0.0f);
+ svfloat32_t sum3_hi = svdup_n_f32(0.0f);
+ svfloat32_t sum4_lo = svdup_n_f32(0.0f);
+ svfloat32_t sum4_hi = svdup_n_f32(0.0f);
+
+ for (int i = 0; i < np; i += ggml_f16_step) {
+ ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0), GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0));
+ ggml_sve_f16_fma_widened(&sum2_lo, &sum2_hi, GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1), GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1));
+ ggml_sve_f16_fma_widened(&sum3_lo, &sum3_hi, GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2), GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2));
+ ggml_sve_f16_fma_widened(&sum4_lo, &sum4_hi, GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3), GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3));
+ ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4), GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4));
+ ggml_sve_f16_fma_widened(&sum2_lo, &sum2_hi, GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5), GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5));
+ ggml_sve_f16_fma_widened(&sum3_lo, &sum3_hi, GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6), GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6));
+ ggml_sve_f16_fma_widened(&sum4_lo, &sum4_hi, GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7), GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7));
+ }
+
+ for (int i = np; i < np2; i += ggml_f16_epr) {
+ ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, GGML_F16x_VEC_LOAD(x + i, 0), GGML_F16x_VEC_LOAD(y + i, 0));
+ }
+
+ if (np2 < n) {
+ const svbool_t pg = svwhilelt_b16(np2, n);
+ const svfloat16_t rx = svld1_f16(pg, (const __fp16 *)(x + np2));
+ const svfloat16_t ry = svld1_f16(pg, (const __fp16 *)(y + np2));
+
+ ggml_sve_f16_fma_widened(&sum1_lo, &sum1_hi, rx, ry);
+ }
+
+ sum1_lo = svadd_f32_m(DEFAULT_PG32, sum1_lo, sum2_lo);
+ sum1_hi = svadd_f32_m(DEFAULT_PG32, sum1_hi, sum2_hi);
+ sum3_lo = svadd_f32_m(DEFAULT_PG32, sum3_lo, sum4_lo);
+ sum3_hi = svadd_f32_m(DEFAULT_PG32, sum3_hi, sum4_hi);
+ sum1_lo = svadd_f32_m(DEFAULT_PG32, sum1_lo, sum3_lo);
+ sum1_hi = svadd_f32_m(DEFAULT_PG32, sum1_hi, sum3_hi);
+
+ sumf = ggml_sve_sum_f32x2(sum1_lo, sum1_hi);
+ #elif defined(__riscv_v_intrinsic)
+ #if defined(__riscv_zvfh)
+ int vl = __riscv_vsetvlmax_e32m2();
+ vfloat32m1_t vs = __riscv_vfmv_v_f_f32m1(0.0f, 1);
+ vfloat32m2_t vsum;
+ vfloat16m1_t ax;
+ vfloat16m1_t ay;
+ vsum = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vmv_v_x_u32m2(0, vl));
+ for (int i = 0; i < n; i += vl) {
+ vl = __riscv_vsetvl_e16m1(n - i);
+ ax = __riscv_vle16_v_f16m1_tu(ax, (const _Float16 *)&x[i], vl);
+ ay = __riscv_vle16_v_f16m1_tu(ay, (const _Float16 *)&y[i], vl);
+ vsum = __riscv_vfwmacc_vv_f32m2_tu(vsum, ax, ay, vl);
+ }
+ vl = __riscv_vsetvlmax_e32m1();
+ vfloat32m1_t ac0 = __riscv_vfadd_vv_f32m1(__riscv_vget_v_f32m2_f32m1(vsum, 0), __riscv_vget_v_f32m2_f32m1(vsum, 1), vl);
+ vs = __riscv_vfredusum_vs_f32m1_f32m1(ac0, vs, vl);
+ sumf += __riscv_vfmv_f_s_f32m1_f32(vs);
+ #else
+ for (int i = 0; i < n; ++i) {
+ sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
+ }
+ #endif // __riscv_zvfh
+ #else
+ const int np = (n & ~(GGML_F16_DOT_STEP - 1));
+
+ GGML_F16_DOT_VEC sum[GGML_F16_DOT_ARR] = { GGML_F16_DOT_VEC_ZERO };
+
+ GGML_F16_DOT_VEC ax[GGML_F16_DOT_ARR];
+ GGML_F16_DOT_VEC ay[GGML_F16_DOT_ARR];
+
+ for (int i = 0; i < np; i += GGML_F16_DOT_STEP) {
+ for (int j = 0; j < GGML_F16_DOT_ARR; j++) {
+ ax[j] = GGML_F16_DOT_VEC_LOAD(x + i + j*GGML_F16_DOT_EPR, j);
+ ay[j] = GGML_F16_DOT_VEC_LOAD(y + i + j*GGML_F16_DOT_EPR, j);
+
+ sum[j] = GGML_F16_DOT_VEC_FMA(sum[j], ax[j], ay[j]);
+ }
+ }
+
+ // reduce sum0..sum3 to sum0
+ GGML_F16_DOT_VEC_REDUCE(sumf, sum);
+
+ // leftovers
+ for (int i = np; i < n; ++i) {
+ sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
+ }
+ // if you hit this, you are likely running outside the FP range
+ assert(!isnan(sumf) && !isinf(sumf));
+ #endif
+#else
+ for (int i = 0; i < n; ++i) {
+ sumf += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[i])*GGML_CPU_FP16_TO_FP32(y[i]));
+ }
+#endif // GGML_SIMD
+
+ *s = sumf;
+}
+
+void ggml_vec_silu_f32(const int n, float * y, const float * x) {
+ int i = 0;
+#if defined(__AVX512F__) && defined(__AVX512DQ__)
+ for (; i + 15 < n; i += 16) {
+ _mm512_storeu_ps(y + i, ggml_v_silu(_mm512_loadu_ps(x + i)));
+ }
+#elif defined(__AVX2__) && defined(__FMA__)
+ for (; i + 7 < n; i += 8) {
+ _mm256_storeu_ps(y + i, ggml_v_silu(_mm256_loadu_ps(x + i)));
+ }
+#elif defined(__SSE2__)
+ for (; i + 3 < n; i += 4) {
+ _mm_storeu_ps(y + i, ggml_v_silu(_mm_loadu_ps(x + i)));
+ }
+#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+ const int vlen = svcntw();
+ for (; i < n; i += vlen) {
+ const svbool_t pg = svwhilelt_b32_s32(i, n);
+ svst1_f32(pg, y + i, ggml_v_silu(pg, svld1_f32(pg, x + i)));
+ }
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+ for (; i + 3 < n; i += 4) {
+ vst1q_f32(y + i, ggml_v_silu(vld1q_f32(x + i)));
+ }
+#elif defined(__riscv_v_intrinsic)
+ for (int vl; i < n; i += vl) {
+ vl = __riscv_vsetvl_e32m2(n - i);
+ vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
+ vfloat32m2_t vy = ggml_v_silu_m2(vx, vl);
+ __riscv_vse32_v_f32m2(&y[i], vy, vl);
+ }
+#endif
+ for (; i < n; ++i) {
+ y[i] = ggml_silu_f32(x[i]);
+ }
+}
+
+void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float * g) {
+ int i = 0;
+#if defined(__AVX512F__) && defined(__AVX512DQ__)
+ for (; i + 15 < n; i += 16) {
+ _mm512_storeu_ps(y + i, _mm512_mul_ps(ggml_v_silu(_mm512_loadu_ps(x + i)), _mm512_loadu_ps(g + i)));
+ }
+#elif defined(__AVX2__) && defined(__FMA__)
+ for (; i + 7 < n; i += 8) {
+ _mm256_storeu_ps(y + i, _mm256_mul_ps(ggml_v_silu(_mm256_loadu_ps(x + i)), _mm256_loadu_ps(g + i)));
+ }
+#elif defined(__SSE2__)
+ for (; i + 3 < n; i += 4) {
+ _mm_storeu_ps(y + i, _mm_mul_ps(ggml_v_silu(_mm_loadu_ps(x + i)), _mm_loadu_ps(g + i)));
+ }
+#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+ const int vlen = svcntw();
+ for (; i < n; i += vlen) {
+ const svbool_t pg = svwhilelt_b32_s32(i, n);
+ svst1_f32(pg, y + i, svmul_f32_x(pg, ggml_v_silu(pg, svld1_f32(pg, x + i)), svld1_f32(pg, g + i)));
+ }
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+ for (; i + 3 < n; i += 4) {
+ vst1q_f32(y + i, vmulq_f32(ggml_v_silu(vld1q_f32(x + i)), vld1q_f32(g + i)));
+ }
+#elif defined(__riscv_v_intrinsic)
+ for (int vl; i < n; i += vl) {
+ vl = __riscv_vsetvl_e32m2(n - i);
+ vfloat32m2_t vx = __riscv_vle32_v_f32m2(&x[i], vl);
+ vfloat32m2_t vg = __riscv_vle32_v_f32m2(&g[i], vl);
+ vfloat32m2_t vy = __riscv_vfmul_vv_f32m2(ggml_v_silu_m2(vx, vl), vg, vl);
+ __riscv_vse32_v_f32m2(&y[i], vy, vl);
+ }
+#endif
+ for (; i < n; ++i) {
+ y[i] = ggml_silu_f32(x[i]) * g[i];
+ }
+}
+
+ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const float mean) {
+ int i = 0;
+ ggml_float sum = 0;
+// TODO: optimize to process the remaining elements in groups using the smaller vector sizes from AVX2 and SSE
+// ref: https://github.com/ggml-org/llama.cpp/pull/15953#pullrequestreview-3310928344
+#if defined(__AVX512F__) && defined(__AVX512DQ__)
+ for (; i + 15 < n; i += 16) {
+ __m512 val = _mm512_sub_ps(_mm512_loadu_ps(x + i),
+ _mm512_set1_ps(mean));
+ _mm512_storeu_ps(y + i, val);
+ sum += (ggml_float)_mm512_reduce_add_ps(_mm512_mul_ps(val, val));
+ }
+#elif defined(__AVX2__) && defined(__FMA__)
+ for (; i + 7 < n; i += 8) {
+ __m256 val = _mm256_sub_ps(_mm256_loadu_ps(x + i),
+ _mm256_set1_ps(mean));
+ _mm256_storeu_ps(y + i, val);
+ val = _mm256_mul_ps(val,val);
+ __m128 val2 = _mm_add_ps(_mm256_extractf128_ps(val, 1),
+ _mm256_castps256_ps128(val));
+ val2 = _mm_add_ps(val2, _mm_movehl_ps(val2, val2));
+ val2 = _mm_add_ss(val2, _mm_movehdup_ps(val2));
+ sum += (ggml_float)_mm_cvtss_f32(val2);
+ }
+#elif defined(__SSE2__)
+ for (; i + 3 < n; i += 4) {
+ __m128 val = _mm_sub_ps(_mm_loadu_ps(x + i),
+ _mm_set1_ps(mean));
+ _mm_storeu_ps(y + i, val);
+ val = _mm_mul_ps(val, val);
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+ val = _mm_add_ps(val, _mm_movehl_ps(val, val));
+ val = _mm_add_ss(val, _mm_movehdup_ps(val));
+#else
+ __m128 tmp = _mm_shuffle_ps(val, val, _MM_SHUFFLE(2, 3, 0, 1));
+ val = _mm_add_ps(val, tmp);
+ tmp = _mm_movehl_ps(tmp, val);
+ val = _mm_add_ss(val, tmp);
+#endif // __AVX__ || __AVX2__ || __AVX512F__
+ sum += (ggml_float)_mm_cvtss_f32(val);
+ }
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+ for (; i + 3 < n; i += 4) {
+ float32x4_t val = vsubq_f32(vld1q_f32(x + i),
+ vdupq_n_f32(mean));
+ vst1q_f32(y + i, val);
+ val = vmulq_f32(val, val);
+ sum += (ggml_float)vaddvq_f32(val);
+ }
+#elif defined(__VXE__) || defined(__VXE2__)
+ for (; i + 3 < n; i += 4) {
+ float32x4_t val = vec_sub(vec_xl(0, x + i), vec_splats(mean));
+ vec_xst(val, 0, y + i);
+ val = vec_mul(val, val);
+ sum += (ggml_float)vec_hsum_f32x4(val);
+ }
+#elif defined(__riscv_v_intrinsic)
+ vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
+ for (int vl; i < n; i += vl) {
+ vl = __riscv_vsetvl_e32m2(n - i);
+ vfloat32m2_t val = __riscv_vfsub_vf_f32m2(__riscv_vle32_v_f32m2(&x[i], vl), mean, vl);
+ __riscv_vse32_v_f32m2(&y[i], val, vl);
+ val = __riscv_vfmul_vv_f32m2(val, val, vl);
+ vsum = __riscv_vfwredusum_vs_f32m2_f64m1(val, vsum, vl);
+ }
+ sum = (ggml_float)__riscv_vfmv_f_s_f64m1_f64(vsum);
+#endif
+ for (; i < n; ++i) {
+ float val = x[i] - mean;
+ y[i] = val;
+ val *= val;
+ sum += (ggml_float)val;
+ }
+ return sum/n;
+}
+
+ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max) {
+ int i = 0;
+ ggml_float sum = 0;
+#if defined(__AVX512F__) && defined(__AVX512DQ__)
+ for (; i + 15 < n; i += 16) {
+ __m512 val = ggml_v_expf(_mm512_sub_ps(_mm512_loadu_ps(x + i),
+ _mm512_set1_ps(max)));
+ _mm512_storeu_ps(y + i, val);
+ sum += (ggml_float)_mm512_reduce_add_ps(val);
+ }
+#elif defined(__AVX2__) && defined(__FMA__)
+ for (; i + 7 < n; i += 8) {
+ __m256 val = ggml_v_expf(_mm256_sub_ps(_mm256_loadu_ps(x + i),
+ _mm256_set1_ps(max)));
+ _mm256_storeu_ps(y + i, val);
+ __m128 val2 = _mm_add_ps(_mm256_extractf128_ps(val, 1),
+ _mm256_castps256_ps128(val));
+ val2 = _mm_add_ps(val2, _mm_movehl_ps(val2, val2));
+ val2 = _mm_add_ss(val2, _mm_movehdup_ps(val2));
+ sum += (ggml_float)_mm_cvtss_f32(val2);
+ }
+#elif defined(__SSE2__)
+ for (; i + 3 < n; i += 4) {
+ __m128 val = ggml_v_expf(_mm_sub_ps(_mm_loadu_ps(x + i),
+ _mm_set1_ps(max)));
+ _mm_storeu_ps(y + i, val);
+#if defined(__AVX__) || defined(__AVX2__) || defined(__AVX512F__)
+ val = _mm_add_ps(val, _mm_movehl_ps(val, val));
+ val = _mm_add_ss(val, _mm_movehdup_ps(val));
+#else
+ __m128 tmp = _mm_shuffle_ps(val, val, _MM_SHUFFLE(2, 3, 0, 1));
+ val = _mm_add_ps(val, tmp);
+ tmp = _mm_movehl_ps(tmp, val);
+ val = _mm_add_ss(val, tmp);
+#endif
+ sum += (ggml_float)_mm_cvtss_f32(val);
+ }
+#elif defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+ const int vlen = svcntw();
+ for (; i < n; i += vlen) {
+ const svbool_t pg = svwhilelt_b32_s32(i, n);
+ svfloat32_t val = ggml_v_expf(pg, svsub_f32_x(pg, svld1_f32(pg, x + i),
+ svdup_n_f32_x(pg, max)));
+ svst1_f32(pg, y + i, val);
+ sum += (ggml_float)svaddv_f32(pg, val);
+ }
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+ for (; i + 3 < n; i += 4) {
+ float32x4_t val = ggml_v_expf(vsubq_f32(vld1q_f32(x + i),
+ vdupq_n_f32(max)));
+ vst1q_f32(y + i, val);
+ sum += (ggml_float)vaddvq_f32(val);
+ }
+#elif defined(__riscv_v_intrinsic)
+ vfloat64m1_t vsum = __riscv_vfmv_v_f_f64m1(0, 1);
+ for (int avl; i < n; i += avl) {
+ avl = __riscv_vsetvl_e32m2(n - i);
+ vfloat32m2_t val = ggml_v_expf_m2(__riscv_vfsub_vf_f32m2(__riscv_vle32_v_f32m2(&x[i], avl), max, avl), avl);
+ __riscv_vse32_v_f32m2(&y[i], val, avl);
+ vsum = __riscv_vfwredusum_vs_f32m2_f64m1(val, vsum, avl);
+ }
+ return (ggml_float)__riscv_vfmv_f_s_f64m1_f64(vsum);
+#endif
+ for (; i < n; ++i) {
+ float val = expf(x[i] - max);
+ sum += (ggml_float)val;
+ y[i] = val;
+ }
+ return sum;
+}
+
+ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max) {
+ // log(soft_max) = log(soft_max_i / soft_max_sum) = log(soft_max_i) - log(soft_max_sum) = (logit_i - max) - log(soft_max_i)
+
+ int i = 0;
+ ggml_float sum = 0;
+ for (; i < n; ++i) {
+ float val = x[i] - max;
+ y[i] = val;
+ sum += (ggml_float)expf(val);
+ }
+ return sum = (ggml_float)logf(sum);
+}
diff --git a/python/freetoken/kernel/csrc/ggml/ggml-cpu/vec.h b/python/freetoken/kernel/csrc/ggml/ggml-cpu/vec.h
new file mode 100644
index 00000000..59fa88f2
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/ggml-cpu/vec.h
@@ -0,0 +1,1630 @@
+// Vectorized functions for fundamental operations
+
+#pragma once
+
+#include "ggml-impl.h"
+#include "simd-mappings.h"
+#include "ggml.h"
+#include "ggml-cpu.h"
+
+#if defined(GGML_USE_ACCELERATE)
+#include <Accelerate/Accelerate.h>
+#endif
+
+// floating point type used to accumulate sums
+typedef double ggml_float;
+
+#if defined(__ARM_FEATURE_SVE)
+inline static void ggml_sve_f16_fma_widened(
+ svfloat32_t * acc_lo,
+ svfloat32_t * acc_hi,
+ svfloat16_t x,
+ svfloat16_t y) {
+#if defined(__ARM_FEATURE_SVE2)
+ *acc_lo = svmlalb_f32(*acc_lo, x, y);
+ *acc_hi = svmlalt_f32(*acc_hi, x, y);
+#else
+ // Plain SVE fallback path if SVE2 instructions not available
+ svfloat16_t x_even = svtrn1_f16(x, x);
+ svfloat16_t x_odd = svtrn2_f16(x, x);
+
+ svfloat16_t y_even = svtrn1_f16(y, y);
+ svfloat16_t y_odd = svtrn2_f16(y, y);
+
+ svbool_t pg = svptrue_b32();
+
+ *acc_lo = svmla_f32_x(pg, *acc_lo, svcvt_f32_f16_x(pg, x_even), svcvt_f32_f16_x(pg, y_even));
+ *acc_hi = svmla_f32_x(pg, *acc_hi, svcvt_f32_f16_x(pg, x_odd), svcvt_f32_f16_x(pg, y_odd));
+#endif
+}
+
+inline static ggml_float ggml_sve_sum_f32x2(svfloat32_t sum_lo, svfloat32_t sum_hi) {
+ return (ggml_float) (svaddv_f32(svptrue_b32(), sum_lo) + svaddv_f32(svptrue_b32(), sum_hi));
+}
+#endif
+
+#define GGML_GELU_FP16
+#define GGML_GELU_QUICK_FP16
+
+#define GGML_SOFT_MAX_UNROLL 4
+#define GGML_VEC_DOT_UNROLL 2
+#define GGML_VEC_MAD_UNROLL 32
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+//
+// global data
+//
+
+// precomputed gelu table for f16 (128 KB)
+extern ggml_fp16_t ggml_table_gelu_f16[1 << 16];
+
+// precomputed quick gelu table for f16 (128 KB)
+extern ggml_fp16_t ggml_table_gelu_quick_f16[1 << 16];
+
+//
+// fundamental operations
+//
+
+void ggml_vec_dot_f32(int n, float * GGML_RESTRICT s, size_t bs, const float * GGML_RESTRICT x, size_t bx, const float * GGML_RESTRICT y, size_t by, int nrc);
+void ggml_vec_dot_bf16(int n, float * GGML_RESTRICT s, size_t bs, ggml_bf16_t * GGML_RESTRICT x, size_t bx, ggml_bf16_t * GGML_RESTRICT y, size_t by, int nrc);
+void ggml_vec_dot_f16(int n, float * GGML_RESTRICT s, size_t bs, ggml_fp16_t * GGML_RESTRICT x, size_t bx, ggml_fp16_t * GGML_RESTRICT y, size_t by, int nrc);
+
+void ggml_vec_silu_f32(const int n, float * y, const float * x);
+ggml_float ggml_vec_cvar_f32(const int n, float * y, const float * x, const float mean); //it will also center y ( y = y - mean )
+ggml_float ggml_vec_soft_max_f32(const int n, float * y, const float * x, float max);
+ggml_float ggml_vec_log_soft_max_f32(const int n, float * y, const float * x, float max);
+
+inline static void ggml_vec_set_i8(const int n, int8_t * x, const int8_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+inline static void ggml_vec_set_i16(const int n, int16_t * x, const int16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+
+inline static void ggml_vec_set_i32(const int n, int32_t * x, const int32_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+inline static void ggml_vec_cpy_i32(const int n, int32_t * y, const int32_t * x) { for (int i = 0; i < n; ++i) y[i] = x[i]; }
+
+inline static void ggml_vec_set_f16(const int n, ggml_fp16_t * x, const ggml_fp16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+inline static void ggml_vec_set_bf16(const int n, ggml_bf16_t * x, const ggml_bf16_t v) { for (int i = 0; i < n; ++i) x[i] = v; }
+
+inline static void ggml_vec_add_f32 (const int n, float * z, const float * x, const float * y) {
+ int i = 0;
+#if defined(__AVX2__)
+ for (; i + 7 < n; i += 8) {
+ __m256 vx = _mm256_loadu_ps(x + i);
+ __m256 vy = _mm256_loadu_ps(y + i);
+ __m256 vz = _mm256_add_ps(vx, vy);
+ _mm256_storeu_ps(z + i, vz);
+ }
+#endif
+ for (; i < n; ++i) {
+ z[i] = x[i] + y[i];
+ }
+}
+
+inline static void ggml_vec_add_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
+ for (int i = 0; i < n; ++i) {
+ z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) + GGML_CPU_FP16_TO_FP32(y[i]));
+ }
+}
+inline static void ggml_vec_add1_f32(const int n, float * z, const float * x, const float v) { for (int i = 0; i < n; ++i) z[i] = x[i] + v; }
+inline static void ggml_vec_acc_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] += x[i]; }
+inline static void ggml_vec_acc1_f32(const int n, float * y, const float v) { for (int i = 0; i < n; ++i) y[i] += v; }
+inline static void ggml_vec_sub_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i] - y[i]; }
+inline static void ggml_vec_sub_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
+ for (int i = 0; i < n; ++i) {
+ z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) - GGML_CPU_FP16_TO_FP32(y[i]));
+ }
+}
+inline static void ggml_vec_set_f32 (const int n, float * x, const float v) { for (int i = 0; i < n; ++i) x[i] = v; }
+inline static void ggml_vec_cpy_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]; }
+inline static void ggml_vec_neg_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = -x[i]; }
+inline static void ggml_vec_neg_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16(-GGML_CPU_FP16_TO_FP32(x[i]));
+ }
+}
+
+inline static void ggml_vec_mul_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]*y[i]; }
+inline static void ggml_vec_mul_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
+ for (int i = 0; i < n; ++i) {
+ z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) * GGML_CPU_FP16_TO_FP32(y[i]));
+ }
+}
+inline static void ggml_vec_div_f32 (const int n, float * z, const float * x, const float * y) { for (int i = 0; i < n; ++i) z[i] = x[i]/y[i]; }
+inline static void ggml_vec_div_f16 (const int n, ggml_fp16_t * z, const ggml_fp16_t * x, const ggml_fp16_t * y) {
+ for (int i = 0; i < n; ++i) {
+ z[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(x[i]) / GGML_CPU_FP16_TO_FP32(y[i]));
+ }
+}
+
+// compute GGML_VEC_DOT_UNROLL dot products at once
+// xs - x row stride in bytes
+inline static void ggml_vec_dot_f16_unroll(const int n, const int xs, float * GGML_RESTRICT s, void * GGML_RESTRICT xv, ggml_fp16_t * GGML_RESTRICT y) {
+ ggml_float sumf[GGML_VEC_DOT_UNROLL] = { 0.0 };
+
+ ggml_fp16_t * GGML_RESTRICT x[GGML_VEC_DOT_UNROLL];
+
+ for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) {
+ x[i] = (ggml_fp16_t *) ((char *) xv + i*xs);
+ }
+
+#if defined(GGML_SIMD)
+ #if defined(__ARM_FEATURE_SVE)
+
+ const int ggml_f16_epr = svcnth();
+ const int ggml_f16_step = 2 * ggml_f16_epr;
+ int np = n - (n % ggml_f16_step);
+ int np2 = n - (n % ggml_f16_epr);
+
+ svfloat32_t sum_0_0_lo = svdup_n_f32(0.0f);
+ svfloat32_t sum_0_0_hi = svdup_n_f32(0.0f);
+ svfloat32_t sum_0_1_lo = svdup_n_f32(0.0f);
+ svfloat32_t sum_0_1_hi = svdup_n_f32(0.0f);
+ svfloat32_t sum_1_0_lo = svdup_n_f32(0.0f);
+ svfloat32_t sum_1_0_hi = svdup_n_f32(0.0f);
+ svfloat32_t sum_1_1_lo = svdup_n_f32(0.0f);
+ svfloat32_t sum_1_1_hi = svdup_n_f32(0.0f);
+
+ for (int i = 0; i < np; i += ggml_f16_step) {
+ const svfloat16_t ay0 = GGML_F16x_VEC_LOAD(y + i, 0);
+ const svfloat16_t ax00 = GGML_F16x_VEC_LOAD(x[0] + i, 0);
+ const svfloat16_t ax01 = GGML_F16x_VEC_LOAD(x[1] + i, 0);
+
+ ggml_sve_f16_fma_widened(&sum_0_0_lo, &sum_0_0_hi, ax00, ay0);
+ ggml_sve_f16_fma_widened(&sum_1_0_lo, &sum_1_0_hi, ax01, ay0);
+
+ const svfloat16_t ay1 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 0);
+ const svfloat16_t ax10 = GGML_F16x_VEC_LOAD(x[0] + i + 1 * ggml_f16_epr, 0);
+ const svfloat16_t ax11 = GGML_F16x_VEC_LOAD(x[1] + i + 1 * ggml_f16_epr, 0);
+
+ ggml_sve_f16_fma_widened(&sum_0_1_lo, &sum_0_1_hi, ax10, ay1);
+ ggml_sve_f16_fma_widened(&sum_1_1_lo, &sum_1_1_hi, ax11, ay1);
+ }
+
+ for (int i = np; i < np2; i += ggml_f16_epr) {
+ const svfloat16_t ry = GGML_F16x_VEC_LOAD(y + i, 0);
+ const svfloat16_t rx0 = GGML_F16x_VEC_LOAD(x[0] + i, 0);
+ const svfloat16_t rx1 = GGML_F16x_VEC_LOAD(x[1] + i, 0);
+
+ ggml_sve_f16_fma_widened(&sum_0_0_lo, &sum_0_0_hi, rx0, ry);
+ ggml_sve_f16_fma_widened(&sum_1_0_lo, &sum_1_0_hi, rx1, ry);
+ }
+
+ if (np2 < n) {
+ const svbool_t pg = svwhilelt_b16(np2, n);
+ const svfloat16_t ay = svld1_f16(pg, (const __fp16 *)(y + np2));
+ const svfloat16_t ax0 = svld1_f16(pg, (const __fp16 *)(x[0] + np2));
+ const svfloat16_t ax1 = svld1_f16(pg, (const __fp16 *)(x[1] + np2));
+
+ ggml_sve_f16_fma_widened(&sum_0_0_lo, &sum_0_0_hi, ax0, ay);
+ ggml_sve_f16_fma_widened(&sum_1_0_lo, &sum_1_0_hi, ax1, ay);
+ }
+
+ svfloat32_t sum_0_lo = svadd_f32_x(DEFAULT_PG32, sum_0_0_lo, sum_0_1_lo);
+ svfloat32_t sum_0_hi = svadd_f32_x(DEFAULT_PG32, sum_0_0_hi, sum_0_1_hi);
+ svfloat32_t sum_1_lo = svadd_f32_x(DEFAULT_PG32, sum_1_0_lo, sum_1_1_lo);
+ svfloat32_t sum_1_hi = svadd_f32_x(DEFAULT_PG32, sum_1_0_hi, sum_1_1_hi);
+ sumf[0] = ggml_sve_sum_f32x2(sum_0_lo, sum_0_hi);
+ sumf[1] = ggml_sve_sum_f32x2(sum_1_lo, sum_1_hi);
+ np = n;
+ #elif defined(__riscv_v_intrinsic)
+ #if defined(__riscv_zvfh)
+ size_t vl = __riscv_vsetvlmax_e32m4();
+
+ // initialize accumulators to all zeroes
+ vfloat32m4_t vsum0_0 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+ vfloat32m4_t vsum0_1 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+ vfloat32m4_t vsum1_0 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+ vfloat32m4_t vsum1_1 = __riscv_vfmv_v_f_f32m4(0.0f, vl);
+
+ // calculate step size
+ const size_t epr = __riscv_vsetvlmax_e16m2();
+ const size_t step = epr * 2;
+ int np = (n & ~(step - 1));
+
+ // unroll by 2 along the row dimension
+ for (int i = 0; i < np; i += step) {
+ vfloat16m2_t ay0 = __riscv_vle16_v_f16m2((const _Float16 *)(y + i), epr);
+ vfloat16m2_t ax0_0 = __riscv_vle16_v_f16m2((const _Float16 *)(x[0] + i), epr);
+ vfloat16m2_t ax1_0 = __riscv_vle16_v_f16m2((const _Float16 *)(x[1] + i), epr);
+ vsum0_0 = __riscv_vfwmacc_vv_f32m4(vsum0_0, ax0_0, ay0, epr);
+ vsum1_0 = __riscv_vfwmacc_vv_f32m4(vsum1_0, ax1_0, ay0, epr);
+
+ vfloat16m2_t ay1 = __riscv_vle16_v_f16m2((const _Float16 *)(y + i + epr), epr);
+ vfloat16m2_t ax0_1 = __riscv_vle16_v_f16m2((const _Float16 *)(x[0] + i + epr), epr);
+ vfloat16m2_t ax1_1 = __riscv_vle16_v_f16m2((const _Float16 *)(x[1] + i + epr), epr);
+ vsum0_1 = __riscv_vfwmacc_vv_f32m4(vsum0_1, ax0_1, ay1, epr);
+ vsum1_1 = __riscv_vfwmacc_vv_f32m4(vsum1_1, ax1_1, ay1, epr);
+ }
+
+ vfloat32m4_t vsum0 = __riscv_vfadd_vv_f32m4(vsum0_0, vsum0_1, vl);
+ vfloat32m4_t vsum1 = __riscv_vfadd_vv_f32m4(vsum1_0, vsum1_1, vl);
+
+ // leftovers
+ for (int i = np; i < n; i += vl) {
+ vl = __riscv_vsetvl_e16m2(n - i);
+ vfloat16m2_t ay = __riscv_vle16_v_f16m2((const _Float16 *)(y + i), vl);
+ vfloat16m2_t ax0 = __riscv_vle16_v_f16m2((const _Float16 *)(x[0] + i), vl);
+ vfloat16m2_t ax1 = __riscv_vle16_v_f16m2((const _Float16 *)(x[1] + i), vl);
+
+ vsum0 = __riscv_vfwmacc_vv_f32m4(vsum0, ax0, ay, vl);
+ vsum1 = __riscv_vfwmacc_vv_f32m4(vsum1, ax1, ay, vl);
+ }
+
+ // reduce
+ vl = __riscv_vsetvlmax_e32m2();
+ vfloat32m2_t acc0_0 = __riscv_vfadd_vv_f32m2(__riscv_vget_v_f32m4_f32m2(vsum0, 0),
+ __riscv_vget_v_f32m4_f32m2(vsum0, 1), vl);
+ vl = __riscv_vsetvlmax_e32m1();
+ vfloat32m1_t acc0_1 = __riscv_vfadd_vv_f32m1(__riscv_vget_v_f32m2_f32m1(acc0_0, 0),
+ __riscv_vget_v_f32m2_f32m1(acc0_0, 1), vl);
+ vfloat32m1_t redsum0 = __riscv_vfredusum_vs_f32m1_f32m1(
+ acc0_1, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
+
+ vl = __riscv_vsetvlmax_e32m2();
+ vfloat32m2_t acc1_0 = __riscv_vfadd_vv_f32m2(__riscv_vget_v_f32m4_f32m2(vsum1, 0),
+ __riscv_vget_v_f32m4_f32m2(vsum1, 1), vl);
+ vl = __riscv_vsetvlmax_e32m1();
+ vfloat32m1_t acc1_1 = __riscv_vfadd_vv_f32m1(__riscv_vget_v_f32m2_f32m1(acc1_0, 0),
+ __riscv_vget_v_f32m2_f32m1(acc1_0, 1), vl);
+ vfloat32m1_t redsum1 = __riscv_vfredusum_vs_f32m1_f32m1(
+ acc1_1, __riscv_vfmv_v_f_f32m1(0.0f, 1), vl);
+ sumf[0] = __riscv_vfmv_f_s_f32m1_f32(redsum0);
+ sumf[1] = __riscv_vfmv_f_s_f32m1_f32(redsum1);
+ np = n;
+ #else
+ const int np = 0;
+ #endif
+ #else
+ const int np = (n & ~(GGML_F16_DOT_STEP - 1));
+
+ GGML_F16_DOT_VEC sum[GGML_VEC_DOT_UNROLL][GGML_F16_DOT_ARR] = { { GGML_F16_DOT_VEC_ZERO } };
+
+ GGML_F16_DOT_VEC ax[GGML_F16_DOT_ARR];
+ GGML_F16_DOT_VEC ay[GGML_F16_DOT_ARR];
+
+ for (int i = 0; i < np; i += GGML_F16_DOT_STEP) {
+ for (int j = 0; j < GGML_F16_DOT_ARR; j++) {
+ ay[j] = GGML_F16_DOT_VEC_LOAD(y + i + j*GGML_F16_DOT_EPR, j);
+
+ for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
+ ax[j] = GGML_F16_DOT_VEC_LOAD(x[k] + i + j*GGML_F16_DOT_EPR, j);
+
+ sum[k][j] = GGML_F16_DOT_VEC_FMA(sum[k][j], ax[j], ay[j]);
+ }
+ }
+ }
+
+ // reduce sum0..sum3 to sum0
+ for (int k = 0; k < GGML_VEC_DOT_UNROLL; ++k) {
+ GGML_F16_DOT_VEC_REDUCE(sumf[k], sum[k]);
+ }
+ #endif
+#else
+ // scalar path
+ const int np = 0;
+#endif
+ // scalar and leftovers
+ for (int i = np; i < n; ++i) {
+ for (int j = 0; j < GGML_VEC_DOT_UNROLL; ++j) {
+ sumf[j] += (ggml_float)(GGML_CPU_FP16_TO_FP32(x[j][i])*GGML_CPU_FP16_TO_FP32(y[i]));
+ }
+ }
+
+ for (int i = 0; i < GGML_VEC_DOT_UNROLL; ++i) {
+ s[i] = (float)sumf[i];
+ }
+}
+
+inline static void ggml_vec_mad_f32(const int n, float * GGML_RESTRICT y, const float * GGML_RESTRICT x, const float v) {
+#if defined(GGML_SIMD)
+ #if defined(__ARM_FEATURE_SVE)
+
+ const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
+ const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
+ const int ggml_f32_step = 8 * ggml_f32_epr; // choose 8 SVE registers
+ GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+
+ const int np = (n & ~(ggml_f32_step - 1));
+ svfloat32_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
+ svfloat32_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
+ for (int i = 0; i < np; i += ggml_f32_step) {
+
+ ax1 = GGML_F32_VEC_LOAD(x + i);
+ ay1 = GGML_F32_VEC_LOAD(y + i);
+ ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx);
+
+ GGML_F32_VEC_STORE(y + i, ay1);
+
+ ax2 = GGML_F32_VEC_LOAD(x + i + 1*ggml_f32_epr);
+ ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
+ ay2 = GGML_F32_VEC_FMA(ay2, ax2, vx);
+
+ GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2);
+
+ ax3 = GGML_F32_VEC_LOAD(x + i + 2*ggml_f32_epr);
+ ay3 = GGML_F32_VEC_LOAD(y + i + 2*ggml_f32_epr);
+ ay3 = GGML_F32_VEC_FMA(ay3, ax3, vx);
+
+ GGML_F32_VEC_STORE(y + i + 2*ggml_f32_epr, ay3);
+
+ ax4 = GGML_F32_VEC_LOAD(x + i + 3*ggml_f32_epr);
+ ay4 = GGML_F32_VEC_LOAD(y + i + 3*ggml_f32_epr);
+ ay4 = GGML_F32_VEC_FMA(ay4, ax4, vx);
+
+ GGML_F32_VEC_STORE(y + i + 3*ggml_f32_epr, ay4);
+
+ ax5 = GGML_F32_VEC_LOAD(x + i + 4*ggml_f32_epr);
+ ay5 = GGML_F32_VEC_LOAD(y + i + 4*ggml_f32_epr);
+ ay5 = GGML_F32_VEC_FMA(ay5, ax5, vx);
+
+ GGML_F32_VEC_STORE(y + i + 4*ggml_f32_epr, ay5);
+
+ ax6 = GGML_F32_VEC_LOAD(x + i + 5*ggml_f32_epr);
+ ay6 = GGML_F32_VEC_LOAD(y + i + 5*ggml_f32_epr);
+ ay6 = GGML_F32_VEC_FMA(ay6, ax6, vx);
+
+ GGML_F32_VEC_STORE(y + i + 5*ggml_f32_epr, ay6);
+
+ ax7 = GGML_F32_VEC_LOAD(x + i + 6*ggml_f32_epr);
+ ay7 = GGML_F32_VEC_LOAD(y + i + 6*ggml_f32_epr);
+ ay7 = GGML_F32_VEC_FMA(ay7, ax7, vx);
+
+ GGML_F32_VEC_STORE(y + i + 6*ggml_f32_epr, ay7);
+
+ ax8 = GGML_F32_VEC_LOAD(x + i + 7*ggml_f32_epr);
+ ay8 = GGML_F32_VEC_LOAD(y + i + 7*ggml_f32_epr);
+ ay8 = GGML_F32_VEC_FMA(ay8, ax8, vx);
+
+ GGML_F32_VEC_STORE(y + i + 7*ggml_f32_epr, ay8);
+ }
+ // leftovers
+ // Since 8 unrolls are done in above loop, leftovers lie in range [0, ggml_f32_step] which is handled in below loop
+ const int np2 = (n & ~(ggml_f32_epr - 1));
+ for (int i = np; i < np2; i += ggml_f32_epr) {
+ ax1 = GGML_F32_VEC_LOAD(x + i);
+ ay1 = GGML_F32_VEC_LOAD(y + i);
+ ay1 = GGML_F32_VEC_FMA(ay1, ax1, vx);
+
+ GGML_F32_VEC_STORE(y + i, ay1);
+ }
+ // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
+ if (np2 < n) {
+ svbool_t pg =svwhilelt_b32(np2, n);
+ ax1 = svld1_f32(pg, x + np2);
+ ay1 = svld1_f32(pg, y + np2);
+ ay1 = svmad_f32_m(pg, ax1, vx, ay1);
+
+ svst1_f32(pg, y + np2, ay1);
+ }
+ #elif defined(__riscv_v_intrinsic)
+ for (int i = 0, avl; i < n; i += avl) {
+ avl = __riscv_vsetvl_e32m8(n - i);
+ vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
+ vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
+ vfloat32m8_t ny = __riscv_vfmadd_vf_f32m8(ax, v, ay, avl);
+ __riscv_vse32_v_f32m8(&y[i], ny, avl);
+ }
+ #else
+ const int np = (n & ~(GGML_F32_STEP - 1));
+
+ GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+
+ GGML_F32_VEC ax[GGML_F32_ARR];
+ GGML_F32_VEC ay[GGML_F32_ARR];
+
+ for (int i = 0; i < np; i += GGML_F32_STEP) {
+ for (int j = 0; j < GGML_F32_ARR; j++) {
+ ax[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
+ ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+ ay[j] = GGML_F32_VEC_FMA(ay[j], ax[j], vx);
+
+ GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+ }
+ }
+
+ // leftovers
+ for (int i = np; i < n; ++i) {
+ y[i] += x[i]*v;
+ }
+ #endif
+#else
+ // scalar
+ for (int i = 0; i < n; ++i) {
+ y[i] += x[i]*v;
+ }
+#endif
+}
+
+inline static void ggml_vec_mad_f16(const int n, ggml_fp16_t * GGML_RESTRICT y, const ggml_fp16_t * GGML_RESTRICT x, const float v) {
+#if defined(GGML_SIMD) && defined(__ARM_FEATURE_SVE)
+ const int sve_register_length = svcntb() * 8;
+ const int ggml_f16_epr = sve_register_length / 16;
+ const int ggml_f16_step = 8 * ggml_f16_epr;
+
+ GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
+
+ int np = (n & ~(ggml_f16_step - 1));
+
+ svfloat16_t ax1, ax2, ax3, ax4, ax5, ax6, ax7, ax8;
+ svfloat16_t ay1, ay2, ay3, ay4, ay5, ay6, ay7, ay8;
+ for (int i = 0; i < np; i += ggml_f16_step) {
+ ax1 = GGML_F16x_VEC_LOAD(x + i + 0 * ggml_f16_epr, 0);
+ ay1 = GGML_F16x_VEC_LOAD(y + i + 0 * ggml_f16_epr, 0);
+ ay1 = GGML_F16x_VEC_FMA(ay1, ax1, vx);
+
+ GGML_F16x_VEC_STORE(y + i + 0 * ggml_f16_epr, ay1, 0);
+
+ ax2 = GGML_F16x_VEC_LOAD(x + i + 1 * ggml_f16_epr, 1);
+ ay2 = GGML_F16x_VEC_LOAD(y + i + 1 * ggml_f16_epr, 1);
+ ay2 = GGML_F16x_VEC_FMA(ay2, ax2, vx);
+
+ GGML_F16x_VEC_STORE(y + i + 1 * ggml_f16_epr, ay2, 1);
+
+ ax3 = GGML_F16x_VEC_LOAD(x + i + 2 * ggml_f16_epr, 2);
+ ay3 = GGML_F16x_VEC_LOAD(y + i + 2 * ggml_f16_epr, 2);
+ ay3 = GGML_F16x_VEC_FMA(ay3, ax3, vx);
+
+ GGML_F16x_VEC_STORE(y + i + 2 * ggml_f16_epr, ay3, 2);
+
+ ax4 = GGML_F16x_VEC_LOAD(x + i + 3 * ggml_f16_epr, 3);
+ ay4 = GGML_F16x_VEC_LOAD(y + i + 3 * ggml_f16_epr, 3);
+ ay4 = GGML_F16x_VEC_FMA(ay4, ax4, vx);
+
+ GGML_F16x_VEC_STORE(y + i + 3 * ggml_f16_epr, ay4, 3);
+
+ ax5 = GGML_F16x_VEC_LOAD(x + i + 4 * ggml_f16_epr, 4);
+ ay5 = GGML_F16x_VEC_LOAD(y + i + 4 * ggml_f16_epr, 4);
+ ay5 = GGML_F16x_VEC_FMA(ay5, ax5, vx);
+
+ GGML_F16x_VEC_STORE(y + i + 4 * ggml_f16_epr, ay5, 4);
+
+ ax6 = GGML_F16x_VEC_LOAD(x + i + 5 * ggml_f16_epr, 5);
+ ay6 = GGML_F16x_VEC_LOAD(y + i + 5 * ggml_f16_epr, 5);
+ ay6 = GGML_F16x_VEC_FMA(ay6, ax6, vx);
+
+ GGML_F16x_VEC_STORE(y + i + 5 * ggml_f16_epr, ay6, 5);
+
+ ax7 = GGML_F16x_VEC_LOAD(x + i + 6 * ggml_f16_epr, 6);
+ ay7 = GGML_F16x_VEC_LOAD(y + i + 6 * ggml_f16_epr, 6);
+ ay7 = GGML_F16x_VEC_FMA(ay7, ax7, vx);
+
+ GGML_F16x_VEC_STORE(y + i + 6 * ggml_f16_epr, ay7, 6);
+
+ ax8 = GGML_F16x_VEC_LOAD(x + i + 7 * ggml_f16_epr, 7);
+ ay8 = GGML_F16x_VEC_LOAD(y + i + 7 * ggml_f16_epr, 7);
+ ay8 = GGML_F16x_VEC_FMA(ay8, ax8, vx);
+
+ GGML_F16x_VEC_STORE(y + i + 7 * ggml_f16_epr, ay8, 7);
+ }
+ const int np2 = (n & ~(ggml_f16_epr - 1));
+ for (int k = np; k < np2; k += ggml_f16_epr) {
+ svfloat16_t rx = GGML_F16x_VEC_LOAD(x + k, 0);
+ svfloat16_t ry = GGML_F16x_VEC_LOAD(y + k, 0);
+ ry = GGML_F16x_VEC_FMA(ry, rx, vx);
+
+ GGML_F16x_VEC_STORE(y + k, ry, 0);
+ }
+
+ if (np2 < n) {
+ svbool_t pg = svwhilelt_b16(np2, n);
+ svfloat16_t hx = svld1_f16(pg, (const __fp16 *)(x + np2));
+ svfloat16_t hy = svld1_f16(pg, (const __fp16 *)(y + np2));
+ hy = svmad_f16_x(pg, hx, vx, hy);
+ svst1_f16(pg, (__fp16 *)(y + np2), hy);
+ }
+ np = n;
+#elif defined(__riscv_v_intrinsic) // implies __riscv_v_intrinsic
+ #if defined (__riscv_zvfh)
+ const ggml_fp16_t s = GGML_CPU_FP32_TO_FP16(v);
+ const _Float16 scale = *(const _Float16*)(&s);
+
+ // calculate step size
+ const int epr = __riscv_vsetvlmax_e16m4();
+ const int step = epr * 2;
+ int np = (n & ~(step - 1));
+
+ // unroll by 2
+ for (int i = 0; i < np; i += step) {
+ vfloat16m4_t ax0 = __riscv_vle16_v_f16m4((const _Float16*)x + i, epr);
+ vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, epr);
+ ay0 = __riscv_vfmacc_vf_f16m4(ay0, scale, ax0, epr);
+ __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, epr);
+ __asm__ __volatile__ ("" ::: "memory");
+
+ vfloat16m4_t ax1 = __riscv_vle16_v_f16m4((const _Float16*)x + i + epr, epr);
+ vfloat16m4_t ay1 = __riscv_vle16_v_f16m4((const _Float16*)y + i + epr, epr);
+ ay1 = __riscv_vfmacc_vf_f16m4(ay1, scale, ax1, epr);
+ __riscv_vse16_v_f16m4((_Float16*)y + i + epr, ay1, epr);
+ __asm__ __volatile__ ("" ::: "memory");
+ }
+
+ // leftovers
+ int vl;
+ for (int i = np; i < n; i += vl) {
+ vl = __riscv_vsetvl_e16m4(n - i);
+ vfloat16m4_t ax0 = __riscv_vle16_v_f16m4((const _Float16*)x + i, vl);
+ vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, vl);
+ ay0 = __riscv_vfmacc_vf_f16m4(ay0, scale, ax0, vl);
+ __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, vl);
+ }
+ np = n;
+ #else
+ // fall to scalar path
+ const int np = 0;
+ #endif
+#elif defined(GGML_SIMD)
+ const int np = (n & ~(GGML_F16_STEP - 1));
+
+ GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+
+ GGML_F16_VEC ax[GGML_F16_ARR];
+ GGML_F16_VEC ay[GGML_F16_ARR];
+
+ for (int i = 0; i < np; i += GGML_F16_STEP) {
+ for (int j = 0; j < GGML_F16_ARR; j++) {
+ ax[j] = GGML_F16_VEC_LOAD(x + i + j*GGML_F16_EPR, j);
+ ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+ ay[j] = GGML_F16_VEC_FMA(ay[j], ax[j], vx);
+
+ GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
+ }
+ }
+#else
+ // scalar path
+ const int np = 0;
+#endif
+
+ // scalar and leftovers
+ for (int i = np; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i]) + GGML_CPU_FP16_TO_FP32(x[i])*v);
+ }
+}
+
+// xs and vs are byte strides of x and v
+inline static void ggml_vec_mad_f32_unroll(const int n, const int xs, const int vs, float * GGML_RESTRICT y, const float * GGML_RESTRICT xv, const float * GGML_RESTRICT vv) {
+
+ const float * GGML_RESTRICT x[GGML_VEC_MAD_UNROLL];
+ const float * GGML_RESTRICT v[GGML_VEC_MAD_UNROLL];
+
+ for (int i = 0; i < GGML_VEC_MAD_UNROLL; ++i) {
+ x[i] = (const float *) ((const char *) xv + i*xs);
+ v[i] = (const float *) ((const char *) vv + i*vs);
+ }
+
+#if defined(GGML_SIMD)
+ #if defined(__ARM_FEATURE_SVE)
+ // scalar Route to scalar implementation //TODO: Write SVE code
+ for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+ for (int i = 0; i < n; ++i) {
+ y[i] += x[k][i]*v[k][0];
+ }
+ }
+ #elif defined(__riscv_v_intrinsic)
+ for (int i = 0, avl; i < n; i += avl) {
+ avl = __riscv_vsetvl_e32m8(n - i);
+ vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
+ for (int k = 0; k < GGML_VEC_MAD_UNROLL; k++) {
+ vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[k][i], avl);
+ ay = __riscv_vfmadd_vf_f32m8(ax, v[k][0], ay, avl);
+ }
+ __riscv_vse32_v_f32m8(&y[i], ay, avl);
+ }
+ #else
+ const int np = (n & ~(GGML_F32_STEP - 1));
+
+ GGML_F32_VEC vx[GGML_VEC_MAD_UNROLL];
+
+ for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+ vx[k] = GGML_F32_VEC_SET1(v[k][0]);
+ }
+
+ GGML_F32_VEC ax[GGML_VEC_MAD_UNROLL][GGML_F32_ARR];
+ GGML_F32_VEC ay[GGML_F32_ARR];
+
+ for (int i = 0; i < np; i += GGML_F32_STEP) {
+ for (int j = 0; j < GGML_F32_ARR; j++) {
+ ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+
+ for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+ ax[k][j] = GGML_F32_VEC_LOAD(x[k] + i + j*GGML_F32_EPR);
+ ay[j] = GGML_F32_VEC_FMA(ay[j], ax[k][j], vx[k]);
+ }
+
+ GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+ }
+ }
+
+ // leftovers
+ for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+ for (int i = np; i < n; ++i) {
+ y[i] += x[k][i]*v[k][0];
+ }
+ }
+ #endif
+#else
+ // scalar
+ for (int k = 0; k < GGML_VEC_MAD_UNROLL; ++k) {
+ for (int i = 0; i < n; ++i) {
+ y[i] += x[k][i]*v[k][0];
+ }
+ }
+#endif
+}
+
+inline static void ggml_vec_mad1_f32(const int n, float * y, const float * x, const float s, const float b) {
+#if defined(GGML_USE_ACCELERATE)
+ vDSP_vsmsa(x, 1, &s, &b, y, 1, n);
+#elif defined(GGML_SIMD)
+ #if defined(__ARM_FEATURE_SVE)
+ // scalar ; TODO: Write SVE code
+ for (int i = 0; i < n; ++i) {
+ y[i] = x[i]*s + b;
+ }
+ #elif defined(__riscv_v_intrinsic)
+ for (int i = 0, avl; i < n; i += avl) {
+ avl = __riscv_vsetvl_e32m8(n - i);
+ vfloat32m8_t ax = __riscv_vle32_v_f32m8(&x[i], avl);
+ vfloat32m8_t vb = __riscv_vfmv_v_f_f32m8(b, avl);
+ vfloat32m8_t ny = __riscv_vfmadd_vf_f32m8(ax, s, vb, avl);
+ __riscv_vse32_v_f32m8(&y[i], ny, avl);
+ }
+ #else
+ const int np = (n & ~(GGML_F32_STEP - 1));
+
+ GGML_F32_VEC vs = GGML_F32_VEC_SET1(s);
+ GGML_F32_VEC vb = GGML_F32_VEC_SET1(b);
+
+ GGML_F32_VEC ay[GGML_F32_ARR];
+
+ for (int i = 0; i < np; i += GGML_F32_STEP) {
+ for (int j = 0; j < GGML_F32_ARR; j++) {
+ ay[j] = GGML_F32_VEC_LOAD(x + i + j*GGML_F32_EPR);
+ ay[j] = GGML_F32_VEC_FMA(vb, ay[j], vs);
+
+ GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+ }
+ }
+
+ // leftovers
+ for (int i = np; i < n; ++i) {
+ y[i] = x[i]*s + b;
+ }
+ #endif
+#else
+ // scalar
+ for (int i = 0; i < n; ++i) {
+ y[i] = x[i]*s + b;
+ }
+#endif
+}
+
+//inline static void ggml_vec_scale_f32(const int n, float * y, const float v) { for (int i = 0; i < n; ++i) y[i] *= v; }
+inline static void ggml_vec_scale_f32(const int n, float * y, const float v) {
+#if defined(GGML_USE_ACCELERATE)
+ vDSP_vsmul(y, 1, &v, y, 1, n);
+#elif defined(GGML_SIMD)
+ #if defined(__ARM_FEATURE_SVE)
+ const int sve_register_length = ggml_cpu_get_sve_cnt() * 8;
+ const int ggml_f32_epr = sve_register_length / 32;//8;//svcntw(); // SVE128:4, SVE256:8, SVE512:16
+ const int ggml_f32_step = 2 * ggml_f32_epr;
+
+ GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+ const int np = (n & ~(ggml_f32_step - 1));
+ svfloat32_t ay1;
+ svfloat32_t ay2;
+ for (int i = 0; i < np; i += ggml_f32_step) {
+ ay1 = GGML_F32_VEC_LOAD(y + i);
+ ay1 = GGML_F32_VEC_MUL(ay1, vx);
+ GGML_F32_VEC_STORE(y + i, ay1);
+
+ ay2 = GGML_F32_VEC_LOAD(y + i + 1*ggml_f32_epr);
+ ay2 = GGML_F32_VEC_MUL(ay2, vx);
+ GGML_F32_VEC_STORE(y + i + 1*ggml_f32_epr, ay2);
+ }
+ // leftovers
+ // maximum number of leftover elements will be less that ggml_f32_epr. Apply predicated svmad on available elements only
+ for (int i = np; i < n; i += ggml_f32_epr) {
+ svbool_t pg = svwhilelt_b32(i, n);
+ ay1 = svld1_f32(pg, y + i);
+ ay1 = svmul_f32_m(pg, ay1, vx);
+ svst1_f32(pg, y + i, ay1);
+ }
+ #elif defined(__riscv_v_intrinsic)
+ for (int i = 0, avl; i < n; i += avl) {
+ avl = __riscv_vsetvl_e32m8(n - i);
+ vfloat32m8_t ay = __riscv_vle32_v_f32m8(&y[i], avl);
+ vfloat32m8_t ny = __riscv_vfmul_vf_f32m8(ay, v, avl);
+ __riscv_vse32_v_f32m8(&y[i], ny, avl);
+ }
+ #else
+ const int np = (n & ~(GGML_F32_STEP - 1));
+
+ GGML_F32_VEC vx = GGML_F32_VEC_SET1(v);
+
+ GGML_F32_VEC ay[GGML_F32_ARR];
+
+ for (int i = 0; i < np; i += GGML_F32_STEP) {
+ for (int j = 0; j < GGML_F32_ARR; j++) {
+ ay[j] = GGML_F32_VEC_LOAD(y + i + j*GGML_F32_EPR);
+ ay[j] = GGML_F32_VEC_MUL(ay[j], vx);
+
+ GGML_F32_VEC_STORE(y + i + j*GGML_F32_EPR, ay[j]);
+ }
+ }
+
+ // leftovers
+ for (int i = np; i < n; ++i) {
+ y[i] *= v;
+ }
+ #endif
+#else
+ // scalar
+ for (int i = 0; i < n; ++i) {
+ y[i] *= v;
+ }
+#endif
+}
+
+inline static void ggml_vec_scale_f16(const int n, ggml_fp16_t * y, const float v) {
+#if defined(GGML_SIMD) && defined(__ARM_FEATURE_SVE)
+ const int sve_register_length = svcntb() * 8;
+ const int ggml_f16_epr = sve_register_length / 16;
+ const int ggml_f16_step = 2 * ggml_f16_epr;
+
+ GGML_F16x_VEC vx = GGML_F16x_VEC_SET1(v);
+ int np = (n & ~(ggml_f16_step - 1));
+ svfloat16_t ay1, ay2;
+
+ for (int i = 0; i < np; i += ggml_f16_step) {
+ ay1 = GGML_F16x_VEC_LOAD(y + i + 0*ggml_f16_epr, 0);
+ ay1 = GGML_F16x_VEC_MUL(ay1, vx);
+ GGML_F16x_VEC_STORE(y + i + 0*ggml_f16_epr, ay1, 0);
+
+ ay2 = GGML_F16x_VEC_LOAD(y + i + 1*ggml_f16_epr, 1);
+ ay2 = GGML_F16x_VEC_MUL(ay2, vx);
+ GGML_F16x_VEC_STORE(y + i + 1*ggml_f16_epr, ay2, 1);
+ }
+ // leftovers
+ // maximum number of leftover elements will be less that ggmlF_16x_epr. Apply predicated svmad on available elements only
+ if (np < n) {
+ svbool_t pg = svwhilelt_b16(np, n);
+ svfloat16_t hy = svld1_f16(pg, (__fp16 *)(y + np));
+ svfloat16_t out = svmul_f16_m(pg, hy, vx);
+ svst1_f16(pg, (__fp16 *)(y + np), out);
+ }
+ np = n;
+#elif defined(__riscv_v_intrinsic)
+ #if defined(__riscv_zvfh)
+ const ggml_fp16_t s = GGML_CPU_FP32_TO_FP16(v);
+ const _Float16 scale = *(const _Float16*)(&s);
+
+ // calculate step size
+ const int epr = __riscv_vsetvlmax_e16m4();
+ const int step = epr * 2;
+ int np = (n & ~(step - 1));
+
+ // unroll by 2
+ for (int i = 0; i < np; i += step) {
+ vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, epr);
+ ay0 = __riscv_vfmul_vf_f16m4(ay0, scale, epr);
+ __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, epr);
