Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 1 | /* |
Georgios Pinitas | 652bde5 | 2018-01-10 15:33:28 +0000 | [diff] [blame] | 2 | * Copyright (c) 2017-2018 ARM Limited. |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +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 |
| 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 | |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 26 | #define ADD_OP(a, b) ((a) + (b)) |
| 27 | #define SUB_OP(a, b) ((a) - (b)) |
| 28 | #define MUL_OP(a, b) ((a) * (b)) |
| 29 | #define INVSQRT_OP(a) rsqrt((a)) |
| 30 | #define SQCVT_SAT(a) (a) |
| 31 | |
Georgios Pinitas | c936917 | 2018-09-26 11:25:40 +0100 | [diff] [blame] | 32 | #if defined(VEC_SIZE) && defined(DATA_TYPE) |
| 33 | |
Giorgio Arena | 99ac60b | 2018-02-16 15:17:23 +0000 | [diff] [blame] | 34 | #if defined(FUSED_ACTIVATION) |
| 35 | #include "activation_layer.cl" |
| 36 | #define ACTIVATION_FUNC(x) ACTIVATION_OP(FUSED_ACTIVATION, x) |
| 37 | #else /* defined(FUSED_ACTIVATION) */ |
Giorgio Arena | 1167487 | 2018-02-07 15:38:12 +0000 | [diff] [blame] | 38 | #define ACTIVATION_FUNC(x) (x) |
Giorgio Arena | 99ac60b | 2018-02-16 15:17:23 +0000 | [diff] [blame] | 39 | #endif /* defined(FUSED_ACTIVATION) */ |
Giorgio Arena | 1167487 | 2018-02-07 15:38:12 +0000 | [diff] [blame] | 40 | |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 41 | /** Apply batch normalization. |
| 42 | * |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame] | 43 | * @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16/F32 |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 44 | * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) |
| 45 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 46 | * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) |
| 47 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 48 | * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes) |
| 49 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 50 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 51 | * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 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 tensor |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 59 | * @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p input_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 60 | * @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes) |
| 61 | * @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes) |
| 62 | * @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 63 | * @param[in] var_ptr Pointer to the var tensor. Supported data types: same as @p input_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 64 | * @param[in] var_stride_x Stride of the var tensor in X dimension (in bytes) |
| 65 | * @param[in] var_step_x var_stride_x * number of elements along X processed per workitem(in bytes) |
| 66 | * @param[in] var_offset_first_element_in_bytes The offset of the first element in the var source tensor |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 67 | * @param[in] beta_ptr Pointer to the beta source tensor. Supported data types: same as @p input_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 68 | * @param[in] beta_stride_x Stride of the beta source tensor in X dimension (in bytes) |
| 69 | * @param[in] beta_step_x beta_stride_x * number of elements along X processed per workitem(in bytes) |
