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 --- arm_compute/runtime/NEON/NEFunctions.h | 2 +- arm_compute/runtime/NEON/functions/NEGEMMLowp.h | 94 ---------------------- .../NEON/functions/NEGEMMLowpMatrixMultiplyCore.h | 33 +++++--- .../runtime/NEON/functions/NEGEMMLowpOutputStage.h | 69 ++++++++++++++++ 4 files changed, 92 insertions(+), 106 deletions(-) delete mode 100644 arm_compute/runtime/NEON/functions/NEGEMMLowp.h create mode 100644 arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h (limited to 'arm_compute/runtime/NEON') diff --git a/arm_compute/runtime/NEON/NEFunctions.h b/arm_compute/runtime/NEON/NEFunctions.h index 563ade288a..118603b20b 100644 --- a/arm_compute/runtime/NEON/NEFunctions.h +++ b/arm_compute/runtime/NEON/NEFunctions.h @@ -60,8 +60,8 @@ #include "arm_compute/runtime/NEON/functions/NEFullyConnectedLayer.h" #include "arm_compute/runtime/NEON/functions/NEGEMM.h" #include "arm_compute/runtime/NEON/functions/NEGEMMInterleave4x4.h" -#include "arm_compute/runtime/NEON/functions/NEGEMMLowp.h" #include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h" +#include "arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h" #include "arm_compute/runtime/NEON/functions/NEGEMMTranspose1xW.h" #include "arm_compute/runtime/NEON/functions/NEGaussian3x3.h" #include "arm_compute/runtime/NEON/functions/NEGaussian5x5.h" diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowp.h b/arm_compute/runtime/NEON/functions/NEGEMMLowp.h deleted file mode 100644 index 59c919e161..0000000000 --- a/arm_compute/runtime/NEON/functions/NEGEMMLowp.h +++ /dev/null @@ -1,94 +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. - */ -#ifndef __ARM_COMPUTE_NEGEMMLOWP_H__ -#define __ARM_COMPUTE_NEGEMMLOWP_H__ - -#include "arm_compute/core/NEON/INEKernel.h" -#include "arm_compute/core/NEON/kernels/NEGEMMLowpFinalizeKernel.h" -#include "arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h" -#include "arm_compute/runtime/IFunction.h" -#include "arm_compute/runtime/IMemoryManager.h" -#include "arm_compute/runtime/MemoryGroup.h" -#include "arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h" -#include "arm_compute/runtime/Tensor.h" - -#include - -namespace arm_compute -{ -class ITensor; - -/** Basic function to execute GEMMLowp on NEON. This function calls the following NEON kernels/function: - * - * -# @ref NEGEMMLowpMatrixAReductionKernel - * -# @ref NEGEMMLowpMatrixBReductionKernel - * -# @ref NEGEMMLowpMatrixMultiplyCore - * -# @ref NEGEMMLowpFinalizeKernel - * -*/ -class NEGEMMLowp : public IFunction -{ -public: - /** Constructor */ - NEGEMMLowp(std::shared_ptr memory_manager = nullptr); - /** Initialise the kernel's inputs, output - * - * @note GEMM_LOWP: low precision GEMM kernel - * This kernel performs the following computations: - * - * -# Convert a values from int8 to int32 and add a_offset to each of them. - * -# Convert b values from int8 to int32 and add b_offset to each of them. - * -# Compute the int32 matrix product of the resulting a * b. - * -# Add output_offset to each entry of the result. - * -# Multiply each entry of the result and round to the nearest integer - * -# Clamp the resulting int32 values to the [0..255] range and cast to int8. - * - * @param[in] a First input tensor (Matrix A). Data type supported: S8. - * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a - * @param[out] output Output tensor. Data type supported: same as @p a. - * @param[in] a_offset Offset to be added to each element of the matrix A. - * @param[in] b_offset Offset to be added to each element of the matrix B. - * @param[in] c_offset Offset to be added to each element of the output matrix - * @param[in] output_mult_int Value to be multiplied to each element of the output matrix - * @param[in] shift Number of bits to shift right the result. - */ - void 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); - - // Inherited methods overridden: - void run() override; - -private: - MemoryGroup _memory_group; - NEGEMMLowpMatrixMultiplyCore _mm_func; - NEGEMMLowpMatrixAReductionKernel _mtx_a_reduction_kernel; - NEGEMMLowpMatrixBReductionKernel _mtx_b_reduction_kernel; - NEGEMMLowpFinalizeKernel _finalize_kernel; - Tensor _vector_sum_col; - Tensor _vector_sum_row; - Tensor _mm_output; - int32_t _a_offset; - int32_t _b_offset; -}; -} -#endif /*__ARM_COMPUTE_NEGEMMLOWP_H__ */ diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h index c81a432295..0c441df4b9 100644 --- a/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h +++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpMatrixMultiplyCore.h @@ -25,6 +25,8 @@ #define __ARM_COMPUTE_NEGEMMLOWPMATRIXMULTIPLYCORE_H__ #include "arm_compute/core/NEON/INEKernel.h" +#include "arm_compute/core/NEON/kernels/NEGEMMLowpOffsetContributionKernel.h" +#include "arm_compute/core/NEON/kernels/NEGEMMLowpReductionKernel.h" #include "arm_compute/runtime/IFunction.h" #include "arm_compute/runtime/IMemoryManager.h" #include "arm_compute/runtime/MemoryGroup.h" @@ -41,11 +43,13 @@ class ITensor; * -# @ref NEGEMMInterleave4x4Kernel * -# @ref NEGEMMTranspose1xWKernel * -# @ref NEGEMMLowpMatrixMultiplyKernel + * -# @ref NEGEMMLowpOffsetContributionKernel * * otherwise if the DOT product instruction is available: * * -# @ref NEGEMMInterleaveBlockedKernel * -# @ref NEGEMMLowpAArch64V8P4Kernel + * -# @ref NEGEMMLowpOffsetContributionKernel * */ class NEGEMMLowpMatrixMultiplyCore : public IFunction @@ -58,11 +62,11 @@ public: * @note GEMM_LOWP: low precision GEMM kernel * This kernel performs the following computations: * - * -# Convert a values from uint8 to int32 - * -# Convert b values from uint8 to int32 - * -# Compute the int32 matrix product of the resulting a * b. + * -# Convert a values from QASYMM8 to int32 and add a_offset to each of them. + * -# Convert b values from QASYMM8 to int32 add b_offset to each of them. + * -# Compute the matrix product of the resulting a * b in int32. * - * @param[in] a First input tensor (Matrix A). Data type supported: U8. + * @param[in] a First input tensor (Matrix A). Data type supported: QASYMM8. * @param[in] b Second input tensor (Matrix B). Data type supported: same as @p a * @param[out] output Output tensor. Data type supported: Data type supported: S32 */ @@ -72,13 +76,20 @@ public: void run() override; private: - MemoryGroup _memory_group; - std::unique_ptr _mm_kernel; - std::unique_ptr _mtx_a_reshape_kernel; - std::unique_ptr _mtx_b_reshape_kernel; - Tensor _tmp_a; - Tensor _tmp_b; - Tensor _workspace; + MemoryGroup _memory_group; + std::unique_ptr _mm_kernel; + std::unique_ptr _mtx_a_reshape_kernel; + std::unique_ptr _mtx_b_reshape_kernel; + NEGEMMLowpMatrixAReductionKernel _mtx_a_reduction_kernel; + NEGEMMLowpMatrixBReductionKernel _mtx_b_reduction_kernel; + NEGEMMLowpOffsetContributionKernel _offset_contribution_kernel; + Tensor _vector_sum_col; + Tensor _vector_sum_row; + Tensor _tmp_a; + Tensor _tmp_b; + Tensor _workspace; + int32_t _a_offset; + int32_t _b_offset; }; } #endif /*__ARM_COMPUTE_NEGEMMLOWPMATRIXMULTIPLYCORE_H__ */ diff --git a/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h new file mode 100644 index 0000000000..8557ef42e1 --- /dev/null +++ b/arm_compute/runtime/NEON/functions/NEGEMMLowpOutputStage.h @@ -0,0 +1,69 @@ +/* + * 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. + */ +#ifndef __ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H__ +#define __ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H__ + +#include "arm_compute/runtime/NEON/INESimpleFunction.h" + +/** This file contains all available output stages for GEMMLowp on NEON. + * + * In gemmlowp, the "output stage" is the process that takes a final int32 accumulator value (the output of @ref NEGEMMLowpMatrixMultiplyCore), + * and processes it to obtain the final ASYMM8 value. + * + * More information about the GEMMLowp output stage can be found at https://github.com/google/gemmlowp/blob/master/doc/output.md + */ + +namespace arm_compute +{ +class ITensor; + +/** Basic function to execute NEGEMMLowpQuantizeDownInt32ToUint8Scale on NEON. + * + * NEGEMMLowpQuantizeDownInt32ToUint8Scale depends on 3 parameters: result_offset, result_mult_int, result_shift + * The final result is: + * + * ((input[i][k] + result_offset) * result_mult_int + rounding) >> result_shift + * + * where rounding = (result_shift < 1) ? 0 : (1 << (result_shift - 1)) + * + * This function calls the following NEON kernels: + * + * -# @ref NEGEMMLowpQuantizeDownInt32ToUint8ScaleKernel + * +*/ +class NEGEMMLowpQuantizeDownInt32ToUint8Scale : public INESimpleFunction +{ +public: + /** Initialise the kernel's inputs, output + * + * @param[in] input Input tensor. It is the output of @ref NEGEMMLowpMatrixMultiplyCore function. Data type supported: S32 + * @param[out] output Output tensor. Data type supported: Data type supported: QASYMM8 + * @param[in] result_offset Offset to be added to each element of the input matrix + * @param[in] result_mult_int Value to be multiplied to each element of the input matrix when once the result_offset has been add + * @param[in] result_shift Number of bits to shift right the result before converting back to QASYMM8 + */ + void configure(const ITensor *input, ITensor *output, int result_offset, int result_mult_int, int result_shift); +}; +} +#endif /*__ARM_COMPUTE_NEGEMMLOWPOUTPUTSTAGE_H__ */ \ No newline at end of file -- cgit v1.2.1