LLM inference in C/C++
C++ ggml
sycl: fuse the delta-net alpha gate (add + unary + mul) (#29687)
Titaniumtownff30363a0e3e
diff --git a/ggml/src/ggml-sycl/element_wise.cpp b/ggml/src/ggml-sycl/element_wise.cppindex 2e926abea7..0d263aa768 100644--- a/ggml/src/ggml-sycl/element_wise.cpp+++ b/ggml/src/ggml-sycl/element_wise.cpp@@ -452,6 +452,46 @@ static void unary_mul_sycl(const T * x, const T * g, T * dst, const int64_t k, c});}+// ADD(bias) + UNARY + MUL(scale) with both broadcast over dim 0, the delta-net alpha gate:+// dst[i] = op(a[i] + bias[i % ne0]) * scale[i % ne0]. k == ne0 makes that the flat index.+template<typename F>+static void add_unary_mul_flat_kernel(const float * a, const float * bias, const float * scale, float * dst,+ const int64_t k, const sycl::nd_item<1> &item_ct1, F op) {+ SYCL_GLOBAL_ID_LOOP(k, item_ct1) {+ dst[i] = op(a[i] + bias[i]) * scale[i];+ }+}++template<typename F>+static void add_unary_mul_bcast_kernel(const float * a, const float * bias, const float * scale, float * dst,+ const int64_t k, const sycl::uint3 ne0_fd, const sycl::nd_item<1> &item_ct1, F op) {+ SYCL_GLOBAL_ID_LOOP(k, item_ct1) {+ const uint32_t h = fastmodulo((uint32_t) i, ne0_fd);+ dst[i] = op(a[i] + bias[h]) * scale[h];+ }+}++template<typename F>+static void add_unary_mul_sycl(const float * a, const float * bias, const float * scale, float * dst,+ const int64_t k, const int64_t ne0, queue_ptr main_stream, F op) {+ const size_t num_blocks = ceil_div((size_t) k, (size_t) SYCL_GLU_BLOCK_SIZE);+ const sycl::nd_range<1> range(num_blocks * sycl::range<1>(SYCL_GLU_BLOCK_SIZE), sycl::range<1>(SYCL_GLU_BLOCK_SIZE));++ if (k == ne0) {+ main_stream->parallel_for(range, [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {+ add_unary_mul_flat_kernel(a, bias, scale, dst, k, item_ct1, op);+ });+ return;+ }++ // 32-bit fastdiv, exact only below 2^31; ggml_sycl_can_fuse() already declined past that+ GGML_ASSERT(k < ((int64_t) 1 << 31));+ const sycl::uint3 ne0_fd = init_fastdiv_values((uint32_t) ne0);+ main_stream->parallel_for(range, [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {+ add_unary_mul_bcast_kernel(a, bias, scale, dst, k, ne0_fd, item_ct1, op);+ });+}+namespace ggml_sycl_detail {static void acc_f32_sycl(const char *x, const char *y, float *dst,const int64_t n_elements,@@ -995,6 +1035,19 @@ static inline void ggml_sycl_op_swiglu(ggml_backend_sycl_context & ctx, ggml_ten});}+// Hands `launch` the functor for the unary op of a fused unary chain. Anything else+// ggml_sycl_can_fuse() has already declined, so the default is a dispatcher bug.+template<typename F>+static void dispatch_fused_unary_op(ggml_unary_op uop, F && launch) {+ switch (uop) {+ case GGML_UNARY_OP_SILU: launch([](float v) { return op_silu(v); }); break;+ case GGML_UNARY_OP_SIGMOID: launch([](float v) { return op_sigmoid(v); }); break;+ case GGML_UNARY_OP_SOFTPLUS: launch([](float v) { return op_softplus(v); }); break;+ default:+ GGML_ABORT("fused unary chain: unsupported unary op %s", ggml_unary_op_name(uop));+ }+}+// dst = op(unary_node->src[0]) * other, written straight to the MUL output, saving the// standalone unary launch. Preconditions come from ggml_sycl_can_fuse(); re-asserted here.void