Vulkan: VK_KHR_cooperative_matrix support to speed up prompt processing#10597

Merged

0cc4m merged 12 commits intomasterfrom

0cc4m/vulkan-coopmat

Dec 7, 2024

Commits on Dec 5, 2024

Vulkan: Implement VK_KHR_cooperative_matrix support in the matrix matrix multiplication shader
0cc4m
committed
Improve performance with better q4_k and q5_k dequant and store unrolling
0cc4m
committed
Add Vulkan MUL_MAT and MUL_MAT_ID accumulator precision selection
0cc4m
committed
Rework mulmat shader selection and compilation logic, avoid compiling shaders that won't get used by device
0cc4m
committed
Vulkan: Implement accumulator switch for specific mul mat mat shaders
0cc4m
committed
Vulkan: Unroll more loops for more mul mat mat performance
0cc4m
committed
Vulkan: Add VK_AMD_shader_core_properties2 support to read Compute Unit count for split_k logic
0cc4m
committed
Disable coopmat support on AMD proprietary driver
0cc4m
committed
Remove redundant checks
0cc4m
committed
Add environment variable GGML_VK_DISABLE_COOPMAT to disable VK_KHR_cooperative_matrix support
0cc4m
committed

Commits on Dec 6, 2024

Fix rebase typo
0cc4m
committed

Commits on Dec 7, 2024

Fix coopmat2 MUL_MAT_ID pipeline selection
0cc4m
committed