From ae54e026c86aec7d6819ee3ef76372c1a3c92467 Mon Sep 17 00:00:00 2001 From: Georgios Pinitas Date: Mon, 16 Jul 2018 15:41:27 +0100 Subject: COMPMID-1364: Add support for NHWC in NEDepthConcatenateLayer Change-Id: I4f8e46d1c79afa9284f2c6dc00383c453a8e7bd5 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/140165 Reviewed-by: Giorgio Arena Reviewed-by: Pablo Tello Tested-by: Jenkins --- src/runtime/NEON/functions/NEConcatenateLayer.cpp | 90 ++++++++++++++++++++ .../NEON/functions/NEDepthConcatenateLayer.cpp | 27 +++++- .../NEON/functions/NEWidthConcatenateLayer.cpp | 96 ++++++++++++++++++++++ 3 files changed, 210 insertions(+), 3 deletions(-) create mode 100644 src/runtime/NEON/functions/NEConcatenateLayer.cpp create mode 100644 src/runtime/NEON/functions/NEWidthConcatenateLayer.cpp (limited to 'src/runtime/NEON') diff --git a/src/runtime/NEON/functions/NEConcatenateLayer.cpp b/src/runtime/NEON/functions/NEConcatenateLayer.cpp new file mode 100644 index 0000000000..21ab47d3fe --- /dev/null +++ b/src/runtime/NEON/functions/NEConcatenateLayer.cpp @@ -0,0 +1,90 @@ +/* + * 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/runtime/NEON/functions/NEConcatenateLayer.h" + +#include "arm_compute/runtime/NEON/functions/NEDepthConcatenateLayer.h" +#include "arm_compute/runtime/NEON/functions/NEWidthConcatenateLayer.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "support/ToolchainSupport.h" + +namespace arm_compute +{ +NEConcatenateLayer::NEConcatenateLayer() + : _concat_function(nullptr) +{ +} + +void NEConcatenateLayer::configure(const std::vector &inputs_vector, ITensor *output, DataLayoutDimension axis) +{ + ARM_COMPUTE_ERROR_ON(output == nullptr); + + switch(get_data_layout_dimension_index(output->info()->data_layout(), axis)) + { + case 0: + { + auto func = support::cpp14::make_unique(); + func->configure(inputs_vector, output); + _concat_function = std::move(func); + break; + } + case 2: + { + auto func = support::cpp14::make_unique(); + func->configure(inputs_vector, output); + _concat_function = std::move(func); + break; + } + default: + ARM_COMPUTE_ERROR("Concatenation is supported across width and depth only!"); + } +} + +Status NEConcatenateLayer::validate(const std::vector &inputs_vector, const ITensorInfo *output, DataLayoutDimension axis) +{ + ARM_COMPUTE_RETURN_ERROR_ON(output == nullptr); + + switch(get_data_layout_dimension_index(output->data_layout(), axis)) + { + case 0: + ARM_COMPUTE_RETURN_ON_ERROR(NEWidthConcatenateLayer::validate(inputs_vector, output)); + break; + case 2: + ARM_COMPUTE_RETURN_ON_ERROR(NEDepthConcatenateLayer::validate(inputs_vector, output)); + break; + default: + ARM_COMPUTE_RETURN_ERROR_MSG("Concatenation is supported across width and depth only!"); + } + return Status{}; +} + +void NEConcatenateLayer::run() +{ + ARM_COMPUTE_ERROR_ON(_concat_function == nullptr); + _concat_function->run(); +} +} // namespace arm_compute diff --git a/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp b/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp index cb0157664b..49db855f21 100644 --- a/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp +++ b/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp @@ -27,6 +27,7 @@ #include "arm_compute/core/Helpers.h" #include "arm_compute/core/ITensor.h" #include "arm_compute/core/PixelValue.h" +#include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/NEON/NEScheduler.h" @@ -42,10 +43,8 @@ NEDepthConcatenateLayer::NEDepthConcatenateLayer() // NOLINT { } -void NEDepthConcatenateLayer::configure(std::vector inputs_vector, ITensor *output) // NOLINT +void NEDepthConcatenateLayer::configure(const std::vector &inputs_vector, ITensor *output) // NOLINT { - ARM_COMPUTE_ERROR_ON(inputs_vector.size() < 2); - _num_inputs = inputs_vector.size(); _concat_kernels_vector = arm_compute::support::cpp14::make_unique(_num_inputs); _border_handlers_vector = arm_compute::support::cpp14::make_unique(_num_inputs); @@ -59,6 +58,7 @@ void