[tosa][mlir] Support dynamic batch dimension for ops where the batch dim is explicit

[NatashaKnk]

Summary

Dynamic batch for rescale, gather, max_pool, avg_pool, conv2D and depthwise_conv2D. Split helper functions into a separate header file.

[rsuderman]

Move mlir/lib/Conversion/TosaToLinalg/TosaToLinalgCommon.h to mlir/include/mlir/Dialect/Tosa/Utils/utils.h. You can check mlir/include/mlir/Dialect/Linalg/Utils/Utils.h for an example on how to declare utilities for dialects.

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

14

This should be a full path and not local.

mlir/lib/Conversion/TosaToLinalg/TosaToLinalgCommon.h

23 ↗

(On Diff #398966)

Add an additional tosa namespace so that these helpers are not placed in the global mlir namespace.

24 ↗

(On Diff #398966)

I believe you can just use static to avoid the namespace collision. This is usually preferred to inline.

57 ↗

(On Diff #398966)

Rename to checkHasDynamicBatchDims. We can likely drop the Only and Has is more commonly used in names.

mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamed.cpp

13

ditto on full path.

[rsuderman]

mlir/lib/Conversion/TosaToLinalg/TosaToLinalgCommon.h

29 ↗

(On Diff #398966)

Also, add documentation for what each function does. Once we move to Utils we can expect other libraries to use these functions.

[NatashaKnk]

Moved header util file and created a .cpp file. Added a couple of fixes

[rsuderman]

mlir/include/mlir/Dialect/Tosa/Utils/CoversionUtils.h

30

Small typo, should be condenseValues

[NatashaKnk]

Fix typo

mlir/include/mlir/Dialect/Tosa/Utils/CoversionUtils.h

+//===- ConversionUtils.h - Helper functions for tosa conversion -*- C++ -*-===//
+//
+// 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
+//
+//===----------------------------------------------------------------------===//
+//
+// Utility functions for TOSA lowering
+//
+//===----------------------------------------------------------------------===//
+
+#ifndef DIALECT_TOSA_UTILS_COVERSION_UTILS_H_
+#define DIALECT_TOSA_UTILS_COVERSION_UTILS_H_
+
+#include "mlir/Dialect/Arithmetic/IR/Arithmetic.h"
+#include "mlir/Dialect/StandardOps/IR/Ops.h"
+#include "mlir/Dialect/Tensor/IR/Tensor.h"
+#include "mlir/Dialect/Utils/StructuredOpsUtils.h"
+#include "mlir/IR/PatternMatch.h"
+
+namespace mlir {
+namespace tosa {
+
+// Creates a SmallVector of Stringrefs for N parallel loops
+SmallVector<StringRef> getNParallelLoopsAttrs(unsigned nParallelLoops);
+
+// Takes a vector of values and condenses them to a vector with no gaps.
+SmallVector<Value> condenseValues(const SmallVector<Value> &values);
+
+// Takes the parameters for a clamp and turns it into a series of ops.
+template <typename T, typename P>
+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);
+}
+
+// Returns the values in an attribute as an array of values.
+template <typename T>
+void getValuesFromIntArrayAttribute(ArrayAttr attr,
+                                    SmallVector<T> &arrayValues) {
+  for (Attribute val : attr.getValue()) {
+    arrayValues.push_back(val.cast<IntegerAttr>().getValue().getSExtValue());
+  }
+}
+
+// Checks for a dynamic batch dim in any of the passed parameters of an op.
+// The batch dimention must be #0 and the rest of the dimensions must be static.
+template <typename Op>
+Optional<SmallVector<Value>> checkHasDynamicBatchDims(PatternRewriter &rewriter,
+                                                      Op op,
+                                                      ArrayRef<Value> params) {
+  SmallVector<ShapedType> dynTypes;
+  SmallVector<Value> dynamicDims;
+  for (const Value &param : params) {
+    auto paramTy = param.getType().cast<ShapedType>();
+    if (!paramTy.hasStaticShape())
+      dynTypes.push_back(paramTy);
+  }
+
+  if (dynTypes.empty())
+    return dynamicDims;
+
+  for (const ShapedType &dynTy : dynTypes) {
+    if (llvm::any_of(dynTy.getShape().drop_front(), ShapedType::isDynamic)) {
+      (void)rewriter.notifyMatchFailure(
+          op, "input can only be dynamic for batch size");
+      return llvm::None;
+    }
+  }
+
+  dynamicDims.push_back(
+      rewriter.create<tensor::DimOp>(op->getLoc(), params[0], 0));
+  return dynamicDims;
+}
+
+} // namespace tosa
+} // namespace mlir
+
+#endif // DIALECT_TOSA_UTILS_COVERSION_UTILS_H_

