[mlir][linalg] Add a test for inferConvolutionDimsImpl #90057

banach-space · 2024-04-25T14:04:10Z

Adds a test for inferConvolutionDimsImpl to exercise the logic for depthwise convs.

ATM, `inferConvolutionDimsImpl` will "remove" "unconvolved" dims from the calculation of the channel dims. However, that's incorrect for depthwise convolutions for which the channel dimension falls into that group (i.e. "unconvolved" dims).

llvmbot · 2024-04-25T14:04:45Z

@llvm/pr-subscribers-mlir-linalg

Author: Andrzej Warzyński (banach-space)

Changes

ATM, inferConvolutionDimsImpl will "remove" "unconvolved" dims from the
calculation of the channel dims. However, that's incorrect for depthwise
convolutions for which the channel dimension falls into that group (i.e.
"unconvolved" dims).

Full diff: https://github.com/llvm/llvm-project/pull/90057.diff

2 Files Affected:

(modified) mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp (-1)
(modified) mlir/test/Dialect/Linalg/match-ops-interpreter.mlir (+22)

diff --git a/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp b/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp
index 3627ff6617eda3..3b92da5ceccd39 100644
--- a/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp
+++ b/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp
@@ -556,7 +556,6 @@ inferConvolutionDimsImpl(LinalgOp linalgOp,
   // filterDims & outputDims - unConvolvedDims are the output channel iterators.
   llvm::SmallDenseSet<int64_t> oc = filterDims;
   llvm::set_intersect(oc, outputDims);
-  llvm::set_subtract(oc, inputExprWalker.unConvolvedDims);
 
   // filterDims & outputDims & unConvolvedDims are the depth iterators.
   llvm::SmallDenseSet<int64_t> depth = filterDims;
diff --git a/mlir/test/Dialect/Linalg/match-ops-interpreter.mlir b/mlir/test/Dialect/Linalg/match-ops-interpreter.mlir
index 24c7bdd9e1050e..c637e1df7efd3e 100644
--- a/mlir/test/Dialect/Linalg/match-ops-interpreter.mlir
+++ b/mlir/test/Dialect/Linalg/match-ops-interpreter.mlir
@@ -1062,6 +1062,28 @@ module attributes { transform.target_tag = "start_here" } {
     return %result : tensor<10x18x15xf64>
   }
 
+  func.func @convolution_depthwise(%input: tensor<1x10x196x48xf32>, %filter: tensor<1x4x48xf32>) -> tensor<1x10x191x48xf32> {
+    %cst = arith.constant 0.0 : f32 
+    %empty = tensor.empty() : tensor<1x10x191x48xf32>
+    %fill = linalg.fill ins(%cst : f32) outs(%empty : tensor<1x10x191x48xf32>) -> tensor<1x10x191x48xf32>
+    // expected-remark @below {{convolution}}
+    // expected-remark @below {{batch dims 0}}
+    // expected-remark @below {{output image dims 1 : i64, 2 : i64}}
+    // expected-remark @below {{output channel dims 3}}
+    // expected-remark @below {{filter loop dims 4 : i64, 5 : i64}}
+    // expected-remark @below {{input channel dims}}
+    // expected-remark @below {{depth dims 3}}
+    // expected-remark @below {{strides 1 : i64, 1 : i64}}
+    // expected-remark @below {{dilations 1 : i64, 1 : i64}}
+    %result = linalg.depthwise_conv_2d_nhwc_hwc {
+      dilations = dense<1> : tensor<2xi64>,
+      strides = dense<1> : tensor<2xi64>}
+      ins(%input, %filter : tensor<1x10x196x48xf32>, tensor<1x4x48xf32>)
+      outs(%fill : tensor<1x10x191x48xf32>) -> tensor<1x10x191x48xf32>
+
+    return %result : tensor<1x10x191x48xf32>
+  }
+
   func.func @convolution_multi_channel(%input: tensor<2x34x68x16xf32>, %filter: tensor<8x2x3x5x16x16xf32>) -> tensor<8x32x32x16xf32> {
     %cst = arith.constant 0.0 : f32
     %empty = tensor.empty() : tensor<8x32x32x16xf32>

llvmbot · 2024-04-25T14:04:45Z

@llvm/pr-subscribers-mlir

Author: Andrzej Warzyński (banach-space)

