[ROCM] Update bounds for large f16 data-tiling ukernel #22481
+30
−2
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
As part of trying to enable the tensor ukernels flag by default (#22318), it was discovered that some ukernels perform worse on some matmul benchmarks, resulting in regressions: https://github.com/iree-org/iree/actions/runs/18651161302/job/53169844154?pr=22318.
Through experimentation I found that the large f16 data-tiling ukernel only starts performing well from M==512, N==+-32832, K==+-512.
Benchmark Results (K=4096, N=16384)
Benchmark Results (K=4096, N=32768)
Benchmark Results (K=4096, N=65536)
I did a more granular sweep around these boundaries to show they're reasonable. As you can see, not using a ukernel performs better on smaller M dimensions. From the table it looks like around 384 would be better, but as you can see from above results, that's not always true, so I am being a bit more conservative here.
Comprehensive Matrix Sweep Results - M×K×N - All UKernels
All values in milliseconds
Matrix Sweep Results - M×K×N - No UKernels
All values in milliseconds