Rewrite shape and size OP (llvm#285)

* add shape inference

* Revert "add shape inference"

This reverts commit f9d42f39e68e14b5648abccfc8617fff00244d16.

* add rewrite rules

* test cases

* format

* add constraint

* response to review

* response to review

src/Dialect/ONNX/ONNXOps.td.inc

@@ -4729,6 +4729,7 @@ def ONNXSequenceLengthOp:ONNX_Op<"SequenceLength",
 
 def ONNXShapeOp:ONNX_Op<"Shape",
   [NoSideEffect, DeclareOpInterfaceMethods<ShapeInferenceOpInterface>]> {
+  let hasCanonicalizer = 1;
   let summary = "ONNX Shape operation";
   let description = [{
   "Takes a tensor as input and outputs an 1D int64 tensor containing the shape of the input tensor."
@@ -4863,6 +4864,7 @@ def ONNXSinhOp:ONNX_Op<"Sinh",
 
 def ONNXSizeOp:ONNX_Op<"Size",
   [NoSideEffect]> {
+  let hasCanonicalizer = 1;
   let summary = "ONNX Size operation";
   let description = [{
   "Takes a tensor as input and outputs a int64 scalar that equals to the total number of elements of the input tensor."

src/Transform/ONNX/Rewrite.cpp

@@ -27,6 +27,29 @@ DenseElementsAttr createDenseElementsAttrFromFloatAttr(
   return mlir::DenseElementsAttr::get(tensorType, llvm::makeArrayRef(values));
 }
 
+// Create a DenseElementsAttr based on the shape of type.
+DenseElementsAttr createDenseElementsAttrFromShape(
+    PatternRewriter &rewriter, Value value) {
+  auto inType = value.getType().cast<ShapedType>();
+  auto shape = inType.getShape();
+  SmallVector<int64_t, 1> dims = {inType.getRank()};
+  SmallVector<int64_t, 4> values(shape.begin(), shape.end());
+  auto tensorType =
+      mlir::RankedTensorType::get(dims, rewriter.getIntegerType(64));
+  return mlir::DenseElementsAttr::get(tensorType, llvm::makeArrayRef(values));
+}
+
+// Create a DenseElementsAttr based on the size of type.
+DenseElementsAttr createDenseElementsAttrFromSize(
+    PatternRewriter &rewriter, Value value) {
+  auto inType = value.getType().cast<ShapedType>();
+  SmallVector<int64_t, 1> dims(1, 1);
+  SmallVector<int64_t, 1> values = {inType.getNumElements()};
+  auto tensorType =
+      mlir::RankedTensorType::get(dims, rewriter.getIntegerType(64));
+  return mlir::DenseElementsAttr::get(tensorType, llvm::makeArrayRef(values));
+}
+
 // If 'lhs' is not NoneType, return 'lhs - rhs'.
 // Otherwise, return '-rhs'.
 Value subtractOrNeg(
@@ -128,3 +151,15 @@ void ONNXBatchNormalizationTestModeOp::getCanonicalizationPatterns(
     OwningRewritePatternList &results, MLIRContext *context) {
   results.insert<FuseBatchNormTestModeConvPattern>(context);
 }
+
+/// on the ONNXShapeOp.
+void ONNXShapeOp::getCanonicalizationPatterns(
+    OwningRewritePatternList &results, MLIRContext *context) {
+  results.insert<ShapeToConstantPattern>(context);
+}
+
+/// on the ONNXSizeOp.
+void ONNXSizeOp::getCanonicalizationPatterns(
+    OwningRewritePatternList &results, MLIRContext *context) {
+  results.insert<SizeToConstantPattern>(context);
+}

src/Transform/ONNX/Rewrite.td

@@ -28,6 +28,14 @@ include "src/Dialect/ONNX/ONNXOps.td"
 def createDenseElementsAttrFromFloatAttr : NativeCodeCall<
   "createDenseElementsAttrFromFloatAttr($_builder, $0.getType().cast<ShapedType>().getElementType(), $1)">;
 
+// Create a DenseElementsAttr from the shape of the type of a value.
+def createDenseElementsAttrFromShape : NativeCodeCall<
+  "createDenseElementsAttrFromShape($_builder, $0)">;
+
+// Create a DenseElementsAttr from the size of the type of a value.
+def createDenseElementsAttrFromSize : NativeCodeCall<
+  "createDenseElementsAttrFromSize($_builder, $0)">;
+
 // If '$1' is not NoneType, do subtraction '$1 - $2'.
 // Otherwise, take the negative of '$2'.
 def subtractOrNeg: NativeCodeCall<
@@ -172,4 +180,25 @@ def FuseBatchNormTestModeConvPattern: Pat<
      $auto_pad, $dilation, $group, $kernel_shape, $pads, $strides)
 >;
 
