diff options
Diffstat (limited to 'src/core/CL/kernels/CLCopyKernel.cpp')
| -rw-r--r-- | src/core/CL/kernels/CLCopyKernel.cpp | 171 |
1 files changed, 33 insertions, 138 deletions
diff --git a/src/core/CL/kernels/CLCopyKernel.cpp b/src/core/CL/kernels/CLCopyKernel.cpp index 769f15de0f..184b80caa8 100644 --- a/src/core/CL/kernels/CLCopyKernel.cpp +++ b/src/core/CL/kernels/CLCopyKernel.cpp @@ -26,11 +26,7 @@ #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/TensorInfo.h" #include "arm_compute/core/Utils.h" -#include "arm_compute/core/utils/misc/ShapeCalculator.h" -#include "src/core/AccessWindowStatic.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" #include "support/StringSupport.h" @@ -39,100 +35,28 @@ namespace arm_compute { namespace { -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, const PaddingList &padding = PaddingList(), Window *output_window = nullptr) +Status validate_arguments(const ITensorInfo *input, const ITensorInfo *output, Window *output_window = nullptr) { ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_ERROR_ON(!padding.empty() && output_window != nullptr); - ARM_COMPUTE_RETURN_ERROR_ON(padding.size() > 4); // Validate output if initialized if(output->total_size() != 0) { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); if(output_window == nullptr) { - 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_DIMENSIONS(input->tensor_shape(), output->tensor_shape()); } else { - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_QUANTIZATION_INFO(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(input->tensor_shape(), output_window->shape()); } - ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); } return Status{}; } -std::pair<Status, Window> configure_window(ITensorInfo *input, ITensorInfo *output) -{ - // Output auto inizialitation if not yet initialized - auto_init_if_empty(*output, *input); - - // Configure window - const unsigned int vec_size_x = adjust_vec_size(16 / input->element_size(), input->dimension(0)); - - const Window win = calculate_max_window(*input, Steps(vec_size_x)); - return std::make_pair(Status{}, win); -} - -std::pair<Status, Window> 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() @@ -140,15 +64,15 @@ CLCopyKernel::CLCopyKernel() { } -void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output, const PaddingList &padding, Window *output_window) +void CLCopyKernel::configure(const ICLTensor *input, ICLTensor *output, Window *output_window) { - configure(CLKernelLibrary::get().get_compile_context(), input, output, padding, output_window); + configure(CLKernelLibrary::get().get_compile_context(), input, output, output_window); } -void CLCopyKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, const PaddingList &padding, Window *output_window) +void CLCopyKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, ICLTensor *output, Window *output_window) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), padding, output_window)); + ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), output->info(), output_window)); auto padding_info = get_padding_info({ input, output }); @@ -159,80 +83,51 @@ void CLCopyKernel::configure(const CLCompileContext &compile_context, const ICLT CLBuildOptions build_opts; build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); - std::pair<Status, Window> win_config; - - const unsigned int vec_size_x = 16 / input->info()->element_size(); + // Output auto inizialitation if not yet initialized + auto_init_if_empty(*(output->info()), *(input->info())); - if(padding.empty()) - { - // Configure window - win_config = configure_window(input->info(), output->info()); + // Configure window + const unsigned int vec_size_x = adjust_vec_size(16 / input->info()->element_size(), input->info()->dimension(0)); - if(output_window != nullptr) - { - _has_output_window = true; - _output_window = Window(*output_window); - const int width_x = output_window->num_iterations(0); - const int vec_size_x_leftover = width_x % vec_size_x; - const bool multi_access_x = width_x >= static_cast<int32_t>(vec_size_x); + const Window win_config = calculate_max_window(*(input->info()), Steps(vec_size_x)); - if(multi_access_x) - { - _output_window.set(Window::DimX, Window::Dimension(output_window->x().start(), ceil_to_multiple(output_window->x().end(), vec_size_x), vec_size_x)); - } + if(output_window != nullptr) + { + _has_output_window = true; + _output_window = Window(*output_window); + const int width_x = output_window->num_iterations(0); + const int vec_size_x_leftover = width_x % vec_size_x; + const bool multi_access_x = width_x >= static_cast<int32_t>(vec_size_x); - build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_x_leftover)); - } - else + if(multi_access_x) { - const int width_x = input->info()->tensor_shape().x(); - const int vec_size_x_leftover = width_x % vec_size_x; - - build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_x_leftover)); + _output_window.set(Window::DimX, Window::Dimension(output_window->x().start(), ceil_to_multiple(output_window->x().end(), vec_size_x), vec_size_x)); } - build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); - - // Build kernel - _kernel = create_kernel(compile_context, "copy_tensor", build_opts.options()); + build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_x_leftover)); } else { - build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); - - // Add compile time options - add_padding_as_build_options(padding, build_opts); + const int width_x = input->info()->tensor_shape().x(); + const int vec_size_x_leftover = width_x % vec_size_x; - // 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_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_x_leftover)); + } - // Build kernel - _kernel = create_kernel(compile_context, "copy_pad_tensor", build_opts.options()); + build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); - // Configure window - win_config = validate_and_configure_window_with_padding(input->info(), output->info(), padding); - } + // Build kernel + _kernel = create_kernel(compile_context, "copy_tensor", build_opts.options()); // Validate and set the window - ARM_COMPUTE_ERROR_THROW_ON(win_config.first); - ICLKernel::configure_internal(win_config.second); + ICLKernel::configure_internal(win_config); ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); } -Status CLCopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output, const PaddingList &padding, Window *output_window) +Status CLCopyKernel::validate(const arm_compute::ITensorInfo *input, const arm_compute::ITensorInfo *output, Window *output_window) { - ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, padding, output_window)); - - if(!padding.empty()) - { - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window_with_padding(input->clone().get(), output->clone().get(), padding).first); - } + ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, output, output_window)); return Status{}; } |