Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
| 2 | * Copyright (c) 2017 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 | |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 26 | #ifdef POOL_AVG |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 27 | #define POOL_OP(x, y) ((x) + (y)) |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 28 | #else /* POOL_AVG */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 29 | #define POOL_OP(x, y) (fmax((x), (y))) |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 30 | #endif /* POOL_AVG */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 31 | |
| 32 | float calculate_avg_scale(const int pool_size, const int upper_bound_w, const int upper_bound_h, |
| 33 | const int pad_x, const int pad_y, const int stride_x, const int stride_y) |
| 34 | { |
| 35 | int start_x = get_global_id(0) * stride_x - pad_x; |
| 36 | int start_y = get_global_id(1) * stride_y - pad_y; |
| 37 | int end_x = min(start_x + pool_size, upper_bound_w); |
| 38 | int end_y = min(start_y + pool_size, upper_bound_h); |
| 39 | return 1.f / ((end_y - start_y) * (end_x - start_x)); |
| 40 | } |
| 41 | |
| 42 | /** Performs a pooling function of pool size equal to 2. |
| 43 | * |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 44 | * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16, F32; |
| 45 | * @note In case of average pooling -DPOOL_AVG must be provided otherwise max pooling will be performed. |
| 46 | * |
| 47 | * @param[in] input_ptr Pointer to the source image. Supported data types: F16, F32 |
| 48 | * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) |
| 49 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 50 | * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) |
| 51 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 52 | * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 53 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 54 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image |
| 55 | * @param[out] output_ptr Pointer to the destination image. Supported data types: F16, F32 |
| 56 | * @param[in] output_stride_x Stride of the destination image in X dimension (in bytes) |
| 57 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 58 | * @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes) |
| 59 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 60 | * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 61 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 62 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image |
| 63 | * @param[in] max_dims The maximum index that can be accessed in x and y dimension (width + pad) |
| 64 | * @param[in] strides The pooling operation strides in each dimension |
| 65 | * @param[in] paddings The pooling operation paddings in each dimension |
| 66 | */ |
| 67 | __kernel void pooling_layer_2( |
| 68 | TENSOR3D_DECLARATION(input), |
| 69 | TENSOR3D_DECLARATION(output) |
| 70 | #ifdef POOL_AVG |
| 71 | , |
| 72 | int2 max_dims, int2 strides, int2 paddings |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 73 | #endif /* POOL_AVG */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 74 | ) |
| 75 | { |
| 76 | // Get pixels pointer |
| 77 | Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); |
| 78 | Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); |
| 79 | |
| 80 | // Load data |
| 81 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 82 | data0 = vload2(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); |
| 83 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 84 | data1 = vload2(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); |
| 85 | |
| 86 | // Perform calculations |
| 87 | data0 = POOL_OP(data0, data1); |
| 88 | DATA_TYPE res = POOL_OP(data0.s0, data0.s1); |
| 89 | |
Georgios Pinitas | ce09314 | 2017-06-19 16:11:53 +0100 | [diff] [blame] | 90 | // Divide by pool region in case of average pooling |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 91 | #ifdef POOL_AVG |
| 92 | res *= calculate_avg_scale(2, max_dims.x, max_dims.y, paddings.x, paddings.y, strides.x, strides.y); |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 93 | #endif /* POOL_AVG */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 94 | |
| 95 | // Store result |
| 96 | *(__global DATA_TYPE *)output.ptr = res; |
| 97 | } |
| 98 | |
| 99 | /** Performs a pooling function of pool size equal to 3. |
| 100 | * |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 101 | * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16, F32; |
| 102 | * @note In case of average pooling -DPOOL_AVG must be provided otherwise max pooling will be performed. |
| 103 | * |
| 104 | * @param[in] input_ptr Pointer to the source image. Supported data types: F16, F32 |
| 105 | * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) |
| 106 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 107 | * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) |
| 108 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 109 | * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 110 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 111 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image |
| 112 | * @param[out] output_ptr Pointer to the destination image. Supported data types: F16, F32 |
| 113 | * @param[in] output_stride_x Stride of the destination image in X dimension (in bytes) |
| 114 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 115 | * @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes) |
| 116 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 117 | * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 118 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 119 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image |
| 120 | * @param[in] max_dims The maximum index that can be accessed in x and y dimension (width + pad) |
| 121 | * @param[in] strides The pooling operation strides in each dimension |
| 122 | * @param[in] paddings The pooling operation paddings in each dimension |
| 123 | */ |
| 124 | __kernel void pooling_layer_3( |
| 125 | TENSOR3D_DECLARATION(input), |
| 126 | TENSOR3D_DECLARATION(output) |
| 127 | #ifdef POOL_AVG |
| 128 | , |
| 129 | int2 max_dims, int2 strides, int2 paddings |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 130 | #endif /* POOL_AVG */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 131 | ) |
