%w = arith.constant dense<127.0> : tensor<1x1x12x512xf32>
  %mm = "tfl.batch_matmul"(%w, %0) {adj_x = false, adj_y = true} : (tensor<1x1x12x512xf32>, tensor<1x1x3x512xf32>) -> tensor<1x1x12x3xf32>
  %mm_s = "quantfork.stats"(%mm) {layerStats = dense<[0.000000e+00, 1.000000e+01]> : tensor<2xf32>} : (tensor<1x1x12x3xf32>) -> tensor<1x1x12x3xf32>
  func.return %mm_s : tensor<1x1x12x3xf32>

    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x512xf32>} : () -> tensor<2x3x3x512xf32>
    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x512xf32>, tensor<2x3x3x512xf32>) -> tensor<*xf32>

    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x512xf32>} : () -> tensor<2x3x3x512xf32>
    %0 = "tf.PartitionedCall"(%arg0, %cst_1) {_tfl_quant_trait = "fully_quantizable", config = "", config_proto = "", executor_type = "", f = @composite_depthwise_conv2d_fn} : (tensor<1x3x4x512xf32>, tensor<2x3x3x512xf32>) -> tensor<*xf32>

    %cst_1 = "tf.Const"() {value = dense<3.000000e+00> : tensor<2x3x3x512xf32>} : () -> tensor<2x3x3x512xf32>
    %0 = "tf.Conv2D"(%arg0, %cst_1) {attr_map = "0:strides,1:use_cudnn_on_gpu,2:padding,3:explicit_paddings,4:dilations", data_format = "NHWC", device = "", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 2, 1], use_cudnn_on_gpu = true} : (tensor<1x3x4x3xf32>, tensor<2x3x3x512xf32>) -> tensor<*xf32>
    // Dequantize added before BiasAdd.

func.func @mean_4d_keepdim(%arg0: tensor<1x48x48x512xf32>) -> tensor<1x1x1x512xf32> {
  %cst = arith.constant dense<[1, 2]> : tensor<2xi32>
  %0 = "tfl.mean"(%arg0, %cst) {keep_dims = true} : (tensor<1x48x48x512xf32>, tensor<2xi32>) -> tensor<1x1x1x512xf32>
  func.return %0 : tensor<1x1x1x512xf32>
}

// CHECK:       func @mean_4d_keepdim([[VAL_0:%.*]]: tensor<1x48x48x512xf32>) -> tensor<1x1x1x512xf32> {

// PerTensor-LABEL: QuantizeMatmulWithActConst
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>

  %1 = "tf.Maximum"(%0, %cst_0) : (tensor<1x3x4x2xf32>, tensor<f32>) -> tensor<1x3x4x2xf32>
  %2 = "tf.Minimum"(%1, %cst_1) : (tensor<1x3x4x2xf32>, tensor<f32>) -> tensor<1x3x4x2xf32>
  func.return %2 : tensor<1x3x4x2xf32>
// CHECK-DAG: %[[CONST_0:.*]] = "tf.Const"() <{value = dense<{{.*}}> : tensor<1x1x3x2xf32>}> : () -> tensor<1x1x3x2xf32>

  // CHECK: %[[VAL_1:.*]] = "tf.Identity"(%[[VAL_0]]) : (tensor<*xf32>) -> tensor<*xf32>
  // CHECK: return %[[VAL_1]]
  %0 = "tf.Conv2D"(%arg2, %arg1) {data_format = "NHWC", dilations = [1, 1, 1, 1], explicit_paddings = [], padding = "SAME", strides = [1, 1, 1, 1], use_cudnn_on_gpu = true} : (tensor<8x32x32x3xf32>, tensor<1x1x3x128xf32>) -> tensor<*xf32>

  // CHECK-DAG: %[[CST:.+]] = mhlo.constant dense<[1.000000e-01, 2.000000e-01]> : tensor<2xf32>
  // CHECK-DAG: %[[CST_BCAST:.+]] = "mhlo.broadcast_in_dim"(%[[CST]]) <{broadcast_dimensions = dense<3> : tensor<1xi64>}> : (tensor<2xf32>) -> tensor<1x1x3x2xf32>
  // CHECK-DAG: %[[NEW_FILTER:.+]] =  mhlo.multiply %[[CST_BCAST]], %[[FILTER]] : tensor<1x1x3x2xf32>

  %r = "mhlo.concatenate"(%0, %1, %2) <{dimension = 0 : i64}> : (tensor<1x1x512xf32>, tensor<1x1x512xf32>, tensor<1x1x512xf32>) -> tensor<3x1x512xf32>
  func.return %r : tensor<3x1x512xf32>

// CHECK: return %arg0 : tensor<3x1x512xf32>
}

// -----

// CHECK-LABEL: testConvertReshapeDotRhsToBatchedDot