[Onnx Zoo Models] expected offsets to be non-negative, but got -1 in TensorPadToTensorInsertSlicePass

[pravg-amd]

What happened?

For the below IR,

module {
  func.func @nasnetalarge_Opset16_timm(%arg0: !torch.vtensor<[?,168,?,?],f32>) ->  !torch.vtensor<[?,?,?,?],f32> attributes {torch.onnx_meta.ir_version = 8 : si64, torch.onnx_meta.opset_version = 21 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "2.1.0"} {
    %1689 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__362> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
    %1690 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__363> : tensor<4xsi64>} : () -> !torch.vtensor<[4],si64>
    %1691 = torch.operator "onnx.ConstantOfShape"(%1689) {torch.onnx.value = dense_resource<__364> : tensor<1xsi64>} : (!torch.vtensor<[1],si64>) -> !torch.vtensor<[4],si64>
    %1692 = torch.operator "onnx.Concat"(%1690, %1691) {torch.onnx.axis = 0 : si64} : (!torch.vtensor<[4],si64>, !torch.vtensor<[4],si64>) -> !torch.vtensor<[8],si64>
    %1693 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__365> : tensor<2xsi64>} : () -> !torch.vtensor<[2],si64>
    %1694 = torch.operator "onnx.Reshape"(%1692, %1693) {torch.onnx.allowzero = 0 : si64} : (!torch.vtensor<[8],si64>, !torch.vtensor<[2],si64>) -> !torch.vtensor<[4,2],si64>
    %1695 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__366> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
    %1696 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__367> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
    %1697 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__368> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
    %1698 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__369> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
    %1699 = torch.operator "onnx.Slice"(%1694, %1696, %1697, %1695, %1698) : (!torch.vtensor<[4,2],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[4,2],si64>
    %1700 = torch.operator "onnx.Transpose"(%1699) {torch.onnx.perm = [1 : si64, 0 : si64]} : (!torch.vtensor<[4,2],si64>) -> !torch.vtensor<[2,4],si64>
    %1701 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__370> : tensor<1xsi64>} : () -> !torch.vtensor<[1],si64>
    %1702 = torch.operator "onnx.Reshape"(%1700, %1701) {torch.onnx.allowzero = 0 : si64} : (!torch.vtensor<[2,4],si64>, !torch.vtensor<[1],si64>) -> !torch.vtensor<[8],si64>
    %1703 = torch.operator "onnx.Cast"(%1702) {torch.onnx.to = 7 : si64} : (!torch.vtensor<[8],si64>) -> !torch.vtensor<[8],si64>
    %cst2 = torch.operator "onnx.Constant"() {torch.onnx.value = dense_resource<__371> : tensor<f32>} : () -> !torch.vtensor<[],f32>
    %0 = torch.operator "onnx.Pad"(%arg0, %1703, %cst2) {torch.onnx.mode = "constant"} : (!torch.vtensor<[?,168,?,?],f32>, !torch.vtensor<[8],si64>, !torch.vtensor<[],f32>) -> !torch.vtensor<[?,?,?,?],f32>
    return %0 : !torch.vtensor<[?,?,?,?],f32>
  }
}

{-#
  dialect_resources: {
    builtin: {
      __362: "0x080000000400000000000000",
      __363: "0x08000000FFFFFFFFFFFFFFFF0100000000000000FFFFFFFFFFFFFFFF0100000000000000",
      __364: "0x080000000000000000000000",
      __365: "0x08000000FFFFFFFFFFFFFFFF0200000000000000",
      __366: "0x080000000000000000000000",
      __367: "0x08000000FFFFFFFFFFFFFFFF",
      __368: "0x080000000100000000000080",
      __369: "0x08000000FFFFFFFFFFFFFFFF",
      __370: "0x08000000FFFFFFFFFFFFFFFF",
      __371: "0x0800000000000000",
      __372: "0x080000000200000000000000"
    }
  }
#-}

Getting the error in TensorPadToTensorInsertSlicePass as

test.mlir:19:10: error: expected offsets to be non-negative, but got -1
    %0 = torch.operator "onnx.Pad"(%arg0, %1703, %cst2) {torch.onnx.mode = "constant"} : (!torch.vtensor<[?,168,?,?],f32>, !torch.vtensor<[8],si64>, !torch.vtensor<[],f32>) -> !torch.vtensor<[?,?,?,?],f32>
         ^
test.mlir:19:10: note: see current operation: %16 = "tensor.insert_slice"(%7, %12, %13, %14, %15) <{operandSegmentSizes = array<i32: 1, 1, 0, 3, 0>, static_offsets = array<i64: 0, 0, -1, -1>, static_sizes = array<i64: -9223372036854775808, 168, -9223372036854775808, -9223372036854775808>, static_strides = array<i64: 1, 1, 1, 1>}> : (tensor<?x168x?x?xf32>, tensor<?x168x?x?xf32>, index, index, index) -> tensor<?x168x?x?xf32>

Steps to reproduce your issue

Command:

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

What component(s) does this issue relate to?

Compiler

Version information

IREE commit : 5f7b471

Additional context

Model : Onnx Zoo Models (nasnetalarge_Opset16_timm / pnasnet5large_Opset16_timm)

Dump IR using : '--mlir-print-ir-after-all --mlir-print-ir-before-all --mlir-disable-threading --mlir-elide-elementsattrs-if-larger=4'

IR Dump: https://gist.github.com/pravg-amd/167c7ecdac7aa2fdc3b8beaeb7152268

[hanhanW]

I think it is because you are padding negative sizes.

// -----// IR Dump Before ConvertTorchToLinalg (convert-torch-to-linalg) //----- //
func.func @nasnetalarge_Opset16_timm(%arg0: !torch.vtensor<[?,168,?,?],f32>) -> !torch.vtensor<[?,168,?,?],f32> attributes {torch.assume_strict_symbolic_shapes, torch.onnx_meta.ir_version = 8 : si64, torch.onnx_meta.opset_version = 21 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "2.1.0"} {
  %int0 = torch.constant.int 0
  %int-1 = torch.constant.int -1
  %int1 = torch.constant.int 1
  %float0.000000e00 = torch.constant.float 0.000000e+00
  %0 = torch.prim.ListConstruct %int-1, %int1, %int-1, %int1, %int0, %int0, %int0, %int0 : (!torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>
  %1 = torch.aten.constant_pad_nd %arg0, %0, %float0.000000e00 : !torch.vtensor<[?,168,?,?],f32>, !torch.list<int>, !torch.float -> !torch.vtensor<[?,168,?,?],f32>
  return %1 : !torch.vtensor<[?,168,?,?],f32>
}

// -----// IR Dump After ConvertTorchToLinalg (convert-torch-to-linalg) //----- //
func.func @nasnetalarge_Opset16_timm(%arg0: !torch.vtensor<[?,168,?,?],f32>) -> !torch.vtensor<[?,168,?,?],f32> attributes {torch.assume_strict_symbolic_shapes, torch.onnx_meta.ir_version = 8 : si64, torch.onnx_meta.opset_version = 21 : si64, torch.onnx_meta.producer_name = "pytorch", torch.onnx_meta.producer_version = "2.1.0"} {
  %0 = torch_c.to_builtin_tensor %arg0 : !torch.vtensor<[?,168,?,?],f32> -> tensor<?x168x?x?xf32>
  %int0 = torch.constant.int 0
  %int-1 = torch.constant.int -1
  %int1 = torch.constant.int 1
  %float0.000000e00 = torch.constant.float 0.000000e+00
  %1 = torch_c.to_f64 %float0.000000e00
  %2 = torch.prim.ListConstruct %int-1, %int1, %int-1, %int1, %int0, %int0, %int0, %int0 : (!torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int, !torch.int) -> !torch.list<int>
  %3 = arith.truncf %1 : f64 to f32
  %padded = tensor.pad %0 low[0, 0, -1, -1] high[0, 0, 1, 1] {
  ^bb0(%arg1: index, %arg2: index, %arg3: index, %arg4: index):
    tensor.yield %3 : f32
  } : tensor<?x168x?x?xf32> to tensor<?x168x?x?xf32>
  %cast = tensor.cast %padded : tensor<?x168x?x?xf32> to tensor<?x168x?x?xf32>
  %4 = torch_c.from_builtin_tensor %cast : tensor<?x168x?x?xf32> -> !torch.vtensor<[?,168,?,?],f32>
  return %4 : !torch.vtensor<[?,168,?,?],f32>
}