ExternalVendorCode/whisper.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/whisper.cpp.git synced 2025-03-18 10:05:59 +00:00
Code Issues Actions 2 Packages Projects Releases Wiki Activity

vulkan: optimize coopmat2 q2_k dequant function (llama/11130)

Browse Source
This commit is contained in:
Jeff Bolz 2025-01-16 15:16:39 -06:00 committed by Georgi Gerganov
parent de49024e49
commit 62e2414620
1 changed files with 12 additions and 6 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

18
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
View File

@ -101,19 +101,25 @@ layout(buffer_reference, std430, buffer_reference_align = 4) buffer decodeBufQ2_
block_q2_K block;
};
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ2_K_packed16 {
block_q2_K_packed16 block;
};
float16_t dequantFuncQ2_K(const in decodeBufQ2_K bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
decodeBufQ2_K_packed16 bl16 = decodeBufQ2_K_packed16(bl);
const f16vec2 d = bl.block.d;
const uint idx = coordInBlock[1];
const uint iqs = idx;
const uint qsi = (iqs / 128) * 32 + (iqs % 32); // 0..31
const uint scalesi = iqs / 16; // 0..15
const uint qsshift = ((iqs % 128) / 32) * 2; // 0,2,4,6
const uint scalesi = (idx & 0xF0) >> 4; // 0..15
const uint qsshift = (idx & 0x60) >> 4; // 0,2,4,6
uint qs = uint32_t(bl16.block.qs[((idx & 0x80) >> 3) + ((idx & 0x1E) >> 1)]);
qs = (qs >> qsshift) & 0x0303;
qs = unpack8(qs)[idx & 1];
uint32_t qs = bl.block.qs[qsi];
const uint scales = bl.block.scales[scalesi];
float16_t ret = d.x * float16_t(scales & 0xF) * float16_t((qs >> qsshift) & 3) - d.y * float16_t(scales >> 4);
float16_t ret = d.x * float16_t(scales & 0xF) * float16_t(qs) - d.y * float16_t(scales >> 4);
return ret;
}