+ __asm__ __volatile__ ("" ::: "memory");
+
+ vfloat16m4_t ay1 = __riscv_vle16_v_f16m4((const _Float16*)y + i + epr, epr);
+ ay1 = __riscv_vfmul_vf_f16m4(ay1, scale, epr);
+ __riscv_vse16_v_f16m4((_Float16*)y + i + epr, ay1, epr);
+ __asm__ __volatile__ ("" ::: "memory");
+ }
+
+ // leftovers
+ int vl;
+ for (int i = np; i < n; i += vl) {
+ vl = __riscv_vsetvl_e16m4(n - i);
+ vfloat16m4_t ay0 = __riscv_vle16_v_f16m4((const _Float16*)y + i, vl);
+ ay0 = __riscv_vfmul_vf_f16m4(ay0, scale, vl);
+ __riscv_vse16_v_f16m4((_Float16*)y + i, ay0, vl);
+ }
+ np = n;
+ #else
+ // fall to scalar path
+ const int np = 0;
+ #endif
+#elif defined(GGML_SIMD)
+ const int np = (n & ~(GGML_F16_STEP - 1));
+
+ GGML_F16_VEC vx = GGML_F16_VEC_SET1(v);
+
+ GGML_F16_VEC ay[GGML_F16_ARR];
+
+ for (int i = 0; i < np; i += GGML_F16_STEP) {
+ for (int j = 0; j < GGML_F16_ARR; j++) {
+ ay[j] = GGML_F16_VEC_LOAD(y + i + j*GGML_F16_EPR, j);
+ ay[j] = GGML_F16_VEC_MUL(ay[j], vx);
+
+ GGML_F16_VEC_STORE(y + i + j*GGML_F16_EPR, ay, j);
+ }
+ }
+#else
+ // scalar path
+ const int np = 0;
+#endif
+ // scalar and leftovers
+ for (int i = np; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(y[i])*v);
+ }
+}
+
+inline static void ggml_vec_norm_f32 (const int n, float * s, const float * x) { ggml_vec_dot_f32(n, s, 0, x, 0, x, 0, 1); *s = sqrtf(*s); }
+inline static void ggml_vec_sqr_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i]*x[i]; }
+inline static void ggml_vec_sqr_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ float v = GGML_CPU_FP16_TO_FP32(x[i]);
+ y[i] = GGML_CPU_FP32_TO_FP16(v*v);
+ }
+}
+inline static void ggml_vec_sqrt_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sqrtf(x[i]); }
+inline static void ggml_vec_sqrt_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16(sqrtf(GGML_CPU_FP16_TO_FP32(x[i])));
+ }
+}
+inline static void ggml_vec_log_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = logf(x[i]); }
+inline static void ggml_vec_log_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16(logf(GGML_CPU_FP16_TO_FP32(x[i])));
+ }
+}
+inline static void ggml_vec_sin_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = sinf(x[i]); }
+inline static void ggml_vec_sin_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16(sinf(GGML_CPU_FP16_TO_FP32(x[i])));
+ }
+}
+inline static void ggml_vec_cos_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = cosf(x[i]); }
+inline static void ggml_vec_cos_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16(cosf(GGML_CPU_FP16_TO_FP32(x[i])));
+ }
+}
+inline static void ggml_vec_abs_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fabsf(x[i]); }
+inline static void ggml_vec_abs_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16(fabsf(GGML_CPU_FP16_TO_FP32(x[i])));
+ }
+}
+inline static void ggml_vec_sgn_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : ((x[i] < 0.f) ? -1.f : 0.f); }
+inline static void ggml_vec_sgn_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ float v = GGML_CPU_FP16_TO_FP32(x[i]);
+ y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? 1.f : ((v < 0.f) ? -1.f : 0.f));
+ }
+}
+inline static void ggml_vec_step_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? 1.f : 0.f; }
+inline static void ggml_vec_step_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16((GGML_CPU_FP16_TO_FP32(x[i]) > 0.f) ? 1.f : 0.f);
+ }
+}
+inline static void ggml_vec_tanh_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = tanhf(x[i]); }
+inline static void ggml_vec_tanh_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16(tanhf(GGML_CPU_FP16_TO_FP32(x[i])));
+ }
+}
+inline static void ggml_vec_elu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : expm1f(x[i]); }
+inline static void ggml_vec_elu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ const float v = GGML_CPU_FP16_TO_FP32(x[i]);
+ y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v : expm1f(v));
+ }
+}
+inline static void ggml_vec_relu_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = (x[i] > 0.f) ? x[i] : 0.f; }
+inline static void ggml_vec_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ float v = GGML_CPU_FP16_TO_FP32(x[i]);
+ y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v : 0.f);
+ }
+}
+inline static void ggml_vec_leaky_relu_f32 (const int n, float * y, const float * x, const float ns) { for (int i = 0; i < n; ++i) y[i] = ((x[i] > 0.f) ? x[i] : 0.f) + ns * ((x[i] < 0.0f) ? x[i] : 0.f); }
+inline static void ggml_vec_leaky_relu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const float ns) {
+ for (int i = 0; i < n; ++i) {
+ float v = GGML_CPU_FP16_TO_FP32(x[i]);
+ y[i] = GGML_CPU_FP32_TO_FP16(((v > 0.f) ? v : 0.f) + ns * ((v < 0.0f) ? v : 0.f));
+ }
+}
+inline static void ggml_vec_sigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = 1.f / (1.f + expf(-x[i])); }
+inline static void ggml_vec_sigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16(1.f / (1.f + expf(-GGML_CPU_FP16_TO_FP32(x[i]))));
+ }
+}
+// TODO: optimize performance
+inline static void ggml_vec_hardswish_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = x[i] * fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
+inline static void ggml_vec_hardswish_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ float v = GGML_CPU_FP16_TO_FP32(x[i]);
+ y[i] = GGML_CPU_FP32_TO_FP16(v * fminf(1.0f, fmaxf(0.0f, (v + 3.0f) / 6.0f)));
+ }
+}
+inline static void ggml_vec_hardsigmoid_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = fminf(1.0f, fmaxf(0.0f, (x[i] + 3.0f) / 6.0f)); }
+inline static void ggml_vec_hardsigmoid_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16(fminf(1.0f, fmaxf(0.0f, (GGML_CPU_FP16_TO_FP32(x[i]) + 3.0f) / 6.0f)));
+ }
+}
+inline static void ggml_vec_exp_f32 (const int n, float * y, const float * x) { for (int i = 0; i < n; ++i) y[i] = expf(x[i]); }
+inline static void ggml_vec_exp_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_CPU_FP32_TO_FP16(expf(GGML_CPU_FP16_TO_FP32(x[i])));
+ }
+}
+
+static const float GELU_COEF_A = 0.044715f;
+static const float GELU_QUICK_COEF = -1.702f;
+static const float SQRT_2_OVER_PI = 0.79788456080286535587989211986876f;
+static const float SQRT_2_INV = 0.70710678118654752440084436210484f;
+
+inline static float ggml_gelu_f32(float x) {
+ return 0.5f*x*(1.0f + tanhf(SQRT_2_OVER_PI*x*(1.0f + GELU_COEF_A*x*x)));
+}
+
+inline static void ggml_vec_gelu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ const uint16_t * i16 = (const uint16_t *) x;
+ for (int i = 0; i < n; ++i) {
+ y[i] = ggml_table_gelu_f16[i16[i]];
+ }
+}
+
+inline static void ggml_vec_gelu_bf16(const int n, ggml_bf16_t * y, const ggml_bf16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_FP32_TO_BF16(ggml_gelu_f32(GGML_BF16_TO_FP32(x[i])));
+ }
+}
+
+inline static void ggml_vec_gelu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ float xi = GGML_CPU_FP16_TO_FP32(x[i]);
+ float res = 0.5f*xi*(1.0f + erff(xi*SQRT_2_INV));
+ y[i] = GGML_CPU_FP32_TO_FP16(res);
+ }
+}
+
+inline static void ggml_vec_gelu_erf_bf16(const int n, ggml_bf16_t * y, const ggml_bf16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ float xi = GGML_BF16_TO_FP32(x[i]);
+ y[i] = GGML_FP32_TO_BF16(0.5f*xi*(1.0f + erff(xi*SQRT_2_INV)));
+ }
+}
+
+#ifdef GGML_GELU_FP16
+inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
+ uint16_t t;
+ for (int i = 0; i < n; ++i) {
+ if (x[i] <= -10.0f) {
+ y[i] = 0.0f;
+ } else if (x[i] >= 10.0f) {
+ y[i] = x[i];
+ } else {
+ ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
+ memcpy(&t, &fp16, sizeof(uint16_t));
+ y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[t]);
+ }
+ }
+}
+#else
+inline static void ggml_vec_gelu_f32(const int n, float * y, const float * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = ggml_gelu_f32(x[i]);
+ }
+}
+#endif
+
+inline static void ggml_vec_gelu_erf_f32(const int n, float * y, const float * x) {
+ for (int i = 0; i < n; ++i) {
+ float xi = x[i];
+ y[i] = 0.5f*xi*(1.0f + erff(xi*SQRT_2_INV));
+ }
+}
+
+inline static float ggml_gelu_quick_f32(float x) {
+ return x*(1.0f/(1.0f+expf(GELU_QUICK_COEF*x)));
+}
+
+inline static void ggml_vec_gelu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ const uint16_t * i16 = (const uint16_t *) x;
+ for (int i = 0; i < n; ++i) {
+ y[i] = ggml_table_gelu_quick_f16[i16[i]];
+ }
+}
+
+inline static void ggml_vec_gelu_quick_bf16(const int n, ggml_bf16_t * y, const ggml_bf16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_FP32_TO_BF16(ggml_gelu_quick_f32(GGML_BF16_TO_FP32(x[i])));
+ }
+}
+
+#ifdef GGML_GELU_QUICK_FP16
+inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) {
+ uint16_t t;
+ for (int i = 0; i < n; ++i) {
+ ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
+ memcpy(&t, &fp16, sizeof(uint16_t));
+ y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]);
+ }
+}
+#else
+inline static void ggml_vec_gelu_quick_f32(const int n, float * y, const float * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = ggml_gelu_quick_f32(x[i]);
+ }
+}
+#endif
+
+// Sigmoid Linear Unit (SiLU) function
+inline static float ggml_silu_f32(float x) {
+ return x/(1.0f + expf(-x));
+}
+inline static ggml_fp16_t ggml_silu_f16(ggml_fp16_t x) {
+ float v = GGML_CPU_FP16_TO_FP32(x);
+ return GGML_CPU_FP32_TO_FP16(v/(1.0f + expf(-v)));
+}
+
+#if __FINITE_MATH_ONLY__
+#error "some routines in ggml.c require non-finite math arithmetics -- pass -fno-finite-math-only to the compiler to fix"
+#error "ref: https://github.com/ggml-org/llama.cpp/pull/7154#issuecomment-2143844461"
+#endif
+
+/* Below function was borrowed from the GitHub repository:
+https://github.com/openvinotoolkit/openvino/blob/master/src/plugins/intel_cpu/src/nodes/kernels/scaled_attn/common.hpp */
+#if defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+ inline static svfloat32_t exp_ps_sve(svbool_t pg, svfloat32_t src) {
+ // Constants
+ const svfloat32_t log2_e = svdup_n_f32(1.4426950409f);
+ const svfloat32_t ln2 = svdup_n_f32(0.6931473921f);
+ const svfloat32_t half_ln2_sq = svdup_n_f32(0.2413862043f);
+ const svuint32_t not_mask17 = svdup_n_u32(~((1u << 17) - 1));
+ const svfloat32_t one = svdup_n_f32(1.0f);
+ const svfloat32_t inactive1 = svdup_n_f32(0.0f);
+ const svint32_t inactive2 = svdup_n_s32(0);
+
+ // Algorithm starts here
+ svfloat32_t t0 = svmul_f32_m(pg, src, log2_e); // y = x * log2(e)
+ svfloat32_t t1 = svrintm_f32_m(inactive1, pg, t0); // rount to int (float)
+ svint32_t t2 = svcvt_s32_f32_m(inactive2, pg, t1); // n
+
+ t1 = svsub_f32_m(pg, t0, t1); // a = y - floor(y)
+ t1 = svadd_f32_m(pg, t1, one); // b = a + 1
+
+ svuint32_t t3 = svlsr_n_u32_m(pg, svreinterpret_u32_f32(t1), 17); // v = b >> 17 (u32)
+ svfloat32_t t4 = svexpa_f32(t3); // c = fexpa(v)
+ t4 = svscale_f32_m(pg, t4, t2); // fexpa(v) * 2^(n)
+
+ // and_(t2.d, t1.d, not_mask17.d)
+ svfloat32_t t5 = svreinterpret_f32_u32(svand_u32_m(pg, svreinterpret_u32_f32(t1), not_mask17));
+ t5 = svsub_f32_m(pg, t1, t5); // z
+ t0 = svmla_f32_m(pg, ln2, t5, half_ln2_sq); // ln2 + half_ln2_sq * z
+ t0 = svmla_f32_m(pg, one, t5, t0); // 1 + (ln2 * z) + (half_ln2_sq * z * z)
+ t0 = svmul_f32_m(pg, t0, t4); // Final result
+
+ return t0;
+ }
+#endif
+
+#if defined(__ARM_FEATURE_SVE) && defined(__aarch64__)
+
+inline static svfloat32_t ggml_v_expf(svbool_t pg, svfloat32_t x) {
+ const svfloat32_t r = svdup_n_f32_x(pg, 0x1.8p23f);
+ const svfloat32_t z = svmla_n_f32_x(pg, r, x, 0x1.715476p+0f);
+ const svfloat32_t n = svsub_f32_x(pg, z, r);
+ const svfloat32_t b = svmls_n_f32_x(pg, svmls_n_f32_x(pg, x, n, 0x1.62e4p-1f), n, 0x1.7f7d1cp-20f);
+ const svuint32_t e = svlsl_n_u32_x(pg, svreinterpret_u32_f32(z), 23);
+ const svfloat32_t k = svreinterpret_f32_u32(svadd_u32_x(pg, e, svreinterpret_u32_f32(svdup_n_f32_x(pg, 1))));
+ const svbool_t c = svacgt_n_f32(pg, n, 126);
+ const svfloat32_t u = svmul_f32_x(pg, b, b);
+ const svfloat32_t j = svmla_f32_x(pg,
+ svmul_n_f32_x(pg, b, 0x1.ffffecp-1f),
+ svmla_f32_x(pg, svmla_f32_x(pg, svdup_n_f32_x(pg, 0x1.fffdb6p-2f), svdup_n_f32_x(pg, 0x1.555e66p-3f), b),
+ svmla_f32_x(pg, svdup_n_f32_x(pg, 0x1.573e2ep-5f), svdup_n_f32_x(pg, 0x1.0e4020p-7f), b), u), u);
+ const svuint32_t d = svdup_n_u32_z(svcmple_n_f32(pg, n, 0.0), 0x82000000);
+ const svfloat32_t s1 = svreinterpret_f32_u32(svadd_n_u32_x(pg, d, 0x7f000000));
+ const svfloat32_t s2 = svreinterpret_f32_u32(svsub_u32_x(pg, e, d));
+ return svsel_f32(svacgt_f32(pg, n, svdup_n_f32_x(pg, 192)), svmul_f32_x(pg, s1, s1),
+ svsel_f32(c, svmul_f32_x(pg, svmla_f32_x(pg, s2, s2, j), s1), svmla_f32_x(pg, k, k, j)));
+}
+
+// computes silu x/(1+exp(-x)) in single precision vector
+inline static svfloat32_t ggml_v_silu(svbool_t pg, svfloat32_t x) {
+ const svfloat32_t one = svdup_n_f32_x(pg, 1.0f);
+ const svfloat32_t zero = svdup_n_f32_x(pg, 0.0f);
+ const svfloat32_t neg_x = svsub_f32_x(pg, zero, x);
+ const svfloat32_t exp_neg_x = ggml_v_expf(pg, neg_x);
+ const svfloat32_t one_plus_exp_neg_x = svadd_f32_x(pg, one, exp_neg_x);
+ return svdiv_f32_x(pg, x, one_plus_exp_neg_x);
+}
+
+#elif defined(__ARM_NEON) && defined(__aarch64__)
+
+// adapted from arm limited optimized routine
+// the maximum error is 1.45358 plus 0.5 ulps
+// numbers above 88.38 will flush to infinity
+// numbers beneath -103.97 will flush to zero
+inline static float32x4_t ggml_v_expf(float32x4_t x) {
+ const float32x4_t r = vdupq_n_f32(0x1.8p23f);
+ const float32x4_t z = vfmaq_f32(r, x, vdupq_n_f32(0x1.715476p+0f));
+ const float32x4_t n = vsubq_f32(z, r);
+ const float32x4_t b = vfmsq_f32(vfmsq_f32(x, n, vdupq_n_f32(0x1.62e4p-1f)), n,
+ vdupq_n_f32(0x1.7f7d1cp-20f));
+ const uint32x4_t e = vshlq_n_u32(vreinterpretq_u32_f32(z), 23);
+ const float32x4_t k = vreinterpretq_f32_u32(vaddq_u32(e, vreinterpretq_u32_f32(vdupq_n_f32(1))));
+ const uint32x4_t c = vcagtq_f32(n, vdupq_n_f32(126));
+ const float32x4_t u = vmulq_f32(b, b);
+ const float32x4_t j = vfmaq_f32(
+ vmulq_f32(vdupq_n_f32(0x1.ffffecp-1f), b),
+ vfmaq_f32(vfmaq_f32(vdupq_n_f32(0x1.fffdb6p-2f), vdupq_n_f32(0x1.555e66p-3f), b),
+ vfmaq_f32(vdupq_n_f32(0x1.573e2ep-5f), vdupq_n_f32(0x1.0e4020p-7f), b), u), u);
+ if (!vpaddd_u64(vreinterpretq_u64_u32(c)))
+ return vfmaq_f32(k, j, k);
+ const uint32x4_t d = vandq_u32(vclezq_f32(n), vdupq_n_u32(0x82000000));
+ const float32x4_t s1 = vreinterpretq_f32_u32(vaddq_u32(d, vdupq_n_u32(0x7f000000)));
+ const float32x4_t s2 = vreinterpretq_f32_u32(vsubq_u32(e, d));
+ return vbslq_f32(vcagtq_f32(n, vdupq_n_f32(192)), vmulq_f32(s1, s1),
+ vbslq_f32(c, vmulq_f32(vfmaq_f32(s2, s2, j), s1), vfmaq_f32(k, k, j)));
+}
+
+// computes silu x/(1+exp(-x)) in single precision vector
+inline static float32x4_t ggml_v_silu(float32x4_t x) {
+ const float32x4_t one = vdupq_n_f32(1.0f);
+ const float32x4_t zero = vdupq_n_f32(0.0f);
+ const float32x4_t neg_x = vsubq_f32(zero, x);
+ const float32x4_t exp_neg_x = ggml_v_expf(neg_x);
+ const float32x4_t one_plus_exp_neg_x = vaddq_f32(one, exp_neg_x);
+ return vdivq_f32(x, one_plus_exp_neg_x);
+}
+
+#elif defined(__AVX512F__) && defined(__AVX512DQ__)
+
+// adapted from arm limited optimized routine
+// the maximum error is 1.45358 plus 0.5 ulps
+// numbers above 88.38 will flush to infinity
+// numbers beneath -103.97 will flush to zero
+inline static __m512 ggml_v_expf(__m512 x) {
+ const __m512 r = _mm512_set1_ps(0x1.8p23f);
+ const __m512 z = _mm512_fmadd_ps(x, _mm512_set1_ps(0x1.715476p+0f), r);
+ const __m512 n = _mm512_sub_ps(z, r);
+ const __m512 b =
+ _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.7f7d1cp-20f),
+ _mm512_fnmadd_ps(n, _mm512_set1_ps(0x1.62e4p-1f), x));
+ const __mmask16 d =
+ _mm512_cmp_ps_mask(_mm512_abs_ps(n), _mm512_set1_ps(192), _CMP_GT_OQ);
+ const __m512 u = _mm512_mul_ps(b, b);
+ const __m512 j = _mm512_fmadd_ps(
+ _mm512_fmadd_ps(_mm512_fmadd_ps(_mm512_set1_ps(0x1.0e4020p-7f), b,
+ _mm512_set1_ps(0x1.573e2ep-5f)),
+ u,
+ _mm512_fmadd_ps(_mm512_set1_ps(0x1.555e66p-3f), b,
+ _mm512_set1_ps(0x1.fffdb6p-2f))),
+ u,
+ _mm512_fmadd_ps(_mm512_set1_ps(0x1.ffffecp-1f), b, _mm512_set1_ps(1.0F)));
+ const __m512 res = _mm512_scalef_ps(j, n);
+ if (_mm512_kortestz(d, d))
+ return res;
+ const __m512 zero = _mm512_setzero_ps();
+ const __m512 alt = _mm512_mask_blend_ps(
+ _mm512_cmp_ps_mask(n, zero, _CMP_LE_OQ), _mm512_set1_ps(INFINITY), zero);
+ return _mm512_mask_blend_ps(d, res, alt);
+}
+
+// computes silu x/(1+exp(-x)) in single precision vector
+inline static __m512 ggml_v_silu(__m512 x) {
+ const __m512 one = _mm512_set1_ps(1);
+ const __m512 zero = _mm512_setzero_ps();
+ const __m512 neg_x = _mm512_sub_ps(zero, x);
+ const __m512 exp_neg_x = ggml_v_expf(neg_x);
+ const __m512 one_plus_exp_neg_x = _mm512_add_ps(one, exp_neg_x);
+ return _mm512_div_ps(x, one_plus_exp_neg_x);
+}
+
+#elif defined(__AVX2__) && defined(__FMA__)
+
+// adapted from arm limited optimized routine
+// the maximum error is 1.45358 plus 0.5 ulps
+// numbers above 88.38 will flush to infinity
+// numbers beneath -103.97 will flush to zero
+inline static __m256 ggml_v_expf(__m256 x) {
+ const __m256 r = _mm256_set1_ps(0x1.8p23f);
+ const __m256 z = _mm256_fmadd_ps(x, _mm256_set1_ps(0x1.715476p+0f), r);
+ const __m256 n = _mm256_sub_ps(z, r);
+ const __m256 b = _mm256_fnmadd_ps(n, _mm256_set1_ps(0x1.7f7d1cp-20f),
+ _mm256_fnmadd_ps(n, _mm256_set1_ps(0x1.62e4p-1f), x));
+ const __m256i e = _mm256_slli_epi32(_mm256_castps_si256(z), 23);
+ const __m256 k = _mm256_castsi256_ps(
+ _mm256_add_epi32(e, _mm256_castps_si256(_mm256_set1_ps(1))));
+ const __m256i c = _mm256_castps_si256(
+ _mm256_cmp_ps(_mm256_andnot_ps(_mm256_set1_ps(-0.f), n),
+ _mm256_set1_ps(126), _CMP_GT_OQ));
+ const __m256 u = _mm256_mul_ps(b, b);
+ const __m256 j = _mm256_fmadd_ps(_mm256_fmadd_ps(_mm256_fmadd_ps(_mm256_set1_ps(0x1.0e4020p-7f), b,
+ _mm256_set1_ps(0x1.573e2ep-5f)), u,
+ _mm256_fmadd_ps(_mm256_set1_ps(0x1.555e66p-3f), b,
+ _mm256_set1_ps(0x1.fffdb6p-2f))),
+ u, _mm256_mul_ps(_mm256_set1_ps(0x1.ffffecp-1f), b));
+ if (!_mm256_movemask_ps(_mm256_castsi256_ps(c)))
+ return _mm256_fmadd_ps(j, k, k);
+ const __m256i g = _mm256_and_si256(
+ _mm256_castps_si256(_mm256_cmp_ps(n, _mm256_setzero_ps(), _CMP_LE_OQ)),
+ _mm256_set1_epi32(0x82000000u));
+ const __m256 s1 =
+ _mm256_castsi256_ps(_mm256_add_epi32(g, _mm256_set1_epi32(0x7f000000u)));
+ const __m256 s2 = _mm256_castsi256_ps(_mm256_sub_epi32(e, g));
+ const __m256i d = _mm256_castps_si256(
+ _mm256_cmp_ps(_mm256_andnot_ps(_mm256_set1_ps(-0.f), n),
+ _mm256_set1_ps(192), _CMP_GT_OQ));
+ return _mm256_or_ps(
+ _mm256_and_ps(_mm256_castsi256_ps(d), _mm256_mul_ps(s1, s1)),
+ _mm256_andnot_ps(
+ _mm256_castsi256_ps(d),
+ _mm256_or_ps(
+ _mm256_and_ps(_mm256_castsi256_ps(c),
+ _mm256_mul_ps(_mm256_fmadd_ps(s2, j, s2), s1)),
+ _mm256_andnot_ps(_mm256_castsi256_ps(c), _mm256_fmadd_ps(k, j, k)))));
+}
+
+// computes silu x/(1+exp(-x)) in single precision vector
+inline static __m256 ggml_v_silu(__m256 x) {
+ const __m256 one = _mm256_set1_ps(1);
+ const __m256 zero = _mm256_setzero_ps();
+ const __m256 neg_x = _mm256_sub_ps(zero, x);
+ const __m256 exp_neg_x = ggml_v_expf(neg_x);
+ const __m256 one_plus_exp_neg_x = _mm256_add_ps(one, exp_neg_x);
+ return _mm256_div_ps(x, one_plus_exp_neg_x);
+}
+
+#elif defined(__SSE2__) // __AVX2__ / __ARM_NEON
+
+#if defined(__FMA__)
+#define MADD128(x, y, z) _mm_fmadd_ps(x, y, z)
+#define NMADD128(x, y, z) _mm_fnmadd_ps(x, y, z)
+#else
+#define MADD128(x, y, z) _mm_add_ps(_mm_mul_ps(x, y), z)
+#define NMADD128(x, y, z) _mm_sub_ps(z, _mm_mul_ps(x, y))
+#endif
+
+// adapted from arm limited optimized routine
+// the maximum error is 1.45358 plus 0.5 ulps
+// numbers above 88.38 will flush to infinity
+// numbers beneath -103.97 will flush to zero
+inline static __m128 ggml_v_expf(__m128 x) {
+ const __m128 r = _mm_set1_ps(0x1.8p23f);
+ const __m128 z = MADD128(x, _mm_set1_ps(0x1.715476p+0f), r);
+ const __m128 n = _mm_sub_ps(z, r);
+ const __m128 b =
+ NMADD128(n, _mm_set1_ps(0x1.7f7d1cp-20f), NMADD128(n, _mm_set1_ps(0x1.62e4p-1f), x));
+ const __m128i e = _mm_slli_epi32(_mm_castps_si128(z), 23);
+ const __m128 k = _mm_castsi128_ps(_mm_add_epi32(e, _mm_castps_si128(_mm_set1_ps(1))));
+ const __m128i c =
+ _mm_castps_si128(_mm_cmpgt_ps(_mm_andnot_ps(_mm_set1_ps(-0.f), n), _mm_set1_ps(126)));
+ const __m128 u = _mm_mul_ps(b, b);
+ const __m128 j =
+ MADD128(MADD128(MADD128(_mm_set1_ps(0x1.0e4020p-7f), b, _mm_set1_ps(0x1.573e2ep-5f)), u,
+ MADD128(_mm_set1_ps(0x1.555e66p-3f), b, _mm_set1_ps(0x1.fffdb6p-2f))),
+ u, _mm_mul_ps(_mm_set1_ps(0x1.ffffecp-1f), b));
+ if (!_mm_movemask_epi8(c))
+ return MADD128(j, k, k);
+ const __m128i g = _mm_and_si128(_mm_castps_si128(_mm_cmple_ps(n, _mm_setzero_ps())),
+ _mm_set1_epi32(0x82000000u));
+ const __m128 s1 = _mm_castsi128_ps(_mm_add_epi32(g, _mm_set1_epi32(0x7f000000u)));
+ const __m128 s2 = _mm_castsi128_ps(_mm_sub_epi32(e, g));
+ const __m128i d =
+ _mm_castps_si128(_mm_cmpgt_ps(_mm_andnot_ps(_mm_set1_ps(-0.f), n), _mm_set1_ps(192)));
+ return _mm_or_ps(
+ _mm_and_ps(_mm_castsi128_ps(d), _mm_mul_ps(s1, s1)),
+ _mm_andnot_ps(_mm_castsi128_ps(d),
+ _mm_or_ps(_mm_and_ps(_mm_castsi128_ps(c), _mm_mul_ps(MADD128(s2, j, s2), s1)),
+ _mm_andnot_ps(_mm_castsi128_ps(c), MADD128(k, j, k)))));
+}
+
+// computes silu x/(1+exp(-x)) in single precision vector
+inline static __m128 ggml_v_silu(__m128 x) {
+ const __m128 one = _mm_set1_ps(1);
+ const __m128 zero = _mm_setzero_ps();
+ const __m128 neg_x = _mm_sub_ps(zero, x);
+ const __m128 exp_neg_x = ggml_v_expf(neg_x);
+ const __m128 one_plus_exp_neg_x = _mm_add_ps(one, exp_neg_x);
+ return _mm_div_ps(x, one_plus_exp_neg_x);
+}
+
+#elif defined(__riscv_v_intrinsic)
+
+// adapted from arm limited optimized routine
+// the maximum error is 1.45358 plus 0.5 ulps
+// numbers above 88.38 will flush to infinity
+// numbers beneath -103.97 will flush to zero
+inline static vfloat32m2_t ggml_v_expf_m2(vfloat32m2_t x, int vl) {
+ const vfloat32m2_t r = __riscv_vfmv_v_f_f32m2(0x1.8p23f, vl);
+#ifdef __riscv_xtheadvector
+ // workaround for compiler bug (gcc 14.3.0: Error: unrecognized opcode `th.vmv1r.v v2,v4')
+ vfloat32m2_t z = __riscv_vfadd_vf_f32m2(r, 0.0f, vl);
+ z = __riscv_vfmacc_vf_f32m2(z, 0x1.715476p+0f, x, vl);
+#else
+ const vfloat32m2_t z = __riscv_vfmacc_vf_f32m2(r, 0x1.715476p+0f, x, vl);
+#endif
+ const vfloat32m2_t n = __riscv_vfsub_vv_f32m2(z, r, vl);
+ const vfloat32m2_t b = __riscv_vfnmsac_vf_f32m2(__riscv_vfnmsac_vf_f32m2(x, 0x1.62e4p-1f, n, vl),
+ 0x1.7f7d1cp-20f, n, vl);
+ const vuint32m2_t e = __riscv_vsll_vx_u32m2(__riscv_vreinterpret_v_f32m2_u32m2(z), 23, vl);
+ const vfloat32m2_t k = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vadd_vx_u32m2(e, 0x3f800000, vl)); // 1.0f
+ const vbool16_t c = __riscv_vmfgt_vf_f32m2_b16(__riscv_vfabs_v_f32m2(n, vl), 126.0f, vl);
+ const vfloat32m2_t u = __riscv_vfmul_vv_f32m2(b, b, vl);
+ const vfloat32m2_t j = __riscv_vfmacc_vv_f32m2(
+ __riscv_vfmul_vf_f32m2(b, 0x1.ffffecp-1f, vl),
+ __riscv_vfmacc_vv_f32m2(
+ __riscv_vfmacc_vf_f32m2(__riscv_vfmv_v_f_f32m2(0x1.fffdb6p-2f, vl), 0x1.555e66p-3f, b, vl),
+ __riscv_vfmacc_vf_f32m2(__riscv_vfmv_v_f_f32m2(0x1.573e2ep-5f, vl), 0x1.0e4020p-7f, b, vl),
+ u, vl), u, vl);
+ if (!__riscv_vcpop_m_b16(c, vl))
+ return __riscv_vfmacc_vv_f32m2(k, j, k, vl);
+ const vbool16_t dm = __riscv_vmfle_vf_f32m2_b16(n, 0.0f, vl);
+ const vuint32m2_t d = __riscv_vmerge_vxm_u32m2(__riscv_vmv_v_x_u32m2(0, vl), 0x82000000, dm, vl);
+ const vfloat32m2_t s1 = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vadd_vx_u32m2(d, 0x7f000000, vl));
+ const vfloat32m2_t s2 = __riscv_vreinterpret_v_u32m2_f32m2(__riscv_vsub_vv_u32m2(e, d, vl));
+ const vfloat32m2_t r1 = __riscv_vmerge_vvm_f32m2(
+ __riscv_vfmacc_vv_f32m2(k, k, j, vl),
+ __riscv_vfmul_vv_f32m2(__riscv_vfmacc_vv_f32m2(s2, s2, j, vl), s1, vl),
+ c, vl);
+ return __riscv_vmerge_vvm_f32m2(
+ r1, __riscv_vfmul_vv_f32m2(s1, s1, vl),
+ __riscv_vmfgt_vf_f32m2_b16(__riscv_vfabs_v_f32m2(n, vl), 192.0f, vl),
+ vl);
+}
+
+// computes silu x/(1+exp(-x)) in single precision vector
+inline static vfloat32m2_t ggml_v_silu_m2(vfloat32m2_t x, int vl) {
+ const vfloat32m2_t neg_x = __riscv_vfneg_v_f32m2(x, vl);
+ const vfloat32m2_t exp_neg_x = ggml_v_expf_m2(neg_x, vl);
+ const vfloat32m2_t one_plus_exp_neg_x = __riscv_vfadd_vf_f32m2(exp_neg_x, 1.0f, vl);
+ return __riscv_vfdiv_vv_f32m2(x, one_plus_exp_neg_x, vl);
+}
+
+#endif // __ARM_NEON / __AVX2__ / __SSE2__ / __riscv_v_intrinsic
+
+inline static void ggml_vec_silu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = ggml_silu_f16(x[i]);
+ }
+}
+
+inline static void ggml_vec_silu_bf16(const int n, ggml_bf16_t * y, const ggml_bf16_t * x) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_FP32_TO_BF16(ggml_silu_f32(GGML_BF16_TO_FP32(x[i])));
+ }
+}
+
+inline static float ggml_silu_backward_f32(float x, float dy) {
+ const float s = 1.0f/(1.0f + expf(-x));
+ return dy*s*(1.0f + x*(1.0f - s));
+}
+
+inline static ggml_fp16_t ggml_silu_backward_f16(ggml_fp16_t x, ggml_fp16_t dy) {
+ const float v = GGML_CPU_FP16_TO_FP32(x);
+ const float s = 1.0f/(1.0f + expf(-v));
+ return GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(dy)*s*(1.0f + v*(1.0f - s)));
+}
+
+inline static void ggml_vec_silu_backward_f32(const int n, float * dx, const float * x, const float * dy) {
+ for (int i = 0; i < n; ++i) {
+ dx[i] = ggml_silu_backward_f32(x[i], dy[i]);
+ }
+}
+
+inline static void ggml_vec_silu_backward_f16(const int n, ggml_fp16_t * dx, const ggml_fp16_t * x, const ggml_fp16_t * dy) {
+ for (int i = 0; i < n; ++i) {
+ dx[i] = ggml_silu_backward_f16(x[i], dy[i]);
+ }
+}
+
+inline static void ggml_vec_reglu_f32 (const int n, float * y, const float * x, const float * g) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = (x[i] > 0.f) ? x[i] * g[i] : 0.f;
+ }
+}
+
+inline static void ggml_vec_reglu_f16 (const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+ for (int i = 0; i < n; ++i) {
+ float v = GGML_CPU_FP16_TO_FP32(x[i]);
+ y[i] = GGML_CPU_FP32_TO_FP16((v > 0.f) ? v * GGML_CPU_FP16_TO_FP32(g[i]) : 0.f);
+ }
+}
+
+inline static void ggml_vec_reglu_bf16(const int n, ggml_bf16_t * y, const ggml_bf16_t * x, const ggml_bf16_t * g) {
+ for (int i = 0; i < n; ++i) {
+ float v = GGML_BF16_TO_FP32(x[i]);
+ y[i] = GGML_FP32_TO_BF16((v > 0.f) ? v * GGML_BF16_TO_FP32(g[i]) : 0.f);
+ }
+}
+
+#ifdef GGML_GELU_FP16
+inline static void ggml_vec_geglu_f32(const int n, float * y, const float * x, const float * g) {
+ uint16_t t;
+ for (int i = 0; i < n; ++i) {
+ if (x[i] <= -10.0f) {
+ y[i] = 0.0f;
+ } else if (x[i] >= 10.0f) {
+ y[i] = x[i] * g[i];
+ } else {
+ ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
+ memcpy(&t, &fp16, sizeof(uint16_t));
+ y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[t]) * g[i];
+ }
+ }
+}
+#else
+inline static void ggml_vec_geglu_f32(const int n, float * y, const float * x, const float * g) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = ggml_gelu_f32(x[i]) * g[i];
+ }
+}
+#endif
+
+inline static void ggml_vec_geglu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+ const uint16_t * i16 = (const uint16_t *) x;
+ for (int i = 0; i < n; ++i) {
+ float v = GGML_CPU_FP16_TO_FP32(g[i]);
+ y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(ggml_table_gelu_f16[i16[i]]) * v);
+ }
+}
+
+inline static void ggml_vec_geglu_bf16(const int n, ggml_bf16_t * y, const ggml_bf16_t * x, const ggml_bf16_t * g) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_FP32_TO_BF16(ggml_gelu_f32(GGML_BF16_TO_FP32(x[i])) * GGML_BF16_TO_FP32(g[i]));
+ }
+}
+
+void ggml_vec_swiglu_f32(const int n, float * y, const float * x, const float * g);
+
+inline static void ggml_vec_swiglu_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+ for (int i = 0; i < n; ++i) {
+ float xi = GGML_CPU_FP16_TO_FP32(x[i]);
+ float gi = GGML_CPU_FP16_TO_FP32(g[i]);
+ y[i] = GGML_CPU_FP32_TO_FP16((xi/(1.0f + expf(-xi))) * gi);
+ }
+}
+
+inline static void ggml_vec_swiglu_bf16(const int n, ggml_bf16_t * y, const ggml_bf16_t * x, const ggml_bf16_t * g) {
+ for (int i = 0; i < n; ++i) {
+ float xi = GGML_BF16_TO_FP32(x[i]);
+ float gi = GGML_BF16_TO_FP32(g[i]);
+ y[i] = GGML_FP32_TO_BF16((xi/(1.0f + expf(-xi))) * gi);
+ }
+}
+
+inline static void ggml_vec_geglu_erf_f32(const int n, float * y, const float * x, const float * g) {
+ for (int i = 0; i < n; ++i) {
+ float xi = x[i];
+ y[i] = 0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * g[i];
+ }
+}
+
+inline static void ggml_vec_geglu_erf_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+ for (int i = 0; i < n; ++i) {
+ float xi = GGML_CPU_FP16_TO_FP32(x[i]);
+ float gi = GGML_CPU_FP16_TO_FP32(g[i]);
+ y[i] = GGML_CPU_FP32_TO_FP16(0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * gi);
+ }
+}
+
+inline static void ggml_vec_geglu_erf_bf16(const int n, ggml_bf16_t * y, const ggml_bf16_t * x, const ggml_bf16_t * g) {
+ for (int i = 0; i < n; ++i) {
+ float xi = GGML_BF16_TO_FP32(x[i]);
+ float gi = GGML_BF16_TO_FP32(g[i]);
+ y[i] = GGML_FP32_TO_BF16(0.5f * xi * (1.0f + erff(xi*SQRT_2_INV)) * gi);
+ }
+}
+
+#ifdef GGML_GELU_QUICK_FP16
+inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
+ uint16_t t;
+ for (int i = 0; i < n; ++i) {
+ ggml_fp16_t fp16 = GGML_CPU_FP32_TO_FP16(x[i]);
+ memcpy(&t, &fp16, sizeof(uint16_t));
+ y[i] = GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[t]) * g[i];
+ }
+}
+#else
+inline static void ggml_vec_geglu_quick_f32(const int n, float * y, const float * x, const float * g) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = ggml_gelu_quick_f32(x[i]) * g[i];
+ }
+}
+#endif
+
+inline static void ggml_vec_geglu_quick_f16(const int n, ggml_fp16_t * y, const ggml_fp16_t * x, const ggml_fp16_t * g) {
+ const uint16_t * i16 = (const uint16_t *) x;
+ for (int i = 0; i < n; ++i) {
+ float v = GGML_CPU_FP16_TO_FP32(g[i]);
+ y[i] = GGML_CPU_FP32_TO_FP16(GGML_CPU_FP16_TO_FP32(ggml_table_gelu_quick_f16[i16[i]]) * v);
+ }
+}
+
+inline static void ggml_vec_geglu_quick_bf16(const int n, ggml_bf16_t * y, const ggml_bf16_t * x, const ggml_bf16_t * g) {
+ for (int i = 0; i < n; ++i) {
+ y[i] = GGML_FP32_TO_BF16(ggml_gelu_quick_f32(GGML_BF16_TO_FP32(x[i])) * GGML_BF16_TO_FP32(g[i]));
+ }
+}
+
+inline static void ggml_vec_sum_f32(const int n, float * s, const float * x) {
+#ifndef GGML_USE_ACCELERATE
+ ggml_float sum = 0.0;
+ for (int i = 0; i < n; ++i) {
+ sum += (ggml_float)x[i];
+ }
+ *s = (float)sum;
+#else
+ vDSP_sve(x, 1, s, n);
+#endif
+}
+
+inline static void ggml_vec_cumsum_f32(const int n, float * y, const float * x) {
+ for (int i = 0; i < n; ++i) {
+ if (i == 0) {
+ y[i] = x[i];
+ } else {
+ y[i] = y[i - 1] + x[i];
+ }
+ }
+}
+
+inline static void ggml_vec_sum_f32_ggf(const int n, ggml_float * s, const float * x) {
+ ggml_float sum = 0.0;
+ for (int i = 0; i < n; ++i) {
+ sum += (ggml_float)x[i];
+ }
+ *s = sum;
+}
+
+inline static void ggml_vec_sum_f16_ggf(const int n, float * s, const ggml_fp16_t * x) {
+ float sum = 0.0f;
+ for (int i = 0; i < n; ++i) {
+ sum += GGML_CPU_FP16_TO_FP32(x[i]);
+ }
+ *s = sum;
+}
+
+inline static void ggml_vec_sum_bf16_ggf(const int n, float * s, const ggml_bf16_t * x) {
+ float sum = 0.0f;
+ for (int i = 0; i < n; ++i) {
+ sum += GGML_BF16_TO_FP32(x[i]);
+ }
+ *s = sum;
+}
+
+inline static void ggml_vec_max_f32(const int n, float * s, const float * x) {
+#ifndef GGML_USE_ACCELERATE
+ float max = -INFINITY;
+ for (int i = 0; i < n; ++i) {
+ max = MAX(max, x[i]);
+ }
+ *s = max;
+#else
+ vDSP_maxv(x, 1, s, n);
+#endif
+}
+
+inline static void ggml_vec_norm_inv_f32(const int n, float * s, const float * x) {
+ ggml_vec_norm_f32(n, s, x);
+ *s = 1.f/(*s);
+}
+
+inline static void ggml_vec_argmax_f32(const int n, int * s, const float * x) {
+ float max = -INFINITY;
+ int idx = 0;
+ for (int i = 0; i < n; ++i) {
+ max = MAX(max, x[i]);
+ if (max == x[i]) { idx = i; }
+ }
+ *s = idx;
+}
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/python/freetoken/kernel/csrc/ggml/include/ggml-cpu.h b/python/freetoken/kernel/csrc/ggml/include/ggml-cpu.h
new file mode 100644
index 00000000..dc6453c6
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/include/ggml-cpu.h
@@ -0,0 +1,152 @@
+#pragma once
+
+#include "ggml.h"
+#include "ggml-backend.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+ // the compute plan that needs to be prepared for ggml_graph_compute()
+ // since https://github.com/ggml-org/ggml/issues/287
+ struct ggml_cplan {
+ size_t work_size; // size of work buffer, calculated by `ggml_graph_plan()`
+ uint8_t * work_data; // work buffer, to be allocated by caller before calling to `ggml_graph_compute()`
+
+ int n_threads;
+ struct ggml_threadpool * threadpool;
+
+ // abort ggml_graph_compute when true
+ ggml_abort_callback abort_callback;
+ void * abort_callback_data;
+
+ // use only reference implementations
+ bool use_ref;
+ };
+
+ // numa strategies
+ enum ggml_numa_strategy {
+ GGML_NUMA_STRATEGY_DISABLED = 0,
+ GGML_NUMA_STRATEGY_DISTRIBUTE = 1,
+ GGML_NUMA_STRATEGY_ISOLATE = 2,
+ GGML_NUMA_STRATEGY_NUMACTL = 3,
+ GGML_NUMA_STRATEGY_MIRROR = 4,
+ GGML_NUMA_STRATEGY_COUNT
+ };
+
+ GGML_BACKEND_API void ggml_numa_init(enum ggml_numa_strategy numa); // call once for better performance on NUMA systems
+ GGML_BACKEND_API bool ggml_is_numa(void); // true if init detected that system has >1 NUMA node
+
+ GGML_BACKEND_API struct ggml_tensor * ggml_new_i32(struct ggml_context * ctx, int32_t value);
+ GGML_BACKEND_API struct ggml_tensor * ggml_new_f32(struct ggml_context * ctx, float value);
+
+ GGML_BACKEND_API struct ggml_tensor * ggml_set_i32 (struct ggml_tensor * tensor, int32_t value);
+ GGML_BACKEND_API struct ggml_tensor * ggml_set_f32 (struct ggml_tensor * tensor, float value);
+
+ GGML_BACKEND_API int32_t ggml_get_i32_1d(const struct ggml_tensor * tensor, int i);
+ GGML_BACKEND_API void ggml_set_i32_1d(const struct ggml_tensor * tensor, int i, int32_t value);
+
+ GGML_BACKEND_API int32_t ggml_get_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+ GGML_BACKEND_API void ggml_set_i32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, int32_t value);
+
+ GGML_BACKEND_API float ggml_get_f32_1d(const struct ggml_tensor * tensor, int i);
+ GGML_BACKEND_API void ggml_set_f32_1d(const struct ggml_tensor * tensor, int i, float value);
+
+ GGML_BACKEND_API float ggml_get_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3);
+ GGML_BACKEND_API void ggml_set_f32_nd(const struct ggml_tensor * tensor, int i0, int i1, int i2, int i3, float value);
+
+ GGML_BACKEND_API struct ggml_threadpool * ggml_threadpool_new (struct ggml_threadpool_params * params);
+ GGML_BACKEND_API void ggml_threadpool_free (struct ggml_threadpool * threadpool);
+ GGML_BACKEND_API int ggml_threadpool_get_n_threads (struct ggml_threadpool * threadpool);
+ GGML_BACKEND_API void ggml_threadpool_pause (struct ggml_threadpool * threadpool);
+ GGML_BACKEND_API void ggml_threadpool_resume (struct ggml_threadpool * threadpool);
+
+ // ggml_graph_plan() has to be called before ggml_graph_compute()
+ // when plan.work_size > 0, caller must allocate memory for plan.work_data
+ GGML_BACKEND_API struct ggml_cplan ggml_graph_plan(
+ const struct ggml_cgraph * cgraph,
+ int n_threads, /* = GGML_DEFAULT_N_THREADS */
+ struct ggml_threadpool * threadpool /* = NULL */ );
+ GGML_BACKEND_API enum ggml_status ggml_graph_compute(struct ggml_cgraph * cgraph, struct ggml_cplan * cplan);
+
+ // same as ggml_graph_compute() but the work data is allocated as a part of the context
+ // note: the drawback of this API is that you must have ensured that the context has enough memory for the work data
+ GGML_BACKEND_API enum ggml_status ggml_graph_compute_with_ctx(struct ggml_context * ctx, struct ggml_cgraph * cgraph, int n_threads);
+
+ //
+ // system info
+ //
+
+ // x86
+ GGML_BACKEND_API int ggml_cpu_has_sse3 (void);
+ GGML_BACKEND_API int ggml_cpu_has_ssse3 (void);
+ GGML_BACKEND_API int ggml_cpu_has_avx (void);
+ GGML_BACKEND_API int ggml_cpu_has_avx_vnni (void);
+ GGML_BACKEND_API int ggml_cpu_has_avx2 (void);
+ GGML_BACKEND_API int ggml_cpu_has_bmi2 (void);
+ GGML_BACKEND_API int ggml_cpu_has_f16c (void);
+ GGML_BACKEND_API int ggml_cpu_has_fma (void);
+ GGML_BACKEND_API int ggml_cpu_has_avx512 (void);
+ GGML_BACKEND_API int ggml_cpu_has_avx512_vbmi(void);
+ GGML_BACKEND_API int ggml_cpu_has_avx512_vnni(void);
+ GGML_BACKEND_API int ggml_cpu_has_avx512_bf16(void);
+ GGML_BACKEND_API int ggml_cpu_has_amx_int8 (void);
+ // ARM
+ GGML_BACKEND_API int ggml_cpu_has_neon (void);
+ GGML_BACKEND_API int ggml_cpu_has_arm_fma (void);
+ GGML_BACKEND_API int ggml_cpu_has_fp16_va (void);
+ GGML_BACKEND_API int ggml_cpu_has_dotprod (void);
+ GGML_BACKEND_API int ggml_cpu_has_matmul_int8(void);
+ GGML_BACKEND_API int ggml_cpu_has_sve (void);
+ GGML_BACKEND_API int ggml_cpu_get_sve_cnt (void); // sve vector length in bytes
+ GGML_BACKEND_API int ggml_cpu_has_sme (void);
+ GGML_BACKEND_API int ggml_cpu_has_sme2 (void);
+ // other
+ GGML_BACKEND_API int ggml_cpu_has_riscv_v (void);
+ GGML_BACKEND_API int ggml_cpu_get_rvv_vlen (void); // risc-v vector length in bytes
+ GGML_BACKEND_API int ggml_cpu_has_vsx (void);
+ GGML_BACKEND_API int ggml_cpu_has_vxe (void);
+ GGML_BACKEND_API int ggml_cpu_has_wasm_simd (void);
+ GGML_BACKEND_API int ggml_cpu_has_llamafile (void);
+
+ // Internal types and functions exposed for tests and benchmarks
+
+ typedef void (*ggml_vec_dot_t) (int n, float * GGML_RESTRICT s, size_t bs, const void * GGML_RESTRICT x, size_t bx,
+ const void * GGML_RESTRICT y, size_t by, int nrc);
+
+ struct ggml_type_traits_cpu {
+ ggml_from_float_t from_float;
+ ggml_vec_dot_t vec_dot;
+ enum ggml_type vec_dot_type;
+ int64_t nrows; // number of rows to process simultaneously
+ };
+
+ GGML_BACKEND_API const struct ggml_type_traits_cpu * ggml_get_type_traits_cpu(enum ggml_type type);
+
+ GGML_BACKEND_API void ggml_cpu_init(void);
+
+ //
+ // CPU backend
+ //
+
+ GGML_BACKEND_API ggml_backend_t ggml_backend_cpu_init(void);
+
+ GGML_BACKEND_API bool ggml_backend_is_cpu (ggml_backend_t backend);
+ GGML_BACKEND_API void ggml_backend_cpu_set_n_threads (ggml_backend_t backend_cpu, int n_threads);
+ GGML_BACKEND_API void ggml_backend_cpu_set_threadpool (ggml_backend_t backend_cpu, ggml_threadpool_t threadpool);
+ GGML_BACKEND_API void ggml_backend_cpu_set_abort_callback(ggml_backend_t backend_cpu, ggml_abort_callback abort_callback, void * abort_callback_data);
+
+ GGML_BACKEND_API void ggml_backend_cpu_set_use_ref(ggml_backend_t backend_cpu, bool use_ref);
+
+ GGML_BACKEND_API ggml_backend_reg_t ggml_backend_cpu_reg(void);
+
+ GGML_BACKEND_API void ggml_cpu_fp32_to_fp32(const float *, float *, int64_t);
+ GGML_BACKEND_API void ggml_cpu_fp32_to_i32 (const float *, int32_t *, int64_t);
+ GGML_BACKEND_API void ggml_cpu_fp32_to_fp16(const float *, ggml_fp16_t *, int64_t);
+ GGML_BACKEND_API void ggml_cpu_fp16_to_fp32(const ggml_fp16_t *, float *, int64_t);
+ GGML_BACKEND_API void ggml_cpu_fp32_to_bf16(const float *, ggml_bf16_t *, int64_t);
+ GGML_BACKEND_API void ggml_cpu_bf16_to_fp32(const ggml_bf16_t *, float *, int64_t);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/python/freetoken/kernel/csrc/ggml/src/ggml-backend-impl.h b/python/freetoken/kernel/csrc/ggml/src/ggml-backend-impl.h
new file mode 100644
index 00000000..dc7367de
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/src/ggml-backend-impl.h
@@ -0,0 +1,292 @@
+#pragma once
+
+// ggml-backend internal header
+
+#include "ggml-backend.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+ #define GGML_BACKEND_API_VERSION 3
+
+ //
+ // Backend buffer type
+ //
+
+ struct ggml_backend_buffer_type_i {
+ const char * (*get_name) (ggml_backend_buffer_type_t buft);
+ // allocate a buffer of this type
+ ggml_backend_buffer_t (*alloc_buffer) (ggml_backend_buffer_type_t buft, size_t size);
+ // (optional) allocate tensors from a list into a buffer of this type (defaults to alloc_buffer + linear allocator)
+ ggml_backend_buffer_t (*alloc_buffer_n) (ggml_backend_buffer_type_t buft, struct ggml_tensor ** tensors, int n_tensors);
+ // tensor alignment
+ size_t (*get_alignment) (ggml_backend_buffer_type_t buft);
+ // (optional) max buffer size that can be allocated (defaults to SIZE_MAX)
+ size_t (*get_max_size) (ggml_backend_buffer_type_t buft);
+ // (optional) data size needed to allocate the tensor, including padding (defaults to ggml_nbytes)
+ size_t (*get_alloc_size) (ggml_backend_buffer_type_t buft, const struct ggml_tensor * tensor);
+ // (optional) total data size needed to allocate the given tensors, including padding and splitting (defaults to per-tensor get_alloc_size)
+ size_t (*get_alloc_size_n)(ggml_backend_buffer_type_t buft, struct ggml_tensor ** tensors, int n_tensors);
+ // (optional) check if tensor data is in host memory and uses standard ggml tensor layout (defaults to false)
+ bool (*is_host) (ggml_backend_buffer_type_t buft);
+ };
+
+ struct ggml_backend_buffer_type {
+ struct ggml_backend_buffer_type_i iface;
+ ggml_backend_dev_t device;
+ void * context;
+ };
+
+ // [TAG_ALLOC_SIZE_EXPAND]
+ // returns true for ops that may require additional memory for fleeting data on some backends,
+ // i.e. the backend buffer type's get_alloc_size may return more than ggml_nbytes for the output tensor
+ GGML_API bool ggml_op_alloc_size_may_expand(enum ggml_op op);
+
+ //
+ // Backend buffer
+ //
+
+ struct ggml_backend_buffer_i {
+ // (optional) free the buffer
+ void (*free_buffer) (ggml_backend_buffer_t buffer);
