module {
  // For XLA weight-only per-channel depthwise convolution, tensor shape should have
  // transformed from [H,W,C,M] to [H,W,1,CxM],
  func.func @depthwise_conv(%arg0: tensor<1x3x4x3xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<6xf32>} : () -> tensor<6xf32>

// RUN: tf-quant-opt %s -split-input-file -quant-preprocess-op | FileCheck %s

module {
  // For UniformQuantized depthwise convolution, tensor shape should have
  // transformed from [H,W,C,M] to [H,W,1,CxM],
  func.func @depthwise_conv(%arg0: tensor<1x3x4x3xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<6xf32>} : () -> tensor<6xf32>

// PerChannel: return %[[out_1]], %[[out_2]]

}

// -----

module {
  func.func @depthwise_conv(%arg0: tensor<1x3x4x3xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<3xf32>} : () -> tensor<3xf32>

// CHECK: return %3 : tensor<*xf32>

// CHECK-LABEL: func private @composite_conv2d_fn_1
// CHECK: %0 = "tf.Conv2D"(%arg0, %arg1)
// CHECK: return %0 : tensor<*xf32>
}

// -----

module {
  func.func @depthwise_conv(%arg0: tensor<1x3x4x512xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>

// CHECK: return %3 : tensor<*xf32>

// CHECK-LABEL: func private @composite_conv2d_fn_1
// CHECK: %0 = "tf.Conv2D"(%arg0, %arg1)
// CHECK: return %0 : tensor<*xf32>
}

// -----

module {
  func.func @depthwise_conv(%arg0: tensor<1x3x4x512xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>

module attributes {tf.versions = {bad_consumers = [], min_consumer = 12 : i32, producer = 1269 : i32}, tf_saved_model.semantics} {

// CHECK-DAG: %[[cst:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<2x3x3x2xf32>}> : () -> tensor<2x3x3x2xf32>
// CHECK: %[[depthwise_conv:.*]] = "tf.DepthwiseConv2dNative"(%arg0, %[[cst]])
// CHECK: %[[identity:.*]] = "tf.IdentityN"(%[[depthwise_conv]]) {device = ""} : (tensor<1x2x2x6xf32>) -> tensor<1x2x2x6xf32>
// CHECK: return %[[identity]] : tensor<1x2x2x6xf32>

      if (!input_type || !input_type.hasStaticShape()) {
        return false;
      }
      total_count += input_type.getNumElements();
    }
    *count = total_count;
    return true;
  }

  // For conv2d/depthwise_conv/fully_connected ops.
  // This algorithm actually comes from TOCO tooling_util.cc
  static bool GetArithmeticCountForConvAndFullyconnectedOp(mlir::Operation* op,

// CHECK-LABEL: func private @quantized_conv2d_fn_1
// CHECK:      %[[CONV2D_0:.*]] = "tf.UniformQuantizedConvolutionHybrid"
// CHECK-SAME: padding = "SAME"

}

// -----

module {
  func.func @depthwise_conv(%arg0: tensor<1x3x4x3xf32>) -> (tensor<*xf32>, tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<3xf32>} : () -> tensor<3xf32>

// CHECK: %[[DEPTHWISE_CONV2D:.*]] = "tf.DepthwiseConv2dNative"(%arg0, %[[DEQUANTIZED]]) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1]}> {attr_map = "0:strides,1:padding,2:explicit_paddings,3:dilations", device = ""} : (tensor<1x3x4x512xf32>, tensor<2x3x3x512xf32>) -> tensor<*xf32>
// CHECK: return %[[DEPTHWISE_CONV2D]] : tensor<*xf32>