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 --- .../kernels/GCGEMMMatrixMultiplyKernel.cpp | 210 +++++++++++++++++++++ 1 file changed, 210 insertions(+) create mode 100644 src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp (limited to 'src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp') diff --git a/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp b/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp new file mode 100644 index 0000000000..ea9b3874b2 --- /dev/null +++ b/src/core/GLES_COMPUTE/kernels/GCGEMMMatrixMultiplyKernel.cpp @@ -0,0 +1,210 @@ +/* + * 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/GCGEMMMatrixMultiplyKernel.h" + +#include "arm_compute/core/AccessWindowStatic.h" +#include "arm_compute/core/AccessWindowTranspose.h" +#include "arm_compute/core/Error.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/Types.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; + +GCGEMMMatrixMultiplyKernel::GCGEMMMatrixMultiplyKernel() + : _input0(nullptr), _input1(nullptr), _output(nullptr) +{ +} + +void GCGEMMMatrixMultiplyKernel::configure(const IGCTensor *input0, const IGCTensor *input1, IGCTensor *output, float alpha, bool is_interleaved_transposed) +{ + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(input0, 1, DataType::F32, DataType::F16); + ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(input0, input1, output); + + if(!is_interleaved_transposed) + { + ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1)); + } + + _input0 = input0; + _input1 = input1; + _output = output; + + std::set build_opts; + Window win; + + 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.emplace("#define COLS_A " + support::cpp11::to_string(input0->info()->dimension(0))); + build_opts.emplace("#define COLS_B " + support::cpp11::to_string(input1->info()->dimension(0))); + build_opts.emplace("#define ALPHA " + float_to_string_with_full_precision(alpha)); + + // Check if the output tensor is a vector. If so,the kernel runs the vector-matrix multiplication + if(is_interleaved_transposed) + { + switch(input0->info()->data_type()) + { + case DataType::F16: + build_opts.emplace("#define DATA_TYPE_FP16"); + break; + + case DataType::F32: + build_opts.emplace("#define DATA_TYPE_FP32"); + break; + + default: + ARM_COMPUTE_ERROR("Current data type is not supported"); + break; + } + + build_opts.emplace("#define GEMM_MM_INTERLEAVED_TRANSPOSED"); + + // Create kernel + _kernel = GCKernelLibrary::get().create_kernel(("gemm_mm_interleaved_transposed"), build_opts); + + // Configure window kernel + const unsigned int num_elems_processed_per_iteration_x = max_gc_vector_width / data_size_from_type(input0->info()->data_type()); + constexpr unsigned int num_elems_processed_per_iteration_y = 4; + + win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); + + AccessWindowRectangle input0_access(input0->info(), 0, 0, num_elems_processed_per_iteration_y, 1, 1.f, 0.25f); + AccessWindowTranspose input1_access(input1->info(), 0, 0, num_elems_processed_per_iteration_x, 1, 0.f, 0.25f); + AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); + + update_window_and_padding(win, input0_access, input1_access, output_access); + + output_access.set_valid_region(win, ValidRegion(Coordinates(0, 0), output->info()->tensor_shape())); + } + else + { + ARM_COMPUTE_ERROR_ON(input0->info()->dimension(0) != input1->info()->dimension(1)); + + // Special case for 1xN, 2xN, 3xN and 4xN input0 tensor + unsigned int num_elems_processed_per_iteration_x; + unsigned int num_elems_processed_per_iteration_y; + + switch(input0->info()->data_type()) + { + case DataType::F16: + num_elems_processed_per_iteration_x = 4; + num_elems_processed_per_iteration_y = 1; + build_opts.emplace("#define DATA_TYPE_FP16"); + break; + + case DataType::F32: + num_elems_processed_per_iteration_x = max_gc_vector_width / data_size_from_type(input0->info()->data_type()); + num_elems_processed_per_iteration_y = std::min(static_cast(output->info()->dimension(1)), 4); + build_opts.emplace("#define DATA_TYPE_FP32"); + break; + + default: + ARM_COMPUTE_ERROR("Current data type is not supported"); + break; + } + + build_opts.emplace("#define GEMM_MM_FLOATING_POINT"); + build_opts.emplace("#define NUM_ELEMS_PROCESSED_PER_THREAD_X " + support::cpp11::to_string(num_elems_processed_per_iteration_x)); + build_opts.emplace("#define NUM_ELEMS_PROCESSED_PER_THREAD_Y " + support::cpp11::to_string(num_elems_processed_per_iteration_y)); + + // Create kernel + _kernel = GCKernelLibrary::get().create_kernel("gemm_mm_floating_point", build_opts); + + win = calculate_max_window(*output->info(), Steps(num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y)); + + AccessWindowStatic input0_access(input0->info(), 0, 0, ceil_to_multiple(input0->info()->dimension(0), num_elems_processed_per_iteration_x), ceil_to_multiple(input0->info()->dimension(1), + num_elems_processed_per_iteration_y)); + AccessWindowStatic input1_access(input1->info(), 0, 0, ceil_to_multiple(input1->info()->dimension(0), num_elems_processed_per_iteration_x), input1->info()->dimension(1)); + AccessWindowRectangle output_access(output->info(), 0, 0, num_elems_processed_per_iteration_x, num_elems_processed_per_iteration_y); + + update_window_and_padding(win, input0_access, input1_access, output_access); + + Coordinates coord; + coord.set_num_dimensions(output->info()->num_dimensions()); + output_access.set_valid_region(win, ValidRegion(coord, output->info()->tensor_shape())); + } + + _kernel.clear_params(); + _kernel.set_shader_params_binding_point(0); + IGCKernel::configure(win); +} + +void GCGEMMMatrixMultiplyKernel::run(const Window &window) +{ + ARM_COMPUTE_ERROR_ON_UNCONFIGURED_KERNEL(this); + ARM_COMPUTE_ERROR_ON_INVALID_SUBWINDOW(IGCKernel::window(), window); + + _kernel.use(); + + Window slice = window.first_slice_window_2D(); + Window slice_matrix_b = slice; + + slice_matrix_b.set(Window::DimX, Window::Dimension(0, 1, 1)); + slice_matrix_b.set(Window::DimY, Window::Dimension(0, 1, 1)); + + do + { + Window slice_b = slice; + // Don't slice matrix B along the z dimension if matrix B has just 2 dimensions and matrix A more than 2 + // This scenario can happen when the the matrix multiplication is used to perform a convolution operation + if(_input1->info()->num_dimensions() < 3) + { + slice_b = slice_matrix_b; + } + + unsigned int idx = 0; + switch(_input0->info()->data_type()) + { + case DataType::F16: + add_2D_tensor_argument(idx, _input0, BufferParam(1, 2), slice); + add_2D_tensor_argument(idx, _input1, BufferParam(2, 3), slice_b); + add_2D_tensor_argument(idx, _output, BufferParam(3, 3), slice); + break; + + case DataType::F32: + add_2D_tensor_argument(idx, _input0, BufferParam(1, 2), slice); + add_2D_tensor_argument(idx, _input1, BufferParam(2, 2), slice_b); + add_2D_tensor_argument(idx, _output, BufferParam(3, 2), slice); + break; + + default: + ARM_COMPUTE_ERROR("Current data type is not supported"); + break; + } + + _kernel.update_shader_params(); + enqueue(*this, slice); + } + while(window.slide_window_slice_2D(slice)); +} -- cgit v1.2.1