diff options
| author | Anthony Barbier <anthony.barbier@arm.com> | 2018-02-22 15:45:35 +0000 |
|---|---|---|
| committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-02-23 11:49:54 +0000 |
| commit | 06ea048f062a50404b1b3998a61a45449c2d1f0f (patch) | |
| tree | aa0dea3b0c49422538df9a5a02578b2c29e6fa67 /tests | |
| parent | 292227986edb37b01061afcad6df18ba9d6ccbeb (diff) | |
| download | armcl-06ea048f062a50404b1b3998a61a45449c2d1f0f.tar.gz armcl-06ea048f062a50404b1b3998a61a45449c2d1f0f.tar.bz2 armcl-06ea048f062a50404b1b3998a61a45449c2d1f0f.zip | |
arm_compute v18.02
Change-Id: I7207aa488e5470f235f39b6c188b4678dc38d1a6
Diffstat (limited to 'tests')
139 files changed, 5183 insertions, 1088 deletions
diff --git a/tests/AssetsLibrary.cpp b/tests/AssetsLibrary.cpp index 00993efc9..5660f3c06 100644 --- a/tests/AssetsLibrary.cpp +++ b/tests/AssetsLibrary.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -48,19 +48,26 @@ namespace template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0> void rgb_to_luminance(const RawTensor &src, RawTensor &dst) { - const size_t min_size = std::min(src.size(), dst.size()); + // Ensure in/out tensors have same image dimensions (independent of element size and number of channels) + ARM_COMPUTE_ERROR_ON_MSG(src.num_elements() != dst.num_elements(), "Input and output images must have equal dimensions"); - for(size_t i = 0, j = 0; i < min_size; i += 3, ++j) + const size_t num_elements = dst.num_elements(); + + // Currently, input is always RGB888 (3 U8 channels per element). Output can be U8, U16/S16 or U32 + // Note that src.data()[i] returns pointer to first channel of element[i], so RGB values have [0,1,2] offsets + for(size_t i = 0, j = 0; j < num_elements; i += 3, ++j) { - reinterpret_cast<T *>(dst.data())[j] = 0.2126f * src.data()[i + 0] + 0.7152f * src.data()[i + 1] + 0.0722f * src.data()[i + 2]; + reinterpret_cast<T *>(dst.data())[j] = 0.2126f * src.data()[i] + 0.7152f * src.data()[i + 1] + 0.0722f * src.data()[i + 2]; } } void extract_r_from_rgb(const RawTensor &src, RawTensor &dst) { - const size_t min_size = std::min(src.size(), dst.size()); + ARM_COMPUTE_ERROR_ON(src.size() != 3 * dst.size()); + + const size_t num_elements = dst.num_elements(); - for(size_t i = 0, j = 0; i < min_size; i += 3, ++j) + for(size_t i = 0, j = 0; j < num_elements; i += 3, ++j) { dst.data()[j] = src.data()[i]; } @@ -68,9 +75,23 @@ void extract_r_from_rgb(const RawTensor &src, RawTensor &dst) void extract_g_from_rgb(const RawTensor &src, RawTensor &dst) { - const size_t min_size = std::min(src.size(), dst.size()); + ARM_COMPUTE_ERROR_ON(src.size() != 3 * dst.size()); + + const size_t num_elements = dst.num_elements(); + + for(size_t i = 1, j = 0; j < num_elements; i += 3, ++j) + { + dst.data()[j] = src.data()[i]; + } +} + +void extract_b_from_rgb(const RawTensor &src, RawTensor &dst) +{ + ARM_COMPUTE_ERROR_ON(src.size() != 3 * dst.size()); + + const size_t num_elements = dst.num_elements(); - for(size_t i = 1, j = 0; i < min_size; i += 3, ++j) + for(size_t i = 2, j = 0; j < num_elements; i += 3, ++j) { dst.data()[j] = src.data()[i]; } @@ -320,7 +341,8 @@ const AssetsLibrary::Extractor &AssetsLibrary::get_extractor(Format format, Chan static std::map<std::pair<Format, Channel>, Extractor> extractors = { { std::make_pair(Format::RGB888, Channel::R), extract_r_from_rgb }, - { std::make_pair(Format::RGB888, Channel::G), extract_g_from_rgb } + { std::make_pair(Format::RGB888, Channel::G), extract_g_from_rgb }, + { std::make_pair(Format::RGB888, Channel::B), extract_b_from_rgb } }; const auto it = extractors.find(std::make_pair(format, channel)); @@ -385,7 +407,7 @@ const RawTensor &AssetsLibrary::find_or_create_raw_tensor(const std::string &nam } const RawTensor &src = get(name, format); - RawTensor dst(src.shape(), get_channel_format(channel)); + RawTensor dst(src.shape(), get_channel_format(channel)); (*get_extractor(format, channel))(src, dst); diff --git a/tests/AssetsLibrary.h b/tests/AssetsLibrary.h index 1fe8c6373..7e2a042e9 100644 --- a/tests/AssetsLibrary.h +++ b/tests/AssetsLibrary.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -416,21 +416,24 @@ void AssetsLibrary::fill_borders_with_garbage(T &&tensor, D &&distribution, std: template <typename T, typename D> void AssetsLibrary::fill(T &&tensor, D &&distribution, std::random_device::result_type seed_offset) const { - Window window; - for(unsigned int d = 0; d < tensor.shape().num_dimensions(); ++d) - { - window.set(d, Window::Dimension(0, tensor.shape()[d], 1)); - } + using ResultType = typename std::remove_reference<D>::type::result_type; std::mt19937 gen(_seed + seed_offset); - execute_window_loop(window, [&](const Coordinates & id) + // Iterate over all elements + for(int element_idx = 0; element_idx < tensor.num_elements(); ++element_idx) { - using ResultType = typename std::remove_reference<D>::type::result_type; - const ResultType value = distribution(gen); - void *const out_ptr = tensor(id); - store_value_with_data_type(out_ptr, value, tensor.data_type()); - }); + const Coordinates id = index2coord(tensor.shape(), element_idx); + + // Iterate over all channels + for(int channel = 0; channel < tensor.num_channels(); ++channel) + { + const ResultType value = distribution(gen); + ResultType &target_value = reinterpret_cast<ResultType *const>(tensor(id))[channel]; + + store_value_with_data_type(&target_value, value, tensor.data_type()); + } + } fill_borders_with_garbage(tensor, distribution, seed_offset); } diff --git a/tests/SimpleTensor.h b/tests/SimpleTensor.h index 902f5b51b..f3155ffea 100644 --- a/tests/SimpleTensor.h +++ b/tests/SimpleTensor.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -297,18 +297,33 @@ int SimpleTensor<T>::num_channels() const switch(_format) { case Format::U8: - case Format::S16: case Format::U16: - case Format::S32: + case Format::S16: case Format::U32: + case Format::S32: + case Format::F16: case Format::F32: return 1; + // Because the U and V channels are subsampled + // these formats appear like having only 2 channels: + case Format::YUYV422: + case Format::UYVY422: + return 2; + case Format::UV88: + return 2; case Format::RGB888: return 3; + case Format::RGBA8888: + return 4; case Format::UNKNOWN: return _num_channels; + //Doesn't make sense for planar formats: + case Format::NV12: + case Format::NV21: + case Format::IYUV: + case Format::YUV444: default: - ARM_COMPUTE_ERROR("NOT SUPPORTED!"); + return 0; } } diff --git a/tests/Utils.h b/tests/Utils.h index 750d90777..27e0397b6 100644 --- a/tests/Utils.h +++ b/tests/Utils.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -500,6 +500,22 @@ inline T create_tensor(const TensorShape &shape, Format format) return tensor; } +/** Create and initialize a multi-image of the given type. + * + * @param[in] shape Tensor shape. + * @param[in] format Format type. + * + * @return Initialized tensor of given type. + */ +template <typename T> +inline T create_multi_image(const TensorShape &shape, Format format) +{ + T multi_image; + multi_image.init(shape.x(), shape.y(), format); + + return multi_image; +} + /** Create and initialize a HOG (Histogram of Oriented Gradients) of the given type. * * @param[in] cell_size Cell size in pixels diff --git a/tests/benchmark/CL/BatchNormalizationLayer.cpp b/tests/benchmark/CL/BatchNormalizationLayer.cpp index 0fc872724..a61e7cc74 100644 --- a/tests/benchmark/CL/BatchNormalizationLayer.cpp +++ b/tests/benchmark/CL/BatchNormalizationLayer.cpp @@ -29,6 +29,7 @@ #include "tests/CL/CLAccessor.h" #include "tests/benchmark/fixtures/BatchNormalizationLayerFixture.h" #include "tests/datasets/system_tests/googlenet/inceptionv4/GoogLeNetInceptionV4BatchNormalizationLayerDataset.h" +#include "tests/datasets/system_tests/mobilenet/MobileNetBatchNormalizationLayerDataset.h" #include "tests/datasets/system_tests/yolo/v2/YOLOV2BatchNormalizationLayerDataset.h" #include "tests/framework/Macros.h" #include "tests/framework/datasets/Datasets.h" @@ -47,24 +48,41 @@ using CLBatchNormalizationLayerFixture = BatchNormalizationLayerFixture<CLTensor TEST_SUITE(CL) +REGISTER_FIXTURE_DATA_TEST_CASE(MobileNetBatchNormalizationLayer, CLBatchNormalizationLayerFixture, framework::DatasetMode::ALL, + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::MobileNetBatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))), + data_types), + framework::dataset::make("Batches", 1))); + REGISTER_FIXTURE_DATA_TEST_CASE(YOLOV2BatchNormalizationLayer, CLBatchNormalizationLayerFixture, framework::DatasetMode::ALL, - framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2BatchNormalizationLayerDataset(), + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2BatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), data_types), framework::dataset::make("Batches", 1))); REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4BatchNormalizationLayer, CLBatchNormalizationLayerFixture, framework::DatasetMode::ALL, - framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4BatchNormalizationLayerDataset(), + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4BatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), data_types), framework::dataset::make("Batches", 1))); TEST_SUITE(NIGHTLY) + +REGISTER_FIXTURE_DATA_TEST_CASE(MobileNetBatchNormalizationLayer, CLBatchNormalizationLayerFixture, framework::DatasetMode::NIGHTLY, + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::MobileNetBatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))), + data_types), + framework::dataset::make("Batches", { 4, 8 }))); + REGISTER_FIXTURE_DATA_TEST_CASE(YOLOV2BatchNormalizationLayer, CLBatchNormalizationLayerFixture, framework::DatasetMode::NIGHTLY, - framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2BatchNormalizationLayerDataset(), + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2BatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), data_types), framework::dataset::make("Batches", { 4, 8 }))); REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4BatchNormalizationLayer, CLBatchNormalizationLayerFixture, framework::DatasetMode::NIGHTLY, - framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4BatchNormalizationLayerDataset(), + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4BatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), data_types), framework::dataset::make("Batches", { 4, 8 }))); TEST_SUITE_END() diff --git a/tests/benchmark/CL/GEMM.cpp b/tests/benchmark/CL/GEMM.cpp index 615b492ee..60362b43e 100644 --- a/tests/benchmark/CL/GEMM.cpp +++ b/tests/benchmark/CL/GEMM.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -41,17 +41,26 @@ namespace test { namespace { -const auto data_types = framework::dataset::make("DataType", { DataType::F16, DataType::F32 }); +const auto data_types = framework::dataset::make("DataType", { DataType::F16, DataType::F32 }); +const auto reshape_b_only_once = framework::dataset::make("ReshapeBOnlyOnce", { false, true }); } // namespace using CLGEMMFixture = GEMMFixture<CLTensor, CLGEMM, CLAccessor>; TEST_SUITE(CL) -REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1GEMM, CLGEMMFixture, framework::DatasetMode::ALL, framework::dataset::combine(datasets::GoogLeNetInceptionV1GEMMDataset(), data_types)); -REGISTER_FIXTURE_DATA_TEST_CASE(MatrixMultiplyGEMM, CLGEMMFixture, framework::DatasetMode::ALL, framework::dataset::combine(datasets::MatrixMultiplyGEMMDataset(), data_types)); -REGISTER_FIXTURE_DATA_TEST_CASE(GoogleNetGEMM, CLGEMMFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(datasets::GoogleNetGEMMDataset(), data_types)); -REGISTER_FIXTURE_DATA_TEST_CASE(AlexNetGEMM, CLGEMMFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(datasets::AlexNetGEMMDataset(), data_types)); +REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1GEMM, CLGEMMFixture, framework::DatasetMode::ALL, + framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV1GEMMDataset(), + data_types), + reshape_b_only_once)); +REGISTER_FIXTURE_DATA_TEST_CASE(MatrixMultiplyGEMM, CLGEMMFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::MatrixMultiplyGEMMDataset(), + data_types), + reshape_b_only_once)); +REGISTER_FIXTURE_DATA_TEST_CASE(GoogleNetGEMM, CLGEMMFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::GoogleNetGEMMDataset(), data_types), + reshape_b_only_once)); +REGISTER_FIXTURE_DATA_TEST_CASE(AlexNetGEMM, CLGEMMFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::AlexNetGEMMDataset(), + data_types), + reshape_b_only_once)); TEST_SUITE_END() } // namespace test diff --git a/tests/benchmark/CL/SYSTEM/AlexNet.cpp b/tests/benchmark/CL/SYSTEM/AlexNet.cpp index 1d38c1a0c..5401cbdd0 100644 --- a/tests/benchmark/CL/SYSTEM/AlexNet.cpp +++ b/tests/benchmark/CL/SYSTEM/AlexNet.cpp @@ -56,7 +56,7 @@ using CLAlexNetFixture = AlexNetFixture<ICLTensor, CLSoftmaxLayer>; TEST_SUITE(CL) -TEST_SUITE(SYSTEM_TEST) +TEST_SUITE(SYSTEM_TESTS) REGISTER_FIXTURE_DATA_TEST_CASE(AlexNet, CLAlexNetFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::make("DataType", { DataType::F16, DataType::F32 }), diff --git a/tests/benchmark/CL/SYSTEM/LeNet5.cpp b/tests/benchmark/CL/SYSTEM/LeNet5.cpp index 21e543294..09c6f65fc 100644 --- a/tests/benchmark/CL/SYSTEM/LeNet5.cpp +++ b/tests/benchmark/CL/SYSTEM/LeNet5.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -49,7 +49,7 @@ using CLLeNet5Fixture = LeNet5Fixture<CLTensor, CLSoftmaxLayer>; TEST_SUITE(CL) -TEST_SUITE(SYSTEM_TEST) +TEST_SUITE(SYSTEM_TESTS) REGISTER_FIXTURE_DATA_TEST_CASE(LeNet5, CLLeNet5Fixture, framework::DatasetMode::ALL, framework::dataset::make("Batches", { 1, 4, 8 })); diff --git a/tests/benchmark/CL/SYSTEM/MobileNet.cpp b/tests/benchmark/CL/SYSTEM/MobileNet.cpp index 4712bc0c8..bc728d9c0 100644 --- a/tests/benchmark/CL/SYSTEM/MobileNet.cpp +++ b/tests/benchmark/CL/SYSTEM/MobileNet.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -51,7 +51,7 @@ using CLMobileNetFixture = MobileNetFixture<CLTensor, CLPoolingLayer>; TEST_SUITE(CL) -TEST_SUITE(SYSTEM_TEST) +TEST_SUITE(SYSTEM_TESTS) REGISTER_FIXTURE_DATA_TEST_CASE(MobileNet, CLMobileNetFixture, framework::DatasetMode::ALL, framework::dataset::make("Batches", { 1, 4, 8 })); diff --git a/tests/benchmark/CL/SYSTEM/MobileNetV1.cpp b/tests/benchmark/CL/SYSTEM/MobileNetV1.cpp index 851148a86..53ab1a7fc 100644 --- a/tests/benchmark/CL/SYSTEM/MobileNetV1.cpp +++ b/tests/benchmark/CL/SYSTEM/MobileNetV1.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -68,7 +68,7 @@ using CLMobileNetV1_128_Fixture = MobileNetV1Fixture<CLTensor, 128>; TEST_SUITE(CL) -TEST_SUITE(SYSTEM_TEST) +TEST_SUITE(SYSTEM_TESTS) REGISTER_FIXTURE_DATA_TEST_CASE(MobileNetV1_224, CLMobileNetV1_224_Fixture, framework::DatasetMode::ALL, framework::dataset::make("Batches", { 1, 4, 8 })); diff --git a/tests/benchmark/GLES_COMPUTE/BatchNormalizationLayer.cpp b/tests/benchmark/GLES_COMPUTE/BatchNormalizationLayer.cpp index 4464ea240..e615860ec 100755..100644 --- a/tests/benchmark/GLES_COMPUTE/BatchNormalizationLayer.cpp +++ b/tests/benchmark/GLES_COMPUTE/BatchNormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -29,6 +29,7 @@ #include "tests/GLES_COMPUTE/GCAccessor.h" #include "tests/benchmark/fixtures/BatchNormalizationLayerFixture.h" #include "tests/datasets/system_tests/googlenet/inceptionv4/GoogLeNetInceptionV4BatchNormalizationLayerDataset.h" +#include "tests/datasets/system_tests/mobilenet/MobileNetBatchNormalizationLayerDataset.h" #include "tests/datasets/system_tests/yolo/v2/YOLOV2BatchNormalizationLayerDataset.h" #include "tests/framework/Macros.h" #include "tests/framework/datasets/Datasets.h" @@ -47,24 +48,41 @@ using GCBatchNormalizationLayerFixture = BatchNormalizationLayerFixture<GCTensor TEST_SUITE(GC) +REGISTER_FIXTURE_DATA_TEST_CASE(MobileNetBatchNormalizationLayer, GCBatchNormalizationLayerFixture, framework::DatasetMode::ALL, + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::MobileNetBatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))), + data_types), + framework::dataset::make("Batches", 1))); + REGISTER_FIXTURE_DATA_TEST_CASE(YOLOV2BatchNormalizationLayer, GCBatchNormalizationLayerFixture, framework::DatasetMode::ALL, - framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2BatchNormalizationLayerDataset(), + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2BatchNormalizationLayerDataset(), framework::dataset::make("ActivationInfo", + ActivationLayerInfo())), data_types), framework::dataset::make("Batches", 1))); REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4BatchNormalizationLayer, GCBatchNormalizationLayerFixture, framework::DatasetMode::ALL, - framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4BatchNormalizationLayerDataset(), + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4BatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), data_types), framework::dataset::make("Batches", 1))); TEST_SUITE(NIGHTLY) + +REGISTER_FIXTURE_DATA_TEST_CASE(MobileNetBatchNormalizationLayer, GCBatchNormalizationLayerFixture, framework::DatasetMode::NIGHTLY, + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::MobileNetBatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))), + data_types), + framework::dataset::make("Batches", { 4, 8 }))); + REGISTER_FIXTURE_DATA_TEST_CASE(YOLOV2BatchNormalizationLayer, GCBatchNormalizationLayerFixture, framework::DatasetMode::NIGHTLY, - framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2BatchNormalizationLayerDataset(), + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2BatchNormalizationLayerDataset(), framework::dataset::make("ActivationInfo", + ActivationLayerInfo())), data_types), framework::dataset::make("Batches", { 4, 8 }))); REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4BatchNormalizationLayer, GCBatchNormalizationLayerFixture, framework::DatasetMode::NIGHTLY, - framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4BatchNormalizationLayerDataset(), + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4BatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), data_types), framework::dataset::make("Batches", { 4, 8 }))); TEST_SUITE_END() diff --git a/tests/benchmark/GLES_COMPUTE/GEMM.cpp b/tests/benchmark/GLES_COMPUTE/GEMM.cpp index 69d8e5491..97c92c679 100644 --- a/tests/benchmark/GLES_COMPUTE/GEMM.cpp +++ b/tests/benchmark/GLES_COMPUTE/GEMM.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -40,16 +40,24 @@ namespace test { namespace { -const auto data_types = framework::dataset::make("DataType", { DataType::F32 }); +const auto data_types = framework::dataset::make("DataType", { DataType::F32 }); +const auto reshape_b_only_once = framework::dataset::make("ReshapeBOnlyOnce", { false }); } // namespace using GCGEMMFixture = GEMMFixture<GCTensor, GCGEMM, GCAccessor>; TEST_SUITE(GC) -REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1GEMM, GCGEMMFixture, framework::DatasetMode::ALL, framework::dataset::combine(datasets::GoogLeNetInceptionV1GEMMDataset(), data_types)); -REGISTER_FIXTURE_DATA_TEST_CASE(MatrixMultiplyGEMM, GCGEMMFixture, framework::DatasetMode::ALL, framework::dataset::combine(datasets::MatrixMultiplyGEMMDataset(), data_types)); -REGISTER_FIXTURE_DATA_TEST_CASE(GoogleNetGEMM, GCGEMMFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(datasets::GoogleNetGEMMDataset(), data_types)); +REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1GEMM, GCGEMMFixture, framework::DatasetMode::ALL, + framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV1GEMMDataset(), + data_types), + reshape_b_only_once)); +REGISTER_FIXTURE_DATA_TEST_CASE(MatrixMultiplyGEMM, GCGEMMFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::MatrixMultiplyGEMMDataset(), + data_types), + reshape_b_only_once)); +REGISTER_FIXTURE_DATA_TEST_CASE(GoogleNetGEMM, GCGEMMFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::GoogleNetGEMMDataset(), + data_types), + reshape_b_only_once)); TEST_SUITE_END() } // namespace test diff --git a/tests/benchmark/NEON/BatchNormalizationLayer.cpp b/tests/benchmark/NEON/BatchNormalizationLayer.cpp index 3a7f2c6b6..25200374f 100644 --- a/tests/benchmark/NEON/BatchNormalizationLayer.cpp +++ b/tests/benchmark/NEON/BatchNormalizationLayer.cpp @@ -33,6 +33,7 @@ #include "tests/benchmark/fixtures/BatchNormalizationLayerFixture.h" #include "tests/datasets/system_tests/googlenet/inceptionv4/GoogLeNetInceptionV4BatchNormalizationLayerDataset.h" +#include "tests/datasets/system_tests/mobilenet/MobileNetBatchNormalizationLayerDataset.h" #include "tests/datasets/system_tests/yolo/v2/YOLOV2BatchNormalizationLayerDataset.h" namespace arm_compute @@ -52,21 +53,39 @@ using NEBatchNormalizationLayerFixture = BatchNormalizationLayerFixture<Tensor, TEST_SUITE(NEON) +REGISTER_FIXTURE_DATA_TEST_CASE(MobileNetBatchNormalizationLayer, NEBatchNormalizationLayerFixture, framework::DatasetMode::ALL, + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::MobileNetBatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), + data_types), + framework::dataset::make("Batches", 1))); REGISTER_FIXTURE_DATA_TEST_CASE(YOLOV2BatchNormalizationLayer, NEBatchNormalizationLayerFixture, framework::DatasetMode::ALL, - framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2BatchNormalizationLayerDataset(), data_types), + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2BatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), + data_types), framework::dataset::make("Batches", 1))); REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4BatchNormalizationLayer, NEBatchNormalizationLayerFixture, framework::DatasetMode::ALL, - framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4BatchNormalizationLayerDataset(), data_types), + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4BatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), + data_types), framework::dataset::make("Batches", 1))); TEST_SUITE(NIGHTLY) +REGISTER_FIXTURE_DATA_TEST_CASE(MobileNetBatchNormalizationLayer, NEBatchNormalizationLayerFixture, framework::DatasetMode::NIGHTLY, + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::MobileNetBatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), + data_types), + framework::dataset::make("Batches", { 4, 8 }))); REGISTER_FIXTURE_DATA_TEST_CASE(YOLOV2BatchNormalizationLayer, NEBatchNormalizationLayerFixture, framework::DatasetMode::NIGHTLY, - framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2BatchNormalizationLayerDataset(), data_types), + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2BatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), + data_types), framework::dataset::make("Batches", { 4, 8 }))); REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4BatchNormalizationLayer, NEBatchNormalizationLayerFixture, framework::DatasetMode::NIGHTLY, - framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4BatchNormalizationLayerDataset(), data_types), + framework::dataset::combine(framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4BatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), + data_types), framework::dataset::make("Batches", { 4, 8 }))); TEST_SUITE_END() TEST_SUITE_END() diff --git a/tests/benchmark/NEON/ConvolutionLayer.cpp b/tests/benchmark/NEON/ConvolutionLayer.cpp index b2aa92917..1be95a50c 100644 --- a/tests/benchmark/NEON/ConvolutionLayer.cpp +++ b/tests/benchmark/NEON/ConvolutionLayer.cpp @@ -47,13 +47,14 @@ namespace test namespace { #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -const auto data_types = framework::dataset::make("DataType", { DataType::F16, DataType::F32 }); -#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -const auto data_types = framework::dataset::make("DataType", { DataType::F32 }); +const auto data_types = framework::dataset::make("DataType", { DataType::F16, DataType::F32, DataType::QASYMM8 }); +#else /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ +const auto data_types = framework::dataset::make("DataType", { DataType::F32, DataType::QASYMM8 }); + #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ } // namespace -using NEConvolutionLayerFixture = ConvolutionLayerFixture<Tensor, NEConvolutionLayer, Accessor>; +using NEGEMMConvolutionLayerFixture = ConvolutionLayerFixture<Tensor, NEGEMMConvolutionLayer, Accessor>; TEST_SUITE(NEON) #if defined(__aarch64__) @@ -76,53 +77,53 @@ REGISTER_FIXTURE_DATA_TEST_CASE(SqueezeNetWinogradLayer, NEWinogradLayerFixture, framework::dataset::make("Batches", 1))); #endif /* __aarch64__ */ -REGISTER_FIXTURE_DATA_TEST_CASE(AlexNetConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::ALL, +REGISTER_FIXTURE_DATA_TEST_CASE(AlexNetConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::AlexNetConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", 1))); -REGISTER_FIXTURE_DATA_TEST_CASE(LeNet5ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::ALL, +REGISTER_FIXTURE_DATA_TEST_CASE(LeNet5ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::LeNet5ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", 1))); -REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::ALL, +REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV1ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", 1))); -REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::ALL, +REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", 1))); -REGISTER_FIXTURE_DATA_TEST_CASE(SqueezeNetConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::ALL, +REGISTER_FIXTURE_DATA_TEST_CASE(SqueezeNetConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::SqueezeNetConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", 1))); TEST_SUITE(NIGHTLY) -REGISTER_FIXTURE_DATA_TEST_CASE(AlexNetConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(AlexNetConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::AlexNetConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 4, 8 }))); -REGISTER_FIXTURE_DATA_TEST_CASE(LeNet5ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(LeNet5ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::LeNet5ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 4, 8 }))); -REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV1ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 4, 8 }))); -REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV4ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV4ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 4, 8 }))); -REGISTER_FIXTURE_DATA_TEST_CASE(SqueezeNetConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(SqueezeNetConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::SqueezeNetConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 4, 8 }))); // 8 batches use about 2GB of memory which is too much for most devices! -REGISTER_FIXTURE_DATA_TEST_CASE(VGG16ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(VGG16ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::VGG16ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 1, 2 }))); -REGISTER_FIXTURE_DATA_TEST_CASE(YOLOV2ConvolutionLayer, NEConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, +REGISTER_FIXTURE_DATA_TEST_CASE(YOLOV2ConvolutionLayer, NEGEMMConvolutionLayerFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::YOLOV2ConvolutionLayerDataset(), data_types), framework::dataset::make("Batches", { 1, 4, 8 }))); diff --git a/tests/benchmark/NEON/GEMM.cpp b/tests/benchmark/NEON/GEMM.cpp index 1d6ea8df7..473d6429f 100644 --- a/tests/benchmark/NEON/GEMM.cpp +++ b/tests/benchmark/NEON/GEMM.cpp @@ -48,15 +48,23 @@ const auto data_types = framework::dataset::make("DataType", #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ DataType::F32 }); +const auto reshape_b_only_once = framework::dataset::make("ReshapeBOnlyOnce", { false, true }); } // namespace using NEGEMMFixture = GEMMFixture<Tensor, NEGEMM, Accessor>; TEST_SUITE(NEON) -REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1GEMM, NEGEMMFixture, framework::DatasetMode::ALL, framework::dataset::combine(datasets::GoogLeNetInceptionV1GEMMDataset(), data_types)); -REGISTER_FIXTURE_DATA_TEST_CASE(MatrixMultiplyGEMM, NEGEMMFixture, framework::DatasetMode::ALL, framework::dataset::combine(datasets::MatrixMultiplyGEMMDataset(), data_types)); -REGISTER_FIXTURE_DATA_TEST_CASE(GoogleNetGEMM, NEGEMMFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(datasets::GoogleNetGEMMDataset(), data_types)); +REGISTER_FIXTURE_DATA_TEST_CASE(GoogLeNetInceptionV1GEMM, NEGEMMFixture, framework::DatasetMode::ALL, + framework::dataset::combine(framework::dataset::combine(datasets::GoogLeNetInceptionV1GEMMDataset(), + data_types), + reshape_b_only_once)); +REGISTER_FIXTURE_DATA_TEST_CASE(MatrixMultiplyGEMM, NEGEMMFixture, framework::DatasetMode::ALL, framework::dataset::combine(framework::dataset::combine(datasets::MatrixMultiplyGEMMDataset(), + data_types), + reshape_b_only_once)); +REGISTER_FIXTURE_DATA_TEST_CASE(GoogleNetGEMM, NEGEMMFixture, framework::DatasetMode::NIGHTLY, framework::dataset::combine(framework::dataset::combine(datasets::GoogleNetGEMMDataset(), + data_types), + reshape_b_only_once)); TEST_SUITE_END() } // namespace test diff --git a/tests/benchmark/NEON/SYSTEM/AlexNet.cpp b/tests/benchmark/NEON/SYSTEM/AlexNet.cpp index 2da61802c..e28f96777 100644 --- a/tests/benchmark/NEON/SYSTEM/AlexNet.cpp +++ b/tests/benchmark/NEON/SYSTEM/AlexNet.cpp @@ -65,7 +65,7 @@ using NEAlexNetFixture = AlexNetFixture<ITensor, NESoftmaxLayer>; TEST_SUITE(NEON) -TEST_SUITE(SYSTEM_TEST) +TEST_SUITE(SYSTEM_TESTS) REGISTER_FIXTURE_DATA_TEST_CASE(AlexNet, NEAlexNetFixture, framework::DatasetMode::ALL, framework::dataset::combine(alex_net_data_types, diff --git a/tests/benchmark/NEON/SYSTEM/LeNet5.cpp b/tests/benchmark/NEON/SYSTEM/LeNet5.cpp index 5075da50c..d8ec5754e 100644 --- a/tests/benchmark/NEON/SYSTEM/LeNet5.cpp +++ b/tests/benchmark/NEON/SYSTEM/LeNet5.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -49,7 +49,7 @@ using NELeNet5Fixture = LeNet5Fixture<Tensor, NESoftmaxLayer>; TEST_SUITE(NEON) -TEST_SUITE(SYSTEM_TEST) +TEST_SUITE(SYSTEM_TESTS) REGISTER_FIXTURE_DATA_TEST_CASE(LeNet5, NELeNet5Fixture, framework::DatasetMode::ALL, framework::dataset::make("Batches", { 1, 4, 8 })); diff --git a/tests/benchmark/fixtures/ActivationLayerFixture.h b/tests/benchmark/fixtures/ActivationLayerFixture.h index 8558527ad..cb8fa7f08 100644 --- a/tests/benchmark/fixtures/ActivationLayerFixture.h +++ b/tests/benchmark/fixtures/ActivationLayerFixture.h @@ -46,9 +46,10 @@ public: shape.set(shape.num_dimensions(), batches); // Create tensors - const int fixed_point_position = 4; - src = create_tensor<TensorType>(shape, data_type, 1, fixed_point_position); - dst = create_tensor<TensorType>(shape, data_type, 1, fixed_point_position); + const int fixed_point_position = 4; + const QuantizationInfo q_info(0.5f, -10); + src = create_tensor<TensorType>(shape, data_type, 1, fixed_point_position, q_info); + dst = create_tensor<TensorType>(shape, data_type, 1, fixed_point_position, q_info); // Create and configure function act_layer.configure(&src, &dst, info); diff --git a/tests/benchmark/fixtures/BatchNormalizationLayerFixture.h b/tests/benchmark/fixtures/BatchNormalizationLayerFixture.h index 38a326320..a031ec6d9 100644 --- a/tests/benchmark/fixtures/BatchNormalizationLayerFixture.h +++ b/tests/benchmark/fixtures/BatchNormalizationLayerFixture.h @@ -40,7 +40,7 @@ class BatchNormalizationLayerFixture : public framework::Fixture { public: template <typename...> - void setup(TensorShape tensor_shape, TensorShape param_shape, float epsilon, DataType data_type, int batches) + void setup(TensorShape tensor_shape, TensorShape param_shape, float epsilon, ActivationLayerInfo act_info, DataType data_type, int batches) { // Set batched in source and destination shapes const unsigned int fixed_point_position = 4; @@ -55,7 +55,7 @@ public: gamma = create_tensor<TensorType>(param_shape, data_type, 1, fixed_point_position); // Create and configure function - batch_norm_layer.configure(&src, &dst, &mean, &variance, &beta, &gamma, epsilon); + batch_norm_layer.configure(&src, &dst, &mean, &variance, &beta, &gamma, epsilon, act_info); // Allocate tensors src.allocator()->allocate(); diff --git a/tests/benchmark/fixtures/ConvolutionLayerFixture.h b/tests/benchmark/fixtures/ConvolutionLayerFixture.h index 0d2c3fd95..7dc7de0dc 100644 --- a/tests/benchmark/fixtures/ConvolutionLayerFixture.h +++ b/tests/benchmark/fixtures/ConvolutionLayerFixture.h @@ -46,11 +46,12 @@ public: const unsigned int fixed_point_position = 4; src_shape.set(3 /* batch */, batches); dst_shape.set(3 /* batch */, batches); + DataType bias_data_type = is_data_type_quantized_asymmetric(data_type) ? DataType::S32 : data_type; // Create tensors src = create_tensor<TensorType>(src_shape, data_type, 1, fixed_point_position); weights = create_tensor<TensorType>(weights_shape, data_type, 1, fixed_point_position); - biases = create_tensor<TensorType>(biases_shape, data_type, 1, fixed_point_position); + biases = create_tensor<TensorType>(biases_shape, bias_data_type, 1, fixed_point_position); dst = create_tensor<TensorType>(dst_shape, data_type, 1, fixed_point_position); // Create and configure function diff --git a/tests/benchmark/fixtures/DepthwiseConvolutionLayerFixture.h b/tests/benchmark/fixtures/DepthwiseConvolutionLayerFixture.h index 8283b4d51..a156f4bc6 100644 --- a/tests/benchmark/fixtures/DepthwiseConvolutionLayerFixture.h +++ b/tests/benchmark/fixtures/DepthwiseConvolutionLayerFixture.h @@ -48,10 +48,10 @@ public: dst_shape.set(3 /* batch */, batches); // Create tensors - src = create_tensor<TensorType>(src_shape, data_type, 1, fixed_point_position); - weights = create_tensor<TensorType>(weights_shape, data_type, 1, fixed_point_position); + src = create_tensor<TensorType>(src_shape, data_type, 1, fixed_point_position, QuantizationInfo(0.5f, 10)); + weights = create_tensor<TensorType>(weights_shape, data_type, 1, fixed_point_position, QuantizationInfo(0.5f, 10)); biases = create_tensor<TensorType>(TensorShape(weights_shape[2]), is_data_type_quantized_asymmetric(data_type) ? DataType::S32 : data_type, 1, fixed_point_position); - dst = create_tensor<TensorType>(dst_shape, data_type, 1, fixed_point_position); + dst = create_tensor<TensorType>(dst_shape, data_type, 1, fixed_point_position, QuantizationInfo(0.5f, 10)); // Create and configure function depth_conv.configure(&src, &weights, &biases, &dst, info); diff --git a/tests/benchmark/fixtures/GEMMFixture.h b/tests/benchmark/fixtures/GEMMFixture.h index e958d4fdd..3a308d9ad 100644 --- a/tests/benchmark/fixtures/GEMMFixture.h +++ b/tests/benchmark/fixtures/GEMMFixture.h @@ -40,7 +40,7 @@ class GEMMFixture : public framework::Fixture { public: template <typename...> - void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, TensorShape shape_dst, float alpha, float beta, DataType data_type) + void setup(TensorShape shape_a, TensorShape shape_b, TensorShape shape_c, TensorShape shape_dst, float alpha, float beta, DataType data_type, bool reshape_b_only_on_first_run) { constexpr int fixed_point_position = 4; @@ -51,7 +51,7 @@ public: dst = create_tensor<TensorType>(shape_dst, data_type, 1, fixed_point_position); // Create and configure function - gemm.configure(&a, &b, &c, &dst, alpha, beta); + gemm.configure(&a, &b, &c, &dst, alpha, beta, GEMMInfo(false, false, reshape_b_only_on_first_run)); // Allocate tensors a.allocator()->allocate(); diff --git a/tests/datasets/DepthwiseConvolutionLayerDataset.h b/tests/datasets/DepthwiseConvolutionLayerDataset.h index c5a9f96a2..b8a16a7c3 100644 --- a/tests/datasets/DepthwiseConvolutionLayerDataset.h +++ b/tests/datasets/DepthwiseConvolutionLayerDataset.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -118,7 +118,7 @@ class SmallDepthwiseConvolutionLayerDataset final : public DepthwiseConvolutionL public: SmallDepthwiseConvolutionLayerDataset() { - add_config(TensorShape(7U, 7U, 3U), TensorShape(3U, 3U, 3U), TensorShape(5U, 5U, 3U), PadStrideInfo(1, 1, 0, 0)); + add_config(TensorShape(7U, 7U, 1U), TensorShape(3U, 3U, 1U), TensorShape(5U, 5U, 1U), PadStrideInfo(1, 1, 0, 0)); add_config(TensorShape(23U, 27U, 5U), TensorShape(3U, 5U, 5U), TensorShape(11U, 23U, 5U), PadStrideInfo(2, 1, 0, 0)); add_config(TensorShape(33U, 27U, 7U), TensorShape(7U, 3U, 7U), TensorShape(10U, 13U, 7U), PadStrideInfo(3, 2, 1, 0)); add_config(TensorShape(33U, 27U, 11U), TensorShape(3U, 3U, 11U), TensorShape(31U, 14U, 11U), PadStrideInfo(1, 2, 0, 1)); @@ -154,10 +154,14 @@ class SmallDepthwiseConvolutionLayerDataset3x3 final : public DepthwiseConvoluti public: SmallDepthwiseConvolutionLayerDataset3x3() { + add_config(TensorShape(3U, 3U, 2U), TensorShape(3U, 3U, 2U), TensorShape(1U, 1U, 2U), PadStrideInfo(1, 1, 0, 0)); add_config(TensorShape(7U, 7U, 3U, 2U), TensorShape(3U, 3U, 3U), TensorShape(5U, 5U, 3U, 2U), PadStrideInfo(1, 1, 0, 0)); add_config(TensorShape(33U, 27U, 11U), TensorShape(3U, 3U, 11U), TensorShape(11U, 14U, 11U), PadStrideInfo(3, 2, 1, 1)); add_config(TensorShape(21U, 31U, 9U, 4U), TensorShape(3U, 3U, 9U), TensorShape(21U, 15U, 9U, 4U), PadStrideInfo(1, 2, 1, 0)); add_config(TensorShape(33U, 27U, 11U, 3U), TensorShape(3U, 3U, 11U), TensorShape(31U, 14U, 11U, 3U), PadStrideInfo(1, 2, 0, 1)); + // Asymmetric padding + add_config(TensorShape(33U, 27U, 11U), TensorShape(3U, 3U, 11U), TensorShape(16U, 13U, 11U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)); + add_config(TensorShape(33U, 27U, 11U), TensorShape(3U, 3U, 11U), TensorShape(18U, 14U, 11U), PadStrideInfo(2, 2, 3, 1, 2, 1, DimensionRoundingType::FLOOR)); } }; @@ -174,6 +178,23 @@ public: add_config(TensorShape(177U, 311U, 22U), TensorShape(3U, 3U, 22U), TensorShape(89U, 311U, 22U), PadStrideInfo(2, 1, 1, 1)); } }; +class OptimizedDepthwiseConvolutionLayerDataset3x3 final : public DepthwiseConvolutionLayerDataset +{ +public: + OptimizedDepthwiseConvolutionLayerDataset3x3() + { + // Stride 1 + add_config(TensorShape(7U, 7U, 16U), TensorShape(3U, 3U, 16U), TensorShape(5U, 5U, 16U), PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL)); + add_config(TensorShape(7U, 7U, 16U), TensorShape(3U, 3U, 16U), TensorShape(7U, 7U, 16U), PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL)); + add_config(TensorShape(28U, 28U, 16U), TensorShape(3U, 3U, 16U), TensorShape(26U, 26U, 16U), PadStrideInfo(1, 1, 0, 0, DimensionRoundingType::CEIL)); + add_config(TensorShape(28U, 28U, 16U), TensorShape(3U, 3U, 16U), TensorShape(28U, 28U, 16U), PadStrideInfo(1, 1, 1, 1, DimensionRoundingType::CEIL)); + // Stride 2 + add_config(TensorShape(7U, 7U, 32U), TensorShape(3U, 3U, 32U), TensorShape(3U, 3U, 32U), PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)); + add_config(TensorShape(7U, 7U, 32U), TensorShape(3U, 3U, 32U), TensorShape(4U, 4U, 32U), PadStrideInfo(2, 2, 1, 1, 1, 1, DimensionRoundingType::CEIL)); + add_config(TensorShape(8U, 8U, 32U), TensorShape(3U, 3U, 32U), TensorShape(3U, 3U, 32U), PadStrideInfo(2, 2, 0, 0, DimensionRoundingType::CEIL)); + add_config(TensorShape(8U, 8U, 32U), TensorShape(3U, 3U, 32U), TensorShape(4U, 4U, 32U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::CEIL)); + } +}; } // namespace datasets } // namespace test } // namespace arm_compute diff --git a/tests/datasets/FullyConnectedLayerDataset.h b/tests/datasets/FullyConnectedLayerDataset.h index b2008d604..085e9c76b 100644 --- a/tests/datasets/FullyConnectedLayerDataset.h +++ b/tests/datasets/FullyConnectedLayerDataset.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -114,12 +114,29 @@ private: std::vector<TensorShape> _dst_shapes{}; }; +class TinyFullyConnectedLayerDataset final : public FullyConnectedLayerDataset +{ +public: + TinyFullyConnectedLayerDataset() + { + // Conv -> FC + add_config(TensorShape(8U, 1U, 1U), TensorShape(8U, 16U), TensorShape(16U), TensorShape(16U)); + // Conv -> FC + add_config(TensorShape(1U, 1U, 1U, 3U), TensorShape(1U, 10U), TensorShape(10U), TensorShape(10U, 3U)); + // FC -> FC + add_config(TensorShape(1U), TensorShape(1U, 10U), TensorShape(10U), TensorShape(10U)); + // FC -> FC (batched) + add_config(TensorShape(1U, 3U), TensorShape(1U, 10U), TensorShape(10U), TensorShape(10U, 3U)); + } +}; class SmallFullyConnectedLayerDataset final : public FullyConnectedLayerDataset { public: SmallFullyConnectedLayerDataset() { // Conv -> FC + add_config(TensorShape(8U, 1U, 1U), TensorShape(8U, 16U), TensorShape(16U), TensorShape(16U)); + // Conv -> FC add_config(TensorShape(1U, 1U, 1U, 3U), TensorShape(1U, 10U), TensorShape(10U), TensorShape(10U, 3U)); // Conv -> FC add_config(TensorShape(9U, 5U, 7U), TensorShape(315U, 271U), TensorShape(271U), TensorShape(271U)); diff --git a/tests/datasets/PoolingLayerDataset.h b/tests/datasets/PoolingLayerDataset.h index 56ec3b87d..53e392fe6 100644 --- a/tests/datasets/PoolingLayerDataset.h +++ b/tests/datasets/PoolingLayerDataset.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -105,6 +105,19 @@ private: std::vector<TensorShape> _dst_shapes{}; std::vector<PoolingLayerInfo> _infos{}; }; + +// Special pooling dataset +class PoolingLayerDatasetSpecial final : public PoolingLayerDataset +{ +public: + PoolingLayerDatasetSpecial() + { + // Special cases + add_config(TensorShape(60U, 52U, 3U, 2U), TensorShape(13U, 11U, 32U), PoolingLayerInfo(PoolingType::AVG, Size2D(100, 100), PadStrideInfo(5, 5, 50, 50), true)); + // Asymmetric padding + add_config(TensorShape(112U, 112U, 32U), TensorShape(56U, 56U, 32U), PoolingLayerInfo(PoolingType::MAX, 3, PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR))); + } +}; } // namespace datasets } // namespace test } // namespace arm_compute diff --git a/tests/datasets/ShapeDatasets.h b/tests/datasets/ShapeDatasets.h index a5e03c737..4b563708e 100644 --- a/tests/datasets/ShapeDatasets.h +++ b/tests/datasets/ShapeDatasets.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -51,6 +51,19 @@ public: } }; +/** Data set containing tiny 2D tensor shapes. */ +class Tiny2DShapes final : public ShapeDataset +{ +public: + Tiny2DShapes() + : ShapeDataset("Shape", + { + TensorShape{ 7U, 7U }, + TensorShape{ 11U, 13U }, + }) + { + } +}; /** Data set containing small 2D tensor shapes. */ class Small2DShapes final : public ShapeDataset { @@ -66,6 +79,20 @@ public: } }; +/** Data set containing tiny 3D tensor shapes. */ +class Tiny3DShapes final : public ShapeDataset +{ +public: + Tiny3DShapes() + : ShapeDataset("Shape", + { + TensorShape{ 7U, 7U, 5U }, + TensorShape{ 23U, 13U, 9U }, + }) + { + } +}; + /** Data set containing small 3D tensor shapes. */ class Small3DShapes final : public ShapeDataset { @@ -73,7 +100,8 @@ public: Small3DShapes() : ShapeDataset("Shape", { - TensorShape{ 7U, 7U, 5U }, + TensorShape{ 1U, 7U, 7U }, + TensorShape{ 7U, 7U, 5U }, TensorShape{ 27U, 13U, 37U }, TensorShape{ 128U, 64U, 21U } }) @@ -81,6 +109,19 @@ public: } }; +/** Data set containing tiny 4D tensor shapes. */ +class Tiny4DShapes final : public ShapeDataset +{ +public: + Tiny4DShapes() + : ShapeDataset("Shape", + { + TensorShape{ 7U, 7U, 5U, 3U }, + TensorShape{ 17U, 13U, 7U, 2U }, + }) + { + } +}; /** Data set containing small 4D tensor shapes. */ class Small4DShapes final : public ShapeDataset { @@ -88,7 +129,8 @@ public: Small4DShapes() : ShapeDataset("Shape", { - TensorShape{ 7U, 7U, 5U, 3U }, + TensorShape{ 1U, 7U, 1U, 3U }, + TensorShape{ 7U, 7U, 5U, 3U }, TensorShape{ 27U, 13U, 37U, 2U }, TensorShape{ 128U, 64U, 21U, 3U } }) @@ -97,6 +139,20 @@ public: }; /** Data set containing small tensor shapes. */ +class TinyShapes final : public ShapeDataset +{ +public: + TinyShapes() + : ShapeDataset("Shape", + { + // Batch size 1 + TensorShape{ 9U, 9U }, + TensorShape{ 27U, 13U, 2U }, + }) + { + } +}; +/** Data set containing small tensor shapes. */ class SmallShapes final : public ShapeDataset { public: @@ -117,6 +173,34 @@ public: } }; +/** Data set containing pairs of small tensor shapes that are broadcast compatible. */ +class SmallShapesBroadcast final : public framework::dataset::ZipDataset<ShapeDataset, ShapeDataset> +{ +public: + SmallShapesBroadcast() + : ZipDataset<ShapeDataset, ShapeDataset>( + ShapeDataset("Shape0", + { + TensorShape{ 9U, 9U }, + TensorShape{ 27U, 13U, 2U }, + TensorShape{ 128U, 1U, 5U, 3U }, + TensorShape{ 9U, 9U, 3U, 4U }, + TensorShape{ 27U, 13U, 2U, 4U }, + TensorShape{ 1U, 1U, 1U, 5U } + }), + ShapeDataset("Shape1", + { + TensorShape{ 9U, 1U, 2U }, + TensorShape{ 1U, 13U, 2U }, + TensorShape{ 128U, 64U, 1U, 3U }, + TensorShape{ 9U, 1U, 3U }, + TensorShape{ 1U }, + TensorShape{ 9U, 9U, 3U, 5U } + })) + { + } +}; + /** Data set containing medium tensor shapes. */ class MediumShapes final : public ShapeDataset { @@ -172,6 +256,30 @@ public: } }; +/** Data set containing pairs of large tensor shapes that are broadcast compatible. */ +class LargeShapesBroadcast final : public framework::dataset::ZipDataset<ShapeDataset, ShapeDataset> +{ +public: + LargeShapesBroadcast() + : ZipDataset<ShapeDataset, ShapeDataset>( + ShapeDataset("Shape0", + { + TensorShape{ 1921U, 541U }, + TensorShape{ 1U, 485U, 2U, 3U }, + TensorShape{ 4159U, 1U }, + TensorShape{ 799U } + }), + ShapeDataset("Shape1", + { + TensorShape{ 1921U, 1U, 2U }, + TensorShape{ 641U, 1U, 2U, 3U }, + TensorShape{ 1U, 127U, 25U }, + TensorShape{ 799U, 595U, 1U, 4U } + })) + { + } +}; + /** Data set containing large 1D tensor shapes. */ class Large1DShapes final : public ShapeDataset { @@ -239,7 +347,7 @@ public: SmallDeconvolutionShapes() : ShapeDataset("InputShape", { - TensorShape{ 4U, 3U, 3U, 2U }, + TensorShape{ 5U, 4U, 3U, 2U }, TensorShape{ 5U, 5U, 3U }, TensorShape{ 11U, 13U, 4U, 3U } }) @@ -247,6 +355,20 @@ public: } }; +/** Data set containing tiny tensor shapes for direct convolution. */ +class TinyDirectConvolutionShapes final : public ShapeDataset +{ +public: + TinyDirectConvolutionShapes() + : ShapeDataset("InputShape", + { + // Batch size 1 + TensorShape{ 11U, 13U, 3U }, + TensorShape{ 7U, 27U, 3U } + }) + { + } +}; /** Data set containing small tensor shapes for direct convolution. */ class SmallDirectConvolutionShapes final : public ShapeDataset { @@ -261,7 +383,27 @@ public: TensorShape{ 32U, 37U, 3U, 4U }, // Batch size 8 TensorShape{ 32U, 37U, 3U, 8U }, - TensorShape{ 33U, 35U, 8U, 8U }, + TensorShape{ 33U, 35U, 8U, 8U } + }) + { + } +}; + +/** 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 } }) @@ -302,6 +444,19 @@ public: { } }; +/** Data set containing tiny softmax layer shapes. */ +class SoftmaxLayerTinyShapes final : public ShapeDataset +{ +public: + SoftmaxLayerTinyShapes() + : ShapeDataset("Shape", + { + TensorShape{ 9U, 9U }, + TensorShape{ 128U, 10U, 2U }, + }) + { + } +}; /** Data set containing small softmax layer shapes. */ class SoftmaxLayerSmallShapes final : public ShapeDataset diff --git a/tests/datasets/SmallConvolutionLayerDataset.h b/tests/datasets/SmallConvolutionLayerDataset.h index 019dfe1fc..adb61de8e 100644 --- a/tests/datasets/SmallConvolutionLayerDataset.h +++ b/tests/datasets/SmallConvolutionLayerDataset.