From d7647d4ebd0f0b5253b7f31ffcd48a851ba62947 Mon Sep 17 00:00:00 2001 From: Giuseppe Rossini Date: Tue, 17 Jul 2018 18:13:13 +0100 Subject: [COMPMID-1229] Implementing Pad on OpenCL -FP32/FP16 Change-Id: Ideead99410e5e0bda1035030af1bbcd0a65ea15e Reviewed-on: https://eu-gerrit-1.euhpc.arm.com/144792 Tested-by: bsgcomp Reviewed-by: Georgios Pinitas --- src/core/CL/kernels/CLCopyKernel.cpp | 126 +++++++++++++++++++++++++++++---- src/core/CL/kernels/CLMemsetKernel.cpp | 95 +++++++++++++++++++++++++ 2 files changed, 206 insertions(+), 15 deletions(-) create mode 100644 src/core/CL/kernels/CLMemsetKernel.cpp (limited to 'src/core/CL/kernels') diff --git a/src/core/CL/kernels/CLCopyKernel.cpp b/src/core/CL/kernels/CLCopyKernel.cpp index 2da67d2666..e14e5dafab 100644 --- a/src/core/CL/kernels/CLCopyKernel.cpp +++ b/src/core/CL/kernels/CLCopyKernel.cpp @@ -30,21 +30,22 @@ #include "arm_compute/core/Error.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Validate.h" #include "arm_compute/core/Window.h" +#include "arm_compute/core/utils/misc/ShapeCalculator.h" -using namespace arm_compute; - +namespace arm_compute +{ namespace { -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output) +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding = PaddingList()) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); + ARM_COMPUTE_RETURN_ERROR_ON(padding.size() > 4); // Validate output if initialized if(output->total_size() != 0) { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(input->tensor_shape(), output->tensor_shape()); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(misc::shape_calculator::compute_padded_shape(input->tensor_shape(), padding), output->tensor_shape()); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); } @@ -69,6 +70,64 @@ std::pair validate_and_configure_window(ITensorInfo *input, ITen Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; return std::make_pair(err, win); } + +std::pair validate_and_configure_window_with_padding(ITensorInfo *input, ITensorInfo *output, const PaddingList &padding) +{ + TensorShape input_shape = input->tensor_shape(); + TensorShape padded_shape = misc::shape_calculator::compute_padded_shape(input_shape, padding); + + auto_init_if_empty(*output, input->clone()->set_tensor_shape(padded_shape)); + + // Configure window + const unsigned int num_elems_processed_per_iteration = 16 / input->element_size(); + + Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); + + // Pad on the x dimension accounting for the padding offset along the same dimension + AccessWindowHorizontal output_access(output, padding[0].first, num_elems_processed_per_iteration); + AccessWindowHorizontal input_access(input, 0, 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); +} + +/** Generate the string "-DPAD= @p dim @p index @p padding" + * + * @param[in] dim The dimension index + * @param[in] index Can be 0 for the start dimension and 1 for the end dimension + * @param[in] padding The value to pad for that index/dimension pair + * + * @return The correct concatenated string + */ +std::string generate_pad_string(const size_t dim, const size_t index, const size_t padding) +{ + return "-DPAD" + support::cpp11::to_string(dim) + support::cpp11::to_string(index) + "=" + support::cpp11::to_string(padding); +} + +/** Pass the padding as build option to the kernel. + * + * @param[in] tensor The padded tensor + * @param[in] padding The list of the padding for each dimension + * @param[out] build_opts The build option to which adding the padding + */ +void add_padding_as_build_options(const PaddingList &padding, CLBuildOptions &build_opts) +{ + size_t dim = 0; + for(dim = 0; dim < padding.size(); dim++) + { + build_opts.add_option(generate_pad_string(dim, 0, padding[dim].first)); + build_opts.add_option(generate_pad_string(dim, 1, padding[dim].second)); + } + + while(dim < TensorShape::num_max_dimensions) + { + build_opts.add_option(generate_pad_string(dim, 0, 0)); + build_opts.add_option(generate_pad_string(dim, 1, 0)); + dim++; + } +} + } // namespace CLCopyKernel::CLCopyKernel() @@ -76,32 +135,68 @@ CLCopyKernel::CLCopyKernel() { } -void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output) +void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output, const PaddingList &padding) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info())); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), padding)); _input = input; _output = output; - const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size(); - // Create kernel CLBuildOptions build_opts; build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); + + const unsigned int