Lumpiasty/llama.cpp
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Vulkan MMQ Integer Dot Refactor and K-Quant support (#16536)

Browse Source 
* vulkan: add mmq q2_k integer dot support

* Refactor mmq caching

* Reduce mmq register use

* Load 4 quant blocks into shared memory in one step

* Pack q2_k blocks into caches of 32

* Use 32-bit accumulators for integer dot matmul

* Add q4_k mmq

* Add q3_k mmq

* Add q5_k mmq

* Add q6_k mmq

* Add mxfp4 mmq, enable MMQ MUL_MAT_ID

* Fix mmv dm loads
This commit is contained in:
Ruben Ortlam
2025-10-29 14:39:03 +01:00
committed by GitHub
parent 3eb2be1ca5
commit bcf5bda6f5
18 changed files with 941 additions and 418 deletions
Download Patch File Download Diff File Expand all files Collapse all files
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
+1 -71
View File
@@ -120,81 +120,11 @@ shared FLOAT_TYPE_VEC2 buf_b[BN * SHMEM_STRIDE];
#define NUM_WARPS (BLOCK_SIZE / WARP)
#ifdef MUL_MAT_ID
shared u16vec2 row_ids[BN];
uint _ne1;
#ifdef MUL_MAT_ID_USE_SUBGROUPS
shared uvec4 ballots_sh[NUM_WARPS];
void load_row_ids(uint expert_idx, bool nei0_is_pow2, uint ic) {
_ne1 = 0;
uint num_elements = p.nei1 * p.nei0;
uint nei0shift = findLSB(p.nei0);
uint ids[16];
uint iter = 0;
for (uint j = 0; j < num_elements; j += BLOCK_SIZE) {
// prefetch up to 16 elements
if (iter == 0) {
[[unroll]] for (uint k = 0; k < 16; ++k) {
uint i = j + gl_LocalInvocationIndex + k*BLOCK_SIZE;
bool in_range = i < num_elements;
uint ii1;
if (nei0_is_pow2) {
ii1 = i >> nei0shift;
} else {
ii1 = i / p.nei0;
}
uint ii0 = i - ii1 * p.nei0;
ids[k] = in_range ? data_ids[ii1*p.nbi1 + ii0] : 0;
}
}
uint i = j + gl_LocalInvocationIndex;
bool in_range = i < num_elements;
uint ii1;
if (nei0_is_pow2) {
ii1 = i >> nei0shift;
} else {
ii1 = i / p.nei0;
}
uint ii0 = i - ii1 * p.nei0;
uint id = ids[iter++];
uvec4 ballot = subgroupBallot(in_range && id == expert_idx);
ballots_sh[gl_SubgroupID] = ballot;
barrier();
uint subgroup_base = 0;
uint total = 0;
for (uint k = 0; k < gl_NumSubgroups; ++k) {
if (k == gl_SubgroupID) {
subgroup_base = total;
}
total += subgroupBallotBitCount(ballots_sh[k]);
}
barrier();
uint idx = subgroup_base + subgroupBallotExclusiveBitCount(ballot);
if (in_range && id == expert_idx && _ne1 + idx >= ic * BN && _ne1 + idx < (ic + 1) * BN) {
row_ids[_ne1 + idx - ic * BN] = u16vec2(ii0, ii1);
}
_ne1 += total;
iter &= 15;
if (_ne1 >= (ic + 1) * BN) {
break;
}
}
barrier();
}
#endif // MUL_MAT_ID_USE_SUBGROUPS
#endif // MUL_MAT_ID
#ifdef COOPMAT
shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
#endif
#include "mul_mm_id_funcs.glsl"
#include "mul_mm_funcs.glsl"
void main() {