[Mlir-commits] [mlir] 313de31 - [mlir][tosa] Split tosa-to-linalg named ops out of pass

Rob Suderman llvmlistbot at llvm.org
Thu Dec 23 12:24:54 PST 2021


Author: Rob Suderman
Date: 2021-12-23T12:23:19-08:00
New Revision: 313de31fbb757643db13bcb47f8fe515039e298a

URL: https://github.com/llvm/llvm-project/commit/313de31fbb757643db13bcb47f8fe515039e298a
DIFF: https://github.com/llvm/llvm-project/commit/313de31fbb757643db13bcb47f8fe515039e298a.diff

LOG: [mlir][tosa] Split tosa-to-linalg named ops out of pass

Linalg named ops lowering are moved to a separate pass. This allows TOSA
canonicalizers to run between named-ops lowerings and the general TOSA
lowerings. This allows the TOSA canonicalizers to run between lowerings.

Reviewed By: NatashaKnk

Differential Revision: https://reviews.llvm.org/D116057

Added: 
    mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamed.cpp
    mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamedPass.cpp
    mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir

Modified: 
    mlir/include/mlir/Conversion/Passes.td
    mlir/include/mlir/Conversion/TosaToLinalg/TosaToLinalg.h
    mlir/lib/Conversion/TosaToLinalg/CMakeLists.txt
    mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
    mlir/lib/Conversion/TosaToLinalg/TosaToLinalgPass.cpp
    mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir

Removed: 
    


################################################################################
diff  --git a/mlir/include/mlir/Conversion/Passes.td b/mlir/include/mlir/Conversion/Passes.td
index f6e49cc889d15..4d1f383c02294 100644
--- a/mlir/include/mlir/Conversion/Passes.td
+++ b/mlir/include/mlir/Conversion/Passes.td
@@ -645,6 +645,20 @@ def TosaToLinalg : FunctionPass<"tosa-to-linalg"> {
   let constructor = "tosa::createTosaToLinalg()";
 }
 
+//===----------------------------------------------------------------------===//
+// 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.
+  }];
+
+  let constructor = "tosa::createTosaToLinalgNamed()";
+}
+
 //===----------------------------------------------------------------------===//
 // TosaToSCF
 //===----------------------------------------------------------------------===//

diff  --git a/mlir/include/mlir/Conversion/TosaToLinalg/TosaToLinalg.h b/mlir/include/mlir/Conversion/TosaToLinalg/TosaToLinalg.h
index b542833fa0e99..ec44d01065a7a 100644
--- a/mlir/include/mlir/Conversion/TosaToLinalg/TosaToLinalg.h
+++ b/mlir/include/mlir/Conversion/TosaToLinalg/TosaToLinalg.h
@@ -20,6 +20,7 @@ namespace mlir {
 namespace tosa {
 
 std::unique_ptr<Pass> createTosaToLinalg();
+std::unique_ptr<Pass> createTosaToLinalgNamed();
 
 /// 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
@@ -29,6 +30,9 @@ 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
 

diff  --git a/mlir/lib/Conversion/TosaToLinalg/CMakeLists.txt b/mlir/lib/Conversion/TosaToLinalg/CMakeLists.txt
index 5617dd3e0ce08..b98afb2aaadc0 100644
--- a/mlir/lib/Conversion/TosaToLinalg/CMakeLists.txt
+++ b/mlir/lib/Conversion/TosaToLinalg/CMakeLists.txt
@@ -1,5 +1,7 @@
 add_mlir_conversion_library(MLIRTosaToLinalg
   TosaToLinalg.cpp
+  TosaToLinalgNamed.cpp
+  TosaToLinalgNamedPass.cpp
   TosaToLinalgPass.cpp
 
   ADDITIONAL_HEADER_DIRS

diff  --git a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
index abff8b57ccdc2..04262234ceaad 100644
--- a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
+++ b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp
@@ -61,37 +61,6 @@ static mlir::SelectOp clampHelper(Location loc, Value arg,
   return rewriter.create<mlir::SelectOp>(loc, largerThanMax, max, minOrArg);
 }
 
-static mlir::Value applyPad(Location loc, Value input, ArrayRef<int64_t> pad,
-                            Attribute padAttr, OpBuilder &rewriter) {
-  // Input should be padded if necessary.
-  if (llvm::all_of(pad, [](int64_t p) { return p == 0; }))
-    return input;
-
-  ShapedType inputTy = input.getType().cast<ShapedType>();
-  Type inputETy = inputTy.getElementType();
-  auto inputShape = inputTy.getShape();
-
-  assert((inputShape.size() * 2) == pad.size());
-
-  SmallVector<int64_t, 4> paddedShape;
-  SmallVector<OpFoldResult, 8> lowIndices;
-  SmallVector<OpFoldResult, 8> highIndices;
-  for (int i = 0, s = inputShape.size(); i < s; i++) {
-    auto lowPad = pad[i * 2];
-    auto highPad = pad[i * 2 + 1];
-    paddedShape.push_back(inputShape[i] + highPad + lowPad);
-    lowIndices.push_back(rewriter.getIndexAttr(lowPad));
-    highIndices.push_back(rewriter.getIndexAttr(highPad));
-  }
-
-  Value padValue = rewriter.create<arith::ConstantOp>(loc, padAttr);
-
-  return linalg::PadTensorOp::createPadScalarOp(
-             RankedTensorType::get(paddedShape, inputETy), input, padValue,
-             lowIndices, highIndices, /*nofold=*/false, loc, rewriter)
-      .result();
-}
-
 static SmallVector<Value> filterDynamicDims(SmallVector<Value> dynDims) {
   SmallVector<Value> filteredDims;
   for (auto dim : dynDims)
@@ -1065,510 +1034,6 @@ class PointwiseConverter : public OpRewritePattern<SrcOp> {
   }
 };
 
-class ConvConverter : public OpConversionPattern<tosa::Conv2DOp> {
-public:
-  using OpConversionPattern<tosa::Conv2DOp>::OpConversionPattern;
-  LogicalResult
-  matchAndRewrite(tosa::Conv2DOp op, OpAdaptor adaptor,
-                  ConversionPatternRewriter &rewriter) const final {
-    Location loc = op->getLoc();
-    Value input = op->getOperand(0);
-    Value weight = op->getOperand(1);
-    Value bias = op->getOperand(2);
-
-    ShapedType inputTy = input.getType().cast<ShapedType>();
-    ShapedType weightTy = weight.getType().cast<ShapedType>();
-    ShapedType biasTy = bias.getType().cast<ShapedType>();
-    ShapedType resultTy = op->getResult(0).getType().cast<ShapedType>();
-
-    Type inputETy = inputTy.getElementType();
-    Type resultETy = resultTy.getElementType();
-
-    auto padAttr = op->getAttr("pad").cast<ArrayAttr>();
-    auto strideTosaAttr = op->getAttr("stride").cast<ArrayAttr>();
-    auto dilationTosaAttr = op->getAttr("dilation").cast<ArrayAttr>();
-    bool isQuantized = op->hasAttr("quantization_info");
-
-    if (!inputTy.hasStaticShape() || !weightTy.hasStaticShape() ||
-        !biasTy.hasStaticShape() || !resultTy.hasStaticShape())
-      return rewriter.notifyMatchFailure(op,
-                                         "tosa.conv ops require static shapes");
-
-    if (inputETy.isUnsignedInteger())
-      return rewriter.notifyMatchFailure(
-          op, "tosa.conv ops does not support unsigned integer input");
-
-    auto weightShape = weightTy.getShape();
-
-    // Apply padding as necessary.
-    Attribute zeroAttr = rewriter.getZeroAttr(inputETy);
-    if (isQuantized) {
-      auto quantizationInfo =
-          op->getAttr("quantization_info").cast<tosa::ConvOpQuantizationAttr>();
-      auto iZp = quantizationInfo.input_zp().getValue().getSExtValue();
-
-      int64_t intMin =
-          APInt::getSignedMinValue(inputETy.getIntOrFloatBitWidth())
-              .getSExtValue();
-      int64_t intMax =
-          APInt::getSignedMaxValue(inputETy.getIntOrFloatBitWidth())
-              .getSExtValue();
-
-      if (iZp < intMin || iZp > intMax)
-        return rewriter.notifyMatchFailure(
-            op, "tosa.conv op quantization has zp outside of input range");
-
-      zeroAttr = rewriter.getIntegerAttr(inputETy, iZp);
-    }
-
-    llvm::SmallVector<int64_t> pad;
-    pad.resize(2, 0);
-    getValuesFromIntArrayAttribute(padAttr, pad);
-    pad.resize(pad.size() + 2, 0);
-    input = applyPad(loc, input, pad, zeroAttr, rewriter);
-
-    // Transpose the kernel to match dimension ordering of the linalg
-    // convolution operation.
-    // TODO(suderman): See if this can be efficiently folded - check whether
-    // the input is used anywhere else, if not fold the constant.
-    SmallVector<int64_t> weightPerm{1, 2, 3, 0};
-    SmallVector<int64_t> newWeightShape{weightShape[1], weightShape[2],
-                                        weightShape[3], weightShape[0]};
-    auto weightPermAttr = DenseIntElementsAttr::get(
-        RankedTensorType::get({4}, rewriter.getI64Type()), weightPerm);
-    Value weightPermValue =
-        rewriter.create<arith::ConstantOp>(loc, weightPermAttr);
-    Type newWeightTy =
-        RankedTensorType::get(newWeightShape, weightTy.getElementType());
-    weight = rewriter.create<tosa::TransposeOp>(loc, newWeightTy, weight,
-                                                weightPermValue);
-
-    Attribute resultZeroAttr = rewriter.getZeroAttr(resultETy);
-    Value initTensor = rewriter.create<linalg::InitTensorOp>(
-        loc, resultTy.getShape(), resultETy);
-    Value zero = rewriter.create<arith::ConstantOp>(loc, resultZeroAttr);
-    Value zeroTensor =
-        rewriter.create<linalg::FillOp>(loc, zero, initTensor).getResult(0);
-
-    // Extract the attributes for convolution.
-    llvm::SmallVector<int64_t> stride, dilation;
-    getValuesFromIntArrayAttribute(strideTosaAttr, stride);
-    getValuesFromIntArrayAttribute(dilationTosaAttr, dilation);
-
-    // Create the convolution op.
-    auto strideAttr = DenseIntElementsAttr::get(
-        RankedTensorType::get({2}, rewriter.getI64Type()), stride);
-    auto dilationAttr = DenseIntElementsAttr::get(
-        RankedTensorType::get({2}, rewriter.getI64Type()), dilation);
-
-    // Create maps for the bias broadcasting
-    SmallVector<AffineMap, 4> indexingMaps;
-    indexingMaps.push_back(AffineMap::get(
-        /*dimCount=*/resultTy.getRank(), /*symbolCount=*/0,
-        {rewriter.getAffineDimExpr(3)}, rewriter.getContext()));
-    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(resultTy.getRank()));
-    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(resultTy.getRank()));
-
-    Value biasInitTensor = rewriter.create<linalg::InitTensorOp>(
-        loc, resultTy.getShape(), resultETy);
-
-    if (isQuantized) {
-      auto quantizationInfo =
-          op->getAttr("quantization_info").cast<tosa::ConvOpQuantizationAttr>();
-      auto iZp = rewriter.getI32IntegerAttr(
-          quantizationInfo.input_zp().getValue().getSExtValue());
-      auto kZp = rewriter.getI32IntegerAttr(
-          quantizationInfo.weight_zp().getValue().getSExtValue());
-
-      auto iZpVal = rewriter.create<arith::ConstantOp>(loc, iZp);
-      auto kZpVal = rewriter.create<arith::ConstantOp>(loc, kZp);
-      Value conv =
-          rewriter
-              .create<linalg::Conv2DNhwcHwcfQOp>(
-                  loc, resultTy, ValueRange{input, weight, iZpVal, kZpVal},
-                  ValueRange{zeroTensor}, strideAttr, dilationAttr)
-              ->getResult(0);
-
-      Value result =
-          rewriter
-              .create<linalg::GenericOp>(
-                  loc, resultTy, ValueRange({bias, conv}), biasInitTensor,
-                  indexingMaps, getNParallelLoopsAttrs(resultTy.getRank()),
-                  [&](OpBuilder &nestedBuilder, Location nestedLoc,
-                      ValueRange args) {
-                    Value added = nestedBuilder.create<arith::AddIOp>(
-                        loc, args[0], args[1]);
-                    nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
-                  })
-              .getResult(0);
-      rewriter.replaceOp(op, result);
-      return success();
-    }
-
-    Value conv = rewriter
-                     .create<linalg::Conv2DNhwcHwcfOp>(
-                         loc, resultTy, ValueRange{input, weight},
-                         ValueRange{zeroTensor}, strideAttr, dilationAttr)
-                     ->getResult(0);
-
-    Value result =
-        rewriter
-            .create<linalg::GenericOp>(
-                loc, resultTy, ValueRange({bias, conv}), biasInitTensor,
-                indexingMaps, getNParallelLoopsAttrs(resultTy.getRank()),
-                [&](OpBuilder &nestedBuilder, Location nestedLoc,
-                    ValueRange args) {
-                  Value added = nestedBuilder.create<arith::AddFOp>(
-                      loc, args[0], args[1]);
-                  nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
-                })
-            .getResult(0);
-
-    rewriter.replaceOp(op, result);
-    return success();
-  }
-};
-
-class DepthwiseConvConverter
-    : public OpConversionPattern<tosa::DepthwiseConv2DOp> {
-public:
-  using OpConversionPattern<tosa::DepthwiseConv2DOp>::OpConversionPattern;
-  LogicalResult
-  matchAndRewrite(tosa::DepthwiseConv2DOp op, OpAdaptor adaptor,
-                  ConversionPatternRewriter &rewriter) const final {
-    Location loc = op->getLoc();
-    Value input = op->getOperand(0);
-    Value weight = op->getOperand(1);
-    Value bias = op->getOperand(2);
-
-    ShapedType inputTy = input.getType().cast<ShapedType>();
-    ShapedType weightTy = weight.getType().cast<ShapedType>();
-    ShapedType biasTy = bias.getType().cast<ShapedType>();
-    ShapedType resultTy = op->getResult(0).getType().cast<ShapedType>();
-
-    Type inputETy = inputTy.getElementType();
-    Type resultETy = resultTy.getElementType();
-
-    auto padAttr = op->getAttr("pad").cast<ArrayAttr>();
-    auto strideTosaAttr = op->getAttr("stride").cast<ArrayAttr>();
-    auto dilationTosaAttr = op->getAttr("dilation").cast<ArrayAttr>();
-
-    bool isQuantized = op->hasAttr("quantization_info");
-    IntegerAttr iZp;
-    IntegerAttr kZp;
-    if (isQuantized) {
-      auto quantizationInfo =
-          op->getAttr("quantization_info").cast<tosa::ConvOpQuantizationAttr>();
-      iZp = rewriter.getI32IntegerAttr(
-          quantizationInfo.input_zp().getValue().getSExtValue());
-      kZp = rewriter.getI32IntegerAttr(
-          quantizationInfo.weight_zp().getValue().getSExtValue());
-    }
-
-    if (!inputTy.hasStaticShape() || !weightTy.hasStaticShape() ||
-        !biasTy.hasStaticShape() || !resultTy.hasStaticShape())
-      return rewriter.notifyMatchFailure(op,
-                                         "tosa.conv ops require static shapes");
-
-    auto weightShape = weightTy.getShape();
-    auto resultShape = resultTy.getShape();
-
-    // Apply padding as necessary.
-    Attribute zeroAttr = rewriter.getZeroAttr(inputETy);
-    if (isQuantized) {
-      auto quantizationInfo =
-          op->getAttr("quantization_info").cast<tosa::ConvOpQuantizationAttr>();
-      auto iZp = quantizationInfo.input_zp().getValue().getSExtValue();
-
-      int64_t intMin =
-          APInt::getSignedMinValue(inputETy.getIntOrFloatBitWidth())
-              .getSExtValue();
-      int64_t intMax =
-          APInt::getSignedMaxValue(inputETy.getIntOrFloatBitWidth())
-              .getSExtValue();
-
-      if (iZp < intMin || iZp > intMax)
-        return rewriter.notifyMatchFailure(
-            op, "tosa.depthwise_conv op quantization has zp outside of input "
-                "range");
-
-      zeroAttr = rewriter.getIntegerAttr(inputETy, iZp);
-    }
-
-    llvm::SmallVector<int64_t> pad;
-    pad.resize(2, 0);
-    getValuesFromIntArrayAttribute(padAttr, pad);
-    pad.resize(pad.size() + 2, 0);
-
-    input = applyPad(loc, input, pad, zeroAttr, rewriter);
-
-    // Extract the attributes for convolution.
-    llvm::SmallVector<int64_t> stride, dilation;
-    getValuesFromIntArrayAttribute(strideTosaAttr, stride);
-    getValuesFromIntArrayAttribute(dilationTosaAttr, dilation);
-
-    // Create the convolution op.
-    auto strideAttr = DenseIntElementsAttr::get(
-        RankedTensorType::get({2}, rewriter.getI64Type()), stride);
-    auto dilationAttr = DenseIntElementsAttr::get(
-        RankedTensorType::get({2}, rewriter.getI64Type()), dilation);
-    ShapedType linalgConvTy =
-        RankedTensorType::get({resultShape[0], resultShape[1], resultShape[2],
-                               weightShape[2], weightShape[3]},
-                              resultETy);
-
-    // Broadcast the initial value to the output tensor before convolving.
-    SmallVector<AffineMap, 4> indexingMaps;
-    indexingMaps.push_back(AffineMap::get(
-        /*dimCount=*/resultTy.getRank(), /*symbolCount=*/0,
-        {rewriter.getAffineDimExpr(3)}, rewriter.getContext()));
-    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(resultTy.getRank()));
-    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(resultTy.getRank()));
-
-    Attribute resultZeroAttr = rewriter.getZeroAttr(resultETy);
-    Value initTensor = rewriter.create<linalg::InitTensorOp>(
-        loc, linalgConvTy.getShape(), resultETy);
-    Value zero = rewriter.create<arith::ConstantOp>(loc, resultZeroAttr);
-    Value zeroTensor =
-        rewriter.create<linalg::FillOp>(loc, zero, initTensor).getResult(0);
-
-    Value biasInitTensor = rewriter.create<linalg::InitTensorOp>(
-        loc, resultTy.getShape(), resultETy);
-    if (!isQuantized) {
-      Value conv = rewriter
-                       .create<linalg::DepthwiseConv2DNhwcHwcmOp>(
-                           loc, linalgConvTy, ValueRange{input, weight},
-                           ValueRange{zeroTensor}, strideAttr, dilationAttr)
-                       .getResult(0);
-      Value convReshape = rewriter.create<tosa::ReshapeOp>(loc, resultTy, conv);
-      Value result =
-          rewriter
-              .create<linalg::GenericOp>(
-                  loc, resultTy, ValueRange({bias, convReshape}),
-                  biasInitTensor, indexingMaps,
-                  getNParallelLoopsAttrs(resultTy.getRank()),
-                  [&](OpBuilder &nestedBuilder, Location nestedLoc,
-                      ValueRange args) {
-                    Value added = nestedBuilder.create<arith::AddFOp>(
-                        loc, args[0], args[1]);
-                    nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
-                  })
-              .getResult(0);
-      rewriter.replaceOp(op, result);
-    } else {
-      auto iZpVal = rewriter.create<arith::ConstantOp>(loc, iZp);
-      auto kZpVal = rewriter.create<arith::ConstantOp>(loc, kZp);
-      Value conv =
-          rewriter
-              .create<linalg::DepthwiseConv2DNhwcHwcmQOp>(
-                  loc, linalgConvTy, ValueRange{input, weight, iZpVal, kZpVal},
-                  ValueRange{zeroTensor}, strideAttr, dilationAttr)
-              .getResult(0);
-      Value convReshape = rewriter.create<tosa::ReshapeOp>(loc, resultTy, conv);
-      Value result =
-          rewriter
-              .create<linalg::GenericOp>(
-                  loc, resultTy, ValueRange({bias, convReshape}),
-                  biasInitTensor, indexingMaps,
-                  getNParallelLoopsAttrs(resultTy.getRank()),
-                  [&](OpBuilder &nestedBuilder, Location nestedLoc,
-                      ValueRange args) {
-                    Value added = nestedBuilder.create<arith::AddIOp>(
-                        loc, args[0], args[1]);
-                    nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
-                  })
-              .getResult(0);
-      rewriter.replaceOp(op, result);
-    }
-    return success();
-  }
-};
-
-class MatMulConverter : public OpConversionPattern<tosa::MatMulOp> {
-public:
-  using OpConversionPattern<tosa::MatMulOp>::OpConversionPattern;
-  LogicalResult
-  matchAndRewrite(tosa::MatMulOp op, OpAdaptor adaptor,
-                  ConversionPatternRewriter &rewriter) const final {
-    Location loc = op.getLoc();
-
-    auto outputTy = op.getType().cast<ShapedType>();
-    auto outputElementTy = outputTy.getElementType();
-
-    auto firstOperandTy = op->getOperand(0).getType().cast<ShapedType>();
-    auto secondOperandTy = op->getOperand(1).getType().cast<ShapedType>();
-
-    SmallVector<Value> dynDims;
-    dynDims.resize(op->getResult(0).getType().cast<ShapedType>().getRank());
-
-    if (!firstOperandTy.hasRank() || firstOperandTy.isDynamicDim(0)) {
-      dynDims[0] = rewriter.create<tensor::DimOp>(loc, op->getOperand(0), 0);
-    }
-
-    if (!firstOperandTy.hasRank() || firstOperandTy.isDynamicDim(1)) {
-      dynDims[1] = rewriter.create<tensor::DimOp>(loc, op->getOperand(0), 1);
-    }
-
-    if (!secondOperandTy.hasRank() || secondOperandTy.isDynamicDim(2)) {
-      dynDims[2] = rewriter.create<tensor::DimOp>(loc, op->getOperand(1), 2);
-    }
-
-    SmallVector<Value> filteredDims = filterDynamicDims(dynDims);
-
-    auto zeroAttr = rewriter.getZeroAttr(outputElementTy);
-    Value zero = rewriter.create<arith::ConstantOp>(loc, zeroAttr);
-    auto initTensor = rewriter.create<linalg::InitTensorOp>(
-        loc, filteredDims, outputTy.getShape(), outputTy.getElementType());
-    Value zeroTensor =
-        rewriter.create<linalg::FillOp>(loc, zero, initTensor).getResult(0);
-    if (!op.quantization_info()) {
-      rewriter.replaceOpWithNewOp<linalg::BatchMatmulOp>(
-          op, TypeRange{op.getType()}, ValueRange{adaptor.a(), adaptor.b()},
-          ValueRange{zeroTensor});
-      return success();
-    }
-
-    auto quantizationInfo = op.quantization_info().getValue();
-    auto aZp = rewriter.create<arith::ConstantOp>(
-        loc, rewriter.getI32IntegerAttr(
-                 quantizationInfo.a_zp().getValue().getSExtValue()));
-    auto bZp = rewriter.create<arith::ConstantOp>(
-        loc, rewriter.getI32IntegerAttr(
-                 quantizationInfo.b_zp().getValue().getSExtValue()));
-    rewriter.replaceOpWithNewOp<linalg::QuantizedBatchMatmulOp>(
-        op, TypeRange{op.getType()},
-        ValueRange{adaptor.a(), adaptor.b(), aZp, bZp}, zeroTensor);
-
-    return success();
-  }
-};
-
-class FullyConnectedConverter
-    : public OpConversionPattern<tosa::FullyConnectedOp> {
-public:
-  using OpConversionPattern<tosa::FullyConnectedOp>::OpConversionPattern;
-  LogicalResult
-  matchAndRewrite(tosa::FullyConnectedOp op, OpAdaptor adaptor,
-                  ConversionPatternRewriter &rewriter) const final {
-    Location loc = op.getLoc();
-    auto outputTy = op.getType().cast<ShapedType>();
-    auto input = op.input();
-    auto inputTy = input.getType().cast<ShapedType>();
-
-    auto bias = op.bias();
-
-    auto weight = op.weight();
-    auto weightTy = weight.getType().cast<ShapedType>();
-    auto weightShape = weightTy.getShape();
-
-    auto outputETy = outputTy.getElementType();
-
-    SmallVector<Value> dynDims;
-    dynDims.resize(op->getResult(0).getType().cast<ShapedType>().getRank());
-
-    if (!inputTy.hasRank() || inputTy.isDynamicDim(0)) {
-      dynDims[0] = rewriter.create<tensor::DimOp>(loc, input, 0);
-    }
-
-    if (!weightTy.hasRank() || weightTy.isDynamicDim(0)) {
-      dynDims[1] = rewriter.create<tensor::DimOp>(loc, weight, 0);
-    }
-
-    SmallVector<Value> filteredDims = filterDynamicDims(dynDims);
-
-    // Creating maps for the output of MatMul and the bias
-    SmallVector<AffineMap, 4> indexingMaps;
-
-    // Broadcast the bias.
-    indexingMaps.push_back(AffineMap::get(/*dimCount=*/2, /*symbolCount=*/0,
-                                          {rewriter.getAffineDimExpr(1)},
-                                          rewriter.getContext()));
-
-    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(outputTy.getRank()));
-    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(outputTy.getRank()));
-
-    auto initTensor = rewriter.create<linalg::InitTensorOp>(
-        loc, filteredDims, outputTy.getShape(), outputTy.getElementType());
-
-    // When quantized, the input elemeny type is not the same as the output
-    Attribute resultZeroAttr = rewriter.getZeroAttr(outputETy);
-    Value zero = rewriter.create<arith::ConstantOp>(loc, resultZeroAttr);
-    Value zeroTensor =
-        rewriter.create<linalg::FillOp>(loc, zero, initTensor).getResult(0);
-
-    SmallVector<int64_t> permutation{1, 0};
-    auto permutationAttr = DenseIntElementsAttr::get(
-        RankedTensorType::get({2}, rewriter.getI64Type()), permutation);
-    Value permutationValue =
-        rewriter.create<arith::ConstantOp>(loc, permutationAttr);
-
-    SmallVector<int64_t> newWeightShape{weightShape[1], weightShape[0]};
-    Type newWeightTy =
-        RankedTensorType::get(newWeightShape, weightTy.getElementType());
-
-    Value transposedWeight = rewriter.create<tosa::TransposeOp>(
-        loc, newWeightTy, weight, permutationValue);
-
-    auto biasInitTensor =
-        rewriter
-            .create<linalg::InitTensorOp>(loc, filteredDims,
-                                          outputTy.getShape(), outputETy)
-            ->getResults();
-
-    if (!op.quantization_info()) {
-      Value matmul = rewriter
-                         .create<linalg::MatmulOp>(
-                             loc, TypeRange{op.getType()},
-                             ValueRange{input, transposedWeight}, zeroTensor)
-                         ->getResult(0);
-
-      Value result =
-          rewriter
-              .create<linalg::GenericOp>(
-                  loc, outputTy, ValueRange({bias, matmul}), biasInitTensor,
-                  indexingMaps, getNParallelLoopsAttrs(outputTy.getRank()),
-                  [&](OpBuilder &nestedBuilder, Location nestedLoc,
-                      ValueRange args) {
-                    Value added = nestedBuilder.create<arith::AddFOp>(
-                        loc, args[0], args[1]);
-                    nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
-                  })
-              .getResult(0);
-      rewriter.replaceOp(op, result);
-      return success();
-    }
-
-    auto quantizationInfo = op.quantization_info().getValue();
-    auto inputZp = rewriter.create<arith::ConstantOp>(
-        loc, rewriter.getI32IntegerAttr(
-                 quantizationInfo.input_zp().getValue().getSExtValue()));
-    auto outputZp = rewriter.create<arith::ConstantOp>(
-        loc, rewriter.getI32IntegerAttr(
-                 quantizationInfo.weight_zp().getValue().getSExtValue()));
-    Value matmul =
-        rewriter
-            .create<linalg::QuantizedMatmulOp>(
-                loc, TypeRange{op.getType()},
-                ValueRange{input, transposedWeight, inputZp, outputZp},
-                zeroTensor)
-            ->getResult(0);
-    Value result =
-        rewriter
-            .create<linalg::GenericOp>(
-                loc, outputTy, ValueRange({bias, matmul}), biasInitTensor,
-                indexingMaps, getNParallelLoopsAttrs(outputTy.getRank()),
-                [&](OpBuilder &nestedBuilder, Location nestedLoc,
-                    ValueRange args) {
-                  Value added = nestedBuilder.create<arith::AddIOp>(
-                      loc, args[0], args[1]);
-                  nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
-                })
-            .getResult(0);
-    rewriter.replaceOp(op, result);
-    return success();
-  }
-};
-
 class ReshapeConverterCollapse : public OpConversionPattern<tosa::ReshapeOp> {
 public:
   using OpConversionPattern<tosa::ReshapeOp>::OpConversionPattern;
@@ -2810,277 +2275,6 @@ class TableConverter : public OpRewritePattern<tosa::TableOp> {
   }
 };
 
