diff options
author | Georgios Pinitas <georgios.pinitas@arm.com> | 2018-07-16 14:40:09 +0100 |
---|---|---|
committer | Anthony Barbier <anthony.barbier@arm.com> | 2018-11-02 16:54:54 +0000 |
commit | e29acf14f5c3f2d2c20799a1ea3e4aad50dff834 (patch) | |
tree | ca7a01ed169605404b2b954fc0d174ee00e121f2 | |
parent | 76259ca5a74a60ef2b80f29431ab53bbac9fdb63 (diff) | |
download | ComputeLibrary-e29acf14f5c3f2d2c20799a1ea3e4aad50dff834.tar.gz |
COMPMID-1365: Add support for NHWC in CLDepthConcatenateLayer
Change-Id: I3ed55bdb95d888aff0b0b76fb841bf1669659308
Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/139963
Tested-by: Jenkins <bsgcomp@arm.com>
Reviewed-by: Anthony Barbier <anthony.barbier@arm.com>
-rw-r--r-- | arm_compute/core/CL/kernels/CLDepthConcatenateLayerKernel.h | 11 | ||||
-rw-r--r-- | arm_compute/core/CL/kernels/CLWidthConcatenateLayerKernel.h | 2 | ||||
-rw-r--r-- | arm_compute/core/Utils.h | 31 | ||||
-rw-r--r-- | arm_compute/core/utils/misc/ShapeCalculator.h | 25 | ||||
-rw-r--r-- | arm_compute/runtime/CL/CLFunctions.h | 1 | ||||
-rw-r--r-- | arm_compute/runtime/CL/functions/CLConcatenateLayer.h | 81 | ||||
-rw-r--r-- | arm_compute/runtime/CL/functions/CLDepthConcatenateLayer.h | 26 | ||||
-rw-r--r-- | arm_compute/runtime/CL/functions/CLWidthConcatenateLayer.h | 14 | ||||
-rw-r--r-- | src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp | 105 | ||||
-rw-r--r-- | src/runtime/CL/functions/CLConcatenateLayer.cpp | 90 | ||||
-rw-r--r-- | src/runtime/CL/functions/CLDepthConcatenateLayer.cpp | 35 | ||||
-rw-r--r-- | src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp | 8 | ||||
-rw-r--r-- | tests/benchmark/fixtures/DepthConcatenateLayerFixture.h | 3 | ||||
-rw-r--r-- | tests/validation/CL/DepthConcatenateLayer.cpp | 38 | ||||
-rw-r--r-- | tests/validation/fixtures/DepthConcatenateLayerFixture.h | 3 |
15 files changed, 390 insertions, 83 deletions
diff --git a/arm_compute/core/CL/kernels/CLDepthConcatenateLayerKernel.h b/arm_compute/core/CL/kernels/CLDepthConcatenateLayerKernel.h index cbcab8f554..ff8009085f 100644 --- a/arm_compute/core/CL/kernels/CLDepthConcatenateLayerKernel.h +++ b/arm_compute/core/CL/kernels/CLDepthConcatenateLayerKernel.h @@ -52,7 +52,7 @@ public: ~CLDepthConcatenateLayerKernel() = default; /** Initialise the kernel's inputs and output * - * @param[in] input Input tensor. Data types supported: F16/F32. + * @param[in] input Input tensor. Data types supported: QASYMM8/F16/F32. * @param[in] depth_offset The offset on the Z axis. * @param[in,out] output Output tensor. Data types supported: Same as @p input. * @@ -61,6 +61,15 @@ public: * */ void configure(const ICLTensor *input, unsigned int depth_offset, ICLTensor *output); + /** Static function to check if given info will lead to a valid configuration of @ref CLDepthConcatenateLayerKernel + * + * @param[in] input Input tensor info. Data types supported: QASYMM8/F16/F32 + * @param[in] depth_offset The offset on the Z axis. + * @param[in] output Output tensor info. Data types supported: Same as @p input. + * + * @return a status + */ + static Status validate(const ITensorInfo *input, unsigned int depth_offset, const ITensorInfo *output); // Inherited methods overridden: void run(const Window &window, cl::CommandQueue &queue) override; diff --git a/arm_compute/core/CL/kernels/CLWidthConcatenateLayerKernel.h b/arm_compute/core/CL/kernels/CLWidthConcatenateLayerKernel.h index d206eb0da7..7ecd9276aa 100644 --- a/arm_compute/core/CL/kernels/CLWidthConcatenateLayerKernel.h +++ b/arm_compute/core/CL/kernels/CLWidthConcatenateLayerKernel.h @@ -58,7 +58,7 @@ public: * */ void configure(const ICLTensor *input, unsigned int width_offset, ICLTensor *output); - /** Static function to check if given