1 /* 2 * Copyright (c) 2021-2023 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 #ifndef SRC_CORE_CL_CL_KERNELS_TILE_HELPERS 25 #define SRC_CORE_CL_CL_KERNELS_TILE_HELPERS 26 27 // *INDENT-OFF* 28 // clang-format off 29 30 #define TILE_VECTOR_SIZE1 1 31 #define TILE_VECTOR_SIZE2 2 32 #define TILE_VECTOR_SIZE3 3 33 #define TILE_VECTOR_SIZE4 4 34 #define TILE_VECTOR_SIZE5 8 35 #define TILE_VECTOR_SIZE6 8 36 #define TILE_VECTOR_SIZE7 8 37 #define TILE_VECTOR_SIZE8 8 38 #define TILE_VECTOR_SIZE9 16 39 #define TILE_VECTOR_SIZE10 16 40 #define TILE_VECTOR_SIZE11 16 41 #define TILE_VECTOR_SIZE12 16 42 #define TILE_VECTOR_SIZE13 16 43 #define TILE_VECTOR_SIZE14 16 44 #define TILE_VECTOR_SIZE15 16 45 #define TILE_VECTOR_SIZE16 16 46 47 #define TILE_VECTOR_TYPE1(DATA_TYPE) DATA_TYPE##1 48 #define TILE_VECTOR_TYPE2(DATA_TYPE) DATA_TYPE##2 49 #define TILE_VECTOR_TYPE3(DATA_TYPE) DATA_TYPE##3 50 #define TILE_VECTOR_TYPE4(DATA_TYPE) DATA_TYPE##4 51 #define TILE_VECTOR_TYPE5(DATA_TYPE) DATA_TYPE##8 52 #define TILE_VECTOR_TYPE6(DATA_TYPE) DATA_TYPE##8 53 #define TILE_VECTOR_TYPE7(DATA_TYPE) DATA_TYPE##8 54 #define TILE_VECTOR_TYPE8(DATA_TYPE) DATA_TYPE##8 55 #define TILE_VECTOR_TYPE9(DATA_TYPE) DATA_TYPE##16 56 #define TILE_VECTOR_TYPE10(DATA_TYPE) DATA_TYPE##16 57 #define TILE_VECTOR_TYPE11(DATA_TYPE) DATA_TYPE##16 58 #define TILE_VECTOR_TYPE12(DATA_TYPE) DATA_TYPE##16 59 #define TILE_VECTOR_TYPE13(DATA_TYPE) DATA_TYPE##16 60 #define TILE_VECTOR_TYPE14(DATA_TYPE) DATA_TYPE##16 61 #define TILE_VECTOR_TYPE15(DATA_TYPE) DATA_TYPE##16 62 #define TILE_VECTOR_TYPE16(DATA_TYPE) DATA_TYPE##16 63 64 /** Tile object 65 * A tile object is a 2D memory block and can be accessed using the following syntax: 66 * -# a[m0].v = access the the vector at row "m0" (OpenCL vector) 67 * -# dst[m0].s[n0] = access the scalar element at row "m0" and column "n0" (scalar access) 68 * 69 * @param[in] DATA_TYPE Data type of the tile 70 * @param[in] H Number of tile rows 71 * @param[in] W Number of tile colums 72 * @param[in] BASENAME Tile's name 73 */ 74 #define TILE(DATA_TYPE, H, W, BASENAME) TILE_STR(DATA_TYPE, H, W, BASENAME) 75 #define TILE_STR(DATA_TYPE, H, W, BASENAME) \ 76 union { \ 77 DATA_TYPE s[TILE_VECTOR_SIZE##W]; \ 78 TILE_VECTOR_TYPE##W(DATA_TYPE) v; \ 79 } BASENAME[H] 80 81 #define TENSOR4D_IMAGE(name) \ 82 __read_only image2d_t name##_img, \ 83 __global uchar *name##_ptr, \ 84 uint name##_stride_x, \ 85 uint name##_step_x, \ 86 uint name##_stride_y, \ 87 uint name##_step_y, \ 88 uint name##_stride_z, \ 89 uint name##_step_z, \ 90 uint name##_stride_w, \ 91 uint name##_step_w, \ 92 uint name##_offset_first_element_in_bytes 93 94 #define TENSOR4D_BUFFER(name) \ 95 __global uchar *name##_ptr, \ 96 uint name##_stride_x, \ 97 uint name##_step_x, \ 98 uint name##_stride_y, \ 99 uint name##_step_y, \ 100 uint name##_stride_z, \ 101 uint name##_step_z, \ 102 uint name##_stride_w, \ 103 uint name##_step_w, \ 104 uint name##_offset_first_element_in_bytes 105 106 #define TENSOR4D_STR(name, type) TENSOR4D_##type(name) 107 #define TENSOR4D(name, type) TENSOR4D_STR(name, type) 108 109 #define TENSOR4D_T_IMAGE(name) \ 110 __read_only image2d_t name##_img, \ 111 __global uchar *name##_ptr, \ 112 uint name##_stride_y, \ 113 uint name##_stride_z, \ 114 uint name##_stride_w, \ 115 uint name##_c, \ 116 uint name##_w, \ 117 uint name##_h, \ 118 uint name##_n, \ 119 uint name##_offset_first_element_in_bytes 120 121 #define TENSOR4D_T_BUFFER(name) \ 122 __global uchar *name##_ptr, \ 123 uint name##_stride_y, \ 124 uint name##_stride_z, \ 125 uint name##_stride_w, \ 126 uint name##_c, \ 127 uint name##_w, \ 128 uint name##_h, \ 129 uint name##_n, \ 130 uint name##_offset_first_element_in_bytes 131 132 #define TENSOR4D_T_STR(name, type) TENSOR4D_T_##type(name) 133 134 /** Legacy tensor 4D arguments 135 * 136 * @param[in] name Tensor name. The tensor name is the prefix of the tensor components 137 * @param[in] type Tensor type (BUFFER or IMAGE) 138 */ 139 #define