| 132 | { |
| 133 | // Get pixels pointer |
| 134 | Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); |
| 135 | Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); |
| 136 | |
| 137 | // Load data |
| 138 | VEC_DATA_TYPE(DATA_TYPE, 3) |
| 139 | data0 = vload3(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); |
| 140 | VEC_DATA_TYPE(DATA_TYPE, 3) |
| 141 | data1 = vload3(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); |
| 142 | VEC_DATA_TYPE(DATA_TYPE, 3) |
| 143 | data2 = vload3(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); |
| 144 | |
| 145 | // Perform calculations |
| 146 | data0 = POOL_OP(data0, data1); |
| 147 | data0 = POOL_OP(data0, data2); |
| 148 | DATA_TYPE res = POOL_OP(POOL_OP(data0.s0, data0.s1), data0.s2); |
| 149 | |
Georgios Pinitas | ce09314 | 2017-06-19 16:11:53 +0100 | [diff] [blame] | 150 | // Divide by pool region in case of average pooling |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 151 | #ifdef POOL_AVG |
| 152 | res *= calculate_avg_scale(3, max_dims.x, max_dims.y, paddings.x, paddings.y, strides.x, strides.y); |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 153 | #endif /* POOL_AVG */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 154 | |
| 155 | // Store result |
| 156 | *(__global DATA_TYPE *)output.ptr = res; |
| 157 | } |
Georgios Pinitas | ce09314 | 2017-06-19 16:11:53 +0100 | [diff] [blame] | 158 | |
| 159 | /** Performs a pooling function of pool size equal to 7. |
| 160 | * |
| 161 | * @note Datatype must be passed using -DDATA_TYPE e.g. -DDATA_TYPE=float. Supported data types are F16, F32; |
| 162 | * @note In case of average pooling -DPOOL_AVG must be provided otherwise max pooling will be performed. |
| 163 | * |
| 164 | * @param[in] input_ptr Pointer to the source image. Supported data types: F16, F32 |
| 165 | * @param[in] input_stride_x Stride of the source image in X dimension (in bytes) |
| 166 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 167 | * @param[in] input_stride_y Stride of the source image in Y dimension (in bytes) |
| 168 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 169 | * @param[in] input_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 170 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 171 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the source image |
| 172 | * @param[out] output_ptr Pointer to the destination image. Supported data types: F16, F32 |
| 173 | * @param[in] output_stride_x Stride of the destination image in X dimension (in bytes) |
| 174 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 175 | * @param[in] output_stride_y Stride of the destination image in Y dimension (in bytes) |
| 176 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 177 | * @param[in] output_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 178 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 179 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination image |
| 180 | * @param[in] max_dims The maximum index that can be accessed in x and y dimension (width + pad) |
| 181 | * @param[in] strides The pooling operation strides in each dimension |
| 182 | * @param[in] paddings The pooling operation paddings in each dimension |
| 183 | */ |
| 184 | __kernel void pooling_layer_7( |
| 185 | TENSOR3D_DECLARATION(input), |
| 186 | TENSOR3D_DECLARATION(output) |
| 187 | #ifdef POOL_AVG |
| 188 | , |
| 189 | int2 max_dims, int2 strides, int2 paddings |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 190 | #endif /* POOL_AVG */ |
Georgios Pinitas | ce09314 | 2017-06-19 16:11:53 +0100 | [diff] [blame] | 191 | ) |
| 192 | { |
| 193 | // Get pixels pointer |
| 194 | Tensor3D input = CONVERT_TO_TENSOR3D_STRUCT(input); |
| 195 | Tensor3D output = CONVERT_TO_TENSOR3D_STRUCT(output); |
| 196 | |
| 197 | // Load data |
| 198 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 199 | data0 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 0, 0)); |
| 200 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 201 | data1 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 1, 0)); |
| 202 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 203 | data2 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 2, 0)); |
| 204 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 205 | data3 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 3, 0)); |
| 206 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 207 | data4 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 4, 0)); |
| 208 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 209 | data5 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 5, 0)); |
| 210 | VEC_DATA_TYPE(DATA_TYPE, 8) |
| 211 | data6 = vload8(0, (__global DATA_TYPE *)tensor3D_offset(&input, 0, 6, 0)); |
| 212 | |
| 213 | // Pool operation of all rows |
| 214 | data0 = POOL_OP(data0, data1); |
| 215 | data2 = POOL_OP(data2, data3); |
| 216 | data4 = POOL_OP(data4, data5); |
| 217 | data0 = POOL_OP(data0, data2); |
| 218 | data4 = POOL_OP(data4, data6); |
| 219 | data0 = POOL_OP(data0, data4); |
| 220 | |
| 221 | // Set last element |
| 222 | #ifdef POOL_AVG |
| 223 | data0.s7 = 0; |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 224 | #else /* POOL_AVG */ |
Georgios Pinitas | ce09314 | 2017-06-19 16:11:53 +0100 | [diff] [blame] | 225 | data0.s7 = data0.s6; |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 226 | #endif /* POOL_AVG */ |
Georgios Pinitas | ce09314 | 2017-06-19 16:11:53 +0100 | [diff] [blame] | 227 | |
| 228 | // Reduce result |
| 229 | VEC_DATA_TYPE(DATA_TYPE, 4) |
| 230 | reduce4 = POOL_OP(data0.s0123, data0.s4567); |
| 231 | VEC_DATA_TYPE(DATA_TYPE, 2) |
| 232 | reduce2 = POOL_OP(reduce4.s01, reduce4.s23); |
| 233 | DATA_TYPE res = POOL_OP(reduce2.s0, reduce2.s1); |
| 234 | |
| 235 | // Divide by pool region in case of average pooling |
| 236 | #ifdef POOL_AVG |
| 237 | res *= calculate_avg_scale(7, max_dims.x, max_dims.y, paddings.x, paddings.y, strides.x, strides.y); |
Anthony Barbier | ac69aa1 | 2017-07-03 17:39:37 +0100 | [diff] [blame] | 238 | #endif /* POOL_AVG */ |
Georgios Pinitas | ce09314 | 2017-06-19 16:11:53 +0100 | [diff] [blame] | 239 | |
| 240 | // Store result |
| 241 | *(__global DATA_TYPE *)output.ptr = res; |
| 242 | } |