Skip to content

Qwen3.8-Flash-Next on any consumer hardware: one-click install for Windows / Linux. Strata inference engine, OpenAI/Anthropic API on localhost, optional image input.

C++

metal: the IQ2_XXS expert GEMV ported - 87.7 GB/s, and why it cannot go faster as a warp-per-row kernel

Niko12219c1e61a6f37f
diff --git a/metal/tests/dsv41/iq2xxs_gemv_parity.mm b/metal/tests/dsv41/iq2xxs_gemv_parity.mm
new file mode 100644
index 000000000..199563024
--- /dev/null
+++ b/metal/tests/dsv41/iq2xxs_gemv_parity.mm
@@ -0,0 +1,185 @@
+// iq2xxs_probe.mm - the IQ2_XXS expert GEMV in MSL, the 2.25bpw set's dominant expert format (49 GB of the 149 GB).
+//
+// Decode (ggml-quants.c dequantize_row_iq2_xxs, pinned 3cf03257f): a block is 66 bytes - an f16 d and 32 bytes of qs. Eight
+// groups of 32 weights: aux32[0..1] are the group's 8 bytes; db = d * (0.5 + (aux32[1] >> 28)) * 0.25; each of the group's four
+// sub-blocks takes grid = iq2xxs_grid[aux8[l]] (8 bytes) and signs = ksigns_iq2xs[(aux32[1] >> 7l) & 127], giving
+// y[j] = db * grid[j] * (signs & kmask_iq2xs[j] ? -1 : 1).
+//
+// The grid tables are read from the generated MSL header so the CPU reference uses exactly the same numbers as the GPU.
+// clang++ -O2 -fobjc-arc -framework Metal -framework Foundation -o iq2xxs_probe iq2xxs_probe.mm && ./iq2xxs_probe
+#import <Foundation/Foundation.h>
+#import <Metal/Metal.h>
+#include <cstdio>
+#include <cstdlib>
+#include <cmath>
+#include <cstring>
+#include <cstdint>
+#include <fstream>
+#include <regex>
+#include <string>
+#include <vector>
+#include <mach/mach_time.h>
+
+static std::vector<uint64_t> load_msl_table(const std::string& path, const std::string& name, size_t expect) {
+ std::ifstream f(path);
+ std::string all((std::istreambuf_iterator<char>(f)), std::istreambuf_iterator<char>());
+ std::regex re("constant \\w+ " + name + "\\[[^\\]]*\\] = \\{([^}]*)\\}");
+ std::smatch m;
+ std::vector<uint64_t> out;
+ if (!std::regex_search(all, m, re)) { fprintf(stderr, "table %s not found\n", name.c_str()); exit(2); }
+ std::string body = m[1];
+ size_t i = 0;
+ while (i < body.size()) {
+ while (i < body.size() && !isdigit((unsigned char) body[i])) ++i;
+ if (i >= body.size()) break;
+ size_t j = i;
+ while (j < body.size() && (isdigit((unsigned char) body[j]) || body[j] == 'x' || isxdigit((unsigned char) body[j]))) ++j;
+ out.push_back(strtoull(body.substr(i, j - i).c_str(), nullptr, 0));
+ i = j;
+ if (out.size() > expect) break;
+ }
+ if (out.size() != expect) { fprintf(stderr, "table %s: %zu of %zu\n", name.c_str(), out.size(), expect); exit(2); }
+ return out;
+}
+
+int main() {
+ @autoreleasepool {
+ const int n = 65536, k = 5120; // rows and k
+ const int nb = k / 256; // IQ2_XXS blocks a row
+ const size_t bs = 66; // bytes a block
+ auto grid = load_msl_table("ggml_grids.h", "iq2xxs_grid", 256);
