func.func @main(%arg0: tensor<?x19x19x3xf32>) -> tensor<?x9x9x4xf32> {
  %cst = arith.constant dense<1.0> : tensor<4xf32>
  %cst_3 = arith.constant dense<2.0> : tensor<4x3x3x3xf32>

    else_branch = @_functionalize_if_else_branch_00, is_stateless = false,
    then_branch = @_functionalize_if_then_branch_00} :
      (tensor<i1>, tensor<i1>, tensor<3x15x14x3xf32>, tensor<3x15x14x3xf32>) -> tensor<i1>

        ":test_utilities",
    ],
    driver = "@llvm-project//mlir:run_lit.sh",
    size_override = {
        "quant_stats.pbtxt": "medium",
    },
    tags_override = {
        "add.pbtxt": ["no_rocm"],
        "conv_2d.pbtxt": ["no_rocm"],
        "fake_quant_per_channel.pbtxt": ["no_rocm"],
    },
    test_file_exts = [
        "pbtxt",
    ],
)

# Bundle together all of the test utilities that are used by tests.
filegroup(

    }
    func.func private @composite_fn_1(%arg0: tensor<1x4x4x3xf32>, %arg1: tensor<3x1x1x3xf32>, %arg2: tensor<3xf32>) -> tensor<1x4x4x3xf32> attributes {tf_quant.composite_function} {

         tensor<64x3x3x3xf32>
    %conv = "tfl.conv_2d"(%input_act, %w, %bias)

    but if it is supported, it will be rewritten as:

    %q_w = "tfl.pseudo_qconst"() {
         qtype = tensor<64x3x3x3x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>
    %conv = "tfl.conv_2d"(%input_act, %q_w, %bias)

    Note that this is part of reaching feature parity with the old quantizer for

^bb0(%arg0: tensor<256x32x32x3xf32>, %arg1: tensor<16x3x3x3xf32>, %arg2: tensor<16xf32>):
  // CHECK: _arithmetic_count = 230686720 : i64

  %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;

  }
}
node {
  name: "conv_net_2d/conv_2d_0/w/read"
  op: "Identity"
  input: "conv_net_2d/conv_2d_0/w"
  attr {
    key: "T"
    value {
      type: DT_FLOAT
    }
  }
  attr {
    key: "_class"
    value {
      list {
        s: "loc:@conv_net_2d/conv_2d_0/w"
      }
    }
  }
}
node {
  name: "conv_net_2d_1/conv_2d_0/convolution"
  op: "Conv2D"
  input: "input"