Skip to content

Commit

Permalink
Reenable FMA in CUDA builds
Browse files Browse the repository at this point in the history
This might be controversial, but we're kneecapping the CUDA build by
turning off FMA. I think it's high time to turn it back on. Also removes
an inaccurate comment, as FMA doesn't have anything to do with fast
math, as FMA is well specified in IEEE 754-2008.
  • Loading branch information
stephenswat committed Dec 8, 2024
1 parent 033b4fc commit 8e7ed12
Showing 1 changed file with 0 additions and 3 deletions.
3 changes: 0 additions & 3 deletions cmake/detray-compiler-options-cuda.cmake
Original file line number Diff line number Diff line change
Expand Up @@ -29,9 +29,6 @@ if("${CMAKE_CUDA_COMPILER_ID}" MATCHES "NVIDIA")
# Allow to use functions in device code that are constexpr, even if they are
# not marked with __device__.
detray_add_flag( CMAKE_CUDA_FLAGS "--expt-relaxed-constexpr" )

# Turn off fast math for the device code.
detray_add_flag( CMAKE_CUDA_FLAGS "-fmad=false" )
endif()

# Make CUDA generate debug symbols for the device code as well in a debug
Expand Down

0 comments on commit 8e7ed12

Please sign in to comment.