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

vulkan: fix failing dequant shaders (#15862)

Browse Source 
* vulkan: fix failing dequant shaders

* add missing const
This commit is contained in:
Jeff Bolz
2025-09-13 16:29:43 +01:00
committed by GitHub
parent b9c9c9f789
commit aa0c461efe
4 changed files with 13 additions and 9 deletions
Download Patch File Download Diff File Expand all files Collapse all files

11
ggml/src/ggml-vulkan/vulkan-shaders/dequant_iq3_s.comp
View File

@@ -22,15 +22,16 @@ void main() {
const uint b_idx = 256 * ib + 32 * is;
const float d = float(data_a[ib].d);
const float db = d * (1 + 2 * ((data_a[ib].scales[is] >> (4 * (is % 2))) & 0xf));
const float db = d * (1 + 2 * ((data_a[ib].scales[is / 2] >> (4 * (is % 2))) & 0xf));
// We must produce 32 values using 4 sign bytes, 1 qh byte, 8 qs bytes.
uint qh = data_a[ib].qh[is];
[[unroll]] for (uint l = 0; l < 8; ++l) {
uint qs = data_a[ib].qs[8 * is + l];
uint gidx = qs | ((qh << (8 - l)) & 256);
uint8_t signs = data_a[ib].signs[8 * is + l / 2] >> (4 * (l & 1));
u8vec4 grid = unpack8(iq3s_grid[gidx]);
const uint iqs = 8 * is + l;
const uint qs = data_a[ib].qs[iqs];
const uint gidx = qs | ((qh << (8 - l)) & 256);
const uint8_t signs = data_a[ib].signs[iqs / 2] >> (4 * (l & 1));
const u8vec4 grid = unpack8(iq3s_grid[gidx]);
data_b[b_idx + 4 * l + 0] = D_TYPE(db * grid.x * ((signs & 1) != 0 ? -1.0 : 1.0));
data_b[b_idx + 4 * l + 1] = D_TYPE(db * grid.y * ((signs & 2) != 0 ? -1.0 : 1.0));
data_b[b_idx + 4 * l + 2] = D_TYPE(db * grid.z * ((signs & 4) != 0 ? -1.0 : 1.0));