key: "MobilenetV1/MobilenetV1/Conv2d_0/BatchNorm/FusedBatchNorm@"
  value {
    file_line_cols {
      file_index: 5
      line: 362
    }
  }
}
traces {
  key: "MobilenetV1/MobilenetV1/Conv2d_0/Conv2D@"
  value {
    file_line_cols {
      file_index: 2
      line: 27
    }
    file_line_cols {
      file_index: 3
      line: 28
    }
    file_line_cols {
      file_index: 4

def DefinedByConv2D : Constraint<CPred<"llvm::isa_and_nonnull<mlir::TF::Conv2DOp>($0.getDefiningOp())">>;
// Checks if the value has only one user.
def HasOneUse : Constraint<CPred<"$0.hasOneUse()">>;

// If we see a Conv2D op followed by Mul, then multiply the filter
// with the value in Mul.
def FuseMulAndConv2D :
  Pat<(TF_MulOp:$mul (TF_Conv2DOp:$conv $input,
                          (Arith_ConstantOp:$filter F32ElementsAttr:$filter_value),

  key: "MobilenetV1/MobilenetV1/Conv2d_0/BatchNorm/FusedBatchNorm@"
  value {
    file_line_cols {
      file_index: 5
      line: 362
    }
  }
}
traces {
  key: "MobilenetV1/MobilenetV1/Conv2d_0/Conv2D@"
  value {
    file_line_cols {
      file_index: 2
      line: 27
    }
  }
}
traces {
  key: "input@"
  value {
    file_line_cols {
      file_index: 40
      line: 690

  %0 = "tf.Conv2D"(%output, %cst) <{data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1], use_cudnn_on_gpu = true}> {attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations", device...

    %3 = "tf.Cast"(%2) {Truncate = false} : (tensor<960x960x3x512xi8>) -> tensor<960x960x3x512xi32>
    %4 = "tf.Sub"(%3, %arg5) : (tensor<960x960x3x512xi32>, tensor<512xi32>) -> tensor<960x960x3x512xi32>
    %5 = "tf.Conv2D"(%1, %4) {dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x2240x2240x3xi32>, tensor<960x960x3x512xi32>) -> tensor<1x2240x1120x512xi32>

package main

import (
	"./a"
	"fmt"
)

func main() {
	s := "foo"
	x := a.Conv(s)
	if x != s {
		panic(fmt.Sprintf("got %s wanted %s", x, s))
	}
	y, ok := a.Conv2(s)
	if !ok {
		panic("conversion failed")
	}
	if y != s {
		panic(fmt.Sprintf("got %s wanted %s", y, s))
	}
	z := a.Conv3(s)
	if z != s {
		panic(fmt.Sprintf("got %s wanted %s", z, s))
	}

def MultiplyFakeQuantValue : NativeCodeCall<
  "MultiplyFakeQuantValue($_builder, $_loc, $0...)">;

// Convert AddV2Op following an AffineOp to BiasAddOp.
// For Conv3D, even though the Conv3D op has "NDHWC" data format, the BiasAdd
// will still has the data format of "NHWC".
def ConvertAddToBiasAdd : Pat<
  (TF_AddV2Op
    (SupportedAffineOpMatcher $conv_out, $input, $weight),

  %dq_bias = "quantfork.dcast"(%q_bias) : (tensor<2x!quant.uniform<i32:f32, 0.044022349891595126>>) -> tensor<2xf32>

    derived_attrs=['T: {float, int8}'],
    outputs=['o: T'])
def _composite_conv_add_relu(input_, filter_, bias, stride_w, stride_h,
                             dilation_w, dilation_h, padding, act):
  res = tf.raw_ops.Conv2D(
      input=input_,
      filter=filter_,
      strides=[1, stride_w, stride_h, 1],
      dilations=[1, dilation_w, dilation_h, 1],
      padding=padding)
  res = tf.raw_ops.Add(x=res, y=bias)
  if act == 'RELU':

      %outputs_6, %control_7 = tf_executor.island wraps "tf.Const"() {device = "", value = dense<[-1, 16384]> : tensor<2xi32>} : () -> tensor<2xi32>
      %outputs_8, %control_9 = tf_executor.island wraps "tf.Conv2D"(%arg0, %outputs_0) {data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 2, 2, 1], use_cudnn_on_gpu = true} : (tensor<16x224x224x3xf32>, tensor<*xf32>) -> tensor<16x112x112x?xf32>