CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-08 10:07:01 +00:00
Code Issues Packages 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

55
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
View File

@@ -6,6 +6,9 @@
#ifdef FLOAT16
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
#endif
#if defined(DATA_A_IQ1_M)
#extension GL_EXT_shader_explicit_arithmetic_types_float16 : require
#endif
#ifdef COOPMAT
#extension GL_KHR_cooperative_matrix : enable
@@ -95,7 +98,7 @@ shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
#endif
void main() {
#if defined(DATA_A_IQ2_XXS) || defined(DATA_A_IQ2_XS) || defined(DATA_A_IQ2_S) || defined(DATA_A_IQ3_XXS) || defined(DATA_A_IQ3_S) || defined(DATA_A_IQ4_XS) || defined(DATA_A_IQ4_NL)
#ifdef NEEDS_INIT_IQ_SHMEM
init_iq_shmem(gl_WorkGroupSize);
#endif
@@ -437,6 +440,56 @@ void main() {
buf_a[buf_idx ] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi ] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi ] >> qhshift) & 3) << 4)) - 32));
buf_a[buf_idx + 1] = FLOAT_TYPE(dscale * float(int8_t(((data_a[ib].ql[qsi + 1] >> (b * 4)) & 0xF) | (((data_a[ib].qh[qhi + 1] >> qhshift) & 3) << 4)) - 32));
#elif defined(DATA_A_IQ1_S)
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
const uint ib = idx / 128; // 2 values per idx
const uint ib32 = (idx % 128) / 16; // 0..7
const uint ib8 = (idx % 128) / 4;
const int i8 = 2 * int(idx % 4);
const float d = float(data_a[ib].d);
const uint qh = data_a[ib].qh[ib32];
const uint qs = data_a[ib].qs[ib8];
const float dl = d * (2 * bitfieldExtract(qh, 12, 3) + 1);
const float delta = ((qh & 0x8000) != 0) ? -IQ1S_DELTA : IQ1S_DELTA;
const int16_t grid = int16_t(iq1s_grid[qs | (bitfieldExtract(qh, 3 * int(ib8 & 3), 3) << 8)]);
const ivec2 gvec = ivec2(
bitfieldExtract(grid, 2 * (i8), 2),
bitfieldExtract(grid, 2 * (i8 + 1), 2)
);
const vec2 v = dl * (vec2(gvec) + delta);
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_IQ1_M)
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;
const uint ib = idx / 128; // 2 values per idx
const uint ib8 = (idx % 128) / 4;
const uint ib16 = ib8 / 2;
const int i8 = 2 * int(idx % 4);
const uint16_t[4] scales = data_a[ib].scales;
const u16vec4 s = u16vec4(scales[0], scales[1], scales[2], scales[3]) >> 12;
const float d = float(unpackHalf2x16(s.x | (s.y << 4) | (s.z << 8) | (s.w << 12)).x);
const uint sc = scales[ib8 / 8];
const uint qs = data_a[ib].qs[ib8];
const uint qh = data_a[ib].qh[ib16] >> (4 * (ib8 & 1));
const float dl = d * (2 * bitfieldExtract(sc, 3 * int(ib16 & 3), 3) + 1);
const float delta = ((qh & 8) != 0) ? -IQ1M_DELTA : IQ1M_DELTA;
const int16_t grid = int16_t(iq1s_grid[qs | ((qh & 7) << 8)]);
const ivec2 gvec = ivec2(
bitfieldExtract(grid, 2 * (i8), 2),
bitfieldExtract(grid, 2 * (i8 + 1), 2)
);
const vec2 v = dl * (vec2(gvec) + delta);
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_IQ2_XXS)
const uint idx = pos_a + (loadc_a + l) * p.stride_a / LOAD_VEC_A + loadr_a;
const uint buf_idx = (loadc_a + l) * SHMEM_STRIDE + loadr_a * LOAD_VEC_A;