+ // base address of the buffer
+ void * (*get_base) (ggml_backend_buffer_t buffer);
+ // (optional) initialize a tensor in the buffer (eg. add tensor extras)
+ enum ggml_status (*init_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor);
+ // tensor data access
+ void (*memset_tensor)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, uint8_t value, size_t offset, size_t size);
+ void (*set_tensor) (ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+ void (*get_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
+ // (optional) 2d data copies
+ void (*set_tensor_2d)(ggml_backend_buffer_t buffer, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+ void (*get_tensor_2d)(ggml_backend_buffer_t buffer, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+
+ // (optional) tensor copy: dst is in the buffer, src may be in any buffer, including buffers from a different backend (return false if not supported)
+ bool (*cpy_tensor) (ggml_backend_buffer_t buffer, const struct ggml_tensor * src, struct ggml_tensor * dst);
+ // clear the entire buffer
+ void (*clear) (ggml_backend_buffer_t buffer, uint8_t value);
+ // (optional) reset any internal state due to tensor initialization, such as tensor extras
+ void (*reset) (ggml_backend_buffer_t buffer);
+ };
+
+ struct ggml_backend_buffer {
+ struct ggml_backend_buffer_i iface;
+ ggml_backend_buffer_type_t buft;
+ void * context;
+ size_t size;
+ enum ggml_backend_buffer_usage usage;
+ };
+
+ GGML_API ggml_backend_buffer_t ggml_backend_buffer_init(
+ ggml_backend_buffer_type_t buft,
+ struct ggml_backend_buffer_i iface,
+ void * context,
+ size_t size);
+
+ // do not use directly, use ggml_backend_tensor_copy instead
+ GGML_API bool ggml_backend_buffer_copy_tensor(const struct ggml_tensor * src, struct ggml_tensor * dst);
+
+ // multi-buffer
+ // buffer that contains a collection of buffers
+ GGML_API ggml_backend_buffer_t ggml_backend_multi_buffer_alloc_buffer(ggml_backend_buffer_t * buffers, size_t n_buffers);
+ GGML_API bool ggml_backend_buffer_is_multi_buffer(ggml_backend_buffer_t buffer);
+ GGML_API void ggml_backend_multi_buffer_set_usage(ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+ GGML_API void ggml_backend_meta_buffer_set_usage (ggml_backend_buffer_t buffer, enum ggml_backend_buffer_usage usage);
+
+ //
+ // Backend (meta)
+ //
+
+ GGML_API bool ggml_backend_is_meta (ggml_backend_t backend);
+ GGML_API bool ggml_backend_buffer_is_meta(ggml_backend_buffer_t buf);
+ GGML_API bool ggml_backend_buft_is_meta (ggml_backend_buffer_type_t buft);
+
+ GGML_API size_t ggml_backend_meta_n_backends (ggml_backend_t meta_backend);
+ GGML_API ggml_backend_t ggml_backend_meta_simple_backend(ggml_backend_t meta_backend, size_t index);
+
+ //
+ // Backend (stream)
+ //
+
+ // passed to graph_optimize so the backend can add allocation dependencies:
+ // if the backend executes parts of the graph out of order (e.g. on concurrent streams),
+ // it must keep the affected tensors allocated until a node where execution is known to have joined
+ struct ggml_backend_graph_optimize_params {
+ // keep `tensor` allocated at least until `until` (a node of the same graph) has been computed
+ // can be called multiple times for the same tensor: the longest lifetime applies
+ void (*add_alloc_dep)(void * user_data, struct ggml_tensor * tensor, struct ggml_tensor * until);
+ void * user_data;
+ };
+
+ struct ggml_backend_i {
+ const char * (*get_name)(ggml_backend_t backend);
+
+ void (*free)(ggml_backend_t backend);
+
+ // (optional) asynchronous tensor data access
+ void (*set_tensor_async) (ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size);
+ void (*get_tensor_async) (ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size);
+ void (*set_tensor_2d_async)(ggml_backend_t backend, struct ggml_tensor * tensor, const void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+ void (*get_tensor_2d_async)(ggml_backend_t backend, const struct ggml_tensor * tensor, void * data, size_t offset, size_t size, size_t n_copies, size_t stride_tensor, size_t stride_data);
+ bool (*cpy_tensor_async)(ggml_backend_t backend_src, ggml_backend_t backend_dst, const struct ggml_tensor * src, struct ggml_tensor * dst);
+
+ // (optional) complete all pending operations (required if the backend supports async operations)
+ void (*synchronize)(ggml_backend_t backend);
+
+ // (optional) graph plans (not used currently)
+ // compute graph with a plan
+ ggml_backend_graph_plan_t (*graph_plan_create) (ggml_backend_t backend, const struct ggml_cgraph * cgraph);
+ void (*graph_plan_free) (ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+ // update the plan with a new graph - this should be faster than creating a new plan when the graph has the same topology
+ void (*graph_plan_update) (ggml_backend_t backend, ggml_backend_graph_plan_t plan, const struct ggml_cgraph * cgraph);
+ // compute the graph with the plan
+ enum ggml_status (*graph_plan_compute)(ggml_backend_t backend, ggml_backend_graph_plan_t plan);
+
+ // compute graph (always async if supported by the backend)
+ enum ggml_status (*graph_compute) (ggml_backend_t backend, struct ggml_cgraph * cgraph);
+
+ // (optional) event synchronization
+ // record an event on this stream
+ void (*event_record)(ggml_backend_t backend, ggml_backend_event_t event);
+ // wait for an event on on a different stream
+ void (*event_wait) (ggml_backend_t backend, ggml_backend_event_t event);
+
+ // (optional) sort/optimize the nodes in the graph
+ void (*graph_optimize) (ggml_backend_t backend, struct ggml_cgraph * cgraph, struct ggml_backend_graph_optimize_params * params);
+ };
+
+ struct ggml_backend {
+ ggml_guid_t guid;
+ struct ggml_backend_i iface;
+ ggml_backend_dev_t device;
+ void * context;
+ };
+
+ struct ggml_backend_event {
+ struct ggml_backend_device * device;
+ void * context;
+ };
+
+ //
+ // Backend device
+ //
+
+ // Note: if additional properties are needed, we should add a struct with all of them
+ // the current functions to obtain the properties can remain, since they are more convenient for often used properties
+ struct ggml_backend_device_i {
+ // device name: short identifier for this device, such as "CPU" or "CUDA0"
+ const char * (*get_name)(ggml_backend_dev_t dev);
+
+ // device description: short informative description of the device, could be the model name
+ const char * (*get_description)(ggml_backend_dev_t dev);
+
+ // device memory in bytes: 0 bytes to indicate no memory to report
+ void (*get_memory)(ggml_backend_dev_t dev, size_t * free, size_t * total);
+
+ // device type
+ enum ggml_backend_dev_type (*get_type)(ggml_backend_dev_t dev);
+
+ // device properties
+ void (*get_props)(ggml_backend_dev_t dev, struct ggml_backend_dev_props * props);
+
+ // backend (stream) initialization
+ ggml_backend_t (*init_backend)(ggml_backend_dev_t dev, const char * params);
+
+ // preferred buffer type
+ ggml_backend_buffer_type_t (*get_buffer_type)(ggml_backend_dev_t dev);
+
+ // (optional) host buffer type (in system memory, typically this is a pinned memory buffer for faster transfers between host and device)
+ ggml_backend_buffer_type_t (*get_host_buffer_type)(ggml_backend_dev_t dev);
+
+ // (optional) buffer from pointer: create a buffer from a host pointer (useful for memory mapped models and importing data from other libraries)
+ ggml_backend_buffer_t (*buffer_from_host_ptr)(ggml_backend_dev_t dev, void * ptr, size_t size, size_t max_tensor_size);
+
+ // check if the backend can compute an operation
+ bool (*supports_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op);
+
+ // check if the backend can use tensors allocated in a buffer type
+ bool (*supports_buft)(ggml_backend_dev_t dev, ggml_backend_buffer_type_t buft);
+
+ // (optional) check if the backend wants to run an operation, even if the weights are allocated in an incompatible buffer
+ // these should be expensive operations that may benefit from running on this backend instead of the CPU backend
+ bool (*offload_op)(ggml_backend_dev_t dev, const struct ggml_tensor * op);
+
+ // (optional) event synchronization
+ ggml_backend_event_t (*event_new) (ggml_backend_dev_t dev);
+ void (*event_free) (ggml_backend_dev_t dev, ggml_backend_event_t event);
+ void (*event_synchronize) (ggml_backend_dev_t dev, ggml_backend_event_t event);
+ };
+
+ struct ggml_backend_device {
+ struct ggml_backend_device_i iface;
+ ggml_backend_reg_t reg;
+ void * context;
+ };
+
+ //
+ // Backend (reg)
+ //
+
+ struct ggml_backend_reg_i {
+ const char * (*get_name)(ggml_backend_reg_t reg);
+
+ // enumerate available devices
+ size_t (*get_device_count)(ggml_backend_reg_t reg);
+ ggml_backend_dev_t (*get_device)(ggml_backend_reg_t reg, size_t index);
+
+ // (optional) get a pointer to a function in the backend
+ // backends can add custom functions that are not part of the standard ggml-backend interface
+ void * (*get_proc_address)(ggml_backend_reg_t reg, const char * name);
+ };
+
+ struct ggml_backend_reg {
+ int api_version; // initialize to GGML_BACKEND_API_VERSION
+ struct ggml_backend_reg_i iface;
+ void * context;
+ };
+
+ // Add backend dynamic loading support to the backend
+
+ // Initialize the backend
+ typedef ggml_backend_reg_t (*ggml_backend_init_t)(void);
+ // Optional: obtain a score for the backend based on the system configuration
+ // Higher scores are preferred, 0 means the backend is not supported in the current system
+ typedef int (*ggml_backend_score_t)(void);
+
+#ifdef GGML_BACKEND_DL
+# ifdef __cplusplus
+# define GGML_BACKEND_DL_IMPL(reg_fn) \
+ extern "C" { \
+ GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
+ } \
+ ggml_backend_reg_t ggml_backend_init(void) { \
+ return reg_fn(); \
+ }
+# define GGML_BACKEND_DL_SCORE_IMPL(score_fn) \
+ extern "C" { \
+ GGML_BACKEND_API int ggml_backend_score(void); \
+ } \
+ int ggml_backend_score(void) { \
+ return score_fn(); \
+ }
+# else
+# define GGML_BACKEND_DL_IMPL(reg_fn) \
+ GGML_BACKEND_API ggml_backend_reg_t ggml_backend_init(void); \
+ ggml_backend_reg_t ggml_backend_init(void) { \
+ return reg_fn(); \
+ }
+# define GGML_BACKEND_DL_SCORE_IMPL(score_fn) \
+ GGML_BACKEND_API int ggml_backend_score(void); \
+ int ggml_backend_score(void) { \
+ return score_fn(); \
+ }
+# endif
+#else
+# define GGML_BACKEND_DL_IMPL(reg_fn)
+# define GGML_BACKEND_DL_SCORE_IMPL(score_fn)
+#endif
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/python/freetoken/kernel/csrc/ggml/src/ggml-quants.c b/python/freetoken/kernel/csrc/ggml/src/ggml-quants.c
new file mode 100644
index 00000000..7750a72c
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/src/ggml-quants.c
@@ -0,0 +1,5639 @@
+#define GGML_COMMON_IMPL_C
+#include "ggml-common.h"
+
+#include "ggml-quants.h"
+#include "ggml-impl.h"
+#include "ggml-cpu/ggml-cpu-impl.h"
+#include "ggml-cpu.h"
+
+#include <math.h>
+#include <string.h>
+#include <assert.h>
+#include <float.h>
+#include <stdlib.h> // for qsort
+#include <stdio.h> // for GGML_ASSERT
+
+#ifdef GGML_USE_OPENMP
+#include <omp.h>
+#endif
+
+#define GROUP_MAX_EPS 1e-15f
+#define GROUP_MAX_EPS_IQ3_XXS 1e-8f
+#define GROUP_MAX_EPS_IQ2_S 1e-8f
+#define GROUP_MAX_EPS_IQ1_M 1e-7f
+#define GROUP_MAX_EPS_IQ1_S 1e-12f
+
+#define UNUSED GGML_UNUSED
+
+static inline int best_index_int8(int n, const int8_t * val, float x) {
+ if (x <= val[0]) return 0;
+ if (x >= val[n-1]) return n-1;
+ int ml = 0, mu = n-1;
+ while (mu-ml > 1) {
+ int mav = (ml+mu)/2;
+ if (x < val[mav]) mu = mav; else ml = mav;
+ }
+ return x - val[mu-1] < val[mu] - x ? mu-1 : mu;
+}
+
+// reference implementation for deterministic creation of model files
+void quantize_row_q1_0_ref(const float * GGML_RESTRICT x, block_q1_0 * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK1_0;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ float sum_abs = 0.0f;
+ for (int j = 0; j < qk; j++) {
+ sum_abs += fabsf(x[i*qk + j]);
+ }
+ const float d = sum_abs / qk;
+
+ y[i].d = GGML_FP32_TO_FP16(d);
+
+ // Clear all bits first
+ for (int j = 0; j < qk / 8; ++j) {
+ y[i].qs[j] = 0;
+ }
+
+ // Just store sign of each weight directly (no normalization)
+ for (int j = 0; j < qk; ++j) {
+ const int bit_index = j;
+ const int byte_index = bit_index / 8;
+ const int bit_offset = bit_index % 8;
+
+ if (x[i*qk + j] >= 0.0f) {
+ y[i].qs[byte_index] |= (1 << bit_offset);
+ }
+ }
+ }
+}
+
+void quantize_row_q2_0_ref(const float * GGML_RESTRICT x, block_q2_0 * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK2_0;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ // Compute scale as max absolute value in the block
+ float amax = 0.0f;
+ for (int j = 0; j < qk; j++) {
+ const float a = fabsf(x[i*qk + j]);
+ if (a > amax) amax = a;
+ }
+ const float d = amax;
+ const float id = d > 0.0f ? 1.0f / d : 0.0f;
+
+ y[i].d = GGML_FP32_TO_FP16(d);
+
+ // Clear quant bytes
+ for (int j = 0; j < qk / 4; ++j) {
+ y[i].qs[j] = 0;
+ }
+
+ // Encode 2-bit values: round(w/d) clamped to [-1, 2], then add 1
+ // 00 (-1) = -scale, 01 (0) = 0, 10 (+1) = +scale, 11 (+2) = 2*scale
+ for (int j = 0; j < qk; ++j) {
+ const float w = x[i*qk + j];
+ int q = (int)roundf(w * id) + 1;
+ if (q < 0) q = 0;
+ if (q > 3) q = 3;
+ const int byte_index = j / 4;
+ const int bit_offset = (j % 4) * 2;
+ y[i].qs[byte_index] |= ((uint8_t)q << bit_offset);
+ }
+ }
+}
+
+// reference implementation for deterministic creation of model files
+void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK4_0;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ float amax = 0.0f; // absolute max
+ float max = 0.0f;
+
+ for (int j = 0; j < qk; j++) {
+ const float v = x[i*qk + j];
+ if (amax < fabsf(v)) {
+ amax = fabsf(v);
+ max = v;
+ }
+ }
+
+ const float d = max / -8;
+ const float id = d ? 1.0f/d : 0.0f;
+
+ y[i].d = GGML_FP32_TO_FP16(d);
+
+ for (int j = 0; j < qk/2; ++j) {
+ const float x0 = x[i*qk + 0 + j]*id;
+ const float x1 = x[i*qk + qk/2 + j]*id;
+
+ const uint8_t xi0 = MIN(15, (int8_t)(x0 + 8.5f));
+ const uint8_t xi1 = MIN(15, (int8_t)(x1 + 8.5f));
+
+ y[i].qs[j] = xi0;
+ y[i].qs[j] |= xi1 << 4;
+ }
+ }
+}
+
+void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k) {
+ const int qk = QK4_1;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ float min = FLT_MAX;
+ float max = -FLT_MAX;
+
+ for (int j = 0; j < qk; j++) {
+ const float v = x[i*qk + j];
+
+ if (v < min) min = v;
+ if (v > max) max = v;
+ }
+
+ const float d = (max - min) / ((1 << 4) - 1);
+ const float id = d ? 1.0f/d : 0.0f;
+
+ y[i].d = GGML_FP32_TO_FP16(d);
+ y[i].m = GGML_FP32_TO_FP16(min);
+
+ for (int j = 0; j < qk/2; ++j) {
+ const float x0 = (x[i*qk + 0 + j] - min)*id;
+ const float x1 = (x[i*qk + qk/2 + j] - min)*id;
+
+ const uint8_t xi0 = MIN(15, (int8_t)(x0 + 0.5f));
+ const uint8_t xi1 = MIN(15, (int8_t)(x1 + 0.5f));
+
+ y[i].qs[j] = xi0;
+ y[i].qs[j] |= xi1 << 4;
+ }
+ }
+}
+
+void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK5_0;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ float amax = 0.0f; // absolute max
+ float max = 0.0f;
+
+ for (int j = 0; j < qk; j++) {
+ const float v = x[i*qk + j];
+ if (amax < fabsf(v)) {
+ amax = fabsf(v);
+ max = v;
+ }
+ }
+
+ const float d = max / -16;
+ const float id = d ? 1.0f/d : 0.0f;
+
+ y[i].d = GGML_FP32_TO_FP16(d);
+
+ uint32_t qh = 0;
+
+ for (int j = 0; j < qk/2; ++j) {
+ const float x0 = x[i*qk + 0 + j]*id;
+ const float x1 = x[i*qk + qk/2 + j]*id;
+
+ const uint8_t xi0 = MIN(31, (int8_t)(x0 + 16.5f));
+ const uint8_t xi1 = MIN(31, (int8_t)(x1 + 16.5f));
+
+ y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
+
+ // get the 5-th bit and store it in qh at the right position
+ qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+ qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2);
+ }
+
+ memcpy(&y[i].qh, &qh, sizeof(qh));
+ }
+}
+
+void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k) {
+ const int qk = QK5_1;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ float min = FLT_MAX;
+ float max = -FLT_MAX;
+
+ for (int j = 0; j < qk; j++) {
+ const float v = x[i*qk + j];
+
+ if (v < min) min = v;
+ if (v > max) max = v;
+ }
+
+ const float d = (max - min) / ((1 << 5) - 1);
+ const float id = d ? 1.0f/d : 0.0f;
+
+ y[i].d = GGML_FP32_TO_FP16(d);
+ y[i].m = GGML_FP32_TO_FP16(min);
+
+ uint32_t qh = 0;
+
+ for (int j = 0; j < qk/2; ++j) {
+ const float x0 = (x[i*qk + 0 + j] - min)*id;
+ const float x1 = (x[i*qk + qk/2 + j] - min)*id;
+
+ const uint8_t xi0 = (uint8_t)(x0 + 0.5f);
+ const uint8_t xi1 = (uint8_t)(x1 + 0.5f);
+
+ y[i].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
+
+ // get the 5-th bit and store it in qh at the right position
+ qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+ qh |= ((xi1 & 0x10u) >> 4) << (j + qk/2);
+ }
+
+ memcpy(&y[i].qh, &qh, sizeof(y[i].qh));
+ }
+}
+
+// reference implementation for deterministic creation of model files
+void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK8_0 == 0);
+ const int nb = k / QK8_0;
+
+ for (int i = 0; i < nb; i++) {
+ float amax = 0.0f; // absolute max
+
+ for (int j = 0; j < QK8_0; j++) {
+ const float v = x[i*QK8_0 + j];
+ amax = MAX(amax, fabsf(v));
+ }
+
+ const float d = amax / ((1 << 7) - 1);
+ const float id = d ? 1.0f/d : 0.0f;
+
+ y[i].d = GGML_FP32_TO_FP16(d);
+
+ for (int j = 0; j < QK8_0; ++j) {
+ const float x0 = x[i*QK8_0 + j]*id;
+
+ y[i].qs[j] = roundf(x0);
+ }
+ }
+}
+
+// reference implementation for deterministic creation of model files
+void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k) {
+ assert(QK8_1 == 32);
+ assert(k % QK8_1 == 0);
+ const int nb = k / QK8_1;
+
+ for (int i = 0; i < nb; i++) {
+ float amax = 0.0f; // absolute max
+
+ for (int j = 0; j < QK8_1; j++) {
+ const float v = x[i*QK8_1 + j];
+ amax = MAX(amax, fabsf(v));
+ }
+
+ const float d = amax / ((1 << 7) - 1);
+ const float id = d ? 1.0f/d : 0.0f;
+
+ y[i].d = GGML_FP32_TO_FP16(d);
+
+ int sum = 0;
+
+ for (int j = 0; j < QK8_1/2; ++j) {
+ const float v0 = x[i*QK8_1 + j]*id;
+ const float v1 = x[i*QK8_1 + QK8_1/2 + j]*id;
+
+ y[i].qs[ j] = roundf(v0);
+ y[i].qs[QK8_1/2 + j] = roundf(v1);
+
+ sum += y[i].qs[ j];
+ sum += y[i].qs[QK8_1/2 + j];
+ }
+
+ y[i].s = GGML_FP32_TO_FP16(sum*d);
+ }
+}
+
+static inline int best_index_mxfp4(float x, float e) {
+ int best_index = 0;
+ float best_err = fabsf(kvalues_mxfp4[0]*e - x);
+ for (int i = 1; i < 16; i++) {
+ float err = fabsf(kvalues_mxfp4[i]*e - x);
+ if (err < best_err) {
+ best_index = i;
+ best_err = err;
+ }
+ }
+ return best_index;
+}
+
+void quantize_row_mxfp4_ref(const float * GGML_RESTRICT x, block_mxfp4 * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK_MXFP4;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ float amax = 0.0f; // absolute max
+
+ for (int j = 0; j < qk; j++) {
+ const float v = x[i*qk + j];
+
+ if (amax < fabsf(v)) {
+ amax = fabsf(v);
+ }
+ }
+
+ const uint8_t e = amax > 0.0f ? (uint8_t) (floorf(log2f(amax)) - 2 + 127) : 0;
+
+ const float d = GGML_E8M0_TO_FP32_HALF(e);
+
+ y[i].e = e;
+
+ for (int j = 0; j < qk/2; ++j) {
+ const uint8_t x0 = best_index_mxfp4(x[i*qk + 0 + j], d);
+ const uint8_t x1 = best_index_mxfp4(x[i*qk + qk/2 + j], d);
+
+ y[i].qs[j] = x0;
+ y[i].qs[j] |= x1 << 4;
+ }
+ }
+}
+
+void quantize_row_nvfp4_ref(const float * GGML_RESTRICT x, block_nvfp4 * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK_NVFP4;
+ static const int qk_sub = QK_NVFP4_SUB;
+ static const int n_sub = QK_NVFP4 / QK_NVFP4_SUB;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ for (int s = 0; s < n_sub; s++) {
+ const float * xb = x + i*qk + s*qk_sub;
+
+ float amax = 0.0f;
+ for (int j = 0; j < qk_sub; j++) {
+ if (amax < fabsf(xb[j])) {
+ amax = fabsf(xb[j]);
+ }
+ }
+
+ // UE4M3 scale: amax / 6.0 maps the max E2M1 value (6.0) to amax
+ const uint8_t ue = ggml_fp32_to_ue4m3(amax / 6.0f);
+ y[i].d[s] = ue;
+ const float d = ggml_ue4m3_to_fp32(ue);
+
+ for (int j = 0; j < qk_sub/2; ++j) {
+ const uint8_t x0 = best_index_mxfp4(xb[0 + j], d);
+ const uint8_t x1 = best_index_mxfp4(xb[qk_sub/2 + j], d);
+
+ y[i].qs[s*(qk_sub/2) + j] = x0 | (x1 << 4);
+ }
+ }
+ }
+}
+
+void dequantize_row_q1_0(const block_q1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK1_0;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+ const float neg_d = -d;
+
+ for (int j = 0; j < qk; ++j) {
+ const int byte_index = j / 8;
+ const int bit_offset = j % 8;
+ const uint8_t bit = (x[i].qs[byte_index] >> bit_offset) & 1;
+ y[i*qk + j] = bit ? d : neg_d;
+ }
+ }
+}
+
+void dequantize_row_q2_0(const block_q2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK2_0;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+
+ for (int j = 0; j < qk; ++j) {
+ const int byte_index = j / 4;
+ const int bit_offset = (j % 4) * 2;
+ const uint8_t q = (x[i].qs[byte_index] >> bit_offset) & 0x03;
+ // 00=-1, 01=0, 10=+1, 11=+2
+ y[i*qk + j] = ((int)q - 1) * d;
+ }
+ }
+}
+
+void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK4_0;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+
+ for (int j = 0; j < qk/2; ++j) {
+ const int x0 = (x[i].qs[j] & 0x0F) - 8;
+ const int x1 = (x[i].qs[j] >> 4) - 8;
+
+ y[i*qk + j + 0 ] = x0*d;
+ y[i*qk + j + qk/2] = x1*d;
+ }
+ }
+}
+
+void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK4_1;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+ const float m = GGML_FP16_TO_FP32(x[i].m);
+
+ for (int j = 0; j < qk/2; ++j) {
+ const int x0 = (x[i].qs[j] & 0x0F);
+ const int x1 = (x[i].qs[j] >> 4);
+
+ y[i*qk + j + 0 ] = x0*d + m;
+ y[i*qk + j + qk/2] = x1*d + m;
+ }
+ }
+}
+
+void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK5_0;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+
+ uint32_t qh;
+ memcpy(&qh, x[i].qh, sizeof(qh));
+
+ for (int j = 0; j < qk/2; ++j) {
+ const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10;
+ const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10;
+
+ const int32_t x0 = ((x[i].qs[j] & 0x0F) | xh_0) - 16;
+ const int32_t x1 = ((x[i].qs[j] >> 4) | xh_1) - 16;
+
+ y[i*qk + j + 0 ] = x0*d;
+ y[i*qk + j + qk/2] = x1*d;
+ }
+ }
+}
+
+void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK5_1;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+ const float m = GGML_FP16_TO_FP32(x[i].m);
+
+ uint32_t qh;
+ memcpy(&qh, x[i].qh, sizeof(qh));
+
+ for (int j = 0; j < qk/2; ++j) {
+ const uint8_t xh_0 = ((qh >> (j + 0)) << 4) & 0x10;
+ const uint8_t xh_1 = ((qh >> (j + 12)) ) & 0x10;
+
+ const int x0 = (x[i].qs[j] & 0x0F) | xh_0;
+ const int x1 = (x[i].qs[j] >> 4) | xh_1;
+
+ y[i*qk + j + 0 ] = x0*d + m;
+ y[i*qk + j + qk/2] = x1*d + m;
+ }
+ }
+}
+
+void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK8_0;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+
+ for (int j = 0; j < qk; ++j) {
+ y[i*qk + j] = x[i].qs[j]*d;
+ }
+ }
+}
+
+void dequantize_row_mxfp4(const block_mxfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK_MXFP4;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ const float d = GGML_E8M0_TO_FP32_HALF(x[i].e);
+
+ for (int j = 0; j < qk/2; ++j) {
+ const int8_t x0 = kvalues_mxfp4[x[i].qs[j] & 0x0F];
+ const int8_t x1 = kvalues_mxfp4[x[i].qs[j] >> 4];
+
+ y[i*qk + j + 0 ] = x0*d;
+ y[i*qk + j + qk/2] = x1*d;
+ }
+ }
+}
+
+void dequantize_row_nvfp4(const block_nvfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ static const int qk = QK_NVFP4;
+ static const int qk_sub = QK_NVFP4_SUB;
+ static const int n_sub = QK_NVFP4 / QK_NVFP4_SUB;
+
+ assert(k % qk == 0);
+
+ const int nb = k / qk;
+
+ for (int i = 0; i < nb; i++) {
+ for (int s = 0; s < n_sub; s++) {
+ const float d = ggml_ue4m3_to_fp32(x[i].d[s]);
+ float * yb = y + i*qk + s*qk_sub;
+
+ for (int j = 0; j < qk_sub/2; ++j) {
+ const int8_t v0 = kvalues_mxfp4[x[i].qs[s*(qk_sub/2) + j] & 0x0F];
+ const int8_t v1 = kvalues_mxfp4[x[i].qs[s*(qk_sub/2) + j] >> 4];
+
+ yb[j + 0 ] = v0*d;
+ yb[j + qk_sub/2] = v1*d;
+ }
+ }
+ }
+}
+
+//
+// 2-6 bit quantization in super-blocks
+//
+
+//
+// ===================== Helper functions
+//
+static inline int nearest_int(float fval) {
+ assert(fabsf(fval) <= 4194303.f);
+ float val = fval + 12582912.f;
+ int i; memcpy(&i, &val, sizeof(int));
+ return (i & 0x007fffff) - 0x00400000;
+}
+
+static float make_qx_quants(int n, int nmax, const float * GGML_RESTRICT x, int8_t * GGML_RESTRICT L, int rmse_type,
+ const float * GGML_RESTRICT qw) {
+ float max = 0;
+ float amax = 0;
+ for (int i = 0; i < n; ++i) {
+ float ax = fabsf(x[i]);
+ if (ax > amax) { amax = ax; max = x[i]; }
+ }
+ if (amax < GROUP_MAX_EPS) { // all zero
+ for (int i = 0; i < n; ++i) {
+ L[i] = 0;
+ }
+ return 0.f;
+ }
+ float iscale = -nmax / max;
+ if (rmse_type == 0) {
+ for (int i = 0; i < n; ++i) {
+ int l = nearest_int(iscale * x[i]);
+ L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
+ }
+ return 1/iscale;
+ }
+ bool return_early = false;
+ if (rmse_type < 0) {
+ rmse_type = -rmse_type;
+ return_early = true;
+ }
+ float sumlx = 0;
+ float suml2 = 0;
+#ifdef HAVE_BUGGY_APPLE_LINKER
+ // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7
+ for (volatile int i = 0; i < n; ++i) {
+#else
+ for (int i = 0; i < n; ++i) {
+#endif
+ int l = nearest_int(iscale * x[i]);
+ l = MAX(-nmax, MIN(nmax-1, l));
+ L[i] = l + nmax;
+ float w = qw ? qw[i] : rmse_type == 1 ? x[i] * x[i] : rmse_type == 2 ? 1 : rmse_type == 3 ? fabsf(x[i]) : sqrtf(fabsf(x[i]));
+ sumlx += w*x[i]*l;
+ suml2 += w*l*l;
+ }
+ float scale = suml2 ? sumlx/suml2 : 0.0f;
+ if (return_early) return suml2 > 0 ? 0.5f*(scale + 1/iscale) : 1/iscale;
+ float best = scale * sumlx;
+ for (int is = -9; is <= 9; ++is) {
+ if (is == 0) {
+ continue;
+ }
+ iscale = -(nmax + 0.1f*is) / max;
+ sumlx = suml2 = 0;
+ for (int i = 0; i < n; ++i) {
+ int l = nearest_int(iscale * x[i]);
+ l = MAX(-nmax, MIN(nmax-1, l));
+ float w = qw ? qw[i] : rmse_type == 1 ? x[i] * x[i] : rmse_type == 2 ? 1 : rmse_type == 3 ? fabsf(x[i]) : sqrtf(fabsf(x[i]));
+ sumlx += w*x[i]*l;
+ suml2 += w*l*l;
+ }
+ if (suml2 > 0 && sumlx*sumlx > best*suml2) {
+ for (int i = 0; i < n; ++i) {
+ int l = nearest_int(iscale * x[i]);
+ L[i] = nmax + MAX(-nmax, MIN(nmax-1, l));
+ }
+ scale = sumlx/suml2; best = scale*sumlx;
+ }
+ }
+ return scale;
+}
+
+static float make_q3_quants(int n, int nmax, const float * GGML_RESTRICT x, int8_t * GGML_RESTRICT L, bool do_rmse) {
+ float max = 0;
+ float amax = 0;
+ for (int i = 0; i < n; ++i) {
+ float ax = fabsf(x[i]);
+ if (ax > amax) { amax = ax; max = x[i]; }
+ }
+ if (amax < GROUP_MAX_EPS) { // all zero
+ for (int i = 0; i < n; ++i) { L[i] = 0; }
+ return 0.f;
+ }
+ float iscale = -nmax / max;
+ if (do_rmse) {
+ float sumlx = 0;
+ float suml2 = 0;
+ for (int i = 0; i < n; ++i) {
+ int l = nearest_int(iscale * x[i]);
+ l = MAX(-nmax, MIN(nmax-1, l));
+ L[i] = l;
+ float w = x[i]*x[i];
+ sumlx += w*x[i]*l;
+ suml2 += w*l*l;
+ }
+ for (int itry = 0; itry < 5; ++itry) {
+ int n_changed = 0;
+ for (int i = 0; i < n; ++i) {
+ float w = x[i]*x[i];
+ float slx = sumlx - w*x[i]*L[i];
+ if (slx > 0) {
+ float sl2 = suml2 - w*L[i]*L[i];
+ int new_l = nearest_int(x[i] * sl2 / slx);
+ new_l = MAX(-nmax, MIN(nmax-1, new_l));
+ if (new_l != L[i]) {
+ slx += w*x[i]*new_l;
+ sl2 += w*new_l*new_l;
+ if (sl2 > 0 && slx*slx*suml2 > sumlx*sumlx*sl2) {
+ L[i] = new_l; sumlx = slx; suml2 = sl2;
+ ++n_changed;
+ }
+ }
+ }
+ }
+ if (!n_changed) {
+ break;
+ }
+ }
+ for (int i = 0; i < n; ++i) {
+ L[i] += nmax;
+ }
+ return suml2 > 0.0f ? sumlx / suml2 : 0.0f;
+ }
+ for (int i = 0; i < n; ++i) {
+ int l = nearest_int(iscale * x[i]);
+ l = MAX(-nmax, MIN(nmax-1, l));
+ L[i] = l + nmax;
+ }
+ return 1/iscale;
+}
+
+static float make_qkx1_quants(int n, int nmax, const float * GGML_RESTRICT x, uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min,
+ int ntry, float alpha) {
+ float min = x[0];
+ float max = x[0];
+ for (int i = 1; i < n; ++i) {
+ if (x[i] < min) min = x[i];
+ if (x[i] > max) max = x[i];
+ }
+ if (max == min) {
+ for (int i = 0; i < n; ++i) L[i] = 0;
+ *the_min = 0;
+ return 0.f;
+ }
+ if (min > 0) min = 0;
+ float iscale = nmax/(max - min);
+ float scale = 1/iscale;
+ for (int itry = 0; itry < ntry; ++itry) {
+ float sumlx = 0; int suml2 = 0;
+ bool did_change = false;
+ for (int i = 0; i < n; ++i) {
+ int l = nearest_int(iscale*(x[i] - min));
+ l = MAX(0, MIN(nmax, l));
+ if (l != L[i]) {
+ L[i] = l;
+ did_change = true;
+ }
+ sumlx += (x[i] - min)*l;
+ suml2 += l*l;
+ }
+ scale = sumlx/suml2;
+ float sum = 0;
+ for (int i = 0; i < n; ++i) {
+ sum += x[i] - scale*L[i];
+ }
+ min = alpha*min + (1 - alpha)*sum/n;
+ if (min > 0) min = 0;
+ iscale = 1/scale;
+ if (!did_change) break;
+ }
+ *the_min = -min;
+ return scale;
+}
+
+static float make_qkx2_quants(int n, int nmax, const float * GGML_RESTRICT x, const float * GGML_RESTRICT weights,
+ uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min, uint8_t * GGML_RESTRICT Laux,
+ float rmin, float rdelta, int nstep, bool use_mad) {
+ float min = x[0];
+ float max = x[0];
+ float sum_w = weights[0];
+ float sum_x = sum_w * x[0];
+#ifdef HAVE_BUGGY_APPLE_LINKER
+ // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7
+ for (volatile int i = 1; i < n; ++i) {
+#else
+ for (int i = 1; i < n; ++i) {
+#endif
+ if (x[i] < min) min = x[i];
+ if (x[i] > max) max = x[i];
+ float w = weights[i];
+ sum_w += w;
+ sum_x += w * x[i];
+ }
+ if (min > 0) min = 0;
+ if (max == min) {
+ for (int i = 0; i < n; ++i) L[i] = 0;
+ *the_min = -min;
+ return 0.f;
+ }
+ float iscale = nmax/(max - min);
+ float scale = 1/iscale;
+ float best_error = 0;
+ for (int i = 0; i < n; ++i) {
+ int l = nearest_int(iscale*(x[i] - min));
+ L[i] = MAX(0, MIN(nmax, l));
+ float diff = scale * L[i] + min - x[i];
+ diff = use_mad ? fabsf(diff) : diff * diff;
+ float w = weights[i];
+ best_error += w * diff;
+ }
+ if (nstep < 1) {
+ *the_min = -min;
+ return scale;
+ }
+ for (int is = 0; is <= nstep; ++is) {
+ iscale = (rmin + rdelta*is + nmax)/(max - min);
+ float sum_l = 0, sum_l2 = 0, sum_xl = 0;
+ for (int i = 0; i < n; ++i) {
+ int l = nearest_int(iscale*(x[i] - min));
+ l = MAX(0, MIN(nmax, l));
+ Laux[i] = l;
+ float w = weights[i];
+ sum_l += w*l;
+ sum_l2 += w*l*l;
+ sum_xl += w*l*x[i];
+ }
+ float D = sum_w * sum_l2 - sum_l * sum_l;
+ if (D > 0) {
+ float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D;
+ float this_min = (sum_l2 * sum_x - sum_l * sum_xl)/D;
+ if (this_min > 0) {
+ this_min = 0;
+ this_scale = sum_xl / sum_l2;
+ }
+ float cur_error = 0;
+ for (int i = 0; i < n; ++i) {
+ float diff = this_scale * Laux[i] + this_min - x[i];
+ diff = use_mad ? fabsf(diff) : diff * diff;
+ float w = weights[i];
+ cur_error += w * diff;
+ }
+ if (cur_error < best_error) {
+ for (int i = 0; i < n; ++i) {
+ L[i] = Laux[i];
+ }
+ best_error = cur_error;
+ scale = this_scale;
+ min = this_min;
+ }
+ }
+ }
+ *the_min = -min;
+ return scale;
+}
+
+static inline void get_scale_min_k4(int j, const uint8_t * GGML_RESTRICT q, uint8_t * GGML_RESTRICT d, uint8_t * GGML_RESTRICT m) {
+ if (j < 4) {
+ *d = q[j] & 63; *m = q[j + 4] & 63;
+ } else {
+ *d = (q[j+4] & 0xF) | ((q[j-4] >> 6) << 4);
+ *m = (q[j+4] >> 4) | ((q[j-0] >> 6) << 4);
+ }
+}
+
+//========================- 2-bit (de)-quantization
+
+void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int nb = k / QK_K;
+
+ uint8_t L[QK_K];
+ uint8_t Laux[16];
+ float weights[16];
+ float mins[QK_K/16];
+ float scales[QK_K/16];
+
+ const float q4scale = 15.f;
+
+ for (int i = 0; i < nb; i++) {
+ float max_scale = 0; // as we are deducting the min, scales are always positive
+ float max_min = 0;
+ for (int j = 0; j < QK_K/16; ++j) {
+ for (int l = 0; l < 16; ++l) weights[l] = fabsf(x[16*j + l]);
+ scales[j] = make_qkx2_quants(16, 3, x + 16*j, weights, L + 16*j, &mins[j], Laux, -0.5f, 0.1f, 15, true);
+ float scale = scales[j];
+ if (scale > max_scale) {
+ max_scale = scale;
+ }
+ float min = mins[j];
+ if (min > max_min) {
+ max_min = min;
+ }
+ }
+
+ if (max_scale > 0) {
+ float iscale = q4scale/max_scale;
+ for (int j = 0; j < QK_K/16; ++j) {
+ int l = nearest_int(iscale*scales[j]);
+ y[i].scales[j] = l;
+ }
+ y[i].d = GGML_FP32_TO_FP16(max_scale/q4scale);
+ } else {
+ for (int j = 0; j < QK_K/16; ++j) y[i].scales[j] = 0;
+ y[i].d = GGML_FP32_TO_FP16(0.f);
+ }
+ if (max_min > 0) {
+ float iscale = q4scale/max_min;
+ for (int j = 0; j < QK_K/16; ++j) {
+ int l = nearest_int(iscale*mins[j]);
+ y[i].scales[j] |= (l << 4);
+ }
+ y[i].dmin = GGML_FP32_TO_FP16(max_min/q4scale);
+ } else {
+ y[i].dmin = GGML_FP32_TO_FP16(0.f);
+ }
+ for (int j = 0; j < QK_K/16; ++j) {
+ const float d = GGML_FP16_TO_FP32(y[i].d) * (y[i].scales[j] & 0xF);
+ if (!d) continue;
+ const float dm = GGML_FP16_TO_FP32(y[i].dmin) * (y[i].scales[j] >> 4);
+ for (int ii = 0; ii < 16; ++ii) {
+ int l = nearest_int((x[16*j + ii] + dm)/d);
+ l = MAX(0, MIN(3, l));
+ L[16*j + ii] = l;
+ }
+ }
+
+ for (int j = 0; j < QK_K; j += 128) {
+ for (int l = 0; l < 32; ++l) {
+ y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
+ }
+ }
+
+ x += QK_K;
+ }
+}
+
+void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int nb = k / QK_K;
+
+ for (int i = 0; i < nb; i++) {
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+ const float min = GGML_FP16_TO_FP32(x[i].dmin);
+
+ const uint8_t * q = x[i].qs;
+
+ int is = 0;
+ float dl, ml;
+ for (int n = 0; n < QK_K; n += 128) {
+ int shift = 0;
+ for (int j = 0; j < 4; ++j) {
+
+ uint8_t sc = x[i].scales[is++];
+ dl = d * (sc & 0xF); ml = min * (sc >> 4);
+ for (int l = 0; l < 16; ++l) *y++ = dl * ((int8_t)((q[l] >> shift) & 3)) - ml;
+
+ sc = x[i].scales[is++];
+ dl = d * (sc & 0xF); ml = min * (sc >> 4);
+ for (int l = 0; l < 16; ++l) *y++ = dl * ((int8_t)((q[l+16] >> shift) & 3)) - ml;
+
+ shift += 2;
+ }
+ q += 32;
+ }
+ }
+}
+
+static float make_qkx3_quants(int n, int nmax, const float * GGML_RESTRICT x, const float * GGML_RESTRICT weights,
+ uint8_t * GGML_RESTRICT L, float * GGML_RESTRICT the_min, uint8_t * GGML_RESTRICT Laux,
+ float rmin, float rdelta, int nstep, bool use_mad) {
+ float min = x[0];
+ float max = x[0];
+ float sum_w = weights ? weights[0] : x[0]*x[0];
+ float sum_x = sum_w * x[0];
+#ifdef HAVE_BUGGY_APPLE_LINKER
+ // use 'volatile' to prevent unroll and work around a bug in Apple ld64 1015.7
+ for (volatile int i = 1; i < n; ++i) {
+#else
+ for (int i = 1; i < n; ++i) {
+#endif
+ if (x[i] < min) min = x[i];
+ if (x[i] > max) max = x[i];
+ float w = weights ? weights[i] : x[i]*x[i];
+ sum_w += w;
+ sum_x += w * x[i];
+ }
+ if (min > 0) {
+ min = 0;
+ }
+ if (max <= min) {
+ memset(L, 0, n);
+ *the_min = -min;
+ return 0.f;
+ }
+ float iscale = nmax/(max - min);
+ float scale = 1/iscale;
+ float best_mad = 0;
+ for (int i = 0; i < n; ++i) {
+ int l = nearest_int(iscale*(x[i] - min));
+ L[i] = MAX(0, MIN(nmax, l));
+ float diff = scale * L[i] + min - x[i];
+ diff = use_mad ? fabsf(diff) : diff*diff;
+ float w = weights ? weights[i] : x[i]*x[i];
+ best_mad += w * diff;
+ }
+ if (nstep < 1) {
+ *the_min = -min;
+ return scale;
+ }
+ for (int is = 0; is <= nstep; ++is) {
+ iscale = (rmin + rdelta*is + nmax)/(max - min);
+ float sum_l = 0, sum_l2 = 0, sum_xl = 0;
+ for (int i = 0; i < n; ++i) {
+ // min is the best fit so far and can be at or near max, so v can be inf, nan or out of range for nearest_int
+ const float v = iscale*(x[i] - min);
+ const int l = v > 0 ? nearest_int(MIN(v, nmax)) : 0;
+ Laux[i] = l;
+ float w = weights ? weights[i] : x[i]*x[i];
+ sum_l += w*l;
+ sum_l2 += w*l*l;
+ sum_xl += w*l*x[i];
+ }
+ float D = sum_w * sum_l2 - sum_l * sum_l;
+ if (D > 0) {
+ float this_scale = (sum_w * sum_xl - sum_x * sum_l)/D;
+ float this_min = (sum_l2 * sum_x - sum_l * sum_xl)/D;
+ if (this_min > 0) {
+ this_min = 0;
+ this_scale = sum_xl / sum_l2;
+ }
+ float mad = 0;
+ for (int i = 0; i < n; ++i) {
+ float diff = this_scale * Laux[i] + this_min - x[i];
+ diff = use_mad ? fabsf(diff) : diff*diff;
+ float w = weights ? weights[i] : x[i]*x[i];
+ mad += w * diff;
+ }
+ if (mad < best_mad) {
+ for (int i = 0; i < n; ++i) {
+ L[i] = Laux[i];
+ }
+ best_mad = mad;
+ scale = this_scale;
+ min = this_min;
+ }
+ }
+ }
+ *the_min = -min;
+ return scale;
+}
+
+static float make_qp_quants(int n, int nmax, const float * GGML_RESTRICT x, uint8_t * GGML_RESTRICT L, const float * quant_weights) {
+ float max = 0;
+ for (int i = 0; i < n; ++i) {
+ max = MAX(max, x[i]);
+ }
+ if (max < GROUP_MAX_EPS) { // all zero
+ for (int i = 0; i < n; ++i) { L[i] = 0; }
+ return 0.f;
+ }
+ float iscale = nmax / max;
+ for (int i = 0; i < n; ++i) {
+ L[i] = nearest_int(iscale * x[i]);
+ }
+ float scale = 1/iscale;
+ float best_mse = 0;
+ for (int i = 0; i < n; ++i) {
+ float diff = x[i] - scale*L[i];
+ float w = quant_weights[i];
+ best_mse += w*diff*diff;
+ }
+ for (int is = -4; is <= 4; ++is) {
+ if (is == 0) continue;
+ float iscale_is = (0.1f*is + nmax)/max;
+ float scale_is = 1/iscale_is;
+ float mse = 0;
+ for (int i = 0; i < n; ++i) {
+ int l = nearest_int(iscale_is*x[i]);
+ l = MIN(nmax, l);
+ float diff = x[i] - scale_is*l;
+ float w = quant_weights[i];
+ mse += w*diff*diff;
+ }
+ if (mse < best_mse) {
+ best_mse = mse;
+ iscale = iscale_is;
+ }
+ }
+ float sumlx = 0;
+ float suml2 = 0;
+ for (int i = 0; i < n; ++i) {
+ int l = nearest_int(iscale * x[i]);
+ l = MIN(nmax, l);
+ L[i] = l;
+ float w = quant_weights[i];
+ sumlx += w*x[i]*l;
+ suml2 += w*l*l;
+ }
+ for (int itry = 0; itry < 5; ++itry) {
+ int n_changed = 0;
+ for (int i = 0; i < n; ++i) {
+ float w = quant_weights[i];
+ float slx = sumlx - w*x[i]*L[i];
+ float sl2 = suml2 - w*L[i]*L[i];
+ if (slx > 0 && sl2 > 0) {
+ int new_l = nearest_int(x[i] * sl2 / slx);
+ new_l = MIN(nmax, new_l);
+ if (new_l != L[i]) {
+ slx += w*x[i]*new_l;
+ sl2 += w*new_l*new_l;
+ if (slx*slx*suml2 > sumlx*sumlx*sl2) {
+ L[i] = new_l; sumlx = slx; suml2 = sl2;
+ ++n_changed;
+ }
+ }
+ }
+ }
+ if (!n_changed) {
+ break;
+ }
+ }
+ return suml2 > 0.0f ? sumlx / suml2 : 0.0f;
+}
+
+static void quantize_row_q2_K_impl(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int k, const float * GGML_RESTRICT quant_weights) {
+ GGML_ASSERT(quant_weights);
+ assert(k % QK_K == 0);
+ const int nb = k / QK_K;
+ const bool requantize = true;
+
+ uint8_t L[QK_K];
+ uint8_t Laux[16];
+ float mins[QK_K/16];
+ float scales[QK_K/16];
+ float sw[QK_K/16];
+ float weight[16];
+ uint8_t Ls[QK_K/16], Lm[QK_K/16];
+
+ for (int i = 0; i < nb; i++) {
+ memset(sw, 0, QK_K/16*sizeof(float));
+ float sumx2 = 0;
+ for (int j = 0; j < QK_K; ++j) sumx2 += x[j]*x[j];
+ float sigma2 = sumx2/QK_K;
+ for (int j = 0; j < QK_K/16; ++j) {
+ const float * GGML_RESTRICT qw = quant_weights + QK_K * i + 16*j;
+ for (int l = 0; l < 16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j + l]*x[16*j + l]);
+ for (int l = 0; l < QK_K/16; ++l) sw[j] += weight[l];
+ scales[j] = make_qkx3_quants(16, 3, x + 16*j, weight, L + 16*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
+ }
+
+ float dm, mm;
+ dm = make_qp_quants(QK_K/16, 15, scales, Ls, sw);
+ mm = make_qp_quants(QK_K/16, 15, mins, Lm, sw);
+
+ y[i].d = GGML_FP32_TO_FP16(dm);
+ y[i].dmin = GGML_FP32_TO_FP16(mm);
+ dm = GGML_FP16_TO_FP32(y[i].d);
+ mm = GGML_FP16_TO_FP32(y[i].dmin);
+
+ for (int j = 0; j < QK_K/16; ++j) {
+ y[i].scales[j] = Ls[j] | (Lm[j] << 4);
+ }
+
+ if (requantize) {
+ for (int j = 0; j < QK_K/16; ++j) {
+ const float d = dm * (y[i].scales[j] & 0xF);
+ if (!d) continue;
+ const float m = mm * (y[i].scales[j] >> 4);
+ for (int ii = 0; ii < 16; ++ii) {
+ int l = nearest_int((x[16*j + ii] + m)/d);
+ l = MAX(0, MIN(3, l));
+ L[16*j + ii] = l;
+ }
+ }
+ }
+
+ for (int j = 0; j < QK_K; j += 128) {
+ for (int l = 0; l < 32; ++l) {
+ y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
+ }
+ }
+
+ x += QK_K;
+ }
+}
+
+size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ size_t row_size = ggml_row_size(GGML_TYPE_Q2_K, n_per_row);
+ if (!quant_weights) {
+ quantize_row_q2_K_ref(src, dst, (int64_t)nrow*n_per_row);
+ }
+ else {
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_q2_K_impl(src, (block_q2_K*)qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += row_size;
+ }
+ }
+ return nrow * row_size;
+}
+
+//========================= 3-bit (de)-quantization
+
+void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int nb = k / QK_K;
+
+ int8_t L[QK_K];
+ float scales[QK_K / 16];
+
+ for (int i = 0; i < nb; i++) {
+
+ float max_scale = 0;
+ float amax = 0;
+ for (int j = 0; j < QK_K/16; ++j) {
+ scales[j] = make_q3_quants(16, 4, x + 16*j, L + 16*j, true);
+ float scale = fabsf(scales[j]);
+ if (scale > amax) {
+ amax = scale; max_scale = scales[j];
+ }
+ }
+
+ memset(y[i].scales, 0, 12);
+ if (max_scale) {
+ float iscale = -32.f/max_scale;
+ for (int j = 0; j < QK_K/16; ++j) {
+ int8_t l = nearest_int(iscale*scales[j]);
+ l = MAX(-32, MIN(31, l)) + 32;
+ if (j < 8) {
+ y[i].scales[j] = l & 0xF;
+ } else {
+ y[i].scales[j-8] |= ((l & 0xF) << 4);
+ }
+ l >>= 4;
+ y[i].scales[j%4 + 8] |= (l << (2*(j/4)));
+ }
+ y[i].d = GGML_FP32_TO_FP16(1/iscale);
+ } else {
+ y[i].d = GGML_FP32_TO_FP16(0.f);
+ }
+
+ int8_t sc;
+ for (int j = 0; j < QK_K/16; ++j) {
+ sc = j < 8 ? y[i].scales[j] & 0xF : y[i].scales[j-8] >> 4;
+ sc = (sc | (((y[i].scales[8 + j%4] >> (2*(j/4))) & 3) << 4)) - 32;
+ float d = GGML_FP16_TO_FP32(y[i].d) * sc;
+ if (!d) {
+ continue;
+ }
+ for (int ii = 0; ii < 16; ++ii) {
+ int l = nearest_int(x[16*j + ii]/d);
+ l = MAX(-4, MIN(3, l));
+ L[16*j + ii] = l + 4;
+ }
+ }
+
+ memset(y[i].hmask, 0, QK_K/8);
+ // We put the high-bit for the 1st 8 quants into bit 0, the next 8 into bit 1, etc.
+ int m = 0;
+ uint8_t hm = 1;
+ for (int j = 0; j < QK_K; ++j) {
+ if (L[j] > 3) {
+ y[i].hmask[m] |= hm;
+ L[j] -= 4;
+ }
+ if (++m == QK_K/8) {
+ m = 0; hm <<= 1;
+ }
+ }
+ for (int j = 0; j < QK_K; j += 128) {
+ for (int l = 0; l < 32; ++l) {
+ y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
+ }
+ }
+
+ x += QK_K;
+ }
+}
+
+void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int nb = k / QK_K;
+
+ const uint32_t kmask1 = 0x03030303;
+ const uint32_t kmask2 = 0x0f0f0f0f;
+
+ uint32_t aux[4];
+ const int8_t * scales = (const int8_t*)aux;
+
+ for (int i = 0; i < nb; i++) {
+
+ const float d_all = GGML_FP16_TO_FP32(x[i].d);
+
+ const uint8_t * GGML_RESTRICT q = x[i].qs;
+ const uint8_t * GGML_RESTRICT hm = x[i].hmask;
+ uint8_t m = 1;
+
+ memcpy(aux, x[i].scales, 12);
+ uint32_t tmp = aux[2];
+ aux[2] = ((aux[0] >> 4) & kmask2) | (((tmp >> 4) & kmask1) << 4);
+ aux[3] = ((aux[1] >> 4) & kmask2) | (((tmp >> 6) & kmask1) << 4);
+ aux[0] = (aux[0] & kmask2) | (((tmp >> 0) & kmask1) << 4);
+ aux[1] = (aux[1] & kmask2) | (((tmp >> 2) & kmask1) << 4);
+
+ int is = 0;
+ float dl;
+ for (int n = 0; n < QK_K; n += 128) {
+ int shift = 0;
+ for (int j = 0; j < 4; ++j) {
+
+ dl = d_all * (scales[is++] - 32);
+ for (int l = 0; l < 16; ++l) {
+ *y++ = dl * ((int8_t)((q[l+ 0] >> shift) & 3) - ((hm[l+ 0] & m) ? 0 : 4));
+ }
+
+ dl = d_all * (scales[is++] - 32);
+ for (int l = 0; l < 16; ++l) {
+ *y++ = dl * ((int8_t)((q[l+16] >> shift) & 3) - ((hm[l+16] & m) ? 0 : 4));
+ }
+
+ shift += 2;
+ m <<= 1;
+ }
+ q += 32;
+ }
+
+ }
+}
+
+static void quantize_row_q3_K_impl(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t n_per_row, const float * GGML_RESTRICT quant_weights) {
+ assert(n_per_row % QK_K == 0);
+ const int nb = n_per_row / QK_K;
+
+ int8_t L[QK_K];
+ float scales[QK_K / 16];
+ float weight[16];
+ float sw[QK_K / 16];
+ int8_t Ls[QK_K / 16];
+
+ for (int i = 0; i < nb; i++) {
+
+ float sumx2 = 0;
+ for (int j = 0; j < QK_K; ++j) sumx2 += x[j]*x[j];
+ float sigma2 = 2*sumx2/QK_K;
+
+ for (int j = 0; j < QK_K/16; ++j) {
+ if (quant_weights) {
+ const float * qw = quant_weights + QK_K * i + 16*j;
+ for (int l = 0; l < 16; ++l) weight[l] = qw[l] * sqrtf(sigma2 + x[16*j+l]*x[16*j+l]);
+ } else {
+ for (int l = 0; l < 16; ++l) weight[l] = x[16*j+l]*x[16*j+l];
+ }
+ float sumw = 0;
+ for (int l = 0; l < 16; ++l) sumw += weight[l];
+ sw[j] = sumw;
+
+ scales[j] = make_qx_quants(16, 4, x + 16*j, L + 16*j, 1, weight);
+
+ }
+
+ memset(y[i].scales, 0, 12);
+
+ float d_block = make_qx_quants(QK_K/16, 32, scales, Ls, 1, sw);
+ for (int j = 0; j < QK_K/16; ++j) {
+ int l = Ls[j];
+ if (j < 8) {
+ y[i].scales[j] = l & 0xF;
+ } else {
+ y[i].scales[j-8] |= ((l & 0xF) << 4);
+ }
+ l >>= 4;
+ y[i].scales[j%4 + 8] |= (l << (2*(j/4)));
+ }
+ y[i].d = GGML_FP32_TO_FP16(d_block);
+
+ int8_t sc;
+ for (int j = 0; j < QK_K/16; ++j) {
+ sc = j < 8 ? y[i].scales[j] & 0xF : y[i].scales[j-8] >> 4;
+ sc = (sc | (((y[i].scales[8 + j%4] >> (2*(j/4))) & 3) << 4)) - 32;
+ float d = GGML_FP16_TO_FP32(y[i].d) * sc;
+ if (!d) {
+ continue;
+ }
+ for (int ii = 0; ii < 16; ++ii) {
+ int l = nearest_int(x[16*j + ii]/d);
+ l = MAX(-4, MIN(3, l));
+ L[16*j + ii] = l + 4;
+ }
+ }
+
+ memset(y[i].hmask, 0, QK_K/8);
+ // We put the high-bit for the 1st 8 quants into bit 0, the next 8 into bit 1, etc.
+ int m = 0;
+ uint8_t hm = 1;
+ for (int j = 0; j < QK_K; ++j) {
+ if (L[j] > 3) {
+ y[i].hmask[m] |= hm;
+ L[j] -= 4;
+ }
+ if (++m == QK_K/8) {
+ m = 0; hm <<= 1;
+ }
+ }
+ for (int j = 0; j < QK_K; j += 128) {
+ for (int l = 0; l < 32; ++l) {
+ y[i].qs[j/4 + l] = L[j + l] | (L[j + l + 32] << 2) | (L[j + l + 64] << 4) | (L[j + l + 96] << 6);
+ }
+ }
+
+ x += QK_K;
+ }
+}
+
+size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ size_t row_size = ggml_row_size(GGML_TYPE_Q3_K, n_per_row);
+ if (!quant_weights) {
+ quantize_row_q3_K_ref(src, dst, (int64_t)nrow*n_per_row);
+ }
+ else {
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_q3_K_impl(src, (block_q3_K*)qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += row_size;
+ }
+ }
+ return nrow * row_size;
+}
+
+// ====================== 4-bit (de)-quantization
+
+void quantize_row_q4_K_ref(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int nb = k / QK_K;
+
+ uint8_t L[QK_K];
+ uint8_t Laux[32];
+ float weights[32];
+ float mins[QK_K/32];
+ float scales[QK_K/32];
+
+ for (int i = 0; i < nb; i++) {
+ float max_scale = 0; // as we are deducting the min, scales are always positive
+ float max_min = 0;
+ for (int j = 0; j < QK_K/32; ++j) {
+ //scales[j] = make_qkx1_quants(32, 15, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
+ float sum_x2 = 0;
+ for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
+ float av_x = sqrtf(sum_x2/32);
+ for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
+ scales[j] = make_qkx2_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -1.f, 0.1f, 20, false);
+ float scale = scales[j];
+ if (scale > max_scale) {
+ max_scale = scale;
+ }
+ float min = mins[j];
+ if (min > max_min) {
+ max_min = min;
+ }
+ }
+
+ float inv_scale = max_scale > 0 ? 63.f/max_scale : 0.f;
+ float inv_min = max_min > 0 ? 63.f/max_min : 0.f;
+ for (int j = 0; j < QK_K/32; ++j) {
+ uint8_t ls = nearest_int(inv_scale*scales[j]);
+ uint8_t lm = nearest_int(inv_min*mins[j]);
+ ls = MIN(63, ls);
+ lm = MIN(63, lm);
+ if (j < 4) {
+ y[i].scales[j] = ls;
+ y[i].scales[j+4] = lm;
+ } else {
+ y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4);
+ y[i].scales[j-4] |= ((ls >> 4) << 6);
+ y[i].scales[j-0] |= ((lm >> 4) << 6);
+ }
+ }
+ y[i].d = GGML_FP32_TO_FP16(max_scale/63.f);
+ y[i].dmin = GGML_FP32_TO_FP16(max_min/63.f);
+
+ uint8_t sc, m;
+ for (int j = 0; j < QK_K/32; ++j) {
+ get_scale_min_k4(j, y[i].scales, &sc, &m);
+ const float d = GGML_FP16_TO_FP32(y[i].d) * sc;
+ if (!d) continue;
+ const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m;
+ for (int ii = 0; ii < 32; ++ii) {
+ int l = nearest_int((x[32*j + ii] + dm)/d);
+ l = MAX(0, MIN(15, l));
+ L[32*j + ii] = l;
+ }
+ }
+
+ uint8_t * q = y[i].qs;
+ for (int j = 0; j < QK_K; j += 64) {
+ for (int l = 0; l < 32; ++l) q[l] = L[j + l] | (L[j + l + 32] << 4);
+ q += 32;
+ }
+
+ x += QK_K;
+ }
+}
+
+void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int nb = k / QK_K;
+
+ for (int i = 0; i < nb; i++) {
+ const uint8_t * q = x[i].qs;
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+ const float min = GGML_FP16_TO_FP32(x[i].dmin);
+
+ int is = 0;
+ uint8_t sc, m;
+ for (int j = 0; j < QK_K; j += 64) {
+ get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
+ const float d1 = d * sc; const float m1 = min * m;
+ get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
+ const float d2 = d * sc; const float m2 = min * m;
+ for (int l = 0; l < 32; ++l) *y++ = d1 * (q[l] & 0xF) - m1;
+ for (int l = 0; l < 32; ++l) *y++ = d2 * (q[l] >> 4) - m2;
+ q += 32; is += 2;
+ }
+ }
+}
+
+static void quantize_row_q4_K_impl(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
+ assert(n_per_row % QK_K == 0);
+ const int64_t nb = n_per_row / QK_K;
+
+ uint8_t L[QK_K];
+ uint8_t Laux[32];
+ uint8_t Ls[QK_K/32];
+ uint8_t Lm[QK_K/32];
+ float weights[32];
+ float sw[QK_K/32];
+ float mins[QK_K/32];
+ float scales[QK_K/32];
+
+ for (int i = 0; i < nb; i++) {
+
+ float sum_x2 = 0;
+ for (int l = 0; l < QK_K; ++l) sum_x2 += x[l] * x[l];
+ float sigma2 = 2*sum_x2/QK_K;
+ float av_x = sqrtf(sigma2);
+
+ for (int j = 0; j < QK_K/32; ++j) {
+ if (quant_weights) {
+ const float * qw = quant_weights + QK_K*i + 32*j;
+ for (int l = 0; l < 32; ++l) weights[l] = qw[l] * sqrtf(sigma2 + x[32*j + l]*x[32*j + l]);
+ } else {
+ for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
+ }
+ float sumw = 0;
+ for (int l = 0; l < 32; ++l) sumw += weights[l];
+ sw[j] = sumw;
+ scales[j] = make_qkx3_quants(32, 15, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
+ }
+
+ float d_block = make_qp_quants(QK_K/32, 63, scales, Ls, sw);
+ float m_block = make_qp_quants(QK_K/32, 63, mins, Lm, sw);
+ for (int j = 0; j < QK_K/32; ++j) {
+ uint8_t ls = Ls[j];
+ uint8_t lm = Lm[j];
+ if (j < 4) {
+ y[i].scales[j] = ls;
+ y[i].scales[j+4] = lm;
+ } else {
+ y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4);
+ y[i].scales[j-4] |= ((ls >> 4) << 6);
+ y[i].scales[j-0] |= ((lm >> 4) << 6);
+ }
+ }
+ y[i].d = GGML_FP32_TO_FP16(d_block);
+ y[i].dmin = GGML_FP32_TO_FP16(m_block);
+
+ uint8_t sc, m;
+ for (int j = 0; j < QK_K/32; ++j) {
+ get_scale_min_k4(j, y[i].scales, &sc, &m);
+ const float d = GGML_FP16_TO_FP32(y[i].d) * sc;
+ if (!d) continue;
+ const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m;
+ for (int ii = 0; ii < 32; ++ii) {
+ int l = nearest_int((x[32*j + ii] + dm)/d);
+ l = MAX(0, MIN(15, l));
+ L[32*j + ii] = l;
+ }
+ }
+ uint8_t * q = y[i].qs;
+ for (int j = 0; j < QK_K; j += 64) {
+ for (int l = 0; l < 32; ++l) q[l] = L[j + l] | (L[j + l + 32] << 4);
+ q += 32;
+ }
+
+ x += QK_K;
+
+ }
+}
+
+size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ size_t row_size = ggml_row_size(GGML_TYPE_Q4_K, n_per_row);
+ if (!quant_weights) {
+ quantize_row_q4_K_ref(src, dst, (int64_t)nrow*n_per_row);
+ }
+ else {
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_q4_K_impl(src, (block_q4_K*)qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += row_size;
+ }
+ }
+ return nrow * row_size;
+}
+
+// ====================== 5-bit (de)-quantization
+
+void quantize_row_q5_K_ref(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ uint8_t L[QK_K];
+ float mins[QK_K/32];
+ float scales[QK_K/32];
+ float weights[32];
+ uint8_t Laux[32];
+
+ for (int i = 0; i < nb; i++) {
+ float max_scale = 0; // as we are deducting the min, scales are always positive
+ float max_min = 0;
+ for (int j = 0; j < QK_K/32; ++j) {
+ //scales[j] = make_qkx1_quants(32, 31, x + 32*j, L + 32*j, &mins[j], 9, 0.5f);
+ float sum_x2 = 0;
+ for (int l = 0; l < 32; ++l) sum_x2 += x[32*j + l] * x[32*j + l];
+ float av_x = sqrtf(sum_x2/32);
+ for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
+ scales[j] = make_qkx2_quants(32, 31, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.5f, 0.1f, 15, false);
+ float scale = scales[j];
+ if (scale > max_scale) {
+ max_scale = scale;
+ }
+ float min = mins[j];
+ if (min > max_min) {
+ max_min = min;
+ }
+ }
+
+ float inv_scale = max_scale > 0 ? 63.f/max_scale : 0.f;
+ float inv_min = max_min > 0 ? 63.f/max_min : 0.f;
+ for (int j = 0; j < QK_K/32; ++j) {
+ uint8_t ls = nearest_int(inv_scale*scales[j]);
+ uint8_t lm = nearest_int(inv_min*mins[j]);
+ ls = MIN(63, ls);
+ lm = MIN(63, lm);
+ if (j < 4) {
+ y[i].scales[j] = ls;
+ y[i].scales[j+4] = lm;
+ } else {
+ y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4);
+ y[i].scales[j-4] |= ((ls >> 4) << 6);
+ y[i].scales[j-0] |= ((lm >> 4) << 6);
+ }
+ }
+ y[i].d = GGML_FP32_TO_FP16(max_scale/63.f);
+ y[i].dmin = GGML_FP32_TO_FP16(max_min/63.f);
+
+ uint8_t sc, m;
+ for (int j = 0; j < QK_K/32; ++j) {
+ get_scale_min_k4(j, y[i].scales, &sc, &m);
+ const float d = GGML_FP16_TO_FP32(y[i].d) * sc;
+ if (!d) continue;
+ const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m;
+ for (int ii = 0; ii < 32; ++ii) {
+ int l = nearest_int((x[32*j + ii] + dm)/d);
+ l = MAX(0, MIN(31, l));
+ L[32*j + ii] = l;
+ }
+ }
+
+ uint8_t * GGML_RESTRICT qh = y[i].qh;
+ uint8_t * GGML_RESTRICT ql = y[i].qs;
+ memset(qh, 0, QK_K/8);
+
+ uint8_t m1 = 1, m2 = 2;
+ for (int n = 0; n < QK_K; n += 64) {
+ for (int j = 0; j < 32; ++j) {
+ int l1 = L[n + j];
+ if (l1 > 15) {
+ l1 -= 16; qh[j] |= m1;
+ }
+ int l2 = L[n + j + 32];
+ if (l2 > 15) {
+ l2 -= 16; qh[j] |= m2;
+ }
+ ql[j] = l1 | (l2 << 4);
+ }
+ m1 <<= 2; m2 <<= 2;
+ ql += 32;
+ }
+
+ x += QK_K;
+ }
+}
+
+void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ for (int i = 0; i < nb; i++) {
+ const uint8_t * ql = x[i].qs;
+ const uint8_t * qh = x[i].qh;
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+ const float min = GGML_FP16_TO_FP32(x[i].dmin);
+
+ int is = 0;
+ uint8_t sc, m;
+ uint8_t u1 = 1, u2 = 2;
+ for (int j = 0; j < QK_K; j += 64) {
+ get_scale_min_k4(is + 0, x[i].scales, &sc, &m);
+ const float d1 = d * sc; const float m1 = min * m;
+ get_scale_min_k4(is + 1, x[i].scales, &sc, &m);
+ const float d2 = d * sc; const float m2 = min * m;
+ for (int l = 0; l < 32; ++l) *y++ = d1 * ((ql[l] & 0xF) + (qh[l] & u1 ? 16 : 0)) - m1;
+ for (int l = 0; l < 32; ++l) *y++ = d2 * ((ql[l] >> 4) + (qh[l] & u2 ? 16 : 0)) - m2;
+ ql += 32; is += 2;
+ u1 <<= 2; u2 <<= 2;
+ }
+ }
+}
+
+static void quantize_row_q5_K_impl(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
+ assert(n_per_row % QK_K == 0);
+ const int64_t nb = n_per_row / QK_K;
+
+ uint8_t L[QK_K];
+ uint8_t Laux[32];
+ uint8_t Ls[QK_K/32];
+ uint8_t Lm[QK_K/32];
+ float mins[QK_K/32];
+ float scales[QK_K/32];
+ float sw[QK_K/32];
+ float weights[32];
+
+ for (int i = 0; i < nb; i++) {
+
+ float sum_x2 = 0;
+ for (int l = 0; l < QK_K; ++l) sum_x2 += x[l] * x[l];
+ float sigma2 = 2*sum_x2/QK_K;
+ float av_x = sqrtf(sigma2);
+
+ for (int j = 0; j < QK_K/32; ++j) {
+ if (quant_weights) {
+ const float * qw = quant_weights + QK_K*i + 32*j;
+ for (int l = 0; l < 32; ++l) weights[l] = qw[l] * sqrtf(sigma2 + x[32*j + l]*x[32*j + l]);
+ } else {
+ for (int l = 0; l < 32; ++l) weights[l] = av_x + fabsf(x[32*j + l]);
+ }
+ float sumw = 0;
+ for (int l = 0; l < 32; ++l) sumw += weights[l];
+ sw[j] = sumw;
+
+ scales[j] = make_qkx3_quants(32, 31, x + 32*j, weights, L + 32*j, &mins[j], Laux, -0.9f, 0.05f, 36, false);
+ }
+
+ float d_block = make_qp_quants(QK_K/32, 63, scales, Ls, sw);
+ float m_block = make_qp_quants(QK_K/32, 63, mins, Lm, sw);
+
+ for (int j = 0; j < QK_K/32; ++j) {
+ uint8_t ls = Ls[j];
+ uint8_t lm = Lm[j];
+ ls = MIN(63, ls);
+ lm = MIN(63, lm);
+ if (j < 4) {
+ y[i].scales[j] = ls;
+ y[i].scales[j+4] = lm;
+ } else {
+ y[i].scales[j+4] = (ls & 0xF) | ((lm & 0xF) << 4);
+ y[i].scales[j-4] |= ((ls >> 4) << 6);
+ y[i].scales[j-0] |= ((lm >> 4) << 6);
+ }
+ }
+ y[i].d = GGML_FP32_TO_FP16(d_block);
+ y[i].dmin = GGML_FP32_TO_FP16(m_block);
+
+ uint8_t sc, m;
+ for (int j = 0; j < QK_K/32; ++j) {
+ get_scale_min_k4(j, y[i].scales, &sc, &m);
+ const float d = GGML_FP16_TO_FP32(y[i].d) * sc;
+ if (!d) continue;
+ const float dm = GGML_FP16_TO_FP32(y[i].dmin) * m;
+ for (int ii = 0; ii < 32; ++ii) {
+ int l = nearest_int((x[32*j + ii] + dm)/d);
+ l = MAX(0, MIN(31, l));
+ L[32*j + ii] = l;
+ }
+ }
+
+ uint8_t * GGML_RESTRICT qh = y[i].qh;
+ uint8_t * GGML_RESTRICT ql = y[i].qs;
+ memset(qh, 0, QK_K/8);
+
+ uint8_t m1 = 1, m2 = 2;
+ for (int n = 0; n < QK_K; n += 64) {
+ for (int j = 0; j < 32; ++j) {
+ int l1 = L[n + j];
+ if (l1 > 15) {
+ l1 -= 16; qh[j] |= m1;
+ }
+ int l2 = L[n + j + 32];
+ if (l2 > 15) {
+ l2 -= 16; qh[j] |= m2;
+ }
+ ql[j] = l1 | (l2 << 4);
+ }
+ m1 <<= 2; m2 <<= 2;
+ ql += 32;
+ }
+
+ x += QK_K;
+
+ }
+}
+
+size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ size_t row_size = ggml_row_size(GGML_TYPE_Q5_K, n_per_row);
+ if (!quant_weights) {
+ quantize_row_q5_K_ref(src, dst, (int64_t)nrow*n_per_row);
+ }
+ else {
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_q5_K_impl(src, (block_q5_K*)qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += row_size;
+ }
+ }
+ return nrow * row_size;
+}
+
+// ====================== 6-bit (de)-quantization
+
+void quantize_row_q6_K_ref(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ int8_t L[QK_K];
+ float scales[QK_K/16];
+
+ for (int i = 0; i < nb; i++) {
+
+ float max_scale = 0;
+ float max_abs_scale = 0;
+
+ for (int ib = 0; ib < QK_K/16; ++ib) {
+
+ const float scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1, NULL);
+ scales[ib] = scale;
+
+ const float abs_scale = fabsf(scale);
+ if (abs_scale > max_abs_scale) {
+ max_abs_scale = abs_scale;
+ max_scale = scale;
+ }
+
+ }
+
+ if (max_abs_scale < GROUP_MAX_EPS) {
+ memset(&y[i], 0, sizeof(block_q6_K));
+ y[i].d = GGML_FP32_TO_FP16(0.f);
+ x += QK_K;
+ continue;
+ }
+
+ float iscale = -128.f/max_scale;
+ y[i].d = GGML_FP32_TO_FP16(1/iscale);
+ for (int ib = 0; ib < QK_K/16; ++ib) {
+ y[i].scales[ib] = MIN(127, nearest_int(iscale*scales[ib]));
+ }
+
+ for (int j = 0; j < QK_K/16; ++j) {
+ float d = GGML_FP16_TO_FP32(y[i].d) * y[i].scales[j];
+ if (!d) {
+ continue;
+ }
+ for (int ii = 0; ii < 16; ++ii) {
+ int l = nearest_int(x[16*j + ii]/d);
+ l = MAX(-32, MIN(31, l));
+ L[16*j + ii] = l + 32;
+ }
+ }
+
+ uint8_t * GGML_RESTRICT ql = y[i].ql;
+ uint8_t * GGML_RESTRICT qh = y[i].qh;
+ for (int j = 0; j < QK_K; j += 128) {
+ for (int l = 0; l < 32; ++l) {
+ const uint8_t q1 = L[j + l + 0] & 0xF;
+ const uint8_t q2 = L[j + l + 32] & 0xF;
+ const uint8_t q3 = L[j + l + 64] & 0xF;
+ const uint8_t q4 = L[j + l + 96] & 0xF;
+ ql[l+ 0] = q1 | (q3 << 4);
+ ql[l+32] = q2 | (q4 << 4);
+ qh[l] = (L[j + l] >> 4) | ((L[j + l + 32] >> 4) << 2) | ((L[j + l + 64] >> 4) << 4) | ((L[j + l + 96] >> 4) << 6);
+ }
+ ql += 64;
+ qh += 32;
+ }
+
+ x += QK_K;
+ }
+}
+
+void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ for (int i = 0; i < nb; i++) {
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+
+ const uint8_t * GGML_RESTRICT ql = x[i].ql;
+ const uint8_t * GGML_RESTRICT qh = x[i].qh;
+ const int8_t * GGML_RESTRICT sc = x[i].scales;
+
+ for (int n = 0; n < QK_K; n += 128) {
+ for (int l = 0; l < 32; ++l) {
+ int is = l/16;
+ const int8_t q1 = (int8_t)((ql[l + 0] & 0xF) | (((qh[l] >> 0) & 3) << 4)) - 32;
+ const int8_t q2 = (int8_t)((ql[l + 32] & 0xF) | (((qh[l] >> 2) & 3) << 4)) - 32;
+ const int8_t q3 = (int8_t)((ql[l + 0] >> 4) | (((qh[l] >> 4) & 3) << 4)) - 32;
+ const int8_t q4 = (int8_t)((ql[l + 32] >> 4) | (((qh[l] >> 6) & 3) << 4)) - 32;
+ y[l + 0] = d * sc[is + 0] * q1;
+ y[l + 32] = d * sc[is + 2] * q2;
+ y[l + 64] = d * sc[is + 4] * q3;
+ y[l + 96] = d * sc[is + 6] * q4;
+ }
+ y += 128;
+ ql += 64;
+ qh += 32;
+ sc += 8;
+ }
+ }
+}
+
+static void quantize_row_q6_K_impl(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
+ assert(n_per_row % QK_K == 0);
+ const int64_t nb = n_per_row / QK_K;
+
+ int8_t L[QK_K];
+ float scales[QK_K/16];
+ //float weights[16];
+
+ for (int i = 0; i < nb; i++) {
+
+ //float sum_x2 = 0;
+ //for (int j = 0; j < QK_K; ++j) sum_x2 += x[j]*x[j];
+ //float sigma2 = sum_x2/QK_K;
+
+ float max_scale = 0;
+ float max_abs_scale = 0;
+
+ for (int ib = 0; ib < QK_K/16; ++ib) {
+
+ float scale;
+ if (quant_weights) {
+ const float * qw = quant_weights + QK_K*i + 16*ib;
+ //for (int j = 0; j < 16; ++j) weights[j] = qw[j] * sqrtf(sigma2 + x[16*ib + j]*x[16*ib + j]);
+ //scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1, weights);
+ scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1, qw);
+ } else {
+ scale = make_qx_quants(16, 32, x + 16*ib, L + 16*ib, 1, NULL);
+ }
+ scales[ib] = scale;
+
+ const float abs_scale = fabsf(scale);
+ if (abs_scale > max_abs_scale) {
+ max_abs_scale = abs_scale;
+ max_scale = scale;
+ }
+
+ }
+
+ if (max_abs_scale < GROUP_MAX_EPS) {
+ memset(&y[i], 0, sizeof(block_q6_K));
+ y[i].d = GGML_FP32_TO_FP16(0.f);
+ x += QK_K;
+ continue;
+ }
+
+ float iscale = -128.f/max_scale;
+ y[i].d = GGML_FP32_TO_FP16(1/iscale);
+ for (int ib = 0; ib < QK_K/16; ++ib) {
+ y[i].scales[ib] = MIN(127, nearest_int(iscale*scales[ib]));
+ }
+
+ for (int j = 0; j < QK_K/16; ++j) {
+ float d = GGML_FP16_TO_FP32(y[i].d) * y[i].scales[j];
+ if (!d) {
+ continue;
+ }
+ for (int ii = 0; ii < 16; ++ii) {
+ int l = nearest_int(x[16*j + ii]/d);
+ l = MAX(-32, MIN(31, l));
+ L[16*j + ii] = l + 32;
+ }
+ }
+
+ uint8_t * GGML_RESTRICT ql = y[i].ql;
+ uint8_t * GGML_RESTRICT qh = y[i].qh;
+ for (int j = 0; j < QK_K; j += 128) {
+ for (int l = 0; l < 32; ++l) {
+ const uint8_t q1 = L[j + l + 0] & 0xF;
+ const uint8_t q2 = L[j + l + 32] & 0xF;
+ const uint8_t q3 = L[j + l + 64] & 0xF;
+ const uint8_t q4 = L[j + l + 96] & 0xF;
+ ql[l+ 0] = q1 | (q3 << 4);
+ ql[l+32] = q2 | (q4 << 4);
+ qh[l] = (L[j + l] >> 4) | ((L[j + l + 32] >> 4) << 2) | ((L[j + l + 64] >> 4) << 4) | ((L[j + l + 96] >> 4) << 6);
+ }
+ ql += 64;
+ qh += 32;
+ }
+
+ x += QK_K;
+
+ }
+}
+
+size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ size_t row_size = ggml_row_size(GGML_TYPE_Q6_K, n_per_row);
+ if (!quant_weights) {
+ quantize_row_q6_K_ref(src, dst, (int64_t)nrow*n_per_row);
+ }
+ else {
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_q6_K_impl(src, (block_q6_K*)qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += row_size;
+ }
+ }
+ return nrow * row_size;
+}
+
+static void quantize_row_q4_0_impl(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
+ static_assert(QK4_0 == 32, "QK4_0 must be 32");
+
+ if (!quant_weights) {
+ quantize_row_q4_0_ref(x, y, n_per_row);
+ return;
+ }
+
+ float weight[QK4_0];
+ int8_t L[QK4_0];
+
+ float sum_x2 = 0;
+ for (int j = 0; j < n_per_row; ++j) sum_x2 += x[j]*x[j];
+ float sigma2 = sum_x2/n_per_row;
+
+ const int64_t nb = n_per_row/QK4_0;
+ for (int ib = 0; ib < nb; ++ib) {
+ const float * xb = x + QK4_0 * ib;
+ const float * qw = quant_weights + QK4_0 * ib;
+ for (int j = 0; j < QK4_0; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
+ float d = make_qx_quants(QK4_0, 8, xb, L, 1, weight);
+ y[ib].d = GGML_FP32_TO_FP16(d);
+ for (int j = 0; j < 16; ++j) {
+ y[ib].qs[j] = L[j] | (L[j+16] << 4);
+ }
+ }
+}
+
+size_t quantize_q1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ if (!quant_weights) {
+ quantize_row_q1_0_ref(src, dst, (int64_t)nrow*n_per_row);
+ return nrow * ggml_row_size(GGML_TYPE_Q1_0, n_per_row);
+ }
+ size_t row_size = ggml_row_size(GGML_TYPE_Q1_0, n_per_row);
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_q1_0_ref(src, (block_q1_0*)qrow, n_per_row);
+ src += n_per_row;
+ qrow += row_size;
+ }
+ return nrow * row_size;
+}
+
+size_t quantize_q2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ if (!quant_weights) {
+ quantize_row_q2_0_ref(src, dst, (int64_t)nrow*n_per_row);
+ return nrow * ggml_row_size(GGML_TYPE_Q2_0, n_per_row);
+ }
+ size_t row_size = ggml_row_size(GGML_TYPE_Q2_0, n_per_row);
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_q2_0_ref(src, (block_q2_0*)qrow, n_per_row);
+ src += n_per_row;
+ qrow += row_size;
+ }
+ return nrow * row_size;
+}
+
+size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ if (!quant_weights) {
+ quantize_row_q4_0_ref(src, dst, (int64_t)nrow*n_per_row);
+ return nrow * ggml_row_size(GGML_TYPE_Q4_0, n_per_row);
+ }
+ size_t row_size = ggml_row_size(GGML_TYPE_Q4_0, n_per_row);
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_q4_0_impl(src, (block_q4_0*)qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += row_size;
+ }
+ return nrow * row_size;
+}
+
+static void quantize_row_q4_1_impl(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
+ static_assert(QK4_1 == 32, "QK4_1 must be 32");
+
+ if (!quant_weights) {
+ quantize_row_q4_1_ref(x, y, n_per_row);
+ return;
+ }
+
+ float weight[QK4_1];
+ uint8_t L[QK4_1], Laux[QK4_1];
+
+ float sum_x2 = 0;
+ for (int j = 0; j < n_per_row; ++j) sum_x2 += x[j]*x[j];
+ float sigma2 = sum_x2/n_per_row;
+
+ const int64_t nb = n_per_row/QK4_1;
+ for (int ib = 0; ib < nb; ++ib) {
+ const float * xb = x + QK4_1 * ib;
+ const float * qw = quant_weights + QK4_1 * ib;
+ for (int j = 0; j < QK4_1; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
+ float min;
+ float d = make_qkx3_quants(QK4_1, 15, xb, weight, L, &min, Laux, -0.9f, 0.05f, 36, false);
+ y[ib].d = GGML_FP32_TO_FP16(d);
+ y[ib].m = GGML_FP32_TO_FP16(-min);
+ for (int j = 0; j < 16; ++j) {
+ y[ib].qs[j] = L[j] | (L[j+16] << 4);
+ }
+ }
+}
+
+size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ if (!quant_weights) {
+ quantize_row_q4_1_ref(src, dst, (int64_t)nrow*n_per_row);
+ return nrow * ggml_row_size(GGML_TYPE_Q4_1, n_per_row);
+ }
+ size_t row_size = ggml_row_size(GGML_TYPE_Q4_1, n_per_row);
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_q4_1_impl(src, (block_q4_1*)qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += row_size;
+ }
+ return nrow * row_size;
+}
+
+static void quantize_row_q5_0_impl(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
+ static_assert(QK5_0 == 32, "QK5_0 must be 32");
+
+ if (!quant_weights) {
+ quantize_row_q5_0_ref(x, y, n_per_row);
+ return;
+ }
+
+ float weight[QK5_0];
+ int8_t L[QK5_0];
+
+ float sum_x2 = 0;
+ for (int j = 0; j < n_per_row; ++j) sum_x2 += x[j]*x[j];
+ float sigma2 = sum_x2/n_per_row;
+
+ const int64_t nb = n_per_row/QK5_0;
+ for (int ib = 0; ib < nb; ++ib) {
+ const float * xb = x + QK5_0 * ib;
+ const float * qw = quant_weights + QK5_0 * ib;
+ for (int j = 0; j < QK5_0; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
+ float d = make_qx_quants(QK5_0, 16, xb, L, 1, weight);
+ y[ib].d = GGML_FP32_TO_FP16(d);
+
+ uint32_t qh = 0;
+
+ for (int j = 0; j < 16; ++j) {
+ const uint8_t xi0 = L[j];
+ const uint8_t xi1 = L[j+16];
+ y[ib].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
+
+ // get the 5-th bit and store it in qh at the right position
+ qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+ qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+ }
+
+ memcpy(&y[ib].qh, &qh, sizeof(qh));
+ }
+}
+
+size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ if (!quant_weights) {
+ quantize_row_q5_0_ref(src, dst, (int64_t)nrow*n_per_row);
+ return nrow * ggml_row_size(GGML_TYPE_Q5_0, n_per_row);
+ }
+ size_t row_size = ggml_row_size(GGML_TYPE_Q5_0, n_per_row);
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_q5_0_impl(src, (block_q5_0*)qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += row_size;
+ }
+ return nrow * row_size;
+}
+
+static void quantize_row_q5_1_impl(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t n_per_row, const float * quant_weights) {
+ static_assert(QK5_1 == 32, "QK5_1 must be 32");
+
+ if (!quant_weights) {
+ quantize_row_q5_1_ref(x, y, n_per_row);
+ return;
+ }
+
+ float weight[QK5_1];
+ uint8_t L[QK5_1], Laux[QK5_1];
+
+ float sum_x2 = 0;
+ for (int j = 0; j < n_per_row; ++j) sum_x2 += x[j]*x[j];
+ float sigma2 = sum_x2/n_per_row;
+
+ const int64_t nb = n_per_row/QK5_1;
+ for (int ib = 0; ib < nb; ++ib) {
+ const float * xb = x + QK5_1 * ib;
+ const float * qw = quant_weights + QK5_1 * ib;
+ for (int j = 0; j < QK5_1; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
+ float min;
+ float d = make_qkx3_quants(QK5_1, 31, xb, weight, L, &min, Laux, -0.9f, 0.05f, 36, false);
+ y[ib].d = GGML_FP32_TO_FP16(d);
+ y[ib].m = GGML_FP32_TO_FP16(-min);
+
+ uint32_t qh = 0;
+ for (int j = 0; j < 16; ++j) {
+ const uint8_t xi0 = L[j];
+ const uint8_t xi1 = L[j+16];
+ y[ib].qs[j] = (xi0 & 0x0F) | ((xi1 & 0x0F) << 4);
+ // get the 5-th bit and store it in qh at the right position
+ qh |= ((xi0 & 0x10u) >> 4) << (j + 0);
+ qh |= ((xi1 & 0x10u) >> 4) << (j + QK5_0/2);
+ }
+ memcpy(&y[ib].qh, &qh, sizeof(qh));
+ }
+}
+
+size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ if (!quant_weights) {
+ quantize_row_q5_1_ref(src, dst, (int64_t)nrow*n_per_row);
+ return nrow * ggml_row_size(GGML_TYPE_Q5_1, n_per_row);
+ }
+ size_t row_size = ggml_row_size(GGML_TYPE_Q5_1, n_per_row);
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_q5_1_impl(src, (block_q5_1*)qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += row_size;
+ }
+ return nrow * row_size;
+}
+
+size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ (void)quant_weights; // not used
+ const size_t row_size = ggml_row_size(GGML_TYPE_Q8_0, n_per_row);
+ quantize_row_q8_0_ref(src, dst, (int64_t)nrow*n_per_row);
+ return nrow * row_size;
+}
+
+size_t quantize_mxfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ GGML_UNUSED(quant_weights);
+ quantize_row_mxfp4_ref(src, dst, (int64_t)nrow*n_per_row);
+ return nrow * ggml_row_size(GGML_TYPE_MXFP4, n_per_row);
+}
+
+size_t quantize_nvfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ GGML_UNUSED(quant_weights);
+ quantize_row_nvfp4_ref(src, dst, (int64_t)nrow*n_per_row);
+ return nrow * ggml_row_size(GGML_TYPE_NVFP4, n_per_row);
+}
+
+// ====================== Ternary (de)-quantization (BitNet b1.58 and TriLMs)
+
+void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ for (int64_t i = 0; i < nb; i++) {
+ float amax = 0.0f; // absolute max
+
+ for (int j = 0; j < QK_K; j++) {
+ const float v = x[j];
+ amax = MAX(amax, fabsf(v));
+ }
+
+ const float d = amax;
+ const float id = d ? 1.0f/d : 0.0f;
+
+ y[i].d = GGML_FP32_TO_FP16(d);
+
+ // 5 elements per byte, along 32 bytes
+ for (size_t j = 0; j < sizeof(y->qs) - sizeof(y->qs) % 32; j += 32) {
+ for (size_t m = 0; m < 32; ++m) {
+ uint8_t q = 0;
+ for (size_t n = 0; n < 5; ++n) {
+ int xi = lroundf(x[m + n*32] * id) + 1; // -1, 0, 1 -> 0, 1, 2
+ q *= 3;
+ q += xi;
+ }
+ // ceiling division (243 == pow(3, 5))
+ q = ((uint16_t)q * 256 + (243 - 1)) / 243;
+ y[i].qs[j + m] = q;
+ }
+ x += 5*32;
+ }
+ // along 16 bytes
+ for (size_t j = sizeof(y->qs) - sizeof(y->qs) % 32; j < sizeof(y->qs); j += 16) {
+ for (size_t m = 0; m < 16; ++m) {
+ uint8_t q = 0;
+ for (size_t n = 0; n < 5; ++n) {
+ int xi = lroundf(x[m + n*16] * id) + 1; // -1, 0, 1 -> 0, 1, 2
+ q *= 3;
+ q += xi;
+ }
+ // ceiling division (243 == pow(3, 5))
+ q = ((uint16_t)q * 256 + (243 - 1)) / 243;
+ y[i].qs[j + m] = q;
+ }
+ x += 5*16;
+ }
+ // 4 elements per byte
+ for (size_t j = 0; j < sizeof(y->qh); ++j) {
+ uint8_t q = 0;
+ for (size_t m = 0; m < 4; ++m) {
+ // -1, 0, 1 -> 0, 1, 2
+ int xi = lroundf(x[j + m*sizeof(y->qh)] * id) + 1;
+ q *= 3;
+ q += xi;
+ }
+ // shift the first value to the most significant trit
+ q *= 3;
+ // ceiling division (243 == pow(3, 5))
+ q = ((uint16_t)q * 256 + (243 - 1)) / 243;
+ y[i].qh[j] = q;
+ }
+ x += 4*sizeof(y->qh);
+ }
+}
+
+void quantize_row_tq2_0_ref(const float * GGML_RESTRICT x, block_tq2_0 * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ for (int64_t i = 0; i < nb; i++) {
+ float amax = 0.0f; // absolute max
+
+ for (int j = 0; j < QK_K; j++) {
+ const float v = x[j];
+ amax = MAX(amax, fabsf(v));
+ }
+
+ const float d = amax;
+ const float id = d ? 1.0f/d : 0.0f;
+
+ y[i].d = GGML_FP32_TO_FP16(d);
+
+ for (size_t j = 0; j < sizeof(y->qs); j += 32) {
+ for (size_t m = 0; m < 32; ++m) {
+ uint8_t q = 0;
+ for (size_t n = 0; n < 4; ++n) {
+ // -1, 0, 1 -> 0, 1, 2
+ int xi = lroundf(x[m + n*32] * id) + 1;
+ q += (xi & 3) << (2*n);
+ }
+ y[i].qs[j + m] = q;
+ }
+ x += 4*32;
+ }
+ }
+}
+
+size_t quantize_tq1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ (void)quant_weights; // not used
+ const size_t row_size = ggml_row_size(GGML_TYPE_TQ1_0, n_per_row);
+ quantize_row_tq1_0_ref(src, dst, (int64_t)nrow*n_per_row);
+ return nrow * row_size;
+}
+
+size_t quantize_tq2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ (void)quant_weights; // not used
+ const size_t row_size = ggml_row_size(GGML_TYPE_TQ2_0, n_per_row);
+ quantize_row_tq2_0_ref(src, dst, (int64_t)nrow*n_per_row);
+ return nrow * row_size;
+}
+
+void dequantize_row_tq1_0(const block_tq1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ const uint8_t pow3[6] = {1, 3, 9, 27, 81, 243};
+
+ for (int64_t i = 0; i < nb; ++i) {
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+
+ for (size_t j = 0; j < sizeof(x->qs) - sizeof(x->qs) % 32; j += 32) {
+ for (size_t n = 0; n < 5; ++n) {
+ for (size_t m = 0; m < 32; ++m) {
+ uint8_t q = x[i].qs[j + m] * pow3[n];
+ int16_t xi = ((uint16_t) q * 3) >> 8;
+ *y++ = (float) (xi - 1) * d;
+ }
+ }
+ }
+ for (size_t j = sizeof(x->qs) - sizeof(x->qs) % 32; j < sizeof(x->qs); j += 16) {
+ for (size_t n = 0; n < 5; ++n) {
+ for (size_t m = 0; m < 16; ++m) {
+ uint8_t q = x[i].qs[j + m] * pow3[n];
+ int16_t xi = ((uint16_t) q * 3) >> 8;
+ *y++ = (float) (xi - 1) * d;
+ }
+ }
+ }
+
+ for (size_t n = 0; n < 4; ++n) {
+ for (size_t j = 0; j < sizeof(x->qh); ++j) {
+ uint8_t q = x[i].qh[j] * pow3[n];
+ int16_t xi = ((uint16_t) q * 3) >> 8;
+ *y++ = (float) (xi - 1) * d;
+ }
+ }
+ }
+}
+
+void dequantize_row_tq2_0(const block_tq2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ for (int64_t i = 0; i < nb; ++i) {
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+
+ for (size_t j = 0; j < sizeof(x->qs); j += 32) {
+ for (size_t l = 0; l < 4; ++l) {
+ for (size_t m = 0; m < 32; ++m) {
+ int8_t q = (x[i].qs[j + m] >> (l*2)) & 3;
+ *y++ = (float) (q - 1) * d;
+ }
+ }
+ }
+ }
+}
+
+// ====================== "True" 2-bit (de)-quantization
+
+void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ uint32_t aux32[2];
+ const uint8_t * aux8 = (const uint8_t *)aux32;
+
+ for (int i = 0; i < nb; i++) {
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+
+ for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+ memcpy(aux32, x[i].qs + 4*ib32, 2*sizeof(uint32_t));
+ const float db = d * (0.5f + (aux32[1] >> 28)) * 0.25f;
+ for (int l = 0; l < 4; ++l) {
+ const uint8_t * grid = (const uint8_t *)(iq2xxs_grid + aux8[l]);
+ const uint8_t signs = ksigns_iq2xs[(aux32[1] >> 7*l) & 127];
+ for (int j = 0; j < 8; ++j) {
+ y[j] = db * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+ }
+ y += 8;
+ }
+ }
+ }
+}
+
+// ====================== 2.3125 bpw (de)-quantization
+
+void dequantize_row_iq2_xs(const block_iq2_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ float db[2];
+
+ for (int i = 0; i < nb; i++) {
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+
+ for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+ db[0] = d * (0.5f + (x[i].scales[ib32] & 0xf)) * 0.25f;
+ db[1] = d * (0.5f + (x[i].scales[ib32] >> 4)) * 0.25f;
+ for (int l = 0; l < 4; ++l) {
+ const uint8_t * grid = (const uint8_t *)(iq2xs_grid + (x[i].qs[4*ib32 + l] & 511));
+ const uint8_t signs = ksigns_iq2xs[x[i].qs[4*ib32 + l] >> 9];
+ for (int j = 0; j < 8; ++j) {
+ y[j] = db[l/2] * grid[j] * (signs & kmask_iq2xs[j] ? -1.f : 1.f);
+ }
+ y += 8;
+ }
+ }
+ }
+}
+
+// ====================== 2.5625 bpw (de)-quantization
+
+void dequantize_row_iq2_s(const block_iq2_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ float db[2];
+
+ for (int i = 0; i < nb; i++) {
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+ const uint8_t * qs = x[i].qs;
+ const uint8_t * qh = x[i].qh;
+ const uint8_t * signs = qs + QK_K/8;
+
+ for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+ db[0] = d * (0.5f + (x[i].scales[ib32] & 0xf)) * 0.25f;
+ db[1] = d * (0.5f + (x[i].scales[ib32] >> 4)) * 0.25f;
+ for (int l = 0; l < 4; ++l) {
+ const float dl = db[l/2];
+ const uint8_t * grid = (const uint8_t *)(iq2s_grid + (qs[l] | (qh[ib32] << (8-2*l) & 0x300)));
+ for (int j = 0; j < 8; ++j) {
+ y[j] = dl * grid[j] * (signs[l] & kmask_iq2xs[j] ? -1.f : 1.f);
+ }
+ y += 8;
+ }
+ qs += 4;
+ signs += 4;
+ }
+ }
+}
+
+// ====================== 3.0625 bpw (de)-quantization
+
+void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ uint32_t aux32;
+
+ for (int i = 0; i < nb; i++) {
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+ const uint8_t * qs = x[i].qs;
+ const uint8_t * scales_and_signs = qs + QK_K/4;
+
+ for (int ib32 = 0; ib32 < QK_K/32; ++ib32) {
