// CHECK-NEXT:    version: 1
  // CHECK-NEXT:    builtin_code: DEQUANTIZE
  // CHECK-NEXT:  }, {
  // CHECK-NEXT:    deprecated_builtin_code: 4,
  // CHECK-NEXT:    version: 1
  // CHECK-NEXT:    builtin_code: DEPTHWISE_CONV_2D
  // CHECK-NEXT:  } ],
  // CHECK-NEXT:  subgraphs: [ {
  // CHECK-NEXT:    tensors: [ {
  // CHECK-NEXT:      shape: [ 1, 224, 224, 3 ],
  // CHECK-NEXT:      buffer: 1,
  // CHECK-NEXT:      name: "arg0",

  // CHECK-NEXT:    version: 1,
  // CHECK-NEXT:    builtin_code: DEQUANTIZE
  // CHECK-NEXT:  }, {
  // CHECK-NEXT:    deprecated_builtin_code: 4,
  // CHECK-NEXT:    version: 2,
  // CHECK-NEXT:    builtin_code: DEPTHWISE_CONV_2D
  // CHECK-NEXT:  } ],
  // CHECK-NEXT:  subgraphs: [ {
  // CHECK-NEXT:    tensors: [ {
  // CHECK-NEXT:      shape: [ 1, 224, 224, 3 ],
  // CHECK-NEXT:      buffer: 1,
  // CHECK-NEXT:      name: "arg0",

func.func @QuantizeDepthwiseConv2D(%arg0: tensor<1x224x224x3xf32>) -> tensor<1x112x112x64xf32> {
  %w = arith.constant dense<127.0> : tensor<64x3x3x3xf32>
  %b = arith.constant dense<0.0> : tensor<64xf32>

  %w = arith.constant dense<127.0> : tensor<64x3x3x3xf32>
  %b = arith.constant dense<0.0> : tensor<64xf32>

  func.return %0 : tensor<256x30x30x12xbf16>
  // CHECK: "tfl.depthwise_conv_2d"
}

// CHECK-LABEL: conv_2d_f16
func.func @conv_2d_f16(%arg0 : tensor<256x32x32x3xf16>, %arg1 : tensor<3x3x3x16xf16>) -> tensor<256x8x7x16xf16> {

// CHECK: %[[conv:.*]] = "tfl.depthwise_conv_2d"(%arg0, %[[dq]]

// PerTensor: %[[cst:.*]] = arith.constant dense<1.270000e+02> : tensor<32x3x3x3xf32>
// PerTensor: %[[q:.*]] = "tfl.quantize"(%[[cst]]) <{qtype = tensor<32x3x3x3x!quant.uniform<i8<-127:127>:f32,
// PerTensor: %[[dq:.*]] = "tfl.dequantize"(%[[q]])
// PerTensor: %[[conv:.*]] = "tfl.depthwise_conv_2d"(%arg0, %[[dq]]
}

  // CHECK-DAG: %[[VAL3:.+]] = "tfl.dequantize"(%[[VAL1]]) : (tensor<32x!quant.uniform<{{.+}})
  // CHECK-DAG: %[[VAL4:.+]] = "tfl.depthwise_conv_2d"(%arg0, %[[VAL2]], %[[VAL3]]) <{depth_multiplier = 4 : i32, dilation_h_factor = 1 : i32, dilation_w_factor = 1 : i32, fused_activation_function = "NONE", padding = "VALID", stride_h = 4 : i32, stride_w = 5 : i32}>

// CHECK: %[[QUANTIZE:.*]] = "tfl.quantize"(%[[CONSTANT0]]) <{qtype = tensor<1x3x3x48x!quant.uniform<u8:f32, 1.000000e+00>>}>
// CHECK: %[[DEQUANTIZE:.*]] = "tfl.dequantize"(%[[QUANTIZE]])
// CHECK: %[[CONV:.*]] = "tfl.depthwise_conv_2d"(%arg0, %[[DEQUANTIZE]], %[[CONSTANT]])
// CHECK: return %[[CONV]]
}

// CHECK-LABEL: perChannelFakeQuantWithDepthwiseConv2D

  TargetHardwareOpRegistration<CpuHardware, Op> Op##_CpuHardware_hardware( \
      Create);

// Operation costs on CPU

// Currently used for these ops:
// tfl.conv_2d / tfl.depthwise_conv_2d / tfl.fully_connected
class CpuConvOp : public TargetHardwareOperation {
  double GetOpCost(mlir::Operation* op) const override {
    float cost = 0.0;
    int64_t arithmetic_count;

// -----

func.func @testDepthwiseConv(%arg0: tensor<1x112x112x32xf32>, %arg1: tensor<1x3x3x32xf32>, %arg2: tensor<32xf32>) -> tensor<1x112x112x32xf32> {
  // CHECK: tac.device = "GPU", tac.inference_type = "FLOAT"