MIOpenTensile Part2 : Support low precision types #556
Conversation
|
@atamazov @DrizztDoUrden @asroy @pfultz2 @JehandadKhan |
I thought we agreed that MIOpenTensile tests will be OFF by default. Could you please remind why do you need to always test "Latest"? |
|
@ce1adon I would like to merge this today; this question can be postponed for a day or two, I guess. But let me do some final fixes before merging -- that would be faster than discussions. |
@atamazov In our meeting, My original proposal was to disable both tensile stages. Paul @pfultz2 mentioned we have to keep one tensile test stage in the CI to justify other PRs. Here I leave the latest tensile version stage ON. |
…m the rest. Remove non-MIOpenTensile related changes.
# RESOLVED Conflicts: # Jenkinsfile
What if I add some smoke tests for Tensile and keep them always enabled, but disable all Full MIOpenTensile tests by default? That would save ~1.5-2 hours for each CI job. |
I've already disabled all the tests I can and I didn't even enable conv3d for gfx908 as I recommended in the meeting. You are welcome to prune the test into a new smoke test in a new PR later. For this one I just hope it can be wrapped. The longer we wait, the more merge conflicts will emerge. |
| @@ -188,11 +188,11 @@ pipeline { | |||
| description: "") | |||
| booleanParam( | |||
| name: "MIOPENTENSILE", | |||
| defaultValue: off, | |||
| defaultValue: false, | |||
There was a problem hiding this comment.
What's the point to test this PR for so long if both tensile stages are disable?
There was a problem hiding this comment.
Everyone can manually enable tests when needed:
There was a problem hiding this comment.
It's OK if you want to expand it to a new feature. But please aware that this PR is currently a blocker of many consecutive tasks and PRs. Please treat it as high priority and urgent as the tag says.
There was a problem hiding this comment.
I am ready to merge it as soon as CI testing done (if no problems). However, our CI is overloaded with other jobs, including unnecessary ones initiated by DevOps.
Ok, the "Latest" stage is enabled back. I'll prepare separate PR. |
In fact, I welcome any changes that improve the quality of testing. We just need to make such changes in separate PR. I will be glad if you begin (or even participate somehow in) this work. Thanks. |
@ce1adon ㊗️ |
The purpose of this PR is to add MIOpenTensile path in GEMM algorithm for low precision types on hip backend.
This PR depends on #139.
Set "on hold" till #139 merged.Update: The targetID env changes are in PR #674