mlir/lib/Conversion/TosaToLinalg/TosaToLinalg.cpp

 //===- TosaToLinalg.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
 //
 //===----------------------------------------------------------------------===//
 //
 // These rewriters lower from the Tosa to the Linalg dialect.
 //
 //===----------------------------------------------------------------------===//
 
 #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/Utils/CoversionUtils.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());
-}
+using namespace mlir::tosa;
 
 template <typename T>
 static arith::ConstantOp
 createConstFromIntAttribute(Operation *op, const std::string &attrName,
                             Type requiredAttrType, OpBuilder &rewriter) {
   auto castedN = static_cast<T>(
       op->getAttr(attrName).cast<IntegerAttr>().getValue().getSExtValue());
   return rewriter.create<arith::ConstantOp>(
       op->getLoc(), IntegerAttr::get(requiredAttrType, castedN));
 }
 
-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 SmallVector<Value> filterDynamicDims(const SmallVector<Value> &dynDims) {
-  SmallVector<Value> filteredDims;
-  for (auto dim : dynDims)
-    if (dim)
-      filteredDims.push_back(dim);
-  return filteredDims;
-}
-
 static Value
 createLinalgBodyCalculationForElementwiseOp(Operation *op, ValueRange args,
                                             ArrayRef<Type> resultTypes,
                                             PatternRewriter &rewriter) {
   Location loc = op->getLoc();
   auto elementTy =
       op->getOperand(0).getType().cast<ShapedType>().getElementType();
 
elementwiseMatchAndRewriteHelper(Operation *operation,
   for (auto arg : operation->getOperands()) {
     auto operandTy = arg.getType().cast<ShapedType>();
     for (int i = 0; i < operandTy.getRank(); i++) {
       if (operandTy.isDynamicDim(i) && !dynDims[i])
         dynDims[i] = rewriter.create<tensor::DimOp>(loc, arg, i);
     }
   }
 
-  SmallVector<Value> filteredDims = filterDynamicDims(dynDims);
+  SmallVector<Value> filteredDims = condenseValues(dynDims);
 
   for (auto result : results) {
     auto resultTy = result.getType().template cast<ShapedType>();
     initTensors.push_back(rewriter.create<linalg::InitTensorOp>(
         loc, filteredDims, resultTy.getShape(), resultTy.getElementType()));
     opResultTypes.push_back(result.getType());
   }
 
for (const auto &permutation : llvm::enumerate(perms.getValues<APInt>())) {
       auto index = permutation.index();
       auto value = permutation.value().getZExtValue();
       if (!operandTy.hasRank() || operandTy.isDynamicDim(index)) {
         dynDims[value] = rewriter.create<tensor::DimOp>(loc, input, index);
       }
       inputExprs[value] = rewriter.getAffineDimExpr(index);
     }
 
-    SmallVector<Value> filteredDims = filterDynamicDims(dynDims);
+    SmallVector<Value> filteredDims = condenseValues(dynDims);
 
     auto initTensor = rewriter.create<linalg::InitTensorOp>(
         loc, filteredDims, resultTy.getShape(), resultTy.getElementType());
 
     SmallVector<AffineMap, 2> affineMaps = {
         AffineMap::get(resultTy.getRank(), /*symbolCount=*/0, inputExprs,
                        rewriter.getContext()),
         rewriter.getMultiDimIdentityMap(resultTy.getRank())};
LogicalResult matchAndRewrite(tosa::RescaleOp op,
     auto outputTy = op.output().getType().cast<ShapedType>();
     unsigned rank = inputTy.getRank();
 
     // This is an illegal configuration. terminate and log an error
     if (op.double_round() && !op.scale32())
       return rewriter.notifyMatchFailure(
           op, "tosa.rescale requires scale32 for double_round to be true");
 
