diff options
Diffstat (limited to 'src/core/CL/kernels/CLTileKernel.cpp')
| -rw-r--r-- | src/core/CL/kernels/CLTileKernel.cpp | 47 |
1 files changed, 26 insertions, 21 deletions
diff --git a/src/core/CL/kernels/CLTileKernel.cpp b/src/core/CL/kernels/CLTileKernel.cpp index c0c3d2e2ee..fa996c4008 100644 --- a/src/core/CL/kernels/CLTileKernel.cpp +++ b/src/core/CL/kernels/CLTileKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -22,8 +22,11 @@ * SOFTWARE. */ #include "src/core/CL/kernels/CLTileKernel.h" + #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/utils/misc/ShapeCalculator.h" +#include "arm_compute/core/utils/StringUtils.h" + #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" #include "support/StringSupport.h" @@ -38,15 +41,13 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c ARM_COMPUTE_RETURN_ERROR_ON(input->data_type() == DataType::UNKNOWN); ARM_COMPUTE_RETURN_ERROR_ON(multiples.size() > 4); ARM_COMPUTE_RETURN_ERROR_ON(multiples.empty()); - ARM_COMPUTE_RETURN_ERROR_ON(std::any_of(multiples.begin(), multiples.end(), [](uint32_t e) - { - return e == 0; - })); + ARM_COMPUTE_RETURN_ERROR_ON(std::any_of(multiples.begin(), multiples.end(), [](uint32_t e) { return e == 0; })); // Validate output if initialized - if(output->total_size() != 0) + if (output->total_size() != 0) { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(misc::shape_calculator::compute_tiled_shape(input->tensor_shape(), multiples), output->tensor_shape()); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS( + misc::shape_calculator::compute_tiled_shape(input->tensor_shape(), multiples), output->tensor_shape()); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); } @@ -54,9 +55,9 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, c } } // namespace -CLTileKernel::CLTileKernel() - : _input(nullptr), _output(nullptr) +CLTileKernel::CLTileKernel() : _input(nullptr), _output(nullptr) { + _type = CLKernelType::ELEMENTWISE; } void CLTileKernel::configure(const ICLTensor *input, ICLTensor *output, const Multiples &multiples) @@ -64,7 +65,10 @@ void CLTileKernel::configure(const ICLTensor *input, ICLTensor *output, const Mu configure(CLKernelLibrary::get().get_compile_context(), input, output, multiples); } -void CLTileKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const Multiples &multiples) +void CLTileKernel::configure(const CLCompileContext &compile_context, + const ICLTensor *input, + ICLTensor *output, + const Multiples &multiples) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); @@ -78,11 +82,13 @@ void CLTileKernel::configure(const CLCompileContext &compile_context, const ICLT _input = input; _output = output; - const DataType data_type = input->info()->data_type(); - const int vec_size_x = 16 / input->info()->element_size(); - const int input_width_x = input->info()->tensor_shape().x(); - const unsigned int offset = ceil_to_multiple(input_width_x, vec_size_x) - input_width_x; - const bool multi_access_x = (input_width_x / vec_size_x > 0); + const DataType data_type = input->info()->data_type(); + const int vec_size_x = 16 / input->info()->element_size(); + const int input_width_x = input->info()->tensor_shape().x(); + const unsigned int input_width_ceil = ceil_to_multiple(input_width_x, vec_size_x); + const unsigned int input_width_tiles = input_width_ceil / vec_size_x; + const unsigned int offset = input_width_ceil - input_width_x; + const bool multi_access_x = (input_width_x / vec_size_x > 0); // Create kernel CLBuildOptions build_opts; @@ -94,20 +100,20 @@ void CLTileKernel::configure(const CLCompileContext &compile_context, const ICLT build_opts.add_option("-DDST_DEPTH=" + support::cpp11::to_string(output->info()->dimension(2))); build_opts.add_option_if(multi_access_x, "-DOFFSET=" + support::cpp11::to_string(offset)); 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, "-DSRC_WIDTH_TILES=" + support::cpp11::to_string(input_width_tiles)); _kernel = create_kernel(compile_context, "tile", build_opts.options()); // Configure window without padding Window win = calculate_max_window(*output->info()); - if(multi_access_x) + if (multi_access_x) { // If multi-access is enabled, no thread should cross the tile boundaries. This means we need // as many threads as those to cover a single tile times multiples[0]. Note that if threads // do not cross the boundaries of the tiles, they won't cross the boundaries of the last tile, and // we don't need to pad the output const unsigned int size_win_x = ceil_to_multiple(input->info()->dimension(0), vec_size_x) * multiples[0]; - win.set(Window::DimX, - Window::Dimension(win.x().start(), size_win_x, vec_size_x)); + win.set(Window::DimX, Window::Dimension(win.x().start(), size_win_x, vec_size_x)); } ICLKernel::configure_internal(win); @@ -116,7 +122,7 @@ void CLTileKernel::configure(const CLCompileContext &compile_context, const ICLT _config_id = "tile"; _config_id += "_"; _config_id += lower_string(string_from_data_type(input->info()->data_type())); - for(unsigned int i = 0; i < multiples.size(); ++i) + for (unsigned int i = 0; i < multiples.size(); ++i) { _config_id += "_"; _config_id += support::cpp11::to_string(input->info()->dimension(i)); @@ -145,7 +151,6 @@ void CLTileKernel::run(const Window &window, cl::CommandQueue &queue) add_4D_tensor_argument(idx, _input, slice); add_4D_tensor_argument(idx, _output, slice); enqueue(queue, *this, slice, lws_hint()); - } - while(collapsed.slide_window_slice_4D(slice)); + } while (collapsed.slide_window_slice_4D(slice)); } } // namespace arm_compute |