root/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2024-12-27 20:04:35 +00:00
Code Issues Actions 3 Packages Projects Releases Wiki Activity

Vulkan: Unroll more loops for more mul mat mat performance

Browse Source
This commit is contained in:
0cc4m 2024-12-03 19:52:36 +00:00
parent a0deeeed28
commit 9622fbe373
1 changed files with 6 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

12
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
View File

@ -196,7 +196,7 @@ void main() {
coopmat<float16_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b; coopmat<float16_t, gl_ScopeSubgroup, TK, TN, gl_MatrixUseB> cache_b;
coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col]; coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator> sums[cms_per_row * cms_per_col];
for (uint i = 0; i < cms_per_row * cms_per_col; i++) { [[unroll]] for (uint i = 0; i < cms_per_row * cms_per_col; i++) {
sums[i] = coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0f); sums[i] = coopmat<ACC_TYPE, gl_ScopeSubgroup, TM, TN, gl_MatrixUseAccumulator>(0.0f);
} }
#else #else
@ -209,7 +209,7 @@ void main() {
} }
#endif #endif
[[dont_unroll]] for (uint block = start_k; block < end_k; block += BK) { for (uint block = start_k; block < end_k; block += BK) {
[[unroll]] for (uint l = 0; l < BM; l += loadstride_a) { [[unroll]] for (uint l = 0; l < BM; l += loadstride_a) {
#if defined(DATA_A_F32) || defined(DATA_A_F16) #if defined(DATA_A_F32) || defined(DATA_A_F16)
@ -506,12 +506,12 @@ void main() {
pos_b += BK / LOAD_VEC_B; pos_b += BK / LOAD_VEC_B;
#ifdef COOPMAT #ifdef COOPMAT
for (uint i = 0; i < BK; i += TK) { [[unroll]] for (uint i = 0; i < BK; i += TK) {
for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) { [[unroll]] for (uint cm_row = 0; cm_row < cms_per_row; cm_row++) {
// Load from shared into cache // Load from shared into cache
coopMatLoad(cache_a, buf_a, (warp_r * WM + cm_row * TM) * SHMEM_STRIDE + i, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor); coopMatLoad(cache_a, buf_a, (warp_r * WM + cm_row * TM) * SHMEM_STRIDE + i, SHMEM_STRIDE, gl_CooperativeMatrixLayoutRowMajor);
for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) { [[unroll]] for (uint cm_col = 0; cm_col < cms_per_col; cm_col++) {
coopMatLoad(cache_b, buf_b, (warp_c * WN + cm_col * TN) * SHMEM_STRIDE + i, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor); coopMatLoad(cache_b, buf_b, (warp_c * WN + cm_col * TN) * SHMEM_STRIDE + i, SHMEM_STRIDE, gl_CooperativeMatrixLayoutColumnMajor);
sums[cm_col * cms_per_row + cm_row] = coopMatMulAdd(cache_a, cache_b, sums[cm_col * cms_per_row + cm_row]); sums[cm_col * cms_per_row + cm_row] = coopMatMulAdd(cache_a, cache_b, sums[cm_col * cms_per_row + cm_row]);
@ -519,7 +519,7 @@ void main() {
} }
} }
#else #else
for (uint i = 0; i < BK; i++) { [[unroll]] for (uint i = 0; i < BK; i++) {
// Load from shared into cache // Load from shared into cache
[[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) { [[unroll]] for (uint wsir = 0; wsir < WMITER; wsir++) {
[[unroll]] for (uint j = 0; j < TM; j++) { [[unroll]] for (uint j = 0; j < TM; j++) {