-    if (!outputTy.hasStaticShape())
-      return rewriter.notifyMatchFailure(
-          op, "tosa to linalg conversion expects statically shaped tensors");
+    auto dynamicDimsOr =
+        checkHasDynamicBatchDims(rewriter, op, {input, op.output()});
+    if (!dynamicDimsOr.hasValue())
+      return failure();
+    SmallVector<Value> dynamicDims = dynamicDimsOr.getValue();
 
     // The shift and multiplier values.
     SmallVector<int32_t> multiplierValues;
     getValuesFromIntArrayAttribute(op.multiplier(), multiplierValues);
 
     SmallVector<int8_t> shiftValues;
     getValuesFromIntArrayAttribute(op.shift(), shiftValues);
 
if (shiftValues.size() == 1) {
       shiftArg = indexingMaps.size() - 1;
     }
 
     // Indexing maps for output values.
     indexingMaps.push_back(rewriter.getMultiDimIdentityMap(rank));
 
     // Construct the indexing maps needed for linalg.generic ops.
     Value initTensor = rewriter.create<linalg::InitTensorOp>(
-        loc, ArrayRef<Value>({}), outputTy.getShape(),
-        outputTy.getElementType());
+        loc, dynamicDims, outputTy.getShape(), outputTy.getElementType());
 
     auto linalgOp = rewriter.create<linalg::GenericOp>(
         loc, outputTy, genericInputs, ValueRange{initTensor}, indexingMaps,
         getNParallelLoopsAttrs(rank),
         [&](OpBuilder &nestedBuilder, Location nestedLoc,
             ValueRange blockArgs) {
           Value value = blockArgs[0];
           Type valueTy = value.getType();
LogicalResult matchAndRewrite(tosa::ResizeOp op,
     auto input = op.input();
     auto inputTy = input.getType().cast<ShapedType>();
     auto resultTy = op.getType().cast<ShapedType>();
     auto resultElementTy = resultTy.getElementType();
 
     auto imageH = inputTy.getShape()[1];
     auto imageW = inputTy.getShape()[2];
 
-    if (!resultTy.hasStaticShape())
+    auto dynamicDimsOr =
+        checkHasDynamicBatchDims(rewriter, op, {input, op.output()});
+    if (!dynamicDimsOr.hasValue())
       return failure();
+    SmallVector<Value> dynamicDims = dynamicDimsOr.getValue();
+
     if (op.mode() != "NEAREST_NEIGHBOR" && op.mode() != "BILINEAR")
       return failure();
 
-    auto initTensor =
-        rewriter
-            .create<linalg::InitTensorOp>(loc, ArrayRef<Value>{},
-                                          resultTy.getShape(), resultElementTy)
-            .result();
+    auto initTensor = rewriter.create<linalg::InitTensorOp>(
+        loc, dynamicDims, resultTy.getShape(), resultElementTy);
 
     SmallVector<AffineMap, 2> affineMaps = {
         rewriter.getMultiDimIdentityMap(resultTy.getRank())};
 
     auto genericOp = rewriter.create<linalg::GenericOp>(
         loc, resultTy, ValueRange({}), ValueRange{initTensor}, affineMaps,
         getNParallelLoopsAttrs(resultTy.getRank()));
     rewriter.replaceOp(op, genericOp.getResult(0));

 public:
   using OpConversionPattern<tosa::GatherOp>::OpConversionPattern;
   LogicalResult
   matchAndRewrite(tosa::GatherOp op, OpAdaptor adaptor,
                   ConversionPatternRewriter &rewriter) const final {
     auto input = adaptor.getOperands()[0];
     auto indices = adaptor.getOperands()[1];
 
-    auto inputTy = input.getType().cast<ShapedType>();
-    auto indicesTy = indices.getType().cast<ShapedType>();
     auto resultTy = op.getType().cast<ShapedType>();
 
-    if (!inputTy.hasStaticShape() || !indicesTy.hasStaticShape())
-      return rewriter.notifyMatchFailure(
-          op, "require input type to have static shape");
+    auto dynamicDimsOr =
+        checkHasDynamicBatchDims(rewriter, op, {input, indices, op.output()});
+    if (!dynamicDimsOr.hasValue())
+      return failure();
+    SmallVector<Value> dynamicDims = dynamicDimsOr.getValue();
 
     auto resultElementTy = resultTy.getElementType();
 
     auto loc = op.getLoc();
 
     auto initTensor =
         rewriter
-            .create<linalg::InitTensorOp>(loc, ArrayRef<Value>{},
-                                          resultTy.getShape(), resultElementTy)
+            .create<linalg::InitTensorOp>(loc, dynamicDims, resultTy.getShape(),
+                                          resultElementTy)
             .result();
 
