Dynamic Fusion Pooling Layer 2d

- Adds Dynamic fusion PoolingLayer2D as Unfusable Operator
- Indices are not supported
- Adds tests for F32/F16 Datatypes

Resolves : [COMPMID-5520]

Signed-off-by: Mohammed Suhail Munshi <MohammedSuhail.Munshi@arm.com>
Change-Id: I0d112545eb9209c836bf9ea153069f8627531e0a
Reviewed-on: https://review.mlplatform.org/c/ml/ComputeLibrary/+/8893
Reviewed-by: Gunes Bayir <gunes.bayir@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>
Benchmark: Arm Jenkins <bsgcomp@arm.com>

4 files changed, 546 insertions, 2 deletions

diff --git a/tests/datasets/dynamic_fusion/PoolingLayerDataset.h b/tests/datasets/dynamic_fusion/PoolingLayerDataset.h
new file mode 100644
index 0000000000..c4911f4940
--- /dev/null
+++ b/tests/datasets/dynamic_fusion/PoolingLayerDataset.h
@@ -0,0 +1,122 @@
+/*
+ * Copyright (c) 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "utils/TypePrinter.h"
+#include "arm_compute/dynamic_fusion/sketch/attributes/Pool2dAttributes.h"
+
+
+using Pool2dAttributes = arm_compute::experimental::dynamic_fusion::Pool2dAttributes;
+
+namespace arm_compute
+{
+namespace test
+{
+namespace datasets
+{
+
+class DynamicFusionPoolingLayerDataset
+{
+public:
+    using type = std::tuple<TensorShape, Pool2dAttributes>;
+
+    struct iterator
+    {
+        iterator(std::vector<TensorShape>::const_iterator      src_it,
+                 std::vector<Pool2dAttributes>::const_iterator infos_it)
+            : _src_it{ std::move(src_it) },
+              _infos_it{ std::move(infos_it) }
+        {
+        }
+
+        std::string description() const
+        {
+            std::stringstream description;
+            description << "In=" << *_src_it << ":";
+            description << "Info=" << *_infos_it << ":";
+            return description.str();
+        }
+
+        DynamicFusionPoolingLayerDataset::type operator*() const
+        {
+            return std::make_tuple(*_src_it, *_infos_it);
+        }
+
+        iterator &operator++()
+        {
+            ++_src_it;
+            ++_infos_it;
+
+            return *this;
+        }
+
+    private:
+        std::vector<TensorShape>::const_iterator      _src_it;
+        std::vector<Pool2dAttributes>::const_iterator _infos_it;
+    };
+
+    iterator begin() const
+    {
+        return iterator(_src_shapes.begin(), _infos.begin());
+    }
+
+    int size() const
+    {
+        return std::min(_src_shapes.size(), _infos.size());
+    }
+
+    void add_config(TensorShape src, Pool2dAttributes info)
+    {
+        _src_shapes.emplace_back(std::move(src));
+        _infos.emplace_back(std::move(info));
+    }
+
+protected:
+    DynamicFusionPoolingLayerDataset()                       = default;
+    DynamicFusionPoolingLayerDataset(DynamicFusionPoolingLayerDataset &&) = default;
+
+private:
+    std::vector<TensorShape>      _src_shapes{};
+    std::vector<Pool2dAttributes> _infos{};
+};
+
+// Special pooling dataset
+class PoolingLayerDatasetSpecialDynamicFusion final : public DynamicFusionPoolingLayerDataset
+{
+public:
+    PoolingLayerDatasetSpecialDynamicFusion()
+    {
+        // NCHW DataLayout 
+        // Special cases
+        add_config(TensorShape(2U, 3U, 4U, 1U), Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(2,2)).stride(Size2D(3,3)));
+        add_config(TensorShape(60U, 52U, 3U, 2U), Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(100,100)).stride(Size2D(5,5)).pad(Padding2D(50,50,50,50)));
+        // Asymmetric padding
+        add_config(TensorShape(112U, 112U, 32U), Pool2dAttributes().pool_type(PoolingType::MAX).pool_size(Size2D(3,3)).pad(Padding2D(0,1,0,1)).stride(Size2D(2,2)));
