[WARP] Update warp LKG version to use 1.0.17 #671
Conversation
bob80905
added
the
test-all
test/WaveOps/WaveActiveMax.test
Outdated
| # not treat -INF as an identity. | ||
|
|
||
| # XFAIL: DirectX && WARP | ||
| # XFAIL: !x64 && DirectX && WARP |
There was a problem hiding this comment.
It is probably better to make this arm64\aarch64 (not sure what we call it) Since its the arm64 warp that is still failing. Doing Not x64 does not imply the same thing because there are still x86 (32bit) builds of warp and even though we aren't testing/using that one in CI we could cause failures for someone that is using.
https://nuget.info/packages/Microsoft.Direct3D.WARP/1.0.17
test/WaveOps/WaveActiveMax.test
Outdated
There was a problem hiding this comment.
I think this might need the GIS Target because it is failing the same way on Nvidia for both clang-dxc and dxc
|
|
||
| # XFAIL: DirectX && WARP | ||
| # XFAIL: arm64 && DirectX && WARP | ||
|
|
There was a problem hiding this comment.
Add an XFAIL for DirectX and Nvidia.
We are expecting [ inf, inf, inf, inf ] but Nvidia it is [ inf, 0, inf, inf ]. First try the GIS flag if that doesn't work we should look into filing a vendor bug.
There was a problem hiding this comment.
I think this test is completely broken.
For one, this line Mix[TID.x % 8] = WaveActiveMax(Mix[TID.x % 8]); looks like it's accessing out of bounds memory.
Secondly, @llvm-beanz has had a PR to fix this specific test for a couple months now, here: #347
I think we should just update the relevant XFAIL, so as not to regress further, but leave this one out of it, so that it can be properly addressed separately.
3 participants
This PR updates the warp version used by the offload test suite to the latest, 1.0.17, from 1.0.15.
Compared to 1.0.15, the latest version fixes bugs with reserved resources and other wave operations.
Fixes #670