diff options
Diffstat (limited to 'src/core/CL/kernels/CLROIPoolingLayerKernel.cpp')
-rw-r--r-- | src/core/CL/kernels/CLROIPoolingLayerKernel.cpp | 156 |
1 files changed, 88 insertions, 68 deletions
diff --git a/src/core/CL/kernels/CLROIPoolingLayerKernel.cpp b/src/core/CL/kernels/CLROIPoolingLayerKernel.cpp index a5b80eb5ef..1b2c414a49 100644 --- a/src/core/CL/kernels/CLROIPoolingLayerKernel.cpp +++ b/src/core/CL/kernels/CLROIPoolingLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 ARM Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,85 +21,86 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/core/CL/kernels/CLROIPoolingLayerKernel.h" +#include "src/core/CL/kernels/CLROIPoolingLayerKernel.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/CLValidate.h" -#include "arm_compute/core/CL/ICLArray.h" #include "arm_compute/core/CL/ICLTensor.h" #include "arm_compute/core/CL/OpenCL.h" #include "arm_compute/core/Helpers.h" #include "arm_compute/core/TensorInfo.h" #include "arm_compute/core/Utils.h" -#include "arm_compute/core/Window.h" +#include "arm_compute/core/utils/StringUtils.h" + +#include "src/core/CL/CLValidate.h" +#include "src/core/helpers/AutoConfiguration.h" +#include "src/core/helpers/WindowHelpers.h" #include "support/StringSupport.h" +#include <cfloat> #include <cmath> -#include <set> #include <string> namespace arm_compute { -namespace -{ -std::pair<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *rois, ITensorInfo *output, const ROIPoolingLayerInfo &pool_info) +CLROIPoolingLayerKernel::CLROIPoolingLayerKernel() + : _input(nullptr), _rois(nullptr), _output(nullptr), _pool_info(0, 0, 0.f) { - ARM_COMPUTE_ERROR_ON_NULLPTR(input, output); - - // Output auto initialization if not yet initialized - TensorShape output_shape(pool_info.pooled_width(), pool_info.pooled_height(), input->dimension(2), rois->dimension(1)); - auto_init_if_empty((*output), output_shape, 1, input->data_type()); - - // Configure kernel window - constexpr unsigned int num_elems_processed_per_iteration = 1; - Window win = calculate_max_window(*output, Steps(num_elems_processed_per_iteration)); - - AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); - AccessWindowHorizontal input_access(input, input->valid_region().start(0), num_elems_processed_per_iteration); - - bool window_changed = update_window_and_padding(win, input_access, output_access); - output_access.set_valid_region(win, ValidRegion(Coordinates(), output->tensor_shape())); - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - return std::make_pair(err, win); + _type = CLKernelType::ELEMENTWISE; } -} // namespace -CLROIPoolingLayerKernel::CLROIPoolingLayerKernel() - : _input(nullptr), _rois(nullptr), _output(nullptr), _pool_info(0, 0, 0.f) +Status CLROIPoolingLayerKernel::validate(const ITensorInfo *input, + const ITensorInfo *rois, + const ITensorInfo *output, + const ROIPoolingLayerInfo &pool_info) { + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, rois, output); + + //Validate arguments + ARM_COMPUTE_RETURN_ERROR_ON_NULLPTR(input, rois, output); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(rois, 1, DataType::U16); + ARM_COMPUTE_RETURN_ERROR_ON(rois->dimension(0) != 5); + ARM_COMPUTE_RETURN_ERROR_ON(rois->num_dimensions() > 2); + ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); + ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16, DataType::QASYMM8); + ARM_COMPUTE_RETURN_ERROR_ON((pool_info.pooled_width() == 0) || (pool_info.pooled_height() == 0)); + + if (output->total_size() != 0) + { + ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_RETURN_ERROR_ON((output->dimension(0) != pool_info.pooled_width()) || + (output->dimension(1) != pool_info.pooled_height())); + ARM_COMPUTE_RETURN_ERROR_ON(input->dimension(2) != output->dimension(2)); + ARM_COMPUTE_RETURN_ERROR_ON(rois->dimension(1) != output->dimension(3)); + } + + return Status{}; } -void CLROIPoolingLayerKernel::configure(const ICLTensor *input, const ICLTensor *rois, ICLTensor *output, const ROIPoolingLayerInfo &pool_info) +void CLROIPoolingLayerKernel::configure(const ICLTensor *input, + const ICLTensor *rois, + ICLTensor *output, + const ROIPoolingLayerInfo &pool_info) { configure(CLKernelLibrary::get().get_compile_context(), input, rois, output, pool_info); } -void CLROIPoolingLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *rois, ICLTensor *output, const ROIPoolingLayerInfo &pool_info) +void CLROIPoolingLayerKernel::configure(const CLCompileContext &compile_context, + const ICLTensor *input, + const ICLTensor *rois, + const ICLTensor *output, + const ROIPoolingLayerInfo &pool_info) { - ARM_COMPUTE_ERROR_ON_NULLPTR(input, rois, output); + ARM_COMPUTE_ERROR_THROW_ON( + CLROIPoolingLayerKernel::validate(input->info(), rois->info(), output->info(), pool_info)); - //Validate arguments - ARM_COMPUTE_ERROR_ON_NULLPTR(input->info(), rois->info(), output->info()); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(rois, 1, DataType::U16); - ARM_COMPUTE_ERROR_ON(rois->info()->dimension(0) != 5); - ARM_COMPUTE_ERROR_ON(rois->info()->num_dimensions() > 2); - ARM_COMPUTE_ERROR_ON_F16_UNSUPPORTED(input); - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F32, DataType::F16); - ARM_COMPUTE_ERROR_ON((pool_info.pooled_width() == 0) || (pool_info.pooled_height() == 0)); - - if(output->info()->total_size() != 0) - { - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); - ARM_COMPUTE_ERROR_ON((output->info()->dimension(0) != pool_info.pooled_width()) || (output->info()->dimension(1) != pool_info.pooled_height())); - ARM_COMPUTE_ERROR_ON(input->info()->dimension(2) != output->info()->dimension(2)); - ARM_COMPUTE_ERROR_ON(rois->info()->dimension(1) != output->info()->dimension(3)); - } + auto padding_info = get_padding_info({input, rois, output}); - // Configure kernel window - auto win_config = validate_and_configure_window(input->info(), rois->info(), output->info(), pool_info); - ARM_COMPUTE_ERROR_THROW_ON(win_config.first); + // Output auto initialization if not yet initialized + TensorShape output_shape(pool_info.pooled_width(), pool_info.pooled_height(), input->info()->dimension(2), + rois->info()->dimension(1)); + auto_init_if_empty(*(output->info()), output_shape, 1, input->info()->data_type(), + output->info()->quantization_info()); // Set instance variables _input = input; @@ -107,27 +108,46 @@ void CLROIPoolingLayerKernel::configure(const CLCompileContext &compile_context, _output = output; _pool_info = pool_info; + const DataType data_type = input->info()->data_type(); + const bool is_qasymm = is_data_type_quantized_asymmetric(data_type); + // Set build options - std::set<std::string> build_opts; - build_opts.emplace(("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type()))); - build_opts.emplace(("-DDATA_SIZE=" + get_data_size_from_data_type(input->info()->data_type()))); - build_opts.emplace(("-DMAX_DIM_X=" + support::cpp11::to_string(_input->info()->dimension(Window::DimX)))); - build_opts.emplace(("-DMAX_DIM_Y=" + support::cpp11::to_string(_input->info()->dimension(Window::DimY)))); - build_opts.emplace(("-DMAX_DIM_Z=" + support::cpp11::to_string(_input->info()->dimension(Window::DimZ)))); - build_opts.emplace(("-DPOOLED_DIM_X=" + support::cpp11::to_string(pool_info.pooled_width()))); - build_opts.emplace(("-DPOOLED_DIM_Y=" + support::cpp11::to_string(pool_info.pooled_height()))); - build_opts.emplace(("-DSPATIAL_SCALE=" + support::cpp11::to_string(pool_info.spatial_scale()))); + CLBuildOptions build_opts; + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(data_type)); + build_opts.add_option("-DDATA_SIZE=" + get_data_size_from_data_type(data_type)); + build_opts.add_option("-DMAX_DIM_X=" + support::cpp11::to_string(_input->info()->dimension(Window::DimX))); + build_opts.add_option("-DMAX_DIM_Y=" + support::cpp11::to_string(_input->info()->dimension(Window::DimY))); + build_opts.add_option("-DMAX_DIM_Z=" + support::cpp11::to_string(_input->info()->dimension(Window::DimZ))); + build_opts.add_option("-DPOOLED_DIM_X=" + support::cpp11::to_string(pool_info.pooled_width())); + build_opts.add_option("-DPOOLED_DIM_Y=" + support::cpp11::to_string(pool_info.pooled_height())); + build_opts.add_option("-DSPATIAL_SCALE=" + support::cpp11::to_string(pool_info.spatial_scale())); + + if (is_qasymm) + { + // Determine quantization info scale, offset + UniformQuantizationInfo uqinfo = UniformQuantizationInfo(); + uqinfo = compute_requantization_scale_offset(_input->info()->quantization_info().uniform(), + _output->info()->quantization_info().uniform()); + build_opts.add_option("-DOFFSET_OUT=" + float_to_string_with_full_precision(uqinfo.offset)); + build_opts.add_option("-DSCALE_OUT=" + float_to_string_with_full_precision(uqinfo.scale)); + + // Specify minimum possible value of datatype + build_opts.add_option("-DMIN_VALUE=" + support::cpp11::to_string(0)); + } + else + { + // Specify min value of F32 datatype + build_opts.add_option("-DMIN_VALUE=" + support::cpp11::to_string(-FLT_MAX)); + } + + Window win = calculate_max_window(*(output->info()), Steps()); + ICLKernel::configure_internal(win); // Create kernel std::string kernel_name = "roi_pooling_layer"; - _kernel = create_kernel(compile_context, kernel_name, build_opts); - - // Set static kernel arguments - unsigned int idx = 2 * num_arguments_per_3D_tensor() + num_arguments_per_1D_array(); - add_argument<cl_uint>(idx, _input->info()->strides_in_bytes()[3]); - add_argument<cl_uint>(idx, _output->info()->strides_in_bytes()[3]); + _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); - ICLKernel::configure_internal(win_config.second); + ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); } void CLROIPoolingLayerKernel::run(const Window &window, cl::CommandQueue &queue) |