LLM inference in C/C++
C++ ggml
sycl : implement xielu op (#25550)
Andrew Smith22b208b1cacb
diff --git a/docs/ops.md b/docs/ops.mdindex f138753854..c560152369 100644--- a/docs/ops.md+++ b/docs/ops.md@@ -120,4 +120,4 @@ Legend:| TRI | ❌ | ❌ | ✅ | ✅ | ✅ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ || TRUNC | ❌ | ❌ | ✅ | 🟡 | 🟡 | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ || UPSCALE | ❌ | 🟡 | ✅ | ✅ | ❌ | ✅ | 🟡 | ✅ | ✅ | ✅ | ❌ | ❌ |-| XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ✅ | ❌ | ❌ |+| XIELU | ❌ | ❌ | ✅ | ❌ | ❌ | ✅ | ❌ | ✅ | ✅ | ✅ | ❌ | ❌ |diff --git a/docs/ops/SYCL.csv b/docs/ops/SYCL.csvindex 8c94d14b58..b563e76a87 100644--- a/docs/ops/SYCL.csv+++ b/docs/ops/SYCL.csv@@ -11600,10 +11600,10 @@ zjy 2"SYCL0","CUMSUM","type=f32,ne=[242004,1,1,1]","support","1","yes","SYCL""SYCL0","CUMSUM","type=f32,ne=[375960,1,1,1]","support","1","yes","SYCL""SYCL0","CUMSUM","type=f32,ne=[20481,4,1,1]","support","1","yes","SYCL"-"SYCL0","XIELU","type=f32,ne=[10,5,4,3]","support","0","no","SYCL"-"SYCL0","XIELU","type=f16,ne=[10,5,4,3]","support","0","no","SYCL"-"SYCL0","XIELU","type=f32,ne=[512,16,1,1]","support","0","no","SYCL"-"SYCL0","XIELU","type=f16,ne=[512,16,1,1]","support","0","no","SYCL"+"SYCL0","XIELU","type=f32,ne=[10,5,4,3]","support","1","yes","SYCL"+"SYCL0","XIELU","type=f16,ne=[10,5,4,3]","support","1","yes","SYCL"+"SYCL0","XIELU","type=f32,ne=[512,16,1,1]","support","1","yes","SYCL"+"SYCL0","XIELU","type=f16,ne=[512,16,1,1]","support","1","yes","SYCL""SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=3","support","1","yes","SYCL""SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=2","support","1","yes","SYCL""SYCL0","TRI","type=f32,ne=[10,10,4,3],tri_type=1","support","1","yes","SYCL"diff --git a/ggml/src/ggml-sycl/element_wise.cpp b/ggml/src/ggml-sycl/element_wise.cppindex bae157a487..b2406e11b5 100644--- a/ggml/src/ggml-sycl/element_wise.cpp+++ b/ggml/src/ggml-sycl/element_wise.cpp@@ -247,6 +247,17 @@ static __dpct_inline__ T op_leaky_relu(T x, float negative_slope) {}}+template<typename T>+static __dpct_inline__ T op_xielu(T x, float alpha_n, float alpha_p, float beta, float eps) {+ const float xi = static_cast<float>(x);+ const float gate_pos = (xi > 0.0f);+ const float y_pos = alpha_p * xi * xi + beta * xi;+ const float min_v_eps = sycl::fmin(xi, eps);+ const float y_neg = (sycl::expm1(min_v_eps) - xi) * alpha_n + beta * xi;+ const float out = gate_pos * y_pos + (1.0f - gate_pos) * y_neg;+ return static_cast<T>(out);+}+template<typename T>static __dpct_inline__ T op_sqr(T x) {return x * x;@@ -359,6 +370,13 @@ static void unary_op_leaky_relu_kernel(const T * x, T * dst, const int k, float}}+template<typename T>+static void unary_op_xielu_kernel(const T * x, T * dst, const int k, float alpha_n, float alpha_p, float beta, float eps, const sycl::nd_item<1> &item_ct1) {+ SYCL_GLOBAL_ID_LOOP(k, item_ct1) {+ dst[i] = op_xielu(x[i], alpha_n, alpha_p, beta, eps);+ }+}+template<typename T>static void unary_op_sqr_kernel(const T * x, T * dst, const int k, const sycl::nd_item<1> &item_ct1) {SYCL_GLOBAL_ID_LOOP(k, item_ct1) {@@ -836,6 +854,23 @@ static inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tens}, min_val, max_val);}+static inline void ggml_sycl_op_xielu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {+ const float alpha_n = ggml_get_op_params_f32(dst, 1);+ const float alpha_p = ggml_get_op_params_f32(dst, 2);+ const float beta = ggml_get_op_params_f32(dst, 3);+ const float eps = ggml_get_op_params_f32(dst, 4);+ ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,+ [](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream, float alpha_n_arg, float alpha_p_arg, float beta_arg, float eps_arg) {+ const int num_blocks = ceil_div(k_elements, SYCL_RELU_BLOCK_SIZE);+ stream->parallel_for(+ sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_RELU_BLOCK_SIZE),+ sycl::range<1>(SYCL_RELU_BLOCK_SIZE)),+ [=](sycl::nd_item<1> item_ct1) [[sycl::reqd_sub_group_size(WARP_SIZE)]] {+ unary_op_xielu_kernel(src, dst_ptr, k_elements, alpha_n_arg, alpha_p_arg, beta_arg, eps_arg, item_ct1);+ });+ }, alpha_n, alpha_p, beta, eps);+}+static inline void ggml_sycl_op_floor(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {ggml_sycl_detail::ggml_sycl_op_unary(ctx, dst, [](auto x) {return op_floor(x);@@ -1153,6 +1188,11 @@ void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {ggml_sycl_op_clamp(ctx, dst);}+void ggml_sycl_xielu(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {+ scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);+ ggml_sycl_op_xielu(ctx, dst);+}+void ggml_sycl_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {scope_op_debug_print scope_dbg_print(__func__, dst, /*num_src=*/1);ggml_sycl_op_sgn(ctx, dst);diff --git a/ggml/src/ggml-sycl/element_wise.hpp b/ggml/src/ggml-sycl/element_wise.hppindex 3bdc385968..beea052cf0 100644--- a/ggml/src/ggml-sycl/element_wise.hpp+++ b/ggml/src/ggml-sycl/element_wise.hpp@@ -75,6 +75,8 @@ void ggml_sycl_sqr(ggml_backend_sycl_context & ctx, ggml_tensor * dst);void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst);+void ggml_sycl_xielu(ggml_backend_sycl_context & ctx, ggml_tensor * dst);+void ggml_sycl_sgn(ggml_backend_sycl_context & ctx, ggml_tensor * dst);void ggml_sycl_abs(ggml_backend_sycl_context & ctx, ggml_tensor * dst);diff --git a/ggml/src/ggml-sycl/ggml-sycl.cpp b/ggml/src/ggml-sycl/ggml-sycl.cppindex f534d237d7..cb8974eedb 100644--- a/ggml/src/ggml-sycl/ggml-sycl.cpp+++ b/ggml/src/ggml-sycl/ggml-sycl.cpp@@ -5011,6 +5011,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggcase GGML_UNARY_OP_ELU:ggml_sycl_elu(ctx, dst);break;+ case GGML_UNARY_OP_XIELU:+ ggml_sycl_xielu(ctx, dst);+ break;case GGML_UNARY_OP_FLOOR:ggml_sycl_floor(ctx, dst);break;@@ -5673,6 +5676,7 @@ static bool do_ggml_backend_sycl_device_supports_op(ggml_backend_dev_t dev, conscase GGML_UNARY_OP_EXPM1:case GGML_UNARY_OP_SOFTPLUS:case GGML_UNARY_OP_ELU:+ case GGML_UNARY_OP_XIELU:case GGML_UNARY_OP_CEIL:return true;case GGML_UNARY_OP_FLOOR: