CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-20 12:07:33 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: vec dot matrix multiplication fix (#16151)

Browse Source 
* vulkan: fix matrix multiplication index calculation for odd m/n and odd k in combination with batching

* add odd m/n + odd k test with batching
This commit is contained in:
Ruben Ortlam
2025-09-22 07:22:43 +02:00
committed by GitHub
parent 51f5a45fbe
commit 9073a73d82
4 changed files with 31 additions and 18 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm_funcs.comp
View File

@@ -14,8 +14,8 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
FLOAT_TYPE_VEC4 aa = FLOAT_TYPE_VEC4(data_a[idx]);
buf_a[buf_idx ] = aa.xy;
buf_a[buf_idx + 1] = aa.zw;
#else // LOAD_VEC_A == 2
const uint idx = pos_a * 2 + col * p.stride_a + row * 2;
#else // LOAD_VEC_BATCH_A == 2
const uint idx = pos_a + col * p.stride_a + row * 2;
const uint buf_idx = col * SHMEM_STRIDE + row;
if (idx_m < p.M && block + row * 2 + 1 < end_k) {
buf_a[buf_idx] = FLOAT_TYPE_VEC2(data_a[idx],
@@ -33,8 +33,8 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
FLOAT_TYPE_VEC4 aa = FLOAT_TYPE_VEC4(TO_FLOAT_TYPE(data_a[idx]));
buf_a[buf_idx ] = aa.xy;
buf_a[buf_idx + 1] = aa.zw;
#else // LOAD_VEC_A == 2
const uint idx = pos_a * 2 + col * p.stride_a + row * 2;
#else // LOAD_VEC_BATCH_A == 2
const uint idx = pos_a + col * p.stride_a + row * 2;
const uint buf_idx = col * SHMEM_STRIDE + row;
if (idx_m < p.M && block + row * 2 + 1 < end_k) {
buf_a[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_a[idx]),
@@ -500,8 +500,8 @@ void load_b_to_shmem(const uint pos_b, const uint row, const uint col, const uin
#endif
buf_b[buf_idx + 0] = bb.xy;
buf_b[buf_idx + 1] = bb.zw;
#else // LOAD_VEC_B == 2
const uint idx = pos_b * 2 + col * p.stride_b + row * 2;
#else // LOAD_VEC_BATCH_B == 2
const uint idx = pos_b + col * p.stride_b + row * 2;
const uint buf_idx = col * SHMEM_STRIDE + row;
if (idx_n < p.N && block + row * 2 + 1 < end_k) {
buf_b[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_b[idx]),
@@ -536,17 +536,17 @@ void load_b_to_shmem(const uint pos_b, const uint row, const uint col, const uin
#endif
buf_b[buf_idx + 0] = bb.xy;
buf_b[buf_idx + 1] = bb.zw;
#else // LOAD_VEC_B == 2
#else // LOAD_VEC_BATCH_B == 2
const uint row_i = ic * BN + col;
const uint buf_idx = col * SHMEM_STRIDE + row;
if (row_i < _ne1 && block + row * 2 + 1 < end_k) {
const u16vec2 row_idx = row_ids[col];
const uint idx = pos_b * 2 + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
const uint idx = pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
buf_b[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_b[idx]),
TO_FLOAT_TYPE(data_b[idx + 1]));
} else if (row_i < _ne1 && block + row * 2 < end_k) {
const u16vec2 row_idx = row_ids[col];
const uint idx = pos_b * 2 + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
const uint idx = pos_b + row_idx.y * p.batch_stride_b + (row_idx.x % p.ne11) * p.stride_b + row * 2;
buf_b[buf_idx] = FLOAT_TYPE_VEC2(TO_FLOAT_TYPE(data_b[idx]), 0.0f);
} else {
buf_b[buf_idx] = FLOAT_TYPE_VEC2(0.0f);