%1 = "quantfork.stats"(%arg0) {layerStats = dense<[1.27501142, 149.824783]> : tensor<2xf32>} : (tensor<1x3x4x3xf32>) -> tensor<1x3x4x3xf32>

  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

  %2 = "tfl.pseudo_qconst"() {qtype = tensor<32x!quant.uniform<i32:f32, 1.7052092479439231E-4>>, value = dense<0> : tensor<32xi32>} : () -> tensor<32x!quant.uniform<i32:f32, 1.7052092479439231E-4>>

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

  %4 = "tfl.dequantize"(%3) : (tensor<32x3x3x3x!quant.uniform<u8<1:255>:f32, 0.1>>) -> tensor<32x3x3x3xf32>

// TF-DAG: "tf.DumpTensor"(%[[conv0_float]]) <{enabled = true, file_name = "unquantized_tensor_data.pb", func_name = "conv_with_dump", log_dir_path = "/tmp/dumps/composite_conv2d_with_bias_and_relu6_fn_2", node_name = "Conv2D"}> {device = ""}

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),

In this pass, every op will be targeted with the user specified targets based on
the device capabilites. For example, If the user specified the desired targets
are "GPU", "CPU", `conv2d` can run on both "GPU" and "CPU", we will annotate
the op `conv2d` with "GPU" since it's preferred; `pack` can only run on "CPU",
so we will annotate the op with "CPU" since "GPU" does not support this op.

#### Raise Target Subgraphs Pass

    func.return %arg0 : tensor<*xi32>
  }

  // Test conv2d inferReturnTypes can infer some information when input or
  // filter does not have fully static shape.

  // CHECK-LABEL: func @conv2d_unranked_input_and_filter
  func.func @conv2d_unranked_input_and_filter(%arg0: tensor<*xf32>, %arg1: tensor<*xf32>) -> tensor<*xf32> {
    // CHECK: "tf.Conv2D"
    // CHECK-SAME: -> tensor<?x?x?x?xf32>

                           op.input(2), strides, padding, filter_attrs));

  Conv2D::Attrs conv_attrs;
  conv_attrs.use_cudnn_on_gpu_ = use_cudnn_on_gpu;
  conv_attrs.explicit_paddings_ = explicit_paddings;
  conv_attrs.data_format_ = data_format;
  conv_attrs.dilations_ = dilations;
  grad_outputs->push_back(
      Conv2D(scope, grad_inputs[0], op.input(1), strides, padding, conv_attrs));
  return scope.status();
}