diff options
author | Gian Marco Iodice <gianmarco.iodice@arm.com> | 2018-02-22 16:17:20 +0000 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:49:16 +0000 |
commit | 7e4b23953e885e58d655a7d9f35a1afcc38365e4 (patch) | |
tree | 4f5a3f6535aae10a36482bd4f996d3427ac77080 /tests/validation | |
parent | 66c656a1d10831d8311f7797b285faa2c30bcb3f (diff) | |
download | ComputeLibrary-7e4b23953e885e58d655a7d9f35a1afcc38365e4.tar.gz |
COMPMID-935 - Implementing Convolution with Winograd on OpenCL (part 2)
Implemented Winograd Filter Transform 3x3 on OpenCL
Change-Id: I8f2b2dd938c5c000ef7ce392a37fb7b8b4202a4e
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/122708
Reviewed-by: Georgios Pinitas <georgios.pinitas@arm.com>
Tested-by: Jenkins <bsgcomp@arm.com>
Diffstat (limited to 'tests/validation')
-rw-r--r-- | tests/validation/CL/Winograd.cpp | 85 | ||||
-rw-r--r--[-rwxr-xr-x] | tests/validation/Helpers.h | 0 | ||||
-rw-r--r-- | tests/validation/fixtures/WinogradLayerFixture.h | 84 | ||||
-rw-r--r-- | tests/validation/reference/Winograd.cpp | 105 | ||||
-rw-r--r-- | tests/validation/reference/Winograd.h | 5 |
5 files changed, 272 insertions, 7 deletions
diff --git a/tests/validation/CL/Winograd.cpp b/tests/validation/CL/Winograd.cpp index 664b3f4ef8..0b21ed2577 100644 --- a/tests/validation/CL/Winograd.cpp +++ b/tests/validation/CL/Winograd.cpp @@ -18,15 +18,20 @@ * 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 CONCLCTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ +#include "arm_compute/core/CL/kernels/CLWinogradFilterTransformKernel.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/CL/CLTensor.h" #include "arm_compute/runtime/CL/CLTensorAllocator.h" #include "arm_compute/runtime/CL/functions/CLWinogradInputTransform.h" #include "tests/CL/CLAccessor.h" +#include "tests/CL/Helper.h" +#include "tests/PaddingCalculator.h" +#include "tests/datasets/ShapeDatasets.h" +#include "tests/datasets/WinogradFilterTransformDataset.h" #include "tests/datasets/WinogradInputTransformDataset.h" #include "tests/framework/Asserts.h" #include "tests/framework/Macros.h" @@ -40,6 +45,13 @@ namespace test { namespace validation { +namespace +{ +constexpr AbsoluteTolerance<float> tolerance_f32(0.0001f); +} // namespace + +using namespace arm_compute::misc::shape_calculator; + TEST_SUITE(CL) TEST_SUITE(Winograd) @@ -125,11 +137,76 @@ FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradInputTransformFixture, framework::Dat { validate(CLAccessor(_target), _reference); } +TEST_SUITE_END() // InputTransform + +TEST_SUITE(FilterTransform) +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip( + framework::dataset::make("InputInfo",{ + TensorInfo(TensorShape(3U, 3U, 5U, 3U), 1, DataType::F16), // F16 not supported + TensorInfo(TensorShape(3U, 3U, 5U, 3U), 1, DataType::QASYMM8), // QASYMM8 not supported + TensorInfo(TensorShape(5U, 5U, 5U, 3U), 1, DataType::F32), // Kernel size not supported + TensorInfo(TensorShape(3U, 3U), 1, DataType::F32), // valid + TensorInfo(TensorShape(3U, 3U, 5U, 3U), 1, DataType::F32), // valid + TensorInfo(TensorShape(3U, 3U, 37U, 2U), 1, DataType::F32), // valid + TensorInfo(TensorShape(3U, 3U, 37U, 22U), 1, DataType::F32) // valid + }), + framework::dataset::make("OutputInfo", { + TensorInfo(TensorShape(3U, 5U, 16U), 1, DataType::F16), + TensorInfo(TensorShape(3U, 5U, 16U), 1, DataType::QASYMM8), + TensorInfo(TensorShape(3U, 5U, 16U), 1, DataType::F32), + TensorInfo(TensorShape(1U, 1U, 16U), 1, DataType::F32), + TensorInfo(TensorShape(3U, 5U, 16U), 1, DataType::F32), + TensorInfo(TensorShape(2U, 37U, 16U), 1, DataType::F32), + TensorInfo(TensorShape(22U, 