+        add_config(TensorShape(14U, 14U, 832U), Pool2dAttributes().pool_type(PoolingType::MAX).pool_size(Size2D(2,2)).stride(Size2D(1,1)).pad(Padding2D(0,0,0,0)));
+
+    }
+};
+} // namespace datasets
+} // namespace test
+} // namespace arm_compute
\ No newline at end of file
diff --git a/tests/validation/dynamic_fusion/gpu/cl/Pool2d.cpp b/tests/validation/dynamic_fusion/gpu/cl/Pool2d.cpp
new file mode 100644
index 0000000000..a7772aef4d
--- /dev/null
+++ b/tests/validation/dynamic_fusion/gpu/cl/Pool2d.cpp
@@ -0,0 +1,233 @@
+/*
+ * Copyright (c) 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+
+#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuPool2d.h"
+
+#include "tests/CL/CLAccessor.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/datasets/dynamic_fusion/PoolingLayerDataset.h"
+#include "tests/framework/Fixture.h"
+#include "tests/framework/Macros.h"
+#include "tests/framework/datasets/Datasets.h"
+#include "tests/validation/Validation.h"
+#include "tests/validation/fixtures/dynamic_fusion/gpu/cl/Pool2dFixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+TEST_SUITE(CL)
+TEST_SUITE(DYNAMIC_FUSION)
+TEST_SUITE(POOL2D)
+
+constexpr AbsoluteTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for 32-bit floating-point type */
+constexpr AbsoluteTolerance<float> tolerance_f16(0.01f);  /**< Tolerance value for comparing reference's output against implementation's output for 16-bit floating-point type */
+
+const auto PoolingLayerDatasetFP = combine(combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { Size2D(2, 2), Size2D(3, 3) })),
+                                                           framework::dataset::make("Pad", { Padding2D() })),
+                                                   framework::dataset::make("Stride", { Size2D(1, 1), Size2D(2, 1), Size2D(5, 7) })),
+                                           framework::dataset::make("ExcludePadding", { true }));
+
+const auto pool_fp_mixed_precision_dataset = framework::dataset::make("FpMixedPrecision", { true, false });
+
+template <typename T>
+using DynamicFusionGpuPool2dFixture = DynamicFusionGpuPool2dValidationFixture<CLTensor, CLAccessor, GpuPool2d, T>;
+
+template <typename T>
+using DFSpecialGpuPool2dFixture = DynamicFusionGpuPool2dSpecialValidationFixture<CLTensor, CLAccessor, GpuPool2d, T>;
+
+template <typename T>
+using DFPoolMixedPrecisionFixture = DynamicFusionGpuPool2dMixedPrecisionValidationFixture<CLTensor, CLAccessor, GpuPool2d, T>;
+// *INDENT-OFF*
+// clang-format off
+
+DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(
+            framework::dataset::make("InputInfo", { TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),     // Mismatching data type
+                                                    TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),     // Invalid pad/size combination
+                                                    TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::F32, DataLayout::NHWC),     // Invalid pad/size combination
+                                                    TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::QASYMM8, DataLayout::NHWC), // Invalid parameters, unsupported pooling
+                                                    TensorInfo(TensorShape(5U, 15U, 13U), 1, DataType::F32, DataLayout::NHWC),     // Valid Non-rectangular Global Pooling
+                                                    TensorInfo(TensorShape(5U, 13U, 13U), 1, DataType::F32, DataLayout::NHWC),     // Invalid output Global Pooling
+                                                    TensorInfo(TensorShape(5U, 13U, 13U), 1, DataType::QASYMM8, DataLayout::NHWC), // Invalid - Quantized not supported.
