Michalis Spyrou | 6a8d3b6 | 2018-08-31 10:07:09 +0100 | [diff] [blame] | 1 | /* |
Michele Di Giorgio | d9eaf61 | 2020-07-08 11:12:57 +0100 | [diff] [blame] | 2 | * Copyright (c) 2018-2020 Arm Limited. |
Michalis Spyrou | 6a8d3b6 | 2018-08-31 10:07:09 +0100 | [diff] [blame] | 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 |
Sheri Zhang | 0de45d0 | 2020-04-17 14:59:13 +0100 | [diff] [blame] | 8 | * deal in the Software without restriction, including without limitation the |
Michalis Spyrou | 6a8d3b6 | 2018-08-31 10:07:09 +0100 | [diff] [blame] | 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) |
Michalis Spyrou | f1addb6 | 2018-09-11 11:16:47 +0100 | [diff] [blame] | 27 | /** Batch to space transformation. (NCHW) |
Michalis Spyrou | 6a8d3b6 | 2018-08-31 10:07:09 +0100 | [diff] [blame] | 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 | * |
Sheri Zhang | 0de45d0 | 2020-04-17 14:59:13 +0100 | [diff] [blame] | 33 | * @param[in] input_ptr Pointer to the source tensor. Supported data types: All |
Michalis Spyrou | 6a8d3b6 | 2018-08-31 10:07:09 +0100 | [diff] [blame] | 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 | */ |
Michalis Spyrou | f1addb6 | 2018-09-11 11:16:47 +0100 | [diff] [blame] | 57 | __kernel void batch_to_space_nchw( |
Michalis Spyrou | 6a8d3b6 | 2018-08-31 10:07:09 +0100 | [diff] [blame] | 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(0); |
| 72 | const int y = get_global_id(1); |
| 73 | const int z = get_global_id(2); |
| 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, out_x, out_y, z, w)) = *((__global DATA_TYPE *)in.ptr); |
| 80 | } |
Michalis Spyrou | f1addb6 | 2018-09-11 11:16:47 +0100 | [diff] [blame] | 81 | /** Batch to space transformation. (NHWC) |
| 82 | * |
| 83 | * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float |
| 84 | * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float |
| 85 | * @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2 |
| 86 | * |
Sheri Zhang | 0de45d0 | 2020-04-17 14:59:13 +0100 | [diff] [blame] | 87 | * @param[in] input_ptr Pointer to the source tensor. Supported data types: All |
Michalis Spyrou | f1addb6 | 2018-09-11 11:16:47 +0100 | [diff] [blame] | 88 | * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) |
| 89 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 90 | * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 91 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 92 | * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 93 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 94 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor |
| 95 | * @param[in] batch_id The input tensor batch id |
| 96 | * @param[in] block_shape_ptr Pointer to the source tensor. Supported data types: S32 |
| 97 | * @param[in] block_shape_stride_x Stride of the source tensor in X dimension (in bytes) |
| 98 | * @param[in] block_shape_step_x block_shape_stride_x * number of elements along X processed per workitem(in bytes) |
| 99 | * @param[in] block_shape_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 100 | * @param[in] block_shape_step_y block_shape_stride_y * number of elements along Y processed per workitem(in bytes) |
| 101 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor |
| 102 | * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr |
| 103 | * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 104 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 105 | * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 106 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 107 | * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 108 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 109 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 110 | */ |
| 111 | __kernel void batch_to_space_nhwc( |
| 112 | TENSOR3D_DECLARATION(input), |
| 113 | const int batch_id, |
| 114 | VECTOR_DECLARATION(block_shape), |
| 115 | TENSOR4D_DECLARATION(output)) |
| 116 | { |
| 117 | Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input); |
| 118 | Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0); |
| 119 | Vector block = CONVERT_TO_VECTOR_STRUCT_NO_STEP(block_shape); |
| 120 | |
| 121 | const int block_x = *((__global int *)vector_offset(&block, 0)); |
| 122 | const int block_y = *((__global int *)vector_offset(&block, 1)); |
| 123 | |
| 124 | const int r = (BATCH_SIZE / (block_x * block_y)); |
| 125 | const int x = get_global_id(1); |
| 126 | const int y = get_global_id(2); |
| 127 | const int z = get_global_id(0); |
| 128 | const int w = batch_id % r; |
| 129 | |
| 130 | const int out_x = x * block_x + (batch_id / r) % block_x; |
| 131 | const int out_y = y * block_y + (batch_id / r) / block_x; |
| 132 | |
| 133 | *((__global DATA_TYPE *)tensor4D_offset(&out, z, out_x, out_y, w)) = *((__global DATA_TYPE *)in.ptr); |
| 134 | } |
Michalis Spyrou | 6a8d3b6 | 2018-08-31 10:07:09 +0100 | [diff] [blame] | 135 | #endif // defined(DATA_TYPE) && defined(BATCH_SIZE) |
| 136 | |
| 137 | #if defined(DATA_TYPE) && defined(BATCH_SIZE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y) |
Michalis Spyrou | f1addb6 | 2018-09-11 11:16:47 +0100 | [diff] [blame] | 138 | /** Batch to space transformation. (NCHW) |
Michalis Spyrou | 6a8d3b6 | 2018-08-31 10:07:09 +0100 | [diff] [blame] | 139 | * |
| 140 | * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float |
| 141 | * @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2 |
