tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-08-07 09:22:39 -04:00
Code Issues Packages Projects Releases Wiki Activity

[SYCL] Fix WARP_SIZE=16 bug of Intel GPU (#8266)

Browse Source 
* fix group_norm ut

* split softmax

* fix softmax

* add concat support condition

* revert debug code

* move QK_WARP_SIZE to presets.hpp
This commit is contained in:
luoyu-intel
2024-07-05 05:06:13 +00:00
committed by GitHub
parent e235b267a2
commit a9554e20b6
8 changed files with 301 additions and 257 deletions
Download Patch File Download Diff File Expand all files Collapse all files

8
ggml/src/ggml-sycl/norm.cpp
View File

@@ -57,6 +57,7 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con
const int nwarps = nthreads / WARP_SIZE;
assert(nwarps % WARP_SIZE == 0);
start += item_ct1.get_local_id(2);
int nreduce = nwarps / WARP_SIZE;
if (end >= ne_elements) {
end = ne_elements;
@@ -87,7 +88,6 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con
*/
item_ct1.barrier();
tmp = 0.f;
int nreduce = nwarps / WARP_SIZE;
for (size_t i = 0; i < nreduce; i += 1)
{
tmp += s_sum[lane_id + i * WARP_SIZE];
@@ -122,7 +122,11 @@ static void group_norm_f32(const float* x, float* dst, const int group_size, con
better performance if there is no access to global memory.
*/
item_ct1.barrier();
tmp = s_sum[lane_id];
tmp = 0.f;
for (size_t i = 0; i < nreduce; i += 1)
{
tmp += s_sum[lane_id + i * WARP_SIZE];
}
tmp = warp_reduce_sum(tmp, item_ct1);
}