Skip to content

benchmark for gpu compute capability using vulkan

Notifications You must be signed in to change notification settings

junka/vkcompbench

Repository files navigation

vkcompbench

Do benchmark for vulkan compute pipeline

We will need glslc from shaderc projec.

  • For Macos, It can be installed by: brew install shaderc
  • For Linux, It can be installed by: sudo apt install glslc

Then compile and run:

cmake -S . -B build
cmake --build build
./build/vkcompbench

Note for fp64 benchmark:

  • _Float64 did not get supported in clang, use double instead

Note for fp16 benchmark:

  • _Float16 need gcc-12 at least

C23 features will support std::float16_t and std::float64_t, but not all current active compilers will work with new features.

testcases cover

  • fp64: float point 64
  • int64: integer 64
  • fp32: float point 32
  • int32: integer 32
  • fp16: float point 16
  • int16: integer 16
  • int8: integer 8
  • int8dot: integer 8 dot product
  • int8dotaccsat: integer 8 dot product with accumulate and saturation
  • int8dot4x8packed: integer 8 dot product with 4x8 packed

About

benchmark for gpu compute capability using vulkan

Topics

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published