| 142 | * @note The block shape x must be passed at compile time using -DBLOCK_SHAPE_X. e.g. -DBLOCK_SHAPE_X=2 |
| 143 | * @note The block shape y must be passed at compile time using -DBLOCK_SHAPE_Y. e.g. -DBLOCK_SHAPE_Y=2 |
| 144 | * |
Sheri Zhang | 0de45d0 | 2020-04-17 14:59:13 +0100 | [diff] [blame] | 145 | * @param[in] input_ptr Pointer to the source tensor. Supported data types: All |
Michalis Spyrou | 6a8d3b6 | 2018-08-31 10:07:09 +0100 | [diff] [blame] | 146 | * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) |
| 147 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 148 | * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 149 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 150 | * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 151 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 152 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor |
| 153 | * @param[in] batch_id The input tensor batch id |
| 154 | * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr |
| 155 | * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 156 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 157 | * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 158 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 159 | * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 160 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 161 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 162 | */ |
Michalis Spyrou | f1addb6 | 2018-09-11 11:16:47 +0100 | [diff] [blame] | 163 | __kernel void batch_to_space_static_nchw( |
Michalis Spyrou | 6a8d3b6 | 2018-08-31 10:07:09 +0100 | [diff] [blame] | 164 | TENSOR3D_DECLARATION(input), |
| 165 | const int batch_id, |
| 166 | TENSOR4D_DECLARATION(output)) |
| 167 | { |
| 168 | Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input); |
| 169 | Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0); |
| 170 | |
| 171 | const int block_x = BLOCK_SHAPE_X; |
| 172 | const int block_y = BLOCK_SHAPE_Y; |
| 173 | |
| 174 | const int r = (BATCH_SIZE / (block_x * block_y)); |
| 175 | const int x = get_global_id(0); |
| 176 | const int y = get_global_id(1); |
| 177 | const int z = get_global_id(2); |
| 178 | const int w = batch_id % r; |
| 179 | |
| 180 | const int out_x = x * block_x + (batch_id / r) % block_x; |
| 181 | const int out_y = y * block_y + (batch_id / r) / block_x; |
| 182 | |
| 183 | *((__global DATA_TYPE *)tensor4D_offset(&out, out_x, out_y, z, w)) = *((__global DATA_TYPE *)in.ptr); |
| 184 | } |
Michalis Spyrou | f1addb6 | 2018-09-11 11:16:47 +0100 | [diff] [blame] | 185 | /** Batch to space transformation. (NHWC) |
| 186 | * |
| 187 | * @note Datatype should be given as a preprocessor argument using -DDATA_TYPE=type. e.g. -DDATA_TYPE=float |
| 188 | * @note The input tensor batch size must be passed at compile time using -DBATCH_SIZE. e.g. -DBATCH_SIZE=2 |
| 189 | * @note The block shape x must be passed at compile time using -DBLOCK_SHAPE_X. e.g. -DBLOCK_SHAPE_X=2 |
| 190 | * @note The block shape y must be passed at compile time using -DBLOCK_SHAPE_Y. e.g. -DBLOCK_SHAPE_Y=2 |
| 191 | * |
Sheri Zhang | 0de45d0 | 2020-04-17 14:59:13 +0100 | [diff] [blame] | 192 | * @param[in] input_ptr Pointer to the source tensor. Supported data types: All |
Michalis Spyrou | f1addb6 | 2018-09-11 11:16:47 +0100 | [diff] [blame] | 193 | * @param[in] input_stride_x Stride of the source tensor in X dimension (in bytes) |
| 194 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 195 | * @param[in] input_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 196 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 197 | * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 198 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 199 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor |
| 200 | * @param[in] batch_id The input tensor batch id |
| 201 | * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr |
| 202 | * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 203 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 204 | * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 205 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 206 | * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 207 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 208 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 209 | */ |
| 210 | __kernel void batch_to_space_static_nhwc( |
| 211 | TENSOR3D_DECLARATION(input), |
| 212 | const int batch_id, |
| 213 | TENSOR4D_DECLARATION(output)) |
| 214 | { |
| 215 | Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input); |
| 216 | Tensor4D out = CONVERT_TO_TENSOR4D_STRUCT_NO_STEP(output, 0); |
| 217 | |
| 218 | const int block_x = BLOCK_SHAPE_X; |
| 219 | const int block_y = BLOCK_SHAPE_Y; |
| 220 | |
| 221 | const int r = (BATCH_SIZE / (block_x * block_y)); |
| 222 | const int x = get_global_id(1); |
| 223 | const int y = get_global_id(2); |
| 224 | const int z = get_global_id(0); |
| 225 | const int w = batch_id % r; |
| 226 | |
| 227 | const int out_x = x * block_x + (batch_id / r) % block_x; |
| 228 | const int out_y = y * block_y + (batch_id / r) / block_x; |
| 229 | |
| 230 | *((__global DATA_TYPE *)tensor4D_offset(&out, z, out_x, out_y, w)) = *((__global DATA_TYPE *)in.ptr); |
| 231 | } |
Michalis Spyrou | 6a8d3b6 | 2018-08-31 10:07:09 +0100 | [diff] [blame] | 232 | #endif // defined(DATA_TYPE) && defined(BATCH_SIZE) && defined(BLOCK_SHAPE_X) && defined(BLOCK_SHAPE_Y) |