Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests 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
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
+20 -8
View File
@@ -31,10 +31,22 @@
#include "types.comp"
#ifndef LOAD_VEC_A
#define LOAD_VEC_A 2
#define LOAD_VEC_A 1
#endif
#ifndef LOAD_VEC_B
#define LOAD_VEC_B 2
#define LOAD_VEC_B 1
#endif
// Load 2 values at once without affecting index calculations through LOAD_VEC
#if (defined(DATA_A_F32) || defined(DATA_A_F16) || defined(DATA_A_BF16)) && !defined(ALIGNED)
#define LOAD_VEC_BATCH_A 2
#else
#define LOAD_VEC_BATCH_A 1
#endif
#if !defined(ALIGNED)
#define LOAD_VEC_BATCH_B 2
#else
#define LOAD_VEC_BATCH_B 1
#endif
#if !defined(TO_FLOAT_TYPE)
@@ -236,13 +248,13 @@ void main() {
const uint warp_r = warp_i % (BM / WM);
const uint warp_c = warp_i / (BM / WM);
const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A);
const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A);
const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B);
const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B);
const uint loadr_a = gl_LocalInvocationID.x % (BK / LOAD_VEC_A / LOAD_VEC_BATCH_A);
const uint loadc_a = gl_LocalInvocationID.x / (BK / LOAD_VEC_A / LOAD_VEC_BATCH_A);
const uint loadr_b = gl_LocalInvocationID.x % (BK / LOAD_VEC_B / LOAD_VEC_BATCH_B);
const uint loadc_b = gl_LocalInvocationID.x / (BK / LOAD_VEC_B / LOAD_VEC_BATCH_B);
const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A / BK;
const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B / BK;
const uint loadstride_a = gl_WorkGroupSize.x * LOAD_VEC_A * LOAD_VEC_BATCH_A / BK;
const uint loadstride_b = gl_WorkGroupSize.x * LOAD_VEC_B * LOAD_VEC_BATCH_B / BK;
#ifdef MUL_MAT_ID
#ifdef MUL_MAT_ID_USE_SUBGROUPS