From 33ff9ef467153eef05b700820d859515a52481f4 Mon Sep 17 00:00:00 2001 From: Xinghang Zhou Date: Wed, 17 Jan 2018 11:23:39 +0800 Subject: APPBROWSER-400: Implement the tensorshift kernel for fixing DC's alignment issue on OpenGL ES Change-Id: I7a8489bb0fddc72899ea165e414ee87bdbfb45b3 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/118106 Reviewed-by: Joel Liang Tested-by: Jenkins Reviewed-by: Anthony Barbier --- tests/datasets/ShapeDatasets.h | 22 ++ .../DirectConvolutionLayerTensorShift.cpp | 90 +++++++ .../DirectConvolutionLayerTensorShiftFixture.h | 269 +++++++++++++++++++++ 3 files changed, 381 insertions(+) create mode 100644 tests/validation/GLES_COMPUTE/DirectConvolutionLayerTensorShift.cpp create mode 100644 tests/validation/fixtures/DirectConvolutionLayerTensorShiftFixture.h (limited to 'tests') diff --git a/tests/datasets/ShapeDatasets.h b/tests/datasets/ShapeDatasets.h index dbcd9d5000..9114f514aa 100644 --- a/tests/datasets/ShapeDatasets.h +++ b/tests/datasets/ShapeDatasets.h @@ -387,6 +387,28 @@ public: } }; +/** Data set containing small tensor shapes for direct convolution. */ +class SmallDirectConvolutionTensorShiftShapes final : public ShapeDataset +{ +public: + SmallDirectConvolutionTensorShiftShapes() + : ShapeDataset("InputShape", + { + // Batch size 1 + TensorShape{ 35U, 35U, 3U }, + TensorShape{ 32U, 37U, 3U }, + // Batch size 4 + TensorShape{ 32U, 37U, 3U, 4U }, + // Batch size 8 + TensorShape{ 32U, 37U, 3U, 8U }, + TensorShape{ 33U, 35U, 3U, 8U }, + // Arbitrary batch size + TensorShape{ 32U, 37U, 3U, 8U } + }) + { + } +}; + /** Data set containing 2D tensor shapes for DepthConcatenateLayer. */ class DepthConcatenateLayerShapes final : public ShapeDataset { diff --git a/tests/validation/GLES_COMPUTE/DirectConvolutionLayerTensorShift.cpp b/tests/validation/GLES_COMPUTE/DirectConvolutionLayerTensorShift.cpp new file mode 100644 index 0000000000..45fb76cad5 --- /dev/null +++ b/tests/validation/GLES_COMPUTE/DirectConvolutionLayerTensorShift.cpp @@ -0,0 +1,90 @@ +/* + * Copyright (c) 2017-2018 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/GLES_COMPUTE/GCTensor.h" +#include "arm_compute/runtime/GLES_COMPUTE/GCTensorAllocator.h" +#include "arm_compute/runtime/GLES_COMPUTE/functions/GCDirectConvolutionLayer.h" +#include "tests/GLES_COMPUTE/GCAccessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Macros.h" +#include "tests/framework/datasets/Datasets.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/DirectConvolutionLayerTensorShiftFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +RelativeTolerance tolerance_fp16(half(0.2)); /**< Tolerance for floating point tests */ +RelativeTolerance tolerance_fp32(0.02f); /**< Tolerance for floating point tests */ +constexpr float tolerance_num = 0.07f; /**< Tolerance number */ + +/** Direct convolution data set. */ +const auto data = combine(datasets::SmallDirectConvolutionTensorShiftShapes(), + combine(framework::dataset::make("StrideX", 1, 3), + combine(framework::dataset::make("StrideY", 1, 3), + combine(concat(combine(framework::dataset::make("PadX", 0), + combine(framework::dataset::make("PadY", 0), + framework::dataset::make("KernelSize", 1))), + combine(framework::dataset::make("PadX", 0, 2), + combine(framework::dataset::make("PadY", 0, 2), + framework::dataset::make("KernelSize", { 3, 5 })))), + framework::dataset::make("NumKernels", { 3 }))))); +} // namespace + +TEST_SUITE(GC) +TEST_SUITE(DirectConvolutionLayerTensorShift) + +template +using GCDirectConvolutionLayerTensorShiftFixture = DirectConvolutionValidationTensorShiftFixture; + +TEST_SUITE(Float) +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(Run, GCDirectConvolutionLayerTensorShiftFixture, framework::DatasetMode::ALL, combine(data, framework::dataset::make("DataType", DataType::F16))) +{ + // Validate output + validate(GCAccessor(_target), _reference, tolerance_fp16, tolerance_num); +} +TEST_SUITE_END() + +TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(Run, GCDirectConvolutionLayerTensorShiftFixture, framework::DatasetMode::ALL, combine(data, framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(GCAccessor(_target), _reference, tolerance_fp32); +} +TEST_SUITE_END() +TEST_SUITE_END() + +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/fixtures/DirectConvolutionLayerTensorShiftFixture.h b/tests/validation/fixtures/DirectConvolutionLayerTensorShiftFixture.h new file mode 100644 index 0000000000..d810a765cb --- /dev/null +++ b/tests/validation/fixtures/DirectConvolutionLayerTensorShiftFixture.h @@ -0,0 +1,269 @@ +/* + * Copyright (c) 2017-2018 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 "arm_compute/runtime/GLES_COMPUTE/GCScheduler.h" +#include "tests/AssetsLibrary.h" +#include "tests/Globals.h" +#include "tests/IAccessor.h" +#include "tests/framework/Asserts.h" +#include "tests/framework/Fixture.h" +#include "tests/validation/Helpers.h" +#include "tests/validation/fixtures/ConvolutionLayerFixture.h" +#include "tests/validation/reference/ConvolutionLayer.h" + +#include + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template +class DirectConvolutionValidationGenericTensorShiftFixture : public framework::Fixture +{ +public: + using TBias = typename std::conditional::type, uint8_t>::value, int32_t, T>::type; + +public: + template + void setup(TensorShape input_shape, int stride_x, int stride_y, int pad_x, int pad_y, unsigned int kernel_size, unsigned int num_kernels, + DataType data_type, int fractional_bits, QuantizationInfo quantization_info) + { + _fractional_bits = fractional_bits; + _quantization_info = quantization_info; + _data_type = data_type; + + const TensorShape weights_shape(kernel_size, kernel_size, input_shape.z(), num_kernels); + const TensorShape bias_shape(num_kernels); + const PadStrideInfo info(stride_x, stride_y, pad_x, pad_y, DimensionRoundingType::FLOOR); + const TensorShape output_shape = get_output_shape(input_shape, weights_shape, info); + const DataType bias_data_type = is_data_type_quantized_asymmetric(data_type) ? DataType::S32 : data_type; + + _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, fractional_bits, quantization_info); + _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, fractional_bits, quantization_info); + } + + template + void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, + DataType data_type, int fractional_bits, QuantizationInfo quantization_info) + { + _fractional_bits = fractional_bits; + _quantization_info = quantization_info; + _data_type = data_type; + + const DataType bias_data_type = is_data_type_quantized_asymmetric(data_type) ? DataType::S32 : data_type; + + _target = compute_target(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, fractional_bits, quantization_info); + _reference = compute_reference(input_shape, weights_shape, bias_shape, output_shape, info, data_type, bias_data_type, fractional_bits, quantization_info); + } + +protected: + template + void fill(U &&tensor, int i) + { + switch(tensor.data_type()) + { + case DataType::QASYMM8: + { + std::uniform_int_distribution distribution(0, 50); + library->fill(tensor, distribution, i); + break; + } + case DataType::F16: + case DataType::F32: + { + std::uniform_real_distribution<> distribution(-1.0f, 1.0f); + library->fill(tensor, distribution, i); + break; + } + case DataType::S32: + { + std::uniform_int_distribution distribution(-5, 5); + library->fill(tensor, distribution, i); + break; + } + default: + library->fill_tensor_uniform(tensor, i); + } + } + + TensorType compute_target(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info, + DataType data_type, DataType bias_data_type, int fixed_point_position, QuantizationInfo quantization_info) + { + // Create tensors + TensorType src = create_tensor(input_shape, data_type, 1, fixed_point_position, quantization_info); + TensorType weights = create_tensor(weights_shape, data_type, 1, fixed_point_position, quantization_info); + TensorType bias = create_tensor(bias_shape, bias_data_type, 1, fixed_point_position, quantization_info); + TensorType dst = create_tensor(output_shape, data_type, 1, fixed_point_position, quantization_info); + + TensorShape output_shape1 = get_output_shape(output_shape, weights_shape, info); + TensorType dst1 = create_tensor(output_shape1, data_type, 1, fixed_point_position, quantization_info); + + // Create and configure function + FunctionType conv; + conv.configure(&src, &weights, &bias, &dst, info); + FunctionType conv1; + conv1.configure(&dst, &weights, &bias, &dst1, info); + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(bias.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(dst1.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Allocate tensors + src.allocator()->allocate(); + weights.allocator()->allocate(); + bias.allocator()->allocate(); + dst.allocator()->allocate(); + dst1.allocator()->allocate(); + + ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!weights.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!bias.