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_funcs.glsl
+7 -7
View File
@@ -134,15 +134,15 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
const uint ib = idx / 128; // 2 values per idx
const uint iqs = idx % 128; // 0..127
const uint qsi = (iqs / 64) * 32 + (iqs % 16) * 2; // 0,2,4..30
const uint qsi = (iqs / 64) * 16 + (iqs % 16); // 0..15
const uint scalesi = iqs / 8; // 0..15
const uint qsshift = ((iqs % 64) / 16) * 2; // 0,2,4,6
const uvec2 qs = uvec2(data_a[ib].qs[qsi], data_a[ib].qs[qsi + 1]);
const uvec2 qs = uvec2(unpack8(data_a_packed16[ib].qs[qsi]));
const uint scales = data_a[ib].scales[scalesi];
const vec2 d = vec2(data_a[ib].d);
const vec2 dm = vec2(data_a[ib].dm);
const vec2 v = d.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - d.y * float(scales >> 4);
const vec2 v = dm.x * float(scales & 0xF) * vec2((qs >> qsshift) & 3) - dm.y * float(scales >> 4);
buf_a[buf_idx] = FLOAT_TYPE_VEC2(v.xy);
#elif defined(DATA_A_Q3_K)
@@ -179,7 +179,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
const uint is = 2 * n + b; // 0..7
const uint qsi = n * 32 + (iqs % 16) * 2; // 0,2,4..126
const vec2 loadd = vec2(data_a[ib].d);
const vec2 loadd = vec2(data_a[ib].dm);
const uint scidx0 = (is < 4) ? is : (is + 4);
const uint scidx1 = (is < 4) ? is : (is - 4);
@@ -215,7 +215,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
const uint8_t hm = uint8_t(1 << (iqs / 16));
const vec2 loadd = vec2(data_a[ib].d);
const vec2 loadd = vec2(data_a[ib].dm);
const uint scidx0 = (is < 4) ? is : (is + 4);
const uint scidx1 = (is < 4) ? is : (is - 4);
@@ -468,7 +468,7 @@ void load_a_to_shmem(const uint pos_a, const uint row, const uint col, const uin
const uint ib = idx / 8;
const uint iqs = (idx & 0x07) * 2;
const float d = e8m0_to_fp32(data_a[ib].e);
const float d = e8m0_to_fp32(data_a[ib].e) * 0.5;
const uint vui = uint(data_a[ib].qs[iqs]);
const uint vui2 = uint(data_a[ib].qs[iqs+1]);