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -42,10 +42,16 @@ class SmallWinogradLayerDataset final : public ConvolutionLayerDataset public: SmallWinogradLayerDataset() { + // Kernel size 3 // Batch size 1 add_config(TensorShape(8U, 8U, 2U), TensorShape(3U, 3U, 2U, 1U), TensorShape(1U), TensorShape(6U, 6U, 1U), PadStrideInfo(1, 1, 0, 0)); // Batch size 4 add_config(TensorShape(23U, 27U, 5U, 4U), TensorShape(3U, 3U, 5U, 21U), TensorShape(21U), TensorShape(21U, 25U, 21U, 4U), PadStrideInfo(1, 1, 0, 0)); + add_config(TensorShape(8U, 8U, 2U), TensorShape(3U, 3U, 2U, 1U), TensorShape(1U), TensorShape(8U, 8U, 1U), PadStrideInfo(1, 1, 1, 1)); + + // Kernel size 5 + add_config(TensorShape(8U, 8U, 2U), TensorShape(5U, 5U, 2U, 1U), TensorShape(1U), TensorShape(4U, 4U, 1U), PadStrideInfo(1, 1, 0, 0)); + add_config(TensorShape(8U, 8U, 2U), TensorShape(5U, 5U, 2U), TensorShape(1U), TensorShape(8U, 8U, 1U), PadStrideInfo(1, 1, 2, 2)); } }; diff --git a/tests/datasets/TinyConvolutionLayerDataset.h b/tests/datasets/TinyConvolutionLayerDataset.h new file mode 100644 index 000000000..178ebd33f --- /dev/null +++ b/tests/datasets/TinyConvolutionLayerDataset.h @@ -0,0 +1,57 @@ +/* + * Copyright (c) 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. + */ +#ifndef ARM_COMPUTE_TEST_TINY_CONVOLUTION_LAYER_DATASET +#define ARM_COMPUTE_TEST_TINY_CONVOLUTION_LAYER_DATASET + +#include "tests/datasets/ConvolutionLayerDataset.h" + +#include "utils/TypePrinter.h" + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" + +namespace arm_compute +{ +namespace test +{ +namespace datasets +{ +class TinyConvolutionLayerDataset final : public ConvolutionLayerDataset +{ +public: + TinyConvolutionLayerDataset() + { + // Batch size 1 + add_config(TensorShape(23U, 27U, 5U), TensorShape(3U, 3U, 5U, 21U), TensorShape(21U), TensorShape(11U, 25U, 21U), PadStrideInfo(2, 1, 0, 0)); + add_config(TensorShape(33U, 27U, 7U), TensorShape(5U, 5U, 7U, 16U), TensorShape(16U), TensorShape(11U, 12U, 16U), PadStrideInfo(3, 2, 1, 0)); + // Batch size 4 + add_config(TensorShape(17U, 31U, 2U, 4U), TensorShape(5U, 5U, 2U, 19U), TensorShape(19U), TensorShape(15U, 15U, 19U, 4U), PadStrideInfo(1, 2, 1, 1)); + // FC convolution + add_config(TensorShape(1U, 1U, 1024U), TensorShape(1U, 1U, 1024U, 1001U), TensorShape(1001U), TensorShape(1U, 1U, 1001U), PadStrideInfo(1, 1, 0, 0)); + } +}; +} // namespace datasets +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_TINY_CONVOLUTION_LAYER_DATASET */ diff --git a/tests/datasets/TinyGEMMDataset.h b/tests/datasets/TinyGEMMDataset.h new file mode 100644 index 000000000..83af8f009 --- /dev/null +++ b/tests/datasets/TinyGEMMDataset.h @@ -0,0 +1,52 @@ +/* + * 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. + */ +#ifndef ARM_COMPUTE_TEST_TINY_GEMM_DATASET +#define ARM_COMPUTE_TEST_TINY_GEMM_DATASET + +#include "tests/datasets/GEMMDataset.h" + +#include "utils/TypePrinter.h" + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" + +namespace arm_compute +{ +namespace test +{ +namespace datasets +{ +class TinyGEMMDataset final : public GEMMDataset +{ +public: + TinyGEMMDataset() + { + add_config(TensorShape(21U, 13U), TensorShape(33U, 21U), TensorShape(33U, 13U), TensorShape(33U, 13U), 1.0f, 0.0f); + add_config(TensorShape(31U, 1U), TensorShape(23U, 31U), TensorShape(23U, 1U), TensorShape(23U, 1U), 1.0f, 0.0f); + } +}; +} // namespace datasets +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_TINY_GEMM_DATASET */ diff --git a/tests/datasets/system_tests/mobilenet/MobileNetBatchNormalizationLayerDataset.h b/tests/datasets/system_tests/mobilenet/MobileNetBatchNormalizationLayerDataset.h new file mode 100644 index 000000000..d09ff02ef --- /dev/null +++ b/tests/datasets/system_tests/mobilenet/MobileNetBatchNormalizationLayerDataset.h @@ -0,0 +1,68 @@ +/* + * 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. + */ +#ifndef ARM_COMPUTE_TEST_MOBILENET_BATCHNORMALIZATION_LAYER_DATASET +#define ARM_COMPUTE_TEST_MOBILENET_BATCHNORMALIZATION_LAYER_DATASET + +#include "tests/datasets/BatchNormalizationLayerDataset.h" + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.h" + +namespace arm_compute +{ +namespace test +{ +namespace datasets +{ +class MobileNetBatchNormalizationLayerDataset final : public BatchNormalizationLayerDataset +{ +public: + MobileNetBatchNormalizationLayerDataset() + { + // conv1_bn, dwc0_bn + add_config(TensorShape(112U, 112U, 32U), TensorShape(32U), 0.001f); + // pwc0_bn + add_config(TensorShape(112U, 112U, 64U), TensorShape(64U), 0.001f); + // dwc1_bn + add_config(TensorShape(56U, 56U, 64U), TensorShape(64U), 0.001f); + // dwc2_bn, pwc1_bn, pwc2_bn + add_config(TensorShape(56U, 56U, 128U), TensorShape(128U), 0.001f); + // dwc3_bn + add_config(TensorShape(28U, 28U, 128U), TensorShape(128U), 0.001f); + // dwc4_bn, pwc3_bn, pwc4_bn + add_config(TensorShape(28U, 28U, 256U), TensorShape(256U), 0.001f); + // dwc5_bn + add_config(TensorShape(14U, 14U, 256U), TensorShape(256U), 0.001f); + // dwc6_bn, dwc7_bn, dwc8_bn, dwc9_bn, dwc10_bn, pwc5_bn, pwc6_bn, pwc7_bn, pwc8_bn, pwc9_bn, pwc10_bn + add_config(TensorShape(14U, 14U, 512U), TensorShape(512U), 0.001f); + // dwc11_bn + add_config(TensorShape(7U, 7U, 512U), TensorShape(512U), 0.001f); + // dwc12_bn, pwc11_bn, pwc12_bn + add_config(TensorShape(7U, 7U, 1024U), TensorShape(1024U), 0.001f); + } +}; +} // namespace datasets +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_MOBILENET_BATCHNORMALIZATION_LAYER_DATASET */ diff --git a/tests/datasets/system_tests/mobilenet/MobileNetDepthwiseConvolutionLayerDataset.h b/tests/datasets/system_tests/mobilenet/MobileNetDepthwiseConvolutionLayerDataset.h index bd44600e4..c4c199a23 100644 --- a/tests/datasets/system_tests/mobilenet/MobileNetDepthwiseConvolutionLayerDataset.h +++ b/tests/datasets/system_tests/mobilenet/MobileNetDepthwiseConvolutionLayerDataset.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -42,8 +42,9 @@ class MobileNetDepthwiseConvolutionLayerDataset final : public DepthwiseConvolut public: MobileNetDepthwiseConvolutionLayerDataset() { - add_config(TensorShape(7U, 7U, 1024U), TensorShape(3U, 3U, 1024U), TensorShape(3U, 3U, 1024U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)); + add_config(TensorShape(7U, 7U, 1024U), TensorShape(3U, 3U, 1024U), TensorShape(7U, 7U, 1024U), PadStrideInfo(1, 1, 1, 1)); add_config(TensorShape(14U, 14U, 512U), TensorShape(3U, 3U, 512U), TensorShape(7U, 7U, 512U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)); + add_config(TensorShape(14U, 14U, 512U), TensorShape(3U, 3U, 512U), TensorShape(14U, 14U, 512U), PadStrideInfo(1, 1, 1, 1)); add_config(TensorShape(28U, 28U, 256U), TensorShape(3U, 3U, 256U), TensorShape(14U, 14U, 256U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)); add_config(TensorShape(28U, 28U, 256U), TensorShape(3U, 3U, 256U), TensorShape(28U, 28U, 256U), PadStrideInfo(1, 1, 1, 1)); add_config(TensorShape(56U, 56U, 128U), TensorShape(3U, 3U, 128U), TensorShape(28U, 28U, 128U), PadStrideInfo(2, 2, 0, 1, 0, 1, DimensionRoundingType::FLOOR)); diff --git a/tests/framework/Framework.cpp b/tests/framework/Framework.cpp index 79a77d9e5..3091b6677 100644 --- a/tests/framework/Framework.cpp +++ b/tests/framework/Framework.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -41,6 +41,9 @@ Framework::Framework() _available_instruments.emplace(std::pair<InstrumentType, ScaleFactor>(InstrumentType::WALL_CLOCK_TIMER, ScaleFactor::NONE), Instrument::make_instrument<WallClockTimer, ScaleFactor::NONE>); _available_instruments.emplace(std::pair<InstrumentType, ScaleFactor>(InstrumentType::WALL_CLOCK_TIMER, ScaleFactor::TIME_MS), Instrument::make_instrument<WallClockTimer, ScaleFactor::TIME_MS>); _available_instruments.emplace(std::pair<InstrumentType, ScaleFactor>(InstrumentType::WALL_CLOCK_TIMER, ScaleFactor::TIME_S), Instrument::make_instrument<WallClockTimer, ScaleFactor::TIME_S>); + _available_instruments.emplace(std::pair<InstrumentType, ScaleFactor>(InstrumentType::SCHEDULER_TIMER, ScaleFactor::NONE), Instrument::make_instrument<SchedulerTimer, ScaleFactor::NONE>); + _available_instruments.emplace(std::pair<InstrumentType, ScaleFactor>(InstrumentType::SCHEDULER_TIMER, ScaleFactor::TIME_MS), Instrument::make_instrument<SchedulerTimer, ScaleFactor::TIME_MS>); + _available_instruments.emplace(std::pair<InstrumentType, ScaleFactor>(InstrumentType::SCHEDULER_TIMER, ScaleFactor::TIME_S), Instrument::make_instrument<SchedulerTimer, ScaleFactor::TIME_S>); #ifdef PMU_ENABLED _available_instruments.emplace(std::pair<InstrumentType, ScaleFactor>(InstrumentType::PMU, ScaleFactor::NONE), Instrument::make_instrument<PMUCounter, ScaleFactor::NONE>); _available_instruments.emplace(std::pair<InstrumentType, ScaleFactor>(InstrumentType::PMU, ScaleFactor::SCALE_1K), Instrument::make_instrument<PMUCounter, ScaleFactor::SCALE_1K>); diff --git a/tests/framework/command_line/CommonOptions.cpp b/tests/framework/command_line/CommonOptions.cpp index 631981be3..e0b93f683 100644 --- a/tests/framework/command_line/CommonOptions.cpp +++ b/tests/framework/command_line/CommonOptions.cpp @@ -26,6 +26,7 @@ #include "../Framework.h" #include "../printers/Printers.h" #include "CommandLineParser.h" +#include <unistd.h> namespace arm_compute { @@ -42,7 +43,7 @@ CommonOptions::CommonOptions(CommandLineParser &parser) log_file(parser.add_option<SimpleOption<std::string>>("log-file")), log_level(), throw_errors(parser.add_option<ToggleOption>("throw-errors")), - color_output(parser.add_option<ToggleOption>("color-output", true)), + color_output(parser.add_option<ToggleOption>("color-output", isatty(STDOUT_FILENO))), // Only enable colors by default if we're running in a terminal pretty_console(parser.add_option<ToggleOption>("pretty-console", false)), json_file(parser.add_option<SimpleOption<std::string>>("json-file")), pretty_file(parser.add_option<SimpleOption<std::string>>("pretty-file")), diff --git a/tests/framework/datasets/ContainerDataset.h b/tests/framework/datasets/ContainerDataset.h index bdca97cba..80616c46f 100644 --- a/tests/framework/datasets/ContainerDataset.h +++ b/tests/framework/datasets/ContainerDataset.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -72,7 +72,8 @@ public: { } - ContainerDataset(ContainerDataset &&) = default; + ContainerDataset(const ContainerDataset &) = default; + ContainerDataset(ContainerDataset &&) = default; /** Type of the dataset. */ using type = std::tuple<container_value_type>; diff --git a/tests/framework/instruments/Instrument.h b/tests/framework/instruments/Instrument.h index 9156d7001..e25939a28 100644 --- a/tests/framework/instruments/Instrument.h +++ b/tests/framework/instruments/Instrument.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -88,6 +88,7 @@ inline std::unique_ptr<Instrument> Instrument::make_instrument() { return support::cpp14::make_unique<T>(scale); } + } // namespace framework } // namespace test } // namespace arm_compute diff --git a/tests/framework/instruments/Instruments.cpp b/tests/framework/instruments/Instruments.cpp index 641e6305c..64e87f9cc 100644 --- a/tests/framework/instruments/Instruments.cpp +++ b/tests/framework/instruments/Instruments.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -44,6 +44,9 @@ InstrumentsDescription instrument_type_from_name(const std::string &name) { "wall_clock_timer", std::pair<InstrumentType, ScaleFactor>(InstrumentType::WALL_CLOCK_TIMER, ScaleFactor::NONE) }, { "wall_clock_timer_ms", std::pair<InstrumentType, ScaleFactor>(InstrumentType::WALL_CLOCK_TIMER, ScaleFactor::TIME_MS) }, { "wall_clock_timer_s", std::pair<InstrumentType, ScaleFactor>(InstrumentType::WALL_CLOCK_TIMER, ScaleFactor::TIME_S) }, + { "scheduler_timer", std::pair<InstrumentType, ScaleFactor>(InstrumentType::SCHEDULER_TIMER, ScaleFactor::NONE) }, + { "scheduler_timer_ms", std::pair<InstrumentType, ScaleFactor>(InstrumentType::SCHEDULER_TIMER, ScaleFactor::TIME_MS) }, + { "scheduler_timer_s", std::pair<InstrumentType, ScaleFactor>(InstrumentType::SCHEDULER_TIMER, ScaleFactor::TIME_S) }, { "pmu", std::pair<InstrumentType, ScaleFactor>(InstrumentType::PMU, ScaleFactor::NONE) }, { "pmu_k", std::pair<InstrumentType, ScaleFactor>(InstrumentType::PMU, ScaleFactor::SCALE_1K) }, { "pmu_m", std::pair<InstrumentType, ScaleFactor>(InstrumentType::PMU, ScaleFactor::SCALE_1M) }, diff --git a/tests/framework/instruments/Instruments.h b/tests/framework/instruments/Instruments.h index 651f0f567..fe4c71931 100644 --- a/tests/framework/instruments/Instruments.h +++ b/tests/framework/instruments/Instruments.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -27,6 +27,7 @@ #include "MaliCounter.h" #include "OpenCLTimer.h" #include "PMUCounter.h" +#include "SchedulerTimer.h" #include "WallClockTimer.h" #include <sstream> @@ -48,6 +49,7 @@ enum class InstrumentType : unsigned int PMU_INSTRUCTION_COUNTER = 0x0202, MALI = 0x0300, OPENCL_TIMER = 0x0400, + SCHEDULER_TIMER = 0x0500, }; using InstrumentsDescription = std::pair<InstrumentType, ScaleFactor>; @@ -82,6 +84,22 @@ inline ::std::stringstream &operator<<(::std::stringstream &stream, InstrumentsD throw std::invalid_argument("Unsupported instrument scale"); } break; + case InstrumentType::SCHEDULER_TIMER: + switch(instrument.second) + { + case ScaleFactor::NONE: + stream << "SCHEDULER_TIMER"; + break; + case ScaleFactor::TIME_MS: + stream << "SCHEDULER_TIMER_MS"; + break; + case ScaleFactor::TIME_S: + stream << "SCHEDULER_TIMER_S"; + break; + default: + throw std::invalid_argument("Unsupported instrument scale"); + } + break; case InstrumentType::PMU: switch(instrument.second) { diff --git a/tests/framework/instruments/InstrumentsStats.cpp b/tests/framework/instruments/InstrumentsStats.cpp new file mode 100644 index 000000000..6fad8f36e --- /dev/null +++ b/tests/framework/instruments/InstrumentsStats.cpp @@ -0,0 +1,62 @@ +/* + * Copyright (c) 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 "InstrumentsStats.h" +#include "arm_compute/core/utils/misc/utility.h" + +namespace arm_compute +{ +namespace test +{ +namespace framework +{ +InstrumentsStats::InstrumentsStats(const std::vector<Measurement> &measurements) + : _min(nullptr), _max(nullptr), _median(nullptr), _mean(measurements.begin()->value().is_floating_point), _stddev(0.0) +{ + auto add_measurements = [](Measurement::Value a, const Measurement & b) + { + return a + b.value(); + }; + + //Calculate min, max & median values + auto indices = arm_compute::utility::sort_indices(measurements); + _median = &measurements[indices[measurements.size() / 2]]; + _min = &measurements[indices[0]]; + _max = &measurements[indices[measurements.size() - 1]]; + + Measurement::Value sum_values = std::accumulate(measurements.begin(), measurements.end(), Measurement::Value(_min->value().is_floating_point), add_measurements); + + // Calculate the relative standard deviation + _mean = sum_values / measurements.size(); + std::vector<Measurement::Value> diff(measurements.size(), _min->value().is_floating_point); + std::transform(measurements.begin(), measurements.end(), diff.begin(), [&](const Measurement & x) + { + return x.value() - _mean; + }); + auto sq_sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), Measurement::Value(_min->value().is_floating_point)); + auto variance = sq_sum / measurements.size(); + _stddev = Measurement::Value::relative_standard_deviation(variance, _mean); +} +} // namespace framework +} // namespace test +} // namespace arm_compute diff --git a/tests/framework/instruments/InstrumentsStats.h b/tests/framework/instruments/InstrumentsStats.h new file mode 100644 index 000000000..f1085aafb --- /dev/null +++ b/tests/framework/instruments/InstrumentsStats.h @@ -0,0 +1,89 @@ +/* + * Copyright (c) 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. + */ +#ifndef ARM_COMPUTE_TEST_INSTRUMENTSMAP +#define ARM_COMPUTE_TEST_INSTRUMENTSMAP + +#include "Measurement.h" + +#include <vector> + +namespace arm_compute +{ +namespace test +{ +namespace framework +{ +/** Generate common statistics for a set of measurements + */ +class InstrumentsStats +{ +public: + /** Compute statistics for the passed set of measurements + * + * @param[in] measurements The measurements to process + */ + InstrumentsStats(const std::vector<Measurement> &measurements); + /** The measurement with the minimum value + */ + const Measurement &min() const + { + return *_min; + } + /** The measurement with the maximum value + */ + const Measurement &max() const + { + return *_max; + } + /** The median measurement + */ + const Measurement &median() const + { + return *_median; + } + /** The average of all the measurements + */ + const Measurement::Value &mean() const + { + return _mean; + } + /** The relative standard deviation of the measurements + */ + double relative_standard_deviation() const + { + return _stddev; + } + +private: + const Measurement *_min; + const Measurement *_max; + const Measurement *_median; + Measurement::Value _mean; + double _stddev; +}; + +} // namespace framework +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_INSTRUMENTSMAP */ diff --git a/tests/framework/instruments/Measurement.h b/tests/framework/instruments/Measurement.h index 9fb75d7b9..1beacf6cc 100644 --- a/tests/framework/instruments/Measurement.h +++ b/tests/framework/instruments/Measurement.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -208,6 +208,17 @@ struct Measurement bool is_floating_point; /**< Is the stored value floating point or integer ? */ }; + /** Compare the stored value with another value + * + * @param[in] b Value to compare against + * + * @return The result of stored value < b + */ + bool operator<(const Measurement &b) const + { + return _value < b.value(); + } + /** Stream output operator to print the measurement. * * Prints value and unit. diff --git a/tests/framework/instruments/OpenCLTimer.cpp b/tests/framework/instruments/OpenCLTimer.cpp index 3de953fbe..9743015ce 100644 --- a/tests/framework/instruments/OpenCLTimer.cpp +++ b/tests/framework/instruments/OpenCLTimer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -142,20 +142,10 @@ Instrument::MeasurementsMap OpenCLTimer::measurements() const unsigned int kernel_number = 0; for(auto kernel : kernels) { - //cl_int status = kernel.event.getInfo(<CL_EVENT_COMMAND_EXECUTION_STATUS>(); - cl_ulong queued = kernel.event.getProfilingInfo<CL_PROFILING_COMMAND_QUEUED>(); - cl_ulong submit = kernel.event.getProfilingInfo<CL_PROFILING_COMMAND_SUBMIT>(); - cl_ulong start = kernel.event.getProfilingInfo<CL_PROFILING_COMMAND_START>(); - cl_ulong end = kernel.event.getProfilingInfo<CL_PROFILING_COMMAND_END>(); + cl_ulong start = kernel.event.getProfilingInfo<CL_PROFILING_COMMAND_START>(); + cl_ulong end = kernel.event.getProfilingInfo<CL_PROFILING_COMMAND_END>(); - std::list<std::string> raw_data = - { - "queued", support::cpp11::to_string(queued), - "submit", support::cpp11::to_string(submit), - "start", support::cpp11::to_string(start), - "end", support::cpp11::to_string(end), - }; - measurements.emplace(kernel.name + " #" + support::cpp11::to_string(kernel_number++), Measurement((end - start) / _scale_factor, _unit, raw_data)); + measurements.emplace(kernel.name + " #" + support::cpp11::to_string(kernel_number++), Measurement((end - start) / _scale_factor, _unit)); } return measurements; diff --git a/tests/framework/instruments/PMU.h b/tests/framework/instruments/PMU.h index e0a187053..c069a6366 100644 --- a/tests/framework/instruments/PMU.h +++ b/tests/framework/instruments/PMU.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -27,6 +27,7 @@ #include "arm_compute/core/Error.h" #include <cstdint> +#include <errno.h> #include <linux/perf_event.h> #include <stdexcept> #include <sys/syscall.h> diff --git a/tests/framework/instruments/SchedulerTimer.cpp b/tests/framework/instruments/SchedulerTimer.cpp new file mode 100644 index 000000000..e42cebde2 --- /dev/null +++ b/tests/framework/instruments/SchedulerTimer.cpp @@ -0,0 +1,116 @@ +/* + * 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 "SchedulerTimer.h" + +#include "WallClockTimer.h" +#include "arm_compute/core/CPP/ICPPKernel.h" + +namespace arm_compute +{ +namespace test +{ +namespace framework +{ +std::string SchedulerTimer::id() const +{ + return "SchedulerTimer"; +} + +class Interceptor final : public IScheduler +{ +public: + /** Default constructor. */ + Interceptor(std::list<SchedulerTimer::kernel_info> &kernels, IScheduler &real_scheduler, ScaleFactor scale_factor) + : _kernels(kernels), _real_scheduler(real_scheduler), _timer(scale_factor) + { + } + + void set_num_threads(unsigned int num_threads) override + { + _real_scheduler.set_num_threads(num_threads); + } + + unsigned int num_threads() const override + { + return _real_scheduler.num_threads(); + } + + void schedule(ICPPKernel *kernel, unsigned int split_dimension) override + { + _timer.start(); + _real_scheduler.schedule(kernel, split_dimension); + _timer.stop(); + + SchedulerTimer::kernel_info info; + info.name = kernel->name(); + info.measurements = _timer.measurements(); + _kernels.push_back(std::move(info)); + } + +private: + std::list<SchedulerTimer::kernel_info> &_kernels; + IScheduler &_real_scheduler; + WallClockTimer _timer; +}; + +SchedulerTimer::SchedulerTimer(ScaleFactor scale_factor) + : _kernels(), _real_scheduler(nullptr), _real_scheduler_type(), _scale_factor(scale_factor) +{ +} + +void SchedulerTimer::start() +{ + ARM_COMPUTE_ERROR_ON(_real_scheduler != nullptr); + _real_scheduler_type = Scheduler::get_type(); + //Note: We can't currently replace a custom scheduler + if(_real_scheduler_type != Scheduler::Type::CUSTOM) + { + _real_scheduler = &Scheduler::get(); + auto interceptor = std::make_shared<Interceptor>(_kernels, *_real_scheduler, _scale_factor); + Scheduler::set(std::static_pointer_cast<IScheduler>(interceptor)); + } + _kernels.clear(); +} + +void SchedulerTimer::stop() +{ + // Restore real scheduler + Scheduler::set(_real_scheduler_type); + _real_scheduler = nullptr; +} + +Instrument::MeasurementsMap SchedulerTimer::measurements() const +{ + MeasurementsMap measurements; + unsigned int kernel_number = 0; + for(auto kernel : _kernels) + { + measurements.emplace(kernel.name + " #" + support::cpp11::to_string(kernel_number++), kernel.measurements.begin()->second); + } + + return measurements; +} +} // namespace framework +} // namespace test +} // namespace arm_compute diff --git a/tests/framework/instruments/SchedulerTimer.h b/tests/framework/instruments/SchedulerTimer.h new file mode 100644 index 000000000..446506ad7 --- /dev/null +++ b/tests/framework/instruments/SchedulerTimer.h @@ -0,0 +1,63 @@ +/* + * 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. + */ +#ifndef ARM_COMPUTE_TEST_SCHEDULER_TIMER +#define ARM_COMPUTE_TEST_SCHEDULER_TIMER + +#include "Instrument.h" +#include "arm_compute/runtime/Scheduler.h" +#include <list> + +namespace arm_compute +{ +namespace test +{ +namespace framework +{ +/** Instrument creating measurements based on the information returned by clGetEventProfilingInfo for each OpenCL kernel executed*/ +class SchedulerTimer : public Instrument +{ +public: + SchedulerTimer(const SchedulerTimer &) = delete; + SchedulerTimer &operator=(const SchedulerTimer &) = delete; + SchedulerTimer(ScaleFactor scale_factor); + std::string id() const override; + void start() override; + void stop() override; + Instrument::MeasurementsMap measurements() const override; + struct kernel_info + { + Instrument::MeasurementsMap measurements{}; /**< Time it took the kernel to run */ + std::string name{}; /**< Kernel name */ + }; + +private: + std::list<kernel_info> _kernels; + IScheduler *_real_scheduler; + Scheduler::Type _real_scheduler_type; + ScaleFactor _scale_factor; +}; +} // namespace framework +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_SCHEDULER_TIMER */ diff --git a/tests/framework/printers/JSONPrinter.cpp b/tests/framework/printers/JSONPrinter.cpp index 676ec6933..6b982f5bb 100644 --- a/tests/framework/printers/JSONPrinter.cpp +++ b/tests/framework/printers/JSONPrinter.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -179,26 +179,6 @@ void JSONPrinter::print_measurements(const Profiler::MeasurementsMap &measuremen { *_stream << R"(")" << i_it->first << R"(" : {)"; - auto add_measurements = [](Measurement::Value a, const Measurement & b) - { - return a + b.value(); - }; - - auto cmp_measurements = [](const Measurement & a, const Measurement & b) - { - return a.value() < b.value(); - }; - - int num_values = i_it->second.size(); - const auto minmax_values = std::minmax_element(i_it->second.begin(), i_it->second.end(), cmp_measurements); - - Measurement::Value sum_values = std::accumulate(i_it->second.cbegin(), i_it->second.cend(), Measurement::Value(minmax_values.first->value().is_floating_point), add_measurements); - if(num_values > 2) - { - sum_values -= minmax_values.first->value() + minmax_values.second->value(); - num_values -= 2; - } - auto measurement_to_string = [](const Measurement & measurement) { if(measurement.raw_data().size() == 1) @@ -214,14 +194,8 @@ void JSONPrinter::print_measurements(const Profiler::MeasurementsMap &measuremen return str.str(); } }; - *_stream << R"("avg" : )" << (sum_values / num_values) << ","; - if(num_values > 1) - { - *_stream << R"("min" : )" << minmax_values.first->value() << ","; - *_stream << R"("max" : )" << minmax_values.second->value() << ","; - } *_stream << R"("raw" : [)" << join(i_it->second.begin(), i_it->second.end(), ",", measurement_to_string) << "],"; - *_stream << R"("unit" : ")" << minmax_values.first->unit() << R"(")"; + *_stream << R"("unit" : ")" << i_it->second.begin()->unit() << R"(")"; *_stream << "}"; if(++i_it != i_end) diff --git a/tests/framework/printers/PrettyPrinter.cpp b/tests/framework/printers/PrettyPrinter.cpp index 085b2b900..ef8f91a79 100644 --- a/tests/framework/printers/PrettyPrinter.cpp +++ b/tests/framework/printers/PrettyPrinter.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -24,6 +24,7 @@ #include "PrettyPrinter.h" #include "../Framework.h" +#include "../instruments/InstrumentsStats.h" #include "../instruments/Measurement.h" #include <algorithm> @@ -121,50 +122,16 @@ void PrettyPrinter::print_measurements(const Profiler::MeasurementsMap &measurem { *_stream << begin_color("3") << " " << instrument.first << ":"; - auto add_measurements = [](Measurement::Value a, const Measurement & b) - { - return a + b.value(); - }; - - auto cmp_measurements = [](const Measurement & a, const Measurement & b) - { - return a.value() < b.value(); - }; - - int num_values = instrument.second.size(); - const auto minmax_values = std::minmax_element(instrument.second.begin(), instrument.second.end(), cmp_measurements); - Measurement::Value sum_values = std::accumulate(instrument.second.begin(), instrument.second.end(), Measurement::Value(minmax_values.first->value().is_floating_point), add_measurements); - - // Calculate the median value - auto measurements = instrument.second; - std::nth_element(measurements.begin(), measurements.begin() + (num_values / 2), measurements.end(), cmp_measurements); - const auto median_value = measurements[num_values / 2]; - - // Calculate the relative standard deviation - auto mean_value = sum_values / num_values; - std::vector<Measurement::Value> diff(measurements.size(), minmax_values.first->value().is_floating_point); - std::transform(measurements.begin(), measurements.end(), diff.begin(), [mean_value](const Measurement & x) - { - return x.value() - mean_value; - }); - auto sq_sum = std::inner_product(diff.begin(), diff.end(), diff.begin(), Measurement::Value(minmax_values.first->value().is_floating_point)); - auto variance = sq_sum / measurements.size(); - auto rsd = Measurement::Value::relative_standard_deviation(variance, mean_value); - - if(num_values > 2) - { - sum_values -= minmax_values.first->value() + minmax_values.second->value(); - num_values -= 2; - } + InstrumentsStats stats(instrument.second); *_stream << " "; - *_stream << "MEDIAN=" << median_value.value() << " " << median_value.unit() << ", "; - *_stream << "AVG=" << (sum_values / num_values) << " " << minmax_values.second->unit() << ", "; - *_stream << "STDDEV=" << arithmetic_to_string(rsd, 2) << " %, "; - if(num_values > 1) + *_stream << "AVG=" << stats.mean() << " " << stats.max().unit(); + if(instrument.second.size() > 1) { - *_stream << "MIN=" << *minmax_values.first << ", "; - *_stream << "MAX=" << *minmax_values.second; + *_stream << ", STDDEV=" << arithmetic_to_string(stats.relative_standard_deviation(), 2) << " %"; + *_stream << ", MIN=" << stats.min() << ", "; + *_stream << ", MAX=" << stats.max() << ", "; + *_stream << ", MEDIAN=" << stats.median().value() << " " << stats.median().unit(); } *_stream << end_color() << "\n"; } diff --git a/tests/validation/CL/ActivationLayer.cpp b/tests/validation/CL/ActivationLayer.cpp index 5ceaecaaa..5fd0855de 100644 --- a/tests/validation/CL/ActivationLayer.cpp +++ b/tests/validation/CL/ActivationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -183,7 +183,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU), - ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), @@ -241,15 +241,15 @@ TEST_SUITE(FixedPoint) TEST_SUITE(QS8) // We test for fixed point precision [3,5] because [1,2] and [6,7] ranges cause // overflowing issues in most of the transcendentals functions. -FIXTURE_DATA_TEST_CASE(RunSmall, CLActivationLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ActivationDataset), - framework::dataset::make("DataType", - DataType::QS8)), - framework::dataset::make("FractionalBits", 3, 6))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLActivationLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), ActivationDataset), + framework::dataset::make("DataType", + DataType::QS8)), + framework::dataset::make("FractionalBits", 3, 6))) { // Validate output validate(CLAccessor(_target), _reference, tolerance(_function, _data_type)); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLActivationLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ActivationDataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLActivationLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), ActivationDataset), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 3, 6))) @@ -261,15 +261,15 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) as reciprocal limits the maximum fixed point position to 14 -FIXTURE_DATA_TEST_CASE(RunSmall, CLActivationLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ActivationDataset), - framework::dataset::make("DataType", - DataType::QS16)), - framework::dataset::make("FractionalBits", 1, 14))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLActivationLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), ActivationDataset), + framework::dataset::make("DataType", + DataType::QS16)), + framework::dataset::make("FractionalBits", 1, 14))) { // Validate output validate(CLAccessor(_target), _reference, tolerance(_function, _data_type)); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLActivationLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ActivationDataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLActivationLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), ActivationDataset), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) @@ -284,7 +284,12 @@ template <typename T> using CLActivationLayerQuantizedFixture = ActivationValidationQuantizedFixture<CLTensor, CLAccessor, CLActivationLayer, T>; /** Input data sets. */ -const auto QuantizedActivationDataset = combine(combine(framework::dataset::make("InPlace", { false, true }), framework::dataset::make("ActivationFunction", { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU })), +const auto QuantizedActivationFunctionsDataset = framework::dataset::make("ActivationFunction", { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, + ActivationLayerInfo::ActivationFunction::RELU, + ActivationLayerInfo::ActivationFunction::BOUNDED_RELU + }); + +const auto QuantizedActivationDataset = combine(combine(framework::dataset::make("InPlace", { false, true }), QuantizedActivationFunctionsDataset), framework::dataset::make("AlphaBeta", { 0.5f, 1.f })); TEST_SUITE(Quantized) diff --git a/tests/validation/CL/ArithmeticAddition.cpp b/tests/validation/CL/ArithmeticAddition.cpp index 787b1b986..e9a892d74 100644 --- a/tests/validation/CL/ArithmeticAddition.cpp +++ b/tests/validation/CL/ArithmeticAddition.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -176,7 +176,7 @@ using CLArithmeticAdditionFixedPointFixture = ArithmeticAdditionValidationFixedP TEST_SUITE(Quantized) TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionQS8Dataset), +FIXTURE_DATA_TEST_CASE(RunTiny, CLArithmeticAdditionFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), ArithmeticAdditionQS8Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 7))) { @@ -184,7 +184,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFixedPointFixture<int8_t>, validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticAdditionFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ArithmeticAdditionQS8Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionQS8Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 7))) { @@ -194,7 +194,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticAdditionFixedPointFixture<int8_t>, TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionQS16Dataset), +FIXTURE_DATA_TEST_CASE(RunTiny, CLArithmeticAdditionFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), ArithmeticAdditionQS16Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 15))) { @@ -202,7 +202,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFixedPointFixture<int16_t>, validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticAdditionFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ArithmeticAdditionQS16Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticAdditionFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionQS16Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 15))) { @@ -259,6 +259,25 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticAdditionFixture<float>, framework:: // Validate output validate(CLAccessor(_target), _reference); } + +template <typename T> +using CLArithmeticAdditionBroadcastFixture = ArithmeticAdditionBroadcastValidationFixture<CLTensor, CLAccessor, CLArithmeticAddition, T>; + +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, CLArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapesBroadcast(), + ArithmeticAdditionFP32Dataset), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + +FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, CLArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapesBroadcast(), + ArithmeticAdditionFP32Dataset), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} TEST_SUITE_END() TEST_SUITE_END() diff --git a/tests/validation/CL/ArithmeticSubtraction.cpp b/tests/validation/CL/ArithmeticSubtraction.cpp index 34fdb0b93..43c94a7b9 100644 --- a/tests/validation/CL/ArithmeticSubtraction.cpp +++ b/tests/validation/CL/ArithmeticSubtraction.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -246,7 +246,7 @@ using CLArithmeticSubtractionFixedPointFixture = ArithmeticSubtractionValidation TEST_SUITE(Quantized) TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTiny, CLArithmeticSubtractionFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), ArithmeticSubtractionQS8Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 7))) @@ -255,7 +255,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixedPointFixture<int8_t validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticSubtractionFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionQS8Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 7))) @@ -266,7 +266,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticSubtractionFixedPointFixture<int8_t TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTiny, CLArithmeticSubtractionFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), ArithmeticSubtractionQS16Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 15))) @@ -275,7 +275,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixedPointFixture<int16_ validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLArithmeticSubtractionFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLArithmeticSubtractionFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionQS16Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 15))) diff --git a/tests/validation/CL/BatchNormalizationLayer.cpp b/tests/validation/CL/BatchNormalizationLayer.cpp index 30dd70a66..ef535153f 100644 --- a/tests/validation/CL/BatchNormalizationLayer.cpp +++ b/tests/validation/CL/BatchNormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -47,6 +47,12 @@ constexpr AbsoluteTolerance<float> tolerance_f32(0.00001f); /**< Tolerance value constexpr AbsoluteTolerance<float> tolerance_f16(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ constexpr AbsoluteTolerance<float> tolerance_qs8(3.0f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::QS8 */ constexpr AbsoluteTolerance<float> tolerance_qs16(6.0f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::QS16 */ +const auto act_infos = framework::dataset::make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f), +}); } // namespace TEST_SUITE(CL) @@ -80,13 +86,16 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::Ran // *INDENT-OFF* // clang-format off -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Window shrink TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching data types TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching data types TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Invalid mean/var/beta/gamma shape TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2), // Mismatching fixed point position + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2), // Fused activation with fixed point not supported + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Unsupported fused activation + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Fused activation's a < b TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2), }), @@ -96,6 +105,9 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 3), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 3), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2), TensorInfo(), })), @@ -106,16 +118,31 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(5U), 1, DataType::F32), TensorInfo(TensorShape(2U), 1, DataType::QS8, 2), TensorInfo(TensorShape(2U), 1, DataType::QS8, 2), + TensorInfo(TensorShape(2U), 1, DataType::F32), + TensorInfo(TensorShape(2U), 1, DataType::F32), TensorInfo(TensorShape(2U), 1, DataType::QS8, 2), + TensorInfo(TensorShape(2U), 1, DataType::QS8, 2), + })), + framework::dataset::make("ActivationLayerInfo",{ ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f, 2.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f, 2.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::TANH), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 2.f, 6.f), + ActivationLayerInfo(), + ActivationLayerInfo(), })), - framework::dataset::make("Expected", { true, false, false, false, false, false, true, true})), - input_info, output_info, mvbg_info, expected) + framework::dataset::make("Expected", { true, false, false, false, false, false, false, false, false, true, true})), + input_info, output_info, mvbg_info, act_info, expected) { const auto &mean_info = mvbg_info; const auto &var_info = mvbg_info; const auto &beta_info = mvbg_info; const auto &gamma_info = mvbg_info; - bool has_error = bool(CLBatchNormalizationLayer::validate(&input_info.clone()->set_is_resizable(false), (output_info.total_size() == 0) ? nullptr : &output_info.clone()->set_is_resizable(false), &mean_info.clone()->set_is_resizable(false), &var_info.clone()->set_is_resizable(false), &beta_info.clone()->set_is_resizable(false), &gamma_info.clone()->set_is_resizable(false), 1.f)); + bool has_error = bool(CLBatchNormalizationLayer::validate(&input_info.clone()->set_is_resizable(false), (output_info.total_size() == 0) ? nullptr : &output_info.clone()->set_is_resizable(false), &mean_info.clone()->set_is_resizable(false), &var_info.clone()->set_is_resizable(false), &beta_info.clone()->set_is_resizable(false), &gamma_info.clone()->set_is_resizable(false), 1.f, act_info)); ARM_COMPUTE_EXPECT(has_error == expected, framework::LogLevel::ERRORS); } // clang-format on @@ -123,7 +150,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TEST_SUITE(Float) TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::RandomBatchNormalizationLayerDataset(), +FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::RandomBatchNormalizationLayerDataset(), + act_infos), framework::dataset::make("DataType", DataType::F32))) { // Validate output @@ -132,7 +160,8 @@ FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixture<float>, framewor TEST_SUITE_END() TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::RandomBatchNormalizationLayerDataset(), +FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::RandomBatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f))), framework::dataset::make("DataType", DataType::F16))) { // Validate output @@ -146,7 +175,8 @@ template <typename T> using CLBatchNormalizationLayerFixedPointFixture = BatchNormalizationLayerValidationFixedPointFixture<CLTensor, CLAccessor, CLBatchNormalizationLayer, T>; TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::RandomBatchNormalizationLayerDataset(), +FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::RandomBatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 1, 6))) { @@ -156,7 +186,8 @@ FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixedPointFixture<int8_t TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::RandomBatchNormalizationLayerDataset(), +FIXTURE_DATA_TEST_CASE(Random, CLBatchNormalizationLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::RandomBatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) { diff --git a/tests/validation/CL/ChannelExtract.cpp b/tests/validation/CL/ChannelExtract.cpp new file mode 100644 index 000000000..926e16bd4 --- /dev/null +++ b/tests/validation/CL/ChannelExtract.cpp @@ -0,0 +1,151 @@ +/* + * 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/CL/CLMultiImage.h" +#include "arm_compute/runtime/CL/CLTensor.h" +#include "arm_compute/runtime/CL/CLTensorAllocator.h" +#include "arm_compute/runtime/CL/functions/CLChannelExtract.h" +#include "tests/CL/CLAccessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/ConvertPolicyDataset.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/ChannelExtractFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +// Input data sets +const auto ChannelExtractRGBADataset = combine(framework::dataset::make("FormatType", { Format::RGBA8888 }), + framework::dataset::make("ChannelType", { Channel::R, Channel::G, Channel::B, Channel::A })); +const auto ChannelExtractYUVDataset = combine(framework::dataset::make("FormatType", { Format::YUYV422, Format::UYVY422 }), + framework::dataset::make("ChannelType", { Channel::Y, Channel::U, Channel::V })); +const auto ChannelExtractYUVPlanarDataset = combine(framework::dataset::make("FormatType", { Format::IYUV, Format::YUV444, Format::NV12, Format::NV21 }), + framework::dataset::make("ChannelType", { Channel::Y, Channel::U, Channel::V })); + +inline void validate_configuration(const TensorShape &shape, Format format, Channel channel) +{ + const unsigned int num_planes = num_planes_from_format(format); + + TensorShape dst_shape = adjust_odd_shape(shape, format); + dst_shape = calculate_subsampled_shape(dst_shape, format, channel); + + // Create tensors + CLMultiImage ref_src = create_multi_image<CLMultiImage>(shape, format); + CLTensor dst = create_tensor<CLTensor>(dst_shape, Format::U8); + + // Create and Configure function + CLChannelExtract channel_extract; + + if(1U == num_planes) + { + const CLTensor *plane_src = ref_src.cl_plane(0); + + channel_extract.configure(plane_src, channel, &dst); + } + else + { + channel_extract.configure(&ref_src, channel, &dst); + } +} +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(ChannelExtract) + +template <typename T> +using CLChannelExtractFixture = ChannelExtractValidationFixture<CLMultiImage, CLTensor, CLAccessor, CLChannelExtract, T>; + +TEST_SUITE(Configuration) +DATA_TEST_CASE(RGBA, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), ChannelExtractRGBADataset), + shape, format, channel) +{ + validate_configuration(shape, format, channel); +} +DATA_TEST_CASE(YUV, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), ChannelExtractYUVDataset), + shape, format, channel) +{ + validate_configuration(shape, format, channel); +} + +DATA_TEST_CASE(YUVPlanar, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), ChannelExtractYUVPlanarDataset), + shape, format, channel) +{ + validate_configuration(shape, format, channel); +} +TEST_SUITE_END() + +TEST_SUITE(RGBA) +FIXTURE_DATA_TEST_CASE(RunSmall, CLChannelExtractFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), ChannelExtractRGBADataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLChannelExtractFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::Large2DShapes(), ChannelExtractRGBADataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() + +TEST_SUITE(YUV) +FIXTURE_DATA_TEST_CASE(RunSmall, CLChannelExtractFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), ChannelExtractYUVDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLChannelExtractFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::Large2DShapes(), ChannelExtractYUVDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() + +TEST_SUITE(YUVPlanar) +FIXTURE_DATA_TEST_CASE(RunSmall, CLChannelExtractFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), ChannelExtractYUVPlanarDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLChannelExtractFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::Large2DShapes(), ChannelExtractYUVPlanarDataset)) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} +TEST_SUITE_END() + +TEST_SUITE_END() +TEST_SUITE_END() + +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/Convolution.cpp b/tests/validation/CL/Convolution.cpp index ccb0abcbd..eb95d1a29 100644 --- a/tests/validation/CL/Convolution.cpp +++ b/tests/validation/CL/Convolution.