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(DATA_TYPE) && defined(NUM_GROUPS) 27/** This kernel reshapes the tensor's low three dimensions to single column 28 * 29 * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=short 30 * @note The number of groups should be given as a preprocessor argument using -DNUM_GROUPS=number. e.g. -DNUM_GROUPS=2 31 * 32 * @param[in] src_ptr Pointer to the source tensor. Supported data types: All 33 * @param[in] src_stride_x Stride of the source tensor in X dimension (in bytes) 34 * @param[in] src_step_x src_stride_x * number of elements along X processed per workitem(in bytes) 35 * @param[in] src_stride_y Stride of the source tensor in Y dimension (in bytes) 36 * @param[in] src_step_y src_stride_y * number of elements along Y processed per workitem(in bytes) 37 * @param[in] src_stride_z Stride of the source tensor in Z dimension (in bytes) 38 * @param[in] src_step_z src_stride_z * number of elements along Y processed per workitem(in bytes) 39 * @param[in] src_offset_first_element_in_bytes The offset of the first element in the source tensor 40 * @param[out] dst_ptr Pointer to the destination tensor. Same as @p src_ptr 41 * @param[in] dst_stride_x Stride of the destination tensor in X dimension (in bytes) 42 * @param[in] dst_step_x dst_stride_x * number of elements along X processed per workitem(in bytes) 43 * @param[in] dst_stride_y Stride of the destination tensor in Y dimension (in bytes) 44 * @param[in] dst_step_y dst_stride_y * number of elements along Y processed per workitem(in bytes) 45 * @param[in] dst_offset_first_element_in_bytes The offset of the first element in the destination tensor 46 * @param[in] bias_ptr Pointer to the bias tensor. Supported data types: F16/F32, for quantized types this must be nullptr 47 * @param[in] bias_stride_x Stride of the bias tensor in X dimension (in bytes) 48 * @param[in] bias_step_x bias_stride_x * number of elements along X processed per workitem(in bytes) 49 * @param[in] bias_offset_first_element_in_bytes The offset of the first element in the source tensor 50 * @param[in] width The width of the input tensor 51 * @param[in] height The height of the input tensor 52 * @param[in] depth The depth of the input tensor 53 * @param[in] total_filters Total number of filters. 4th dimension of the weights matrix 54 * @param[in] dst_stride_z Stride of the destination tensor in Z dimension (in bytes) 55 */ 56__kernel void reshape_to_columns( 57 TENSOR3D_DECLARATION(src), 58 IMAGE_DECLARATION(dst), 59#ifdef HAS_BIAS 60 VECTOR_DECLARATION(bias), 61#endif /* HAS_BIAS */ 62 uint width, uint height, uint depth, uint total_filters, uint dst_stride_z) 63{ 64 Tensor3D src = CONVERT_TO_TENSOR3D_STRUCT(src); 65 bool is_last_thread = (get_global_id(0) == (get_global_size(0) - 1) && get_global_id(1) == (get_global_size(1) - 1) && get_global_id(2) == (get_global_size(2) - 1)); 66 67 __global uchar *tmp_src_ptr = src.ptr; 68 __global uchar *tmp_dst_ptr = dst_ptr + dst_offset_first_element_in_bytes + get_global_id(0) * dst_stride_y + get_global_id(1) * width * dst_stride_y + get_global_id( 69 2) * width * height * dst_stride_y; 70#ifdef HAS_BIAS 71 __global uchar *tmp_bias_ptr = bias_ptr + bias_offset_first_element_in_bytes; 72#endif /* HAS_BIAS */ 73 74 if(is_last_thread) 75 { 76 for(uint g = 0; g < NUM_GROUPS; ++g) 77 { 78 __global uchar *curr_group_dst = tmp_dst_ptr; 79 80 for(uint i = 0; i < total_filters / NUM_GROUPS; ++i) 81 { 82 *((__global DATA_TYPE *)curr_group_dst) = *((__global DATA_TYPE *)tmp_src_ptr); 83 84#ifdef HAS_BIAS 85 *((__global DATA_TYPE *)(curr_group_dst + dst_stride_y)) = *((__global DATA_TYPE *)(tmp_bias_ptr)); 86 tmp_bias_ptr += bias_stride_x; 87#endif /* HAS_BIAS */ 88 tmp_src_ptr += depth * src_stride_z; 89 curr_group_dst += dst_stride_x; 90 } 91 92 tmp_dst_ptr += dst_stride_z; 93 } 94 } 95 else 96 { 97 for(uint g = 0; g < NUM_GROUPS; ++g) 98 { 99 __global uchar *curr_group_dst = tmp_dst_ptr; 100 101 for(uint i = 0; i < total_filters / NUM_GROUPS; ++i) 102 { 103 *((__global DATA_TYPE *)curr_group_dst) = *((__global DATA_TYPE *)tmp_src_ptr); 104 tmp_src_ptr += depth * src_stride_z; 105 curr_group_dst += dst_stride_x; 106 } 107 108 tmp_dst_ptr += dst_stride_z; 109 } 110 } 111} 112#endif // defined(DATA_TYPE) 113