info will lead to a valid configuration of @ref CLDepthConcatenateLayerKernel + /** Static function to check if given info will lead to a valid configuration of @ref CLWidthConcatenateLayerKernel * * @param[in] input Input tensor info. Data types supported: U8/S8/QASYMM8/U16/S16/F16/U32/S32/F32 * @param[in] width_offset The offset on the X axis. diff --git a/arm_compute/core/Utils.h b/arm_compute/core/Utils.h index 729a46fe3f..1cdfd389db 100644 --- a/arm_compute/core/Utils.h +++ b/arm_compute/core/Utils.h @@ -630,37 +630,6 @@ inline uint32_t calculate_matrix_scale(const int16_t *matrix, unsigned int matri return std::max(1, std::abs(std::accumulate(matrix, matrix + size, 0))); } -/** Calculate the output shapes of the depth concatenate function. - * - * @param[in] inputs_vector The vector that stores all the pointers to input. - * - * @return the output shape - */ -template <typename T> -TensorShape calculate_depth_concatenate_shape(const std::vector<T *> &inputs_vector) -{ - TensorShape out_shape = inputs_vector[0]->info()->tensor_shape(); - - size_t max_x = 0; - size_t max_y = 0; - size_t depth = 0; - - for(const auto &tensor : inputs_vector) - { - ARM_COMPUTE_ERROR_ON(tensor == nullptr); - const TensorShape shape = tensor->info()->tensor_shape(); - max_x = std::max(shape.x(), max_x); - max_y = std::max(shape.y(), max_y); - depth += shape.z(); - } - - out_shape.set(0, max_x); - out_shape.set(1, max_y); - out_shape.set(2, depth); - - return out_shape; -} - /** Adjust tensor shape size if width or height are odd for a given multi-planar format. No modification is done for other formats. * * @note Adding here a few links discussing the issue of odd size and sharing the same solution: diff --git a/arm_compute/core/utils/misc/ShapeCalculator.h b/arm_compute/core/utils/misc/ShapeCalculator.h index 9bf6b046b4..e5516ba154 100644 --- a/arm_compute/core/utils/misc/ShapeCalculator.h +++ b/arm_compute/core/utils/misc/ShapeCalculator.h @@ -414,6 +414,31 @@ inline TensorShape get_shape_from_info(ITensorInfo *info) } template <typename T> +inline TensorShape calculate_depth_concatenate_shape(const std::vector<T *> &inputs_vector) +{ + TensorShape out_shape = get_shape_from_info(inputs_vector[0]); + + size_t max_x = 0; + size_t max_y = 0; + size_t depth = 0; + + for(const auto &tensor : inputs_vector) + { + ARM_COMPUTE_ERROR_ON(tensor == nullptr); + const TensorShape shape = get_shape_from_info(tensor); + max_x = std::max(shape.x(), max_x); + max_y = std::max(shape.y(), max_y); + depth += shape.z(); + } + + out_shape.set(0, max_x); + out_shape.set(1, max_y); + out_shape.set(2, depth); + + return out_shape; +} + +template <typename T> inline TensorShape calculate_width_concatenate_shape(const std::vector<T *> &inputs_vector) { TensorShape out_shape = get_shape_from_info(inputs_vector[0]); diff --git a/arm_compute/runtime/CL/CLFunctions.h b/arm_compute/runtime/CL/CLFunctions.h index 0b69c96673..5e42715c2f 100644 --- a/arm_compute/runtime/CL/CLFunctions.h +++ b/arm_compute/runtime/CL/CLFunctions.h @@ -42,6 +42,7 @@ #include "arm_compute/runtime/CL/functions/CLChannelExtract.h" #include "arm_compute/runtime/CL/functions/CLChannelShuffleLayer.h" #include "arm_compute/runtime/CL/functions/CLColorConvert.h" +#include "arm_compute/runtime/CL/functions/CLConcatenateLayer.h" #include "arm_compute/runtime/CL/functions/CLConvertFullyConnectedWeights.h" #include "arm_compute/runtime/CL/functions/CLConvolution.h" #include "arm_compute/runtime/CL/functions/CLConvolutionLayer.h" diff --git a/arm_compute/runtime/CL/functions/CLConcatenateLayer.h b/arm_compute/runtime/CL/functions/CLConcatenateLayer.h new file mode 100644 index 0000000000..018c58942f --- /dev/null +++ b/arm_compute/runtime/CL/functions/CLConcatenateLayer.h @@ -0,0 +1,81 @@ +/* + * 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_CLCONCATENATELAYER_H__ +#define __ARM_COMPUTE_CLCONCATENATELAYER_H__ + +#include "arm_compute/runtime/IFunction.h" + +#include "arm_compute/core/Types.h" + +#include <memory> +#include <vector> + +namespace arm_compute +{ +// Forward declarations +class ICLTensor; +class ITensorInfo; +class Status; + +/** Basic function to execute concatenate tensors along a given axis. This function calls the following kernels: + * + * -# @ref CLWidthConcatenateLayer (if underlying concatenation axis is 0). + * -# @ref CLDepthConcatenateLayer (if underlying concatenation axis is 2). + */ +class CLConcatenateLayer : public IFunction +{ +public: + /** Default constructor */ + CLConcatenateLayer(); + /** Initialise the kernel's inputs vector and output. + * + * @note Input and output tensor dimensions preconditions defer depending on the concatenation axis. + * @note Preconditions can be found respectively at @ref CLWidthConcatenateLayer and @ref CLDepthConcatenateLayer. + * + * @param[in,out] inputs_vector The vectors containing all the tensors to concatenate. Data types supported: QASYMM8/F16/F32. + * @param[out] output Output tensor. Data types supported: Same as @p input. + * @param[in] axis Concatenation axis. Supported underlying concatenation axis are 0 and 2. + */ + void configure(std::vector<ICLTensor *> inputs_vector, ICLTensor *output, DataLayoutDimension axis); + /** Static function to check if given info will lead to a valid configuration of @ref CLConcatenateLayer + * + * @note Input and output tensor dimensions preconditions defer depending on the concatenation axis. + * @note Preconditions can be found respectively at @ref CLWidthConcatenateLayer and @ref CLDepthConcatenateLayer. + * + * @param[in] inputs_vector The vectors containing all the tensors info to concatenate. Data types supported: QASYMM8/F16/F32. + * @param[in] output Output tensor info. Data types supported: Same as @p input. + * @param[in] axis Concatenation axis. Supported underlying concatenation axis are 0 and 2. + * + * @return a status + */ + static Status validate(const std::vector<ITensorInfo *> &inputs_vector, const ITensorInfo *output, DataLayoutDimension axis); + + // Inherited methods overridden: + void run() override; + +private: + std::unique_ptr<IFunction> _concat_function; +}; +} +#endif /* __ARM_COMPUTE_CLCONCATENATELAYER_H__ */ diff --git a/arm_compute/runtime/CL/functions/CLDepthConcatenateLayer.h b/arm_compute/runtime/CL/functions/CLDepthConcatenateLayer.h index d505814e73..bafce1c66f 100644 --- a/arm_compute/runtime/CL/functions/CLDepthConcatenateLayer.h +++ b/arm_compute/runtime/CL/functions/CLDepthConcatenateLayer.h @@ -52,10 +52,34 @@ public: CLDepthConcatenateLayer(); /** Initialise the kernel's inputs vector and output. * - * @param[in,out] inputs_vector The vectors containing all the tensors to concatenate. Data types supported: F16/F32. + * @param[in,out] inputs_vector The vectors containing all the tensors to concatenate. Data types supported: QASYMM8/F16/F32. + * Input dimensions might differ for each input for the first three dimensions (width, height, depth) + * and must match for the rest. + * Note that the difference between the minimum and maximum width and height among the input tensors + * must be divisible by 2 otherwise it is not clear how padding should be added on the inputs' width and + * height when they are less than the maximum input sizes. * @param[out] output Output tensor. Data types supported: Same as @p input. + * Output tensor dimensions match the inputs' ones from the fourth dimension and above, + * while width and height are the maximum width and height of the input