+                                                    TensorInfo(TensorShape(5U, 13U, 13U), 1, DataType::F32, DataLayout::NHWC),     // Valid global pooling
+                                                    TensorInfo(TensorShape(13U, 13U, 5U), 1, DataType::F32, DataLayout::NCHW),     // Unsupported data layout
+                                                }),
+            framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(2U, 25U, 11U), 1, DataType::F16, DataLayout::NHWC),
+                                                    TensorInfo(TensorShape(2U, 30U, 11U), 1, DataType::F32, DataLayout::NHWC),
+                                                    TensorInfo(TensorShape(2U, 25U, 16U), 1, DataType::F32, DataLayout::NHWC),
+                                                    TensorInfo(TensorShape(2U, 27U, 13U), 1, DataType::QASYMM8, DataLayout::NHWC),
+                                                    TensorInfo(TensorShape(5U, 1U, 1U), 1, DataType::F32, DataLayout::NHWC),
+                                                    TensorInfo(TensorShape(5U, 2U, 2U), 1, DataType::F32, DataLayout::NHWC),
+                                                    TensorInfo(TensorShape(5U, 12U, 12U), 1, DataType::QASYMM8, DataLayout::NHWC),
+                                                    TensorInfo(TensorShape(5U, 1U, 1U), 1, DataType::F32, DataLayout::NHWC),
+                                                    TensorInfo(TensorShape(1U, 1U, 5U), 1, DataType::F32, DataLayout::NHWC),
+                                                })),
+            framework::dataset::make("Pool2dAttributes", {
+                                                    Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(3,3)).pad(Padding2D(0,0,0,0)).stride(Size2D(1,1)),
+                                                    Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(2,2)).pad(Padding2D(2,2,0,0)).stride(Size2D(1,1)),
+                                                    Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(2,2)).pad(Padding2D(0,0,2,2)).stride(Size2D(1,1)),
+                                                    Pool2dAttributes().pool_type(PoolingType::L2).pool_size(Size2D(3,3)).pad(Padding2D(0,0,0,0)).stride(Size2D(1,1)),
+                                                    Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(15U, 13U)),
+                                                    Pool2dAttributes().pool_type(PoolingType::MAX).pool_size(Size2D(13U, 13U)),
+                                                    Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(2,2)).pad(Padding2D()).stride(Size2D(1,1)),
+                                                    Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(13U,13U)),
+                                                    Pool2dAttributes().pool_type(PoolingType::AVG).pool_size(Size2D(13U,13U)),
+                                                })),
+            framework::dataset::make("Expected", { false, false, false, false, true, false, false, true, false })),
+            input_info, output_info, pool2d_attr, expected)
+{
+    // Create a new workload sketch
+    auto              cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
+    auto              gpu_ctx        = GpuWorkloadContext{ &cl_compile_ctx };
+    GpuWorkloadSketch sketch{ &gpu_ctx };
+
+    // Declare GpuPool2d settings
+    const GpuPool2dSettings &settings = GpuPool2dSettings().mixed_precision(false);
+
+    // Validate Pool2d Configuration
+    auto                   src_info    = sketch.create_tensor_info(input_info);
+    auto                   dst_info    = sketch.create_tensor_info(output_info);
+    bool                   res         = bool(GpuPool2d::validate_op(sketch, &src_info, &dst_info, pool2d_attr, settings));
+    ARM_COMPUTE_EXPECT(res == expected, framework::LogLevel::ERRORS);
+}
+
+// clang-format on
+// *INDENT-ON*
+
+TEST_SUITE(Float)
+TEST_SUITE(FP32)
+FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuPool2dFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallNoneUnitShapes(), PoolingLayerDatasetFP),
+                                                                                                                  framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuPool2dFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), PoolingLayerDatasetFP),
+                                                                                                                framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+FIXTURE_DATA_TEST_CASE(RunSpecial, DFSpecialGpuPool2dFixture<float>, framework::DatasetMode::ALL, combine(datasets::PoolingLayerDatasetSpecialDynamicFusion(),
+                                                                                                                  framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+
+TEST_SUITE(GlobalPooling)
+FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuPool2dFixture<float>, framework::DatasetMode::ALL,
+                       combine(combine(combine(combine(combine(combine(
+                                                                   framework::dataset::make("InputShape", { TensorShape(27U, 13U, 2U),
+                                                                                                            TensorShape(27U, 13U, 2U, 4U)
+                                                                                                          }),
+                                                                   framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::MAX })),
+                                                               framework::dataset::make("PoolingSize", { Size2D(27, 13) })),
