From 7068f9900d136312318ff430aef588b14e0c87ad Mon Sep 17 00:00:00 2001 From: Anthony Barbier Date: Thu, 26 Oct 2017 15:23:08 +0100 Subject: COMPMID-631: Merge branches/gles_compute branch Last commit: commit b25c5f68042b0c81bf611d59a1bb8535e1c42497 Author: Xinghang Zhou Date: Wed Oct 25 18:48:10 2017 +0800 Synced validation's tolerances of GCSoftmax from cl side Change-Id: Ibe72054205c1c8721845d679a31af7ed0a7c5cf6 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/93283 Reviewed-by: Anthony Barbier Tested-by: Kaizen --- .../GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.cpp | 353 +++++++++++++++++++++ 1 file changed, 353 insertions(+) create mode 100644 src/core/GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.cpp (limited to 'src/core/GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.cpp') diff --git a/src/core/GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.cpp new file mode 100644 index 0000000000..09a0f79ab2 --- /dev/null +++ b/src/core/GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.cpp @@ -0,0 +1,353 @@ +/* + * Copyright (c) 2017 ARM Limited. + * + * SPDX-License-Identifier: MIT + * + * Permission is hereby granted, free of charge, to any person obtaining a copy + * of this software and associated documentation files (the "Software"), to + * deal in the Software without restriction, including without limitation the + * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or + * sell copies of the Software, and to permit persons to whom the Software is + * furnished to do so, subject to the following conditions: + * + * The above copyright notice and this permission notice shall be included in all + * copies or substantial portions of the Software. + * + * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR + * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, + * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE + * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER + * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, + * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE + * SOFTWARE. + */ +#include "arm_compute/core/GLES_COMPUTE/kernels/GCSoftmaxLayerKernel.h" + +#include "arm_compute/core/AccessWindowStatic.h" +#include "arm_compute/core/GLES_COMPUTE/GCHelpers.h" +#include "arm_compute/core/GLES_COMPUTE/GCKernelLibrary.h" +#include "arm_compute/core/GLES_COMPUTE/IGCTensor.h" +#include "arm_compute/core/GLES_COMPUTE/OpenGLES.h" +#include "arm_compute/core/Helpers.h" +#include "arm_compute/core/TensorInfo.h" +#include "arm_compute/core/Utils.h" +#include "arm_compute/core/Validate.h" +#include "arm_compute/core/Window.h" + +#include +#include + +using namespace arm_compute; + +void GCLogits1DMaxKernel::configure(const IGCTensor *input, IGCTensor *output) +{ + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_NULLPTR(output); + + // Softmax across the x dimension + TensorShape output_shape{ input->info()->tensor_shape() }; + output_shape.set(0, 1); + + // Output auto initialization if not yet initialized + auto_init_if_empty(*output->info(), output_shape, 1, input->info()->data_type(), input->info()->fixed_point_position()); + + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DIMENSIONS(output->info()->tensor_shape(), output_shape); + + _input = input; + _output = output; + + // Set build options + std::set build_opts; + std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16"; + build_opts.insert("#define " + dt_name); + build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1)); + build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1)); + build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1)); + build_opts.insert("#define SOFTMAX_LAYER_MAX"); + + // Tell the kernel that the width is not a multiple of 4 + if((input->info()->dimension(0) % 4) != 0) + { + build_opts.insert("#define NON_MULTIPLE_OF_4"); + } + + // Create kernel + _kernel = static_cast(GCKernelLibrary::get().create_kernel("softmax_layer_max", build_opts)); + + _kernel.clear_params(); + + // Set fixed arguments + unsigned int idx = 2 * num_arguments_per_3D_tensor(); //Skip the input and output parameters + _kernel.set_params(idx++, input->info()->dimension(0)); + + // Configure kernel window + // The kernel loops over all elements in steps of 4 + const unsigned int num_elems_processed_per_iteration = ceil_to_multiple(input->info()->dimension(0), 4); + unsigned int num_elems_written_per_iteration = 