-
Notifications
You must be signed in to change notification settings - Fork 901
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
[REVIEW] Use CMAKE_CUDA_ARCHITECTURES #7391
[REVIEW] Use CMAKE_CUDA_ARCHITECTURES #7391
Conversation
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Things generally look good. Should we consider to move to using the CUDA_ARCHITECTURES
target property instead of manually passing flags to the CUDA compiler?
I think that move is reasonable. Maybe after / part of #7107? It would allow better compilation flags control. If we ever needed targets built for a subset of archs we could. |
Codecov Report
@@ Coverage Diff @@
## branch-0.19 #7391 +/- ##
==============================================
Coverage ? 82.19%
==============================================
Files ? 100
Lines ? 16968
Branches ? 0
==============================================
Hits ? 13947
Misses ? 3021
Partials ? 0 Continue to review full report at Codecov.
|
rerun tests |
473169d
to
db5c7d3
Compare
db5c7d3
to
d319394
Compare
Pinged some folks offline regarding rapids-compose and devops scripts to make sure no one is using |
@gpucibot merge |
Based on #706 as they both modify `cmake/Modules/SetGPUArchs.cmake` This brings rmm's handling of `CMAKE_CUDA_ARCHITECTURES` to match that is proposed for cudf in rapidsai/cudf#7391 Authors: - Robert Maynard (@robertmaynard) Approvers: - Mark Harris (@harrism) - Keith Kraus (@kkraus14) URL: #709
This eliminates the `Policy CMP0104 is not set` warning during the JNI build by using `CMAKE_CUDA_ARCHITECTURES` to specify the targeted CUDA architectures during the build. This eliminates a lot of code replicated from the cpp build and reuses the `ConfigureCUDA` module added in #7391. The architectures being used by the build are visible in the `mvn` output, e.g.: ``` [exec] -- CUDF: Building CUDF for GPU architectures: 60-real;70-real;75 ``` This also configures the CPU compiler to use the same flags as the cpp build which required fixing a number of warnings (e.g.: sign mismatch comparisons, unused variables, etc.) in the JNI code since warnings are errors in the build. Authors: - Jason Lowe (@jlowe) Approvers: - Robert (Bobby) Evans (@revans2) - Alessandro Bellina (@abellina) - MithunR (@mythrocks) URL: #7425
When using CMake >= 3.18 you now don't get
CMP0104
warnings for each target that use CUDA.To correct this issue I have cherry-picked a collection of changes from #7107 which updated cudf to use
CMAKE_CUDA_ARCHITECTURES
.When using CMake < 3.18 cudf will transform
CMAKE_CUDA_ARCHITECTURES
into the correct flags and store it intoCMAKE_CUDA_FLAGS
.When I was porting this over from !7107 I noticed that when cudf is built for multi arch ( sm60, sm70, sm80, ... ) it only compiles compute for the newest arch. This behavior is preserved, and !7107 will need to be updated with this change.