tensors. + * Finally, depth is the sum of the input depths. */ void configure(std::vector<ICLTensor *> inputs_vector, ICLTensor *output); + /** Static function to check if given info will lead to a valid configuration of @ref CLDepthConcatenateLayer + * + * @param[in] inputs_vector The vectors containing all the tensors to concatenate. Data types supported: QASYMM8/F16/F32. + * Input dimensions might differ for each input for the first three dimensions (width, height, depth) + * and must match for the rest. + * Note that the difference between the minimum and maximum width and height among the input tensors + * must be divisible by 2 otherwise it is not clear how padding should be added on the inputs' width and + * height when they are less than the maximum input sizes. + * @param[in] output Output tensor. Data types supported: Same as @p input. + * Output tensor dimensions match the inputs' ones from the fourth dimension and above, + * while width and height are the maximum width and height of the input tensors. + * Finally, depth is the sum of the input depths. + * + * @return a status + */ + static Status validate(const std::vector<ITensorInfo *> &inputs_vector, const ITensorInfo *output); // Inherited methods overridden: void run() override; diff --git a/arm_compute/runtime/CL/functions/CLWidthConcatenateLayer.h b/arm_compute/runtime/CL/functions/CLWidthConcatenateLayer.h index 289191e030..44462b02b2 100644 --- a/arm_compute/runtime/CL/functions/CLWidthConcatenateLayer.h +++ b/arm_compute/runtime/CL/functions/CLWidthConcatenateLayer.h @@ -50,14 +50,20 @@ public: CLWidthConcatenateLayer(); /** Initialise the kernel's inputs vector and output. * - * @param[in,out] inputs_vector The vectors containing all the tensors to concatenate. Data types supported: QASYMM8/F16/F32. - * @param[out] output Output tensor. Data types supported: Same as @p input. + * @param[in] inputs_vector The vectors containing all the tensors to concatenate. Data types supported: QASYMM8/F16/F32. + * Dimensions of all the inputs should match apart for the width which can differ. + * @param[out] output Output tensor. Data types supported: Same as @p input. + * Output tensor dimensions are the same with the inputs from the second dimension and above. + * The first dimension (width) is the sum of the input tensors' widths. */ void configure(std::vector<ICLTensor *> inputs_vector, ICLTensor *output); /** Static function to check if given info will lead to a valid configuration of @ref CLDepthConcatenateLayerKernel * - * @param[in] inputs_vector The vectors containing all the tensors info to concatenate. Data types supported: QASYMM8/F16/F32. - * @param[in] output Output tensor info. Data types supported: Same as @p input. + * @param[in] inputs_vector The vectors containing all the tensors to concatenate. Data types supported: QASYMM8/F16/F32. + * Dimensions of all the inputs should match apart for the width which can differ. + * @param[in] output Output tensor. Data types supported: Same as @p input. + * Output tensor dimensions are the same with the inputs from the second dimension and above. + * The first dimension (width) is the sum of the input tensors' widths. * * @return a status */ diff --git a/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp b/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp index 72dc21197d..4055d1c7ab 100644 --- a/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp +++ b/src/core/CL/kernels/CLDepthConcatenateLayerKernel.cpp @@ -41,6 +41,53 @@ using namespace arm_compute; +namespace +{ +std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, unsigned int depth_offset, ITensorInfo *output) +{ + ARM_COMPUTE_UNUSED(depth_offset); + + // Configure kernel window + const int