+ memcpy(&aux32, scales_and_signs + 4*ib32, sizeof(uint32_t));
+ const float db = d * (0.5f + (aux32 >> 28)) * 0.5f;
+ for (int l = 0; l < 4; ++l) {
+ const uint8_t signs = ksigns_iq2xs[(aux32 >> 7*l) & 127];
+ const uint8_t * grid1 = (const uint8_t *)(iq3xxs_grid + qs[2*l+0]);
+ const uint8_t * grid2 = (const uint8_t *)(iq3xxs_grid + qs[2*l+1]);
+ for (int j = 0; j < 4; ++j) {
+ y[j+0] = db * grid1[j] * (signs & kmask_iq2xs[j+0] ? -1.f : 1.f);
+ y[j+4] = db * grid2[j] * (signs & kmask_iq2xs[j+4] ? -1.f : 1.f);
+ }
+ y += 8;
+ }
+ qs += 8;
+ }
+ }
+}
+
+// ====================== 3.3125 bpw (de)-quantization
+
+void dequantize_row_iq3_s(const block_iq3_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ for (int i = 0; i < nb; i++) {
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+ const uint8_t * qs = x[i].qs;
+ const uint8_t * qh = x[i].qh;
+ const uint8_t * signs = x[i].signs;
+
+ for (int ib32 = 0; ib32 < QK_K/32; ib32 += 2) {
+ const float db1 = d * (1 + 2*(x[i].scales[ib32/2] & 0xf));
+ const float db2 = d * (1 + 2*(x[i].scales[ib32/2] >> 4));
+ for (int l = 0; l < 4; ++l) {
+ const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[0] << (8-2*l)) & 256)));
+ const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[0] << (7-2*l)) & 256)));
+ for (int j = 0; j < 4; ++j) {
+ y[j+0] = db1 * grid1[j] * (signs[l] & kmask_iq2xs[j+0] ? -1.f : 1.f);
+ y[j+4] = db1 * grid2[j] * (signs[l] & kmask_iq2xs[j+4] ? -1.f : 1.f);
+ }
+ y += 8;
+ }
+ qs += 8;
+ signs += 4;
+ for (int l = 0; l < 4; ++l) {
+ const uint8_t * grid1 = (const uint8_t *)(iq3s_grid + (qs[2*l+0] | ((qh[1] << (8-2*l)) & 256)));
+ const uint8_t * grid2 = (const uint8_t *)(iq3s_grid + (qs[2*l+1] | ((qh[1] << (7-2*l)) & 256)));
+ for (int j = 0; j < 4; ++j) {
+ y[j+0] = db2 * grid1[j] * (signs[l] & kmask_iq2xs[j+0] ? -1.f : 1.f);
+ y[j+4] = db2 * grid2[j] * (signs[l] & kmask_iq2xs[j+4] ? -1.f : 1.f);
+ }
+ y += 8;
+ }
+ qh += 2;
+ qs += 8;
+ signs += 4;
+ }
+ }
+}
+
+// ====================== 1.5625 bpw (de)-quantization
+
+void dequantize_row_iq1_s(const block_iq1_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ for (int i = 0; i < nb; i++) {
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+ const uint8_t * qs = x[i].qs;
+ const uint16_t * qh = x[i].qh;
+
+ for (int ib = 0; ib < QK_K/32; ++ib) {
+ const float dl = d * (2*((qh[ib] >> 12) & 7) + 1);
+ const float delta = qh[ib] & 0x8000 ? -IQ1S_DELTA : IQ1S_DELTA;
+ for (int l = 0; l < 4; ++l) {
+ const int8_t * grid = (const int8_t *)(iq1s_grid + (qs[l] | (((qh[ib] >> 3*l) & 7) << 8)));
+ for (int j = 0; j < 8; ++j) {
+ y[j] = dl * (grid[j] + delta);
+ }
+ y += 8;
+ }
+ qs += 4;
+ }
+ }
+}
+
+void dequantize_row_iq1_m(const block_iq1_m * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ float delta[4];
+ uint16_t idx[4];
+
+ iq1m_scale_t scale;
+
+ for (int i = 0; i < nb; i++) {
+
+ const uint16_t * sc = (const uint16_t *)x[i].scales;
+ scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+ const float d = GGML_FP16_TO_FP32(scale.f16);
+
+ const uint8_t * qs = x[i].qs;
+ const uint8_t * qh = x[i].qh;
+
+ for (int ib = 0; ib < QK_K/32; ++ib) {
+ const float dl1 = d * (2*((sc[ib/2] >> (6*(ib%2)+0)) & 0x7) + 1);
+ const float dl2 = d * (2*((sc[ib/2] >> (6*(ib%2)+3)) & 0x7) + 1);
+
+ idx[0] = qs[0] | ((qh[0] << 8) & 0x700);
+ idx[1] = qs[1] | ((qh[0] << 4) & 0x700);
+ idx[2] = qs[2] | ((qh[1] << 8) & 0x700);
+ idx[3] = qs[3] | ((qh[1] << 4) & 0x700);
+ delta[0] = qh[0] & 0x08 ? -IQ1S_DELTA : IQ1S_DELTA;
+ delta[1] = qh[0] & 0x80 ? -IQ1S_DELTA : IQ1S_DELTA;
+ delta[2] = qh[1] & 0x08 ? -IQ1S_DELTA : IQ1S_DELTA;
+ delta[3] = qh[1] & 0x80 ? -IQ1S_DELTA : IQ1S_DELTA;
+ for (int l = 0; l < 2; ++l) {
+ const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]);
+ for (int j = 0; j < 8; ++j) {
+ y[j] = dl1 * (grid[j] + delta[l]);
+ }
+ y += 8;
+ }
+ for (int l = 2; l < 4; ++l) {
+ const int8_t * grid = (const int8_t *)(iq1s_grid + idx[l]);
+ for (int j = 0; j < 8; ++j) {
+ y[j] = dl2 * (grid[j] + delta[l]);
+ }
+ y += 8;
+ }
+ qs += 4;
+ qh += 2;
+ }
+ }
+}
+
+void dequantize_row_iq4_nl(const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK4_NL == 0);
+ const int64_t nb = k / QK4_NL;
+
+ for (int i = 0; i < nb; i++) {
+
+ const uint8_t * qs = x[i].qs;
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+ for (int j = 0; j < QK4_NL/2; ++j) {
+ y[j+ 0] = d * kvalues_iq4nl[qs[j] & 0xf];
+ y[j+QK4_NL/2] = d * kvalues_iq4nl[qs[j] >> 4];
+ }
+ y += QK4_NL;
+ qs += QK4_NL/2;
+ }
+}
+
+void dequantize_row_iq4_xs(const block_iq4_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ for (int i = 0; i < nb; i++) {
+
+ const uint8_t * qs = x[i].qs;
+
+ const float d = GGML_FP16_TO_FP32(x[i].d);
+
+ for (int ib = 0; ib < QK_K/32; ++ib) {
+ const int ls = ((x[i].scales_l[ib/2] >> 4*(ib%2)) & 0xf) | (((x[i].scales_h >> 2*ib) & 3) << 4);
+ const float dl = d * (ls - 32);
+ for (int j = 0; j < 16; ++j) {
+ y[j+ 0] = dl * kvalues_iq4nl[qs[j] & 0xf];
+ y[j+16] = dl * kvalues_iq4nl[qs[j] >> 4];
+ }
+ y += 32;
+ qs += 16;
+ }
+ }
+}
+
+//===================================== Q8_K ==============================================
+
+void quantize_row_q8_K_ref(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ for (int i = 0; i < nb; i++) {
+
+ float max = 0;
+ float amax = 0;
+ for (int j = 0; j < QK_K; ++j) {
+ float ax = fabsf(x[j]);
+ if (ax > amax) {
+ amax = ax; max = x[j];
+ }
+ }
+ if (!amax) {
+ y[i].d = 0;
+ memset(y[i].qs, 0, QK_K);
+ x += QK_K;
+ continue;
+ }
+ //const float iscale = -128.f/max;
+ // We need this change for IQ2_XXS, else the AVX implementation becomes very awkward
+ const float iscale = -127.f/max;
+ for (int j = 0; j < QK_K; ++j) {
+ int v = nearest_int(iscale*x[j]);
+ y[i].qs[j] = MIN(127, v);
+ }
+ for (int j = 0; j < QK_K/16; ++j) {
+ int sum = 0;
+ for (int ii = 0; ii < 16; ++ii) {
+ sum += y[i].qs[j*16 + ii];
+ }
+ y[i].bsums[j] = sum;
+ }
+ y[i].d = 1/iscale;
+ x += QK_K;
+ }
+}
+
+void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ const int64_t nb = k / QK_K;
+
+ for (int i = 0; i < nb; i++) {
+ for (int j = 0; j < QK_K; ++j) {
+ *y++ = x[i].d * x[i].qs[j];
+ }
+ }
+}
+
+// ================================ IQ2 quantization =============================================
+
+typedef struct {
+ uint64_t * grid;
+ int * map;
+ uint16_t * neighbours;
+} iq2_entry_t;
+
+static iq2_entry_t iq2_data[4] = {
+ {NULL, NULL, NULL},
+ {NULL, NULL, NULL},
+ {NULL, NULL, NULL},
+ {NULL, NULL, NULL},
+};
+
+static inline int iq2_data_index(enum ggml_type type) {
+ GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M || type == GGML_TYPE_IQ2_S);
+ return type == GGML_TYPE_IQ2_XXS ? 0 :
+ type == GGML_TYPE_IQ2_XS ? 1 :
+ type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? 2 : 3;
+}
+
+static inline int iq2_grid_size(enum ggml_type type) {
+ GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M || type == GGML_TYPE_IQ2_S);
+ return type == GGML_TYPE_IQ2_XXS ? 256 :
+ type == GGML_TYPE_IQ2_XS ? 512 :
+ type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? NGRID_IQ1S : 1024;
+}
+
+static int iq2_compare_func(const void * left, const void * right) {
+ const int * l = (const int *)left;
+ const int * r = (const int *)right;
+ return l[0] < r[0] ? -1 : l[0] > r[0] ? 1 : l[1] < r[1] ? -1 : l[1] > r[1] ? 1 : 0;
+}
+
+void iq2xs_init_impl(enum ggml_type type) {
+ const int gindex = iq2_data_index(type);
+ const int grid_size = iq2_grid_size(type);
+ if (iq2_data[gindex].grid) {
+ return;
+ }
+ static const uint16_t kgrid_2bit_256[256] = {
+ 0, 2, 5, 8, 10, 17, 20, 32, 34, 40, 42, 65, 68, 80, 88, 97,
+ 100, 128, 130, 138, 162, 257, 260, 272, 277, 320, 388, 408, 512, 514, 546, 642,
+ 1025, 1028, 1040, 1057, 1060, 1088, 1090, 1096, 1120, 1153, 1156, 1168, 1188, 1280, 1282, 1288,
+ 1312, 1350, 1385, 1408, 1425, 1545, 1552, 1600, 1668, 1700, 2048, 2053, 2056, 2068, 2088, 2113,
+ 2116, 2128, 2130, 2184, 2308, 2368, 2562, 2580, 4097, 4100, 4112, 4129, 4160, 4192, 4228, 4240,
+ 4245, 4352, 4360, 4384, 4432, 4442, 4480, 4644, 4677, 5120, 5128, 5152, 5157, 5193, 5248, 5400,
+ 5474, 5632, 5654, 6145, 6148, 6160, 6208, 6273, 6400, 6405, 6560, 6737, 8192, 8194, 8202, 8260,
+ 8289, 8320, 8322, 8489, 8520, 8704, 8706, 9217, 9220, 9232, 9280, 9302, 9472, 9537, 9572, 9872,
+ 10248, 10272, 10388, 10820, 16385, 16388, 16400, 16408, 16417, 16420, 16448, 16456, 16470, 16480, 16513, 16516,
+ 16528, 16640, 16672, 16737, 16768, 16773, 16897, 16912, 16968, 16982, 17000, 17408, 17416, 17440, 17536, 17561,
+ 17682, 17700, 17920, 18433, 18436, 18448, 18496, 18501, 18688, 18776, 18785, 18818, 19013, 19088, 20480, 20488,
+ 20497, 20505, 20512, 20608, 20616, 20740, 20802, 20900, 21137, 21648, 21650, 21770, 22017, 22100, 22528, 22545,
+ 22553, 22628, 22848, 23048, 24580, 24592, 24640, 24680, 24832, 24917, 25112, 25184, 25600, 25605, 25872, 25874,
+ 25988, 26690, 32768, 32770, 32778, 32833, 32898, 33028, 33048, 33088, 33297, 33793, 33796, 33808, 33813, 33856,
+ 33888, 34048, 34118, 34196, 34313, 34368, 34400, 34818, 35076, 35345, 36868, 36880, 36900, 36928, 37025, 37142,
+ 37248, 37445, 37888, 37922, 37956, 38225, 39041, 39200, 40962, 41040, 41093, 41225, 41472, 42008, 43088, 43268,
+ };
+ static const uint16_t kgrid_2bit_512[512] = {
+ 0, 2, 5, 8, 10, 17, 20, 22, 25, 32, 34, 37, 40, 65, 68, 70,
+ 73, 80, 82, 85, 88, 97, 100, 128, 130, 133, 136, 145, 148, 153, 160, 257,
+ 260, 262, 265, 272, 274, 277, 280, 282, 289, 292, 320, 322, 325, 328, 337, 340,
+ 352, 360, 385, 388, 400, 512, 514, 517, 520, 529, 532, 544, 577, 580, 592, 597,
+ 640, 650, 1025, 1028, 1030, 1033, 1040, 1042, 1045, 1048, 1057, 1060, 1088, 1090, 1093, 1096,
+ 1105, 1108, 1110, 1120, 1153, 1156, 1168, 1280, 1282, 1285, 1288, 1297, 1300, 1312, 1345, 1348,
+ 1360, 1377, 1408, 1537, 1540, 1552, 1574, 1600, 1602, 1668, 2048, 2050, 2053, 2056, 2058, 2065,
+ 2068, 2080, 2085, 2113, 2116, 2128, 2136, 2176, 2208, 2218, 2305, 2308, 2320, 2368, 2433, 2441,
+ 2560, 2592, 2600, 2710, 2720, 4097, 4100, 4102, 4105, 4112, 4114, 4117, 4120, 4129, 4132, 4160,
+ 4162, 4165, 4168, 4177, 4180, 4192, 4202, 4225, 4228, 4240, 4352, 4354, 4357, 4360, 4369, 4372,
+ 4384, 4417, 4420, 4432, 4480, 4500, 4502, 4609, 4612, 4614, 4624, 4672, 4704, 5120, 5122, 5125,
+ 5128, 5137, 5140, 5152, 5185, 5188, 5193, 5200, 5220, 5248, 5377, 5380, 5392, 5440, 5632, 5652,
+ 5705, 6145, 6148, 6160, 6162, 6208, 6228, 6278, 6400, 6405, 6502, 6737, 6825, 8192, 8194, 8197,
+ 8200, 8202, 8209, 8212, 8224, 8257, 8260, 8272, 8320, 8352, 8449, 8452, 8464, 8512, 8520, 8549,
+ 8704, 8738, 8832, 8872, 9217, 9220, 9232, 9257, 9280, 9472, 9537, 9554, 9625, 9729, 9754, 9894,
+ 10240, 10248, 10250, 10272, 10325, 10376, 10402, 10600, 10640, 10760, 10784, 10882, 10888, 10890, 16385, 16388,
+ 16390, 16393, 16400, 16402, 16405, 16408, 16417, 16420, 16448, 16450, 16453, 16456, 16458, 16465, 16468, 16480,
+ 16485, 16513, 16516, 16528, 16640, 16642, 16645, 16648, 16657, 16660, 16672, 16705, 16708, 16720, 16768, 16773,
+ 16802, 16897, 16900, 16912, 16914, 16937, 16960, 17408, 17410, 17413, 17416, 17425, 17428, 17433, 17440, 17473,
+ 17476, 17488, 17536, 17556, 17665, 17668, 17680, 17700, 17728, 17818, 17920, 17930, 17988, 18000, 18433, 18436,
+ 18448, 18496, 18501, 18516, 18530, 18688, 18705, 18756, 18768, 18793, 18948, 20480, 20482, 20485, 20488, 20497,
+ 20500, 20512, 20520, 20545, 20548, 20560, 20608, 20737, 20740, 20752, 20757, 20800, 20802, 20992, 21060, 21162,
+ 21505, 21508, 21520, 21537, 21568, 21600, 21633, 21665, 21760, 21768, 21888, 21896, 22049, 22120, 22177, 22528,
+ 22548, 22593, 22608, 22681, 22810, 22848, 22850, 23173, 24577, 24580, 24592, 24640, 24660, 24674, 24710, 24745,
+ 24832, 25124, 25162, 25234, 25600, 25622, 25872, 25920, 25925, 26020, 26625, 26730, 26917, 27142, 27220, 27234,
+ 32768, 32770, 32773, 32776, 32785, 32788, 32800, 32810, 32833, 32836, 32848, 32896, 32898, 32936, 32938, 33025,
+ 33028, 33030, 33040, 33088, 33105, 33113, 33280, 33312, 33408, 33410, 33440, 33448, 33793, 33796, 33808, 33810,
+ 33813, 33856, 33888, 33929, 34048, 34116, 34213, 34328, 34410, 34816, 34824, 34853, 34906, 34944, 34946, 34984,
+ 35078, 35362, 35456, 35464, 35478, 35496, 36865, 36868, 36880, 36928, 36950, 36996, 37120, 37154, 37220, 37462,
+ 37513, 37888, 37893, 37956, 37968, 37976, 38185, 38288, 38290, 38465, 38993, 39078, 39241, 39445, 39520, 40960,
+ 40962, 40968, 40970, 40992, 41002, 41120, 41297, 41305, 41382, 41472, 41474, 41480, 41514, 41600, 41632, 42048,
+ 42133, 42597, 42648, 43018, 43040, 43042, 43048, 43168, 43176, 43268, 43396, 43398, 43560, 43562, 43665, 43690,
+ };
+ static const uint16_t kgrid_1bit_2048[NGRID_IQ1S] = {
+ 0, 2, 5, 8, 10, 17, 21, 32, 34, 40, 42, 69, 81, 84, 86, 101,
+ 128, 130, 136, 138, 149, 160, 162, 168, 170, 260, 261, 273, 276, 278, 281, 282,
+ 293, 321, 326, 329, 338, 341, 346, 353, 356, 358, 360, 389, 401, 404, 406, 421,
+ 512, 514, 520, 522, 533, 544, 546, 552, 554, 581, 593, 601, 612, 617, 640, 642,
+ 648, 650, 657, 661, 665, 672, 674, 680, 682, 1041, 1044, 1046, 1061, 1089, 1097, 1109,
+ 1114, 1124, 1125, 1169, 1177, 1189, 1281, 1284, 1285, 1286, 1301, 1304, 1306, 1321, 1344, 1349,
+ 1354, 1360, 1361, 1364, 1365, 1366, 1369, 1376, 1378, 1381, 1384, 1386, 1409, 1425, 1429, 1432,
+ 1434, 1441, 1444, 1445, 1446, 1449, 1556, 1561, 1601, 1604, 1616, 1618, 1621, 1624, 1632, 1633,
+ 1638, 1641, 1669, 1681, 1684, 1689, 2048, 2050, 2056, 2058, 2069, 2080, 2082, 2088, 2090, 2117,
+ 2129, 2134, 2149, 2176, 2178, 2184, 2186, 2197, 2208, 2210, 2216, 2218, 2309, 2321, 2324, 2329,
+ 2340, 2341, 2369, 2384, 2385, 2389, 2401, 2404, 2409, 2449, 2452, 2454, 2457, 2469, 2560, 2562,
+ 2568, 2570, 2581, 2592, 2594, 2600, 2602, 2629, 2641, 2649, 2657, 2661, 2688, 2690, 2693, 2696,
+ 2698, 2709, 2720, 2722, 2728, 2730, 4112, 4113, 4116, 4121, 4132, 4133, 4161, 4164, 4176, 4181,
+ 4184, 4193, 4196, 4197, 4201, 4241, 4244, 4246, 4257, 4261, 4353, 4356, 4358, 4361, 4368, 4370,
+ 4373, 4376, 4385, 4388, 4393, 4421, 4426, 4432, 4433, 4434, 4436, 4437, 4438, 4441, 4448, 4453,
+ 4484, 4498, 4501, 4513, 4516, 4625, 4628, 4630, 4645, 4672, 4678, 4681, 4690, 4693, 4696, 4698,
+ 4708, 4710, 4741, 4753, 4756, 4758, 4773, 5121, 5126, 5129, 5140, 5141, 5144, 5145, 5153, 5158,
+ 5185, 5189, 5190, 5192, 5194, 5201, 5204, 5205, 5206, 5209, 5218, 5221, 5224, 5252, 5257, 5264,
+ 5268, 5269, 5272, 5273, 5274, 5281, 5284, 5285, 5289, 5378, 5381, 5386, 5393, 5396, 5397, 5398,
+ 5401, 5408, 5410, 5413, 5416, 5418, 5441, 5444, 5445, 5446, 5457, 5458, 5460, 5461, 5462, 5465,
+ 5466, 5473, 5476, 5477, 5478, 5481, 5504, 5506, 5508, 5509, 5512, 5514, 5520, 5521, 5524, 5525,
+ 5526, 5529, 5530, 5536, 5538, 5541, 5633, 5636, 5637, 5638, 5653, 5654, 5656, 5658, 5665, 5670,
+ 5696, 5698, 5700, 5701, 5704, 5706, 5713, 5717, 5718, 5720, 5721, 5729, 5732, 5733, 5736, 5737,
+ 5738, 5766, 5770, 5778, 5781, 5796, 5801, 6161, 6166, 6181, 6209, 6212, 6214, 6217, 6224, 6229,
+ 6232, 6234, 6240, 6241, 6244, 6246, 6249, 6277, 6289, 6292, 6309, 6416, 6418, 6421, 6426, 6433,
+ 6437, 6466, 6468, 6469, 6472, 6481, 6484, 6485, 6486, 6489, 6490, 6496, 6501, 6506, 6537, 6545,
+ 6546, 6549, 6552, 6561, 6566, 6569, 6665, 6678, 6692, 6694, 6724, 6726, 6729, 6736, 6738, 6741,
+ 6744, 6753, 6758, 6761, 6789, 6801, 6806, 6810, 8192, 8194, 8200, 8202, 8213, 8224, 8226, 8229,
+ 8232, 8234, 8261, 8273, 8281, 8289, 8293, 8320, 8322, 8328, 8330, 8341, 8352, 8354, 8357, 8360,
+ 8362, 8453, 8465, 8468, 8473, 8485, 8514, 8516, 8521, 8533, 8536, 8538, 8545, 8548, 8549, 8550,
+ 8581, 8592, 8598, 8601, 8613, 8705, 8712, 8714, 8721, 8725, 8736, 8738, 8744, 8746, 8773, 8785,
+ 8790, 8793, 8805, 8833, 8840, 8842, 8849, 8853, 8864, 8866, 8872, 8874, 9221, 9236, 9238, 9241,
+ 9253, 9284, 9285, 9286, 9289, 9298, 9301, 9304, 9306, 9318, 9349, 9361, 9364, 9369, 9377, 9381,
+ 9481, 9493, 9505, 9513, 9536, 9541, 9544, 9553, 9556, 9557, 9561, 9570, 9573, 9576, 9609, 9616,
+ 9620, 9621, 9624, 9626, 9633, 9636, 9638, 9641, 9733, 9744, 9746, 9753, 9765, 9793, 9801, 9813,
+ 9824, 9825, 9833, 9860, 9862, 9872, 9882, 10240, 10242, 10248, 10250, 10261, 10272, 10274, 10280, 10282,
+ 10309, 10321, 10324, 10341, 10368, 10370, 10376, 10378, 10400, 10402, 10408, 10410, 10505, 10513, 10516, 10521,
+ 10533, 10566, 10569, 10578, 10581, 10593, 10596, 10598, 10601, 10629, 10640, 10646, 10649, 10660, 10661, 10752,
+ 10754, 10760, 10762, 10784, 10786, 10792, 10794, 10821, 10833, 10838, 10841, 10853, 10880, 10882, 10888, 10890,
+ 10901, 10912, 10914, 10920, 10922, 16389, 16401, 16406, 16421, 16457, 16466, 16469, 16472, 16474, 16481, 16484,
+ 16486, 16532, 16537, 16545, 16550, 16640, 16641, 16644, 16646, 16649, 16658, 16661, 16662, 16664, 16666, 16673,
+ 16678, 16681, 16709, 16712, 16714, 16721, 16724, 16725, 16726, 16729, 16730, 16741, 16744, 16746, 16769, 16772,
+ 16774, 16784, 16786, 16789, 16800, 16801, 16802, 16901, 16913, 16916, 16918, 16933, 16961, 16978, 16981, 16986,
+ 16996, 17001, 17033, 17044, 17061, 17409, 17429, 17433, 17449, 17477, 17480, 17482, 17489, 17492, 17493, 17494,
+ 17505, 17506, 17509, 17512, 17514, 17537, 17542, 17545, 17552, 17554, 17557, 17568, 17569, 17577, 17665, 17666,
+ 17669, 17674, 17681, 17684, 17685, 17686, 17689, 17696, 17701, 17706, 17729, 17732, 17733, 17734, 17737, 17744,
+ 17745, 17748, 17749, 17750, 17752, 17753, 17761, 17764, 17765, 17766, 17769, 17794, 17796, 17797, 17800, 17809,
+ 17812, 17813, 17814, 17817, 17818, 17829, 17832, 17834, 17921, 17925, 17929, 17940, 17941, 17944, 17946, 17953,
+ 17956, 17961, 17984, 17986, 17989, 17992, 18000, 18001, 18002, 18005, 18006, 18009, 18018, 18021, 18024, 18049,
+ 18053, 18058, 18068, 18069, 18081, 18084, 18086, 18437, 18449, 18453, 18458, 18469, 18498, 18505, 18512, 18517,
+ 18520, 18529, 18532, 18534, 18537, 18565, 18577, 18580, 18582, 18585, 18597, 18689, 18693, 18694, 18698, 18704,
+ 18708, 18709, 18712, 18721, 18724, 18726, 18752, 18757, 18762, 18769, 18770, 18772, 18773, 18774, 18777, 18784,
+ 18786, 18789, 18790, 18794, 18822, 18825, 18834, 18837, 18838, 18840, 18849, 18852, 18854, 18857, 18966, 19012,
+ 19014, 19017, 19029, 19032, 19034, 19044, 19049, 19092, 19109, 20481, 20484, 20485, 20486, 20489, 20498, 20501,
+ 20506, 20513, 20516, 20521, 20544, 20549, 20552, 20561, 20564, 20565, 20566, 20569, 20581, 20584, 20614, 20617,
+ 20629, 20632, 20640, 20641, 20646, 20649, 20741, 20744, 20745, 20746, 20753, 20756, 20757, 20758, 20760, 20761,
+ 20768, 20773, 20774, 20776, 20778, 20801, 20804, 20805, 20806, 20809, 20816, 20817, 20818, 20820, 20821, 20822,
+ 20824, 20825, 20826, 20833, 20836, 20837, 20838, 20841, 20866, 20869, 20881, 20884, 20885, 20886, 20889, 20896,
+ 20901, 20906, 20993, 20998, 21010, 21013, 21018, 21025, 21028, 21058, 21061, 21066, 21073, 21076, 21077, 21078,
+ 21081, 21090, 21093, 21125, 21136, 21138, 21141, 21145, 21146, 21156, 21508, 21509, 21521, 21524, 21525, 21526,
+ 21528, 21529, 21537, 21541, 21544, 21546, 21569, 21572, 21573, 21574, 21577, 21578, 21584, 21585, 21588, 21589,
+ 21590, 21592, 21593, 21594, 21601, 21602, 21604, 21605, 21606, 21609, 21632, 21640, 21642, 21649, 21652, 21653,
+ 21654, 21657, 21665, 21668, 21669, 21674, 21761, 21762, 21764, 21765, 21766, 21769, 21776, 21777, 21778, 21780,
+ 21781, 21782, 21785, 21786, 21793, 21796, 21797, 21798, 21801, 21824, 21825, 21826, 21828, 21829, 21830, 21832,
+ 21833, 21840, 21841, 21842, 21844, 21845, 21846, 21848, 21849, 21850, 21856, 21857, 21860, 21861, 21862, 21864,
+ 21865, 21866, 21889, 21892, 21893, 21897, 21898, 21904, 21905, 21908, 21909, 21910, 21912, 21913, 21921, 21924,
+ 21925, 21926, 21929, 22016, 22017, 22018, 22020, 22022, 22024, 22025, 22033, 22036, 22037, 22040, 22041, 22048,
+ 22049, 22050, 22052, 22053, 22054, 22056, 22057, 22081, 22085, 22086, 22088, 22089, 22090, 22096, 22097, 22098,
+ 22100, 22101, 22102, 22104, 22105, 22106, 22113, 22116, 22117, 22121, 22146, 22149, 22150, 22152, 22153, 22154,
+ 22161, 22165, 22170, 22178, 22181, 22182, 22184, 22185, 22532, 22533, 22534, 22537, 22544, 22549, 22552, 22561,
+ 22570, 22597, 22600, 22602, 22609, 22612, 22613, 22614, 22616, 22617, 22624, 22626, 22628, 22629, 22658, 22665,
+ 22672, 22674, 22677, 22680, 22689, 22697, 22785, 22786, 22789, 22794, 22801, 22804, 22805, 22806, 22809, 22821,
+ 22849, 22852, 22853, 22854, 22857, 22864, 22865, 22866, 22868, 22869, 22870, 22872, 22873, 22874, 22881, 22884,
+ 22885, 22886, 22889, 22913, 22917, 22921, 22929, 22932, 22933, 22934, 22936, 22937, 22949, 23044, 23048, 23061,
+ 23066, 23072, 23077, 23078, 23081, 23109, 23112, 23113, 23121, 23125, 23126, 23128, 23129, 23138, 23141, 23144,
+ 23146, 23169, 23178, 23186, 23189, 23190, 23192, 23194, 23201, 24581, 24596, 24598, 24601, 24613, 24644, 24656,
+ 24661, 24662, 24664, 24666, 24673, 24676, 24678, 24681, 24705, 24726, 24741, 24833, 24836, 24838, 24841, 24850,
+ 24853, 24865, 24866, 24870, 24873, 24901, 24905, 24913, 24917, 24918, 24921, 24933, 24934, 24938, 24964, 24970,
+ 24978, 24981, 24993, 24998, 25001, 25105, 25110, 25113, 25152, 25153, 25158, 25173, 25174, 25176, 25184, 25221,
+ 25233, 25238, 25253, 25617, 25618, 25621, 25622, 25626, 25633, 25638, 25641, 25664, 25666, 25669, 25672, 25674,
+ 25681, 25684, 25685, 25686, 25689, 25690, 25696, 25698, 25701, 25732, 25733, 25737, 25744, 25746, 25748, 25749,
+ 25750, 25752, 25754, 25761, 25764, 25769, 25861, 25864, 25866, 25873, 25877, 25878, 25881, 25924, 25925, 25926,
+ 25929, 25936, 25937, 25940, 25941, 25942, 25945, 25953, 25956, 25957, 25958, 25961, 25990, 25993, 25994, 26001,
+ 26005, 26006, 26009, 26010, 26018, 26021, 26022, 26024, 26114, 26121, 26133, 26144, 26150, 26152, 26153, 26176,
+ 26181, 26184, 26186, 26193, 26196, 26197, 26198, 26200, 26202, 26208, 26213, 26216, 26240, 26242, 26245, 26250,
+ 26260, 26262, 26264, 26265, 26272, 26276, 26278, 26282, 26646, 26649, 26661, 26689, 26706, 26709, 26714, 26721,
+ 26729, 26757, 26769, 26776, 26790, 26881, 26884, 26896, 26901, 26913, 26916, 26918, 26921, 26944, 26945, 26949,
+ 26950, 26952, 26961, 26964, 26965, 26966, 26969, 26976, 26981, 26986, 27010, 27012, 27018, 27029, 27041, 27044,
+ 27045, 27049, 27153, 27158, 27160, 27201, 27204, 27209, 27216, 27221, 27224, 27226, 27236, 27237, 27241, 27270,
+ 27284, 27288, 27290, 27302, 32768, 32770, 32776, 32778, 32800, 32802, 32808, 32810, 32837, 32848, 32849, 32852,
+ 32854, 32857, 32869, 32896, 32898, 32904, 32906, 32917, 32928, 32930, 32936, 32938, 33029, 33041, 33044, 33046,
+ 33049, 33061, 33089, 33092, 33097, 33104, 33106, 33109, 33110, 33112, 33113, 33124, 33126, 33129, 33157, 33161,
+ 33172, 33174, 33177, 33189, 33280, 33282, 33288, 33290, 33301, 33312, 33314, 33320, 33322, 33361, 33364, 33369,
+ 33381, 33408, 33410, 33416, 33418, 33429, 33440, 33442, 33448, 33450, 33812, 33817, 33857, 33860, 33873, 33877,
+ 33882, 33889, 33892, 33897, 33940, 33945, 34049, 34057, 34066, 34069, 34074, 34086, 34089, 34112, 34113, 34117,
+ 34120, 34129, 34132, 34133, 34134, 34137, 34138, 34149, 34150, 34152, 34154, 34177, 34180, 34182, 34185, 34192,
+ 34194, 34197, 34200, 34214, 34321, 34326, 34329, 34341, 34369, 34372, 34377, 34378, 34384, 34389, 34393, 34394,
+ 34401, 34406, 34410, 34437, 34449, 34458, 34468, 34816, 34818, 34824, 34826, 34837, 34848, 34850, 34856, 34858,
+ 34881, 34885, 34897, 34900, 34905, 34917, 34921, 34944, 34946, 34952, 34954, 34965, 34976, 34978, 34984, 34986,
+ 35077, 35078, 35089, 35092, 35094, 35109, 35137, 35140, 35142, 35145, 35152, 35154, 35157, 35162, 35169, 35172,
+ 35205, 35222, 35225, 35237, 35328, 35330, 35336, 35338, 35349, 35360, 35362, 35368, 35370, 35397, 35409, 35412,
+ 35414, 35456, 35458, 35464, 35466, 35477, 35488, 35490, 35496, 35498, 36869, 36881, 36886, 36888, 36889, 36901,
+ 36929, 36934, 36937, 36949, 36952, 36954, 36969, 36970, 36997, 37009, 37012, 37014, 37017, 37029, 37121, 37124,
+ 37126, 37129, 37136, 37141, 37144, 37146, 37153, 37156, 37158, 37161, 37184, 37189, 37200, 37201, 37204, 37205,
+ 37206, 37209, 37218, 37221, 37252, 37254, 37266, 37269, 37272, 37281, 37284, 37286, 37289, 37381, 37393, 37396,
+ 37401, 37413, 37444, 37446, 37449, 37456, 37458, 37461, 37464, 37478, 37481, 37509, 37524, 37526, 37545, 37889,
+ 37892, 37894, 37904, 37909, 37912, 37926, 37952, 37962, 37969, 37972, 37973, 37974, 37976, 37977, 37984, 37985,
+ 37986, 37989, 38020, 38022, 38034, 38036, 38037, 38040, 38049, 38057, 38144, 38149, 38152, 38154, 38160, 38161,
+ 38164, 38165, 38166, 38169, 38177, 38181, 38185, 38186, 38209, 38212, 38213, 38214, 38217, 38224, 38225, 38226,
+ 38228, 38229, 38230, 38232, 38233, 38234, 38241, 38244, 38245, 38246, 38249, 38273, 38277, 38280, 38289, 38290,
+ 38292, 38293, 38294, 38297, 38298, 38304, 38306, 38309, 38312, 38314, 38401, 38404, 38416, 38421, 38425, 38432,
+ 38438, 38441, 38469, 38472, 38473, 38481, 38482, 38485, 38486, 38489, 38501, 38504, 38530, 38532, 38537, 38538,
+ 38546, 38548, 38549, 38564, 38566, 38569, 38917, 38934, 38937, 38949, 38977, 38982, 38992, 38994, 38997, 38998,
+ 39002, 39012, 39013, 39045, 39057, 39062, 39065, 39077, 39172, 39174, 39177, 39184, 39186, 39189, 39192, 39194,
+ 39200, 39201, 39204, 39206, 39232, 39234, 39237, 39240, 39242, 39249, 39252, 39253, 39254, 39257, 39266, 39269,
+ 39270, 39274, 39297, 39300, 39312, 39314, 39317, 39322, 39329, 39334, 39429, 39445, 39461, 39492, 39494, 39497,
+ 39504, 39509, 39512, 39521, 39557, 39569, 39572, 39573, 39574, 40960, 40962, 40968, 40970, 40981, 40992, 40994,
+ 41000, 41002, 41029, 41041, 41044, 41046, 41049, 41088, 41090, 41096, 41098, 41109, 41120, 41122, 41128, 41130,
+ 41221, 41225, 41233, 41236, 41238, 41241, 41242, 41286, 41289, 41297, 41301, 41304, 41306, 41313, 41316, 41349,
+ 41360, 41362, 41366, 41369, 41474, 41480, 41482, 41488, 41497, 41506, 41512, 41514, 41541, 41553, 41558, 41561,
+ 41573, 41600, 41602, 41608, 41610, 41621, 41632, 41634, 41640, 41642, 42009, 42021, 42049, 42052, 42064, 42068,
+ 42069, 42072, 42074, 42081, 42085, 42086, 42088, 42089, 42117, 42246, 42249, 42256, 42258, 42261, 42264, 42278,
+ 42281, 42306, 42309, 42321, 42324, 42325, 42326, 42329, 42341, 42346, 42369, 42372, 42373, 42374, 42377, 42386,
+ 42389, 42392, 42501, 42513, 42518, 42522, 42529, 42533, 42564, 42566, 42570, 42578, 42581, 42582, 42584, 42592,
+ 42594, 42630, 42640, 42645, 42646, 42649, 42657, 42660, 42662, 43008, 43010, 43016, 43018, 43040, 43042, 43048,
+ 43050, 43089, 43092, 43094, 43097, 43136, 43138, 43144, 43146, 43157, 43168, 43170, 43176, 43178, 43269, 43284,
+ 43289, 43297, 43301, 43329, 43344, 43349, 43354, 43361, 43366, 43369, 43408, 43414, 43520, 43522, 43528, 43530,
+ 43552, 43554, 43560, 43562, 43601, 43604, 43606, 43648, 43650, 43656, 43658, 43669, 43680, 43682, 43688, 43690,
+ };
+ static const uint16_t kgrid_2bit_1024[1024] = {
+ 0, 2, 5, 8, 10, 17, 20, 22, 25, 32, 34, 37, 40, 65, 68, 70,
+ 73, 80, 82, 85, 88, 97, 100, 102, 105, 128, 130, 133, 136, 145, 148, 160,
+ 165, 170, 257, 260, 262, 265, 272, 274, 277, 280, 289, 292, 320, 322, 325, 328,
+ 337, 340, 342, 345, 352, 357, 360, 385, 388, 400, 402, 405, 417, 420, 512, 514,
+ 517, 520, 529, 532, 544, 554, 577, 580, 582, 585, 592, 597, 640, 645, 650, 660,
+ 674, 1025, 1028, 1030, 1033, 1040, 1042, 1045, 1048, 1057, 1060, 1062, 1065, 1088, 1090, 1093,
+ 1096, 1098, 1105, 1108, 1110, 1113, 1120, 1122, 1125, 1153, 1156, 1158, 1161, 1168, 1173, 1176,
+ 1185, 1188, 1280, 1282, 1285, 1288, 1290, 1297, 1300, 1302, 1305, 1312, 1317, 1320, 1345, 1348,
+ 1350, 1353, 1360, 1362, 1365, 1368, 1377, 1380, 1408, 1410, 1413, 1416, 1425, 1428, 1440, 1537,
+ 1540, 1542, 1545, 1552, 1557, 1600, 1605, 1608, 1617, 1620, 1632, 1665, 1668, 1680, 2048, 2050,
+ 2053, 2056, 2065, 2068, 2070, 2073, 2080, 2085, 2090, 2113, 2116, 2118, 2121, 2128, 2130, 2133,
+ 2136, 2145, 2148, 2176, 2181, 2196, 2218, 2305, 2308, 2320, 2322, 2325, 2328, 2337, 2368, 2373,
+ 2376, 2385, 2388, 2400, 2433, 2448, 2560, 2577, 2580, 2594, 2600, 2602, 2640, 2713, 4097, 4100,
+ 4102, 4105, 4112, 4114, 4117, 4120, 4129, 4132, 4134, 4160, 4162, 4165, 4168, 4177, 4180, 4182,
+ 4185, 4192, 4194, 4197, 4200, 4225, 4228, 4230, 4240, 4245, 4248, 4257, 4260, 4352, 4354, 4357,
+ 4360, 4362, 4369, 4372, 4374, 4377, 4384, 4386, 4389, 4392, 4417, 4420, 4422, 4425, 4432, 4434,
+ 4437, 4440, 4449, 4452, 4480, 4482, 4485, 4488, 4497, 4500, 4609, 4612, 4617, 4624, 4629, 4641,
+ 4644, 4672, 4677, 4689, 4692, 4737, 4740, 4752, 5120, 5122, 5125, 5128, 5137, 5140, 5142, 5145,
+ 5152, 5157, 5160, 5185, 5188, 5190, 5193, 5200, 5202, 5205, 5208, 5217, 5220, 5248, 5250, 5253,
+ 5256, 5265, 5268, 5280, 5377, 5380, 5382, 5385, 5392, 5394, 5397, 5400, 5409, 5412, 5440, 5442,
+ 5445, 5448, 5457, 5460, 5472, 5505, 5508, 5520, 5632, 5637, 5640, 5649, 5652, 5664, 5697, 5700,
+ 5712, 5760, 5802, 6145, 6148, 6150, 6153, 6160, 6165, 6168, 6177, 6208, 6210, 6213, 6216, 6225,
+ 6228, 6240, 6273, 6276, 6400, 6402, 6405, 6408, 6417, 6420, 6432, 6465, 6468, 6480, 6505, 6562,
+ 6660, 6672, 6720, 6742, 8192, 8194, 8197, 8200, 8209, 8212, 8214, 8217, 8224, 8229, 8234, 8257,
+ 8260, 8272, 8274, 8277, 8292, 8320, 8330, 8340, 8362, 8449, 8452, 8464, 8466, 8469, 8481, 8512,
+ 8514, 8517, 8529, 8532, 8544, 8577, 8580, 8592, 8704, 8714, 8738, 8744, 8746, 8772, 8784, 8840,
+ 8842, 8872, 9217, 9220, 9222, 9225, 9232, 9237, 9240, 9249, 9252, 9280, 9282, 9285, 9288, 9297,
+ 9300, 9312, 9345, 9348, 9360, 9472, 9477, 9480, 9489, 9492, 9504, 9537, 9540, 9552, 9574, 9600,
+ 9729, 9732, 9744, 9792, 9817, 10240, 10245, 10257, 10260, 10305, 10308, 10320, 10378, 10410, 10497, 10500,
+ 10512, 10645, 10762, 10786, 10852, 10888, 10890, 16385, 16388, 16390, 16393, 16400, 16402, 16405, 16408, 16410,
+ 16417, 16420, 16422, 16448, 16450, 16453, 16456, 16458, 16465, 16468, 16470, 16473, 16480, 16482, 16485, 16513,
+ 16516, 16528, 16533, 16536, 16545, 16548, 16640, 16642, 16645, 16648, 16657, 16660, 16662, 16665, 16672, 16674,
+ 16677, 16705, 16708, 16710, 16713, 16720, 16722, 16725, 16728, 16737, 16740, 16768, 16770, 16773, 16776, 16785,
+ 16788, 16800, 16897, 16900, 16912, 16914, 16917, 16920, 16932, 16960, 16965, 16968, 16977, 16980, 16992, 17025,
+ 17028, 17408, 17410, 17413, 17416, 17418, 17425, 17428, 17430, 17433, 17440, 17442, 17445, 17448, 17473, 17476,
+ 17478, 17481, 17488, 17490, 17493, 17496, 17505, 17508, 17536, 17538, 17541, 17544, 17553, 17556, 17568, 17665,
+ 17668, 17670, 17673, 17680, 17682, 17685, 17688, 17697, 17700, 17728, 17730, 17733, 17736, 17745, 17748, 17760,
+ 17770, 17793, 17796, 17808, 17920, 17922, 17925, 17928, 17937, 17940, 17952, 17985, 17988, 18000, 18048, 18085,
+ 18433, 18436, 18441, 18448, 18450, 18453, 18456, 18465, 18468, 18496, 18498, 18501, 18504, 18513, 18516, 18528,
+ 18564, 18576, 18688, 18690, 18693, 18696, 18705, 18708, 18720, 18753, 18756, 18768, 18816, 18838, 18945, 18948,
+ 18960, 19008, 20480, 20482, 20485, 20488, 20497, 20500, 20502, 20505, 20512, 20514, 20517, 20520, 20545, 20548,
+ 20550, 20553, 20560, 20562, 20565, 20568, 20577, 20580, 20608, 20610, 20613, 20616, 20625, 20628, 20737, 20740,
+ 20742, 20745, 20752, 20754, 20757, 20760, 20769, 20772, 20800, 20802, 20805, 20808, 20817, 20820, 20832, 20865,
+ 20868, 20880, 20992, 20997, 21000, 21009, 21012, 21024, 21057, 21060, 21072, 21097, 21120, 21505, 21508, 21510,
+ 21513, 21520, 21522, 21525, 21528, 21537, 21540, 21568, 21570, 21573, 21576, 21585, 21588, 21600, 21633, 21636,
+ 21648, 21760, 21762, 21765, 21768, 21777, 21780, 21792, 21825, 21828, 21840, 21888, 22017, 22020, 22032, 22054,
+ 22080, 22528, 22530, 22533, 22536, 22545, 22548, 22560, 22593, 22596, 22608, 22618, 22656, 22785, 22788, 22800,
+ 22848, 23040, 23065, 23173, 23208, 24577, 24580, 24582, 24592, 24594, 24597, 24600, 24609, 24612, 24640, 24645,
+ 24648, 24657, 24660, 24672, 24708, 24720, 24832, 24834, 24837, 24840, 24849, 24852, 24864, 24897, 24900, 24912,
+ 24960, 24985, 25092, 25104, 25152, 25174, 25249, 25600, 25605, 25608, 25617, 25620, 25632, 25665, 25668, 25680,
+ 25728, 25857, 25860, 25872, 25920, 25930, 25960, 26002, 26112, 26260, 26625, 26628, 26640, 26725, 26776, 26880,
+ 26922, 27202, 27297, 32768, 32770, 32773, 32776, 32785, 32788, 32793, 32800, 32805, 32833, 32836, 32848, 32850,
+ 32853, 32856, 32865, 32896, 32901, 32913, 32916, 33025, 33028, 33033, 33040, 33042, 33045, 33048, 33057, 33060,
+ 33088, 33090, 33093, 33096, 33105, 33108, 33153, 33156, 33168, 33193, 33280, 33285, 33290, 33297, 33300, 33345,
+ 33348, 33360, 33793, 33796, 33798, 33801, 33808, 33810, 33813, 33816, 33825, 33856, 33858, 33861, 33864, 33873,
+ 33876, 33888, 33921, 33924, 33936, 34048, 34050, 34053, 34056, 34065, 34068, 34080, 34113, 34116, 34128, 34176,
+ 34186, 34305, 34308, 34320, 34345, 34368, 34816, 34821, 34833, 34836, 34881, 34884, 34896, 34978, 35073, 35076,
+ 35136, 35173, 35362, 35416, 35418, 35458, 35490, 36865, 36868, 36873, 36880, 36882, 36885, 36888, 36900, 36928,
+ 36930, 36933, 36936, 36945, 36948, 36960, 36993, 36996, 37008, 37120, 37125, 37137, 37140, 37185, 37188, 37200,
+ 37210, 37377, 37380, 37392, 37440, 37542, 37888, 37890, 37893, 37896, 37905, 37908, 37920, 37953, 37956, 37968,
+ 38016, 38038, 38145, 38148, 38160, 38208, 38296, 38305, 38400, 38470, 38500, 38913, 38916, 38928, 38950, 38976,
+ 39081, 39168, 39241, 39250, 39568, 40960, 40965, 40970, 40980, 40994, 41002, 41025, 41028, 41040, 41122, 41130,
+ 41280, 41317, 41474, 41482, 41506, 41512, 41514, 41602, 41608, 41610, 41640, 41985, 41988, 42000, 42048, 42121,
+ 42148, 42240, 42265, 42577, 43018, 43048, 43170, 43348, 43398, 43528, 43530, 43552, 43554, 43560, 43656, 43690,
+ };
+
+ const int kmap_size = 43692;
+ //const int nwant = type == GGML_TYPE_IQ1_S ? 3 : 2;
+ const int nwant = type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? 3 : type == GGML_TYPE_IQ2_S ? 1 : 2;
+ const uint16_t * kgrid = type == GGML_TYPE_IQ2_XXS ? kgrid_2bit_256 :
+ type == GGML_TYPE_IQ2_XS ? kgrid_2bit_512 :
+ type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M ? kgrid_1bit_2048 : kgrid_2bit_1024;
+ uint64_t * kgrid_q2xs;
+ int * kmap_q2xs;
+ uint16_t * kneighbors_q2xs;
+
+ //printf("================================================================= %s(grid_size = %d)\n", __func__, grid_size);
+ uint64_t * the_grid = (uint64_t *)malloc(grid_size*sizeof(uint64_t));
+ for (int k = 0; k < grid_size; ++k) {
+ int8_t * pos = (int8_t *)(the_grid + k);
+ for (int i = 0; i < 8; ++i) {
+ int l = (kgrid[k] >> 2*i) & 0x3;
+ pos[i] = 2*l + 1;
+ }
+ }
+ kgrid_q2xs = the_grid;
+ iq2_data[gindex].grid = the_grid;
+ kmap_q2xs = (int *)malloc(kmap_size*sizeof(int));
+ iq2_data[gindex].map = kmap_q2xs;
+ for (int i = 0; i < kmap_size; ++i) kmap_q2xs[i] = -1;
+ uint64_t aux64;
+ uint8_t * aux8 = (uint8_t *)&aux64;
+ for (int i = 0; i < grid_size; ++i) {
+ aux64 = kgrid_q2xs[i];
+ uint16_t index = 0;
+ for (int k=0; k<8; ++k) {
+ uint16_t q = (aux8[k] - 1)/2;
+ index |= (q << 2*k);
+ }
+ kmap_q2xs[index] = i;
+ }
+ // The neighbour search runs in three passes:
+ // 1. Parallel: for each i, qsort and count its neighbours into n_per_i,
+ // and reduce the totals (num_neighbors, num_not_in_map).
