diff options
Diffstat (limited to 'src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp')
-rw-r--r-- | src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp | 98 |
1 files changed, 47 insertions, 51 deletions
diff --git a/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp b/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp index e04950d0a2..9ed9d7c5b0 100644 --- a/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp +++ b/src/core/CL/kernels/CLL2NormalizeLayerKernel.cpp @@ -1,5 +1,5 @@ /* - * Copyright (c) 2017-2020 ARM Limited. + * Copyright (c) 2017-2021, 2023 Arm Limited. * * SPDX-License-Identifier: MIT * @@ -21,18 +21,20 @@ * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE * SOFTWARE. */ -#include "arm_compute/core/CL/kernels/CLL2NormalizeLayerKernel.h" +#include "src/core/CL/kernels/CLL2NormalizeLayerKernel.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/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/helpers/AdjustVecSize.h" #include "arm_compute/core/Validate.h" -#include "arm_compute/core/Window.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 @@ -41,9 +43,8 @@ namespace { constexpr int max_input_tensor_dim = 3; -constexpr unsigned int num_elems_processed_per_iteration = 16; - -Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon) +Status +validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon) { ARM_COMPUTE_UNUSED(epsilon); @@ -53,14 +54,15 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, cons ARM_COMPUTE_RETURN_ERROR_ON_F16_UNSUPPORTED(input); ARM_COMPUTE_RETURN_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis > 2, "Actual axis greater than 2 is not supported"); - ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis >= TensorShape::num_max_dimensions, "Actual normalization axis greater than max number of dimensions"); + ARM_COMPUTE_RETURN_ERROR_ON_MSG(actual_axis >= TensorShape::num_max_dimensions, + "Actual normalization axis greater than max number of dimensions"); // Reduce shape on axis TensorShape sum_shape = input->tensor_shape(); sum_shape.set(actual_axis, 1); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DIMENSIONS(sum->tensor_shape(), sum_shape); - if(output->total_size() != 0) + if (output->total_size() != 0) { ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_SHAPES(input, output); ARM_COMPUTE_RETURN_ERROR_ON_MISMATCHING_DATA_LAYOUT(input, output); @@ -70,40 +72,30 @@ Status validate_arguments(const ITensorInfo *input, const ITensorInfo *sum, cons return Status{}; } - -std::tuple<Status, Window> validate_and_configure_window(ITensorInfo *input, ITensorInfo *output) -{ - Window win = calculate_max_window(*input, Steps(num_elems_processed_per_iteration)); - - // Output tensor auto initialization if not yet initialized - auto_init_if_empty(*output, input->tensor_shape(), 1, input->data_type()); - - AccessWindowHorizontal input_access(input, 0, num_elems_processed_per_iteration); - AccessWindowHorizontal output_access(output, 0, num_elems_processed_per_iteration); - - bool window_changed = update_window_and_padding(win, input_access, output_access); - output_access.set_valid_region(win, input->valid_region()); - - Status err = (window_changed) ? ARM_COMPUTE_CREATE_ERROR(ErrorCode::RUNTIME_ERROR, "Insufficient Padding!") : Status{}; - - return std::make_tuple(err, win); -} } // namespace CLL2NormalizeLayerKernel::CLL2NormalizeLayerKernel() : _input(nullptr), _sum(nullptr), _output(nullptr), _actual_axis(0), _epsilon(1e-12) { + _type = CLKernelType::ELEMENTWISE; } -void CLL2NormalizeLayerKernel::configure(const ICLTensor *input, const ICLTensor *sum, ICLTensor *output, int axis, float epsilon) +void CLL2NormalizeLayerKernel::configure( + const ICLTensor *input, const ICLTensor *sum, ICLTensor *output, int axis, float epsilon) { configure(CLKernelLibrary::get().get_compile_context(), input, sum, output, axis, epsilon); } -void CLL2NormalizeLayerKernel::configure(const CLCompileContext &compile_context, const ICLTensor *input, const ICLTensor *sum, ICLTensor *output, int axis, float epsilon) +void CLL2NormalizeLayerKernel::configure(const CLCompileContext &compile_context, + const ICLTensor *input, + const ICLTensor *sum, + ICLTensor *output, + int axis, + float epsilon) { ARM_COMPUTE_ERROR_ON_NULLPTR(input, sum, output); ARM_COMPUTE_ERROR_THROW_ON(validate_arguments(input->info(), sum->info(), output->info(), axis, epsilon)); + auto