ggml_sycl_op_unary_mul_fused(ggml_backend_sycl_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node) {@@ -1032,13 +1085,41 @@ void ggml_sycl_op_unary_mul_fused(ggml_backend_sycl_context & ctx, ggml_tensor *}};- switch (ggml_get_unary_op(unary_node)) {- case GGML_UNARY_OP_SILU: dispatch_type([](float v) { return op_silu(v); }); break;- case GGML_UNARY_OP_SIGMOID: dispatch_type([](float v) { return op_sigmoid(v); }); break;- case GGML_UNARY_OP_SOFTPLUS: dispatch_type([](float v) { return op_softplus(v); }); break;- default:- GGML_ABORT("fused unary+mul: unsupported unary op %s", ggml_unary_op_name(ggml_get_unary_op(unary_node)));- }+ dispatch_fused_unary_op(ggml_get_unary_op(unary_node), dispatch_type);+}++// dst = op(a + bias) * scale for an ADD + UNARY + MUL chain whose bias and scale broadcast+// over dim 0. Preconditions come from ggml_sycl_can_fuse(); re-asserted here.+void ggml_sycl_op_add_unary_mul_fused(ggml_backend_sycl_context & ctx, ggml_tensor * add_node,+ ggml_tensor * unary_node, ggml_tensor * mul_node) {+ // the dst-arity convention the other fusions follow; a and bias live on add_node+ scope_op_debug_print scope_dbg_print(__func__, mul_node, /*num_src=*/2);++ const ggml_tensor * a = add_node->src[0];+ const ggml_tensor * bias = add_node->src[1];+ const ggml_tensor * scale = (mul_node->src[0] == unary_node) ? mul_node->src[1] : mul_node->src[0];++ // scale is picked by elimination; ggml_can_fuse()'s single-use rule rules out MUL(unary, unary)+ GGML_ASSERT(scale != unary_node);+ GGML_ASSERT(a->type == GGML_TYPE_F32 && bias->type == GGML_TYPE_F32);+ GGML_ASSERT(scale->type == GGML_TYPE_F32 && mul_node->type == GGML_TYPE_F32);+ GGML_ASSERT(ggml_are_same_shape(a, mul_node));+ // a and dst are indexed flat+ GGML_ASSERT(ggml_is_contiguous(a) && ggml_is_contiguous(mul_node));+ // bias and scale are one contiguous ne0-length row each, broadcast over the outer dims+ GGML_ASSERT(bias->ne[0] == a->ne[0] && scale->ne[0] == a->ne[0]);+ GGML_ASSERT(ggml_nrows(bias) == 1 && ggml_nrows(scale) == 1);+ GGML_ASSERT(ggml_is_contiguous(bias) && ggml_is_contiguous(scale));++ queue_ptr main_stream = ctx.stream();+ SYCL_CHECK(ggml_sycl_set_device(ctx.device));++ const auto dispatch_op = [&](auto op) {+ add_unary_mul_sycl((const float *) a->data, (const float *) bias->data, (const float *) scale->data,+ (float *) mul_node->data, ggml_nelements(mul_node), mul_node->ne[0], main_stream, op);+ };++ dispatch_fused_unary_op(ggml_get_unary_op(unary_node), dispatch_op);}__dpct_inline__ float ggml_sycl_op_swiglu_oai_single(float x, float g, float alpha = 1.702f, float limit = 7.0f) {diff --git a/ggml/src/ggml-sycl/element_wise.hpp b/ggml/src/ggml-sycl/element_wise.hppindex d280066efb..0a5bfa9568 100644--- a/ggml/src/ggml-sycl/element_wise.hpp+++ b/ggml/src/ggml-sycl/element_wise.hpp@@ -132,4 +132,9 @@ void ggml_sycl_arange(ggml_backend_sycl_context & ctx, ggml_tensor * dst);// fused UNARY(silu|sigmoid|softplus) + MUL; see ggml_sycl_can_fuse() for the accepted shapesvoid ggml_sycl_op_unary_mul_fused(ggml_backend_sycl_context & ctx, ggml_tensor * unary_node, ggml_tensor * mul_node);+// fused f32 ADD + UNARY(silu|sigmoid|softplus) + MUL with the bias and the scale broadcast+// over