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -41,23 +41,17 @@ namespace test { namespace validation { -namespace -{ -} // namespace - TEST_SUITE(CL) TEST_SUITE(CustomConvolution) -TEST_SUITE(CustomConvolutionSquare) -TEST_SUITE(CustomConvolution3x3) - -DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", DataType::U8)), +TEST_SUITE(Square3x3) +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", { DataType::U8, DataType::S16 })), datasets::BorderModes()), framework::dataset::make("filter_size", { 3 })), - shape, data_type, border_mode, filter_size) + shape, output_data_type, border_mode, filter_size) { // Create tensors - CLTensor src = create_tensor<CLTensor>(shape, data_type); - CLTensor dst = create_tensor<CLTensor>(shape, data_type); + CLTensor src = create_tensor<CLTensor>(shape, DataType::U8); + CLTensor dst = create_tensor<CLTensor>(shape, output_data_type); // Create conv matrix int16_t conv[9]; @@ -92,6 +86,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi template <typename T> using CLConvolutionFixture = ConvolutionSquareValidationFixture<CLTensor, CLAccessor, CLConvolution3x3, T>; +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -109,18 +104,37 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<uint8_t>, framework::Datas // Validate output validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); } -TEST_SUITE_END() /* Custom_Convolution 3x3 */ +TEST_SUITE_END() -TEST_SUITE(CustomConvolution5x5) +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 3 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 3 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Square 3x3 */ -DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", DataType::U8)), +TEST_SUITE(Square5x5) +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", { DataType::U8, DataType::S16 })), datasets::BorderModes()), framework::dataset::make("filter_size", { 5 })), - shape, data_type, border_mode, filter_size) + shape, output_data_type, border_mode, filter_size) { // Create tensors - CLTensor src = create_tensor<CLTensor>(shape, data_type); - CLTensor dst = create_tensor<CLTensor>(shape, data_type); + CLTensor src = create_tensor<CLTensor>(shape, DataType::U8); + CLTensor dst = create_tensor<CLTensor>(shape, output_data_type); // Create conv matrix int16_t conv[25]; @@ -154,6 +168,8 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi template <typename T> using CLConvolutionFixture = ConvolutionSquareValidationFixture<CLTensor, CLAccessor, CLConvolution5x5, T>; + +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -171,17 +187,38 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<uint8_t>, framework::Datas // Validate output validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); } -TEST_SUITE_END() /* Custom Convolution 5x5 */ +TEST_SUITE_END() + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 5 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 5 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Square5x5 */ -TEST_SUITE(CustomConvolution7x7) -DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", DataType::U8)), +TEST_SUITE(Square7x7) +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", { DataType::U8, DataType::S16 })), datasets::BorderModes()), framework::dataset::make("filter_size", { 7 })), - shape, data_type, border_mode, filter_size) + shape, output_data_type, border_mode, filter_size) { // Create tensors - CLTensor src = create_tensor<CLTensor>(shape, data_type); - CLTensor dst = create_tensor<CLTensor>(shape, data_type); + CLTensor src = create_tensor<CLTensor>(shape, DataType::U8); + CLTensor dst = create_tensor<CLTensor>(shape, output_data_type); // Create conv matrix int16_t conv[49]; @@ -216,6 +253,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi template <typename T> using CLConvolutionFixture = ConvolutionSquareValidationFixture<CLTensor, CLAccessor, CLConvolution7x7, T>; +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -233,17 +271,38 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<uint8_t>, framework::Datas // Validate output validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); } -TEST_SUITE_END() /* Custom Convolution 7x7 */ +TEST_SUITE_END() + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 7 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 7 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Square7x7 */ -TEST_SUITE(CustomConvolution9x9) -DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", DataType::U8)), +TEST_SUITE(Square9x9) +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", { DataType::U8, DataType::S16 })), datasets::BorderModes()), framework::dataset::make("filter_size", { 9 })), - shape, data_type, border_mode, filter_size) + shape, output_data_type, border_mode, filter_size) { // Create tensors - CLTensor src = create_tensor<CLTensor>(shape, data_type); - CLTensor dst = create_tensor<CLTensor>(shape, data_type); + CLTensor src = create_tensor<CLTensor>(shape, DataType::U8); + CLTensor dst = create_tensor<CLTensor>(shape, output_data_type); // Create conv matrix int16_t conv[81]; @@ -278,6 +337,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi template <typename T> using CLConvolutionFixture = ConvolutionSquareValidationFixture<CLTensor, CLAccessor, CLConvolution9x9, T>; +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -295,21 +355,40 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<uint8_t>, framework::Datas // Validate output validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); } -TEST_SUITE_END() /* Custom Convolution 9x9 */ -TEST_SUITE_END() /* Custom Convolution Square */ +TEST_SUITE_END() + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 9 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} -TEST_SUITE(CustomConvolutionRectangle) +FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 9 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Square9x9 */ +TEST_SUITE(Rectangle) DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", - DataType::U8)), - datasets::BorderModes()), - framework::dataset::make("filter_width", { 3, 5, 7, 9 })), - framework::dataset::make("filter_height", { 3, 5, 7, 9 })), - shape, data_type, border_mode, filter_width, filter_height) +{ DataType::U8, DataType::S16 })), +datasets::BorderModes()), +framework::dataset::make("filter_width", { 3, 5, 7, 9 })), +framework::dataset::make("filter_height", { 3, 5, 7, 9 })), +shape, output_data_type, border_mode, filter_width, filter_height) { // Create tensors - CLTensor src = create_tensor<CLTensor>(shape, data_type); - CLTensor dst = create_tensor<CLTensor>(shape, data_type); + CLTensor src = create_tensor<CLTensor>(shape, DataType::U8); + CLTensor dst = create_tensor<CLTensor>(shape, output_data_type); // Create conv matrix int16_t conv[filter_width * filter_height]; @@ -348,6 +427,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi template <typename T> using CLConvolutionFixture = ConvolutionRectangleValidationFixture<CLTensor, CLAccessor, CLConvolutionRectangle, T>; +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -367,13 +447,36 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<uint8_t>, framework::Datas // Validate output validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); } -TEST_SUITE_END() /* Custom Convolution Rectangle */ +TEST_SUITE_END() -TEST_SUITE(CustomConvolutionSeparable) -TEST_SUITE(CustomConvolutionSeparable5x5) +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_width", { 3, 5, 7, 9 })), + framework::dataset::make("filter_height", { 3, 5, 7, 9 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_width", { 3, 5, 7, 9 })), + framework::dataset::make("filter_height", { 3, 5, 7, 9 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Rectangle */ + +TEST_SUITE(Separable5x5) template <typename T> using CLConvolutionFixture = ConvolutionSeparableValidationFixture<CLTensor, CLAccessor, CLConvolution5x5, T>; +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -391,12 +494,34 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<uint8_t>, framework::Datas // Validate output validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); } -TEST_SUITE_END() /* Custom Convolution Separable 5x5 */ +TEST_SUITE_END() + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 5 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 5 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Separable5x5 */ -TEST_SUITE(CustomConvolutionSeparablex7x7) +TEST_SUITE(Separable7x7) template <typename T> using CLConvolutionFixture = ConvolutionSeparableValidationFixture<CLTensor, CLAccessor, CLConvolution7x7, T>; +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -414,12 +539,34 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<uint8_t>, framework::Datas // Validate output validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); } -TEST_SUITE_END() /* Custom Convolution Separable 7x7 */ +TEST_SUITE_END() + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 7 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 7 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Separable7x7 */ -TEST_SUITE(CustomConvolutionSeparable9x9) +TEST_SUITE(Separable9x9) template <typename T> using CLConvolutionFixture = ConvolutionSeparableValidationFixture<CLTensor, CLAccessor, CLConvolution9x9, T>; +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -437,9 +584,29 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<uint8_t>, framework::Datas // Validate output validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); } -TEST_SUITE_END() /* Custom Convolution Separable 9x9 */ +TEST_SUITE_END() + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 9 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 9 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2))); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Separable9x9 */ -TEST_SUITE_END() /* Custom Convolution Separable */ TEST_SUITE_END() /* Custom Convolution */ TEST_SUITE_END() } // namespace validation diff --git a/tests/validation/CL/ConvolutionLayer.cpp b/tests/validation/CL/ConvolutionLayer.cpp index 42d2e9f1d..9b857c81a 100644 --- a/tests/validation/CL/ConvolutionLayer.cpp +++ b/tests/validation/CL/ConvolutionLayer.cpp @@ -25,10 +25,12 @@ #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" #include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h" +#include "arm_compute/runtime/CL/functions/CLGEMMConvolutionLayer.h" #include "tests/CL/CLAccessor.h" #include "tests/PaddingCalculator.h" #include "tests/datasets/LargeConvolutionLayerDataset.h" #include "tests/datasets/SmallConvolutionLayerDataset.h" +#include "tests/datasets/TinyConvolutionLayerDataset.h" #include "tests/framework/Asserts.h" #include "tests/framework/Macros.h" #include "tests/framework/datasets/Datasets.h" @@ -63,6 +65,57 @@ const auto CNNDataTypes = framework::dataset::make("DataType", TEST_SUITE(CL) TEST_SUITE(ConvolutionLayer) +DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(17U, 31U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(23U, 27U, 5U, 4U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U, 1U), 1, DataType::F32, 0), + TensorInfo(TensorShape(33U, 27U, 7U, 4U), 1, DataType::F32, 0) + }), + framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(5U, 5U, 2U, 19U), 1, DataType::F32, 0), + TensorInfo(TensorShape(5U, 5U, 2U, 19U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 5U, 21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 5U, 21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(5U, 5U, 7U, 16U), 1, DataType::F16, 0) + })), + framework::dataset::make("BiasesInfo", { TensorInfo(TensorShape(19U), 1, DataType::F32, 0), + TensorInfo(TensorShape(19U), 1, DataType::F32, 0), + TensorInfo(TensorShape(21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(16U), 1, DataType::F32, 0) + })), + framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(15U, 15U, 19U), 1, DataType::F32, 0), + TensorInfo(TensorShape(15U, 15U, 19U), 1, DataType::F32, 0), + TensorInfo(TensorShape(21U, 25U, 21U, 4U), 1, DataType::F32, 0), + TensorInfo(TensorShape(11U, 25U, 21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(11U, 12U, 16U, 4U), 1, DataType::F32, 0) + })), + framework::dataset::make("ConvInfo", { PadStrideInfo(1, 2, 1, 1), + PadStrideInfo(1, 2, 1, 1), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(2, 1, 0, 0), + PadStrideInfo(3, 2, 1, 0) + })), + framework::dataset::make("GpuTarget", { GPUTarget::BIFROST, + GPUTarget::MIDGARD, + GPUTarget::G70, + GPUTarget::MIDGARD, + GPUTarget::BIFROST + })), + + framework::dataset::make("Expected", { ConvolutionMethod::GEMM, ConvolutionMethod::GEMM, ConvolutionMethod::GEMM, ConvolutionMethod::GEMM, ConvolutionMethod::GEMM })), + input_info, weights_info, biases_info, output_info, conv_info, gpu_target, expected) +{ + ConvolutionMethod is_valid = CLConvolutionLayer::get_convolution_method(&input_info.clone()->set_is_resizable(false), + &weights_info.clone()->set_is_resizable(false), + &biases_info.clone()->set_is_resizable(false), + &output_info.clone()->set_is_resizable(false), conv_info, WeightsInfo(), gpu_target); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} +TEST_SUITE_END() + +TEST_SUITE(GEMMConvolutionLayer) + DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallConvolutionLayerDataset(), datasets::LargeConvolutionLayerDataset()), CNNDataTypes), input_shape, weights_shape, bias_shape, output_shape, info, data_type) { @@ -86,7 +139,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da const QuantizationInfo weights_quantization_info = weights.info()->quantization_info(); // Create and configure function - CLConvolutionLayer conv; + CLGEMMConvolutionLayer conv; conv.configure(&src, &weights, &bias, &dst, info); // Validate valid region @@ -106,22 +159,22 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da } template <typename T> -using CLConvolutionLayerFixture = ConvolutionValidationFixture<CLTensor, CLAccessor, CLConvolutionLayer, T>; +using CLGEMMConvolutionLayerFixture = ConvolutionValidationFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T>; TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::F16))) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::F16))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::F16))) +FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::F16))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f16, tolerance_num); @@ -129,18 +182,18 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionLayerFixture<half>, framework::Dat TEST_SUITE_END() TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::F32))) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::F32))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::F32))) +FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::F32))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); @@ -149,12 +202,12 @@ TEST_SUITE_END() TEST_SUITE_END() template <typename T> -using CLConvolutionLayerFixedPointFixture = ConvolutionValidationFixedPointFixture<CLTensor, CLAccessor, CLConvolutionLayer, T>; +using CLGEMMConvolutionLayerFixedPointFixture = ConvolutionValidationFixedPointFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T>; TEST_SUITE(FixedPoint) TEST_SUITE(QS8) // We test for fixed point precision [4,6] -FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QS8)), @@ -163,11 +216,11 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerFixedPointFixture<int8_t>, fr // Validate output validate(CLAccessor(_target), _reference, tolerance_fixed); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::QS8)), - framework::dataset::make("FractionalBits", 4, 7))) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::QS8)), + framework::dataset::make("FractionalBits", 4, 7))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fixed); @@ -176,7 +229,7 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) -FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QS16)), @@ -185,11 +238,11 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerFixedPointFixture<int16_t>, f // Validate output validate(CLAccessor(_target), _reference, tolerance_fixed); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", - DataType::QS16)), - framework::dataset::make("FractionalBits", 1, 14))) +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", + DataType::QS16)), + framework::dataset::make("FractionalBits", 1, 14))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fixed); @@ -198,11 +251,11 @@ TEST_SUITE_END() TEST_SUITE_END() template <typename T> -using CLConvolutionLayerQuantizedFixture = ConvolutionValidationQuantizedFixture<CLTensor, CLAccessor, CLConvolutionLayer, T>; +using CLGEMMConvolutionLayerQuantizedFixture = ConvolutionValidationQuantizedFixture<CLTensor, CLAccessor, CLGEMMConvolutionLayer, T>; TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true })), framework::dataset::make("DataType", DataType::QASYMM8)), framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) }))) @@ -210,10 +263,10 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLConvolutionLayerQuantizedFixture<uint8_t>, fr // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true })), - framework::dataset::make("DataType", DataType::QASYMM8)), - framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 0) }))) +FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 0) }))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qasymm8); diff --git a/tests/validation/CL/DeconvolutionLayer.cpp b/tests/validation/CL/DeconvolutionLayer.cpp index 59e85537e..58a20268e 100644 --- a/tests/validation/CL/DeconvolutionLayer.cpp +++ b/tests/validation/CL/DeconvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -45,6 +45,9 @@ namespace { constexpr AbsoluteTolerance<float> tolerance_fp32(0.001f); /**< Tolerance for floating point tests */ +const auto data4x4 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 3) + * framework::dataset::make("PadY", 0, 3) * framework::dataset::make("ax", 0) * framework::dataset::make("ay", 0) * framework::dataset::make("NumKernels", { 1, 3 }); + const auto data3x3 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 2) * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("ax", 0) * framework::dataset::make("ay", 0) * framework::dataset::make("NumKernels", { 1, 3 }); @@ -157,6 +160,9 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip(zi // *INDENT-ON* template <typename T> +using CLDeconvolutionLayerFixture4x4 = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 4, 4>; + +template <typename T> using CLDeconvolutionLayerFixture3x3 = DeconvolutionValidationFixture<CLTensor, CLAccessor, CLDeconvolutionLayer, T, 3, 3>; template <typename T> @@ -165,6 +171,15 @@ using CLDeconvolutionLayerFixture1x1 = DeconvolutionValidationFixture<CLTensor, TEST_SUITE(Float) TEST_SUITE(FP32) +TEST_SUITE(W4x4) + +FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerFixture4x4<float>, framework::DatasetMode::ALL, combine(data4x4, framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_fp32); +} +TEST_SUITE_END() + TEST_SUITE(W3x3) FIXTURE_DATA_TEST_CASE(Run, CLDeconvolutionLayerFixture3x3<float>, framework::DatasetMode::ALL, combine(data3x3, framework::dataset::make("DataType", DataType::F32))) diff --git a/tests/validation/CL/DepthConcatenateLayer.cpp b/tests/validation/CL/DepthConcatenateLayer.cpp index 723cd94f1..bed2a4519 100644 --- a/tests/validation/CL/DepthConcatenateLayer.cpp +++ b/tests/validation/CL/DepthConcatenateLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -79,14 +79,14 @@ TEST_SUITE_END() TEST_SUITE(Quantized) TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthConcatenateLayerFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), - framework::dataset::make("DataType", - DataType::QS8))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLDepthConcatenateLayerFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Tiny2DShapes(), + framework::dataset::make("DataType", + DataType::QS8))) { // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthConcatenateLayerFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::DepthConcatenateLayerShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthConcatenateLayerFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::DepthConcatenateLayerShapes(), framework::dataset::make("DataType", DataType::QS8))) { @@ -96,14 +96,14 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthConcatenateLayerFixture<int8_t>, framewo TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthConcatenateLayerFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), - framework::dataset::make("DataType", - DataType::QS16))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLDepthConcatenateLayerFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Tiny2DShapes(), + framework::dataset::make("DataType", + DataType::QS16))) { // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthConcatenateLayerFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(datasets::DepthConcatenateLayerShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthConcatenateLayerFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(datasets::DepthConcatenateLayerShapes(), framework::dataset::make("DataType", DataType::QS16))) { diff --git a/tests/validation/CL/DepthConvertLayer.cpp b/tests/validation/CL/DepthConvertLayer.cpp index 9c6cc46ca..603f4a07d 100644 --- a/tests/validation/CL/DepthConvertLayer.cpp +++ b/tests/validation/CL/DepthConvertLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -392,7 +392,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi validate(src.info()->padding(), padding); validate(dst.info()->padding(), padding); } -FIXTURE_DATA_TEST_CASE(RunSmallQS8, CLDepthConvertLayerToFP32FixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTinyQS8, CLDepthConvertLayerToFP32FixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), DepthConvertLayerQS8toFP32Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -400,7 +400,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallQS8, CLDepthConvertLayerToFP32FixedPointFixture<i // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunSmallQS16, CLDepthConvertLayerToFP32FixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTinyQS16, CLDepthConvertLayerToFP32FixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), DepthConvertLayerQS16toFP32Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -408,7 +408,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallQS16, CLDepthConvertLayerToFP32FixedPointFixture< // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLargeQS8, CLDepthConvertLayerToFP32FixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmallQS8, CLDepthConvertLayerToFP32FixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerQS8toFP32Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -416,7 +416,7 @@ FIXTURE_DATA_TEST_CASE(RunLargeQS8, CLDepthConvertLayerToFP32FixedPointFixture<i // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLargeQS16, CLDepthConvertLayerToFP32FixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmallQS16, CLDepthConvertLayerToFP32FixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerQS16toFP32Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -451,7 +451,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi validate(src.info()->padding(), padding); validate(dst.info()->padding(), padding); } -FIXTURE_DATA_TEST_CASE(RunSmallQS8, CLDepthConvertLayerToQS8FixedPointFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTinyQS8, CLDepthConvertLayerToQS8FixedPointFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), DepthConvertLayerFP32toQS8Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -459,7 +459,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallQS8, CLDepthConvertLayerToQS8FixedPointFixture<fl // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunSmallQS16, CLDepthConvertLayerToQS16FixedPointFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTinyQS16, CLDepthConvertLayerToQS16FixedPointFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), DepthConvertLayerFP32toQS16Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -467,7 +467,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallQS16, CLDepthConvertLayerToQS16FixedPointFixture< // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLargeQS8, CLDepthConvertLayerToQS8FixedPointFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmallQS8, CLDepthConvertLayerToQS8FixedPointFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerFP32toQS8Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -475,7 +475,7 @@ FIXTURE_DATA_TEST_CASE(RunLargeQS8, CLDepthConvertLayerToQS8FixedPointFixture<fl // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLargeQS16, CLDepthConvertLayerToQS16FixedPointFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmallQS16, CLDepthConvertLayerToQS16FixedPointFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerFP32toQS16Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) diff --git a/tests/validation/CL/DepthwiseConvolutionLayer.cpp b/tests/validation/CL/DepthwiseConvolutionLayer.cpp index 43e04fbf0..8ac882cc6 100644 --- a/tests/validation/CL/DepthwiseConvolutionLayer.cpp +++ b/tests/validation/CL/DepthwiseConvolutionLayer.cpp @@ -42,8 +42,9 @@ namespace validation { namespace { -constexpr RelativeTolerance<float> tolerance_f32(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ -constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for DataType::QASYMM8 */ +RelativeTolerance<half_float::half> tolerance_f16(half_float::half(0.001)); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ +constexpr RelativeTolerance<float> tolerance_f32(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance value for comparing reference's output against implementation's output for DataType::QASYMM8 */ } // namespace TEST_SUITE(CL) @@ -70,6 +71,23 @@ template <typename T> using CLDepthwiseConvolutionLayerFixture3x3 = DepthwiseConvolutionLayerValidationFixture<CLTensor, CLAccessor, CLDepthwiseConvolutionLayer3x3, T>; TEST_SUITE(Float) +TEST_SUITE(F16) +TEST_SUITE(W3x3) +FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture3x3<half>, framework::DatasetMode::ALL, combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), + framework::dataset::make("DataType", + DataType::F16))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerFixture3x3<half>, framework::DatasetMode::NIGHTLY, combine(datasets::LargeDepthwiseConvolutionLayerDataset3x3(), + framework::dataset::make("DataType", + DataType::F16))) +{ + validate(CLAccessor(_target), _reference, tolerance_f16); +} +TEST_SUITE_END() +TEST_SUITE_END() + TEST_SUITE(FP32) TEST_SUITE(W3x3) FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerFixture3x3<float>, framework::DatasetMode::ALL, combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), @@ -89,10 +107,26 @@ TEST_SUITE_END() TEST_SUITE_END() template <typename T> +using CLDepthwiseConvolutionLayerQuantizedFixture = DepthwiseConvolutionLayerValidationQuantizedFixture<CLTensor, CLAccessor, CLDepthwiseConvolutionLayer, T>; +template <typename T> using CLDepthwiseConvolutionLayerQuantizedFixture3x3 = DepthwiseConvolutionLayerValidationQuantizedFixture<CLTensor, CLAccessor, CLDepthwiseConvolutionLayer3x3, T>; TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) +TEST_SUITE(Generic) +FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) }))) +{ + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeDepthwiseConvolutionLayerDataset(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) }))) +{ + validate(CLAccessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() TEST_SUITE(W3x3) FIXTURE_DATA_TEST_CASE(RunSmall, CLDepthwiseConvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), framework::dataset::make("DataType", DataType::QASYMM8)), diff --git a/tests/validation/CL/DirectConvolutionLayer.cpp b/tests/validation/CL/DirectConvolutionLayer.cpp index 94e9b9520..12f3d3da3 100644 --- a/tests/validation/CL/DirectConvolutionLayer.cpp +++ b/tests/validation/CL/DirectConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -43,7 +43,6 @@ namespace validation { namespace { -// COMPMID-517 Invesitgate the mismatch to see whether it is a real bug RelativeTolerance<half> tolerance_fp16(half(0.2)); /**< Tolerance for floating point tests */ RelativeTolerance<float> tolerance_fp32(0.02f); /**< Tolerance for floating point tests */ constexpr float tolerance_num = 0.07f; /**< Tolerance number */ @@ -56,14 +55,14 @@ constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); /**< Tolerance for qu const auto data = combine(datasets::SmallDirectConvolutionShapes(), 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), + combine(concat(combine(framework::dataset::make("PadX", 0, 1), + combine(framework::dataset::make("PadY", 0, 1), 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", { 1, 4, 8, 16 }))))); -const auto data_fixed_point = combine(datasets::SmallDirectConvolutionShapes(), +const auto data_fixed_point = combine(datasets::TinyDirectConvolutionShapes(), combine(framework::dataset::make("StrideX", 1, 3), combine(framework::dataset::make("StrideY", 1, 3), combine(concat(combine(framework::dataset::make("PadX", 0), diff --git a/tests/validation/CL/EqualizeHistogram.cpp b/tests/validation/CL/EqualizeHistogram.cpp new file mode 100644 index 000000000..de78013e6 --- /dev/null +++ b/tests/validation/CL/EqualizeHistogram.cpp @@ -0,0 +1,87 @@ +/* + * 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/runtime/CL/functions/CLEqualizeHistogram.h" +#include "tests/CL/CLAccessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Macros.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/EqualizeHistogramFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +TEST_SUITE(CL) +TEST_SUITE(EqualizeHistogram) + +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), framework::dataset::make("DataType", DataType::U8)), shape, data_type) +{ + // Create tensors + CLTensor src = create_tensor<CLTensor>(shape, data_type); + CLTensor dst = create_tensor<CLTensor>(shape, data_type); + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Create and configure function + CLEqualizeHistogram equalize_histogram; + equalize_histogram.configure(&src, &dst); + + // Validate valid region + const ValidRegion valid_region = shape_to_valid_region(shape); + validate(src.info()->valid_region(), valid_region); + validate(dst.info()->valid_region(), valid_region); + + // Validate padding + const PaddingSize padding = PaddingCalculator(shape.x(), 8).required_padding(); + validate(src.info()->padding(), padding); + validate(dst.info()->padding(), padding); +} + +template <typename T> +using CLEqualizeHistogramFixture = EqualizeHistogramValidationFixture<CLTensor, CLAccessor, CLEqualizeHistogram, T>; + +FIXTURE_DATA_TEST_CASE(RunSmall, CLEqualizeHistogramFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), framework::dataset::make("DataType", + DataType::U8))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLEqualizeHistogramFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::Large2DShapes(), framework::dataset::make("DataType", + DataType::U8))) +{ + // Validate output + validate(CLAccessor(_target), _reference); +} + +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/FastCorners.cpp b/tests/validation/CL/FastCorners.cpp new file mode 100644 index 000000000..93af59d84 --- /dev/null +++ b/tests/validation/CL/FastCorners.cpp @@ -0,0 +1,116 @@ +/* + * 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/CL/functions/CLFastCorners.h" +#include "arm_compute/runtime/Tensor.h" +#include "arm_compute/runtime/TensorAllocator.h" +#include "tests/CL/CLAccessor.h" +#include "tests/CL/CLArrayAccessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/ImageFileDatasets.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/FastCornersFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +/* Radius of the Bresenham circle around the candidate point */ +const unsigned int bresenham_radius = 3; +/* Tolerance used to compare corner strengths */ +const AbsoluteTolerance<float> tolerance(0.5f); +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(FastCorners) + +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), + framework::dataset::make("Format", Format::U8)), + framework::dataset::make("SuppressNonMax", { false, true })), + framework::dataset::make("BorderMode", BorderMode::UNDEFINED)), + shape, format, suppress_nonmax, border_mode) +{ + std::mt19937 gen(library->seed()); + std::uniform_int_distribution<uint8_t> int_dist(0, 255); + std::uniform_real_distribution<float> real_dist(0, 255); + + const uint8_t constant_border_value = int_dist(gen); + const float threshold = real_dist(gen); + + // Create tensors + CLTensor src = create_tensor<CLTensor>(shape, data_type_from_format(format)); + src.info()->set_format(format); + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + + CLKeyPointArray corners; + unsigned int num_corners; + + // Create and configure function + CLFastCorners fast_corners; + fast_corners.configure(&src, threshold, suppress_nonmax, &corners, &num_corners, border_mode, constant_border_value); + + // Validate padding + PaddingCalculator calculator(shape.x(), 1); // elems_processed + + calculator.set_border_size(bresenham_radius); + calculator.set_access_offset(-bresenham_radius); + calculator.set_accessed_elements(7); // elems_read + + validate(src.info()->padding(), calculator.required_padding()); +} + +template <typename T> +using CLFastCornersFixture = FastCornersValidationFixture<CLTensor, CLAccessor, CLKeyPointArray, CLFastCorners, T>; + +FIXTURE_DATA_TEST_CASE(RunSmall, CLFastCornersFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallImageFiles(), framework::dataset::make("Format", Format::U8)), + framework::dataset::make("SuppressNonMax", { false, true })), + framework::dataset::make("BorderMode", BorderMode::UNDEFINED))) +{ + // Validate output + CLArrayAccessor<KeyPoint> array(_target); + validate_keypoints(array.buffer(), array.buffer() + array.num_values(), _reference.begin(), _reference.end(), tolerance); +} +FIXTURE_DATA_TEST_CASE(RunLarge, CLFastCornersFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeImageFiles(), framework::dataset::make("Format", Format::U8)), + framework::dataset::make("SuppressNonMax", { false, true })), + framework::dataset::make("BorderMode", BorderMode::UNDEFINED))) +{ + // Validate output + CLArrayAccessor<KeyPoint> array(_target); + validate_keypoints(array.buffer(), array.buffer() + array.num_values(), _reference.begin(), _reference.end(), tolerance); +} + +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/CL/Flatten.cpp b/tests/validation/CL/Flatten.cpp index 3350a7291..04fbef6d7 100644 --- a/tests/validation/CL/Flatten.cpp +++ b/tests/validation/CL/Flatten.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -80,13 +80,13 @@ TEST_SUITE_END() TEST_SUITE(Quantized) TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLFlattenLayerFixture<int8_t>, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::Small3DShapes(), datasets::Small4DShapes()), - framework::dataset::make("DataType", DataType::QS8))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLFlattenLayerFixture<int8_t>, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::Tiny3DShapes(), datasets::Tiny4DShapes()), + framework::dataset::make("DataType", DataType::QS8))) { // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLFlattenLayerFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(datasets::Large3DShapes(), datasets::Large4DShapes()), +FIXTURE_DATA_TEST_CASE(RunSmall, CLFlattenLayerFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(datasets::Small3DShapes(), datasets::Small4DShapes()), framework::dataset::make("DataType", DataType::QS8))) { // Validate output @@ -95,13 +95,13 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLFlattenLayerFixture<int8_t>, framework::Datas TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLFlattenLayerFixture<int16_t>, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::Small3DShapes(), datasets::Small4DShapes()), - framework::dataset::make("DataType", DataType::QS16))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLFlattenLayerFixture<int16_t>, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::Tiny3DShapes(), datasets::Tiny4DShapes()), + framework::dataset::make("DataType", DataType::QS16))) { // Validate output validate(CLAccessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLFlattenLayerFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(datasets::Large3DShapes(), datasets::Large4DShapes()), +FIXTURE_DATA_TEST_CASE(RunSmall, CLFlattenLayerFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(datasets::Small3DShapes(), datasets::Small4DShapes()), framework::dataset::make("DataType", DataType::QS16))) { // Validate output diff --git a/tests/validation/CL/FullyConnectedLayer.cpp b/tests/validation/CL/FullyConnectedLayer.cpp index aba92f14d..7db9cf5b7 100644 --- a/tests/validation/CL/FullyConnectedLayer.cpp +++ b/tests/validation/CL/FullyConnectedLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -124,15 +124,13 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( TensorInfo(TensorShape(8U, 4U, 6U, 4U), 1, DataType::F32), TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Invalid weights dimensions TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Wrongly reshaped weights - TensorInfo(TensorShape(8U, 4U, 6U, 4U), 1, DataType::F32), }), framework::dataset::make("WeightsInfo",{ TensorInfo(TensorShape(315U, 271U), 1, DataType::F16), TensorInfo(TensorShape(315U, 271U), 1, DataType::QS8, 3), TensorInfo(TensorShape(192U, 192U), 1, DataType::F32), TensorInfo(TensorShape(192U, 192U), 1, DataType::F32), + TensorInfo(TensorShape(217U, 231U), 1, DataType::F32), TensorInfo(TensorShape(217U, 315U), 1, DataType::F32), - TensorInfo(TensorShape(217U, 315U), 1, DataType::F32), - TensorInfo(TensorShape(192U, 192U), 1, DataType::F32), })), framework::dataset::make("BiasInfo",{ TensorInfo(TensorShape(271U), 1, DataType::F32), TensorInfo(TensorShape(271U), 1, DataType::QS8, 2), @@ -140,7 +138,6 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( TensorInfo(TensorShape(192U), 1, DataType::F32), TensorInfo(TensorShape(271U), 1, DataType::F32), TensorInfo(TensorShape(271U), 1, DataType::F32), - TensorInfo(TensorShape(192U), 1, DataType::F32), })), framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), TensorInfo(TensorShape(271U, 3U), 1, DataType::QS8, 3), @@ -148,11 +145,10 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( TensorInfo(TensorShape(192U, 4U), 1, DataType::F32), TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), - TensorInfo(TensorShape(192U, 4U), 1, DataType::F32), })), - framework::dataset::make("TransposeWeights",{ true, true, true, false, true, true, true })), - framework::dataset::make("ReshapedWeights",{ false, false, false, false, false, false , false})), - framework::dataset::make("Expected", { false, false, true, true, false, false, true })), + framework::dataset::make("TransposeWeights",{ true, true, true, false, true, true })), + framework::dataset::make("ReshapedWeights",{ false, false, false, false, false, false})), + framework::dataset::make("Expected", { false, false, true, true, false, false })), input_info, weights_info, bias_info, output_info, transpose_weights, reshaped_weights, expected) { Status status = CLFullyConnectedLayer::validate(&input_info.clone()->set_is_resizable(false), @@ -209,7 +205,7 @@ using CLFullyConnectedLayerFixedPointFixture = FullyConnectedLayerValidationFixe TEST_SUITE(FixedPoint) TEST_SUITE(QS8) // Testing for fixed point position [1,6) as reciprocal limits the maximum fixed point position to 5 -FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunTiny, CLFullyConnectedLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyFullyConnectedLayerDataset(), FullyConnectedParameters), framework::dataset::make("DataType", DataType::QS8)), @@ -218,7 +214,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixedPointFixture<int8_t>, // Validate output validate(CLAccessor(_target), _reference, tolerance_fixed_point); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters), framework::dataset::make("DataType", DataType::QS8)), @@ -231,7 +227,7 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) as reciprocal limits the maximum fixed point position to 14 -FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunTiny, CLFullyConnectedLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyFullyConnectedLayerDataset(), FullyConnectedParameters), framework::dataset::make("DataType", DataType::QS16)), @@ -240,7 +236,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixedPointFixture<int16_t> // Validate output validate(CLAccessor(_target), _reference, tolerance_fixed_point); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLFullyConnectedLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLFullyConnectedLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters), framework::dataset::make("DataType", DataType::QS16)), diff --git a/tests/validation/CL/GEMM.cpp b/tests/validation/CL/GEMM.cpp index 4e7b24e16..5f53d6fdd 100644 --- a/tests/validation/CL/GEMM.cpp +++ b/tests/validation/CL/GEMM.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -32,6 +32,7 @@ #include "tests/PaddingCalculator.h" #include "tests/datasets/LargeGEMMDataset.h" #include "tests/datasets/SmallGEMMDataset.h" +#include "tests/datasets/TinyGEMMDataset.h" #include "tests/framework/Asserts.h" #include "tests/framework/Macros.h" #include "tests/framework/datasets/Datasets.h" @@ -84,7 +85,7 @@ TEST_SUITE_END() // FP32 TEST_SUITE(Quantized) TEST_SUITE(QS8) using CLGEMMInterleave4x4Fixture = GEMMInterleave4x4ValidationFixedPointFixture<CLTensor, CLAccessor, CLGEMMInterleave4x4, int8_t>; -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMInterleave4x4Fixture, framework::DatasetMode::PRECOMMIT, data_interleave * +FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMInterleave4x4Fixture, framework::DatasetMode::PRECOMMIT, data_interleave * framework::dataset::make("DataType", DataType::QS8) * framework::dataset::make("FractionalBits", 1, 7)) { @@ -95,7 +96,7 @@ TEST_SUITE_END() TEST_SUITE(QS16) using CLGEMMInterleave4x4Fixture = GEMMInterleave4x4ValidationFixedPointFixture<CLTensor, CLAccessor, CLGEMMInterleave4x4, int16_t>; -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMInterleave4x4Fixture, framework::DatasetMode::PRECOMMIT, data_interleave * +FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMInterleave4x4Fixture, framework::DatasetMode::PRECOMMIT, data_interleave * framework::dataset::make("DataType", DataType::QS16) * framework::dataset::make("FractionalBits", 1, 14)) { @@ -148,7 +149,7 @@ TEST_SUITE(Quantized) TEST_SUITE(QS8) using CLGEMMTranspose1xW = CLSynthetizeFunctionWithZeroConstantBorder<CLGEMMTranspose1xWKernel, 16>; using CLGEMMTranspose1xWFixture = GEMMTranspose1xWValidationFixedPointFixture<CLTensor, CLAccessor, CLGEMMTranspose1xW, int8_t>; -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMTranspose1xWFixture, framework::DatasetMode::PRECOMMIT, data_transpose * +FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMTranspose1xWFixture, framework::DatasetMode::PRECOMMIT, data_transpose * framework::dataset::make("DataType", DataType::QS8) * framework::dataset::make("FractionalBits", 1, 7)) { @@ -160,7 +161,7 @@ TEST_SUITE_END() TEST_SUITE(QS16) using CLGEMMTranspose1xW = CLSynthetizeFunctionWithZeroConstantBorder<CLGEMMTranspose1xWKernel, 8>; using CLGEMMTranspose1xWFixture = GEMMTranspose1xWValidationFixedPointFixture<CLTensor, CLAccessor, CLGEMMTranspose1xW, int16_t>; -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMTranspose1xWFixture, framework::DatasetMode::PRECOMMIT, data_transpose * +FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMTranspose1xWFixture, framework::DatasetMode::PRECOMMIT, data_transpose * framework::dataset::make("DataType", DataType::QS16) * framework::dataset::make("FractionalBits", 1, 14)) { @@ -207,15 +208,15 @@ using CLGEMMFixedPointFixture = GEMMValidationFixedPointFixture<CLTensor, CLAcce TEST_SUITE(Quantized) TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallGEMMDataset(), - framework::dataset::make("DataType", - DataType::QS8)), - framework::dataset::make("FractionalBits", 1, 7))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::TinyGEMMDataset(), + framework::dataset::make("DataType", + DataType::QS8)), + framework::dataset::make("FractionalBits", 1, 7))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_q); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeGEMMDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SmallGEMMDataset(), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 1, 7))) @@ -226,15 +227,15 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMFixedPointFixture<int8_t>, framework::Dat TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallGEMMDataset(), - framework::dataset::make("DataType", - DataType::QS16)), - framework::dataset::make("FractionalBits", 1, 14))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLGEMMFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::TinyGEMMDataset(), + framework::dataset::make("DataType", + DataType::QS16)), + framework::dataset::make("FractionalBits", 1, 14))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_q); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLGEMMFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeGEMMDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLGEMMFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SmallGEMMDataset(), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) diff --git a/tests/validation/CL/HOGDescriptor.cpp b/tests/validation/CL/HOGDescriptor.cpp index aef265a65..f5a00ba1d 100644 --- a/tests/validation/CL/HOGDescriptor.cpp +++ b/tests/validation/CL/HOGDescriptor.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,7 @@ namespace validation { namespace { -AbsoluteTolerance<float> tolerance(1e-2f); +RelativeTolerance<float> tolerance(0.001f); } // namespace TEST_SUITE(CL) diff --git a/tests/validation/CL/HarrisCorners.cpp b/tests/validation/CL/HarrisCorners.cpp index 00b691022..890367c16 100644 --- a/tests/validation/CL/HarrisCorners.cpp +++ b/tests/validation/CL/HarrisCorners.