vbatts/llama.cpp
1
0
Fork 0
Code Issues Pull requests Projects Releases Packages Wiki Activity Actions

force 16 sequential threads per block

Browse source
This commit is contained in:
Eve 2024-11-28 17:38:12 -05:00
parent 97e0c686a3
commit 2bca812230
1 changed files with 4 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

8
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec_q6_k.comp
View file

@ -23,11 +23,11 @@ void main() {
const uint num_blocks_per_row = p.ncols / QUANT_K; const uint num_blocks_per_row = p.ncols / QUANT_K;
const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row; const uint ib0 = a_offset / QUANT_K + row*num_blocks_per_row;
uint it_size = gl_WorkGroupSize.x/16; // 16 threads are used to process each block
const uint it_size = gl_WorkGroupSize.x/16;
const uint tid = gl_LocalInvocationID.x; const uint tid = gl_LocalInvocationID.x;
const uint itid = tid/it_size; // 0...16 const uint itid = tid%16; // 0...16
const uint ix = tid%it_size; const uint ix = tid/16;
const uint step = 8; const uint step = 8;