| 70 | * @param[in] beta_offset_first_element_in_bytes The offset of the first element in the beta source tensor |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 71 | * @param[in] gamma_ptr Pointer to the gamma source tensor. Supported data types: same as @p input_ptr |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 72 | * @param[in] gamma_stride_x Stride of the gamma source tensor in X dimension (in bytes) |
| 73 | * @param[in] gamma_step_x gamma_stride_x * number of elements along X processed per workitem(in bytes) |
| 74 | * @param[in] gamma_offset_first_element_in_bytes The offset of the first element in the gamma source tensor |
| 75 | * @param[in] epsilon Epsilon parameter in the batch normalization equation |
| 76 | */ |
Michele Di Giorgio | bf3c662 | 2018-03-08 11:52:27 +0000 | [diff] [blame] | 77 | __kernel void batchnormalization_layer_nchw(TENSOR3D_DECLARATION(input), |
Georgios Pinitas | 409ee0a | 2017-08-18 10:16:09 +0100 | [diff] [blame] | 78 | #ifndef IN_PLACE |
Michele Di Giorgio | bf3c662 | 2018-03-08 11:52:27 +0000 | [diff] [blame] | 79 | TENSOR3D_DECLARATION(output), |
Georgios Pinitas | 409ee0a | 2017-08-18 10:16:09 +0100 | [diff] [blame] | 80 | #endif /* not IN_PLACE */ |
Michele Di Giorgio | bf3c662 | 2018-03-08 11:52:27 +0000 | [diff] [blame] | 81 | VECTOR_DECLARATION(mean), |
| 82 | VECTOR_DECLARATION(var), |
Michele Di Giorgio | 4d33630 | 2018-03-02 09:43:54 +0000 | [diff] [blame] | 83 | #ifndef USE_DEFAULT_BETA |
Michele Di Giorgio | bf3c662 | 2018-03-08 11:52:27 +0000 | [diff] [blame] | 84 | VECTOR_DECLARATION(beta), |
Michele Di Giorgio | 4d33630 | 2018-03-02 09:43:54 +0000 | [diff] [blame] | 85 | #endif /* USE_DEFAULT_BETA */ |
| 86 | #ifndef USE_DEFAULT_GAMMA |
Michele Di Giorgio | bf3c662 | 2018-03-08 11:52:27 +0000 | [diff] [blame] | 87 | VECTOR_DECLARATION(gamma), |
Michele Di Giorgio | 4d33630 | 2018-03-02 09:43:54 +0000 | [diff] [blame] | 88 | #endif /* USE_DEFAULT_GAMMA */ |
Michele Di Giorgio | bf3c662 | 2018-03-08 11:52:27 +0000 | [diff] [blame] | 89 | float epsilon) |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 90 | { |
Georgios Pinitas | 409ee0a | 2017-08-18 10:16:09 +0100 | [diff] [blame] | 91 | Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input); |
| 92 | #ifdef IN_PLACE |
| 93 | Tensor3D out = in; |
| 94 | #else /* IN_PLACE */ |
| 95 | Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output); |
| 96 | #endif /* IN_PLACE */ |
Michele Di Giorgio | 4d33630 | 2018-03-02 09:43:54 +0000 | [diff] [blame] | 97 | Vector mean = CONVERT_TO_VECTOR_STRUCT(mean); |
| 98 | Vector var = CONVERT_TO_VECTOR_STRUCT(var); |
| 99 | #ifndef USE_DEFAULT_BETA |
| 100 | Vector beta = CONVERT_TO_VECTOR_STRUCT(beta); |
| 101 | #endif /* USE_DEFAULT_BETA */ |
| 102 | #ifndef USE_DEFAULT_GAMMA |
Georgios Pinitas | 409ee0a | 2017-08-18 10:16:09 +0100 | [diff] [blame] | 103 | Vector gamma = CONVERT_TO_VECTOR_STRUCT(gamma); |
Michele Di Giorgio | 4d33630 | 2018-03-02 09:43:54 +0000 | [diff] [blame] | 104 | #endif /* USE_DEFAULT_GAMMA */ |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 105 | |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 106 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
Gian Marco Iodice | 349feef | 2017-09-28 11:21:29 +0100 | [diff] [blame] | 107 | data = 0; |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 108 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 109 | denominator = 0; |
