MLCE-129: NEPad 30x slower than TensorFlow's implementation

Change-Id: I44770e6a3134c70c4bd58f890d06cb43c9bd8bff
Signed-off-by: Manuel Bottini <manuel.bottini@arm.com>
Reviewed-on: https://review.mlplatform.org/c/1853
Reviewed-by: Giorgio Arena <giorgio.arena@arm.com>
Comments-Addressed: Arm Jenkins <bsgcomp@arm.com>
Tested-by: Arm Jenkins <bsgcomp@arm.com>

3 files changed, 196 insertions, 1 deletions

diff --git a/tests/benchmark/NEON/PadLayer.cpp b/tests/benchmark/NEON/PadLayer.cpp
new file mode 100644
index 0000000000..c55c93b8a6
--- /dev/null
+++ b/tests/benchmark/NEON/PadLayer.cpp
@@ -0,0 +1,87 @@
+/*
+ * Copyright (c) 2019 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/Types.h"
+#include "arm_compute/runtime/NEON/functions/NEPadLayer.h"
+#include "arm_compute/runtime/Tensor.h"
+#include "arm_compute/runtime/TensorAllocator.h"
+#include "tests/NEON/Accessor.h"
+#include "tests/benchmark/fixtures/PadLayerFixture.h"
+#include "tests/datasets/ShapeDatasets.h"
+#include "tests/datasets/SplitDataset.h"
+#include "tests/framework/Asserts.h"
+#include "tests/framework/Macros.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace benchmark
+{
+namespace
+{
+const auto Fssd_25_8bit_ShapesDataset = framework::dataset::make("TensorShape",
+{
+    TensorShape{ 320U, 320U, 3U },
+    TensorShape{ 160U, 160U, 16U },
+    TensorShape{ 80U, 80U, 32U },
+    TensorShape{ 40U, 40U, 64U },
+    TensorShape{ 20U, 20U, 128U },
+    TensorShape{ 10U, 10U, 256U },
+    TensorShape{ 10U, 10U, 64U },
+    TensorShape{ 5U, 5U, 32U },
+    TensorShape{ 3U, 3U, 32U },
+    TensorShape{ 2U, 2U, 32U }
+});
+
+const auto PaddingSizesDataset = framework::dataset::make("PaddingSize",
+{
+    PaddingList{ { 1, 1 }, { 1, 1 } },
+});
+} // namespace
+
+TEST_SUITE(NEON)
+TEST_SUITE(PadLayer)
+
+template <typename T>
+using NEPaddingFixture = PaddingFixture<Tensor, Accessor, NEPadLayer, T>;
+
+REGISTER_FIXTURE_DATA_TEST_CASE(RunF32, NEPaddingFixture<float>, framework::DatasetMode::ALL,
+                                combine(combine(combine(
+                                                    Fssd_25_8bit_ShapesDataset,
+                                                    framework::dataset::make("DataType", { DataType::F32 })),
+                                                PaddingSizesDataset),
+                                        framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT, PaddingMode::REFLECT })));
+
+REGISTER_FIXTURE_DATA_TEST_CASE(RunQASYMM8, NEPaddingFixture<uint8_t>, framework::DatasetMode::ALL,
+                                combine(combine(combine(
+                                                    Fssd_25_8bit_ShapesDataset,
+                                                    framework::dataset::make("DataType", { DataType::QASYMM8 })),
+                                                PaddingSizesDataset),
+                                        framework::dataset::make("PaddingMode", { PaddingMode::CONSTANT, PaddingMode::REFLECT })));
+
+TEST_SUITE_END() // PadLayer
+TEST_SUITE_END() // NEON
+} // namespace benchmark
+} // namespace test
+} // namespace arm_compute
diff --git a/tests/benchmark/fixtures/PadLayerFixture.h b/tests/benchmark/fixtures/PadLayerFixture.h
new file mode 100644
index 0000000000..2f482a0abf
--- /dev/null
+++ b/tests/benchmark/fixtures/PadLayerFixture.h
@@ -0,0 +1,109 @@
+/*
+ * Copyright (c) 2019 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 ARM_COMPUTE_TEST_PADLAYERFIXTURE
+#define ARM_COMPUTE_TEST_PADLAYERFIXTURE
+
+#include "arm_compute/core/TensorShape.h"
+#include "arm_compute/core/Types.h"
+#include "arm_compute/core/utils/misc/ShapeCalculator.h"
+#include "tests/Globals.h"
+#include "tests/Utils.h"
+#include "tests/framework/Fixture.h"
+
+namespace arm_compute
+{
+namespace test
+{
+namespace benchmark
+{
+/** Fixture that can be used for NEON and CL */
+
+template <typename TensorType, typename Accessor, typename Function, typename T>
+class PaddingFixture : public framework::Fixture
+{
+public:
+    template <typename...>
+    void setup(TensorShape shape, DataType data_type, const PaddingList &paddings, const PaddingMode mode)
+    {
+        PaddingList clamped_padding = paddings;
+        if(mode != PaddingMode::CONSTANT)
+        {
+            // Clamp padding to prevent applying more than is possible.
+            for(uint32_t i = 0; i < paddings.size(); ++i)
+            {
+                if(mode == PaddingMode::REFLECT)
+                {
+                    clamped_padding[i].first  = std::min(static_cast<uint64_t>(paddings[i].first), static_cast<uint64_t>(shape[i] - 1));
+                    clamped_padding[i].second = std::min(static_cast<uint64_t>(paddings[i].second), static_cast<uint64_t>(shape[i] - 1));
+                }
+                else
+                {
+                    clamped_padding[i].first  = std::min(static_cast<uint64_t>(paddings[i].first), static_cast<uint64_t>(shape[i]));
+                    clamped_padding[i].second = std::min(static_cast<uint64_t>(paddings[i].second), static_cast<uint64_t>(shape[i]));
+                }
+            }
+        }
+
+        const PixelValue const_value = PixelValue(static_cast<T>(0));
+
+        TensorShape output_shape = arm_compute::misc::shape_calculator::compute_padded_shape(shape, paddings);
+
+        // Create tensors
+        src = create_tensor<TensorType>(shape, data_type);
+        dst = create_tensor<TensorType>(output_shape, data_type);
+
+        // Create and configure function
+        pad_layer.configure(&src, &dst, paddings, const_value, mode);
+
+        // Allocate tensors
+        src.allocator()->allocate();
+        dst.allocator()->allocate();
+    }
+
+    void run()
+    {
+        pad_layer.run();
+    }
+
+    void sync()
+    {
+        sync_if_necessary<TensorType>();
+        sync_tensor_if_necessary<TensorType>(dst);
+    }
+
+    void teardown()
+    {
+        src.allocator()->free();
+        dst.allocator()->free();
+    }
+
+private:
+    TensorType src{};
+    TensorType dst{};
+    Function   pad_layer{};
+};
+} // namespace benchmark
+} // namespace test
+} // namespace arm_compute
+#endif /* ARM_COMPUTE_TEST_PADLAYERFIXTURE */
diff --git a/tests/validation/fixtures/PadLayerFixture.h b/tests/validation/fixtures/PadLayerFixture.h
index 3538cabfeb..58ca81f0df 100644
--- a/tests/validation/fixtures/PadLayerFixture.h
+++ b/tests/validation/fixtures/PadLayerFixture.h
@@ -109,7 +109,6 @@ protected:
 
         // Compute function
         padding.run();
-
         return dst;
     }