ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-06-22 16:38:58 +00:00
Code Issues Actions Packages Projects Releases Wiki Activity

SYCL: Remove misleading ggml_sycl_op_flatten function (llama/12387)

Browse Source 
* SYCL: Remove misleading ggml_sycl_op_flatten function

* remove trailing whitespace

* Fix L2 norm from rebase

* remove try catch block from element_wise.cpp

* remove comment from common.hp

* ggml-sycl.cpp: Add try catch sycl::exception block in compute_forward

* norm.cpp: remove try catch exception block
This commit is contained in:
Akarshan Biswas
2025-03-31 14:55:24 +05:30
committed by Georgi Gerganov
parent 93631b2be6
commit 2e2f0f954b
12 changed files with 369 additions and 586 deletions
Download Patch File Download Diff File Expand all files Collapse all files

212
ggml/src/ggml-sycl/ggml-sycl.cpp
View File

@ -1988,16 +1988,8 @@ catch (sycl::exception const &exc) {
std::exit(1);
}
static void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_d, const float *src1_d,
float *dst_d,
const queue_ptr &main_stream) {
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_repeat>>(ctx, dst, src0, dst, nullptr, src0_d, dst_d, main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(src1_d);
static void ggml_sycl_op_repeat(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
ggml_sycl_op_bin_bcast<bin_bcast_sycl<op_repeat>>(ctx, dst, dst->src[0], dst);
}
@ -2132,13 +2124,14 @@ catch (sycl::exception const &exc) {
std::exit(1);
}
static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd, const queue_ptr &main_stream) {
static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
const int32_t * opts = (const int32_t *)dst->op_params;
enum ggml_op_pool op = static_cast<ggml_op_pool>(opts[0]);
@ -2149,8 +2142,8 @@ static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tens
const int p0 = opts[5];
const int p1 = opts[6];
const int64_t IH = src0->ne[1];
const int64_t IW = src0->ne[0];
const int64_t IH = dst->src[0]->ne[1];
const int64_t IW = dst->src[0]->ne[0];
const int64_t N = dst->ne[3];
const int64_t OC = dst->ne[2];
@ -2169,163 +2162,125 @@ static void ggml_sycl_op_pool2d(ggml_backend_sycl_context & ctx, const ggml_tens
parallel_elements, src0_dd, dst_dd, op,
item_ct1);
});
GGML_UNUSED(src1);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
inline void ggml_sycl_op_sum(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
const int64_t ne = ggml_nelements(src0);
const int64_t ne = ggml_nelements(dst->src[0]);
sum_rows_f32_sycl(src0_dd, dst_dd, ne, 1, main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_sum_rows(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
const int64_t ncols = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
const int64_t ncols = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);
sum_rows_f32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_argsort(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT(dst->type == GGML_TYPE_I32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
int32_t * dst_dd = static_cast<int32_t *>(dst->data);
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_I32);
const int64_t ncols = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
const int64_t ncols = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);
enum ggml_sort_order order = (enum ggml_sort_order) dst->op_params[0];
argsort_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, order, main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
argsort_f32_i32_sycl(src0_dd, (int *) dst_dd, ncols, nrows, order, main_stream);
}
inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1, ggml_tensor *dst,
const float *src0_dd, const float *src1_dd,
float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_argmax(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_I32);
const int64_t ncols = src0->ne[0];
const int64_t nrows = ggml_nrows(src0);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
int32_t * dst_dd = static_cast<int32_t *>(dst->data);
argmax_f32_i32_sycl(src0_dd, (int *)dst_dd, ncols, nrows, main_stream);
const int64_t ncols = dst->src[0]->ne[0];
const int64_t nrows = ggml_nrows(dst->src[0]);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
argmax_f32_i32_sycl(src0_dd, dst_dd, ncols, nrows, main_stream);
}
inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx, const ggml_tensor *src0,
const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_diag_mask_inf(ggml_backend_sycl_context & ctx,ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
const int64_t ne00 = src0->ne[0];
const int64_t ne01 = src0->ne[1];
const int nrows0 = ggml_nrows(src0);
const int64_t ne00 = dst->src[0]->ne[0];
const int64_t ne01 = dst->src[0]->ne[1];
const int nrows0 = ggml_nrows(dst->src[0]);
const int n_past = ((int32_t *) dst->op_params)[0];
diag_mask_inf_f32_sycl(src0_dd, dst_dd, ne00, nrows0, ne01, n_past, main_stream);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_scale(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
dpct::queue_ptr main_stream = ctx.stream();
SYCL_CHECK(ggml_sycl_set_device(ctx.device));
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
float scale;
