You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Is your feature request related to a problem? Please describe.
We want to announce behavior changes coming of the NATIVE case of CMAKE_CUDA_ARCHITECTURES from rapids-cmake.
Describe the solution you'd like
Unlike #316 the swtich from NATIVE to native does have an observable side-effect on the project. The rapids-cmake NATIVE resolves down to the actual architecture value ( 75-real ) after the project() call, while native stays native as now nvcc supports flags like -arch=native.
What we want to do for 22.03 is add a behavior notice to rapids-cmake that NATIVE will map to native starting in 23.04.
The text was updated successfully, but these errors were encountered:
Was thinking about this again and wanted to measure the performance cost of using -arch=native compared to what rapids-cmake does.
For libcudf total compile times at -j20 go from 8min 42sec to 8min 44sec so basically noise. We shouldn't consider performance of -arch=native to a issue when switching rapids-cmake NATIVE over to native.
Is your feature request related to a problem? Please describe.
We want to announce behavior changes coming of the
NATIVE
case of CMAKE_CUDA_ARCHITECTURES from rapids-cmake.Describe the solution you'd like
Unlike #316 the swtich from
NATIVE
tonative
does have an observable side-effect on the project. The rapids-cmakeNATIVE
resolves down to the actual architecture value ( 75-real ) after theproject()
call, whilenative
stays native as nownvcc
supports flags like-arch=native
.What we want to do for 22.03 is add a behavior notice to rapids-cmake that
NATIVE
will map tonative
starting in 23.04.The text was updated successfully, but these errors were encountered: