From 55b3d1216b4011d86d5f06335e518dc924987ae5 Mon Sep 17 00:00:00 2001 From: Michalis Spyrou Date: Wed, 9 May 2018 09:59:23 +0100 Subject: COMPMID-1137 OpenCL concatenate width Change-Id: I40faba421281b1cf080fa6a825d04a4366cdaeb0 Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/130700 Reviewed-by: Anthony Barbier Tested-by: Jenkins Reviewed-by: Georgios Pinitas --- src/core/CL/CLKernelLibrary.cpp | 1 + src/core/CL/cl_kernels/concatenate.cl | 37 +++++- .../CL/kernels/CLWidthConcatenateLayerKernel.cpp | 136 +++++++++++++++++++++ 3 files changed, 173 insertions(+), 1 deletion(-) create mode 100644 src/core/CL/kernels/CLWidthConcatenateLayerKernel.cpp (limited to 'src/core/CL') diff --git a/src/core/CL/CLKernelLibrary.cpp b/src/core/CL/CLKernelLibrary.cpp index 220c7490f3..bdb26f8b0f 100644 --- a/src/core/CL/CLKernelLibrary.cpp +++ b/src/core/CL/CLKernelLibrary.cpp @@ -172,6 +172,7 @@ const std::map CLKernelLibrary::_kernel_program_map = { "combine_gradients_L1", "canny.cl" }, { "combine_gradients_L2", "canny.cl" }, { "concatenate_depth", "concatenate.cl" }, + { "concatenate_width", "concatenate.cl" }, { "convolution_rectangle", "convolution_rectangle.cl" }, { "col2im", "col2im.cl" }, { "convert_depth_down", "depth_convert.cl" }, diff --git a/src/core/CL/cl_kernels/concatenate.cl b/src/core/CL/cl_kernels/concatenate.cl index a92ab5bdad..f97ae13a9a 100644 --- a/src/core/CL/cl_kernels/concatenate.cl +++ b/src/core/CL/cl_kernels/concatenate.cl @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017 ARM Limited. + * Copyright (c) 2017-2018 ARM Limited. * * SPDX-License-Identifier: MIT * @@ -23,6 +23,41 @@ */ #include "helpers.h" +/** This kernel concatenates the input tensor into the output tensor along the first dimension + * + * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8, QASYMM8, QS16, F16, F32 + * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) + * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) + * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] src_step_z src_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor + * @param[out] dst_ptr Pointer to the destination tensor. Supported data types: same as @p src_ptr + * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) + * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) + * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) + * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) + * @param[in] dst_stride_z Stride of the source tensor in Z dimension (in bytes) + * @param[in] dst_step_z dst_stride_z * number of elements along Z processed per workitem(in bytes) + * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor + * @param[in] offset The offset to the first valid element of the output tensor in bytes + */ +__kernel void concatenate_width( + TENSOR3D_DECLARATION(src), + TENSOR3D_DECLARATION(dst), + int offset) +{ + Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); + Tensor3D dst = CONVERT_TO_TENSOR3D_STRUCT(dst); + + VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) + source_values = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)src.ptr); + + VSTORE(VEC_SIZE) + (source_values, 0, (__global DATA_TYPE *)(dst.ptr + offset)); +} + /** This kernel concatenates the input tensor into the output tensor along the third dimension * * @param[in] src_ptr Pointer to the source tensor. Supported data types: QS8, QS16, F16, F32 diff --git a/src/core/CL/kernels/CLWidthConcatenateLayerKernel.cpp b/src/core/CL/kernels/CLWidthConcatenateLayerKernel.cpp new file mode 100644 index 0000000000..b8bce38cad --- /dev/null +++ b/src/core/CL/kernels/CLWidthConcatenateLayerKernel.cpp @@ -0,0 +1,136 @@ +/* + * 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/CL/kernels/CLWidthConcatenateLayerKernel.h" + +#include "arm_compute/core/CL/CLHelpers.h" +#include "arm_compute/core/CL/CLKernelLibrary.h" +#include "arm_compute/core/CL/ICLTensor.h" +#include "arm_compute/core/CL/OpenCL.h" +#include "arm_compute/core/Error.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/IAccessWindow.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/Window.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" + +#include "support/ToolchainSupport.h" + +#include + +using namespace arm_compute; +namespace +{ +std::pair validate_and_configure_window(ITensorInfo *input, unsigned int width_offset, ITensorInfo *output) +{ + const unsigned int num_elems_processed_per_iteration = 16; + + // The window needs to be based on input as we copy all the widths of input + Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); + AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); + AccessWindowHorizontal output_access(output, width_offset, num_elems_processed_per_iteration); + bool window_changed = update_window_and_padding(win, input_access, output_access); + + 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 width_offset, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::U8, DataType::S8, DataType::QS8, DataType::QASYMM8, DataType::U16, DataType::S16, DataType::QS16, DataType::F16, DataType::U32, + DataType::F32); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, output); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(0) + width_offset > output->dimension(0)); + + for(size_t i = 1; i < Coordinates::num_max_dimensions; ++i) + { + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(i) != output->dimension(i)); + } + ARM_COMPUTE_RETURN_ERROR_ON(input->num_dimensions() > 3); + + return Status{}; +} +} // namespace + +CLWidthConcatenateLayerKernel::CLWidthConcatenateLayerKernel() + : _input(nullptr), _output(nullptr), _width_offset(0) +{ +} + +Status CLWidthConcatenateLayerKernel::validate(const ITensorInfo *input, unsigned int width_offset, const ITensorInfo *output) +{ + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, width_offset, output)); + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), width_offset, output->clone().get()).first); + return Status{}; +} + +void CLWidthConcatenateLayerKernel::configure(const ICLTensor *input, unsigned int width_offset, ICLTensor *output) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), width_offset, output->info())); + + _input = input; + _output = output; + _width_offset = width_offset; + + const unsigned int num_elems_processed_per_iteration = 16; + + // Add build options + 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(CLKernelLibrary::get().create_kernel("concatenate_width", build_opts.options())); + + const int offset_to_first_elements_in_bytes = _width_offset * _output->info()->strides_in_bytes()[0]; + + unsigned int idx = 2 * num_arguments_per_3D_tensor(); // Skip the input and output parameters + _kernel.setArg(idx, offset_to_first_elements_in_bytes); + + // Configure kernel window + auto win_config = validate_and_configure_window(input->info(), width_offset, output->info()); + ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); + + ICLKernel::configure(std::get<1>(win_config)); +} + +void CLWidthConcatenateLayerKernel::run(const Window &window, cl::CommandQueue &queue) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); + + Window slice = window.first_slice_window_3D(); + + do + { + unsigned int idx = 0; + add_3D_tensor_argument(idx, _input, slice); + add_3D_tensor_argument(idx, _output, slice); + enqueue(queue, *this, slice); + } + while(window.slide_window_slice_3D(slice)); +} -- cgit v1.2.1