Allow to move collapse/expand_shape through linalg.generic if its a projected permuatation with zeros #387
Conversation
This file contains 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
```
module attributes {
llvm.data_layout = "e-m:e-p270:32:32-p271:32:32-p272:64:64-i64:64-i128:128-f80:128-n8:16:32:64-S128",
llvm.target_triple = "x86_64-unknown-linux-gnu"} {
func.func @forward(%arg0: tensor<1x255x40x40xi8> {func.orig_type = tensor<1x255x40x40xi8>, onnx.name = "inp3in"} loc(unknown), %arg1: tensor<1x255x20x20xi8> {func.orig_type = tensor<1x255x20x20xi8>, onnx.name = "inp2in"} loc(unknown), %arg2: tensor<1x255x80x80xi8> {func.orig_type = tensor<1x255x80x80xi8>, onnx.name = "inp1in"} loc(unknown)) -> (tensor<1x3x40x40x85xf32> {func.orig_type = tensor<1x3x40x40x85xf32>, onnx.name = "onnx::Sigmoid_564"}) {
%cst = arith.constant dense<3.100000e+00> : tensor<1x1x1x1xf32> loc(#loc)
%0 = tensor.empty() : tensor<1x255x40x40xf32> loc(#loc1)
%1 = linalg.generic {indexing_maps = [#map, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg0 : tensor<1x255x40x40xi8>) outs(%0 : tensor<1x255x40x40xf32>) {
^bb0(%in: i8 loc(unknown), %out: f32 loc("")):
%5 = arith.sitofp %in : i8 to f32 loc(#loc1)
linalg.yield %5 : f32 loc(#loc1)
} -> tensor<1x255x40x40xf32> loc(#loc1)
%2 = tensor.empty() : tensor<1x255x40x40xf32> loc(#loc1)
%3 = linalg.generic {indexing_maps = [#map, #map2, #map1], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%1, %cst : tensor<1x255x40x40xf32>, tensor<1x1x1x1xf32>) outs(%2 : tensor<1x255x40x40xf32>) {
^bb0(%in: f32 loc(""), %in_0: f32 loc(unknown), %out: f32 loc("")):
%5 = arith.mulf %in, %in_0 : f32 loc(#loc1)
linalg.yield %5 : f32 loc(#loc1)
} -> tensor<1x255x40x40xf32> loc(#loc1)
%expanded = tensor.expand_shape %3 [[0], [1, 2], [3], [4]] output_shape [1, 3, 85, 40, 40] : tensor<1x255x40x40xf32> into tensor<1x3x85x40x40xf32> loc(#loc2)
%4 = tensor.empty() : tensor<1x3x40x40x85xf32> loc(#loc3)
%transposed = linalg.transpose ins(%expanded : tensor<1x3x85x40x40xf32>) outs(%4 : tensor<1x3x40x40x85xf32>) permutation = [0, 1, 3, 4, 2] loc(#loc3)
return %transposed : tensor<1x3x40x40x85xf32> loc(#loc)
} loc(#loc)
} loc(#loc)
```
would not move the tensor.expand_shape to the top because the check `isProjectedPermutation` on `#map`
would fail. It passes if changing `#map` to the equivalent `affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>` (considering that the tensor size is 1 in that dimension).
From my understanding, the transformation doesn't strictly require latter map,
and there is already an option in isProjectedPermutation to allow zeros.
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.
This PR still crashes.
would not move the tensor.expand_shape to the top because the check
isProjectedPermutationon#mapwould fail. It passes if changing#mapto the equivalentaffine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>(considering that the tensor size is 1 in that dimension). From my understanding, the transformation doesn't strictly require latter map, and there is already an option in isProjectedPermutation to allow zeros.