+ auto signs_t = load_msl_table("ggml_grids.h", "ksigns_iq2xs", 128);
+
+ // a synthetic pack of blocks (the decoder does not care where the bytes come from)
+ std::vector<uint8_t> w((size_t) n * nb * bs);
+ srand(3);
+ for (size_t i = 0; i < w.size(); ++i) w[i] = (uint8_t) (rand() & 0xff);
+ for (int r = 0; r < n; ++r)
+ for (int b = 0; b < nb; ++b) {
+ // a sane f16 scale in the block's first two bytes
+ _Float16 d = (_Float16) (((rand() % 900) + 100) / 8000.0f);
+ memcpy(&w[((size_t) r * nb + b) * bs], &d, 2);
+ }
+ std::vector<float> x(k), y(n);
+ for (int i = 0; i < k; ++i) x[i] = ((rand() % 20001) - 10000) / 8192.0f;
+
+ // pull the grid table into the MSL by textual include at compile time
+ std::string grids;
+ {
+ std::ifstream f("ggml_grids.h");
+ std::string line;
+ while (std::getline(f, line)) {
+ if (line.rfind("#pragma once", 0) == 0 || line.rfind("//", 0) == 0 || line.empty()) continue;
+ grids += line + "\n";
+ }
+ }
+ std::string src = "#include <metal_stdlib>\nusing namespace metal;\n" + grids + R"MSL(
+kernel void iq2xxs_gemv(device const uchar* w [[buffer(0)]], device const float* x [[buffer(1)]],
+ device float* y [[buffer(2)]], constant uint& k [[buffer(3)]], constant uint& n [[buffer(4)]],
+ uint gid [[thread_position_in_grid]], uint lane [[thread_index_in_threadgroup]]) {
+ const uint r = gid / 32u;
+ if (r >= n) return;
+ const uint nb = k / 256u;
+ const device uchar* base = w + (size_t) r * nb * 66u;
+ float acc = 0.0f;
+ for (uint b = lane; b < nb; b += 32u) {
+ const device uchar* blk = base + (size_t) b * 66u;
+ const float d = float(as_type<half>(*(const device ushort*) blk));
+ for (int ib = 0; ib < 8; ++ib) {
+ // groups advance 4 bytes and read 8, so an 8-byte vector load is misaligned on odd groups: read bytes
+ const uint aux0 = uint(blk[2 + 4 * ib]) | (uint(blk[3 + 4 * ib]) << 8) | (uint(blk[4 + 4 * ib]) << 16) | (uint(blk[5 + 4 * ib]) << 24);
+ const uint aux1 = uint(blk[6 + 4 * ib]) | (uint(blk[7 + 4 * ib]) << 8) | (uint(blk[8 + 4 * ib]) << 16) | (uint(blk[9 + 4 * ib]) << 24);
+ const float db = d * (0.5f + float(aux1 >> 28)) * 0.25f;
+ for (int l = 0; l < 4; ++l) {
+ const ulong g = iq2xxs_grid[(aux0 >> (8 * l)) & 0xff];
+ const uint sg = uint(ksigns_iq2xs[(aux1 >> (7 * l)) & 127]);
+ const uint xoff = (uint) ib * 32u + (uint) l * 8u;
+ for (int j = 0; j < 8; ++j) {
+ const float wv = db * float((g >> (8 * j)) & 0xff) * ((sg & kmask_iq2xs[j]) ? -1.0f : 1.0f);
+ acc += wv * x[xoff + j];
+ }
+ }
+ }
+ }
+ acc += simd_shuffle_xor(acc, 16);
+ acc += simd_shuffle_xor(acc, 8);
+ acc += simd_shuffle_xor(acc, 4);
+ acc += simd_shuffle_xor(acc, 2);
+ acc += simd_shuffle_xor(acc, 1);
+ if (lane == 0) y[r] = acc;
+}
+)MSL";
+ const uint8_t kmask[8] = {1, 2, 4, 8, 16, 32, 64, 128};
+ // kmask_iq2xs comes from ggml_grids.h
+ std::string src_with_km = src;
+
+ id<MTLDevice> dev = MTLCreateSystemDefaultDevice();
+ NSError* err = nil;
