[mlir][linalg] Add canonicalizers for depthwise conv

[rsuderman]

Summary

There are two main versions of depthwise conv depending whether the multiplier
is 1 or not. In cases where m == 1 we should use the version without the
multiplier channel as it can perform greater optimization.

Add lowering for the quantized/float versions to have a multiplier of one.

[rsuderman]

Rebased on head to retrigger presubmits

[nicolasvasilache]

This feels like you want the inverse of the generalize transformation ..
The way it would work is you'd:

1. replace the named op by the generic impl
2. drop the unit dims you want out
3. compare that to the generic coming from a subset of candidate ops.

Basically that would avoid and inverse problem and replace it by a search and 2 forward problems (i.e. just compare the generated version from a database of named ops).
I could see how to get that in the absence of attributes but attributes are definitely more tricky (unless you know they are not going to be modified which seems to be the case here).

We had discussed this in the past with @antiagainst ; do you see an opportunity to make progress on the above?

[antiagainst]

We had discussed this in the past with @antiagainst ; do you see an opportunity to make progress on the above?

Sorry for the late reply. Missed this one. I haven't been able to find time to work on the inverse of Linalg generalization. Likely also true for the near future. So I'd +1 to land this for now given Rob already have it to serve the current use cases.

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

3014

Can cast to LinalgOp and then check hasTensorSemantics.

[rsuderman]

Fixed for antiagainst@ comments. Checking clang-tidy passes.

mlir/lib/Dialect/Linalg/IR/LinalgOps.cpp

for (auto result : llvm::zip(op->getResults(), newOp->getResults())) {
         replacements.push_back(newResult);
       }
     }
     rewriter.replaceOp(op, replacements);
 
