PR with opt params
- PR is online
- Added two architectures to FindO2GPU.cmake
- Changed Ampere detection from cc 8.6 to 8.0 (why it was like this?)
if(CUDA_FIRST_TARGET GREATER_EQUAL 120)
set(CUDA_TARGET BLACKWELL)
elseif(CUDA_FIRST_TARGET GREATER_EQUAL 90)
set(CUDA_TARGET HOPPER)
elseif(CUDA_FIRST_TARGET GREATER_EQUAL 89)
set(CUDA_TARGET ADA)
elseif(CUDA_FIRST_TARGET GREATER_EQUAL 80) # was 86 before
set(CUDA_TARGET AMPERE)
elseif(CUDA_FIRST_TARGET GREATER_EQUAL 75)
set(CUDA_TARGET TURING)
elseif(CUDA_FIRST_TARGET GREATER_EQUAL 60)
set(CUDA_TARGET PASCAL)
elseif(CUDA_FIRST_TARGET GREATER_EQUAL 30)
set(CUDA_TARGET KEPLER)
elseif(CUDA_FIRST_TARGET GREATER_EQUAL 20)
set(CUDA_TARGET FERMI)
else()
set(CUDA_TARGET TESLA)
endif()
Best parameters comparison - Nvidia
Kernels in single stream steps




Kernels in multi stream steps



Next steps
- Validate against manual tuning