| 110 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 111 | numerator = 0; |
| 112 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 113 | x_bar = 0; |
| 114 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
Michele Di Giorgio | 4d33630 | 2018-03-02 09:43:54 +0000 | [diff] [blame] | 115 | res = 0; |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 116 | |
| 117 | const int current_slice = get_global_id(2); |
| 118 | |
Gian Marco Iodice | 349feef | 2017-09-28 11:21:29 +0100 | [diff] [blame] | 119 | data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in.ptr); |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 120 | denominator = *((__global DATA_TYPE *)(var.ptr + current_slice * var.stride_x)); |
Gian Marco Iodice | 349feef | 2017-09-28 11:21:29 +0100 | [diff] [blame] | 121 | denominator = INVSQRT_OP(ADD_OP(denominator, ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(epsilon)))); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 122 | |
| 123 | // Calculate x bar and store results |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 124 | numerator = *((__global DATA_TYPE *)(mean.ptr + current_slice * mean.stride_x)); |
Gian Marco Iodice | 349feef | 2017-09-28 11:21:29 +0100 | [diff] [blame] | 125 | numerator = SUB_OP(data, numerator); |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 126 | x_bar = MUL_OP(numerator, denominator); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 127 | |
Michele Di Giorgio | 4d33630 | 2018-03-02 09:43:54 +0000 | [diff] [blame] | 128 | #ifndef USE_DEFAULT_GAMMA |
Giorgio Arena | 1167487 | 2018-02-07 15:38:12 +0000 | [diff] [blame] | 129 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
Michele Di Giorgio | 4d33630 | 2018-03-02 09:43:54 +0000 | [diff] [blame] | 130 | gamma_vec = *((__global DATA_TYPE *)(gamma.ptr + current_slice * gamma.stride_x)); |
| 131 | |
| 132 | res = MUL_OP(gamma_vec, x_bar); |
| 133 | #else /* USE_DEFAULT_GAMMA */ |
| 134 | // gamma is equal to 1, no need to perform multiplications |
Michele Di Giorgio | bf3c662 | 2018-03-08 11:52:27 +0000 | [diff] [blame] | 135 | res = x_bar; |
Michele Di Giorgio | 4d33630 | 2018-03-02 09:43:54 +0000 | [diff] [blame] | 136 | #endif /* USE_DEFAULT_GAMMA */ |
| 137 | |
| 138 | #ifndef USE_DEFAULT_BETA |
| 139 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 140 | beta_vec = *((__global DATA_TYPE *)(beta.ptr + current_slice * beta.stride_x)); |
| 141 | // beta is not zero, hence we need to perform the addition |
| 142 | res = ADD_OP(res, beta_vec); |
| 143 | #endif /* USE_DEFAULT_BETA */ |
Giorgio Arena | 1167487 | 2018-02-07 15:38:12 +0000 | [diff] [blame] | 144 | |
| 145 | res = ACTIVATION_FUNC(res); |
| 146 | |
Michalis Spyrou | 172e570 | 2017-06-26 14:18:47 +0100 | [diff] [blame] | 147 | VSTORE(VEC_SIZE) |
Giorgio Arena | 1167487 | 2018-02-07 15:38:12 +0000 | [diff] [blame] | 148 | (res, 0, (__global DATA_TYPE *)out.ptr); |
Anthony Barbier | 6ff3b19 | 2017-09-04 18:44:23 +0100 | [diff] [blame] | 149 | } |
Giorgio Arena | 1167487 | 2018-02-07 15:38:12 +0000 | [diff] [blame] | 150 | |
Michele Di Giorgio | bf3c662 | 2018-03-08 11:52:27 +0000 | [diff] [blame] | 151 | /** Apply batch normalization on tensors with NHWC format. |
| 152 | * |
Vidhya Sudhan Loganathan | 7485d5a | 2018-07-04 09:34:00 +0100 | [diff] [blame] | 153 | * @param[in] input_ptr Pointer to the first source tensor. Supported data types: F16/F32 |
Michele Di Giorgio | bf3c662 | 2018-03-08 11:52:27 +0000 | [diff] [blame] | 154 | * @param[in] input_stride_x Stride of the first source tensor in X dimension (in bytes) |
| 155 | * @param[in] input_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 156 | * @param[in] input_stride_y Stride of the first source tensor in Y dimension (in bytes) |
| 157 | * @param[in] input_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 158 | * @param[in] input_stride_z Stride of the first source tensor in Z dimension (in bytes) |
| 159 | * @param[in] input_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
| 160 | * @param[in] input_offset_first_element_in_bytes The offset of the first element in the first source tensor |
| 161 | * @param[out] output_ptr Pointer to the destination tensor. Supported data types: same as @p input_ptr |
| 162 | * @param[in] output_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 163 | * @param[in] output_step_x output_stride_x * number of elements along X processed per workitem(in bytes) |
| 164 | * @param[in] output_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 165 | * @param[in] output_step_y output_stride_y * number of elements along Y processed per workitem(in bytes) |
| 166 | * @param[in] output_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 167 | * @param[in] output_step_z output_stride_z * number of elements along Z processed per workitem(in bytes) |
| 168 | * @param[in] output_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 169 | * @param[in] mean_ptr Pointer to the mean source tensor. Supported data types: same as @p input_ptr |
| 170 | * @param[in] mean_stride_x Stride of the mean source tensor in X dimension (in bytes) |
| 171 | * @param[in] mean_step_x mean_stride_x * number of elements along X processed per workitem(in bytes) |
| 172 | * @param[in] mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor |
| 173 | * @param[in] var_ptr Pointer to the var tensor. Supported data types: same as @p input_ptr |
| 174 | * @param[in] var_stride_x Stride of the var tensor in X dimension (in bytes) |
| 175 | * @param[in] var_step_x var_stride_x * number of elements along X processed per workitem(in bytes) |
| 176 | * @param[in] var_offset_first_element_in_bytes The offset of the first element in the var source tensor |
| 177 | * @param[in] beta_ptr Pointer to the beta source tensor. Supported data types: same as @p input_ptr |
| 178 | * @param[in] beta_stride_x Stride of the beta source tensor in X dimension (in bytes) |
| 179 | * @param[in] beta_step_x beta_stride_x * number of elements along X processed per workitem(in bytes) |
| 180 | * @param[in] beta_offset_first_element_in_bytes The offset of the first element in the beta source tensor |
| 181 | * @param[in] gamma_ptr Pointer to the gamma source tensor. Supported data types: same as @p input_ptr |
| 182 | * @param[in] gamma_stride_x Stride of the gamma source tensor in X dimension (in bytes) |
| 183 | * @param[in] gamma_step_x gamma_stride_x * number of elements along X processed per workitem(in bytes) |
| 184 | * @param[in] gamma_offset_first_element_in_bytes The offset of the first element in the gamma source tensor |
| 185 | * @param[in] epsilon Epsilon parameter in the batch normalization equation |
| 186 | */ |