dim 0; see ggml_sycl_can_fuse() for the accepted shapes+void ggml_sycl_op_add_unary_mul_fused(ggml_backend_sycl_context & ctx, ggml_tensor * add_node,+ ggml_tensor * unary_node, ggml_tensor * mul_node);+#endif // GGML_SYCL_ELEMENTWISE_HPPdiff --git a/ggml/src/ggml-sycl/fusion.cpp b/ggml/src/ggml-sycl/fusion.cppindex d3e995233c..79fe13a1d0 100644--- a/ggml/src/ggml-sycl/fusion.cpp+++ b/ggml/src/ggml-sycl/fusion.cpp@@ -64,6 +64,12 @@ static bool ggml_sycl_should_fuse_mul_mat_glu(const ggml_tensor * gate, const ggreturn true;}+// the unary ops the fused unary chains in element_wise.cpp have a functor for+static bool ggml_sycl_fused_unary_has_kernel(ggml_unary_op unary_op) {+ return unary_op == GGML_UNARY_OP_SILU || unary_op == GGML_UNARY_OP_SIGMOID ||+ unary_op == GGML_UNARY_OP_SOFTPLUS;+}+bool ggml_sycl_can_fuse(const ggml_cgraph * cgraph, int node_idx, std::initializer_list<enum ggml_op> ops,std::initializer_list<enum ggml_unary_op> unary_ops) {#ifndef NDEBUG@@ -184,9 +190,7 @@ bool ggml_sycl_can_fuse(const ggml_cgraph * cgraph, int node_idx, std::initializreturn false;}- // the ops ggml_sycl_op_unary_mul_fused() has a kernel for- if (unary_op != GGML_UNARY_OP_SILU && unary_op != GGML_UNARY_OP_SIGMOID &&- unary_op != GGML_UNARY_OP_SOFTPLUS) {+ if (!ggml_sycl_fused_unary_has_kernel(unary_op)) {return false;}@@ -233,6 +237,55 @@ bool ggml_sycl_can_fuse(const ggml_cgraph * cgraph, int node_idx, std::initializreturn true;}+ // ADD(bias) + UNARY + MUL(scale): the delta-net alpha gate, softplus(alpha + dt) * a.+ // The broadcast is what stops the same-shape UNARY + MUL branch above firing past one token.+ if (ops.size() == 3 && ops.begin()[0] == GGML_OP_ADD && ops.begin()[1] == GGML_OP_UNARY &&+ ops.begin()[2] == GGML_OP_MUL && unary_ops.size() == 1) {+ const ggml_tensor * add = cgraph->nodes[node_idx];+ const ggml_tensor * unary = cgraph->nodes[node_idx + 1];+ const ggml_tensor * mul = cgraph->nodes[node_idx + 2];++ const ggml_unary_op unary_op = ggml_get_unary_op(unary);+ if (unary_op != unary_ops.begin()[0]) {+ return false;+ }++ if (!ggml_sycl_fused_unary_has_kernel(unary_op)) {+ return false;+ }++ // ggml_can_fuse() has already pinned the chain: unary consumes add, mul consumes+ // unary, add and unary have one use each, and all three have the same shape+ const ggml_tensor * a = add->src[0];+ const ggml_tensor * bias = add->src[1];+ const ggml_tensor * scale = (mul->src[0] == unary) ? mul->src[1] : mul->src[0];++ if (a->type != GGML_TYPE_F32 || bias->type != GGML_TYPE_F32 ||+ scale->type != GGML_TYPE_F32 || mul->type != GGML_TYPE_F32) {+ return false;+ }++ // the activation and the destination are indexed flat+ if (!ggml_is_contiguous(a) || !ggml_is_contiguous(mul) || !ggml_are_same_shape(a, mul)) {+ return false;+ }++ // the kernel reads the bias and the scale as v[col], so each must be a single+ // contiguous row spanning ne0+ if (bias->ne[0] != a->ne[0] || scale->ne[0] != a->ne[0] ||+ ggml_nrows(bias) != 1 || ggml_nrows(scale) != 1 ||+ !ggml_is_contiguous(bias) || !ggml_is_contiguous(scale)) {+ return false;+ }++ // the 32-bit fastdiv is inexact past 2^31; decline, the unfused path handles it+ if (ggml_nelements(mul) >= ((int64_t) 1 << 31)) {+ return