-class MaxPool2dConverter : public OpRewritePattern<tosa::MaxPool2dOp> {
-public:
-  using OpRewritePattern<tosa::MaxPool2dOp>::OpRewritePattern;
-
-  LogicalResult matchAndRewrite(tosa::MaxPool2dOp op,
-                                PatternRewriter &rewriter) const final {
-    Location loc = op.getLoc();
-    Value input = op.input();
-    ShapedType inputTy = input.getType().cast<ShapedType>();
-
-    ShapedType resultTy = op.getType().template cast<ShapedType>();
-    Type resultETy = inputTy.getElementType();
-
-    if (!inputTy.hasStaticShape())
-      return failure();
-
-    // Determine what the initial value needs to be for the max pool op.
-    Attribute initialAttr;
-    if (resultETy.isF32())
-      initialAttr = rewriter.getFloatAttr(
-          resultETy,
-          APFloat::getLargest(resultETy.cast<FloatType>().getFloatSemantics(),
-                              true));
-
-    if (resultETy.isa<IntegerType>())
-      initialAttr = rewriter.getIntegerAttr(
-          resultETy,
-          APInt::getSignedMinValue(resultETy.getIntOrFloatBitWidth()));
-
-    if (!initialAttr)
-      return rewriter.notifyMatchFailure(
-          op, "Unsupported initial value for tosa.maxpool_2d op");
-
-    // Apply padding as necessary.
-    llvm::SmallVector<int64_t> pad;
-    pad.resize(2, 0);
-    getValuesFromIntArrayAttribute(op.pad(), pad);
-    pad.resize(pad.size() + 2, 0);
-    Value paddedInput = applyPad(loc, input, pad, initialAttr, rewriter);
-
-    Value initialValue = rewriter.create<arith::ConstantOp>(loc, initialAttr);
-
-    SmallVector<int64_t> kernel, stride;
-    getValuesFromIntArrayAttribute(op.kernel(), kernel);
-    getValuesFromIntArrayAttribute(op.stride(), stride);
-
-    Attribute strideAttr = rewriter.getI64VectorAttr(stride);
-    Attribute dilationAttr = rewriter.getI64VectorAttr({1, 1});
-
-    // Create the linalg op that performs pooling.
-    Value initTensor = rewriter.create<linalg::InitTensorOp>(
-        loc, resultTy.getShape(), resultTy.getElementType());
-
-    Value filledInitTensor =
-        rewriter.create<linalg::FillOp>(loc, initialValue, initTensor).result();
-
-    Value fakeWindowDims =
-        rewriter.create<linalg::InitTensorOp>(loc, kernel, resultETy);
-
-    rewriter.replaceOpWithNewOp<linalg::PoolingNhwcMaxOp>(
-        op, ArrayRef<Type>{resultTy}, ValueRange{paddedInput, fakeWindowDims},
-        filledInitTensor, strideAttr, dilationAttr);
-    return success();
-  }
-};
-
-class AvgPool2dConverter : public OpRewritePattern<tosa::AvgPool2dOp> {
-public:
-  using OpRewritePattern<tosa::AvgPool2dOp>::OpRewritePattern;
-
-  LogicalResult matchAndRewrite(tosa::AvgPool2dOp op,
-                                PatternRewriter &rewriter) const final {
-    Location loc = op.getLoc();
-    Value input = op.input();
-    ShapedType inputTy = input.getType().cast<ShapedType>();
-    Type inElementTy = inputTy.getElementType();
-
-    ShapedType resultTy = op.getType().template cast<ShapedType>();
-    Type resultETy = op.getType().cast<ShapedType>().getElementType();
-
-    Type accETy =
-        inElementTy.isa<IntegerType>() ? rewriter.getI32Type() : inElementTy;
-    ShapedType accTy = resultTy.clone(accETy);
-
-    if (!inputTy.hasStaticShape())
-      return failure();
-
-    // Apply padding as necessary.
-    llvm::SmallVector<int64_t> pad;
-    pad.resize(2, 0);
-    getValuesFromIntArrayAttribute(op.pad(), pad);
-    pad.resize(pad.size() + 2, 0);
-    Attribute padAttr = rewriter.getZeroAttr(inElementTy);
-    Value paddedInput = applyPad(loc, input, pad, padAttr, rewriter);
-
-    Attribute initialAttr = rewriter.getZeroAttr(accETy);
-    Value initialValue = rewriter.create<arith::ConstantOp>(loc, initialAttr);
-
-    SmallVector<int64_t> kernel, stride;
-    getValuesFromIntArrayAttribute(op.kernel(), kernel);
-    getValuesFromIntArrayAttribute(op.stride(), stride);
-
-    Attribute strideAttr = rewriter.getI64VectorAttr(stride);
-    Attribute dilationAttr = rewriter.getI64VectorAttr({1, 1});
-
-    // Create the linalg op that performs pooling.
-    Value poolInitTensor =
-        rewriter.create<linalg::InitTensorOp>(loc, accTy.getShape(), accETy);
-
-    Value filledInitTensor =
-        rewriter.create<linalg::FillOp>(loc, initialValue, poolInitTensor)
-            .result();
-
-    Value fakeWindowDims =
-        rewriter.create<linalg::InitTensorOp>(loc, kernel, accETy);
-
-    // Sum across the pooled region.
-    Value poolingOp = rewriter
-                          .create<linalg::PoolingNhwcSumOp>(
-                              loc, ArrayRef<Type>{accTy},
-                              ValueRange{paddedInput, fakeWindowDims},
-                              filledInitTensor, strideAttr, dilationAttr)
-                          .getResult(0);
-
-    // Normalize the summed value by the number of elements grouped in each
-    // pool.
-    auto poolingOpTy = poolingOp.getType().cast<ShapedType>();
-    auto affineMap = rewriter.getMultiDimIdentityMap(resultTy.getRank());
-
-    Value genericInitTensor = rewriter.create<linalg::InitTensorOp>(
-        loc, resultTy.getShape(), resultETy);
-
-    auto genericOp = rewriter.create<linalg::GenericOp>(
-        loc, ArrayRef<Type>({resultTy}), ValueRange{poolingOp},
-        ValueRange{genericInitTensor},
-        ArrayRef<AffineMap>({affineMap, affineMap}),
-        getNParallelLoopsAttrs(resultTy.getRank()),
-        [&](OpBuilder &b, Location loc, ValueRange args) {
-          auto zero = rewriter.create<arith::ConstantIndexOp>(loc, 0);
-          auto one = rewriter.create<arith::ConstantIndexOp>(loc, 1);
-          auto iH = rewriter.create<arith::ConstantIndexOp>(
-              loc, poolingOpTy.getDimSize(1) - 1);
-          auto iW = rewriter.create<arith::ConstantIndexOp>(
-              loc, poolingOpTy.getDimSize(2) - 1);
-
-          // Compute the indices from either end.
-          auto y0 = rewriter.create<linalg::IndexOp>(loc, 1);
-          auto x0 = rewriter.create<linalg::IndexOp>(loc, 2);
-          auto y1 = rewriter.create<arith::SubIOp>(loc, iH, y0);
-          auto x1 = rewriter.create<arith::SubIOp>(loc, iW, x0);
-
-          // Determines what the portion of valid input is covered by the
-          // kernel.
-          auto padFn = [&](Value v, Value x, int64_t pad) -> Value {
-            if (pad == 0)
-              return v;
-
-            auto padVal = rewriter.create<arith::ConstantIndexOp>(loc, pad);
-            Value dx = rewriter.create<arith::SubIOp>(loc, x, padVal);
-
-            Value cmp = rewriter.create<arith::CmpIOp>(
-                loc, arith::CmpIPredicate::slt, dx, zero);
-            Value offset = rewriter.create<mlir::SelectOp>(loc, cmp, dx, zero);
-            return rewriter.create<arith::AddIOp>(loc, v, offset)->getResult(0);
-          };
-
-          // Compute the vertical component of coverage.
-          auto kH0 = rewriter.create<arith::ConstantIndexOp>(loc, kernel[0]);
-          auto kH1 = padFn(kH0, y0, pad[2]);
-          auto kH2 = padFn(kH1, y1, pad[3]);
-          auto kHCmp = rewriter.create<arith::CmpIOp>(
-              loc, arith::CmpIPredicate::slt, kH2, one);
-          auto kH3 = rewriter.create<SelectOp>(loc, kHCmp, one, kH2);
-
-          // compute the horizontal component of coverage.
-          auto kW0 = rewriter.create<arith::ConstantIndexOp>(loc, kernel[1]);
-          auto kW1 = padFn(kW0, x0, pad[4]);
-          auto kW2 = padFn(kW1, x1, pad[5]);
-          auto kWCmp = rewriter.create<arith::CmpIOp>(
-              loc, arith::CmpIPredicate::slt, kW2, one);
-          auto kW3 = rewriter.create<SelectOp>(loc, kWCmp, one, kW2);
-
-          // Compute the total number of elements and normalize.
-          Value count = rewriter.create<arith::MulIOp>(loc, kH3, kW3);
-          auto countI = rewriter.create<arith::IndexCastOp>(
-              loc, rewriter.getI32Type(), count);
-
-          // Divide by the number of summed values. For floats this is just
-          // a div however for quantized values input normalization had
-          // to be applied.
-          Value poolVal = args[0];
-          if (accETy.isa<FloatType>()) {
-            auto countF = rewriter.create<arith::SIToFPOp>(loc, accETy, countI);
-            poolVal = rewriter.create<arith::DivFOp>(loc, poolVal, countF)
-                          ->getResult(0);
-          } else {
-
-            // If we have quantization information we need to apply an offset
-            // for the input zp value.
-            if (op.quantization_info()) {
-              auto quantizationInfo = op.quantization_info().getValue();
-              auto inputZp = rewriter.create<arith::ConstantOp>(
-                  loc, quantizationInfo.input_zp());
-              Value offset =
-                  rewriter.create<arith::MulIOp>(loc, accETy, countI, inputZp);
-              poolVal =
-                  rewriter.create<arith::SubIOp>(loc, accETy, poolVal, offset);
-            }
-
-            // Compute the multiplier and shift values for the quantization
-            // normalization. Preferably we would want to compute more bits
-            // however 32-bits should be enough for compute. Honestly we
-            // should probably straight divide.
-            int64_t numerator = ((1 << 30) + 1);
-            int64_t shift = 30;
-
-            Value numeratorVal = rewriter.create<arith::ConstantOp>(
-                loc, rewriter.getI32IntegerAttr(numerator));
-            Value multiplierVal =
-                rewriter
-                    .create<arith::DivUIOp>(loc, rewriter.getI32Type(),
-                                            numeratorVal, countI)
-                    .getResult();
-            Value shiftVal = rewriter.create<arith::ConstantOp>(
-                loc, rewriter.getI8IntegerAttr(shift));
-
-            auto scaled =
-                rewriter
-                    .create<tosa::ApplyScaleOp>(
-                        loc, rewriter.getI32Type(), poolVal, multiplierVal,
-                        shiftVal, rewriter.getBoolAttr(false))
-                    .getResult();
-
-            // If we have quantization information we need to apply output
-            // zeropoint.
-            if (op.quantization_info()) {
-              auto quantizationInfo = op.quantization_info().getValue();
-              auto outputZp = rewriter.create<arith::ConstantOp>(
-                  loc, quantizationInfo.output_zp());
-              scaled = rewriter.create<arith::AddIOp>(loc, scaled, outputZp)
-                           .getResult();
-            }
-
-            // Apply Clip.
-            int64_t outBitwidth = resultETy.getIntOrFloatBitWidth();
-
-            auto min = rewriter.create<arith::ConstantIntOp>(
-                loc, APInt::getSignedMinValue(outBitwidth).getSExtValue(),
-                accETy);
-            auto max = rewriter.create<arith::ConstantIntOp>(
-                loc, APInt::getSignedMaxValue(outBitwidth).getSExtValue(),
-                accETy);
-            auto clamp = clampHelper<arith::CmpIOp>(
-                loc, scaled, min, max, arith::CmpIPredicate::slt, rewriter);
-
-            poolVal = clamp;
-            // Convert type.
-            if (resultETy != clamp.getType()) {
-              poolVal =
-                  rewriter.create<arith::TruncIOp>(loc, resultETy, poolVal);
-            }
-          }
-
-          rewriter.create<linalg::YieldOp>(loc, poolVal);
-        });
-
-    rewriter.replaceOp(op, genericOp.getResult(0));
-    return success();
-  }
-};
-
 } // namespace
 
 void mlir::tosa::populateTosaToLinalgConversionPatterns(
@@ -3132,8 +2326,6 @@ void mlir::tosa::populateTosaToLinalgConversionPatterns(
       ReduceConverter<tosa::ReduceProdOp>,
       ArgMaxConverter,
       ConcatConverter,
-      ConvConverter,
-      DepthwiseConvConverter,
       GatherConverter,
       PadConverter,
       ReshapeConverterCollapse,
@@ -3144,10 +2336,6 @@ void mlir::tosa::populateTosaToLinalgConversionPatterns(
       ReverseConverter,
       TableConverter,
       TileConverter,
-      TransposeConverter,
-      MatMulConverter,
-      MaxPool2dConverter,
-      AvgPool2dConverter,
-      FullyConnectedConverter>(patterns->getContext());
+      TransposeConverter>(patterns->getContext());
   // clang-format on
 }

diff  --git a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamed.cpp b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamed.cpp
new file mode 100644
index 0000000000000..90220ef44e974
--- /dev/null
+++ b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamed.cpp
@@ -0,0 +1,885 @@
+//===- TosaToLinalgNamed.cpp - Lowering Tosa to Linalg Named Ops ----------===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+//
+// These rewriters lower from the Tosa to the Linalg named ops.
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Conversion/TosaToLinalg/TosaToLinalg.h"
+#include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
+#include "mlir/Dialect/Math/IR/Math.h"
+#include "mlir/Dialect/SCF/SCF.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/Utils/ReshapeOpsUtils.h"
+#include "mlir/IR/Matchers.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/Transforms/DialectConversion.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+
+#include <numeric>
+
+using namespace mlir;
+
+static SmallVector<StringRef> getNParallelLoopsAttrs(unsigned nParallelLoops) {
+  return SmallVector<StringRef>(nParallelLoops, getParallelIteratorTypeName());
+}
+
+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 T, typename P>
+static mlir::SelectOp clampHelper(Location loc, Value arg,
+                                  arith::ConstantOp min, arith::ConstantOp max,
+                                  P pred, OpBuilder &rewriter) {
+  auto smallerThanMin = rewriter.create<T>(loc, pred, arg, min);
+  auto minOrArg =
+      rewriter.create<mlir::SelectOp>(loc, smallerThanMin, min, arg);
+  auto largerThanMax = rewriter.create<T>(loc, pred, max, arg);
+  return rewriter.create<mlir::SelectOp>(loc, largerThanMax, max, minOrArg);
+}
+
+static mlir::Value applyPad(Location loc, Value input, ArrayRef<int64_t> pad,
+                            Attribute padAttr, OpBuilder &rewriter) {
+  // Input should be padded if necessary.
+  if (llvm::all_of(pad, [](int64_t p) { return p == 0; }))
+    return input;
+
+  ShapedType inputTy = input.getType().cast<ShapedType>();
+  Type inputETy = inputTy.getElementType();
+  auto inputShape = inputTy.getShape();
+
+  assert((inputShape.size() * 2) == pad.size());
+
+  SmallVector<int64_t, 4> paddedShape;
+  SmallVector<OpFoldResult, 8> lowIndices;
+  SmallVector<OpFoldResult, 8> highIndices;
+  for (int i = 0, s = inputShape.size(); i < s; i++) {
+    auto lowPad = pad[i * 2];
+    auto highPad = pad[i * 2 + 1];
+    paddedShape.push_back(inputShape[i] + highPad + lowPad);
+    lowIndices.push_back(rewriter.getIndexAttr(lowPad));
+    highIndices.push_back(rewriter.getIndexAttr(highPad));
+  }
+
+  Value padValue = rewriter.create<arith::ConstantOp>(loc, padAttr);
+
+  return linalg::PadTensorOp::createPadScalarOp(
+             RankedTensorType::get(paddedShape, inputETy), input, padValue,
+             lowIndices, highIndices, /*nofold=*/false, loc, rewriter)
+      .result();
+}
+
+static SmallVector<Value> filterDynamicDims(SmallVector<Value> dynDims) {
+  SmallVector<Value> filteredDims;
+  for (auto dim : dynDims)
+    if (dim)
+      filteredDims.push_back(dim);
+  return filteredDims;
+}
+
+namespace {
+
+class ConvConverter : public OpConversionPattern<tosa::Conv2DOp> {
+public:
+  using OpConversionPattern<tosa::Conv2DOp>::OpConversionPattern;
+  LogicalResult
+  matchAndRewrite(tosa::Conv2DOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const final {
+    Location loc = op->getLoc();
+    Value input = op->getOperand(0);
+    Value weight = op->getOperand(1);
+    Value bias = op->getOperand(2);
+
+    ShapedType inputTy = input.getType().cast<ShapedType>();
+    ShapedType weightTy = weight.getType().cast<ShapedType>();
+    ShapedType biasTy = bias.getType().cast<ShapedType>();
+    ShapedType resultTy = op->getResult(0).getType().cast<ShapedType>();
+
+    Type inputETy = inputTy.getElementType();
+    Type resultETy = resultTy.getElementType();
+
+    auto padAttr = op->getAttr("pad").cast<ArrayAttr>();
+    auto strideTosaAttr = op->getAttr("stride").cast<ArrayAttr>();
+    auto dilationTosaAttr = op->getAttr("dilation").cast<ArrayAttr>();
+    bool isQuantized = op->hasAttr("quantization_info");
+
+    if (!inputTy.hasStaticShape() || !weightTy.hasStaticShape() ||
+        !biasTy.hasStaticShape() || !resultTy.hasStaticShape())
+      return rewriter.notifyMatchFailure(op,
+                                         "tosa.conv ops require static shapes");
+
+    if (inputETy.isUnsignedInteger())
+      return rewriter.notifyMatchFailure(
+          op, "tosa.conv ops does not support unsigned integer input");
+
+    auto weightShape = weightTy.getShape();
+
+    // Apply padding as necessary.
+    Attribute zeroAttr = rewriter.getZeroAttr(inputETy);
+    if (isQuantized) {
+      auto quantizationInfo =
+          op->getAttr("quantization_info").cast<tosa::ConvOpQuantizationAttr>();
+      auto iZp = quantizationInfo.input_zp().getValue().getSExtValue();
+
+      int64_t intMin =
+          APInt::getSignedMinValue(inputETy.getIntOrFloatBitWidth())
+              .getSExtValue();
+      int64_t intMax =
+          APInt::getSignedMaxValue(inputETy.getIntOrFloatBitWidth())
+              .getSExtValue();
+
+      if (iZp < intMin || iZp > intMax)
+        return rewriter.notifyMatchFailure(
+            op, "tosa.conv op quantization has zp outside of input range");
+
+      zeroAttr = rewriter.getIntegerAttr(inputETy, iZp);
+    }
+
+    llvm::SmallVector<int64_t> pad;
+    pad.resize(2, 0);
+    getValuesFromIntArrayAttribute(padAttr, pad);
+    pad.resize(pad.size() + 2, 0);
+    input = applyPad(loc, input, pad, zeroAttr, rewriter);
+
+    // Transpose the kernel to match dimension ordering of the linalg
+    // convolution operation.
+    // TODO(suderman): See if this can be efficiently folded - check whether
+    // the input is used anywhere else, if not fold the constant.
+    SmallVector<int64_t> weightPerm{1, 2, 3, 0};
+    SmallVector<int64_t> newWeightShape{weightShape[1], weightShape[2],
+                                        weightShape[3], weightShape[0]};
+    auto weightPermAttr = DenseIntElementsAttr::get(
+        RankedTensorType::get({4}, rewriter.getI64Type()), weightPerm);
+    Value weightPermValue =
+        rewriter.create<arith::ConstantOp>(loc, weightPermAttr);
+    Type newWeightTy =
+        RankedTensorType::get(newWeightShape, weightTy.getElementType());
+    weight = rewriter.create<tosa::TransposeOp>(loc, newWeightTy, weight,
+                                                weightPermValue);
+
+    Attribute resultZeroAttr = rewriter.getZeroAttr(resultETy);
+    Value initTensor = rewriter.create<linalg::InitTensorOp>(
+        loc, resultTy.getShape(), resultETy);
+    Value zero = rewriter.create<arith::ConstantOp>(loc, resultZeroAttr);
+    Value zeroTensor =
+        rewriter.create<linalg::FillOp>(loc, zero, initTensor).getResult(0);
+
+    // Extract the attributes for convolution.
+    llvm::SmallVector<int64_t> stride, dilation;
+    getValuesFromIntArrayAttribute(strideTosaAttr, stride);
+    getValuesFromIntArrayAttribute(dilationTosaAttr, dilation);
+
+    // Create the convolution op.
+    auto strideAttr = DenseIntElementsAttr::get(
+        RankedTensorType::get({2}, rewriter.getI64Type()), stride);
+    auto dilationAttr = DenseIntElementsAttr::get(
+        RankedTensorType::get({2}, rewriter.getI64Type()), dilation);
+
+    // Create maps for the bias broadcasting
+    SmallVector<AffineMap, 4> indexingMaps;
+    indexingMaps.push_back(AffineMap::get(
+        /*dimCount=*/resultTy.getRank(), /*symbolCount=*/0,
+        {rewriter.getAffineDimExpr(3)}, rewriter.getContext()));
+    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(resultTy.getRank()));
+    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(resultTy.getRank()));
+
+    Value biasInitTensor = rewriter.create<linalg::InitTensorOp>(
+        loc, resultTy.getShape(), resultETy);
+
+    if (isQuantized) {
+      auto quantizationInfo =
+          op->getAttr("quantization_info").cast<tosa::ConvOpQuantizationAttr>();
+      auto iZp = rewriter.getI32IntegerAttr(
+          quantizationInfo.input_zp().getValue().getSExtValue());
+      auto kZp = rewriter.getI32IntegerAttr(
+          quantizationInfo.weight_zp().getValue().getSExtValue());
+
+      auto iZpVal = rewriter.create<arith::ConstantOp>(loc, iZp);
+      auto kZpVal = rewriter.create<arith::ConstantOp>(loc, kZp);
+      Value conv =
+          rewriter
+              .create<linalg::Conv2DNhwcHwcfQOp>(
+                  loc, resultTy, ValueRange{input, weight, iZpVal, kZpVal},
+                  ValueRange{zeroTensor}, strideAttr, dilationAttr)
+              ->getResult(0);
+
+      Value result =
+          rewriter
+              .create<linalg::GenericOp>(
+                  loc, resultTy, ValueRange({bias, conv}), biasInitTensor,
+                  indexingMaps, getNParallelLoopsAttrs(resultTy.getRank()),
+                  [&](OpBuilder &nestedBuilder, Location nestedLoc,
+                      ValueRange args) {
+                    Value added = nestedBuilder.create<arith::AddIOp>(
+                        loc, args[0], args[1]);
+                    nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
+                  })
+              .getResult(0);
+      rewriter.replaceOp(op, result);
+      return success();
+    }
+
+    Value conv = rewriter
+                     .create<linalg::Conv2DNhwcHwcfOp>(
+                         loc, resultTy, ValueRange{input, weight},
+                         ValueRange{zeroTensor}, strideAttr, dilationAttr)
+                     ->getResult(0);
+
+    Value result =
+        rewriter
+            .create<linalg::GenericOp>(
+                loc, resultTy, ValueRange({bias, conv}), biasInitTensor,
+                indexingMaps, getNParallelLoopsAttrs(resultTy.getRank()),
+                [&](OpBuilder &nestedBuilder, Location nestedLoc,
+                    ValueRange args) {
+                  Value added = nestedBuilder.create<arith::AddFOp>(
+                      loc, args[0], args[1]);
+                  nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
+                })
+            .getResult(0);
+
+    rewriter.replaceOp(op, result);
+    return success();
+  }
+};
+
+class DepthwiseConvConverter
+    : public OpConversionPattern<tosa::DepthwiseConv2DOp> {
+public:
+  using OpConversionPattern<tosa::DepthwiseConv2DOp>::OpConversionPattern;
+  LogicalResult
+  matchAndRewrite(tosa::DepthwiseConv2DOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const final {
+    Location loc = op->getLoc();
+    Value input = op->getOperand(0);
+    Value weight = op->getOperand(1);
+    Value bias = op->getOperand(2);
+
+    ShapedType inputTy = input.getType().cast<ShapedType>();
+    ShapedType weightTy = weight.getType().cast<ShapedType>();
+    ShapedType biasTy = bias.getType().cast<ShapedType>();
+    ShapedType resultTy = op->getResult(0).getType().cast<ShapedType>();
+
+    Type inputETy = inputTy.getElementType();
+    Type resultETy = resultTy.getElementType();
+
+    auto padAttr = op->getAttr("pad").cast<ArrayAttr>();
+    auto strideTosaAttr = op->getAttr("stride").cast<ArrayAttr>();
+    auto dilationTosaAttr = op->getAttr("dilation").cast<ArrayAttr>();
+
+    bool isQuantized = op->hasAttr("quantization_info");
+    IntegerAttr iZp;
+    IntegerAttr kZp;
+    if (isQuantized) {
+      auto quantizationInfo =
+          op->getAttr("quantization_info").cast<tosa::ConvOpQuantizationAttr>();
+      iZp = rewriter.getI32IntegerAttr(
+          quantizationInfo.input_zp().getValue().getSExtValue());
+      kZp = rewriter.getI32IntegerAttr(
+          quantizationInfo.weight_zp().getValue().getSExtValue());
+    }
+
+    if (!inputTy.hasStaticShape() || !weightTy.hasStaticShape() ||
+        !biasTy.hasStaticShape() || !resultTy.hasStaticShape())
+      return rewriter.notifyMatchFailure(op,
+                                         "tosa.conv ops require static shapes");
+
+    auto weightShape = weightTy.getShape();
+    auto resultShape = resultTy.getShape();
+
+    // Apply padding as necessary.
+    Attribute zeroAttr = rewriter.getZeroAttr(inputETy);
+    if (isQuantized) {
+      auto quantizationInfo =
+          op->getAttr("quantization_info").cast<tosa::ConvOpQuantizationAttr>();
+      auto iZp = quantizationInfo.input_zp().getValue().getSExtValue();
+
+      int64_t intMin =
+          APInt::getSignedMinValue(inputETy.getIntOrFloatBitWidth())
+              .getSExtValue();
+      int64_t intMax =
+          APInt::getSignedMaxValue(inputETy.getIntOrFloatBitWidth())
+              .getSExtValue();
+
+      if (iZp < intMin || iZp > intMax)
+        return rewriter.notifyMatchFailure(
+            op, "tosa.depthwise_conv op quantization has zp outside of input "
+                "range");
+
+      zeroAttr = rewriter.getIntegerAttr(inputETy, iZp);
+    }
+
+    llvm::SmallVector<int64_t> pad;
+    pad.resize(2, 0);
+    getValuesFromIntArrayAttribute(padAttr, pad);
+    pad.resize(pad.size() + 2, 0);
+
+    input = applyPad(loc, input, pad, zeroAttr, rewriter);
+
+    // Extract the attributes for convolution.
+    llvm::SmallVector<int64_t> stride, dilation;
+    getValuesFromIntArrayAttribute(strideTosaAttr, stride);
+    getValuesFromIntArrayAttribute(dilationTosaAttr, dilation);
+
+    // Create the convolution op.
+    auto strideAttr = DenseIntElementsAttr::get(
+        RankedTensorType::get({2}, rewriter.getI64Type()), stride);
+    auto dilationAttr = DenseIntElementsAttr::get(
+        RankedTensorType::get({2}, rewriter.getI64Type()), dilation);
+    ShapedType linalgConvTy =
+        RankedTensorType::get({resultShape[0], resultShape[1], resultShape[2],
+                               weightShape[2], weightShape[3]},
+                              resultETy);
+
+    // Broadcast the initial value to the output tensor before convolving.
+    SmallVector<AffineMap, 4> indexingMaps;
+    indexingMaps.push_back(AffineMap::get(
+        /*dimCount=*/resultTy.getRank(), /*symbolCount=*/0,
+        {rewriter.getAffineDimExpr(3)}, rewriter.getContext()));
+    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(resultTy.getRank()));
+    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(resultTy.getRank()));
+
+    Attribute resultZeroAttr = rewriter.getZeroAttr(resultETy);
+    Value initTensor = rewriter.create<linalg::InitTensorOp>(
+        loc, linalgConvTy.getShape(), resultETy);
+    Value zero = rewriter.create<arith::ConstantOp>(loc, resultZeroAttr);
+    Value zeroTensor =
+        rewriter.create<linalg::FillOp>(loc, zero, initTensor).getResult(0);
+
+    Value biasInitTensor = rewriter.create<linalg::InitTensorOp>(
+        loc, resultTy.getShape(), resultETy);
+    if (!isQuantized) {
+      Value conv = rewriter
+                       .create<linalg::DepthwiseConv2DNhwcHwcmOp>(
+                           loc, linalgConvTy, ValueRange{input, weight},
+                           ValueRange{zeroTensor}, strideAttr, dilationAttr)
+                       .getResult(0);
+      Value convReshape = rewriter.create<tosa::ReshapeOp>(
+          loc, resultTy, conv, rewriter.getI64ArrayAttr(resultTy.getShape()));
+      Value result =
+          rewriter
+              .create<linalg::GenericOp>(
+                  loc, resultTy, ValueRange({bias, convReshape}),
+                  biasInitTensor, indexingMaps,
+                  getNParallelLoopsAttrs(resultTy.getRank()),
+                  [&](OpBuilder &nestedBuilder, Location nestedLoc,
+                      ValueRange args) {
+                    Value added = nestedBuilder.create<arith::AddFOp>(
+                        loc, args[0], args[1]);
+                    nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
+                  })
+              .getResult(0);
+      rewriter.replaceOp(op, result);
+    } else {
+      auto iZpVal = rewriter.create<arith::ConstantOp>(loc, iZp);
+      auto kZpVal = rewriter.create<arith::ConstantOp>(loc, kZp);
+      Value conv =
+          rewriter
+              .create<linalg::DepthwiseConv2DNhwcHwcmQOp>(
+                  loc, linalgConvTy, ValueRange{input, weight, iZpVal, kZpVal},
+                  ValueRange{zeroTensor}, strideAttr, dilationAttr)
+              .getResult(0);
+      Value convReshape = rewriter.create<tosa::ReshapeOp>(
+          loc, resultTy, conv, rewriter.getI64ArrayAttr(resultTy.getShape()));
+      Value result =
+          rewriter
+              .create<linalg::GenericOp>(
+                  loc, resultTy, ValueRange({bias, convReshape}),
+                  biasInitTensor, indexingMaps,
+                  getNParallelLoopsAttrs(resultTy.getRank()),
+                  [&](OpBuilder &nestedBuilder, Location nestedLoc,
+                      ValueRange args) {
+                    Value added = nestedBuilder.create<arith::AddIOp>(
+                        loc, args[0], args[1]);
+                    nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
+                  })
+              .getResult(0);
+      rewriter.replaceOp(op, result);
+    }
+    return success();
+  }
+};
+
+class MatMulConverter : public OpConversionPattern<tosa::MatMulOp> {
+public:
+  using OpConversionPattern<tosa::MatMulOp>::OpConversionPattern;
+  LogicalResult
+  matchAndRewrite(tosa::MatMulOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const final {
+    Location loc = op.getLoc();
+
+    auto outputTy = op.getType().cast<ShapedType>();
+    auto outputElementTy = outputTy.getElementType();
+
+    auto firstOperandTy = op->getOperand(0).getType().cast<ShapedType>();
+    auto secondOperandTy = op->getOperand(1).getType().cast<ShapedType>();
+
+    SmallVector<Value> dynDims;
+    dynDims.resize(op->getResult(0).getType().cast<ShapedType>().getRank());
+
+    if (!firstOperandTy.hasRank() || firstOperandTy.isDynamicDim(0)) {
+      dynDims[0] = rewriter.create<tensor::DimOp>(loc, op->getOperand(0), 0);
+    }
+
+    if (!firstOperandTy.hasRank() || firstOperandTy.isDynamicDim(1)) {
+      dynDims[1] = rewriter.create<tensor::DimOp>(loc, op->getOperand(0), 1);
+    }
+
+    if (!secondOperandTy.hasRank() || secondOperandTy.isDynamicDim(2)) {
+      dynDims[2] = rewriter.create<tensor::DimOp>(loc, op->getOperand(1), 2);
+    }
+
+    SmallVector<Value> filteredDims = filterDynamicDims(dynDims);
+
+    auto zeroAttr = rewriter.getZeroAttr(outputElementTy);
+    Value zero = rewriter.create<arith::ConstantOp>(loc, zeroAttr);
+    auto initTensor = rewriter.create<linalg::InitTensorOp>(
+        loc, filteredDims, outputTy.getShape(), outputTy.getElementType());
+    Value zeroTensor =
+        rewriter.create<linalg::FillOp>(loc, zero, initTensor).getResult(0);
+    if (!op.quantization_info()) {
+      rewriter.replaceOpWithNewOp<linalg::BatchMatmulOp>(
+          op, TypeRange{op.getType()}, ValueRange{adaptor.a(), adaptor.b()},
+          ValueRange{zeroTensor});
+      return success();
+    }
+
+    auto quantizationInfo = op.quantization_info().getValue();
+    auto aZp = rewriter.create<arith::ConstantOp>(
+        loc, rewriter.getI32IntegerAttr(
+                 quantizationInfo.a_zp().getValue().getSExtValue()));
+    auto bZp = rewriter.create<arith::ConstantOp>(
+        loc, rewriter.getI32IntegerAttr(
+                 quantizationInfo.b_zp().getValue().getSExtValue()));
+    rewriter.replaceOpWithNewOp<linalg::QuantizedBatchMatmulOp>(
+        op, TypeRange{op.getType()},
+        ValueRange{adaptor.a(), adaptor.b(), aZp, bZp}, zeroTensor);
