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

vulkan: im2col and matmul optimizations for stable diffusion (#10942)

Browse source 
* tests: Add im2col perf tests

* vulkan: optimize im2col, more elements per thread

* vulkan: increase small tile size for NV_coopmat2

* vulkan: change im2col to 512 elements per workgroup
This commit is contained in:
Jeff Bolz 2024-12-29 03:16:34 -06:00 • committed by GitHub
parent fdd2188912
commit a813badbbd
No known key found for this signature in database
GPG key ID: B5690EEEBB952194
3 changed files with 66 additions and 29 deletions
Download patch file Download diff file Expand all files Collapse all files

12
tests/test-backend-ops.cpp
View file

@ -3945,6 +3945,18 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_perf() {
}
}
for (int K : {3, 5}) {
for (int IC : {256, 2560}) {
for (int IW_IH : {32, 64, 256}) {
if (IC == 2560 && IW_IH == 256) {
// too big
continue;
}
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {IW_IH, IW_IH, IC, 1}, {K, K, IC, 1}, 1, 1, 1, 1, 1, 1, true));
}
}
}
return test_cases;
}