%cst = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
  %0 = "tf.FakeQuantWithMinMaxArgs"(%arg1) {device = "", max = 2.000000e+00 : f32, min = -1.000000e+00 : f32, narrow_range = false, num_bits = 8 : i64} : (tensor<2x3x3x2xf32>) -> tensor<*xf32>

    %1 = stablehlo.constant dense<1.000000e+03> : tensor<1x1x1x1xf32>  // Input inverse scale.
    %2 = stablehlo.constant dense<-128> : tensor<1x1x1x1xi8>  // Input zero point.
    %3 = stablehlo.constant dense<2.000000e+01> : tensor<3x3x4x4xf32>  // Quantized filter tensor.
    %4 = stablehlo.constant dense<3.000000e+03> : tensor<1x1x1x4xf32>
    %5 = stablehlo.constant dense<4.000000e+03> : tensor<1x1x1x1xf32>  // Output inverse scale.

// CHECK: return %0 : tensor<*xf32>
}

// -----

module {
  func.func @conv2d(%arg0: tensor<1x3x4x3xf32>) -> (tensor<*xf32>) {
    %cst_0 = "tf.Const"() {value = dense<0.000000e+00> : tensor<2xf32>} : () -> tensor<2xf32>
    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x512xf32>} : () -> tensor<2x3x3x512xf32>

// CHECK: %[[BITCAST_CONVERT_0:.*]] = stablehlo.bitcast_convert %[[UQ]] : (tensor<1x!quant.uniform<i8:f32, 1.000000e+00:3>>) -> tensor<1xi8>
// CHECK: %[[BITCAST_CONVERT_1:.*]] = stablehlo.bitcast_convert %[[BITCAST_CONVERT_0]] : (tensor<1xi8>) -> tensor<1x!quant.uniform<i8:f32, 1.000000e+00:3>>

// CHECK: %[[q1:.*]] = "quantfork.qcast"(%[[maxpool]])
// CHECK-SAME: quant.uniform<i8:f32, 5.000000e-02:-10>
// CHECK: %[[dq1:.*]] = "quantfork.dcast"(%[[q1]])
// CHECK-SAME: quant.uniform<i8:f32, 5.000000e-02:-10>
// CHECK: %[[reshape:.*]] = "tf.Reshape"(%[[dq1]]
// CHECK: %[[q2:.*]] = "quantfork.qcast"(%[[reshape]])
// CHECK-SAME: quant.uniform<i8:f32, 5.000000e-02:-10>
// CHECK: %[[dq2:.*]] = "quantfork.dcast"(%[[q2]])

00000020  00 00 00 44 4f 57 4e 47  52 44 00 00 00 05 00 00  |...DOWNGRD......|
00000030  09 ff 01 00 01 00 00 17  00 00 16 03 02 02 59 0b  |..............Y.|
00000040  00 02 55 00 02 52 00 02  4f 30 82 02 4b 30 82 01  |..U..R..O0..K0..|
00000050  b4 a0 03 02 01 02 02 09  00 e8 f0 9d 3f e2 5b ea  |............?.[.|
00000060  a6 30 0d 06 09 2a 86 48  86 f7 0d 01 01 0b 05 00  |.0...*.H........|
00000070  30 1f 31 0b 30 09 06 03  55 04 0a 13 02 47 6f 31  |0.1.0...U....Go1|

  func.return %0 : tensor<384x128x!quant.uniform<i8:f32, 0.09:-4>>
}

// PARTIAL:   func @quantization_test(%[[VAL_0:.*]]: tensor<384x512x!quant.uniform<i8:f32, 1.000000e-01>>, %[[VAL_1:.*]]: tensor<128x512x!quant.uniform<i8<-127:127>:f32, 1.000000e-02>>, %[[VAL_2:.*]]: tensor<128x!quant.uniform<i32:f32, 0.69999999999999996>>) -> tensor<384x128x!quant.uniform<i8:f32, 0.089999999999999996:-4>> {

  %cst = arith.constant dense<0.000000e+00> : tensor<4x2xf32>
  %cst_0 = arith.constant dense<0.000000e+00> : tensor<8xf32>
  %cst_1 = arith.constant dense<[1, 0, 2]> : tensor<3xi32>
  %cst_2 = arith.constant dense<0.000000e+00> : tensor<4x4x2xf32>
  %cst_3 = arith.constant dense<4> : tensor<i32>
  %cst_4 = arith.constant dense<1.000000e+00> : tensor<f32>
  %cst_5 = arith.constant dense<1> : tensor<i32>

  %0 = "tf.Const"() <{value = dense<0.5> : tensor<1024x3xf32>}> : () -> tensor<1024x3xf32>
  %1:4 = "tf.CustomAggregator"(%arg0) <{id = "1", calibration_method = 1 : i32, num_bins = 0 : i32, max_percentile = 0.000000e+00 : f32, min_percentile = 0.000000e+00 : f32}> {min = 7.547870e-07 : f32, max = 0.999992311 : f32} : (tensor<1x1024xf32>) -> (tensor<1x1024xf32>, tensor<f32>, tensor<f32>, tensor<*xi64>)

func.func @nchw_conv_with_bias_add_relu(%arg0: tensor<1x2x5x5xf32>) -> tensor<1x4x5x5xf32> {
  %0 = stablehlo.constant dense<2.000000e+00> : tensor<4x2x3x3xf32>
  %5 = stablehlo.constant dense<0.000000e+00> : tensor<1x4x5x5xf32>
  %1 = stablehlo.constant dense<3.000000e+00> : tensor<1x4x5x5xf32>