+
+    return success();
+  }
+};
+
+class FullyConnectedConverter
+    : public OpConversionPattern<tosa::FullyConnectedOp> {
+public:
+  using OpConversionPattern<tosa::FullyConnectedOp>::OpConversionPattern;
+  LogicalResult
+  matchAndRewrite(tosa::FullyConnectedOp op, OpAdaptor adaptor,
+                  ConversionPatternRewriter &rewriter) const final {
+    Location loc = op.getLoc();
+    auto outputTy = op.getType().cast<ShapedType>();
+    auto input = op.input();
+    auto inputTy = input.getType().cast<ShapedType>();
+
+    auto bias = op.bias();
+
+    auto weight = op.weight();
+    auto weightTy = weight.getType().cast<ShapedType>();
+    auto weightShape = weightTy.getShape();
+
+    auto outputETy = outputTy.getElementType();
+
+    SmallVector<Value> dynDims;
+    dynDims.resize(op->getResult(0).getType().cast<ShapedType>().getRank());
+
+    if (!inputTy.hasRank() || inputTy.isDynamicDim(0)) {
+      dynDims[0] = rewriter.create<tensor::DimOp>(loc, input, 0);
+    }
+
+    if (!weightTy.hasRank() || weightTy.isDynamicDim(0)) {
+      dynDims[1] = rewriter.create<tensor::DimOp>(loc, weight, 0);
+    }
+
+    SmallVector<Value> filteredDims = filterDynamicDims(dynDims);
+
+    // Creating maps for the output of MatMul and the bias
+    SmallVector<AffineMap, 4> indexingMaps;
+
+    // Broadcast the bias.
+    indexingMaps.push_back(AffineMap::get(/*dimCount=*/2, /*symbolCount=*/0,
+                                          {rewriter.getAffineDimExpr(1)},
+                                          rewriter.getContext()));
+
+    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(outputTy.getRank()));
+    indexingMaps.push_back(rewriter.getMultiDimIdentityMap(outputTy.getRank()));
+
+    auto initTensor = rewriter.create<linalg::InitTensorOp>(
+        loc, filteredDims, outputTy.getShape(), outputTy.getElementType());
+
+    // When quantized, the input elemeny type is not the same as the output
+    Attribute resultZeroAttr = rewriter.getZeroAttr(outputETy);
+    Value zero = rewriter.create<arith::ConstantOp>(loc, resultZeroAttr);
+    Value zeroTensor =
+        rewriter.create<linalg::FillOp>(loc, zero, initTensor).getResult(0);
+
+    SmallVector<int64_t> permutation{1, 0};
+    auto permutationAttr = DenseIntElementsAttr::get(
+        RankedTensorType::get({2}, rewriter.getI64Type()), permutation);
+    Value permutationValue =
+        rewriter.create<arith::ConstantOp>(loc, permutationAttr);
+
+    SmallVector<int64_t> newWeightShape{weightShape[1], weightShape[0]};
+    Type newWeightTy =
+        RankedTensorType::get(newWeightShape, weightTy.getElementType());
+
+    Value transposedWeight = rewriter.create<tosa::TransposeOp>(
+        loc, newWeightTy, weight, permutationValue);
+
+    auto biasInitTensor =
+        rewriter
+            .create<linalg::InitTensorOp>(loc, filteredDims,
+                                          outputTy.getShape(), outputETy)
+            ->getResults();
+
+    if (!op.quantization_info()) {
+      Value matmul = rewriter
+                         .create<linalg::MatmulOp>(
+                             loc, TypeRange{op.getType()},
+                             ValueRange{input, transposedWeight}, zeroTensor)
+                         ->getResult(0);
+
+      Value result =
+          rewriter
+              .create<linalg::GenericOp>(
+                  loc, outputTy, ValueRange({bias, matmul}), biasInitTensor,
+                  indexingMaps, getNParallelLoopsAttrs(outputTy.getRank()),
+                  [&](OpBuilder &nestedBuilder, Location nestedLoc,
+                      ValueRange args) {
+                    Value added = nestedBuilder.create<arith::AddFOp>(
+                        loc, args[0], args[1]);
+                    nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
+                  })
+              .getResult(0);
+      rewriter.replaceOp(op, result);
+      return success();
+    }
+
+    auto quantizationInfo = op.quantization_info().getValue();
+    auto inputZp = rewriter.create<arith::ConstantOp>(
+        loc, rewriter.getI32IntegerAttr(
+                 quantizationInfo.input_zp().getValue().getSExtValue()));
+    auto outputZp = rewriter.create<arith::ConstantOp>(
+        loc, rewriter.getI32IntegerAttr(
+                 quantizationInfo.weight_zp().getValue().getSExtValue()));
+    Value matmul =
+        rewriter
+            .create<linalg::QuantizedMatmulOp>(
+                loc, TypeRange{op.getType()},
+                ValueRange{input, transposedWeight, inputZp, outputZp},
+                zeroTensor)
+            ->getResult(0);
+    Value result =
+        rewriter
+            .create<linalg::GenericOp>(
+                loc, outputTy, ValueRange({bias, matmul}), biasInitTensor,
+                indexingMaps, getNParallelLoopsAttrs(outputTy.getRank()),
+                [&](OpBuilder &nestedBuilder, Location nestedLoc,
+                    ValueRange args) {
+                  Value added = nestedBuilder.create<arith::AddIOp>(
+                      loc, args[0], args[1]);
+                  nestedBuilder.create<linalg::YieldOp>(nestedLoc, added);
+                })
+            .getResult(0);
+    rewriter.replaceOp(op, result);
+    return success();
+  }
+};
+
+class MaxPool2dConverter : public OpRewritePattern<tosa::MaxPool2dOp> {
+public:
+  using OpRewritePattern<tosa::MaxPool2dOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(tosa::MaxPool2dOp op,
+                                PatternRewriter &rewriter) const final {
+    Location loc = op.getLoc();
+    Value input = op.input();
+    ShapedType inputTy = input.getType().cast<ShapedType>();
+
+    ShapedType resultTy = op.getType().template cast<ShapedType>();
+    Type resultETy = inputTy.getElementType();
+
+    if (!inputTy.hasStaticShape())
+      return failure();
+
+    // Determine what the initial value needs to be for the max pool op.
+    Attribute initialAttr;
+    if (resultETy.isF32())
+      initialAttr = rewriter.getFloatAttr(
+          resultETy,
+          APFloat::getLargest(resultETy.cast<FloatType>().getFloatSemantics(),
+                              true));
+
+    if (resultETy.isa<IntegerType>())
+      initialAttr = rewriter.getIntegerAttr(
+          resultETy,
+          APInt::getSignedMinValue(resultETy.getIntOrFloatBitWidth()));
+
+    if (!initialAttr)
+      return rewriter.notifyMatchFailure(
+          op, "Unsupported initial value for tosa.maxpool_2d op");
+
+    // Apply padding as necessary.
+    llvm::SmallVector<int64_t> pad;
+    pad.resize(2, 0);
+    getValuesFromIntArrayAttribute(op.pad(), pad);
+    pad.resize(pad.size() + 2, 0);
+    Value paddedInput = applyPad(loc, input, pad, initialAttr, rewriter);
+
+    Value initialValue = rewriter.create<arith::ConstantOp>(loc, initialAttr);
+
+    SmallVector<int64_t> kernel, stride;
+    getValuesFromIntArrayAttribute(op.kernel(), kernel);
+    getValuesFromIntArrayAttribute(op.stride(), stride);
+
+    Attribute strideAttr = rewriter.getI64VectorAttr(stride);
+    Attribute dilationAttr = rewriter.getI64VectorAttr({1, 1});
+
+    // Create the linalg op that performs pooling.
+    Value initTensor = rewriter.create<linalg::InitTensorOp>(
+        loc, resultTy.getShape(), resultTy.getElementType());
+
+    Value filledInitTensor =
+        rewriter.create<linalg::FillOp>(loc, initialValue, initTensor).result();
+
+    Value fakeWindowDims =
+        rewriter.create<linalg::InitTensorOp>(loc, kernel, resultETy);
+
+    rewriter.replaceOpWithNewOp<linalg::PoolingNhwcMaxOp>(
+        op, ArrayRef<Type>{resultTy}, ValueRange{paddedInput, fakeWindowDims},
+        filledInitTensor, strideAttr, dilationAttr);
+    return success();
+  }
+};
+
+class AvgPool2dConverter : public OpRewritePattern<tosa::AvgPool2dOp> {
+public:
+  using OpRewritePattern<tosa::AvgPool2dOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(tosa::AvgPool2dOp op,
+                                PatternRewriter &rewriter) const final {
+    Location loc = op.getLoc();
+    Value input = op.input();
+    ShapedType inputTy = input.getType().cast<ShapedType>();
+    Type inElementTy = inputTy.getElementType();
+
+    ShapedType resultTy = op.getType().template cast<ShapedType>();
+    Type resultETy = op.getType().cast<ShapedType>().getElementType();
+
+    Type accETy =
+        inElementTy.isa<IntegerType>() ? rewriter.getI32Type() : inElementTy;
+    ShapedType accTy = resultTy.clone(accETy);
+
+    if (!inputTy.hasStaticShape())
+      return failure();
+
+    // Apply padding as necessary.
+    llvm::SmallVector<int64_t> pad;
+    pad.resize(2, 0);
+    getValuesFromIntArrayAttribute(op.pad(), pad);
+    pad.resize(pad.size() + 2, 0);
+    Attribute padAttr = rewriter.getZeroAttr(inElementTy);
+    Value paddedInput = applyPad(loc, input, pad, padAttr, rewriter);
+
+    Attribute initialAttr = rewriter.getZeroAttr(accETy);
+    Value initialValue = rewriter.create<arith::ConstantOp>(loc, initialAttr);
+
+    SmallVector<int64_t> kernel, stride;
+    getValuesFromIntArrayAttribute(op.kernel(), kernel);
+    getValuesFromIntArrayAttribute(op.stride(), stride);
+
+    Attribute strideAttr = rewriter.getI64VectorAttr(stride);
+    Attribute dilationAttr = rewriter.getI64VectorAttr({1, 1});
+
+    // Create the linalg op that performs pooling.
+    Value poolInitTensor =
+        rewriter.create<linalg::InitTensorOp>(loc, accTy.getShape(), accETy);
+
+    Value filledInitTensor =
+        rewriter.create<linalg::FillOp>(loc, initialValue, poolInitTensor)
+            .result();
+
+    Value fakeWindowDims =
+        rewriter.create<linalg::InitTensorOp>(loc, kernel, accETy);
+
+    // Sum across the pooled region.
+    Value poolingOp = rewriter
+                          .create<linalg::PoolingNhwcSumOp>(
+                              loc, ArrayRef<Type>{accTy},
+                              ValueRange{paddedInput, fakeWindowDims},
+                              filledInitTensor, strideAttr, dilationAttr)
+                          .getResult(0);
+
+    // Normalize the summed value by the number of elements grouped in each
+    // pool.
+    auto poolingOpTy = poolingOp.getType().cast<ShapedType>();
+    auto affineMap = rewriter.getMultiDimIdentityMap(resultTy.getRank());
+
+    Value genericInitTensor = rewriter.create<linalg::InitTensorOp>(
+        loc, resultTy.getShape(), resultETy);
+
+    auto genericOp = rewriter.create<linalg::GenericOp>(
+        loc, ArrayRef<Type>({resultTy}), ValueRange{poolingOp},
+        ValueRange{genericInitTensor},
+        ArrayRef<AffineMap>({affineMap, affineMap}),
+        getNParallelLoopsAttrs(resultTy.getRank()),
+        [&](OpBuilder &b, Location loc, ValueRange args) {
+          auto zero = rewriter.create<arith::ConstantIndexOp>(loc, 0);
+          auto one = rewriter.create<arith::ConstantIndexOp>(loc, 1);
+          auto iH = rewriter.create<arith::ConstantIndexOp>(
+              loc, poolingOpTy.getDimSize(1) - 1);
+          auto iW = rewriter.create<arith::ConstantIndexOp>(
+              loc, poolingOpTy.getDimSize(2) - 1);
+
+          // Compute the indices from either end.
+          auto y0 = rewriter.create<linalg::IndexOp>(loc, 1);
+          auto x0 = rewriter.create<linalg::IndexOp>(loc, 2);
+          auto y1 = rewriter.create<arith::SubIOp>(loc, iH, y0);
+          auto x1 = rewriter.create<arith::SubIOp>(loc, iW, x0);
+
+          // Determines what the portion of valid input is covered by the
+          // kernel.
+          auto padFn = [&](Value v, Value x, int64_t pad) -> Value {
+            if (pad == 0)
+              return v;
+
+            auto padVal = rewriter.create<arith::ConstantIndexOp>(loc, pad);
+            Value dx = rewriter.create<arith::SubIOp>(loc, x, padVal);
+
+            Value cmp = rewriter.create<arith::CmpIOp>(
+                loc, arith::CmpIPredicate::slt, dx, zero);
+            Value offset = rewriter.create<mlir::SelectOp>(loc, cmp, dx, zero);
+            return rewriter.create<arith::AddIOp>(loc, v, offset)->getResult(0);
+          };
+
+          // Compute the vertical component of coverage.
+          auto kH0 = rewriter.create<arith::ConstantIndexOp>(loc, kernel[0]);
+          auto kH1 = padFn(kH0, y0, pad[2]);
+          auto kH2 = padFn(kH1, y1, pad[3]);
+          auto kHCmp = rewriter.create<arith::CmpIOp>(
+              loc, arith::CmpIPredicate::slt, kH2, one);
+          auto kH3 = rewriter.create<SelectOp>(loc, kHCmp, one, kH2);
+
+          // compute the horizontal component of coverage.
+          auto kW0 = rewriter.create<arith::ConstantIndexOp>(loc, kernel[1]);
+          auto kW1 = padFn(kW0, x0, pad[4]);
+          auto kW2 = padFn(kW1, x1, pad[5]);
+          auto kWCmp = rewriter.create<arith::CmpIOp>(
+              loc, arith::CmpIPredicate::slt, kW2, one);
+          auto kW3 = rewriter.create<SelectOp>(loc, kWCmp, one, kW2);
+
+          // Compute the total number of elements and normalize.
+          Value count = rewriter.create<arith::MulIOp>(loc, kH3, kW3);
+          auto countI = rewriter.create<arith::IndexCastOp>(
+              loc, rewriter.getI32Type(), count);
+
+          // Divide by the number of summed values. For floats this is just
+          // a div however for quantized values input normalization had
+          // to be applied.
+          Value poolVal = args[0];
+          if (accETy.isa<FloatType>()) {
+            auto countF = rewriter.create<arith::SIToFPOp>(loc, accETy, countI);
+            poolVal = rewriter.create<arith::DivFOp>(loc, poolVal, countF)
+                          ->getResult(0);
+          } else {
+
+            // If we have quantization information we need to apply an offset
+            // for the input zp value.
+            if (op.quantization_info()) {
+              auto quantizationInfo = op.quantization_info().getValue();
+              auto inputZp = rewriter.create<arith::ConstantOp>(
+                  loc, quantizationInfo.input_zp());
+              Value offset =
+                  rewriter.create<arith::MulIOp>(loc, accETy, countI, inputZp);
+              poolVal =
+                  rewriter.create<arith::SubIOp>(loc, accETy, poolVal, offset);
+            }
+
+            // Compute the multiplier and shift values for the quantization
+            // normalization. Preferably we would want to compute more bits
+            // however 32-bits should be enough for compute. Honestly we
+            // should probably straight divide.
+            int64_t numerator = ((1 << 30) + 1);
+            int64_t shift = 30;
+
+            Value numeratorVal = rewriter.create<arith::ConstantOp>(
+                loc, rewriter.getI32IntegerAttr(numerator));
+            Value multiplierVal =
+                rewriter
+                    .create<arith::DivUIOp>(loc, rewriter.getI32Type(),
+                                            numeratorVal, countI)
+                    .getResult();
+            Value shiftVal = rewriter.create<arith::ConstantOp>(
+                loc, rewriter.getI8IntegerAttr(shift));
+
+            auto scaled =
+                rewriter
+                    .create<tosa::ApplyScaleOp>(
+                        loc, rewriter.getI32Type(), poolVal, multiplierVal,
+                        shiftVal, rewriter.getBoolAttr(false))
+                    .getResult();
+
+            // If we have quantization information we need to apply output
+            // zeropoint.
+            if (op.quantization_info()) {
+              auto quantizationInfo = op.quantization_info().getValue();
+              auto outputZp = rewriter.create<arith::ConstantOp>(
+                  loc, quantizationInfo.output_zp());
+              scaled = rewriter.create<arith::AddIOp>(loc, scaled, outputZp)
+                           .getResult();
+            }
+
+            // Apply Clip.
+            int64_t outBitwidth = resultETy.getIntOrFloatBitWidth();
+
+            auto min = rewriter.create<arith::ConstantIntOp>(
+                loc, APInt::getSignedMinValue(outBitwidth).getSExtValue(),
+                accETy);
+            auto max = rewriter.create<arith::ConstantIntOp>(
+                loc, APInt::getSignedMaxValue(outBitwidth).getSExtValue(),
+                accETy);
+            auto clamp = clampHelper<arith::CmpIOp>(
+                loc, scaled, min, max, arith::CmpIPredicate::slt, rewriter);
+
+            poolVal = clamp;
+            // Convert type.
+            if (resultETy != clamp.getType()) {
+              poolVal =
+                  rewriter.create<arith::TruncIOp>(loc, resultETy, poolVal);
+            }
+          }
+
+          rewriter.create<linalg::YieldOp>(loc, poolVal);
+        });
+
+    rewriter.replaceOp(op, genericOp.getResult(0));
+    return success();
+  }
+};
+
+} // namespace
+
+void mlir::tosa::populateTosaToLinalgNamedConversionPatterns(
+    RewritePatternSet *patterns) {
+  patterns->add<
+      // clang-format off
+      ConvConverter,
+      DepthwiseConvConverter,
+      MatMulConverter,
+      MaxPool2dConverter,
+      AvgPool2dConverter,
+      FullyConnectedConverter>(patterns->getContext());
+  // clang-format on
+}

diff  --git a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamedPass.cpp b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamedPass.cpp
new file mode 100644
index 0000000000000..f5f6ac1a5469c
--- /dev/null
+++ b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamedPass.cpp
@@ -0,0 +1,68 @@
+//===- TosaToLinalgPass.cpp - Lowering Tosa to Linalg Dialect -------------===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+//
+// This transformation pass legalizes Tosa operations to the Linalg dialect.
+//
+//===----------------------------------------------------------------------===//
+
+#include "../PassDetail.h"
+#include "mlir/Conversion/TosaToLinalg/TosaToLinalg.h"
+#include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
+#include "mlir/Dialect/Linalg/IR/Linalg.h"
+#include "mlir/Dialect/Math/IR/Math.h"
+#include "mlir/Dialect/SCF/SCF.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/QuantUtils.h"
+#include "mlir/IR/PatternMatch.h"
+#include "mlir/Pass/PassManager.h"
+#include "mlir/Transforms/DialectConversion.h"
+#include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+
+using namespace mlir;
+
+namespace {
+struct TosaToLinalgNamed : public TosaToLinalgNamedBase<TosaToLinalgNamed> {
+public:
+  void getDependentDialects(DialectRegistry &registry) const override {
+    registry.insert<arith::ArithmeticDialect, linalg::LinalgDialect,
+                    math::MathDialect, StandardOpsDialect,
+                    tensor::TensorDialect, scf::SCFDialect>();
+  }
+
+  void runOnFunction() override {
+    RewritePatternSet patterns(&getContext());
+    ConversionTarget target(getContext());
+    target.addLegalDialect<linalg::LinalgDialect, StandardOpsDialect,
+                           tosa::TosaDialect, tensor::TensorDialect,
+                           scf::SCFDialect>();
+
+    // Not every TOSA op can be legalized to linalg.
+    target.addIllegalOp<tosa::Conv2DOp>();
+    target.addIllegalOp<tosa::DepthwiseConv2DOp>();
+    target.addIllegalOp<tosa::MaxPool2dOp>();
+    target.addIllegalOp<tosa::AvgPool2dOp>();
+    target.addIllegalOp<tosa::MatMulOp>();
+    target.addIllegalOp<tosa::FullyConnectedOp>();
+
+    target.markUnknownOpDynamicallyLegal([](Operation *) { return true; });
+
+    FuncOp func = getFunction();
+    mlir::tosa::populateTosaToLinalgNamedConversionPatterns(&patterns);
+    if (failed(applyFullConversion(func, target, std::move(patterns))))
+      signalPassFailure();
+  }
+};
+} // namespace
+
+std::unique_ptr<Pass> mlir::tosa::createTosaToLinalgNamed() {
+  return std::make_unique<TosaToLinalgNamed>();
+}

diff  --git a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalgPass.cpp b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalgPass.cpp
index 8f4f872c88606..3813ba3451374 100644
--- a/mlir/lib/Conversion/TosaToLinalg/TosaToLinalgPass.cpp
+++ b/mlir/lib/Conversion/TosaToLinalg/TosaToLinalgPass.cpp
@@ -26,6 +26,7 @@
 #include "mlir/Pass/PassManager.h"
 #include "mlir/Transforms/DialectConversion.h"
 #include "mlir/Transforms/GreedyPatternRewriteDriver.h"
+#include "mlir/Transforms/Passes.h"
 
 using namespace mlir;
 
@@ -67,6 +68,9 @@ std::unique_ptr<Pass> mlir::tosa::createTosaToLinalg() {
 }
 
 void mlir::tosa::addTosaToLinalgPasses(OpPassManager &pm) {
+  pm.addNestedPass<FuncOp>(createTosaMakeBroadcastablePass());
+  pm.addNestedPass<FuncOp>(createTosaToLinalgNamed());
+  pm.addNestedPass<FuncOp>(mlir::createCanonicalizerPass());
   pm.addNestedPass<FuncOp>(createTosaMakeBroadcastablePass());
   pm.addNestedPass<FuncOp>(createTosaToLinalg());
 }

diff  --git a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir
new file mode 100644
index 0000000000000..f5814883cc49c
--- /dev/null
+++ b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir
@@ -0,0 +1,448 @@
+// RUN: mlir-opt --split-input-file --tosa-to-linalg-named %s -verify-diagnostics -o -| FileCheck %s
+
+// CHECK-LABEL: @matmul
+func @matmul(%arg0: tensor<1x5x3xf32>, %arg1: tensor<1x3x6xf32>) -> (tensor<1x5x6xf32>) {
+  // CHECK: [[C0:%.+]] = arith.constant 0
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 5, 6]
+  // CHECK: [[FILLED:%.+]] = linalg.fill([[C0]], [[INIT]]) : f32, tensor<1x5x6xf32> -> tensor<1x5x6xf32>
+  // CHECK: linalg.batch_matmul ins(%arg0, %arg1 : tensor<1x5x3xf32>, tensor<1x3x6xf32>) outs([[FILLED]] : tensor<1x5x6xf32>) -> tensor<1x5x6xf32>
+  %0 = "tosa.matmul"(%arg0, %arg1) : (tensor<1x5x3xf32>, tensor<1x3x6xf32>)  -> (tensor<1x5x6xf32>)
+  return %0 : tensor<1x5x6xf32>
+}
+
+// -----
+
+
+// CHECK-LABEL: @matmul_quantized
+func @matmul_quantized(%arg0: tensor<1x5x3xi8>, %arg1: tensor<1x3x6xi8>) -> (tensor<1x5x6xi32>) {
+  // CHECK: [[C0:%.+]] = arith.constant 0
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 5, 6]
+  // CHECK: [[FILLED:%.+]] = linalg.fill([[C0]], [[INIT]]) : i32, tensor<1x5x6xi32> -> tensor<1x5x6xi32>
+  // CHECK: [[ONE:%.+]] = arith.constant 1
+  // CHECK: [[TWO:%.+]] = arith.constant 2
+  // CHECK: linalg.quantized_batch_matmul ins(%arg0, %arg1, [[ONE]], [[TWO]] : tensor<1x5x3xi8>, tensor<1x3x6xi8>, i32, i32) outs([[FILLED]] : tensor<1x5x6xi32>) -> tensor<1x5x6xi32>
+  %0 = "tosa.matmul"(%arg0, %arg1) {quantization_info = {a_zp = 1 : i32, b_zp = 2 : i32}} : (tensor<1x5x3xi8>, tensor<1x3x6xi8>) -> (tensor<1x5x6xi32>)
+  return %0 : tensor<1x5x6xi32>
+}
+
+// -----
+
+// CHECK-LABEL: @matmul_dyn_batch
+func @matmul_dyn_batch(%arg0: tensor<?x5x3xf32>, %arg1: tensor<?x3x6xf32>) -> (tensor<?x5x6xf32>) {
+  // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: %[[DIM:.+]] = tensor.dim %arg0, %[[C0]]
+  // CHECK: %[[C0_0:.+]] = arith.constant 0
+  // CHECK: %[[INIT:.+]] = linalg.init_tensor [%[[DIM]], 5, 6]
+  // CHECK: %[[FILLED:.+]] = linalg.fill(%[[C0_0]], %[[INIT]]) : f32, tensor<?x5x6xf32> -> tensor<?x5x6xf32>
+  // CHECK: linalg.batch_matmul ins(%arg0, %arg1 : tensor<?x5x3xf32>, tensor<?x3x6xf32>) outs(%[[FILLED]] : tensor<?x5x6xf32>) -> tensor<?x5x6xf32>
+  %0 = "tosa.matmul"(%arg0, %arg1) : (tensor<?x5x3xf32>, tensor<?x3x6xf32>)  -> (tensor<?x5x6xf32>)
+  return %0 : tensor<?x5x6xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @matmul_dyn_independent_dim
+func @matmul_dyn_independent_dim(%arg0: tensor<1x5x3xf32>, %arg1: tensor<1x3x?xf32>) -> (tensor<1x5x?xf32>) {
+  // CHECK: %[[C2:.+]] = arith.constant 2
+  // CHECK: %[[DIM:.+]] = tensor.dim %arg1, %[[C2]]
+  // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: %[[INIT:.+]] = linalg.init_tensor [1, 5, %[[DIM]]]
+  // CHECK: %[[FILLED:.+]] = linalg.fill(%[[C0]], %[[INIT]]) : f32, tensor<1x5x?xf32> -> tensor<1x5x?xf32>
+  // CHECK: linalg.batch_matmul ins(%arg0, %arg1 : tensor<1x5x3xf32>, tensor<1x3x?xf32>) outs(%[[FILLED]] : tensor<1x5x?xf32>) -> tensor<1x5x?xf32>
+  %0 = "tosa.matmul"(%arg0, %arg1) : (tensor<1x5x3xf32>, tensor<1x3x?xf32>)  -> (tensor<1x5x?xf32>)
+  return %0 : tensor<1x5x?xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @matmul_dyn_independent_dim
+func @matmul_dyn_independent_dim(%arg0: tensor<1x5x?xf32>, %arg1: tensor<1x?x6xf32>) -> (tensor<1x5x6xf32>) {
+  // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: %[[INIT:.+]] = linalg.init_tensor [1, 5, 6]
+  // CHECK: %[[FILLED:.+]] = linalg.fill(%[[C0]], %[[INIT]]) : f32, tensor<1x5x6xf32> -> tensor<1x5x6xf32>
+  // CHECK: linalg.batch_matmul ins(%arg0, %arg1 : tensor<1x5x?xf32>, tensor<1x?x6xf32>) outs(%[[FILLED]] : tensor<1x5x6xf32>) -> tensor<1x5x6xf32>
+  %0 = "tosa.matmul"(%arg0, %arg1) : (tensor<1x5x?xf32>, tensor<1x?x6xf32>)  -> (tensor<1x5x6xf32>)
+  return %0 : tensor<1x5x6xf32>
+}
+
+// -----
+
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d1)>
+// CHECK: #[[$MAP2:.*]] = affine_map<(d0, d1) -> (d0, d1)>
+
+// CHECK-LABEL: @fully_connected
+func @fully_connected(%arg0: tensor<5x3xf32>, %arg1: tensor<6x3xf32>, %arg2: tensor<6xf32>) -> (tensor<5x6xf32>) {
+  // CHECK: [[INITT:%.+]] = linalg.init_tensor [5, 6]
+  // CHECK: [[ZERO:%.+]] = arith.constant 0
+  // CHECK: [[FILL:%.+]] = linalg.fill([[ZERO]], [[INITT]])
+  // CHECK: [[PERM:%.+]] = arith.constant dense<[1, 0]>
+  // CHECK: [[TRANSPOSE:%.+]] = "tosa.transpose"(%arg1, [[PERM]])
+  // CHECK: [[INITB:%.+]] = linalg.init_tensor [5, 6]
+  // CHECK: [[MATMUL:%.+]] = linalg.matmul ins(%arg0, [[TRANSPOSE]] : tensor<5x3xf32>, tensor<3x6xf32>) outs([[FILL]] : tensor<5x6xf32>) -> tensor<5x6xf32>
+  // CHECK: [[ADDED:%.+]] = linalg.generic {indexing_maps = [#[[$MAP1]], #[[$MAP2]], #[[$MAP2]]], iterator_types = ["parallel", "parallel"]} ins(%arg2, [[MATMUL]] : tensor<6xf32>, tensor<5x6xf32>) outs([[INITB]] : tensor<5x6xf32>) {
+  // CHECK: ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
+  // CHECK:   [[ADD:%.+]] = arith.addf %arg3, %arg4 : f32
+  // CHECK:   linalg.yield [[ADD]] : f32
+
+  %0 = "tosa.fully_connected"(%arg0, %arg1, %arg2) : (tensor<5x3xf32>, tensor<6x3xf32>, tensor<6xf32>)  -> (tensor<5x6xf32>)
+  return %0 : tensor<5x6xf32>
+}
+
+// -----
+
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d1)>
+// CHECK: #[[$MAP2:.*]] = affine_map<(d0, d1) -> (d0, d1)>
+
+// CHECK-LABEL: @quantized_fully_connected
+func @quantized_fully_connected(%arg0: tensor<5x3xi8>, %arg1: tensor<6x3xi8>, %arg2: tensor<6xi32>) -> (tensor<5x6xi32>) {
+  // CHECK: [[INITT:%.+]] = linalg.init_tensor [5, 6]
+  // CHECK: [[ZERO:%.+]] = arith.constant 0
+  // CHECK: [[FILL:%.+]] = linalg.fill([[ZERO]], [[INITT]])
+  // CHECK: [[PERM:%.+]] = arith.constant dense<[1, 0]>
+  // CHECK: [[TRANSPOSE:%.+]] = "tosa.transpose"(%arg1, [[PERM]])
+  // CHECK: [[INITB:%.+]] = linalg.init_tensor [5, 6]
+  // CHECK: [[ONE:%.+]] = arith.constant 1
+  // CHECK: [[TWO:%.+]] = arith.constant 2
+  // CHECK: [[MATMUL:%.+]] = linalg.quantized_matmul ins(%arg0, [[TRANSPOSE]], [[ONE]], [[TWO]] : tensor<5x3xi8>, tensor<3x6xi8>, i32, i32) outs([[FILL]] : tensor<5x6xi32>) -> tensor<5x6xi32>
+  // CHECK: [[ADDED:%.+]] = linalg.generic {indexing_maps = [#[[$MAP1]], #[[$MAP2]], #[[$MAP2]]], iterator_types = ["parallel", "parallel"]} ins(%arg2, [[MATMUL]] : tensor<6xi32>, tensor<5x6xi32>) outs([[INITB]]
+  // CHECK: ^bb0([[IN1:%.+]]: i32, [[IN2:%.+]]: i32, [[UNUSED:%.+]]: i32):
+  // CHECK:   [[ADD:%.+]] = arith.addi
+  // CHECK:   linalg.yield [[ADD]] : i32
+  %0 = "tosa.fully_connected"(%arg0, %arg1, %arg2) {quantization_info = {input_zp = 1:i32, weight_zp = 2:i32}} : (tensor<5x3xi8>, tensor<6x3xi8>, tensor<6xi32>)  -> (tensor<5x6xi32>)
+  return %0 : tensor<5x6xi32>
+}
+
+// -----
+
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d1)>
+// CHECK: #[[$MAP2:.*]] = affine_map<(d0, d1) -> (d0, d1)>
+
+// CHECK-LABEL: @fully_connected_dyn
+func @fully_connected_dyn(%arg0: tensor<?x3xf32>, %arg1: tensor<6x3xf32>, %arg2: tensor<6xf32>) -> (tensor<?x6xf32>) {
+  // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: %[[DIM:.+]] = tensor.dim %arg0, %[[C0]]
+  // CHECK: %[[INITT:.+]] = linalg.init_tensor [%[[DIM]], 6]
+  // CHECK: %[[ZERO:.+]] = arith.constant 0
+  // CHECK: %[[FILL:.+]] = linalg.fill(%[[ZERO]], %[[INITT]])
+  // CHECK: %[[PERM:.+]] = arith.constant dense<[1, 0]>
+  // CHECK: %[[TRANSPOSE:.+]] = "tosa.transpose"(%arg1, %[[PERM]])
+  // CHECK: %[[INITB:.+]] = linalg.init_tensor [%[[DIM]], 6]
+  // CHECK: %[[MATMUL:.+]] = linalg.matmul ins(%arg0, %[[TRANSPOSE]] : tensor<?x3xf32>, tensor<3x6xf32>) outs(%[[FILL]] : tensor<?x6xf32>) -> tensor<?x6xf32>
+  // CHECK: %[[ADDED:.+]] = linalg.generic {indexing_maps = [#[[$MAP1]], #[[$MAP2]], #[[$MAP2]]], iterator_types = ["parallel", "parallel"]} ins(%arg2, %[[MATMUL]] : tensor<6xf32>, tensor<?x6xf32>) outs(%[[INITB]] : tensor<?x6xf32>) {
+  // CHECK: ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
+  // CHECK:   %[[ADD:.+]] = arith.addf %arg3, %arg4 : f32
+  // CHECK:   linalg.yield %[[ADD]] : f32
+
+  %0 = "tosa.fully_connected"(%arg0, %arg1, %arg2) : (tensor<?x3xf32>, tensor<6x3xf32>, tensor<6xf32>)  -> (tensor<?x6xf32>)
+  return %0 : tensor<?x6xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @max_pool
+func @max_pool(%arg0: tensor<1x6x34x62xf32>) -> () {
+  // CHECK-DAG: [[CONST:%.+]] = arith.constant -3.40282347E+38
+  // CHECK-DAG: [[INIT:%.+]] = linalg.init_tensor [1, 4, 32, 62]
+  // CHECK-DAG: [[FILL:%.+]] = linalg.fill([[CONST]], [[INIT]])
+  // CHECK-DAG: [[KERNEL:%.+]] = linalg.init_tensor [3, 3]
+  // CHECK: linalg.pooling_nhwc_max {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%arg0, [[KERNEL]] : tensor<1x6x34x62xf32>, tensor<3x3xf32>) outs([[FILL]] : tensor<1x4x32x62xf32>)
+  %0 = "tosa.max_pool2d"(%arg0) {pad = [0, 0, 0, 0], kernel = [3, 3], stride = [1, 1]} : (tensor<1x6x34x62xf32>)  -> (tensor<1x4x32x62xf32>)
+  return
+}
+
+// CHECK-LABEL: @max_pool_padded
+func @max_pool_padded(%arg0: tensor<1x6x34x62xf32>) -> () {
+  // CHECK-DAG: [[CONST:%.+]] = arith.constant -3.40282347E+38 : f32
+  // CHECK-DAG: [[PAD:%.+]] = linalg.pad_tensor %arg0 low[0, 0, 0, 0] high[0, 0, 1, 0]
+  // CHECK-DAG:   linalg.yield [[CONST]]
+  // CHECK-DAG: [[INITVAL:%.+]] = arith.constant -3.40282347E+38 : f32
+  // CHECK-DAG: [[INIT:%.+]] = linalg.init_tensor [1, 4, 33, 62]
+  // CHECK-DAG: [[FILL:%.+]] = linalg.fill([[INITVAL]], [[INIT]])
+  // CHECK-DAG: [[KERNEL:%.+]] = linalg.init_tensor [3, 3]
+  // CHECK: linalg.pooling_nhwc_max {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins([[PAD]], [[KERNEL]] : tensor<1x6x35x62xf32>, tensor<3x3xf32>) outs([[FILL]] : tensor<1x4x33x62xf32>)
+  %0 = "tosa.max_pool2d"(%arg0) {pad = [0, 0, 0, 1], kernel = [3, 3], stride = [1, 1]} : (tensor<1x6x34x62xf32>)  -> (tensor<1x4x33x62xf32>)
+  return
+}
+
+// CHECK-LABEL: @max_pool_i8
+func @max_pool_i8(%arg0: tensor<1x6x34x62xi8>) -> () {
+  // CHECK: arith.constant -128
+  // CHECK: linalg.pooling_nhwc_max
+  %0 = "tosa.max_pool2d"(%arg0) {pad = [0, 0, 0, 0], kernel = [3, 3], stride = [1, 1]} : (tensor<1x6x34x62xi8>)  -> (tensor<1x4x32x62xi8>)
+  return
+}
+
+// CHECK-LABEL: @max_pool_i16
+func @max_pool_i16(%arg0: tensor<1x6x34x62xi16>) -> () {
+  // CHECK: arith.constant -32768
+  // CHECK: linalg.pooling_nhwc_max
+  %0 = "tosa.max_pool2d"(%arg0) {pad = [0, 0, 0, 0], kernel = [3, 3], stride = [1, 1]} : (tensor<1x6x34x62xi16>)  -> (tensor<1x4x32x62xi16>)
+  return
+}
+
+// CHECK-LABEL: @max_pool_i32
+func @max_pool_i32(%arg0: tensor<1x6x34x62xi32>) -> () {
+  // CHECK: arith.constant -2147483648
+  // CHECK: linalg.pooling_nhwc_max
+  %0 = "tosa.max_pool2d"(%arg0) {pad = [0, 0, 0, 0], kernel = [3, 3], stride = [1, 1]} : (tensor<1x6x34x62xi32>)  -> (tensor<1x4x32x62xi32>)
+  return
+}
+// -----
+
+// CHECK-LABEL: @avg_pool
+func @avg_pool(%arg0: tensor<1x6x34x62xf32>) -> (tensor<1x5x33x62xf32>) {
+  // Initial piece computes the sum of the pooling region, with appropriate padding.
+  // CHECK: [[CONST:%.+]] = arith.constant 0
+  // CHECK: [[PAD:%.+]] = linalg.pad_tensor %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0]
+  // CHECK: [[CONST:%.+]] = arith.constant 0
+  // CHECK: [[POOLINIT:%.+]] = linalg.init_tensor [1, 5, 33, 62]
+  // CHECK: [[FILL:%.+]] = linalg.fill([[CONST]], [[POOLINIT]])
+  // CHECK: [[KERNEL:%.+]] = linalg.init_tensor [4, 4]
+  // CHECK: [[POOL:%.+]] = linalg.pooling_nhwc_sum {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins([[PAD]], [[KERNEL]] : tensor<1x8x36x62xf32>, tensor<4x4xf32>) outs([[FILL]] : tensor<1x5x33x62xf32>)
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 5, 33, 62]
+  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins([[POOL]] : tensor<1x5x33x62xf32>) outs([[INIT]] : tensor<1x5x33x62xf32>)
+  // CHECK:   [[ZERO:%.0]] = arith.constant 0
+  // CHECK:   [[ONE:%.+]] = arith.constant 1
+  // CHECK:   [[HEIGHT:%.+]] = arith.constant 4
+  // CHECK:   [[WIDTH:%.+]] = arith.constant 32
+  // CHECK:   [[IDX1:%.+]] = linalg.index 1
+  // CHECK:   [[IDX2:%.+]] = linalg.index 2
+
+  // The large block below computes what portion of the kernel is within non-padded input.
+  // CHECK:   [[NY:%.+]] = arith.subi [[HEIGHT]], [[IDX1]]
+  // CHECK:   [[NX:%.+]] = arith.subi [[WIDTH]], [[IDX2]]
+  // CHECK:   [[KH:%.+]] = arith.constant 4
+  // CHECK:   [[PAD0:%.+]] = arith.constant 1
+  // CHECK:   [[SUBP0:%.+]] = arith.subi [[IDX1]], [[PAD0]]
+  // CHECK:   [[P0CMP:%.+]] = arith.cmpi slt, [[SUBP0]], [[ZERO]]
+  // CHECK:   [[SELP0:%.+]] = select [[P0CMP]], [[SUBP0]], [[ZERO]]
+  // CHECK:   [[ADDP0:%.+]] = arith.addi [[KH]], [[SELP0]]
+  // CHECK:   [[PAD1:%.+]] = arith.constant 1
+  // CHECK:   [[SUBP1:%.+]] = arith.subi [[NY]], [[PAD1]]
+  // CHECK:   [[P1CMP:%.+]] = arith.cmpi slt, [[SUBP1]], [[ZERO]]
+  // CHECK:   [[SELP1:%.+]] = select [[P1CMP]], [[SUBP1]], [[ZERO]]
+  // CHECK:   [[ADDP1:%.+]] = arith.addi [[ADDP0]], [[SELP1]]
+  // CHECK:   [[YCMP:%.+]] = arith.cmpi slt, [[ADDP1]], [[ONE]]