Changes

ATM, inferConvolutionDimsImpl will "remove" "unconvolved" dims from the
calculation of the channel dims. However, that's incorrect for depthwise
convolutions for which the channel dimension falls into that group (i.e.
"unconvolved" dims).

Full diff: https://github.com/llvm/llvm-project/pull/90057.diff

2 Files Affected:

(modified) mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp (-1)
(modified) mlir/test/Dialect/Linalg/match-ops-interpreter.mlir (+22)

diff --git a/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp b/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp
index 3627ff6617eda3..3b92da5ceccd39 100644
--- a/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp
+++ b/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp
@@ -556,7 +556,6 @@ inferConvolutionDimsImpl(LinalgOp linalgOp,
   // filterDims & outputDims - unConvolvedDims are the output channel iterators.
   llvm::SmallDenseSet<int64_t> oc = filterDims;
   llvm::set_intersect(oc, outputDims);
-  llvm::set_subtract(oc, inputExprWalker.unConvolvedDims);
 
   // filterDims & outputDims & unConvolvedDims are the depth iterators.
   llvm::SmallDenseSet<int64_t> depth = filterDims;
diff --git a/mlir/test/Dialect/Linalg/match-ops-interpreter.mlir b/mlir/test/Dialect/Linalg/match-ops-interpreter.mlir
index 24c7bdd9e1050e..c637e1df7efd3e 100644
--- a/mlir/test/Dialect/Linalg/match-ops-interpreter.mlir
+++ b/mlir/test/Dialect/Linalg/match-ops-interpreter.mlir
@@ -1062,6 +1062,28 @@ module attributes { transform.target_tag = "start_here" } {
     return %result : tensor<10x18x15xf64>
   }
 
+  func.func @convolution_depthwise(%input: tensor<1x10x196x48xf32>, %filter: tensor<1x4x48xf32>) -> tensor<1x10x191x48xf32> {
+    %cst = arith.constant 0.0 : f32 
+    %empty = tensor.empty() : tensor<1x10x191x48xf32>
+    %fill = linalg.fill ins(%cst : f32) outs(%empty : tensor<1x10x191x48xf32>) -> tensor<1x10x191x48xf32>
+    // expected-remark @below {{convolution}}
+    // expected-remark @below {{batch dims 0}}
+    // expected-remark @below {{output image dims 1 : i64, 2 : i64}}
+    // expected-remark @below {{output channel dims 3}}
+    // expected-remark @below {{filter loop dims 4 : i64, 5 : i64}}
+    // expected-remark @below {{input channel dims}}
+    // expected-remark @below {{depth dims 3}}
+    // expected-remark @below {{strides 1 : i64, 1 : i64}}
+    // expected-remark @below {{dilations 1 : i64, 1 : i64}}
+    %result = linalg.depthwise_conv_2d_nhwc_hwc {
+      dilations = dense<1> : tensor<2xi64>,
+      strides = dense<1> : tensor<2xi64>}
+      ins(%input, %filter : tensor<1x10x196x48xf32>, tensor<1x4x48xf32>)
+      outs(%fill : tensor<1x10x191x48xf32>) -> tensor<1x10x191x48xf32>
+
+    return %result : tensor<1x10x191x48xf32>
+  }
+
   func.func @convolution_multi_channel(%input: tensor<2x34x68x16xf32>, %filter: tensor<8x2x3x5x16x16xf32>) -> tensor<8x32x32x16xf32> {
     %cst = arith.constant 0.0 : f32
     %empty = tensor.empty() : tensor<8x32x32x16xf32>

banach-space · 2024-04-25T14:06:08Z

@qedawkins I've not worked much with non-depthwise convs and inferConvolutionDimsImpl is quite ... convolved 😅 So, please could you double check this and let me know if I missed sth?