TENSOR4D_T(name, type) TENSOR4D_T_STR(name, type) 140 141 #define TENSOR4D_RO_T_IMAGE(name) \ 142 __read_only image2d_t name##_img, \ 143 TENSOR4D_T_BUFFER(name) 144 145 #define TENSOR4D_RO_T_BUFFER(name) TENSOR4D_T_BUFFER(name) 146 147 #define TENSOR4D_RO_T_STR(name, type) TENSOR4D_RO_T_##type(name) 148 149 /** Read-Only (RO) tensor 4D. 150 * 151 * @param[in] name Tensor name. The tensor name is the prefix of the tensor components 152 * @param[in] type Tensor type (BUFFER or IMAGE) 153 */ 154 #define TENSOR4D_RO_T(name, type) TENSOR4D_RO_T_STR(name, type) 155 156 #define TENSOR4D_WO_T_IMAGE(name) \ 157 __write_only image2d_t name##_img, \ 158 TENSOR4D_T_BUFFER(name) 159 160 #define TENSOR4D_WO_T_BUFFER(name) TENSOR4D_T_BUFFER(name) 161 162 #define TENSOR4D_WO_T_STR(name, type) TENSOR4D_WO_T_##type(name) 163 164 /** Write-Only (WO) tensor 4D. 165 * 166 * @param[in] name Tensor name. The tensor name is the prefix of the tensor components 167 * @param[in] type Tensor type (BUFFER or IMAGE) 168 */ 169 #define TENSOR4D_WO_T(name, type) TENSOR4D_WO_T_STR(name, type) 170 171 #define TENSOR3D_T_IMAGE(name) \ 172 __read_only image2d_t name##_img, \ 173 __global uchar *name##_ptr, \ 174 uint name##_stride_y, \ 175 uint name##_stride_z, \ 176 uint name##_w, \ 177 uint name##_h, \ 178 uint name##_n, \ 179 uint name##_offset_first_element_in_bytes 180 181 #define TENSOR3D_T_BUFFER(name) \ 182 __global uchar *name##_ptr, \ 183 uint name##_stride_y, \ 184 uint name##_stride_z, \ 185 uint name##_w, \ 186 uint name##_h, \ 187 uint name##_n, \ 188 uint name##_offset_first_element_in_bytes 189 190 #define TENSOR3D_T_STR(name, type) TENSOR3D_T_##type(name) 191 #define TENSOR3D_T(name, type) TENSOR3D_T_STR(name, type) 192 193 #if !defined(UNROLL_WITH_PRAGMA) 194 #define UNROLL_INCR(idx, step, macro) idx += (step); (macro) 195 196 #define LOOP_UNROLLING_1(idx, step, macro) (macro) 197 #define LOOP_UNROLLING_2(idx, step, macro) LOOP_UNROLLING_1(idx, step, macro); UNROLL_INCR(idx, step, macro) 198 #define LOOP_UNROLLING_3(idx, step, macro) LOOP_UNROLLING_2(idx, step, macro); UNROLL_INCR(idx, step, macro) 199 #define LOOP_UNROLLING_4(idx, step, macro) LOOP_UNROLLING_3(idx, step, macro); UNROLL_INCR(idx, step, macro) 200 #define LOOP_UNROLLING_5(idx, step, macro) LOOP_UNROLLING_4(idx, step, macro); UNROLL_INCR(idx, step, macro) 201 #define LOOP_UNROLLING_6(idx, step, macro) LOOP_UNROLLING_5(idx, step, macro); UNROLL_INCR(idx, step, macro) 202 #define LOOP_UNROLLING_7(idx, step, macro) LOOP_UNROLLING_6(idx, step, macro); UNROLL_INCR(idx, step, macro) 203 #define LOOP_UNROLLING_8(idx, step, macro) LOOP_UNROLLING_7(idx, step, macro); UNROLL_INCR(idx, step, macro) 204 #define LOOP_UNROLLING_9(idx, step, macro) LOOP_UNROLLING_8(idx, step, macro); UNROLL_INCR(idx, step, macro) 205 #define LOOP_UNROLLING_10(idx, step, macro) LOOP_UNROLLING_9(idx, step, macro); UNROLL_INCR(idx, step, macro) 206 #define LOOP_UNROLLING_11(idx, step, macro) LOOP_UNROLLING_10(idx, step, macro); UNROLL_INCR(idx, step, macro) 207 #define LOOP_UNROLLING_12(idx, step, macro) LOOP_UNROLLING_11(idx, step, macro); UNROLL_INCR(idx, step, macro) 208 #define LOOP_UNROLLING_13(idx, step, macro) LOOP_UNROLLING_12(idx, step, macro); UNROLL_INCR(idx, step, macro) 209 #define LOOP_UNROLLING_14(idx, step, macro) LOOP_UNROLLING_13(idx, step, macro); UNROLL_INCR(idx, step, macro) 210 #define LOOP_UNROLLING_15(idx, step, macro) LOOP_UNROLLING_14(idx, step, macro); UNROLL_INCR(idx, step, macro) 211 #define LOOP_UNROLLING_16(idx, step, macro) LOOP_UNROLLING_15(idx, step, macro); UNROLL_INCR(idx, step, macro) 212 #define LOOP_UNROLLING_17(idx, step, macro) LOOP_UNROLLING_16(idx, step, macro); UNROLL_INCR(idx, step, macro) 213 #define LOOP_UNROLLING_18(idx, step, macro) LOOP_UNROLLING_17(idx, step, macro); UNROLL_INCR(idx, step, macro) 214 #define LOOP_UNROLLING_19(idx, step, macro) LOOP_UNROLLING_18(idx, step, macro); UNROLL_INCR(idx, step, macro) 215 #define