| 187 | __kernel void batchnormalization_layer_nhwc(TENSOR3D_DECLARATION(input), |
| 188 | #ifndef IN_PLACE |
| 189 | TENSOR3D_DECLARATION(output), |
| 190 | #endif /* not IN_PLACE */ |
| 191 | VECTOR_DECLARATION(mean), |
| 192 | VECTOR_DECLARATION(var), |
| 193 | #ifndef USE_DEFAULT_BETA |
| 194 | VECTOR_DECLARATION(beta), |
| 195 | #endif /* USE_DEFAULT_BETA */ |
| 196 | #ifndef USE_DEFAULT_GAMMA |
| 197 | VECTOR_DECLARATION(gamma), |
| 198 | #endif /* USE_DEFAULT_GAMMA */ |
| 199 | float epsilon) |
| 200 | { |
| 201 | Tensor3D in = CONVERT_TO_TENSOR3D_STRUCT(input); |
| 202 | #ifdef IN_PLACE |
| 203 | Tensor3D out = in; |
| 204 | #else /* IN_PLACE */ |
| 205 | Tensor3D out = CONVERT_TO_TENSOR3D_STRUCT(output); |
| 206 | #endif /* IN_PLACE */ |
| 207 | Vector mean = CONVERT_TO_VECTOR_STRUCT(mean); |
| 208 | Vector var = CONVERT_TO_VECTOR_STRUCT(var); |
| 209 | #ifndef USE_DEFAULT_BETA |
| 210 | Vector beta = CONVERT_TO_VECTOR_STRUCT(beta); |
| 211 | #endif /* USE_DEFAULT_BETA */ |
| 212 | #ifndef USE_DEFAULT_GAMMA |
| 213 | Vector gamma = CONVERT_TO_VECTOR_STRUCT(gamma); |
| 214 | #endif /* USE_DEFAULT_GAMMA */ |
| 215 | |
| 216 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 217 | data = 0; |
| 218 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 219 | denominator = 0; |
| 220 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 221 | numerator = 0; |
| 222 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 223 | x_bar = 0; |
| 224 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 225 | res = 0; |
| 226 | |
| 227 | const int current_slice = get_global_id(0); |
| 228 | |
| 229 | data = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)in.ptr); |
| 230 | denominator = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(var.ptr + current_slice * VEC_SIZE * var.stride_x)); |
| 231 | denominator = INVSQRT_OP(ADD_OP(denominator, ((VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE))SQCVT_SAT(epsilon)))); |
| 232 | |
| 233 | // Calculate x bar and store results |
| 234 | numerator = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(mean.ptr + current_slice * VEC_SIZE * mean.stride_x)); |
| 235 | numerator = SUB_OP(data, numerator); |
| 236 | x_bar = MUL_OP(numerator, denominator); |
| 237 | |
| 238 | #ifndef USE_DEFAULT_GAMMA |
| 239 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 240 | gamma_vec = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(gamma.ptr + current_slice * VEC_SIZE * gamma.stride_x)); |
| 241 | |
| 242 | res = MUL_OP(gamma_vec, x_bar); |
| 243 | #else /* USE_DEFAULT_GAMMA */ |
| 244 | // gamma is equal to 1, no need to perform multiplications |
| 245 | res = x_bar; |
| 246 | #endif /* USE_DEFAULT_GAMMA */ |
| 247 | |
| 248 | #ifndef USE_DEFAULT_BETA |
| 249 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 250 | beta_vec = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)(beta.ptr + current_slice * VEC_SIZE * beta.stride_x)); |
| 251 | // beta is not zero, hence we need to perform the addition |
| 252 | res = ADD_OP(res, beta_vec); |
| 253 | #endif /* USE_DEFAULT_BETA */ |
| 254 | |
| 255 | res = ACTIVATION_FUNC(res); |
| 256 | |
| 257 | VSTORE(VEC_SIZE) |
| 258 | (res, 0, (__global DATA_TYPE *)out.ptr); |
| 259 | } |
Michele Di Giorgio | 4d33630 | 2018-03-02 09:43:54 +0000 | [diff] [blame] | 260 | #endif /* defined(VEC_SIZE) && defined(DATA_TYPE) */ |
