// CHECK-SAME: %[[X:[^:[:space:]]+]]
// CHECK-SAME: %[[SCALE:[^:[:space:]]+]]
// CHECK-SAME: %[[OFFSET:[^:[:space:]]+]]
// CHECK-SAME: %[[MEAN:[^:[:space:]]+]]
// CHECK-SAME: %[[VARIANCE:[^:[:space:]]+]]
func.func @unfuse_batch_norm(
    %x: tensor<4x256xf32>, %scale: tensor<256xf32>, %offset: tensor<256xf32>,
    %mean: tensor<256xf32>, %variance: tensor<256xf32>)
    -> (tensor<4x256xf32>) {

  // CHECK: %[[RED_IDX:.*]] = "tf.Const"() <{value = dense<[2, 3]> : tensor<2xi32>}>
  // CHECK: %[[MEAN:[0-9]*]] = "tf.Mean"(%arg0, %[[RED_IDX]])
  // CHECK-SAME: (tensor<1x64x112x112xf32>, tensor<2xi32>) -> tensor<1x64xf32>
  // CHECK: return %[[MEAN]]

  // NOFOLD: %[[CST:.*]] = "tf.Const"() <{value = dense<[0, 2, 3, 1]> : tensor<4xi32>}>

  // Reduce spatial dimensions
  %16 = "tf.Mean"(%15, %1) : (tensor<?x256x56x56xf32>, tensor<2xi32>) -> tensor<?x256xf32>

  // Mean should compute reduction over NHWC spatial dimensions.
  // CHECK: %[[MEAN_DIMS:.*]] = "tf.Const"() <{value = dense<[1, 2]> : tensor<2xi32>}>
  // CHECK: %[[MEAN:[0-9]*]] = "tf.Mean"(%[[RELU]], %[[MEAN_DIMS]])

//   %y:6 = "tf.FusedBatchNormV3"(%x, %scale, %offset, %mean, %variance)
//   %0 = "tf.Relu"(%y#0)
// ->
//   %y:6 = "tf._FusedBatchNormEx"(%x, %scale, %offset, %mean, %variance)
//
// Or:
//   %y:6 = "tf.FusedBatchNormV3"(%x, %scale, %offset, %mean, %variance)
//   %0 = "tf.AddV2"(%y#0, %side_input)
//   %1 = "tf.Relu"(%0)
// ->
//  %y:6 = "tf._FusedBatchNormEx"(%x, %scale, %offset, %mean, %variance,

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>
}

// operations. Specifically, performs the following calculation:
//
//   (x - mean) * scale / sqrt(variance + epsilon) + offset
//
// Let multiplier = scale / sqrt(variance + epsilon),
// to compute
//   (x - mean) * scale / sqrt(variance + epsilon) + offset,
// is then to compute
//   (x * multiplier) + (offset - mean * multiplier).
//
// TODO(b/228916181): There is a known issue with this DDR rule that it doesn't

  "tf.EnqueueTPUEmbeddingSparseTensorBatch"(%arg0, %arg1, %arg2, %arg3, %arg4, %arg5, %0, %0, %0, %arg7) {_tpu_embedding_layer = "call1", _xla_outside_compilation = "0", combiners = ["mean", "sum"], device_ordinal = -1 : i64, max_sequence_lengths = [0, 0, 0], table_ids = [1, 1, 0]} : (tensor<?x2xi32>, tensor<?x2xi32>, tensor<?x2xi32>, tensor<?xi32>, tensor<?xi32>, tensor<?xi32>, tensor<0xf32>, tensor<0xf32>, tensor<0xf32>, tensor<!tf_type.string>)...

	// len(fractions) > 2: calculate std based on the well-known formula - root square of Σ((fraction(i)-mean)^2)/len(fractions)
	// Otherwise, set the std to zero is enough.
	if len(resourceToFractions) == 2 {
		std = math.Abs((resourceToFractions[0] - resourceToFractions[1]) / 2)

	} else if len(resourceToFractions) > 2 {
		mean := totalFraction / float64(len(resourceToFractions))
		var sum float64

  snprintf(buf, kBufSize, "Mean with %2.0f%% trimmed:", trim_ratio * 100);
  std::string label_trimmed(buf);
  snprintf(buf, kBufSize, "Mean of %2.0f%% best:", best_ratio * 100);
  std::string label_best(buf);
  std::vector<std::pair<std::string, double>> groups = {
      {"Best:", sorted_us.front()},
      {"Worst:", sorted_us.back()},
      {"Median:", sorted_us[count_us / 2]},
      {"Mean:", sum_us / count_us},

        when:
        repository.get('unknown')

        then:
        UnknownDomainObjectException e = thrown()
        e.message == 'No meta-data is available for class \'unknown\'. Did you mean? [unkown]'
    }

    def getFailsForWrongPackage() {
        given:
        repository.put('org.gradle.jvm.test.JUnitTestSuiteSpec', new TestDomainObject('org.gradle.jvm.test.JUnitTestSuiteSpec'))

        when: