Create ClPRelu operator

Make the class that was in experimental namespace
as ClOperator to prepare porting to new interface.
The followings are added as a part of this change

Also, in-place computation is now correctly considered
to be aligned with the class description. Test cases
to test in-place computation are added.

Partially Implements: COMPMID-4184

Signed-off-by: Sang-Hoon Park <sang-hoon.park@arm.com>
Change-Id: I71c18ab47fe0370a2060d5303a58ff3650c0093f
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/5201
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Reviewed-by: Michele Di Giorgio <michele.digiorgio@arm.com>
diff --git a/tests/Utils.h b/tests/Utils.h
index ab25603..2569c41 100644
--- a/tests/Utils.h
+++ b/tests/Utils.h
@@ -514,6 +514,21 @@
 
 /** Create and initialize a tensor of the given type.
  *
+ * @param[in] info Tensor information to be used to create the tensor
+ * @param[in] ctx  (Optional) Pointer to the runtime context.
+ *
+ * @return Initialized tensor of given type.
+ */
+template <typename T>
+inline T create_tensor(const TensorInfo &info, IRuntimeContext *ctx = nullptr)
+{
+    T tensor(ctx);
+    tensor.allocator()->init(info);
+    return tensor;
+}
+
+/** Create and initialize a tensor of the given type.
+ *
  * @param[in] shape             Tensor shape.
  * @param[in] data_type         Data type.
  * @param[in] num_channels      (Optional) Number of channels.
@@ -531,9 +546,8 @@
     TensorInfo info(shape, num_channels, data_type);
     info.set_quantization_info(quantization_info);
     info.set_data_layout(data_layout);
-    tensor.allocator()->init(info);
 
-    return tensor;
+    return create_tensor<T>(info, ctx);
 }
 
 /** Create and initialize a tensor of the given type.
@@ -549,10 +563,7 @@
 {
     TensorInfo info(shape, format);
 
-    T tensor(ctx);
-    tensor.allocator()->init(info);
-
-    return tensor;
+    return create_tensor<T>(info, ctx);
 }
 
 /** Create and initialize a multi-image of the given type.
diff --git a/tests/validation/CL/PReluLayer.cpp b/tests/validation/CL/PReluLayer.cpp
index 82436a9..82f3e4f 100644
--- a/tests/validation/CL/PReluLayer.cpp
+++ b/tests/validation/CL/PReluLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2019-2020 Arm Limited.
+ * Copyright (c) 2019-2021 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -93,6 +93,63 @@
 // clang-format on
 // *INDENT-ON*
 
+TEST_SUITE(InPlace)
+TEST_CASE(Validate, framework::DatasetMode::ALL)
+{
+    // PRelu operaotr should be able to take nullptr as output and do the in-place computation.
+    // Shape and data type are selected randomly since they shouldn't matter
+    const auto tensor_info = TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32);
+    const auto result      = arm_compute::CLPReluLayer::validate(&tensor_info, &tensor_info, nullptr);
+    ARM_COMPUTE_EXPECT(bool(result) == true, framework::LogLevel::ERRORS);
+}
+
+SimpleTensor<float> compute_float_reference(const TensorInfo &tensor_info)
+{
+    SimpleTensor<float> ref_src1{ tensor_info.tensor_shape(), tensor_info.data_type() };
+    SimpleTensor<float> ref_src2{ tensor_info.tensor_shape(), tensor_info.data_type() };
+    SimpleTensor<float> ref_dst{ tensor_info.tensor_shape(), tensor_info.data_type() };
+
+    library->fill_tensor_uniform(ref_src1, 0);
+    library->fill_tensor_uniform(ref_src2, 1);
+
+    return reference::arithmetic_operation<float>(ArithmeticOperation::PRELU, ref_src1, ref_src2, ref_dst);
+}
+
+void compute_float_target_in_place(CLTensor &src1, CLTensor &src2, bool use_nullptr_output)
+{
+    auto fn = arm_compute::CLPReluLayer{};
+    fn.configure(&src1, &src2, use_nullptr_output ? nullptr : &src1);
+
+    src1.allocator()->allocate();
+    src2.allocator()->allocate();
+
+    library->fill_tensor_uniform(CLAccessor(src1), 0);
+    library->fill_tensor_uniform(CLAccessor(src2), 1);
+
+    fn.run();
+}
+
+TEST_CASE(ComputeWithNullPtr, framework::DatasetMode::ALL)
+{
+    const auto tensor_info = TensorInfo(TensorShape(33U, 13U, 2U), 1, DataType::F32);
+
+    auto src1 = create_tensor<CLTensor>(tensor_info);
+    auto src2 = create_tensor<CLTensor>(tensor_info);
+    compute_float_target_in_place(src1, src2, true);
+    validate(CLAccessor(src1), compute_float_reference(tensor_info));
+}
+
+TEST_CASE(ComputeWithSameTensor, framework::DatasetMode::ALL)
+{
+    const auto tensor_info = TensorInfo(TensorShape(33U, 13U, 2U), 1, DataType::F32);
+
+    auto src1 = create_tensor<CLTensor>(tensor_info);
+    auto src2 = create_tensor<CLTensor>(tensor_info);
+    compute_float_target_in_place(src1, src2, false);
+    validate(CLAccessor(src1), compute_float_reference(tensor_info));
+}
+TEST_SUITE_END() // InPlace
+
 template <typename T>
 using CLPReluLayerFixture = PReluLayerValidationFixture<CLTensor, CLAccessor, CLPReluLayer, T>;