Georgios Pinitas | c936917 | 2018-09-26 11:25:40 +0100 | [diff] [blame] | 261 | |
| 262 | #if defined(NUM_CHANNELS) && defined(DATA_TYPE) && defined(EPSILON) |
| 263 | /** Fuse batchnorm parameters to convolution layer parameters |
| 264 | * |
| 265 | * @attention Data type should be passed using the -DDATA_TYPE compile flag, e.g. -DDATA_TYPE=float |
| 266 | * @attention Input tensor depth should be given as a preprocessor argument using -DNUM_CHANNELS=size. e.g. -DNUM_CHANNELS=16 |
| 267 | * @attention Batch normalization epsilon parameter should be given as a preprocessor argument with -DEPSILON=value. e.g. -DEPSILON=0.001f |
| 268 | * |
| 269 | * @param[in] conv_w_ptr Pointer to the source tensor. Supported data types: F16/F32 |
| 270 | * @param[in] conv_w_stride_x Stride of the source tensor in X dimension (in bytes) |
| 271 | * @param[in] conv_w_step_x input_stride_x * number of elements along X processed per workitem(in bytes) |
| 272 | * @param[in] conv_w_stride_y Stride of the source tensor in Y dimension (in bytes) |
| 273 | * @param[in] conv_w_step_y input_stride_y * number of elements along Y processed per workitem(in bytes) |
| 274 | * @param[in] conv_w_stride_z Stride of the source tensor in Z dimension (in bytes) |
| 275 | * @param[in] conv_w_step_z input_stride_z * number of elements along Z processed per workitem(in bytes) |
Georgios Pinitas | c55beee | 2018-10-23 15:23:23 +0100 | [diff] [blame] | 276 | * @param[in] conv_w_stride_w Stride of the source tensor in W dimension (in bytes) |
| 277 | * @param[in] conv_w_step_w input_stride_w * number of elements along W processed per workitem(in bytes) |
Georgios Pinitas | c936917 | 2018-09-26 11:25:40 +0100 | [diff] [blame] | 278 | * @param[in] conv_w_offset_first_element_in_bytes The offset of the first element in the source tensor |
| 279 | * @param[in] bn_mean_ptr Pointer to the mean source tensor. Supported data types: same as @p input_ptr |
| 280 | * @param[in] bn_mean_stride_x Stride of the mean source tensor in X dimension (in bytes) |
| 281 | * @param[in] bn_mean_step_x bn_mean_stride_x * number of elements along X processed per workitem(in bytes) |
| 282 | * @param[in] bn_mean_offset_first_element_in_bytes The offset of the first element in the mean source tensor |
| 283 | * @param[in] bn_var_ptr Pointer to the var tensor. Supported data types: same as @p input_ptr |
| 284 | * @param[in] bn_var_stride_x Stride of the var tensor in X dimension (in bytes) |
| 285 | * @param[in] bn_var_step_x bn_var_stride_x * number of elements along X processed per workitem(in bytes) |
| 286 | * @param[in] bn_var_offset_first_element_in_bytes The offset of the first element in the var source tensor |
| 287 | * @param[out] fused_w_ptr Pointer to the destination weights tensors. Supported data types: same as @p input_ptr |
| 288 | * @param[in] fused_w_stride_x Stride of the destination tensor in X dimension (in bytes) |
| 289 | * @param[in] fused_w_step_x fused_w_stride_x * number of elements along X processed per workitem(in bytes) |
| 290 | * @param[in] fused_w_stride_y Stride of the destination tensor in Y dimension (in bytes) |
| 291 | * @param[in] fused_w_step_y fused_w_stride_y * number of elements along Y processed per workitem(in bytes) |
| 292 | * @param[in] fused_w_stride_z Stride of the destination tensor in Z dimension (in bytes) |
| 293 | * @param[in] fused_w_step_z fused_w_stride_z * number of elements along Z processed per workitem(in bytes) |