     SmallVector<AffineMap, 2> affineMaps = {
         AffineMap::get(
             /*dimCount=*/resultTy.getRank(), /*symbolCount=*/0,
             {rewriter.getAffineDimExpr(0), rewriter.getAffineDimExpr(1)},
             rewriter.getContext()),
         rewriter.getMultiDimIdentityMap(resultTy.getRank())};

mlir/lib/Conversion/TosaToLinalg/TosaToLinalgNamed.cpp

 //===- 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/Tosa/Utils/CoversionUtils.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);
-}
+using namespace mlir::tosa;
 
 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>();
static mlir::Value applyPad(Location loc, Value input, ArrayRef<int64_t> pad,
   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(const 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 {
matchAndRewrite(tosa::Conv2DOp op, OpAdaptor adaptor,
     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 (!weightTy.hasStaticShape() || !biasTy.hasStaticShape())
+      return rewriter.notifyMatchFailure(
+          op, "tosa.conv ops require static shapes for weight and bias");
+
+    auto dynamicDimsOr =
+        checkHasDynamicBatchDims(rewriter, op, {input, op.output()});
+    if (!dynamicDimsOr.hasValue())
+      return failure();
+    SmallVector<Value> dynamicDims = dynamicDimsOr.getValue();
 
     if (inputETy.isUnsignedInteger())
       return rewriter.notifyMatchFailure(
           op, "tosa.conv ops does not support unsigned integer input");
 
     auto weightShape = weightTy.getShape();
 
     // Apply padding as necessary.
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);
+        loc, dynamicDims, 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);
+        loc, dynamicDims, 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());
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");
+    if (!weightTy.hasStaticShape() || !biasTy.hasStaticShape())
+      return rewriter.notifyMatchFailure(
+          op, "tosa.depthwise_conv ops require static shapes");
+
+    auto dynamicDimsOr =
+        checkHasDynamicBatchDims(rewriter, op, {input, op.output()});
+    if (!dynamicDimsOr.hasValue())
+      return failure();
+    SmallVector<Value> dynamicDims = dynamicDimsOr.getValue();
 
     auto weightShape = weightTy.getShape();
     auto resultShape = resultTy.getShape();
 
     // Apply padding as necessary.
     Attribute zeroAttr = rewriter.getZeroAttr(inputETy);
     if (isQuantized) {
       auto quantizationInfo =
matchAndRewrite(tosa::DepthwiseConv2DOp op, OpAdaptor adaptor,
     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);
+        loc, dynamicDims, 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);
+        loc, dynamicDims, 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()));
matchAndRewrite(tosa::MatMulOp op, OpAdaptor adaptor,
     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);
+    SmallVector<Value> filteredDims = condenseValues(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()) {
matchAndRewrite(tosa::FullyConnectedOp op, OpAdaptor adaptor,
     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);
+    SmallVector<Value> filteredDims = condenseValues(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()));
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())
+    auto dynamicDimsOr =
+        checkHasDynamicBatchDims(rewriter, op, {input, op.output()});
+    if (!dynamicDimsOr.hasValue())
       return failure();
+    SmallVector<Value> dynamicDims = dynamicDimsOr.getValue();
 
     // 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));
LogicalResult matchAndRewrite(tosa::MaxPool2dOp op,
     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());
+        loc, dynamicDims, 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>(
LogicalResult matchAndRewrite(tosa::AvgPool2dOp op,
 
     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())
+    auto dynamicDimsOr =
+        checkHasDynamicBatchDims(rewriter, op, {input, op.output()});
+    if (!dynamicDimsOr.hasValue())
       return failure();
+    SmallVector<Value> dynamicDims = dynamicDimsOr.getValue();
 
     // 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 poolInitTensor = rewriter.create<linalg::InitTensorOp>(
+        loc, dynamicDims, 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);
+        loc, dynamicDims, 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 genericOp = rewriter.create<linalg::GenericOp>(
           };
 
