Summary: [mlir][tosa] Move all TOSA Operations Decomposition Patterns into populatePatterns() functions, that can be used by external rewrite engines

[mehdi_amini]

LG but please run clang-format.

mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeTransposeConv.cpp

376–379

Please install clang-format on your machine, arc will run it when you upload patches.
I also recommend setting up the pre-push hook: https://llvm.org/docs/GettingStarted.html#git-pre-push-hook

[aardeb]

Apply clang-format to: mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeTransposeConv.cpp

[aardeb]

mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeTransposeConv.cpp

376–379

I have no idea how this got through with arc as I had clang-format on my machine. Will pay more attention next time.

[aardeb]

mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeTransposeConv.cpp

39

I am using clang-format version 10.0.0-4ubuntu1~18.04.2. I'm not its why these are showing up.

[aardeb]

Fix Lint Pre-merge checks

[mehdi_amini]

mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeTransposeConv.cpp

39

In general we're using the last LLVM release, but clang-10 should catch these :)

[rsuderman]

Please update TosaToLinalg lowerings to include transpose decomposition:

https://github.com/llvm/llvm-project/blob/main/mlir/lib/Conversion/TosaToLinalg/TosaToLinalgPass.cpp#L70

mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeConv2D.cpp

2

This line is incorrect, please update.

10

Please detail what it is replaced by.

21

Please remove unused dependencies. I am quite certain BlockAndValueMapping is not being used.

mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeDepthwise.cpp

1

Please fix. Also, rename file something more specific. TosaDepthwiseToMul would be more apt.

9

Should include what the pattern decomposes into.

mlir/test/lib/Dialect/Tosa/TosaTestDecomposePass.cpp

10

Detail more than helper functions. Overly vague.

[aardeb]

Fix Comments and Update TosatoLinalg Conversion

[aardeb]

mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeDepthwise.cpp

1

I am not 100% sure about renaming. Creating a file per specific rewrite seems unscalable. This file should be the location we place where we put any Depth wise decomposition rewrites. Just like the TransposeConv has multiple rewrites in the future we may have multiple rewrites here.

[aardeb]

Fix Linting Issue

[rsuderman]

Some minor changes on my end. Waiting on @mehdi_amini to review.

mlir/include/mlir/Conversion/Passes.td

652

Rewrites is superfluous as the majority of passes rewrite IR. I suggest something such as tosa-optional-decompositions.

mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeDepthwise.cpp

1

I agree this is an issue. I would have preferred having one TosaOptionalRewriters file that have groupings of useful rewriters. In this case, the file name and comment header did not match, which was the primary issue.

mlir/test/lib/Dialect/Tosa/TosaTestDecomposePass.cpp

2

Should match filename.

[mehdi_amini]

It's not clear to me why this is in Conversion when it seems to me a transformation that operates TOSA->TOSA?
Conversion/XXX directories are really intended to be trimmed to the conversion code that spans two dialects (or more).

[aardeb]

Fix Naming issues

[aardeb]

Fix Spelling mistake

[aardeb]

Fix comment

[aardeb]

In D116526#3224212, @mehdi_amini wrote:

It's not clear to me why this is in Conversion when it seems to me a transformation that operates TOSA->TOSA?
Conversion/XXX directories are really intended to be trimmed to the conversion code that spans two dialects (or more).

I believe we agreed that we would expose these optional rewrites as populate() functions within the TOSA Dialect folder and then allow the consumer to add them to the their rewrite engines. Otherwise we would need to create a pass in the TOSA folder which would take us back to the original solution this patch is trying to refactor.

[mehdi_amini]

In D116526#3224896, @aardeb wrote:

In D116526#3224212, @mehdi_amini wrote:

It's not clear to me why this is in Conversion when it seems to me a transformation that operates TOSA->TOSA?
Conversion/XXX directories are really intended to be trimmed to the conversion code that spans two dialects (or more).

I believe we agreed that we would expose these optional rewrites as populate() functions within the TOSA Dialect folder and then allow the consumer to add them to the their rewrite engines. Otherwise we would need to create a pass in the TOSA folder which would take us back to the original solution this patch is trying to refactor.

What I'm saying is that I'm not sure why the TosaOptionalDecompositions pass is not just defined in the Transforms directory?

Also, why do you have TosaTestDecompose if you need anyway a public TosaOptionalDecompositions pass? I can't spot the difference between the two right now.

[aardeb]

Remove multiple similar passes

[aardeb]

Fix Cmake File

[aardeb]

Remove redundant new line

[aardeb]

In D116526#3228931, @mehdi_amini wrote:

In D116526#3224896, @aardeb wrote:

In D116526#3224212, @mehdi_amini wrote:

It's not clear to me why this is in Conversion when it seems to me a transformation that operates TOSA->TOSA?
Conversion/XXX directories are really intended to be trimmed to the conversion code that spans two dialects (or more).

I believe we agreed that we would expose these optional rewrites as populate() functions within the TOSA Dialect folder and then allow the consumer to add them to the their rewrite engines. Otherwise we would need to create a pass in the TOSA folder which would take us back to the original solution this patch is trying to refactor.

What I'm saying is that I'm not sure why the TosaOptionalDecompositions pass is not just defined in the Transforms directory?

Also, why do you have TosaTestDecompose if you need anyway a public TosaOptionalDecompositions pass? I can't spot the difference between the two right now.

I understand your point we have ended with similar passes in the test/ and linalg/ folders. Given that I have reverted to a single pass in the tosa/transforms folder but with naming and organisation fixed as previously discussed.

[mehdi_amini]

LG, thanks!

mlir/include/mlir/Conversion/Passes.td

let description = [{
     Pass that converts TOSA operations to the equivalent operations using the
     tensor operations in LinAlg.
   }];
 
   let constructor = "tosa::createTosaToLinalg()";
 }
 
 //===----------------------------------------------------------------------===//
+// TosaOptionalDecompositions
+//===----------------------------------------------------------------------===//
+
+def TosaOptionalDecompositions : FunctionPass<"tosa-optional-decompositions"> {
+  let summary = "Applies Tosa opeerations optional decompositions";
+  let description = [{
+    Pass to apply the Tosa operations decomposition 
+    exposed as populate functions in include/mlir/Dialect/Tosa/Transforms/Passes.h
+  }];
+
+  let constructor = "tosa::createTosaOptionalDecompositions()";
+}
+
+//===----------------------------------------------------------------------===//
 // TosaToLinalgNamed
 //===----------------------------------------------------------------------===//
 
 def TosaToLinalgNamed : FunctionPass<"tosa-to-linalg-named"> {
   let summary = "Lower TOSA to LinAlg named operations";
   let description = [{
     Pass that converts TOSA operations to the equivalent operations using the
     Linalg named operations.

mlir/include/mlir/Conversion/TosaToLinalg/TosaToLinalg.h

 
 #include "mlir/Pass/Pass.h"
 
 namespace mlir {
 namespace tosa {
 
 std::unique_ptr<Pass> createTosaToLinalg();
 std::unique_ptr<Pass> createTosaToLinalgNamed();
+std::unique_ptr<Pass> createTosaOptionalDecompositions();
 
 /// Populates passes to convert from TOSA to Linalg on buffers. At the end of
 /// the pass, the function will only contain linalg ops or standard ops if the
 /// pipeline succeeds.
 void addTosaToLinalgPasses(OpPassManager &pm);
 
 /// Populates conversion passes from TOSA dialect to Linalg dialect.
 void populateTosaToLinalgConversionPatterns(RewritePatternSet *patterns);
 
 /// Populates conversion passes from TOSA dialect to Linalg named operations.
 void populateTosaToLinalgNamedConversionPatterns(RewritePatternSet *patterns);
 
 } // namespace tosa
 } // namespace mlir
 
 #endif // MLIR_CONVERSION_TOSATOLINALG_TOSATOLINALG_H

mlir/include/mlir/Dialect/Tosa/Transforms/Passes.h

 #define MLIR_DIALECT_TOSA_TRANSFORMS_PASSES_H
 
 #include "mlir/Dialect/Tensor/IR/Tensor.h"
 #include "mlir/Pass/Pass.h"
 
 namespace mlir {
 namespace tosa {
 
-std::unique_ptr<Pass> createTosaDecomposeTransposeConvPass();
+// Expose Rewrite Functions that decompose TOSA Ops into further TOSA Ops.
+// The rewrites can be selectively added to a conversion pass.
+void populateTosaDecomposeConv2D(MLIRContext *ctx, RewritePatternSet &patterns);
+void populateTosaDecomposeTransposeConv(MLIRContext *ctx,
+                                        RewritePatternSet &patterns);
+void populateTosaDecomposeDepthwise(MLIRContext *ctx,
+                                    RewritePatternSet &patterns);
+
 std::unique_ptr<Pass> createTosaInferShapesPass();
 std::unique_ptr<Pass> createTosaMakeBroadcastablePass();
-std::unique_ptr<Pass> createTosaOptimizationPass();
 std::unique_ptr<Pass> createTosaTestQuantUtilAPIPass();
 
 #define GEN_PASS_REGISTRATION
 #include "mlir/Dialect/Tosa/Transforms/Passes.h.inc"
 
 } // namespace tosa
 } // namespace mlir
 
 #endif // MLIR_DIALECT_TOSA_TRANSFORMS_PASSES_H

mlir/include/mlir/Dialect/Tosa/Transforms/Passes.td

 //
 //===----------------------------------------------------------------------===//
 
 #ifndef MLIR_DIALECT_TOSA_TRANSFORMS_PASSES
 #define MLIR_DIALECT_TOSA_TRANSFORMS_PASSES
 
 include "mlir/Pass/PassBase.td"
 
-def TosaDecomposeTransposeConv : FunctionPass<"tosa-decompose-transpose-conv"> {
-  let summary = "Deompose transpose convolutiions into standard convolutions.";
-  let description = [{
-    Pass that uses shape manipulation and convolution operations to transform
-    a transpose convolution into a regular convolution.
-  }];
-
-  let constructor = "createTosaDecomposeTransposeConvPass()";
-  let dependentDialects = [
-    "StandardOpsDialect",
-    "tensor::TensorDialect",
-    "tosa::TosaDialect",
-  ];
-}
-
 def TosaInferShapes : FunctionPass<"tosa-infer-shapes"> {
   let summary = "Propagate shapes across TOSA operations";
   let description = [{
     Pass that uses operand types and propagates shapes to TOSA operations.
     This includes legalizing rankless and dynamic shapes towards static.
   }];
 
   let constructor = "createTosaInferShapesPass()";
let description = [{
     of size one until the number of dimensions is equal. Implements
     approach similar to step 1 of Numpy 4-step broadcasting:
     https://numpy.org/doc/stable/reference/ufuncs.html#broadcasting
   }];
 
   let constructor = "createTosaMakeBroadcastablePass()";
 }
 
-def TosaOptimization : FunctionPass<"tosa-optimization"> {
-  let summary = "TOSA operation optimizations";
-  let description = [{
-    "Pass to perform optimizations on TOSA operations"
-  }];
-
-  let constructor = "createTosaOptimizationPass()";
-}
-
 #endif // MLIR_DIALECT_TOSA_TRANSFORMS_PASSES

mlir/lib/Conversion/TosaToLinalg/TosaToLinalgPass.cpp

void runOnFunction() override {
     target.markUnknownOpDynamicallyLegal([](Operation *) { return true; });
 
     FuncOp func = getFunction();
     mlir::tosa::populateTosaToLinalgConversionPatterns(&patterns);
     if (failed(applyFullConversion(func, target, std::move(patterns))))
       signalPassFailure();
   }
 };
+
+struct TosaOptionalDecompositions
+    : public TosaOptionalDecompositionsBase<TosaOptionalDecompositions> {
+  void runOnFunction() {
+    auto *ctx = &getContext();
+    RewritePatternSet patterns(ctx);
+    auto func = getFunction();
+
+    mlir::tosa::populateTosaDecomposeConv2D(ctx, patterns);
+    mlir::tosa::populateTosaDecomposeTransposeConv(ctx, patterns);
+    mlir::tosa::populateTosaDecomposeDepthwise(ctx, patterns);
+
+    if (applyPatternsAndFoldGreedily(func, std::move(patterns)).failed())
+      signalPassFailure();
+  }
+};
+
 } // namespace
 
 std::unique_ptr<Pass> mlir::tosa::createTosaToLinalg() {
   return std::make_unique<TosaToLinalg>();
 }
 
+std::unique_ptr<Pass> mlir::tosa::createTosaOptionalDecompositions() {
+  return std::make_unique<TosaOptionalDecompositions>();
+}
+
 void mlir::tosa::addTosaToLinalgPasses(OpPassManager &pm) {
+  pm.addNestedPass<FuncOp>(createTosaOptionalDecompositions());
   pm.addNestedPass<FuncOp>(createTosaMakeBroadcastablePass());
   pm.addNestedPass<FuncOp>(createTosaToLinalgNamed());
   pm.addNestedPass<FuncOp>(mlir::createCanonicalizerPass());
   pm.addNestedPass<FuncOp>(createTosaMakeBroadcastablePass());
   pm.addNestedPass<FuncOp>(createTosaToLinalg());
 }

mlir/lib/Dialect/Tosa/Transforms/CMakeLists.txt

 add_mlir_dialect_library(MLIRTosaTransforms
   TosaDecomposeTransposeConv.cpp
+  TosaDecomposeConv2D.cpp
+  TosaDecomposeDepthwise.cpp
   TosaInferShapes.cpp
   TosaMakeBroadcastable.cpp
-  TosaOptimization.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/Tosa/Transforms
 
