Suppress divby assume on tile-of-pointers from offset.#219
Merged
Conversation
For gather/scatter, `Intrinsics.offset(base_ptr, tile_offset)` produces a tile-of-pointers whose per-element divisor is `gcd(ptr_align, off_div * sizeof(elem))`. When that gcd collapses to the natural pointee alignment (e.g. 2 for f16), stamping it onto the IR as `assume div_by<2>` is vacuously true and actively harmful: tileiras's vectorizer trusts the explicit assume as the source of truth and skips its own offset-SSA walk, emitting `STG.E.U16` (2-byte stores) where the structurally-equivalent Python kernel gets `STG.E.128` (16-byte stores). 8x narrower stores, ~9x wall-time regression on the MoE scatter (121 GB/s vs Python's 1111 GB/s on the isolated MWE). Fix the divby transfer for `Intrinsics.offset` to mirror cuTile Python's `PointerOffset` rule: only propagate alignment when the result is 0-D (`Array.slice` style); for tile-shaped offsets, return 1 and let tileiras handle vectorization. MoE bench: 1.27 ms -> 0.95 ms (now 25% faster than Python). Other benchmarks unchanged.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
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Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be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.
For gather/scatter,
Intrinsics.offset(base_ptr, tile_offset)produces a tile-of-pointers whose per-element divisor isgcd(ptr_align, off_div * sizeof(elem)). When that gcd collapses to the natural pointee alignment (e.g. 2 for f16), stamping it onto the IR asassume div_by<2>is vacuously true and actively harmful: tileiras's vectorizer trusts the explicit assume as the source of truth and skips its own offset-SSA walk, emittingSTG.E.U16(2-byte stores) where the structurally-equivalent Python kernel getsSTG.E.128(16-byte stores). 8x narrower stores, ~9x wall-time regression on the MoE scatter (121 GB/s vs Python's 1111 GB/s on the isolated MWE).Fix the divby transfer for
Intrinsics.offsetto mirror cuTile Python'sPointerOffsetrule: only propagate alignment when the result is 0-D (Array.slicestyle); for tile-shaped offsets, return 1 and let tileiras handle vectorization. MoE bench: 1.27 ms -> 0.95 ms (now 25% faster than Python). Other benchmarks unchanged.