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * diff --git a/tests/validation/CL/NormalizationLayer.cpp b/tests/validation/CL/NormalizationLayer.cpp index c133df001..bc1f44da9 100644 --- a/tests/validation/CL/NormalizationLayer.cpp +++ b/tests/validation/CL/NormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -56,6 +56,10 @@ const auto NormalizationDataset = combine(combine(combine(combine(datasets::Smal framework::dataset::make("NormalizationSize", 3, 9, 2)), framework::dataset::make("Beta", { 0.5f, 1.f, 2.f })), framework::dataset::make("IsScaled", { true })); +const auto NormalizationDatasetQS = combine(combine(combine(combine(datasets::TinyShapes(), datasets::NormalizationTypes()), + framework::dataset::make("NormalizationSize", 3, 9, 2)), + framework::dataset::make("Beta", { 0.5f, 1.f, 2.f })), + framework::dataset::make("IsScaled", { true })); const auto NormalizationDatasetFP16 = combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("NormType", { NormType::IN_MAP_1D, NormType::CROSS_MAP })), framework::dataset::make("NormalizationSize", 3, 9, 2)), framework::dataset::make("Beta", { 0.5f, 1.f, 2.f })), @@ -142,14 +146,14 @@ using CLNormalizationLayerFixedPointFixture = NormalizationValidationFixedPointF TEST_SUITE(Quantized) TEST_SUITE(QS8) // Testing for fixed point position [1,6) as reciprocal limits the maximum fixed point position to 5 -FIXTURE_DATA_TEST_CASE(RunSmall, CLNormalizationLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(NormalizationDataset, framework::dataset::make("DataType", +FIXTURE_DATA_TEST_CASE(RunTiny, CLNormalizationLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(NormalizationDatasetQS, framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 1, 6))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qs8); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLNormalizationLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(NormalizationDataset, framework::dataset::make("DataType", +FIXTURE_DATA_TEST_CASE(RunSmall, CLNormalizationLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(NormalizationDataset, framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 1, 6))) { @@ -160,14 +164,14 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) as reciprocal limits the maximum fixed point position to 5 -FIXTURE_DATA_TEST_CASE(RunSmall, CLNormalizationLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(NormalizationDataset, framework::dataset::make("DataType", +FIXTURE_DATA_TEST_CASE(RunTiny, CLNormalizationLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(NormalizationDatasetQS, framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qs16); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLNormalizationLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(NormalizationDataset, framework::dataset::make("DataType", +FIXTURE_DATA_TEST_CASE(RunSmall, CLNormalizationLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(NormalizationDataset, framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) { diff --git a/tests/validation/CL/Permute.cpp b/tests/validation/CL/Permute.cpp index 6c31ccce3..bdd8f6e44 100644 --- a/tests/validation/CL/Permute.cpp +++ b/tests/validation/CL/Permute.cpp @@ -50,6 +50,56 @@ const auto PermuteParametersLarge = combine(datasets::Large4DShapes(), TEST_SUITE(CL) TEST_SUITE(Permute) +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( + framework::dataset::make("InputInfo",{ + TensorInfo(TensorShape(7U, 7U, 5U, 3U), 1, DataType::U16), // permutation not supported + TensorInfo(TensorShape(7U, 7U, 5U, 3U), 1, DataType::U16), // permutation not supported + TensorInfo(TensorShape(7U, 7U, 5U, 3U), 1, DataType::U16), // permutation not supported + TensorInfo(TensorShape(1U, 7U), 1, DataType::U8), // invalid input size + TensorInfo(TensorShape(7U, 7U, 5U, 3U), 1, DataType::U16), // valid + TensorInfo(TensorShape(27U, 13U, 37U, 2U), 1, DataType::F32), // valid + TensorInfo(TensorShape(27U, 13U, 37U, 2U), 1, DataType::F32), // valid + TensorInfo(TensorShape(128U, 64U, 21U, 2U), 1, DataType::QASYMM8), // permutation not supported + TensorInfo(TensorShape(128U, 64U, 21U, 2U), 1, DataType::F32), // permutation not supported + TensorInfo(TensorShape(128U, 64U, 21U, 2U), 1, DataType::F32), // permutation not supported + TensorInfo(TensorShape(128U, 64U, 21U, 2U), 1, DataType::U16), // permutation not supported + }), + framework::dataset::make("OutputInfo", { + TensorInfo(TensorShape(5U, 7U, 7U, 3U), 1, DataType::U16), + TensorInfo(TensorShape(5U, 5U, 7U, 3U), 1, DataType::U16), + TensorInfo(TensorShape(7U, 7U, 7U, 3U), 1, DataType::U16), + TensorInfo(TensorShape(5U, 7U), 1, DataType::U8), + TensorInfo(TensorShape(5U, 7U, 7U, 3U), 1, DataType::U16), + TensorInfo(TensorShape(13U, 37U, 27U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(2U, 37U, 27U, 13U), 1, DataType::F32), + TensorInfo(TensorShape(128U, 64U, 21U, 2U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(128U, 64U, 21U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(21U, 64U, 2U, 128U), 1, DataType::F32), + TensorInfo(TensorShape(2U, 21U, 64U, 128U), 1, DataType::U16), + })), + framework::dataset::make("PermutationVector", { + PermutationVector(2U, 1U, 0U), + PermutationVector(2U, 2U, 1U), + PermutationVector(1U, 1U, 1U), + PermutationVector(2U, 0U, 1U), + PermutationVector(2U, 0U, 1U), + PermutationVector(1U, 2U, 0U), + PermutationVector(3U, 2U, 0U, 1U), + PermutationVector(2U, 3U, 1U, 0U), + PermutationVector(1U, 1U, 1U, 1U), + PermutationVector(2U, 1U, 3U, 0U), + PermutationVector(3U, 2U, 1U, 0U), + })), + framework::dataset::make("Expected", { false, false, false, false, true, true, true, false, false, false, false })), + input_info, output_info, perm_vect, expected) +{ + ARM_COMPUTE_EXPECT(bool(CLPermute::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), perm_vect)) == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::Small4DShapes(), framework::dataset::make("DataType", { DataType::S8, DataType::U8, DataType::S16, DataType::U16, DataType::U32, DataType::S32, DataType::F16, DataType::F32 })), shape, data_type) { diff --git a/tests/validation/CL/PixelWiseMultiplication.cpp b/tests/validation/CL/PixelWiseMultiplication.cpp index 031f10f1c..6a71175f5 100644 --- a/tests/validation/CL/PixelWiseMultiplication.cpp +++ b/tests/validation/CL/PixelWiseMultiplication.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -86,6 +86,8 @@ template <typename T> using CLPixelWiseMultiplicationToQS16Fixture = PixelWiseMultiplicationValidationFixture<CLTensor, CLAccessor, CLPixelWiseMultiplication, T, qint16_t>; template <typename T> using CLFixedPointPixelWiseMultiplicationFixture = FixedPointPixelWiseMultiplicationValidationFixture<CLTensor, CLAccessor, CLPixelWiseMultiplication, T>; +template <typename T> +using CLPixelWiseMultiplicationBroadcastFixture = PixelWiseMultiplicationBroadcastValidationFixture<CLTensor, CLAccessor, CLPixelWiseMultiplication, T, float>; TEST_SUITE(CL) TEST_SUITE(PixelWiseMultiplication) @@ -156,7 +158,7 @@ TEST_SUITE(FixedPointPixelWiseMultiplication) TEST_SUITE(QS8) TEST_SUITE(ScaleUnity) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, Fixture<qint8_t>, PRECOMMIT, SmallShapes(), QS8, QS8, scale_unity, TO_ZERO, 1, 7) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunTiny, Fixture<qint8_t>, PRECOMMIT, TinyShapes(), QS8, QS8, scale_unity, TO_ZERO, 1, 7) TEST_SUITE_END() // ScaleUnity TEST_SUITE_END() // QS8 @@ -164,11 +166,15 @@ TEST_SUITE_END() // QS8 TEST_SUITE(QS16) TEST_SUITE(ScaleUnity) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, scale_unity, TO_ZERO, 1, 15) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunTiny, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, scale_unity, TO_ZERO, 1, 15) TEST_SUITE_END() // ScaleUnity TEST_SUITE_END() // QS16 +TEST_SUITE(Broadcast) +PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, BroadcastFixture<float>, PRECOMMIT, SmallShapesBroadcast(), F32, F32, scale_255, TO_NEAREST_UP, VALIDATE(float, 1.f)) +TEST_SUITE_END() // Broadcast + TEST_SUITE_END() // FixedPointPixelWiseMultiplication TEST_SUITE_END() } // namespace validation diff --git a/tests/validation/CL/PoolingLayer.cpp b/tests/validation/CL/PoolingLayer.cpp index 4e5e5aa2e..9da4c55c7 100644 --- a/tests/validation/CL/PoolingLayer.cpp +++ b/tests/validation/CL/PoolingLayer.cpp @@ -27,6 +27,7 @@ #include "arm_compute/runtime/CL/functions/CLPoolingLayer.h" #include "tests/CL/CLAccessor.h" #include "tests/PaddingCalculator.h" +#include "tests/datasets/PoolingLayerDataset.h" #include "tests/datasets/PoolingTypesDataset.h" #include "tests/datasets/ShapeDatasets.h" #include "tests/framework/Asserts.h" @@ -43,24 +44,18 @@ namespace validation { namespace { -/** Failing data set */ -const auto PoolingLayerDatasetSpecial = ((((framework::dataset::make("Shape", TensorShape{ 60U, 52U, 3U, 5U }) - * framework::dataset::make("PoolType", PoolingType::AVG)) - * framework::dataset::make("PoolingSize", 100)) - * framework::dataset::make("PadStride", PadStrideInfo(5, 5, 50, 50))) - * framework::dataset::make("ExcludePadding", true)); /** Input data set for floating-point data types */ -const auto PoolingLayerDatasetFP = combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { 2, 3, 4, 7, 9 })), +const auto PoolingLayerDatasetFP = combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { Size2D(2, 2), Size2D(3, 3), Size2D(7, 7), Size2D(9, 9), Size2D(5, 7), Size2D(7, 9) })), framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(1, 2, 1, 1), PadStrideInfo(2, 2, 1, 0) })), framework::dataset::make("ExcludePadding", { true, false })); /** Input data set for fixed-point data types */ -const auto PoolingLayerDatasetQS = combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { 2, 3 })), +const auto PoolingLayerDatasetQS = combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { Size2D(2, 2), Size2D(3, 3) })), framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(1, 2, 1, 1), PadStrideInfo(2, 2, 1, 0) })), framework::dataset::make("ExcludePadding", { true, false })); /** Input data set for asymmetric data type */ -const auto PoolingLayerDatasetQASYMM8 = combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { 2, 3 })), +const auto PoolingLayerDatasetQASYMM8 = combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { Size2D(2, 2), Size2D(3, 3), Size2D(5, 7), Size2D(8, 9) })), framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(1, 2, 1, 1), PadStrideInfo(2, 2, 1, 0) })), framework::dataset::make("ExcludePadding", { true, false })); @@ -110,7 +105,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( PoolingLayerInfo(PoolingType::MAX), PoolingLayerInfo(PoolingType::AVG), })), - framework::dataset::make("Expected", { false, false, false, true, false, false, false, false, false, true })), + framework::dataset::make("Expected", { false, false, false, true, false, false, false, true, false, true })), input_info, output_info, pool_info, expected) { ARM_COMPUTE_EXPECT(bool(CLPoolingLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), pool_info)) == expected, framework::LogLevel::ERRORS); @@ -121,9 +116,12 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( template <typename T> using CLPoolingLayerFixture = PoolingLayerValidationFixture<CLTensor, CLAccessor, CLPoolingLayer, T>; +template <typename T> +using CLSpecialPoolingLayerFixture = SpecialPoolingLayerValidationFixture<CLTensor, CLAccessor, CLPoolingLayer, T>; + TEST_SUITE(Float) TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSpecial, CLPoolingLayerFixture<float>, framework::DatasetMode::ALL, PoolingLayerDatasetSpecial * framework::dataset::make("DataType", DataType::F32)) +FIXTURE_DATA_TEST_CASE(RunSpecial, CLSpecialPoolingLayerFixture<float>, framework::DatasetMode::ALL, datasets::PoolingLayerDatasetSpecial() * framework::dataset::make("DataType", DataType::F32)) { // Validate output validate(CLAccessor(_target), _reference, tolerance_f32); @@ -163,14 +161,14 @@ using CLPoolingLayerFixedPointFixture = PoolingLayerValidationFixedPointFixture< TEST_SUITE(FixedPoint) TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerFixedPointFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetQS, - framework::dataset::make("DataType", DataType::QS8))), - framework::dataset::make("FractionalBits", 1, 4))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLPoolingLayerFixedPointFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), combine(PoolingLayerDatasetQS, + framework::dataset::make("DataType", DataType::QS8))), + framework::dataset::make("FractionalBits", 1, 4))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qs8); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLPoolingLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), combine(PoolingLayerDatasetQS, +FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetQS, framework::dataset::make("DataType", DataType::QS8))), framework::dataset::make("FractionalBits", 1, 4))) { @@ -180,14 +178,14 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLPoolingLayerFixedPointFixture<int8_t>, framew TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerFixedPointFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetQS, - framework::dataset::make("DataType", DataType::QS16))), - framework::dataset::make("FractionalBits", 1, 12))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLPoolingLayerFixedPointFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), combine(PoolingLayerDatasetQS, + framework::dataset::make("DataType", DataType::QS16))), + framework::dataset::make("FractionalBits", 1, 12))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_qs16); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLPoolingLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), combine(PoolingLayerDatasetQS, +FIXTURE_DATA_TEST_CASE(RunSmall, CLPoolingLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetQS, framework::dataset::make("DataType", DataType::QS16))), framework::dataset::make("FractionalBits", 1, 12))) { diff --git a/tests/validation/CL/SoftmaxLayer.cpp b/tests/validation/CL/SoftmaxLayer.cpp index 62a689de1..f9204f942 100644 --- a/tests/validation/CL/SoftmaxLayer.cpp +++ b/tests/validation/CL/SoftmaxLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -93,17 +93,8 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datase CLLogits1DMaxShiftExpSumKernel::ParallelReductionInfo reduction_info = CLLogits1DMaxShiftExpSumKernel::is_parallel_reduction(shape.x()); // Validate src padding - // Legacy path used only by quantized asymmetric data type - if(is_data_type_quantized_asymmetric(data_type)) - { - const PaddingSize padding_src = PaddingCalculator(shape.x(), 16).required_padding(); - validate(src.info()->padding(), padding_src); - } - else - { - const PaddingSize padding_src = PaddingCalculator(shape.x(), std::get<1>(reduction_info)).required_padding(); - validate(src.info()->padding(), padding_src); - } + const PaddingSize padding_src = PaddingCalculator(shape.x(), std::get<1>(reduction_info)).required_padding(); + validate(src.info()->padding(), padding_src); // Validate dst padding const PaddingSize padding_dst = PaddingCalculator(shape.x(), 16).required_padding(); @@ -188,14 +179,14 @@ using CLSoftmaxLayerFixedPointFixture = SoftmaxValidationFixedPointFixture<CLTen TEST_SUITE(FixedPoint) TEST_SUITE(QS8) // Testing for fixed point position [1,6) as reciprocal limits the maximum fixed point position to 5 -FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerFixedPointFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType", - DataType::QS8)), - framework::dataset::make("FractionalBits", 1, 6))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLSoftmaxLayerFixedPointFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(datasets::SoftmaxLayerTinyShapes(), framework::dataset::make("DataType", + DataType::QS8)), + framework::dataset::make("FractionalBits", 1, 6))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fixed_point); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLSoftmaxLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SoftmaxLayerLargeShapes(), framework::dataset::make("DataType", +FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 1, 6))) { @@ -206,15 +197,15 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) as reciprocal limits the maximum fixed point position to 14 -FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerFixedPointFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::SoftmaxLayerSmallShapes(), - framework::dataset::make("DataType", - DataType::QS16)), - framework::dataset::make("FractionalBits", 1, 14))) +FIXTURE_DATA_TEST_CASE(RunTiny, CLSoftmaxLayerFixedPointFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::SoftmaxLayerTinyShapes(), + framework::dataset::make("DataType", + DataType::QS16)), + framework::dataset::make("FractionalBits", 1, 14))) { // Validate output validate(CLAccessor(_target), _reference, tolerance_fixed_point); } -FIXTURE_DATA_TEST_CASE(RunLarge, CLSoftmaxLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SoftmaxLayerLargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, CLSoftmaxLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) diff --git a/tests/validation/CPP/Permute.cpp b/tests/validation/CPP/Permute.cpp index 3341da3ac..0a97041f8 100644 --- a/tests/validation/CPP/Permute.cpp +++ b/tests/validation/CPP/Permute.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -42,7 +42,7 @@ namespace validation { namespace { -const auto PermuteParametersSmall = combine(datasets::Small4DShapes(), +const auto PermuteParametersSmall = combine(concat(concat(datasets::Small2DShapes(), datasets::Small3DShapes()), datasets::Small4DShapes()), framework::dataset::make("PermutationVector", { PermutationVector(2U, 0U, 1U), PermutationVector(1U, 2U, 0U), PermutationVector(3U, 2U, 0U, 1U) })); const auto PermuteParametersLarge = combine(datasets::Large4DShapes(), framework::dataset::make("PermutationVector", { PermutationVector(2U, 0U, 1U), PermutationVector(1U, 2U, 0U), PermutationVector(3U, 2U, 0U, 1U) })); diff --git a/tests/validation/GLES_COMPUTE/BatchNormalizationLayer.cpp b/tests/validation/GLES_COMPUTE/BatchNormalizationLayer.cpp index a82149bdc..d817fc0e6 100644 --- a/tests/validation/GLES_COMPUTE/BatchNormalizationLayer.cpp +++ b/tests/validation/GLES_COMPUTE/BatchNormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -45,6 +45,12 @@ namespace { constexpr AbsoluteTolerance<float> tolerance_f(0.00001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ constexpr AbsoluteTolerance<float> tolerance_f16(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ +const auto act_infos = framework::dataset::make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f), +}); } // namespace TEST_SUITE(GC) @@ -78,7 +84,8 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::Ran TEST_SUITE(Float) TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(Random, GCBatchNormalizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::RandomBatchNormalizationLayerDataset(), +FIXTURE_DATA_TEST_CASE(Random, GCBatchNormalizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::RandomBatchNormalizationLayerDataset(), + act_infos), framework::dataset::make("DataType", DataType::F16))) { // Validate output @@ -87,7 +94,8 @@ FIXTURE_DATA_TEST_CASE(Random, GCBatchNormalizationLayerFixture<half>, framework TEST_SUITE_END() TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(Random, GCBatchNormalizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::RandomBatchNormalizationLayerDataset(), +FIXTURE_DATA_TEST_CASE(Random, GCBatchNormalizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::RandomBatchNormalizationLayerDataset(), + act_infos), framework::dataset::make("DataType", DataType::F32))) { // Validate output diff --git a/tests/validation/GLES_COMPUTE/ConvolutionLayer.cpp b/tests/validation/GLES_COMPUTE/ConvolutionLayer.cpp index a5d1b6992..ddb597658 100644 --- a/tests/validation/GLES_COMPUTE/ConvolutionLayer.cpp +++ b/tests/validation/GLES_COMPUTE/ConvolutionLayer.cpp @@ -98,9 +98,6 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da // Validate QuantizationInfo ARM_COMPUTE_EXPECT(src.info()->quantization_info() == src_quantization_info, framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(weights.info()->quantization_info() == weights_quantization_info, framework::LogLevel::ERRORS); - - //Validate padding - //TODO(COMPMID-415) Need to validate padding? } template <typename T> @@ -125,10 +122,10 @@ FIXTURE_DATA_TEST_CASE(RunLarge, GCConvolutionLayerFixture<half>, framework::Dat validate(GCAccessor(_target), _reference, tolerance_f16, tolerance_num); } TEST_SUITE_END() +TEST_SUITE_END() TEST_SUITE_END() TEST_SUITE_END() -} } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/GLES_COMPUTE/DirectConvolutionLayerTensorShift.cpp b/tests/validation/GLES_COMPUTE/DirectConvolutionLayerTensorShift.cpp new file mode 100644 index 000000000..45fb76cad --- /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<half> tolerance_fp16(half(0.2)); /**< Tolerance for floating point tests */ +RelativeTolerance<float> 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 <typename T> +using GCDirectConvolutionLayerTensorShiftFixture = DirectConvolutionValidationTensorShiftFixture<GCTensor, GCAccessor, GCDirectConvolutionLayer, T>; + +TEST_SUITE(Float) +TEST_SUITE(FP16) +FIXTURE_DATA_TEST_CASE(Run, GCDirectConvolutionLayerTensorShiftFixture<half_float::half>, 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<float>, 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/GLES_COMPUTE/PoolingLayer.cpp b/tests/validation/GLES_COMPUTE/PoolingLayer.cpp index e789dbab0..1496ceec1 100644 --- a/tests/validation/GLES_COMPUTE/PoolingLayer.cpp +++ b/tests/validation/GLES_COMPUTE/PoolingLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -44,7 +44,7 @@ namespace validation namespace { /** Input data set for floating-point data types */ -const auto PoolingLayerDatasetFP = combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { 2, 3, 4, 7, 9 })), +const auto PoolingLayerDatasetFP = combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { Size2D(2, 2), Size2D(3, 3), Size2D(4, 4), Size2D(7, 7), Size2D(9, 9) })), framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(1, 2, 1, 1), PadStrideInfo(2, 2, 1, 0) })), framework::dataset::make("ExcludePadding", { true, false })); diff --git a/tests/validation/NEON/ActivationLayer.cpp b/tests/validation/NEON/ActivationLayer.cpp index 8a918b259..e77fe5a29 100644 --- a/tests/validation/NEON/ActivationLayer.cpp +++ b/tests/validation/NEON/ActivationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -213,15 +213,15 @@ TEST_SUITE(FixedPoint) TEST_SUITE(QS8) // We test for fixed point precision [3,5] because [1,2] and [6,7] ranges cause // overflowing issues in most of the transcendentals functions. -FIXTURE_DATA_TEST_CASE(RunSmall, NEActivationLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ActivationDataset), - framework::dataset::make("DataType", - DataType::QS8)), - framework::dataset::make("FractionalBits", 3, 6))) +FIXTURE_DATA_TEST_CASE(RunTiny, NEActivationLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), ActivationDataset), + framework::dataset::make("DataType", + DataType::QS8)), + framework::dataset::make("FractionalBits", 3, 6))) { // Validate output validate(Accessor(_target), _reference, tolerance(_data_type, _function)); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEActivationLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ActivationDataset), +FIXTURE_DATA_TEST_CASE(RunSmall, NEActivationLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), ActivationDataset), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 3, 6))) @@ -233,15 +233,15 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) as reciprocal limits the maximum fixed point position to 14 -FIXTURE_DATA_TEST_CASE(RunSmall, NEActivationLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ActivationDataset), - framework::dataset::make("DataType", - DataType::QS16)), - framework::dataset::make("FractionalBits", 1, 14))) +FIXTURE_DATA_TEST_CASE(RunTiny, NEActivationLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), ActivationDataset), + framework::dataset::make("DataType", + DataType::QS16)), + framework::dataset::make("FractionalBits", 1, 14))) { // Validate output validate(Accessor(_target), _reference, tolerance(_data_type, _function)); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEActivationLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ActivationDataset), +FIXTURE_DATA_TEST_CASE(RunSmall, NEActivationLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), ActivationDataset), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) @@ -256,7 +256,11 @@ template <typename T> using NEActivationLayerQuantizedFixture = ActivationValidationQuantizedFixture<Tensor, Accessor, NEActivationLayer, T>; /** Input data sets. */ -const auto QuantizedActivationDataset = combine(combine(framework::dataset::make("InPlace", { false, true }), framework::dataset::make("ActivationFunction", { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU })), +const auto QuantizedActivationFunctionsDataset = framework::dataset::make("ActivationFunction", { ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, + ActivationLayerInfo::ActivationFunction::RELU + }); + +const auto QuantizedActivationDataset = combine(combine(framework::dataset::make("InPlace", { false, true }), QuantizedActivationFunctionsDataset), framework::dataset::make("AlphaBeta", { 0.5f, 1.f })); TEST_SUITE(Quantized) diff --git a/tests/validation/NEON/ArithmeticAddition.cpp b/tests/validation/NEON/ArithmeticAddition.cpp index e20e8df66..98eeaa2f6 100644 --- a/tests/validation/NEON/ArithmeticAddition.cpp +++ b/tests/validation/NEON/ArithmeticAddition.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -178,7 +178,7 @@ using NEArithmeticAdditionFixedPointFixture = ArithmeticAdditionValidationFixedP TEST_SUITE(Quantized) TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionQS8Dataset), +FIXTURE_DATA_TEST_CASE(RunTiny, NEArithmeticAdditionFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), ArithmeticAdditionQS8Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 7))) { @@ -186,7 +186,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixedPointFixture<int8_t>, validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ArithmeticAdditionQS8Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionQS8Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 7))) { @@ -196,7 +196,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixedPointFixture<int8_t>, TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionQS16Dataset), +FIXTURE_DATA_TEST_CASE(RunTiny, NEArithmeticAdditionFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), ArithmeticAdditionQS16Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 15))) { @@ -204,7 +204,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixedPointFixture<int16_t>, validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), ArithmeticAdditionQS16Dataset), +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticAdditionFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), ArithmeticAdditionQS16Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 15))) { @@ -263,6 +263,25 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticAdditionFixture<float>, framework:: // Validate output validate(Accessor(_target), _reference); } + +template <typename T> +using NEArithmeticAdditionBroadcastFixture = ArithmeticAdditionBroadcastValidationFixture<Tensor, Accessor, NEArithmeticAddition, T>; + +FIXTURE_DATA_TEST_CASE(RunSmallBroadcast, NEArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallShapesBroadcast(), + ArithmeticAdditionFP32Dataset), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +{ + // Validate output + validate(Accessor(_target), _reference); +} + +FIXTURE_DATA_TEST_CASE(RunLargeBroadcast, NEArithmeticAdditionBroadcastFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapesBroadcast(), + ArithmeticAdditionFP32Dataset), + framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP }))) +{ + // Validate output + validate(Accessor(_target), _reference); +} TEST_SUITE_END() TEST_SUITE_END() diff --git a/tests/validation/NEON/ArithmeticSubtraction.cpp b/tests/validation/NEON/ArithmeticSubtraction.cpp index f5a50335d..01a107c17 100644 --- a/tests/validation/NEON/ArithmeticSubtraction.cpp +++ b/tests/validation/NEON/ArithmeticSubtraction.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -248,7 +248,7 @@ using NEArithmeticSubtractionFixedPointFixture = ArithmeticSubtractionValidation TEST_SUITE(Quantized) TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTiny, NEArithmeticSubtractionFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), ArithmeticSubtractionQS8Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 7))) @@ -257,7 +257,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixedPointFixture<int8_t validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionQS8Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 7))) @@ -268,7 +268,7 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionFixedPointFixture<int8_t TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTiny, NEArithmeticSubtractionFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), ArithmeticSubtractionQS16Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 15))) @@ -277,7 +277,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixedPointFixture<int16_ validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEArithmeticSubtractionFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEArithmeticSubtractionFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), ArithmeticSubtractionQS16Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), framework::dataset::make("FractionalBits", 1, 15))) diff --git a/tests/validation/NEON/BatchNormalizationLayer.cpp b/tests/validation/NEON/BatchNormalizationLayer.cpp index dfa32bbb0..054ed278a 100644 --- a/tests/validation/NEON/BatchNormalizationLayer.cpp +++ b/tests/validation/NEON/BatchNormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -49,6 +49,12 @@ constexpr AbsoluteTolerance<float> tolerance_f16(0.01f); /**< Tolerance value fo #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ constexpr AbsoluteTolerance<float> tolerance_qs8(3.0f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::QS8 */ constexpr AbsoluteTolerance<float> tolerance_qs16(6.0f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::QS16 */ +const auto act_infos = framework::dataset::make("ActivationInfo", +{ + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 8.f, 2.f), +}); } // namespace TEST_SUITE(NEON) @@ -82,13 +88,15 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::Ran // *INDENT-OFF* // clang-format off -DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip( framework::dataset::make("InputInfo", { TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(27U, 13U, 2U), 1, DataType::F32), // Window shrink TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching data types TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Mismatching data types TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Invalid mean/var/beta/gamma shape TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2), // Mismatching fixed point position + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2), // Fused activation with fixed point not supported + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), // Fused activation's a < b TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2), }), @@ -98,6 +106,8 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F16), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 3), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::F32), + TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 3), TensorInfo(TensorShape(32U, 13U, 2U), 1, DataType::QS8, 2), TensorInfo(), })), @@ -108,10 +118,23 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(5U), 1, DataType::F32), TensorInfo(TensorShape(2U), 1, DataType::QS8, 2), TensorInfo(TensorShape(2U), 1, DataType::QS8, 2), + TensorInfo(TensorShape(2U), 1, DataType::F32), + TensorInfo(TensorShape(2U), 1, DataType::QS8, 2), TensorInfo(TensorShape(2U), 1, DataType::QS8, 2), })), - framework::dataset::make("Expected", { true, false, false, false, false, false, true, true})), - input_info, output_info, mvbg_info, expected) + framework::dataset::make("ActivationLayerInfo",{ ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::RELU), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::BOUNDED_RELU, 6.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f, 2.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 6.f, 2.f), + ActivationLayerInfo(ActivationLayerInfo::ActivationFunction::LU_BOUNDED_RELU, 2.f, 6.f), + ActivationLayerInfo(), + ActivationLayerInfo(), + })), + framework::dataset::make("Expected", { true, false, false, false, false, false, false, false, true, true})), + input_info, output_info, mvbg_info, act_info, expected) { const auto &mean_info = mvbg_info; const auto &var_info = mvbg_info; @@ -120,14 +143,15 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( bool has_error = bool(NEBatchNormalizationLayer::validate( &input_info.clone()->set_is_resizable(false), output_info.total_size() ? &output_info.clone()->set_is_resizable(false) : nullptr, &mean_info.clone()->set_is_resizable(false), &var_info.clone()->set_is_resizable(false), - &beta_info.clone()->set_is_resizable(false), &gamma_info.clone()->set_is_resizable(false), 1.f)); + &beta_info.clone()->set_is_resizable(false), &gamma_info.clone()->set_is_resizable(false), 1.f, act_info)); ARM_COMPUTE_EXPECT(has_error == expected, framework::LogLevel::ERRORS); } // clang-format on // *INDENT-ON* TEST_SUITE(Float) -FIXTURE_DATA_TEST_CASE(Random, NEBatchNormalizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(datasets::RandomBatchNormalizationLayerDataset(), +FIXTURE_DATA_TEST_CASE(Random, NEBatchNormalizationLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::RandomBatchNormalizationLayerDataset(), + act_infos), framework::dataset::make("DataType", DataType::F32))) { // Validate output @@ -137,7 +161,8 @@ TEST_SUITE_END() #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(Float16) -FIXTURE_DATA_TEST_CASE(Random, NEBatchNormalizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(datasets::RandomBatchNormalizationLayerDataset(), +FIXTURE_DATA_TEST_CASE(Random, NEBatchNormalizationLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::RandomBatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), framework::dataset::make("DataType", DataType::F16))) { // Validate output @@ -151,7 +176,8 @@ template <typename T> using NEBatchNormalizationLayerFixedPointFixture = BatchNormalizationLayerValidationFixedPointFixture<Tensor, Accessor, NEBatchNormalizationLayer, T>; TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(Random, NEBatchNormalizationLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::RandomBatchNormalizationLayerDataset(), +FIXTURE_DATA_TEST_CASE(Random, NEBatchNormalizationLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::RandomBatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 1, 6))) { @@ -161,7 +187,8 @@ FIXTURE_DATA_TEST_CASE(Random, NEBatchNormalizationLayerFixedPointFixture<int8_t TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(Random, NEBatchNormalizationLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::RandomBatchNormalizationLayerDataset(), +FIXTURE_DATA_TEST_CASE(Random, NEBatchNormalizationLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::RandomBatchNormalizationLayerDataset(), + framework::dataset::make("ActivationInfo", ActivationLayerInfo())), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) { diff --git a/tests/validation/NEON/ChannelExtract.cpp b/tests/validation/NEON/ChannelExtract.cpp new file mode 100644 index 000000000..8bf43d43e --- /dev/null +++ b/tests/validation/NEON/ChannelExtract.cpp @@ -0,0 +1,151 @@ +/* + * 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/MultiImage.h" +#include "arm_compute/runtime/NEON/functions/NEChannelExtract.h" +#include "arm_compute/runtime/Tensor.h" +#include "arm_compute/runtime/TensorAllocator.h" +#include "tests/NEON/Accessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/ConvertPolicyDataset.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/ChannelExtractFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +// Input data sets +const auto ChannelExtractRGBADataset = combine(framework::dataset::make("FormatType", { Format::RGBA8888 }), + framework::dataset::make("ChannelType", { Channel::R, Channel::G, Channel::B, Channel::A })); +const auto ChannelExtractYUVDataset = combine(framework::dataset::make("FormatType", { Format::YUYV422, Format::UYVY422 }), + framework::dataset::make("ChannelType", { Channel::Y, Channel::U, Channel::V })); +const auto ChannelExtractYUVPlanarDataset = combine(framework::dataset::make("FormatType", { Format::IYUV, Format::YUV444, Format::NV12, Format::NV21 }), + framework::dataset::make("ChannelType", { Channel::Y, Channel::U, Channel::V })); + +inline void validate_configuration(const TensorShape &shape, Format format, Channel channel) +{ + const unsigned int num_planes = num_planes_from_format(format); + + TensorShape dst_shape = adjust_odd_shape(shape, format); + dst_shape = calculate_subsampled_shape(dst_shape, format, channel); + + // Create tensors + MultiImage ref_src = create_multi_image<MultiImage>(shape, format); + Tensor dst = create_tensor<Tensor>(dst_shape, Format::U8); + + // Create and Configure function + NEChannelExtract channel_extract; + + if(1U == num_planes) + { + const Tensor *plane_src = ref_src.plane(0); + + channel_extract.configure(plane_src, channel, &dst); + } + else + { + channel_extract.configure(&ref_src, channel, &dst); + } +} +} // namespace + +TEST_SUITE(NEON) +TEST_SUITE(ChannelExtract) + +template <typename T> +using NEChannelExtractFixture = ChannelExtractValidationFixture<MultiImage, Tensor, Accessor, NEChannelExtract, T>; + +TEST_SUITE(Configuration) +DATA_TEST_CASE(RGBA, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), ChannelExtractRGBADataset), + shape, format, channel) +{ + validate_configuration(shape, format, channel); +} +DATA_TEST_CASE(YUV, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), ChannelExtractYUVDataset), + shape, format, channel) +{ + validate_configuration(shape, format, channel); +} + +DATA_TEST_CASE(YUVPlanar, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), ChannelExtractYUVPlanarDataset), + shape, format, channel) +{ + validate_configuration(shape, format, channel); +} +TEST_SUITE_END() + +TEST_SUITE(RGBA) +FIXTURE_DATA_TEST_CASE(RunSmall, NEChannelExtractFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), ChannelExtractRGBADataset)) +{ + // Validate output + validate(Accessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, NEChannelExtractFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::Large2DShapes(), ChannelExtractRGBADataset)) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() + +TEST_SUITE(YUV) +FIXTURE_DATA_TEST_CASE(RunSmall, NEChannelExtractFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), ChannelExtractYUVDataset)) +{ + // Validate output + validate(Accessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, NEChannelExtractFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::Large2DShapes(), ChannelExtractYUVDataset)) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() + +TEST_SUITE(YUVPlanar) +FIXTURE_DATA_TEST_CASE(RunSmall, NEChannelExtractFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), ChannelExtractYUVPlanarDataset)) +{ + // Validate output + validate(Accessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, NEChannelExtractFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::Large2DShapes(), ChannelExtractYUVPlanarDataset)) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() + +TEST_SUITE_END() +TEST_SUITE_END() + +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/NEON/Convolution.cpp b/tests/validation/NEON/Convolution.cpp index 5af803010..5545b7f57 100644 --- a/tests/validation/NEON/Convolution.cpp +++ b/tests/validation/NEON/Convolution.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -49,24 +49,24 @@ namespace * while reference performs direct division with scale. */ constexpr AbsoluteTolerance<uint8_t> tolerance_u8(1); +constexpr AbsoluteTolerance<int16_t> tolerance_s16(1); } // namespace TEST_SUITE(NEON) TEST_SUITE(CustomConvolution) -TEST_SUITE(CustomConvolutionSuqare) -TEST_SUITE(CustomConvolution3x3) +TEST_SUITE(Square3x3) -DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", DataType::U8)), +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", { DataType::U8, DataType::S16 })), datasets::BorderModes()), framework::dataset::make("filter_size", { 3 })), - shape, data_type, border_mode, filter_size) + shape, output_data_type, border_mode, filter_size) { // Create tensors - Tensor src = create_tensor<Tensor>(shape, data_type); - Tensor dst = create_tensor<Tensor>(shape, data_type); + Tensor src = create_tensor<Tensor>(shape, DataType::U8); + Tensor dst = create_tensor<Tensor>(shape, output_data_type); // Create conv matrix - int16_t conv[9]; + int16_t conv[9] = {}; ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); @@ -98,6 +98,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi template <typename T> using NEConvolutionFixture = ConvolutionSquareValidationFixture<Tensor, Accessor, NEConvolution3x3, T>; +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -115,20 +116,41 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<uint8_t>, framework::Datas // Validate output validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_u8); } -TEST_SUITE_END() /* Custom Convolution3x3 */ +TEST_SUITE_END() + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 3 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 3 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Square3x3 */ -TEST_SUITE(CustomConvolution5x5) -DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", DataType::U8)), +TEST_SUITE(Square5x5) +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", { DataType::U8, DataType::S16 })), datasets::BorderModes()), framework::dataset::make("filter_size", { 5 })), - shape, data_type, border_mode, filter_size) + shape, output_data_type, border_mode, filter_size) { // Create tensors - Tensor src = create_tensor<Tensor>(shape, data_type); - Tensor dst = create_tensor<Tensor>(shape, data_type); + Tensor src = create_tensor<Tensor>(shape, DataType::U8); + Tensor dst = create_tensor<Tensor>(shape, output_data_type); // Create conv matrix - int16_t conv[25]; + int16_t conv[25] = {}; ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); @@ -160,6 +182,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi template <typename T> using NEConvolutionFixture = ConvolutionSquareValidationFixture<Tensor, Accessor, NEConvolution5x5, T>; +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -177,20 +200,41 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<uint8_t>, framework::Datas // Validate output validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_u8); } -TEST_SUITE_END() /* Custom Convolution 5x5 */ +TEST_SUITE_END() -TEST_SUITE(CustomConvolution7x7) -DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", DataType::U8)), +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 5 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 5 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Square5x5 */ + +TEST_SUITE(Square7x7) +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", { DataType::U8, DataType::S16 })), datasets::BorderModes()), framework::dataset::make("filter_size", { 7 })), - shape, data_type, border_mode, filter_size) + shape, output_data_type, border_mode, filter_size) { // Create tensors - Tensor src = create_tensor<Tensor>(shape, data_type); - Tensor dst = create_tensor<Tensor>(shape, data_type); + Tensor src = create_tensor<Tensor>(shape, DataType::U8); + Tensor dst = create_tensor<Tensor>(shape, output_data_type); // Create conv matrix - int16_t conv[49]; + int16_t conv[49] = {}; ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); @@ -222,6 +266,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi template <typename T> using NEConvolutionFixture = ConvolutionSquareValidationFixture<Tensor, Accessor, NEConvolution7x7, T>; +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -239,20 +284,41 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<uint8_t>, framework::Datas // Validate output validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_u8); } -TEST_SUITE_END() /* Custom Convolution 7x7 */ +TEST_SUITE_END() -TEST_SUITE(CustomConvolution9x9) -DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", DataType::U8)), +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 7 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 7 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Square7x7 */ + +TEST_SUITE(Square9x9) +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", { DataType::U8, DataType::S16 })), datasets::BorderModes()), framework::dataset::make("filter_size", { 9 })), - shape, data_type, border_mode, filter_size) + shape, output_data_type, border_mode, filter_size) { // Create tensors - Tensor src = create_tensor<Tensor>(shape, data_type); - Tensor dst = create_tensor<Tensor>(shape, data_type); + Tensor src = create_tensor<Tensor>(shape, DataType::U8); + Tensor dst = create_tensor<Tensor>(shape, output_data_type); // Create conv matrix - int16_t conv[81]; + int16_t conv[81] = {}; ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); @@ -284,6 +350,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi template <typename T> using NEConvolutionFixture = ConvolutionSquareValidationFixture<Tensor, Accessor, NEConvolution9x9, T>; +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -301,31 +368,50 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<uint8_t>, framework::Datas // Validate output validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_u8); } -TEST_SUITE_END() /* Custom Convolution 9x9 */ -TEST_SUITE_END() /* Custom Convolution Square */ +TEST_SUITE_END() + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 9 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} -TEST_SUITE(CustomConvolutionRectangle) +FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 9 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Square9x9 */ +TEST_SUITE(Rectangle) DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(combine(concat(datasets::SmallShapes(), datasets::LargeShapes()), framework::dataset::make("DataType", - DataType::U8)), - datasets::BorderModes()), - framework::dataset::make("filter_width", { 3, 5, 7, 9 })), - framework::dataset::make("filter_height", { 3, 5, 7, 9 })), - shape, data_type, border_mode, filter_width, filter_height) +{ DataType::U8, DataType::S16 })), +datasets::BorderModes()), +framework::dataset::make("filter_width", { 3, 5, 7, 9 })), +framework::dataset::make("filter_height", { 3, 5, 7, 9 })), +shape, output_data_type, border_mode, filter_width, filter_height) { // Create tensors - Tensor src = create_tensor<Tensor>(shape, data_type); - Tensor dst = create_tensor<Tensor>(shape, data_type); + Tensor src = create_tensor<Tensor>(shape, DataType::U8); + Tensor dst = create_tensor<Tensor>(shape, output_data_type); // Create conv matrix - int16_t conv[filter_width * filter_height]; + std::vector<int16_t> conv(filter_height * filter_width); ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); // Create and configure function NEConvolutionRectangle convolution; - convolution.configure(&src, &dst, conv, filter_width, filter_height, 1, border_mode); + convolution.configure(&src, &dst, conv.data(), filter_width, filter_height, 1, border_mode); // Validate valid region const ValidRegion dst_valid_region = shape_to_valid_region(shape, (border_mode == BorderMode::UNDEFINED), BorderSize(filter_height / 2, filter_width / 2)); @@ -354,6 +440,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi template <typename T> using NEConvolutionFixture = ConvolutionRectangleValidationFixture<Tensor, Accessor, NEConvolutionRectangle, T>; +TEST_SUITE(U8) FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", DataType::U8)), datasets::BorderModes()), @@ -374,7 +461,166 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<uint8_t>, framework::Datas validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_u8); } TEST_SUITE_END() + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_width", { 3, 5, 7, 9 })), + framework::dataset::make("filter_height", { 3, 5, 7, 9 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_width", { 3, 5, 7, 9 })), + framework::dataset::make("filter_height", { 3, 5, 7, 9 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Rectangle */ + +TEST_SUITE(Separable5x5) +template <typename T> +using NEConvolutionFixture = ConvolutionSeparableValidationFixture<Tensor, Accessor, NEConvolution5x5, T>; + +TEST_SUITE(U8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::U8)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 5 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_u8); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::U8)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 5 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_u8); +} +TEST_SUITE_END() + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 5 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 5 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Separable5x5 */ + +TEST_SUITE(Separable7x7) +template <typename T> +using NEConvolutionFixture = ConvolutionSeparableValidationFixture<Tensor, Accessor, NEConvolution7x7, T>; + +TEST_SUITE(U8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::U8)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 7 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_u8); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::U8)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 7 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_u8); +} +TEST_SUITE_END() + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 7 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 7 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} +TEST_SUITE_END() +TEST_SUITE_END() /* Separable7x7 */ + +TEST_SUITE(Separable9x9) +template <typename T> +using NEConvolutionFixture = ConvolutionSeparableValidationFixture<Tensor, Accessor, NEConvolution9x9, T>; + +TEST_SUITE(U8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::U8)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 9 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_u8); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::U8)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 9 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_u8); +} +TEST_SUITE_END() + +TEST_SUITE(S16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 9 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), framework::dataset::make("DataType", + DataType::S16)), + datasets::BorderModes()), + framework::dataset::make("filter_size", { 9 }))) +{ + // Validate output + validate(Accessor(_target), _reference, shape_to_valid_region(_reference.shape(), (_border_mode == BorderMode::UNDEFINED), BorderSize(_height / 2, _width / 2)), tolerance_s16); +} TEST_SUITE_END() +TEST_SUITE_END() /* Separable9x9 */ + +TEST_SUITE_END() /* Custom Convolution */ TEST_SUITE_END() } // namespace validation } // namespace test diff --git a/tests/validation/NEON/ConvolutionLayer.cpp b/tests/validation/NEON/ConvolutionLayer.cpp index 1fd28df40..eabc6ad59 100644 --- a/tests/validation/NEON/ConvolutionLayer.cpp +++ b/tests/validation/NEON/ConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -23,6 +23,7 @@ */ #include "arm_compute/core/Types.h" #include "arm_compute/runtime/NEON/functions/NEConvolutionLayer.h" +#include "arm_compute/runtime/NEON/functions/NEGEMMConvolutionLayer.h" #include "arm_compute/runtime/NEON/functions/NEWinogradLayer.h" #include "arm_compute/runtime/Tensor.h" #include "arm_compute/runtime/TensorAllocator.h" @@ -30,6 +31,7 @@ #include "tests/PaddingCalculator.h" #include "tests/datasets/LargeConvolutionLayerDataset.h" #include "tests/datasets/SmallConvolutionLayerDataset.h" +#include "tests/datasets/TinyConvolutionLayerDataset.h" #include "tests/framework/Asserts.h" #include "tests/framework/Macros.h" #include "tests/framework/datasets/Datasets.h" @@ -47,9 +49,10 @@ namespace { const AbsoluteTolerance<float> tolerance_f32(0.001f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F32 */ #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC -const AbsoluteTolerance<float> tolerance_f16(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ -#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ -const AbsoluteTolerance<float> tolerance_q(1.0f); /**< Tolerance value for comparing reference's output against implementation's output for fixed point data types */ +const AbsoluteTolerance<float> tolerance_f16(0.01f); /**< Tolerance value for comparing reference's output against implementation's output for DataType::F16 */ +#endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ +const AbsoluteTolerance<float> tolerance_q(1.0f); /**< Tolerance value for comparing reference's output against implementation's output for fixed point data types */ +constexpr AbsoluteTolerance<float> tolerance_qasymm8(0.0); /**< Tolerance value for comparing reference's output against implementation's output for quantized data types */ /** CNN data types */ const auto CNNDataTypes = framework::dataset::make("DataType", @@ -60,12 +63,50 @@ const auto CNNDataTypes = framework::dataset::make("DataType", DataType::F32, DataType::QS8, DataType::QS16, + DataType::QASYMM8, }); } // namespace TEST_SUITE(NEON) -#if defined(__aarch64__) +TEST_SUITE(ConvolutionLayer) +DATA_TEST_CASE(ValidateConvolutionMethod, framework::DatasetMode::ALL, zip(zip(zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(8U, 8U, 2U), 1, DataType::F32, 0), + TensorInfo(TensorShape(23U, 27U, 5U, 4U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 2U, 1U), 1, DataType::F32, 0), + TensorInfo(TensorShape(33U, 27U, 7U, 4U), 1, DataType::F32, 0) + }), + framework::dataset::make("WeightsInfo", { TensorInfo(TensorShape(3U, 3U, 5U, 21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 5U, 21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(3U, 3U, 5U, 21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(5U, 5U, 7U, 16U), 1, DataType::F16, 0) + })), + framework::dataset::make("BiasesInfo", { TensorInfo(TensorShape(1U), 1, DataType::F32, 0), + TensorInfo(TensorShape(21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(16U), 1, DataType::F32, 0) + })), + framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(6U, 6U, 1U), 1, DataType::F32, 0), + TensorInfo(TensorShape(21U, 25U, 21U, 4U), 1, DataType::F32, 0), + TensorInfo(TensorShape(11U, 25U, 21U), 1, DataType::F32, 0), + TensorInfo(TensorShape(11U, 12U, 16U, 4U), 1, DataType::F32, 0) + })), + framework::dataset::make("ConvInfo", { PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(1, 1, 0, 0), + PadStrideInfo(2, 1, 0, 0), + PadStrideInfo(3, 2, 1, 0) + })), + framework::dataset::make("Expected", { ConvolutionMethod::WINOGRAD, ConvolutionMethod::WINOGRAD, ConvolutionMethod::GEMM, ConvolutionMethod::GEMM })), + input_info, weights_info, biases_info, output_info, conv_info, expected) +{ + ConvolutionMethod is_valid = NEConvolutionLayer::get_convolution_method(&input_info.clone()->set_is_resizable(false), + &weights_info.clone()->set_is_resizable(false), + &biases_info.clone()->set_is_resizable(false), + &output_info.clone()->set_is_resizable(false), conv_info); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} +TEST_SUITE_END() + TEST_SUITE(WinogradLayer) template <typename T> using NEWinogradLayerFixture = WinogradLayerValidationFixture<Tensor, Accessor, NEWinogradLayer, T>; @@ -79,9 +120,8 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEWinogradLayerFixture<float>, framework::Datas TEST_SUITE_END() TEST_SUITE_END() -#endif /* __aarch64__ */ -TEST_SUITE(ConvolutionLayer) +TEST_SUITE(GEMMConvolutionLayer) DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallConvolutionLayerDataset(), datasets::LargeConvolutionLayerDataset()), CNNDataTypes), input_shape, weights_shape, bias_shape, output_shape, info, data_type) @@ -89,19 +129,24 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da // Set fixed point position data type allowed int fixed_point_position = is_data_type_fixed_point(data_type) ? 3 : 0; + auto bias_data_type = is_data_type_quantized_asymmetric(data_type) ? DataType::S32 : data_type; + // Create tensors - Tensor src = create_tensor<Tensor>(input_shape, data_type, 1, fixed_point_position); - Tensor weights = create_tensor<Tensor>(weights_shape, data_type, 1, fixed_point_position); - Tensor bias = create_tensor<Tensor>(bias_shape, data_type, 1, fixed_point_position); - Tensor dst = create_tensor<Tensor>(output_shape, data_type, 1, fixed_point_position); + Tensor src = create_tensor<Tensor>(input_shape, data_type, 1, fixed_point_position, QuantizationInfo(2.f / 255.f, 127)); + Tensor weights = create_tensor<Tensor>(weights_shape, data_type, 1, fixed_point_position, QuantizationInfo(2.f / 255.f, 127)); + Tensor bias = create_tensor<Tensor>(bias_shape, bias_data_type, 1, fixed_point_position, QuantizationInfo(2.f / 255.f, 127)); + Tensor dst = create_tensor<Tensor>(output_shape, data_type, 1, fixed_point_position, QuantizationInfo(2.f / 255.f, 127)); 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); + const QuantizationInfo src_quantization_info = src.info()->quantization_info(); + const QuantizationInfo weights_quantization_info = weights.info()->quantization_info(); + // Create and configure function - NEConvolutionLayer conv; + NEGEMMConvolutionLayer conv; conv.configure(&src, &weights, &bias, &dst, info); // Validate valid region @@ -114,24 +159,28 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da validate(weights.info()->valid_region(), weights_valid_region); validate(bias.info()->valid_region(), bias_valid_region); validate(dst.info()->valid_region(), dst_valid_region); + + // Validate QuantizationInfo + ARM_COMPUTE_EXPECT(src.info()->quantization_info() == src_quantization_info, framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(weights.info()->quantization_info() == weights_quantization_info, framework::LogLevel::ERRORS); } template <typename T> -using NEConvolutionLayerFixture = ConvolutionValidationFixture<Tensor, Accessor, NEConvolutionLayer, T>; +using NEGEMMConvolutionLayerFixture = ConvolutionValidationFixture<Tensor, Accessor, NEConvolutionLayer, T>; TEST_SUITE(Float) #ifdef __ARM_FEATURE_FP16_VECTOR_ARITHMETIC TEST_SUITE(FP16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::F16))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture<half>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::F16))) { // Validate output validate(Accessor(_target), _reference, tolerance_f16); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::F16))) +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMConvolutionLayerFixture<half>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::F16))) { // Validate output validate(Accessor(_target), _reference, tolerance_f16); @@ -140,16 +189,16 @@ TEST_SUITE_END() #endif /* __ARM_FEATURE_FP16_VECTOR_ARITHMETIC */ TEST_SUITE(FP32) -FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::F32))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::F32))) { // Validate output validate(Accessor(_target), _reference, tolerance_f32); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::F32))) +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMConvolutionLayerFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::F32))) { // Validate output validate(Accessor(_target), _reference, tolerance_f32); @@ -158,12 +207,12 @@ TEST_SUITE_END() TEST_SUITE_END() template <typename T> -using NEConvolutionLayerFixedPointFixture = ConvolutionValidationFixedPointFixture<Tensor, Accessor, NEConvolutionLayer, T>; +using NEGEMMConvolutionLayerFixedPointFixture = ConvolutionValidationFixedPointFixture<Tensor, Accessor, NEGEMMConvolutionLayer, T>; -TEST_SUITE(Quantized) +TEST_SUITE(FixedPoint) TEST_SUITE(QS8) // We test for fixed point precision [4,6] -FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true, false })), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 4, 7))) @@ -171,10 +220,10 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionLayerFixedPointFixture<int8_t>, fr // Validate output validate(Accessor(_target), _reference, tolerance_q); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::QS8)), - framework::dataset::make("FractionalBits", 4, 7))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::QS8)), + framework::dataset::make("FractionalBits", 4, 7))) { // Validate output validate(Accessor(_target), _reference, tolerance_q); @@ -183,7 +232,7 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) -FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyConvolutionLayerDataset(), framework::dataset::make("ReshapeWeights", { true, false })), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) @@ -191,10 +240,10 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEConvolutionLayerFixedPointFixture<int16_t>, f // Validate output validate(Accessor(_target), _reference, tolerance_q); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), - framework::dataset::make("ReshapeWeights", { true, false })), - framework::dataset::make("DataType", DataType::QS16)), - framework::dataset::make("FractionalBits", 1, 14))) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true, false })), + framework::dataset::make("DataType", DataType::QS16)), + framework::dataset::make("FractionalBits", 1, 14))) { // Validate output validate(Accessor(_target), _reference, tolerance_q); @@ -202,6 +251,30 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEConvolutionLayerFixedPointFixture<int16_t>, f TEST_SUITE_END() TEST_SUITE_END() +template <typename T> +using NEGEMMConvolutionLayerQuantizedFixture = ConvolutionValidationQuantizedFixture<Tensor, Accessor, NEGEMMConvolutionLayer, T>; + +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeConvolutionLayerDataset(), + framework::dataset::make("ReshapeWeights", { true })), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(2.f / 255.f, 10) }))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() +TEST_SUITE_END() + TEST_SUITE_END() TEST_SUITE_END() } // namespace validation diff --git a/tests/validation/NEON/DeconvolutionLayer.cpp b/tests/validation/NEON/DeconvolutionLayer.cpp index 9573784d8..566b75a82 100644 --- a/tests/validation/NEON/DeconvolutionLayer.cpp +++ b/tests/validation/NEON/DeconvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -44,6 +44,9 @@ namespace { constexpr AbsoluteTolerance<float> tolerance_fp32(0.001f); /**< Tolerance for floating point tests */ +const auto data4x4 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 3) + * framework::dataset::make("PadY", 0, 3) * framework::dataset::make("ax", 0) * framework::dataset::make("ay", 0) * framework::dataset::make("NumKernels", { 1, 3 }); + const auto data3x3 = datasets::SmallDeconvolutionShapes() * framework::dataset::make("StrideX", 1, 4) * framework::dataset::make("StrideY", 1, 4) * framework::dataset::make("PadX", 0, 2) * framework::dataset::make("PadY", 0, 2) * framework::dataset::make("ax", 0) * framework::dataset::make("ay", 0) * framework::dataset::make("NumKernels", { 1, 3 }); @@ -56,6 +59,9 @@ TEST_SUITE(NEON) TEST_SUITE(DeconvolutionLayer) template <typename T> +using NEDeconvolutionLayerFixture4x4 = DeconvolutionValidationFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 4, 4>; + +template <typename T> using NEDeconvolutionLayerFixture3x3 = DeconvolutionValidationFixture<Tensor, Accessor, NEDeconvolutionLayer, T, 3, 3>; template <typename T> @@ -64,6 +70,15 @@ using NEDeconvolutionLayerFixture1x1 = DeconvolutionValidationFixture<Tensor, Ac TEST_SUITE(Float) TEST_SUITE(FP32) +TEST_SUITE(W4x4) + +FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerFixture4x4<float>, framework::DatasetMode::ALL, combine(data4x4, framework::dataset::make("DataType", DataType::F32))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_fp32); +} +TEST_SUITE_END() + TEST_SUITE(W3x3) FIXTURE_DATA_TEST_CASE(Run, NEDeconvolutionLayerFixture3x3<float>, framework::DatasetMode::ALL, combine(data3x3, framework::dataset::make("DataType", DataType::F32))) diff --git a/tests/validation/NEON/DepthConcatenateLayer.cpp b/tests/validation/NEON/DepthConcatenateLayer.cpp index c79c7dbb6..f6df1fb7b 100644 --- a/tests/validation/NEON/DepthConcatenateLayer.cpp +++ b/tests/validation/NEON/DepthConcatenateLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -81,14 +81,14 @@ TEST_SUITE_END() TEST_SUITE(Quantized) TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConcatenateLayerFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), - framework::dataset::make("DataType", - DataType::QS8))) +FIXTURE_DATA_TEST_CASE(RunTiny, NEDepthConcatenateLayerFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Tiny2DShapes(), + framework::dataset::make("DataType", + DataType::QS8))) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthConcatenateLayerFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::DepthConcatenateLayerShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConcatenateLayerFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::DepthConcatenateLayerShapes(), framework::dataset::make("DataType", DataType::QS8))) { @@ -98,14 +98,14 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthConcatenateLayerFixture<int8_t>, framewo TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConcatenateLayerFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), - framework::dataset::make("DataType", - DataType::QS16))) +FIXTURE_DATA_TEST_CASE(RunTiny, NEDepthConcatenateLayerFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Tiny2DShapes(), + framework::dataset::make("DataType", + DataType::QS16))) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthConcatenateLayerFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(datasets::DepthConcatenateLayerShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthConcatenateLayerFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(datasets::DepthConcatenateLayerShapes(), framework::dataset::make("DataType", DataType::QS16))) { diff --git a/tests/validation/NEON/DepthConvertLayer.cpp b/tests/validation/NEON/DepthConvertLayer.cpp index a56298bab..ea63a5f52 100644 --- a/tests/validation/NEON/DepthConvertLayer.cpp +++ b/tests/validation/NEON/DepthConvertLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -392,7 +392,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi validate(src.info()->padding(), padding); validate(dst.info()->padding(), padding); } -FIXTURE_DATA_TEST_CASE(RunSmallQS8, NEDepthConvertLayerToFP32FixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTinyQS8, NEDepthConvertLayerToFP32FixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), DepthConvertLayerQS8toFP32Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -400,7 +400,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallQS8, NEDepthConvertLayerToFP32FixedPointFixture<i // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunSmallQS16, NEDepthConvertLayerToFP32FixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTinyQS16, NEDepthConvertLayerToFP32FixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), DepthConvertLayerQS16toFP32Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -408,7 +408,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallQS16, NEDepthConvertLayerToFP32FixedPointFixture< // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLargeQS8, NEDepthConvertLayerToFP32FixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmallQS8, NEDepthConvertLayerToFP32FixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerQS8toFP32Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -416,7 +416,7 @@ FIXTURE_DATA_TEST_CASE(RunLargeQS8, NEDepthConvertLayerToFP32FixedPointFixture<i // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLargeQS16, NEDepthConvertLayerToFP32FixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmallQS16, NEDepthConvertLayerToFP32FixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerQS16toFP32Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -451,7 +451,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combi validate(src.info()->padding(), padding); validate(dst.info()->padding(), padding); } -FIXTURE_DATA_TEST_CASE(RunSmallQS8, NEDepthConvertLayerToQS8FixedPointFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTinyQS8, NEDepthConvertLayerToQS8FixedPointFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), DepthConvertLayerFP32toQS8Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -459,7 +459,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallQS8, NEDepthConvertLayerToQS8FixedPointFixture<fl // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunSmallQS16, NEDepthConvertLayerToQS16FixedPointFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallShapes(), +FIXTURE_DATA_TEST_CASE(RunTinyQS16, NEDepthConvertLayerToQS16FixedPointFixture<float>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyShapes(), DepthConvertLayerFP32toQS16Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -467,7 +467,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallQS16, NEDepthConvertLayerToQS16FixedPointFixture< // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLargeQS8, NEDepthConvertLayerToQS8FixedPointFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmallQS8, NEDepthConvertLayerToQS8FixedPointFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerFP32toQS8Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) @@ -475,7 +475,7 @@ FIXTURE_DATA_TEST_CASE(RunLargeQS8, NEDepthConvertLayerToQS8FixedPointFixture<fl // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLargeQS16, NEDepthConvertLayerToQS16FixedPointFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmallQS16, NEDepthConvertLayerToQS16FixedPointFixture<float>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallShapes(), DepthConvertLayerFP32toQS16Dataset), framework::dataset::make("ConvertPolicy", { ConvertPolicy::SATURATE, ConvertPolicy::WRAP })), DepthConvertLayerFixedPointQuantizedDataset)) diff --git a/tests/validation/NEON/DepthwiseConvolutionLayer.cpp b/tests/validation/NEON/DepthwiseConvolutionLayer.cpp index e8c771595..0cdd4c029 100644 --- a/tests/validation/NEON/DepthwiseConvolutionLayer.cpp +++ b/tests/validation/NEON/DepthwiseConvolutionLayer.cpp @@ -82,8 +82,11 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::da validate(bias.info()->valid_region(), bias_valid_region); // Validate padding - const int step = 16 >> info.stride().first; - const PaddingSize padding = PaddingCalculator(output_shape.x(), step).required_padding(); + bool is_optimized_run = NEDepthwiseConvolutionLayer3x3Kernel::is_optimized_execution_possible(input_shape, info, data_type, DataLayout::NCHW); + const int step_non_opt_dwc = 16 >> info.stride().first; + const int step_bias_add = 16 / src.info()->element_size(); + const int step = is_optimized_run ? step_bias_add : std::max(step_non_opt_dwc, step_bias_add); + const PaddingSize padding = PaddingCalculator(output_shape.x(), step).required_padding(); validate(dst.info()->padding(), padding); } @@ -121,6 +124,12 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEDepthwiseConvolutionLayerFixture3x3<float>, f { validate(Accessor(_target), _reference, tolerance_f32); } +FIXTURE_DATA_TEST_CASE(RunOptimized, NEDepthwiseConvolutionLayerFixture3x3<float>, framework::DatasetMode::ALL, combine(datasets::OptimizedDepthwiseConvolutionLayerDataset3x3(), + framework::dataset::make("DataType", + DataType::F32))) +{ + validate(Accessor(_target), _reference, tolerance_f32); +} TEST_SUITE_END() TEST_SUITE_END() @@ -128,9 +137,19 @@ TEST_SUITE_END() template <typename T> using NEDepthwiseConvolutionLayerQuantizedFixture3x3 = DepthwiseConvolutionLayerValidationQuantizedFixture<Tensor, Accessor, NEDepthwiseConvolutionLayer3x3, T>; +template <typename T> +using NEDepthwiseConvolutionLayerQuantizedFixture = DepthwiseConvolutionLayerValidationQuantizedFixture<Tensor, Accessor, NEDepthwiseConvolutionLayer, T>; TEST_SUITE(Quantized) TEST_SUITE(QASYMM8) +TEST_SUITE(Generic) +FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset(), + framework::dataset::make("DataType", DataType::QASYMM8)), + framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, 10) }))) +{ + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() TEST_SUITE(W3x3) FIXTURE_DATA_TEST_CASE(RunSmall, NEDepthwiseConvolutionLayerQuantizedFixture3x3<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallDepthwiseConvolutionLayerDataset3x3(), framework::dataset::make("DataType", DataType::QASYMM8)), diff --git a/tests/validation/NEON/DirectConvolutionLayer.cpp b/tests/validation/NEON/DirectConvolutionLayer.cpp index 9982b4984..77f28924d 100644 --- a/tests/validation/NEON/DirectConvolutionLayer.cpp +++ b/tests/validation/NEON/DirectConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -49,8 +49,8 @@ constexpr AbsoluteTolerance<float> tolerance_fp16(0.01f); /**< Tolerance for ha constexpr AbsoluteTolerance<float> tolerance_fp32(0.001f); /**< Tolerance for floating point tests */ /** Direct convolution data set. */ -const auto data_pad_f32 = concat(concat(combine(framework::dataset::make("PadX", 0), - combine(framework::dataset::make("PadY", 0), +const auto data_pad_f32 = concat(concat(combine(framework::dataset::make("PadX", 0, 1), + combine(framework::dataset::make("PadY", 0, 1), framework::dataset::make("KernelSize", 1))), combine(framework::dataset::make("PadX", 0, 2), combine(framework::dataset::make("PadY", 0, 2), @@ -72,14 +72,14 @@ const auto data_f32 = combine(datasets::SmallDirectConvolutionShapes(), combine(data_pad_f32, framework::dataset::make("NumKernels", { 1, 4, 8, 16 }))))); -const auto data_qs8 = combine(datasets::SmallDirectConvolutionShapes(), +const auto data_qs8 = combine(datasets::TinyDirectConvolutionShapes(), combine(framework::dataset::make("StrideX", 1, 3), combine(framework::dataset::make("StrideY", 1, 3), combine(data_pad_qs8, framework::dataset::make("NumKernels", { 1, 4, 8, 16 }))))); /** Direct convolution QS16 data set. */ -const auto data_qs16 = combine(datasets::SmallDirectConvolutionShapes(), +const auto data_qs16 = combine(datasets::TinyDirectConvolutionShapes(), combine(framework::dataset::make("StrideX", 1, 3), combine(framework::dataset::make("StrideY", 1, 3), combine(framework::dataset::make("PadX", 0), diff --git a/tests/validation/NEON/EqualizeHistogram.cpp b/tests/validation/NEON/EqualizeHistogram.cpp new file mode 100644 index 000000000..633890e24 --- /dev/null +++ b/tests/validation/NEON/EqualizeHistogram.cpp @@ -0,0 +1,87 @@ +/* + * 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/runtime/NEON/functions/NEEqualizeHistogram.h" +#include "tests/NEON/Accessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/framework/Macros.h" +#include "tests/validation/Validation.h" +#include "tests/validation/fixtures/EqualizeHistogramFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +TEST_SUITE(NEON) +TEST_SUITE(EqualizeHistogram) + +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), framework::dataset::make("DataType", DataType::U8)), shape, data_type) +{ + // Create tensors + Tensor src = create_tensor<Tensor>(shape, data_type); + Tensor dst = create_tensor<Tensor>(shape, data_type); + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Create and configure function + NEEqualizeHistogram equalize_histogram; + equalize_histogram.configure(&src, &dst); + + // Validate valid region + const ValidRegion valid_region = shape_to_valid_region(shape); + validate(src.info()->valid_region(), valid_region); + validate(dst.info()->valid_region(), valid_region); + + // Validate padding + const PaddingSize padding = PaddingCalculator(shape.x(), 16).required_padding(); + validate(src.info()->padding(), padding); + validate(dst.info()->padding(), padding); +} + +template <typename T> +using NEEqualizeHistogramFixture = EqualizeHistogramValidationFixture<Tensor, Accessor, NEEqualizeHistogram, T>; + +FIXTURE_DATA_TEST_CASE(RunSmall, NEEqualizeHistogramFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(datasets::Small2DShapes(), framework::dataset::make("DataType", + DataType::U8))) +{ + // Validate output + validate(Accessor(_target), _reference); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, NEEqualizeHistogramFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(datasets::Large2DShapes(), framework::dataset::make("DataType", + DataType::U8))) +{ + // Validate output + validate(Accessor(_target), _reference); +} + +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/NEON/FastCorners.cpp b/tests/validation/NEON/FastCorners.cpp new file mode 100644 index 000000000..4416662fe --- /dev/null +++ b/tests/validation/NEON/FastCorners.cpp @@ -0,0 +1,115 @@ +/* + * 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/NEON/functions/NEFastCorners.h" +#include "arm_compute/runtime/Tensor.h" +#include "arm_compute/runtime/TensorAllocator.h" +#include "tests/NEON/Accessor.h" +#include "tests/NEON/ArrayAccessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/ImageFileDatasets.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/FastCornersFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +/* Radius of the Bresenham circle around the candidate point */ +const unsigned int bresenham_radius = 3; +/* Tolerance used to compare corner strengths */ +const AbsoluteTolerance<float> tolerance(0.5f); +} // namespace + +TEST_SUITE(NEON) +TEST_SUITE(FastCorners) + +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(combine(concat(datasets::Small2DShapes(), datasets::Large2DShapes()), + framework::dataset::make("Format", Format::U8)), + framework::dataset::make("SuppressNonMax", { false, true })), + framework::dataset::make("BorderMode", BorderMode::UNDEFINED)), + shape, format, suppress_nonmax, border_mode) +{ + std::mt19937 gen(library->seed()); + std::uniform_int_distribution<uint8_t> int_dist(0, 255); + std::uniform_real_distribution<float> real_dist(0, 255); + + const uint8_t constant_border_value = int_dist(gen); + const float threshold = real_dist(gen); + + // Create tensors + Tensor src = create_tensor<Tensor>(shape, data_type_from_format(format)); + src.info()->set_format(format); + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + + KeyPointArray corners; + + // Create and configure function + NEFastCorners fast_corners; + fast_corners.configure(&src, threshold, suppress_nonmax, &corners, border_mode, constant_border_value); + + // Validate padding + PaddingCalculator calculator(shape.x(), 1); // elems_processed + + calculator.set_border_size(bresenham_radius); + calculator.set_access_offset(-bresenham_radius); + calculator.set_accessed_elements(8); // elems_read + + validate(src.info()->padding(), calculator.required_padding()); +} + +template <typename T> +using NEFastCornersFixture = FastCornersValidationFixture<Tensor, Accessor, KeyPointArray, NEFastCorners, T>; + +FIXTURE_DATA_TEST_CASE(RunSmall, NEFastCornersFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallImageFiles(), framework::dataset::make("Format", Format::U8)), + framework::dataset::make("SuppressNonMax", { false, true })), + framework::dataset::make("BorderMode", BorderMode::UNDEFINED))) +{ + // Validate output + ArrayAccessor<KeyPoint> array(_target); + validate_keypoints(array.buffer(), array.buffer() + array.num_values(), _reference.begin(), _reference.end(), tolerance); +} +FIXTURE_DATA_TEST_CASE(RunLarge, NEFastCornersFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeImageFiles(), framework::dataset::make("Format", Format::U8)), + framework::dataset::make("SuppressNonMax", { false, true })), + framework::dataset::make("BorderMode", BorderMode::UNDEFINED))) +{ + // Validate output + ArrayAccessor<KeyPoint> array(_target); + validate_keypoints(array.buffer(), array.buffer() + array.num_values(), _reference.begin(), _reference.end(), tolerance); +} + +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/NEON/FixedPoint/FixedPoint.cpp b/tests/validation/NEON/FixedPoint/FixedPoint.cpp index 871143bc1..a583b587a 100644 --- a/tests/validation/NEON/FixedPoint/FixedPoint.cpp +++ b/tests/validation/NEON/FixedPoint/FixedPoint.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -43,14 +43,12 @@ namespace validation { namespace { -constexpr AbsoluteTolerance<float> tolerance_exp_qs8(0.0f); /**< Tolerance value for comparing reference's output against implementation's output (exponential) for DataType::QS8 */ -constexpr AbsoluteTolerance<float> tolerance_exp_qs16(1.0f); /**< Tolerance value for comparing reference's output against implementation's output (exponential) for DataType::QS16 */ -constexpr AbsoluteTolerance<float> tolerance_invsqrt_qs8(4.0f); /**< Tolerance value for comparing reference's output against implementation's output (inverse square-root) for DataType::QS8 */ -constexpr AbsoluteTolerance<float> tolerance_invsqrt_qs16(5.0f); /**< Tolerance value for comparing reference's output against implementation's output (inverse square-root) for DataType::QS16 */ -constexpr AbsoluteTolerance<float> tolerance_log_qs8(5.0f); /**< Tolerance value for comparing reference's output against implementation's output (logarithm) for DataType::QS8 */ -constexpr AbsoluteTolerance<float> tolerance_log_qs16(7.0f); /**< Tolerance value for comparing reference's output against implementation's output (logarithm) for DataType::QS16 */ -constexpr AbsoluteTolerance<float> tolerance_reciprocal_qs8(3); /**< Tolerance value for comparing reference's output against implementation's output (reciprocal) for DataType::QS8 */ -constexpr AbsoluteTolerance<float> tolerance_reciprocal_qs16(11.0f); /**< Tolerance value for comparing reference's output against implementation's output (reciprocal) for DataType::QS16 */ +constexpr AbsoluteTolerance<float> tolerance_exp_qs8(0.0f); /**< Tolerance value for comparing reference's output against implementation's output (exponential) for DataType::QS8 */ +constexpr AbsoluteTolerance<float> tolerance_exp_qs16(1.0f); /**< Tolerance value for comparing reference's output against implementation's output (exponential) for DataType::QS16 */ +constexpr AbsoluteTolerance<float> tolerance_invsqrt_qs8(4.0f); /**< Tolerance value for comparing reference's output against implementation's output (inverse square-root) for DataType::QS8 */ +constexpr AbsoluteTolerance<float> tolerance_invsqrt_qs16(5.0f); /**< Tolerance value for comparing reference's output against implementation's output (inverse square-root) for DataType::QS16 */ +constexpr AbsoluteTolerance<float> tolerance_log_qs8(5.0f); /**< Tolerance value for comparing reference's output against implementation's output (logarithm) for DataType::QS8 */ +constexpr AbsoluteTolerance<float> tolerance_log_qs16(7.0f); /**< Tolerance value for comparing reference's output against implementation's output (logarithm) for DataType::QS16 */ } // namespace TEST_SUITE(NEON) @@ -92,17 +90,6 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEFixedPointFixture<int8_t>, framework::Dataset validate(Accessor(_target), _reference, tolerance_log_qs8, 0); } TEST_SUITE_END() - -TEST_SUITE(Reciprocal) -FIXTURE_DATA_TEST_CASE(RunSmall, NEFixedPointFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::Small1DShapes(), framework::dataset::make("DataType", - DataType::QS8)), - framework::dataset::make("FixedPointOp", FixedPointOp::RECIPROCAL)), - framework::dataset::make("FractionalBits", 1, 6))) -{ - // Validate output - validate(Accessor(_target), _reference, tolerance_reciprocal_qs8, 0); -} -TEST_SUITE_END() TEST_SUITE_END() TEST_SUITE(QS16) @@ -139,17 +126,6 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEFixedPointFixture<int16_t>, framework::Datase validate(Accessor(_target), _reference, tolerance_log_qs16, 0); } TEST_SUITE_END() - -TEST_SUITE(Reciprocal) -FIXTURE_DATA_TEST_CASE(RunSmall, NEFixedPointFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(combine(datasets::Small1DShapes(), framework::dataset::make("DataType", - DataType::QS16)), - framework::dataset::make("FixedPointOp", FixedPointOp::RECIPROCAL)), - framework::dataset::make("FractionalBits", 1, 14))) -{ - // Validate output - validate(Accessor(_target), _reference, tolerance_reciprocal_qs16, 0); -} -TEST_SUITE_END() TEST_SUITE_END() TEST_SUITE_END() diff --git a/tests/validation/NEON/FixedPointPixelWiseMultiplication.cpp b/tests/validation/NEON/FixedPointPixelWiseMultiplication.cpp index f0a499d8d..d7954dec6 100644 --- a/tests/validation/NEON/FixedPointPixelWiseMultiplication.cpp +++ b/tests/validation/NEON/FixedPointPixelWiseMultiplication.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -83,29 +83,29 @@ TEST_SUITE(FixedPointPixelWiseMultiplication) TEST_SUITE(QS8) TEST_SUITE(Scale255) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, Fixture<qint8_t>, PRECOMMIT, SmallShapes(), QS8, QS8, scale_255, TO_NEAREST_UP, 1, 7) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunLarge, Fixture<qint8_t>, NIGHTLY, LargeShapes(), QS8, QS8, scale_255, TO_NEAREST_UP, 1, 7) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunTiny, Fixture<qint8_t>, PRECOMMIT, TinyShapes(), QS8, QS8, scale_255, TO_NEAREST_UP, 1, 7) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, Fixture<qint8_t>, NIGHTLY, SmallShapes(), QS8, QS8, scale_255, TO_NEAREST_UP, 1, 7) TEST_SUITE_END() // Scale255 TEST_SUITE(ScaleUnity) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, Fixture<qint8_t>, PRECOMMIT, SmallShapes(), QS8, QS8, scale_unity, TO_ZERO, 1, 7) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunLarge, Fixture<qint8_t>, NIGHTLY, LargeShapes(), QS8, QS8, scale_unity, TO_ZERO, 1, 7) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunTiny, Fixture<qint8_t>, PRECOMMIT, TinyShapes(), QS8, QS8, scale_unity, TO_ZERO, 1, 7) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, Fixture<qint8_t>, NIGHTLY, SmallShapes(), QS8, QS8, scale_unity, TO_ZERO, 1, 7) TEST_SUITE_END() // ScaleUnity TEST_SUITE(ScaleOther) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther1, Fixture<qint8_t>, PRECOMMIT, SmallShapes(), QS8, QS8, TO_ZERO, 1, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther2, Fixture<qint8_t>, PRECOMMIT, SmallShapes(), QS8, QS8, TO_ZERO, 2, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther3, Fixture<qint8_t>, PRECOMMIT, SmallShapes(), QS8, QS8, TO_ZERO, 3, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther4, Fixture<qint8_t>, PRECOMMIT, SmallShapes(), QS8, QS8, TO_ZERO, 4, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther5, Fixture<qint8_t>, PRECOMMIT, SmallShapes(), QS8, QS8, TO_ZERO, 5, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther6, Fixture<qint8_t>, PRECOMMIT, SmallShapes(), QS8, QS8, TO_ZERO, 6, tolerance) - -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunLargeOther1, Fixture<qint8_t>, NIGHTLY, LargeShapes(), QS8, QS8, TO_ZERO, 1, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunLargeOther2, Fixture<qint8_t>, NIGHTLY, LargeShapes(), QS8, QS8, TO_ZERO, 2, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunLargeOther3, Fixture<qint8_t>, NIGHTLY, LargeShapes(), QS8, QS8, TO_ZERO, 3, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunLargeOther4, Fixture<qint8_t>, NIGHTLY, LargeShapes(), QS8, QS8, TO_ZERO, 4, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunLargeOther5, Fixture<qint8_t>, NIGHTLY, LargeShapes(), QS8, QS8, TO_ZERO, 5, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunLargeOther6, Fixture<qint8_t>, NIGHTLY, LargeShapes(), QS8, QS8, TO_ZERO, 6, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther1, Fixture<qint8_t>, PRECOMMIT, TinyShapes(), QS8, QS8, TO_ZERO, 1, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther2, Fixture<qint8_t>, PRECOMMIT, TinyShapes(), QS8, QS8, TO_ZERO, 2, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther3, Fixture<qint8_t>, PRECOMMIT, TinyShapes(), QS8, QS8, TO_ZERO, 3, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther4, Fixture<qint8_t>, PRECOMMIT, TinyShapes(), QS8, QS8, TO_ZERO, 4, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther5, Fixture<qint8_t>, PRECOMMIT, TinyShapes(), QS8, QS8, TO_ZERO, 5, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther6, Fixture<qint8_t>, PRECOMMIT, TinyShapes(), QS8, QS8, TO_ZERO, 6, tolerance) + +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther1, Fixture<qint8_t>, NIGHTLY, SmallShapes(), QS8, QS8, TO_ZERO, 1, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther2, Fixture<qint8_t>, NIGHTLY, SmallShapes(), QS8, QS8, TO_ZERO, 2, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther3, Fixture<qint8_t>, NIGHTLY, SmallShapes(), QS8, QS8, TO_ZERO, 3, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther4, Fixture<qint8_t>, NIGHTLY, SmallShapes(), QS8, QS8, TO_ZERO, 4, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther5, Fixture<qint8_t>, NIGHTLY, SmallShapes(), QS8, QS8, TO_ZERO, 5, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther6, Fixture<qint8_t>, NIGHTLY, SmallShapes(), QS8, QS8, TO_ZERO, 6, tolerance) TEST_SUITE_END() // ScaleOther TEST_SUITE_END() // QS8 @@ -113,29 +113,29 @@ TEST_SUITE_END() // QS8 TEST_SUITE(QS16) TEST_SUITE(Scale255) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, scale_255, TO_NEAREST_UP, 1, 15) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunTiny, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, scale_255, TO_NEAREST_UP, 1, 15) TEST_SUITE_END() // Scale255 TEST_SUITE(ScaleUnity) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, scale_unity, TO_ZERO, 1, 15) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunLarge, Fixture<qint16_t>, NIGHTLY, LargeShapes(), QS16, QS16, scale_unity, TO_ZERO, 1, 15) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunTiny, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, scale_unity, TO_ZERO, 1, 15) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE(RunSmall, Fixture<qint16_t>, NIGHTLY, SmallShapes(), QS16, QS16, scale_unity, TO_ZERO, 1, 15) TEST_SUITE_END() // ScaleUnity TEST_SUITE(ScaleOther) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther1, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 1, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther2, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 2, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther3, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 3, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther4, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 4, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther5, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 5, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther6, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 6, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther7, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 7, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther8, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 8, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther9, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 9, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther10, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 10, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther11, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 11, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther12, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 12, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther13, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 13, tolerance) -FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunSmallOther14, Fixture<qint16_t>, PRECOMMIT, SmallShapes(), QS16, QS16, TO_ZERO, 14, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther1, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 1, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther2, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 2, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther3, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 3, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther4, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 4, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther5, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 5, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther6, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 6, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther7, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 7, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther8, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 8, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther9, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 9, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther10, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 10, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther11, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 11, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther12, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 12, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther13, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 13, tolerance) +FP_PIXEL_WISE_MULTIPLICATION_FIXTURE_DATA_TEST_CASE_OTHER(RunTinyOther14, Fixture<qint16_t>, PRECOMMIT, TinyShapes(), QS16, QS16, TO_ZERO, 14, tolerance) TEST_SUITE_END() // ScaleOther TEST_SUITE_END() // QS16 diff --git a/tests/validation/NEON/Flatten.cpp b/tests/validation/NEON/Flatten.cpp index d80faf49d..332716925 100644 --- a/tests/validation/NEON/Flatten.cpp +++ b/tests/validation/NEON/Flatten.