diff options
Diffstat (limited to 'src/core/CL/kernels/CLSelectKernel.cpp')
-rw-r--r-- | src/core/CL/kernels/CLSelectKernel.cpp | 95 |
1 files changed, 32 insertions, 63 deletions
diff --git a/src/core/CL/kernels/CLSelectKernel.cpp b/src/core/CL/kernels/CLSelectKernel.cpp index 53e5414c88..703c64d8d3 100644 --- a/src/core/CL/kernels/CLSelectKernel.cpp +++ b/src/core/CL/kernels/CLSelectKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2018-2020 Arm Limited. + * Copyright (c) 2018-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -29,10 +29,11 @@ #include "arm_compute/core/Helpers.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Utils.h" +#include "arm_compute/core/utils/helpers/AdjustVecSize.h" + #include "src/core/CL/CLValidate.h" #include "src/core/helpers/AutoConfiguration.h" #include "src/core/helpers/WindowHelpers.h" - #include "support/StringSupport.h" namespace arm_compute @@ -41,7 +42,7 @@ namespace { Status validate_arguments(const ITensorInfo *c, const ITensorInfo *x, const ITensorInfo *y, const ITensorInfo *output) { - ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(c, x, y); + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(c, x, y, output); ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(x); ARM_COMPUTE_RETURN_ERROR_ON(x->data_type() == DataType::UNKNOWN); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(x, y); @@ -50,9 +51,11 @@ Status validate_arguments(const ITensorInfo *c, const ITensorInfo *x, const ITen const bool is_same_rank = (c->tensor_shape().num_dimensions() == x->tensor_shape().num_dimensions()); ARM_COMPUTE_RETURN_ERROR_ON(is_same_rank && (x->tensor_shape() != c->tensor_shape())); - ARM_COMPUTE_RETURN_ERROR_ON(!is_same_rank && ((c->tensor_shape().num_dimensions() > 1) || (c->tensor_shape().x() != x->tensor_shape()[x->tensor_shape().num_dimensions() - 1]))); + ARM_COMPUTE_RETURN_ERROR_ON(!is_same_rank && + ((c->tensor_shape().num_dimensions() > 1) || + (c->tensor_shape().x() != x->tensor_shape()[x->tensor_shape().num_dimensions() - 1]))); - if(output != nullptr && output->total_size() != 0) + if (output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(x, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(x, output); @@ -60,55 +63,18 @@ Status validate_arguments(const ITensorInfo *c, const ITensorInfo *x, const ITen return Status{}; } - -std::pair<Status, Window> validate_and_configure_window(ITensorInfo *c, ITensorInfo *x, ITensorInfo *y, ITensorInfo *output) -{ - if(output != nullptr) - { - // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*output, *x->clone()); - } - - const bool is_same_rank = (c->tensor_shape().num_dimensions() == x->tensor_shape().num_dimensions()); - - const unsigned int num_elems_processed_per_iteration = 16 / x->element_size(); - - // Configure kernel window - Window win = calculate_max_window(*x, Steps(num_elems_processed_per_iteration)); - AccessWindowHorizontal x_access(x, 0, num_elems_processed_per_iteration); - AccessWindowHorizontal y_access(y, 0, num_elems_processed_per_iteration); - bool window_changed = update_window_and_padding(win, x_access, y_access); - - // Update window for condition - if(is_same_rank) - { - AccessWindowHorizontal c_access(c, 0, num_elems_processed_per_iteration); - window_changed = window_changed || update_window_and_padding(win, c_access); - } - - // Update window for output - if(output != nullptr) - { - AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); - window_changed = window_changed || update_window_and_padding(win, output_access); - output_access.set_valid_region(win, x->valid_region()); - } - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_pair(err, win); -} } // namespace -CLSelectKernel::CLSelectKernel() - : _c(nullptr), _x(nullptr), _y(nullptr), _output(nullptr), _has_same_rank(false) +CLSelectKernel::CLSelectKernel() : _c(nullptr), _x(nullptr), _y(nullptr), _output(nullptr), _has_same_rank(false) { -} -void CLSelectKernel::configure(const ICLTensor *c, const ICLTensor *x, const ICLTensor *y, ICLTensor *output) -{ - configure(CLKernelLibrary::get().get_compile_context(), c, x, y, output); + _type = CLKernelType::ELEMENTWISE; } -void CLSelectKernel::configure(const CLCompileContext &compile_context, const ICLTensor *c, const