LOOP_UNROLLING_20(idx, step, macro) LOOP_UNROLLING_19(idx, step, macro); UNROLL_INCR(idx, step, macro) 216 #define LOOP_UNROLLING_21(idx, step, macro) LOOP_UNROLLING_20(idx, step, macro); UNROLL_INCR(idx, step, macro) 217 #define LOOP_UNROLLING_22(idx, step, macro) LOOP_UNROLLING_21(idx, step, macro); UNROLL_INCR(idx, step, macro) 218 #define LOOP_UNROLLING_23(idx, step, macro) LOOP_UNROLLING_22(idx, step, macro); UNROLL_INCR(idx, step, macro) 219 #define LOOP_UNROLLING_24(idx, step, macro) LOOP_UNROLLING_23(idx, step, macro); UNROLL_INCR(idx, step, macro) 220 #define LOOP_UNROLLING_25(idx, step, macro) LOOP_UNROLLING_24(idx, step, macro); UNROLL_INCR(idx, step, macro) 221 #define LOOP_UNROLLING_26(idx, step, macro) LOOP_UNROLLING_25(idx, step, macro); UNROLL_INCR(idx, step, macro) 222 #define LOOP_UNROLLING_27(idx, step, macro) LOOP_UNROLLING_26(idx, step, macro); UNROLL_INCR(idx, step, macro) 223 #define LOOP_UNROLLING_28(idx, step, macro) LOOP_UNROLLING_27(idx, step, macro); UNROLL_INCR(idx, step, macro) 224 #define LOOP_UNROLLING_29(idx, step, macro) LOOP_UNROLLING_28(idx, step, macro); UNROLL_INCR(idx, step, macro) 225 #define LOOP_UNROLLING_30(idx, step, macro) LOOP_UNROLLING_29(idx, step, macro); UNROLL_INCR(idx, step, macro) 226 #define LOOP_UNROLLING_31(idx, step, macro) LOOP_UNROLLING_30(idx, step, macro); UNROLL_INCR(idx, step, macro) 227 #define LOOP_UNROLLING_32(idx, step, macro) LOOP_UNROLLING_31(idx, step, macro); UNROLL_INCR(idx, step, macro) 228 #define LOOP_UNROLLING_33(idx, step, macro) LOOP_UNROLLING_32(idx, step, macro); UNROLL_INCR(idx, step, macro) 229 #define LOOP_UNROLLING_34(idx, step, macro) LOOP_UNROLLING_33(idx, step, macro); UNROLL_INCR(idx, step, macro) 230 #define LOOP_UNROLLING_35(idx, step, macro) LOOP_UNROLLING_34(idx, step, macro); UNROLL_INCR(idx, step, macro) 231 #define LOOP_UNROLLING_36(idx, step, macro) LOOP_UNROLLING_35(idx, step, macro); UNROLL_INCR(idx, step, macro) 232 #define LOOP_UNROLLING_37(idx, step, macro) LOOP_UNROLLING_36(idx, step, macro); UNROLL_INCR(idx, step, macro) 233 #define LOOP_UNROLLING_38(idx, step, macro) LOOP_UNROLLING_37(idx, step, macro); UNROLL_INCR(idx, step, macro) 234 #define LOOP_UNROLLING_39(idx, step, macro) LOOP_UNROLLING_38(idx, step, macro); UNROLL_INCR(idx, step, macro) 235 #define LOOP_UNROLLING_40(idx, step, macro) LOOP_UNROLLING_39(idx, step, macro); UNROLL_INCR(idx, step, macro) 236 #define LOOP_UNROLLING_41(idx, step, macro) LOOP_UNROLLING_40(idx, step, macro); UNROLL_INCR(idx, step, macro) 237 #define LOOP_UNROLLING_42(idx, step, macro) LOOP_UNROLLING_41(idx, step, macro); UNROLL_INCR(idx, step, macro) 238 #define LOOP_UNROLLING_43(idx, step, macro) LOOP_UNROLLING_42(idx, step, macro); UNROLL_INCR(idx, step, macro) 239 #define LOOP_UNROLLING_44(idx, step, macro) LOOP_UNROLLING_43(idx, step, macro); UNROLL_INCR(idx, step, macro) 240 #define LOOP_UNROLLING_45(idx, step, macro) LOOP_UNROLLING_44(idx, step, macro); UNROLL_INCR(idx, step, macro) 241 #define LOOP_UNROLLING_46(idx, step, macro) LOOP_UNROLLING_45(idx, step, macro); UNROLL_INCR(idx, step, macro) 242 #define LOOP_UNROLLING_47(idx, step, macro) LOOP_UNROLLING_46(idx, step, macro); UNROLL_INCR(idx, step, macro) 243 #define LOOP_UNROLLING_48(idx, step, macro) LOOP_UNROLLING_47(idx, step, macro); UNROLL_INCR(idx, step, macro) 244 #define LOOP_UNROLLING_49(idx, step, macro) LOOP_UNROLLING_48(idx, step, macro); UNROLL_INCR(idx, step, macro) 245 #define LOOP_UNROLLING_50(idx, step, macro) LOOP_UNROLLING_49(idx, step, macro); UNROLL_INCR(idx, step, macro) 246 #define LOOP_UNROLLING_51(idx, step, macro) LOOP_UNROLLING_50(idx, step, macro); UNROLL_INCR(idx, step, macro) 247 #define LOOP_UNROLLING_52(idx, step, macro) LOOP_UNROLLING_51(idx, step, macro); UNROLL_INCR(idx, step, macro) 248 #define LOOP_UNROLLING_53(idx, step, macro) LOOP_UNROLLING_52(idx, step, macro); UNROLL_INCR(idx, step, macro) 249 #define LOOP_UNROLLING_54(idx, step, macro) LOOP_UNROLLING_53(idx, step, macro); UNROLL_INCR(idx, step, macro) 250 #define