EngineX-Ascend/enginex-ascend-910-llama.cpp
10
0
Fork 0
Code Issues Pull Requests Actions 4 Projects Releases Wiki Activity

vulkan: initial support for IQ1_S and IQ1_M quantizations (#11528)

Browse Source 
* vulkan: initial support for IQ1_S and IQ1_M quantizations

* vulkan: define MMV kernels for IQ1 quantizations

* devops: increase timeout of Vulkan tests again

* vulkan: simplify ifdef for init_iq_shmem
This commit is contained in:
Rémy O
2025-02-15 09:01:40 +01:00
committed by GitHub
parent 89daa2564f
commit fc1b0d0936
17 changed files with 711 additions and 30 deletions
Download Patch File Download Diff File Expand all files Collapse all files

54
ggml/src/ggml-vulkan/vulkan-shaders/dequant_funcs_cm2.comp
View File

@@ -301,6 +301,56 @@ float16_t dequantFuncQ6_K(const in decodeBufQ6_K bl, const in uint blockCoords[2
return ret;
}
#if defined(DATA_A_IQ1_S)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1_S {
block_iq1_s block;
};
float16_t dequantFuncIQ1_S(const in decodeBufIQ1_S bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const float16_t d = bl.block.d;
const uint idx = coordInBlock[1];
const uint ib32 = idx / 32;
const uint ib8 = idx / 8;
const uint qh = bl.block.qh[ib32];
const uint qs = bl.block.qs[ib8];
const float dl = d * float(2 * bitfieldExtract(qh, 12, 3) + 1);
const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
const uint grid = iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)];
float16_t ret = float16_t(dl) * (float16_t(bitfieldExtract(int(grid), 2 * int(idx % 8), 2)) + float16_t(delta));
return ret;
}
#endif
#if defined(DATA_A_IQ1_M)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ1_M {
block_iq1_m block;
};
float16_t dequantFuncIQ1_M(const in decodeBufIQ1_M bl, const in uint blockCoords[2], const in uint coordInBlock[2])
{
const u16vec4 scales = u16vec4(bl.block.scales[0], bl.block.scales[1], bl.block.scales[2], bl.block.scales[3]) >> 12;
const float16_t d = uint16BitsToHalf(scales.x | (scales.y << 4) | (scales.z << 8) | (scales.w << 12));
const uint idx = coordInBlock[1];
const uint ib8 = idx / 8;
const uint ib16 = idx / 16;
const int i8 = int(idx % 8);
const uint sc = bl.block.scales[ib8 / 8];
const uint qs = bl.block.qs[ib8];
const uint qh = bl.block.qh[ib16] >> (4 * (ib8 & 1));
const float dl = 2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1;
const float delta = ((qh & 8) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
const uint grid = iq1s_grid[qs | ((qh & 7) << 8)];
float16_t ret = d * float16_t(dl) * (float16_t(bitfieldExtract(int(grid), 2 * i8, 2)) + float16_t(delta));
return ret;
}
#endif
#if defined(DATA_A_IQ2_XXS)
layout(buffer_reference, std430, buffer_reference_align = 2) buffer decodeBufIQ2_XXS {
block_iq2_xxs block;
@@ -512,6 +562,10 @@ float16_t dequantFuncIQ4_NL(const in decodeBufIQ4_NL bl, const in uint blockCoor
#define dequantFuncA dequantFuncQ5_K
#elif defined(DATA_A_Q6_K)
#define dequantFuncA dequantFuncQ6_K
#elif defined(DATA_A_IQ1_S)
#define dequantFuncA dequantFuncIQ1_S
#elif defined(DATA_A_IQ1_M)
#define dequantFuncA dequantFuncIQ1_M
#elif defined(DATA_A_IQ2_XXS)
#define dequantFuncA dequantFuncIQ2_XXS
#elif defined(DATA_A_IQ2_XS)