1/* 2 * Copyright (c) 2016-2021 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#include "activation_quant_helpers.h" 25 26#define VEC_FLOAT VEC_DATA_TYPE(float, VEC_SIZE) 27 28#if defined(FLOAT_DOMAIN) 29// Activations performed in the float domain 30 31#include "activation_float_helpers.h" 32 33/** This performs an activation function on quantized inputs with float transformations. 34 * 35 * @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time 36 * 37 * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short 38 * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 39 * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE 40 * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively. 41 * @note Quantization scales of the input/output tensors are passed in with -DS1_VAL= and -DS2_VAL= respectively. 42 * @note Quantization offsets of the input/output tensors are passed in only if asymmetric with -DO1_VAL= and -DO2_VAL= respectively. 43 * @note Quantized value of constant zero should be given as a preprocessor argument using -DCONST_0=value. e.g. -DCONST_0=128. 44 * 45 * @param[in] input_ptr Pointer to the source image. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM16 46 * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) 47 * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) 48 * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) 49 * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) 50 * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) 51 * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) 52 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image 53 * @param[out] output_ptr (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr 54 * @param[in] output_stride_x (Optional) Stride of the destination image in X dimension (in bytes) 55 * @param[in] output_step_x (Optional) output_stride_x * number of elements along X processed per workitem(in bytes) 56 * @param[in] output_stride_y (Optional) Stride of the destination image in Y dimension (in bytes) 57 * @param[in] output_step_y (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes) 58 * @param[in] output_stride_z (Optional) Stride of the source tensor in Z dimension (in bytes) 59 * @param[in] output_step_z (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes) 60 * @param[in] output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image 61 */ 62__kernel void activation_layer_quant_f32( 63 TENSOR3D_DECLARATION(input) 64#ifndef IN_PLACE 65 , 66 TENSOR3D_DECLARATION(output) 67#endif /* not IN_PLACE */ 68) 69{ 70 uint x_offs = max((int)(get_global_id(0) * VEC_SIZE * sizeof(DATA_TYPE) - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE * sizeof(DATA_TYPE)), 0); 71 72 // Get pixels pointer 73 __global uchar *input_addr = input_ptr + input_offset_first_element_in_bytes + x_offs + get_global_id(1) * input_stride_y + get_global_id(2) * input_stride_z; 74#ifdef IN_PLACE 75 __global uchar *output_addr = input_addr; 76#else /* IN_PLACE */ 77 __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs + get_global_id(1) * output_stride_y + get_global_id(2) * output_stride_z; 78#endif /* IN_PLACE */ 79 80 // Load data 81 TYPE data0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr); 82 83 VEC_FLOAT data_flt = CONVERT(data0, VEC_FLOAT); 84#if defined(O1_VAL) 85 data_flt = round(data_flt - (float)O1_VAL) * ((float)S1_VAL); 86#else // defined(O1_VAL) 87 data_flt = round(data_flt) * ((float)S1_VAL); 88#endif // defined(O1_VAL) 89 data_flt = ACTIVATION(ACT, float, VEC_SIZE, data_flt, A_VAL, B_VAL); 90 91#if defined(O2_VAL) 92 data0 = CONVERT_SAT(round(data_flt / ((float)S2_VAL)) + (float)O2_VAL, TYPE); 93#else // defined(O2_VAL) 94 data0 = CONVERT_SAT(round(data_flt / ((float)S2_VAL)), TYPE); 95#endif // defined(O2_VAL) 96 97 // Store result 98 STORE_VECTOR_SELECT(data, DATA_TYPE, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) 99} 100 101#else // defined(FLOAT_DOMAIN) 102// Activations performed in the quantized domain 103 104#if defined(ACT) 105/** This performs an activation function on quantized inputs. 106 * 107 * @note In order to perform the activation function "in-place", the pre-processor -DIN_PLACE must be passed at compile time 108 * 109 * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short 110 * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 111 * @note Leftover vector size has to be passed at compile time using -DVEC_SIZE_LEFTOVER. e.g. -DVEC_SIZE_LEFTOVER=3. It is defined as the remainder between the input's first dimension and VEC_SIZE 112 * @note Activation function should be given as a preprocessor argument using -DACT=name. e.g. -DACT=TANH 113 * @note A, B variables required by some activation functions are set using -DA_VAL= and -DB_VAL= respectively. 114 * @note Quantization scales of the input/output tensors are passed in with -DS1_VAL= and -DS2_VAL= respectively. 115 * @note Quantization offsets of the input/output tensors are passed in with -DO1_VAL= and -DO2_VAL= respectively. 116 * @note Quantized value of constant zero should be given as a preprocessor argument using -DCONST_0=value. e.g. -DCONST_0=128. 117 * 118 * @param[in] input_ptr Pointer to the source image. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM16 119 * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) 120 * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) 121 * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) 122 * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) 123 * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) 124 * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) 125 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image 126 * @param[out] output_ptr (Optional) Pointer to the destination image. Supported data types: same as @p input_ptr 127 * @param[in] output_stride_x (Optional) Stride of the destination image in X dimension (in bytes) 128 * @param[in] output_step_x (Optional) output_stride_x * number of elements along X processed per workitem(in bytes) 129 * @param[in] output_stride_y (Optional) Stride of the destination image in Y dimension (in bytes) 130 * @param[in] output_step_y (Optional) output_stride_y * number of elements along Y processed per workitem(in bytes) 131 * @param[in] output_stride_z (Optional) Stride of the source tensor in Z dimension (in bytes) 132 * @param[in] output_step_z (Optional) output_stride_z * number of elements along Z processed per workitem(in bytes) 133 * @param[in] output_offset_first_element_in_bytes (Optional) The offset of the first element in the destination image 134 */ 135__kernel void activation_layer_quant( 136 TENSOR3D_DECLARATION(input) 137#ifndef IN_PLACE 138 , 139 TENSOR3D_DECLARATION(output) 140#endif /* not IN_PLACE */ 141) 142{ 143 uint x_offs = max((int)(get_global_id(0) * VEC_SIZE * sizeof(DATA_TYPE) - (VEC_SIZE - VEC_SIZE_LEFTOVER) % VEC_SIZE * sizeof(DATA_TYPE)), 0); 144 145 // Get pixels pointer 146 __global uchar *input_addr = input_ptr + input_offset_first_element_in_bytes + x_offs + get_global_id(1) * input_stride_y + get_global_id(2) * input_stride_z; 147#ifdef IN_PLACE 148 __global uchar *output_addr = input_addr; 149#else /* IN_PLACE */ 150 __global uchar *output_addr = output_ptr + output_offset_first_element_in_bytes + x_offs + get_global_id(1) * output_stride_y + get_global_id(2) * output_stride_z; 151#endif /* IN_PLACE */ 152 153 // Load data 154 TYPE data0 = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)input_addr); 155 156 data0 = PERFORM_ACTIVATION_QUANT(ACT, data0); 157 158 // Store result 159 STORE_VECTOR_SELECT(data, DATA_TYPE, output_addr, VEC_SIZE, VEC_SIZE_LEFTOVER, VEC_SIZE_LEFTOVER != 0 && get_global_id(0) == 0) 160} 161#endif // defined(ACT) 162#endif // defined(FLOAT_DOMAIN) 163