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -82,13 +82,13 @@ TEST_SUITE_END() TEST_SUITE(Quantized) TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEFlattenLayerFixture<int8_t>, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::Small3DShapes(), datasets::Small4DShapes()), - framework::dataset::make("DataType", DataType::QS8))) +FIXTURE_DATA_TEST_CASE(RunTiny, NEFlattenLayerFixture<int8_t>, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::Tiny3DShapes(), datasets::Tiny4DShapes()), + framework::dataset::make("DataType", DataType::QS8))) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEFlattenLayerFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(datasets::Large3DShapes(), datasets::Large4DShapes()), +FIXTURE_DATA_TEST_CASE(RunSmall, NEFlattenLayerFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(datasets::Small3DShapes(), datasets::Small4DShapes()), framework::dataset::make("DataType", DataType::QS8))) { // Validate output @@ -97,13 +97,13 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEFlattenLayerFixture<int8_t>, framework::Datas TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEFlattenLayerFixture<int16_t>, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::Small3DShapes(), datasets::Small4DShapes()), - framework::dataset::make("DataType", DataType::QS16))) +FIXTURE_DATA_TEST_CASE(RunTiny, NEFlattenLayerFixture<int16_t>, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::Tiny3DShapes(), datasets::Tiny4DShapes()), + framework::dataset::make("DataType", DataType::QS16))) { // Validate output validate(Accessor(_target), _reference); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEFlattenLayerFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(datasets::Large3DShapes(), datasets::Large4DShapes()), +FIXTURE_DATA_TEST_CASE(RunSmall, NEFlattenLayerFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(framework::dataset::concat(datasets::Small3DShapes(), datasets::Small4DShapes()), framework::dataset::make("DataType", DataType::QS16))) { // Validate output diff --git a/tests/validation/NEON/FullyConnectedLayer.cpp b/tests/validation/NEON/FullyConnectedLayer.cpp index afdcc0504..ed0edcd3b 100644 --- a/tests/validation/NEON/FullyConnectedLayer.cpp +++ b/tests/validation/NEON/FullyConnectedLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -115,6 +115,52 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(combine(frame validate(dst.info()->valid_region(), dst_valid_region); } +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip(zip(zip(zip( + framework::dataset::make("InputInfo", { TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Mismatching data types + TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::QS8, 2), // Mismatching fixed point position + TensorInfo(TensorShape(8U, 4U, 6U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(8U, 4U, 6U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Invalid weights dimensions + TensorInfo(TensorShape(9U, 5U, 7U, 3U), 1, DataType::F32), // Wrongly reshaped weights + TensorInfo(TensorShape(8U, 4U, 6U, 4U), 1, DataType::F32), + }), + framework::dataset::make("WeightsInfo",{ TensorInfo(TensorShape(315U, 271U), 1, DataType::F16), + TensorInfo(TensorShape(315U, 271U), 1, DataType::QS8, 3), + TensorInfo(TensorShape(192U, 192U), 1, DataType::F32), + TensorInfo(TensorShape(192U, 192U), 1, DataType::F32), + TensorInfo(TensorShape(217U, 315U), 1, DataType::F32), + TensorInfo(TensorShape(217U, 315U), 1, DataType::F32), + TensorInfo(TensorShape(192U, 192U), 1, DataType::F32), + })), + framework::dataset::make("BiasInfo",{ TensorInfo(TensorShape(271U), 1, DataType::F32), + TensorInfo(TensorShape(271U), 1, DataType::QS8, 2), + TensorInfo(TensorShape(192U), 1, DataType::F32), + TensorInfo(TensorShape(192U), 1, DataType::F32), + TensorInfo(TensorShape(271U), 1, DataType::F32), + TensorInfo(TensorShape(271U), 1, DataType::F32), + TensorInfo(TensorShape(192U), 1, DataType::F32), + })), + framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), + TensorInfo(TensorShape(271U, 3U), 1, DataType::QS8, 3), + TensorInfo(TensorShape(192U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(192U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), + TensorInfo(TensorShape(271U, 3U), 1, DataType::F32), + TensorInfo(TensorShape(192U, 4U), 1, DataType::F32), + })), + framework::dataset::make("TransposeWeights",{ true, true, true, false, true, true, true })), + framework::dataset::make("ReshapedWeights",{ false, false, false, false, false, false , false})), + framework::dataset::make("Expected", { false, false, true, true, false, false, true })), + input_info, weights_info, bias_info, output_info, transpose_weights, reshaped_weights, expected) +{ + Status status = NEFullyConnectedLayer::validate(&input_info.clone()->set_is_resizable(false), &weights_info.clone()->set_is_resizable(false), &bias_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), transpose_weights, reshaped_weights); + ARM_COMPUTE_EXPECT(bool(status) == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + template <typename T> using NEFullyConnectedLayerFixture = FullyConnectedLayerValidationFixture<Tensor, Accessor, NEFullyConnectedLayer, T, true>; @@ -160,7 +206,7 @@ using NEFullyConnectedLayerFixedPointFixture = FullyConnectedLayerValidationFixe TEST_SUITE(FixedPoint) TEST_SUITE(QS8) // Testing for fixed point position [1,6) as reciprocal limits the maximum fixed point position to 5 -FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunTiny, NEFullyConnectedLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyFullyConnectedLayerDataset(), FullyConnectedParameters), framework::dataset::make("DataType", DataType::QS8)), @@ -169,7 +215,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixedPointFixture<int8_t>, // Validate output validate(Accessor(_target), _reference, tolerance_fixed_point); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters), framework::dataset::make("DataType", DataType::QS8)), @@ -182,7 +228,7 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) as reciprocal limits the maximum fixed point position to 14 -FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunTiny, NEFullyConnectedLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(combine(datasets::TinyFullyConnectedLayerDataset(), FullyConnectedParameters), framework::dataset::make("DataType", DataType::QS16)), @@ -191,7 +237,7 @@ FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixedPointFixture<int16_t> // Validate output validate(Accessor(_target), _reference, tolerance_fixed_point); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEFullyConnectedLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::LargeFullyConnectedLayerDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEFullyConnectedLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(combine(datasets::SmallFullyConnectedLayerDataset(), FullyConnectedParameters), framework::dataset::make("DataType", DataType::QS16)), diff --git a/tests/validation/NEON/GEMM.cpp b/tests/validation/NEON/GEMM.cpp index cc8279a58..0c9f20c67 100644 --- a/tests/validation/NEON/GEMM.cpp +++ b/tests/validation/NEON/GEMM.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -32,6 +32,7 @@ #include "tests/PaddingCalculator.h" #include "tests/datasets/LargeGEMMDataset.h" #include "tests/datasets/SmallGEMMDataset.h" +#include "tests/datasets/TinyGEMMDataset.h" #include "tests/framework/Asserts.h" #include "tests/framework/Macros.h" #include "tests/framework/datasets/Datasets.h" @@ -210,15 +211,15 @@ using NEGEMMFixedPointFixture = GEMMValidationFixedPointFixture<Tensor, Accessor TEST_SUITE(Quantized) TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallGEMMDataset(), - framework::dataset::make("DataType", - DataType::QS8)), - framework::dataset::make("FractionalBits", 1, 7))) +FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::TinyGEMMDataset(), + framework::dataset::make("DataType", + DataType::QS8)), + framework::dataset::make("FractionalBits", 1, 7))) { // Validate output validate(Accessor(_target), _reference, tolerance_q); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeGEMMDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SmallGEMMDataset(), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 1, 7))) @@ -229,15 +230,15 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMFixedPointFixture<int8_t>, framework::Dat TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SmallGEMMDataset(), - framework::dataset::make("DataType", - DataType::QS16)), - framework::dataset::make("FractionalBits", 1, 14))) +FIXTURE_DATA_TEST_CASE(RunTiny, NEGEMMFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::TinyGEMMDataset(), + framework::dataset::make("DataType", + DataType::QS16)), + framework::dataset::make("FractionalBits", 1, 14))) { // Validate output validate(Accessor(_target), _reference, tolerance_q); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEGEMMFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeGEMMDataset(), +FIXTURE_DATA_TEST_CASE(RunSmall, NEGEMMFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SmallGEMMDataset(), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) diff --git a/tests/validation/NEON/GaussianPyramid.cpp b/tests/validation/NEON/GaussianPyramid.cpp index a9130cfcf..6fee0dd16 100644 --- a/tests/validation/NEON/GaussianPyramid.cpp +++ b/tests/validation/NEON/GaussianPyramid.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -50,7 +50,7 @@ const auto small_gaussian_pyramid_levels = combine(datasets::Medium2DShapes(), d const auto large_gaussian_pyramid_levels = combine(datasets::Large2DShapes(), datasets::BorderModes()) * framework::dataset::make("numlevels", 2, 5); template <typename T, typename U> -inline void validate_gaussian_pyramid(const Pyramid &target, const std::vector<SimpleTensor<T>> &reference, BorderMode border_mode, U tolerance) +inline void validate_gaussian_pyramid(const Pyramid &target, const std::vector<SimpleTensor<T>> &reference, BorderMode border_mode, U tolerance, float tolerance_number = 0.0f) { ValidRegion prev_valid_region = shape_to_valid_region(reference[0].shape()); @@ -59,7 +59,7 @@ inline void validate_gaussian_pyramid(const Pyramid &target, const std::vector<S const ValidRegion valid_region = shape_to_valid_region_gaussian_pyramid_half(reference[i - 1].shape(), prev_valid_region, (border_mode == BorderMode::UNDEFINED)); // Validate outputs - validate(Accessor(*(target.get_pyramid_level(i))), reference[i], valid_region, tolerance); + validate(Accessor(*(target.get_pyramid_level(i))), reference[i], valid_region, tolerance, tolerance_number); // Keep the valid region for the next level prev_valid_region = valid_region; @@ -102,7 +102,7 @@ FIXTURE_DATA_TEST_CASE(RunSmallGaussianPyramidHalf, NEGaussianPyramidHalfFixture FIXTURE_DATA_TEST_CASE(RunLargeGaussianPyramidHalf, NEGaussianPyramidHalfFixture<uint8_t>, framework::DatasetMode::NIGHTLY, large_gaussian_pyramid_levels) { - validate_gaussian_pyramid(_target, _reference, _border_mode, tolerance_fp32); + validate_gaussian_pyramid(_target, _reference, _border_mode, tolerance_fp32, 0.01f); } TEST_SUITE_END() TEST_SUITE_END() diff --git a/tests/validation/NEON/HOGDescriptor.cpp b/tests/validation/NEON/HOGDescriptor.cpp index 5f31773aa..29663676c 100644 --- a/tests/validation/NEON/HOGDescriptor.cpp +++ b/tests/validation/NEON/HOGDescriptor.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -44,8 +44,8 @@ namespace validation { namespace { -AbsoluteTolerance<float> tolerance(0.5f); -constexpr float tolerance_number = 0.01f; +RelativeTolerance<float> tolerance(0.1f); +constexpr float tolerance_number = 0.05f; } // namespace TEST_SUITE(NEON) diff --git a/tests/validation/NEON/HarrisCorners.cpp b/tests/validation/NEON/HarrisCorners.cpp index b8d791774..3474a96f8 100644 --- a/tests/validation/NEON/HarrisCorners.cpp +++ b/tests/validation/NEON/HarrisCorners.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -46,9 +46,9 @@ namespace validation namespace { /* Allowed percentage of keypoints missing for target */ -float allowed_missing_percentage = 10.f; +const float allowed_missing_percentage = 10.f; /* Allowed percentage of keypoints mismatching between target and reference */ -float allowed_mismatch_percentage = 10.f; +const float allowed_mismatch_percentage = 10.f; const auto use_fp16 = framework::dataset::make("UseFP16", { diff --git a/tests/validation/NEON/Im2Col.cpp b/tests/validation/NEON/Im2Col.cpp index f8e474b6c..96dd6f86a 100644 --- a/tests/validation/NEON/Im2Col.cpp +++ b/tests/validation/NEON/Im2Col.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -44,6 +44,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(10U, 12U, 2U), 1, DataType::F32), // Mismatching data type TensorInfo(TensorShape(10U, 12U, 2U), 1, DataType::QS8, 2), // Mismatching fixed point TensorInfo(TensorShape(10U, 12U, 2U), 1, DataType::QASYMM8), // Bias not supported with QASYMM8 + TensorInfo(TensorShape(10U, 12U, 2U), 1, DataType::QASYMM8), // Mismatching shapes TensorInfo(TensorShape(10U, 12U, 2U), 1, DataType::QASYMM8), }), framework::dataset::make("OutputInfo",{ TensorInfo(TensorShape(3U, 4U, 10U, 2U), 1, DataType::F16), @@ -51,12 +52,13 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( TensorInfo(TensorShape(3U, 4U, 10U, 2U), 1, DataType::QS8, 3), TensorInfo(TensorShape(3U, 3U, 10U, 2U), 1, DataType::QASYMM8), TensorInfo(TensorShape(3U, 4U, 10U, 2U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(18U, 80U, 1U, 2U), 1, DataType::QASYMM8), })), - framework::dataset::make("HasBias", { true, true, true, true, false })), - framework::dataset::make("Expected", { false, false, false, false, true })), + framework::dataset::make("HasBias", { true, true, true, true, false, false })), + framework::dataset::make("Expected", { false, false, false, false, false, true })), input_info, output_info, has_bias, expected) { - bool status = bool(NEIm2Col::validate(&input_info, &output_info, Size2D(3U, 3U), PadStrideInfo(), has_bias)); + bool status = bool(NEIm2Col::validate(&input_info, &output_info, Size2D(3U, 3U), PadStrideInfo(), has_bias, false)); ARM_COMPUTE_EXPECT(status == expected, framework::LogLevel::ERRORS); } // clang-format on diff --git a/tests/validation/NEON/NormalizationLayer.cpp b/tests/validation/NEON/NormalizationLayer.cpp index 365447327..f5918db62 100644 --- a/tests/validation/NEON/NormalizationLayer.cpp +++ b/tests/validation/NEON/NormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -53,6 +53,9 @@ constexpr AbsoluteTolerance<int8_t> tolerance_qs8(2); constexpr AbsoluteTolerance<int16_t> tolerance_qs16(4); /** Input data set. */ +const auto NormalizationDatasetQS = combine(combine(combine(combine(datasets::TinyShapes(), datasets::NormalizationTypes()), framework::dataset::make("NormalizationSize", 3, 9, 2)), + framework::dataset::make("Beta", { 0.5f, 1.f, 2.f })), + framework::dataset::make("IsScaled", { true })); const auto NormalizationDataset = combine(combine(combine(combine(datasets::SmallShapes(), datasets::NormalizationTypes()), framework::dataset::make("NormalizationSize", 3, 9, 2)), framework::dataset::make("Beta", { 0.5f, 1.f, 2.f })), framework::dataset::make("IsScaled", { true })); @@ -139,14 +142,14 @@ using NENormalizationLayerFixedPointFixture = NormalizationValidationFixedPointF TEST_SUITE(Quantized) TEST_SUITE(QS8) // Testing for fixed point position [1,6) as reciprocal limits the maximum fixed point position to 5 -FIXTURE_DATA_TEST_CASE(RunSmall, NENormalizationLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(NormalizationDataset, framework::dataset::make("DataType", +FIXTURE_DATA_TEST_CASE(RunTiny, NENormalizationLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(NormalizationDatasetQS, framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 1, 6))) { // Validate output validate(Accessor(_target), _reference, tolerance_qs8); } -FIXTURE_DATA_TEST_CASE(RunLarge, NENormalizationLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(NormalizationDataset, framework::dataset::make("DataType", +FIXTURE_DATA_TEST_CASE(RunSmall, NENormalizationLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(NormalizationDataset, framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 1, 6))) { @@ -157,14 +160,14 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) as reciprocal limits the maximum fixed point position to 14 -FIXTURE_DATA_TEST_CASE(RunSmall, NENormalizationLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(NormalizationDataset, framework::dataset::make("DataType", +FIXTURE_DATA_TEST_CASE(RunTiny, NENormalizationLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(NormalizationDatasetQS, framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) { // Validate output validate(Accessor(_target), _reference, tolerance_qs16); } -FIXTURE_DATA_TEST_CASE(RunLarge, NENormalizationLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(NormalizationDataset, framework::dataset::make("DataType", +FIXTURE_DATA_TEST_CASE(RunSmall, NENormalizationLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(NormalizationDataset, framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) { diff --git a/tests/validation/NEON/Permute.cpp b/tests/validation/NEON/Permute.cpp new file mode 100644 index 000000000..872a16b35 --- /dev/null +++ b/tests/validation/NEON/Permute.cpp @@ -0,0 +1,157 @@ +/* + * Copyright (c) 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/NEON/functions/NEPermute.h" +#include "arm_compute/runtime/Tensor.h" +#include "arm_compute/runtime/TensorAllocator.h" +#include "tests/NEON/Accessor.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/PermuteFixture.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +const auto PermuteParametersSmall = combine(concat(concat(datasets::Small2DShapes(), datasets::Small3DShapes()), datasets::Small4DShapes()), + framework::dataset::make("PermutationVector", { PermutationVector(2U, 0U, 1U), PermutationVector(1U, 2U, 0U) })); +const auto PermuteParametersLarge = combine(datasets::Large4DShapes(), + framework::dataset::make("PermutationVector", { PermutationVector(2U, 0U, 1U), PermutationVector(1U, 2U, 0U) })); +} // namespace +TEST_SUITE(NEON) +TEST_SUITE(Permute) + +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( + framework::dataset::make("InputInfo",{ + TensorInfo(TensorShape(7U, 7U, 5U, 3U), 1, DataType::U16), // permutation not supported + TensorInfo(TensorShape(7U, 7U, 5U, 3U), 1, DataType::U16), // permutation not supported + TensorInfo(TensorShape(7U, 7U, 5U, 3U), 1, DataType::U16), // permutation not supported + TensorInfo(TensorShape(1U, 7U), 1, DataType::U8), // invalid input size + TensorInfo(TensorShape(7U, 7U, 5U, 3U), 1, DataType::U16), // valid + TensorInfo(TensorShape(27U, 13U, 37U, 2U), 1, DataType::F32), // valid + }), + framework::dataset::make("OutputInfo", { + TensorInfo(TensorShape(5U, 7U, 7U, 3U), 1, DataType::U16), + TensorInfo(TensorShape(7U, 7U, 5U, 3U), 1, DataType::U16), + TensorInfo(TensorShape(7U, 7U, 5U, 3U), 1, DataType::U16), + TensorInfo(TensorShape(5U, 7U), 1, DataType::U8), + TensorInfo(TensorShape(5U, 7U, 7U, 3U), 1, DataType::U16), + TensorInfo(TensorShape(13U, 37U, 27U, 2U), 1, DataType::F32), + })), + framework::dataset::make("PermutationVector", { + PermutationVector(2U, 1U, 0U), + PermutationVector(2U, 2U, 1U), + PermutationVector(1U, 1U, 1U), + PermutationVector(2U, 0U, 1U), + PermutationVector(2U, 0U, 1U), + PermutationVector(1U, 2U, 0U), + })), + framework::dataset::make("Expected", { false, false, false, false, true, true })), + input_info, output_info, perm_vect, expected) +{ + ARM_COMPUTE_EXPECT(bool(NEPermute::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), perm_vect)) == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(datasets::Small4DShapes(), framework::dataset::make("DataType", { DataType::S8, DataType::U8, DataType::S16, DataType::U16, DataType::U32, DataType::S32, DataType::F16, DataType::F32 })), + shape, data_type) +{ + // Define permutation vector + const PermutationVector perm(2U, 0U, 1U); + + // Permute shapes + TensorShape output_shape = shape; + permute(output_shape, perm); + + // Create tensors + Tensor ref_src = create_tensor<Tensor>(shape, data_type); + Tensor dst = create_tensor<Tensor>(output_shape, data_type); + + // Create and Configure function + NEPermute perm_func; + perm_func.configure(&ref_src, &dst, perm); + + // Validate valid region + const ValidRegion valid_region = shape_to_valid_region(output_shape); + validate(dst.info()->valid_region(), valid_region); +} + +template <typename T> +using NEPermuteFixture = PermuteValidationFixture<Tensor, Accessor, NEPermute, T>; + +TEST_SUITE(U8) +FIXTURE_DATA_TEST_CASE(RunSmall, NEPermuteFixture<uint8_t>, framework::DatasetMode::PRECOMMIT, combine(PermuteParametersSmall, framework::dataset::make("DataType", DataType::U8))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, NEPermuteFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(PermuteParametersLarge, framework::dataset::make("DataType", DataType::U8))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() + +TEST_SUITE(U16) +FIXTURE_DATA_TEST_CASE(RunSmall, NEPermuteFixture<uint16_t>, framework::DatasetMode::PRECOMMIT, combine(PermuteParametersSmall, framework::dataset::make("DataType", DataType::U16))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, NEPermuteFixture<uint16_t>, framework::DatasetMode::NIGHTLY, combine(PermuteParametersLarge, framework::dataset::make("DataType", DataType::U16))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() + +TEST_SUITE(U32) +FIXTURE_DATA_TEST_CASE(RunSmall, NEPermuteFixture<uint32_t>, framework::DatasetMode::PRECOMMIT, combine(PermuteParametersSmall, framework::dataset::make("DataType", DataType::U32))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +FIXTURE_DATA_TEST_CASE(RunLarge, NEPermuteFixture<uint32_t>, framework::DatasetMode::NIGHTLY, combine(PermuteParametersLarge, framework::dataset::make("DataType", DataType::U32))) +{ + // Validate output + validate(Accessor(_target), _reference); +} +TEST_SUITE_END() + +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/NEON/PoolingLayer.cpp b/tests/validation/NEON/PoolingLayer.cpp index e1c4ed590..2a86c10a9 100644 --- a/tests/validation/NEON/PoolingLayer.cpp +++ b/tests/validation/NEON/PoolingLayer.cpp @@ -27,6 +27,7 @@ #include "arm_compute/runtime/TensorAllocator.h" #include "tests/NEON/Accessor.h" #include "tests/PaddingCalculator.h" +#include "tests/datasets/PoolingLayerDataset.h" #include "tests/datasets/PoolingTypesDataset.h" #include "tests/datasets/ShapeDatasets.h" #include "tests/framework/Asserts.h" @@ -44,17 +45,19 @@ namespace validation namespace { /** Input data set for float data types */ -const auto PoolingLayerDatasetFP = combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { 2, 3, 7, 9 })), + +const auto PoolingLayerDatasetFP = combine(combine(combine(datasets::PoolingTypes(), framework::dataset::make("PoolingSize", { Size2D(2, 2), Size2D(3, 3), Size2D(7, 7), Size2D(9, 9), Size2D(4, 4), Size2D(3, 7), Size2D(7, 8) })), framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(1, 2, 1, 1), PadStrideInfo(2, 2, 1, 0) })), framework::dataset::make("ExcludePadding", { true, false })); /** Input data set for quantized data types */ -const auto PoolingLayerDatasetQS = combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { 2, 3 })), +const auto PoolingLayerDatasetQS = combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { Size2D(2, 2), Size2D(3, 3) })), framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(1, 2, 1, 1), PadStrideInfo(2, 2, 1, 0) })), framework::dataset::make("ExcludePadding", { false })); /** Input data set for asymmetric data type */ -const auto PoolingLayerDatasetQASYMM8 = combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { 2, 3, 9 })), + +const auto PoolingLayerDatasetQASYMM8 = combine(combine(combine(framework::dataset::make("PoolingType", { PoolingType::MAX, PoolingType::AVG }), framework::dataset::make("PoolingSize", { Size2D(2, 2), Size2D(3, 3), Size2D(4, 4), Size2D(9, 9), Size2D(3, 7), Size2D(7, 8) })), framework::dataset::make("PadStride", { PadStrideInfo(1, 1, 0, 0), PadStrideInfo(2, 1, 0, 0), PadStrideInfo(1, 2, 1, 1), PadStrideInfo(2, 2, 1, 0) })), framework::dataset::make("ExcludePadding", { true, false })); @@ -103,7 +106,7 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( PoolingLayerInfo(PoolingType::MAX), PoolingLayerInfo(PoolingType::AVG), })), - framework::dataset::make("Expected", { false, false, false, false, false, false, false, false, false, true })), + framework::dataset::make("Expected", { false, false, false, false, false, false, true, false, false, true })), input_info, output_info, pool_info, expected) { bool is_valid = bool(NEPoolingLayer::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false), pool_info)); @@ -115,8 +118,16 @@ DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( template <typename T> using NEPoolingLayerFixture = PoolingLayerValidationFixture<Tensor, Accessor, NEPoolingLayer, T>; +template <typename T> +using NESpecialPoolingLayerFixture = SpecialPoolingLayerValidationFixture<Tensor, Accessor, NEPoolingLayer, T>; + TEST_SUITE(Float) TEST_SUITE(FP32) +FIXTURE_DATA_TEST_CASE(RunSpecial, NESpecialPoolingLayerFixture<float>, framework::DatasetMode::ALL, datasets::PoolingLayerDatasetSpecial() * framework::dataset::make("DataType", DataType::F32)) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_f32); +} FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerFixture<float>, framework::DatasetMode::ALL, combine(datasets::SmallShapes(), combine(PoolingLayerDatasetFP, framework::dataset::make("DataType", DataType::F32)))) { @@ -154,14 +165,14 @@ using NEPoolingLayerFixedPointFixture = PoolingLayerValidationFixedPointFixture< TEST_SUITE(FixedPoint) TEST_SUITE(QS8) -FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerFixedPointFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetQS, - framework::dataset::make("DataType", DataType::QS8))), - framework::dataset::make("FractionalBits", 1, 5))) +FIXTURE_DATA_TEST_CASE(RunTiny, NEPoolingLayerFixedPointFixture<int8_t>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), combine(PoolingLayerDatasetQS, + framework::dataset::make("DataType", DataType::QS8))), + framework::dataset::make("FractionalBits", 1, 5))) { // Validate output validate(Accessor(_target), _reference, tolerance_qs8); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEPoolingLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), combine(PoolingLayerDatasetQS, +FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetQS, framework::dataset::make("DataType", DataType::QS8))), framework::dataset::make("FractionalBits", 1, 5))) { @@ -171,14 +182,14 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NEPoolingLayerFixedPointFixture<int8_t>, framew TEST_SUITE_END() TEST_SUITE(QS16) -FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerFixedPointFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetQS, - framework::dataset::make("DataType", DataType::QS16))), - framework::dataset::make("FractionalBits", 1, 13))) +FIXTURE_DATA_TEST_CASE(RunTiny, NEPoolingLayerFixedPointFixture<int16_t>, framework::DatasetMode::ALL, combine(combine(datasets::TinyShapes(), combine(PoolingLayerDatasetQS, + framework::dataset::make("DataType", DataType::QS16))), + framework::dataset::make("FractionalBits", 1, 13))) { // Validate output validate(Accessor(_target), _reference, tolerance_qs16); } -FIXTURE_DATA_TEST_CASE(RunLarge, NEPoolingLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::LargeShapes(), combine(PoolingLayerDatasetQS, +FIXTURE_DATA_TEST_CASE(RunSmall, NEPoolingLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SmallShapes(), combine(PoolingLayerDatasetQS, framework::dataset::make("DataType", DataType::QS16))), framework::dataset::make("FractionalBits", 1, 13))) { diff --git a/tests/validation/NEON/Remap.cpp b/tests/validation/NEON/Remap.cpp index 6e58000d5..2e54b1152 100644 --- a/tests/validation/NEON/Remap.cpp +++ b/tests/validation/NEON/Remap.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -43,8 +43,8 @@ namespace validation { namespace { -constexpr AbsoluteTolerance<uint8_t> tolerance_value(1); -constexpr float tolerance_number = 0.2f; +constexpr AbsoluteTolerance<uint8_t> tolerance_value(0); +constexpr float tolerance_number = 0.f; } // namespace TEST_SUITE(NEON) diff --git a/tests/validation/NEON/SoftmaxLayer.cpp b/tests/validation/NEON/SoftmaxLayer.cpp index 0b688dfd1..e091c0268 100644 --- a/tests/validation/NEON/SoftmaxLayer.cpp +++ b/tests/validation/NEON/SoftmaxLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -50,6 +50,9 @@ constexpr AbsoluteTolerance<float> tolerance_f16(0.0001f); /** Tolerance for fixed point operations */ constexpr AbsoluteTolerance<int16_t> tolerance_fixed_point(2); +/** Tolerance for quantized operations */ +constexpr AbsoluteTolerance<uint8_t> tolerance_qasymm8(1); + /** CNN data types */ const auto CNNDataTypes = framework::dataset::make("DataType", { @@ -90,7 +93,7 @@ DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(concat(datase const int step = 16 / data_size_from_type(data_type); const PaddingSize padding = PaddingCalculator(shape.x(), step).required_padding(); validate(src.info()->padding(), padding); - validate(dst.info()->padding(), padding); + validate(dst.info()->padding(), PaddingSize()); } // *INDENT-OFF* @@ -159,17 +162,17 @@ TEST_SUITE_END() template <typename T> using NESoftmaxLayerFixedPointFixture = SoftmaxValidationFixedPointFixture<Tensor, Accessor, NESoftmaxLayer, T>; -TEST_SUITE(Quantized) +TEST_SUITE(FixedPoint) TEST_SUITE(QS8) // Testing for fixed point position [1,6) as reciprocal limits the maximum fixed point position to 5 -FIXTURE_DATA_TEST_CASE(RunSmall, NESoftmaxLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType", - DataType::QS8)), - framework::dataset::make("FractionalBits", 1, 6))) +FIXTURE_DATA_TEST_CASE(RunTiny, NESoftmaxLayerFixedPointFixture<int8_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SoftmaxLayerTinyShapes(), framework::dataset::make("DataType", + DataType::QS8)), + framework::dataset::make("FractionalBits", 1, 6))) { // Validate output validate(Accessor(_target), _reference, tolerance_fixed_point); } -FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SoftmaxLayerLargeShapes(), framework::dataset::make("DataType", +FIXTURE_DATA_TEST_CASE(RunSmall, NESoftmaxLayerFixedPointFixture<int8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType", DataType::QS8)), framework::dataset::make("FractionalBits", 1, 6))) { @@ -180,15 +183,15 @@ TEST_SUITE_END() TEST_SUITE(QS16) // Testing for fixed point position [1,14) as reciprocal limits the maximum fixed point position to 14 -FIXTURE_DATA_TEST_CASE(RunSmall, NESoftmaxLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SoftmaxLayerSmallShapes(), - framework::dataset::make("DataType", - DataType::QS16)), - framework::dataset::make("FractionalBits", 1, 14))) +FIXTURE_DATA_TEST_CASE(RunTiny, NESoftmaxLayerFixedPointFixture<int16_t>, framework::DatasetMode::PRECOMMIT, combine(combine(datasets::SoftmaxLayerTinyShapes(), + framework::dataset::make("DataType", + DataType::QS16)), + framework::dataset::make("FractionalBits", 1, 14))) { // Validate output validate(Accessor(_target), _reference, tolerance_fixed_point); } -FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SoftmaxLayerLargeShapes(), +FIXTURE_DATA_TEST_CASE(RunSmall, NESoftmaxLayerFixedPointFixture<int16_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SoftmaxLayerSmallShapes(), framework::dataset::make("DataType", DataType::QS16)), framework::dataset::make("FractionalBits", 1, 14))) @@ -199,6 +202,30 @@ FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerFixedPointFixture<int16_t>, frame TEST_SUITE_END() TEST_SUITE_END() +template <typename T> +using NESoftmaxLayerQuantizedFixture = SoftmaxValidationQuantizedFixture<Tensor, Accessor, NESoftmaxLayer, T>; + +TEST_SUITE(Quantized) +TEST_SUITE(QASYMM8) +FIXTURE_DATA_TEST_CASE(RunSmall, NESoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::ALL, combine(combine(datasets::SoftmaxLayerSmallShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }), + framework::dataset::make("Beta", { 1.0f, 2.0f })))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +FIXTURE_DATA_TEST_CASE(RunLarge, NESoftmaxLayerQuantizedFixture<uint8_t>, framework::DatasetMode::NIGHTLY, combine(combine(datasets::SoftmaxLayerLargeShapes(), + framework::dataset::make("DataType", DataType::QASYMM8)), + combine(framework::dataset::make("QuantizationInfo", { QuantizationInfo(0.5f, -10) }), + framework::dataset::make("Beta", { 1.0f, 2.0f })))) +{ + // Validate output + validate(Accessor(_target), _reference, tolerance_qasymm8); +} +TEST_SUITE_END() +TEST_SUITE_END() + TEST_SUITE_END() TEST_SUITE_END() } // namespace validation diff --git a/tests/validation/UNIT/Utils.cpp b/tests/validation/UNIT/Utils.cpp index 969b2136a..398067417 100644 --- a/tests/validation/UNIT/Utils.cpp +++ b/tests/validation/UNIT/Utils.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * diff --git a/tests/validation/Validation.h b/tests/validation/Validation.h index b12d7de97..8ed22c29d 100644 --- a/tests/validation/Validation.h +++ b/tests/validation/Validation.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -363,7 +363,10 @@ void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const V const T &target_value = reinterpret_cast<const T *>(tensor(id))[c]; const T &reference_value = reinterpret_cast<const T *>(reference(id))[c]; - if(!compare<U>(target_value, reference_value, tolerance_value)) + // Truncate numbers to the 4th decimal + const T target_truncated_value = static_cast<T>(static_cast<int>(target_value * 10000) / 10000); + const T reference_truncated_value = static_cast<T>(static_cast<int>(target_value * 10000) / 10000); + if(!compare<U>(target_truncated_value, reference_truncated_value, tolerance_value)) { ARM_COMPUTE_TEST_INFO("id = " << id); ARM_COMPUTE_TEST_INFO("channel = " << c); @@ -473,76 +476,6 @@ void validate_wrap(const IAccessor &tensor, const SimpleTensor<T> &reference, co } } -/** Check which keypoints from [first1, last1) are missing in [first2, last2) */ -template <typename T, typename U, typename V> -std::pair<int64_t, int64_t> compare_keypoints(T first1, T last1, U first2, U last2, V tolerance) -{ - int64_t num_missing = 0; - int64_t num_mismatches = 0; - - while(first1 != last1) - { - const auto point = std::find_if(first2, last2, [&](KeyPoint point) - { - return point.x == first1->x && point.y == first1->y; - }); - - if(point == last2) - { - ++num_missing; - ARM_COMPUTE_TEST_INFO("Key point not found" << *first1) - ARM_COMPUTE_TEST_INFO("keypoint1 = " << *first1) - } - else if(!validate(point->tracking_status, first1->tracking_status) || !validate(point->strength, first1->strength, tolerance) || !validate(point->scale, first1->scale) - || !validate(point->orientation, first1->orientation) || !validate(point->error, first1->error)) - { - ++num_mismatches; - ARM_COMPUTE_TEST_INFO("Mismatching keypoint") - ARM_COMPUTE_TEST_INFO("keypoint1 = " << *first1) - ARM_COMPUTE_TEST_INFO("keypoint2 = " << *point) - } - - ++first1; - } - - return std::make_pair(num_missing, num_mismatches); -} - -template <typename T, typename U, typename V> -void validate_keypoints(T target_first, T target_last, U reference_first, U reference_last, V tolerance, - float allowed_missing_percentage, float allowed_mismatch_percentage) -{ - const int64_t num_elements_target = std::distance(target_first, target_last); - const int64_t num_elements_reference = std::distance(reference_first, reference_last); - - int64_t num_missing = 0; - int64_t num_mismatches = 0; - - if(num_elements_reference > 0) - { - std::tie(num_missing, num_mismatches) = compare_keypoints(reference_first, reference_last, target_first, target_last, tolerance); - - const float percent_missing = static_cast<float>(num_missing) / num_elements_reference * 100.f; - const float percent_mismatches = static_cast<float>(num_mismatches) / num_elements_reference * 100.f; - - ARM_COMPUTE_TEST_INFO(num_missing << " keypoints (" << std::fixed << std::setprecision(2) << percent_missing << "%) are missing in target"); - ARM_COMPUTE_EXPECT(percent_missing <= allowed_missing_percentage, framework::LogLevel::ERRORS); - - ARM_COMPUTE_TEST_INFO(num_mismatches << " keypoints (" << std::fixed << std::setprecision(2) << percent_mismatches << "%) mismatched"); - ARM_COMPUTE_EXPECT(percent_mismatches <= allowed_mismatch_percentage, framework::LogLevel::ERRORS); - } - - if(num_elements_target > 0) - { - std::tie(num_missing, num_mismatches) = compare_keypoints(target_first, target_last, reference_first, reference_last, tolerance); - - const float percent_missing = static_cast<float>(num_missing) / num_elements_target * 100.f; - - ARM_COMPUTE_TEST_INFO(num_missing << " keypoints (" << std::fixed << std::setprecision(2) << percent_missing << "%) are not part of target"); - ARM_COMPUTE_EXPECT(percent_missing <= allowed_missing_percentage, framework::LogLevel::ERRORS); - } -} - template <typename T, typename U> void validate(const IAccessor &tensor, const SimpleTensor<T> &reference, const SimpleTensor<T> &valid_mask, U tolerance_value, float tolerance_number) { @@ -652,6 +585,123 @@ void validate_min_max_loc(const MinMaxLocationValues<T> &target, const MinMaxLoc } } +/** Check which keypoints from [first1, last1) are missing in [first2, last2) */ +template <typename T, typename U, typename V> +std::pair<int64_t, int64_t> compare_keypoints(T first1, T last1, U first2, U last2, V tolerance, bool check_mismatches = true) +{ + /* Keypoint (x,y) should have similar strength (within tolerance) and other properties in both reference and target */ + const auto compare_props_eq = [&](const KeyPoint & lhs, const KeyPoint & rhs) + { + return compare<V>(lhs.strength, rhs.strength, tolerance) + && lhs.tracking_status == rhs.tracking_status + && lhs.scale == rhs.scale + && lhs.orientation == rhs.orientation + && lhs.error == rhs.error; + }; + + /* Used to sort KeyPoints by coordinates (x, y) */ + const auto compare_coords_lt = [](const KeyPoint & lhs, const KeyPoint & rhs) + { + return std::tie(lhs.x, lhs.y) < std::tie(rhs.x, rhs.y); + }; + + std::sort(first1, last1, compare_coords_lt); + std::sort(first2, last2, compare_coords_lt); + + if(check_mismatches) + { + ARM_COMPUTE_TEST_INFO("Checking for mismatches: ref count = " << std::distance(first1, last1) << " \ttarget count = " << std::distance(first2, last2)); + } + + int64_t num_missing = 0; + int64_t num_mismatches = 0; + bool rest_missing = false; + + while(first1 != last1) + { + if(first2 == last2) + { + rest_missing = true; + break; + } + + if(compare_coords_lt(*first1, *first2)) + { + ++num_missing; + ARM_COMPUTE_TEST_INFO("Key point not found"); + ARM_COMPUTE_TEST_INFO("keypoint1 = " << *first1++); + } + else + { + if(!compare_coords_lt(*first2, *first1)) // Equal coordinates + { + if(check_mismatches && !compare_props_eq(*first1, *first2)) // Check other properties + { + ++num_mismatches; + ARM_COMPUTE_TEST_INFO("Mismatching keypoint"); + ARM_COMPUTE_TEST_INFO("keypoint1 [ref] = " << *first1); + ARM_COMPUTE_TEST_INFO("keypoint2 [tgt] = " << *first2); + } + ++first1; + } + ++first2; + } + } + + if(rest_missing) + { + while(first1 != last1) + { + ++num_missing; + ARM_COMPUTE_TEST_INFO("Key point not found"); + ARM_COMPUTE_TEST_INFO("keypoint1 = " << *first1++); + } + } + + return std::make_pair(num_missing, num_mismatches); +} + +template <typename T, typename U, typename V> +void validate_keypoints(T target_first, T target_last, U reference_first, U reference_last, V tolerance, float allowed_missing_percentage, float allowed_mismatch_percentage) +{ + const int64_t num_elements_target = std::distance(target_first, target_last); + const int64_t num_elements_reference = std::distance(reference_first, reference_last); + + int64_t num_missing = 0; + int64_t num_mismatches = 0; + + if(num_elements_reference > 0) + { + std::tie(num_missing, num_mismatches) = compare_keypoints(reference_first, reference_last, target_first, target_last, tolerance); + + const float percent_missing = static_cast<float>(num_missing) / num_elements_reference * 100.f; + const float percent_mismatches = static_cast<float>(num_mismatches) / num_elements_reference * 100.f; + + ARM_COMPUTE_TEST_INFO(num_missing << " keypoints (" << std::fixed << std::setprecision(2) << percent_missing << "%) in ref are missing from target"); + ARM_COMPUTE_TEST_INFO("Missing (not in tgt): " << num_missing << "/" << num_elements_reference << " = " << std::fixed << std::setprecision(2) << percent_missing + << "% \tMax allowed: " << allowed_missing_percentage << "%"); + ARM_COMPUTE_EXPECT(percent_missing <= allowed_missing_percentage, framework::LogLevel::ERRORS); + + ARM_COMPUTE_TEST_INFO(num_mismatches << " keypoints (" << std::fixed << std::setprecision(2) << percent_mismatches << "%) mismatched"); + ARM_COMPUTE_TEST_INFO("Mismatched keypoints: " << num_mismatches << "/" << num_elements_reference << " = " << std::fixed << std::setprecision(2) << percent_mismatches + << "% \tMax allowed: " << allowed_mismatch_percentage << "%"); + ARM_COMPUTE_EXPECT(percent_mismatches <= allowed_mismatch_percentage, framework::LogLevel::ERRORS); + } + + if(num_elements_target > 0) + { + // Note: no need to check for mismatches a second time (last argument is 'false') + std::tie(num_missing, num_mismatches) = compare_keypoints(target_first, target_last, reference_first, reference_last, tolerance, false); + + const float percent_missing = static_cast<float>(num_missing) / num_elements_target * 100.f; + + ARM_COMPUTE_TEST_INFO(num_missing << " keypoints (" << std::fixed << std::setprecision(2) << percent_missing << "%) in target are missing from ref"); + ARM_COMPUTE_TEST_INFO("Missing (not in ref): " << num_missing << "/" << num_elements_target << " = " << std::fixed << std::setprecision(2) << percent_missing + << "% \tMax allowed: " << allowed_missing_percentage << "%"); + ARM_COMPUTE_EXPECT(percent_missing <= allowed_missing_percentage, framework::LogLevel::ERRORS); + } +} + } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/fixtures/ArithmeticAdditionFixture.h b/tests/validation/fixtures/ArithmeticAdditionFixture.h index c3a51b97d..f3888ae56 100644 --- a/tests/validation/fixtures/ArithmeticAdditionFixture.h +++ b/tests/validation/fixtures/ArithmeticAdditionFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -41,15 +41,14 @@ namespace test namespace validation { template <typename TensorType, typename AccessorType, typename FunctionType, typename T> -class ArithmeticAdditionValidationFixedPointFixture : public framework::Fixture +class ArithmeticAdditionBroadcastValidationFixedPointFixture : public framework::Fixture { public: template <typename...> - void setup(TensorShape shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fractional_bits) + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fractional_bits) { - _fractional_bits = fractional_bits; - _target = compute_target(shape, data_type0, data_type1, output_data_type, convert_policy, fractional_bits); - _reference = compute_reference(shape, data_type0, data_type1, output_data_type, convert_policy, fractional_bits); + _target = compute_target(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, fractional_bits); + _reference = compute_reference(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, fractional_bits); } protected: @@ -59,12 +58,13 @@ protected: library->fill_tensor_uniform(tensor, i); } - TensorType compute_target(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fixed_point_position) + TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, + int fixed_point_position) { // Create tensors - TensorType ref_src1 = create_tensor<TensorType>(shape, data_type0, 1, fixed_point_position); - TensorType ref_src2 = create_tensor<TensorType>(shape, data_type1, 1, fixed_point_position); - TensorType dst = create_tensor<TensorType>(shape, output_data_type, 1, fixed_point_position); + TensorType ref_src1 = create_tensor<TensorType>(shape0, data_type0, 1, fixed_point_position); + TensorType ref_src2 = create_tensor<TensorType>(shape1, data_type1, 1, fixed_point_position); + TensorType dst = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), output_data_type, 1, fixed_point_position); // Create and configure function FunctionType add; @@ -93,11 +93,12 @@ protected: return dst; } - SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fixed_point_position) + SimpleTensor<T> compute_reference(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, + int fixed_point_position) { // Create reference - SimpleTensor<T> ref_src1{ shape, data_type0, 1, fixed_point_position }; - SimpleTensor<T> ref_src2{ shape, data_type1, 1, fixed_point_position }; + SimpleTensor<T> ref_src1{ shape0, data_type0, 1, fixed_point_position }; + SimpleTensor<T> ref_src2{ shape1, data_type1, 1, fixed_point_position }; // Fill reference fill(ref_src1, 0); @@ -108,14 +109,36 @@ protected: TensorType _target{}; SimpleTensor<T> _reference{}; - int _fractional_bits{}; }; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class ArithmeticAdditionBroadcastValidationFixture : public ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + template <typename...> + void setup(const TensorShape &shape0, const TensorShape &shape1, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy) + { + ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, data_type0, data_type1, output_data_type, convert_policy, 0); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class ArithmeticAdditionValidationFixedPointFixture : public ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + template <typename...> + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy, int fractional_bits) + { + ArithmeticAdditionBroadcastValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, shape, data_type0, data_type1, output_data_type, convert_policy, fractional_bits); + } +}; + template <typename TensorType, typename AccessorType, typename FunctionType, typename T> class ArithmeticAdditionValidationFixture : public ArithmeticAdditionValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T> { public: template <typename...> - void setup(TensorShape shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy) + void setup(const TensorShape &shape, DataType data_type0, DataType data_type1, DataType output_data_type, ConvertPolicy convert_policy) { ArithmeticAdditionValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type0, data_type1, output_data_type, convert_policy, 0); } diff --git a/tests/validation/fixtures/BatchNormalizationLayerFixture.h b/tests/validation/fixtures/BatchNormalizationLayerFixture.h index 298c9ca41..e02c61924 100644 --- a/tests/validation/fixtures/BatchNormalizationLayerFixture.h +++ b/tests/validation/fixtures/BatchNormalizationLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -45,12 +45,12 @@ class BatchNormalizationLayerValidationFixedPointFixture : public framework::Fix { public: template <typename...> - void setup(TensorShape shape0, TensorShape shape1, float epsilon, DataType dt, int fractional_bits) + void setup(TensorShape shape0, TensorShape shape1, float epsilon, ActivationLayerInfo act_info, DataType dt, int fractional_bits) { _fractional_bits = fractional_bits; _data_type = dt; - _target = compute_target(shape0, shape1, epsilon, dt, fractional_bits); - _reference = compute_reference(shape0, shape1, epsilon, dt, fractional_bits); + _target = compute_target(shape0, shape1, epsilon, act_info, dt, fractional_bits); + _reference = compute_reference(shape0, shape1, epsilon, act_info, dt, fractional_bits); } protected: @@ -85,7 +85,7 @@ protected: } } - TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, float epsilon, DataType dt, int fixed_point_position) + TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, float epsilon, ActivationLayerInfo act_info, DataType dt, int fixed_point_position) { // Create tensors TensorType src = create_tensor<TensorType>(shape0, dt, 1, fixed_point_position); @@ -97,7 +97,7 @@ protected: // Create and configure function FunctionType norm; - norm.configure(&src, &dst, &mean, &var, &beta, &gamma, epsilon); + norm.configure(&src, &dst, &mean, &var, &beta, &gamma, epsilon, act_info); ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); @@ -130,7 +130,7 @@ protected: return dst; } - SimpleTensor<T> compute_reference(const TensorShape &shape0, const TensorShape &shape1, float epsilon, DataType dt, int fixed_point_position) + SimpleTensor<T> compute_reference(const TensorShape &shape0, const TensorShape &shape1, float epsilon, ActivationLayerInfo act_info, DataType dt, int fixed_point_position) { // Create reference SimpleTensor<T> ref_src{ shape0, dt, 1, fixed_point_position }; @@ -142,7 +142,7 @@ protected: // Fill reference fill(ref_src, ref_mean, ref_var, ref_beta, ref_gamma); - return reference::batch_normalization_layer(ref_src, ref_mean, ref_var, ref_beta, ref_gamma, epsilon, fixed_point_position); + return reference::batch_normalization_layer(ref_src, ref_mean, ref_var, ref_beta, ref_gamma, epsilon, act_info, fixed_point_position); } TensorType _target{}; @@ -156,9 +156,9 @@ class BatchNormalizationLayerValidationFixture : public BatchNormalizationLayerV { public: template <typename...