Remove redundant depth-1 exception from objectFifo lowering#2986
Draft
Remove redundant depth-1 exception from objectFifo lowering#2986
Conversation
The `findObjectFifoSize()` function in AIEObjectFifoStatefulTransform had a special case that returned 1 buffer when both the declared scalar depth and the max acquire count were 1, suppressing the usual `maxAcquire + 1` ping-pong increment. This exception was redundant because the array-depth form (e.g. `[1, 1]`) already provides a way to explicitly enforce per-endpoint depths. Remove the exception so that a scalar depth-1 objectFifo consistently lowers to two buffers like any other scalar depth, and document that `[1, 1]` is the correct way to enforce a single buffer. Changes: - Remove the `maxAcquire == 1 && objFifo.size() == 1` guard from `findObjectFifoSize()` in AIEObjectFifoStatefulTransform.cpp - Update `depth_one_objectfifo_test.mlir` to use `[1, 1]` array depth so the test continues to verify single-buffer lowering via the explicit array-depth mechanism - Add `depth_one_scalar_test.mlir`: scalar depth 1 now lowers to 2 buffers (ping-pong) - Add `depth_one_array_test.mlir`: array depth `[1, 1]` still lowers to 1 buffer, confirming the enforcement mechanism works - Add `depth_two_scalar_test.mlir`: scalar depth 2 with max acquire 1 lowers to 2 buffers (not 3), confirming buffer count is driven by maxAcquire rather than the declared size - Update `01_single_double_buffer/` example designs to use depth 2 (ping-pong) and add a note that `[1, 1]` enforces a single buffer Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- depth_one_array/scalar and depth_two_scalar tests: shim tile consumers have no local memory, so no consumer buffers are created; update CHECKs to expect producer-side buffers only - repeat_count tests: scalar depth-1 now applies the prefetch increment, so switch objectfifos to array depth [1, 1] to preserve single-buffer semantics required by the repeat_count feature on compute tiles Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Contributor
Coverage ReportCreated: 2026-03-23 08:42Click here for information about interpreting this report.
Generated by llvm-cov -- llvm version 18.1.3 |
Three issues introduced after removing the depth-1 scalar exception:
1. Segfault in printObjectFifoInitValues when elemNumber is an ArrayAttr
(e.g. [1, 1]) and initValues is set: dyn_cast to IntegerAttr returned
null, crashing on .getInt(). Mirror the array-check already present
in parseObjectFifoInitValues.
2. Memory overflow in bottleneck/resnet ML examples: weight objectFifos
declared with scalar depth=1 now go through findObjectFifoSize() which
returns maxAcquire+1=2, doubling weight buffer allocations and
exceeding tile memory limits. Fix by using array-form depth [1, 1] so
findObjectFifoSize() is bypassed.
- IRON API: _get_depths() no longer collapses depth-1 to scalar,
so ObjectFifo(depth=1) emits [1, 1] MLIR array form automatically.
- Low-level API (_placed.py): explicitly pass [1, 1] / [1, 1, 1].
- MLIR: change 1 : i32 to [1, 1] / [1, 1, 1] in resnet aie.mlir.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
The rule is: - Scalar depth in IRON emits scalar depth in MLIR, letting findObjectFifoSize() run during lowering (so depth=1 correctly becomes double-buffered, depth=2 stays 2, etc.) - List depth in IRON emits array depth in MLIR, bypassing findObjectFifoSize() so the declared depths are used as-is. Implement this by: - Accepting list[int] for ObjectFifo(depth=...), storing an _array_depth flag and normalising _depth to a scalar for per-handle bounds checks. - Restoring _get_depths() to collapse uniform scalar depths to a single int; only returning a list when _array_depth is set or depths differ across endpoints. - Widening split(depths=...) to accept list[int | list[int]]. Update bottleneck.py and resnet.py weight fifos to depth=[1, 1] so they stay single-buffered rather than being double-buffered by findObjectFifoSize() after the depth-1 exception removal. Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
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.
The
findObjectFifoSize()function in AIEObjectFifoStatefulTransform had a special case that returned 1 buffer when both the declared scalar depth and the max acquire count were 1, suppressing the usualmaxAcquire + 1ping-pong increment. This exception was redundant because the array-depth form (e.g.[1, 1]) already provides a way to explicitly enforce per-endpoint depths. Remove the exception so that a scalar depth-1 objectFifo consistently lowers to two buffers like any other scalar depth, and document that[1, 1]is the correct way to enforce a single buffer.Changes:
maxAcquire == 1 && objFifo.size() == 1guard fromfindObjectFifoSize()in AIEObjectFifoStatefulTransform.cppdepth_one_objectfifo_test.mlirto use[1, 1]array depth so the test continues to verify single-buffer lowering via the explicit array-depth mechanismdepth_one_scalar_test.mlir: scalar depth 1 now lowers to 2 buffers (ping-pong)depth_one_array_test.mlir: array depth[1, 1]still lowers to 1 buffer, confirming the enforcement mechanism worksdepth_two_scalar_test.mlir: scalar depth 2 with max acquire 1 lowers to 2 buffers (not 3), confirming buffer count is driven by maxAcquire rather than the declared size01_single_double_buffer/example designs to use depth 2 (ping-pong) and add a note that[1, 1]enforces a single buffer