LOOP_UNROLLING_55(idx, step, macro) LOOP_UNROLLING_54(idx, step, macro); UNROLL_INCR(idx, step, macro) 251 #define LOOP_UNROLLING_56(idx, step, macro) LOOP_UNROLLING_55(idx, step, macro); UNROLL_INCR(idx, step, macro) 252 #define LOOP_UNROLLING_57(idx, step, macro) LOOP_UNROLLING_56(idx, step, macro); UNROLL_INCR(idx, step, macro) 253 #define LOOP_UNROLLING_58(idx, step, macro) LOOP_UNROLLING_57(idx, step, macro); UNROLL_INCR(idx, step, macro) 254 #define LOOP_UNROLLING_59(idx, step, macro) LOOP_UNROLLING_58(idx, step, macro); UNROLL_INCR(idx, step, macro) 255 #define LOOP_UNROLLING_60(idx, step, macro) LOOP_UNROLLING_59(idx, step, macro); UNROLL_INCR(idx, step, macro) 256 #define LOOP_UNROLLING_61(idx, step, macro) LOOP_UNROLLING_60(idx, step, macro); UNROLL_INCR(idx, step, macro) 257 #define LOOP_UNROLLING_62(idx, step, macro) LOOP_UNROLLING_61(idx, step, macro); UNROLL_INCR(idx, step, macro) 258 #define LOOP_UNROLLING_63(idx, step, macro) LOOP_UNROLLING_62(idx, step, macro); UNROLL_INCR(idx, step, macro) 259 #define LOOP_UNROLLING_64(idx, step, macro) LOOP_UNROLLING_63(idx, step, macro); UNROLL_INCR(idx, step, macro) 260 #define LOOP_UNROLLING_65(idx, step, macro) LOOP_UNROLLING_64(idx, step, macro); UNROLL_INCR(idx, step, macro) 261 #define LOOP_UNROLLING_66(idx, step, macro) LOOP_UNROLLING_65(idx, step, macro); UNROLL_INCR(idx, step, macro) 262 #define LOOP_UNROLLING_67(idx, step, macro) LOOP_UNROLLING_66(idx, step, macro); UNROLL_INCR(idx, step, macro) 263 #define LOOP_UNROLLING_68(idx, step, macro) LOOP_UNROLLING_67(idx, step, macro); UNROLL_INCR(idx, step, macro) 264 #define LOOP_UNROLLING_69(idx, step, macro) LOOP_UNROLLING_68(idx, step, macro); UNROLL_INCR(idx, step, macro) 265 #define LOOP_UNROLLING_70(idx, step, macro) LOOP_UNROLLING_69(idx, step, macro); UNROLL_INCR(idx, step, macro) 266 #define LOOP_UNROLLING_71(idx, step, macro) LOOP_UNROLLING_70(idx, step, macro); UNROLL_INCR(idx, step, macro) 267 #define LOOP_UNROLLING_72(idx, step, macro) LOOP_UNROLLING_71(idx, step, macro); UNROLL_INCR(idx, step, macro) 268 #define LOOP_UNROLLING_73(idx, step, macro) LOOP_UNROLLING_72(idx, step, macro); UNROLL_INCR(idx, step, macro) 269 #define LOOP_UNROLLING_74(idx, step, macro) LOOP_UNROLLING_73(idx, step, macro); UNROLL_INCR(idx, step, macro) 270 #define LOOP_UNROLLING_75(idx, step, macro) LOOP_UNROLLING_74(idx, step, macro); UNROLL_INCR(idx, step, macro) 271 #define LOOP_UNROLLING_76(idx, step, macro) LOOP_UNROLLING_75(idx, step, macro); UNROLL_INCR(idx, step, macro) 272 #define LOOP_UNROLLING_77(idx, step, macro) LOOP_UNROLLING_76(idx, step, macro); UNROLL_INCR(idx, step, macro) 273 #define LOOP_UNROLLING_78(idx, step, macro) LOOP_UNROLLING_77(idx, step, macro); UNROLL_INCR(idx, step, macro) 274 #define LOOP_UNROLLING_79(idx, step, macro) LOOP_UNROLLING_78(idx, step, macro); UNROLL_INCR(idx, step, macro) 275 #define LOOP_UNROLLING_80(idx, step, macro) LOOP_UNROLLING_79(idx, step, macro); UNROLL_INCR(idx, step, macro) 276 #define LOOP_UNROLLING_81(idx, step, macro) LOOP_UNROLLING_80(idx, step, macro); UNROLL_INCR(idx, step, macro) 277 #define LOOP_UNROLLING_82(idx, step, macro) LOOP_UNROLLING_81(idx, step, macro); UNROLL_INCR(idx, step, macro) 278 #define LOOP_UNROLLING_83(idx, step, macro) LOOP_UNROLLING_82(idx, step, macro); UNROLL_INCR(idx, step, macro) 279 #define LOOP_UNROLLING_84(idx, step, macro) LOOP_UNROLLING_83(idx, step, macro); UNROLL_INCR(idx, step, macro) 280 #define LOOP_UNROLLING_85(idx, step, macro) LOOP_UNROLLING_84(idx, step, macro); UNROLL_INCR(idx, step, macro) 281 #define LOOP_UNROLLING_86(idx, step, macro) LOOP_UNROLLING_85(idx, step, macro); UNROLL_INCR(idx, step, macro) 282 #define LOOP_UNROLLING_87(idx, step, macro) LOOP_UNROLLING_86(idx, step, macro); UNROLL_INCR(idx, step, macro) 283 #define LOOP_UNROLLING_88(idx, step, macro) LOOP_UNROLLING_87(idx, step, macro); UNROLL_INCR(idx, step, macro) 284 #define LOOP_UNROLLING_89(idx, step, macro) LOOP_UNROLLING_88(idx, step, macro); UNROLL_INCR(idx, step, macro) 285 #define LOOP_UNROLLING_90(idx, step, macro) LOOP_UNROLLING_89(idx, step, macro); UNROLL_INCR(idx, step, macro) 286 #define LOOP_UNROLLING_91(idx, step, macro) LOOP_UNROLLING_90(idx, step, macro); UNROLL_INCR(idx, step, macro) 287 #define LOOP_UNROLLING_92(idx, step, macro) LOOP_UNROLLING_91(idx, step, macro); UNROLL_INCR(idx, step, macro) 288 #define LOOP_UNROLLING_93(idx, step, macro) LOOP_UNROLLING_92(idx, step, macro); UNROLL_INCR(idx, step, macro) 289 #define LOOP_UNROLLING_94(idx, step, macro) LOOP_UNROLLING_93(idx, step, macro); UNROLL_INCR(idx, step, macro) 290 #define LOOP_UNROLLING_95(idx, step, macro) LOOP_UNROLLING_94(idx, step, macro); UNROLL_INCR(idx, step, macro) 291 #define LOOP_UNROLLING_96(idx, step, macro) LOOP_UNROLLING_95(idx, step, macro); UNROLL_INCR(idx, step, macro) 292 #define LOOP_UNROLLING_97(idx, step, macro) LOOP_UNROLLING_96(idx, step, macro); UNROLL_INCR(idx, step, macro) 293 #define LOOP_UNROLLING_98(idx, step, macro) LOOP_UNROLLING_97(idx, step, macro); UNROLL_INCR(idx, step, macro) 294 #define LOOP_UNROLLING_99(idx, step, macro) LOOP_UNROLLING_98(idx, step, macro); UNROLL_INCR(idx, step, macro) 295 #define LOOP_UNROLLING_100(idx, step, macro) LOOP_UNROLLING_99(idx, step, macro); UNROLL_INCR(idx, step, macro) 296 #define LOOP_UNROLLING_101(idx, step, macro) LOOP_UNROLLING_100(idx, step, macro); UNROLL_INCR(idx, step, macro) 297 #define LOOP_UNROLLING_102(idx, step, macro) LOOP_UNROLLING_101(idx, step, macro); UNROLL_INCR(idx, step, macro) 298 #define LOOP_UNROLLING_103(idx, step, macro) LOOP_UNROLLING_102(idx, step, macro); UNROLL_INCR(idx, step, macro) 299 #define LOOP_UNROLLING_104(idx, step, macro) LOOP_UNROLLING_103(idx, step, macro); UNROLL_INCR(idx, step, macro) 300 #define LOOP_UNROLLING_105(idx, step, macro) LOOP_UNROLLING_104(idx, step, macro); UNROLL_INCR(idx, step, macro) 301 #define LOOP_UNROLLING_106(idx, step, macro) LOOP_UNROLLING_105(idx, step, macro); UNROLL_INCR(idx, step, macro) 302 #define LOOP_UNROLLING_107(idx, step, macro) LOOP_UNROLLING_106(idx, step, macro); UNROLL_INCR(idx, step, macro) 303 #define LOOP_UNROLLING_108(idx, step, macro) LOOP_UNROLLING_107(idx, step, macro); UNROLL_INCR(idx, step, macro) 304 #define LOOP_UNROLLING_109(idx, step, macro) LOOP_UNROLLING_108(idx, step, macro); UNROLL_INCR(idx, step, macro) 305 #define LOOP_UNROLLING_110(idx, step, macro) LOOP_UNROLLING_109(idx, step, macro); UNROLL_INCR(idx, step, macro) 306 #define LOOP_UNROLLING_111(idx, step, macro) LOOP_UNROLLING_110(idx, step, macro); UNROLL_INCR(idx, step, macro) 307 #define LOOP_UNROLLING_112(idx, step, macro) LOOP_UNROLLING_111(idx, step, macro); UNROLL_INCR(idx, step, macro) 308 #define LOOP_UNROLLING_113(idx, step, macro) LOOP_UNROLLING_112(idx, step, macro); UNROLL_INCR(idx, step, macro) 309 #define LOOP_UNROLLING_114(idx, step, macro) LOOP_UNROLLING_113(idx, step, macro); UNROLL_INCR(idx, step, macro) 310 #define LOOP_UNROLLING_115(idx, step, macro) LOOP_UNROLLING_114(idx, step, macro); UNROLL_INCR(idx, step, macro) 311 #define LOOP_UNROLLING_116(idx, step, macro) LOOP_UNROLLING_115(idx, step, macro); UNROLL_INCR(idx, step, macro) 312 #define LOOP_UNROLLING_117(idx, step, macro) LOOP_UNROLLING_116(idx, step, macro); UNROLL_INCR(idx, step, macro) 313 #define LOOP_UNROLLING_118(idx, step, macro) LOOP_UNROLLING_117(idx, step, macro); UNROLL_INCR(idx, step, macro) 314 #define LOOP_UNROLLING_119(idx, step, macro) LOOP_UNROLLING_118(idx, step, macro); UNROLL_INCR(idx, step, macro) 315 #define LOOP_UNROLLING_120(idx, step, macro) LOOP_UNROLLING_119(idx, step, macro); UNROLL_INCR(idx, step, macro) 316 #define LOOP_UNROLLING_121(idx, step, macro) LOOP_UNROLLING_120(idx, step, macro); UNROLL_INCR(idx, step, macro) 317 #define LOOP_UNROLLING_122(idx, step, macro) LOOP_UNROLLING_121(idx, step, macro); UNROLL_INCR(idx, step, macro) 318 #define LOOP_UNROLLING_123(idx, step, macro) LOOP_UNROLLING_122(idx, step, macro); UNROLL_INCR(idx, step, macro) 319 #define LOOP_UNROLLING_124(idx, step, macro) LOOP_UNROLLING_123(idx, step, macro); UNROLL_INCR(idx, step, macro) 320 #define LOOP_UNROLLING_125(idx, step, macro) LOOP_UNROLLING_124(idx, step, macro); UNROLL_INCR(idx, step, macro) 321 #define LOOP_UNROLLING_126(idx, step, macro) LOOP_UNROLLING_125(idx, step, macro); UNROLL_INCR(idx, step, macro) 322 #define LOOP_UNROLLING_127(idx, step, macro) LOOP_UNROLLING_126(idx, step, macro); UNROLL_INCR(idx, step, macro) 323 #define LOOP_UNROLLING_128(idx, step, macro) LOOP_UNROLLING_127(idx, step, macro); UNROLL_INCR(idx, step, macro) 324 325 #define LOOP_UNROLLING_STR(type, idx, start, step, num, macro) \ 326 { \ 327 type idx = start; \ 328 LOOP_UNROLLING_##num(idx, step, macro); \ 329 } 330 #else // !defined(UNROLL_WITH_PRAGMA) 331 #define LOOP_UNROLLING_STR(type, idx, start, step, num, macro) \ 332 { \ 333 _Pragma("unroll") \ 334 for(type idx = start; idx < (num * step); idx += step) \ 335 { \ 336 (macro); \ 337 } \ 338 } 339 #endif // !defined(UNROLL_WITH_PRAGMA) 340 #define LOOP_UNROLLING(type, idx, start, step, num, macro) LOOP_UNROLLING_STR(type, idx, start, step, num, macro) 341 342 /** Get the get_global_id with partial N0. This function is useful when the dimension is not multiple of N0 and we need to use a partial N0 343 * to avoid out-of-bound read/write 344 * 345 * @note PARTIAL_N0 is used for get_global_id(n) = 0. 346 * 347 * @param[in] IDX get_global_id index (0,1 and 2 only) 348 * @param[in] N0 Number of elements read/written on the IDX direction 349 * @param[in] PARTIAL_N0 Number of elements read/written on the IDX direction for get_global_id(IDX) = 0. If zero, 350 * the Number of elements read/written on the IDX direction for get_global_id(IDX) = 0 is N0 351 */ 352 #define GET_SPATIAL_IDX(IDX, N0, PARTIAL_N0) (max((int)(get_global_id(IDX) * N0 - (N0 - PARTIAL_N0) % N0), 0)) 353 354 /** Dot product integet 8bit function 355 * 356 * @note Performs: c += dot(a, b) 357 * 358 * @param[in] A_DATA_TYPE A (lhs) data type 359 * @param[in] B_DATA_TYPE B (rhs) data type 360 * @param[in] C_DATA_TYPE C (accumulator) data type 361 * @param[in] K0 Number of accumulations 362 * @param[in] a OpenCL vector a 363 * @param[in] b OpenCL vector b 364 * @param[in] c Scalar variable c 365 */ 366 #define DOT_PRODUCT_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, b, c) DOT_PRODUCT_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, b, c) 367 #define DOT_PRODUCT_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, b, c) DOT_PRODUCT##K0##_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) 368 #define DOT_PRODUCT1_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 369 ({ \ 370 c += (C_DATA_TYPE)(a) * (C_DATA_TYPE)(b); \ 371 }) 372 #if defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_khr_integer_dot_product) 373 #define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += dot((A_DATA_TYPE##4)((a).s01, (A_DATA_TYPE##2)(0)), (B_DATA_TYPE##4)(((b).s01), (B_DATA_TYPE##2)(0))); 374 #define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += dot((A_DATA_TYPE##4)((a).s012, (A_DATA_TYPE)0), (B_DATA_TYPE##4)(((b).s012), (B_DATA_TYPE)0)); 375 #define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += dot((a), (b)); 376 #elif defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_khr_integer_dot_product) 377 #define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c = arm_dot_acc((A_DATA_TYPE##4)((a).s01, (A_DATA_TYPE##2)(0)), (B_DATA_TYPE##4)(((b).s01), (B_DATA_TYPE##2)(0)), (c)); 378 #define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c = arm_dot_acc((A_DATA_TYPE##4)((a).s012, (A_DATA_TYPE)0), (B_DATA_TYPE##4)(((b).s012), (B_DATA_TYPE)0), (c)); 379 #define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c = arm_dot_acc((a), (b), (c)); 380 #elif defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) // defined(ARM_COMPUTE_OPENCL_DOT8_ENABLED) && defined(cl_arm_integer_dot_product_int8) 381 #define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += arm_dot((A_DATA_TYPE##4)((a).s01, (A_DATA_TYPE##2)(0)), (B_DATA_TYPE##4)(((b).s01), (B_DATA_TYPE##2)(0))); 