> - void setup(TensorShape shape0, TensorShape shape1, float epsilon, DataType dt) + void setup(TensorShape shape0, TensorShape shape1, float epsilon, ActivationLayerInfo act_info, DataType dt) { - BatchNormalizationLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, epsilon, dt, 0); + BatchNormalizationLayerValidationFixedPointFixture<TensorType, AccessorType, FunctionType, T>::setup(shape0, shape1, epsilon, act_info, dt, 0); } }; } // namespace validation diff --git a/tests/validation/fixtures/ChannelExtractFixture.h b/tests/validation/fixtures/ChannelExtractFixture.h new file mode 100644 index 000000000..c3c2e17d0 --- /dev/null +++ b/tests/validation/fixtures/ChannelExtractFixture.h @@ -0,0 +1,192 @@ +/* + * 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. + */ +#ifndef ARM_COMPUTE_TEST_CHANNEL_EXTRACT_FIXTURE +#define ARM_COMPUTE_TEST_CHANNEL_EXTRACT_FIXTURE + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.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/reference/ChannelExtract.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename MultiImageType, typename TensorType, typename AccessorType, typename FunctionType, typename T> +class ChannelExtractValidationFixture : public framework::Fixture +{ +public: + template <typename...> + void setup(TensorShape shape, Format format, Channel channel) + { + shape = adjust_odd_shape(shape, format); + + _target = compute_target(shape, format, channel); + _reference = compute_reference(shape, format, channel); + } + +protected: + template <typename U> + void fill(U &&tensor, int i) + { + library->fill_tensor_uniform(tensor, i); + } + + std::vector<SimpleTensor<T>> create_tensor_planes_reference(const TensorShape &shape, Format format) + { + TensorShape input = adjust_odd_shape(shape, format); + + std::vector<SimpleTensor<T>> tensor_planes; + + switch(format) + { + case Format::RGB888: + case Format::RGBA8888: + case Format::YUYV422: + case Format::UYVY422: + { + tensor_planes.emplace_back(input, format); + break; + } + case Format::NV12: + case Format::NV21: + { + const TensorShape shape_uv88 = calculate_subsampled_shape(shape, Format::UV88); + + tensor_planes.emplace_back(input, Format::U8); + tensor_planes.emplace_back(shape_uv88, Format::UV88); + break; + } + case Format::IYUV: + { + const TensorShape shape_sub2 = calculate_subsampled_shape(shape, Format::IYUV); + + tensor_planes.emplace_back(input, Format::U8); + tensor_planes.emplace_back(shape_sub2, Format::U8); + tensor_planes.emplace_back(shape_sub2, Format::U8); + break; + } + case Format::YUV444: + tensor_planes.emplace_back(input, Format::U8); + tensor_planes.emplace_back(input, Format::U8); + tensor_planes.emplace_back(input, Format::U8); + break; + default: + ARM_COMPUTE_ERROR("Not supported"); + break; + } + + return tensor_planes; + } + + TensorType compute_target(const TensorShape &shape, Format format, Channel channel) + { + const unsigned int num_planes = num_planes_from_format(format); + + TensorShape dst_shape = calculate_subsampled_shape(shape, format, channel); + + // Create tensors + MultiImageType ref_src = create_multi_image<MultiImageType>(shape, format); + TensorType dst = create_tensor<TensorType>(dst_shape, Format::U8); + + // Create and configure function + FunctionType channel_extract; + + if(1U == num_planes) + { + const TensorType *plane_src = static_cast<TensorType *>(ref_src.plane(0)); + + channel_extract.configure(plane_src, channel, &dst); + } + else + { + channel_extract.configure(&ref_src, channel, &dst); + } + + for(unsigned int plane_idx = 0; plane_idx < num_planes; ++plane_idx) + { + const TensorType *src_plane = static_cast<const TensorType *>(ref_src.plane(plane_idx)); + + ARM_COMPUTE_EXPECT(src_plane->info()->is_resizable(), framework::LogLevel::ERRORS); + } + + ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Allocate tensors + ref_src.allocate(); + dst.allocator()->allocate(); + + for(unsigned int plane_idx = 0; plane_idx < num_planes; ++plane_idx) + { + const TensorType *src_plane = static_cast<const TensorType *>(ref_src.plane(plane_idx)); + + ARM_COMPUTE_EXPECT(!src_plane->info()->is_resizable(), framework::LogLevel::ERRORS); + } + + ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Fill tensor planes + for(unsigned int plane_idx = 0; plane_idx < num_planes; ++plane_idx) + { + TensorType *src_plane = static_cast<TensorType *>(ref_src.plane(plane_idx)); + + fill(AccessorType(*src_plane), plane_idx); + } + + // Compute function + channel_extract.run(); + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &shape, Format format, Channel channel) + { + const unsigned int num_planes = num_planes_from_format(format); + + // Create reference + std::vector<SimpleTensor<T>> ref_src = create_tensor_planes_reference(shape, format); + + // Fill references + for(unsigned int plane_idx = 0; plane_idx < num_planes; ++plane_idx) + { + fill(ref_src[plane_idx], plane_idx); + } + + return reference::channel_extract<T>(shape, ref_src, format, channel); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_CHANNEL_EXTRACT_FIXTURE */ diff --git a/tests/validation/fixtures/ConvolutionFixture.h b/tests/validation/fixtures/ConvolutionFixture.h index 85070cff8..8ebb924c6 100644 --- a/tests/validation/fixtures/ConvolutionFixture.h +++ b/tests/validation/fixtures/ConvolutionFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -46,7 +46,7 @@ class ConvolutionValidationFixture : public framework::Fixture { protected: template <typename...> - void setup(TensorShape shape, DataType data_type, BorderMode border_mode, const unsigned int width, const unsigned int height, const bool is_separable = false) + void setup(TensorShape shape, DataType output_data_type, BorderMode border_mode, const unsigned int width, const unsigned int height, const bool is_separable = false) { std::mt19937 gen(library->seed()); std::uniform_int_distribution<uint8_t> distribution(0, 255); @@ -59,22 +59,22 @@ protected: ARM_COMPUTE_ERROR_ON(3 != width && 5 != width && 7 != width && 9 != width); ARM_COMPUTE_ERROR_ON(3 != height && 5 != height && 7 != height && 9 != height); - int16_t conv[width * height]; + std::vector<int16_t> conv(width * height); _width = width; _height = height; if(is_separable) { - create_separable_conv(conv); + create_separable_conv(conv.data()); } else { - create_conv(conv); + create_conv(conv.data()); } - _target = compute_target(shape, data_type, conv, scale, border_mode, constant_border_value); - _reference = compute_reference(shape, data_type, conv, scale, border_mode, constant_border_value); + _target = compute_target(shape, output_data_type, conv.data(), scale, border_mode, constant_border_value); + _reference = compute_reference(shape, output_data_type, conv.data(), scale, border_mode, constant_border_value); } void @@ -122,21 +122,19 @@ protected: library->fill_tensor_uniform(tensor, i); } - SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value) + SimpleTensor<T> compute_reference(const TensorShape &shape, DataType output_data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value) { - ARM_COMPUTE_ERROR_ON(data_type != DataType::U8); - // Create reference - SimpleTensor<T> src{ shape, data_type }; + SimpleTensor<uint8_t> src{ shape, DataType::U8 }; // Fill reference fill(src, 0); // Compute reference - return reference::convolution<T>(src, conv, scale, border_mode, constant_border_value, _width, _height); + return reference::convolution<T>(src, output_data_type, conv, scale, border_mode, constant_border_value, _width, _height); } - virtual TensorType compute_target(const TensorShape &shape, DataType data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value) = 0; + virtual TensorType compute_target(const TensorShape &shape, DataType output_data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value) = 0; BorderMode _border_mode{}; TensorType _target{}; @@ -150,17 +148,17 @@ class ConvolutionSquareValidationFixture : public ConvolutionValidationFixture<T { public: template <typename...> - void setup(TensorShape shape, DataType data_type, BorderMode border_mode, const unsigned int width) + void setup(TensorShape shape, DataType output_data_type, BorderMode border_mode, const unsigned int width) { - ConvolutionValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, border_mode, width, width); + ConvolutionValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, output_data_type, border_mode, width, width); } protected: - TensorType compute_target(const TensorShape &shape, DataType data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value) + TensorType compute_target(const TensorShape &shape, DataType output_data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value) { // Create tensors - TensorType src = create_tensor<TensorType>(shape, data_type); - TensorType dst = create_tensor<TensorType>(shape, data_type); + TensorType src = create_tensor<TensorType>(shape, DataType::U8); + TensorType dst = create_tensor<TensorType>(shape, output_data_type); // Create and configure function FunctionType convolution; @@ -192,17 +190,17 @@ class ConvolutionSeparableValidationFixture : public ConvolutionValidationFixtur { public: template <typename...> - void setup(TensorShape shape, DataType data_type, BorderMode border_mode, const unsigned int width) + void setup(TensorShape shape, DataType output_data_type, BorderMode border_mode, const unsigned int width) { - ConvolutionValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, border_mode, width, width, true); + ConvolutionValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, output_data_type, border_mode, width, width, true); } protected: - TensorType compute_target(const TensorShape &shape, DataType data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value) + TensorType compute_target(const TensorShape &shape, DataType output_data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value) { // Create tensors - TensorType src = create_tensor<TensorType>(shape, data_type); - TensorType dst = create_tensor<TensorType>(shape, data_type); + TensorType src = create_tensor<TensorType>(shape, DataType::U8); + TensorType dst = create_tensor<TensorType>(shape, output_data_type); // Create and configure function FunctionType convolution; @@ -234,17 +232,17 @@ class ConvolutionRectangleValidationFixture : public ConvolutionValidationFixtur { public: template <typename...> - void setup(TensorShape shape, DataType data_type, BorderMode border_mode, const unsigned int width, const unsigned int height) + void setup(TensorShape shape, DataType output_data_type, BorderMode border_mode, const unsigned int width, const unsigned int height) { - ConvolutionValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, data_type, border_mode, width, height); + ConvolutionValidationFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, output_data_type, border_mode, width, height); } protected: - TensorType compute_target(const TensorShape &shape, DataType data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value) + TensorType compute_target(const TensorShape &shape, DataType output_data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value) { // Create tensors - TensorType src = create_tensor<TensorType>(shape, data_type); - TensorType dst = create_tensor<TensorType>(shape, data_type); + TensorType src = create_tensor<TensorType>(shape, DataType::U8); + TensorType dst = create_tensor<TensorType>(shape, output_data_type); // Create and configure function FunctionType convolution; diff --git a/tests/validation/fixtures/DirectConvolutionLayerTensorShiftFixture.h b/tests/validation/fixtures/DirectConvolutionLayerTensorShiftFixture.h new file mode 100644 index 000000000..d810a765c --- /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 <random> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DirectConvolutionValidationGenericTensorShiftFixture : public framework::Fixture +{ +public: + using TBias = typename std::conditional<std::is_same<typename std::decay<T>::type, uint8_t>::value, int32_t, T>::type; + +public: + template <typename...> + 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 <typename...> + 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 <typename U> + void fill(U &&tensor, int i) + { + switch(tensor.data_type()) + { + case DataType::QASYMM8: + { + std::uniform_int_distribution<uint8_t> 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<int32_t> 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<TensorType>(input_shape, data_type, 1, fixed_point_position, quantization_info); + TensorType weights = create_tensor<TensorType>(weights_shape, data_type, 1, fixed_point_position, quantization_info); + TensorType bias = create_tensor<TensorType>(bias_shape, bias_data_type, 1, fixed_point_position, quantization_info); + TensorType dst = create_tensor<TensorType>(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<TensorType>(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<T> 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<T> src{ input_shape, data_type, 1, fixed_point_position, quantization_info }; + SimpleTensor<T> weights{ weights_shape, data_type, 1, fixed_point_position, quantization_info }; + SimpleTensor<TBias> bias{ bias_shape, bias_data_type, 1, fixed_point_position, quantization_info }; + + SimpleTensor<T> 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<T>(src, weights, bias, output_shape, info); + return reference::convolution_layer<T>(dst, weights, bias, output_shape1, info); + } + + TensorType _target{}; + SimpleTensor<T> _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<unsigned int, unsigned int> 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 <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DirectConvolutionValidationTensorShiftFixture : public DirectConvolutionValidationGenericTensorShiftFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + template <typename...> + 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<TensorType, AccessorType, FunctionType, T>::setup(input_shape, stride_x, stride_y, pad_x, pad_y, kernel_size, num_kernels, data_type, 0, + QuantizationInfo()); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DirectConvolutionValidationFixedPointTensorShiftFixture : public DirectConvolutionValidationGenericTensorShiftFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + template <typename...> + 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<TensorType, AccessorType, FunctionType, T>::setup(input_shape, stride_x, stride_y, pad_x, pad_y, kernel_size, num_kernels, data_type, + fractional_bits, + QuantizationInfo()); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DirectConvolutionValidationQuantizedTensorShiftFixture : public DirectConvolutionValidationGenericTensorShiftFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + template <typename...> + 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<TensorType, AccessorType, FunctionType, T>::setup(input_shape, stride_x, stride_y, pad_x, pad_y, kernel_size, num_kernels, data_type, 0, + quantization_info); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DirectConvolutionValidationWithTensorShapesQuantizedTensorShiftFixture : public DirectConvolutionValidationGenericTensorShiftFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + template <typename...> + void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, + DataType data_type, QuantizationInfo quantization_info) + { + DirectConvolutionValidationGenericTensorShiftFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, data_type, 0, quantization_info); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class DirectConvolutionValidationWithTensorShapesTensorShiftFixture : public DirectConvolutionValidationGenericTensorShiftFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + template <typename...> + void setup(TensorShape input_shape, TensorShape weights_shape, TensorShape bias_shape, TensorShape output_shape, PadStrideInfo info, + DataType data_type) + { + DirectConvolutionValidationGenericTensorShiftFixture<TensorType, AccessorType, FunctionType, T>::setup(input_shape, weights_shape, bias_shape, output_shape, info, data_type, 0, QuantizationInfo()); + } +}; + +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/fixtures/EqualizeHistogramFixture.h b/tests/validation/fixtures/EqualizeHistogramFixture.h new file mode 100644 index 000000000..0d91e6d50 --- /dev/null +++ b/tests/validation/fixtures/EqualizeHistogramFixture.h @@ -0,0 +1,106 @@ +/* + * 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. + */ +#ifndef ARM_COMPUTE_TEST_EQUALIZE_HISTOGRAM_FIXTURE +#define ARM_COMPUTE_TEST_EQUALIZE_HISTOGRAM_FIXTURE + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.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/reference/EqualizeHistogram.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class EqualizeHistogramValidationFixture : public framework::Fixture +{ +public: + template <typename...> + void setup(TensorShape shape, DataType data_type) + { + _target = compute_target(shape, data_type); + _reference = compute_reference(shape, data_type); + } + +protected: + template <typename U> + void fill(U &&tensor) + { + library->fill_tensor_uniform(tensor, 0); + } + + TensorType compute_target(const TensorShape &shape, DataType data_type) + { + // Create tensors + TensorType src = create_tensor<TensorType>(shape, data_type); + TensorType dst = create_tensor<TensorType>(shape, data_type); + + // Create and configure function + FunctionType equalize_histogram; + + equalize_histogram.configure(&src, &dst); + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Allocate tensors + src.allocator()->allocate(); + dst.allocator()->allocate(); + ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Fill tensors + fill(AccessorType(src)); + + // Compute function + equalize_histogram.run(); + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &shape, DataType data_type) + { + // Create reference + SimpleTensor<T> src{ shape, data_type }; + + // Fill reference + fill(src); + + return reference::equalize_histogram<T>(src); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_EQUALIZE_HISTOGRAM_FIXTURE */ diff --git a/tests/validation/fixtures/FastCornersFixture.h b/tests/validation/fixtures/FastCornersFixture.h new file mode 100644 index 000000000..6f2add621 --- /dev/null +++ b/tests/validation/fixtures/FastCornersFixture.h @@ -0,0 +1,138 @@ +/* + * 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. + */ +#ifndef ARM_COMPUTE_TEST_FAST_CORNERS_FIXTURE +#define ARM_COMPUTE_TEST_FAST_CORNERS_FIXTURE + +#include "arm_compute/core/TensorShape.h" +#include "arm_compute/core/Types.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/reference/FastCorners.h" + +#include <random> + +namespace arm_compute +{ +class CLFastCorners; +class NEFastCorners; + +namespace test +{ +namespace validation +{ +template <typename TensorType, typename AccessorType, typename ArrayType, typename FunctionType, typename T> +class FastCornersValidationFixture : public framework::Fixture +{ +public: + template <typename...> + void setup(std::string image, Format format, bool suppress_nonmax, BorderMode border_mode) + { + std::mt19937 gen(library->seed()); + std::uniform_int_distribution<uint8_t> int_dist(0, 255); + std::uniform_real_distribution<float> real_dist(0, 255); + + const uint8_t constant_border_value = int_dist(gen); + const float threshold = real_dist(gen); + + _target = compute_target(image, format, threshold, suppress_nonmax, border_mode, constant_border_value); + _reference = compute_reference(image, format, threshold, suppress_nonmax, border_mode, constant_border_value); + } + +protected: + template <typename U> + void fill(U &&tensor, RawTensor raw) + { + library->fill(tensor, raw); + } + + template <typename F, typename std::enable_if<std::is_same<F, CLFastCorners>::value, int>::type = 0> + void configure_target(F &func, TensorType &src, ArrayType &corners, unsigned int *num_corners, float threshold, bool suppress_nonmax, BorderMode border_mode, uint8_t constant_border_value) + { + func.configure(&src, threshold, suppress_nonmax, &corners, num_corners, border_mode, constant_border_value); + } + + template <typename F, typename std::enable_if<std::is_same<F, NEFastCorners>::value, int>::type = 0> + void configure_target(F &func, TensorType &src, ArrayType &corners, unsigned int *num_corners, float threshold, bool suppress_nonmax, BorderMode border_mode, uint8_t constant_border_value) + { + ARM_COMPUTE_UNUSED(num_corners); + func.configure(&src, threshold, suppress_nonmax, &corners, border_mode, constant_border_value); + } + + ArrayType compute_target(const std::string &image, Format format, float threshold, bool suppress_nonmax, BorderMode border_mode, uint8_t constant_border_value) + { + // Load the image (cached by the library if loaded before) + const RawTensor &raw = library->get(image, format); + + // Create tensors + TensorType src = create_tensor<TensorType>(raw.shape(), format); + + // Create array of keypoints + ArrayType corners(raw.shape().total_size()); + unsigned int num_corners = raw.shape().total_size(); + + // Create and configure function + FunctionType fast_corners; + configure_target<FunctionType>(fast_corners, src, corners, &num_corners, threshold, suppress_nonmax, border_mode, constant_border_value); + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Allocate tensors + src.allocator()->allocate(); + + ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Fill tensors + fill(AccessorType(src), raw); + + // Compute function + fast_corners.run(); + + return corners; + } + + std::vector<KeyPoint> compute_reference(const std::string &image, Format format, float threshold, bool suppress_nonmax, BorderMode border_mode, uint8_t constant_border_value) + { + // Load the image (cached by the library if loaded before) + const RawTensor &raw = library->get(image, format); + + // Create reference + SimpleTensor<T> src{ raw.shape(), format }; + + // Fill reference + fill(src, raw); + + // Compute reference + return reference::fast_corners<T>(src, threshold, suppress_nonmax, border_mode, constant_border_value); + } + + ArrayType _target{}; + std::vector<KeyPoint> _reference{}; +}; +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* ARM_COMPUTE_TEST_FAST_CORNERS_FIXTURE */ diff --git a/tests/validation/fixtures/GEMMLowpAssemblyFixture.h b/tests/validation/fixtures/GEMMLowpAssemblyFixture.h index ff33c9db3..d6b94a197 100644 --- a/tests/validation/fixtures/GEMMLowpAssemblyFixture.h +++ b/tests/validation/fixtures/GEMMLowpAssemblyFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -98,8 +98,8 @@ protected: } else { - fill(AccessorType(a), 0, 0, 128); - fill(AccessorType(b), 1, 0, 128); + fill(AccessorType(a), 0, 0, 255); + fill(AccessorType(b), 1, 0, 255); } fill(AccessorType(c), 2, 0, 0); @@ -124,8 +124,8 @@ protected: } else { - fill(a, 0, 0, 128); - fill(b, 1, 0, 128); + fill(a, 0, 0, 255); + fill(b, 1, 0, 255); } return reference::gemmlowp<int32_t, T2>(a, b); diff --git a/tests/validation/fixtures/HarrisCornersFixture.h b/tests/validation/fixtures/HarrisCornersFixture.h index e3c29aed1..ae262af66 100644 --- a/tests/validation/fixtures/HarrisCornersFixture.h +++ b/tests/validation/fixtures/HarrisCornersFixture.h @@ -51,7 +51,7 @@ public: { HarrisCornersParameters params = harris_corners_parameters(); - _target = compute_target(image, gradient_size, block_size, border_mode, use_fp16, format, params); + _target = compute_target(image, gradient_size, block_size, border_mode, use_fp16, format, params); _reference = compute_reference(image, gradient_size, block_size, border_mode, format, params); } diff --git a/tests/validation/fixtures/NormalizePlanarYUVLayerFixture.h b/tests/validation/fixtures/NormalizePlanarYUVLayerFixture.h index ae5c53a05..4ec4ae647 100644 --- a/tests/validation/fixtures/NormalizePlanarYUVLayerFixture.h +++ b/tests/validation/fixtures/NormalizePlanarYUVLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -62,7 +62,7 @@ protected: float max_bound = 0.f; std::tie(min_bound, max_bound) = get_normalize_planar_yuv_layer_test_bounds<T>(); std::uniform_real_distribution<> distribution(min_bound, max_bound); - std::uniform_real_distribution<> distribution_sd(0, max_bound); + std::uniform_real_distribution<> distribution_sd(0.1, max_bound); library->fill(src_tensor, distribution, 0); library->fill(mean_tensor, distribution, 1); library->fill(sd_tensor, distribution_sd, 2); diff --git a/tests/validation/fixtures/PixelWiseMultiplicationFixture.h b/tests/validation/fixtures/PixelWiseMultiplicationFixture.h index 7428fb5cb..b9f19f3e7 100644 --- a/tests/validation/fixtures/PixelWiseMultiplicationFixture.h +++ b/tests/validation/fixtures/PixelWiseMultiplicationFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -40,19 +40,20 @@ namespace test namespace validation { template <typename TensorType, typename AccessorType, typename FunctionType, typename T1, typename T2> -class PixelWiseMultiplicationValidationFixture : public framework::Fixture +class PixelWiseMultiplicationBroadcastValidationFixture : public framework::Fixture { public: template <typename...> - void setup(TensorShape shape, - DataType dt_in1, - DataType dt_in2, - float scale, - ConvertPolicy convert_policy, - RoundingPolicy rounding_policy) + void setup(const TensorShape &shape0, + const TensorShape &shape1, + DataType dt_in1, + DataType dt_in2, + float scale, + ConvertPolicy convert_policy, + RoundingPolicy rounding_policy) { - _target = compute_target(shape, dt_in1, dt_in2, scale, convert_policy, rounding_policy); - _reference = compute_reference(shape, dt_in1, dt_in2, scale, convert_policy, rounding_policy); + _target = compute_target(shape0, shape1, dt_in1, dt_in2, scale, convert_policy, rounding_policy); + _reference = compute_reference(shape0, shape1, dt_in1, dt_in2, scale, convert_policy, rounding_policy); } protected: @@ -62,12 +63,13 @@ protected: library->fill_tensor_uniform(tensor, seed_offset); } - TensorType compute_target(const TensorShape &shape, DataType dt_in1, DataType dt_in2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy) + TensorType compute_target(const TensorShape &shape0, const TensorShape &shape1, DataType dt_in1, DataType dt_in2, + float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy) { // Create tensors - TensorType src1 = create_tensor<TensorType>(shape, dt_in1); - TensorType src2 = create_tensor<TensorType>(shape, dt_in2); - TensorType dst = create_tensor<TensorType>(shape, dt_in2); + TensorType src1 = create_tensor<TensorType>(shape0, dt_in1); + TensorType src2 = create_tensor<TensorType>(shape1, dt_in2); + TensorType dst = create_tensor<TensorType>(TensorShape::broadcast_shape(shape0, shape1), dt_in2); // Create and configure function FunctionType multiply; @@ -96,11 +98,12 @@ protected: return dst; } - SimpleTensor<T2> compute_reference(const TensorShape &shape, DataType dt_in1, DataType dt_in2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy) + SimpleTensor<T2> compute_reference(const TensorShape &shape0, const TensorShape &shape1, DataType dt_in1, DataType dt_in2, + float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy) { // Create reference - SimpleTensor<T1> src1{ shape, dt_in1 }; - SimpleTensor<T2> src2{ shape, dt_in2 }; + SimpleTensor<T1> src1{ shape0, dt_in1 }; + SimpleTensor<T2> src2{ shape1, dt_in2 }; // Fill reference fill(src1, 0); @@ -112,6 +115,18 @@ protected: TensorType _target{}; SimpleTensor<T2> _reference{}; }; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T1, typename T2> +class PixelWiseMultiplicationValidationFixture : public PixelWiseMultiplicationBroadcastValidationFixture<TensorType, AccessorType, FunctionType, T1, T2> +{ +public: + template <typename...> + void setup(const TensorShape &shape, DataType dt_in1, DataType dt_in2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy) + { + PixelWiseMultiplicationBroadcastValidationFixture<TensorType, AccessorType, FunctionType, T1, T2>::setup(shape, shape, dt_in1, dt_in2, scale, convert_policy, rounding_policy); + } +}; + } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/fixtures/PoolingLayerFixture.h b/tests/validation/fixtures/PoolingLayerFixture.h index 890eef2d4..3bbb403ae 100644 --- a/tests/validation/fixtures/PoolingLayerFixture.h +++ b/tests/validation/fixtures/PoolingLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -132,7 +132,7 @@ class PoolingLayerValidationFixture : public PoolingLayerValidationGenericFixtur { public: template <typename...> - void setup(TensorShape shape, PoolingType pool_type, int pool_size, PadStrideInfo pad_stride_info, bool exclude_padding, DataType data_type) + void setup(TensorShape shape, PoolingType pool_type, Size2D pool_size, PadStrideInfo pad_stride_info, bool exclude_padding, DataType data_type) { PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, PoolingLayerInfo(pool_type, pool_size, pad_stride_info, exclude_padding), data_type, 0, QuantizationInfo()); @@ -144,7 +144,7 @@ class PoolingLayerValidationFixedPointFixture : public PoolingLayerValidationGen { public: template <typename...> - void setup(TensorShape shape, PoolingType pool_type, int pool_size, PadStrideInfo pad_stride_info, bool exclude_padding, DataType data_type, int fractional_bits) + void setup(TensorShape shape, PoolingType pool_type, Size2D pool_size, PadStrideInfo pad_stride_info, bool exclude_padding, DataType data_type, int fractional_bits) { PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, PoolingLayerInfo(pool_type, pool_size, pad_stride_info, exclude_padding), data_type, fractional_bits, QuantizationInfo()); @@ -156,7 +156,7 @@ class PoolingLayerValidationQuantizedFixture : public PoolingLayerValidationGene { public: template <typename...> - void setup(TensorShape shape, PoolingType pool_type, int pool_size, PadStrideInfo pad_stride_info, bool exclude_padding, DataType data_type, QuantizationInfo quantization_info) + void setup(TensorShape shape, PoolingType pool_type, Size2D pool_size, PadStrideInfo pad_stride_info, bool exclude_padding, DataType data_type, QuantizationInfo quantization_info) { PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, PoolingLayerInfo(pool_type, pool_size, pad_stride_info, exclude_padding), data_type, 0, quantization_info); @@ -164,6 +164,18 @@ public: }; template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class SpecialPoolingLayerValidationFixture : public PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> +{ +public: + template <typename...> + void setup(TensorShape src_shape, TensorShape dst_shape, PoolingLayerInfo pool_info, DataType data_type) + { + ARM_COMPUTE_UNUSED(dst_shape); + PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(src_shape, pool_info, data_type, 0, QuantizationInfo()); + } +}; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> class GlobalPoolingLayerValidationFixture : public PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T> { public: @@ -173,6 +185,7 @@ public: PoolingLayerValidationGenericFixture<TensorType, AccessorType, FunctionType, T>::setup(shape, PoolingLayerInfo(pool_type), data_type, 0, QuantizationInfo()); } }; + } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/fixtures/RemapFixture.h b/tests/validation/fixtures/RemapFixture.h index 846ebf44a..78b30151a 100644 --- a/tests/validation/fixtures/RemapFixture.h +++ b/tests/validation/fixtures/RemapFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -50,7 +50,7 @@ public: { std::mt19937 gen(library->seed()); std::uniform_int_distribution<uint8_t> distribution(0, 255); - T constant_border_value = static_cast<T>(distribution(gen)); + const T constant_border_value = static_cast<T>(distribution(gen)); _target = compute_target(shape, policy, data_type, border_mode, constant_border_value); _reference = compute_reference(shape, policy, data_type, border_mode, constant_border_value); @@ -58,9 +58,10 @@ public: protected: template <typename U> - void fill(U &&tensor, int i) + void fill(U &&tensor, int i, float min, float max) { - library->fill_tensor_uniform(tensor, i); + std::uniform_int_distribution<> distribution((int)min, (int)max); + library->fill(tensor, distribution, i); } TensorType compute_target(const TensorShape &shape, InterpolationPolicy policy, DataType data_type, BorderMode border_mode, T constant_border_value) @@ -92,9 +93,9 @@ protected: ARM_COMPUTE_EXPECT(!dst.info()->is_resizable(), framework::LogLevel::ERRORS); // Fill tensors - fill(AccessorType(src), 0); - fill(AccessorType(map_x), 1); - fill(AccessorType(map_y), 2); + fill(AccessorType(src), 0, 0, 255); + fill(AccessorType(map_x), 1, -5, shape.x() + 5); + fill(AccessorType(map_y), 2, -5, shape.y() + 5); // Compute function remap.run(); @@ -115,9 +116,9 @@ protected: _valid_mask = SimpleTensor<T> { shape, data_type }; // Fill reference - fill(src, 0); - fill(map_x, 1); - fill(map_y, 2); + fill(src, 0, 0, 255); + fill(map_x, 1, -5, shape.x() + 5); + fill(map_y, 2, -5, shape.y() + 5); // Compute reference return reference::remap<T>(src, map_x, map_y, _valid_mask, policy, border_mode, constant_border_value); diff --git a/tests/validation/fixtures/WinogradLayerFixture.h b/tests/validation/fixtures/WinogradLayerFixture.h index 7aa26c714..d7f0cbfdf 100644 --- a/tests/validation/fixtures/WinogradLayerFixture.h +++ b/tests/validation/fixtures/WinogradLayerFixture.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -88,7 +88,7 @@ protected: // Create and configure function FunctionType conv; - conv.configure(&src, &weights, nullptr, &dst, info); + conv.configure(&src, &weights, &bias, &dst, info); ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(weights.info()->is_resizable(), framework::LogLevel::ERRORS); @@ -98,8 +98,8 @@ protected: // Allocate tensors src.allocator()->allocate(); weights.allocator()->allocate(); - bias.allocator()->allocate(); dst.allocator()->allocate(); + bias.allocator()->allocate(); ARM_COMPUTE_EXPECT(!src.info()->is_resizable(), framework::LogLevel::ERRORS); ARM_COMPUTE_EXPECT(!weights.info()->is_resizable(), framework::LogLevel::ERRORS); @@ -109,7 +109,7 @@ protected: // Fill tensors fill(AccessorType(src), 0, -1.f, 1.f); fill(AccessorType(weights), 1, -1.f, 1.f); - fill(AccessorType(bias), 2, 0.f, 0.f); + fill(AccessorType(bias), 2, -1.f, 1.f); fill(AccessorType(dst), 3, -1.f, 1.f); // Compute NEWinogradLayer function @@ -128,7 +128,7 @@ protected: // Fill reference fill(src, 0, -1.f, 1.f); fill(weights, 1, -1.f, 1.f); - fill(bias, 2, 0.f, 0.f); + fill(bias, 2, -1.f, 1.f); return reference::convolution_layer<T>(src, weights, bias, output_shape, info); } diff --git a/tests/validation/reference/ArithmeticAddition.cpp b/tests/validation/reference/ArithmeticAddition.cpp index 82dd1437c..17020a627 100644 --- a/tests/validation/reference/ArithmeticAddition.cpp +++ b/tests/validation/reference/ArithmeticAddition.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -35,27 +35,72 @@ namespace validation { namespace reference { +namespace +{ template <typename T> -SimpleTensor<T> arithmetic_addition(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, DataType dst_data_type, ConvertPolicy convert_policy) +T add(T src1, T src2, ConvertPolicy convert_policy) { - SimpleTensor<T> result(src1.shape(), dst_data_type); - using intermediate_type = typename common_promoted_signed_type<T>::intermediate_type; - for(int i = 0; i < src1.num_elements(); ++i) + intermediate_type val = static_cast<intermediate_type>(src1) + static_cast<intermediate_type>(src2); + + T result = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T>(val) : static_cast<T>(val); + + return result; +} + +template <size_t dim> +struct BroadcastUnroll +{ + template <typename T> + static void unroll(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst, + ConvertPolicy convert_policy, Coordinates &id_src1, Coordinates &id_src2, Coordinates &id_dst) { - intermediate_type val = static_cast<intermediate_type>(src1[i]) + static_cast<intermediate_type>(src2[i]); - result[i] = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T>(val) : static_cast<T>(val); + const bool src1_is_broadcast = (src1.shape()[dim - 1] != dst.shape()[dim - 1]); + const bool src2_is_broadcast = (src2.shape()[dim - 1] != dst.shape()[dim - 1]); + + id_src1.set(dim - 1, 0); + id_src2.set(dim - 1, 0); + id_dst.set(dim - 1, 0); + + for(size_t i = 0; i < dst.shape()[dim - 1]; ++i, ++id_dst[dim - 1]) + { + BroadcastUnroll < dim - 1 >::unroll(src1, src2, dst, convert_policy, id_src1, id_src2, id_dst); + + id_src1[dim - 1] += !src1_is_broadcast; + id_src2[dim - 1] += !src2_is_broadcast; + } } +}; - return result; +template <> +struct BroadcastUnroll<0> +{ + template <typename T> + static void unroll(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, SimpleTensor<T> &dst, + ConvertPolicy convert_policy, Coordinates &id_src1, Coordinates &id_src2, Coordinates &id_dst) + { + dst[coord2index(dst.shape(), id_dst)] = add(src1[coord2index(src1.shape(), id_src1)], src2[coord2index(src2.shape(), id_src2)], convert_policy); + } +}; +} // namespace + +template <typename T> +SimpleTensor<T> arithmetic_addition(const SimpleTensor<T> &src1, const SimpleTensor<T> &src2, DataType dst_data_type, ConvertPolicy convert_policy) +{ + SimpleTensor<T> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), dst_data_type); + + Coordinates id_src1, id_src2, id_dst; + + BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, convert_policy, id_src1, id_src2, id_dst); + + return dst; } template SimpleTensor<uint8_t> arithmetic_addition(const SimpleTensor<uint8_t> &src1, const SimpleTensor<uint8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy); template SimpleTensor<int16_t> arithmetic_addition(const SimpleTensor<int16_t> &src1, const SimpleTensor<int16_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy); template SimpleTensor<int8_t> arithmetic_addition(const SimpleTensor<int8_t> &src1, const SimpleTensor<int8_t> &src2, DataType dst_data_type, ConvertPolicy convert_policy); -template SimpleTensor<half> arithmetic_addition(const SimpleTensor<half> &src1, const SimpleTensor<half> &src2, DataType dst_data_type, - ConvertPolicy convert_policy); +template SimpleTensor<half> arithmetic_addition(const SimpleTensor<half> &src1, const SimpleTensor<half> &src2, DataType dst_data_type, ConvertPolicy convert_policy); template SimpleTensor<float> arithmetic_addition(const SimpleTensor<float> &src1, const SimpleTensor<float> &src2, DataType dst_data_type, ConvertPolicy convert_policy); } // namespace reference } // namespace validation diff --git a/tests/validation/reference/BatchNormalizationLayer.cpp b/tests/validation/reference/BatchNormalizationLayer.cpp index e4446d169..a9d9f0320 100644 --- a/tests/validation/reference/BatchNormalizationLayer.cpp +++ b/tests/validation/reference/BatchNormalizationLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -23,6 +23,8 @@ */ #include "BatchNormalizationLayer.h" +#include "ActivationLayer.h" + #include "tests/validation/FixedPoint.h" #include "tests/validation/Helpers.h" @@ -37,8 +39,9 @@ namespace reference // Batch Normalization Layer for fixed point type template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type *> SimpleTensor<T> batch_normalization_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &mean, const SimpleTensor<T> &var, const SimpleTensor<T> &beta, const SimpleTensor<T> &gamma, float epsilon, - int fixed_point_position) + ActivationLayerInfo act_info, int fixed_point_position) { + ARM_COMPUTE_UNUSED(act_info); SimpleTensor<T> result(src.shape(), src.data_type()); const auto cols = static_cast<int>(src.shape()[0]); @@ -79,7 +82,7 @@ SimpleTensor<T> batch_normalization_layer(const SimpleTensor<T> &src, const Simp // Batch Normalization Layer for floating point type template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type *> SimpleTensor<T> batch_normalization_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &mean, const SimpleTensor<T> &var, const SimpleTensor<T> &beta, const SimpleTensor<T> &gamma, float epsilon, - int fixed_point_position) + ActivationLayerInfo act_info, int fixed_point_position) { ARM_COMPUTE_UNUSED(fixed_point_position); @@ -103,21 +106,28 @@ SimpleTensor<T> batch_normalization_layer(const SimpleTensor<T> &src, const Simp const float numerator = src[pos] - mean[i]; const float x_bar = numerator / denominator; result[pos] = beta[i] + x_bar * gamma[i]; + ; } } } } + + if(act_info.enabled()) + { + result = activation_layer(result, act_info); + } + return result; } template SimpleTensor<float> batch_normalization_layer(const SimpleTensor<float> &src, const SimpleTensor<float> &mean, const SimpleTensor<float> &var, const SimpleTensor<float> &beta, - const SimpleTensor<float> &gamma, float epsilon, int fixed_point_position); + const SimpleTensor<float> &gamma, float epsilon, ActivationLayerInfo act_info, int fixed_point_position); template SimpleTensor<int8_t> batch_normalization_layer(const SimpleTensor<int8_t> &src, const SimpleTensor<int8_t> &mean, const SimpleTensor<int8_t> &var, const SimpleTensor<int8_t> &beta, - const SimpleTensor<int8_t> &gamma, float epsilon, int fixed_point_position); + const SimpleTensor<int8_t> &gamma, float epsilon, ActivationLayerInfo act_info, int fixed_point_position); template SimpleTensor<int16_t> batch_normalization_layer(const SimpleTensor<int16_t> &src, const SimpleTensor<int16_t> &mean, const SimpleTensor<int16_t> &var, const SimpleTensor<int16_t> &beta, - const SimpleTensor<int16_t> &gamma, float epsilon, int fixed_point_position); + const SimpleTensor<int16_t> &gamma, float epsilon, ActivationLayerInfo act_info, int fixed_point_position); template SimpleTensor<half> batch_normalization_layer(const SimpleTensor<half> &src, const SimpleTensor<half> &mean, const SimpleTensor<half> &var, const SimpleTensor<half> &beta, - const SimpleTensor<half> &gamma, float epsilon, int fixed_point_position); + const SimpleTensor<half> &gamma, float epsilon, ActivationLayerInfo act_info, int fixed_point_position); } // namespace reference } // namespace validation diff --git a/tests/validation/reference/BatchNormalizationLayer.h b/tests/validation/reference/BatchNormalizationLayer.h index 1a554adf7..329909dab 100644 --- a/tests/validation/reference/BatchNormalizationLayer.h +++ b/tests/validation/reference/BatchNormalizationLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -37,11 +37,13 @@ namespace reference { template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type * = nullptr> SimpleTensor<T> batch_normalization_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &mean, const SimpleTensor<T> &var, const SimpleTensor<T> &beta, const SimpleTensor<T> &gamma, float epsilon, - int fixed_point_position); + ActivationLayerInfo act_info, + int fixed_point_position); template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type * = nullptr> SimpleTensor<T> batch_normalization_layer(const SimpleTensor<T> &src, const SimpleTensor<T> &mean, const SimpleTensor<T> &var, const SimpleTensor<T> &beta, const SimpleTensor<T> &gamma, float epsilon, - int fixed_point_position); + ActivationLayerInfo act_info, + int fixed_point_position); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/ChannelExtract.cpp b/tests/validation/reference/ChannelExtract.cpp new file mode 100644 index 000000000..595bb1309 --- /dev/null +++ b/tests/validation/reference/ChannelExtract.cpp @@ -0,0 +1,76 @@ +/* + * 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 "ChannelExtract.h" + +#include "arm_compute/core/Types.h" +#include "tests/validation/FixedPoint.h" +#include "tests/validation/Helpers.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +template <typename T> +SimpleTensor<uint8_t> channel_extract(const TensorShape &shape, const std::vector<SimpleTensor<T>> &tensor_planes, Format format, Channel channel) +{ + // Find plane and channel index + const unsigned int plane_idx = plane_idx_from_channel(format, channel); + const unsigned int channel_idx = channel_idx_from_format(format, channel); + + // Create dst and get src tensor + SimpleTensor<T> src = tensor_planes[plane_idx]; + SimpleTensor<T> dst{ calculate_subsampled_shape(shape, format, channel), Format::U8 }; + + // Single planar formats with subsampling require a double horizontal step + const int step_x = ((Format::YUYV422 == format || Format::UYVY422 == format) && Channel::Y != channel) ? 2 : 1; + const int width = dst.shape().x(); + const int height = dst.shape().y(); + + // Loop over each pixel and extract channel + for(int y = 0; y < height; ++y) + { + for(int x = 0; x < width; ++x) + { + const Coordinates src_coord{ x * step_x, y }; + const Coordinates dst_coord{ x, y }; + + const auto *src_pixel = reinterpret_cast<const T *>(src(src_coord)); + auto *dst_pixel = reinterpret_cast<T *>(dst(dst_coord)); + + dst_pixel[0] = src_pixel[channel_idx]; + } + } + + return dst; +} + +template SimpleTensor<uint8_t> channel_extract(const TensorShape &shape, const std::vector<SimpleTensor<uint8_t>> &tensor_planes, Format format, Channel channel); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/reference/ChannelExtract.h b/tests/validation/reference/ChannelExtract.h new file mode 100644 index 000000000..ac7aefbde --- /dev/null +++ b/tests/validation/reference/ChannelExtract.h @@ -0,0 +1,43 @@ +/* + * 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. + */ +#ifndef __ARM_COMPUTE_TEST_CHANNEL_EXTRACT_H__ +#define __ARM_COMPUTE_TEST_CHANNEL_EXTRACT_H__ + +#include "tests/SimpleTensor.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +template <typename T> +SimpleTensor<uint8_t> channel_extract(const TensorShape &shape, const std::vector<SimpleTensor<T>> &tensor_planes, Format format, Channel channel); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* __ARM_COMPUTE_TEST_CHANNEL_EXTRACT_H__ */ diff --git a/tests/validation/reference/Convolution.cpp b/tests/validation/reference/Convolution.cpp index 777e2df40..e3d4b4ac1 100644 --- a/tests/validation/reference/Convolution.cpp +++ b/tests/validation/reference/Convolution.