           // 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);
+          auto kH3 = rewriter.create<mlir::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);
+          auto kW3 = rewriter.create<mlir::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

mlir/lib/Dialect/Tosa/CMakeLists.txt

 add_mlir_dialect_library(MLIRTosa
+  Utils/ConversionUtils.cpp
   Utils/QuantUtils.cpp
   IR/TosaOps.cpp
 
   ADDITIONAL_HEADER_DIRS
   ${MLIR_MAIN_INCLUDE_DIR}/mlir/Dialect/Tosa
 
   DEPENDS
   MLIRStandardOpsIncGen

mlir/lib/Dialect/Tosa/Utils/ConversionUtils.cpp

+//===- ConversionUtils.cpp ------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+//
+// Utility functions for TOSA lowering
+//
+//===----------------------------------------------------------------------===//
+
+#include "mlir/Dialect/Tosa/Utils/CoversionUtils.h"
+
+using namespace mlir;
+using namespace mlir::tosa;
+
+SmallVector<StringRef>
+mlir::tosa::getNParallelLoopsAttrs(unsigned nParallelLoops) {
+  return SmallVector<StringRef>(nParallelLoops, getParallelIteratorTypeName());
+}
+
+SmallVector<Value>
+mlir::tosa::condenseValues(const SmallVector<Value> &values) {
+  SmallVector<Value> condensedValues;
+  for (auto value : values)
+    if (value)
+      condensedValues.push_back(value);
+  return condensedValues;
+}

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg-named.mlir

func @max_pool_padded(%arg0: tensor<1x6x34x62xf32>) -> () {
   // 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_dyn
+func @max_pool_dyn(%arg0: tensor<?x6x34x62xf32>) -> () {
+  // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: %[[BATCH:.+]] = tensor.dim %arg0, %[[C0]]
+  // CHECK: %[[CONST:.+]] = arith.constant -3.40282347E+38
+  // CHECK: %[[INIT:.+]] = linalg.init_tensor [%[[BATCH]], 4, 32, 62]
+  // CHECK: %[[FILL:.+]] = linalg.fill(%[[CONST]], %[[INIT]])
+  // CHECK: %[[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<?x6x34x62xf32>, tensor<3x3xf32>) outs(%[[FILL]] : tensor<?x4x32x62xf32>)
+  %0 = "tosa.max_pool2d"(%arg0) {pad = [0, 0, 0, 0], kernel = [3, 3], stride = [1, 1]} : (tensor<?x6x34x62xf32>)  -> (tensor<?x4x32x62xf32>)
+  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
 }
 