+ // 2. Serial: prefix-sum n_per_i into offsets[], so each i has a
+ // pre-assigned slice of kneighbors_q2xs to write into.
+ // 3. Parallel: redo the qsort and write each i's neighbour list at
+ // offsets[i].
+ int * n_per_i = (int *)malloc(kmap_size*sizeof(int));
+ GGML_ASSERT(n_per_i);
+ int num_neighbors = 0, num_not_in_map = 0;
+#ifdef GGML_USE_OPENMP
+ #pragma omp parallel reduction(+:num_neighbors,num_not_in_map)
+#endif
+ {
+ int * dist2 = (int *)malloc(2*grid_size*sizeof(int));
+ GGML_ASSERT(dist2);
+ int8_t pos[8];
+ int i;
+#ifdef GGML_USE_OPENMP
+ #pragma omp for schedule(dynamic, 64)
+#endif
+ for (i = 0; i < kmap_size; ++i) {
+ if (kmap_q2xs[i] >= 0) {
+ n_per_i[i] = 0;
+ continue;
+ }
+ ++num_not_in_map;
+ for (int k = 0; k < 8; ++k) {
+ int l = (i >> 2*k) & 0x3;
+ pos[k] = 2*l + 1;
+ }
+ for (int j = 0; j < grid_size; ++j) {
+ const int8_t * pg = (const int8_t *)(kgrid_q2xs + j);
+ int d2 = 0;
+ for (int k = 0; k < 8; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]);
+ dist2[2*j+0] = d2;
+ dist2[2*j+1] = j;
+ }
+ qsort(dist2, grid_size, 2*sizeof(int), iq2_compare_func);
+ int n = 0; int d2 = dist2[0];
+ int nhave = 1;
+ for (int j = 0; j < grid_size; ++j) {
+ if (dist2[2*j] > d2) {
+ if (nhave == nwant) break;
+ d2 = dist2[2*j];
+ ++nhave;
+ }
+ ++n;
+ }
+ n_per_i[i] = n;
+ num_neighbors += n;
+ }
+ free(dist2);
+ }
+ //printf("%s: %d neighbours in total\n", __func__, num_neighbors);
+ kneighbors_q2xs = (uint16_t *)malloc((num_neighbors + num_not_in_map)*sizeof(uint16_t));
+ iq2_data[gindex].neighbours = kneighbors_q2xs;
+
+ int * offsets = (int *)malloc(kmap_size*sizeof(int));
+ GGML_ASSERT(offsets);
+ int counter = 0;
+ for (int i = 0; i < kmap_size; ++i) {
+ if (kmap_q2xs[i] >= 0) {
+ offsets[i] = -1;
+ continue;
+ }
+ offsets[i] = counter;
+ counter += 1 + n_per_i[i];
+ }
+
+#ifdef GGML_USE_OPENMP
+ #pragma omp parallel
+#endif
+ {
+ int * dist2 = (int *)malloc(2*grid_size*sizeof(int));
+ GGML_ASSERT(dist2);
+ int8_t pos[8];
+ int i;
+#ifdef GGML_USE_OPENMP
+ #pragma omp for schedule(dynamic, 64)
+#endif
+ for (i = 0; i < kmap_size; ++i) {
+ if (kmap_q2xs[i] >= 0) continue;
+ for (int k = 0; k < 8; ++k) {
+ int l = (i >> 2*k) & 0x3;
+ pos[k] = 2*l + 1;
+ }
+ for (int j = 0; j < grid_size; ++j) {
+ const int8_t * pg = (const int8_t *)(kgrid_q2xs + j);
+ int d2 = 0;
+ for (int k = 0; k < 8; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]);
+ dist2[2*j+0] = d2;
+ dist2[2*j+1] = j;
+ }
+ qsort(dist2, grid_size, 2*sizeof(int), iq2_compare_func);
+ int local_counter = offsets[i];
+ kmap_q2xs[i] = -(local_counter + 1);
+ int d2 = dist2[0];
+ uint16_t * start = &kneighbors_q2xs[local_counter++];
+ int n = 0, nhave = 1;
+ for (int j = 0; j < grid_size; ++j) {
+ if (dist2[2*j] > d2) {
+ if (nhave == nwant) break;
+ d2 = dist2[2*j];
+ ++nhave;
+ }
+ kneighbors_q2xs[local_counter++] = dist2[2*j+1];
+ ++n;
+ }
+ *start = n;
+ }
+ free(dist2);
+ }
+ free(offsets);
+ free(n_per_i);
+}
+
+void iq2xs_free_impl(enum ggml_type type) {
+ GGML_ASSERT(type == GGML_TYPE_IQ2_XXS || type == GGML_TYPE_IQ2_XS || type == GGML_TYPE_IQ1_S || type == GGML_TYPE_IQ1_M || type == GGML_TYPE_IQ2_S);
+ const int gindex = iq2_data_index(type);
+ if (iq2_data[gindex].grid) {
+ free(iq2_data[gindex].grid); iq2_data[gindex].grid = NULL;
+ free(iq2_data[gindex].map); iq2_data[gindex].map = NULL;
+ free(iq2_data[gindex].neighbours); iq2_data[gindex].neighbours = NULL;
+ }
+}
+
+static int iq2_find_best_neighbour(const uint16_t * GGML_RESTRICT neighbours, const uint64_t * GGML_RESTRICT grid,
+ const float * GGML_RESTRICT xval, const float * GGML_RESTRICT weight, float scale, int8_t * GGML_RESTRICT L) {
+ int num_neighbors = neighbours[0];
+ GGML_ASSERT(num_neighbors > 0);
+ float best_d2 = FLT_MAX;
+ int grid_index = -1;
+ for (int j = 1; j <= num_neighbors; ++j) {
+ const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+ float d2 = 0;
+ for (int i = 0; i < 8; ++i) {
+ float q = pg[i];
+ float diff = scale*q - xval[i];
+ d2 += weight[i]*diff*diff;
+ }
+ if (d2 < best_d2) {
+ best_d2 = d2; grid_index = neighbours[j];
+ }
+ }
+ GGML_ASSERT(grid_index >= 0);
+ const int8_t * pg = (const int8_t *)(grid + grid_index);
+ for (int i = 0; i < 8; ++i) L[i] = (pg[i] - 1)/2;
+ return grid_index;
+}
+
+static void quantize_row_iq2_xxs_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights) {
+
+ const int gindex = iq2_data_index(GGML_TYPE_IQ2_XXS);
+
+ const uint64_t * kgrid_q2xs = iq2_data[gindex].grid;
+ const int * kmap_q2xs = iq2_data[gindex].map;
+ const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+ GGML_ASSERT(quant_weights && "missing quantization weights");
+ GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(n%QK_K == 0);
+
+ const int kMaxQ = 3;
+
+ const int64_t nbl = n/QK_K;
+
+ block_iq2_xxs * y = vy;
+
+ float scales[QK_K/32];
+ float weight[32];
+ float xval[32];
+ int8_t L[32];
+ int8_t Laux[32];
+ float waux[32];
+ uint8_t block_signs[4];
+ uint32_t q2[2*(QK_K/32)];
+
+ for (int ibl = 0; ibl < nbl; ++ibl) {
+
+ y[ibl].d = GGML_FP32_TO_FP16(0.f);
+ memset(q2, 0, QK_K/4);
+
+ float max_scale = 0;
+
+ const float * xbl = x + QK_K*ibl;
+ float sumx2 = 0;
+ for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+ float sigma2 = sumx2/QK_K;
+
+ for (int ib = 0; ib < QK_K/32; ++ib) {
+ const float * xb = xbl + 32*ib;
+ const float * qw = quant_weights + QK_K*ibl + 32*ib;
+ for (int i = 0; i < 32; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+ for (int i = 0; i < 32; ++i) waux[i] = sqrtf(weight[i]);
+ for (int k = 0; k < 4; ++k) {
+ int nflip = 0;
+ uint8_t s = 0;
+ for (int i = 0; i < 8; ++i) {
+ if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+ else {
+ xval[8*k + i] = -xb[8*k + i]; ++nflip; s |= (1 << i);
+ }
+ }
+ if (nflip%2) {
+ int imin = 0; float min = weight[8*k+imin]*xb[8*k+imin]*xb[8*k+imin];
+ for (int i = 1; i < 8; ++i) {
+ float ax = weight[8*k+i]*xb[8*k+i]*xb[8*k+i];
+ if (ax < min) {
+ min = ax; imin = i;
+ }
+ }
+ xval[8*k+imin] = -xval[8*k+imin];
+ s ^= (1 << imin);
+ }
+ block_signs[k] = s & 127;
+ }
+ float max = xval[0];
+ for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
+ if (max < GROUP_MAX_EPS) {
+ scales[ib] = 0;
+ memset(L, 0, 32);
+ continue;
+ }
+ float scale = make_qp_quants(32, kMaxQ+1, xval, (uint8_t*)L, weight);
+ float eff_max = scale*kMaxQ;
+ if (eff_max <= 0) {
+ scales[ib] = 0;
+ memset(L, 0, 32);
+ continue;
+ }
+ float best = 0;
+ for (int is = -6; is <= 6; ++is) {
+ float id = (2*kMaxQ-1+is*0.1f)/eff_max;
+ float this_scale = 1/id;
+ for (int k = 0; k < 4; ++k) {
+ for (int i = 0; i < 8; ++i) {
+ int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+ Laux[8*k+i] = MAX(0, MIN(kMaxQ-1, l));
+ }
+ uint16_t u = 0;
+ for (int i = 0; i < 8; ++i) u |= (Laux[8*k+i] << 2*i);
+ int grid_index = kmap_q2xs[u];
+ if (grid_index < 0) {
+ const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+ grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, this_scale, Laux + 8*k);
+ }
+ }
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < 32; ++i) {
+ float w = weight[i];
+ float q = 2*Laux[i] + 1;
+ sumqx += w*xval[i]*q;
+ sumq2 += w*q*q;
+ }
+ if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+ scale = sumqx/sumq2; best = scale*sumqx;
+ memcpy(L, Laux, 32);
+ }
+ }
+ if (scale > 0) {
+ float id = 1/scale;
+ for (int k = 0; k < 4; ++k) {
+ uint16_t u = 0;
+ for (int i = 0; i < 8; ++i) {
+ int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+ l = MAX(0, MIN(kMaxQ-1, l));
+ u |= (l << 2*i);
+ }
+ int grid_index = kmap_q2xs[u];
+ if (grid_index < 0) {
+ const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+ grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, scale, L + 8*k);
+ }
+ const int8_t * pg = (const int8_t *)(kgrid_q2xs + grid_index);
+ for (int i = 0; i < 8; ++i) L[8*k+i] = (pg[i] - 1)/2;
+ }
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < 32; ++i) {
+ float w = weight[i];
+ float q = 2*L[i] + 1;
+ sumqx += w*xval[i]*q;
+ sumq2 += w*q*q;
+ }
+ if (sumq2 > 0) scale = sumqx/sumq2;
+ }
+ if (scale < 0) {
+ // This should never happen, but just in case, flip scale so that it is positive (we use uint's to encode the scale)
+ // and correspondingly flip quant signs.
+ scale = -scale;
+ for (int k = 0; k < 4; ++k) block_signs[k] = (~block_signs[k]) & 127;
+ }
+ for (int k = 0; k < 4; ++k) {
+ uint16_t u = 0;
+ for (int i = 0; i < 8; ++i) u |= (L[8*k+i] << 2*i);
+ int grid_index = kmap_q2xs[u];
+ if (grid_index < 0) {
+ printf("Oops: found point %u not on grid:", u);
+ for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
+ printf("\n");
+ GGML_ABORT("fatal error");
+ }
+ q2[2*ib+0] |= ((uint32_t) grid_index << 8*k);
+ q2[2*ib+1] |= (block_signs[k] << 7*k);
+ }
+ GGML_ASSERT(scale >= 0);
+ scales[ib] = scale;
+ max_scale = MAX(max_scale, scale);
+ }
+
+ if (!max_scale) {
+ memset(y[ibl].qs, 0, QK_K/4);
+ continue;
+ }
+
+ float d = max_scale/31;
+ y[ibl].d = GGML_FP32_TO_FP16(d);
+ float id = 1/d;
+ for (int ib = 0; ib < QK_K/32; ++ib) {
+ int l = nearest_int(0.5f*(id*scales[ib]-1));
+ l = MAX(0, MIN(15, l));
+ q2[2*ib+1] |= ((uint32_t)l << 28);
+ }
+ memcpy(y[ibl].qs, q2, QK_K/4);
+ }
+}
+
+static void quantize_row_iq2_xs_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights) {
+
+ const int gindex = iq2_data_index(GGML_TYPE_IQ2_XS);
+
+ const uint64_t * kgrid_q2xs = iq2_data[gindex].grid;
+ const int * kmap_q2xs = iq2_data[gindex].map;
+ const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+ GGML_ASSERT(quant_weights && "missing quantization weights");
+ GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(n%QK_K == 0);
+
+ const int kMaxQ = 3;
+
+ const int64_t nbl = n/QK_K;
+
+ block_iq2_xs * y = vy;
+
+ float scales[QK_K/16];
+ float weight[16];
+ float xval[16];
+ int8_t L[16];
+ int8_t Laux[16];
+ float waux[16];
+ bool is_on_grid[2];
+ bool is_on_grid_aux[2];
+ uint8_t block_signs[2];
+ uint16_t q2[2*(QK_K/16)];
+
+ for (int ibl = 0; ibl < nbl; ++ibl) {
+
+ y[ibl].d = GGML_FP32_TO_FP16(0.f);
+ memset(q2, 0, QK_K/4);
+ memset(y[ibl].scales, 0, QK_K/32);
+
+ float max_scale = 0;
+
+ const float * xbl = x + QK_K*ibl;
+ float sumx2 = 0;
+ for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+ float sigma2 = sumx2/QK_K;
+
+ for (int ib = 0; ib < QK_K/16; ++ib) {
+ const float * xb = xbl + 16*ib;
+ const float * qw = quant_weights + QK_K*ibl + 16*ib;
+ for (int i = 0; i < 16; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+ for (int i = 0; i < 16; ++i) waux[i] = sqrtf(weight[i]);
+ for (int k = 0; k < 2; ++k) {
+ int nflip = 0;
+ uint8_t s = 0;
+ for (int i = 0; i < 8; ++i) {
+ if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+ else {
+ xval[8*k + i] = -xb[8*k + i]; ++nflip; s |= (1 << i);
+ }
+ }
+ if (nflip%2) {
+ int imin = 0; float min = weight[8*k+imin]*xb[8*k+imin]*xb[8*k+imin];
+ for (int i = 1; i < 8; ++i) {
+ float ax = weight[8*k+i]*xb[8*k+i]*xb[8*k+i];
+ if (ax < min) {
+ min = ax; imin = i;
+ }
+ }
+ xval[8*k+imin] = -xval[8*k+imin];
+ s ^= (1 << imin);
+ }
+ block_signs[k] = s & 127;
+ }
+ float max = xval[0];
+ for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
+ memset(L, 0, 16);
+ if (max < GROUP_MAX_EPS) {
+ scales[ib] = 0;
+ continue;
+ }
+ float best = 0;
+ float scale = max/(2*kMaxQ-1);
+ is_on_grid[0] = is_on_grid[1] = true;
+ for (int is = -9; is <= 9; ++is) {
+ float id = (2*kMaxQ-1+is*0.1f)/max;
+ float this_scale = 1/id;
+ for (int k = 0; k < 2; ++k) {
+ for (int i = 0; i < 8; ++i) {
+ int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+ Laux[8*k+i] = MAX(0, MIN(kMaxQ-1, l));
+ }
+ uint16_t u = 0;
+ for (int i = 0; i < 8; ++i) u |= (Laux[8*k+i] << 2*i);
+ int grid_index = kmap_q2xs[u];
+ is_on_grid_aux[k] = true;
+ if (grid_index < 0) {
+ is_on_grid_aux[k] = false;
+ const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+ grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, this_scale, Laux + 8*k);
+ }
+ }
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < 16; ++i) {
+ float w = weight[i];
+ float q = 2*Laux[i] + 1;
+ sumqx += w*xval[i]*q;
+ sumq2 += w*q*q;
+ }
+ if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+ scale = sumqx/sumq2; best = scale*sumqx;
+ for (int i = 0; i < 16; ++i) L[i] = Laux[i];
+ for (int k = 0; k < 2; ++k) is_on_grid[k] = is_on_grid_aux[k];
+ }
+ }
+ int n_not_ongrid = 0;
+ for (int k = 0; k < 2; ++k) if (!is_on_grid[k]) ++n_not_ongrid;
+ if (n_not_ongrid > 0 && scale > 0) {
+ float id = 1/scale;
+ for (int k = 0; k < 2; ++k) {
+ if (is_on_grid[k]) continue;
+ uint16_t u = 0;
+ for (int i = 0; i < 8; ++i) {
+ int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+ l = MAX(0, MIN(kMaxQ-1, l));
+ u |= (l << 2*i);
+ L[8*k + i] = l;
+ }
+ int grid_index = kmap_q2xs[u];
+ if (grid_index < 0) {
+ const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+ grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, scale, L + 8*k);
+ }
+ }
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < 16; ++i) {
+ float w = weight[i];
+ float q = 2*L[i] + 1;
+ sumqx += w*xval[i]*q;
+ sumq2 += w*q*q;
+ }
+ if (sumq2 > 0) scale = sumqx/sumq2;
+ }
+ if (scale < 0) {
+ scale = -scale;
+ for (int k = 0; k < 2; ++k) block_signs[k] = (~block_signs[k]) & 127;
+ }
+ for (int k = 0; k < 2; ++k) {
+ uint16_t u = 0;
+ for (int i = 0; i < 8; ++i) u |= (L[8*k+i] << 2*i);
+ int grid_index = kmap_q2xs[u];
+ if (grid_index < 0) {
+ printf("Oops: found point %u not on grid:", u);
+ for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
+ printf("\n");
+ GGML_ABORT("fatal error");
+ }
+ q2[2*ib+k] = grid_index | (block_signs[k] << 9);
+ }
+ GGML_ASSERT(scale >= 0);
+ scales[ib] = scale;
+ max_scale = MAX(max_scale, scale);
+ }
+
+ if (!max_scale) {
+ memset(y[ibl].qs, 0, QK_K/4);
+ continue;
+ }
+
+ float d = max_scale/31;
+ y[ibl].d = GGML_FP32_TO_FP16(d);
+ float id = 1/d;
+ for (int ib = 0; ib < QK_K/16; ++ib) {
+ int l = nearest_int(0.5f*(id*scales[ib]-1));
+ l = MAX(0, MIN(15, l));
+ if (ib%2 == 0) y[ibl].scales[ib/2] = l;
+ else y[ibl].scales[ib/2] |= (l << 4);
+ }
+ memcpy(y[ibl].qs, q2, QK_K/4);
+
+ }
+}
+
+size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ GGML_ASSERT(n_per_row%QK_K == 0);
+ int64_t nblock = n_per_row/QK_K;
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_iq2_xxs_impl(src, qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += nblock*sizeof(block_iq2_xxs);
+ }
+ return nrow * nblock * sizeof(block_iq2_xxs);
+}
+
+size_t quantize_iq2_xs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ GGML_ASSERT(n_per_row%QK_K == 0);
+ int64_t nblock = n_per_row/QK_K;
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_iq2_xs_impl(src, qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += nblock*sizeof(block_iq2_xs);
+ }
+ return nrow * nblock * sizeof(block_iq2_xs);
+}
+
+//
+// ============================================= 3-bit using D4 lattice
+//
+
+typedef struct {
+ uint32_t * grid;
+ int * map;
+ uint16_t * neighbours;
+} iq3_entry_t;
+
+static iq3_entry_t iq3_data[2] = {
+ {NULL, NULL, NULL},
+ {NULL, NULL, NULL},
+};
+
+static inline int iq3_data_index(int grid_size) {
+ (void)grid_size;
+ GGML_ASSERT(grid_size == 256 || grid_size == 512);
+ return grid_size == 256 ? 0 : 1;
+}
+
+static int iq3_compare_func(const void * left, const void * right) {
+ const int * l = (const int *)left;
+ const int * r = (const int *)right;
+ return l[0] < r[0] ? -1 : l[0] > r[0] ? 1 : l[1] < r[1] ? -1 : l[1] > r[1] ? 1 : 0;
+}
+
+void iq3xs_init_impl(int grid_size) {
+ const int gindex = iq3_data_index(grid_size);
+ if (iq3_data[gindex].grid) {
+ return;
+ }
+ static const uint16_t kgrid_256[256] = {
+ 0, 2, 4, 9, 11, 15, 16, 18, 25, 34, 59, 61, 65, 67, 72, 74,
+ 81, 85, 88, 90, 97, 108, 120, 128, 130, 132, 137, 144, 146, 153, 155, 159,
+ 169, 175, 189, 193, 199, 200, 202, 213, 248, 267, 287, 292, 303, 315, 317, 321,
+ 327, 346, 362, 413, 436, 456, 460, 462, 483, 497, 513, 515, 520, 522, 529, 531,
+ 536, 538, 540, 551, 552, 576, 578, 585, 592, 594, 641, 643, 648, 650, 657, 664,
+ 698, 704, 706, 720, 729, 742, 758, 769, 773, 808, 848, 852, 870, 889, 901, 978,
+ 992, 1024, 1026, 1033, 1035, 1040, 1042, 1046, 1049, 1058, 1089, 1091, 1093, 1096, 1098, 1105,
+ 1112, 1139, 1143, 1144, 1152, 1154, 1161, 1167, 1168, 1170, 1183, 1184, 1197, 1217, 1224, 1228,
+ 1272, 1276, 1309, 1323, 1347, 1367, 1377, 1404, 1473, 1475, 1486, 1509, 1537, 1544, 1546, 1553,
+ 1555, 1576, 1589, 1594, 1600, 1602, 1616, 1625, 1636, 1638, 1665, 1667, 1672, 1685, 1706, 1722,
+ 1737, 1755, 1816, 1831, 1850, 1856, 1862, 1874, 1901, 1932, 1950, 1971, 2011, 2032, 2052, 2063,
+ 2077, 2079, 2091, 2095, 2172, 2192, 2207, 2208, 2224, 2230, 2247, 2277, 2308, 2345, 2356, 2389,
+ 2403, 2424, 2501, 2504, 2506, 2520, 2570, 2593, 2616, 2624, 2630, 2646, 2669, 2700, 2714, 2746,
+ 2754, 2795, 2824, 2835, 2839, 2874, 2882, 2905, 2984, 3028, 3042, 3092, 3108, 3110, 3124, 3153,
+ 3185, 3215, 3252, 3288, 3294, 3364, 3397, 3434, 3483, 3523, 3537, 3587, 3589, 3591, 3592, 3610,
+ 3626, 3670, 3680, 3722, 3749, 3754, 3776, 3789, 3803, 3824, 3857, 3873, 3904, 3906, 3924, 3992,
+ };
+ static const uint16_t kgrid_512[512] = {
+ 0, 1, 2, 5, 7, 8, 9, 10, 12, 14, 16, 17, 21, 27, 32, 34,
+ 37, 39, 41, 43, 48, 50, 57, 60, 63, 64, 65, 66, 68, 72, 73, 77,
+ 80, 83, 87, 89, 93, 100, 113, 117, 122, 128, 129, 133, 135, 136, 139, 142,
+ 145, 149, 152, 156, 162, 165, 167, 169, 171, 184, 187, 195, 201, 205, 208, 210,
+ 217, 219, 222, 228, 232, 234, 247, 249, 253, 256, 267, 271, 273, 276, 282, 288,
+ 291, 297, 312, 322, 324, 336, 338, 342, 347, 353, 357, 359, 374, 379, 390, 393,
+ 395, 409, 426, 441, 448, 450, 452, 464, 466, 470, 475, 488, 492, 512, 513, 514,
+ 516, 520, 521, 523, 525, 527, 528, 530, 537, 540, 542, 556, 558, 561, 570, 576,
+ 577, 579, 582, 584, 588, 593, 600, 603, 609, 616, 618, 632, 638, 640, 650, 653,
+ 655, 656, 660, 666, 672, 675, 685, 688, 698, 705, 708, 711, 712, 715, 721, 727,
+ 728, 732, 737, 754, 760, 771, 773, 778, 780, 793, 795, 802, 806, 808, 812, 833,
+ 840, 843, 849, 856, 858, 873, 912, 916, 919, 932, 934, 961, 963, 968, 970, 977,
+ 989, 993, 1010, 1016, 1024, 1025, 1027, 1029, 1031, 1032, 1034, 1036, 1038, 1041, 1043, 1047,
+ 1048, 1050, 1057, 1059, 1061, 1064, 1066, 1079, 1080, 1083, 1085, 1088, 1090, 1096, 1099, 1103,
+ 1106, 1109, 1113, 1116, 1122, 1129, 1153, 1156, 1159, 1169, 1171, 1176, 1183, 1185, 1195, 1199,
+ 1209, 1212, 1216, 1218, 1221, 1225, 1234, 1236, 1241, 1243, 1250, 1256, 1270, 1281, 1287, 1296,
+ 1299, 1306, 1309, 1313, 1338, 1341, 1348, 1353, 1362, 1375, 1376, 1387, 1400, 1408, 1410, 1415,
+ 1425, 1453, 1457, 1477, 1481, 1494, 1496, 1507, 1512, 1538, 1545, 1547, 1549, 1551, 1554, 1561,
+ 1563, 1565, 1570, 1572, 1575, 1577, 1587, 1593, 1601, 1603, 1605, 1612, 1617, 1619, 1632, 1648,
+ 1658, 1662, 1664, 1674, 1680, 1690, 1692, 1704, 1729, 1736, 1740, 1745, 1747, 1751, 1752, 1761,
+ 1763, 1767, 1773, 1787, 1795, 1801, 1806, 1810, 1817, 1834, 1840, 1844, 1857, 1864, 1866, 1877,
+ 1882, 1892, 1902, 1915, 1934, 1953, 1985, 1987, 2000, 2002, 2013, 2048, 2052, 2058, 2064, 2068,
+ 2071, 2074, 2081, 2088, 2104, 2114, 2119, 2121, 2123, 2130, 2136, 2141, 2147, 2153, 2157, 2177,
+ 2179, 2184, 2189, 2193, 2203, 2208, 2223, 2226, 2232, 2244, 2249, 2251, 2256, 2258, 2265, 2269,
+ 2304, 2306, 2324, 2335, 2336, 2361, 2373, 2375, 2385, 2418, 2443, 2460, 2480, 2504, 2509, 2520,
+ 2531, 2537, 2562, 2568, 2572, 2578, 2592, 2596, 2599, 2602, 2614, 2620, 2625, 2627, 2629, 2634,
+ 2641, 2650, 2682, 2688, 2697, 2707, 2712, 2718, 2731, 2754, 2759, 2760, 2775, 2788, 2793, 2805,
+ 2811, 2817, 2820, 2832, 2842, 2854, 2890, 2902, 2921, 2923, 2978, 3010, 3012, 3026, 3081, 3083,
+ 3085, 3097, 3099, 3120, 3136, 3152, 3159, 3188, 3210, 3228, 3234, 3245, 3250, 3256, 3264, 3276,
+ 3281, 3296, 3349, 3363, 3378, 3392, 3395, 3420, 3440, 3461, 3488, 3529, 3531, 3584, 3588, 3591,
+ 3600, 3602, 3614, 3616, 3628, 3634, 3650, 3657, 3668, 3683, 3685, 3713, 3716, 3720, 3726, 3729,
+ 3736, 3753, 3778, 3802, 3805, 3819, 3841, 3845, 3851, 3856, 3880, 3922, 3938, 3970, 3993, 4032,
+ };
+
+ const int kmap_size = 4096;
+ const int nwant = grid_size == 256 ? 2 : 3;
+ const uint16_t * kgrid = grid_size == 256 ? kgrid_256 : kgrid_512;
+ uint32_t * kgrid_q3xs;
+ int * kmap_q3xs;
+ uint16_t * kneighbors_q3xs;
+
+ //printf("================================================================= %s(grid_size = %d)\n", __func__, grid_size);
+ uint32_t * the_grid = (uint32_t *)malloc(grid_size*sizeof(uint32_t));
+ for (int k = 0; k < grid_size; ++k) {
+ int8_t * pos = (int8_t *)(the_grid + k);
+ for (int i = 0; i < 4; ++i) {
+ int l = (kgrid[k] >> 3*i) & 0x7;
+ pos[i] = 2*l + 1;
+ }
+ }
+ kgrid_q3xs = the_grid;
+ iq3_data[gindex].grid = the_grid;
+ kmap_q3xs = (int *)malloc(kmap_size*sizeof(int));
+ iq3_data[gindex].map = kmap_q3xs;
+ for (int i = 0; i < kmap_size; ++i) kmap_q3xs[i] = -1;
+ uint32_t aux32;
+ uint8_t * aux8 = (uint8_t *)&aux32;
+ for (int i = 0; i < grid_size; ++i) {
+ aux32 = kgrid_q3xs[i];
+ uint16_t index = 0;
+ for (int k=0; k<4; ++k) {
+ uint16_t q = (aux8[k] - 1)/2;
+ index |= (q << 3*k);
+ }
+ kmap_q3xs[index] = i;
+ }
+ // See explanation of parallelism in iq2xs_init_impl
+ int * n_per_i = (int *)malloc(kmap_size*sizeof(int));
+ GGML_ASSERT(n_per_i);
+ int num_neighbors = 0, num_not_in_map = 0;
+#ifdef GGML_USE_OPENMP
+ #pragma omp parallel reduction(+:num_neighbors,num_not_in_map)
+#endif
+ {
+ int * dist2 = (int *)malloc(2*grid_size*sizeof(int));
+ GGML_ASSERT(dist2);
+ int8_t pos[4];
+ int i;
+#ifdef GGML_USE_OPENMP
+ #pragma omp for schedule(dynamic, 64)
+#endif
+ for (i = 0; i < kmap_size; ++i) {
+ if (kmap_q3xs[i] >= 0) {
+ n_per_i[i] = 0;
+ continue;
+ }
+ ++num_not_in_map;
+ for (int k = 0; k < 4; ++k) {
+ int l = (i >> 3*k) & 0x7;
+ pos[k] = 2*l + 1;
+ }
+ for (int j = 0; j < grid_size; ++j) {
+ const int8_t * pg = (const int8_t *)(kgrid_q3xs + j);
+ int d2 = 0;
+ for (int k = 0; k < 4; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]);
+ dist2[2*j+0] = d2;
+ dist2[2*j+1] = j;
+ }
+ qsort(dist2, grid_size, 2*sizeof(int), iq3_compare_func);
+ int n = 0; int d2 = dist2[0];
+ int nhave = 1;
+ for (int j = 0; j < grid_size; ++j) {
+ if (dist2[2*j] > d2) {
+ if (nhave == nwant) break;
+ d2 = dist2[2*j];
+ ++nhave;
+ }
+ ++n;
+ }
+ n_per_i[i] = n;
+ num_neighbors += n;
+ }
+ free(dist2);
+ }
+ //printf("%s: %d neighbours in total\n", __func__, num_neighbors);
+ kneighbors_q3xs = (uint16_t *)malloc((num_neighbors + num_not_in_map)*sizeof(uint16_t));
+ iq3_data[gindex].neighbours = kneighbors_q3xs;
+
+ int * offsets = (int *)malloc(kmap_size*sizeof(int));
+ GGML_ASSERT(offsets);
+ int counter = 0;
+ for (int i = 0; i < kmap_size; ++i) {
+ if (kmap_q3xs[i] >= 0) {
+ offsets[i] = -1;
+ continue;
+ }
+ offsets[i] = counter;
+ counter += 1 + n_per_i[i];
+ }
+
+#ifdef GGML_USE_OPENMP
+ #pragma omp parallel
+#endif
+ {
+ int * dist2 = (int *)malloc(2*grid_size*sizeof(int));
+ GGML_ASSERT(dist2);
+ int8_t pos[4];
+ int i;
+#ifdef GGML_USE_OPENMP
+ #pragma omp for schedule(dynamic, 64)
+#endif
+ for (i = 0; i < kmap_size; ++i) {
+ if (kmap_q3xs[i] >= 0) continue;
+ for (int k = 0; k < 4; ++k) {
+ int l = (i >> 3*k) & 0x7;
+ pos[k] = 2*l + 1;
+ }
+ for (int j = 0; j < grid_size; ++j) {
+ const int8_t * pg = (const int8_t *)(kgrid_q3xs + j);
+ int d2 = 0;
+ for (int k = 0; k < 4; ++k) d2 += (pg[k] - pos[k])*(pg[k] - pos[k]);
+ dist2[2*j+0] = d2;
+ dist2[2*j+1] = j;
+ }
+ qsort(dist2, grid_size, 2*sizeof(int), iq3_compare_func);
+ int local_counter = offsets[i];
+ kmap_q3xs[i] = -(local_counter + 1);
+ int d2 = dist2[0];
+ uint16_t * start = &kneighbors_q3xs[local_counter++];
+ int n = 0, nhave = 1;
+ for (int j = 0; j < grid_size; ++j) {
+ if (dist2[2*j] > d2) {
+ if (nhave == nwant) break;
+ d2 = dist2[2*j];
+ ++nhave;
+ }
+ kneighbors_q3xs[local_counter++] = dist2[2*j+1];
+ ++n;
+ }
+ *start = n;
+ }
+ free(dist2);
+ }
+ free(offsets);
+ free(n_per_i);
+}
+
+void iq3xs_free_impl(int grid_size) {
+ GGML_ASSERT(grid_size == 256 || grid_size == 512);
+ const int gindex = iq3_data_index(grid_size);
+ if (iq3_data[gindex].grid) {
+ free(iq3_data[gindex].grid); iq3_data[gindex].grid = NULL;
+ free(iq3_data[gindex].map); iq3_data[gindex].map = NULL;
+ free(iq3_data[gindex].neighbours); iq3_data[gindex].neighbours = NULL;
+ }
+}
+
+static int iq3_find_best_neighbour(const uint16_t * GGML_RESTRICT neighbours, const uint32_t * GGML_RESTRICT grid,
+ const float * GGML_RESTRICT xval, const float * GGML_RESTRICT weight, float scale, int8_t * GGML_RESTRICT L) {
+ int num_neighbors = neighbours[0];
+ GGML_ASSERT(num_neighbors > 0);
+ float best_d2 = FLT_MAX;
+ int grid_index = -1;
+ for (int j = 1; j <= num_neighbors; ++j) {
+ const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+ float d2 = 0;
+ for (int i = 0; i < 4; ++i) {
+ float q = pg[i];
+ float diff = scale*q - xval[i];
+ d2 += weight[i]*diff*diff;
+ }
+ if (d2 < best_d2) {
+ best_d2 = d2; grid_index = neighbours[j];
+ }
+ }
+ GGML_ASSERT(grid_index >= 0);
+ const int8_t * pg = (const int8_t *)(grid + grid_index);
+ for (int i = 0; i < 4; ++i) L[i] = (pg[i] - 1)/2;
+ return grid_index;
+}
+
+static void quantize_row_iq3_xxs_impl(int grid_size, const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n,
+ const float * GGML_RESTRICT quant_weights) {
+
+ const int gindex = iq3_data_index(grid_size);
+
+ const uint32_t * kgrid_q3xs = iq3_data[gindex].grid;
+ const int * kmap_q3xs = iq3_data[gindex].map;
+ const uint16_t * kneighbors_q3xs = iq3_data[gindex].neighbours;
+
+ //GGML_ASSERT(quant_weights && "missing quantization weights");
+ GGML_ASSERT(kgrid_q3xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kmap_q3xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kneighbors_q3xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(n%QK_K == 0);
+
+ const int kMaxQ = 8;
+
+ const int64_t nbl = n/QK_K;
+
+ ggml_fp16_t * dh;
+ uint8_t * qs;
+ int block_size;
+ if (grid_size == 256) {
+ block_iq3_xxs * y = vy;
+ dh = &y->d;
+ qs = y->qs;
+ block_size = sizeof(block_iq3_xxs);
+ } else {
+ block_iq3_s * y = vy;
+ dh = &y->d;
+ qs = y->qs;
+ block_size = sizeof(block_iq3_s);
+ }
+ int quant_size = block_size - sizeof(ggml_fp16_t);
+
+ float scales[QK_K/32];
+ float weight[32];
+ float xval[32];
+ int8_t L[32];
+ int8_t Laux[32];
+ float waux[32];
+ bool is_on_grid[8];
+ bool is_on_grid_aux[8];
+ uint8_t block_signs[8];
+ uint8_t q3[3*(QK_K/8)+QK_K/32];
+ uint32_t * scales_and_signs = (uint32_t *)(q3 + QK_K/4);
+ uint8_t * qh = q3 + 3*(QK_K/8);
+
+ for (int ibl = 0; ibl < nbl; ++ibl) {
+
+ dh[0] = GGML_FP32_TO_FP16(0.f);
+ memset(q3, 0, 3*QK_K/8+QK_K/32);
+
+ float max_scale = 0;
+
+ const float * xbl = x + QK_K*ibl;
+ float sumx2 = 0;
+ for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+ float sigma2 = 2*sumx2/QK_K;
+
+ for (int ib = 0; ib < QK_K/32; ++ib) {
+ const float * xb = xbl + 32*ib;
+ if (quant_weights) {
+ const float * qw = quant_weights + QK_K*ibl + 32*ib;
+ for (int i = 0; i < 32; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+ } else {
+ for (int i = 0; i < 32; ++i) weight[i] = xb[i]*xb[i];
+ }
+ for (int i = 0; i < 32; ++i) waux[i] = sqrtf(weight[i]);
+ for (int k = 0; k < 4; ++k) {
+ int nflip = 0;
+ uint8_t s = 0;
+ for (int i = 0; i < 8; ++i) {
+ if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+ else {
+ xval[8*k + i] = -xb[8*k + i]; ++nflip; s |= (1 << i);
+ }
+ }
+ if (nflip%2) {
+ int imin = 0; float min = weight[8*k+imin]*xb[8*k+imin]*xb[8*k+imin];
+ for (int i = 1; i < 8; ++i) {
+ float ax = weight[8*k+i]*xb[8*k+i]*xb[8*k+i];
+ if (ax < min) {
+ min = ax; imin = i;
+ }
+ }
+ xval[8*k+imin] = -xval[8*k+imin];
+ s ^= (1 << imin);
+ }
+ block_signs[k] = s & 127;
+ }
+ float max = xval[0];
+ for (int i = 1; i < 32; ++i) max = MAX(max, xval[i]);
+ memset(L, 0, 32);
+ if (max < GROUP_MAX_EPS_IQ3_XXS) {
+ scales[ib] = 0;
+ continue;
+ }
+ float best = 0;
+ float scale = max/(2*kMaxQ-1);
+ for (int k = 0; k < 8; ++k) is_on_grid[k] = true;
+ for (int is = -15; is <= 15; ++is) {
+ float id = (2*kMaxQ-1+is*0.2f)/max;
+ float this_scale = 1/id;
+ for (int k = 0; k < 8; ++k) {
+ for (int i = 0; i < 4; ++i) {
+ int l = nearest_int(0.5f*(id*xval[4*k+i]-1));
+ Laux[4*k+i] = MAX(0, MIN(kMaxQ-1, l));
+ }
+ uint16_t u = 0;
+ for (int i = 0; i < 4; ++i) u |= (Laux[4*k+i] << 3*i);
+ int grid_index = kmap_q3xs[u];
+ is_on_grid_aux[k] = true;
+ if (grid_index < 0) {
+ is_on_grid_aux[k] = false;
+ const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1;
+ grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, this_scale, Laux + 4*k);
+ }
+ }
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < 32; ++i) {
+ float w = weight[i];
+ float q = 2*Laux[i] + 1;
+ sumqx += w*xval[i]*q;
+ sumq2 += w*q*q;
+ }
+ if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+ scale = sumqx/sumq2; best = scale*sumqx;
+ for (int i = 0; i < 32; ++i) L[i] = Laux[i];
+ for (int k = 0; k < 8; ++k) is_on_grid[k] = is_on_grid_aux[k];
+ }
+ }
+ int n_not_ongrid = 0;
+ for (int k = 0; k < 8; ++k) if (!is_on_grid[k]) ++n_not_ongrid;
+ if (n_not_ongrid > 0 && scale > 0) {
+ float id = 1/scale;
+ for (int k = 0; k < 8; ++k) {
+ if (is_on_grid[k]) continue;
+ uint16_t u = 0;
+ for (int i = 0; i < 4; ++i) {
+ int l = nearest_int(0.5f*(id*xval[4*k+i]-1));
+ l = MAX(0, MIN(kMaxQ-1, l));
+ u |= (l << 3*i);
+ }
+ int grid_index = kmap_q3xs[u];
+ if (grid_index < 0) {
+ const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1;
+ grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, scale, L + 4*k);
+ }
+ const int8_t * pg = (const int8_t *)(kgrid_q3xs + grid_index);
+ for (int i = 0; i < 4; ++i) L[4*k+i] = (pg[i] - 1)/2;
+ }
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < 32; ++i) {
+ float w = weight[i];
+ float q = 2*L[i] + 1;
+ sumqx += w*xval[i]*q;
+ sumq2 += w*q*q;
+ }
+ if (sumq2 > 0) scale = sumqx/sumq2;
+ }
+ if (scale < 0) {
+ // This should never happen, but just in case, flip scale so that it is positive (we use uint's to encode the scale)
+ // and correspondingly flip quant signs.