37U, 16U), 1, DataType::F32) + })), + framework::dataset::make("Expected", { false, false, false, true, true, true, true })), + input_info, output_info, expected) +{ + ARM_COMPUTE_EXPECT(bool(CLWinogradFilterTransformKernel::validate(&input_info.clone()->set_is_resizable(false), &output_info.clone()->set_is_resizable(false))) == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* -TEST_SUITE_END() +using CLWinogradFilterTransform = CLSynthetizeFunctionWithZeroConstantBorder<CLWinogradFilterTransformKernel, 0>; +using CLWinogradFilterTransformFixture = WinogradFilterTransformValidationFixture<CLTensor, CLAccessor, CLWinogradFilterTransform, float>; + +DATA_TEST_CASE(Configuration, framework::DatasetMode::ALL, combine(framework::dataset::concat(datasets::SmallWinogradFilterTransformDataset(), datasets::LargeWinogradFilterTransformDataset()), + framework::dataset::make("DataType", { DataType::F32 })), + shape_a, is_nchw_format, data_type) +{ + ARM_COMPUTE_UNUSED(is_nchw_format); + + TensorShape shape_b = compute_winograd_filter_transform_shape(TensorInfo(shape_a, 1, data_type)); + + // Create tensors + CLTensor a = create_tensor<CLTensor>(shape_a, data_type); + CLTensor b = create_tensor<CLTensor>(shape_b, data_type); + + ARM_COMPUTE_EXPECT(a.info()->is_resizable(), framework::LogLevel::ERRORS); + ARM_COMPUTE_EXPECT(b.info()->is_resizable(), framework::LogLevel::ERRORS); + + // Create and configure function + CLWinogradFilterTransform winograd_filter_transform; + winograd_filter_transform.configure(&a, &b); +} + +FIXTURE_DATA_TEST_CASE(RunSmall, CLWinogradFilterTransformFixture, framework::DatasetMode::ALL, combine(datasets::SmallWinogradFilterTransformDataset(), framework::dataset::make("DataType", { DataType::F32 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} + +FIXTURE_DATA_TEST_CASE(RunLarge, CLWinogradFilterTransformFixture, framework::DatasetMode::NIGHTLY, combine(datasets::LargeWinogradFilterTransformDataset(), framework::dataset::make("DataType", { DataType::F32 }))) +{ + // Validate output + validate(CLAccessor(_target), _reference, tolerance_f32); +} +TEST_SUITE_END() // FilterTransform -TEST_SUITE_END() -TEST_SUITE_END() +TEST_SUITE_END() // Winograd +TEST_SUITE_END() // CL } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/Helpers.h b/tests/validation/Helpers.h index b192f317b4..b192f317b4 100755..100644 --- a/tests/validation/Helpers.h +++ b/tests/validation/Helpers.h diff --git a/tests/validation/fixtures/WinogradLayerFixture.h b/tests/validation/fixtures/WinogradLayerFixture.h index 95e331560d..bfe1efce3b 100644 --- a/tests/validation/fixtures/WinogradLayerFixture.h +++ b/tests/validation/fixtures/WinogradLayerFixture.h @@ -27,7 +27,6 @@ #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "arm_compute/runtime/NEON/NEScheduler.h" #include "tests/AssetsLibrary.h" #include "tests/Globals.h" #include "tests/IAccessor.h" @@ -42,8 +41,6 @@ namespace arm_compute { -class NEWinogradLayer; - namespace test { namespace validation @@ -224,6 +221,87 @@ protected: TensorType _target{}; SimpleTensor<T> _reference{}; }; + +template <typename TensorType, typename AccessorType, typename FunctionType, typename T> +class WinogradFilterTransformValidationFixture : public framework::Fixture +{ +public: + template <typename...