memcpy(&scale, dst->op_params, sizeof(float));
scale_f32_sycl(src0_dd, dst_dd, scale, ggml_nelements(src0), main_stream);
scale_f32_sycl(src0_dd, dst_dd, scale, ggml_nelements(dst->src[0]), main_stream);
/*
DPCT1010:87: SYCL uses exceptions to report errors and does not use the
error codes. The call was replaced with 0. You need to rewrite this code.
*/
SYCL_CHECK(0);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, const ggml_tensor *src0, const ggml_tensor *src1,
ggml_tensor *dst, const float *src0_dd,
const float *src1_dd, float *dst_dd,
const queue_ptr &main_stream) {
inline void ggml_sycl_op_clamp(ggml_backend_sycl_context & ctx, ggml_tensor *dst) {
GGML_ASSERT(src0->type == GGML_TYPE_F32);
GGML_ASSERT(dst->src[0]->type == GGML_TYPE_F32);
GGML_ASSERT( dst->type == GGML_TYPE_F32);
const float * src0_dd = static_cast<const float *>(dst->src[0]->data);
float * dst_dd = static_cast<float *>(dst->data);
float min;
float max;
memcpy(&min, dst->op_params, sizeof(float));
memcpy(&max, (float *) dst->op_params + 1, sizeof(float));
clamp_f32_sycl(src0_dd, dst_dd, min, max, ggml_nelements(src0), main_stream);
clamp_f32_sycl(src0_dd, dst_dd, min, max, ggml_nelements(dst->src[0]), ctx.stream());
/*
DPCT1010:88: SYCL uses exceptions to report errors and does not use the
error codes. The call was replaced with 0. You need to rewrite this code.
*/
SYCL_CHECK(0);
GGML_UNUSED(src1);
GGML_UNUSED(dst);
GGML_UNUSED(src1_dd);
GGML_UNUSED(ctx);
}
static void ggml_sycl_set_peer_access(const int n_tokens, int main_device) {
@ -2695,37 +2650,37 @@ catch (sycl::exception const &exc) {
static void ggml_sycl_repeat(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_repeat);
ggml_sycl_op_repeat(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
static void ggml_sycl_get_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_get_rows);
ggml_sycl_op_get_rows(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
static void ggml_sycl_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_norm);
ggml_sycl_op_norm(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
static void ggml_sycl_rms_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_rms_norm);
ggml_sycl_op_rms_norm(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
static void ggml_sycl_l2_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_l2_norm);
ggml_sycl_op_l2_norm(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
static void ggml_sycl_group_norm(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_SYCL_DEBUG("call %s\n", __func__);
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_group_norm);
ggml_sycl_op_group_norm(ctx, dst);
GGML_SYCL_DEBUG("call %s done\n", __func__);
}
@ -3269,48 +3224,48 @@ catch (sycl::exception const &exc) {
}
static void ggml_sycl_scale(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_scale);
ggml_sycl_op_scale(ctx, dst);
}
static void ggml_sycl_clamp(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_clamp);
ggml_sycl_op_clamp(ctx, dst);
}
static void ggml_sycl_diag_mask_inf(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_diag_mask_inf);
ggml_sycl_op_diag_mask_inf(ctx, dst);
}
static void ggml_sycl_rope(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(dst->src[0])); // TODO: this restriction is temporary until non-cont support is implemented
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_rope);
ggml_sycl_op_rope(ctx, dst);
}
static void ggml_sycl_pool2d(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_pool2d);
ggml_sycl_op_pool2d(ctx, dst);
}
static void ggml_sycl_im2col(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_im2col);
ggml_sycl_op_im2col(ctx, dst);
}
static void ggml_sycl_sum(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(dst->src[0]));
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sum);
ggml_sycl_op_sum(ctx, dst);
}
static void ggml_sycl_sum_rows(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(dst->src[0]));
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_sum_rows);
ggml_sycl_op_sum_rows(ctx, dst);
}
static void ggml_sycl_argsort(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(dst->src[0]));
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_argsort);
ggml_sycl_op_argsort(ctx, dst);
}
static void ggml_sycl_argmax(ggml_backend_sycl_context & ctx, ggml_tensor * dst) {
GGML_ASSERT(ggml_is_contiguous(dst->src[0]));
ggml_sycl_op_flatten(ctx, dst->src[0], dst->src[1], dst, ggml_sycl_op_argmax);
ggml_sycl_op_argmax(ctx, dst);
}
@ -3335,7 +3290,7 @@ catch (sycl::exception const &exc) {
std::exit(1);
}
static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tensor * dst) {
static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct ggml_tensor * dst) try {
if (!g_sycl_loaded) return false;
if (dst->src[0] != nullptr && ggml_backend_buffer_is_sycl_split(dst->src[0]->buffer)) {
@ -3528,6 +3483,9 @@ static bool ggml_sycl_compute_forward(ggml_backend_sycl_context & ctx, struct gg
}
return true;
} catch (sycl::exception & e) {
std::cerr << e.what() << "Exception caught at file:" << __FILE__ << ", line:" << __LINE__ << std::endl;
std::exit(1);
}
GGML_API void ggml_backend_sycl_get_device_description(int device, char *description,