     return success();
   }
 };
+
+static llvm::SmallVector<int64_t> GetIndicesVector(int start, int end) {
+  return llvm::to_vector<2>(llvm::seq<int64_t>(start, end));
+}
+
+LogicalResult MatchAndReplaceDepthwiseConv(Operation *operation, Value input,
+                                           Value kernel, Value iZp, Value kZp,
+                                           Value init, Attribute stride,
+                                           Attribute dilation,
+                                           PatternRewriter &rewriter) {
+  Location loc = operation->getLoc();
+  // Exit out on the memref version of this operation.
+  if (operation->getNumResults() != 1)
+    return failure();
+  auto result = operation->getResult(0);
+
+  auto kernelTy = kernel.getType().dyn_cast<RankedTensorType>();
+  auto initTy = init.getType().dyn_cast<RankedTensorType>();
+  auto resultTy = result.getType().template dyn_cast<RankedTensorType>();
+  if (!kernelTy || !initTy || !resultTy)
+    return failure();
+
+  if (kernelTy.getDimSize(3) != 1)
+    return failure();
+
+  // Collapse kernel dims.
+  SmallVector<ReassociationIndices, 4> collapsedKernelDims = {
+      GetIndicesVector(0, 1), GetIndicesVector(1, 2), GetIndicesVector(2, 4)};
+  auto newKernelTy = RankedTensorType::get(
+      {kernelTy.getDimSize(0), kernelTy.getDimSize(1), kernelTy.getDimSize(2)},
+      kernelTy.getElementType());
+  auto collapsedKernel = rewriter.create<linalg::TensorCollapseShapeOp>(
+      loc, newKernelTy, kernel, collapsedKernelDims);
+
+  // Collapse init dims.
+  SmallVector<ReassociationIndices, 4> collapsedInitDims = {
+      GetIndicesVector(0, 1), GetIndicesVector(1, 2), GetIndicesVector(2, 3),
+      GetIndicesVector(3, 5)};
+  auto newInitTy =
+      RankedTensorType::get({initTy.getDimSize(0), initTy.getDimSize(1),
+                             initTy.getDimSize(2), initTy.getDimSize(3)},
+                            initTy.getElementType());
+  auto collapsedInit = rewriter.create<linalg::TensorCollapseShapeOp>(
+      loc, newInitTy, init, collapsedInitDims);
+
+  Value newConv;
+  if (isa<DepthwiseConv2DNhwcOp>(operation)) {
+    newConv = rewriter
+                  .create<DepthwiseConv2DNhwOp>(
+                      loc, newInitTy, ValueRange{input, collapsedKernel},
+                      ValueRange{collapsedInit}, stride, dilation)
+                  .getResult(0);
+  } else if (isa<DepthwiseConv2DNhwcQOp>(operation)) {
+    newConv =
+        rewriter
+            .create<DepthwiseConv2DNhwQOp>(
+                loc, newInitTy, ValueRange{input, collapsedKernel, iZp, kZp},
+                ValueRange{collapsedInit}, stride, dilation)
+            .getResult(0);
+  }
+
+  if (!newConv)
+    return failure();
+
+  // Expand dimensions back out to
+  rewriter.replaceOpWithNewOp<linalg::TensorExpandShapeOp>(
+      operation, resultTy, newConv, collapsedInitDims);
+  return success();
+}
+
+struct SimplifyDepthwiseConvOp
+    : public OpRewritePattern<DepthwiseConv2DNhwcOp> {
+  using OpRewritePattern<DepthwiseConv2DNhwcOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(DepthwiseConv2DNhwcOp op,
+                                PatternRewriter &rewriter) const override {
+    Operation *operation = op.getOperation();
+    Value input = op.getInputOperand(0)->get();
+    Value kernel = op.getInputOperand(1)->get();
+    Value init = op.getOutputOperand(0)->get();
+
+    auto stride = op.strides();
+    auto dilation = op.dilations();
+
+    return MatchAndReplaceDepthwiseConv(operation, input, kernel, nullptr,
+                                        nullptr, init, stride, dilation,
+                                        rewriter);
+  }
+};
+
+struct SimplifyDepthwiseConvQOp
+    : public OpRewritePattern<DepthwiseConv2DNhwcQOp> {
+  using OpRewritePattern<DepthwiseConv2DNhwcQOp>::OpRewritePattern;
+
+  LogicalResult matchAndRewrite(DepthwiseConv2DNhwcQOp op,
+                                PatternRewriter &rewriter) const override {
+    Operation *operation = op.getOperation();
+    Value input = op.getInputOperand(0)->get();
+    Value kernel = op.getInputOperand(1)->get();
+    Value iZp = op.getInputOperand(2)->get();
+    Value kZp = op.getInputOperand(3)->get();
+    Value init = op.getOutputOperand(0)->get();
+
+    auto stride = op.strides();
+    auto dilation = op.dilations();
+
+    return MatchAndReplaceDepthwiseConv(operation, input, kernel, iZp, kZp,
+                                        init, stride, dilation, rewriter);
+  }
+};
+
 } // namespace
 
 #define LINALGOP_FOLDERS(XXX)                                                  \
   LogicalResult XXX::fold(ArrayRef<Attribute>,                                 \
                           SmallVectorImpl<OpFoldResult> &) {                   \
     return foldMemRefCast(*this);                                              \
   }
 

 // .cpp.inc.
 
 //===----------------------------------------------------------------------===//
 // LinalgDialect
 //===----------------------------------------------------------------------===//
 
 void LinalgDialect::getCanonicalizationPatterns(
     RewritePatternSet &results) const {
-  results.add<EraseDeadLinalgOp, FoldTensorCastOp>(getContext());
+  results.add<EraseDeadLinalgOp, FoldTensorCastOp, SimplifyDepthwiseConvOp,
+              SimplifyDepthwiseConvQOp>(getContext());
 }