ICLTensor *x, const ICLTensor *y, ICLTensor *output) +void CLSelectKernel::configure(const CLCompileContext &compile_context, + const ICLTensor *c, + const ICLTensor *x, + const ICLTensor *y, + ICLTensor *output) { ARM_COMPUTE_ERROR_ON_NULLPTR(c, x, y, output); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(c->info(), x->info(), y->info(), output->info())); @@ -119,23 +85,26 @@ void CLSelectKernel::configure(const CLCompileContext &compile_context, const IC _output = output; _has_same_rank = (c->info()->tensor_shape().num_dimensions() == x->info()->tensor_shape().num_dimensions()); - const unsigned int num_elems_processed_per_iteration = 16 / x->info()->element_size(); + auto padding_info = get_padding_info({c, x, y, output}); + const unsigned int vec_size_x = adjust_vec_size(16 / x->info()->element_size(), x->info()->dimension(0)); + const int vec_size_x_leftovers = output->info()->dimension(0) % vec_size_x; // Set build options CLBuildOptions build_opts; - build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(x->info()->data_type())); - build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration)); + build_opts.add_option("-DDATA_TYPE=" + get_cl_unsigned_type_from_element_size(x->info()->element_size())); + build_opts.add_option("-DVEC_SIZE=" + support::cpp11::to_string(vec_size_x)); + build_opts.add_option("-DVEC_SIZE_LEFTOVER=" + support::cpp11::to_string(vec_size_x_leftovers)); // Create kernel std::string kernel_name = "select"; - if(_has_same_rank) + if (_has_same_rank) { kernel_name += "_same_rank"; } else { const bool is_input_rank_greater_than_two = x->info()->tensor_shape().num_dimensions() > 2; - if(is_input_rank_greater_than_two) + if (is_input_rank_greater_than_two) { const size_t width = x->info()->tensor_shape().x(); const size_t height = x->info()->tensor_shape().y(); @@ -149,9 +118,9 @@ void CLSelectKernel::configure(const CLCompileContext &compile_context, const IC _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); // Configure kernel window - auto win_config = validate_and_configure_window(c->info(), x->info(), y->info(), output->info()); - ARM_COMPUTE_ERROR_THROW_ON(win_config.first); - ICLKernel::configure_internal(win_config.second); + auto_init_if_empty(*output->info(), *x->info()->clone()); + Window win = calculate_max_window(*x->info(), Steps(vec_size_x)); + ICLKernel::configure_internal(win); _config_id = "select_"; _config_id += string_from_data_type(x->info()->data_type()); @@ -161,12 +130,13 @@ void CLSelectKernel::configure(const CLCompileContext &compile_context, const IC _config_id += support::cpp11::to_string(x->info()->dimension(1)); _config_id += "_"; _config_id += support::cpp11::to_string(x->info()->dimension(2)); + ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); } -Status CLSelectKernel::validate(const ITensorInfo *c, const ITensorInfo *x, const ITensorInfo *y, const ITensorInfo *output) +Status +CLSelectKernel::validate(const ITensorInfo *c, const ITensorInfo *x, const ITensorInfo *y, const ITensorInfo *output) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(c, x, y, output)); - ARM_COMPUTE_RETURN_ON_ERROR(validate_and_configure_window(c->clone().get(), x->clone().get(), y->clone().get(), output->clone().get()).first); return Status{}; } @@ -178,7 +148,7 @@ void CLSelectKernel::run(const arm_compute::Window &window, cl::CommandQueue &qu Window collapsed = window.collapse_if_possible(ICLKernel::window(), Window::DimZ); Window slice = collapsed.first_slice_window_3D(); - if(!_has_same_rank) + if (!_has_same_rank) { Window vector_slice = window.first_slice_window_1D(); vector_slice.set(Window::DimX, Window::Dimension(0, 0, 0)); @@ -189,7 +159,7 @@ void CLSelectKernel::run(const arm_compute::Window &window, cl::CommandQueue &qu do { unsigned int idx = _has_same_rank ? 0 : num_arguments_per_1D_tensor(); - if(_has_same_rank) + if (_has_same_rank) { add_3D_tensor_argument(idx, _c, slice); } @@ -198,7 +168,6 @@ void CLSelectKernel::run(const arm_compute::Window &window, cl::CommandQueue &qu add_3D_tensor_argument(idx, _output, slice); enqueue(queue, *this, slice, lws_hint()); - } - while(collapsed.slide_window_slice_3D(slice)); + } while (collapsed.slide_window_slice_3D(slice)); } } // namespace arm_compute |