+                                                       framework::dataset::make("Pad", { Padding2D() })),
+                                               framework::dataset::make("Stride", { Size2D(1, 1) })),
+                                       framework::dataset::make("ExcludePadding", true)),
+                               framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuPool2dFixture<float>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(combine(combine(combine(
+                                                                   framework::dataset::make("InputShape", { TensorShape(79U, 37U, 11U),
+                                                                                                            TensorShape(79U, 37U, 11U, 4U)
+                                                                                                          }),
+                                                                   framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::MAX })),
+                                                               framework::dataset::make("PoolingSize", { Size2D(79, 37) })),
+                                                       framework::dataset::make("Pad", { Padding2D() })),
+                                               framework::dataset::make("Stride", { Size2D(1, 1) })),
+                                       framework::dataset::make("ExcludePadding", true)),
+                               framework::dataset::make("DataType", DataType::F32)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f32);
+}
+TEST_SUITE_END() // GlobalPooling
+TEST_SUITE_END() // FP32
+
+TEST_SUITE(FP16)
+FIXTURE_DATA_TEST_CASE(RunSmall, DFPoolMixedPrecisionFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallNoneUnitShapes(), PoolingLayerDatasetFP),
+                                                                                                                       framework::dataset::make("DataType", DataType::F16)),
+                                                                                                               pool_fp_mixed_precision_dataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+FIXTURE_DATA_TEST_CASE(RunLarge, DFPoolMixedPrecisionFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), PoolingLayerDatasetFP),
+                                                                                                                     framework::dataset::make("DataType", DataType::F16)),
+                                                                                                             pool_fp_mixed_precision_dataset))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+
+TEST_SUITE(GlobalPooling)
+FIXTURE_DATA_TEST_CASE(RunSmall, DynamicFusionGpuPool2dFixture<half>, framework::DatasetMode::ALL,
+                       combine(combine(combine(combine(combine(combine(
+                                                                   framework::dataset::make("InputShape", { TensorShape(27U, 13U, 2U),
+                                                                                                            TensorShape(27U, 13U, 2U, 4U)
+                                                                                                          }),
+                                                                   framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::MAX })),
+                                                               framework::dataset::make("PoolingSize", { Size2D(27, 13) })),
+                                                       framework::dataset::make("Pad", { Padding2D() })),
+                                               framework::dataset::make("Stride", { Size2D(1, 1) })),
+                                       framework::dataset::make("ExcludePadding", true)),
+                               framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+
+FIXTURE_DATA_TEST_CASE(RunLarge, DynamicFusionGpuPool2dFixture<half>, framework::DatasetMode::NIGHTLY,
+                       combine(combine(combine(combine(combine(combine(
+                                                                   framework::dataset::make("InputShape", { TensorShape(79U, 37U, 11U),
+                                                                                                            TensorShape(79U, 37U, 11U, 4U)
+                                                                                                          }),
+                                                                   framework::dataset::make("PoolingType", { PoolingType::AVG, PoolingType::MAX })),
+                                                               framework::dataset::make("PoolingSize", { Size2D(79, 37) })),
+                                                       framework::dataset::make("Pad", { Padding2D() })),
+                                               framework::dataset::make("Stride", { Size2D(1, 1) })),
+                                       framework::dataset::make("ExcludePadding", true)),
+                               framework::dataset::make("DataType", DataType::F16)))
+{
+    // Validate output
+    validate(CLAccessor(_target), _reference, tolerance_f16);
+}
+TEST_SUITE_END() // GlobalPooling
+TEST_SUITE_END() // FP16
+TEST_SUITE_END() // FLOAT
+
+TEST_SUITE_END() // POOL2D
+TEST_SUITE_END() // DYNAMIC_FUSION
+TEST_SUITE_END() // CL
+}
+}
+}
diff --git a/tests/validation/fixtures/dynamic_fusion/gpu/cl/Pool2dFixture.h b/tests/validation/fixtures/dynamic_fusion/gpu/cl/Pool2dFixture.h
new file mode 100644
index 0000000000..efb67f8b11
--- /dev/null
+++ b/tests/validation/fixtures/dynamic_fusion/gpu/cl/Pool2dFixture.h
@@ -0,0 +1,190 @@
+/*
+ * Copyright (c) 2023 Arm Limited.
