%1 = "tf.Sub"(%0, %cst_0) {data_format = "NHWC"} : (tensor<1x3x2x2xf32>, tensor<1x1x1x2xf32>) -> tensor<1x3x2x2xf32>
  %2 = "tf.Mul"(%1, %cst_1) : (tensor<1x3x2x2xf32>, tensor<1x1x1x2xf32>) -> tensor<1x3x2x2xf32>
  %3 = "tf.AddV2"(%2, %cst_2) : (tensor<1x3x2x2xf32>, tensor<1x1x1x2xf32>) -> tensor<1x3x2x2xf32>
  func.return %3 : tensor<1x3x2x2xf32>
}

  %27 = mhlo.subtract %26, %19 : tensor<1x64x32x32xf32>
  %28 = "mhlo.broadcast_in_dim"(%1) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<32xf32>) -> tensor<1x64x32x32xf32>
  %29 = mhlo.multiply %27, %28 : tensor<1x64x32x32xf32>
  %30 = mhlo.add %19, %29 : tensor<1x64x32x32xf32>

  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<1x3x4x3x3xf32>, tensor<2x3x3x3x2xf32>) -> tensor<1x3x2x3x2xf32>
  %7 = "tf.BiasAdd"(%6, %cst_1) {data_format = "NHWC", device = ""} : (tensor<1x3x2x3x2xf32>, tensor<2xf32>) -> tensor<1x3x2x3x2xf32>

  return %2, %5, %7 : tensor<1x3x2x3x2xf32>, tensor<1x3x2x3x2xf32>, tensor<1x3x2x3x2xf32>

// CHECK-DAG: %[[CST:.*]] = "tf.Const"() {{.*}} : () -> tensor<2x3x3x3x2xf32>

// CHECK-DAG: %[[BROADCAST:.*]] = stablehlo.broadcast_in_dim %[[CONST_0]], dims = [3] : (tensor<2xf32>) -> tensor<1x3x2x2xf32>
// CHECK-DAG: %[[ADD:.*]] = stablehlo.add %[[CONV]], %[[BROADCAST]] : tensor<1x3x2x2xf32>
// CHECK-DAG: %[[RELU6:.*]] = stablehlo.clamp %[[CONST_2]], %[[ADD]], %[[CONST_1]] : (tensor<f32>, tensor<1x3x2x2xf32>, tensor<f32>) -> tensor<1x3x2x2xf32>

    %6 = "quantfork.qcast"(%5) : (tensor<1x3x2x2xf32>) -> tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0054901962186775953:-19>>
    %7 = "quantfork.dcast"(%6) : (tensor<1x3x2x2x!quant.uniform<i8:f32, 0.0054901962186775953:-19>>) -> tensor<1x3x2x2xf32>
    return %7 : tensor<1x3x2x2xf32>
  }

    has_token_input_output = false, module = "", platforms = [],
    version = 5 : i64
  } : (tensor<1x3x2x3xf32>, tensor<2x3x3x2xf32>) -> tensor<1x2x2x2xf32>
  return %0 : tensor<1x2x2x2xf32>
}

// CHECK-LABEL: func.func @qdq_for_conv_weight_empty
// CHECK-SAME: (%[[ARG_0:.+]]: tensor<1x3x2x3xf32>) -> tensor<1x2x2x2xf32>

func.func @QuantizeMatmulWithActConst(%arg0: tensor<1x3x3x512xf32>) -> tensor<1x3x3x12xf32> {
  %w = arith.constant dense<127.0> : tensor<512x12xf32>
  %mm = "tfl.batch_matmul"(%arg0, %w) {adj_x = false, adj_y = false} : (tensor<1x3x3x512xf32>, tensor<512x12xf32>) -> tensor<1x3x3x12xf32>
  func.return %mm : tensor<1x3x3x12xf32>

    rhs_quantization_min_val = -128 : i64,
    rhs_quantization_max_val = 127 : i64
  } : (tensor<1x2x2x3xf32>, tensor<2x3x3x2x!tf_type.qint8>, tensor<f32>, tensor<i32>) -> tensor<1x3x3x2xf32>
  func.return %0 : tensor<1x3x3x2xf32>
}

//===----------------------------------------------------------------------===//
// tf.UniformQuantize and tf.UniformDequantize legalization