+  // CHECK:   [[YSEL:%.+]] = select [[YCMP]], [[ONE]], [[ADDP1]]
+  // CHECK:   [[KW:%.+]] = arith.constant 4 : index
+  // CHECK:   [[PAD2:%.+]] = arith.constant 1 : index
+  // CHECK:   [[SUBP2:%.+]] = arith.subi [[IDX2]], [[PAD2]]
+  // CHECK:   [[P2CMP:%.+]] = arith.cmpi slt, [[SUBP2]], [[ZERO]]
+  // CHECK:   [[SELP2:%.+]] = select [[P2CMP]], [[SUBP2]], [[ZERO]]
+  // CHECK:   [[ADDP2:%.+]] = arith.addi [[KW]], [[SELP2]]
+  // CHECK:   [[PAD3:%.+]] = arith.constant 1 : index
+  // CHECK:   [[SUBP3:%.+]] = arith.subi [[NX]], [[PAD3]]
+  // CHECK:   [[P3CMP:%.+]] = arith.cmpi slt, [[SUBP3]], [[ZERO]]
+  // CHECK:   [[SELP3:%.+]] = select [[P3CMP]], [[SUBP3]], [[ZERO]]
+  // CHECK:   [[ADDP3:%.+]] = arith.addi [[ADDP2]], [[SELP3]]
+  // CHECK:   [[XCMP:%.+]] = arith.cmpi slt, [[ADDP3]], [[ONE]]
+  // CHECK:   [[XSEL:%.+]] = select [[XCMP]], [[ONE]], [[ADDP3]]
+
+  // Given the valid coverage of the pooling region, normalize the summation.
+  // CHECK:   [[C:%.+]] = arith.muli [[YSEL]], [[XSEL]]
+  // CHECK:   [[CI:%.+]] = arith.index_cast [[C]]
+  // CHECK:   [[CF:%.+]] = arith.sitofp [[CI]]
+  // CHECK:   [[RESULT:%.+]] = arith.divf %arg1, [[CF]]
+  // CHECK:   linalg.yield [[RESULT]]
+  %0 = "tosa.avg_pool2d"(%arg0) {pad = [1, 1, 1, 1], kernel = [4, 4], stride = [1, 1]} : (tensor<1x6x34x62xf32>)  -> (tensor<1x5x33x62xf32>)
+  return %0 : tensor<1x5x33x62xf32>
+}
+
+// -----
+
+// CHECK-LABEL: @avg_pool_i8
+func @avg_pool_i8(%arg0 : tensor<1x128x128x2xi8>) -> () {
+
+  // CHECK: linalg.pooling_nhwc_sum
+  // CHECK: linalg.generic
+
+  // CHECK: %[[INZP:.+]] = arith.constant -128
+  // CHECK: %[[INZP_OFF:.+]] = arith.muli %{{.+}}, %[[INZP]]
+  // CHECK: %[[OFFSETED:.+]] = arith.subi %arg1, %[[INZP_OFF]]
+  // CHECK: %[[NUMERATOR:.+]] = arith.constant 1073741825
+  // CHECK: %[[MULTIPLIER:.+]] = arith.divui %[[NUMERATOR]], %{{.+}}
+  // CHECK: %[[SHIFT:.+]] = arith.constant 30
+  // CHECK: %[[SCALE:.+]] = "tosa.apply_scale"(%{{.+}}, %[[MULTIPLIER]], %[[SHIFT]]) {double_round = false}
+  // CHECK: %[[OUTZP:.+]] = arith.constant -128
+  // CHECK: %[[OUT:.+]] = arith.addi %[[SCALE]], %[[OUTZP]]
+  // CHECK: %[[MIN:.+]] = arith.constant -128
+  // CHECK: %[[MAX:.+]] = arith.constant 127
+  // CHECK: %[[CMP_MIN:.+]] = arith.cmpi slt, %[[OUT]], %[[MIN]]
+  // CHECK: %[[CLMP_MIN:.+]] = select %[[CMP_MIN]], %[[MIN]], %[[OUT]]
+  // CHECK: %[[CMP_MAX:.+]] = arith.cmpi slt, %[[MAX]], %[[OUT]]
+  // CHECK: %[[CLMP_MAX:.+]] = select %[[CMP_MAX]], %[[MAX]], %[[CLMP_MIN]]
+  // CHECK: %[[TRUNC:.+]] = arith.trunci %[[CLMP_MAX]]
+  // CHECK: linalg.yield %[[TRUNC]]
+  %0 = "tosa.avg_pool2d"(%arg0) {kernel = [4, 4], pad = [0, 0, 0, 0], quantization_info = {input_zp = -128 : i32, output_zp = -128 : i32}, stride = [4, 4]} : (tensor<1x128x128x2xi8>) -> tensor<1x32x32x2xi8>
+  return
+}
+
+// -----
+
+// CHECK-LABEL: @avg_pool_i16
+func @avg_pool_i16(%arg0 : tensor<1x128x128x2xi16>) -> () {
+
+  // CHECK: linalg.pooling_nhwc_sum
+  // CHECK: linalg.generic
+
+  // CHECK: %[[INZP:.+]] = arith.constant -128
+  // CHECK: %[[INZP_OFF:.+]] = arith.muli %{{.+}}, %[[INZP]]
+  // CHECK: %[[OFFSETED:.+]] = arith.subi %arg1, %[[INZP_OFF]]
+  // CHECK: %[[NUMERATOR:.+]] = arith.constant 1073741825
+  // CHECK: %[[MULTIPLIER:.+]] = arith.divui %[[NUMERATOR]], %{{.+}}
+  // CHECK: %[[SHIFT:.+]] = arith.constant 30
+  // CHECK: %[[SCALE:.+]] = "tosa.apply_scale"(%{{.+}}, %[[MULTIPLIER]], %[[SHIFT]]) {double_round = false}
+  // CHECK: %[[OUTZP:.+]] = arith.constant -128
+  // CHECK: %[[OUT:.+]] = arith.addi %[[SCALE]], %[[OUTZP]]
+  // CHECK: %[[MIN:.+]] = arith.constant -32768
+  // CHECK: %[[MAX:.+]] = arith.constant 32767
+  // CHECK: %[[CMP_MIN:.+]] = arith.cmpi slt, %[[OUT]], %[[MIN]]
+  // CHECK: %[[CLMP_MIN:.+]] = select %[[CMP_MIN]], %[[MIN]], %[[OUT]]
+  // CHECK: %[[CMP_MAX:.+]] = arith.cmpi slt, %[[MAX]], %[[OUT]]
+  // CHECK: %[[CLMP_MAX:.+]] = select %[[CMP_MAX]], %[[MAX]], %[[CLMP_MIN]]
+  // CHECK: %[[TRUNC:.+]] = arith.trunci %[[CLMP_MAX]]
+  // CHECK: linalg.yield %[[TRUNC]]
+  %0 = "tosa.avg_pool2d"(%arg0) {kernel = [4, 4], pad = [0, 0, 0, 0], quantization_info = {input_zp = -128 : i32, output_zp = -128 : i32}, stride = [4, 4]} : (tensor<1x128x128x2xi16>) -> tensor<1x32x32x2xi16>
+  return
+}
+
+// -----
+
+// CHECK: #[[$MAP1:.+]] = affine_map<(d0, d1, d2, d3) -> (d3)>
+// CHECK: #[[$MAP2:.+]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
+
+// CHECK-LABEL: @conv2d_f32
+func @conv2d_f32(%input: tensor<1x49x42x27xf32>, %weights: tensor<28x3x3x27xf32>, %bias: tensor<28xf32>) -> () {
+  // CHECK: %[[PERM:.+]] = arith.constant dense<[1, 2, 3, 0]>
+  // CHECK: %[[W:.+]] = "tosa.transpose"(%arg1, %[[PERM]])
+  // CHECK: %[[M_IN:.+]] = linalg.init_tensor [1, 45, 40, 28]
+  // CHECK: %[[CST:.+]] = arith.constant 0
+  // CHECK: %[[FILL:.+]] = linalg.fill
+  // CHECK: %[[B_IN:.+]] = linalg.init_tensor [1, 45, 40, 28]
+  // CHECK: %[[CONV:.+]] = linalg.conv_2d_nhwc_hwcf {dilations = dense<[2, 1]> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %[[W]] : tensor<1x49x42x27xf32>, tensor<3x3x27x28xf32>) outs(%[[FILL]] : tensor<1x45x40x28xf32>)
+  // CHECK: %[[B:.+]] = linalg.generic {indexing_maps = [#[[$MAP1]], #[[$MAP2]], #[[$MAP2]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, %[[CONV]] : tensor<28xf32>, tensor<1x45x40x28xf32>) outs(%[[B_IN]] : tensor<1x45x40x28xf32>)
+  // CHECK:   arith.addf
+  // CHECK:   linalg.yield
+  %0 = "tosa.conv2d"(%input, %weights, %bias) {pad = [0, 0, 0, 0], stride = [1, 1], dilation = [2, 1]} : (tensor<1x49x42x27xf32>, tensor<28x3x3x27xf32>, tensor<28xf32>)  -> (tensor<1x45x40x28xf32>)
+  return
+}
+
+// -----
+
+// CHECK-LABEL: @conv2d_padded_f32
+func @conv2d_padded_f32(%input: tensor<1x47x40x28xf32>, %weights: tensor<28x3x3x28xf32>, %bias: tensor<28xf32>) -> () {
+  // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: linalg.pad_tensor %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0]
+  // CHECK:   linalg.yield %[[C0]]
+  // CHECK: linalg.conv_2d_nhwc_hwcf
+  %0 = "tosa.conv2d"(%input, %weights, %bias) {pad = [1, 1, 1, 1], stride = [1, 1], dilation = [2, 1]} : (tensor<1x47x40x28xf32>, tensor<28x3x3x28xf32>, tensor<28xf32>)  -> (tensor<1x45x40x28xf32>)
+  return
+}
+
+// -----
+
+// CHECK-LABEL: @conv2d_quant
+func @conv2d_quant(%arg0 : tensor<1x12x12x1xi8>, %arg1 : tensor<1024x3x3x1xi8>, %arg2 : tensor<1024xi32>) -> () {
+  // CHECK:   %[[C22:.+]] = arith.constant -22
+  // CHECK: linalg.pad_tensor %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0]
+  // CHECK:   linalg.yield %[[C22]]
+  // CHECK: linalg.conv_2d_nhwc_hwcf_q
+  %0 = "tosa.conv2d"(%arg0, %arg1, %arg2) {dilation = [1, 1], pad = [1, 1, 1, 1], quantization_info = {input_zp = -22 : i32, weight_zp = 42 : i32}, stride = [1, 1]} : (tensor<1x12x12x1xi8>, tensor<1024x3x3x1xi8>, tensor<1024xi32>) -> tensor<1x12x12x1024xi32>
+  return
+}
+
+// -----
+
+// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d3)>
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
+
+// CHECK-LABEL: @depthwise_conv
+func @depthwise_conv(%arg0 : tensor<1x7x5x3xf32>, %arg1 : tensor<3x1x3x11xf32>, %arg2 : tensor<33xf32>) -> () {
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 5, 5, 3, 11]
+  // CHECK: [[CST0:%.+]] = arith.constant 0
+  // CHECK: [[FILL:%.+]] = linalg.fill([[CST0]], [[INIT]])
+  // CHECK: [[OUT:%.+]] = linalg.init_tensor [1, 5, 5, 33]
+  // CHECK: [[DEPTH:%.+]] = linalg.depthwise_conv_2d_nhwc_hwcm {dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<1x7x5x3xf32>, tensor<3x1x3x11xf32>) outs([[FILL]] : tensor<1x5x5x3x11xf32>)
+  // CHECK: [[COLLAPSED:%.+]] = "tosa.reshape"([[DEPTH]]) {new_shape = [1, 5, 5, 33]}
+  // CHECK: [[BIAS:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, [[COLLAPSED]] : tensor<33xf32>, tensor<1x5x5x33xf32>) outs([[OUT]] : tensor<1x5x5x33xf32>) {
+  // CHECK: ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):  // no predecessors
+  // CHECK:   [[ADD:%.+]] = arith.addf %arg3, %arg4 : f32
+  // CHECK:   linalg.yield [[ADD]] : f32
+  // CHECK: } -> tensor<1x5x5x33xf32>
+  %2 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) { pad = [0, 0, 0, 0], stride = [1, 1], dilation = [1, 1] } : (tensor<1x7x5x3xf32>, tensor<3x1x3x11xf32>, tensor<33xf32>)  -> (tensor<1x5x5x33xf32>)
+  return
+}
+
+// -----
+
+// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d3)>
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
+
+// CHECK-LABEL: @depthwise_conv_strides
+func @depthwise_conv_strides(%arg0 : tensor<1x11x9x3xf32>, %arg1 : tensor<3x1x3x11xf32>, %arg2 : tensor<33xf32>) -> () {
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 5, 5, 3, 11]
+  // CHECK: [[CST0:%.+]] = arith.constant 0
+  // CHECK: [[FILL:%.+]] = linalg.fill([[CST0]], [[INIT]])
+  // CHECK: [[OUT:%.+]] = linalg.init_tensor [1, 5, 5, 33]
+  // CHECK: [[DEPTH:%.+]] = linalg.depthwise_conv_2d_nhwc_hwcm {dilations = dense<1> : tensor<2xi64>, strides = dense<2> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<1x11x9x3xf32>, tensor<3x1x3x11xf32>) outs([[FILL]] : tensor<1x5x5x3x11xf32>)
+  // CHECK: [[COLLAPSED:%.+]] = "tosa.reshape"([[DEPTH]]) {new_shape = [1, 5, 5, 33]}
+  // CHECK: [[BIAS:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, [[COLLAPSED]] : tensor<33xf32>, tensor<1x5x5x33xf32>) outs([[OUT]] : tensor<1x5x5x33xf32>) {
+  // CHECK: ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):  // no predecessors
+  // CHECK:   [[ADD:%.+]] = arith.addf %arg3, %arg4 : f32
+  // CHECK:   linalg.yield [[ADD]] : f32
+  // CHECK: } -> tensor<1x5x5x33xf32>
+  %2 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) { pad = [0, 0, 0, 0], stride = [2, 2], dilation = [1, 1] } : (tensor<1x11x9x3xf32>, tensor<3x1x3x11xf32>, tensor<33xf32>)  -> (tensor<1x5x5x33xf32>)
+  return
+}
+
+// -----
+
+// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d3)>
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
+
+// CHECK-LABEL: @depthwise_conv_quant
+func @depthwise_conv_quant(%arg0 : tensor<1x12x12x4xi8>, %arg1 : tensor<3x3x4x128xi8>, %arg2 : tensor<512xi32>) -> () {
+  // CHECK: [[PADV:%.+]] = arith.constant -128
+  // CHECK: [[PAD:%.+]] = linalg.pad_tensor %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0]
+  // CHECK:   linalg.yield [[PADV]]
+
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 12, 12, 4, 128]
+  // CHECK: [[CST0:%.+]] = arith.constant 0
+  // CHECK: [[FILL:%.+]] = linalg.fill([[CST0]], [[INIT]])
+  // CHECK: [[OUT:%.+]] = linalg.init_tensor [1, 12, 12, 512]
+  // CHECK: [[C128:%.+]] = arith.constant -128
+  // CHECK: [[C42:%.+]] = arith.constant 42
+  // CHECK: [[DEPTH:%.+]] = linalg.depthwise_conv_2d_nhwc_hwcm_q {dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins([[PAD]], %arg1, [[C128]], [[C42]] : tensor<1x14x14x4xi8>, tensor<3x3x4x128xi8>, i32, i32) outs([[FILL]] : tensor<1x12x12x4x128xi32>)
+  // CHECK: [[COLLAPSED:%.+]] = "tosa.reshape"([[DEPTH]]) {new_shape = [1, 12, 12, 512]}
+  // CHECK: [[BIAS:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, [[COLLAPSED]] : tensor<512xi32>, tensor<1x12x12x512xi32>) outs([[OUT]] : tensor<1x12x12x512xi32>) {
+  // CHECK: ^bb0(%arg3: i32, %arg4: i32, %arg5: i32):  // no predecessors
+  // CHECK:   [[ADD:%.+]] = arith.addi %arg3, %arg4 : i32
+  // CHECK:   linalg.yield [[ADD]] : i32
+  // CHECK: } -> tensor<1x12x12x512xi32>
+  %0 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) {pad = [1, 1, 1, 1], quantization_info = {input_zp = -128 : i32, weight_zp = 42 : i32}, stride = [1, 1], dilation = [1, 1] } : (tensor<1x12x12x4xi8>, tensor<3x3x4x128xi8>, tensor<512xi32>)  -> tensor<1x12x12x512xi32>
+  return
+}
+
+// -----
+
+// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d3)>
+// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
+
+// CHECK-LABEL: @depthwise_conv_quant_dilations
+func @depthwise_conv_quant_dilations(%arg0 : tensor<1x14x14x4xi8>, %arg1 : tensor<3x3x4x128xi8>, %arg2 : tensor<512xi32>) -> () {
+  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 10, 10, 4, 128]
+  // CHECK: [[CST0:%.+]] = arith.constant 0
+  // CHECK: [[FILL:%.+]] = linalg.fill([[CST0]], [[INIT]])
+  // CHECK: [[OUT:%.+]] = linalg.init_tensor [1, 10, 10, 512]
+  // CHECK: [[C128:%.+]] = arith.constant -128
+  // CHECK: [[C42:%.+]] = arith.constant 42
+  // CHECK: [[DEPTH:%.+]] = linalg.depthwise_conv_2d_nhwc_hwcm_q {dilations = dense<2> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1, [[C128]], [[C42]] : tensor<1x14x14x4xi8>, tensor<3x3x4x128xi8>, i32, i32) outs([[FILL]] : tensor<1x10x10x4x128xi32>)
+  // CHECK: [[COLLAPSED:%.+]] = "tosa.reshape"([[DEPTH]]) {new_shape = [1, 10, 10, 512]}
+  // CHECK: [[BIAS:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, [[COLLAPSED]] : tensor<512xi32>, tensor<1x10x10x512xi32>) outs([[OUT]] : tensor<1x10x10x512xi32>) {
+  // CHECK: ^bb0(%arg3: i32, %arg4: i32, %arg5: i32):  // no predecessors
+  // CHECK:   [[ADD:%.+]] = arith.addi %arg3, %arg4 : i32
+  // CHECK:   linalg.yield [[ADD]] : i32
+  // CHECK: } -> tensor<1x10x10x512xi32>
+  %0 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) {pad = [0, 0, 0, 0], quantization_info = {input_zp = -128 : i32, weight_zp = 42 : i32}, stride = [1, 1], dilation = [2, 2] } : (tensor<1x14x14x4xi8>, tensor<3x3x4x128xi8>, tensor<512xi32>)  -> tensor<1x10x10x512xi32>
+  return
+}

diff  --git a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
index f8dc2e0bbd08d..e68e76c67ef98 100644
--- a/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
+++ b/mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir
@@ -1064,154 +1064,6 @@ func @tile(%arg0 : tensor<2x3xi8>) -> () {
 
