Closed bangtianliu closed 1 week ago
LGTM. I can see this improves performance on SDXL convs. Thanks for all the fixes.
Thanks for your comments and help, appreciate it very much!
Hi @bangtianliu, can you also update the spec in this PR: https://github.com/iree-org/iree/blob/main/build_tools/pkgci/external_test_suite/attention_and_matmul_spec.mlir? We need to keep it in sync with sdxl-scripts repo so we pick up the same perf boosts here as well
This pull request introduces the ability to reorder workgroups based on translation information specified in the configuration entries.
For instance, setting reorder_workgroups = "transpose" will enable the "transpose" method for workgroup reordering.
Notably, this pull request addresses a scenario that is somewhat the reverse of ISSUE #17340.
The "transpose" reordering method has been found to be faster than the natural order for convolution kernels.