%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>

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-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

  func.return %result : tensor<1x1x512xf32>

// CHECK:           %[[VAL_9:.*]] = "tfl.reshape"(%[[VAL_8]], %[[VAL_5]]) : (tensor<1x1x384x512xf32>, tensor<2xi32>) -> tensor<384x512xf32>
// CHECK:           return %[[VAL_9]] : tensor<384x512xf32>
// CHECK:         }

// -----

// CHECK:           %[[VAL_11:.*]] = "tfl.reshape"(%[[VAL_9]], %[[VAL_6]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<384x512xf32>, tensor<4xi32>) -> tensor<1x1x384x512xf32>
// CHECK:           %[[VAL_12:.*]] = "tfl.reshape"(%[[VAL_10]], %[[VAL_7]]) {tac.device = "GPU", tac.inference_type = "FLOAT"} : (tensor<128x512xf32>, tensor<4xi32>) -> tensor<128x1x1x512xf32>

// CHECK:  %1 = "tfl.depthwise_conv_2d"(%arg0, %0, %[[CONSTANT]]) <{depth_multiplier = 4 : i32, dilation_h_factor = 2 : i32, dilation_w_factor = 3 : i32, fused_activation_function = "NONE", padding = "SAME", stride_h = 4 : i32, stride_w = 5 : i32}> : (tensor<256x32x32x3xf32>, tensor<1x3x3x12xf32>, tensor<12xf32>) -> tensor<256x30x30x12xf32>