   DEPENDS
   MLIRTosaPassIncGen
 
   LINK_LIBS PUBLIC
   MLIRPass
   MLIRTosa
   )

mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeConv2D.cpp

+//===- TosaDecomposeConv2D.cpp ------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Decompose TOSA Conv2D operation to a series of TOSA Ops specifically
+// (1) Convert a 1x1 Convolution to a Reshape->FC->Reshape
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Tosa/IR/TosaOps.h"
+#include "mlir/Dialect/Tosa/Transforms/Passes.h"
+#include "mlir/Pass/Pass.h"
+
+using namespace mlir;
+using namespace mlir::tosa;
+
+namespace {
+
+struct Conv2DIsFullyConnected : public OpRewritePattern<tosa::Conv2DOp> {
+  explicit Conv2DIsFullyConnected(MLIRContext *context)
+      : OpRewritePattern(context) {}
+
+  LogicalResult matchAndRewrite(tosa::Conv2DOp op,
+                                PatternRewriter &rewriter) const override {
+    Value input = op.input();
+    Value weight = op.weight();
+    ShapedType inputType = input.getType().cast<ShapedType>();
+    ShapedType weightType = weight.getType().cast<ShapedType>();
+    ShapedType resultType = op.getType().cast<ShapedType>();
+
+    if (!inputType.hasStaticShape() || !weightType.hasRank()) {
+      return failure();
+    }
+
+    // Stride must be 1 for this optimization.
+    for (Attribute stride : op.stride().getValue()) {
+      if (!stride.cast<IntegerAttr>().getValue().isOne()) {
+        return failure();
+      }
+    }
+
+    // Only works for a 1x1 kernel.
+    ArrayRef<int64_t> weightShape = weightType.getShape();
+    if (weightShape[1] != 1 || weightShape[2] != 1) {
+      return failure();
+    }
+
+    // Reshape input to [N,IH,IW,IC] -> [N * IH * IW, IC].
+    ArrayRef<int64_t> inputShape = inputType.getShape();
+    llvm::SmallVector<int64_t, 2> revisedInputShape{
+        inputShape[0] * inputShape[1] * inputShape[2], inputShape[3]};
+    auto revisedInputShapeType = RankedTensorType::get(
+        revisedInputShape,
+        input.getType().dyn_cast<RankedTensorType>().getElementType());
+    auto reshapedInput = rewriter
+                             .create<tosa::ReshapeOp>(
+                                 op.getLoc(), revisedInputShapeType, input,
+                                 rewriter.getI64ArrayAttr(revisedInputShape))
+                             .getResult();
+
+    // Reshape kernel to [OC,KH,KW,IC] -> [OC, IC].
+    llvm::SmallVector<int64_t, 2> revisedWeightShape{weightShape[0],
+                                                     weightShape[3]};
+    auto revisedWeightShapeType = RankedTensorType::get(
+        revisedWeightShape,
+        weight.getType().dyn_cast<RankedTensorType>().getElementType());
+    auto reshapedWeight = rewriter
+                              .create<tosa::ReshapeOp>(
+                                  op.getLoc(), revisedWeightShapeType, weight,
+                                  rewriter.getI64ArrayAttr(revisedWeightShape))
+                              .getResult();
+
+    // Perform a fully connected network over the reshaped input and weight.
+    llvm::SmallVector<int64_t, 2> fullyConnectedShape{
+        inputShape[0] * inputShape[1] * inputShape[2], weightShape[0]};
+    auto fullyConnectedShapeType = RankedTensorType::get(
+        fullyConnectedShape,
+        resultType.dyn_cast<ShapedType>().getElementType());
+
+    Value fullyConnectedValue;
+    if (op.quantization_info()) {
+      fullyConnectedValue =
+          rewriter
+              .create<tosa::FullyConnectedOp>(
+                  op.getLoc(), fullyConnectedShapeType, reshapedInput,
+                  reshapedWeight, op.bias(), op.quantization_info().getValue())
+              .getResult();
+    } else {
+      fullyConnectedValue = rewriter
+                                .create<tosa::FullyConnectedOp>(
+                                    op.getLoc(), fullyConnectedShapeType,
+                                    reshapedInput, reshapedWeight, op.bias())
+                                .getResult();
+    }
+
+    // Reshape output to [N, IH, IW, OC].
+    llvm::SmallVector<int64_t, 4> outputShape{inputShape[0], inputShape[1],
+                                              inputShape[2], weightShape[0]};
+    rewriter.replaceOpWithNewOp<tosa::ReshapeOp>(
+        op, resultType, fullyConnectedValue,
+        rewriter.getI64ArrayAttr(outputShape));
+    return success();
+  }
+};
+
+} // namespace
+
+void mlir::tosa::populateTosaDecomposeConv2D(MLIRContext *ctx,
+                                             RewritePatternSet &patterns) {
+  patterns.insert<Conv2DIsFullyConnected>(ctx);
+}

mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeDepthwise.cpp

+//===- TosaDecomposeDepthwise.cpp
+//------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Decompose TOSA Depthwise operation to a series of TOSA Ops specifically
+// (1) Convert a 1x1 Depthwise to Reshape -> Mul -> Reshape -> Add
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Tosa/IR/TosaOps.h"
+#include "mlir/Dialect/Tosa/Transforms/Passes.h"
+#include "mlir/Pass/Pass.h"
+
+using namespace mlir;
+using namespace mlir::tosa;
+
+namespace {
+
+struct DepthwiseConv2DIsMul : public OpRewritePattern<tosa::DepthwiseConv2DOp> {
+  explicit DepthwiseConv2DIsMul(MLIRContext *context)
+      : OpRewritePattern(context) {}
+
+  LogicalResult matchAndRewrite(tosa::DepthwiseConv2DOp op,
+                                PatternRewriter &rewriter) const override {
+    Value input = op.input();
+    Value weight = op.weight();
+    ShapedType inputType = input.getType().cast<ShapedType>();
+    ShapedType weightType = weight.getType().cast<ShapedType>();
+    ShapedType resultType = op.output().getType().cast<ShapedType>();
+    Type inputEType = inputType.getElementType();
+
+    if (!(inputType.hasStaticShape() && weightType.hasStaticShape() &&
+          resultType.hasStaticShape())) {
+      return failure();
+    }
+
+    // Quantization information needs to still be performed.
+    if (op.quantization_info() || !inputEType.isa<FloatType>()) {
+      return failure();
+    }
+
+    // Stride must be 1 for this optimization.
+    for (Attribute stride : op.stride().getValue()) {
+      if (!stride.cast<IntegerAttr>().getValue().isOne()) {
+        return failure();
+      }
+    }
+
+    // Only works for a 1x1 kernel.
+    ArrayRef<int64_t> weightShape = weightType.getShape();
+    if (weightShape[0] != 1 || weightShape[1] != 1) {
+      return failure();
+    }
+
+    // Reshape input to [N, H, W, C] -> [N, H, W, C, 1].
+    ArrayRef<int64_t> inputShape = inputType.getShape();
+    llvm::SmallVector<int64_t, 2> revisedInputShape{
+        inputShape[0], inputShape[1], inputShape[2], inputShape[3], 1};
+    auto revisedInputShapeType = RankedTensorType::get(
+        revisedInputShape,
+        input.getType().dyn_cast<RankedTensorType>().getElementType());
+    auto reshapedInput = rewriter
+                             .create<tosa::ReshapeOp>(
+                                 op.getLoc(), revisedInputShapeType, input,
+                                 rewriter.getI64ArrayAttr(revisedInputShape))
+                             .getResult();
+
+    // Reshape kernel to [KH, KW, C, M] -> [1, 1, 1, C, M].
+    llvm::SmallVector<int64_t, 2> revisedWeightShape{1, 1, 1, weightShape[2],
+                                                     weightShape[3]};
+    auto revisedWeightShapeType = RankedTensorType::get(
+        revisedWeightShape,
+        weight.getType().dyn_cast<RankedTensorType>().getElementType());
+    auto reshapedWeight = rewriter
+                              .create<tosa::ReshapeOp>(
+                                  op.getLoc(), revisedWeightShapeType, weight,
+                                  rewriter.getI64ArrayAttr(revisedWeightShape))
+                              .getResult();
+
+    // Perform an elementwise mul over the reshaped input and weight.
+    llvm::SmallVector<int64_t, 2> mulShape{inputShape[0], inputShape[1],
+                                           inputShape[2], inputShape[3],
+                                           weightShape[3]};
+    auto mulShapeType = RankedTensorType::get(
+        mulShape,
+        weight.getType().dyn_cast<RankedTensorType>().getElementType());
+    Value mulValue =
+        rewriter
+            .create<tosa::MulOp>(op.getLoc(), mulShapeType, reshapedInput,
+                                 reshapedWeight, /*shift=*/0)
+            .getResult();
+
+    // Reshape output to [N, H, W, C * M].
+    auto outputShape = op.output().getType().cast<ShapedType>().getShape();
+    auto outputShapeType = RankedTensorType::get(
+        outputShape,
+        input.getType().dyn_cast<RankedTensorType>().getElementType());
+    auto outputValue =
+        rewriter.create<tosa::ReshapeOp>(op.getLoc(), outputShapeType, mulValue,
+                                         rewriter.getI64ArrayAttr(outputShape));
+
+    // Add in the bias.
+    rewriter
+        .replaceOpWithNewOp<tosa::AddOp>(op, outputShapeType, outputValue,
+                                         op.bias())
+        .getResult();
+    return success();
+  }
+};
+
+} // namespace
+
+void mlir::tosa::populateTosaDecomposeDepthwise(MLIRContext *ctx,
+                                                RewritePatternSet &patterns) {
+  patterns.insert<DepthwiseConv2DIsMul>(ctx);
+}

mlir/lib/Dialect/Tosa/Transforms/TosaDecomposeTransposeConv.cpp

 //===- TosaDecomposeTransposeConv.cpp
 //------------------------------------------===//
 //
 // Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
 // See https://llvm.org/LICENSE.txt for license information.
 // SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
 //
 //===----------------------------------------------------------------------===//
 //
-// Insert reshape to binary op's input if needed to match rank
+// Decompose TOSA TransposeConv operation to a series of TOSA Ops specifically
+// (1) Convert a Dilated TransposeConv2D to Conv2D including reversing/reshaping
+// etc.. of the weights (2) Convert a Strided TransposeConv2D to Conv2D
+// including transposing/reversing/reshaping etc..
+//     of the weights and input/output tenors and reversing/reshaping etc .. of
+//     the weights
 //
 //===----------------------------------------------------------------------===//
 
-#include "mlir/Dialect/StandardOps/IR/Ops.h"
-#include "mlir/Dialect/Tosa/IR//TosaOps.h"
-#include "mlir/Dialect/Tosa/Transforms/PassDetail.h"
+#include "mlir/Dialect/Tosa/IR/TosaOps.h"
 #include "mlir/Dialect/Tosa/Transforms/Passes.h"
 #include "mlir/Dialect/Tosa/Utils/ShapeUtils.h"
 #include "mlir/Pass/Pass.h"
-#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
 
 using namespace mlir;
 using namespace mlir::tosa;
 
 namespace {
 
 template <typename T>
 static void getValuesFromIntArrayAttribute(ArrayAttr attr,
                                            SmallVector<T> &arrayValues) {
   for (Attribute val : attr.getValue()) {
     arrayValues.push_back(val.cast<IntegerAttr>().getValue().getSExtValue());
   }
 }
 
 template <typename TosaOp, typename... Args>
 TosaOp createOpAndInfer(PatternRewriter &rewriter, Location loc, Type resultTy,
                         Args &&...args) {
   auto op = rewriter.create<TosaOp>(loc, resultTy, args...);
 
   InferShapedTypeOpInterface shapeInterface =
       dyn_cast<InferShapedTypeOpInterface>(op.getOperation());
   if (!shapeInterface)
     return op;
 
   SmallVector<ShapedTypeComponents> returnedShapes;
auto addBias =
         createOpAndInfer<tosa::AddOp>(rewriter, loc, op.getType(), slice, bias);
 
     rewriter.replaceOp(op, addBias.getResult());
 
     return success();
   }
 };
 
-/// Pass that enables broadcast by making all input arrays have the same
-/// number of dimensions. Insert RESHAPE operations to lower rank operand
-struct TosaDecomposeTransposeConv
-    : public TosaDecomposeTransposeConvBase<TosaDecomposeTransposeConv> {
-public:
-  void runOnFunction() override {
-    auto func = getFunction();
-    RewritePatternSet patterns(func.getContext());
-    patterns
-        .insert<TransposeConvDilatedConverter, TransposeConvStridedConverter>(
-            func.getContext());
-    (void)applyPatternsAndFoldGreedily(func, std::move(patterns));
-  }
-};
 } // namespace
 
-std::unique_ptr<Pass> mlir::tosa::createTosaDecomposeTransposeConvPass() {
-  return std::make_unique<TosaDecomposeTransposeConv>();
+void mlir::tosa::populateTosaDecomposeTransposeConv(
+    MLIRContext *ctx, RewritePatternSet &patterns) {
+  patterns.insert<TransposeConvDilatedConverter>(ctx);
+  patterns.insert<TransposeConvStridedConverter>(ctx);
 }