| 294 | * @param[in] fused_w_stride_w Stride of the destination tensor in W dimension (in bytes) |
| 295 | * @param[in] fused_w_step_w fused_w_stride_w * number of elements along W processed per workitem(in bytes) |
| 296 | * @param[in] fused_w_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 297 | * @param[in] fused_b_ptr Pointer to the destination bias tensor. Supported data types: same as @p input_ptr |
| 298 | * @param[in] fused_b_stride_x Stride of the bias source tensor in X dimension (in bytes) |
| 299 | * @param[in] fused_b_step_x fused_b_stride_x * number of elements along X processed per workitem(in bytes) |
| 300 | * @param[in] fused_b_offset_first_element_in_bytes The offset of the first element in the destination tensor |
| 301 | * @param[in] conv_b_ptr Pointer to the source bias tensor. Supported data types: same as @p input_ptr |
| 302 | * @param[in] conv_b_stride_x Stride of the beta source tensor in X dimension (in bytes) |
| 303 | * @param[in] conv_b_step_x conv_b_beta_stride_x * number of elements along X processed per workitem(in bytes) |
| 304 | * @param[in] conv_b_offset_first_element_in_bytes The offset of the first element in the source bias tensor |
| 305 | * @param[in] bn_beta_ptr Pointer to the beta source tensor. Supported data types: same as @p input_ptr |
| 306 | * @param[in] bn_beta_stride_x Stride of the beta source tensor in X dimension (in bytes) |
| 307 | * @param[in] bn_beta_step_x bn_beta_stride_x * number of elements along X processed per workitem(in bytes) |
| 308 | * @param[in] bn_beta_offset_first_element_in_bytes The offset of the first element in the beta source tensor |
| 309 | * @param[in] bn_gamma_ptr Pointer to the gamma source tensor. Supported data types: same as @p input_ptr |
| 310 | * @param[in] bn_gamma_stride_x Stride of the gamma source tensor in X dimension (in bytes) |
| 311 | * @param[in] bn_gamma_step_x bn_gamma_stride_x * number of elements along X processed per workitem(in bytes) |
| 312 | * @param[in] bn_gamma_offset_first_element_in_bytes The offset of the first element in the gamma source tensor |
| 313 | * @param[in] epsilon Epsilon parameter in the batch normalization equation |
| 314 | */ |
| 315 | __kernel void fuse_batchnormalization_layer(TENSOR4D_DECLARATION(conv_w), |
| 316 | VECTOR_DECLARATION(bn_mean), |
| 317 | VECTOR_DECLARATION(bn_var) |
| 318 | #ifndef IN_PLACE_W |
| 319 | , |
| 320 | TENSOR4D_DECLARATION(fused_w) |
| 321 | #endif /* not IN_PLACE_W */ |
| 322 | #ifndef IN_PLACE_B |
| 323 | , |
| 324 | VECTOR_DECLARATION(fused_b) |
| 325 | #endif /* not IN_PLACE_B */ |
| 326 | #ifdef HAS_BIAS |
| 327 | , |
| 328 | VECTOR_DECLARATION(conv_b) |
| 329 | #endif /* HAS_BIAS */ |
| 330 | #ifndef USE_DEFAULT_BETA |
| 331 | , |
| 332 | VECTOR_DECLARATION(bn_beta) |
| 333 | #endif /* USE_DEFAULT_BETA */ |
| 334 | #ifndef USE_DEFAULT_GAMMA |
| 335 | , |
| 336 | VECTOR_DECLARATION(bn_gamma) |
| 337 | #endif /* USE_DEFAULT_GAMMA */ |
| 338 | ) |
| 339 | { |
| 340 | Tensor4D conv_w = CONVERT_TO_TENSOR4D_STRUCT(conv_w, NUM_CHANNELS); |
| 341 | Vector bn_mean = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bn_mean); |
| 342 | Vector bn_var = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bn_var); |
| 343 | |
| 344 | // In-place ops |
| 345 | #ifdef IN_PLACE_W |
| 346 | Tensor4D fused_w = conv_w; |
