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#if defined(VEC_SIZE) && defined(DATA_TYPE_SRC) && defined(DATA_TYPE_DST) && defined(SCALE) && defined(OFFSET) 27 28/** This performs the dequantization of 8-bit unsigned integers to floating point. 29 * 30 * @note Source datatype should be given as a preprocessor argument using -DDATA_TYPE_SRC=type. e.g. -DDATA_TYPE_SRC=char 31 * @note Destination datatype should be given as a preprocessor argument using -DDATA_TYPE_DST=type. e.g. -DDATA_TYPE_DST=float 32 * @note Vector size should be given as a preprocessor argument using -DVEC_SIZE=size. e.g. -DVEC_SIZE=16 33 * @note Quantization scale of input tensor is passed in with -DSCALE=scale. 34 * @note Quantization offset of input tensor is passed in with -DOFFSET=offset. 35 * 36 * @param[in] input_ptr Pointer to the source tensor. Supported data types: QASYMM8/QASYMM8_SIGNED/QSYMM8 37 * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) 38 * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) 39 * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) 40 * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) 41 * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) 42 * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) 43 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source tensor 44 * @param[out] output_ptr Pointer to the destination tensor. Supported data types: F16/F32 45 * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) 46 * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) 47 * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) 48 * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) 49 * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) 50 * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) 51 * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor 52 */ 53__kernel void dequantization_layer( 54 TENSOR3D_DECLARATION(input), 55 TENSOR3D_DECLARATION(output)) 56{ 57 // Get pixels pointer 58 Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); 59 Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); 60 61#if defined(LAST_ACCESSED_X) 62 // Check if access on width gets out of bounds 63 // If it does shift access vector to access elements within bounds 64 const int xi = (int)(get_global_id(0) * VEC_SIZE); 65 input.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * input_stride_x; 66 output.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * output_stride_x; 67 68 // Load data 69 VEC_DATA_TYPE(int, VEC_SIZE) 70 val = CONVERT(VLOAD(VEC_SIZE)(0, (__global DATA_TYPE_SRC *)input.ptr), VEC_DATA_TYPE(int, VEC_SIZE)); 71 72 // Create scale and offset vectors 73 const VEC_DATA_TYPE(float, VEC_SIZE) 74 vscale = SCALE; 75 76 const VEC_DATA_TYPE(int, VEC_SIZE) 77 voffset = OFFSET; 78 79 // Dequantize 80 VEC_DATA_TYPE(float, VEC_SIZE) 81 res = vscale * CONVERT((val - voffset), VEC_DATA_TYPE(float, VEC_SIZE)); 82 83 // Store result 84 VSTORE(VEC_SIZE) 85 (CONVERT(res, VEC_DATA_TYPE(DATA_TYPE_DST, VEC_SIZE)), 0, (__global DATA_TYPE_DST *)output.ptr); 86#else // !defined(LAST_ACCESSED_X) 87 *((__global DATA_TYPE_DST *)(output.ptr)) = (DATA_TYPE_DST)((float)((int)(*((__global DATA_TYPE_SRC *)(input.ptr))) - (int)(OFFSET)) * (float)(SCALE)); 88#endif // defined(LAST_ACCESSED_X) 89} 90#endif // defined(VEC_SIZE) && defined(DATA_TYPE_SRC) && defined(DATA_TYPE_DST) && defined(SCALE) && defined(OFFSET)