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!dst1.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Fill tensors + fill(AccessorType(src), 0); + fill(AccessorType(weights), 1); + fill(AccessorType(bias), 2); + + // Compute NEConvolutionLayer function + GCScheduler::get().memory_barrier(); + conv.run(); + GCScheduler::get().memory_barrier(); + conv1.run(); + + return dst1; + } + + SimpleTensor compute_reference(const TensorShape &input_shape, const TensorShape &weights_shape, const TensorShape &bias_shape, const TensorShape &output_shape, const PadStrideInfo &info, + DataType data_type, DataType bias_data_type, int fixed_point_position, QuantizationInfo quantization_info) + { + // Create reference + SimpleTensor src{ input_shape, data_type, 1, fixed_point_position, quantization_info }; + SimpleTensor weights{ weights_shape, data_type, 1, fixed_point_position, quantization_info }; + SimpleTensor bias{ bias_shape, bias_data_type, 1, fixed_point_position, quantization_info }; + + SimpleTensor dst{ output_shape, data_type, 1, fixed_point_position, quantization_info }; + TensorShape output_shape1 = get_output_shape(output_shape, weights_shape, info); + + // Fill reference + fill(src, 0); + fill(weights, 1); + fill(bias, 2); + + dst = reference::convolution_layer(src, weights, bias, output_shape, info); + return reference::convolution_layer(dst, weights, bias, output_shape1, info); + } + + TensorType _target{}; + SimpleTensor _reference{}; + int _fractional_bits{}; + QuantizationInfo _quantization_info{}; + DataType _data_type{}; + +private: + TensorShape get_output_shape(TensorShape in_shape, TensorShape kernel_shape, const PadStrideInfo &info) + { + TensorShape out_shape(in_shape); + const std::pair scaled_dims = scaled_dimensions(in_shape.x(), + in_shape.y(), + kernel_shape.x(), + kernel_shape.y(), + info); + out_shape.set(0, scaled_dims.first); + out_shape.set(1, scaled_dims.second); + out_shape.set(2, kernel_shape[3]); + return out_shape; + } +}; + +template +class DirectConvolutionValidationTensorShiftFixture : public DirectConvolutionValidationGenericTensorShiftFixture +{ +public: + template + void setup(TensorShape input_shape, int stride_x, int stride_y, int pad_x, int pad_y, unsigned int kernel_size, unsigned int num_kernels, DataType data_type) + { + DirectConvolutionValidationGenericTensorShiftFixture::setup(input_shape, stride_x, stride_y, pad_x, pad_y, kernel_size, num_kernels, data_type, 0, + QuantizationInfo()); + } +}; + +template +class DirectConvolutionValidationFixedPointTensorShiftFixture : public DirectConvolutionValidationGenericTensorShiftFixture +{ +public: + template + void setup(TensorShape input_shape, int stride_x, int stride_y, int pad_x, int pad_y, unsigned int kernel_size, unsigned int num_kernels, DataType data_type, int fractional_bits) + { + DirectConvolutionValidationGenericTensorShiftFixture::setup(input_shape, stride_x, stride_y, pad_x, pad_y, kernel_size, num_kernels, data_type, + fractional_bits, + QuantizationInfo()); + } +}; + +template +class DirectConvolutionValidationQuantizedTensorShiftFixture : public DirectConvolutionValidationGenericTensorShiftFixture +{ +public: + template + void setup(TensorShape input_shape, int stride_x, int stride_y, int pad_x, int pad_y, unsigned int kernel_size, unsigned int num_kernels, DataType data_type, QuantizationInfo quantization_info) + { + DirectConvolutionValidationGenericTensorShiftFixture::setup(input_shape, stride_x, stride_y, pad_x, pad_y, kernel_size, num_kernels, data_type, 0, + quantization_info); + } +}; + +template +class DirectConvolutionValidationWithTensorShapesQuantizedTensorShiftFixture : public DirectConvolutionValidationGenericTensorShiftFixture +{ +public: + template + void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, + DataType data_type, QuantizationInfo quantization_info) + { + DirectConvolutionValidationGenericTensorShiftFixture::setup(input_shape, weights_shape, bias_shape, output_shape, info, data_type, 0, quantization_info); + } +}; + +template +class DirectConvolutionValidationWithTensorShapesTensorShiftFixture : public DirectConvolutionValidationGenericTensorShiftFixture +{ +public: + template + void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, + DataType data_type) + { + DirectConvolutionValidationGenericTensorShiftFixture::setup(input_shape, weights_shape, bias_shape, output_shape, info, data_type, 0, QuantizationInfo()); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute -- cgit v1.2.1