tqcq/llama.cpp
0
0
Fork 0
mirror of https://github.com/ggml-org/llama.cpp.git synced 2025-06-28 20:25:20 +00:00
Code Issues Packages Projects Releases Wiki Activity

vulkan: fix coopmat2 validation failures (#11284)

Browse Source 
mul mat and flash attention shaders were loading f32 types directly into
A/B matrices, which happens to work but is technically invalid usage.
For FA, we can load it as an Accumulator matrix and convert and this
is not in the inner loop and is cheap enough. For mul mat, it's more
efficient to do this conversion in a separate pass and have the input(s)
be f16.

coopmat2 requires SPIR-V 1.6 (related using to LocalSizeId). LocalSizeId
requires maintenance4 be enabled, and SPIR-V 1.6 requires Vulkan 1.3.
This commit is contained in:
Jeff Bolz
2025-01-20 10:38:32 -06:00
committed by GitHub
parent 9f7add1cde
commit aea8ddd516
4 changed files with 53 additions and 56 deletions
Download Patch File Download Diff File Expand all files Collapse all files

7
ggml/src/ggml-vulkan/vulkan-shaders/vulkan-shaders-gen.cpp
View File

@ -316,8 +316,11 @@ void matmul_shaders(bool fp16, bool matmul_id, bool coopmat, bool coopmat2, bool
// For aligned matmul loads
std::string load_vec_a = (coopmat2 || tname == "f32" || tname == "f16") ? load_vec : "2";
string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
// don't generate f32 variants for coopmat2
if (!coopmat2) {
string_to_spv(shader_name + "_" + tname + "_f32", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc);
string_to_spv(shader_name + "_" + tname + "_f32_aligned", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a}, {"LOAD_VEC_B", load_vec}, {"B_TYPE", aligned_b_type_f32}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}, {"ALIGNED", "1"}}), fp16, coopmat, coopmat2, f16acc);
}
if (tname != "f16" && tname != "f32") {
string_to_spv(shader_name + "_" + tname + "_f16", source_name, merge_maps(base_dict, {{data_a_key, "1"}, {"LOAD_VEC_A", load_vec_a_unaligned}, {"B_TYPE", "float16_t"}, {"D_TYPE", "float"}, {"B_IS_FLOAT", "1"}}), fp16, coopmat, coopmat2, f16acc);