mlir/lib/Dialect/Tosa/Transforms/TosaOptimization.cpp

-//===- TosaOptimization.cpp ------------------------------------------===//
-//
-// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
-// See https://llvm.org/LICENSE.txt for license information.
-// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
-//
-//===----------------------------------------------------------------------===//
-//
-// Pass to perform optimizations on TOSA operations
-//
-//===----------------------------------------------------------------------===//
-
-#include "mlir/Analysis/DataFlowAnalysis.h"
-#include "mlir/Dialect/StandardOps/IR/Ops.h"
-#include "mlir/Dialect/Tensor/IR/Tensor.h"
-#include "mlir/Dialect/Tosa/IR/TosaOps.h"
-#include "mlir/Dialect/Tosa/Transforms/PassDetail.h"
-#include "mlir/Dialect/Tosa/Transforms/Passes.h"
-#include "mlir/Dialect/Tosa/Utils/ShapeUtils.h"
-#include "mlir/IR/BlockAndValueMapping.h"
-#include "mlir/IR/Builders.h"
-#include "mlir/IR/BuiltinOps.h"
-#include "mlir/IR/Matchers.h"
-#include "mlir/Pass/Pass.h"
-#include "mlir/Transforms/DialectConversion.h"
-#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
-#include "llvm/Support/FormatVariadic.h"
-
-using namespace mlir;
-using namespace mlir::tosa;
-
-#define PASS_NAME "tosa-optimization"
-#define DEBUG_TYPE PASS_NAME
-
-namespace {
-
-struct Conv2DIsFullyConnected : public OpRewritePattern<tosa::Conv2DOp> {
-  explicit Conv2DIsFullyConnected(MLIRContext *context)
-      : OpRewritePattern(context) {}
-
-  LogicalResult matchAndRewrite(tosa::Conv2DOp op,
-                                PatternRewriter &rewriter) const override {
-    Value input = op.input();
-    Value weight = op.weight();
-    ShapedType inputType = input.getType().cast<ShapedType>();
-    ShapedType weightType = weight.getType().cast<ShapedType>();
-    ShapedType resultType = op.getType().cast<ShapedType>();
-
-    if (!inputType.hasStaticShape() || !weightType.hasRank()) {
-      return failure();
-    }
-
-    // Stride must be 1 for this optimization.
-    for (Attribute stride : op.stride().getValue()) {
-      if (!stride.cast<IntegerAttr>().getValue().isOne()) {
-        return failure();
-      }
-    }
-
-    // Only works for a 1x1 kernel.
-    ArrayRef<int64_t> weightShape = weightType.getShape();
-    if (weightShape[1] != 1 || weightShape[2] != 1) {
-      return failure();
-    }
-
-    // Reshape input to [N,IH,IW,IC] -> [N * IH * IW, IC].
-    ArrayRef<int64_t> inputShape = inputType.getShape();
-    llvm::SmallVector<int64_t, 2> revisedInputShape{
-        inputShape[0] * inputShape[1] * inputShape[2], inputShape[3]};
-    auto revisedInputShapeType = RankedTensorType::get(
-        revisedInputShape,
-        input.getType().dyn_cast<RankedTensorType>().getElementType());
-    auto reshapedInput = rewriter
-                             .create<tosa::ReshapeOp>(
-                                 op.getLoc(), revisedInputShapeType, input,
-                                 rewriter.getI64ArrayAttr(revisedInputShape))
-                             .getResult();
-
-    // Reshape kernel to [OC,KH,KW,IC] -> [OC, IC].
-    llvm::SmallVector<int64_t, 2> revisedWeightShape{weightShape[0],
-                                                     weightShape[3]};
-    auto revisedWeightShapeType = RankedTensorType::get(
-        revisedWeightShape,
-        weight.getType().dyn_cast<RankedTensorType>().getElementType());
-    auto reshapedWeight = rewriter
-                              .create<tosa::ReshapeOp>(
-                                  op.getLoc(), revisedWeightShapeType, weight,
-                                  rewriter.getI64ArrayAttr(revisedWeightShape))
-                              .getResult();
-
-    // Perform a fully connected network over the reshaped input and weight.
-    llvm::SmallVector<int64_t, 2> fullyConnectedShape{
-        inputShape[0] * inputShape[1] * inputShape[2], weightShape[0]};
-    auto fullyConnectedShapeType = RankedTensorType::get(
-        fullyConnectedShape,
-        resultType.dyn_cast<ShapedType>().getElementType());
-
-    Value fullyConnectedValue;
-    if (op.quantization_info()) {
-      fullyConnectedValue =
-          rewriter
-              .create<tosa::FullyConnectedOp>(
-                  op.getLoc(), fullyConnectedShapeType, reshapedInput,
-                  reshapedWeight, op.bias(), op.quantization_info().getValue())
-              .getResult();
-    } else {
-      fullyConnectedValue = rewriter
-                                .create<tosa::FullyConnectedOp>(
-                                    op.getLoc(), fullyConnectedShapeType,
-                                    reshapedInput, reshapedWeight, op.bias())
-                                .getResult();
-    }
-
-    // Reshape output to [N, IH, IW, OC].
-    llvm::SmallVector<int64_t, 4> outputShape{inputShape[0], inputShape[1],
-                                              inputShape[2], weightShape[0]};
-    rewriter.replaceOpWithNewOp<tosa::ReshapeOp>(