NEDepthConcatenateLayer::configure(std::vector inputs_vector, IT // Output auto inizialitation if not yet initialized auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type()); + ARM_COMPUTE_ERROR_THROW_ON(NEDepthConcatenateLayer::validate(inputs_vector_info, output->info())); unsigned int depth_offset = 0; for(unsigned int i = 0; i < _num_inputs; ++i) @@ -73,6 +73,27 @@ void NEDepthConcatenateLayer::configure(std::vector inputs_vector, IT output->info()->set_valid_region(ValidRegion(Coordinates(), output_shape)); } +Status NEDepthConcatenateLayer::validate(const std::vector &inputs_vector, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); + ARM_COMPUTE_RETURN_ERROR_ON(inputs_vector.size() < 2); + + // Output auto inizialitation if not yet initialized + TensorInfo tmp_output_info = *output->clone(); + TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_depth_concatenate_shape(inputs_vector); + auto_init_if_empty(tmp_output_info, output_shape, 1, inputs_vector[0]->data_type()); + + unsigned int depth_offset = 0; + for(const auto &input : inputs_vector) + { + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); + ARM_COMPUTE_RETURN_ON_ERROR(NEDepthConcatenateLayerKernel::validate(input, depth_offset, &tmp_output_info)); + depth_offset += input->dimension(2); + } + + return Status{}; +} + void NEDepthConcatenateLayer::run() { for(unsigned i = 0; i < _num_inputs; ++i) diff --git a/src/runtime/NEON/functions/NEWidthConcatenateLayer.cpp b/src/runtime/NEON/functions/NEWidthConcatenateLayer.cpp new file mode 100644 index 0000000000..097605c062 --- /dev/null +++ b/src/runtime/NEON/functions/NEWidthConcatenateLayer.cpp @@ -0,0 +1,96 @@ +/* + * 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/runtime/NEON/functions/NEWidthConcatenateLayer.h" + +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/runtime/NEON/NEScheduler.h" +#include "arm_compute/runtime/Tensor.h" +#include "support/ToolchainSupport.h" + +using namespace arm_compute; + +NEWidthConcatenateLayer::NEWidthConcatenateLayer() + : _concat_kernels_vector(), + _num_inputs(0) +{ +} + +Status NEWidthConcatenateLayer::validate(const std::vector &inputs_vector, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(output); + ARM_COMPUTE_RETURN_ERROR_ON(inputs_vector.size() < 2); + + // Output auto inizialitation if not yet initialized + TensorInfo tmp_output_info = *output->clone(); + TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(inputs_vector); + auto_init_if_empty(tmp_output_info, output_shape, 1, inputs_vector[0]->data_type()); + + unsigned int width_offset = 0; + for(const auto &input : inputs_vector) + { + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input); + ARM_COMPUTE_RETURN_ON_ERROR(NEWidthConcatenateLayerKernel::validate(input, width_offset, &tmp_output_info)); + width_offset += input->dimension(0); + } + + return Status{}; +} + +void NEWidthConcatenateLayer::configure(std::vector inputs_vector, ITensor *output) +{ + _num_inputs = inputs_vector.size(); + + std::vector inputs_vector_info; + for(unsigned int i = 0; i < _num_inputs; i++) + { + inputs_vector_info.emplace_back(inputs_vector.at(i)->info()); + } + TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_width_concatenate_shape(inputs_vector); + + // Output auto inizialitation if not yet initialized + auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type()); + ARM_COMPUTE_ERROR_THROW_ON(NEWidthConcatenateLayer::validate(inputs_vector_info, output->info())); + + unsigned int width_offset = 0; + + _concat_kernels_vector = arm_compute::support::cpp14::make_unique(_num_inputs); + + for(unsigned int i = 0; i < _num_inputs; i++) + { + _concat_kernels_vector[i].configure(inputs_vector.at(i), width_offset, output); + width_offset += inputs_vector.at(i)->info()->dimension(0); + } +} + +void NEWidthConcatenateLayer::run() +{ + for(unsigned i = 0; i < _num_inputs; i++) + { + NEScheduler::get().schedule(_concat_kernels_vector.get() + i, Window::DimY); + } +} -- cgit v1.2.1