 // -----
 
-
-// CHECK-LABEL: @matmul
-func @matmul(%arg0: tensor<1x5x3xf32>, %arg1: tensor<1x3x6xf32>) -> (tensor<1x5x6xf32>) {
-  // CHECK: [[C0:%.+]] = arith.constant 0
-  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 5, 6]
-  // CHECK: [[FILLED:%.+]] = linalg.fill([[C0]], [[INIT]]) : f32, tensor<1x5x6xf32> -> tensor<1x5x6xf32>
-  // CHECK: linalg.batch_matmul ins(%arg0, %arg1 : tensor<1x5x3xf32>, tensor<1x3x6xf32>) outs([[FILLED]] : tensor<1x5x6xf32>) -> tensor<1x5x6xf32>
-  %0 = "tosa.matmul"(%arg0, %arg1) : (tensor<1x5x3xf32>, tensor<1x3x6xf32>)  -> (tensor<1x5x6xf32>)
-  return %0 : tensor<1x5x6xf32>
-}
-
-// -----
-
-
-// CHECK-LABEL: @matmul_quantized
-func @matmul_quantized(%arg0: tensor<1x5x3xi8>, %arg1: tensor<1x3x6xi8>) -> (tensor<1x5x6xi32>) {
-  // CHECK: [[C0:%.+]] = arith.constant 0
-  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 5, 6]
-  // CHECK: [[FILLED:%.+]] = linalg.fill([[C0]], [[INIT]]) : i32, tensor<1x5x6xi32> -> tensor<1x5x6xi32>
-  // CHECK: [[ONE:%.+]] = arith.constant 1
-  // CHECK: [[TWO:%.+]] = arith.constant 2
-  // CHECK: linalg.quantized_batch_matmul ins(%arg0, %arg1, [[ONE]], [[TWO]] : tensor<1x5x3xi8>, tensor<1x3x6xi8>, i32, i32) outs([[FILLED]] : tensor<1x5x6xi32>) -> tensor<1x5x6xi32>
-  %0 = "tosa.matmul"(%arg0, %arg1) {quantization_info = {a_zp = 1 : i32, b_zp = 2 : i32}} : (tensor<1x5x3xi8>, tensor<1x3x6xi8>) -> (tensor<1x5x6xi32>)
-  return %0 : tensor<1x5x6xi32>
-}
-
-// -----
-
-// CHECK-LABEL: @matmul_dyn_batch
-func @matmul_dyn_batch(%arg0: tensor<?x5x3xf32>, %arg1: tensor<?x3x6xf32>) -> (tensor<?x5x6xf32>) {
-  // CHECK: %[[C0:.+]] = arith.constant 0
-  // CHECK: %[[DIM:.+]] = tensor.dim %arg0, %[[C0]]
-  // CHECK: %[[C0_0:.+]] = arith.constant 0
-  // CHECK: %[[INIT:.+]] = linalg.init_tensor [%[[DIM]], 5, 6]
-  // CHECK: %[[FILLED:.+]] = linalg.fill(%[[C0_0]], %[[INIT]]) : f32, tensor<?x5x6xf32> -> tensor<?x5x6xf32>
-  // CHECK: linalg.batch_matmul ins(%arg0, %arg1 : tensor<?x5x3xf32>, tensor<?x3x6xf32>) outs(%[[FILLED]] : tensor<?x5x6xf32>) -> tensor<?x5x6xf32>
-  %0 = "tosa.matmul"(%arg0, %arg1) : (tensor<?x5x3xf32>, tensor<?x3x6xf32>)  -> (tensor<?x5x6xf32>)
-  return %0 : tensor<?x5x6xf32>
-}
-
-// -----
-
-// CHECK-LABEL: @matmul_dyn_independent_dim
-func @matmul_dyn_independent_dim(%arg0: tensor<1x5x3xf32>, %arg1: tensor<1x3x?xf32>) -> (tensor<1x5x?xf32>) {
-  // CHECK: %[[C2:.+]] = arith.constant 2
-  // CHECK: %[[DIM:.+]] = tensor.dim %arg1, %[[C2]]
-  // CHECK: %[[C0:.+]] = arith.constant 0
-  // CHECK: %[[INIT:.+]] = linalg.init_tensor [1, 5, %[[DIM]]]
-  // CHECK: %[[FILLED:.+]] = linalg.fill(%[[C0]], %[[INIT]]) : f32, tensor<1x5x?xf32> -> tensor<1x5x?xf32>
-  // CHECK: linalg.batch_matmul ins(%arg0, %arg1 : tensor<1x5x3xf32>, tensor<1x3x?xf32>) outs(%[[FILLED]] : tensor<1x5x?xf32>) -> tensor<1x5x?xf32>
-  %0 = "tosa.matmul"(%arg0, %arg1) : (tensor<1x5x3xf32>, tensor<1x3x?xf32>)  -> (tensor<1x5x?xf32>)
-  return %0 : tensor<1x5x?xf32>
-}
-
-// -----
-
-// CHECK-LABEL: @matmul_dyn_independent_dim
-func @matmul_dyn_independent_dim(%arg0: tensor<1x5x?xf32>, %arg1: tensor<1x?x6xf32>) -> (tensor<1x5x6xf32>) {
-  // CHECK: %[[C0:.+]] = arith.constant 0
-  // CHECK: %[[INIT:.+]] = linalg.init_tensor [1, 5, 6]
-  // CHECK: %[[FILLED:.+]] = linalg.fill(%[[C0]], %[[INIT]]) : f32, tensor<1x5x6xf32> -> tensor<1x5x6xf32>
-  // CHECK: linalg.batch_matmul ins(%arg0, %arg1 : tensor<1x5x?xf32>, tensor<1x?x6xf32>) outs(%[[FILLED]] : tensor<1x5x6xf32>) -> tensor<1x5x6xf32>
-  %0 = "tosa.matmul"(%arg0, %arg1) : (tensor<1x5x?xf32>, tensor<1x?x6xf32>)  -> (tensor<1x5x6xf32>)
-  return %0 : tensor<1x5x6xf32>
-}
-
-// -----
-
-// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d1, d0)>
-// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d0, d1)>
-// CHECK: #[[$MAP2:.*]] = affine_map<(d0, d1) -> (d1)>
-
-// CHECK-LABEL: @fully_connected
-func @fully_connected(%arg0: tensor<5x3xf32>, %arg1: tensor<6x3xf32>, %arg2: tensor<6xf32>) -> (tensor<5x6xf32>) {
-  // CHECK: [[INITT:%.+]] = linalg.init_tensor [5, 6]
-  // CHECK: [[ZERO:%.+]] = arith.constant 0
-  // CHECK: [[FILL:%.+]] = linalg.fill([[ZERO]], [[INITT]])
-  // CHECK: [[PERM:%.+]] = arith.constant dense<[1, 0]>
-  // CHECK: [[INITT:%.+]] = linalg.init_tensor [3, 6]
-  // CHECK: [[TRANSPOSE:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["parallel", "parallel"]} ins(%arg1 : tensor<6x3xf32>) outs([[INITT]] : tensor<3x6xf32>) {
-  // CHECK: ^bb0([[IN:%.+]]: f32, [[UNUSED:%.+]]: f32):
-  // CHECK:   linalg.yield [[IN]] : f32
-  // CHECK: [[INITB:%.+]] = linalg.init_tensor [5, 6]
-  // CHECK: [[MATMUL:%.+]] = linalg.matmul ins(%arg0, [[TRANSPOSE]] : tensor<5x3xf32>, tensor<3x6xf32>) outs([[FILL]] : tensor<5x6xf32>) -> tensor<5x6xf32>
-  // CHECK: [[ADDED:%.+]] = linalg.generic {indexing_maps = [#[[$MAP2]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel"]} ins(%arg2, [[MATMUL]] : tensor<6xf32>, tensor<5x6xf32>) outs([[INITB]] : tensor<5x6xf32>) {
-  // CHECK: ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
-  // CHECK:   [[ADD:%.+]] = arith.addf %arg3, %arg4 : f32
-  // CHECK:   linalg.yield [[ADD]] : f32
-
-  %0 = "tosa.fully_connected"(%arg0, %arg1, %arg2) : (tensor<5x3xf32>, tensor<6x3xf32>, tensor<6xf32>)  -> (tensor<5x6xf32>)
-  return %0 : tensor<5x6xf32>
-}
-
-// -----
-
-// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d1, d0)>
-// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1) -> (d0, d1)>
-// CHECK: #[[$MAP2:.*]] = affine_map<(d0, d1) -> (d1)>
-
-// CHECK-LABEL: @quantized_fully_connected
-func @quantized_fully_connected(%arg0: tensor<5x3xi8>, %arg1: tensor<6x3xi8>, %arg2: tensor<6xi32>) -> (tensor<5x6xi32>) {
-  // CHECK: [[INITT:%.+]] = linalg.init_tensor [5, 6]
-  // CHECK: [[ZERO:%.+]] = arith.constant 0
-  // CHECK: [[FILL:%.+]] = linalg.fill([[ZERO]], [[INITT]])
-  // CHECK: [[PERM:%.+]] = arith.constant dense<[1, 0]>
-  // CHECK: [[INITT:%.+]] = linalg.init_tensor [3, 6]
-  // CHECK: [[TRANSPOSE:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["parallel", "parallel"]} ins(%arg1 : tensor<6x3xi8>) outs([[INITT]] : tensor<3x6xi8>) {
-  // CHECK: ^bb0([[IN:%.+]]: i8, [[UNUSED:%.+]]: i8):
-  // CHECK:   linalg.yield [[IN]] : i8
-  // CHECK: [[INITB:%.+]] = linalg.init_tensor [5, 6]
-  // CHECK: [[ONE:%.+]] = arith.constant 1
-  // CHECK: [[TWO:%.+]] = arith.constant 2
-  // CHECK: [[MATMUL:%.+]] = linalg.quantized_matmul ins(%arg0, [[TRANSPOSE]], [[ONE]], [[TWO]] : tensor<5x3xi8>, tensor<3x6xi8>, i32, i32) outs([[FILL]] : tensor<5x6xi32>) -> tensor<5x6xi32>
-  // CHECK: [[ADDED:%.+]] = linalg.generic {indexing_maps = [#[[$MAP2]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel"]} ins(%arg2, [[MATMUL]] : tensor<6xi32>, tensor<5x6xi32>) outs([[INITB]]
-  // CHECK: ^bb0([[IN1:%.+]]: i32, [[IN2:%.+]]: i32, [[UNUSED:%.+]]: i32):
-  // CHECK:   [[ADD:%.+]] = arith.addi
-  // CHECK:   linalg.yield [[ADD]] : i32
-  %0 = "tosa.fully_connected"(%arg0, %arg1, %arg2) {quantization_info = {input_zp = 1:i32, weight_zp = 2:i32}} : (tensor<5x3xi8>, tensor<6x3xi8>, tensor<6xi32>)  -> (tensor<5x6xi32>)
-  return %0 : tensor<5x6xi32>
-}
-
-// -----
-
-// CHECK-LABEL: @fully_connected_dyn
-func @fully_connected_dyn(%arg0: tensor<?x3xf32>, %arg1: tensor<6x3xf32>, %arg2: tensor<6xf32>) -> (tensor<?x6xf32>) {
-  // CHECK: %[[C0:.+]] = arith.constant 0
-  // CHECK: %[[DIM:.+]] = tensor.dim %arg0, %[[C0]]
-  // CHECK: %[[INITT:.+]] = linalg.init_tensor [%[[DIM]], 6]
-  // CHECK: %[[ZERO:.+]] = arith.constant 0
-  // CHECK: %[[FILL:.+]] = linalg.fill(%[[ZERO]], %[[INITT]])
-  // CHECK: %[[PERM:.+]] = arith.constant dense<[1, 0]>
-  // CHECK: %[[INITT:.+]] = linalg.init_tensor [3, 6]
-  // CHECK: %[[TRANSPOSE:.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["parallel", "parallel"]} ins(%arg1 : tensor<6x3xf32>) outs(%[[INITT]] : tensor<3x6xf32>) {
-  // CHECK: ^bb0(%[[IN:.+]]: f32, %[[UNUSED:.+]]: f32):
-  // CHECK:   linalg.yield %[[IN]] : f32
-  // CHECK: %[[INITB:.+]] = linalg.init_tensor [%[[DIM]], 6]
-  // CHECK: %[[MATMUL:.+]] = linalg.matmul ins(%arg0, %[[TRANSPOSE]] : tensor<?x3xf32>, tensor<3x6xf32>) outs(%[[FILL]] : tensor<?x6xf32>) -> tensor<?x6xf32>
-  // CHECK: %[[ADDED:.+]] = linalg.generic {indexing_maps = [#[[$MAP2]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel"]} ins(%arg2, %[[MATMUL]] : tensor<6xf32>, tensor<?x6xf32>) outs(%[[INITB]] : tensor<?x6xf32>) {
-  // CHECK: ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):
-  // CHECK:   %[[ADD:.+]] = arith.addf %arg3, %arg4 : f32
-  // CHECK:   linalg.yield %[[ADD]] : f32
-
-  %0 = "tosa.fully_connected"(%arg0, %arg1, %arg2) : (tensor<?x3xf32>, tensor<6x3xf32>, tensor<6xf32>)  -> (tensor<?x6xf32>)
-  return %0 : tensor<?x6xf32>
-}
-
-// -----
-
 func @pad_float(%arg0 : tensor<1x2xf32>) -> (tensor<4x9xf32>) {
   %0 = arith.constant dense<[[1, 2], [3, 4]]> : tensor<2x2xi32>
   // TODO: Output contains multiple "arith.constant 1 : index".
@@ -1395,318 +1247,6 @@ func @table16(%arg0: tensor<6xi16>, %arg1: tensor<513xi16>) -> () {
 
 // -----
 
-// CHECK-LABEL: @max_pool
-func @max_pool(%arg0: tensor<1x6x34x62xf32>) -> () {
-  // CHECK-DAG: [[CONST:%.+]] = arith.constant -3.40282347E+38
-  // CHECK-DAG: [[INIT:%.+]] = linalg.init_tensor [1, 4, 32, 62]
-  // CHECK-DAG: [[FILL:%.+]] = linalg.fill([[CONST]], [[INIT]])
-  // CHECK-DAG: [[KERNEL:%.+]] = linalg.init_tensor [3, 3]
-  // CHECK: linalg.pooling_nhwc_max {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins(%arg0, [[KERNEL]] : tensor<1x6x34x62xf32>, tensor<3x3xf32>) outs([[FILL]] : tensor<1x4x32x62xf32>)
-  %0 = "tosa.max_pool2d"(%arg0) {pad = [0, 0, 0, 0], kernel = [3, 3], stride = [1, 1]} : (tensor<1x6x34x62xf32>)  -> (tensor<1x4x32x62xf32>)
-  return
-}
-
-// CHECK-LABEL: @max_pool_padded
-func @max_pool_padded(%arg0: tensor<1x6x34x62xf32>) -> () {
-  // CHECK-DAG: [[CONST:%.+]] = arith.constant -3.40282347E+38 : f32
-  // CHECK-DAG: [[PAD:%.+]] = linalg.pad_tensor %arg0 low[0, 0, 0, 0] high[0, 0, 1, 0]
-  // CHECK-DAG:   linalg.yield [[CONST]]
-  // CHECK-DAG: [[INITVAL:%.+]] = arith.constant -3.40282347E+38 : f32
-  // CHECK-DAG: [[INIT:%.+]] = linalg.init_tensor [1, 4, 33, 62]
-  // CHECK-DAG: [[FILL:%.+]] = linalg.fill([[INITVAL]], [[INIT]])
-  // CHECK-DAG: [[KERNEL:%.+]] = linalg.init_tensor [3, 3]
-  // CHECK: linalg.pooling_nhwc_max {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins([[PAD]], [[KERNEL]] : tensor<1x6x35x62xf32>, tensor<3x3xf32>) outs([[FILL]] : tensor<1x4x33x62xf32>)
-  %0 = "tosa.max_pool2d"(%arg0) {pad = [0, 0, 0, 1], kernel = [3, 3], stride = [1, 1]} : (tensor<1x6x34x62xf32>)  -> (tensor<1x4x33x62xf32>)
-  return
-}
-
-// CHECK-LABEL: @max_pool_i8
-func @max_pool_i8(%arg0: tensor<1x6x34x62xi8>) -> () {
-  // CHECK: arith.constant -128
-  // CHECK: linalg.pooling_nhwc_max
-  %0 = "tosa.max_pool2d"(%arg0) {pad = [0, 0, 0, 0], kernel = [3, 3], stride = [1, 1]} : (tensor<1x6x34x62xi8>)  -> (tensor<1x4x32x62xi8>)
-  return
-}
-
-// CHECK-LABEL: @max_pool_i16
-func @max_pool_i16(%arg0: tensor<1x6x34x62xi16>) -> () {
-  // CHECK: arith.constant -32768
-  // CHECK: linalg.pooling_nhwc_max
-  %0 = "tosa.max_pool2d"(%arg0) {pad = [0, 0, 0, 0], kernel = [3, 3], stride = [1, 1]} : (tensor<1x6x34x62xi16>)  -> (tensor<1x4x32x62xi16>)
-  return
-}
-
-// CHECK-LABEL: @max_pool_i32
-func @max_pool_i32(%arg0: tensor<1x6x34x62xi32>) -> () {
-  // CHECK: arith.constant -2147483648
-  // CHECK: linalg.pooling_nhwc_max
-  %0 = "tosa.max_pool2d"(%arg0) {pad = [0, 0, 0, 0], kernel = [3, 3], stride = [1, 1]} : (tensor<1x6x34x62xi32>)  -> (tensor<1x4x32x62xi32>)
-  return
-}
-// -----
-
-// CHECK-LABEL: @avg_pool
-func @avg_pool(%arg0: tensor<1x6x34x62xf32>) -> (tensor<1x5x33x62xf32>) {
-  // Initial piece computes the sum of the pooling region, with appropriate padding.
-  // CHECK: [[CONST:%.+]] = arith.constant 0
-  // CHECK: [[PAD:%.+]] = linalg.pad_tensor %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0]
-  // CHECK: [[CONST:%.+]] = arith.constant 0
-  // CHECK: [[POOLINIT:%.+]] = linalg.init_tensor [1, 5, 33, 62]
-  // CHECK: [[FILL:%.+]] = linalg.fill([[CONST]], [[POOLINIT]])
-  // CHECK: [[KERNEL:%.+]] = linalg.init_tensor [4, 4]
-  // CHECK: [[POOL:%.+]] = linalg.pooling_nhwc_sum {dilations = dense<1> : vector<2xi64>, strides = dense<1> : vector<2xi64>} ins([[PAD]], [[KERNEL]] : tensor<1x8x36x62xf32>, tensor<4x4xf32>) outs([[FILL]] : tensor<1x5x33x62xf32>)
-  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 5, 33, 62]
-  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins([[POOL]] : tensor<1x5x33x62xf32>) outs([[INIT]] : tensor<1x5x33x62xf32>)
-  // CHECK:   [[ZERO:%.0]] = arith.constant 0
-  // CHECK:   [[ONE:%.+]] = arith.constant 1
-  // CHECK:   [[HEIGHT:%.+]] = arith.constant 4
-  // CHECK:   [[WIDTH:%.+]] = arith.constant 32
-  // CHECK:   [[IDX1:%.+]] = linalg.index 1
-  // CHECK:   [[IDX2:%.+]] = linalg.index 2
-
-  // The large block below computes what portion of the kernel is within non-padded input.
-  // CHECK:   [[NY:%.+]] = arith.subi [[HEIGHT]], [[IDX1]]
-  // CHECK:   [[NX:%.+]] = arith.subi [[WIDTH]], [[IDX2]]
-  // CHECK:   [[KH:%.+]] = arith.constant 4
-  // CHECK:   [[PAD0:%.+]] = arith.constant 1
-  // CHECK:   [[SUBP0:%.+]] = arith.subi [[IDX1]], [[PAD0]]
-  // CHECK:   [[P0CMP:%.+]] = arith.cmpi slt, [[SUBP0]], [[ZERO]]
-  // CHECK:   [[SELP0:%.+]] = select [[P0CMP]], [[SUBP0]], [[ZERO]]
-  // CHECK:   [[ADDP0:%.+]] = arith.addi [[KH]], [[SELP0]]
-  // CHECK:   [[PAD1:%.+]] = arith.constant 1
-  // CHECK:   [[SUBP1:%.+]] = arith.subi [[NY]], [[PAD1]]
-  // CHECK:   [[P1CMP:%.+]] = arith.cmpi slt, [[SUBP1]], [[ZERO]]
-  // CHECK:   [[SELP1:%.+]] = select [[P1CMP]], [[SUBP1]], [[ZERO]]
-  // CHECK:   [[ADDP1:%.+]] = arith.addi [[ADDP0]], [[SELP1]]
-  // CHECK:   [[YCMP:%.+]] = arith.cmpi slt, [[ADDP1]], [[ONE]]
-  // CHECK:   [[YSEL:%.+]] = select [[YCMP]], [[ONE]], [[ADDP1]]
-  // CHECK:   [[KW:%.+]] = arith.constant 4 : index
-  // CHECK:   [[PAD2:%.+]] = arith.constant 1 : index
-  // CHECK:   [[SUBP2:%.+]] = arith.subi [[IDX2]], [[PAD2]]
-  // CHECK:   [[P2CMP:%.+]] = arith.cmpi slt, [[SUBP2]], [[ZERO]]
-  // CHECK:   [[SELP2:%.+]] = select [[P2CMP]], [[SUBP2]], [[ZERO]]
-  // CHECK:   [[ADDP2:%.+]] = arith.addi [[KW]], [[SELP2]]
-  // CHECK:   [[PAD3:%.+]] = arith.constant 1 : index
-  // CHECK:   [[SUBP3:%.+]] = arith.subi [[NX]], [[PAD3]]
-  // CHECK:   [[P3CMP:%.+]] = arith.cmpi slt, [[SUBP3]], [[ZERO]]
-  // CHECK:   [[SELP3:%.+]] = select [[P3CMP]], [[SUBP3]], [[ZERO]]
-  // CHECK:   [[ADDP3:%.+]] = arith.addi [[ADDP2]], [[SELP3]]
-  // CHECK:   [[XCMP:%.+]] = arith.cmpi slt, [[ADDP3]], [[ONE]]
-  // CHECK:   [[XSEL:%.+]] = select [[XCMP]], [[ONE]], [[ADDP3]]
-
-  // Given the valid coverage of the pooling region, normalize the summation.
-  // CHECK:   [[C:%.+]] = arith.muli [[YSEL]], [[XSEL]]
-  // CHECK:   [[CI:%.+]] = arith.index_cast [[C]]
-  // CHECK:   [[CF:%.+]] = arith.sitofp [[CI]]
-  // CHECK:   [[RESULT:%.+]] = arith.divf %arg1, [[CF]]
-  // CHECK:   linalg.yield [[RESULT]]
-  %0 = "tosa.avg_pool2d"(%arg0) {pad = [1, 1, 1, 1], kernel = [4, 4], stride = [1, 1]} : (tensor<1x6x34x62xf32>)  -> (tensor<1x5x33x62xf32>)
-  return %0 : tensor<1x5x33x62xf32>
-}
-
-// -----
-
-// CHECK-LABEL: @avg_pool_i8
-func @avg_pool_i8(%arg0 : tensor<1x128x128x2xi8>) -> () {
-
-  // CHECK: linalg.pooling_nhwc_sum
-  // CHECK: linalg.generic
-
-  // CHECK: %[[INZP:.+]] = arith.constant -128
-  // CHECK: %[[INZP_OFF:.+]] = arith.muli %{{.+}}, %[[INZP]]
-  // CHECK: %[[OFFSETED:.+]] = arith.subi %arg1, %[[INZP_OFF]]
-  // CHECK: %[[NUMERATOR:.+]] = arith.constant 1073741825
-  // CHECK: %[[MULTIPLIER:.+]] = arith.divui %[[NUMERATOR]], %{{.+}}
-  // CHECK: %[[SHIFT:.+]] = arith.constant 30
-  // CHECK: %[[SCALE:.+]] = "tosa.apply_scale"(%{{.+}}, %[[MULTIPLIER]], %[[SHIFT]]) {double_round = false}
-  // CHECK: %[[OUTZP:.+]] = arith.constant -128
-  // CHECK: %[[OUT:.+]] = arith.addi %[[SCALE]], %[[OUTZP]]
-  // CHECK: %[[MIN:.+]] = arith.constant -128
-  // CHECK: %[[MAX:.+]] = arith.constant 127
-  // CHECK: %[[CMP_MIN:.+]] = arith.cmpi slt, %[[OUT]], %[[MIN]]
-  // CHECK: %[[CLMP_MIN:.+]] = select %[[CMP_MIN]], %[[MIN]], %[[OUT]]
-  // CHECK: %[[CMP_MAX:.+]] = arith.cmpi slt, %[[MAX]], %[[OUT]]
-  // CHECK: %[[CLMP_MAX:.+]] = select %[[CMP_MAX]], %[[MAX]], %[[CLMP_MIN]]
-  // CHECK: %[[TRUNC:.+]] = arith.trunci %[[CLMP_MAX]]
-  // CHECK: linalg.yield %[[TRUNC]]
-  %0 = "tosa.avg_pool2d"(%arg0) {kernel = [4, 4], pad = [0, 0, 0, 0], quantization_info = {input_zp = -128 : i32, output_zp = -128 : i32}, stride = [4, 4]} : (tensor<1x128x128x2xi8>) -> tensor<1x32x32x2xi8>
-  return
-}
-
-// -----
-
-// CHECK-LABEL: @avg_pool_i16
-func @avg_pool_i16(%arg0 : tensor<1x128x128x2xi16>) -> () {
-
-  // CHECK: linalg.pooling_nhwc_sum
-  // CHECK: linalg.generic
-
-  // CHECK: %[[INZP:.+]] = arith.constant -128
-  // CHECK: %[[INZP_OFF:.+]] = arith.muli %{{.+}}, %[[INZP]]
-  // CHECK: %[[OFFSETED:.+]] = arith.subi %arg1, %[[INZP_OFF]]
-  // CHECK: %[[NUMERATOR:.+]] = arith.constant 1073741825
-  // CHECK: %[[MULTIPLIER:.+]] = arith.divui %[[NUMERATOR]], %{{.+}}
-  // CHECK: %[[SHIFT:.+]] = arith.constant 30
-  // CHECK: %[[SCALE:.+]] = "tosa.apply_scale"(%{{.+}}, %[[MULTIPLIER]], %[[SHIFT]]) {double_round = false}
-  // CHECK: %[[OUTZP:.+]] = arith.constant -128
-  // CHECK: %[[OUT:.+]] = arith.addi %[[SCALE]], %[[OUTZP]]
-  // CHECK: %[[MIN:.+]] = arith.constant -32768
-  // CHECK: %[[MAX:.+]] = arith.constant 32767
-  // CHECK: %[[CMP_MIN:.+]] = arith.cmpi slt, %[[OUT]], %[[MIN]]
-  // CHECK: %[[CLMP_MIN:.+]] = select %[[CMP_MIN]], %[[MIN]], %[[OUT]]
-  // CHECK: %[[CMP_MAX:.+]] = arith.cmpi slt, %[[MAX]], %[[OUT]]
-  // CHECK: %[[CLMP_MAX:.+]] = select %[[CMP_MAX]], %[[MAX]], %[[CLMP_MIN]]
-  // CHECK: %[[TRUNC:.+]] = arith.trunci %[[CLMP_MAX]]
-  // CHECK: linalg.yield %[[TRUNC]]
-  %0 = "tosa.avg_pool2d"(%arg0) {kernel = [4, 4], pad = [0, 0, 0, 0], quantization_info = {input_zp = -128 : i32, output_zp = -128 : i32}, stride = [4, 4]} : (tensor<1x128x128x2xi16>) -> tensor<1x32x32x2xi16>
-  return
-}
-
-// -----
-
-// CHECK: #[[$MAP0:.+]] = affine_map<(d0, d1, d2, d3) -> (d3, d0, d1, d2)>
-// CHECK: #[[$MAP1:.+]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
-// CHECK: #[[$MAP2:.+]] = affine_map<(d0, d1, d2, d3) -> (d3)>
-
-// CHECK-LABEL: @conv2d_f32
-func @conv2d_f32(%input: tensor<1x49x42x27xf32>, %weights: tensor<28x3x3x27xf32>, %bias: tensor<28xf32>) -> () {
-  // CHECK: %[[W_IN:.+]] = linalg.init_tensor [3, 3, 27, 28]
-  // CHECK: %[[W:.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg1 : tensor<28x3x3x27xf32>) outs(%[[W_IN]] : tensor<3x3x27x28xf32>)
-  // CHECK:   linalg.yield %arg3 : f32
-  // CHECK: %[[M_IN:.+]] = linalg.init_tensor [1, 45, 40, 28]
-  // CHECK: %[[CST:.+]] = arith.constant 0
-  // CHECK: %[[FILL:.+]] = linalg.fill
-  // CHECK: %[[B_IN:.+]] = linalg.init_tensor [1, 45, 40, 28]
-  // CHECK: %[[CONV:.+]] = linalg.conv_2d_nhwc_hwcf {dilations = dense<[2, 1]> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %[[W]] : tensor<1x49x42x27xf32>, tensor<3x3x27x28xf32>) outs(%[[FILL]] : tensor<1x45x40x28xf32>)
-  // CHECK: %[[B:.+]] = linalg.generic {indexing_maps = [#[[$MAP2]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, %[[CONV]] : tensor<28xf32>, tensor<1x45x40x28xf32>) outs(%[[B_IN]] : tensor<1x45x40x28xf32>)
-  // CHECK:   arith.addf
-  // CHECK:   linalg.yield %7 : f32
-  %0 = "tosa.conv2d"(%input, %weights, %bias) {pad = [0, 0, 0, 0], stride = [1, 1], dilation = [2, 1]} : (tensor<1x49x42x27xf32>, tensor<28x3x3x27xf32>, tensor<28xf32>)  -> (tensor<1x45x40x28xf32>)
-  return
-}
-
-// -----
-
-// CHECK-LABEL: @conv2d_padded_f32
-func @conv2d_padded_f32(%input: tensor<1x47x40x28xf32>, %weights: tensor<28x3x3x28xf32>, %bias: tensor<28xf32>) -> () {
-  // CHECK: %[[C0:.+]] = arith.constant 0
-  // CHECK: linalg.pad_tensor %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0]
-  // CHECK:   linalg.yield %[[C0]]
-  // CHECK: linalg.conv_2d_nhwc_hwcf
-  %0 = "tosa.conv2d"(%input, %weights, %bias) {pad = [1, 1, 1, 1], stride = [1, 1], dilation = [2, 1]} : (tensor<1x47x40x28xf32>, tensor<28x3x3x28xf32>, tensor<28xf32>)  -> (tensor<1x45x40x28xf32>)
-  return
-}
-
-// -----
-
-// CHECK-LABEL: @conv2d_quant
-func @conv2d_quant(%arg0 : tensor<1x12x12x1xi8>, %arg1 : tensor<1024x3x3x1xi8>, %arg2 : tensor<1024xi32>) -> () {
-  // CHECK:   %[[C22:.+]] = arith.constant -22
-  // CHECK: linalg.pad_tensor %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0]
-  // CHECK:   linalg.yield %[[C22]]
-  // CHECK: linalg.conv_2d_nhwc_hwcf_q
-  %0 = "tosa.conv2d"(%arg0, %arg1, %arg2) {dilation = [1, 1], pad = [1, 1, 1, 1], quantization_info = {input_zp = -22 : i32, weight_zp = 42 : i32}, stride = [1, 1]} : (tensor<1x12x12x1xi8>, tensor<1024x3x3x1xi8>, tensor<1024xi32>) -> tensor<1x12x12x1024xi32>
-  return
-}
-
-// -----
-
-// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d3)>
-// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
-
-// CHECK-LABEL: @depthwise_conv
-func @depthwise_conv(%arg0 : tensor<1x7x5x3xf32>, %arg1 : tensor<3x1x3x11xf32>, %arg2 : tensor<33xf32>) -> () {
-  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 5, 5, 3, 11]
-  // CHECK: [[CST0:%.+]] = arith.constant 0
-  // CHECK: [[FILL:%.+]] = linalg.fill([[CST0]], [[INIT]])
-  // CHECK: [[OUT:%.+]] = linalg.init_tensor [1, 5, 5, 33]
-  // CHECK: [[DEPTH:%.+]] = linalg.depthwise_conv_2d_nhwc_hwcm {dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<1x7x5x3xf32>, tensor<3x1x3x11xf32>) outs([[FILL]] : tensor<1x5x5x3x11xf32>)
-  // CHECK: [[COLLAPSED:%.+]] = tensor.collapse_shape [[DEPTH]] {{\[}}[0], [1], [2], [3, 4]]
-  // CHECK: [[BIAS:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, [[COLLAPSED]] : tensor<33xf32>, tensor<1x5x5x33xf32>) outs([[OUT]] : tensor<1x5x5x33xf32>) {
-  // CHECK: ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):  // no predecessors
-  // CHECK:   [[ADD:%.+]] = arith.addf %arg3, %arg4 : f32
-  // CHECK:   linalg.yield [[ADD]] : f32
-  // CHECK: } -> tensor<1x5x5x33xf32>
-  %2 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) { pad = [0, 0, 0, 0], stride = [1, 1], dilation = [1, 1] } : (tensor<1x7x5x3xf32>, tensor<3x1x3x11xf32>, tensor<33xf32>)  -> (tensor<1x5x5x33xf32>)
-  return
-}
-
-// -----
-
-// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d3)>
-// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
-
-// CHECK-LABEL: @depthwise_conv_strides
-func @depthwise_conv_strides(%arg0 : tensor<1x11x9x3xf32>, %arg1 : tensor<3x1x3x11xf32>, %arg2 : tensor<33xf32>) -> () {
-  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 5, 5, 3, 11]
-  // CHECK: [[CST0:%.+]] = arith.constant 0
-  // CHECK: [[FILL:%.+]] = linalg.fill([[CST0]], [[INIT]])
-  // CHECK: [[OUT:%.+]] = linalg.init_tensor [1, 5, 5, 33]
-  // CHECK: [[DEPTH:%.+]] = linalg.depthwise_conv_2d_nhwc_hwcm {dilations = dense<1> : tensor<2xi64>, strides = dense<2> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<1x11x9x3xf32>, tensor<3x1x3x11xf32>) outs([[FILL]] : tensor<1x5x5x3x11xf32>)
-  // CHECK: [[COLLAPSED:%.+]] = tensor.collapse_shape [[DEPTH]] {{\[}}[0], [1], [2], [3, 4]]
-  // CHECK: [[BIAS:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, [[COLLAPSED]] : tensor<33xf32>, tensor<1x5x5x33xf32>) outs([[OUT]] : tensor<1x5x5x33xf32>) {
-  // CHECK: ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):  // no predecessors
-  // CHECK:   [[ADD:%.+]] = arith.addf %arg3, %arg4 : f32
-  // CHECK:   linalg.yield [[ADD]] : f32
-  // CHECK: } -> tensor<1x5x5x33xf32>
-  %2 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) { pad = [0, 0, 0, 0], stride = [2, 2], dilation = [1, 1] } : (tensor<1x11x9x3xf32>, tensor<3x1x3x11xf32>, tensor<33xf32>)  -> (tensor<1x5x5x33xf32>)
-  return
-}
-
-// -----
-
-// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d3)>
-// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
-
-// CHECK-LABEL: @depthwise_conv_quant
-func @depthwise_conv_quant(%arg0 : tensor<1x12x12x4xi8>, %arg1 : tensor<3x3x4x128xi8>, %arg2 : tensor<512xi32>) -> () {
-  // CHECK: [[PADV:%.+]] = arith.constant -128
-  // CHECK: [[PAD:%.+]] = linalg.pad_tensor %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0]
-  // CHECK:   linalg.yield [[PADV]]
-
-  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 12, 12, 4, 128]
-  // CHECK: [[CST0:%.+]] = arith.constant 0
-  // CHECK: [[FILL:%.+]] = linalg.fill([[CST0]], [[INIT]])
-  // CHECK: [[OUT:%.+]] = linalg.init_tensor [1, 12, 12, 512]
-  // CHECK: [[C128:%.+]] = arith.constant -128
-  // CHECK: [[C42:%.+]] = arith.constant 42
-  // CHECK: [[DEPTH:%.+]] = linalg.depthwise_conv_2d_nhwc_hwcm_q {dilations = dense<1> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins([[PAD]], %arg1, [[C128]], [[C42]] : tensor<1x14x14x4xi8>, tensor<3x3x4x128xi8>, i32, i32) outs([[FILL]] : tensor<1x12x12x4x128xi32>)
-  // CHECK: [[COLLAPSED:%.+]] = tensor.collapse_shape [[DEPTH]] {{\[}}[0], [1], [2], [3, 4]]
-  // CHECK: [[BIAS:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, [[COLLAPSED]] : tensor<512xi32>, tensor<1x12x12x512xi32>) outs([[OUT]] : tensor<1x12x12x512xi32>) {
-  // CHECK: ^bb0(%arg3: i32, %arg4: i32, %arg5: i32):  // no predecessors
-  // CHECK:   [[ADD:%.+]] = arith.addi %arg3, %arg4 : i32
-  // CHECK:   linalg.yield [[ADD]] : i32
-  // CHECK: } -> tensor<1x12x12x512xi32>
-  %0 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) {pad = [1, 1, 1, 1], quantization_info = {input_zp = -128 : i32, weight_zp = 42 : i32}, stride = [1, 1], dilation = [1, 1] } : (tensor<1x12x12x4xi8>, tensor<3x3x4x128xi8>, tensor<512xi32>)  -> tensor<1x12x12x512xi32>
-  return
-}
-
-// -----
-
-// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1, d2, d3) -> (d3)>
-// CHECK: #[[$MAP1:.*]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
-
-// CHECK-LABEL: @depthwise_conv_quant_dilations
-func @depthwise_conv_quant_dilations(%arg0 : tensor<1x14x14x4xi8>, %arg1 : tensor<3x3x4x128xi8>, %arg2 : tensor<512xi32>) -> () {
-  // CHECK: [[INIT:%.+]] = linalg.init_tensor [1, 10, 10, 4, 128]
-  // CHECK: [[CST0:%.+]] = arith.constant 0
-  // CHECK: [[FILL:%.+]] = linalg.fill([[CST0]], [[INIT]])
-  // CHECK: [[OUT:%.+]] = linalg.init_tensor [1, 10, 10, 512]
-  // CHECK: [[C128:%.+]] = arith.constant -128
-  // CHECK: [[C42:%.+]] = arith.constant 42
-  // CHECK: [[DEPTH:%.+]] = linalg.depthwise_conv_2d_nhwc_hwcm_q {dilations = dense<2> : tensor<2xi64>, strides = dense<1> : tensor<2xi64>} ins(%arg0, %arg1, [[C128]], [[C42]] : tensor<1x14x14x4xi8>, tensor<3x3x4x128xi8>, i32, i32) outs([[FILL]] : tensor<1x10x10x4x128xi32>)
-  // CHECK: [[COLLAPSED:%.+]] = tensor.collapse_shape [[DEPTH]] {{\[}}[0], [1], [2], [3, 4]]
-  // CHECK: [[BIAS:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, [[COLLAPSED]] : tensor<512xi32>, tensor<1x10x10x512xi32>) outs([[OUT]] : tensor<1x10x10x512xi32>) {
-  // CHECK: ^bb0(%arg3: i32, %arg4: i32, %arg5: i32):  // no predecessors
-  // CHECK:   [[ADD:%.+]] = arith.addi %arg3, %arg4 : i32
-  // CHECK:   linalg.yield [[ADD]] : i32
-  // CHECK: } -> tensor<1x10x10x512xi32>
-  %0 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) {pad = [0, 0, 0, 0], quantization_info = {input_zp = -128 : i32, weight_zp = 42 : i32}, stride = [1, 1], dilation = [2, 2] } : (tensor<1x14x14x4xi8>, tensor<3x3x4x128xi8>, tensor<512xi32>)  -> tensor<1x10x10x512xi32>
-  return
-}
-
-// -----
-
 // CHECK-LABEL: @resize_nearest
 func @resize_nearest(%input: tensor<1x2x2x1xf32>) -> () {
   // CHECK: %[[INIT:.+]] = linalg.init_tensor [1, 4, 4, 1]


        


More information about the Mlir-commits mailing list