From e75a02b60736f37c34388c23c0ccee230f65da59 Mon Sep 17 00:00:00 2001 From: Gian Marco Date: Wed, 8 Nov 2017 12:24:09 +0000 Subject: COMPMID-675 - Reworked NEGEMMLowp interface/function The new interface makes NEGEMMLowp able to work with ASYMM8 data types. Implemented 2 new functions: - NEGEMMLowpMatrixMultiplyCore - NEGEMMLowpOutputStage These functions should make the integration in android NN doable For more information about GEMMLowp: https://github.com/google/gemmlowp/blob/master/doc/low-precision.md Change-Id: Ie2c775f45234f68ca53dba644b3a912b997fd890 Reviewed-on: http://mpd-gerrit.cambridge.arm.com/95504 Tested-by: Kaizen Reviewed-by: Pablo Tello --- src/runtime/NEON/functions/NEGEMMLowp.cpp | 134 --------------------- .../functions/NEGEMMLowpMatrixMultiplyCore.cpp | 84 ++++++++++++- .../NEON/functions/NEGEMMLowpOutputStage.cpp | 37 ++++++ 3 files changed, 116 insertions(+), 139 deletions(-) delete mode 100644 src/runtime/NEON/functions/NEGEMMLowp.cpp create mode 100644 src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp (limited to 'src/runtime/NEON/functions') diff --git a/src/runtime/NEON/functions/NEGEMMLowp.cpp b/src/runtime/NEON/functions/NEGEMMLowp.cpp deleted file mode 100644 index 90bc6a205b..0000000000 --- a/src/runtime/NEON/functions/NEGEMMLowp.cpp +++ /dev/null @@ -1,134 +0,0 @@ -/* - * 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/runtime/NEON/functions/NEGEMMLowp.h" - -#include "arm_compute/core/Error.h" -#include "arm_compute/core/Helpers.h" -#include "arm_compute/core/ITensor.h" -#include "arm_compute/core/NEON/kernels/arm64/NEGEMMLowpAArch64V8P4Kernel.h" -#include "arm_compute/core/TensorInfo.h" -#include "arm_compute/core/Types.h" -#include "arm_compute/core/Validate.h" -#include "arm_compute/runtime/NEON/NEScheduler.h" -#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h" -#include "arm_compute/runtime/TensorAllocator.h" -#include "support/ToolchainSupport.h" - -using namespace arm_compute; - -NEGEMMLowp::NEGEMMLowp(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _mm_func(), _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(), _finalize_kernel(), _vector_sum_col(), _vector_sum_row(), _mm_output(), _a_offset(0), - _b_offset(0) -{ -} - -void NEGEMMLowp::configure(const ITensor *a, const ITensor *b, ITensor *output, int32_t a_offset, int32_t b_offset, int32_t c_offset, int32_t output_mult_int, int32_t shift) -{ - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN((a), 1, DataType::S8); - ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b, output); - ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(0) != (b)->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); - ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(1) != (output)->info()->dimension(1), "The output matrix must have the same number of rows as the matrix A"); - ARM_COMPUTE_ERROR_ON_MSG((b)->info()->dimension(0) != (output)->info()->dimension(0), "The output matrix must have the same number of columns as the matrix B"); - - _a_offset = a_offset; - _b_offset = b_offset; - - // Initialize matrix multiply output tensor - const TensorShape &shape_mm_output = output->info()->tensor_shape(); - TensorInfo info_mm_output(shape_mm_output, 1, DataType::S32); - _mm_output.allocator()->init(info_mm_output); - _memory_group.manage(&_mm_output); - - // Initialize Matrix B reduction kernel only if _a_offset is not equal to 0 - if(_a_offset != 0) - { - TensorShape shape_vector_sum_col = b->info()->tensor_shape(); - shape_vector_sum_col.remove_dimension(1); - TensorInfo info_vector_sum_col(shape_vector_sum_col, 1, DataType::S32); - _vector_sum_col.allocator()->init(info_vector_sum_col); - _memory_group.manage(&_vector_sum_col); - - // Configure Matrix B reduction kernel - _mtx_b_reduction_kernel.configure(b, &_vector_sum_col, a->info()->dimension(0), false); - } - - // Initialize Matrix A reduction kernel only if _b_offset is not equal to 0 - if(_b_offset != 0) - { - TensorShape shape_vector_sum_row = a->info()->tensor_shape(); - shape_vector_sum_row.set(Window::DimX, a->info()->dimension(1)); - shape_vector_sum_row.remove_dimension(1); - TensorInfo info_vector_sum_row(shape_vector_sum_row, 1, DataType::S32); - _vector_sum_row.allocator()->init(info_vector_sum_row); - _memory_group.manage(&_vector_sum_row); - - // Configure Matrix A reduction kernel - _mtx_a_reduction_kernel.configure(a, &_vector_sum_row, a->info()->dimension(0), false); - } - - // Configure matrix multiply function - _mm_func.configure(a, b, &_mm_output); - - // Configure finalize kernel - _finalize_kernel.configure(_a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, &_mm_output, output, a->info()->dimension(0), a_offset, b_offset, c_offset, - output_mult_int, shift); - - // Allocate tensors - _mm_output.allocator()->allocate(); - - if(_a_offset != 0) - { - _vector_sum_col.allocator()->allocate(); - } - - if(_b_offset != 0) - { - _vector_sum_row.allocator()->allocate(); - } -} - -void NEGEMMLowp::run() -{ - _memory_group.acquire(); - - // Run matrix A reduction kernel only if _b_offset is not equal to 0 - if(_b_offset != 0) - { - NEScheduler::get().schedule(&_mtx_a_reduction_kernel, Window::DimX); - } - - // Run matrix B reduction kernel only if _a_offset is not equal to 0 - if(_a_offset != 0) - { - NEScheduler::get().schedule(&_mtx_b_reduction_kernel, Window::DimX); - } - - // Run matrix multiply core function - _mm_func.run(); - - // Run finalise kernel - NEScheduler::get().schedule(&_finalize_kernel, Window::DimY); - - _memory_group.release(); -} diff --git a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp index 29104cc378..929ee41220 100644 --- a/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp +++ b/src/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.cpp @@ -47,19 +47,25 @@ namespace arm_compute using namespace arm_compute; NEGEMMLowpMatrixMultiplyCore::NEGEMMLowpMatrixMultiplyCore(std::shared_ptr memory_manager) - : _memory_group(std::move(memory_manager)), _mm_kernel(nullptr), _mtx_a_reshape_kernel(nullptr), _mtx_b_reshape_kernel(nullptr), _tmp_a(), _tmp_b(), _workspace() + : _memory_group(std::move(memory_manager)), _mm_kernel(nullptr), _mtx_a_reshape_kernel(nullptr), _mtx_b_reshape_kernel(nullptr), _mtx_a_reduction_kernel(), _mtx_b_reduction_kernel(), + _offset_contribution_kernel(), _vector_sum_col(), _vector_sum_row(), _tmp_a(), _tmp_b(), _workspace(), _a_offset(0), _b_offset(0) { } void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, ITensor *output) { - ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::S8); + ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(a, 1, DataType::QASYMM8); ARM_COMPUTE_ERROR_ON_DATA_TYPE_CHANNEL_NOT_IN(output, 1, DataType::S32); ARM_COMPUTE_ERROR_ON_MISMATCHING_DATA_TYPES(a, b); ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(0) != (b)->info()->dimension(1), "The product AB is defined only if the number of columns in A is equal to the number of rows in B"); ARM_COMPUTE_ERROR_ON_MSG((a)->info()->dimension(1) != (output)->info()->dimension(1), "The output matrix must have the same number of rows as the matrix A"); ARM_COMPUTE_ERROR_ON_MSG((b)->info()->dimension(0) != (output)->info()->dimension(0), "The output matrix must have the same number of columns as the matrix B"); + bool dot_product_path = false; + + _a_offset = a->info()->quantization_info().offset; + _b_offset = b->info()->quantization_info().offset; + #ifdef ARM_COMPUTE_AARCH64_V8_2 // Check for DOT product instruction const struct CPUInfo ci = NEScheduler::get().cpu_info(); @@ -67,6 +73,13 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, if(cpu_has_dotprod != 0) { + dot_product_path = true; + + // If the DOT product instruction is available, the computation will be performed in int8_t + // In order to take into account this, we need to subtract -128 from a_offset and b_offset + _a_offset -= 128; + _b_offset -= 128; + // Configure matrix multiply kernel struct CPUInfo ci = NEScheduler::get().cpu_info(); const int M = output->info()->tensor_shape().y(); @@ -77,12 +90,11 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, constexpr size_t alignment = 4096; _workspace.allocator()->init(TensorInfo(TensorShape{ (gemm.get_working_size() + alignment - 1) * NEScheduler::get().num_threads() }, 1, DataType::U8)); _memory_group.manage(&_workspace); + // Configure matrix multiplication kernel auto k = arm_compute::support::cpp14::make_unique(); k->configure(a, b, output, &_workspace, 1.f, 1.f); _mm_kernel = std::move(k); - - _workspace.allocator()->allocate(); } else #endif /* ARM_COMPUTE_AARCH64_V8_2 */ @@ -124,11 +136,58 @@ void NEGEMMLowpMatrixMultiplyCore::configure(const ITensor *a, const ITensor *b, k->configure(&_tmp_a, &_tmp_b, output); _mm_kernel = std::move(k); } + } - // Allocate tensors + // Initialize matrix B reduction kernel only if _a_offset is not equal to 0 + if(_a_offset != 0) + { + TensorShape shape_vector_sum_col = b->info()->tensor_shape(); + shape_vector_sum_col.remove_dimension(1); + TensorInfo info_vector_sum_col(shape_vector_sum_col, 1, DataType::S32); + _vector_sum_col.allocator()->init(info_vector_sum_col); + _memory_group.manage(&_vector_sum_col); + + // Configure Matrix B reduction kernel + _mtx_b_reduction_kernel.configure(b, &_vector_sum_col, a->info()->dimension(0), false); + } + + // Initialize Matrix A reduction kernel only if _b_offset is not equal to 0 + if(_b_offset != 0) + { + TensorShape shape_vector_sum_row = a->info()->tensor_shape(); + shape_vector_sum_row.set(Window::DimX, a->info()->dimension(1)); + shape_vector_sum_row.remove_dimension(1); + TensorInfo info_vector_sum_row(shape_vector_sum_row, 1, DataType::S32); + _vector_sum_row.allocator()->init(info_vector_sum_row); + _memory_group.manage(&_vector_sum_row); + + // Configure matrix A reduction kernel + _mtx_a_reduction_kernel.configure(a, &_vector_sum_row, a->info()->dimension(0), false); + } + + // Configure offset contribution kernel + _offset_contribution_kernel.configure(output, _a_offset == 0 ? nullptr : &_vector_sum_col, _b_offset == 0 ? nullptr : &_vector_sum_row, a->info()->dimension(0), _a_offset, _b_offset); + + // Allocate tensors + if(!dot_product_path) + { _tmp_a.allocator()->allocate(); _tmp_b.allocator()->allocate(); } + else + { + _workspace.allocator()->allocate(); + } + + if(_a_offset != 0) + { + _vector_sum_col.allocator()->allocate(); + } + + if(_b_offset != 0) + { + _vector_sum_row.allocator()->allocate(); + } } void NEGEMMLowpMatrixMultiplyCore::run() @@ -147,5 +206,20 @@ void NEGEMMLowpMatrixMultiplyCore::run() NEScheduler::get().schedule(_mm_kernel.get(), Window::DimY); + // Run matrix A reduction kernel only if _b_offset is not equal to 0 + if(_b_offset != 0) + { + NEScheduler::get().schedule(&_mtx_a_reduction_kernel, Window::DimX); + } + + // Run matrix B reduction kernel only if _a_offset is not equal to 0 + if(_a_offset != 0) + { + NEScheduler::get().schedule(&_mtx_b_reduction_kernel, Window::DimX); + } + + // Run offset contribution kernel + NEScheduler::get().schedule(&_offset_contribution_kernel, Window::DimY); + _memory_group.release(); } diff --git a/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp b/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp new file mode 100644 index 0000000000..d09827f908 --- /dev/null +++ b/src/runtime/NEON/functions/NEGEMMLowpOutputStage.cpp @@ -0,0 +1,37 @@ +/* + * 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/runtime/NEON/functions/NEGEMMLowpOutputStage.h" + +#include "arm_compute/core/ITensor.h" +#include "arm_compute/core/NEON/kernels/NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel.h" +#include "support/ToolchainSupport.h" + +using namespace arm_compute; + +void NEGEMMLowpQuantizeDownInt32ToUint8Scale::configure(const ITensor *input, ITensor *output, int result_offset, int result_mult_int, int result_shift) +{ + auto k = arm_compute::support::cpp14::make_unique(); + k->configure(input, output, result_offset, result_mult_int, result_shift); + _kernel = std::move(k); +} \ No newline at end of file -- cgit v1.2.1