| 347 | #else /* IN_PLACE_W */ |
| 348 | Tensor4D fused_w = CONVERT_TO_TENSOR4D_STRUCT(fused_w, NUM_CHANNELS); |
| 349 | #endif /* IN_PLACE */ |
| 350 | #ifdef IN_PLACE_B |
| 351 | Vector fused_b = conv_b; |
| 352 | #else /* IN_PLACE_W */ |
| 353 | Vector fused_b = CONVERT_TO_VECTOR_STRUCT_NO_STEP(fused_b); |
| 354 | #endif /* IN_PLACE */ |
| 355 | |
| 356 | // Conditional ops |
| 357 | #ifdef HAS_BIAS |
| 358 | Vector conv_b = CONVERT_TO_VECTOR_STRUCT_NO_STEP(conv_b); |
| 359 | #endif /* USE_DEFAULT_BETA */ |
| 360 | #ifndef USE_DEFAULT_BETA |
| 361 | Vector bn_beta = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bn_beta); |
| 362 | #endif /* USE_DEFAULT_BETA */ |
| 363 | #ifndef USE_DEFAULT_GAMMA |
| 364 | Vector bn_gamma = CONVERT_TO_VECTOR_STRUCT_NO_STEP(bn_gamma); |
| 365 | #endif /* USE_DEFAULT_GAMMA */ |
| 366 | |
| 367 | const int current_slice = get_global_id(2) / NUM_CHANNELS; |
| 368 | |
| 369 | #if defined(VEC_SIZE) && defined(LAST_ACCESSED_X) |
| 370 | // Check if access on width gets out of bounds |
| 371 | // If it does shift access vector to access elements within bounds |
| 372 | const int xi = (int)(get_global_id(0) * VEC_SIZE); |
| 373 | conv_w.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * conv_w_stride_x; |
| 374 | fused_w.ptr -= max(xi - (int)LAST_ACCESSED_X, 0) * fused_w_stride_x; |
| 375 | |
| 376 | // Load W |
| 377 | VEC_DATA_TYPE(DATA_TYPE, VEC_SIZE) |
| 378 | wn = VLOAD(VEC_SIZE)(0, (__global DATA_TYPE *)conv_w.ptr); |
| 379 | #else // !defined(VEC_SIZE) || !defined(LAST_ACCESSED_X) |
| 380 | DATA_TYPE wn = *((__global DATA_TYPE *)(conv_w.ptr)); |
| 381 | #endif // defined(VEC_SIZE) && defined(LAST_ACCESSED_X) |
| 382 | |
| 383 | // rvar = 1 / sqrt(var + epsilon) |
| 384 | const DATA_TYPE var = *((__global DATA_TYPE *)(bn_var.ptr + current_slice * bn_var.stride_x)); |
| 385 | const DATA_TYPE rvar = INVSQRT_OP(ADD_OP(var, SQCVT_SAT((float)EPSILON))); |
| 386 | wn *= rvar; |
| 387 | |
| 388 | // Load b |
| 389 | const DATA_TYPE mean = *((__global DATA_TYPE *)(bn_mean.ptr + current_slice * bn_mean.stride_x)); |
| 390 | DATA_TYPE bn = 0; |
| 391 | #ifdef HAS_BIAS |
| 392 | bn = *((__global DATA_TYPE *)(conv_b.ptr + current_slice * conv_b.stride_x)); |
| 393 | #endif /* HAS_BIAS */ |
| 394 | bn = (bn - mean) * rvar; |
| 395 | |
| 396 | #ifndef USE_DEFAULT_GAMMA |
| 397 | const DATA_TYPE gamma_scalar = *((__global DATA_TYPE *)(bn_gamma.ptr + current_slice * bn_gamma.stride_x)); |
| 398 | wn *= gamma_scalar; |
| 399 | bn *= gamma_scalar; |
| 400 | #endif /* USE_DEFAULT_GAMMA */ |
| 401 | |
| 402 | #ifndef USE_DEFAULT_BETA |
| 403 | const DATA_TYPE beta_scalar = *((__global DATA_TYPE *)(bn_beta.ptr + current_slice * bn_beta.stride_x)); |
| 404 | bn += beta_scalar; |
| 405 | #endif /* USE_DEFAULT_BETA */ |
| 406 | |
| 407 | #if defined(VEC_SIZE) && defined(LAST_ACCESSED_X) |
| 408 | // Store updated weights |
| 409 | VSTORE(VEC_SIZE) |
| 410 | (wn, 0, (__global DATA_TYPE *)fused_w.ptr); |
| 411 | #else // !defined(VEC_SIZE) || !defined(LAST_ACCESSED_X) |
| 412 | *((__global DATA_TYPE *)(fused_w.ptr)) = wn; |
| 413 | #endif // defined(VEC_SIZE) && defined(LAST_ACCESSED_X) |
| 414 | |
| 415 | // Store updated bias |
| 416 | *((__global DATA_TYPE *)(fused_b.ptr + current_slice * fused_b.stride_x)) = bn; |
| 417 | } |
| 418 | #endif /* defined(NUM_CHANNELS) && defined(DATA_TYPE) && defined(EPSILON) */ |