EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

Revert "sycl: add usage of enqueue_functions extension (#14244)" (#15910)

Browse Source 
* Revert "sycl: add usage of enqueue_functions extension (#14244)"

This reverts commit 8308f98c7f.

* fix missed revert code, format the code
This commit is contained in:
Neo Zhang Jianyu
2025-09-12 09:15:12 +08:00
committed by GitHub
parent 360d6533db
commit 704d90c987
20 changed files with 845 additions and 674 deletions
Download Patch File Download Diff File Expand all files Collapse all files

62
ggml/src/ggml-sycl/element_wise.cpp
View File

@@ -407,7 +407,7 @@ static void acc_f32_sycl(const float *x, const float *y, float *dst,
const int ne12, const int nb1, const int nb2,
const int offset, queue_ptr stream) {
int num_blocks = ceil_div(n_elements, SYCL_ACC_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) *
sycl::range<1>(SYCL_ACC_BLOCK_SIZE),
sycl::range<1>(SYCL_ACC_BLOCK_SIZE)),
@@ -425,8 +425,8 @@ static void upscale_sycl(const T *x, T *dst, const int nb00, const int nb01,
int dst_size = ne10 * ne11 * ne12 * ne13;
int num_blocks = ceil_div(dst_size, SYCL_UPSCALE_BLOCK_SIZE);
sycl::range<1> gridDim(num_blocks * SYCL_UPSCALE_BLOCK_SIZE);
sycl_parallel_for<1>(
stream, sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
stream->parallel_for(
sycl::nd_range<1>(gridDim, sycl::range<1>(SYCL_UPSCALE_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
upscale(x, dst, nb00, nb01, nb02, nb03, ne10, ne11, ne12, ne13, sf0, sf1, sf2, sf3, item_ct1);
});
}
@@ -437,7 +437,7 @@ static void pad_sycl(const T *x, T *dst, const int ne00,
const int ne1, const int ne2, queue_ptr stream) {
int num_blocks = ceil_div(ne0, SYCL_PAD_BLOCK_SIZE);
sycl::range<3> gridDim(ne2, ne1, num_blocks);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<3>(gridDim * sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE),
sycl::range<3>(1, 1, SYCL_PAD_BLOCK_SIZE)),
[=](sycl::nd_item<3> item_ct1) { pad(x, dst, ne0, ne00, ne01, ne02, item_ct1); });
@@ -639,7 +639,7 @@ static inline void ggml_sycl_op_sgn(ggml_backend_sycl_context & ctx, ggml_tensor
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, 256);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
sycl::range<1>(256)),
[=](sycl::nd_item<1> item_ct1) {
@@ -652,7 +652,7 @@ static inline void ggml_sycl_op_abs(ggml_backend_sycl_context & ctx, ggml_tensor
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, 256);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
sycl::range<1>(256)),
[=](sycl::nd_item<1> item_ct1) {
@@ -665,7 +665,7 @@ static inline void ggml_sycl_op_elu(ggml_backend_sycl_context & ctx, ggml_tensor
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, 256);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(256),
sycl::range<1>(256)),
[=](sycl::nd_item<1> item_ct1) {
@@ -678,7 +678,7 @@ static inline void ggml_sycl_op_silu(ggml_backend_sycl_context & ctx, ggml_tenso
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_SILU_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SILU_BLOCK_SIZE),
sycl::range<1>(SYCL_SILU_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -691,7 +691,7 @@ static inline void ggml_sycl_op_gelu(ggml_backend_sycl_context & ctx, ggml_tenso
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_GELU_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_GELU_BLOCK_SIZE),
sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -704,7 +704,7 @@ static inline void ggml_sycl_op_gelu_quick(ggml_backend_sycl_context & ctx, ggml
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_GELU_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_GELU_BLOCK_SIZE),
sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -717,7 +717,7 @@ static inline void ggml_sycl_op_gelu_erf(ggml_backend_sycl_context & ctx, ggml_t
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_GELU_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_GELU_BLOCK_SIZE),
sycl::range<1>(SYCL_GELU_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -730,7 +730,7 @@ static inline void ggml_sycl_op_tanh(ggml_backend_sycl_context & ctx, ggml_tenso
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_TANH_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_TANH_BLOCK_SIZE),
sycl::range<1>(SYCL_TANH_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -743,7 +743,7 @@ static inline void ggml_sycl_op_relu(ggml_backend_sycl_context & ctx, ggml_tenso
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_RELU_BLOCK_SIZE);
sycl_parallel_for(stream,
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) {
@@ -756,7 +756,7 @@ static inline void ggml_sycl_op_hardsigmoid(ggml_backend_sycl_context & ctx, ggm
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_HARDSIGMOID_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_HARDSIGMOID_BLOCK_SIZE),
sycl::range<1>(SYCL_HARDSIGMOID_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -769,7 +769,7 @@ static inline void ggml_sycl_op_hardswish(ggml_backend_sycl_context & ctx, ggml_
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_HARDSWISH_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_HARDSWISH_BLOCK_SIZE),
sycl::range<1>(SYCL_HARDSWISH_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -782,7 +782,7 @@ static inline void ggml_sycl_op_exp(ggml_backend_sycl_context & ctx, ggml_tensor
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_EXP_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_EXP_BLOCK_SIZE),
sycl::range<1>(SYCL_EXP_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -795,7 +795,7 @@ static inline void ggml_sycl_op_log(ggml_backend_sycl_context & ctx, ggml_tensor
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_EXP_BLOCK_SIZE); // Using EXP block size
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_EXP_BLOCK_SIZE),
