Handle mix of structured and unstructured accesses in loops #290
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Outdated suggestions cannot be applied.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.
Previously, before the conversion to the PtrDialect for fallback, the structured-to-memref pass has to convert loop's iter-args with triton pointer to unranked memref. This conversion ensures all types coming out of triton-shared are mlir built-in types and therefore allows the CPU backend to correctly lower the IR to llvm. However, in reality, structured ops do not need to use the loop iter-args since ptr-analysis generates load/store ops that directly use the kernel arguments as source; this means the conversion is mostly unnecessary.
With the introduction of the fallback using the PtrDialect (triton-to-ptr), we also convert the loop iter-args of triton pointer type to PtrDialect's ptr type. This conversion, along with the conversion to unranked memref above, means we will end up with
unrealized_conversion_castops that convert back and forth between these two types when handling triton programs that have mixed uses of structured and unstructured accesses in loops.To solve this issue, we: