// CHECK-SAME: src_format = "NHWC"

  // CHECK: %[[ARG_PERM:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi64>}>
  // CHECK: %[[IN_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%arg0, %[[ARG_PERM]])
  // CHECK: %[[OUT_BP_TRANSPOSE:[0-9]*]] = "tf.Transpose"(%arg2, %[[ARG_PERM]])

  // CHECK: %[[CONV2D_BACKPROP:[0-9]*]] = "tf.Conv2DBackpropFilter"
  // CHECK-SAME: (%[[IN_TRANSPOSE]], %[[FILTER_PERM]], %[[OUT_BP_TRANSPOSE]])

  // CHECK: %[[CST:.*]] = "tf.Const"() <{value = dense<[0, 2, 3, 1]> : tensor<4xi64>}>
  // CHECK: %[[R0:.*]] = "tf.Transpose"(%[[ARG0]], %[[CST]])
  // CHECK: %[[R1:.*]] = "tf.BiasAdd"(%[[R0]], %[[ARG1]]) <{data_format = "NHWC"}> {device = ""}
  // CHECK: %[[CST_0:.*]] = "tf.Const"() <{value = dense<[0, 3, 1, 2]> : tensor<4xi64>}>
  // CHECK: "tf.Transpose"(%[[R1]], %[[CST_0]])

  // CHECK-DAG: %[[cst_0:.*]] = arith.constant dense<[0, 2, 1]> : tensor<3xi32>
  // CHECK: %[[v0:.*]] = "tf.Transpose"(%arg0, %[[cst]]) : (tensor<2x5x7xf32>, tensor<3xi32>) -> tensor<5x7x2xf32>
  // CHECK: %[[v1:.*]] = "tf.Transpose"(%arg1, %[[cst_0]]) : (tensor<5x3x2xf32>, tensor<3xi32>) -> tensor<5x2x3xf32>

    // results for depthwise transpose convolutions with non-1 channel
    // multiplier.
    if ((kernel_output_channels / feature_group_count) != 1) {
      return rewriter.notifyMatchFailure(
          conv_op,
          "Unsupported detphwise transpose convolution with non-1 channel "
          "multiplier");
    }

    // Slicing with dynamic offsets (helper method advised)

  // Convert result back: NHWC -> NCHW
  %3 = "tf.Const"() {value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %4 = "tf.Transpose"(%2, %3) : (tensor<1x32x32x8xf32>, tensor<4xi32>) -> tensor<1x8x32x32xf32>

  // Check that Conv2D computed in NCHW format, and all redundant transpose
  // operations removed from the function.

  // CHECK: %[[CONV:[0-9]*]] = "tf.Conv2D"(%arg0, %arg1)
  // CHECK-SAME: data_format = "NCHW"

      // out only isn't supported
      return std::nullopt;
    }
  }
  return dnums;
}

// Function to replace a unary einsum op, that can undergo simple transpose, to
// an explicit transpose op.
LogicalResult rewriteToReduceSumAndTranspose(TF::EinsumOp op,
                                             EinsumDimensionNumbers dnums,

  Output in1 = ops::Placeholder(on_gpu.WithOpName("in1"), DT_FLOAT);

  Output perm = ops::Const(on_cpu.WithOpName("perm"), {3, 1, 2, 0});

  {
    Output tr0 = ops::Transpose(on_gpu.WithOpName("tr0"), in0, perm);
    Output tr1 = ops::Transpose(on_gpu.WithOpName("tr1"), in1, perm);
  }

  std::unique_ptr<Graph> result;
  TF_ASSERT_OK(CloneConstantsForBetterClustering(root, &result));

  OutputTensor tr0_perm;

 ReverseSequence 1
 Slice 2
 Transpose 3
cluster 5 size 21
 All 1
 ConcatV2 1
 Const 11
 Equal 1
 ExpandDims 1
 ReverseSequence 1
 Shape 1
 StridedSlice 1
 Transpose 3
cluster 6 size 11
 Cast 1
 Const 5
 GatherV2 1
 Shape 1
 StridedSlice 1
 Transpose 1
 ZerosLike 1
cluster 7 size 33
 All 2
 Cast 1
 Const 17
 Equal 2
 ExpandDims 2
 GatherV2 1

  // NOFOLD: %[[TRANSPOSE:[0-9]*]] = "tf.Transpose"(%arg0, %[[CST]]) : (tensor<?x224x224x3xf32>, tensor<4xi32>) -> tensor<?x3x224x224xf32>

  // Pad input with new paddings.
  // CHECK: %[[PAD:[0-9]*]] = "tf.Pad"(%arg0, %[[PADDINGS]])
  // CHECK-SAME: (tensor<?x224x224x3xf32>, tensor<4x2xi32>) -> tensor<?x230x230x3xf32>
  // NOFOLD: %[[PAD:[0-9]*]] = "tf.Pad"(%[[TRANSPOSE]], %[[PADDING]])

// CHECK-SAME:                            %[[VAL_0:.*]]: tensor<1x3x6x6xf32>) -> tensor<*xf32> {
// CHECK:           %[[VAL_1:.*]] = arith.constant dense<[0, 2, 3, 1]> : tensor<4xi32>
// CHECK:           %[[VAL_2:.*]] = "tfl.transpose"(%[[VAL_0]], %[[VAL_1]]) : (tensor<1x3x6x6xf32>, tensor<4xi32>) -> tensor<1x6x6x3xf32>
// CHECK:           %[[VAL_3:.*]] = arith.constant dense<0> : tensor<4x2xi32>