CS348Project/llama.cpp
5
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-10-27 08:21:30 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: 64-bit im2col (#16135)

Browse Source 
* vulkan: 64-bit im2col

Add variants of the im2col shaders that use buffer_device_address/buffer_reference,
and use 64-bit address calculations. This is needed for large convolutions used in
stable-diffusion.cpp.

* fix validation error for large im2col
This commit is contained in:
Jeff Bolz
2025-09-28 01:38:37 -05:00
committed by GitHub
parent 6a2c6145a0
commit d8359f5fde
6 changed files with 117 additions and 26 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
tests/test-backend-ops.cpp
View File

@@ -5753,6 +5753,13 @@ static std::vector<std::unique_ptr<test_case>> make_test_cases_eval() {
}
}
#if 0
// >4GB im2col destination. Too slow to run by default.
// Test cases taken from Wan2.1 T2V 1.3B.
test_cases.emplace_back(new test_im2col (GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {832, 480, 192, 4}, {3, 3, 192, 96}, 1, 1, 1, 1, 1, 1, true));
test_cases.emplace_back(new test_im2col_3d(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {834, 482, 6, 96}, {3, 3,3, 9216}, 96, 1, 1, 1, 0, 0, 0, 1, 1, 1, false));
#endif
// im2col 1D
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F32, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));
test_cases.emplace_back(new test_im2col(GGML_TYPE_F32, GGML_TYPE_F16, GGML_TYPE_F32, {3000, 128, 1, 1}, {3, 128, 1280, 1}, 1, 0, 1, 0, 1, 0, false));