Code

Vulkan Optimizations and Fixes#8959

Code

Merged

0cc4m merged 9 commits intomasterggerganov/llama.cpp:masterfrom 0cc4m/vulkan-optimizationggerganov/llama.cpp:0cc4m/vulkan-optimizationCopy head branch name to clipboardAug 14, 2024

+775-845

Commits

Commits on Aug 7, 2024

• Optimize Vulkan REPEAT performance

  0cc4m

  committed
  0645ed5

• Use Vulkan GLSL fused multiply-add instruction where possible

  0cc4m

  committed
  f78487b

• Add GGML_VULKAN_PERF option to output performance data per operator

  0cc4m

  committed
  4c6a7bb

Commits on Aug 9, 2024

• Rework and fix Vulkan descriptor set and descriptor pool handling

  0cc4m

  committed
  efe6aca

• Fix float32 concat f16 shader validation error

  0cc4m

  committed
  9e0ac98

Commits on Aug 10, 2024

• Add Vulkan GROUP_NORM eps parameter

  0cc4m

  committed
  61d8388

Commits on Aug 11, 2024

• Merge upstream changes, fix conflicts

  0cc4m

  committed
  4f197e3

• Fix validation error with transfer queue memory barrier flags

  0cc4m

  committed
  5ae33eb

Commits on Aug 14, 2024

• Remove trailing whitespaces

  0cc4m

  committed
  12d214f