+ *
+ * SPDX-License-Identifier: MIT
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a copy
+ * of this software and associated documentation files (the "Software"), to
+ * deal in the Software without restriction, including without limitation the
+ * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or
+ * sell copies of the Software, and to permit persons to whom the Software is
+ * furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice shall be included in all
+ * copies or substantial portions of the Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+ * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+ * SOFTWARE.
+ */
+#ifndef TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_POOL2DFIXTURE
+#define TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_POOL2DFIXTURE
+
+#include "arm_compute/core/CL/CLKernelLibrary.h"
+#include "arm_compute/core/TensorInfo.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+
+#include "arm_compute/dynamic_fusion/runtime/gpu/cl/ClWorkloadRuntime.h"
+#include "arm_compute/dynamic_fusion/sketch/attributes/Pool2dAttributes.h"
+#include "arm_compute/dynamic_fusion/sketch/gpu/GpuWorkloadSketch.h"
+#include "arm_compute/dynamic_fusion/sketch/gpu/operators/GpuPool2d.h"
+#include "src/dynamic_fusion/utils/Utils.h"
+
+#include "tests/CL/CLAccessor.h"
+#include "tests/framework/Fixture.h"
+#include "tests/validation/reference/PoolingLayer.h"
+
+using namespace arm_compute::experimental::dynamic_fusion;
+
+namespace arm_compute
+{
+namespace test
+{
+namespace validation
+{
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class DynamicFusionGpuPool2dValidationGenericFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, const Pool2dAttributes &pool_attr, DataType data_type, bool mixed_precision)
+    {
+        _target    = compute_target(input_shape, pool_attr, data_type, mixed_precision);
+        _reference = compute_reference(input_shape, convert_pool_attr_to_pool_info(pool_attr, mixed_precision), data_type);
+    }
+
+protected:
+    template <typename U>
+    void fill(U &&tensor, int i)
+    {
+        switch(tensor.data_type())
+        {
+            case DataType::F16:
+            {
+                arm_compute::utils::uniform_real_distribution_16bit<half> distribution{ -1.0f, 1.0f };
+                library->fill(tensor, distribution, i);
+                break;
+            }
+            case DataType::F32:
+            {
+                std::uniform_real_distribution<float> distribution(-1.0f, 1.0f);
+                library->fill(tensor, distribution, i);
+                break;
+            }
+            default:
+                library->fill_tensor_uniform(tensor, i);
+        }
+    }
+
+    // Given input is in nchw format
+    TensorType compute_target(TensorShape input_shape, const Pool2dAttributes &pool_attr, const DataType data_type, bool mixed_precision)
+    {
+        CLScheduler::get().default_reinit();
+
+        // Change shape due to NHWC data layout, test shapes are NCHW
+        permute(input_shape, PermutationVector(2U, 0U, 1U));
+
+        // Create a new workload sketch
+        auto              cl_compile_ctx = CLKernelLibrary::get().get_compile_context();
+        auto              gpu_ctx        = GpuWorkloadContext{ &cl_compile_ctx };
+        GpuWorkloadSketch sketch{ &gpu_ctx };
+
+        // Create sketch tensors
+        auto input_info = sketch.create_tensor_info(TensorInfo(input_shape, 1, data_type, DataLayout::NHWC));
+        auto dst_info   = sketch.create_tensor_info();
+
+        // Create Pool2dSettings
+        GpuPool2dSettings pool_settings = GpuPool2dSettings().mixed_precision(mixed_precision);
+
+        FunctionType::create_op(sketch, &input_info, &dst_info, pool_attr, pool_settings);
+
+        // Configure runtime
+        ClWorkloadRuntime runtime;
+        runtime.configure(sketch);
+        // (Important) Allocate auxiliary tensor memory if there are any
+        for(auto &data : runtime.get_auxiliary_tensors())
+        {
+            auto       tensor      = data.first;
+            const auto aux_mem_req = data.second;
+            tensor->allocator()->init(*data.first->info(), aux_mem_req.alignment);
+            tensor->allocator()->allocate(); // Use ACL allocated memory