false;+ }++ return true;+ }+if (ops.size() == 3 && ops.begin()[0] == GGML_OP_SSM_CONV && ops.begin()[1] == GGML_OP_ADD &&ops.begin()[2] == GGML_OP_UNARY && unary_ops.size() == 1 && unary_ops.begin()[0] == GGML_UNARY_OP_SILU) {const ggml_tensor * ssm_conv = cgraph->nodes[node_idx];diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cppindex 13fc023251..2a5184e012 100644--- a/ggml/src/ggml-sycl/ggml-sycl.cpp+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp@@ -6312,6 +6312,16 @@ static void ggml_backend_sycl_graph_compute_impl(ggml_backend_sycl_context * syci++;continue;}+ // ADD(bias) + UNARY + MUL(scale) with both broadcast over dim 0, the form the branch+ // above cannot take; ggml_get_unary_op() asserts, so check the op first.+ if (node->op == GGML_OP_ADD && i + 2 < cgraph->n_nodes &&+ cgraph->nodes[i + 1]->op == GGML_OP_UNARY &&+ ggml_sycl_can_fuse(cgraph, i, { GGML_OP_ADD, GGML_OP_UNARY, GGML_OP_MUL },+ { ggml_get_unary_op(cgraph->nodes[i + 1]) })) {+ ggml_sycl_op_add_unary_mul_fused(*sycl_ctx, node, cgraph->nodes[i + 1], cgraph->nodes[i + 2]);+ i += 2;+ continue;+ }// Batch consecutive independent same-shape F32 L2_NORM siblings (the GDN q/k// norms) into one launch; sources are strided views of the fused qkv buffer, sodiff --git a/tests/test-backend-ops.cpp b/tests/test-backend-ops.cppindex cfe5b73e7b..f31016bbb6 100644--- a/tests/test-backend-ops.cpp+++ b/tests/test-backend-ops.cpp@@ -4159,6 +4159,93 @@ struct test_unary_mul : public test_case {}};+// GGML_OP_ADD + GGML_OP_UNARY(SILU|SIGMOID|SOFTPLUS) + GGML_OP_MUL with the ADD's bias and+// the MUL's scale broadcast over dim 0: the delta-net alpha gate, softplus(alpha + dt) * a_coeff.+struct test_add_unary_mul : public test_case {+ const ggml_unary_op op;+ const ggml_type type;+ const std::array<int64_t, 4> ne;+ const bool swap; // unary result is the second MUL operand+ const std::string layout; // bias/scale layout, see build_graph()+ const std::string tail; // extra consumer past the MUL, see build_graph()++ std::string op_desc(ggml_tensor * t) override {+ GGML_UNUSED(t);+ return "ADD_" + std::string(ggml_unary_op_name(op)) + "_MUL";+ }+ bool run_whole_graph() override { return true; }++ double max_nmse_err() override {+ switch (type) {+ // f16 never fuses (the kernel is f32-only), so this bound is the unfused+ // chain's own f16 rounding drift, as in test_unary_mul+ case GGML_TYPE_F16: return 5e-5;+ // gelu never fuses either, and the backends' exp form drifts from the CPU's tanhf+ default: return op == GGML_UNARY_OP_GELU ? 5e-7 : 1e-7;+ }+ }++ std::string vars() override {+ return VARS_TO_STR5(type, ne, swap, layout, tail);+ }++ test_add_unary_mul(ggml_unary_op op,+ ggml_type type = GGML_TYPE_F32,+ std::array<int64_t, 4> ne = {32, 7, 1, 1},+ bool swap = false,+ std::string layout = "bcast",+ std::string tail = "")+ : op(op), type(type), ne(ne), swap(swap), layout(std::move(layout)), tail(std::move(tail)) {}++ ggml_tensor * build_graph(ggml_context * ctx) override {+ ggml_tensor * a = ggml_new_tensor(ctx, type, 4, ne.data());+ ggml_set_name(a, "a");++ std::array<int64_t, 4> ne_v = { ne[0], 1, 1, 1 };+ if (layout == "bcast") {+ // one ne0 row each, broadcast over the outer dims, which