left_right = (output->dimension(0) - input->dimension(0)) / 2; + const int top_bottom = (output->dimension(1) - input->dimension(1)) / 2; + + const unsigned int num_elems_processed_per_iteration = 16 / input->element_size(); + const unsigned int num_elems_read_per_iteration = 16 / input->element_size(); + const unsigned int num_rows_read_per_iteration = 1; + + // The window needs to be based on input as we copy all the depths of input + Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration)); + win.set(Window::DimZ, Window::Dimension(0, input->tensor_shape().z(), 1)); + + AccessWindowRectangle input_access(input, -left_right, -top_bottom, num_elems_read_per_iteration, num_rows_read_per_iteration); + AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); + bool window_changed = update_window_and_padding(win, input_access, output_access); + output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); + + Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; + return std::make_pair(err, win); +} +Status validate_arguments(const ITensorInfo *input, unsigned int depth_offset, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::QASYMM8, DataType::F16, DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) + depth_offset > output->dimension(2)); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) > output->dimension(0)); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(1) > output->dimension(1)); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(3, input, output); + + // The gaps between the two lowest dimensions of input and output need to be divisible by 2 + // Otherwise it is not clear how the padding should be added onto the input tensor + ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(0) - input->dimension(0)) % 2); + ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(1) - input->dimension(1)) % 2); + + return Status{}; +} +} // namespace + CLDepthConcatenateLayerKernel::CLDepthConcatenateLayerKernel() : _input(nullptr), _output(nullptr), _top_bottom(0), _left_right(0), _depth_offset(0) { @@ -53,59 +100,41 @@ BorderSize CLDepthConcatenateLayerKernel::border_size() const void CLDepthConcatenateLayerKernel::configure(const ICLTensor *input, unsigned int depth_offset, ICLTensor *output) { - static std::map<int, std::pair<std::string, int>> configs_map = - { - { 1, { "uchar", 16 } }, - { 2, { "ushort", 8 } }, - { 4, { "uint", 4 } }, - { 8, { "ulong", 2 } }, - }; - - ARM_COMPUTE_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) + depth_offset > output->info()->dimension(2)); - ARM_COMPUTE_ERROR_ON(input->info()->dimension(0) > output->info()->dimension(0)); - ARM_COMPUTE_ERROR_ON(input->info()->dimension(1) > output->info()->dimension(1)); - ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(3, input, output); - ARM_COMPUTE_ERROR_ON(configs_map.find(input->info()->element_size()) == configs_map.end()); - - // The gaps between the two lowest dimensions of input and output need to be divisible by 2 - // Otherwise it is not clear how the padding should be added onto the input tensor - ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) - input->info()->dimension(0)) % 2); - ARM_COMPUTE_ERROR_ON((output->info()->dimension(1) - input->info()->dimension(1)) % 2); + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), depth_offset, output->info())); _input = input; _output = output; _depth_offset = depth_offset; + const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size(); + // Add build options - auto config = configs_map.find(static_cast<int>(input->info()->element_size())); - std::set<std::string> build_opts; - build_opts.emplace(("-DDATA_TYPE=" + config->second.first)); - build_opts.emplace(("-DVEC_SIZE=" + support::cpp11::to_string(config->second.second))); + CLBuildOptions build_opts; + build_opts.add_option("-DDATA_TYPE=" + get_underlying_cl_type_from_data_type(input->info()->data_type())); + build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); // Create kernel - _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("concatenate_depth", build_opts)); + _kernel = static_cast<cl::Kernel>(CLKernelLibrary::get().create_kernel("concatenate_depth", build_opts.options())); // Configure kernel window _left_right = (output->info()->dimension(0) - input->info()->dimension(0)) / 2; _top_bottom = (output->info()->dimension(1) - input->info()->dimension(1)) / 2; - const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size(); - const unsigned int num_elems_read_per_iteration = 16 / input->info()->element_size(); - const unsigned int num_rows_read_per_iteration = 1; - - // The window needs to be based on input as we copy all the depths of input - Window win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration)); - win.set(Window::DimZ, Window::Dimension(0, input->info()->tensor_shape().z(), 1)); + // Configure kernel window + auto win_config = validate_and_configure_window(input->info(), depth_offset, output->info()); + ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); - AccessWindowRectangle input_access(input->info(), -_left_right, -_top_bottom, num_elems_read_per_iteration, num_rows_read_per_iteration); - AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); - update_window_and_padding(win, input_access, output_access); - output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); + ICLKernel::configure(std::get<1>(win_config)); +} - ICLKernel::configure(win); +Status CLDepthConcatenateLayerKernel::validate(const arm_compute::ITensorInfo *input, + unsigned int depth_offset, + const arm_compute::ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, depth_offset, output)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), depth_offset, output->clone().get()).first); + return Status{}; } void CLDepthConcatenateLayerKernel::run(const Window &window, cl::CommandQueue &queue) diff --git a/src/runtime/CL/functions/CLConcatenateLayer.cpp b/src/runtime/CL/functions/CLConcatenateLayer.cpp new file mode 100644 index 0000000000..f4bc1ff4ac --- /dev/null +++ b/src/runtime/CL/functions/CLConcatenateLayer.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/CL/functions/CLConcatenateLayer.h" + +#include "arm_compute/runtime/CL/functions/CLDepthConcatenateLayer.h" +#include "arm_compute/runtime/CL/functions/CLWidthConcatenateLayer.h" + +#include "arm_compute/core/CL/ICLTensor.h" +#include "arm_compute/core/Error.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Types.h" +#include "support/ToolchainSupport.h" + +namespace arm_compute +{ +CLConcatenateLayer::CLConcatenateLayer() + : _concat_function(nullptr) +{ +} + +void CLConcatenateLayer::configure(std::vector<ICLTensor *> inputs_vector, ICLTensor *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<CLWidthConcatenateLayer>(); + func->configure(std::move(inputs_vector), output); + _concat_function = std::move(func); + break; + } + case 2: + { + auto func = support::cpp14::make_unique<CLDepthConcatenateLayer>(); + func->configure(std::move(inputs_vector), output); + _concat_function = std::move(func); + break; + } + default: + ARM_COMPUTE_ERROR("Concatenation is supported across width and depth only!"); + } +} + +Status CLConcatenateLayer::validate(const std::vector<ITensorInfo *> &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(CLWidthConcatenateLayer::validate(inputs_vector, output)); + break; + case 2: + ARM_COMPUTE_RETURN_ON_ERROR(CLDepthConcatenateLayer::validate(inputs_vector, output)); + break; + default: + ARM_COMPUTE_RETURN_ERROR_MSG("Concatenation