mirrors/llama.cpp
1
0
Fork 0
mirror of https://github.com/ggerganov/llama.cpp.git synced 2025-04-16 03:26:08 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: use fp32 in coopmat2 q4_k dequant function (#12309)

Browse Source
This commit is contained in:
Jeff Bolz 2025-03-17 04:43:35 -05:00 committed by GitHub
parent 891c63956d
commit f07690c930
No known key found for this signature in database
GPG Key ID: B5690EEEBB952194
1 changed files with 5 additions and 5 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

10
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
View File

@ -178,7 +178,7 @@ float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2
uvec4 v = bl128.block.q4k[0];
const f16vec2 loadd = unpackFloat2x16(v.x);
const vec2 loadd = vec2(unpackFloat2x16(v.x));
uint32_t sc;
uint32_t mbyte;
@ -199,15 +199,15 @@ float16_t dequantFuncQ4_K(const in decodeBufQ4_K bl, const in uint blockCoords[2
sc &= 0x3F;
mbyte &= 0x3F;
const float16_t d = loadd.x * float16_t(sc);
const float16_t m = loadd.y * float16_t(mbyte);
const float d = loadd.x * float(sc);
const float m = loadd.y * float(mbyte);
uint qs = uint32_t(bl16.block.qs[((idx & 0xC0) >> 2) + ((idx & 0x1E) >> 1)]);
qs = (qs >> (b * 4 + 8 * (idx & 1))) & 0xF;
float16_t ret = d * float16_t(qs) - m;
float ret = d * float(qs) - m;
return ret;
return float16_t(ret);
}
layout(buffer_reference, std430, buffer_reference_align = 16) buffer decodeBufQ5_K {