[MLIR][ONNX] Add OnnxToTorch support for GlobalAveragePool op. (#2692)

This commit adds the OnnxToTorch support for GlobalAveragePool op.

Signed-Off By: [email protected]

Author: vivekkhandelwal1

lib/Conversion/TorchOnnxToTorch/DefaultDomainGtoP.cpp

@@ -242,6 +242,77 @@ void mlir::torch::onnx_c::populateDefaultDomainGtoP(
             binder.op, resultType, mm, c, constBeta);
         return success();
       });
+  patterns.onOp(
+      "GlobalAveragePool", 1,
+      [](OpBinder binder, ConversionPatternRewriter &rewriter) {
+        Torch::ValueTensorType resultType;
+        Value operand;
+        if (binder.tensorOperand(operand) ||
+            binder.tensorResultType(resultType))
+          return failure();
+
+        auto inputTensorType = operand.getType().cast<Torch::ValueTensorType>();
+        if (!inputTensorType || !inputTensorType.hasSizes()) {
+          return rewriter.notifyMatchFailure(
+              binder.op, "Expected input type having sizes");
+        }
+        ArrayRef<int64_t> inputShape = inputTensorType.getSizes();
+        unsigned inputRank = inputShape.size();
+        if (!resultType || !resultType.hasSizes()) {
+          return rewriter.notifyMatchFailure(
+              binder.op, "Expected result type having sizes");
+        }
+        ArrayRef<int64_t> resultShape = resultType.getSizes();
+
+        SmallVector<Value> cstKernel, cstPadding, cstStrides;
+        Value cstZero = rewriter.create<Torch::ConstantIntOp>(
+            binder.getLoc(), rewriter.getI64IntegerAttr(0));
+        Value cstOne = rewriter.create<Torch::ConstantIntOp>(
+            binder.getLoc(), rewriter.getI64IntegerAttr(1));
+        for (unsigned i = 2; i < inputRank; i++) {
+          int64_t kernelSize = inputShape[i] - resultShape[i] + 1;
+          cstKernel.push_back(rewriter.create<Torch::ConstantIntOp>(
+              binder.getLoc(), rewriter.getI64IntegerAttr(kernelSize)));
+          cstPadding.push_back(cstZero);
+          cstStrides.push_back(cstOne);
+        }
+        Value kernelSizeList = rewriter.create<Torch::PrimListConstructOp>(
+            binder.getLoc(),
+            Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
+            cstKernel);
+        Value paddingList = rewriter.create<Torch::PrimListConstructOp>(
+            binder.getLoc(),
+            Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
+            cstPadding);
+        Value stridesList = rewriter.create<Torch::PrimListConstructOp>(
+            binder.getLoc(),
+            Torch::ListType::get(Torch::IntType::get(binder.op->getContext())),
+            cstStrides);
+        Value cstFalse = rewriter.create<Torch::ConstantBoolOp>(binder.getLoc(), false);
+        Value cstCeilMode = cstFalse;
+        Value cstCountIncludePad = cstFalse;
+        Value cstNone = rewriter.create<Torch::ConstantNoneOp>(binder.getLoc());
+
+        if (inputRank == 3) {
+          rewriter.replaceOpWithNewOp<Torch::AtenAvgPool1dOp>(
+              binder.op, resultType, operand, kernelSizeList, stridesList,
+              paddingList, cstCeilMode, cstCountIncludePad);
+          return success();
+        } else if (inputRank == 4) {
+          rewriter.replaceOpWithNewOp<Torch::AtenAvgPool2dOp>(
+              binder.op, resultType, operand, kernelSizeList, stridesList,
+              paddingList, cstCeilMode, cstCountIncludePad,
+              /*divisor_override=*/cstNone);
+          return success();
+        } else if (inputRank == 5) {
+          rewriter.replaceOpWithNewOp<Torch::AtenAvgPool3dOp>(
+              binder.op, resultType, operand, kernelSizeList, stridesList,
+              paddingList, cstCeilMode, cstCountIncludePad,
+              /*divisor_override=*/cstNone);
+          return success();
+        }
+        return failure();
+      });
   patterns.onOp("LeakyRelu", 16,
                 [](OpBinder binder, ConversionPatternRewriter &rewriter) {
                   Torch::ValueTensorType resultType;