+def IsStaticShapeTensor:
+   Constraint<
+     CPred<
+       "$_self.getType().cast<::mlir::ShapedType>().hasStaticShape()">,
+     "hasStaticShape">;
+
+def ShapeToConstantPattern: Pat<
+     (ONNXShapeOp $A),
+     (ONNXConstantOp
+        (GetNullAttr),
+        (createDenseElementsAttrFromShape $A)),
+     [(IsStaticShapeTensor:$A)]
+>;
+
+def SizeToConstantPattern: Pat<
+     (ONNXSizeOp $A),
+     (ONNXConstantOp
+        (GetNullAttr),
+        (createDenseElementsAttrFromSize $A)),
+     [(IsStaticShapeTensor:$A)]
+>;
 #endif // ONNX_REWRITE

test/mlir/onnx/onnx_canonicalization.mlir

@@ -222,3 +222,49 @@ func @test_transpose_fusion_removal(%arg0: tensor<10x11x12x13xf32>) -> tensor<10
   // CHECK-NEXT: return %arg0 : tensor<10x11x12x13xf32>
   "std.return"(%1) : (tensor<10x11x12x13xf32>) -> ()
 }
+
+// -----
+
+func @test_shape1(%arg0 : tensor<2x4x8x16xf32>) -> tensor<*xi64> {
+  %0 = "onnx.Shape"(%arg0) : (tensor<2x4x8x16xf32>) -> tensor<*xi64>
+  return %0 : tensor<*xi64>
+
+  // CHECK-LABEL: @test_shape1
+  // CHECK-NEXT: %0 = "onnx.Constant"() {value = dense<[2, 4, 8, 16]> : tensor<4xi64>} : () -> tensor<*xi64>
+  // CHECK-NEXT: %0 : tensor<*xi64>
+}
+
+// -----
+
+func @test_shape2(%arg0 : tensor<?x4x8x16xf32>) -> tensor<*xi64> {
+  %0 = "onnx.Shape"(%arg0) : (tensor<?x4x8x16xf32>) -> tensor<*xi64>
+  return %0 : tensor<*xi64>
+
+  // CHECK-LABEL: @test_shape2
+  // CHECK-NEXT: %0 = "onnx.Shape"(%arg0) : (tensor<?x4x8x16xf32>) -> tensor<*xi64>
+  // CHECK-NEXT: return %0 : tensor<*xi64>
+}
+
+
+// -----
+
+func @test_size1(%arg0 : tensor<2x4x8x16xf32>) -> tensor<*xi64> {
+  %0 = "onnx.Size"(%arg0) : (tensor<2x4x8x16xf32>) -> tensor<*xi64>
+  return %0 : tensor<*xi64>
+
+  // CHECK-LABEL: @test_size1
+  // CHECK-NEXT: %0 = "onnx.Constant"() {value = dense<1024> : tensor<1xi64>} : () -> tensor<*xi64>
+  // CHECK-NEXT: %0 : tensor<*xi64>
+}
+
+// -----
+
+func @test_size2(%arg0 : tensor<*xf32>) -> tensor<*xi64> {
+  %0 = "onnx.Size"(%arg0) : (tensor<*xf32>) -> tensor<*xi64>
+  return %0 : tensor<*xi64>
+
+  // CHECK-LABEL: @test_size2
+  // CHECK-NEXT: %0 = "onnx.Size"(%arg0) : (tensor<*xf32>) -> tensor<*xi64>
+  // CHECK-NEXT: return %0 : tensor<*xi64>
+}
+

utils/gen_onnx_mlir.py

@@ -321,7 +321,7 @@
 ]
 
 # Operations supporting canonicalization.
-OpsWithCanonicalizer = ['Add', 'Identity', 'Gemm', 'Conv', 'Cast', 'Transpose', 'Dropout']
+OpsWithCanonicalizer = ['Add', 'Identity', 'Gemm', 'Conv', 'Cast', 'Transpose', 'Dropout', 'Shape', 'Size']
 
 # Operations who have operands that, if produced by constant operations, should
 # be promoted to become an attribute (via attribute promotion).