qedawkins

Yeah simplifications to the code would be welcome. There is a depth dim classification for this though, maybe would it make more sense if we renamed it to depthChannel? Also I'm not necessarily opposed to a change in classification, I just wrote the current impl to be more in line with the inferContractionDims impl and I haven't worked much with depthwise convs.

qedawkins · 2024-04-25T14:37:26Z

mlir/test/Dialect/Linalg/match-ops-interpreter.mlir

+    // expected-remark @below {{output channel dims 3}}
+    // expected-remark @below {{filter loop dims 4 : i64, 5 : i64}}
+    // expected-remark @below {{input channel dims}}
+    // expected-remark @below {{depth dims 3}}


This doesn't look like what I would expect because it is classifying dim 3 twice. The naming might need to be improved for depthwise convolutions, but the idea with these classifications is that they tell users exactly how each different dimension appears in the indexing maps (and thus input/filter/output tensors). Right now the impl uses this to imply that all dims classified as outputChannel are only featured in the filter and output maps. Dims classified as depth appear in all three of the input, filter, and output.

I'm noticing that there is no description for depth in this comment though:

llvm-project/mlir/lib/Dialect/Linalg/IR/LinalgInterfaces.cpp

Lines 622 to 645 in 76a3be7

/// Find at least 1 parallel (output_image) and reduction (filter_loop)

/// dimension candidates that form a convolution subcomputation within

/// `linalgOp`. The LHS is assumed to be the convolution input while the

/// RHS is assumed as the filter.

/// These dimensions are such that:

/// 1. Optional batch dimensions that appear in the input and filter.

/// 2. The output_image dimension is involved in a cross-correlation along LHS

/// (i.e. it is a permutation on RES and LHS and has an associated

/// filter_loop in RHS).

/// 3. Optional output_channel dimension is involved in an outer-product along

/// RHS (i.e. it is a permutation on RES and RHS and does not appear in

/// LHS).

/// 4. Optional input_channel dimension appears as a permutation on LHS and

/// RHS.

/// 5. The filter_loop dimension appears as a permutation on the RHS and

/// represents the shape of the kernel cross-correlated along a

/// corresponding output_image dim.

/// 6. The input_channel dimension appears as a permutation on LHS and RHS.

/// 7. All dimensions appear only once in any given indexing map.

/// This allows e.g. detecting that some convolution is embedded within

/// `linalgOp` with some orthogonal heuristic.

/// When multiple dimension occurrences exist that match any classification

/// indices are returned in sorted order.

/// Returns a failure if `output_image` (and implicitly `filter_loop`) is empty.

(also the comment is kind of convoluted and could probably be rewritten, maybe with an example).

Thanks for pointing this out - now I see what's happening here. IIUC, "channel" is only used for non-depthwise convolutions. For depthwise convs (HWC), this logic uses the term "depth" instead. Now, "depth" is:

a parallel dim

identical for all inputs.

It looks like this PR is breaking rather than fixing things 😂 Let me update and rephrase.

Revert

banach-space · 2024-04-25T16:59:23Z

Thanks for taking a look, Quinn! I've rephrased this patch - adding a test for depthwise convs should be sufficient.

Also I'm not necessarily opposed to a change in classification, I just wrote the current impl to be more in line with the inferContractionDims impl and I haven't worked much with depthwise convs.