test/Conversion/TorchOnnxToTorch/simple_ops_g_to_p.mlir

@@ -265,3 +265,39 @@ func.func @test_hardsigmoid_default(%arg0: !torch.vtensor<[3,4,5],f32>) -> !torc
   %0 = torch.operator "onnx.HardSigmoid"(%arg0) : (!torch.vtensor<[3,4,5],f32>) -> !torch.vtensor<[3,4,5],f32>
   return %0 : !torch.vtensor<[3,4,5],f32>
 }
+
+// -----
+
+// CHECK-LABEL: @test_globalaveragepool
+func.func @test_globalaveragepool(%arg0: !torch.vtensor<[1,3,5,5],f32>) -> !torch.vtensor<[1,3,1,1],f32> attributes {torch.onnx_meta.ir_version = 3 : si64, torch.onnx_meta.opset_version = 1 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
+  // CHECK: %[[C0:.*]] = torch.constant.int 0
+  // CHECK: %[[C1:.*]] = torch.constant.int 1
+  // CHECK: %[[C5:.*]] = torch.constant.int 5
+  // CHECK: %[[C5_0:.*]] = torch.constant.int 5
+  // CHECK: %[[KERNELSIZE:.*]] = torch.prim.ListConstruct %[[C5]], %[[C5_0]] : (!torch.int, !torch.int) -> !torch.list<int>
+  // CHECK: %[[PADDING:.*]] = torch.prim.ListConstruct %[[C0]], %[[C0]] : (!torch.int, !torch.int) -> !torch.list<int>
+  // CHECK: %[[STRIDE:.*]] = torch.prim.ListConstruct %[[C1]], %[[C1]] : (!torch.int, !torch.int) -> !torch.list<int>
+  // CHECK: %[[FALSE:.*]] = torch.constant.bool false
+  // CHECK: %[[NONE:.*]] = torch.constant.none
+  // CHECK: torch.aten.avg_pool2d %arg0, %[[KERNELSIZE]], %[[STRIDE]], %[[PADDING]], %[[FALSE]], %[[FALSE]], %[[NONE]] : !torch.vtensor<[1,3,5,5],f32>, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.bool, !torch.bool, !torch.none -> !torch.vtensor<[1,3,1,1],f32>
+  %0 = torch.operator "onnx.GlobalAveragePool"(%arg0) : (!torch.vtensor<[1,3,5,5],f32>) -> !torch.vtensor<[1,3,1,1],f32>
+  return %0 : !torch.vtensor<[1,3,1,1],f32>
+}
+
+// -----
+
+// CHECK-LABEL: @test_globalaveragepool_precomputed
+func.func @test_globalaveragepool_precomputed(%arg0: !torch.vtensor<[1,1,3,3],f32>) -> !torch.vtensor<[1,1,1,1],f32> attributes {torch.onnx_meta.ir_version = 3 : si64, torch.onnx_meta.opset_version = 1 : si64, torch.onnx_meta.producer_name = "backend-test", torch.onnx_meta.producer_version = ""} {
+  // CHECK: %[[C0:.*]] = torch.constant.int 0
+  // CHECK: %[[C1:.*]] = torch.constant.int 1
+  // CHECK: %[[C3:.*]] = torch.constant.int 3
+  // CHECK: %[[C3_0:.*]] = torch.constant.int 3
+  // CHECK: %[[KERNELSIZE:.*]] = torch.prim.ListConstruct %[[C3]], %[[C3_0]] : (!torch.int, !torch.int) -> !torch.list<int>
+  // CHECK: %[[PADDING:.*]] = torch.prim.ListConstruct %[[C0]], %[[C0]] : (!torch.int, !torch.int) -> !torch.list<int>
+  // CHECK: %[[STRIDE:.*]] = torch.prim.ListConstruct %[[C1]], %[[C1]] : (!torch.int, !torch.int) -> !torch.list<int>
+  // CHECK: %[[FALSE:.*]] = torch.constant.bool false
+  // CHECK: %[[NONE:.*]] = torch.constant.none
+  // CHECK: torch.aten.avg_pool2d %arg0, %[[KERNELSIZE]], %[[STRIDE]], %[[PADDING]], %[[FALSE]], %[[FALSE]], %[[NONE]] : !torch.vtensor<[1,1,3,3],f32>, !torch.list<int>, !torch.list<int>, !torch.list<int>, !torch.bool, !torch.bool, !torch.none -> !torch.vtensor<[1,1,1,1],f32>
+  %0 = torch.operator "onnx.GlobalAveragePool"(%arg0) : (!torch.vtensor<[1,1,3,3],f32>) -> !torch.vtensor<[1,1,1,1],f32>
+  return %0 : !torch.vtensor<[1,1,1,1],f32>
+}