From 1cd41495153c4e89d6195b42f870967339c1a13b Mon Sep 17 00:00:00 2001 From: Sanghoon Lee Date: Thu, 15 Mar 2018 11:48:48 +0000 Subject: COMPMID-577: Implement CL validation for GaussianPyramid Change-Id: If879cbe15b14d97818c24d44b29fc69b6c8cb686 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/127601 Reviewed-by: Gian Marco Iodice Tested-by: Jenkins Reviewed-by: Anthony Barbier --- .../core/CL/kernels/CLGaussianPyramidKernel.h | 19 ++-- .../core/NEON/kernels/NEGaussianPyramidKernel.h | 4 +- .../runtime/CL/functions/CLGaussianPyramid.h | 5 +- src/core/CL/kernels/CLGaussianPyramidKernel.cpp | 39 +++----- src/runtime/CL/functions/CLGaussianPyramid.cpp | 29 ++++-- tests/validation/CL/GaussianPyramid.cpp | 111 +++++++++++++++++++++ tests/validation/NEON/GaussianPyramid.cpp | 12 +-- tests/validation/reference/GaussianPyramidHalf.cpp | 2 +- 8 files changed, 163 insertions(+), 58 deletions(-) create mode 100644 tests/validation/CL/GaussianPyramid.cpp diff --git a/arm_compute/core/CL/kernels/CLGaussianPyramidKernel.h b/arm_compute/core/CL/kernels/CLGaussianPyramidKernel.h index 43d66c39c9..425f8478c6 100644 --- a/arm_compute/core/CL/kernels/CLGaussianPyramidKernel.h +++ b/arm_compute/core/CL/kernels/CLGaussianPyramidKernel.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -49,19 +49,17 @@ public: /** Initialise the kernel's source, destination and border mode. * - * @param[in] input Source tensor. Data types supported: U8. - * @param[out] output Destination tensor. Output should have half the input width. Data types supported: U16. - * @param[in] border_undefined True if the border mode is undefined. False if it's replicate or constant. + * @param[in] input Source tensor. Data types supported: U8. + * @param[out] output Destination tensor. Output should have half the input width. Data types supported: U16. */ - void configure(const ICLTensor *input, ICLTensor *output, bool border_undefined); + void configure(const ICLTensor *input, ICLTensor *output); // Inherited methods overridden: void run(const Window &window, cl::CommandQueue &queue) override; BorderSize border_size() const override; private: - BorderSize _border_size; - int _l2_load_offset; + int _l2_load_offset; }; /** OpenCL kernel to perform a Gaussian filter and half scaling across height (vertical pass) */ @@ -83,11 +81,10 @@ public: /** Initialise the kernel's source, destination and border mode. * - * @param[in] input Source tensor. Data types supported: U16. - * @param[out] output Destination tensor. Output should have half the input height. Data types supported: U8. - * @param[in] border_undefined True if the border mode is undefined. False if it's replicate or constant. + * @param[in] input Source tensor. Data types supported: U16. + * @param[out] output Destination tensor. Output should have half the input height. Data types supported: U8. */ - void configure(const ICLTensor *input, ICLTensor *output, bool border_undefined); + void configure(const ICLTensor *input, ICLTensor *output); // Inherited methods overridden: void run(const Window &window, cl::CommandQueue &queue) override; diff --git a/arm_compute/core/NEON/kernels/NEGaussianPyramidKernel.h b/arm_compute/core/NEON/kernels/NEGaussianPyramidKernel.h index ac22934786..1446ca8c52 100644 --- a/arm_compute/core/NEON/kernels/NEGaussianPyramidKernel.h +++ b/arm_compute/core/NEON/kernels/NEGaussianPyramidKernel.h @@ -54,7 +54,7 @@ public: /** Initialise the kernel's source, destination and border mode. * * @param[in] input Source tensor. Data type supported: U8. - * @param[out] output Destination tensor. Data type supported: S16. + * @param[out] output Destination tensor. Output should have half the input width. Data type supported: S16. */ void configure(const ITensor *input, ITensor *output); @@ -90,7 +90,7 @@ public: /** Initialise the kernel's source, destination and border mode. * * @param[in] input Source tensor. Data type supported: S16. - * @param[out] output Destination tensor. Data type supported: U8. + * @param[out] output Destination tensor. Output should have half the input height. Data type supported: U8. */ void configure(const ITensor *input, ITensor *output); diff --git a/arm_compute/runtime/CL/functions/CLGaussianPyramid.h b/arm_compute/runtime/CL/functions/CLGaussianPyramid.h index 97935193dc..0110adfa20 100644 --- a/arm_compute/runtime/CL/functions/CLGaussianPyramid.h +++ b/arm_compute/runtime/CL/functions/CLGaussianPyramid.h @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -90,7 +90,8 @@ public: void run() override; private: - std::unique_ptr _border_handler; + std::unique_ptr _horizontal_border_handler; + std::unique_ptr _vertical_border_handler; std::unique_ptr _horizontal_reduction; std::unique_ptr _vertical_reduction; }; diff --git a/src/core/CL/kernels/CLGaussianPyramidKernel.cpp b/src/core/CL/kernels/CLGaussianPyramidKernel.cpp index 34a228c717..a4fda364e3 100644 --- a/src/core/CL/kernels/CLGaussianPyramidKernel.cpp +++ b/src/core/CL/kernels/CLGaussianPyramidKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -32,20 +32,19 @@ using namespace arm_compute; CLGaussianPyramidHorKernel::CLGaussianPyramidHorKernel() - : _border_size(0), _l2_load_offset(0) + : _l2_load_offset(0) { } BorderSize CLGaussianPyramidHorKernel::border_size() const { - return _border_size; + return BorderSize(0, 2); } -void CLGaussianPyramidHorKernel::configure(const ICLTensor *input, ICLTensor *output, bool border_undefined) +void CLGaussianPyramidHorKernel::configure(const ICLTensor *input, ICLTensor *output) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8); ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U16); - ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != 2 * output->info()->dimension(0)); ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) != output->info()->dimension(1)); for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i) @@ -53,9 +52,8 @@ void CLGaussianPyramidHorKernel::configure(const ICLTensor *input, ICLTensor *ou ARM_COMPUTE_ERROR_ON(input->info()->dimension(i) != output->info()->dimension(i)); } - _input = input; - _output = output; - _border_size = BorderSize(border_undefined ? 0 : 2, 2); + _input = input; + _output = output; // Create kernel _kernel = static_cast(CLKernelLibrary::get().create_kernel("gaussian1x5_sub_x")); @@ -64,9 +62,9 @@ void CLGaussianPyramidHorKernel::configure(const ICLTensor *input, ICLTensor *ou constexpr unsigned int num_elems_processed_per_iteration = 16; constexpr unsigned int num_elems_read_per_iteration = 20; constexpr unsigned int num_elems_written_per_iteration = 8; - constexpr float scale_x = 0.5f; + const float scale_x = static_cast(output->info()->dimension(0)) / input->info()->dimension(0); - Window win = calculate_max_window_horizontal(*input->info(), Steps(num_elems_processed_per_iteration), border_undefined, border_size()); + Window win = calculate_max_window_horizontal(*input->info(), Steps(num_elems_processed_per_iteration)); AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration, scale_x); // Sub sampling selects odd pixels (1, 3, 5, ...) for images with even @@ -95,11 +93,7 @@ void CLGaussianPyramidHorKernel::configure(const ICLTensor *input, ICLTensor *ou AccessWindowHorizontal(input->info(), _l2_load_offset, num_elems_read_per_iteration), output_access); - ValidRegion valid_region = input->info()->valid_region(); - valid_region.anchor.set(0, std::ceil((valid_region.anchor[0] + (border_undefined ? border_size().left : 0)) / 2.f)); - valid_region.shape.set(0, (valid_region.shape[0] - (border_undefined ? border_size().right : 0)) / 2 - valid_region.anchor[0]); - - output_access.set_valid_region(win, valid_region); + output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape())); ICLKernel::configure(win); } @@ -139,12 +133,11 @@ BorderSize CLGaussianPyramidVertKernel::border_size() const return BorderSize(2, 0); } -void CLGaussianPyramidVertKernel::configure(const ICLTensor *input, ICLTensor *output, bool border_undefined) +void CLGaussianPyramidVertKernel::configure(const ICLTensor *input, ICLTensor *output) { ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U16); ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::U8); ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) != output->info()->dimension(0)); - ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) != 2 * output->info()->dimension(1)); for(size_t i = 2; i < Coordinates::num_max_dimensions; ++i) { @@ -163,10 +156,10 @@ void CLGaussianPyramidVertKernel::configure(const ICLTensor *input, ICLTensor *o constexpr unsigned int num_elems_written_per_iteration = 8; constexpr unsigned int num_elems_read_per_iteration = 8; constexpr unsigned int num_rows_per_iteration = 5; - constexpr float scale_y = 0.5f; - Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration, num_rows_processed_per_iteration), - border_undefined, border_size()); + const float scale_y = static_cast(output->info()->dimension(1)) / input->info()->dimension(1); + + Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration, num_rows_processed_per_iteration)); AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_written_per_iteration, num_rows_per_iteration, 1.f, scale_y); // Determine whether we need to load even or odd rows. See above for a @@ -182,11 +175,7 @@ void CLGaussianPyramidVertKernel::configure(const ICLTensor *input, ICLTensor *o AccessWindowRectangle(input->info(), 0, _t2_load_offset, num_elems_read_per_iteration, num_rows_per_iteration), output_access); - ValidRegion valid_region = input->info()->valid_region(); - valid_region.anchor.set(1, std::ceil((valid_region.anchor[1] + (border_undefined ? border_size().top : 0)) / 2.f)); - valid_region.shape.set(1, (valid_region.shape[1] - (border_undefined ? border_size().bottom : 0)) / 2 - valid_region.anchor[1]); - - output_access.set_valid_region(win, valid_region); + output->info()->set_valid_region(ValidRegion(Coordinates(), output->info()->tensor_shape())); ICLKernel::configure(win); } diff --git a/src/runtime/CL/functions/CLGaussianPyramid.cpp b/src/runtime/CL/functions/CLGaussianPyramid.cpp index 4b32954d91..ddce5fb8c3 100644 --- a/src/runtime/CL/functions/CLGaussianPyramid.cpp +++ b/src/runtime/CL/functions/CLGaussianPyramid.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -49,7 +49,8 @@ CLGaussianPyramid::CLGaussianPyramid() } CLGaussianPyramidHalf::CLGaussianPyramidHalf() // NOLINT - : _border_handler(), + : _horizontal_border_handler(), + _vertical_border_handler(), _horizontal_reduction(), _vertical_reduction() { @@ -64,6 +65,9 @@ void CLGaussianPyramidHalf::configure(ICLTensor *input, CLPyramid *pyramid, Bord ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) != pyramid->info()->height()); ARM_COMPUTE_ERROR_ON(SCALE_PYRAMID_HALF != pyramid->info()->scale()); + // Constant value to use for vertical fill border when the border mode is CONSTANT + const uint16_t pixel_value_u16 = static_cast(constant_border_value) * 2 + static_cast(constant_border_value) * 8 + static_cast(constant_border_value) * 6; + /* Get number of pyramid levels */ const size_t num_levels = pyramid->info()->num_levels(); @@ -72,28 +76,31 @@ void CLGaussianPyramidHalf::configure(ICLTensor *input, CLPyramid *pyramid, Bord if(num_levels > 1) { - _border_handler = arm_compute::support::cpp14::make_unique(num_levels - 1); - _horizontal_reduction = arm_compute::support::cpp14::make_unique(num_levels - 