1; + if(input->info()->data_type() == DataType::F16) + { + num_elems_written_per_iteration = 2; + } + + Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); + AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration); + AccessWindowHorizontal output_access(output->info(), 0, num_elems_written_per_iteration); + + update_window_and_padding(win, input_access, output_access); + + output_access.set_valid_region(win, ValidRegion(Coordinates(), output->info()->tensor_shape())); + + // set shader params binding point + _kernel.set_shader_params_binding_point(0); + + IGCKernel::configure(win); +} + +void GCLogits1DMaxKernel::run(const Window &window) +{ + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); + + Window slice = window.first_slice_window_3D(); + + _kernel.use(); + + do + { + unsigned int idx1 = 0; + switch(_input->info()->data_type()) + { + case DataType::F16: + add_3D_tensor_argument(idx1, _input, BufferParam(1, 2), slice); + add_3D_tensor_argument(idx1, _output, BufferParam(2, 2), slice); + break; + + case DataType::F32: + add_3D_tensor_argument(idx1, _input, BufferParam(1, 2), slice); + add_3D_tensor_argument(idx1, _output, BufferParam(2, 2), slice); + break; + + default: + ARM_COMPUTE_ERROR("Current data type is mot supported"); + break; + } + + _kernel.update_shader_params(); + enqueue(*this, slice); + } + while(window.slide_window_slice_3D(slice)); +} + +GCLogits1DShiftExpSumKernel::GCLogits1DShiftExpSumKernel() + : _input(nullptr), _max(nullptr), _output(nullptr), _sum(nullptr) +{ +} + +void GCLogits1DShiftExpSumKernel::configure(const IGCTensor *input, const IGCTensor *max, IGCTensor *output, IGCTensor *sum) +{ + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_NULLPTR(max, sum, output); + + // Output auto initialization if not yet initialized + auto_init_if_empty(*sum->info(), max->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); + auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); + + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, output, max, sum); + ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); + ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(max, sum); + + _input = input; + _max = max; + _output = output; + _sum = sum; + + // Set build options + std::set build_opts; + std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16"; + build_opts.insert("#define " + dt_name); + build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1)); + build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1)); + build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1)); + build_opts.insert("#define SOFTMAX_LAYER_SHIFT_EXP_SUM"); + + // Tell the kernel that the width is not a multiple of 4 + if((input->info()->dimension(0) % 4) != 0) + { + build_opts.insert("#define NON_MULTIPLE_OF_4"); + } + + // Create kernel + _kernel = static_cast(GCKernelLibrary::get().create_kernel("softmax_layer_shift_exp_sum", build_opts)); + + _kernel.clear_params(); + + // Set fixed arguments + unsigned int idx = 4 * num_arguments_per_3D_tensor(); //Skip the input and output parameters + _kernel.set_params(idx++, input->info()->dimension(0)); + + // Configure window + // The kernel loops over all elements in steps of 4 + const unsigned int num_elems_processed_per_iteration = ceil_to_multiple(input->info()->dimension(0), 4); + unsigned int num_elems_written_per_iteration = 1; + if(input->info()->data_type() == DataType::F16) + { + num_elems_written_per_iteration = 2; + } + + Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); + + AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration); + AccessWindowHorizontal max_access(max->info(), 0, num_elems_written_per_iteration); + AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); + AccessWindowHorizontal sum_access(sum->info(), 0, num_elems_written_per_iteration); + + update_window_and_padding(win, input_access, max_access, output_access, sum_access); + + output_access.set_valid_region(win, input->info()->valid_region()); + sum_access.set_valid_region(win, ValidRegion(Coordinates(), sum->info()->tensor_shape())); + + // set shader params binding point + _kernel.set_shader_params_binding_point(0); + + IGCKernel::configure(win); +} + +void GCLogits1DShiftExpSumKernel::run(const Window &window) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); + + Window window_collapsed = window.collapse_if_possible(IGCKernel::window(), Window::DimZ); + Window slice = window_collapsed.first_slice_window_3D(); + + _kernel.use(); + + do + { + unsigned int idx = 0; + switch(_input->info()->data_type()) + { + case DataType::F16: + add_3D_tensor_argument(idx, _input, BufferParam(1, 2), slice); + add_3D_tensor_argument(idx, _max, BufferParam(2, 2), slice); + add_3D_tensor_argument(idx, _output, BufferParam(3, 2), slice); + add_3D_tensor_argument(idx, _sum, BufferParam(4, 2), slice); + break; + + case DataType::F32: + add_3D_tensor_argument(idx, _input, BufferParam(1, 2), slice); + add_3D_tensor_argument(idx, _max, BufferParam(2, 2), slice); + add_3D_tensor_argument(idx, _output, BufferParam(3, 2), slice); + add_3D_tensor_argument(idx, _sum, BufferParam(4, 2), slice); + break; + + default: + ARM_COMPUTE_ERROR("Current data type is mot supported"); + break; + } + + _kernel.update_shader_params(); + enqueue(*this, slice); + } + while(window_collapsed.slide_window_slice_3D(slice)); +} + +GCLogits1DNormKernel::GCLogits1DNormKernel() + : _input(nullptr), _sum(nullptr), _output(nullptr) +{ +} + +void GCLogits1DNormKernel::configure(const IGCTensor *input, const IGCTensor *sum, IGCTensor *output) +{ + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input, 1, DataType::F16, DataType::F32); + ARM_COMPUTE_ERROR_ON_NULLPTR(sum, output); + + // Output auto initialization if not yet initialized + auto_init_if_empty(*output->info(), input->info()->tensor_shape(), 1, input->info()->data_type(), input->info()->fixed_point_position()); + + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input, sum, output); + ARM_COMPUTE_ERROR_ON_MISMATCHING_FIXED_POINT_POSITION(input, sum, output); + ARM_COMPUTE_ERROR_ON_MISMATCHING_SHAPES(input, output); + + _input = input; + _sum = sum; + _output = output; + + // Set build options + std::set build_opts; + std::string dt_name = (input->info()->data_type() == DataType::F32) ? "DATA_TYPE_FP32" : "DATA_TYPE_FP16"; + build_opts.insert("#define " + dt_name); + build_opts.emplace("#define LOCAL_SIZE_X " + support::cpp11::to_string(1)); + build_opts.emplace("#define LOCAL_SIZE_Y " + support::cpp11::to_string(1)); + build_opts.emplace("#define LOCAL_SIZE_Z " + support::cpp11::to_string(1)); + build_opts.insert("#define SOFTMAX_LAYER_NORM"); + + // Create kernel + _kernel = static_cast(GCKernelLibrary::get().create_kernel("softmax_layer_norm", build_opts)); + + // Configure window + constexpr unsigned int num_elems_processed_per_iteration = 4; + unsigned int num_elems_written_per_iteration = 1; + if(input->info()->data_type() == DataType::F16) + { + num_elems_written_per_iteration = 2; + } + + Window win = calculate_max_window(*input->info(), Steps(num_elems_processed_per_iteration)); + + AccessWindowHorizontal input_access(input->info(), 0, num_elems_processed_per_iteration); + AccessWindowStatic sum_access(sum->info(), 0, 0, num_elems_written_per_iteration, sum->info()->dimension(1)); + AccessWindowHorizontal output_access(output->info(), 0, num_elems_processed_per_iteration); + + update_window_and_padding(win, input_access, sum_access, output_access); + + output_access.set_valid_region(win, input->info()->valid_region()); + + _kernel.clear_params(); + + // set shader params binding point + _kernel.set_shader_params_binding_point(0); + + IGCKernel::configure(win); +} + +void GCLogits1DNormKernel::run(const Window &window) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IKernel::window(), window); + + Window window_collapsed = window.collapse_if_possible(IGCKernel::window(), Window::DimZ); + Window slice = window_collapsed.first_slice_window_3D(); + + _kernel.use(); + + do + { + Window sum_slice = slice; + sum_slice.set(Window::DimX, Window::Dimension(0, 1, 1)); + + unsigned int idx1 = 0; + switch(_input->info()->data_type()) + { + case DataType::F16: + add_3D_tensor_argument(idx1, _input, BufferParam(1, 2), slice); + add_3D_tensor_argument(idx1, _sum, BufferParam(2, 2), slice); + add_3D_tensor_argument(idx1, _output, BufferParam(3, 2), slice); + break; + + case DataType::F32: + add_3D_tensor_argument(idx1, _input, BufferParam(1, 2), slice); + add_3D_tensor_argument(idx1, _sum, BufferParam(2, 2), slice); + add_3D_tensor_argument(idx1, _output, BufferParam(3, 2), slice); + break; + + default: + ARM_COMPUTE_ERROR("Current data type is mot supported"); + break; + } + + _kernel.update_shader_params(); + enqueue(*this, slice); + } + while(window_collapsed.slide_window_slice_3D(slice)); +} -- cgit v1.2.1