]> git.djapps.eu Git - pkg/ggml/sources/ggml/commit
vulkan: mul_mat_id coopmat2 optimizations (llama/15546)
authorJeff Bolz <redacted>
Sun, 31 Aug 2025 07:06:43 +0000 (02:06 -0500)
committerGeorgi Gerganov <redacted>
Fri, 5 Sep 2025 09:54:08 +0000 (12:54 +0300)
commit26580b60ca884268b79acf7b19f7b1e8cdd62663
tree082dc288082410dfa400cd927812152a54c8cb10
parentd58ab2404fbb59128b1d6563f85d0b997b7feb54
vulkan: mul_mat_id coopmat2 optimizations (llama/15546)

* vulkan: mul_mat_id coopmat2 optimizations

Add a path for when the tile fits in BN/2, similar to what we have for mul_mat.

Only call fetch_scales/store_scales once per QUANT_K block, and once at the
beginning in case start_k is not aligned.

* Also add a path for BN/4 - worth a couple more percent
src/ggml-vulkan/ggml-vulkan.cpp
src/ggml-vulkan/vulkan-shaders/mul_mm_cm2.comp