FunctionLibraryDefinition flib_def(OpRegistry::Global(), library);
  Output in = ops::Placeholder(root, DT_FLOAT);
  Output perm = ops::Const(root, {3, 1, 2, 0});

  NameAttrList b_name_attr;
  b_name_attr.set_name("TransposeCall");
  ops::PartitionedCall call(root.WithOpName("call"), {in, perm}, {DT_FLOAT},
                            b_name_attr);
  call.output.front().node()->AddAttr(kXlaMustCompileAttr, true);

  %cst_2 = arith.constant dense<1> : tensor<i32>
  %0 = "tf.AddV2"(%arg2, %cst_2) {T = i32, device = ""} : (tensor<*xi32>, tensor<i32>) -> tensor<*xi32>
  %1 = "tf.Transpose"(%arg3, %cst) {T = f32, Tperm = i32, device = ""} : (tensor<*xf32>, tensor<2xi32>) -> tensor<?x?xf32>
  %2 = "tf.Rank"(%arg3) : (tensor<*xf32>) -> tensor<i32>
  %3 = "tf.Range"(%2, %cst_0, %cst_1) : (tensor<i32>, tensor<i32>, tensor<i32>) -> tensor<?xi32>

  // expected-error @+1 {{expected perm to be a 1-D Tensor, got perm of rank 2}}
  %0 = "tf.Transpose"(%arg0, %arg1) {T = "tfdtype$DT_FLOAT", Tperm = "tfdtype$DT_INT32"} : (tensor<2x3xf32>, tensor<1x2xi32>) -> tensor<3x2xf32>
  func.return %0 : tensor<3x2xf32>
}

// -----

// Test tf.Transpose with invalid size of perm
func.func @testTranspose(tensor<2x3xf32>) -> tensor<3x2xf32> {
^bb0(%arg0: tensor<2x3xf32>):

	ADD    $0x40, INP, INP
#ifdef GOARCH_ppc64le
	VPERM  IN0, IN0, LEMASK, IN0
	VPERM  IN1, IN1, LEMASK, IN1
	VPERM  IN2, IN2, LEMASK, IN2
	VPERM  IN3, IN3, LEMASK, IN3
#endif

	VXOR IN0, XL, XH

	VPMSUMD IN1, H3L, XL1
	VPMSUMD IN1, H3, XM1
	VPMSUMD IN1, H3H, XH1

	VPERM   H2, H, HIPERM, H21L
	VPERM   IN2, IN3, LOPERM, T0
	VPERM   H2, H, LOPERM, H21H
	VPERM   IN2, IN3, HIPERM, T1

      type.getShape(), isInvert ? invertPermutationVector(perm) : perm);
  return RankedTensorType::get(permutedShape, type.getElementType());
}

static RankedTensorType GetInvertPermutedTensorType(RankedTensorType type,
                                                    ArrayRef<int64_t> perm) {
  return GetPermutedTensorTypeHelper(type, perm, true /*isInvert*/);
}

static Value CreateTranspose(OpBuilder& builder, Value source,

  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);
  }

  Output in_place_add =
      ops::InplaceAdd(on_cpu.WithOpName("tr0"), perm,

// CHECK-LABEL: pushTposeAfterAddSimple
func.func @pushTposeAfterAddSimple(%arg0: tensor<2x3x4x5xf32>) -> tensor<5x2x3x4xf32> {
  %perm = arith.constant dense<[3, 0, 1, 2]> : tensor<4xi32>
  %0 = "tfl.transpose"(%arg0, %perm) : (tensor<2x3x4x5xf32>, tensor<4xi32>) -> tensor<5x2x3x4xf32>
  %cst = arith.constant dense<1.0> : tensor<5x2x3x4xf32>

	[]int{4, 2, 1, 3, 0},                // Perm(5)
	[]int{0, 2, 3, 1, 5, 4, 6, 7},       // Perm(8)
	[]int{2, 0, 8, 3, 4, 7, 6, 5, 1},    // Perm(9)
	[]int{0, 6, 5, 3, 8, 4, 1, 2, 9, 7}, // Perm(10)
	[]int{9, 14, 4, 11, 13, 8, 0, 6, 2, 12, 3, 7, 1, 10, 5, 15}, // Perm(16)
	[]int{},                             // Perm(0)
	[]int{0},                            // Perm(1)
	[]int{2, 4, 0, 3, 1},                // Perm(5)

	show("Int32N(10)", r.Int32N(10), r.Int32N(10), r.Int32N(10))
	show("Int64N(10)", r.Int64N(10), r.Int64N(10), r.Int64N(10))

	// Perm generates a random permutation of the numbers [0, n).
	show("Perm", r.Perm(5), r.Perm(5), r.Perm(5))
	// Output:
	// Float32     0.95955694          0.8076733            0.8135684
	// Float64     0.4297927436037299  0.797802349388613    0.3883664855410056

	VADDUDM	V16, V6, V6
	VADDUDM	V17, V7, V7

	CMPU	INP, END
	BLT	loop

#ifdef GOARCH_ppc64le
	VPERM	V0, V1, KI, V0
	VPERM	V2, V3, KI, V2
	VPERM	V4, V5, KI, V4
	VPERM	V6, V7, KI, V6
#else
	VPERM	V1, V0, KI, V0
	VPERM	V3, V2, KI, V2
	VPERM	V5, V4, KI, V4
	VPERM	V7, V6, KI, V6
#endif
	STXVD2X	VS32, (CTX+R_x000)	// v0 = vs32
	STXVD2X	VS34, (CTX+R_x010)	// v2 = vs34