> + void setup(TensorShape input_shape, bool is_nchw_format, DataType data_type) + { + TensorShape output_shape = compute_winograd_filter_transform_shape(TensorInfo(input_shape, 1, data_type)); + + _target = compute_target(input_shape, output_shape, is_nchw_format, data_type); + _reference = compute_reference(input_shape, output_shape, is_nchw_format, data_type); + } + +protected: + template <typename U> + void fill(U &&tensor, int i, float min, float max) + { + switch(tensor.data_type()) + { + case DataType::F32: + { + std::uniform_real_distribution<> distribution(min, max); + library->fill(tensor, distribution, i); + break; + } + default: + { + ARM_COMPUTE_ERROR("Not supported"); + library->fill_tensor_uniform(tensor, i); + break; + } + } + } + + TensorType compute_target(const TensorShape &input_shape, const TensorShape &output_shape, bool is_nchw_format, DataType data_type) + { + ARM_COMPUTE_UNUSED(is_nchw_format); + + // Create tensors + TensorType src = create_tensor<TensorType>(input_shape, data_type); + TensorType dst = create_tensor<TensorType>(output_shape, data_type); + + // Create and configure function + FunctionType filter_transform; + filter_transform.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), 0, -1.f, 1.f); + + filter_transform.run(); + + return dst; + } + + SimpleTensor<T> compute_reference(const TensorShape &input_shape, const TensorShape &output_shape, bool is_nchw_format, DataType data_type) + { + ARM_COMPUTE_ERROR_ON(!is_nchw_format); + + // Create reference + SimpleTensor<T> src{ input_shape, data_type, 1 }; + + // Fill reference + fill(src, 0, -1.f, 1.f); + + return reference::winograd_filter_transform<T>(src, output_shape); + } + + TensorType _target{}; + SimpleTensor<T> _reference{}; +}; } // namespace validation } // namespace test } // namespace arm_compute diff --git a/tests/validation/reference/Winograd.cpp b/tests/validation/reference/Winograd.cpp index 371bb6348e..3ed55fb9fc 100644 --- a/tests/validation/reference/Winograd.cpp +++ b/tests/validation/reference/Winograd.cpp @@ -26,6 +26,8 @@ #include "tests/validation/Helpers.h" #include "tests/validation/reference/Utils.h" +#include "arm_compute/core/Types.h" + namespace arm_compute { namespace test @@ -108,6 +110,87 @@ void winograd_input_transform3x3(const SimpleTensor<T> &src, SimpleTensor<T> &ds } } } + +template <typename T> +void winograd_filter_transform3x3(const SimpleTensor<T> &in, SimpleTensor<T> &out) +{ + // Simple tensor for the 3x3 input tile + SimpleTensor<T> input_tile{ TensorShape(3u, 3u), in.data_type(), 1 }; + + // Simple tensor for the transformation matrix + SimpleTensor<T> trans_matrix{ TensorShape(3u, 4u), in.data_type(), 1 }; + + // Simple tensor for the transformation matrix transpose + SimpleTensor<T> trans_matrix_transposed{ TensorShape(4u, 3u), in.data_type(), 1 }; + + // Simple tensor for the 4x3 temporary tile + SimpleTensor<T> tmp_tile{ TensorShape(3u, 4u), in.data_type(), 1 }; + + // Simple tensor for the 4x4 output tile + SimpleTensor<T> output_tile{ TensorShape(4u, 4u), in.data_type(), 1 }; + + // Initialize transformation matrix + // 1 | 0 | 0 + // 0.5 | 0.5 | 0.5 + // 0.5 |-0.5 | 0.5 + // 0 | 0 | 1 + trans_matrix[0 + 0 * 3] = 1.0f; + trans_matrix[1 + 0 * 3] = 0.0f; + trans_matrix[2 + 0 * 3] = 0.0f; + trans_matrix[0 + 1 * 3] = 0.5f; + trans_matrix[1 + 1 * 3] = 0.5f; + trans_matrix[2 + 1 * 3] = 0.5f; + trans_matrix[0 + 2 * 3] = 0.5f; + trans_matrix[1 + 2 * 3] = -0.5f; + trans_matrix[2 + 2 * 3] = 0.5f; + trans_matrix[0 + 3 * 3] = 0.0f; + trans_matrix[1 + 3 * 3] = 0.0f; + trans_matrix[2 + 3 * 3] = 1.0f; + + // Transpose the transformation matrix + transpose_matrix(trans_matrix, trans_matrix_transposed); + + const int num_channels = in.shape()[2]; + const int num_filters = in.shape()[3]; + const int num_batches = in.shape().total_size() / (9 * num_channels * num_filters); + + for(int n = 0; n < num_batches; ++n) + { + for(int