+ id<MTLLibrary> lib = [dev newLibraryWithSource:[NSString stringWithUTF8String:src_with_km.c_str()] options:nil error:&err];
+ if (!lib) { printf("MSL compile failed: %s\n", [[err localizedDescription] UTF8String]); return 1; }
+ id<MTLComputePipelineState> pso = [dev newComputePipelineStateWithFunction:[lib newFunctionWithName:@"iq2xxs_gemv"] error:&err];
+ if (!pso) { printf("pipeline failed: %s\n", [[err localizedDescription] UTF8String]); return 1; }
+
+ id<MTLBuffer> bw = [dev newBufferWithBytes:w.data() length:w.size() options:MTLResourceStorageModeShared];
+ id<MTLBuffer> bx = [dev newBufferWithBytes:x.data() length:k * 4 options:MTLResourceStorageModeShared];
+ id<MTLBuffer> by = [dev newBufferWithLength:n * 4 options:MTLResourceStorageModeShared];
+ id<MTLCommandQueue> q = [dev newCommandQueue];
+ unsigned uk = k, un = n;
+ double best = 1e9;
+ mach_timebase_info_data_t tb; mach_timebase_info(&tb);
+ for (int it = 0; it < 5; ++it) {
+ id<MTLCommandBuffer> cb = [q commandBuffer];
+ id<MTLComputeCommandEncoder> ce = [cb computeCommandEncoder];
+ [ce setComputePipelineState:pso];
+ [ce setBuffer:bw offset:0 atIndex:0];
+ [ce setBuffer:bx offset:0 atIndex:1];
+ [ce setBuffer:by offset:0 atIndex:2];
+ [ce setBytes:&uk length:4 atIndex:3];
+ [ce setBytes:&un length:4 atIndex:4];
+ [ce dispatchThreads:MTLSizeMake((NSUInteger) n * 32, 1, 1) threadsPerThreadgroup:MTLSizeMake(32, 1, 1)];
+ [ce endEncoding];
+ const uint64_t t0 = mach_absolute_time();
+ [cb commit];
+ [cb waitUntilCompleted];
+ const double dt = (double) (mach_absolute_time() - t0) * tb.numer / tb.denom / 1e9;
+ if (dt < best) best = dt;
+ }
+ printf("throughput: %.2f GB/s (%.3f ms for %.1f MB)\n", (double) w.size() / best / 1e9, best * 1e3,
+ (double) w.size() / 1e6);
+
+ // the CPU reference: the same decode, from the same tables
+ const float* gy = (const float*) [by contents];
+ double num = 0, den = 0;
+ for (int r = 0; r < n; ++r) {
+ double acc = 0;
+ for (int b = 0; b < nb; ++b) {
+ const uint8_t* blk = &w[((size_t) r * nb + b) * bs];
+ _Float16 dh;
+ memcpy(&dh, blk, 2);
+ const double d = (double) (float) dh;
+ for (int ib = 0; ib < 8; ++ib) {
+ uint32_t aux[2];
+ memcpy(aux, blk + 2 + 4 * ib, 8);
+ const double db = d * (0.5 + (double) (aux[1] >> 28)) * 0.25;
+ const uint8_t* a8 = (const uint8_t*) aux;
+ for (int l = 0; l < 4; ++l) {
+ const uint64_t g = grid[a8[l]];
+ const uint32_t sg = (uint32_t) signs_t[(aux[1] >> (7 * l)) & 127];
+ const int xoff = ib * 32 + l * 8;
+ for (int j = 0; j < 8; ++j) {
+ const double wv = db * (double) ((g >> (8 * j)) & 0xff) * ((sg & kmask[j]) ? -1.0 : 1.0);
+ acc += wv * (double) x[xoff + j];
+ }
+ }
+ }
+ }
+ const double diff = (double) gy[r] - acc;
+ num += diff * diff;
+ den += acc * acc;
+ }
+ const double l2 = sqrt(num / den);
+ printf("relative L2 %.3e\n", l2);
+ printf("%s\n", l2 < 1e-4 ? "PASS" : "FAIL");
+ }
+ return 0;
+}