padding_info = get_padding_info({input, sum, output}); _input = input; _sum = sum; @@ -111,35 +103,40 @@ void CLL2NormalizeLayerKernel::configure(const CLCompileContext &compile_context _actual_axis = wrap_around(axis, max_input_tensor_dim); _epsilon = epsilon; + const unsigned int vec_size_x = + adjust_vec_size(max_cl_vector_width / input->info()->element_size(), input->info()->dimension(0)); + const int vec_size_x_leftovers = input->info()->dimension(0) % vec_size_x; + // 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(("-DVEC_SIZE=" + support::cpp11::to_string(num_elems_processed_per_iteration))); + CLBuildOptions build_opts; + build_opts.add_option("-DDATA_TYPE=" + get_cl_type_from_data_type(input->info()->data_type())); + build_opts.add_option("-DVEC_SIZE_X=" + support::cpp11::to_string(vec_size_x)); + build_opts.add_option("-DVEC_SIZE_LEFTOVER_X=" + support::cpp11::to_string(vec_size_x_leftovers)); // Create kernel std::string kernel_name; unsigned int idx = 0; - switch(_actual_axis) + switch (_actual_axis) { case 0: - kernel_name = "x"; + kernel_name = "l2_normalize_x"; idx = num_arguments_per_2D_tensor() * 3; break; case 1: - kernel_name = "y"; + kernel_name = "l2_normalize_y"; idx = num_arguments_per_2D_tensor() * 3; break; case 2: - kernel_name = "z"; + kernel_name = "l2_normalize_z"; idx = num_arguments_per_3D_tensor() * 3; break; default: ARM_COMPUTE_ERROR("Axis not supported"); } - _kernel = create_kernel(compile_context, "l2_normalize_" + kernel_name, build_opts); + _kernel = create_kernel(compile_context, kernel_name, build_opts.options()); // Set epsilon argument - if(input->info()->data_type() == DataType::F32) + if (input->info()->data_type() == DataType::F32) { _kernel.setArg<cl_float>(idx, _epsilon); } @@ -149,17 +146,19 @@ void CLL2NormalizeLayerKernel::configure(const CLCompileContext &compile_context } // Configure kernel window - auto win_config = validate_and_configure_window(_input->info(), _output->info()); - ARM_COMPUTE_ERROR_THROW_ON(std::get<0>(win_config)); + Window win = calculate_max_window(*input->info(), Steps(vec_size_x)); - ICLKernel::configure_internal(std::get<1>(win_config)); + // Output tensor auto initialization if not yet initialized + auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type()); + + ICLKernel::configure_internal(win); + ARM_COMPUTE_ERROR_ON(has_padding_changed(padding_info)); } -Status CLL2NormalizeLayerKernel::validate(const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon) +Status CLL2NormalizeLayerKernel::validate( + const ITensorInfo *input, const ITensorInfo *sum, const ITensorInfo *output, int axis, float epsilon) { ARM_COMPUTE_RETURN_ON_ERROR(validate_arguments(input, sum, output, axis, epsilon)); - ARM_COMPUTE_RETURN_ON_ERROR(std::get<0>(validate_and_configure_window(input->clone().get(), output->clone().get()))); - return Status{}; } @@ -170,7 +169,7 @@ void CLL2NormalizeLayerKernel::run(const Window &window, cl::CommandQueue &queue Window window_sum(window); - switch(_actual_axis) + switch (_actual_axis) { case 0: { @@ -184,8 +183,7 @@ void CLL2NormalizeLayerKernel::run(const Window &window, cl::CommandQueue &queue add_2D_tensor_argument(idx, _sum, sum_slice); add_2D_tensor_argument(idx, _output, in_slice); enqueue(queue, *this, in_slice, lws_hint()); - } - while(window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(sum_slice)); + } while (window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(sum_slice)); } break; case 1: @@ -200,8 +198,7 @@ void CLL2NormalizeLayerKernel::run(const Window &window, cl::CommandQueue &queue add_2D_tensor_argument(idx, _sum, sum_slice); add_2D_tensor_argument(idx, _output, in_slice); enqueue(queue, *this, in_slice, lws_hint()); - } - while(window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(sum_slice)); + } while (window.slide_window_slice_2D(in_slice) && window.slide_window_slice_2D(sum_slice)); } break; case 2: @@ -216,8 +213,7 @@ void CLL2NormalizeLayerKernel::run(const Window &window, cl::CommandQueue &queue add_3D_tensor_argument(idx, _sum, sum_slice); add_3D_tensor_argument(idx, _output, in_slice); enqueue(queue, *this, in_slice, lws_hint()); - } - while(window.slide_window_slice_3D(in_slice) && window.slide_window_slice_3D(sum_slice)); + } while (window.slide_window_slice_3D(in_slice) && window.slide_window_slice_3D(sum_slice)); } break; default: |