Your implementation is perfectly fine. I think that what we are really missing is some high-level reference doc that would classify all the convs in Linalg (and normalize the terminology). I don't have the right overview as I've been mostly working with depthwise convs. But this is on my TODO list and I would also be more than happy to review PRs for this. We'll get there eventually!

qedawkins

Nice, thanks for pushing on this! Tests and documentation improvements are always welcome :)

banach-space requested a review from qedawkins April 25, 2024 14:04

llvmbot added mlir:linalg mlir labels Apr 25, 2024

banach-space mentioned this pull request Apr 25, 2024

[CPU] Add scalable flags to configs for Conv iree-org/iree#17101

Merged

qedawkins reviewed Apr 25, 2024

View reviewed changes

fixup! [mlir][linalg] Fix inferConvolutionDimsImpl (depthwise convs)

825b49a

Revert

banach-space changed the title ~~[mlir][linalg] Fix inferConvolutionDimsImpl (depthwise convs)~~ [mlir][linalg] Add a test for inferConvolutionDimsImpl Apr 25, 2024

qedawkins approved these changes Apr 25, 2024

View reviewed changes

banach-space merged commit 3005ca2 into llvm:main Apr 25, 2024

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

[mlir][linalg] Add a test for inferConvolutionDimsImpl #90057

[mlir][linalg] Add a test for inferConvolutionDimsImpl #90057

Uh oh!

banach-space commented Apr 25, 2024 •

edited

Loading

Uh oh!

llvmbot commented Apr 25, 2024

Uh oh!

llvmbot commented Apr 25, 2024

Uh oh!

banach-space commented Apr 25, 2024

Uh oh!

qedawkins left a comment

Uh oh!

qedawkins Apr 25, 2024 •

edited

Loading

Uh oh!

banach-space Apr 25, 2024

Uh oh!

banach-space commented Apr 25, 2024

Uh oh!

qedawkins left a comment •

edited

Loading

Uh oh!

Uh oh!

	/// Find at least 1 parallel (output_image) and reduction (filter_loop)
	/// dimension candidates that form a convolution subcomputation within
	/// `linalgOp`. The LHS is assumed to be the convolution input while the
	/// RHS is assumed as the filter.
	/// These dimensions are such that:
	/// 1. Optional batch dimensions that appear in the input and filter.
	/// 2. The output_image dimension is involved in a cross-correlation along LHS
	/// (i.e. it is a permutation on RES and LHS and has an associated
	/// filter_loop in RHS).
	/// 3. Optional output_channel dimension is involved in an outer-product along
	/// RHS (i.e. it is a permutation on RES and RHS and does not appear in
	/// LHS).
	/// 4. Optional input_channel dimension appears as a permutation on LHS and
	/// RHS.
	/// 5. The filter_loop dimension appears as a permutation on the RHS and
	/// represents the shape of the kernel cross-correlated along a
	/// corresponding output_image dim.
	/// 6. The input_channel dimension appears as a permutation on LHS and RHS.
	/// 7. All dimensions appear only once in any given indexing map.
	/// This allows e.g. detecting that some convolution is embedded within
	/// `linalgOp` with some orthogonal heuristic.
	/// When multiple dimension occurrences exist that match any classification
	/// indices are returned in sorted order.
	/// Returns a failure if `output_image` (and implicitly `filter_loop`) is empty.

[mlir][linalg] Add a test for inferConvolutionDimsImpl #90057

[mlir][linalg] Add a test for inferConvolutionDimsImpl #90057

Uh oh!

Conversation

banach-space commented Apr 25, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

llvmbot commented Apr 25, 2024

Uh oh!

llvmbot commented Apr 25, 2024

Uh oh!

banach-space commented Apr 25, 2024

Uh oh!

qedawkins left a comment

Choose a reason for hiding this comment

Uh oh!

qedawkins Apr 25, 2024 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

banach-space Apr 25, 2024

Choose a reason for hiding this comment

Uh oh!

banach-space commented Apr 25, 2024

Uh oh!

qedawkins left a comment • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Uh oh!

banach-space commented Apr 25, 2024 •

edited

Loading

qedawkins Apr 25, 2024 •

edited

Loading

qedawkins left a comment •

edited

Loading