is supported across width and depth only!"); + } + return Status{}; +} + +void CLConcatenateLayer::run() +{ + ARM_COMPUTE_ERROR_ON(_concat_function == nullptr); + _concat_function->run(); +} +} // namespace arm_compute diff --git a/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp b/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp index 0b26f55a29..174be94410 100644 --- a/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp +++ b/src/runtime/CL/functions/CLDepthConcatenateLayer.cpp @@ -27,7 +27,9 @@ #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.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/CL/CLScheduler.h" #include "support/ToolchainSupport.h" @@ -43,20 +45,24 @@ CLDepthConcatenateLayer::CLDepthConcatenateLayer() // NOLINT void CLDepthConcatenateLayer::configure(std::vector<ICLTensor *> inputs_vector, ICLTensor *output) // NOLINT { - ARM_COMPUTE_ERROR_ON(inputs_vector.size() < 2); - _num_inputs = inputs_vector.size(); - unsigned int depth_offset = 0; + std::vector<ITensorInfo *> inputs_vector_info; + for(unsigned int i = 0; i < _num_inputs; i++) + { + inputs_vector_info.emplace_back(inputs_vector.at(i)->info()); + } _concat_kernels_vector = arm_compute::support::cpp14::make_unique<CLDepthConcatenateLayerKernel[]>(_num_inputs); _border_handlers_vector = arm_compute::support::cpp14::make_unique<CLFillBorderKernel[]>(_num_inputs); - TensorShape output_shape = calculate_depth_concatenate_shape(inputs_vector); + TensorShape output_shape = arm_compute::misc::shape_calculator::calculate_depth_concatenate_shape(inputs_vector_info); // 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(CLDepthConcatenateLayer::validate(inputs_vector_info, output->info())); + unsigned int depth_offset = 0; for(unsigned int i = 0; i < _num_inputs; i++) { _concat_kernels_vector[i].configure(inputs_vector.at(i), depth_offset, output); @@ -69,6 +75,27 @@ void CLDepthConcatenateLayer::configure(std::vector<ICLTensor *> inputs_vector, output->info()->set_valid_region(ValidRegion(Coordinates(), output_shape)); } +Status CLDepthConcatenateLayer::validate(const std::vector<ITensorInfo *> &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(CLDepthConcatenateLayerKernel::validate(input, depth_offset, &tmp_output_info)); + depth_offset += input->dimension(2); + } + + return Status{}; +} + void CLDepthConcatenateLayer::run() { cl::CommandQueue q = CLScheduler::get().queue(); diff --git a/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp b/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp index 3d47ec2ac2..cb0157664b 100644 --- a/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp +++ b/src/runtime/NEON/functions/NEDepthConcatenateLayer.cpp @@ -28,6 +28,7 @@ #include "arm_compute/core/ITensor.h" #include "arm_compute/core/PixelValue.h" #include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "arm_compute/runtime/NEON/NEScheduler.h" #include "support/ToolchainSupport.h" @@ -49,7 +50,12 @@ void NEDepthConcatenateLayer::configure(std::vector<ITensor *> inputs_vector, IT _concat_kernels_vector = arm_compute::support::cpp14::make_unique<NEDepthConcatenateLayerKernel[]>(_num_inputs); _border_handlers_vector = arm_compute::support::cpp14::make_unique<NEFillBorderKernel[]>(_num_inputs); - TensorShape output_shape = calculate_depth_concatenate_shape(inputs_vector); + std::vector<ITensorInfo *> 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_depth_concatenate_shape(inputs_vector_info); // Output auto inizialitation if not yet initialized auto_init_if_empty(*output->info(), output_shape, 1, inputs_vector[0]->info()->data_type()); diff --git a/tests/benchmark/fixtures/DepthConcatenateLayerFixture.h b/tests/benchmark/fixtures/DepthConcatenateLayerFixture.h index 292adde49f..7f7ed8b20b 100644 --- a/tests/benchmark/fixtures/DepthConcatenateLayerFixture.h +++ b/tests/benchmark/fixtures/DepthConcatenateLayerFixture.h @@ -27,6 +27,7 @@ #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" #include "arm_compute/core/Utils.