// RUN: tf-quant-opt %s -split-input-file -quant-insert-quantized-functions -quant-quantize-composite-functions | FileCheck %s

module {
  func.func @conv(%arg0: tensor<1x2x2x3xf32>) -> (tensor<*xf32>, tensor<*xf32>) {

  %3 = "tfl.quantize"(%2) {qtype = tensor<1x2x2x3x!quant.uniform<i8:f32, 0.1>>, volatile} : (tensor<1x2x2x3xf32>) -> tensor<1x2x2x3x!quant.uniform<i8:f32, 0.1>>
  %4 = "tfl.dequantize"(%3) : (tensor<1x2x2x3x!quant.uniform<i8:f32, 0.1>>) -> tensor<1x2x2x3xf32>

    %1 = "tf.PartitionedCall"(%arg0, %cst, %cst_0) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_conv2d_with_bias_and_relu6_fn_1} : (tensor<1x2x2x3xf32>, tensor<2x2x3x2xf32>, tensor<2xf32>) -> tensor<*xf32> loc(callsit...

func.func @attribute_cast(%arg0: tensor<1x4x4x1xf32>) -> tensor<1x2x2x1xf32> {
  %0 = "tfr.cast"(%arg0) : (tensor<1x4x4x1xf32>) -> !tfr.tensor
  %stride_i32 = arith.constant 2 : i32
  %1 = tfr.call @tf__my_max_pool(%0, %stride_i32, %stride_i32) : (!tfr.tensor, i32, i32) -> !tfr.tensor
  %2 = "tfr.cast"(%1) : (!tfr.tensor) -> tensor<1x2x2x1xf32>
  func.return %2 : tensor<1x2x2x1xf32>
// CHECK: tf__max_pool
}

    rhs_quantization_max_val = 127 : i64
  } : (tensor<1x2x2x3xf32>, tensor<2x3x3x2x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<1x2x1x2xf32>
  func.return %0 : tensor<1x2x1x2xf32>
}

// -----

// CHECK-LABEL: func @uniform_quantized_convolution_hybrid_explicit
func.func @uniform_quantized_convolution_hybrid_explicit(%input: tensor<1x2x2x3xf32>) -> tensor<1x3x3x2xf32> {

module {
  func.func @conv_with_single_layer(%arg0: tensor<1x2x2x3xf32>) -> (tensor<*xf32>) {

// CHECK: stablehlo.maximum
// CHECK: (tensor<1x5x5x4xf32>, tensor<f32>) -> tensor<1x2x2x4xf32>
// CHECK: %[[TRANSPOSE_1:.+]] = stablehlo.transpose %[[REDUCE_WINDOW_MAX]], dims = [0, 3, 1, 2] : (tensor<1x2x2x4xf32>) -> tensor<1x4x2x2xf32>
// CHECK: return %[[TRANSPOSE_1]]

// -----

// Tests that a `maximum(add(convolution(%activation, %weight), %bias), %zero)`

  %1 = mhlo.constant dense<0> : tensor<i32>
  %2 = "mhlo.iota"() <{iota_dimension = 0 : i64}> : () -> tensor<32xi32>
  %3 = "mhlo.reshape"(%2) : (tensor<32xi32>) -> tensor<1x32x1xi32>
  %4:2 = mhlo.reduce(%arg0 init: %0), (%3 init: %1) across dimensions = [1] : (tensor<1x32x1xf32>, tensor<1x32x1xi32>, tensor<f32>, tensor<i32>) -> (tensor<1x1xf32>, tensor<1x1xi32>)
   reducer(%arg1: tensor<f32>, %arg3: tensor<f32>) (%arg2: tensor<i32>, %arg4: tensor<i32>)  {