is the alpha-gate form+ } else if (layout == "same_shape") {+ // no broadcast at all; fuses only while the activation is a single row+ ne_v = ne;+ } else if (layout == "rep_ne0") {+ // repeat on dim 0, which bias[col] cannot address, so this must not fuse+ ne_v[0] = ne[0] / 4;+ } else {+ GGML_ABORT("unknown layout %s", layout.c_str());+ }++ ggml_tensor * bias = ggml_new_tensor(ctx, type, 4, ne_v.data());+ ggml_set_name(bias, "bias");++ ggml_tensor * scale = ggml_new_tensor(ctx, type, 4, ne_v.data());+ ggml_set_name(scale, "scale");++ ggml_tensor * s = ggml_add(ctx, a, bias);+ ggml_set_name(s, "add");++ ggml_tensor * u = ggml_unary(ctx, s, op);+ ggml_set_name(u, "unary");++ // a broadcasting operand can only be the second one, so swap needs same-shape operands+ ggml_tensor * out = swap ? ggml_mul(ctx, scale, u) : ggml_mul(ctx, u, scale);++ if (tail == "reuse") {+ // a second read of the add result must block the fusion+ ggml_set_name(out, "mul");+ out = ggml_add(ctx, out, s);+ } else if (tail == "consumer") {+ // fusion still applies; catches a dispatcher that skips one node too many+ ggml_set_name(out, "mul");+ out = ggml_add(ctx, out, scale);+ } else if (!tail.empty()) {+ GGML_ABORT("unknown tail %s", tail.c_str());+ }+ ggml_set_name(out, "out");++ return out;+ }+};+// SNAKE activation fusion: y = x + sin(a*x)^2 * inv_b// CUDA backend matches the naive 5-op chain (mul, sin, sqr, mul, add)// and dispatches a single fused kernel.@@ -9393,6 +9480,24 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {}}+ // fused add + unary + mul: the delta-net alpha gate, bias and scale broadcast over dim 0+ for (ggml_unary_op op : { GGML_UNARY_OP_SILU, GGML_UNARY_OP_SIGMOID, GGML_UNARY_OP_SOFTPLUS }) {+ test_cases.emplace_back(new test_add_unary_mul(op, GGML_TYPE_F32, { 32, 7, 1, 1 }));+ test_cases.emplace_back(new test_add_unary_mul(op, GGML_TYPE_F32, { 32, 512, 1, 1 }));+ test_cases.emplace_back(new test_add_unary_mul(op, GGML_TYPE_F32, { 5, 7, 11, 13 }));+ // one token: no broadcast left, and the unary result may be either MUL operand+ test_cases.emplace_back(new test_add_unary_mul(op, GGML_TYPE_F32, { 32, 1, 1, 1 }, false, "same_shape"));+ test_cases.emplace_back(new test_add_unary_mul(op, GGML_TYPE_F32, { 32, 1, 1, 1 }, true, "same_shape"));+ test_cases.emplace_back(new test_add_unary_mul(op, GGML_TYPE_F32, { 32, 7, 1, 1 }, false, "bcast", "consumer"));+ // must not fuse+ test_cases.emplace_back(new test_add_unary_mul(op, GGML_TYPE_F32, { 32, 7, 1, 1 }, false, "same_shape"));+ test_cases.emplace_back(new test_add_unary_mul(op, GGML_TYPE_F32, { 32, 7, 1, 1 }, false, "rep_ne0"));+ test_cases.emplace_back(new test_add_unary_mul(op, GGML_TYPE_F32, { 32, 7, 1, 1 }, false, "bcast", "reuse"));+ test_cases.emplace_back(new test_add_unary_mul(op, GGML_TYPE_F16, { 32, 7, 1, 1 }));+ }+ // a unary op with no fused kernel must fall back to the three-op chain+ test_cases.emplace_back(new test_add_unary_mul(GGML_UNARY_OP_GELU, GGML_TYPE_F32, { 32, 7, 1, 1 }));+// SNAKE activation fusion: x + sin(a*x)^2 * inv_bfor (ggml_type type : { GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_BF16 }) {test_cases.emplace_back(new test_snake_fuse(type, { 5, 7, 1, 1})); // primes sub-block