1/* 2 * Copyright (c) 2018-2021 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(BATCH_SIZE) 27/** Batch to space transformation. (NHWC) 28 * 29 * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float 30 * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float 31 * @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2 32 * 33 * @param[in] input_ptr Pointer to the source tensor. Supported data types: All 34 * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) 35 * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) 36 * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) 37 * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) 38 * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) 39 * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) 40 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor 41 * @param[in] batch_id The input tensor batch id 42 * @param[in] block_shape_ptr Pointer to the source tensor. Supported data types: S32 43 * @param[in] block_shape_stride_x Stride of the source tensor in X dimension (in bytes) 44 * @param[in] block_shape_step_x block_shape_stride_x * number of elements along X processed per workitem(in bytes) 45 * @param[in] block_shape_stride_y Stride of the source tensor in Y dimension (in bytes) 46 * @param[in] block_shape_step_y block_shape_stride_y * number of elements along Y processed per workitem(in bytes) 47 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor 48 * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr 49 * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) 50 * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) 51 * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) 52 * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) 53 * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) 54 * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) 55 * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor 56 */ 57__kernel void batch_to_space_nhwc( 58 TENSOR3D_DECLARATION(input), 59 const int batch_id, 60 VECTOR_DECLARATION(block_shape), 61 TENSOR4D_DECLARATION(output)) 62{ 63 Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input); 64 Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0); 65 Vector block = CONVERT_TO_VECTOR_STRUCT_NO_STEP(block_shape); 66 67 const int block_x = *((__global int *)vector_offset(&block, 0)); 68 const int block_y = *((__global int *)vector_offset(&block, 1)); 69 70 const int r = (BATCH_SIZE / (block_x * block_y)); 71 const int x = get_global_id(1); 72 const int y = get_global_id(2); 73 const int z = get_global_id(0); 74 const int w = batch_id % r; 75 76 const int out_x = x * block_x + (batch_id / r) % block_x; 77 const int out_y = y * block_y + (batch_id / r) / block_x; 78 79 *((__global DATA_TYPE *)tensor4D_offset(&out, z, out_x, out_y, w)) = *((__global DATA_TYPE *)in.ptr); 80} 81#endif // defined(DATA_TYPE) && defined(BATCH_SIZE) 82 83#if defined(DATA_TYPE) && defined(BATCH_SIZE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y) 84/** Batch to space transformation. (NHWC) 85 * 86 * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float 87 * @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2 88 * @note The block shape x must be passed at compile time using -DBLOCK_SHAPE_X. e.g. -DBLOCK_SHAPE_X=2 89 * @note The block shape y must be passed at compile time using -DBLOCK_SHAPE_Y. e.g. -DBLOCK_SHAPE_Y=2 90 * 91 * @param[in] input_ptr Pointer to the source tensor. Supported data types: All 92 * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) 93 * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) 94 * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) 95 * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) 96 * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) 97 * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) 98 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor 99 * @param[in] batch_id The input tensor batch id 100 * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr 101 * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) 102 * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) 103 * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) 104 * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) 105 * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) 106 * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) 107 * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor 108 */ 109__kernel void batch_to_space_static_nhwc( 110 TENSOR3D_DECLARATION(input), 111 const int batch_id, 112 TENSOR4D_DECLARATION(output)) 113{ 114 Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input); 115 Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0); 116 117 const int block_x = BLOCK_SHAPE_X; 118 const int block_y = BLOCK_SHAPE_Y; 119 120 const int r = (BATCH_SIZE / (block_x * block_y)); 121 const int x = get_global_id(1); 122 const int y = get_global_id(2); 123 const int z = get_global_id(0); 124 const int w = batch_id % r; 125 126 const int out_x = x * block_x + (batch_id / r) % block_x; 127 const int out_y = y * block_y + (batch_id / r) / block_x; 128 129 *((__global DATA_TYPE *)tensor4D_offset(&out, z, out_x, out_y, w)) = *((__global DATA_TYPE *)in.ptr); 130} 131#endif // defined(DATA_TYPE) && defined(BATCH_SIZE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y)