func @avg_pool(%arg0: tensor<1x6x34x62xf32>) -> (tensor<1x5x33x62xf32>) {
   // 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_dyn
+func @avg_pool_dyn(%arg0: tensor<?x6x34x62xf32>) -> (tensor<?x5x33x62xf32>) {
+  // The calculations remain the same as above, only testing for dyn behavior
+  // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: %[[BATCH:.+]] = tensor.dim %arg0, %[[C0]]
+  // CHECK: %[[PAD:.+]] = linalg.pad_tensor %arg0 low[0, 1, 1, 0] high[0, 1, 1, 0]
+  // CHECK: %[[POOLINIT:.+]] = linalg.init_tensor [%[[BATCH]], 5, 33, 62]
+  // CHECK: %[[FILL:.+]] = linalg.fill
+  // 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<?x8x36x62xf32>, tensor<4x4xf32>) outs(%[[FILL]] : tensor<?x5x33x62xf32>)
+  // CHECK: %[[INIT:.+]] = linalg.init_tensor [%[[BATCH]], 5, 33, 62]
+  // CHECK: %[[GENERIC:.+]] = linalg.generic {indexing_maps = [#map, #map], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%[[POOL]] : tensor<?x5x33x62xf32>) outs(%[[INIT]] : tensor<?x5x33x62xf32>)
+  %0 = "tosa.avg_pool2d"(%arg0) {pad = [1, 1, 1, 1], kernel = [4, 4], stride = [1, 1]} : (tensor<?x6x34x62xf32>)  -> (tensor<?x5x33x62xf32>)
+  return %0 : tensor<?x5x33x62xf32>
+}
+
+// -----
+
 // 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]]
func @conv2d_f32(%input: tensor<1x49x42x27xf32>, %weights: tensor<28x3x3x27xf32>, %bias: tensor<28xf32>) -> () {
   // 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: #[[$MAP1:.+]] = affine_map<(d0, d1, d2, d3) -> (d3)>
+// CHECK: #[[$MAP2:.+]] = affine_map<(d0, d1, d2, d3) -> (d0, d1, d2, d3)>
+
+// CHECK-LABEL: @conv2d_dyn
+func @conv2d_dyn(%input: tensor<?x49x42x27xf32>, %weights: tensor<28x3x3x27xf32>, %bias: tensor<28xf32>) -> () {
+  // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: %[[BATCH:.+]] = tensor.dim %arg0, %[[C0]]
+  // CHECK: %[[PERM:.+]] = arith.constant dense<[1, 2, 3, 0]>
+  // CHECK: %[[W:.+]] = "tosa.transpose"(%arg1, %[[PERM]])
+  // CHECK: %[[M_IN:.+]] = linalg.init_tensor [%[[BATCH]], 45, 40, 28]
+  // CHECK: %[[CST:.+]] = arith.constant 0
+  // CHECK: %[[FILL:.+]] = linalg.fill
+  // CHECK: %[[B_IN:.+]] = linalg.init_tensor [%[[BATCH]], 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<?x49x42x27xf32>, tensor<3x3x27x28xf32>) outs(%[[FILL]] : tensor<?x45x40x28xf32>)
+  // CHECK: %[[B:.+]] = linalg.generic {indexing_maps = [#[[$MAP1]], #[[$MAP2]], #[[$MAP2]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, %[[CONV]] : tensor<28xf32>, tensor<?x45x40x28xf32>) outs(%[[B_IN]] : tensor<?x45x40x28xf32>)
+  // CHECK:   %[[ADD:.+]] = arith.addf
+  // CHECK:   linalg.yield %[[ADD]] : f32
+  %0 = "tosa.conv2d"(%input, %weights, %bias) {pad = [0, 0, 0, 0], stride = [1, 1], dilation = [2, 1]} : (tensor<?x49x42x27xf32>, tensor<28x3x3x27xf32>, tensor<28xf32>)  -> (tensor<?x45x40x28xf32>)
+  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
func @depthwise_conv(%arg0 : tensor<1x7x5x3xf32>, %arg1 : tensor<3x1x3x11xf32>, %arg2 : tensor<33xf32>) -> () {
   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_dyn
+func @depthwise_conv_dyn(%arg0 : tensor<?x7x5x3xf32>, %arg1 : tensor<3x1x3x11xf32>, %arg2 : tensor<33xf32>) -> () {
+  // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: %[[BATCH:.+]] = tensor.dim %arg0, %[[C0]]
+  // CHECK: %[[INIT:.+]] = linalg.init_tensor [%[[BATCH]], 5, 5, 3, 11]
+  // CHECK: %[[CST0:.+]] = arith.constant 0
+  // CHECK: %[[FILL:.+]] = linalg.fill
+  // CHECK: %[[OUT:.+]] = linalg.init_tensor [%[[BATCH]], 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<?x7x5x3xf32>, tensor<3x1x3x11xf32>) outs(%[[FILL]] : tensor<?x5x5x3x11xf32>)
+  // CHECK: %[[COLLAPSED:.+]] = "tosa.reshape"(%[[DEPTH]]) {new_shape = [-1, 5, 5, 33]}
+  // CHECK: %[[BIAS:.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP1]], #[[$MAP1]]], iterator_types = ["parallel", "parallel", "parallel", "parallel"]} ins(%arg2, %[[COLLAPSED]] : tensor<33xf32>, tensor<?x5x5x33xf32>) outs(%[[OUT]] : tensor<?x5x5x33xf32>) {
+  // CHECK: ^bb0(%arg3: f32, %arg4: f32, %arg5: f32):  // no predecessors
+  // CHECK:   %[[ADD:.+]] = arith.addf %arg3, %arg4 : f32
+  // CHECK:   linalg.yield %[[ADD]] : f32
+  // CHECK: } -> tensor<?x5x5x33xf32>
+  %2 = "tosa.depthwise_conv2d"(%arg0, %arg1, %arg2) { pad = [0, 0, 0, 0], stride = [1, 1], dilation = [1, 1] } : (tensor<?x7x5x3xf32>, tensor<3x1x3x11xf32>, tensor<33xf32>)  -> (tensor<?x5x5x33xf32>)
+  return
+}
+
+// -----
+
+// CHECK: #[[$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]}

mlir/test/Conversion/TosaToLinalg/tosa-to-linalg.mlir

func @rescale_i8(%arg0 : tensor<2xi8>) -> () {
   %1 = "tosa.rescale"(%arg0) {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = [19689 : i32], shift = [15 : i32], scale32 = false, double_round = false, per_channel = false} : (tensor<2xi8>)  -> (tensor<2xui8>)
 
