key: "narrow_range"
    value {
      b: true
    }
  }
  attr {
    key: "num_bits"
    value {
      i: 8
    }
  }
}
node {
  name: "BoxPredictor_4/ClassPredictor/Conv2D"
  op: "Conv2D"
  input: "input"
  input: "BoxPredictor_4/ClassPredictor/weights_quant/FakeQuantWithMinMaxVarsPerChannel"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {

  // CHECK: %0 = "tfl.conv_2d"(%arg0, %arg1, %cst)
}

// CHECK-LABEL: fuse4DAddIntoConv2d
func.func @fuse4DAddIntoConv2d(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<2x3x3x3xf32>) -> tensor<256x32x32x2xf32> {
  %cst = arith.constant dense<[[[[1.0, 2.0]]]]> : tensor<1x1x1x2xf32>
  %cst_0 = arith.constant dense<[1.0, 2.0]> : tensor<2xf32>
  %0 = "tfl.conv_2d"(%arg0, %arg1, %cst_0) {
    dilation_h_factor = 1 : i32,

// CHECK: %[[QUANTIZE:.*]] = "tfl.quantize"(%[[CONSTANT0]]) <{qtype = tensor<16x3x3x3x!quant.uniform<u4:f32, 1.000000e+00>>}>
// CHECK: %[[DEQUANTIZE:.*]] = "tfl.dequantize"(%[[QUANTIZE]])
// CHECK: %[[CONV:.*]] = "tfl.conv_2d"(%arg0, %[[DEQUANTIZE]], %[[CONSTANT]])
// CHECK: return %[[CONV]]
}

// CHECK-LABEL: perChannelFakeQuantWithConv2D
func.func @perChannelFakeQuantWithConv2D(tensor<256x32x32x3xf32>) -> (tensor<256x8x7x16xf32>) {

  %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>>

  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;

      return false;
    }
    return true;
  }
};
std::unique_ptr<TargetHardwareOperation> CreateConcatOp() {
  return std::make_unique<GpuConcatOp>();
}

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

// CHECK: %[[QUANTIZE:.*]] = "tfl.quantize"(%[[CONSTANT0]]) <{qtype = tensor<16x3x3x3x!quant.uniform<u8:f32, 1.000000e+00>>}>
// CHECK: %[[DEQUANTIZE:.*]] = "tfl.dequantize"(%[[QUANTIZE]])
// CHECK: %[[CONV:.*]] = "tfl.conv_2d"(%arg0, %[[DEQUANTIZE]], %[[CONSTANT]])
// CHECK: return %[[CONV]]
}

// CHECK-LABEL: perChannelFakeQuantWithConv2D
func.func @perChannelFakeQuantWithConv2D(tensor<256x32x32x3xf32>) -> (tensor<256x8x7x16xf32>) {

  %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>>

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

    key: "narrow_range"
    value {
      b: true
    }
  }
  attr {
    key: "num_bits"
    value {
      i: 8
    }
  }
}
node {
  name: "BoxPredictor_4/ClassPredictor/Conv2D"
  op: "Conv2D"
  input: "input"
  input: "BoxPredictor_4/ClassPredictor/weights_quant/FakeQuantWithMinMaxVarsPerChannel"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {