CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-02 09:12:03 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: implement initial support for IQ2 and IQ3 quantizations (#11360)

Browse Source 
* vulkan: initial support for IQ3_S

* vulkan: initial support for IQ3_XXS

* vulkan: initial support for IQ2_XXS

* vulkan: initial support for IQ2_XS

* vulkan: optimize Q3_K by removing branches

* vulkan: implement dequantize variants for coopmat2

* vulkan: initial support for IQ2_S

* vulkan: vertically realign code

* port failing dequant callbacks from mul_mm

* Fix array length mismatches

* vulkan: avoid using workgroup size before it is referenced

* tests: increase timeout for Vulkan llvmpipe backend

---------

Co-authored-by: Jeff Bolz <jbolz@nvidia.com>
This commit is contained in:
Rémy Oudompheng
2025-01-29 18:29:39 +01:00
committed by GitHub
parent e51c47b401
commit 66ee4f297c
19 changed files with 1616 additions and 40 deletions
Download Patch File Download Diff File Expand all files Collapse all files

122
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
View File

@@ -95,8 +95,8 @@ shared ACC_TYPE coopmat_stage[TM * TN * NUM_WARPS];
#endif
void main() {
#if defined(DATA_A_IQ4_NL)
init_iq4nl_shmem();
#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_NL)
init_iq_shmem(gl_WorkGroupSize);
#endif
#ifdef MUL_MAT_ID
@@ -343,10 +343,8 @@ void main() {
const uint qsshift = halfsplit * 2; // 0,2,4,6
const uint m = 1 << (4 * n + halfsplit); // 1,2,4,8,16,32,64,128
const int8_t us = int8_t(is < 4 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+8] >> 0) & 3) << 4) :
is < 8 ? (data_a[ib].scales[is-0] & 0xF) | (((data_a[ib].scales[is+4] >> 2) & 3) << 4) :
is < 12 ? (data_a[ib].scales[is-8] >> 4) | (((data_a[ib].scales[is+0] >> 4) & 3) << 4) :
(data_a[ib].scales[is-8] >> 4) | (((data_a[ib].scales[is-4] >> 6) & 3) << 4));
const int8_t us = int8_t(((data_a[ib].scales[is % 8] >> (4 * int(is / 8))) & 0xF)
| (((data_a[ib].scales[8 + (is % 4)] >> (2 * int(is / 4))) & 3) << 4));
const float dl = float(data_a[ib].d) * float(us - 32);
buf_a[buf_idx ] = FLOAT_TYPE(dl * float(int8_t((data_a[ib].qs[qsi ] >> qsshift) & 3) - (((data_a[ib].hmask[hmi ] & m) != 0) ? 0 : 4)));
@@ -439,6 +437,118 @@ 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_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;
const uint ib = idx / 128; // 2 values per idx
const uint ib32 = (idx % 128) / 16; // 0..7
const uint ib8 = (idx / 4) % 4;
const float d = float(data_a[ib].d);
const uint qs = data_a[ib].qs[8 * ib32 + ib8];
const uint signs = pack32(u8vec4(
data_a[ib].qs[8*ib32 + 4],
data_a[ib].qs[8*ib32 + 5],
data_a[ib].qs[8*ib32 + 6],
data_a[ib].qs[8*ib32 + 7]
));
const float db = d * 0.25 * (0.5 + (signs >> 28));
const uint32_t sign7 = bitfieldExtract(signs, 7 * int(ib8), 7);
const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
const uint grid = iq2xxs_grid[qs][(idx % 4) / 2] >> (16 * (idx & 1));
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_IQ2_XS)
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 / 4) % 4; // 0..3
const float d = float(data_a[ib].d);
const uint scale = (data_a[ib].scales[ib32] >> (2 * (ib8 & 2))) & 0xf;
const float db = d * 0.25 * (0.5 + scale);
const uint qs = data_a[ib].qs[4 * ib32 + ib8];
const uint sign7 = qs >> 9;
const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
const uint grid = iq2xs_grid[qs & 511][(idx % 4) / 2] >> (16 * (idx & 1));
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_IQ2_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 ib8 = (idx % 128) / 4; // 0..31
const uint ib32 = ib8 / 4; // 0..7
const uint scale = (data_a[ib].scales[ib32] >> (2 * (ib8 & 2))) & 0xf;
const uint qs = data_a[ib].qs[ib8];
const uint qh = data_a[ib].qh[ib32];
const uint qhshift = 2 * (ib8 % 4);
const uint sign = data_a[ib].qs[QUANT_K / 8 + ib8] >> (2 * (idx % 4));
const float d = float(data_a[ib].d);
const float db = d * 0.25 * (0.5 + scale);
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
const uint16_t grid = unpack16(iq2s_grid[qs | ((qh << (8 - qhshift)) & 0x300)][(idx & 2) >> 1])[idx & 1];
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid));
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_IQ3_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;
const uint ib = idx / 128; // 2 values per idx
const uint iqs = (idx % 128) / 2; // 0..63
const uint is = QUANT_K / 4 + 4 * (iqs / 8); // 8 values
const float d = float(data_a[ib].d);
const uint qs = data_a[ib].qs[iqs];
const uint signs = pack32(u8vec4(
data_a[ib].qs[is+0],
data_a[ib].qs[is+1],
data_a[ib].qs[is+2],
data_a[ib].qs[is+3]
));
const float db = d * 0.5 * (0.5 + (signs >> 28));
const uint32_t sign7 = bitfieldExtract(signs, 7 * (int(iqs / 2) % 4), 7);
const uint sign = (sign7 | (bitCount(sign7) << 7)) >> (2 * (idx % 4));
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(int8_t(sign << 1), int8_t(sign))));
const uint grid = iq3xxs_grid[qs] >> (16 * (idx & 1));
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_IQ3_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 iqs = (idx % 128) / 2; // 0..63
const uint iqh = iqs / 8;
const float d = float(data_a[ib].d);
const uint qs = data_a[ib].qs[iqs];
const uint qh = data_a[ib].qh[iqh];
const int8_t sign = int8_t(data_a[ib].signs[iqs / 2] >> (2 * (idx % 4)));
const uint scale = data_a[ib].scales[iqs / 16];
const i8vec2 sign01 = i8vec2(1 - (2 & i8vec2(sign << 1, sign)));
const float db = d * (1 + 2 * ((scale >> (4 * (iqh & 1))) & 0xf));
const uint32_t grid = iq3s_grid[qs | ((qh << (8 - (iqs % 8))) & 256)] >> (16 * (idx % 2));
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy);
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
#elif defined(DATA_A_IQ4_NL)
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;