   // CHECK: return
   return
 }
 
 // -----
 
+// CHECK: #[[$MAP0:.*]] = affine_map<(d0, d1) -> (d0, d1)>
+
+// CHECK-LABEL: @rescale_i8_dyn
+func @rescale_i8_dyn(%arg0 : tensor<?x2xi8>) -> () {
+  // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: %[[BATCH:.+]] = tensor.dim %arg0, %[[C0]]
+  // CHECK: %[[INIT:.+]] = linalg.init_tensor [%[[BATCH]], 2]
+  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP0]]], iterator_types = ["parallel", "parallel"]} ins(%arg0 : tensor<?x2xi8>) outs(%[[INIT]] : tensor<?x2xi8>)
+  %0 = "tosa.rescale"(%arg0) {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = [19689 : i32], shift = [15 : i32], scale32 = false, double_round = false, per_channel = false} : (tensor<?x2xi8>)  -> (tensor<?x2xi8>)
+
+  // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: %[[BATCH:.+]] = tensor.dim %arg0, %[[C0]]
+  // CHECK: %[[INIT:.+]] = linalg.init_tensor [%[[BATCH]], 2]
+  // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP0]]], iterator_types = ["parallel", "parallel"]} ins(%arg0 : tensor<?x2xi8>) outs(%[[INIT]] : tensor<?x2xui8>)
+  %1 = "tosa.rescale"(%arg0) {input_zp = 17 : i32, output_zp = 22 : i32, multiplier = [19689 : i32], shift = [15 : i32], scale32 = false, double_round = false, per_channel = false} : (tensor<?x2xi8>)  -> (tensor<?x2xui8>)
+
+  return
+}
+// -----
+
 // CHECK: #[[$MAP0:.*]] = affine_map<(d0) -> (d0)>
 
 // CHECK-LABEL: @rescale_ui8
 func @rescale_ui8(%arg0 : tensor<2xui8>) -> () {
   // CHECK: [[C0:%.+]] = arith.constant 19689
   // CHECK: [[C1:%.+]] = arith.constant 15
   // CHECK: [[INIT:%.+]] = linalg.init_tensor [2]
   // CHECK: [[GENERIC:%.+]] = linalg.generic {indexing_maps = [#[[$MAP0]], #[[$MAP0]]], iterator_types = ["parallel"]} ins(%arg0 : tensor<2xui8>) outs([[INIT]] : tensor<2xi8>)
func @gather_float(%arg0: tensor<2x3x2xf32>, %arg1: tensor<2x3xi32>) -> () {
   // CHECK:   %[[CAST:.+]] = arith.index_cast %[[ARG0]]
   // CHECK:   %[[IDX2:.+]] = linalg.index 2
   // CHECK:   %[[EXTRACT:.+]] = tensor.extract %arg0[%[[IDX0]], %[[CAST]], %[[IDX2]]] : tensor<2x3x2xf32>
   // CHECK:   linalg.yield %[[EXTRACT]]
   %0 = "tosa.gather"(%arg0, %arg1)  : (tensor<2x3x2xf32>, tensor<2x3xi32>)  -> (tensor<2x3x2xf32>)
   return
 }
 
