1/* 2 * Copyright (c) 2017-2021 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 "helpers.h" 25 26#define CONVERT_RTE(x, type) (convert_##type##_rte((x))) 27#define CONVERT_RTE_VEC_STR(x, type, size) (convert_##type##size##_rte((x))) 28#define CONVERT_RTE_VEC(x, type, size) CONVERT_RTE_VEC_STR(x, type, size) 29 30#if defined(VEC_SIZE) && defined(DATA_TYPE_IN) && defined(DATA_TYPE_OUT) && defined(SCALE) && defined(OFFSET) && defined(MIN_QUANT_VAL) && defined(MAX_QUANT_VAL) 31 32/** This performs the quantization of floating point inputs or 8-bit quantized integers to 8-bit integers. 33 * 34 * @note Input data type should be given as a preprocessor argument using -DDATA_TYPE_IN=type. e.g. -DDATA_TYPE=short 35 * @note Output data type should be given as a preprocessor argument using -DDATA_TYPE_OUT=type. e.g. -DDATA_TYPE=short 36 * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 37 * @note Quantization scale should be given as a preprocessor argument using -DSCALE=scale. e.g. -DSCALE=0.125 38 * @note Quantization offset should be given as a preprocessor argument using -DOFFSET=offset. e.g. -DOFFSET=125 39 * @note Minimum value for quantized type should be given as a preprocessor argument using -DMIN_QUANT_VAL=value. e.g. -DMIN_QUANT_VAL=0 40 * @note Maximum value for quantized type should be given as a preprocessor argument using -DMAX_QUANT_VAL=value. e.g. -DMAXIN_QUANT_VAL=255 41 * @note If the input data type if a floating point (F16 or F32) the preprocessor argument should be give as -DIS_FLOAT 42 * 43 * @param[in] input_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/F16/F32 44 * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) 45 * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) 46 * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) 47 * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) 48 * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) 49 * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) 50 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor 51 * @param[out] output_ptr Pointer to the destination tensor. Supported data types: QASYMM8/QASYMM8_SIGNED 52 * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) 53 * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) 54 * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) 55 * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) 56 * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) 57 * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) 58 * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor 59 */ 60__kernel void quantization_layer( 61 TENSOR3D_DECLARATION(input), 62 TENSOR3D_DECLARATION(output)) 63{ 64 // Get pixels pointer 65 Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); 66 Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); 67 68#if defined(VEC_SIZE) && defined(LAST_ACCESSED_X) 69 // Check if access on width gets out of bounds 70 // If it does shift access vector to access elements within bounds 71 const int xi = (int)(get_global_id(0) * VEC_SIZE); 72 input.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * input_stride_x; 73 output.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * output_stride_x; 74 75 // Load data 76#if defined(IS_FLOAT) 77 // Load data 78 VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE) 79 val_float = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)input.ptr); 80 81 // Create scale and offset vectors 82 const VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE) vscale = SCALE; 83 const VEC_DATA_TYPE(int, VEC_SIZE) voffset = OFFSET; 84#else // defined(IS_FLOAT) 85 // Load data 86 VEC_DATA_TYPE(DATA_TYPE_IN, VEC_SIZE) 87 val = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_IN *)input.ptr); 88 89 const VEC_DATA_TYPE(float, VEC_SIZE) 90 val_float = CONVERT(val, VEC_DATA_TYPE(float, VEC_SIZE)); 91 92 // Create scale and offset vectors 93 const VEC_DATA_TYPE(float, VEC_SIZE) vscale = SCALE; 94 const VEC_DATA_TYPE(int, VEC_SIZE) voffset = OFFSET; 95#endif // defined(IS_FLOAT) 96 97 // Quantize 98 VEC_DATA_TYPE(int, VEC_SIZE) 99 res = CLAMP(CONVERT_RTE_VEC(val_float / vscale, int, VEC_SIZE) + voffset, MIN_QUANT_VAL, MAX_QUANT_VAL); 100 101 // Store result 102 VSTORE(VEC_SIZE) 103 (CONVERT(res, VEC_DATA_TYPE(DATA_TYPE_OUT, VEC_SIZE)), 0, (__global DATA_TYPE_OUT *)output.ptr); 104#else //!defined(VEC_SIZE) || !defined(LAST_ACCESSED_X) 105 *((__global DATA_TYPE_OUT *)(output.ptr)) = (DATA_TYPE_OUT)CLAMP(CONVERT_RTE(((float) * (__global DATA_TYPE_IN *)input.ptr) / ((float)SCALE), int) + (int)OFFSET, MIN_QUANT_VAL, MAX_QUANT_VAL); 106#endif // defined(VEC_SIZE) && defined(LAST_ACCESSED_X) 107} 108#endif // defined(VEC_SIZE) && defined(DATA_TYPE_IN) && defined(DATA_TYPE_OUT) && defined(SCALE) && defined(OFFSET) && defined(MIN_QUANT_VAL) && defined(MAX_QUANT_VAL) 109