// CHECK: %0 = tfl.add(%arg0, %cst_0) <{fused_activation_function = "NONE"}> : (tensor<2x3x4x5xf32>, tensor<2x3x4x1xf32>) -> tensor<2x3x4x5xf32>
// CHECK: %1 = "tfl.transpose"(%0, %cst) : (tensor<2x3x4x5xf32>, tensor<4xi32>) -> tensor<5x2x3x4xf32>

// -----

// CHECK-LABEL: pushTposeBcastScalarCstInput
func.func @pushTposeBcastScalarCstInput(%arg0: tensor<2x3x4x5xf32>) -> tensor<5x2x3x4xf32> {

// -----

// CHECK-LABEL: transpose_simple_4d
func.func @transpose_simple_4d() -> tensor<5x2x3x4xf32> {
  %0 = stablehlo.constant dense<1.000000e+0> : tensor<2x3x4x5xf32>
  %1 = stablehlo.transpose %0, dims = [3, 0, 1, 2] : (tensor<2x3x4x5xf32>) -> tensor<5x2x3x4xf32>
  return %1 : tensor<5x2x3x4xf32>
}
// CHECK-DAG: %[[CONST_0:.+]] = stablehlo.constant dense<1.000000e+00> : tensor<5x2x3x4xf32>
// CHECK-NOT: transpose

    return %1 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {

  func.return %dequant : tensor<1x3x4x2xf32>
}

// -----

// CHECK-LABEL: @dont_merge_quantization_followed_by_quantization
// CHECK-SAME: %[[ARG_0:.*]]: tensor<1x3x4x3xf32>
func.func @dont_merge_quantization_followed_by_quantization(%arg0: tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32> {
  // CHECK: %[[QUANT_ARG_0:.*]] = stablehlo.uniform_quantize %[[ARG_0]]

    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<*xf32>

    %0 = "tf.FakeQuantWithMinMaxArgs"(%arg0) {device = "", max = 2.000000e-01 : f32, min = -1.000000e-01 : f32, narrow_range = false, num_bits = 8 : i64} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>

// RUN: tf-quant-opt %s -quant-cast-bf16-ops-to-f32 | FileCheck %s

func.func @cast_bf16_conv_to_fp32(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x3x2x2xf32>) {
  %cst = "tf.Const"() {device = "", value = dense_resource<__elided__> : tensor<2x3x3x2xbf16>} : () -> tensor<2x3x3x2xbf16>
  %0 = "tf.Cast"(%arg0) {Truncate = false, device = ""} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xbf16>

    return %2 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {
    %0 = stablehlo.convolution(%arg0, %arg1) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = [[0, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32>

func.func private @conv(%input: tensor<1x3x4x3xf32> {tf._user_specified_name = "input_tensor"}) -> tensor<*xf32> attributes {tf._construction_context = "kEagerRuntime", tf._input_shapes = [#tf_type.shape<1x3x4x3>]} {
  %weight = arith.constant dense_resource<__elided__> : tensor<2x3x3x2xf32>
  %bias = arith.constant dense<[7.11401462, 7.05456924]> : tensor<2xf32>