-        op, resultType, fullyConnectedValue,
-        rewriter.getI64ArrayAttr(outputShape));
-    return success();
-  }
-};
-
-struct DepthwiseConv2DIsMul : public OpRewritePattern<tosa::DepthwiseConv2DOp> {
-  explicit DepthwiseConv2DIsMul(MLIRContext *context)
-      : OpRewritePattern(context) {}
-
-  LogicalResult matchAndRewrite(tosa::DepthwiseConv2DOp op,
-                                PatternRewriter &rewriter) const override {
-    Value input = op.input();
-    Value weight = op.weight();
-    ShapedType inputType = input.getType().cast<ShapedType>();
-    ShapedType weightType = weight.getType().cast<ShapedType>();
-    ShapedType resultType = op.output().getType().cast<ShapedType>();
-    Type inputEType = inputType.getElementType();
-
-    if (!(inputType.hasStaticShape() && weightType.hasStaticShape() &&
-          resultType.hasStaticShape())) {
-      return failure();
-    }
-
-    // Quantization information needs to still be performed.
-    if (op.quantization_info() || !inputEType.isa<FloatType>()) {
-      return failure();
-    }
-
-    // Stride must be 1 for this optimization.
-    for (Attribute stride : op.stride().getValue()) {
-      if (!stride.cast<IntegerAttr>().getValue().isOne()) {
-        return failure();
-      }
-    }
-
-    // Only works for a 1x1 kernel.
-    ArrayRef<int64_t> weightShape = weightType.getShape();
-    if (weightShape[0] != 1 || weightShape[1] != 1) {
-      return failure();
-    }
-
-    // Reshape input to [N, H, W, C] -> [N, H, W, C, 1].
-    ArrayRef<int64_t> inputShape = inputType.getShape();
-    llvm::SmallVector<int64_t, 2> revisedInputShape{
-        inputShape[0], inputShape[1], inputShape[2], inputShape[3], 1};
-    auto revisedInputShapeType = RankedTensorType::get(
-        revisedInputShape,
-        input.getType().dyn_cast<RankedTensorType>().getElementType());
-    auto reshapedInput = rewriter
-                             .create<tosa::ReshapeOp>(
-                                 op.getLoc(), revisedInputShapeType, input,
-                                 rewriter.getI64ArrayAttr(revisedInputShape))
-                             .getResult();
-
-    // Reshape kernel to [KH, KW, C, M] -> [1, 1, 1, C, M].
-    llvm::SmallVector<int64_t, 2> revisedWeightShape{1, 1, 1, weightShape[2],
-                                                     weightShape[3]};
-    auto revisedWeightShapeType = RankedTensorType::get(
-        revisedWeightShape,
-        weight.getType().dyn_cast<RankedTensorType>().getElementType());
-    auto reshapedWeight = rewriter
-                              .create<tosa::ReshapeOp>(
-                                  op.getLoc(), revisedWeightShapeType, weight,
-                                  rewriter.getI64ArrayAttr(revisedWeightShape))
-                              .getResult();
-
-    // Perform an elementwise mul over the reshaped input and weight.
-    llvm::SmallVector<int64_t, 2> mulShape{inputShape[0], inputShape[1],
-                                           inputShape[2], inputShape[3],
-                                           weightShape[3]};
-    auto mulShapeType = RankedTensorType::get(
-        mulShape,
-        weight.getType().dyn_cast<RankedTensorType>().getElementType());
-    Value mulValue =
-        rewriter
-            .create<tosa::MulOp>(op.getLoc(), mulShapeType, reshapedInput,
-                                 reshapedWeight, /*shift=*/0)
-            .getResult();
-
-    // Reshape output to [N, H, W, C * M].
-    auto outputShape = op.output().getType().cast<ShapedType>().getShape();
-    auto outputShapeType = RankedTensorType::get(
-        outputShape,
-        input.getType().dyn_cast<RankedTensorType>().getElementType());
-    auto outputValue =
-        rewriter.create<tosa::ReshapeOp>(op.getLoc(), outputShapeType, mulValue,
-                                         rewriter.getI64ArrayAttr(outputShape));
-
-    // Add in the bias.
-    rewriter
-        .replaceOpWithNewOp<tosa::AddOp>(op, outputShapeType, outputValue,
-                                         op.bias())
-        .getResult();
-    return success();
-  }
-};
-
-class TosaOptimization : public PassWrapper<TosaOptimization, FunctionPass> {
-public:
-  explicit TosaOptimization() = default;
-  void runOnFunction() override;
-
-  StringRef getArgument() const final { return PASS_NAME; }
-  StringRef getDescription() const final {
-    return "Applies TOSA Operation Optimizations";
-  }
-};
-
-void TosaOptimization::runOnFunction() {
-  OwningRewritePatternList patterns(&getContext());
-
-  patterns.insert<Conv2DIsFullyConnected>(&getContext());
-  patterns.insert<DepthwiseConv2DIsMul>(&getContext());
-
-  auto func = getFunction();
-  if (applyPatternsAndFoldGreedily(func, std::move(patterns)).failed()) {
-    signalPassFailure();
-  }
-}
-
-} // namespace
-
-std::unique_ptr<Pass> mlir::tosa::createTosaOptimizationPass() {
-  return std::make_unique<TosaOptimization>();
-}

