Decompose aten.channel_shuffle op (#4243) #4259
Open
Conversation
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
…issue to handle this. For now filtering out failing tests.
…ffle operator was already included.
alaa-ali
reviewed
Jul 15, 2025
sahas3
requested changes
Jul 16, 2025
| // | ||
| // gets replaced with | ||
| // X = input.split_dim(...) # shape (N, g, C, *) | ||
| // X = X.permute(0, N+1, N, N+2, N+3) |
There was a problem hiding this comment.
I think here by N you actually mean dimN ? Can you update accordingly?
| @annotate_args( | ||
| [ | ||
| None, | ||
| ([1, 8, 4, 4], torch.float32, True), |
There was a problem hiding this comment.
Are dynamic dims input supported?
|
|
||
|
|
||
| @register_test_case(module_factory=lambda: ChannelShuffle1D()) | ||
| def ChannelShuffle1D_basic(module, tu: TestUtils): |
There was a problem hiding this comment.
this is not really 1D data, since it's being reshaped to 3d before passing to the channelShuffle op?
Comment on lines
+864
to
+865
| // CHECK-DAG: %[[PERMUTE:.*]] = torch.aten.permute %[[EXPAND]], %[[PERMLIST]] : !torch.vtensor<[1,4,2,4,4],f32>, !torch.list<int> -> !torch.vtensor<[1,2,4,4,4],f32> | ||
| // CHECK-DAG: %[[COLLAPSE:.*]] = torch.prims.collapse %[[PERMUTE]], %[[C1]], %[[C2]] : !torch.vtensor<[1,2,4,4,4],f32>, !torch.int, !torch.int -> !torch.vtensor<[1,8,4,4],f32> |
There was a problem hiding this comment.
Shouldn't all these be CHECK only since the order has to be maintained?
alaa-ali
reviewed
Jul 17, 2025
| // CHECK-DAG: %[[C1:.*]] = torch.constant.int 1 | ||
| // CHECK-DAG: %[[C3:.*]] = torch.constant.int 3 | ||
| // CHECK-DAG: %[[C4:.*]] = torch.constant.int 4 | ||
| // CHECK-DAG: %[[PERMLIST:.*]] = torch.prim.ListConstruct %[[C0]], %[[C2]], %[[C1]], %[[C3]], %[[C4]] : (!torch.int, !torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int> |
There was a problem hiding this comment.
You should be able to replace 'CHECK-DAG' by 'CHECK' starting from line 858 to the end.
| // CHECK-DAG: %[[EXPAND:.*]] = torch.prims.split_dim %[[ARG0]], %[[C1]], %[[C2]] : !torch.vtensor<[1,8,4,4],f32>, !torch.int, !torch.int -> !torch.vtensor<[1,4,2,4,4],f32> | ||
| // CHECK-DAG: %[[PERMUTE:.*]] = torch.aten.permute %[[EXPAND]], %[[PERMLIST]] : !torch.vtensor<[1,4,2,4,4],f32>, !torch.list<int> -> !torch.vtensor<[1,2,4,4,4],f32> | ||
| // CHECK-DAG: %[[COLLAPSE:.*]] = torch.prims.collapse %[[PERMUTE]], %[[C1]], %[[C2]] : !torch.vtensor<[1,2,4,4,4],f32>, !torch.int, !torch.int -> !torch.vtensor<[1,8,4,4],f32> | ||
| // return %[[COLLAPSE]] : !torch.vtensor<[1,8,4,4],f32> |
There was a problem hiding this comment.
missing // CHECK in line# 862?
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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.
Support for the channel shuffle operator is added by torch dialect level decomposition (similar to the pixel_shuffle operation).
The decomposition is based on this specification:
https://docs.pytorch.org/docs/stable/generated/torch.nn.ChannelShuffle.html
and implementation:
aten/src/ATen/native/ChanelShuffle.cpp
https://github.com/pytorch/pytorch/blob/23491519d288dedb2a54cfad5fef7fcb2ad8eade/aten/src/ATen/native/ChanelShuffle.cpp#L4
Note that the operator consists of an expansion, expanded channel dimensions permute, and contraction of channel dimensions back to the original size. For example, for an input array of shape 1x8x4x4 with a group size of 2 would generate the MLIR linalg code below.
module {
func.func @channel_shuffle(%arg0: !torch.vtensor<[1, 8, 4, 4], f32>) -> !torch.vtensor<[1, 8, 4, 4], f32> {
%c0 = torch.constant.int 0
%c1 = torch.constant.int 1
%c2 = torch.constant.int 2
%c3 = torch.constant.int 3
%c4 = torch.constant.int 4
%dims = torch.prim.ListConstruct %c0, %c2, %c1, %c3, %c4 : (!torch.int, !torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list
}
}
References:
PyTorch ChannelShuffle definition:
https://docs.pytorch.org/docs/stable/generated/torch.nn.ChannelShuffle.html
ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices (2017):
https://arxiv.org/pdf/1707.01083
A Lightweight Dendritic ShuffleNet for Medical Image Classification (2025)
https://www.jstage.jst.go.jp/article/transinf/advpub/0/advpub_2024EDP7059/_pdf
PyTorch implementation:
aten/src/ATen/native/ChanelShuffle.cpp
https://github.com/pytorch/pytorch/blob/23491519d288dedb2a54cfad5fef7fcb2ad8eade/aten/src/ATen/native/ChanelShuffle.cpp#L4
Resolves #4243
@newling @silvasean @rsuderman @zjgarvey @penguin-wwy @rafaelubalmw @sahas3 @vinitdeodhar @alaa-ali @dixinzhou @ramiro050 @qedawkins