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

vulkan: copy iq4_nl LUT into shared memory (#10409)

Browse source
This commit is contained in:
Jeff Bolz 2024-11-20 01:40:18 -06:00 • committed by GitHub
parent 1bacb9f625
commit 8fd4b7fa29
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
6 changed files with 29 additions and 4 deletions
Download patch file Download diff file Expand all files Collapse all files

4
ggml/src/ggml-vulkan/vulkan-shaders/mul_mat_vec.comp
View file

@ -161,6 +161,10 @@ void compute_outputs(const uint32_t first_row, const uint32_t num_rows) {
void main() {
const uint first_row = NUM_ROWS * (gl_WorkGroupID.x + gl_NumWorkGroups.x * gl_WorkGroupID.z);
#if defined(DATA_A_IQ4_NL)
init_iq4nl_shmem();
#endif
// do NUM_ROWS at a time, unless there aren't enough remaining rows
if (first_row + NUM_ROWS <= p.stride_d) {
compute_outputs(first_row, NUM_ROWS);