CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-08 10:07:01 +00:00
Code Issues Packages Projects Releases Wiki Activity

sycl: add usage of enqueue_functions extension (#14244)

Browse Source 
* Add header and namespace to use enqueue_functions extension

* Convert submit and parallel_for to use new extension in convert.cpp

* Convert submit and parallel_for to use extension in ggml-sycl.cpp

* Convert submit and parallel_for to use extension in gla.cpp

* Convert submit and parallel_for in mmq.cpp

* Convert submit and parallel_for in mmvq.cpp

* Convert submit and parallel_for in remaining files

* Convert all simple parallel_for to nd_launch from enqueue_functions
extension

* Wrapping extension in general function

Create a general function that enable the enqueue_functions extension if
it is enable in the compiler, otherwise call the general SYCL function
to launch kernels.

---------

Signed-off-by: nscipione <nicolo.scipione@codeplay.com>
This commit is contained in:
Nicolò Scipione
2025-06-20 15:07:21 +02:00
committed by GitHub
parent 6369be0735
commit 8308f98c7f
19 changed files with 750 additions and 986 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
ggml/src/ggml-sycl/wkv.cpp
View File

@@ -207,12 +207,11 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
// Submit kernel
if (C / H == WKV_BLOCK_SIZE) {
stream->submit([&](sycl::handler& cgh) {
sycl_launch(stream, [&](sycl::handler & cgh) {
sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
cgh.parallel_for(
sycl::nd_range<3>(grid_dims * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
sycl_parallel_for(
cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
rwkv_wkv6_f32_kernel<WKV_BLOCK_SIZE>(
B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d,
item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()
@@ -220,12 +219,11 @@ void ggml_sycl_op_rwkv_wkv6(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
});
});
} else {
stream->submit([&](sycl::handler& cgh) {
sycl_launch(stream, [&](sycl::handler & cgh) {
sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
cgh.parallel_for(
sycl::nd_range<3>(grid_dims * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
sycl_parallel_for(
cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
rwkv_wkv6_f32_kernel<WKV_BLOCK_SIZE * 2>(
B, T, C, H, k_d, v_d, r_d, tf_d, td_d, s_d, dst_d,
item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()
@@ -264,12 +262,11 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
// Submit kernel
if (C / H == WKV_BLOCK_SIZE) {
stream->submit([&](sycl::handler& cgh) {
sycl_launch(stream, [&](sycl::handler & cgh) {
sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
cgh.parallel_for(
sycl::nd_range<3>(grid_dims * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
sycl_parallel_for(
cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
rwkv_wkv7_f32_kernel<WKV_BLOCK_SIZE>(
B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d,
item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()
@@ -277,12 +274,11 @@ void ggml_sycl_op_rwkv_wkv7(ggml_backend_sycl_context& ctx, ggml_tensor* dst) {
});
});
} else {
stream->submit([&](sycl::handler& cgh) {
sycl_launch(stream, [&](sycl::handler & cgh) {
sycl::local_accessor<float, 1> shared_mem_acc(shared_mem_size, cgh);
cgh.parallel_for(
sycl::nd_range<3>(grid_dims * block_dims, block_dims),
[=](sycl::nd_item<3> item_ct1) {
sycl_parallel_for(
cgh, sycl::nd_range<3>(grid_dims * block_dims, block_dims), [=](sycl::nd_item<3> item_ct1) {
rwkv_wkv7_f32_kernel<WKV_BLOCK_SIZE * 2>(
B, T, C, H, r_d, w_d, k_d, v_d, a_d, b_d, s_d, dst_d,
item_ct1, (float*)shared_mem_acc.get_multi_ptr<sycl::access::decorated::no>().get()