failed to legalize operation 'torch.operator' that was explicitly marked illegal : onnx.Resize

[pravg-amd]

For the below IR,

module {
  func.func @resize(%arg0: !torch.vtensor<[1,3,224,224],f32>) -> !torch.vtensor<[1,3,800,800],f32> attributes {torch.onnx_meta.ir_version = 8 : si64, torch.onnx_meta.opset_version = 21 : si64, torch.onnx_meta.opset_versions = {ai.onnx.ml = 5 : si64, ai.onnx.preview.training = 1 : si64, ai.onnx.training = 1 : si64, com.microsoft = 1 : si64, com.microsoft.experimental = 1 : si64, com.microsoft.nchwc = 1 : si64, org.pytorch.aten = 1 : si64}, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "2.1.0"} {
    %none = torch.constant.none
    %cst = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__transform_Concat_1_output_0> : tensor<4xsi64>} : () -> !torch.vtensor<[4],si64>
    %0 = torch.operator "onnx.Resize"(%arg0, %none, %none, %cst) {torch.onnx.antialias = 0 : si64, torch.onnx.coordinate_transformation_mode = "half_pixel", torch.onnx.cubic_coeff_a = -7.500000e-01 : f32, torch.onnx.exclude_outside = 0 : si64, torch.onnx.extrapolation_value = 0.000000e+00 : f32, torch.onnx.keep_aspect_ratio_policy = "stretch", torch.onnx.mode = "linear", torch.onnx.nearest_mode = "floor"} : (!torch.vtensor<[1,3,224,224],f32>, !torch.none, !torch.none, !torch.vtensor<[4],si64>) -> !torch.vtensor<[1,3,800,800],f32>
    return %0 : !torch.vtensor<[1,3,800,800],f32>
  }
}

{-#
  dialect_resources: {
    builtin: {
      __transform_Concat_1_output_0: "0x080000000100000000000000030000000000000020030000000000002003000000000000"
    }
  }
#-}

Getting error as

model_torch_onnx.mlir:5:10: error: failed to legalize operation 'torch.operator' that was explicitly marked illegal
    %0 = torch.operator "onnx.Resize"(%arg0, %none, %none, %cst) {torch.onnx.antialias = 0 : si64, torch.onnx.coordinate_transformation_mode = "half_pixel", torch.onnx.cubic_coeff_a = -7.500000e-01 : f32, torch.onnx.exclude_outside = 0 : si64, torch.onnx.extrapolation_value = 0.000000e+00 : f32, torch.onnx.keep_aspect_ratio_policy = "stretch", torch.onnx.mode = "linear", torch.onnx.nearest_mode = "floor"} : (!torch.vtensor<[1,3,224,224],f32>, !torch.none, !torch.none, !torch.vtensor<[4],si64>) -> !torch.vtensor<[1,3,800,800],f32>
         ^

Command

 iree-compile --iree-hal-target-backends=llvm-cpu --iree-llvmcpu-target-cpu=host -o test.vmfb model_torch_onnx.mlir 

Models Impacted:

fcos_resnet50_fpn_Opset18_torchvision
fasterrcnn_mobilenet_v3_large_320_fpn_Opset18_torchvision
fasterrcnn_mobilenet_v3_large_fpn_Opset18_torchvision
fasterrcnn_resnet50_fpn_v2_Opset18_torchvision
maskrcnn_resnet50_fpn_v2_Opset18_torchvision
retinanet_resnet50_fpn_v2_Opset18_torchvision

[zjgarvey]

The keep aspect ratio policy attribute is mis-handled. The value "stretch" is default, so we shouldn't be erroring out.

The code in https://github.com/llvm/torch-mlir/blob/a265d283357f572c5437f9217ea85fa9770d374c/lib/Conversion/TorchOnnxToTorch/DefaultDomainQtoZ.cpp#L2714-L2719
should be merged into the subsequent if statement as:

std::string keepAspectRatioPolicy;

if (... ||
     ... ||
     binder.customOpNameStringAttr(keepAspectRatioPolicy, "keep_aspect_ratio_policy", "stretch") ||
     ...)
         return failure();

// near other similar checks:

if (keepAspectRatioPolicy != "stretch")
     return rewriter.notifyMatchFailure(binder.op, "unimplemented: non-default keep_aspect_ratio_policy attribute");

[pravg-amd]

Review llvm/torch-mlir#4044