Gian Marco | 05288a2 | 2017-11-21 10:57:50 +0000 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 ARM Limited. |
| 3 | * |
| 4 | * SPDX-License-Identifier: MIT |
| 5 | * |
| 6 | * Permission is hereby granted, free of charge, to any person obtaining a copy |
| 7 | * of this software and associated documentation files (the "Software"), to |
| 8 | * deal in the Software without restriction, including without limitation the |
| 9 | * rights to use, copy, modify, merge, publish, distribute, sublicense, and/or |
| 10 | * sell copies of the Software, and to permit persons to whom the Software is |
| 11 | * furnished to do so, subject to the following conditions: |
| 12 | * |
| 13 | * The above copyright notice and this permission notice shall be included in all |
| 14 | * copies or substantial portions of the Software. |
| 15 | * |
| 16 | * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR |
| 17 | * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, |
| 18 | * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE |
| 19 | * AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER |
| 20 | * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, |
| 21 | * OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE |
| 22 | * SOFTWARE. |
| 23 | */ |
| 24 | #ifndef __ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H__ |
| 25 | #define __ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H__ |
| 26 | |
| 27 | #include "arm_compute/runtime/CL/ICLSimpleFunction.h" |
| 28 | |
| 29 | /** This file contains all available output stages for GEMMLowp on OpenCL. |
| 30 | * |
| 31 | * In gemmlowp, the "output stage" is the process that takes a final int32 accumulator value (the output of @ref CLGEMMLowpMatrixMultiplyCore), |
| 32 | * and processes it to obtain the final ASYMM8 value. |
| 33 | * |
| 34 | * More information about the GEMMLowp output stage can be found at https://github.com/google/gemmlowp/blob/master/doc/output.md |
| 35 | */ |
| 36 | |
| 37 | namespace arm_compute |
| 38 | { |
| 39 | class ITensor; |
| 40 | |
| 41 | /** Basic function to execute CLGEMMLowpQuantizeDownInt32ToUint8Scale on OpenCL. |
| 42 | * |
| 43 | * CLGEMMLowpQuantizeDownInt32ToUint8Scale depends on 3 parameters: result_offset, result_mult_int, result_shift |
| 44 | * The final result is: |
| 45 | * |
| 46 | * ((input[i][k] + result_offset) * result_mult_int) >> result_shift |
| 47 | * |
| 48 | * In case the bias tensor is provided, the final result is: |
| 49 | * |
| 50 | * ((input[i][k] + result_offset) * result_mult_int + bias[k]) >> result_shift |
| 51 | * |
| 52 | * This function calls the following OpenCL kernels: |
| 53 | * |
| 54 | * -# @ref CLGEMMLowpQuantizeDownInt32ToUint8ScaleKernel |
| 55 | * |
| 56 | * @note The function accepts also 2 optional input arguments (min and max) which can be used to implement "rectified linear unit" activation functions |
| 57 | * before the result is shifted right by result_shift |
| 58 | */ |
| 59 | class CLGEMMLowpQuantizeDownInt32ToUint8Scale : public ICLSimpleFunction |
| 60 | { |
| 61 | public: |
| 62 | /** Initialise the kernel's inputs, output |
| 63 | * |
| 64 | * @param[in] input Input tensor. It is the output of @ref CLGEMMLowpMatrixMultiplyCore function. Data type supported: S32 |
| 65 | * @param[in] bias Biases tensor. Only shared biases supported and it can be a nullptr if the addition of biases is not required. |
| 66 | * Biases are 1D tensor with dimensions [OFM]. Data type supported: Same as @p input. |
| 67 | * @param[out] output Output tensor. Data type supported: Data type supported: QASYMM8 |
| 68 | * @param[in] result_offset Offset to be added to each element of the input matrix |
| 69 | * @param[in] result_mult_int Value to be multiplied to each element of the input matrix when once the result_offset has been add |
| 70 | * @param[in] result_shift Number of bits to shift right the result before converting back to QASYMM8 |
| 71 | * @param[in] min (Optional) Min value used to saturate down the output result before converting back to QASYMM8 |
| 72 | * @param[in] max (Optional) Max value used to saturate up the output result before converting back to QASYMM8, |
| 73 | * Along with @p min, this value can be used to implement "rectified linear unit" activation functions |
| 74 | */ |
| 75 | void configure(const ICLTensor *input, const ICLTensor *bias, ICLTensor *output, int result_offset, int result_mult_int, int result_shift, int min = 0, int max = 0); |
| 76 | }; |
| 77 | } |
| 78 | #endif /*__ARM_COMPUTE_CLGEMMLOWPOUTPUTSTAGE_H__ */ |