mlir/test/Dialect/Tosa/tosa-decompose-conv2d.mlir

-// RUN: mlir-opt --split-input-file --tosa-optimization %s | FileCheck %s
-
-// -----
-
-// CHECK-LABEL: @conv2d_as_fully_connected
-func @conv2d_as_fully_connected(%arg0: tensor<4x10x10x2xf32>, %arg1: tensor<3x1x1x2xf32>, %arg2: tensor<3xf32>) -> tensor<4x10x10x3xf32> {
-  // CHECK-NOT: "tosa.conv2d"
-  // CHECK: %[[VAR0:.*]] = "tosa.reshape"(%arg0) {new_shape = [400, 2]}
-  // CHECK-SAME: -> tensor<400x2xf32>
-  // CHECK: %[[VAR1:.*]] = "tosa.reshape"(%arg1) {new_shape = [3, 2]}
-  // CHECK-SAME: -> tensor<3x2xf32>
-  // CHECK: %[[VAR2:.*]] = "tosa.fully_connected"(%[[VAR0]], %[[VAR1]], %arg2)
-  // CHECK-SAME: -> tensor<400x3xf32>
-  // CHECK: %[[VAR3:.*]] = "tosa.reshape"(%[[VAR2]]) {new_shape = [4, 10, 10, 3]}
-  // CHECK-SAME: -> tensor<4x10x10x3xf32>
-  // CHECK: return %[[VAR3]]
-  %0 = "tosa.conv2d"(%arg0, %arg1, %arg2) {pad = [0, 0, 0, 0], stride = [1, 1], dilation = [1, 1]} : (tensor<4x10x10x2xf32>, tensor<3x1x1x2xf32>, tensor<3xf32>) -> tensor<4x10x10x3xf32>
-  return %0 : tensor<4x10x10x3xf32>
-}
-
-// -----
-
-// CHECK-LABEL: @conv2d_as_fully_connected_quant
-func @conv2d_as_fully_connected_quant(%arg0: tensor<4x10x10x2xi8>, %arg1: tensor<3x1x1x2xi8>, %arg2: tensor<3xi32>) -> tensor<4x10x10x3xi32> {
-  // CHECK-NOT: "tosa.conv2d"
-  // CHECK: %[[VAR0:.*]] = "tosa.reshape"(%arg0) {new_shape = [400, 2]}
-  // CHECK-SAME: -> tensor<400x2xi8>
-  // CHECK: %[[VAR1:.*]] = "tosa.reshape"(%arg1) {new_shape = [3, 2]}
-  // CHECK-SAME: -> tensor<3x2xi8>
-  // CHECK: %[[VAR2:.*]] = "tosa.fully_connected"(%[[VAR0]], %[[VAR1]], %arg2)
-  // CHECK-SAME: quantization_info = {input_zp = 42 : i32, weight_zp = 24 : i32}
-  // CHECK-SAME: -> tensor<400x3xi32>
-  // CHECK: %[[VAR3:.*]] = "tosa.reshape"(%[[VAR2]]) {new_shape = [4, 10, 10, 3]}
-  // CHECK-SAME: -> tensor<4x10x10x3xi32>
-  // CHECK: return %[[VAR3]]
-  %0 = "tosa.conv2d"(%arg0, %arg1, %arg2) {pad = [0, 0, 0, 0], stride = [1, 1], dilation = [1, 1], quantization_info = {input_zp = 42 : i32, weight_zp = 24 : i32}} : (tensor<4x10x10x2xi8>, tensor<3x1x1x2xi8>, tensor<3xi32>) -> tensor<4x10x10x3xi32>
-  return %0 : tensor<4x10x10x3xi32>
-}
-
-// -----
-
-// CHECK-LABEL: @depthwise_conv2d_as_mul
-func @depthwise_conv2d_as_mul(%arg0: tensor<4x10x10x2xf32>, %arg1: tensor<1x1x2x3xf32>, %arg2: tensor<6xf32>) -> tensor<4x10x10x6xf32> {
-  // CHECK-NOT: "tosa.depthwise_conv2d"
-  // CHECK: %[[VAR0:.*]] = "tosa.reshape"(%arg0) {new_shape = [4, 10, 10, 2, 1]}
-  // CHECK-SAME: -> tensor<4x10x10x2x1xf32>
-  // CHECK: %[[VAR1:.*]] = "tosa.reshape"(%arg1) {new_shape = [1, 1, 1, 2, 3]}
-  // CHECK-SAME: -> tensor<1x1x1x2x3xf32>
-  // CHECK: %[[VAR2:.*]] = "tosa.mul"(%[[VAR0]], %[[VAR1]])
-  // CHECK-SAME: -> tensor<4x10x10x2x3xf32>
-  // CHECK: %[[VAR3:.*]] = "tosa.reshape"(%[[VAR2]]) {new_shape = [4, 10, 10, 6]}
-  // CHECK-SAME: -> tensor<4x10x10x6xf32>
-  // CHECK: %[[VAR4:.*]] = "tosa.add"(%[[VAR3]], %arg2)
-  // CHECK-SAME: -> tensor<4x10x10x6xf32>
-  // CHECK: return %[[VAR4]]
-  %0 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) {pad = [0, 0, 0, 0], stride = [1, 1], dilation = [1, 1]} : (tensor<4x10x10x2xf32>, tensor<1x1x2x3xf32>, tensor<6xf32>) -> tensor<4x10x10x6xf32>
-  return %0 : tensor<4x10x10x6xf32>
-}
-
-// -----
-
-// CHECK-LABEL: @depthwise_conv2d_as_mul_q
-func @depthwise_conv2d_as_mul_q(%arg0: tensor<4x10x10x2xi8>, %arg1: tensor<1x1x2x3xi8>, %arg2: tensor<6xi32>) -> tensor<4x10x10x6xi32> {
-  // CHECK: "tosa.depthwise_conv2d"
-  %0 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) {pad = [0, 0, 0, 0], stride = [1, 1], dilation = [1, 1], quantization_info = {input_zp = 0 : i32, weight_zp = 0 : i32}} : (tensor<4x10x10x2xi8>, tensor<1x1x2x3xi8>, tensor<6xi32>) -> tensor<4x10x10x6xi32>
-  return %0 : tensor<4x10x10x6xi32>
-}
-
-// -----
+// RUN: mlir-opt --split-input-file --tosa-decompose-test-pass %s | FileCheck %s
+
+// -----
+
+// CHECK-LABEL: @conv2d_as_fully_connected
+func @conv2d_as_fully_connected(%arg0: tensor<4x10x10x2xf32>, %arg1: tensor<3x1x1x2xf32>, %arg2: tensor<3xf32>) -> tensor<4x10x10x3xf32> {
+  // CHECK-NOT: "tosa.conv2d"
+  // CHECK: %[[VAR0:.*]] = "tosa.reshape"(%arg0) {new_shape = [400, 2]}
+  // CHECK-SAME: -> tensor<400x2xf32>
+  // CHECK: %[[VAR1:.*]] = "tosa.reshape"(%arg1) {new_shape = [3, 2]}
+  // CHECK-SAME: -> tensor<3x2xf32>
+  // CHECK: %[[VAR2:.*]] = "tosa.fully_connected"(%[[VAR0]], %[[VAR1]], %arg2)
+  // CHECK-SAME: -> tensor<400x3xf32>
+  // CHECK: %[[VAR3:.*]] = "tosa.reshape"(%[[VAR2]]) {new_shape = [4, 10, 10, 3]}
+  // CHECK-SAME: -> tensor<4x10x10x3xf32>
+  // CHECK: return %[[VAR3]]
+  %0 = "tosa.conv2d"(%arg0, %arg1, %arg2) {pad = [0, 0, 0, 0], stride = [1, 1], dilation = [1, 1]} : (tensor<4x10x10x2xf32>, tensor<3x1x1x2xf32>, tensor<3xf32>) -> tensor<4x10x10x3xf32>
+  return %0 : tensor<4x10x10x3xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @conv2d_as_fully_connected_quant
+func @conv2d_as_fully_connected_quant(%arg0: tensor<4x10x10x2xi8>, %arg1: tensor<3x1x1x2xi8>, %arg2: tensor<3xi32>) -> tensor<4x10x10x3xi32> {
+  // CHECK-NOT: "tosa.conv2d"
+  // CHECK: %[[VAR0:.*]] = "tosa.reshape"(%arg0) {new_shape = [400, 2]}
+  // CHECK-SAME: -> tensor<400x2xi8>
+  // CHECK: %[[VAR1:.*]] = "tosa.reshape"(%arg1) {new_shape = [3, 2]}
+  // CHECK-SAME: -> tensor<3x2xi8>
+  // CHECK: %[[VAR2:.*]] = "tosa.fully_connected"(%[[VAR0]], %[[VAR1]], %arg2)
+  // CHECK-SAME: quantization_info = {input_zp = 42 : i32, weight_zp = 24 : i32}
+  // CHECK-SAME: -> tensor<400x3xi32>
+  // CHECK: %[[VAR3:.*]] = "tosa.reshape"(%[[VAR2]]) {new_shape = [4, 10, 10, 3]}
+  // CHECK-SAME: -> tensor<4x10x10x3xi32>
+  // CHECK: return %[[VAR3]]
+  %0 = "tosa.conv2d"(%arg0, %arg1, %arg2) {pad = [0, 0, 0, 0], stride = [1, 1], dilation = [1, 1], quantization_info = {input_zp = 42 : i32, weight_zp = 24 : i32}} : (tensor<4x10x10x2xi8>, tensor<3x1x1x2xi8>, tensor<3xi32>) -> tensor<4x10x10x3xi32>
+  return %0 : tensor<4x10x10x3xi32>
+}
+
+// -----