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -35,34 +35,29 @@ namespace validation namespace reference { template <typename T> -SimpleTensor<T> convolution(const SimpleTensor<T> &src, const int16_t *conv, uint32_t scale, BorderMode border_mode, T constant_border_value, const unsigned int width, const unsigned int height) +SimpleTensor<T> convolution(const SimpleTensor<uint8_t> &src, DataType output_data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value, + const unsigned int width, + const unsigned int height) { - SimpleTensor<T> dst(src.shape(), src.data_type()); - SimpleTensor<int32_t> sum(src.shape(), src.data_type()); + ARM_COMPUTE_ERROR_ON(scale == 0); + ARM_COMPUTE_ERROR_ON(scale >= std::numeric_limits<int32_t>::max()); + + SimpleTensor<T> dst(src.shape(), output_data_type); + SimpleTensor<int32_t> sum(src.shape(), output_data_type); for(int element_idx = 0; element_idx < src.num_elements(); ++element_idx) { const Coordinates id = index2coord(src.shape(), element_idx); apply_2d_spatial_filter(id, src, sum, TensorShape(width, height), conv, 1, border_mode, constant_border_value); - - if(tensor_elem_at<int32_t>(sum, id, border_mode, constant_border_value) < 0) - { - dst[element_idx] = 0; - } - else if((tensor_elem_at<int32_t>(sum, id, border_mode, constant_border_value) / scale) > 255) - { - dst[element_idx] = 255; - } - else - { - dst[element_idx] = tensor_elem_at<int32_t>(sum, id, border_mode, constant_border_value) / scale; - } + dst[element_idx] = saturate_cast<T>(tensor_elem_at<int32_t>(sum, id, border_mode, constant_border_value) / static_cast<int>(scale)); } return dst; } -template SimpleTensor<uint8_t> convolution(const SimpleTensor<uint8_t> &src, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value, +template SimpleTensor<uint8_t> convolution(const SimpleTensor<uint8_t> &src, DataType output_data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value, + const unsigned int widht, const unsigned int height); +template SimpleTensor<int16_t> convolution(const SimpleTensor<uint8_t> &src, DataType output_data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value, const unsigned int widht, const unsigned int height); } // namespace reference } // namespace validation diff --git a/tests/validation/reference/Convolution.h b/tests/validation/reference/Convolution.h index ea9f4e444..b217da7cd 100644 --- a/tests/validation/reference/Convolution.h +++ b/tests/validation/reference/Convolution.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -35,7 +35,9 @@ namespace validation namespace reference { template <typename T> -SimpleTensor<T> convolution(const SimpleTensor<T> &src, const int16_t *conv, uint32_t scale, BorderMode border_mode, T constant_border_value, const unsigned int width, const unsigned int height); +SimpleTensor<T> convolution(const SimpleTensor<uint8_t> &src, DataType output_data_type, const int16_t *conv, uint32_t scale, BorderMode border_mode, uint8_t constant_border_value, + const unsigned int width, + const unsigned int height); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/ConvolutionLayer.cpp b/tests/validation/reference/ConvolutionLayer.cpp index 567fac0f5..b7ed2f56c 100644 --- a/tests/validation/reference/ConvolutionLayer.cpp +++ b/tests/validation/reference/ConvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -58,8 +58,10 @@ void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<T> &weights, co const TB *b_ptr = bias.data() + b_offset; T *out_ptr = out.data() + o_offset; - const int half_width_weights = width_weights / 2; - const int half_height_weights = height_weights / 2; + const int half_width_weights_start = width_weights / 2; + const int half_width_weights_end = ((width_weights % 2) == 0) ? (half_width_weights_start - 1) : half_width_weights_start; + const int half_height_weights_start = height_weights / 2; + const int half_height_weights_end = ((height_weights % 2) == 0) ? (half_height_weights_start - 1) : half_height_weights_start; // Reset accumulator T acc(0); @@ -71,15 +73,15 @@ void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<T> &weights, co const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in; // Compute 2D convolution - for(int yk = -half_height_weights; yk <= half_height_weights; ++yk) + for(int yk = -half_height_weights_start; yk <= half_height_weights_end; ++yk) { - for(int xk = -half_width_weights; xk <= half_width_weights; ++xk) + for(int xk = -half_width_weights_start; xk <= half_width_weights_end; ++xk) { // Check if the pixel is out-of-bound if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in)) { - const int idx = xk + half_width_weights; - const int idy = yk + half_height_weights; + const int idx = xk + half_width_weights_start; + const int idy = yk + half_height_weights_start; const T i_value = in_ptr[offset_slice_in + xk + yk * width_in]; const T w_value = w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights]; @@ -106,8 +108,10 @@ void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<T> &weights, co T *out_ptr = out.data() + o_offset; int fixed_point_position = in.fixed_point_position(); - const int half_width_weights = width_weights / 2; - const int half_height_weights = height_weights / 2; + const int half_width_weights_start = width_weights / 2; + const int half_width_weights_end = ((width_weights % 2) == 0) ? (half_width_weights_start - 1) : half_width_weights_start; + const int half_height_weights_start = height_weights / 2; + const int half_height_weights_end = ((height_weights % 2) == 0) ? (half_height_weights_start - 1) : half_height_weights_start; using namespace fixed_point_arithmetic; using promoted_type = fixed_point_arithmetic::traits::promote_t<T>; @@ -122,15 +126,15 @@ void convolution3d(const SimpleTensor<T> &in, const SimpleTensor<T> &weights, co const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in; // Compute 2D convolution - for(int yk = -half_height_weights; yk <= half_height_weights; ++yk) + for(int yk = -half_height_weights_start; yk <= half_height_weights_end; ++yk) { - for(int xk = -half_width_weights; xk <= half_width_weights; ++xk) + for(int xk = -half_width_weights_start; xk <= half_width_weights_end; ++xk) { // Check if the pixel is out-of-bound if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in)) { - const int idx = xk + half_width_weights; - const int idy = yk + half_height_weights; + const int idx = xk + half_width_weights_start; + const int idy = yk + half_height_weights_start; const fixed_point<promoted_type> i_value(in_ptr[offset_slice_in + xk + yk * width_in], fixed_point_position, true); const fixed_point<promoted_type> w_value(w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights], fixed_point_position, true); @@ -173,8 +177,10 @@ void convolution3d(const SimpleTensor<uint8_t> &in, const SimpleTensor<uint8_t> const float multiplier = input_scale * weights_scale / output_scale; arm_compute::quantization::calculate_quantized_multiplier_less_than_one(multiplier, &output_multiplier, &output_shift); - const int half_width_weights = width_weights / 2; - const int half_height_weights = height_weights / 2; + const int half_width_weights_start = width_weights / 2; + const int half_width_weights_end = ((width_weights % 2) == 0) ? (half_width_weights_start - 1) : half_width_weights_start; + const int half_height_weights_start = height_weights / 2; + const int half_height_weights_end = ((height_weights % 2) == 0) ? (half_height_weights_start - 1) : half_height_weights_start; // Reset accumulator int32_t acc(0); @@ -186,15 +192,15 @@ void convolution3d(const SimpleTensor<uint8_t> &in, const SimpleTensor<uint8_t> const int offset_slice_in = xi + yi * width_in + ifm * width_in * height_in; // Compute 2D convolution - for(int yk = -half_height_weights; yk <= half_height_weights; ++yk) + for(int yk = -half_height_weights_start; yk <= half_height_weights_end; ++yk) { - for(int xk = -half_width_weights; xk <= half_width_weights; ++xk) + for(int xk = -half_width_weights_start; xk <= half_width_weights_end; ++xk) { // Check if the pixel is out-of-bound if(is_valid_pixel(xi + xk, 0, width_in) && is_valid_pixel(yi + yk, 0, height_in)) { - const int idx = xk + half_width_weights; - const int idy = yk + half_height_weights; + const int idx = xk + half_width_weights_start; + const int idy = yk + half_height_weights_start; const uint8_t i_value = in_ptr[offset_slice_in + xk + yk * width_in]; const uint8_t w_value = w_ptr[idx + idy * width_weights + ifm * width_weights * height_weights]; @@ -233,17 +239,17 @@ SimpleTensor<T> convolution_layer(const SimpleTensor<T> &src, const SimpleTensor const int width_weights = weights.shape().x(); const int height_weights = weights.shape().y(); const int depth_weights = weights.shape().z(); - const int pad_left = std::min(static_cast<int>(info.pad_left()), width_weights / 2); - const int pad_top = std::min(static_cast<int>(info.pad_top()), height_weights / 2); - const int pad_right = std::min(static_cast<int>(info.pad_right()), width_weights / 2); - const int pad_bottom = std::min(static_cast<int>(info.pad_bottom()), height_weights / 2); + const int pad_left = info.pad_left(); + const int pad_top = info.pad_top(); + const int stride_xi = info.stride().first; + const int stride_yi = info.stride().second; + + auto output_wh = scaled_dimensions(width_in, height_in, width_weights, height_weights, info); const int start_xi = width_weights / 2 - pad_left; const int start_yi = height_weights / 2 - pad_top; - const int end_xi = width_in + pad_left - width_weights / 2 + pad_right - width_weights / 2; - const int end_yi = height_in + pad_top - height_weights / 2 + pad_bottom - height_weights / 2; - const int stride_xi = info.stride().first; - const int stride_yi = info.stride().second; + const int end_xi = output_wh.first * stride_xi; + const int end_yi = output_wh.second * stride_yi; const int num_batches = src.shape().total_size() / (width_in * height_in * depth_in); for(int r = 0; r < num_batches; ++r) diff --git a/tests/validation/reference/DeconvolutionLayer.cpp b/tests/validation/reference/DeconvolutionLayer.cpp index 0cf108734..617f6908e 100644 --- a/tests/validation/reference/DeconvolutionLayer.cpp +++ b/tests/validation/reference/DeconvolutionLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * diff --git a/tests/validation/reference/DepthwiseConvolutionLayer.cpp b/tests/validation/reference/DepthwiseConvolutionLayer.cpp index 6ca347f1d..b2a706770 100644 --- a/tests/validation/reference/DepthwiseConvolutionLayer.cpp +++ b/tests/validation/reference/DepthwiseConvolutionLayer.cpp @@ -67,10 +67,10 @@ SimpleTensor<T> depthwise_convolution(const SimpleTensor<T> &src, const SimpleTe const int filter_half_width = filter_width / 2; const int filter_half_height = filter_height / 2; - const int pad_left = std::min(static_cast<int>(conv_info.pad_left()), filter_half_width); - const int pad_top = std::min(static_cast<int>(conv_info.pad_top()), filter_half_height); - const int pad_right = std::min(static_cast<int>(conv_info.pad_right()), filter_half_width); - const int pad_bottom = std::min(static_cast<int>(conv_info.pad_bottom()), filter_half_height); + const int pad_left = conv_info.pad_left(); + const int pad_top = conv_info.pad_top(); + const int pad_right = conv_info.pad_right(); + const int pad_bottom = conv_info.pad_bottom(); const int minimum_x = -pad_left + filter_half_width; const int minimum_y = -pad_top + filter_half_height; @@ -140,11 +140,18 @@ SimpleTensor<uint8_t> depthwise_convolution(const SimpleTensor<uint8_t> &src, co const int input_depth = src.shape().z(); const int num_batches = src.shape().total_size() / (input_width * input_height * input_depth); - const int filter_half_size = filter_width / 2; - const int pad_x = std::min(filter_half_size, static_cast<int>(conv_info.pad().first)); - const int pad_y = std::min(filter_half_size, static_cast<int>(conv_info.pad().second)); - const int minimum_x = -pad_x + filter_half_size; - const int minimum_y = -pad_y + filter_half_size; + const int filter_half_width = filter_width / 2; + const int filter_half_height = filter_height / 2; + + const int pad_left = conv_info.pad_left(); + const int pad_top = conv_info.pad_top(); + const int pad_right = conv_info.pad_right(); + const int pad_bottom = conv_info.pad_bottom(); + + const int minimum_x = -pad_left + filter_half_width; + const int minimum_y = -pad_top + filter_half_height; + const int maximum_x = input_width + pad_left - filter_half_width + pad_right - filter_half_width; + const int maximum_y = input_height + pad_top - filter_half_height + pad_bottom - filter_half_height; int out_pos = 0; for(int r = 0; r < num_batches; ++r) @@ -152,17 +159,17 @@ SimpleTensor<uint8_t> depthwise_convolution(const SimpleTensor<uint8_t> &src, co for(int z = 0; z < input_depth; ++z) { int32_t bias_val = *static_cast<const int32_t *>(biases(Coordinates(z))); - for(int y = minimum_y; y < input_height + pad_y - filter_half_size; y += conv_info.stride().second) + for(int y = minimum_y; y < minimum_y + maximum_y; y += conv_info.stride().second) { - for(int x = minimum_x; x < input_width + pad_x - filter_half_size; x += conv_info.stride().first) + for(int x = minimum_x; x < minimum_x + maximum_x; x += conv_info.stride().first) { Coordinates coords(x, y, z, r); int filter_offset = filter_plane * z; int32_t val = 0; - for(int j = y - filter_half_size; j <= (y + filter_half_size); ++j) + for(int j = y - filter_half_height; j <= (y + filter_half_height); ++j) { - for(int i = x - filter_half_size; i <= (x + filter_half_size); ++i) + for(int i = x - filter_half_width; i <= (x + filter_half_width); ++i) { coords.set(0, i); coords.set(1, j); diff --git a/tests/validation/reference/EqualizeHistogram.cpp b/tests/validation/reference/EqualizeHistogram.cpp new file mode 100644 index 000000000..0e966cd0b --- /dev/null +++ b/tests/validation/reference/EqualizeHistogram.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 "EqualizeHistogram.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +template <typename T> +SimpleTensor<T> equalize_histogram(const SimpleTensor<T> &src) +{ + const size_t num_bins = 256; // 0-255 inclusive + + std::vector<T> lut(num_bins); + std::vector<uint32_t> hist(num_bins); + std::vector<uint32_t> cd(num_bins); // cumulative distribution + + SimpleTensor<T> dst(src.shape(), src.data_type()); + + // Create the histogram + for(int element_idx = 0; element_idx < src.num_elements(); ++element_idx) + { + hist[src[element_idx]]++; + } + + // Calculate cumulative distribution + std::partial_sum(hist.begin(), hist.end(), cd.begin()); + + // Get the number of pixels that have the lowest non-zero value + const uint32_t cd_min = *std::find_if(hist.begin(), hist.end(), [](const uint32_t &x) + { + return x > 0; + }); + + const size_t total_num_pixels = cd.back(); + + // Single color - create linear distribution + if(total_num_pixels == cd_min) + { + std::iota(lut.begin(), lut.end(), 0); + } + else + { + const float diff = total_num_pixels - 1; + + for(size_t i = 0; i < num_bins; ++i) + { + lut[i] = lround((cd[i] - cd_min) / diff * 255.f); + } + } + + // Fill output tensor with equalized values + for(int i = 0; i < src.num_elements(); ++i) + { + dst[i] = lut[src[i]]; + } + + return dst; +} + +template SimpleTensor<uint8_t> equalize_histogram(const SimpleTensor<uint8_t> &src); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/reference/EqualizeHistogram.h b/tests/validation/reference/EqualizeHistogram.h new file mode 100644 index 000000000..286a423ce --- /dev/null +++ b/tests/validation/reference/EqualizeHistogram.h @@ -0,0 +1,43 @@ +/* + * 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. + */ +#ifndef __ARM_COMPUTE_TEST_EQUALIZE_HISTOGRAM_H__ +#define __ARM_COMPUTE_TEST_EQUALIZE_HISTOGRAM_H__ + +#include "tests/SimpleTensor.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +template <typename T> +SimpleTensor<T> equalize_histogram(const SimpleTensor<T> &src); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* __ARM_COMPUTE_TEST_EQUALIZE_HISTOGRAM_H__ */ diff --git a/tests/validation/reference/FastCorners.cpp b/tests/validation/reference/FastCorners.cpp new file mode 100644 index 000000000..bcea3f10b --- /dev/null +++ b/tests/validation/reference/FastCorners.cpp @@ -0,0 +1,240 @@ +/* + * 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 "FastCorners.h" + +#include "Utils.h" +#include "tests/validation/Helpers.h" +#include "tests/validation/reference/NonMaximaSuppression.h" + +#include "tests/framework/Asserts.h" +#include <iomanip> + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +namespace +{ +constexpr unsigned int bresenham_radius = 3; +constexpr unsigned int bresenham_count = 16; + +/* + Offsets of the 16 pixels in the Bresenham circle of radius 3 centered on P + . . . . . . . . . + . . . F 0 1 . . . + . . E . . . 2 . . + . D . . . . . 3 . + . C . . P . . 4 . + . B . . . . . 5 . + . . A . . . 6 . . + . . . 9 8 7 . . . + . . . . . . . . . +*/ +const std::array<std::array<int, 2>, 16> circle_offsets = +{ + { + { { 0, -3 } }, // 0 - pixel #1 + { { 1, -3 } }, // 1 - pixel #2 + { { 2, -2 } }, // 2 - pixel #3 + { { 3, -1 } }, // 3 - pixel #4 + { { 3, 0 } }, // 4 - pixel #5 + { { 3, 1 } }, // 5 - pixel #6 + { { 2, 2 } }, // 6 - pixel #7 + { { 1, 3 } }, // 7 - pixel #8 + { { 0, 3 } }, // 8 - pixel #9 + { { -1, 3 } }, // 9 - pixel #10 + { { -2, 2 } }, // A - pixel #11 + { { -3, 1 } }, // B - pixel #12 + { { -3, 0 } }, // C - pixel #13 + { { -3, -1 } }, // D - pixel #14 + { { -2, -2 } }, // E - pixel #15 + { { -1, -3 } } // F - pixel #16 + } +}; + +/* + FAST-9 bit masks for consecutive points surrounding a corner candidate + Rejection of non-corners is expedited by checking pixels 1, 9, then 5, 13... +*/ +const std::array<uint16_t, 16> fast9_masks = +{ + { + 0x01FF, // 0000 0001 1111 1111 + 0x03FE, // 0000 0011 1111 1110 + 0x07FC, // 0000 0111 1111 1100 + 0x0FF8, // 0000 1111 1111 1000 + 0x1FF0, // 0001 1111 1111 0000 + 0x3FE0, // 0011 1111 1110 0000 + 0x7FC0, // 0111 1111 1100 0000 + 0xFF80, // 1111 1111 1000 0000 + 0xFF01, // 1111 1111 0000 0001 + 0xFE03, // 1111 1110 0000 0011 + 0xFC07, // 1111 1100 0000 0111 + 0xF80F, // 1111 1000 0000 1111 + 0xF01F, // 1111 0000 0001 1111 + 0xE03F, // 1110 0000 0011 1111 + 0xC07F, // 1100 0000 0111 1111 + 0x80FF // 1000 0000 1111 1111 + } +}; + +inline bool in_range(const uint8_t low, const uint8_t high, const uint8_t val) +{ + return low <= val && val <= high; +} + +template <typename T, typename F> +bool is_a_corner(const Coordinates &candidate, const SimpleTensor<T> &src, uint8_t threshold, BorderMode border_mode, T constant_border_value, F intensity_at) +{ + const auto intensity_p = tensor_elem_at(src, candidate, border_mode, constant_border_value); + const auto thresh_bright = intensity_p + threshold; + const auto thresh_dark = intensity_p - threshold; + + // Quicker rejection of non-corner points by checking pixels 1, 9 then 5, 13 around the candidate + const auto p1 = intensity_at(candidate, 0); + const auto p9 = intensity_at(candidate, 8); + const auto p5 = intensity_at(candidate, 4); + const auto p13 = intensity_at(candidate, 12); + + if((in_range(thresh_dark, thresh_bright, p1) && in_range(thresh_dark, thresh_bright, p9)) + || (in_range(thresh_dark, thresh_bright, p5) && in_range(thresh_dark, thresh_bright, p13))) + { + return false; + } + + uint16_t mask_bright = 0; + uint16_t mask_dark = 0; + + // Set bits of the brighter/darker pixels mask accordingly + for(unsigned int n = 0; n < bresenham_count; ++n) + { + T intensity_n = intensity_at(candidate, n); + mask_bright |= (intensity_n > thresh_bright) << n; + mask_dark |= (intensity_n < thresh_dark) << n; + } + + // Mark as corner candidate if brighter/darker pixel sequence satisfies any one of the FAST-9 masks + const auto found = std::find_if(fast9_masks.begin(), fast9_masks.end(), [&](decltype(fast9_masks[0]) mask) + { + return (mask_bright & mask) == mask || (mask_dark & mask) == mask; + }); + + return found != fast9_masks.end(); +} +} // namespace + +template <typename T> +std::vector<KeyPoint> fast_corners(const SimpleTensor<T> &src, float input_thresh, bool suppress_nonmax, BorderMode border_mode, T constant_border_value) +{ + // Get intensity of pixel at given index on the Bresenham circle around a candidate point + const auto intensity_at = [&](const Coordinates & point, const unsigned int idx) + { + const auto offset = circle_offsets[idx]; + Coordinates px{ point.x() + offset[0], point.y() + offset[1] }; + return tensor_elem_at(src, px, border_mode, constant_border_value); + }; + + const auto threshold = static_cast<uint8_t>(input_thresh); + std::vector<KeyPoint> corners; + + // 1. Detect potential corners (the segment test) + std::vector<Coordinates> corner_candidates; + SimpleTensor<uint8_t> scores(src.shape(), DataType::U8); + ValidRegion valid_region = shape_to_valid_region(src.shape(), BorderMode::UNDEFINED == border_mode, BorderSize(bresenham_radius)); + + for(int i = 0; i < src.num_elements(); ++i) + { + Coordinates candidate = index2coord(src.shape(), i); + scores[i] = 0; + if(!is_in_valid_region(valid_region, candidate)) + { + continue; + } + + if(is_a_corner(candidate, src, threshold, border_mode, constant_border_value, intensity_at)) + { + corner_candidates.emplace_back(candidate); + scores[i] = 1; + } + } + + // 2. Calculate corner scores if necessary + if(suppress_nonmax) + { + for(const auto &candidate : corner_candidates) + { + const auto index = coord2index(scores.shape(), candidate); + uint8_t thresh_max = UINT8_MAX; + uint8_t thresh_min = threshold; + uint8_t response = (thresh_min + thresh_max) / 2; + + // Corner score (response) is the largest threshold for which the pixel remains a corner + while(thresh_max - thresh_min > 1) + { + response = (thresh_min + thresh_max) / 2; + if(is_a_corner(candidate, src, response, border_mode, constant_border_value, intensity_at)) + { + thresh_min = response; // raise threshold + } + else + { + thresh_max = response; // lower threshold + } + } + scores[index] = thresh_min; + } + + scores = non_maxima_suppression(scores, border_mode, constant_border_value); + valid_region = shape_to_valid_region(scores.shape(), BorderMode::UNDEFINED == border_mode, BorderSize(bresenham_radius + 1)); + } + + for(const auto &candidate : corner_candidates) + { + const auto index = coord2index(scores.shape(), candidate); + if(scores[index] > 0.f && is_in_valid_region(valid_region, candidate)) + { + KeyPoint corner; + corner.x = candidate.x(); + corner.y = candidate.y(); + corner.strength = scores[index]; + corner.tracking_status = 1; + corner.scale = 0.f; + corner.orientation = 0.f; + corner.error = 0.f; + corners.emplace_back(corner); + } + } + + return corners; +} + +template std::vector<KeyPoint> fast_corners(const SimpleTensor<uint8_t> &src, float threshold, bool suppress_nonmax, BorderMode border_mode, uint8_t constant_border_value); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/reference/FastCorners.h b/tests/validation/reference/FastCorners.h new file mode 100644 index 000000000..3d070d97b --- /dev/null +++ b/tests/validation/reference/FastCorners.h @@ -0,0 +1,44 @@ +/* + * 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. + */ +#ifndef __ARM_COMPUTE_TEST_FAST_CORNERS_H__ +#define __ARM_COMPUTE_TEST_FAST_CORNERS_H__ + +#include "arm_compute/core/Types.h" +#include "tests/SimpleTensor.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace reference +{ +template <typename T> +std::vector<KeyPoint> fast_corners(const SimpleTensor<T> &src, float input_thresh, bool suppress_nonmax, BorderMode border_mode, T constant_border_value = 0); +} // namespace reference +} // namespace validation +} // namespace test +} // namespace arm_compute +#endif /* __ARM_COMPUTE_TEST_FAST_CORNERS_H__ */ diff --git a/tests/validation/reference/HOGDescriptor.cpp b/tests/validation/reference/HOGDescriptor.cpp index 369ac74ed..105eb838e 100644 --- a/tests/validation/reference/HOGDescriptor.cpp +++ b/tests/validation/reference/HOGDescriptor.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017, 2018 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -40,17 +40,16 @@ namespace template <typename T> void hog_orientation_compute(const SimpleTensor<T> &mag, const SimpleTensor<T> &phase, std::vector<T> &bins, const HOGInfo &hog_info) { - const size_t num_bins = hog_info.num_bins(); - const size_t cell_height = hog_info.cell_size().height; - const size_t cell_width = hog_info.cell_size().width; + const Size2D &cell_size = hog_info.cell_size(); + const size_t num_bins = hog_info.num_bins(); float phase_scale = (PhaseType::SIGNED == hog_info.phase_type() ? num_bins / 360.0f : num_bins / 180.0f); phase_scale *= (PhaseType::SIGNED == hog_info.phase_type() ? 360.0f / 255.0f : 1.0f); int row_idx = 0; - for(size_t yc = 0; yc < cell_height; ++yc) + for(size_t yc = 0; yc < cell_size.height; ++yc) { - for(size_t xc = 0; xc < cell_height; xc++) + for(size_t xc = 0; xc < cell_size.width; xc++) { const float mag_value = mag[(row_idx + xc)]; const float phase_value = phase[(row_idx + xc)] * phase_scale + 0.5f; @@ -65,7 +64,7 @@ void hog_orientation_compute(const SimpleTensor<T> &mag, const SimpleTensor<T> & bins[(hidx + 1) % num_bins] += mag_value * w1; } - row_idx += cell_width; + row_idx += cell_size.width; } } @@ -117,31 +116,33 @@ void hog_block_normalization_compute(SimpleTensor<T> &block, SimpleTensor<T> &de template <typename T, typename U, typename V> void hog_orientation_binning(const SimpleTensor<T> &mag, const SimpleTensor<U> &phase, SimpleTensor<V> &hog_space, const HOGInfo &hog_info) { - const size_t cell_width = hog_info.cell_size().width; - const size_t cell_height = hog_info.cell_size().height; + const Size2D &cell_size = hog_info.cell_size(); + + const size_t num_bins = hog_info.num_bins(); const size_t shape_width = hog_space.shape().x() * hog_info.cell_size().width; const size_t shape_height = hog_space.shape().y() * hog_info.cell_size().height; - SimpleTensor<V> mag_cell(TensorShape(cell_width, cell_height), DataType::F32); - SimpleTensor<V> phase_cell(TensorShape(cell_width, cell_height), DataType::F32); + TensorShape cell_shape(cell_size.width, cell_size.height); + + SimpleTensor<V> mag_cell(cell_shape, DataType::F32); + SimpleTensor<V> phase_cell(cell_shape, DataType::F32); int cell_idx = 0; int y_offset = 0; - int x_offset = 0; // Traverse shape - for(auto sy = cell_height - 1; sy < shape_height; sy += cell_height) + for(auto sy = cell_size.height; sy <= shape_height; sy += cell_size.height) { - x_offset = 0; - for(auto sx = cell_width - 1; sx < shape_width; sx += cell_width) + int x_offset = 0; + for(auto sx = cell_size.width; sx <= shape_width; sx += cell_size.width) { int row_idx = 0; int elem_idx = 0; // Traverse cell - for(auto y = 0u; y < cell_height; ++y) + for(auto y = 0u; y < cell_size.height; ++y) { - for(auto x = 0u; x < cell_width; ++x) + for(auto x = 0u; x < cell_size.width; ++x) { int shape_idx = x + row_idx + x_offset + y_offset; mag_cell[elem_idx] = mag[shape_idx]; @@ -153,48 +154,46 @@ void hog_orientation_binning(const SimpleTensor<T> &mag, const SimpleTensor<U> & } // Partition magnitude values into bins based on phase values - std::vector<V> bins(hog_info.num_bins()); + std::vector<V> bins(num_bins); hog_orientation_compute(mag_cell, phase_cell, bins, hog_info); - for(size_t i = 0; i < hog_info.num_bins(); ++i) + for(size_t i = 0; i < num_bins; ++i) { - hog_space[cell_idx * hog_info.num_bins() + i] = bins[i]; + hog_space[cell_idx * num_bins + i] = bins[i]; } - x_offset += cell_width; + x_offset += cell_size.width; cell_idx++; } - y_offset += (cell_height * shape_width); + y_offset += (cell_size.height * shape_width); } } template <typename T> void hog_block_normalization(SimpleTensor<T> &desc, const SimpleTensor<T> &hog_space, const HOGInfo &hog_info) { - const Size2D cells_per_block = hog_info.num_cells_per_block(); - const Size2D cells_per_block_stride = hog_info.num_cells_per_block_stride(); - - const size_t block_width = hog_info.block_size().width; - const size_t block_height = hog_info.block_size().height; - const size_t block_stride_width = hog_info.block_stride().width; - const size_t block_stride_height = hog_info.block_stride().height; - const size_t shape_width = hog_space.shape().x() * hog_info.cell_size().width; - const size_t shape_height = hog_space.shape().y() * hog_info.cell_size().height; + const Size2D cells_per_block = hog_info.num_cells_per_block(); + const Size2D cells_per_block_stride = hog_info.num_cells_per_block_stride(); + const Size2D &block_size = hog_info.block_size(); + const Size2D &block_stride = hog_info.block_stride(); + const size_t num_bins = hog_info.num_bins(); - const size_t num_bins = hog_info.num_bins(); - const size_t num_channels = cells_per_block.area() * num_bins; + const size_t shape_width = hog_space.shape().x() * hog_info.cell_size().width; + const size_t shape_height = hog_space.shape().y() * hog_info.cell_size().height; + const size_t num_bins_per_block_x = cells_per_block.width * num_bins; - SimpleTensor<T> block(TensorShape{ 1u, 1u }, DataType::F32, num_channels); + // Tensor representing single block + SimpleTensor<T> block(TensorShape{ 1u, 1u }, DataType::F32, cells_per_block.area() * num_bins); int block_idx = 0; int block_y_offset = 0; // Traverse shape - for(auto sy = block_width - 1; sy < shape_height; sy += block_stride_height) + for(auto sy = block_size.height; sy <= shape_height; sy += block_stride.height) { int block_x_offset = 0; - for(auto sx = block_height - 1; sx < shape_width; sx += block_stride_width) + for(auto sx = block_size.width; sx <= shape_width; sx += block_stride.width) { int cell_y_offset = 0; int elem_idx = 0; @@ -202,17 +201,11 @@ void hog_block_normalization(SimpleTensor<T> &desc, const SimpleTensor<T> &hog_s // Traverse block for(auto y = 0u; y < cells_per_block.height; ++y) { - int cell_x_offset = 0; - for(auto x = 0u; x < cells_per_block.width; ++x) + for(auto x = 0u; x < num_bins_per_block_x; ++x) { - for(auto bin = 0u; bin < num_bins; ++bin) - { - int idx = bin + cell_x_offset + cell_y_offset + block_x_offset + block_y_offset; - block[elem_idx] = hog_space[idx]; - elem_idx++; - } - - cell_x_offset += num_bins; + int idx = x + cell_y_offset + block_x_offset + block_y_offset; + block[elem_idx] = hog_space[idx]; + elem_idx++; } cell_y_offset += hog_space.shape().x() * num_bins; @@ -232,9 +225,6 @@ void hog_block_normalization(SimpleTensor<T> &desc, const SimpleTensor<T> &hog_s template <typename T, typename U> SimpleTensor<T> hog_descriptor(const SimpleTensor<U> &src, BorderMode border_mode, U constant_border_value, const HOGInfo &hog_info) { - SimpleTensor<int16_t> _mag; - SimpleTensor<uint8_t> _phase; - SimpleTensor<int16_t> grad_x; SimpleTensor<int16_t> grad_y; @@ -253,12 +243,11 @@ SimpleTensor<T> hog_descriptor(const SimpleTensor<U> &src, BorderMode border_mod // Calculate derivative std::tie(grad_x, grad_y) = derivative<int16_t>(src, border_mode, constant_border_value, GradientDimension::GRAD_XY); - // Calculate magnitude and phase - _mag = magnitude(grad_x, grad_y, MagnitudeType::L2NORM); - _phase = phase(grad_x, grad_y, hog_info.phase_type()); - // For each cell create histogram based on magnitude and phase - hog_orientation_binning(_mag, _phase, hog_space, hog_info); + hog_orientation_binning(magnitude(grad_x, grad_y, MagnitudeType::L2NORM), + phase(grad_x, grad_y, hog_info.phase_type()), + hog_space, + hog_info); // Normalize histograms based on block size hog_block_normalization(desc, hog_space, hog_info); diff --git a/tests/validation/reference/NonMaximaSuppression.cpp b/tests/validation/reference/NonMaximaSuppression.cpp index eab5cecfc..34c6c07ab 100644 --- a/tests/validation/reference/NonMaximaSuppression.cpp +++ b/tests/validation/reference/NonMaximaSuppression.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -69,6 +69,7 @@ SimpleTensor<T> non_maxima_suppression(const SimpleTensor<T> &src, BorderMode bo } template SimpleTensor<float> non_maxima_suppression(const SimpleTensor<float> &src, BorderMode border_mode, float constant_border_value); +template SimpleTensor<uint8_t> non_maxima_suppression(const SimpleTensor<uint8_t> &src, BorderMode border_mode, uint8_t constant_border_value); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/PixelWiseMultiplication.cpp b/tests/validation/reference/PixelWiseMultiplication.cpp index b3647fc9c..546a886ac 100644 --- a/tests/validation/reference/PixelWiseMultiplication.cpp +++ b/tests/validation/reference/PixelWiseMultiplication.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -41,46 +41,105 @@ struct is_floating_point { }; +namespace +{ +/** Compute the result of `src1 * src2 * scale`. The result type always matches the type of @p src2. + * + * @param[in] src1 An input value. Data types supported: U8/QS8/QS16/S16/F16/F32. + * @param[in] src2 An input value. Data types supported: same as @p src1. + * @param[in] scale Scale to apply after multiplication. + * Scale must be positive and its value must be either 1/255 or 1/2^n where n is between 0 and 15. For QS8 and QS16 scale must be 1. + * @param[in] convert_policy Overflow policy. Supported overflow policies: Wrap, Saturate + * @param[in] rounding_policy Rounding policy. Supported rounding modes: to zero, to nearest even. + */ template <typename T1, typename T2> -SimpleTensor<T2> pixel_wise_multiplication(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy) +T2 mul(const T1 src1, const T2 src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy) { - SimpleTensor<T2> dst(src2.shape(), src2.data_type()); + using intermediate_type = typename common_promoted_signed_type<T1, T2, T2>::intermediate_type; - if(scale < 0) - { - ARM_COMPUTE_ERROR("Scale of pixel-wise multiplication must be non-negative"); - } + const double val = static_cast<intermediate_type>(src1) * static_cast<intermediate_type>(src2) * static_cast<double>(scale); - using intermediate_type = typename common_promoted_signed_type<T1, T2, T2>::intermediate_type; + if(is_floating_point<T2>::value) + { + const auto result = static_cast<T2>(val); - for(int i = 0; i < src1.num_elements(); ++i) + return result; + } + else { - double val = static_cast<intermediate_type>(src1[i]) * static_cast<intermediate_type>(src2[i]) * static_cast<double>(scale); - if(is_floating_point<T2>::value) + double rounded_val = 0; + switch(rounding_policy) { - dst[i] = val; + case(RoundingPolicy::TO_ZERO): + rounded_val = support::cpp11::trunc(val); + break; + case(RoundingPolicy::TO_NEAREST_UP): + rounded_val = round_half_up(val); + break; + case(RoundingPolicy::TO_NEAREST_EVEN): + rounded_val = round_half_even(val); + break; + default: + ARM_COMPUTE_ERROR("Unsupported rounding policy"); } - else + + const auto result = static_cast<T2>((convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T2>(rounded_val) : rounded_val); + + return result; + } +} + +template <size_t dim> +struct BroadcastUnroll +{ + template <typename T1, typename T2> + static void unroll(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, SimpleTensor<T2> &dst, + float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, + Coordinates &id_src1, Coordinates &id_src2, Coordinates &id_dst) + { + const bool src1_is_broadcast = (src1.shape()[dim - 1] != dst.shape()[dim - 1]); + const bool src2_is_broadcast = (src2.shape()[dim - 1] != dst.shape()[dim - 1]); + + id_src1.set(dim - 1, 0); + id_src2.set(dim - 1, 0); + id_dst.set(dim - 1, 0); + + for(size_t i = 0; i < dst.shape()[dim - 1]; ++i, ++id_dst[dim - 1]) { - double rounded_val = 0; - switch(rounding_policy) - { - case(RoundingPolicy::TO_ZERO): - rounded_val = support::cpp11::trunc(val); - break; - case(RoundingPolicy::TO_NEAREST_UP): - rounded_val = round_half_up(val); - break; - case(RoundingPolicy::TO_NEAREST_EVEN): - rounded_val = round_half_even(val); - break; - default: - ARM_COMPUTE_ERROR("Unsupported rounding policy"); - } - - dst[i] = (convert_policy == ConvertPolicy::SATURATE) ? saturate_cast<T2>(rounded_val) : static_cast<T2>(rounded_val); + BroadcastUnroll < dim - 1 >::unroll(src1, src2, dst, scale, convert_policy, rounding_policy, id_src1, id_src2, id_dst); + + id_src1[dim - 1] += !src1_is_broadcast; + id_src2[dim - 1] += !src2_is_broadcast; } } +}; + +template <> +struct BroadcastUnroll<0> +{ + template <typename T1, typename T2> + static void unroll(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, SimpleTensor<T2> &dst, + float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy, + Coordinates &id_src1, Coordinates &id_src2, Coordinates &id_dst) + { + dst[coord2index(dst.shape(), id_dst)] = mul(src1[coord2index(src1.shape(), id_src1)], src2[coord2index(src2.shape(), id_src2)], scale, convert_policy, rounding_policy); + } +}; +} // namespace + +template <typename T1, typename T2> +SimpleTensor<T2> pixel_wise_multiplication(const SimpleTensor<T1> &src1, const SimpleTensor<T2> &src2, float scale, ConvertPolicy convert_policy, RoundingPolicy rounding_policy) +{ + SimpleTensor<T2> dst(TensorShape::broadcast_shape(src1.shape(), src2.shape()), src2.data_type()); + + if(scale < 0) + { + ARM_COMPUTE_ERROR("Scale of pixel-wise multiplication must be non-negative"); + } + + Coordinates id_src1, id_src2, id_dst; + + BroadcastUnroll<Coordinates::num_max_dimensions>::unroll(src1, src2, dst, scale, convert_policy, rounding_policy, id_src1, id_src2, id_dst); return dst; } diff --git a/tests/validation/reference/PoolingLayer.cpp b/tests/validation/reference/PoolingLayer.cpp index 1a7dd4cbb..c14ab98c2 100644 --- a/tests/validation/reference/PoolingLayer.cpp +++ b/tests/validation/reference/PoolingLayer.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -37,14 +37,15 @@ namespace reference { namespace { -TensorShape calculate_output_shape(TensorShape shape, PoolingLayerInfo info) +TensorShape calculate_output_shape(TensorShape shape, const PoolingLayerInfo &info) { - TensorShape dst_shape = shape; - const int pool_size = info.is_global_pooling() ? shape.x() : info.pool_size(); + TensorShape dst_shape = shape; + const int pool_size_x = info.is_global_pooling() ? shape.x() : info.pool_size().width; + const int pool_size_y = info.is_global_pooling() ? shape.y() : info.pool_size().height; const std::pair<unsigned int, unsigned int> scaled_dims = arm_compute::scaled_dimensions(shape.x(), shape.y(), - pool_size, - pool_size, + pool_size_x, + pool_size_y, info.pad_stride_info()); dst_shape.set(0, scaled_dims.first); dst_shape.set(1, scaled_dims.second); @@ -54,16 +55,19 @@ TensorShape calculate_output_shape(TensorShape shape, PoolingLayerInfo info) } // namespace template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type> -SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, PoolingLayerInfo info) +SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, const PoolingLayerInfo &info) { ARM_COMPUTE_ERROR_ON(info.is_global_pooling() && (src.shape().x() != src.shape().y())); - const int pool_size = info.is_global_pooling() ? src.shape().x() : info.pool_size(); + const int pool_size_x = info.is_global_pooling() ? src.shape().x() : info.pool_size().width; + const int pool_size_y = info.is_global_pooling() ? src.shape().y() : info.pool_size().height; PoolingType type = info.pool_type(); int pool_stride_x = info.pad_stride_info().stride().first; int pool_stride_y = info.pad_stride_info().stride().second; - int pad_x = info.pad_stride_info().pad().first; - int pad_y = info.pad_stride_info().pad().second; + int pad_left = info.pad_stride_info().pad_left(); + int pad_top = info.pad_stride_info().pad_top(); + int pad_right = info.pad_stride_info().pad_right(); + int pad_bottom = info.pad_stride_info().pad_bottom(); bool exclude_padding = info.exclude_padding(); const auto w_src = static_cast<int>(src.shape()[0]); @@ -84,10 +88,10 @@ SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, PoolingLayerInfo info) { for(int w = 0; w < w_dst; ++w) { - int wstart = w * pool_stride_x - pad_x; - int hstart = h * pool_stride_y - pad_y; - int wend = std::min(wstart + pool_size, w_src); - int hend = std::min(hstart + pool_size, h_src); + int wstart = w * pool_stride_x - pad_left; + int hstart = h * pool_stride_y - pad_top; + int wend = std::min(wstart + pool_size_x, w_src); + int hend = std::min(hstart + pool_size_y, h_src); wstart = std::max(wstart, 0); hstart = std::max(hstart, 0); @@ -118,10 +122,10 @@ SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, PoolingLayerInfo info) for(int w = 0; w < w_dst; ++w) { T avg_val(0); - int wstart = w * pool_stride_x - pad_x; - int hstart = h * pool_stride_y - pad_y; - int wend = std::min(wstart + pool_size, w_src + pad_x); - int hend = std::min(hstart + pool_size, h_src + pad_y); + int wstart = w * pool_stride_x - pad_left; + int hstart = h * pool_stride_y - pad_top; + int wend = std::min(wstart + pool_size_x, w_src + pad_right); + int hend = std::min(hstart + pool_size_y, h_src + pad_bottom); int pool = (hend - hstart) * (wend - wstart); wstart = std::max(wstart, 0); hstart = std::max(hstart, 0); @@ -165,16 +169,19 @@ SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, PoolingLayerInfo info) } template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type> -SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, PoolingLayerInfo info) +SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, const PoolingLayerInfo &info) { ARM_COMPUTE_ERROR_ON(info.is_global_pooling() && (src.shape().x() != src.shape().y())); - const int pool_size = info.is_global_pooling() ? src.shape().x() : info.pool_size(); + const int pool_size_x = info.is_global_pooling() ? src.shape().x() : info.pool_size().width; + const int pool_size_y = info.is_global_pooling() ? src.shape().y() : info.pool_size().height; PoolingType type = info.pool_type(); int pool_stride_x = info.pad_stride_info().stride().first; int pool_stride_y = info.pad_stride_info().stride().second; - int pad_x = info.pad_stride_info().pad().first; - int pad_y = info.pad_stride_info().pad().second; + int pad_left = info.pad_stride_info().pad_left(); + int pad_top = info.pad_stride_info().pad_top(); + int pad_right = info.pad_stride_info().pad_right(); + int pad_bottom = info.pad_stride_info().pad_bottom(); bool exclude_padding = info.exclude_padding(); const auto w_src = static_cast<int>(src.shape()[0]); @@ -195,10 +202,10 @@ SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, PoolingLayerInfo info) { for(int w = 0; w < w_dst; ++w) { - int wstart = w * pool_stride_x - pad_x; - int hstart = h * pool_stride_y - pad_y; - int wend = std::min(wstart + pool_size, w_src); - int hend = std::min(hstart + pool_size, h_src); + int wstart = w * pool_stride_x - pad_left; + int hstart = h * pool_stride_y - pad_top; + int wend = std::min(wstart + pool_size_x, w_src); + int hend = std::min(hstart + pool_size_y, h_src); wstart = std::max(wstart, 0); hstart = std::max(hstart, 0); @@ -228,10 +235,10 @@ SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, PoolingLayerInfo info) { for(int w = 0; w < w_dst; ++w) { - int wstart = w * pool_stride_x - pad_x; - int hstart = h * pool_stride_y - pad_y; - int wend = std::min(wstart + pool_size, w_src + pad_x); - int hend = std::min(hstart + pool_size, h_src + pad_y); + int wstart = w * pool_stride_x - pad_left; + int hstart = h * pool_stride_y - pad_top; + int wend = std::min(wstart + pool_size_x, w_src + pad_right); + int hend = std::min(hstart + pool_size_y, h_src + pad_bottom); int pool = (hend - hstart) * (wend - wstart); wstart = std::max(wstart, 0); hstart = std::max(hstart, 0); @@ -284,7 +291,7 @@ SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, PoolingLayerInfo info) } template <> -SimpleTensor<uint8_t> pooling_layer<uint8_t>(const SimpleTensor<uint8_t> &src, PoolingLayerInfo info) +SimpleTensor<uint8_t> pooling_layer<uint8_t>(const SimpleTensor<uint8_t> &src, const PoolingLayerInfo &info) { SimpleTensor<float> src_tmp = convert_from_asymmetric(src); SimpleTensor<float> dst_tmp = pooling_layer<float>(src_tmp, info); @@ -292,10 +299,10 @@ SimpleTensor<uint8_t> pooling_layer<uint8_t>(const SimpleTensor<uint8_t> &src, P return dst; } -template SimpleTensor<float> pooling_layer(const SimpleTensor<float> &src, PoolingLayerInfo info); -template SimpleTensor<half> pooling_layer(const SimpleTensor<half> &src, PoolingLayerInfo info); -template SimpleTensor<qint8_t> pooling_layer(const SimpleTensor<qint8_t> &src, PoolingLayerInfo info); -template SimpleTensor<qint16_t> pooling_layer(const SimpleTensor<qint16_t> &src, PoolingLayerInfo info); +template SimpleTensor<float> pooling_layer(const SimpleTensor<float> &src, const PoolingLayerInfo &info); +template SimpleTensor<half> pooling_layer(const SimpleTensor<half> &src, const PoolingLayerInfo &info); +template SimpleTensor<qint8_t> pooling_layer(const SimpleTensor<qint8_t> &src, const PoolingLayerInfo &info); +template SimpleTensor<qint16_t> pooling_layer(const SimpleTensor<qint16_t> &src, const PoolingLayerInfo &info); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/PoolingLayer.h b/tests/validation/reference/PoolingLayer.h index 334054a0e..b0d30af32 100644 --- a/tests/validation/reference/PoolingLayer.h +++ b/tests/validation/reference/PoolingLayer.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -36,10 +36,10 @@ namespace validation namespace reference { template <typename T, typename std::enable_if<is_floating_point<T>::value, int>::type = 0> -SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, PoolingLayerInfo info); +SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, const PoolingLayerInfo &info); template <typename T, typename std::enable_if<std::is_integral<T>::value, int>::type = 0> -SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, PoolingLayerInfo info); +SimpleTensor<T> pooling_layer(const SimpleTensor<T> &src, const PoolingLayerInfo &info); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/Remap.cpp b/tests/validation/reference/Remap.cpp index bef5962fb..f862c1370 100644 --- a/tests/validation/reference/Remap.cpp +++ b/tests/validation/reference/Remap.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -42,16 +42,15 @@ SimpleTensor<T> remap(const SimpleTensor<T> &in, SimpleTensor<float> &map_x, Sim T constant_border_value) { ARM_COMPUTE_ERROR_ON_MSG(border_mode == BorderMode::REPLICATE, "BorderMode not supported"); - SimpleTensor<T> out(in.shape(), in.data_type()); - + ARM_COMPUTE_ERROR_ON(out.num_elements() != map_x.num_elements()); const int width = in.shape().x(); const int height = in.shape().y(); - for(int idx = 0; idx < out.num_elements(); idx++) { - valid_mask[idx] = 1; - Coordinates src_idx; + const Coordinates id_out = index2coord(out.shape(), idx); + valid_mask[idx] = 1; + Coordinates src_idx = id_out; // need to setup all coordinates and not just xy src_idx.set(0, static_cast<int>(std::floor(map_x[idx]))); src_idx.set(1, static_cast<int>(std::floor(map_y[idx]))); if((0 <= map_y[idx]) && (map_y[idx] < height) && (0 <= map_x[idx]) && (map_x[idx] < width)) @@ -59,11 +58,17 @@ SimpleTensor<T> remap(const SimpleTensor<T> &in, SimpleTensor<float> &map_x, Sim switch(policy) { case InterpolationPolicy::NEAREST_NEIGHBOR: + { out[idx] = tensor_elem_at(in, src_idx, border_mode, constant_border_value); break; + } case InterpolationPolicy::BILINEAR: - (valid_bilinear_policy(map_x[idx], map_y[idx], width, height, border_mode)) ? out[idx] = bilinear_policy(in, src_idx, map_x[idx], map_y[idx], border_mode, constant_border_value) : valid_mask[idx] = 0; + { + (valid_bilinear_policy(map_x[idx], map_y[idx], width, height, border_mode)) ? + out[idx] = bilinear_policy(in, src_idx, map_x[idx], map_y[idx], border_mode, constant_border_value) : + valid_mask[idx] = 0; break; + } case InterpolationPolicy::AREA: default: ARM_COMPUTE_ERROR("Interpolation not supported"); |