CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-11-01 09:01:57 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: workaround for AMD Windows driver 16 bit unpack8 bug (#12472)

Browse Source
This commit is contained in:
Eve
2025-03-21 19:27:47 +00:00
committed by GitHub
parent af04481e6b
commit 30c42ef5cb
4 changed files with 12 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

14
ggml/src/ggml-vulkan/vulkan-shaders/mul_mm.comp
View File

@@ -336,8 +336,8 @@ void main() {
const uint iqs = idx & 0x07;
const float d = float(data_a_packed16[ib].d);
const i8vec2 v0 = unpack8(data_a_packed16[ib].qs[2*iqs]);
const i8vec2 v1 = unpack8(data_a_packed16[ib].qs[2*iqs + 1]);
const i8vec2 v0 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs])).xy; // vec4 used due to #12147
const i8vec2 v1 = unpack8(int32_t(data_a_packed16[ib].qs[2*iqs + 1])).xy;
const vec4 v = vec4(v0.x, v0.y, v1.x, v1.y) * d;
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
@@ -544,7 +544,7 @@ void main() {
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);
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -564,7 +564,7 @@ void main() {
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);
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -586,7 +586,7 @@ void main() {
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));
const vec2 v = db * vec2(sign01) * vec2(unpack8(uint32_t(grid)).xy); // vec4 used due to #12147
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -611,7 +611,7 @@ void main() {
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);
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);
@@ -631,7 +631,7 @@ void main() {
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);
const vec2 v = db * vec2(sign01) * vec2(unpack8(grid).xy); // vec4 used due to #12147
buf_a[buf_idx ] = FLOAT_TYPE(v.x);
buf_a[buf_idx + 1] = FLOAT_TYPE(v.y);