Conversation
2a77225 to
d731c38
Compare
cgmb
left a comment
There was a problem hiding this comment.
You can get a pretty big reduction in your build time if you specify an amdgpu_target variant for the mathlibs, particularly rocblas, rocfft, rocsparse and rocsolver. The build time cost of each architecture is basically linear and the default amdgpu_target=auto builds for ~7 architectures. Though, perhaps setting it for rocrand has forced the rest to resolve to amdgpu_target=gfx1030 as well?
|
@cgmb I wasn't sure how to set the I think now that the #30582 is merged I can put that in more easily. |
61c806d to
cc414e9
Compare
cc414e9 to
7dd8035
Compare
|
@scottwittenburg or @srekolam if you guys could give this an approval to get it merged, I think it is ready now! |
b46e530 to
8e4a169
Compare
a2f4f3a to
746f568
Compare
746f568 to
61e5cbf
Compare
|
@srekolam I just rebased, hopefully that fixes the coverage CI. |
|
Is this change waiting on something? A CI run for ROCm would be very useful. |
|
Just waiting on a clean CI. My current plan is to disable the failing packages, but I think there may just need to point at the correct |
|
The failures are because the build is passing |
61e5cbf to
7d02070
Compare
|
Closing this for now. |
Add a pipeline that tests building the ROCm stack. There have been increasingly frequent issues keeping this stack stable as more packages depend on it, so it seems now is a good time to add this pipeline.
@srekolam @chuckatkins