+        }
+        // Construct user tensors
+        TensorType t_input{};
+        TensorType t_dst{};
+
+        // Initialize user tensors
+        t_input.allocator()->init(input_info);
+        t_dst.allocator()->init(dst_info);
+
+        // Allocate and fill user tensors
+        t_input.allocator()->allocate();
+        t_dst.allocator()->allocate();
+
+        fill(AccessorType(t_input), 0);
+
+        // Run runtime
+        runtime.run({ &t_input, &t_dst });
+        return t_dst;
+    }
+
+    SimpleTensor<T> compute_reference(TensorShape shape, PoolingLayerInfo pool_info, DataType data_type)
+    {
+        // Create reference
+        SimpleTensor<T> src(shape, data_type, 1, QuantizationInfo());
+        // Fill reference
+        fill(src, 0);
+        return reference::pooling_layer<T>(src, pool_info, QuantizationInfo(), nullptr, DataLayout::NCHW);
+    }
+
+    TensorType       _target{};
+    SimpleTensor<T>  _reference{};
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class DynamicFusionGpuPool2dValidationFixture : public DynamicFusionGpuPool2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, PoolingType pool_type, Size2D pool_size, Padding2D pad, Size2D stride, bool exclude_padding, DataType data_type)
+    {
+        DynamicFusionGpuPool2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape,
+                                                                                                         Pool2dAttributes().pool_type(pool_type).pool_size(pool_size).pad(pad).stride(stride).exclude_padding(exclude_padding),
+                                                                                                         data_type, false);
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class DynamicFusionGpuPool2dMixedPrecisionValidationFixture : public DynamicFusionGpuPool2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, PoolingType pool_type, Size2D pool_size, Padding2D pad, Size2D stride, bool exclude_padding, DataType data_type, bool mixed_precision)
+    {
+        DynamicFusionGpuPool2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape,
+                                                                                                         Pool2dAttributes().pool_type(pool_type).pool_size(pool_size).pad(pad).stride(stride).exclude_padding(exclude_padding),
+                                                                                                         data_type, mixed_precision);
+    }
+};
+
+template <typename TensorType, typename AccessorType, typename FunctionType, typename T>
+class DynamicFusionGpuPool2dSpecialValidationFixture : public DynamicFusionGpuPool2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>
+{
+public:
+    template <typename...>
+    void setup(TensorShape input_shape, Pool2dAttributes pool_attr, DataType data_type)
+    {
+        DynamicFusionGpuPool2dValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, pool_attr, data_type, false);
+    }
+};
+
+} // namespace validation
+} // namespace test
+} // namespace arm_compute
+
+#endif /* TESTS_VALIDATION_FIXTURES_DYNAMIC_FUSION_GPU_CL_POOL2DFIXTURE */
diff --git a/tests/validation/reference/PoolingLayer.cpp b/tests/validation/reference/PoolingLayer.cpp
index 9e671e3173..378d91d829 100644
--- a/tests/validation/reference/PoolingLayer.cpp
+++ b/tests/validation/reference/PoolingLayer.cpp
@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2017-2021 Arm Limited.
+ * Copyright (c) 2017-2021, 2023 Arm Limited.
  *
  * SPDX-License-Identifier: MIT
  *
@@ -40,7 +40,6 @@ using namespace arm_compute::misc::shape_calculator;
 template <typename T, typename ACC_T, typename std::enable_if<is_floating_point<T>::value, int>::type>
 SimpleTensor<T> pooling_layer_internal(const SimpleTensor<T> &src, const PoolingLayerInfo &info, SimpleTensor<uint32_t> *indices, DataLayout data_layout)
 {
-    ARM_COMPUTE_ERROR_ON(info.is_global_pooling && (src.shape().x() != src.shape().y()));
     // Create reference
     SimpleTensor<T> dst{ compute_pool_shape(TensorInfo(src.shape(), 1, src.data_type()), info), src.data_type(), 1 };
     auto            pooled_shape = compute_pool_shape(TensorInfo(src.shape(), 1, src.data_type()), info);