+ scale = -scale;
+ for (int k = 0; k < 4; ++k) block_signs[k] = (~block_signs[k]) & 127;
+ }
+ for (int k = 0; k < 8; ++k) {
+ uint16_t u = 0;
+ for (int i = 0; i < 4; ++i) u |= (L[4*k+i] << 3*i);
+ int grid_index = kmap_q3xs[u];
+ if (grid_index < 0) {
+ printf("Oops: found point %u not on grid:", u);
+ for (int i = 0; i < 4; ++i) printf(" %d", L[4*k+i]);
+ printf("\n");
+ GGML_ABORT("fatal error");
+ }
+ if (grid_size == 256) {
+ q3[8*ib+k] = grid_index;
+ } else {
+ q3[8*ib+k] = grid_index & 255;
+ qh[ib] |= ((grid_index >> 8) << k);
+ }
+
+ }
+ scales_and_signs[ib] = block_signs[0] | (block_signs[1] << 7) | (block_signs[2] << 14) | (block_signs[3] << 21);
+ GGML_ASSERT(scale >= 0);
+ scales[ib] = scale;
+ max_scale = MAX(max_scale, scale);
+ }
+
+ if (!max_scale) {
+ memset(qs, 0, quant_size);
+ dh += block_size/sizeof(ggml_fp16_t);
+ qs += block_size;
+ continue;
+ }
+
+ float d = max_scale/31;
+ dh[0] = GGML_FP32_TO_FP16(d * 1.0125f); // small improvement via this fudge factor
+ float id = 1/d;
+ for (int ib = 0; ib < QK_K/32; ++ib) {
+ int l = nearest_int(0.5f*(id*scales[ib]-1));
+ l = MAX(0, MIN(15, l));
+ scales_and_signs[ib] |= ((uint32_t)l << 28);
+ }
+ memcpy(qs, q3, quant_size);
+
+ dh += block_size/sizeof(ggml_fp16_t);
+ qs += block_size;
+
+ }
+}
+
+size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ GGML_ASSERT(n_per_row%QK_K == 0);
+ int64_t nblock = n_per_row/QK_K;
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_iq3_xxs_impl(256, src, qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += nblock*sizeof(block_iq3_xxs);
+ }
+ return nrow * nblock * sizeof(block_iq3_xxs);
+}
+
+void quantize_row_iq3_xxs_ref(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ quantize_row_iq3_xxs_impl(256, x, y, k, NULL);
+}
+
+static void quantize_row_iq3_s_impl(int block_size, const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int n,
+ const float * GGML_RESTRICT quant_weights,
+ float * scales,
+ float * weight,
+ float * xval,
+ int8_t * L,
+ int8_t * Laux,
+ float * waux,
+ bool * is_on_grid,
+ bool * is_on_grid_aux,
+ uint8_t * block_signs) {
+
+ const int gindex = iq3_data_index(512);
+
+ const uint32_t * kgrid_q3xs = iq3_data[gindex].grid;
+ const int * kmap_q3xs = iq3_data[gindex].map;
+ const uint16_t * kneighbors_q3xs = iq3_data[gindex].neighbours;
+
+ //GGML_ASSERT(quant_weights && "missing quantization weights");
+ GGML_ASSERT(kgrid_q3xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kmap_q3xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kneighbors_q3xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(n%QK_K == 0);
+
+ const int kMaxQ = 8;
+
+ const int64_t nbl = n/QK_K;
+
+ block_iq3_s * y = vy;
+
+ const int bs4 = block_size/4;
+ const int bs8 = block_size/8;
+
+ for (int ibl = 0; ibl < nbl; ++ibl) {
+
+ memset(&y[ibl], 0, sizeof(block_iq3_s));
+ y[ibl].d = GGML_FP32_TO_FP16(0.f);
+
+ uint8_t * qs = y[ibl].qs;
+ uint8_t * qh = y[ibl].qh;
+ uint8_t * signs = y[ibl].signs;
+
+ float max_scale = 0;
+
+ const float * xbl = x + QK_K*ibl;
+ float sumx2 = 0;
+ for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+ float sigma2 = 2*sumx2/QK_K;
+
+ for (int ib = 0; ib < QK_K/block_size; ++ib) {
+ const float * xb = xbl + block_size*ib;
+ if (quant_weights) {
+ const float * qw = quant_weights + QK_K*ibl + block_size*ib;
+ for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+ } else {
+ for (int i = 0; i < block_size; ++i) weight[i] = xb[i]*xb[i];
+ }
+ for (int i = 0; i < block_size; ++i) waux[i] = sqrtf(weight[i]);
+ for (int k = 0; k < bs8; ++k) {
+ uint8_t s = 0;
+ for (int i = 0; i < 8; ++i) {
+ if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+ else {
+ xval[8*k + i] = -xb[8*k + i]; s |= (1 << i);
+ }
+ }
+ block_signs[k] = s;
+ }
+ float max = xval[0];
+ for (int i = 1; i < block_size; ++i) max = MAX(max, xval[i]);
+ memset(L, 0, block_size);
+ if (!max) {
+ scales[ib] = 0;
+ continue;
+ }
+ float best = 0;
+ float scale = max/(2*kMaxQ-1);
+ for (int k = 0; k < bs4; ++k) is_on_grid[k] = false;
+ for (int is = -9; is <= 9; ++is) {
+ float id = (2*kMaxQ-1+is*0.2f)/max;
+ float this_scale = 1/id;
+ for (int k = 0; k < bs4; ++k) {
+ for (int i = 0; i < 4; ++i) {
+ int l = nearest_int(0.5f*(id*xval[4*k+i]-1));
+ Laux[4*k+i] = MAX(0, MIN(kMaxQ-1, l));
+ }
+ uint16_t u = 0;
+ for (int i = 0; i < 4; ++i) u |= (Laux[4*k+i] << 3*i);
+ int grid_index = kmap_q3xs[u];
+ is_on_grid_aux[k] = true;
+ if (grid_index < 0) {
+ is_on_grid_aux[k] = false;
+ const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1;
+ grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, this_scale, Laux + 4*k);
+ }
+ }
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < block_size; ++i) {
+ float w = weight[i];
+ float q = 2*Laux[i] + 1;
+ sumqx += w*xval[i]*q;
+ sumq2 += w*q*q;
+ }
+ if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+ scale = sumqx/sumq2; best = scale*sumqx;
+ for (int i = 0; i < block_size; ++i) L[i] = Laux[i];
+ for (int k = 0; k < bs4; ++k) is_on_grid[k] = is_on_grid_aux[k];
+ }
+ }
+ int n_not_ongrid = 0;
+ for (int k = 0; k < bs4; ++k) if (!is_on_grid[k]) ++n_not_ongrid;
+ if (n_not_ongrid > 0 && scale > 0) {
+ float id = 1/scale;
+ for (int k = 0; k < bs4; ++k) {
+ //if (is_on_grid[k]) continue;
+ uint16_t u = 0;
+ for (int i = 0; i < 4; ++i) {
+ int l = nearest_int(0.5f*(id*xval[4*k+i]-1));
+ l = MAX(0, MIN(kMaxQ-1, l));
+ u |= (l << 3*i);
+ }
+ int grid_index = kmap_q3xs[u];
+ if (grid_index < 0) {
+ const uint16_t * neighbours = kneighbors_q3xs - kmap_q3xs[u] - 1;
+ grid_index = iq3_find_best_neighbour(neighbours, kgrid_q3xs, xval + 4*k, waux + 4*k, scale, L + 4*k);
+ }
+ const int8_t * pg = (const int8_t *)(kgrid_q3xs + grid_index);
+ for (int i = 0; i < 4; ++i) L[4*k+i] = (pg[i] - 1)/2;
+ }
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < block_size; ++i) {
+ float w = weight[i];
+ float q = 2*L[i] + 1;
+ sumqx += w*xval[i]*q;
+ sumq2 += w*q*q;
+ }
+ if (sumq2 > 0) scale = sumqx/sumq2;
+ }
+ if (scale < 0) {
+ // This should never happen, but just in case, flip scale so that it is positive (we use uint's to encode the scale)
+ // and correspondingly flip quant signs.
+ scale = -scale;
+ for (int k = 0; k < bs8; ++k) block_signs[k] = ~block_signs[k];
+ }
+ for (int k = 0; k < bs4; ++k) {
+ uint16_t u = 0;
+ for (int i = 0; i < 4; ++i) u |= (L[4*k+i] << 3*i);
+ int grid_index = kmap_q3xs[u];
+ if (grid_index < 0) {
+ printf("Oops: found point %u not on grid:", u);
+ for (int i = 0; i < 4; ++i) printf(" %d", L[4*k+i]);
+ printf("\n");
+ GGML_ABORT("fatal error");
+ }
+ qs[k] = grid_index & 255;
+ qh[(ib*bs4+k)/8] |= ((grid_index >> 8) << ((ib*bs4+k)%8));
+ }
+ qs += bs4;
+ for (int k = 0; k < bs8; ++k) signs[k] = block_signs[k];
+ signs += bs8;
+ GGML_ASSERT(scale >= 0);
+ scales[ib] = scale;
+ max_scale = MAX(max_scale, scale);
+ }
+
+ if (!max_scale) {
+ continue;
+ }
+
+ float d = max_scale/31;
+ y[ibl].d = GGML_FP32_TO_FP16(d * 1.033f);
+ float id = 1/d;
+ for (int ib = 0; ib < QK_K/block_size; ib += 2) {
+ int l1 = nearest_int(0.5f*(id*scales[ib+0]-1));
+ l1 = MAX(0, MIN(15, l1));
+ int l2 = nearest_int(0.5f*(id*scales[ib+1]-1));
+ l2 = MAX(0, MIN(15, l2));
+ y[ibl].scales[ib/2] = l1 | (l2 << 4);
+ }
+
+ }
+}
+
+#define IQ3S_BLOCK_SIZE 32
+size_t quantize_iq3_s(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ GGML_ASSERT(n_per_row%QK_K == 0);
+ int64_t nblock = n_per_row/QK_K;
+ float scales[QK_K/IQ3S_BLOCK_SIZE];
+ float weight[IQ3S_BLOCK_SIZE];
+ float xval[IQ3S_BLOCK_SIZE];
+ int8_t L[IQ3S_BLOCK_SIZE];
+ int8_t Laux[IQ3S_BLOCK_SIZE];
+ float waux[IQ3S_BLOCK_SIZE];
+ bool is_on_grid[IQ3S_BLOCK_SIZE/4];
+ bool is_on_grid_aux[IQ3S_BLOCK_SIZE/4];
+ uint8_t block_signs[IQ3S_BLOCK_SIZE/8];
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_iq3_s_impl(IQ3S_BLOCK_SIZE, src, qrow, n_per_row, quant_weights,
+ scales, weight, xval, L, Laux, waux, is_on_grid, is_on_grid_aux, block_signs);
+ src += n_per_row;
+ qrow += nblock*sizeof(block_iq3_s);
+ }
+ return nrow * nblock * sizeof(block_iq3_s);
+}
+
+void quantize_row_iq3_s_ref(const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ quantize_iq3_s(x, y, 1, k, NULL);
+}
+
+
+// =================================== 1.5 bpw ===================================================
+
+static int iq1_find_best_neighbour(const uint16_t * GGML_RESTRICT neighbours, const uint64_t * GGML_RESTRICT grid,
+ const float * GGML_RESTRICT xval, const float * GGML_RESTRICT weight, float * scale, int8_t * GGML_RESTRICT L, int ngrid) {
+ int num_neighbors = neighbours[0];
+ GGML_ASSERT(num_neighbors > 0);
+ float best_score = -FLT_MAX;
+ int grid_index = -1;
+ for (int j = 1; j <= num_neighbors; ++j) {
+ const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < 8; ++i) {
+ float q = (pg[i] - 3)/2;
+ float w = weight[i];
+ sumqx += w*q*xval[i];
+ sumq2 += w*q*q;
+ }
+ if (sumqx > 0 && sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
+ *scale = sumqx/sumq2; best_score = *scale * sumqx;
+ grid_index = neighbours[j];
+ }
+ }
+ if (grid_index < 0) {
+ for (int i = 0; i < ngrid; ++i) {
+ const int8_t * grid_i = (const int8_t *)(grid + i);
+ float sumqx = 0, sumq2 = 0;
+ for (int j = 0; j < 8; ++j) {
+ float w = weight[j];
+ float q = (grid_i[j] - 3)/2;
+ sumqx += w*q*xval[j];
+ sumq2 += w*q*q;
+ }
+ if (sumqx > 0 && sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
+ *scale = sumqx/sumq2; best_score = *scale*sumqx;
+ grid_index = i;
+ }
+ }
+ }
+ if (grid_index < 0) {
+ printf("Oops, did not find grid point\n");
+ printf("Have %d neighbours\n", num_neighbors);
+ for (int j = 1; j <= num_neighbors; ++j) {
+ const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < 8; ++i) {
+ float q = (pg[i] - 3)/2;
+ float w = weight[i];
+ sumqx += w*q*xval[i];
+ sumq2 += w*q*q;
+ }
+ printf(" neighbour %d: sumqx = %g sumq2 = %g\n", j, (double)sumqx, (double)sumq2);
+ }
+ }
+ GGML_ASSERT(grid_index >= 0);
+ //!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+ *scale *= 1.05f; // This is a fudge factor. Don't ask me why it improves the result.
+ //!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!!
+ const int8_t * pg = (const int8_t *)(grid + grid_index);
+ for (int i = 0; i < 8; ++i) L[i] = (pg[i] - 1)/2;
+ return grid_index;
+}
+
+static int iq1_find_best_neighbour2(const uint16_t * GGML_RESTRICT neighbours, const uint64_t * GGML_RESTRICT grid,
+ const float * GGML_RESTRICT xval, const float * GGML_RESTRICT weight, float scale, const float * GGML_RESTRICT xg, int8_t * GGML_RESTRICT L, int ngrid) {
+ int num_neighbors = neighbours[0];
+ GGML_ASSERT(num_neighbors > 0);
+ float best_score = FLT_MAX;
+ int grid_index = -1;
+ for (int j = 1; j <= num_neighbors; ++j) {
+ const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+ float d2 = 0;
+ for (int i = 0; i < 8; ++i) {
+ float q = xg[(pg[i] - 1)/2];
+ float w = weight[i];
+ float diff = scale*q - xval[i];
+ d2 += w*diff*diff;
+ }
+ if (d2 < best_score) {
+ best_score = d2;
+ grid_index = neighbours[j];
+ }
+ }
+ if (grid_index < 0) {
+ for (int i = 0; i < ngrid; ++i) {
+ const int8_t * grid_i = (const int8_t *)(grid + i);
+ float d2 = 0;
+ for (int j = 0; j < 8; ++j) {
+ float w = weight[j];
+ float q = xg[(grid_i[j] - 1)/2];
+ float diff = scale*q - xval[i];
+ d2 += w*diff*diff;
+ }
+ if (d2 < best_score) {
+ best_score = d2;
+ grid_index = i;
+ }
+ }
+ }
+ if (grid_index < 0) {
+ printf("Oops, did not find grid point\n");
+ printf("Have %d neighbours\n", num_neighbors);
+ for (int j = 1; j <= num_neighbors; ++j) {
+ const int8_t * pg = (const int8_t *)(grid + neighbours[j]);
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < 8; ++i) {
+ float q = xg[(pg[i] - 1)/2];
+ float w = weight[i];
+ sumqx += w*q*xval[i];
+ sumq2 += w*q*q;
+ }
+ printf(" neighbour %d: sumqx = %g sumq2 = %g\n", j, (double)sumqx, (double)sumq2);
+ }
+ }
+ GGML_ASSERT(grid_index >= 0);
+ const int8_t * pg = (const int8_t *)(grid + grid_index);
+ for (int i = 0; i < 8; ++i) L[i] = (pg[i] - 1)/2;
+ return grid_index;
+}
+
+static int iq1_sort_helper(const void * left, const void * right) {
+ const float * l = left;
+ const float * r = right;
+ return *l < *r ? -1 : *l > *r ? 1 : 0;
+}
+
+#define IQ1S_BLOCK_SIZE 32
+#define IQ1M_BLOCK_SIZE 16
+static void quantize_row_iq1_s_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights,
+ float * scales,
+ float * weight,
+ float * sumx,
+ float * sumw,
+ float * pairs,
+ int8_t * L,
+ uint16_t * index,
+ int8_t * shifts) {
+
+ const int gindex = iq2_data_index(GGML_TYPE_IQ1_S);
+
+ const uint64_t * kgrid_q2xs = iq2_data[gindex].grid;
+ const int * kmap_q2xs = iq2_data[gindex].map;
+ const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+ GGML_ASSERT(quant_weights && "missing quantization weights");
+ GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(n%QK_K == 0);
+
+ block_iq1_s * y = vy;
+
+ const int64_t nbl = n/QK_K;
+
+ const int block_size = IQ1S_BLOCK_SIZE;
+
+ const float x_p[3] = {-1 + IQ1S_DELTA, IQ1S_DELTA, 1 + IQ1S_DELTA};
+ const float x_m[3] = {-1 - IQ1S_DELTA, -IQ1S_DELTA, 1 - IQ1S_DELTA};
+
+
+ int * idx = (int *)(pairs + 1);
+
+ for (int ibl = 0; ibl < nbl; ++ibl) {
+
+ y[ibl].d = GGML_FP32_TO_FP16(0.f);
+ memset(y[ibl].qs, 0, QK_K/8);
+ memset(y[ibl].qh, 0, QK_K/16);
+
+ float max_scale = 0;
+
+ const float * xbl = x + QK_K*ibl;
+ float sumx2 = 0;
+ for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+ float sigma2 = 2*sumx2/QK_K;
+
+ for (int ib = 0; ib < QK_K/block_size; ++ib) {
+ const float * xb = xbl + block_size*ib;
+ const float * qw = quant_weights + QK_K*ibl + block_size*ib;
+ for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+ float max = fabsf(xb[0]);
+ for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
+ if (max < GROUP_MAX_EPS_IQ1_S) {
+ scales[ib] = 0;
+ shifts[ib] = 1;
+ memset(L, 1, block_size);
+ continue;
+ }
+ // Here we solve exactly the sum of squared difference (SSD) weighted minimization problem.
+ // With just 3 allowed quant values (-1, 0, 1), we can search exhaustively for the two
+ // boundaries that split the weights xb[i] into 3 groups. To do so, we sort the weights
+ // in ascending order, compute Si = sum[weight[j] xb[j], j = 0...i] and
+ // Wi = sum[weight[j], j = 0...i], and use these to quckly get get the optimum scale
+ // for each possible and score for each split.
+ for (int j = 0; j < block_size; ++j) {
+ pairs[2*j] = xb[j];
+ idx[2*j] = j;
+ }
+ qsort(pairs, block_size, 2*sizeof(float), iq1_sort_helper);
+ {
+ sumx[0] = sumw[0] = 0;
+ for (int j = 0; j < block_size; ++j) {
+ int i = idx[2*j];
+ sumx[j+1] = sumx[j] + weight[i]*xb[i];
+ sumw[j+1] = sumw[j] + weight[i];
+ }
+ }
+ float best_score = -FLT_MAX, scale = max;
+ int besti1 = -1, besti2 = -1, best_shift = 0;
+ for (int i1 = 0; i1 <= block_size; ++i1) {
+ for (int i2 = i1; i2 <= block_size; ++i2) {
+ float sumqx = (sumx[i1] - sumx[0])*x_p[0] + (sumx[i2] - sumx[i1])*x_p[1] + (sumx[block_size] - sumx[i2])*x_p[2];
+ float sumq2 = (sumw[i1] - sumw[0])*x_p[0]*x_p[0] + (sumw[i2] - sumw[i1])*x_p[1]*x_p[1] + (sumw[block_size] - sumw[i2])*x_p[2]*x_p[2];
+ if (sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
+ scale = sumqx/sumq2; best_score = scale*sumqx;
+ besti1 = i1; besti2 = i2; best_shift = 1;
+ }
+ sumqx = (sumx[i1] - sumx[0])*x_m[0] + (sumx[i2] - sumx[i1])*x_m[1] + (sumx[block_size] - sumx[i2])*x_m[2];
+ sumq2 = (sumw[i1] - sumw[0])*x_m[0]*x_m[0] + (sumw[i2] - sumw[i1])*x_m[1]*x_m[1] + (sumw[block_size] - sumw[i2])*x_m[2]*x_m[2];
+ if (sumq2 > 0 && sumqx*sumqx > best_score*sumq2) {
+ scale = sumqx/sumq2; best_score = scale*sumqx;
+ besti1 = i1; besti2 = i2; best_shift = -1;
+ }
+ }
+ }
+ if (besti1 < 0 || besti2 < 0 || best_shift == 0) {
+ scales[ib] = 0;
+ shifts[ib] = 1;
+ memset(L, 1, block_size);
+ continue;
+ }
+ for (int j = 0; j < besti1; ++j) L[idx[2*j]] = 0;
+ for (int j = besti1; j < besti2; ++j) L[idx[2*j]] = 1;
+ for (int j = besti2; j < block_size; ++j) L[idx[2*j]] = 2;
+ if (scale < 0) {
+ for (int j = 0; j < block_size; ++j) L[j] = 2 - L[j];
+ scale = -scale; best_shift = -best_shift;
+ }
+ bool all_on_grid = true;
+ const float * xx = best_shift == 1 ? x_p : x_m;
+ for (int k = 0; k < block_size/8; ++k) {
+ uint16_t u = 0;
+ for (int j = 0; j < 8; ++j) u |= (L[8*k+j] << 2*j);
+ int grid_index = kmap_q2xs[u];
+ if (grid_index < 0) {
+ all_on_grid = false;
+ const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+ grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, weight + 8*k, scale, xx, L + 8*k, NGRID_IQ1S);
+ GGML_ASSERT(grid_index >= 0);
+ }
+ index[k] = grid_index;
+ }
+ if (!all_on_grid) {
+ float sumqx = 0, sumq2 = 0;
+ for (int k = 0; k < block_size/8; ++k) {
+ const int8_t * pg = (const int8_t *)(kgrid_q2xs + index[k]);
+ for (int j = 0; j < 8; ++j) {
+ float w = weight[8*k + j];
+ float q = xx[(pg[j] - 1)/2];
+ sumqx += w*q*xb[8*k+j];
+ sumq2 += w*q*q;
+ }
+ }
+ if (sumqx > 0 && sumq2 > 0) scale = sumqx/sumq2;
+ }
+ uint16_t h = 0;
+ for (int k = 0; k < block_size/8; ++k) {
+ y[ibl].qs[(block_size/8)*ib + k] = index[k] & 255;
+ h |= (index[k] >> 8) << 3*k;
+ }
+ y[ibl].qh[ib] = h;
+ GGML_ASSERT(scale >= 0);
+ scales[ib] = scale;
+ shifts[ib] = best_shift;
+ max_scale = MAX(max_scale, scale);
+ }
+
+ if (!max_scale) {
+ continue;
+ }
+
+ float d = max_scale/15;
+ y[ibl].d = GGML_FP32_TO_FP16(d*1.125f); // 1.125f is another fudge factor. Don't ask me why it is needed.
+ float id = 1/d;
+ for (int ib = 0; ib < QK_K/block_size; ++ib) {
+ int l = nearest_int(0.5f*(id*scales[ib]-1));
+ l = MAX(0, MIN(7, l));
+ if (shifts[ib] == -1) l |= 8;
+ y[ibl].qh[ib] |= (l << 12);
+ }
+ }
+}
+
+size_t quantize_iq1_s(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ GGML_ASSERT(n_per_row%QK_K == 0);
+ float scales[QK_K/IQ1S_BLOCK_SIZE];
+ float weight[IQ1S_BLOCK_SIZE];
+ int8_t L[IQ1S_BLOCK_SIZE];
+ float sumx[IQ1S_BLOCK_SIZE+1];
+ float sumw[IQ1S_BLOCK_SIZE+1];
+ float pairs[2*IQ1S_BLOCK_SIZE];
+ uint16_t index[IQ1S_BLOCK_SIZE/8];
+ int8_t shifts[QK_K/IQ1S_BLOCK_SIZE];
+ int64_t nblock = n_per_row/QK_K;
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_iq1_s_impl(src, qrow, n_per_row, quant_weights, scales, weight, sumx, sumw, pairs, L, index, shifts);
+ src += n_per_row;
+ qrow += nblock*sizeof(block_iq1_s);
+ }
+ return nrow * nblock * sizeof(block_iq1_s);
+}
+
+static void quantize_row_iq1_m_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights,
+ float * scales,
+ float * weight,
+ float * pairs,
+ int8_t * L,
+ uint16_t * index,
+ int8_t * shifts) {
+
+ const int gindex = iq2_data_index(GGML_TYPE_IQ1_M);
+
+ const uint64_t * kgrid_q2xs = iq2_data[gindex].grid;
+ const int * kmap_q2xs = iq2_data[gindex].map;
+ const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+ //GGML_ASSERT(quant_weights && "missing quantization weights");
+ GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(n%QK_K == 0);
+
+ block_iq1_m * y = vy;
+
+ const int64_t nbl = n/QK_K;
+
+ const int block_size = IQ1M_BLOCK_SIZE;
+
+ const float x_p[3] = {-1 + IQ1M_DELTA, IQ1M_DELTA, 1 + IQ1M_DELTA};
+ const float x_m[3] = {-1 - IQ1M_DELTA, -IQ1M_DELTA, 1 - IQ1M_DELTA};
+ const uint8_t masks[4] = {0x00, 0x80, 0x08, 0x88};
+
+ int * idx = (int *)(pairs + 1);
+
+ float sumqx[4], sumq2[4];
+
+ iq1m_scale_t s;
+ const float * xx;
+
+ for (int ibl = 0; ibl < nbl; ++ibl) {
+ memset(y[ibl].qs, 0, QK_K/8);
+ memset(y[ibl].qh, 0, QK_K/16);
+ memset(y[ibl].scales, 0, QK_K/32);
+
+ float max_scale = 0;
+
+ const float * xbl = x + QK_K*ibl;
+ float sumx2 = 0;
+ for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+ float sigma2 = 2*sumx2/QK_K;
+
+ for (int ib = 0; ib < QK_K/block_size; ++ib) {
+ const float * xb = xbl + block_size*ib;
+ if (quant_weights) {
+ const float * qw = quant_weights + QK_K*ibl + block_size*ib;
+ for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+ } else {
+ for (int i = 0; i < block_size; ++i) weight[i] = xb[i]*xb[i];
+ }
+ float max = fabsf(xb[0]);
+ for (int i = 1; i < block_size; ++i) max = MAX(max, fabsf(xb[i]));
+ if (max < GROUP_MAX_EPS_IQ1_M) {
+ scales[ib] = 0;
+ shifts[ib] = 0;
+ memset(L, 1, block_size);
+ continue;
+ }
+ // Here we solve exactly the sum of squared difference (SSD) weighted minimization problem.
+ // With just 3 allowed quant values (-1, 0, 1), we can search exhaustively for the two
+ // boundaries that split the weights xb[i] into 3 groups. To do so, we sort the weights
+ // in ascending order, compute Si = sum[weight[j] xb[j], j = 0...i] and
+ // Wi = sum[weight[j], j = 0...i], and use these to quckly get get the optimum scale
+ // for each possible and score for each split.
+ for (int j = 0; j < block_size; ++j) {
+ pairs[2*j] = xb[j];
+ idx[2*j] = j;
+ }
+ qsort(pairs, block_size, 2*sizeof(float), iq1_sort_helper);
+ float best_score = -FLT_MAX, scale = max;
+ int besti1 = -1, besti2 = -1, best_k = -1;
+ // 0: +, +
+ // 1: +, -
+ // 2: -, +
+ // 3: -, -
+ // prefix sums are kept per half of the block because each half can use a different sign (x_p or x_m)
+ // since v[0]-v[1] = v[1]-v[2] = -1 for both x_p and x_m, the 3-group sum for a split collapses to T*v[2] - px[i1] - px[i2]
+ {
+ float px[2][IQ1M_BLOCK_SIZE+1];
+ float pw[2][IQ1M_BLOCK_SIZE+1];
+ px[0][0] = px[1][0] = 0;
+ pw[0][0] = pw[1][0] = 0;
+ for (int j = 0; j < block_size; ++j) {
+ const int i = idx[2*j];
+ const int h = i < block_size/2 ? 0 : 1;
+ px[h][j+1] = px[h][j] + weight[i]*xb[i];
+ px[1-h][j+1] = px[1-h][j];
+ pw[h][j+1] = pw[h][j] + weight[i];
+ pw[1-h][j+1] = pw[1-h][j];
+ }
+ const float txs[2] = {px[0][block_size], px[1][block_size]}; // total weight*x per half
+ const float tws[2] = {pw[0][block_size], pw[1][block_size]}; // total weight per half
+ const float p2 = x_p[2], m2 = x_m[2];
+ const float cp1 = x_p[0]*x_p[0] - x_p[1]*x_p[1];
+ const float cp2 = x_p[1]*x_p[1] - x_p[2]*x_p[2];
+ const float cm1 = x_m[0]*x_m[0] - x_m[1]*x_m[1];
+ const float cm2 = x_m[1]*x_m[1] - x_m[2]*x_m[2];
+ for (int i1 = 0; i1 <= block_size; ++i1) {
+ for (int i2 = i1; i2 <= block_size; ++i2) {
+ float qx_p[2], qx_m[2], q2_p[2], q2_m[2];
+ for (int h = 0; h < 2; ++h) {
+ const float sx = px[h][i1] + px[h][i2];
+ qx_p[h] = txs[h]*p2 - sx;
+ qx_m[h] = txs[h]*m2 - sx;
+ q2_p[h] = tws[h]*p2*p2 + pw[h][i1]*cp1 + pw[h][i2]*cp2;
+ q2_m[h] = tws[h]*m2*m2 + pw[h][i1]*cm1 + pw[h][i2]*cm2;
+ }
+ sumqx[0] = qx_p[0] + qx_p[1];
+ sumqx[1] = qx_p[0] + qx_m[1];
+ sumqx[2] = qx_m[0] + qx_p[1];
+ sumqx[3] = qx_m[0] + qx_m[1];
+ sumq2[0] = q2_p[0] + q2_p[1];
+ sumq2[1] = q2_p[0] + q2_m[1];
+ sumq2[2] = q2_m[0] + q2_p[1];
+ sumq2[3] = q2_m[0] + q2_m[1];
+ for (int k = 0; k < 4; ++k) {
+ if (sumq2[k] > 0 && sumqx[k]*sumqx[k] > best_score*sumq2[k]) {
+ scale = sumqx[k]/sumq2[k]; best_score = scale*sumqx[k];
+ besti1 = i1; besti2 = i2; best_k = k;
+ }
+ }
+ }
+ }
+ }
+ if (besti1 < 0 || besti2 < 0 || best_k < 0) {
+ scales[ib] = 0;
+ shifts[ib] = 0;
+ memset(L, 1, block_size);
+ continue;
+ }
+ for (int j = 0; j < besti1; ++j) L[idx[2*j]] = 0;
+ for (int j = besti1; j < besti2; ++j) L[idx[2*j]] = 1;
+ for (int j = besti2; j < block_size; ++j) L[idx[2*j]] = 2;
+ if (scale < 0) {
+ for (int j = 0; j < block_size; ++j) L[j] = 2 - L[j];
+ scale = -scale;
+ best_k = best_k == 0 ? 3 : best_k == 1 ? 2 : best_k == 2 ? 1 : 0;
+ }
+ bool all_on_grid = true;
+ for (int k = 0; k < block_size/8; ++k) {
+ if (k == 0) xx = best_k < 2 ? x_p : x_m;
+ else xx = best_k%2 == 0 ? x_p : x_m;
+ uint16_t u = 0;
+ for (int j = 0; j < 8; ++j) u |= (L[8*k+j] << 2*j);
+ int grid_index = kmap_q2xs[u];
+ if (grid_index < 0) {
+ all_on_grid = false;
+ const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+ grid_index = iq1_find_best_neighbour2(neighbours, kgrid_q2xs, xb + 8*k, weight + 8*k, scale, xx, L + 8*k, NGRID_IQ1S);
+ GGML_ASSERT(grid_index >= 0);
+ }
+ index[k] = grid_index;
+ }
+ if (!all_on_grid) {
+ float sumqx_f = 0, sumq2_f = 0;
+ for (int k = 0; k < block_size/8; ++k) {
+ if (k == 0) xx = best_k < 2 ? x_p : x_m;
+ else xx = best_k%2 == 0 ? x_p : x_m;
+ const int8_t * pg = (const int8_t *)(kgrid_q2xs + index[k]);
+ for (int j = 0; j < 8; ++j) {
+ float w = weight[8*k + j];
+ float q = xx[(pg[j] - 1)/2];
+ sumqx_f += w*q*xb[8*k+j];
+ sumq2_f += w*q*q;
+ }
+ }
+ if (sumqx_f > 0 && sumq2_f > 0) scale = sumqx_f/sumq2_f;
+ }
+ y[ibl].qs[2*ib + 0] = index[0] & 255;
+ y[ibl].qs[2*ib + 1] = index[1] & 255;
+ y[ibl].qh[ib] = (index[0] >> 8) | ((index[1] >> 8) << 4);
+ GGML_ASSERT(scale >= 0);
+ scales[ib] = scale;
+ shifts[ib] = best_k;
+ max_scale = MAX(max_scale, scale);
+ }
+
+ if (!max_scale) {
+ continue;
+ }
+
+ uint16_t * sc = (uint16_t *)y[ibl].scales;
+ float d = max_scale/15;
+ float id = 1/d;
+ float sumqx_f = 0, sumq2_f = 0;
+ for (int ib = 0; ib < QK_K/block_size; ++ib) {
+ int l = nearest_int(0.5f*(id*scales[ib+0]-1));
+ l = MAX(0, MIN(7, l));
+ sc[ib/4] |= (l << 3*(ib%4));
+ y[ibl].qh[ib] |= masks[shifts[ib]];
+ const float * xb = xbl + block_size*ib;
+ if (quant_weights) {
+ const float * qw = quant_weights + QK_K*ibl + block_size*ib;
+ for (int i = 0; i < block_size; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+ } else {
+ for (int i = 0; i < block_size; ++i) weight[i] = xb[i]*xb[i];
+ }
+ for (int k = 0; k < block_size/8; ++k) {
+ if (k == 0) xx = shifts[ib] < 2 ? x_p : x_m;
+ else xx = shifts[ib]%2 == 0 ? x_p : x_m;
+ const int8_t * pg = (const int8_t *)(kgrid_q2xs + y[ibl].qs[2*ib+k] + ((y[ibl].qh[ib] << (8 - 4*k)) & 0x700));
+ for (int j = 0; j < 8; ++j) {
+ float w = weight[8*k + j];
+ float q = xx[(pg[j] - 1)/2]*(2*l+1);
+ sumqx_f += w*q*xb[8*k+j];
+ sumq2_f += w*q*q;
+ }
+ }
+ }
+ if (sumq2_f > 0) d = sumqx_f/sumq2_f;
+ s.f16 = GGML_FP32_TO_FP16(d*1.1125f); // 1.1125f is another fudge factor. Don't ask me why it is needed.
+ sc[0] |= ((s.u16 & 0x000f) << 12);
+ sc[1] |= ((s.u16 & 0x00f0) << 8);
+ sc[2] |= ((s.u16 & 0x0f00) << 4);
+ sc[3] |= ((s.u16 & 0xf000) << 0);
+ }
+}
+
+size_t quantize_iq1_m(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ GGML_ASSERT(n_per_row%QK_K == 0);
+ float scales[QK_K/IQ1M_BLOCK_SIZE];
+ float weight[IQ1M_BLOCK_SIZE];
+ int8_t L[IQ1M_BLOCK_SIZE];
+ float pairs[2*IQ1M_BLOCK_SIZE];
+ uint16_t index[IQ1M_BLOCK_SIZE/8];
+ int8_t shifts[QK_K/IQ1M_BLOCK_SIZE];
+ int64_t nblock = n_per_row/QK_K;
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_iq1_m_impl(src, qrow, n_per_row, quant_weights, scales, weight, pairs, L, index, shifts);
+ src += n_per_row;
+ qrow += nblock*sizeof(block_iq1_m);
+ }
+ return nrow * nblock * sizeof(block_iq1_m);
+}
+
+// ============================ 4-bit non-linear quants
+
+static void quantize_row_iq4_nl_impl(const int super_block_size, const int block_size, const float * GGML_RESTRICT x,
+ ggml_fp16_t * dh, uint8_t * q4, uint16_t * scales_h, uint8_t * scales_l,
+ float * scales, float * weight, uint8_t * L,
+ const int8_t * values,
+ const float * quant_weights,
+ const int ntry) {
+
+ float sigma2 = 0;
+ for (int j = 0; j < super_block_size; ++j) sigma2 += x[j]*x[j];
+ sigma2 *= 2.f/super_block_size;
+
+ memset(q4, 0, super_block_size/2);
+ dh[0] = GGML_FP32_TO_FP16(0.f);
+
+ float max_scale = 0, amax_scale = 0;
+ for (int ib = 0; ib < super_block_size/block_size; ++ib) {
+ const float * xb = x + ib*block_size;
+ uint8_t * Lb = L + ib*block_size;
+ if (quant_weights) {
+ const float * qw = quant_weights + ib*block_size;
+ for (int j = 0; j < block_size; ++j) weight[j] = qw[j] * sqrtf(sigma2 + xb[j]*xb[j]);
+ } else {
+ for (int j = 0; j < block_size; ++j) weight[j] = xb[j]*xb[j];
+ }
+ float amax = 0, max = 0;
+ for (int j = 0; j < block_size; ++j) {
+ float ax = fabsf(xb[j]);
+ if (ax > amax) {
+ amax = ax; max = xb[j];
+ }
+ }
+ if (amax < GROUP_MAX_EPS) {
+ scales[ib] = 0;
+ continue;
+ }
+ float d = ntry > 0 ? -max/values[0] : max/values[0];
+ float id = 1/d;
+ float sumqx = 0, sumq2 = 0;
+ for (int j = 0; j < block_size; ++j) {
+ float al = id*xb[j];
+ int l = best_index_int8(16, values, al);
+ Lb[j] = l;
+ float q = values[l];
+ float w = weight[j];
+ sumqx += w*q*xb[j];
+ sumq2 += w*q*q;
+ }
+ d = sumq2 > 0 ? sumqx/sumq2 : 0.f;
+ float best = d*sumqx;
+ for (int itry = -ntry; itry <= ntry; ++itry) {
+ id = (itry + values[0])/max;
+ sumqx = sumq2 = 0;
+ for (int j = 0; j < block_size; ++j) {
+ float al = id*xb[j];
+ int l = best_index_int8(16, values, al);
+ float q = values[l];
+ float w = weight[j];
+ sumqx += w*q*xb[j];
+ sumq2 += w*q*q;
+ }
+ if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+ d = sumqx/sumq2; best = d * sumqx;
+ }
+ }
+ scales[ib] = d;
+ float abs_d = fabsf(d);
+ if (abs_d > amax_scale) {
+ amax_scale = abs_d; max_scale = d;
+ }
+ }
+
+ if (super_block_size/block_size > 1) {
+ int nb = super_block_size/block_size;
+ memset(scales_h, 0, ((nb+7)/8)*sizeof(uint16_t));
+ float d = -max_scale/32;
+ dh[0] = GGML_FP32_TO_FP16(d);
+ float id = d ? 1/d : 0.f;
+ for (int ib = 0; ib < super_block_size/block_size; ++ib) {
+ int l = nearest_int(id*scales[ib]);
+ l = MAX(-32, MIN(31, l));
+ float dl = d * l;
+ float idl = dl ? 1/dl : 0.f;
+ uint8_t * Lb = L + ib*block_size;
+ const float * xb = x + ib*block_size;
+ for (int j = 0; j < block_size; ++j) {
+ Lb[j] = best_index_int8(16, values, idl*xb[j]);
+ }
+ l += 32;
+ uint8_t l_l = l & 0xf;
+ uint8_t l_h = l >> 4;
+ if (ib%2 == 0) scales_l[ib/2] = l_l;
+ else scales_l[ib/2] |= (l_l << 4);
+ scales_h[ib/8] |= (l_h << 2*(ib%8));
+ }
+ } else {
+ dh[0] = GGML_FP32_TO_FP16(scales[0]);
+ if (ntry > 0) {
+ float id = scales[0] ? 1/scales[0] : 0;
+ for (int j = 0; j < super_block_size; ++j) {
+ L[j] = best_index_int8(16, values, id*x[j]);
+ }
+ }
+ }
+
+ for (int i = 0; i < super_block_size/32; ++i) {
+ for (int j = 0; j < 16; ++j) {
+ q4[16*i + j] = L[32*i + j] | (L[32*i + 16 + j] << 4);
+ }
+ }
+}
+
+size_t quantize_iq4_nl(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ GGML_ASSERT(n_per_row%QK4_NL == 0);
+ int64_t nblock = n_per_row/QK4_NL;
+ char * qrow = (char *)dst;
+ uint8_t L[QK4_NL];
+ float weight[QK4_NL];
+ uint16_t unused_h;
+ uint8_t * unused_l = NULL;
+ float scale;
+ for (int64_t row = 0; row < nrow; ++row) {
+ block_iq4_nl * iq4 = (block_iq4_nl *)qrow;
+ for (int ibl = 0; ibl < nblock; ++ibl) {
+ const float * qw = quant_weights ? quant_weights + QK4_NL*ibl : NULL;
+ quantize_row_iq4_nl_impl(QK4_NL, 32, src + QK4_NL*ibl, &iq4[ibl].d, iq4[ibl].qs, &unused_h, unused_l,
+ &scale, weight, L, kvalues_iq4nl, qw, 7);
+ }
+ src += n_per_row;
+ qrow += nblock*sizeof(block_iq4_nl);
+ }
+ return nrow * nblock * sizeof(block_iq4_nl);
+}
+
+//void quantize_row_iq4_nl_ref(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t k) {
+void quantize_row_iq4_nl_ref(const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int64_t k) {
+ GGML_ASSERT(k%QK4_NL == 0);
+ int64_t nblock = k/QK4_NL;
+ uint8_t L[QK4_NL];
+ float weight[QK4_NL];
+ uint16_t unused_h;
+ uint8_t * unused_l = NULL;
+ float scale;
+ block_iq4_nl * iq4 = y;
+ for (int ibl = 0; ibl < nblock; ++ibl) {
+ quantize_row_iq4_nl_impl(QK4_NL, 32, x + QK4_NL*ibl, &iq4[ibl].d, iq4[ibl].qs, &unused_h, unused_l,
+ &scale, weight, L, kvalues_iq4nl, NULL, -1);
+ }
+}
+
+size_t quantize_iq4_xs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ GGML_ASSERT(n_per_row%QK_K == 0);
+ int64_t nblock = n_per_row/QK_K;
+ char * qrow = (char *)dst;
+ uint8_t L[QK_K];
+ float weight[32];
+ float scales[QK_K/32];
+ for (int64_t row = 0; row < nrow; ++row) {
+ block_iq4_xs * iq4 = (block_iq4_xs *)qrow;
+ for (int ibl = 0; ibl < nblock; ++ibl) {
+ const float * qw = quant_weights ? quant_weights + QK_K*ibl : NULL;
+ quantize_row_iq4_nl_impl(QK_K, 32, src + QK_K*ibl, &iq4[ibl].d, iq4[ibl].qs, &iq4[ibl].scales_h, iq4[ibl].scales_l,
+ scales, weight, L, kvalues_iq4nl, qw, 7);
+ }
+ src += n_per_row;
+ qrow += nblock*sizeof(block_iq4_xs);
+ }
+ return nrow * nblock * sizeof(block_iq4_xs);
+}
+
+void quantize_row_iq4_xs_ref(const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ quantize_iq4_xs(x, y, 1, k, NULL);
+}
+
+// =============================== 2.5625 bpw
+
+static void quantize_row_iq2_s_impl(const float * GGML_RESTRICT x, void * GGML_RESTRICT vy, int64_t n, const float * GGML_RESTRICT quant_weights) {
+
+ const int gindex = iq2_data_index(GGML_TYPE_IQ2_S);
+
+ const uint64_t * kgrid_q2xs = iq2_data[gindex].grid;
+ const int * kmap_q2xs = iq2_data[gindex].map;
+ const uint16_t * kneighbors_q2xs = iq2_data[gindex].neighbours;
+
+ GGML_ASSERT(kmap_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kgrid_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(kneighbors_q2xs && "forgot to call ggml_quantize_init()?");
+ GGML_ASSERT(n%QK_K == 0);
+
+ const int kMaxQ = 3;
+
+ const int64_t nbl = n/QK_K;
+
+ block_iq2_s * y = vy;
+
+ float scales[QK_K/16];
+ float weight[16];
+ float xval[16];
+ int8_t L[16];
+ int8_t Laux[16];
+ float waux[16];
+ bool is_on_grid[2];
+ bool is_on_grid_aux[2];
+ uint8_t block_signs[2];
+
+ for (int ibl = 0; ibl < nbl; ++ibl) {
+
+ memset(&y[ibl], 0, sizeof(block_iq2_s));
+ y[ibl].d = GGML_FP32_TO_FP16(0.f);
+
+ float max_scale = 0;
+
+ const float * xbl = x + QK_K*ibl;
+ float sumx2 = 0;
+ for (int i = 0; i < QK_K; ++i) sumx2 += xbl[i]*xbl[i];
+ float sigma2 = 2*sumx2/QK_K;
+
+ for (int ib = 0; ib < QK_K/16; ++ib) {
+ const float * xb = xbl + 16*ib;
+ if (quant_weights) {
+ const float * qw = quant_weights + QK_K*ibl + 16*ib;
+ for (int i = 0; i < 16; ++i) weight[i] = qw[i] * sqrtf(sigma2 + xb[i]*xb[i]);
+ } else {
+ for (int i = 0; i < 16; ++i) weight[i] = 0.25f*sigma2 + xb[i]*xb[i];
+ }
+ for (int i = 0; i < 16; ++i) waux[i] = sqrtf(weight[i]);
+ for (int k = 0; k < 2; ++k) {
+ uint8_t s = 0;
+ for (int i = 0; i < 8; ++i) {
+ if (xb[8*k + i] >= 0) xval[8*k + i] = xb[8*k + i];
+ else {
+ xval[8*k + i] = -xb[8*k + i]; s |= (1 << i);
+ }
+ }
+ block_signs[k] = s;
+ }
+ float max = xval[0];
+ for (int i = 1; i < 16; ++i) max = MAX(max, xval[i]);
+ memset(L, 0, 16);
+ if (max < GROUP_MAX_EPS_IQ2_S) {
+ scales[ib] = 0;
+ continue;
+ }
+ float best = 0;
+ float scale = max/(2*kMaxQ-1);
+ is_on_grid[0] = is_on_grid[1] = true;
+ for (int is = -9; is <= 9; ++is) {
+ float id = (2*kMaxQ-1+is*0.1f)/max;
+ float this_scale = 1/id;
+ for (int k = 0; k < 2; ++k) {
+ for (int i = 0; i < 8; ++i) {
+ int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+ Laux[8*k+i] = MAX(0, MIN(kMaxQ-1, l));
+ }
+ uint16_t u = 0;
+ for (int i = 0; i < 8; ++i) u |= (Laux[8*k+i] << 2*i);
+ int grid_index = kmap_q2xs[u];
+ is_on_grid_aux[k] = true;
+ if (grid_index < 0) {
+ is_on_grid_aux[k] = false;
+ const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+ grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, this_scale, Laux + 8*k);
+ }
+ }
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < 16; ++i) {
+ float w = weight[i];
+ float q = 2*Laux[i] + 1;
+ sumqx += w*xval[i]*q;
+ sumq2 += w*q*q;
+ }
+ if (sumq2 > 0 && sumqx*sumqx > best*sumq2) {
+ scale = sumqx/sumq2; best = scale*sumqx;
+ for (int i = 0; i < 16; ++i) L[i] = Laux[i];
+ for (int k = 0; k < 2; ++k) is_on_grid[k] = is_on_grid_aux[k];
+ }
+ }
+ int n_not_ongrid = 0;
+ for (int k = 0; k < 2; ++k) if (!is_on_grid[k]) ++n_not_ongrid;
+ if (n_not_ongrid > 0 && scale > 0) {
+ float id = 1/scale;
+ for (int k = 0; k < 2; ++k) {
+ if (is_on_grid[k]) continue;
+ uint16_t u = 0;
+ for (int i = 0; i < 8; ++i) {
+ int l = nearest_int(0.5f*(id*xval[8*k+i]-1));
+ l = MAX(0, MIN(kMaxQ-1, l));
+ u |= (l << 2*i);
+ L[8*k + i] = l;
+ }
+ int grid_index = kmap_q2xs[u];
+ if (grid_index < 0) {
+ const uint16_t * neighbours = kneighbors_q2xs - kmap_q2xs[u] - 1;
+ grid_index = iq2_find_best_neighbour(neighbours, kgrid_q2xs, xval + 8*k, waux + 8*k, scale, L + 8*k);
+ }
+ }
+ float sumqx = 0, sumq2 = 0;
+ for (int i = 0; i < 16; ++i) {
+ float w = weight[i];
+ float q = 2*L[i] + 1;
+ sumqx += w*xval[i]*q;
+ sumq2 += w*q*q;
+ }
+ if (sumq2 > 0) scale = sumqx/sumq2;
+ }
+ if (scale < 0) {
+ scale = -scale;
+ for (int k = 0; k < 2; ++k) block_signs[k] = ~block_signs[k];
+ }
+ for (int k = 0; k < 2; ++k) {
+ uint16_t u = 0;
+ for (int i = 0; i < 8; ++i) u |= (L[8*k+i] << 2*i);
+ int grid_index = kmap_q2xs[u];
+ if (grid_index < 0) {
+ printf("Oops: found point %u not on grid:", u);
+ for (int i = 0; i < 8; ++i) printf(" %d", L[8*k+i]);
+ printf("\n");
+ GGML_ABORT("fatal error");
+ }
+ const int i8 = 2*ib + k;
+ y[ibl].qs[i8] = grid_index & 255;
+ y[ibl].qh[i8/4] |= ((grid_index >> 8) << 2*(i8%4));
+ y[ibl].qs[QK_K/8 + i8] = block_signs[k];
+ }
+ GGML_ASSERT(scale >= 0);
+ scales[ib] = scale;
+ max_scale = MAX(max_scale, scale);
+ }
+
+ if (!max_scale) {
+ continue;
+ }
+
+ float d = max_scale/31;
+ y[ibl].d = GGML_FP32_TO_FP16(d * 0.9875f);
+ float id = 1/d;
+ for (int ib = 0; ib < QK_K/16; ++ib) {
+ int l = nearest_int(0.5f*(id*scales[ib]-1));
+ l = MAX(0, MIN(15, l));
+ if (ib%2 == 0) y[ibl].scales[ib/2] = l;
+ else y[ibl].scales[ib/2] |= (l << 4);
+ }
+ }
+}
+
+size_t quantize_iq2_s(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrow, int64_t n_per_row, const float * quant_weights) {
+ GGML_ASSERT(n_per_row%QK_K == 0);
+ int64_t nblock = n_per_row/QK_K;
+ char * qrow = (char *)dst;
+ for (int64_t row = 0; row < nrow; ++row) {
+ quantize_row_iq2_s_impl(src, qrow, n_per_row, quant_weights);
+ src += n_per_row;
+ qrow += nblock*sizeof(block_iq2_s);
+ }
+ return nrow * nblock * sizeof(block_iq2_s);
+}
+
+void quantize_row_iq2_s_ref(const float * GGML_RESTRICT x, block_iq2_s * GGML_RESTRICT y, int64_t k) {
+ assert(k % QK_K == 0);
+ quantize_iq2_s(x, y, 1, k, NULL);
+}
+
+// =============================== data validation
+
+static bool validate_float(float f, size_t i) {
+ if (isinf(f)) {
+ fprintf(stderr, "ggml_validate_row_data: found inf value at block %zu\n", i);
+ return false;
+ }
+
+ if (isnan(f)) {
+ fprintf(stderr, "ggml_validate_row_data: found nan value at block %zu\n", i);
+ return false;
+ }
+
+ return true;
+}
+
+static bool isinf_fp16(ggml_fp16_t f) {
+ return (f & 0x7c00) == 0x7c00 && (f & 0x03ff) == 0;
+}
+
+static bool isnan_fp16(ggml_fp16_t f) {
+ return (f & 0x7c00) == 0x7c00 && (f & 0x03ff) != 0;
+}
+
+static bool validate_fp16(ggml_fp16_t f, size_t i) {
+ if (isinf_fp16(f)) {
+ fprintf(stderr, "ggml_validate_row_data: found inf value at block %zu\n", i);
+ return false;
+ }
+
+ if (isnan_fp16(f)) {
+ fprintf(stderr, "ggml_validate_row_data: found nan value at block %zu\n", i);
+ return false;
+ }
+
+ return true;
+}
+
+static bool validate_e_e8m0(uint8_t e, size_t i) {
+ if (e == 0xff) {
+ fprintf(stderr, "ggml_validate_row_data: found invalid e value %d at block %zu\n", e, i);
+ return false;
+ }
+
+ return true;
+}
+
+#define VALIDATE_ROW_DATA_D_F16_IMPL(type, data, nb) \
+ const type * q = (const type *) (data); \
+ for (size_t i = 0; i < (nb); ++i) { \
+ if (!validate_fp16(q[i].d, i)) { \
+ return false; \
+ } \
+ }
+
+#define VALIDATE_ROW_DATA_DM_F16_IMPL(type, data, nb, d, m) \
+ const type * q = (const type *) (data); \
+ for (size_t i = 0; i < (nb); ++i) { \
+ if (!validate_fp16(q[i].d, i) || !validate_fp16(q[i].m, i)) { \
+ return false; \
+ } \
+ }
+
+#define VALIDATE_ROW_DATA_E_E8M0_IMPL(type, data, nb) \
+ const type * q = (const type *) (data); \
+ for (size_t i = 0; i < (nb); ++i) { \
+ if (!validate_e_e8m0(q[i].e, i)) { \
+ return false; \
+ } \
+ }
+
+#define VALIDATE_ROW_DATA_DVEC_F16_IMPL(type, data, nb, nr) \
+ const type * q = (const type *) (data); \
+ for (size_t i = 0; i < (nb); ++i) { \
+ for (size_t j = 0; j < (nr); ++j) { \
+ if (!validate_fp16(q[i].d[j], i)) { \
+ return false; \
+ } \
+ } \
+ }
+
+bool ggml_validate_row_data(enum ggml_type type, const void * data, size_t nbytes) {
+ if (type < 0 || type >= GGML_TYPE_COUNT) {
+ fprintf(stderr, "%s: invalid type %d\n", __func__, type);
+ return false;
+ }