w = 0; w < num_filters; ++w) + { + for(int z = 0; z < num_channels; ++z) + { + // Load the 3x3 tile from the input tensor + get_tile(in, input_tile, Coordinates(0, 0, z, w, n)); + + // First transformation + matrix_multiply(trans_matrix, input_tile, tmp_tile); + + // Second transformation + matrix_multiply(tmp_tile, trans_matrix_transposed, output_tile); + + // Store the 4x4 output tile across the 16 channels + const int output_offset = w + z * num_filters; + out[output_offset + 0 * num_filters * num_channels] = output_tile[0 + 0 * 4]; + out[output_offset + 1 * num_filters * num_channels] = output_tile[1 + 0 * 4]; + out[output_offset + 2 * num_filters * num_channels] = output_tile[2 + 0 * 4]; + out[output_offset + 3 * num_filters * num_channels] = output_tile[3 + 0 * 4]; + out[output_offset + 4 * num_filters * num_channels] = output_tile[0 + 1 * 4]; + out[output_offset + 5 * num_filters * num_channels] = output_tile[1 + 1 * 4]; + out[output_offset + 6 * num_filters * num_channels] = output_tile[2 + 1 * 4]; + out[output_offset + 7 * num_filters * num_channels] = output_tile[3 + 1 * 4]; + out[output_offset + 8 * num_filters * num_channels] = output_tile[0 + 2 * 4]; + out[output_offset + 9 * num_filters * num_channels] = output_tile[1 + 2 * 4]; + out[output_offset + 10 * num_filters * num_channels] = output_tile[2 + 2 * 4]; + out[output_offset + 11 * num_filters * num_channels] = output_tile[3 + 2 * 4]; + out[output_offset + 12 * num_filters * num_channels] = output_tile[0 + 3 * 4]; + out[output_offset + 13 * num_filters * num_channels] = output_tile[1 + 3 * 4]; + out[output_offset + 14 * num_filters * num_channels] = output_tile[2 + 3 * 4]; + out[output_offset + 15 * num_filters * num_channels] = output_tile[3 + 3 * 4]; + } + } + } +} } // namespace template <typename T> @@ -130,7 +213,29 @@ SimpleTensor<T> winograd_input_transform(const SimpleTensor<T> &src, const Tenso return dst; } +template <typename T> +SimpleTensor<T> winograd_filter_transform(const SimpleTensor<T> &in, const TensorShape &output_shape) +{ + ARM_COMPUTE_ERROR_ON_MSG(in.data_layout() != DataLayout::NCHW, "Only supported NCHW data format"); + + // Create reference + SimpleTensor<T> out{ output_shape, in.data_type(), 1 }; + + switch(in.shape()[0]) + { + case 3: + winograd_filter_transform3x3(in, out); + break; + default: + ARM_COMPUTE_ERROR("Only supported 3x3 kernel"); + break; + } + + return out; +} + template SimpleTensor<float> winograd_input_transform(const SimpleTensor<float> &src, const TensorShape &dst_shape, const PadStrideInfo &conv_info, const Size2D &kernel_dims); +template SimpleTensor<float> winograd_filter_transform(const SimpleTensor<float> &in, const TensorShape &output_shape); } // namespace reference } // namespace validation } // namespace test diff --git a/tests/validation/reference/Winograd.h b/tests/validation/reference/Winograd.h index ed95239db3..ba8e5c1cb6 100644 --- a/tests/validation/reference/Winograd.h +++ b/tests/validation/reference/Winograd.h @@ -24,6 +24,8 @@ #ifndef __ARM_COMPUTE_TEST_WINOGRAD_H__ #define __ARM_COMPUTE_TEST_WINOGRAD_H__ +#include "arm_compute/core/TensorShape.h" + #include "tests/SimpleTensor.h" namespace arm_compute @@ -36,6 +38,9 @@ namespace reference { template <typename T> SimpleTensor<T> winograd_input_transform(const SimpleTensor<T> &src, const TensorShape &dst_shape, const PadStrideInfo &conv_info, const Size2D &kernel_dims); + +template <typename T> +SimpleTensor<T> winograd_filter_transform(const SimpleTensor<T> &in, const TensorShape &output_shape); } // namespace reference } // namespace validation } // namespace test |