1); - _vertical_reduction = arm_compute::support::cpp14::make_unique(num_levels - 1); + _horizontal_border_handler = arm_compute::support::cpp14::make_unique(num_levels - 1); + _vertical_border_handler = arm_compute::support::cpp14::make_unique(num_levels - 1); + _horizontal_reduction = arm_compute::support::cpp14::make_unique(num_levels - 1); + _vertical_reduction = arm_compute::support::cpp14::make_unique(num_levels - 1); // Apply half scale to the X dimension of the tensor shape TensorShape tensor_shape = pyramid->info()->tensor_shape(); tensor_shape.set(0, (pyramid->info()->width() + 1) * SCALE_PYRAMID_HALF); PyramidInfo pyramid_info(num_levels - 1, SCALE_PYRAMID_HALF, tensor_shape, Format::U16); - _tmp.init(pyramid_info); for(size_t i = 0; i < num_levels - 1; ++i) { /* Configure horizontal kernel */ - _horizontal_reduction[i].configure(_pyramid->get_pyramid_level(i), _tmp.get_pyramid_level(i), border_mode == BorderMode::UNDEFINED); + _horizontal_reduction[i].configure(_pyramid->get_pyramid_level(i), _tmp.get_pyramid_level(i)); /* Configure vertical kernel */ - _vertical_reduction[i].configure(_tmp.get_pyramid_level(i), _pyramid->get_pyramid_level(i + 1), border_mode == BorderMode::UNDEFINED); + _vertical_reduction[i].configure(_tmp.get_pyramid_level(i), _pyramid->get_pyramid_level(i + 1)); /* Configure border */ - _border_handler[i].configure(_pyramid->get_pyramid_level(i), _horizontal_reduction[i].border_size(), border_mode, PixelValue(constant_border_value)); + _horizontal_border_handler[i].configure(_pyramid->get_pyramid_level(i), _horizontal_reduction[i].border_size(), border_mode, PixelValue(constant_border_value)); + + /* Configure border */ + _vertical_border_handler[i].configure(_tmp.get_pyramid_level(i), _vertical_reduction[i].border_size(), border_mode, PixelValue(pixel_value_u16)); } _tmp.allocate(); } @@ -110,13 +117,15 @@ void CLGaussianPyramidHalf::run() _pyramid->get_pyramid_level(0)->map(CLScheduler::get().queue(), true /* blocking */); _input->map(CLScheduler::get().queue(), true /* blocking */); _pyramid->get_pyramid_level(0)->copy_from(*_input); + _input->unmap(CLScheduler::get().queue()); _pyramid->get_pyramid_level(0)->unmap(CLScheduler::get().queue()); for(unsigned int i = 0; i < num_levels - 1; ++i) { - CLScheduler::get().enqueue(_border_handler[i], false); + CLScheduler::get().enqueue(_horizontal_border_handler[i], false); CLScheduler::get().enqueue(_horizontal_reduction[i], false); + CLScheduler::get().enqueue(_vertical_border_handler[i], false); CLScheduler::get().enqueue(_vertical_reduction[i], false); } } diff --git a/tests/validation/CL/GaussianPyramid.cpp b/tests/validation/CL/GaussianPyramid.cpp new file mode 100644 index 0000000000..2a4596d366 --- /dev/null +++ b/tests/validation/CL/GaussianPyramid.cpp @@ -0,0 +1,111 @@ +/* + * 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/CL/CLTensor.h" +#include "arm_compute/runtime/CL/CLTensorAllocator.h" +#include "arm_compute/runtime/CL/functions/CLGaussianPyramid.h" +#include "tests/CL/CLAccessor.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/BorderModeDataset.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/GaussianPyramidHalfFixture.h" +#include "tests/validation/reference/Utils.h" + +namespace arm_compute +{ +namespace test +{ +namespace validation +{ +namespace +{ +const auto small_gaussian_pyramid_levels = combine(datasets::Medium2DShapes(), datasets::BorderModes()) * framework::dataset::make("numlevels", 2, 4); +const auto large_gaussian_pyramid_levels = combine(datasets::Large2DShapes(), datasets::BorderModes()) * framework::dataset::make("numlevels", 2, 5); + +template +inline void validate_gaussian_pyramid(const CLPyramid &target, const std::vector> &reference, BorderMode border_mode) +{ + ValidRegion prev_valid_region = shape_to_valid_region(reference[0].shape()); + + for(size_t i = 1; i < reference.size(); ++i) + { + 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(CLAccessor(*(target.get_pyramid_level(i))), reference[i], valid_region); + + // Keep the valid region for the next level + prev_valid_region = valid_region; + } +} + +} // namespace + +TEST_SUITE(CL) +TEST_SUITE(GaussianPyramid) +TEST_SUITE(Half) + +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, large_gaussian_pyramid_levels, + shape, border_mode, num_levels) +{ + CLTensor src = create_tensor(shape, DataType::U8); + + // Create pyramid + PyramidInfo pyramid_info(num_levels, SCALE_PYRAMID_HALF, shape, Format::U8); + CLPyramid dst; + dst.init(pyramid_info); + + CLGaussianPyramidHalf gaussian_pyramid_half; + gaussian_pyramid_half.configure(&src, &dst, border_mode, 0); + + ARM_COMPUTE_EXPECT(src.info()->is_resizable(), framework::LogLevel::ERRORS); + + for(size_t level = 0; level < pyramid_info.num_levels(); ++level) + { + ARM_COMPUTE_EXPECT(dst.get_pyramid_level(level)->info()->is_resizable(), framework::LogLevel::ERRORS); + } +} + +template +using CLGaussianPyramidHalfFixture = GaussianPyramidHalfValidationFixture; + +FIXTURE_DATA_TEST_CASE(RunSmallGaussianPyramidHalf, CLGaussianPyramidHalfFixture, framework::DatasetMode::ALL, small_gaussian_pyramid_levels) +{ + validate_gaussian_pyramid(_target, _reference, _border_mode); +} + +FIXTURE_DATA_TEST_CASE(RunLargeGaussianPyramidHalf, CLGaussianPyramidHalfFixture, framework::DatasetMode::NIGHTLY, large_gaussian_pyramid_levels) +{ + validate_gaussian_pyramid(_target, _reference, _border_mode); +} +TEST_SUITE_END() +TEST_SUITE_END() +TEST_SUITE_END() +} // namespace validation +} // namespace test +} // namespace arm_compute diff --git a/tests/validation/NEON/GaussianPyramid.cpp b/tests/validation/NEON/GaussianPyramid.cpp index 0dea57e75b..c646b5011f 100644 --- a/tests/validation/NEON/GaussianPyramid.cpp +++ b/tests/validation/NEON/GaussianPyramid.cpp @@ -44,13 +44,11 @@ namespace validation { namespace { -constexpr AbsoluteTolerance tolerance_fp32(1.0f); /**< Tolerance value for comparing reference's output against implementation's output */ - const auto small_gaussian_pyramid_levels = combine(datasets::Medium2DShapes(), datasets::BorderModes()) * framework::dataset::make("numlevels", 2, 4); const auto large_gaussian_pyramid_levels = combine(datasets::Large2DShapes(), datasets::BorderModes()) * framework::dataset::make("numlevels", 2, 5); -template -inline void validate_gaussian_pyramid(const Pyramid &target, const std::vector> &reference, BorderMode border_mode, U tolerance, float tolerance_number = 0.0f) +template +inline void validate_gaussian_pyramid(const Pyramid &target, const std::vector> &reference, BorderMode border_mode) { ValidRegion prev_valid_region = shape_to_valid_region(reference[0].shape()); @@ -59,7 +57,7 @@ inline void validate_gaussian_pyramid(const Pyramid &target, const std::vector, framework::DatasetMode::ALL, small_gaussian_pyramid_levels) { - validate_gaussian_pyramid(_target, _reference, _border_mode, tolerance_fp32); + validate_gaussian_pyramid(_target, _reference, _border_mode); } FIXTURE_DATA_TEST_CASE(RunLargeGaussianPyramidHalf, NEGaussianPyramidHalfFixture, framework::DatasetMode::NIGHTLY, large_gaussian_pyramid_levels) { - validate_gaussian_pyramid(_target, _reference, _border_mode, tolerance_fp32); + validate_gaussian_pyramid(_target, _reference, _border_mode); } TEST_SUITE_END() TEST_SUITE_END() diff --git a/tests/validation/reference/GaussianPyramidHalf.cpp b/tests/validation/reference/GaussianPyramidHalf.cpp index 0a68dedaee..7d5eb07512 100644 --- a/tests/validation/reference/GaussianPyramidHalf.cpp +++ b/tests/validation/reference/GaussianPyramidHalf.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * -- cgit v1.2.1