num_elems_processed_per_iteration = 16 / input->info()->element_size(); build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); - _kernel = static_cast(CLKernelLibrary::get().create_kernel("copy_tensor", build_opts.options())); - // Configure kernel window - auto win_config = validate_and_configure_window(input->info(), output->info()); + std::pair win_config; + + if(padding.empty()) + { + // Build kernel + _kernel = static_cast(CLKernelLibrary::get().create_kernel("copy_tensor", build_opts.options())); + + // Configure window + win_config = validate_and_configure_window(input->info(), output->info()); + } + else + { + // Add compile time options + add_padding_as_build_options(padding, build_opts); + + // If we are padding in the fourth dimension the kernel needs to know the depth of the + // different cubes + if(padding.size() == 4) + { + const size_t depth = input->info()->tensor_shape()[2]; + build_opts.add_option("-DDEPTH=" + support::cpp11::to_string(depth)); + } + + // Build kernel + _kernel = static_cast(CLKernelLibrary::get().create_kernel("copy_pad_tensor", build_opts.options())); + + // Configure window + win_config = validate_and_configure_window_with_padding(input->info(), output->info(), padding); + } + + // Validate and set the window ARM_COMPUTE_ERROR_THROW_ON(win_config.first); ICLKernel::configure_internal(win_config.second); } -Status CLCopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output) +Status CLCopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output, const PaddingList &padding) { - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first); + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, padding)); + + if(padding.empty()) + { + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(input->clone().get(), output->clone().get()).first); + } + else + { + ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_with_padding(input->clone().get(), output->clone().get(), padding).first); + } return Status{}; } @@ -123,3 +218,4 @@ void CLCopyKernel::run(const Window &window, cl::CommandQueue &queue) } while(collapsed.slide_window_slice_3D(slice)); } +} // namespace arm_compute diff --git a/src/core/CL/kernels/CLMemsetKernel.cpp b/src/core/CL/kernels/CLMemsetKernel.cpp new file mode 100644 index 0000000000..ab53897543 --- /dev/null +++ b/src/core/CL/kernels/CLMemsetKernel.cpp @@ -0,0 +1,95 @@ +/* + * 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/CLMemsetKernel.h" + +#include "arm_compute/core/AccessWindowStatic.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/Helpers.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/Window.h" + +namespace arm_compute +{ +CLMemsetKernel::CLMemsetKernel() + : ICLKernel(), _tensor(nullptr) +{ +} + +void CLMemsetKernel::configure(ICLTensor *tensor, + const PixelValue &constant_value) +{ + ARM_COMPUTE_ERROR_ON_NULLPTR(tensor); + _tensor = tensor; + + const DataType data_type = tensor->info()->data_type(); + const int vec_size_x = 16 / tensor->info()->element_size(); + const int output_width_x = tensor->info()->tensor_shape().x(); + const bool multi_access_x = (output_width_x / vec_size_x > 0); + + // Create and update the window (if needed) + Window win = calculate_max_window(*tensor->info()); + if(multi_access_x) + { + win.set(Window::DimX, + Window::Dimension(win.x().start(), ceil_to_multiple(win.x().end(), vec_size_x), vec_size_x)); + } + ICLKernel::configure_internal(win); + + // Create kernel + CLBuildOptions build_opts; + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); + build_opts.add_option("-DCONSTANT_VALUE=" + string_from_pixel_value(constant_value, data_type)); + build_opts.add_option_if(multi_access_x, "-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); + build_opts.add_option_if(multi_access_x, "-DLAST_ACCESSED_X=" + support::cpp11::to_string(std::max(output_width_x - vec_size_x, 0))); + _kernel = static_cast(CLKernelLibrary::get().create_kernel("memset", build_opts.options())); +} + +Status CLMemsetKernel::validate(const ITensorInfo *tensor, const PixelValue &constant_value) +{ + ARM_COMPUTE_UNUSED(tensor); + ARM_COMPUTE_UNUSED(constant_value); + return Status{}; +} + +void CLMemsetKernel::run(const Window &window, cl::CommandQueue &queue) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(ICLKernel::window(), window); + + // Collapse all the batches on the third + Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimY); + Window slice = collapsed.first_slice_window_2D(); + + do + { + unsigned int idx = 0; + add_2D_tensor_argument(idx, _tensor, slice); + enqueue(queue, *this, slice); + } + while(collapsed.slide_window_slice_2D(slice)); +} +} // namespace arm_compute -- cgit v1.2.1