func.return %mm_s : tensor<1x3x3x2xf32>

// CHECK: %[[w:.*]] = arith.constant dense<1.270000e+02> : tensor<512x2xf32>
// CHECK: %[[q_w:.*]] = "tfl.quantize"(%[[w]]) <{qtype = tensor<512x2x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>}>
// CHECK: %[[dq_w:.*]] = "tfl.dequantize"(%[[q_w]]) : (tensor<512x2x!quant.uniform<i8<-127:127>:f32, 1.000000e+00>>) -> tensor<512x2xf32>

        : (tensor<1x3x4x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x3x2x2xf32> loc(fused["Conv2D:", "test_opt_out"])
    %2 = "tf.BiasAdd"(%1, %cst_0) : (tensor<1x3x2x2xf32>, tensor<2xf32>) -> tensor<1x3x2x2xf32> loc(fused["BiasAdd:", "model/bias_add"])
    %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
    return %3 : tensor<1x3x2x2xf32>
  }
}

  %3 = "tf.BiasAdd"(%2, %cst_0) {data_format = "NHWC", device = ""} : (tensor<1x3x2x2xf32>, tensor<2xf32>) -> tensor<1x3x2x2xf32>
  %4 = "tf.Relu"(%3) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  %5 = "quantfork.qcast"(%4) : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0027450981093387976:-19>>

    return %2 : tensor<1x3x4x2xf32>
  }

  func.func private @composite_conv_fn(%arg0: tensor<1x3x4x3xf32>, %arg1: tensor<2x3x3x2xf32>) -> tensor<1x3x4x2xf32> attributes {_from_xla_call_module} {

  func.return %0#0, %1#0 : tensor<8x8x8x8xf32>, tensor<8x8x8x8xf32>
}
// CHECK: func @customAggregator
// CHECK-NEXT: %[[stats:.*]] = "quantfork.stats"(%arg0) <{layerStats = dense<[-1.000000e-01, 2.000000e-01]> : tensor<2xf32>}> : (tensor<8x8x8x8xf32>) -> tensor<8x8x8x8xf32>
// CHECK-NEXT: return %[[stats]], %arg0

  %2 = "tf.Cast"(%1) {Truncate = false} : (tensor<1x3x2x2xbf16>) -> tensor<1x3x2x2xf32>
  %3 = "tf.IdentityN"(%2) {device = ""} : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2xf32>
  return %3 : tensor<1x3x2x2xf32>
// CHECK: tf.Conv2D
// CHECK-SAME: loc("Model/conv2d")
}

func.func @conv2d_with_valid_loc(%arg0: tensor<1x3x4x3xf32>) -> (tensor<1x3x2x2xf32>) {

  %0 = stablehlo.constant() {value = dense<3> : tensor<3x3x4x2xi8>} : () -> tensor<3x3x4x2x!quant.uniform<i8:f32, 3.000000e-01:-5>>
  %1 = stablehlo.convolution(%arg0, %0) dim_numbers = [b, 0, 1, f]x[0, 1, i, o]->[b, 0, 1, f], window = {pad = [[1, 1], [1, 1]]} {batch_group_count = 1 : i64, feature_group_count = 1 : i64} : (tensor<1x3x3x4xf32>, tensor<3x3x4x2x!quant.uniform<i8:f32, 3.000000e-01:-5>>) -> tensor<1x3x3x2xf32>

  %4 = stablehlo.broadcast_in_dim %1, dims = [3] : (tensor<4xf32>) -> tensor<1x3x3x4xf32>
  %5 = stablehlo.add %3, %4 : tensor<1x3x3x4xf32>
  %6 = stablehlo.maximum %5, %2 : tensor<1x3x3x4xf32>
  func.return %6: tensor<1x3x3x4xf32>
}
// CHECK: %[[CONST_0:.*]] = stablehlo.constant dense<2.000000e+00>

  %9 = "tfl.quantize"(%8) {qtype = tensor<1x3x1x1x!quant.uniform<i8:f32, 1.0:2>>, volatile} : (tensor<1x3x1x1xf32>) -> tensor<1x3x1x1x!quant.uniform<i8:f32, 1.0:2>>
  %10 = "tfl.dequantize"(%9) : (tensor<1x3x1x1x!quant.uniform<i8:f32, 1.0:2>>) -> tensor<1x3x1x1xf32>

  %1 = "tf.AddV2"(%0, %0) : (tensor<1x4x4x8xf32>, tensor<1x4x4x8xf32>) -> tensor<1x4x4x8xf32>
  %2 = "tf.Const"() {value = dense<[0, 3, 1, 2]> : tensor<4xi32>} : () -> tensor<4xi32>
  %3 = "tf.Transpose"(%1, %2) : (tensor<1x4x4x8xf32>, tensor<4xi32>) -> tensor<1x8x4x4xf32>

  func.return %3 : tensor<1x8x4x4xf32>
}

// CHECK-LABEL: move_transpose_handle_broadcast