+// CHECK-LABEL: @gather_float_dyn
+func @gather_float_dyn(%arg0: tensor<?x3x2xf32>, %arg1: tensor<?x3xi32>) -> () {
+  // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: %[[BATCH:.+]] = tensor.dim %arg0, %[[C0]]
+  // CHECK: %[[INIT:.+]] = linalg.init_tensor [%[[BATCH]], 3, 2]
+  // CHECK: %[[GENERIC:.+]] = linalg.generic {indexing_maps = [#map0, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg1 : tensor<?x3xi32>) outs(%[[INIT]] : tensor<?x3x2xf32>)
+  // CHECK: ^bb0(%[[ARG0:.+]]: i32, %[[ARG1:.+]]: f32)
+  // CHECK:   %[[IDX0:.+]] = linalg.index 0
+  // CHECK:   %[[CAST:.+]] = arith.index_cast %[[ARG0]]
+  // CHECK:   %[[IDX2:.+]] = linalg.index 2
+  // CHECK:   %[[EXTRACT:.+]] = tensor.extract %arg0[%[[IDX0]], %[[CAST]], %[[IDX2]]] : tensor<?x3x2xf32>
+  // CHECK:   linalg.yield %[[EXTRACT]]
+  %0 = "tosa.gather"(%arg0, %arg1)  : (tensor<?x3x2xf32>, tensor<?x3xi32>)  -> (tensor<?x3x2xf32>)
+  return
+}
+
 // CHECK-LABEL: @gather_int
 func @gather_int(%arg0: tensor<2x3x2xi32>, %arg1: tensor<2x3xi32>) -> () {
   // CHECK: %[[INIT:.+]] = linalg.init_tensor [2, 3, 2]
   // CHECK: %[[GENERIC:.+]] = linalg.generic {indexing_maps = [#map0, #map1], iterator_types = ["parallel", "parallel", "parallel"]} ins(%arg1 : tensor<2x3xi32>) outs(%[[INIT]] : tensor<2x3x2xi32>)
   // CHECK: ^bb0(%[[ARG0:.+]]: i32, %[[ARG1:.+]]: i32)
   // CHECK:   %[[IDX0:.+]] = linalg.index 0
   // CHECK:   %[[CAST:.+]] = arith.index_cast %[[ARG0]]
   // CHECK:   %[[IDX2:.+]] = linalg.index 2
func @resize_bilinear_int(%input: tensor<1x2x2x1xi8>) -> () {
   // CHECK: %[[NDY:.+]] = arith.subi %[[SCALE]], %[[DY]]
   // CHECK: %[[WLO:.+]] = arith.muli %[[LO]], %[[NDY]]
   // CHECK: %[[WHI:.+]] = arith.muli %[[HI]], %[[DY]]
   // CHECK: %[[RESULT:.+]] = arith.addi %[[WLO]], %[[WHI]]
   // CHECK: linalg.yield %[[RESULT]]
   %output = "tosa.resize"(%input) { output_size = [4, 4], stride = [128, 128], offset = [1, 2], stride_fp = [0. : f32, 0. : f32], offset_fp = [0. : f32, 0. : f32], shift = 8 : i32, mode = "BILINEAR" } : (tensor<1x2x2x1xi8>)  -> (tensor<1x4x4x1xi32>)
   return
 }
+
+// -----
+
+// CHECK-LABEL: @resize_dyn
+func @resize_dyn(%input: tensor<?x2x2x1xi8>) -> () {
+    // CHECK: %[[C0:.+]] = arith.constant 0
+  // CHECK: %[[BATCH:.+]] = tensor.dim %arg0, %[[C0]]
+  // CHECK: %[[INIT:.+]] = linalg.init_tensor [%[[BATCH]], 4, 4, 1]
+  // CHECK: %[[GENERIC:.+]] = linalg.generic
+  %output = "tosa.resize"(%input) { output_size = [4, 4], stride = [128, 128], offset = [1, 2], stride_fp = [0. : f32, 0. : f32], offset_fp = [0. : f32, 0. : f32], shift = 8 : i32, mode = "BILINEAR" } : (tensor<?x2x2x1xi8>)  -> (tensor<?x4x4x1xi32>)
+  return
+}

utils/bazel/llvm-project-overlay/mlir/BUILD.bazel

cc_library(
     hdrs = glob([
         "include/mlir/Dialect/Tosa/IR/*.h",
         "include/mlir/Dialect/Tosa/Utils/*.h",
         "include/mlir/Dialect/Tosa/Transforms/*.h",
     ]),
     includes = ["include"],
     deps = [
         ":Analysis",
+        ":ArithmeticDialect",
         ":Dialect",
+        ":DialectUtils",
         ":IR",
         ":InferTypeOpInterface",
         ":LoopLikeInterface",
         ":Pass",
         ":QuantOps",
-        ":SideEffectInterfaces",
         ":StandardOps",
         ":TensorDialect",
         ":TosaDialectIncGen",
         ":TosaInterfacesIncGen",
         ":TosaPassIncGen",
         ":TransformUtils",
         "//llvm:Support",
     ],