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(BATCH_SIZE) && defined(DATA_TYPE) && defined(WIDTH_IN) && defined(HEIGHT_IN) 27/** Calculate the space to batch conversion. (NHWC) 28 * 29 * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float 30 * @note The block shape tensor rank must be passed at compile time using -DBLOCK_SHAPE_DIM. e.g. -DBLOCK_SHAPE_DIM=2 31 * 32 * @param[in] input_ptr Pointer to the source tensor. Supported data types: All 33 * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) 34 * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) 35 * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) 36 * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) 37 * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) 38 * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) 39 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source image 40 * @param[in] paddings_ptr Pointer to the second source image. Supported data types: S32 41 * @param[in] paddings_stride_x Stride of the paddinds tensor in X dimension (in bytes) 42 * @param[in] paddings_step_x paddings_stride_x * number of elements along X processed per workitem(in bytes) 43 * @param[in] paddings_stride_y Stride of the paddinds tensor in Y dimension (in bytes) 44 * @param[in] paddings_step_y paddings_stride_y * number of elements along Y processed per workitem(in bytes) 45 * @param[in] paddingse_offset_first_element_in_bytes The offset of the first element in the second source image 46 * @param[in] block_shape_ptr Pointer to the block shape tensor. Supported data types: S32 47 * @param[in] block_shape_stride_x Stride of the block shape tensor in X dimension (in bytes) 48 * @param[in] block_shape_step_x block_shape_stride_x * number of elements along X processed per workitem(in bytes) 49 * @param[in] block_shape_offset_first_element_in_bytes The offset of the first element in the block shapetensor 50 * @param[in] batch_id The output tensor batch id 51 * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr 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 destination 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 image 59 */ 60__kernel void space_to_batch_nhwc( 61 TENSOR4D_DECLARATION(input), 62 IMAGE_DECLARATION(paddings), 63 VECTOR_DECLARATION(block_shape), 64 const int batch_id, 65 TENSOR3D_DECLARATION(output)) 66{ 67 Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0); 68 Image pad = CONVERT_TO_IMAGE_STRUCT_NO_STEP(paddings); 69 Vector block = CONVERT_TO_VECTOR_STRUCT_NO_STEP(block_shape); 70 Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output); 71 72 const int pad_left_x = *((__global int *)offset(&pad, 0, 0)); 73 const int pad_right_x = *((__global int *)offset(&pad, 1, 0)); 74 const int pad_left_y = *((__global int *)offset(&pad, 0, 1)); 75 const int pad_right_y = *((__global int *)offset(&pad, 1, 1)); 76 77 int block_x = *((__global int *)vector_offset(&block, 0)); 78 int block_y = *((__global int *)vector_offset(&block, 1)); 79 80 const int out_x = get_global_id(1); 81 const int out_y = get_global_id(2); 82 const int z = get_global_id(0); 83 84 const int pos_x = out_x * block_x + ((batch_id / BATCH_IN) % block_x); 85 const int pos_y = out_y * block_y + ((batch_id / BATCH_IN) / block_x); 86 87 if(((pos_y >= pad_left_y) && (pos_y < pad_left_y + HEIGHT_IN) && (pos_x >= pad_left_x) && (pos_x < pad_left_x + WIDTH_IN))) 88 { 89 const int w = batch_id % BATCH_IN; 90 const int in_x = pos_x - pad_left_x; 91 const int in_y = pos_y - pad_left_y; 92 93 *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, z, in_x, in_y, w)); 94 } 95} 96#endif // defined(BATCH_SIZE) && defined(DATA_TYPE) && defined(WIDTH_IN) && defined(HEIGHT_IN) 97 98#if defined(BATCH_SIZE) && defined(DATA_TYPE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y) && defined(PAD_LEFT_X) && defined(PAD_RIGHT_X) && defined(PAD_LEFT_Y) && defined(PAD_RIGHT_Y) && defined(WIDTH_IN) && defined(HEIGHT_IN) 99/** Calculate the space to batch conversion. (NHWC) 100 * 101 * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float 102 * @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2 103 * @note The block shape x must be passed at compile time using -DBLOCK_SHAPE_X. e.g. -DBLOCK_SHAPE_X=2 104 * @note The block shape y must be passed at compile time using -DBLOCK_SHAPE_Y. e.g. -DBLOCK_SHAPE_Y=2 105 * @note The starting pad value of x must be passed at compile time using -DPAD_LEFT_X. e.g. -DPAD_LEFT_X=2 106 * @note The ending pad value of x must be passed at compile time using -DPAD_RIGHT_X. e.g. -DPAD_RIGHT_X=2 107 * @note The starting pad value of y must be passed at compile time using -DPAD_LEFT_Y. e.g. -DPAD_LEFT_Y=2 108 * @note The ending pad value of y must be passed at compile time using -DPAD_RIGHT_Y. e.g. -DPAD_RIGHT_X=2 109 * 110 * @param[in] input_ptr Pointer to the source tensor. Supported data types: All 111 * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) 112 * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) 113 * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) 114 * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) 115 * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) 116 * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) 117 * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source image 118 * @param[in] batch_id The output tensor batch id 119 * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr 120 * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) 121 * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) 122 * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) 123 * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) 124 * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) 125 * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) 126 * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image 127 */ 128__kernel void space_to_batch_static_nhwc( 129 TENSOR4D_DECLARATION(input), 130 const int batch_id, 131 TENSOR3D_DECLARATION(output)) 132{ 133 Tensor4D in = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(input, 0); 134 Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output); 135 136 int block_x = BLOCK_SHAPE_X; 137 int block_y = BLOCK_SHAPE_Y; 138 139 const int out_x = get_global_id(1); 140 const int out_y = get_global_id(2); 141 const int z = get_global_id(0); 142 143 const int pos_x = out_x * block_x + ((batch_id / BATCH_IN) % block_x); 144 const int pos_y = out_y * block_y + ((batch_id / BATCH_IN) / block_x); 145 146 if(pos_y >= PAD_LEFT_Y && pos_y < PAD_LEFT_Y + HEIGHT_IN && pos_x >= PAD_LEFT_X && pos_x < PAD_LEFT_X + WIDTH_IN) 147 { 148 const int w = batch_id % BATCH_IN; 149 const int in_x = pos_x - PAD_LEFT_X; 150 const int in_y = pos_y - PAD_LEFT_Y; 151 152 *((__global DATA_TYPE *)out.ptr) = *((__global DATA_TYPE *)tensor4D_offset(&in, z, in_x, in_y, w)); 153 } 154} 155#endif // defined(BATCH_SIZE) && defined(DATA_TYPE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y) && defined(PAD_LEFT_X) && defined(PAD_RIGHT_X) && defined(PAD_LEFT_Y) && defined(PAD_RIGHT_Y) && defined(WIDTH_IN) && defined(HEIGHT_IN)