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "tests/AssetsLibrary.h" #include "tests/Globals.h" #include "tests/Utils.h" @@ -99,7 +100,7 @@ public: src_ptrs.emplace_back(&_srcs.back()); } - TensorShape dst_shape = calculate_depth_concatenate_shape(src_ptrs); + TensorShape dst_shape = misc::shape_calculator::calculate_depth_concatenate_shape(src_ptrs); _dst = create_tensor<TensorType>(dst_shape, data_type, 1); _depth_concat.configure(src_ptrs, &_dst); diff --git a/tests/validation/CL/DepthConcatenateLayer.cpp b/tests/validation/CL/DepthConcatenateLayer.cpp index f5a5c230af..a9346dce7d 100644 --- a/tests/validation/CL/DepthConcatenateLayer.cpp +++ b/tests/validation/CL/DepthConcatenateLayer.cpp @@ -42,6 +42,44 @@ namespace validation TEST_SUITE(CL) TEST_SUITE(DepthConcatenateLayer) +// *INDENT-OFF* +// clang-format off +DATA_TEST_CASE(Validate, framework::DatasetMode::ALL, zip(zip(zip( + framework::dataset::make("InputInfo1", { TensorInfo(TensorShape(23U, 27U, 5U), 1, DataType::F32), // Mismatching data type input/output + TensorInfo(TensorShape(24U, 27U, 4U), 1, DataType::F32), // Mismatching x dimension + TensorInfo(TensorShape(23U, 27U, 3U), 1, DataType::F32), // Mismatching total depth + TensorInfo(TensorShape(16U, 27U, 6U), 1, DataType::F32) + }), + framework::dataset::make("InputInfo2", { TensorInfo(TensorShape(23U, 27U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(23U, 27U, 5U), 1, DataType::F32), + TensorInfo(TensorShape(23U, 27U, 4U), 1, DataType::F32), + TensorInfo(TensorShape(16U, 27U, 6U), 1, DataType::F32) + })), + framework::dataset::make("OutputInfo", { TensorInfo(TensorShape(23U, 27U, 9U), 1, DataType::F16), + TensorInfo(TensorShape(25U, 12U, 9U), 1, DataType::F32), + TensorInfo(TensorShape(23U, 27U, 8U), 1, DataType::F32), + TensorInfo(TensorShape(16U, 27U, 12U), 1, DataType::F32) + })), + framework::dataset::make("Expected", { false, false, false, true })), + input_info1, input_info2, output_info,expected) +{ + std::vector<TensorInfo> inputs_vector_info; + inputs_vector_info.emplace_back(std::move(input_info1)); + inputs_vector_info.emplace_back(std::move(input_info2)); + + std::vector<ITensorInfo *> inputs_vector_info_raw; + for(auto &input : inputs_vector_info) + { + inputs_vector_info_raw.emplace_back(&input); + } + + bool is_valid = bool(CLDepthConcatenateLayer::validate(inputs_vector_info_raw, + &output_info.clone()->set_is_resizable(false))); + ARM_COMPUTE_EXPECT(is_valid == expected, framework::LogLevel::ERRORS); +} +// clang-format on +// *INDENT-ON* + TEST_CASE(Configuration, framework::DatasetMode::ALL) { // Create tensors diff --git a/tests/validation/fixtures/DepthConcatenateLayerFixture.h b/tests/validation/fixtures/DepthConcatenateLayerFixture.h index 76b56ad26e..5fdfacbb76 100644 --- a/tests/validation/fixtures/DepthConcatenateLayerFixture.h +++ b/tests/validation/fixtures/DepthConcatenateLayerFixture.h @@ -26,6 +26,7 @@ #include "arm_compute/core/TensorShape.h" #include "arm_compute/core/Types.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" #include "tests/AssetsLibrary.h" #include "tests/Globals.h" #include "tests/IAccessor.h" @@ -106,7 +107,7 @@ protected: src_ptrs.emplace_back(&srcs.back()); } - TensorShape dst_shape = calculate_depth_concatenate_shape(shapes); + TensorShape dst_shape = misc::shape_calculator::calculate_depth_concatenate_shape(src_ptrs); TensorType dst = create_tensor<TensorType>(dst_shape, data_type, 1); // Create and configure function |