mlir/test/Dialect/Tosa/tosa-decompose-depthwise.mlir

+// RUN: mlir-opt --split-input-file --tosa-decompose-test-pass %s | FileCheck %s
+
+// -----
+
+// CHECK-LABEL: @depthwise_conv2d_as_mul
+func @depthwise_conv2d_as_mul(%arg0: tensor<4x10x10x2xf32>, %arg1: tensor<1x1x2x3xf32>, %arg2: tensor<6xf32>) -> tensor<4x10x10x6xf32> {
+  // CHECK-NOT: "tosa.depthwise_conv2d"
+  // CHECK: %[[VAR0:.*]] = "tosa.reshape"(%arg0) {new_shape = [4, 10, 10, 2, 1]}
+  // CHECK-SAME: -> tensor<4x10x10x2x1xf32>
+  // CHECK: %[[VAR1:.*]] = "tosa.reshape"(%arg1) {new_shape = [1, 1, 1, 2, 3]}
+  // CHECK-SAME: -> tensor<1x1x1x2x3xf32>
+  // CHECK: %[[VAR2:.*]] = "tosa.mul"(%[[VAR0]], %[[VAR1]])
+  // CHECK-SAME: -> tensor<4x10x10x2x3xf32>
+  // CHECK: %[[VAR3:.*]] = "tosa.reshape"(%[[VAR2]]) {new_shape = [4, 10, 10, 6]}
+  // CHECK-SAME: -> tensor<4x10x10x6xf32>
+  // CHECK: %[[VAR4:.*]] = "tosa.add"(%[[VAR3]], %arg2)
+  // CHECK-SAME: -> tensor<4x10x10x6xf32>
+  // CHECK: return %[[VAR4]]
+  %0 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) {pad = [0, 0, 0, 0], stride = [1, 1], dilation = [1, 1]} : (tensor<4x10x10x2xf32>, tensor<1x1x2x3xf32>, tensor<6xf32>) -> tensor<4x10x10x6xf32>
+  return %0 : tensor<4x10x10x6xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @depthwise_conv2d_as_mul_q
+func @depthwise_conv2d_as_mul_q(%arg0: tensor<4x10x10x2xi8>, %arg1: tensor<1x1x2x3xi8>, %arg2: tensor<6xi32>) -> tensor<4x10x10x6xi32> {
+  // CHECK: "tosa.depthwise_conv2d"
+  %0 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) {pad = [0, 0, 0, 0], stride = [1, 1], dilation = [1, 1], quantization_info = {input_zp = 0 : i32, weight_zp = 0 : i32}} : (tensor<4x10x10x2xi8>, tensor<1x1x2x3xi8>, tensor<6xi32>) -> tensor<4x10x10x6xi32>
+  return %0 : tensor<4x10x10x6xi32>
+}
+
+// -----

mlir/test/Dialect/Tosa/tosa-decompose-transpose-conv.mlir

-// RUN: mlir-opt --split-input-file --tosa-decompose-transpose-conv %s | FileCheck %s
+// RUN: mlir-opt --split-input-file --tosa-decompose-test-pass %s | FileCheck %s
 
 // CHECK-LABEL: @transpose_conv2d
 func @transpose_conv2d(%arg0: tensor<2x16x14x3xf32>, %arg1: tensor<5x3x6x3xf32>, %arg2: tensor<5xf32>) -> tensor<2x?x?x5xf32> {
   // CHECK: %[[REV1:.+]] = "tosa.reverse"(%arg1) {axis = 1 : i64}
   // CHECK: %[[REV2:.+]] = "tosa.reverse"(%[[REV1]]) {axis = 2 : i64}
   // CHECK: "tosa.conv2d"(%arg0, %[[REV2]], %arg2) {dilation = [1, 1], pad = [2, 2, 5, 5], stride = [1, 1]}
   %0 = "tosa.transpose_conv2d"(%arg0, %arg1, %arg2) {dilation = [1, 1], out_pad = [0, 0], out_shape = [-1, -1, -1, -1], stride = [1, 1]} : (tensor<2x16x14x3xf32>, tensor<5x3x6x3xf32>, tensor<5xf32>) -> tensor<2x18x19x5xf32>
   %1 = tensor.cast %0 : tensor<2x18x19x5xf32> to tensor<2x?x?x5xf32>

mlir/test/lib/Dialect/Tosa/CMakeLists.txt

 add_mlir_dialect_library(MLIRTosaTestPasses
   TosaTestPasses.cpp
+  TosaTestDecomposePass.cpp
 
   EXCLUDE_FROM_LIBMLIR
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/Tosa/Transforms
 
   DEPENDS
   MLIRTosaPassIncGen
 
   LINK_LIBS PUBLIC
   MLIRPass
-  MLIRTosa
   )

mlir/test/lib/Dialect/Tosa/TosaTestDecomposePass.cpp

+//===- TosaTestDecomposePass.cpp
+//-------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Pass to apply the Tosa operations decomposition rewrites
+// exposed as populate functions in
+// include/mlir/Dialect/Tosa/Transforms/Passes.h
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/StandardOps/IR/Ops.h"
+#include "mlir/Dialect/Tensor/IR/Tensor.h"
+#include "mlir/Dialect/Tosa/IR/TosaOps.h"
+#include "mlir/Dialect/Tosa/Transforms/Passes.h"
+#include "mlir/Pass/Pass.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+
+#define PASS_NAME "tosa-decompose-test-pass"
+
+using namespace mlir;
+using namespace mlir::tosa;
+
+namespace {
+
+struct TosaTestDecompose : public PassWrapper<TosaTestDecompose, FunctionPass> {
+  StringRef getArgument() const final { return PASS_NAME; }
+  StringRef getDescription() const final {
+    return "TOSA Test Pass to verify the decomposition patterns";
+  }
+  void runOnFunction() override;
+};
+
+void TosaTestDecompose::runOnFunction() {
+  auto *ctx = &getContext();
+  RewritePatternSet patterns(ctx);
+  auto func = getFunction();
+
+  populateTosaDecomposeConv2D(ctx, patterns);
+  populateTosaDecomposeTransposeConv(ctx, patterns);
+  populateTosaDecomposeDepthwise(ctx, patterns);
+
+  (void)applyPatternsAndFoldGreedily(func, std::move(patterns));
+}
+
+} // namespace
+
+namespace mlir {
+void registerTosaTestDecomposePass() { PassRegistration<TosaTestDecompose>(); }
+} // namespace mlir

mlir/tools/mlir-opt/mlir-opt.cpp

 void registerTestPrintDefUsePass();
 void registerTestPrintNestingPass();
 void registerTestReducer();
 void registerTestSpirvEntryPointABIPass();
 void registerTestSpirvGLSLCanonicalizationPass();
 void registerTestSpirvModuleCombinerPass();
 void registerTestTraitsPass();
 void registerTosaTestQuantUtilAPIPass();
+void registerTosaTestDecomposePass();
 void registerVectorizerTestPass();
 
 namespace test {
 void registerConvertCallOpPass();
 void registerInliner();
 void registerMemRefBoundCheck();
 void registerPatternsTestPass();
 void registerSimpleParametricTilingPass();
void registerTestPasses() {
   registerTestPrintNestingPass();
   registerTestReducer();
   registerTestSpirvEntryPointABIPass();
   registerTestSpirvGLSLCanonicalizationPass();
   registerTestSpirvModuleCombinerPass();
   registerTestTraitsPass();
   registerVectorizerTestPass();
   registerTosaTestQuantUtilAPIPass();
+  registerTosaTestDecomposePass();
 
   mlir::test::registerConvertCallOpPass();
   mlir::test::registerInliner();
   mlir::test::registerMemRefBoundCheck();
   mlir::test::registerPatternsTestPass();
   mlir::test::registerSimpleParametricTilingPass();
   mlir::test::registerTestAffineLoopParametricTilingPass();
   mlir::test::registerTestAliasAnalysisPass();