+
+ if (nbytes % ggml_type_size(type) != 0) {
+ fprintf(stderr, "%s: invalid size %zu for type %s (type size = %zu)\n", __func__, nbytes, ggml_type_name(type), ggml_type_size(type));
+ return false;
+ }
+
+ const size_t nb = nbytes/ggml_type_size(type);
+
+ switch (type) {
+ case GGML_TYPE_BF16:
+ {
+ int nans = 0;
+ int infs = 0;
+ const unsigned short * f = (const unsigned short *) data;
+ for (size_t i = 0; i < nb; ++i) {
+ nans += (f[i] & 0x7fff) > 0x7f80;
+ infs += (f[i] & 0x7fff) == 0x7f80;
+ }
+ if (nans) {
+ fprintf(stderr, "%s: found %d NaNs in row of %zu BF16 values\n", __func__, nans, nb);
+ return false;
+ }
+ if (infs) {
+ fprintf(stderr, "%s: found %d infinities in row of %zu BF16 values\n", __func__, infs, nb);
+ return false;
+ }
+ } break;
+ case GGML_TYPE_F16:
+ {
+ const ggml_fp16_t * f = (const ggml_fp16_t *) data;
+ size_t i = 0;
+#if defined(__AVX2__)
+ for (; i + 15 < nb; i += 16) {
+ __m256i v = _mm256_loadu_si256((const __m256i *)(f + i));
+ __m256i vexp = _mm256_and_si256(v, _mm256_set1_epi16(0x7c00));
+ __m256i cmp = _mm256_cmpeq_epi16(vexp, _mm256_set1_epi16(0x7c00));
+ int mask = _mm256_movemask_epi8(cmp);
+ if (mask) {
+ for (size_t j = 0; j < 16; ++j) {
+ if (!validate_fp16(f[i + j], i + j)) {
+ return false;
+ }
+ }
+ GGML_UNREACHABLE();
+ }
+ }
+#elif defined(__ARM_NEON)
+ for (; i + 7 < nb; i += 8) {
+ uint16x8_t v = vld1q_u16(f + i);
+ uint16x8_t vexp = vandq_u16(v, vdupq_n_u16(0x7c00));
+ uint16x8_t cmp = vceqq_u16(vexp, vdupq_n_u16(0x7c00));
+ uint64_t mask = vget_lane_u64(vreinterpret_u64_u8(vshrn_n_u16(cmp, 4)), 0);
+ if (mask) {
+ for (size_t j = 0; j < 8; ++j) {
+ if (!validate_fp16(f[i + j], i + j)) {
+ return false;
+ }
+ }
+ GGML_UNREACHABLE();
+ }
+ }
+#endif
+ for (; i < nb; ++i) {
+ if (!validate_fp16(f[i], i)) {
+ return false;
+ }
+ }
+ } break;
+ case GGML_TYPE_F32:
+ {
+ const float * f = (const float *) data;
+ size_t i = 0;
+#if defined(__AVX2__)
+ for (; i + 7 < nb; i += 8) {
+ __m256i v = _mm256_loadu_si256((const __m256i *)(f + i));
+ __m256i vexp = _mm256_and_si256(v, _mm256_set1_epi32(0x7f800000));
+ __m256i cmp = _mm256_cmpeq_epi32(vexp, _mm256_set1_epi32(0x7f800000));
+ int mask = _mm256_movemask_epi8(cmp);
+ if (mask) {
+ for (size_t j = 0; j < 8; ++j) {
+ if (!validate_float(f[i + j], i + j)) {
+ return false;
+ }
+ }
+ GGML_UNREACHABLE();
+ }
+ }
+#elif defined(__ARM_NEON)
+ for (; i + 3 < nb; i += 4) {
+ uint32x4_t v = vld1q_u32((const uint32_t *)f + i);
+ uint32x4_t vexp = vandq_u32(v, vdupq_n_u32(0x7f800000));
+ uint32x4_t cmp = vceqq_u32(vexp, vdupq_n_u32(0x7f800000));
+ uint64_t mask = vget_lane_u64(vreinterpret_u64_u16(vshrn_n_u32(cmp, 8)), 0);
+ if (mask) {
+ for (size_t j = 0; j < 4; ++j) {
+ if (!validate_float(f[i + j], i + j)) {
+ return false;
+ }
+ }
+ GGML_UNREACHABLE();
+ }
+ }
+#endif
+ for (; i < nb; ++i) {
+ if (!validate_float(f[i], i)) {
+ return false;
+ }
+ }
+ } break;
+ case GGML_TYPE_F64:
+ {
+ const double * f = (const double *) data;
+ for (size_t i = 0; i < nb; ++i) {
+ if (!validate_float(f[i], i)) {
+ return false;
+ }
+ }
+ } break;
+ case GGML_TYPE_Q1_0:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_q1_0, data, nb);
+ } break;
+ case GGML_TYPE_Q2_0:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_q2_0, data, nb);
+ } break;
+ case GGML_TYPE_Q4_0:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_q4_0, data, nb);
+ } break;
+ case GGML_TYPE_Q4_1:
+ {
+ VALIDATE_ROW_DATA_DM_F16_IMPL(block_q4_1, data, nb, d, m);
+ } break;
+ case GGML_TYPE_Q5_0:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_q5_0, data, nb);
+ } break;
+ case GGML_TYPE_Q5_1:
+ {
+ VALIDATE_ROW_DATA_DM_F16_IMPL(block_q5_1, data, nb, d, m);
+ } break;
+ case GGML_TYPE_Q8_0:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_q8_0, data, nb);
+ } break;
+ case GGML_TYPE_MXFP4:
+ {
+ VALIDATE_ROW_DATA_E_E8M0_IMPL(block_mxfp4, data, nb);
+ } break;
+ case GGML_TYPE_NVFP4:
+ {
+ // UE4M3 scales are uint8_t — all byte values are valid
+ GGML_UNUSED(data);
+ GGML_UNUSED(nb);
+ } break;
+ case GGML_TYPE_Q2_K:
+ {
+ VALIDATE_ROW_DATA_DM_F16_IMPL(block_q2_K, data, nb, d, dmin);
+ } break;
+ case GGML_TYPE_Q3_K:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_q3_K, data, nb);
+ } break;
+ case GGML_TYPE_Q4_K:
+ {
+ VALIDATE_ROW_DATA_DM_F16_IMPL(block_q4_K, data, nb, d, dmin);
+ } break;
+ case GGML_TYPE_Q5_K:
+ {
+ VALIDATE_ROW_DATA_DM_F16_IMPL(block_q5_K, data, nb, d, dmin);
+ } break;
+ case GGML_TYPE_Q6_K:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_q6_K, data, nb);
+ } break;
+ case GGML_TYPE_Q8_K:
+ {
+ const block_q8_K * q = (const block_q8_K *) data;
+ for (size_t i = 0; i < nb; ++i) {
+ if (!validate_float(q[i].d, i)) {
+ return false;
+ }
+ }
+ } break;
+ case GGML_TYPE_TQ1_0:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_tq1_0, data, nb);
+ } break;
+ case GGML_TYPE_TQ2_0:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_tq2_0, data, nb);
+ } break;
+ case GGML_TYPE_IQ1_S:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_iq1_s, data, nb);
+ } break;
+ case GGML_TYPE_IQ1_M:
+ {
+ const block_iq1_m * q = (const block_iq1_m *) data;
+ for (size_t i = 0; i < nb; ++i) {
+ iq1m_scale_t scale;
+ const uint16_t * sc = (const uint16_t *)q[i].scales;
+ scale.u16 = (sc[0] >> 12) | ((sc[1] >> 8) & 0x00f0) | ((sc[2] >> 4) & 0x0f00) | (sc[3] & 0xf000);
+ if (!validate_fp16(scale.f16, i)) {
+ return false;
+ }
+ }
+ } break;
+ case GGML_TYPE_IQ2_XXS:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_iq2_xxs, data, nb);
+ } break;
+ case GGML_TYPE_IQ2_XS:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_iq2_xs, data, nb);
+ } break;
+ case GGML_TYPE_IQ2_S:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_iq2_s, data, nb);
+ } break;
+ case GGML_TYPE_IQ3_XXS:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_iq3_xxs, data, nb);
+ } break;
+
+ case GGML_TYPE_IQ3_S:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_iq3_s, data, nb);
+ } break;
+ case GGML_TYPE_IQ4_XS:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_xs, data, nb);
+ } break;
+ case GGML_TYPE_IQ4_NL:
+ {
+ VALIDATE_ROW_DATA_D_F16_IMPL(block_iq4_nl, data, nb);
+ } break;
+
+ case GGML_TYPE_I8:
+ case GGML_TYPE_I16:
+ case GGML_TYPE_I32:
+ case GGML_TYPE_I64:
+ // nothing to validate
+ break;
+ default:
+ {
+ fprintf(stderr, "%s: invalid type %d\n", __func__, type);
+ return false;
+ }
+ }
+
+ return true;
+}
diff --git a/python/freetoken/kernel/csrc/ggml/src/ggml-quants.h b/python/freetoken/kernel/csrc/ggml/src/ggml-quants.h
new file mode 100644
index 00000000..75188f1a
--- /dev/null
+++ b/python/freetoken/kernel/csrc/ggml/src/ggml-quants.h
@@ -0,0 +1,115 @@
+#pragma once
+
+#define GGML_COMMON_DECL_C
+#include "ggml-common.h"
+
+#include "ggml.h"
+
+// GGML internal header
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// NOTE: these functions are defined as GGML_API because they used by the CPU backend
+
+// Quantization
+GGML_API void quantize_row_q1_0_ref(const float * GGML_RESTRICT x, block_q1_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q2_0_ref(const float * GGML_RESTRICT x, block_q2_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q4_0_ref(const float * GGML_RESTRICT x, block_q4_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q4_1_ref(const float * GGML_RESTRICT x, block_q4_1 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q5_0_ref(const float * GGML_RESTRICT x, block_q5_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q5_1_ref(const float * GGML_RESTRICT x, block_q5_1 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q8_0_ref(const float * GGML_RESTRICT x, block_q8_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q8_1_ref(const float * GGML_RESTRICT x, block_q8_1 * GGML_RESTRICT y, int64_t k);
+
+GGML_API void quantize_row_mxfp4_ref(const float * GGML_RESTRICT x, block_mxfp4 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_nvfp4_ref(const float * GGML_RESTRICT x, block_nvfp4 * GGML_RESTRICT y, int64_t k);
+
+GGML_API void quantize_row_q2_K_ref(const float * GGML_RESTRICT x, block_q2_K * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q3_K_ref(const float * GGML_RESTRICT x, block_q3_K * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q4_K_ref(const float * GGML_RESTRICT x, block_q4_K * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q5_K_ref(const float * GGML_RESTRICT x, block_q5_K * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q6_K_ref(const float * GGML_RESTRICT x, block_q6_K * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_q8_K_ref(const float * GGML_RESTRICT x, block_q8_K * GGML_RESTRICT y, int64_t k);
+
+GGML_API void quantize_row_tq1_0_ref(const float * GGML_RESTRICT x, block_tq1_0 * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_tq2_0_ref(const float * GGML_RESTRICT x, block_tq2_0 * GGML_RESTRICT y, int64_t k);
+
+GGML_API void quantize_row_iq3_xxs_ref(const float * GGML_RESTRICT x, block_iq3_xxs * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_iq4_nl_ref (const float * GGML_RESTRICT x, block_iq4_nl * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_iq4_xs_ref (const float * GGML_RESTRICT x, block_iq4_xs * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_iq3_s_ref (const float * GGML_RESTRICT x, block_iq3_s * GGML_RESTRICT y, int64_t k);
+GGML_API void quantize_row_iq2_s_ref (const float * GGML_RESTRICT x, block_iq2_s * GGML_RESTRICT y, int64_t k);
+
+// Dequantization
+GGML_API void dequantize_row_q1_0(const block_q1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q2_0(const block_q2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q4_0(const block_q4_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q4_1(const block_q4_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q5_0(const block_q5_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q5_1(const block_q5_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q8_0(const block_q8_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+//GGML_API void dequantize_row_q8_1(const block_q8_1 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+GGML_API void dequantize_row_mxfp4(const block_mxfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_nvfp4(const block_nvfp4 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+GGML_API void dequantize_row_q2_K(const block_q2_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q3_K(const block_q3_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q4_K(const block_q4_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q5_K(const block_q5_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q6_K(const block_q6_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_q8_K(const block_q8_K * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+GGML_API void dequantize_row_tq1_0(const block_tq1_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_tq2_0(const block_tq2_0 * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+GGML_API void dequantize_row_iq2_xxs(const block_iq2_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq2_xs (const block_iq2_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq2_s (const block_iq2_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq3_xxs(const block_iq3_xxs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq1_s (const block_iq1_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq1_m (const block_iq1_m * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq4_nl (const block_iq4_nl * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq4_xs (const block_iq4_xs * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+GGML_API void dequantize_row_iq3_s (const block_iq3_s * GGML_RESTRICT x, float * GGML_RESTRICT y, int64_t k);
+
+// Quantization utilizing an importance matrix (a.k.a. "Activation aWare Quantization")
+GGML_API size_t quantize_iq2_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq2_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq2_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq3_xxs(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq1_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq1_m (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq4_nl (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq4_xs (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_iq3_s (const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+GGML_API size_t quantize_tq1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_tq2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+GGML_API size_t quantize_q2_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q3_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q4_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q5_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q6_K(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q1_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q2_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q4_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q4_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q5_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q5_1(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_q8_0(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+GGML_API size_t quantize_mxfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+GGML_API size_t quantize_nvfp4(const float * GGML_RESTRICT src, void * GGML_RESTRICT dst, int64_t nrows, int64_t n_per_row, const float * imatrix);
+
+GGML_API void iq2xs_init_impl(enum ggml_type type);
+GGML_API void iq2xs_free_impl(enum ggml_type type);
+GGML_API void iq3xs_init_impl(int grid_size);
+GGML_API void iq3xs_free_impl(int grid_size);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/python/freetoken/kernel/ggml_cpu/__init__.py b/python/freetoken/kernel/ggml_cpu/__init__.py
new file mode 100644
index 00000000..909c973e
--- /dev/null
+++ b/python/freetoken/kernel/ggml_cpu/__init__.py
@@ -0,0 +1,37 @@
+"""ggml-cpu row kernels for GGUF experts on the CPU MoE executor.
+
+setup.py builds one module per x86 variant of llama.cpp's GGML_CPU_ALL_VARIANTS; select() picks the one this CPU runs best.
+"""
+
+from __future__ import annotations
+
+import functools
+import importlib
+import os
+import pkgutil
+from typing import NamedTuple
+
+
+class Variant(NamedTuple):
+ name: str
+ api: int # address of the variant's kernel table, for _cpu_moe.set_ggml_cpu
+
+
+def select() -> Variant | None:
+ """The highest-scoring variant this CPU runs (upstream's rule), or the one FREETOKEN_GGML_CPU_VARIANT names; None when the build has none."""
+ names = sorted(m.name[1:] for m in pkgutil.iter_modules(__path__) if m.name.startswith("_"))
+ forced = os.environ.get("FREETOKEN_GGML_CPU_VARIANT")
+ if forced:
+ if forced not in names:
+ raise RuntimeError(f"FREETOKEN_GGML_CPU_VARIANT={forced}: this build has the ggml-cpu variants {names}")
+ names = [forced]
+ best, best_score = None, 0
+ for name in names:
+ # importing a variant runs no code built for its ISA extensions
+ mod = importlib.import_module(f"{__name__}._{name}")
+ if (score := mod.score()) > best_score:
+ best, best_score = Variant(name, mod.api()), score
+ if forced and best is None:
+ raise RuntimeError(f"FREETOKEN_GGML_CPU_VARIANT={forced}: this CPU lacks an ISA extension the variant is built for")
+ return best
diff --git a/python/freetoken/layers/quantization/moe/gguf.py b/python/freetoken/layers/quantization/moe/gguf.py
index dfd545be..416c2f4b 100644
--- a/python/freetoken/layers/quantization/moe/gguf.py
+++ b/python/freetoken/layers/quantization/moe/gguf.py
@@ -142,13 +142,9 @@ class GgmlMoEKernel(MoEKernel):
class GGUFMoEMethod(MoEMethod):
candidates = (GgmlMoEKernel,)
- @property
- def cpu_format(self) -> str | None:
- # the CPU executor reads Q4_0 experts in their GGUF rows and BF16 experts as its bf16 banks; other ggml types have no CPU kernel yet
+ def cpu_spec(self) -> tuple[str | None, dict[str, int]]:
gate_up, down = expert_types(self.cfg)
- names = {type_name(gate_up), type_name(down)}
- if names == {"Q4_0"}:
- return "q4_0"
- if names == {"BF16"}:
- return "bf16"
- return self.kernel.cpu_format
+ # BF16 experts keep the executor's bf16 kernels, which use AVX512-BF16 where the CPU has it; ggml-cpu is built for AVX2 only
+ if float_dtype(gate_up) is torch.bfloat16 and float_dtype(down) is torch.bfloat16:
+ return "bf16", {}
+ return "gguf", {"gate_up_type": gate_up, "down_type": down}
diff --git a/python/freetoken/moe/benchbw.py b/python/freetoken/moe/benchbw.py
index 86a03f41..f7ea5aa3 100644
--- a/python/freetoken/moe/benchbw.py
+++ b/python/freetoken/moe/benchbw.py
@@ -63,7 +63,7 @@ from freetoken.utils import init_logger
logger = init_logger(__name__)
# Formats the CPU MoE C++ kernel can compute AND this bench can build banks for; anything
-# else is offload-only here. (The kernel also does q4_0, but this bench has no q4_0 banks.)
+# else is offload-only here.
_CPU_MOE_FORMATS = frozenset({"bf16", "nvfp4", "mxfp4_triton", "ds_fp4", "fp8_block"})
# Formats this bench can build synthetic (correctly-sized) banks for.
_BUILDABLE_FORMATS = frozenset({"bf16", "nvfp4", "fp8_block", "mxfp4_triton", "ds_fp4"})
diff --git a/python/freetoken/moe/cpu_executor.py b/python/freetoken/moe/cpu_executor.py
index 5d9bd3c9..4c0aee7a 100644
--- a/python/freetoken/moe/cpu_executor.py
+++ b/python/freetoken/moe/cpu_executor.py
@@ -147,6 +147,16 @@ def resolve_threads_and_affinity(requested: int) -> tuple[int, list[int]]:
return len(reps), list(reps)
+def _format_label(name: str, opts: tuple) -> str:
+ """One format instance as the startup log names it, e.g. ``gguf(gate_up=IQ2_XXS, down=IQ4_NL)``."""
+ if name == "gguf":
+ from freetoken.layers.quantization.ggml_types import type_name
+
+ gate_up, down = dict(opts)["gate_up_type"], dict(opts)["down_type"]
+ return f"gguf(gate_up={type_name(gate_up)}, down={type_name(down)})"
+ return name + (f"({', '.join(f'{k}={v}' for k, v in opts)})" if opts else "")
+
+
class CpuMoeExecutor:
"""Decode-time CPU expert compute over an ``OffloadMoeCache``'s host banks
(the formats ``_cpu_moe.weight_formats()`` lists — see ``_resolve_banks``)."""
@@ -182,6 +192,15 @@ class CpuMoeExecutor:
f"{sorted(_cpu_moe.weight_formats())} formats, but this checkpoint's experts are "
f"{f!r}; use --moe-strategy offload (GPU-side dequant) instead."
)
+ if f == "gguf":
+ from freetoken.kernel import ggml_cpu
+
+ if (variant := ggml_cpu.select()) is None:
+ raise NotImplementedError(
+ "--moe-strategy cpu/hybrid cannot compute GGUF experts here: this build has no ggml-cpu kernels (x86-64 only); "
+ "use --moe-strategy offload instead."
+ )
+ _cpu_moe.set_ggml_cpu(variant.api)
if activation not in _ACT_IDS:
raise NotImplementedError(f"CPU MoE backend: unsupported activation {activation!r}")
if apply_router_weight_on_input and (output_only := {"mxfp4_triton", "ds_fp4"} & set(names)):
@@ -348,7 +367,7 @@ class CpuMoeExecutor:
self._watchdog.start()
- fmt_label = "; ".join(f"{name}{dict(opts) if opts else ''}: {n} layers" for (name, opts), n in layer_counts.items())
+ fmt_label = "; ".join(f"{_format_label(name, opts)}: {n} layers" for (name, opts), n in layer_counts.items())
logger.info_rank0(
f"CPU MoE executor ready: threads={nthreads} (pinned to cores "
f"{core_ids[0]}..{core_ids[-1]}) isa={self.isa} "
@@ -404,6 +423,8 @@ class CpuMoeExecutor:
(see ``_make_table``), not a single bank's -- the C++ side resolves
``tbl[layer_id]`` per task.
"""
+ if fmt == "gguf":
+ return self._resolve_gguf_banks(banks, rep, int(opts["gate_up_type"]), int(opts["down_type"]))
if fmt == "bf16":
gate_up = banks["gate_up"]
down = banks["down"]
@@ -418,9 +439,6 @@ class CpuMoeExecutor:
banks = self._tables(gate_up=gate_up, down=down)
return banks, {}, (H, I)
- if fmt == "q4_0":
- return self._resolve_q4_0_banks(banks, rep)
-
if fmt == "mxfp4_triton":
return self._resolve_mxfp4_banks(banks, rep)
@@ -447,23 +465,18 @@ class CpuMoeExecutor:
banks = self._tables(gate_up=gup, down=dnp, gate_up_scale=gus, gate_up_global=gug, down_scale=dns, down_global=dng)
return banks, {}, (H, I)
- def _resolve_q4_0_banks(self, banks: dict, rep: int = 0) -> tuple[dict, dict, tuple[int, int]]:
- """Native GGUF Q4_0 schema (gemma4 GGUF): per-32 blocks (fp16 scale + 16 nibble
- bytes), row-major over K -- the *same* packed banks the GPU offload path streams.
- gate_up is [S, 2I, H//32*18], down is [S, H, I//32*18]; the C++ W4A16 GEMV reads a
- row in place (18 bytes / 32 K) and dequantizes weights inside the K-loop."""
+ def _resolve_gguf_banks(self, banks: dict, rep: int, gate_up_type: int, down_type: int) -> tuple[dict, dict, tuple[int, int]]:
+ """GGUF rows read in place: gate_up [E, 2I, row_bytes(H)] (gate rows, then up rows) and down [E, H, row_bytes(I)], one ggml type each; float rows keep their dtype."""
+ from freetoken.layers.quantization.ggml_types import float_dtype, row_bytes, type_name
+
gate_up, down = banks["gate_up"], banks["down"]
- assert gate_up[rep].dtype == torch.uint8 and down[rep].dtype == torch.uint8, (
- gate_up[rep].dtype, down[rep].dtype,
- )
- I = int(gate_up[rep].shape[1] // 2)
- H = int(down[rep].shape[1])
- assert gate_up[rep].shape[1] == 2 * I
- assert H % 32 == 0 and I % 32 == 0, (H, I)
- assert int(gate_up[rep].shape[2]) == (H // 32) * 18, (gate_up[rep].shape, H)
- assert int(down[rep].shape[2]) == (I // 32) * 18, (down[rep].shape, I)
+ H, I = int(down[rep].shape[1]), int(gate_up[rep].shape[1] // 2)
+ for role, t, ggml_type, k in (("gate_up", gate_up[rep], gate_up_type, H), ("down", down[rep], down_type, I)):
+ dtype = float_dtype(ggml_type) if float_dtype(ggml_type) is not None else torch.uint8
+ if t.dtype != dtype or t.shape[-1] * t.element_size() != row_bytes(ggml_type, k):
+ raise ValueError(f"MoE layer {rep} {role} bank {tuple(t.shape)} {t.dtype} is not {type_name(ggml_type)} rows of {k} values")
banks = self._tables(gate_up=gate_up, down=down)
- return banks, {}, (H, I)
+ return banks, {"gate_up_type": gate_up_type, "down_type": down_type}, (H, I)
def _resolve_mxfp4_banks(self, banks: dict, rep: int = 0) -> tuple[dict, dict, tuple[int, int]]:
"""gpt-oss mxfp4 ``mxfp4_triton`` schema: transposed split-K blocks/scales
diff --git a/setup.py b/setup.py
index b9683731..070b8fd5 100644
--- a/setup.py
+++ b/setup.py
@@ -2,11 +2,13 @@ from __future__ import annotations
import importlib.util
import os
+import platform
from pathlib import Path
import sys
-from setuptools import setup
+from setuptools import Extension, setup
+from setuptools.command.build_clib import build_clib
from torch.utils.cpp_extension import BuildExtension, CUDA_HOME, CppExtension, ROCM_HOME
@@ -96,8 +98,154 @@ else:
_check_toolchain()
+# ggml-cpu's row kernels for GGUF experts on the CPU MoE executor: one module per x86 variant of llama.cpp's GGML_CPU_ALL_VARIANTS (ggml/src/CMakeLists.txt), freetoken.kernel.ggml_cpu._<variant>.
+# freetoken.kernel.ggml_cpu.select() imports the best one this CPU runs; each module's entry is built without the variant's ISA flags so that check is safe on any CPU.
+GGML_DIR = "python/freetoken/kernel/csrc/ggml"
+GGML_CPU = platform.machine().lower() in ("x86_64", "amd64")
+_HASWELL = ("SSE42", "AVX", "F16C", "FMA", "AVX2", "BMI2")
+GGML_CPU_VARIANTS = {
+ "x64": (),
+ "sse42": ("SSE42",),
+ "sandybridge": ("SSE42", "AVX"),
+ "ivybridge": ("SSE42", "AVX", "F16C"),
+ "piledriver": ("SSE42", "AVX", "F16C", "FMA"),
+ "haswell": _HASWELL,
+ "skylakex": (*_HASWELL, "AVX512"),
+ "cannonlake": (*_HASWELL, "AVX512", "AVX512_VBMI"),
+ "cascadelake": (*_HASWELL, "AVX512", "AVX512_VNNI"),
+ "icelake": (*_HASWELL, "AVX512", "AVX512_VBMI", "AVX512_VNNI"),
+ "cooperlake": (*_HASWELL, "AVX512", "AVX512_VNNI", "AVX512_BF16"),
+ "zen4": (*_HASWELL, "AVX512", "AVX512_VBMI", "AVX512_VNNI", "AVX512_BF16"),
+ "alderlake": (*_HASWELL, "AVX_VNNI"),
+ "sapphirerapids": (*_HASWELL, "AVX512", "AVX512_VBMI", "AVX512_VNNI", "AVX512_BF16", "AMX_TILE", "AMX_INT8"),
+}
+_GGML_GCC_FLAGS = {
+ "SSE42": ["-msse4.2"], "F16C": ["-mf16c"], "FMA": ["-mfma"], "BMI2": ["-mbmi2"], "AVX": ["-mavx"], "AVX2": ["-mavx2"],
+ "AVX_VNNI": ["-mavxvnni"], "AVX512": ["-mavx512f", "-mavx512cd", "-mavx512vl", "-mavx512dq", "-mavx512bw"],
+ "AVX512_VBMI": ["-mavx512vbmi"], "AVX512_VNNI": ["-mavx512vnni"], "AVX512_BF16": ["-mavx512bf16"],
+ "AMX_TILE": ["-mamx-tile"], "AMX_INT8": ["-mamx-int8"],
+}
+
+
+def _ggml_variant_flags(feats: tuple[str, ...]) -> tuple[list[str], list[str]]:
+ """(ISA flags, macros) of one variant, as upstream's ggml-cpu/CMakeLists.txt sets them; the macros also tell cpu-feats.cpp what to check."""
+ if sys.platform != "win32":
+ return [f for feat in feats for f in _GGML_GCC_FLAGS[feat]], [f"GGML_{feat}" for feat in feats]
+ # cl takes one /arch: level and does not define the macros of the extensions above it
+ if "AVX512" in feats:
+ flags, macros = ["/arch:AVX512"], ["GGML_AVX512"]
+ if "AVX512_VBMI" in feats:
+ macros.append("__AVX512VBMI__")
+ if "AVX512_VNNI" in feats:
+ macros += ["__AVX512VNNI__", "GGML_AVX512_VNNI"]
+ elif "AVX2" in feats:
+ flags, macros = ["/arch:AVX2"], ["GGML_AVX2", "GGML_FMA", "GGML_F16C"]
+ elif "AVX" in feats:
+ flags, macros = ["/arch:AVX"], ["GGML_AVX"]
+ elif "SSE42" in feats:
+ flags, macros = ["/arch:SSE4.2"], ["GGML_SSE42"]
+ else:
+ flags, macros = [], []
+ if "AVX_VNNI" in feats:
+ macros += ["__AVXVNNI__", "GGML_AVX_VNNI"]
+ if "BMI2" in feats:
+ macros += ["__BMI2__", "GGML_BMI2"]
+ return flags, macros
+
+
+def _ggml_cpu_targets() -> tuple[list, list, dict[str, list[str]]]:
+ """(build_clib libraries, variant modules, the static libraries each module links in link order)."""
+ if not GGML_CPU:
+ return [], [], {}
+ if sys.platform == "win32":
+ cflags, c_std, cxx_std, define = ["/O2", "/DNDEBUG", "/utf-8", "/D_CRT_SECURE_NO_WARNINGS"], "/std:c11", "/std:c++20", "/D"
+ # upstream builds these variants only with GCC or Clang
+ variants = {k: v for k, v in GGML_CPU_VARIANTS.items() if k not in ("ivybridge", "piledriver", "cooperlake", "zen4", "sapphirerapids")}
+ else:
+ cflags, c_std, cxx_std, define = ["-O3", "-fPIC", "-fvisibility=hidden", "-DNDEBUG"], "-std=c11", "-std=c++20", "-D"
+ variants = GGML_CPU_VARIANTS
+ build = {
+ "include_dirs": [f"{GGML_DIR}/include", f"{GGML_DIR}/src", f"{GGML_DIR}/ggml-cpu", GGML_DIR],
+ "obj_deps": {"": sorted(str(p.relative_to(ROOT)) for d in ("include", "src", "ggml-cpu", "freetoken") for p in (ROOT / GGML_DIR / d).rglob("*.h"))},
+ }
+ # ggml-base, built once without ISA flags like upstream's; arch/x86/quants.c calls its quantize_row_*_ref
+ libraries = [("freetoken_ggml_base", {"sources": [f"{GGML_DIR}/src/ggml-quants.c"], "cflags": [*cflags, c_std], **build})]
+ modules, module_libs = [], {}
+ for variant, feats in variants.items():
+ flags, macros = _ggml_variant_flags(feats)
+ defines = [define + m for m in macros]
+ libraries += [
+ (f"freetoken_ggml_cpu_{variant}", {
+ "sources": [f"{GGML_DIR}/ggml-cpu/quants.c", f"{GGML_DIR}/ggml-cpu/arch/x86/quants.c", f"{GGML_DIR}/freetoken/cpu_kernels.c"],
+ "cflags": [*cflags, *flags, *defines, c_std],
+ **build,
+ }),
+ # the float row kernels; C++, so a library of their own
+ (f"freetoken_ggml_cpu_vec_{variant}", {"sources": [f"{GGML_DIR}/ggml-cpu/vec.cpp"], "cflags": [*cflags, *flags, *defines, cxx_std], **build}),
+ ]
+ name = f"freetoken.kernel.ggml_cpu._{variant}"
+ module_libs[name] = [f"freetoken_ggml_cpu_{variant}", f"freetoken_ggml_cpu_vec_{variant}", "freetoken_ggml_base"]
+ modules.append(Extension(
+ name=name,
+ sources=[f"{GGML_DIR}/freetoken/variant_module.cpp"],
+ # absolute: ninja compiles from the build directory
+ include_dirs=[str(ROOT / d) for d in (*build["include_dirs"], f"{GGML_DIR}/freetoken")],
+ extra_compile_args=[*cflags, *defines, f"{define}FT_GGML_VARIANT={variant}", cxx_std],
+ ))
+ return libraries, modules, module_libs
+
+
+ggml_cpu_libraries, ggml_cpu_modules, ggml_cpu_module_libs = _ggml_cpu_targets()
+
+
+class BuildClib(build_clib):
+ def build_libraries(self, libraries):
+ # the variants compile the same sources with different flags, so each library gets its own object directory
+ build_temp = self.build_temp
+ try:
+ for lib in libraries:
+ self.build_temp = os.path.join(build_temp, lib[0])
+ super().build_libraries([lib])
+ finally:
+ self.build_temp = build_temp
+
+
+class BuildExt(BuildExtension):
+ def run(self):
+ # build_ext alone (`setup.py build_ext --inplace`) does not run build_clib
+ if self.distribution.has_c_libraries():
+ self.run_command("build_clib")
+ # build_ext would link every build_clib library into every extension; build_extension adds each module's own
+ libraries, self.distribution.libraries = self.distribution.libraries, None
+ try:
+ super().run()
+ finally:
+ self.distribution.libraries = libraries
+
+ def build_extension(self, ext):
+ if ext.name in ggml_cpu_module_libs:
+ build_clib_dir = self.get_finalized_command("build_clib").build_clib
+ if build_clib_dir not in ext.library_dirs:
+ ext.library_dirs.append(build_clib_dir)
+ for name in ggml_cpu_module_libs[ext.name]:
+ if name not in ext.libraries:
+ ext.libraries.append(name)
+ lib = os.path.join(build_clib_dir, self.compiler.library_filename(name))
+ # relink when only a static library changed
+ if lib not in ext.depends:
+ ext.depends.append(lib)
+ # every variant module compiles the same source, so each gets its own object directory
+ build_temp, self.build_temp = self.build_temp, os.path.join(self.build_temp, ext.name)
+ try:
+ super().build_extension(ext)
+ finally:
+ self.build_temp = build_temp
+ return
+ super().build_extension(ext)
+
setup(
+ libraries=ggml_cpu_libraries,
ext_modules=[
CppExtension(
name="freetoken.kernel._pinned_tensor",
@@ -126,7 +274,7 @@ setup(
for p in (ROOT / "python/freetoken/kernel/csrc/cpu_moe/kernels").glob("*.cpp")
),
],
- include_dirs=[KERNEL_INCLUDE, *runtime_include_dirs],
+ include_dirs=[KERNEL_INCLUDE, *runtime_include_dirs, str(ROOT / GGML_DIR / "freetoken")],
library_dirs=runtime_library_dirs,
libraries=[runtime_lib],
extra_compile_args={"cxx": extra_compile + ["-pthread"], "nvcc": device_compile},
@@ -142,6 +290,7 @@ setup(
extra_compile_args=["-O3", "-std=c++20"],
)
] if sys.platform == "linux" else []),
+ *ggml_cpu_modules,
],
- cmdclass={"build_ext": BuildExtension.with_options(use_ninja=True)},
+ cmdclass={"build_clib": BuildClib, "build_ext": BuildExt.with_options(use_ninja=True)},
)
diff --git a/tests/README.md b/tests/README.md
index adc87035..85865673 100644
--- a/tests/README.md
+++ b/tests/README.md
@@ -14,7 +14,7 @@ in its filename (`kvcache/test_dsv4_pool.py`, `models/test_glm4_nvfp4.py`).
| `tokenizer/` | `freetoken.tokenizer` — tokenize/detokenize request plumbing, thinking-mode resolution |
| `engine/` | `freetoken.engine` — cache budget planning, config resolution, backend gating |
| `kernels/` | `freetoken.kernel` / `freetoken.layers` — attention, rope, fused ops, JIT cache, pinned memory |
-| `moe/` | `freetoken.moe` — offload cache, CPU/GPU expert kernels, quantized backends (fp8/nvfp4/mxfp4/q4_0) |
+| `moe/` | `freetoken.moe` — offload cache, CPU/GPU expert kernels, quantized backends (fp8/nvfp4/mxfp4/gguf) |
| `models/` | `freetoken.models` — the registry, and the loading machinery every model shares (sharding, qkv/expert merge, streaming layers into banks) |
| `e2e/` | real-server gates: AIME 24/25/26 generation, runtime cache rebuild |
diff --git a/tests/moe/test_cpu_moe_q4_0.py b/tests/moe/test_cpu_moe_q4_0.py
deleted file mode 100644
index 9452f95c..00000000
--- a/tests/moe/test_cpu_moe_q4_0.py
+++ /dev/null
@@ -1,213 +0,0 @@
-"""CPU MoE executor -- native GGUF Q4_0 experts (``--moe-backend cpu``/``hybrid`` for
-the gemma4 GGUF checkpoint).
-
-The CPU W4A16 GEMV (``q4_0_dot`` in csrc/cpu_moe/kernels/q4_0.cpp) reads the *same*
-packed Q4_0 banks the GPU offload path streams and dequantizes weights inside the
-K-loop. We check it against the reference dequant (gguf-py's numpy dequantize) + the
-production bf16 GPU decode kernel on byte-identical banks: both are W4A16, so the
-only spread is weight bf16-rounding + reduction order -> tight relative tolerance.
-
-Part 2 covers CUDA-graph capture/replay (the cudaLaunchHostFunc submit/sync nodes
-must recompute from the freshly written pinned routing on each replay).
-"""
-
-from __future__ import annotations
-
-from types import SimpleNamespace
-
-import pytest
-import torch
-
-pytestmark = pytest.mark.skipif(not torch.cuda.is_available(), reason="needs CUDA")
-
-
-def _pack_q4_0(nibbles: torch.Tensor, scale: torch.Tensor) -> torch.Tensor:
- """Pack [S, OUT, K] uint8 nibble codes (0..15) + [S, OUT, K//32] fp16 scales into
- the native Q4_0 row layout [S, OUT, K//32*18]: each 32-elem block is a 2-byte fp16
- scale followed by 16 bytes whose byte j holds elem j (low nibble) and elem j+16
- (high nibble) -- the exact layout ``dequant_q4_0`` reads back."""
- S, OUT, K = nibbles.shape
- nb = K // 32
- blk = nibbles.reshape(S, OUT, nb, 32)
- lo = blk[..., :16] # elems 0..15 -> low nibbles of bytes 0..15
- hi = blk[..., 16:] # elems 16..31 -> high nibbles of bytes 0..15
- packed = (lo | (hi << 4)).to(torch.uint8) # [S, OUT, nb, 16]
- d_bytes = scale.to(torch.float16).view(torch.uint8).reshape(S, OUT, nb, 2)
- row = torch.cat([d_bytes, packed], dim=-1) # [S, OUT, nb, 18]
- return row.reshape(S, OUT, nb * 18).contiguous()
-
-
-def _dequant_bank(packed: torch.Tensor, K: int, dev) -> torch.Tensor:
- """[S, OUT, K//32*18] packed Q4_0 -> [S, OUT, K] bf16 (storage order == elem order)."""
- from freetoken.third_party.gguf import GGMLQuantizationType, quants
-
- S, OUT, _ = packed.shape
- values = quants.dequantize(packed.numpy(), GGMLQuantizationType.Q4_0)
- return torch.from_numpy(values).reshape(S, OUT, K).to(dev, torch.bfloat16)
-
-
-def _make_q4_0_cache(L, E, H, I, seed=0):
- """Random but valid native Q4_0 banks for the cpu backend (pinned host tensors)."""
- from freetoken.kernel.pinned import alloc_pinned_tensor
-
- torch.manual_seed(seed)
- S = L * E
-
- def rows(OUT, K):
- nib = torch.randint(0, 16, (S, OUT, K), dtype=torch.uint8)
- scale = 0.02 + 0.03 * torch.rand(S, OUT, K // 32)
- packed = _pack_q4_0(nib, scale) # [S, OUT, K//32*18]
- pinned = alloc_pinned_tensor(*packed.shape, dtype=torch.uint8)
- pinned.copy_(packed)
- return pinned
-
- return SimpleNamespace(
- quant_format="q4_0",
- bank_sources={"gate_up": list(rows(2 * I, H).split(E)), "down": list(rows(H, I).split(E))},
- num_layers=L,
- num_experts=E,
- decode_target="cpu",
- cpu_executor=None,
- )
-
-
[email protected]("bs", [1, 3, 8])
-def test_cpu_decode_q4_0_matches_dequant_then_gpu(bs):
- """CPU inline-dequant Q4_0 GEMV vs. canonical dequant_q4_0 + bf16 GPU decode."""
- from freetoken.moe.cpu_executor import CpuMoeExecutor
- from freetoken.moe.fused import fused_experts_decode_impl
-
- torch.manual_seed(400 + bs)
- L, E, H, I, top_k = 3, 16, 2816, 704, 8 # gemma-4-26B-A4B geometry (H,I % 32 == 0)
- layer = 1
- dev = torch.device("cuda")
- cache = _make_q4_0_cache(L, E, H, I)
-
- ex = CpuMoeExecutor(
- cache,
- top_k=top_k,
- activation="gelu_tanh",
- apply_router_weight_on_input=False,
- num_threads=0,
- max_tokens=bs,
- device=dev,
- )
-
- hidden = torch.randn(bs, H, device=dev, dtype=torch.bfloat16) * 0.5
- ids = torch.stack([torch.randperm(E, device=dev)[:top_k] for _ in range(bs)]).to(torch.int32)
- w = torch.rand(bs, top_k, device=dev, dtype=torch.float32)
-
- cpu_out = ex.decode(layer, hidden, w, ids).float()
- torch.cuda.synchronize()
-
- b = cache.bank_sources
- gate_up_layer = _dequant_bank(b["gate_up"][layer], H, dev) # [E, 2I, H]
- down_layer = _dequant_bank(b["down"][layer], I, dev) # [E, H, I]
- gpu_out = fused_experts_decode_impl(
- hidden, gate_up_layer, down_layer, w, ids.clone(), "gelu_tanh", False
- ).float()
-
- rel = (cpu_out - gpu_out).abs().max() / (gpu_out.abs().max() + 1e-6)
- assert rel < 2e-2, f"q4_0 bs={bs} rel err {rel.item()}"
-
-
-def test_cpu_decode_q4_0_matches_ggml_mmvq():
- """Sanity: the CPU W4A16 GEMV lands close to the GPU ggml MMVQ (W4A8) kernel the
- offload path uses -- a looser tol since MMVQ quantizes activations to int8."""
- from freetoken.kernel import ggml
- from freetoken.layers.activation import gated_act_and_mul
- from freetoken.moe.cpu_executor import CpuMoeExecutor
- from freetoken.third_party.gguf import GGMLQuantizationType
-
- torch.manual_seed(77)
- L, E, H, I, top_k = 2, 16, 2816, 704, 8
- layer, bs = 1, 4
- dev = torch.device("cuda")
- cache = _make_q4_0_cache(L, E, H, I)
-
- ex = CpuMoeExecutor(
- cache, top_k=top_k, activation="gelu_tanh",
- apply_router_weight_on_input=False, num_threads=0, max_tokens=bs, device=dev,
- )
-
- hidden = torch.randn(bs, H, device=dev, dtype=torch.bfloat16) * 0.5
- ids = torch.stack([torch.randperm(E, device=dev)[:top_k] for _ in range(bs)]).to(torch.int32)
- w = torch.rand(bs, top_k, device=dev, dtype=torch.float32)
-
- cpu_out = ex.decode(layer, hidden, w, ids).float()
- torch.cuda.synchronize()
-
- b = cache.bank_sources
- q4_0 = int(GGMLQuantizationType.Q4_0)
- h = ggml.mul_mat_id(b["gate_up"][layer].to(dev), q4_0, hidden.float(), ids)
- inter = torch.empty(bs, top_k, I, device=dev)
- gated_act_and_mul("gelu_tanh", h.view(-1, 2 * I), inter.view(-1, I))
- y = ggml.mul_mat_id(b["down"][layer].to(dev), q4_0, inter, ids)
- gpu_out = ggml.moe_weighted_sum(y, w)
-
- rel = (cpu_out - gpu_out).abs().max() / (gpu_out.abs().max() + 1e-6)
- assert rel < 6e-2, f"q4_0 vs ggml-mmvq rel err {rel.item()}"
-
-
-def test_cpu_moe_decode_q4_0_cuda_graph_replay():
- """Q4_0 CPU path under capture/replay: the host nodes must recompute the GEMV from
- the freshly written pinned routing on each replay (dep flows through pinned buffers)."""
- from freetoken.moe.cpu_executor import CpuMoeExecutor
- from freetoken.moe.fused import fused_experts_decode_impl
-
- torch.manual_seed(9)
- L, E, H, I, top_k = 2, 8, 2816, 704, 8
- layer, bs = 1, 4
- cache = _make_q4_0_cache(L, E, H, I)
-
- dev = torch.device("cuda")
- stream = torch.cuda.Stream()
- torch.cuda.set_stream(stream)
-
- ex = CpuMoeExecutor(
- cache, top_k=top_k, activation="gelu_tanh",
- apply_router_weight_on_input=False, num_threads=8, max_tokens=bs, device=dev,
- )
- b = cache.bank_sources
- gate_up_layer = _dequant_bank(b["gate_up"][layer], H, dev)
- down_layer = _dequant_bank(b["down"][layer], I, dev)
-
- def reference(hidden, ids, w):
- return fused_experts_decode_impl(
- hidden, gate_up_layer, down_layer, w, ids.clone(), "gelu_tanh", False
- ).float()
-
- hidden = torch.randn(bs, H, device=dev, dtype=torch.bfloat16) * 0.5
- ids = torch.randint(0, E, (bs, top_k), device=dev, dtype=torch.int32)
- w = torch.rand(bs, top_k, device=dev, dtype=torch.float32)
-
- ex.decode(layer, hidden, w, ids) # eager warmup: materialize buffers + task
- torch.cuda.synchronize()
-
- g = torch.cuda.CUDAGraph()
- with torch.cuda.graph(g, stream=stream):
- out_cap = ex.decode(layer, hidden, w, ids)
- torch.cuda.synchronize()
-
- for it in range(3):
- torch.manual_seed(600 + it)
- hidden.copy_(torch.randn(bs, H, dtype=torch.bfloat16) * 0.5)
- ids.copy_(torch.randint(0, E, (bs, top_k), dtype=torch.int32))
- w.copy_(torch.rand(bs, top_k, dtype=torch.float32))
- g.replay()
- torch.cuda.synchronize()
- ref = reference(hidden, ids, w)
- rel = (out_cap.float() - ref).abs().max() / (ref.abs().max() + 1e-6)
- assert rel < 2e-2, f"q4_0 replay {it} rel err {rel.item()}"
-
- print("cpu moe cuda graph replay (q4_0) OK")
-
-
-if __name__ == "__main__":
- for bs in (1, 3, 8):
- test_cpu_decode_q4_0_matches_dequant_then_gpu(bs)
- print(f"q4_0 bs={bs} OK")
- test_cpu_decode_q4_0_matches_ggml_mmvq()
- print("q4_0 vs ggml-mmvq OK")
- test_cpu_moe_decode_q4_0_cuda_graph_replay()