mlir/test/Dialect/Linalg/canonicalize.mlir

%r = linalg.tiled_loop (%iv0, %iv1) = (%c0, %c0)
     %inner_dim = tensor.dim %in1, %c0 : tensor<?x?xf32>
     %cast1 = std.index_cast %inner_dim : index to i32
     %cast2 = std.sitofp %cast1 : i32 to f32
     %fill = linalg.fill(%cast2, %out1) : f32, tensor<?x?xf32> -> tensor<?x?xf32>
     linalg.yield %fill : tensor<?x?xf32>
   }
   return %r : tensor<?x?xf32>
 }
+
+// -----
+
+// CHECK-LABEL: @depthwise_conv
+func @depthwise_conv(%arg0: tensor<?x?x?x?xf32>, %arg1: tensor<?x?x?x1xf32>, %arg2: tensor<?x?x?x?x1xf32>) -> tensor<?x?x?x?x1xf32> {
+  // CHECK-DAG: %[[KERNEL:.+]] = linalg.tensor_collapse_shape %arg1 {{\[\[}}0], [1], [2, 3]]
+  // CHECK-DAG: %[[INIT:.+]] = linalg.tensor_collapse_shape %arg2 {{\[\[}}0], [1], [2], [3, 4]]
+  // CHECK-DAG: %[[CONV:.+]] = linalg.depthwise_conv2D_nhw {dilations = dense<1> : tensor<2xi64>, strides = dense<2> : tensor<2xi64>} ins(%arg0, %[[KERNEL]] : tensor<?x?x?x?xf32>, tensor<?x?x?xf32>) outs(%[[INIT]] : tensor<?x?x?x?xf32>)
+  // CHECK: %[[OUT:.+]] = linalg.tensor_expand_shape %[[CONV]] {{\[\[}}0], [1], [2], [3, 4]]
+  %0 = linalg.depthwise_conv2D_nhwc {dilations = dense<1> : tensor<2xi64>, strides = dense<2> : tensor<2xi64>} ins(%arg0, %arg1 : tensor<?x?x?x?xf32>, tensor<?x?x?x1xf32>) outs(%arg2 : tensor<?x?x?x?x1xf32>) -> tensor<?x?x?x?x1xf32>
+  return %0 : tensor<?x?x?x?x1xf32>
+}
+
+
+// -----
+
+// CHECK-LABEL: @depthwise_conv_q
+func @depthwise_conv_q(%arg0: tensor<?x?x?x?xi8>, %arg1: tensor<?x?x?x1xi8>, %arg2: tensor<?x?x?x?x1xi32>, %arg3 : i32, %arg4 : i32) -> tensor<?x?x?x?x1xi32> {
+  // CHECK-DAG: %[[KERNEL:.+]] = linalg.tensor_collapse_shape %arg1 {{\[\[}}0], [1], [2, 3]]
+  // CHECK-DAG: %[[INIT:.+]] = linalg.tensor_collapse_shape %arg2 {{\[\[}}0], [1], [2], [3, 4]]
+  // CHECK-DAG: %[[CONV:.+]] = linalg.depthwise_conv2D_nhw_q {dilations = dense<1> : tensor<2xi64>, strides = dense<2> : tensor<2xi64>} ins(%arg0, %[[KERNEL]], %arg3, %arg4 : tensor<?x?x?x?xi8>, tensor<?x?x?xi8>, i32, i32) outs(%[[INIT]] : tensor<?x?x?x?xi32>)
+  // CHECK: %[[OUT:.+]] = linalg.tensor_expand_shape %[[CONV]] {{\[\[}}0], [1], [2], [3, 4]]
+  %0 = linalg.depthwise_conv2D_nhwc_q {dilations = dense<1> : tensor<2xi64>, strides = dense<2> : tensor<2xi64>} ins(%arg0, %arg1, %arg3, %arg4 : tensor<?x?x?x?xi8>, tensor<?x?x?x1xi8>, i32, i32) outs(%arg2 : tensor<?x?x?x?x1xi32>) -> tensor<?x?x?x?x1xi32>
+  return %0 : tensor<?x?x?x?x1xi32>
+}
+
+