sycl::range<1>(SYCL_EXP_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -808,7 +808,7 @@ static inline void ggml_sycl_op_neg(ggml_backend_sycl_context & ctx, ggml_tensor
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_NEG_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_NEG_BLOCK_SIZE),
sycl::range<1>(SYCL_NEG_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -821,7 +821,7 @@ static inline void ggml_sycl_op_step(ggml_backend_sycl_context & ctx, ggml_tenso
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_NEG_BLOCK_SIZE); // Using NEG block size
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_NEG_BLOCK_SIZE),
sycl::range<1>(SYCL_NEG_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -834,7 +834,7 @@ static inline void ggml_sycl_op_sigmoid(ggml_backend_sycl_context & ctx, ggml_te
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_SIGMOID_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIGMOID_BLOCK_SIZE),
sycl::range<1>(SYCL_SIGMOID_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -847,7 +847,7 @@ static inline void ggml_sycl_op_sqrt(ggml_backend_sycl_context & ctx, ggml_tenso
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_SQRT_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SQRT_BLOCK_SIZE),
sycl::range<1>(SYCL_SQRT_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -860,7 +860,7 @@ static inline void ggml_sycl_op_sin(ggml_backend_sycl_context & ctx, ggml_tensor
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_SIN_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIN_BLOCK_SIZE),
sycl::range<1>(SYCL_SIN_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -873,7 +873,7 @@ static inline void ggml_sycl_op_cos(ggml_backend_sycl_context & ctx, ggml_tensor
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_SIN_BLOCK_SIZE); // Using SIN block size
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SIN_BLOCK_SIZE),
sycl::range<1>(SYCL_SIN_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -888,7 +888,7 @@ static inline void ggml_sycl_op_leaky_relu(ggml_backend_sycl_context & ctx, ggml
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream, float slope) {
const int num_blocks = ceil_div(k_elements, SYCL_RELU_BLOCK_SIZE);
sycl_parallel_for(stream,
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) {
@@ -901,7 +901,7 @@ static inline void ggml_sycl_op_sqr(ggml_backend_sycl_context & ctx, ggml_tensor
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream) {
const int num_blocks = ceil_div(k_elements, SYCL_SQR_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_SQR_BLOCK_SIZE),
sycl::range<1>(SYCL_SQR_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -935,7 +935,7 @@ static inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tens
ggml_sycl_detail::dispatch_ggml_sycl_op_unary(ctx, dst,
[](const auto* src, auto* dst_ptr, int k_elements, queue_ptr stream, float min_arg, float max_arg) {
const int num_blocks = ceil_div(k_elements, SYCL_CLAMP_BLOCK_SIZE);
sycl_parallel_for(stream,
stream->parallel_for(
sycl::nd_range<1>(sycl::range<1>(num_blocks) * sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE),
sycl::range<1>(SYCL_CLAMP_BLOCK_SIZE)),
[=](sycl::nd_item<1> item_ct1) {
@@ -967,7 +967,7 @@ static inline void ggml_sycl_op_geglu(ggml_backend_sycl_context & ctx, ggml_tens
ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
[](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
sycl_parallel_for(main_stream,
main_stream->parallel_for(
sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
gated_op_fused_geglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
});
@@ -978,7 +978,7 @@ static inline void ggml_sycl_op_reglu(ggml_backend_sycl_context & ctx, ggml_tens
ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
[](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
const uint32_t num_blocks = ceil_div((uint32_t)k, SYCL_RELU_BLOCK_SIZE); // Using RELU block size for reglu
sycl_parallel_for(main_stream,
main_stream->parallel_for(
sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), sycl::range<1>(SYCL_RELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
gated_op_fused_reglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
});
@@ -989,7 +989,7 @@ static inline void ggml_sycl_op_swiglu(ggml_backend_sycl_context & ctx, ggml_ten
ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
[](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
const uint32_t num_blocks = ceil_div((uint32_t)k, SYCL_SILU_BLOCK_SIZE); // Using SILU block size for swiglu
sycl_parallel_for(main_stream,
main_stream->parallel_for(
sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), sycl::range<1>(SYCL_SILU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
gated_op_fused_swiglu(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
});
@@ -1000,7 +1000,7 @@ static inline void ggml_sycl_op_geglu_erf(ggml_backend_sycl_context & ctx, ggml_
ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
[](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
sycl_parallel_for(main_stream,
main_stream->parallel_for(
sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
gated_op_fused_geglu_erf(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
});
@@ -1011,7 +1011,7 @@ static inline void ggml_sycl_op_geglu_quick(ggml_backend_sycl_context & ctx, ggm
ggml_sycl_detail::dispatch_ggml_sycl_op_fused_glu(ctx, dst,
[](const auto* x_ptr, const auto* g_ptr, auto* dst_ptr, uint64_t k, uint64_t n, uint64_t o0, uint64_t o1, queue_ptr main_stream) {
const uint32_t num_blocks = ceil_div(k, SYCL_GELU_BLOCK_SIZE);
sycl_parallel_for(main_stream,
main_stream->parallel_for(
sycl::nd_range<1>((num_blocks * sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), sycl::range<1>(SYCL_GELU_BLOCK_SIZE)), [=](sycl::nd_item<1> item_ct1) {
gated_op_fused_geglu_quick(x_ptr, g_ptr, dst_ptr, k, n, o0, o1, item_ct1);
});