382 #define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += arm_dot((A_DATA_TYPE##4)((a).s012, (A_DATA_TYPE)0), (B_DATA_TYPE##4)(((b).s012), (B_DATA_TYPE)0)); 383 #define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) c += arm_dot((a), (b)); 384 #else // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) 385 #define DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 386 ({ \ 387 c += (C_DATA_TYPE)(a).s0 * (C_DATA_TYPE)(b).s0; \ 388 c += (C_DATA_TYPE)(a).s1 * (C_DATA_TYPE)(b).s1; \ 389 }) 390 #define DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 391 ({ \ 392 DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c); \ 393 c += (C_DATA_TYPE)(a).s2 * (C_DATA_TYPE)(b).s2; \ 394 }) 395 #define DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, x, y, val) \ 396 ({ \ 397 val += (C_DATA_TYPE)(x).s0 * (C_DATA_TYPE)(y).s0; \ 398 val += (C_DATA_TYPE)(x).s1 * (C_DATA_TYPE)(y).s1; \ 399 val += (C_DATA_TYPE)(x).s2 * (C_DATA_TYPE)(y).s2; \ 400 val += (C_DATA_TYPE)(x).s3 * (C_DATA_TYPE)(y).s3; \ 401 }) 402 #endif // defined(ARM_COMPUTE_OPENCL_DOT8_ACC_ENABLED) && defined(cl_arm_integer_dot_product_accumulate_int8) 403 #define DOT_PRODUCT5_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 404 ({ \ 405 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c); \ 406 DOT_PRODUCT1_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s4), ((b).s4), c); \ 407 }) 408 #define DOT_PRODUCT6_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 409 ({ \ 410 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c); \ 411 DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s45), ((b).s45), c); \ 412 }) 413 #define DOT_PRODUCT7_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 414 ({ \ 415 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s0123), ((b).s0123), c); \ 416 DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s456), ((b).s456), c); \ 417 }) 418 #define DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 419 ({ \ 420 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).lo), ((b).lo), c); \ 421 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).hi), ((b).hi), c); \ 422 }) 423 #define DOT_PRODUCT9_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 424 ({ \ 425 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ 426 DOT_PRODUCT1_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s8), ((b).s8), c); \ 427 }) 428 #define DOT_PRODUCT10_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 429 ({ \ 430 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ 431 DOT_PRODUCT2_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89), ((b).s89), c); \ 432 }) 433 #define DOT_PRODUCT11_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 434 ({ \ 435 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ 436 DOT_PRODUCT3_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89A), ((b).s89A), c); \ 437 }) 438 #define DOT_PRODUCT12_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 439 ({ \ 440 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ 441 DOT_PRODUCT4_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89AB), ((b).s89AB), c); \ 442 }) 443 #define DOT_PRODUCT13_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 444 ({ \ 445 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ 446 DOT_PRODUCT5_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89ABC), ((b).s89ABC), c); \ 447 }) 448 #define DOT_PRODUCT14_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 449 ({ \ 450 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ 451 DOT_PRODUCT6_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89ABCD), ((b).s89ABCD), c); \ 452 }) 453 #define DOT_PRODUCT15_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 454 ({ \ 455 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s01234567), ((b).s01234567), c); \ 456 DOT_PRODUCT7_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).s89ABCDE), ((b).s89ABCDE), c); \ 457 }) 458 #define DOT_PRODUCT16_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, a, b, c) \ 459 ({ \ 460 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).lo), ((b).lo), c); \ 461 DOT_PRODUCT8_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, ((a).hi), ((b).hi), c); \ 462 }) 463 464 /** Dot product integet 8bit function 465 * 466 * @note Performs: c += dot(a, b) 467 * 468 * @param[in] A_DATA_TYPE A (lhs) data type 469 * @param[in] B_DATA_TYPE B (rhs) data type 470 * @param[in] C_DATA_TYPE C (accumulator) data type 471 * @param[in] K0 Number of accumulations 472 * @param[in] a OpenCL vector a 473 * @param[in] c Scalar variable c 474 */ 475 #define REDUCE_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, c) REDUCE_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, c) 476 #define REDUCE_INTEGER8_STR(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, c) DOT_PRODUCT_INTEGER8(A_DATA_TYPE, B_DATA_TYPE, C_DATA_TYPE, K0, a, (TILE_VECTOR_TYPE##K0(B_DATA_TYPE))1, c) 477 478 /** Load a vector from global memory (tensor) 479 * 480 * @param[in] DATA_TYPE Data type 481 * @param[in] WIDTH Number of dst columns 482 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). 483 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16) 484 * @param[in] TENSOR Tensor basename 485 * @param[in] X Starting X position 486 * @param[in] Y Starting Y position 487 * @param[in] STRIDE_Y Stride Y (in bytes) 488 */ 489 #define V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y) V_LOAD_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y) 490 #define V_LOAD_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y) V_LOAD_##TENSOR_TYPE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) 491 #define V_LOAD_BUFFER(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) \ 492 VLOAD(WIDTH) \ 493 (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (Y) * (STRIDE_Y))) 494 #define V_LOAD_IMAGE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y) READ_IMAGE2D(DATA_TYPE, CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(WIDTH), TENSOR##_img, (X) / 4, (Y)) 495 496 /** Store a vector in global memory (tensor) 497 * 498 * @param[in] DATA_TYPE Data type 499 * @param[in] WIDTH Number of dst columns 500 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). 501 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16) 502 * @param[in] TENSOR Tensor basename 503 * @param[in] X Starting X position 504 * @param[in] Y Starting Y position 505 * @param[in] STRIDE_Y Stride Y (in bytes) 506 * @param[in] VALUES Values to store in memory 507 */ 508 #define V_STORE(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y, VALUES) V_STORE_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y, VALUES) 509 #define V_STORE_STR(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, Y, STRIDE_Y, VALUES) V_STORE_##TENSOR_TYPE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y, VALUES) 510 #define V_STORE_BUFFER(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y, VALUES) \ 511 VSTORE(WIDTH) \ 512 (VALUES, 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (Y) * (STRIDE_Y))) 513 #define V_STORE_IMAGE(DATA_TYPE, WIDTH, TENSOR, X, Y, STRIDE_Y, VALUES) WRITE_IMAGE2D(DATA_TYPE, CONVERT_VECTOR_SIZE_TO_PIXEL_UNIT(WIDTH), TENSOR##_img, (X) / 4, (Y), VALUES) 514 515 /** Load a tile from global memory (tensor) 516 * 517 * @param[in] DATA_TYPE Data type 518 * @param[in] HEIGHT Number of dst rows 519 * @param[in] WIDTH Number of dst columns 520 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). 521 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16) 522 * @param[in] TENSOR Tensor basename 523 * @param[in] X Starting X position 524 * @param[in] Y Starting Y position 525 * @param[in] YI_MULTIPLIER Parameter used to multiply the internal row increment (_i). 526 * In common cases should be 1 but it becomes useful when we want to load rows which are multiple of STRIDE_Y. (e.g. loading the weights of convolution layer). 527 * In this case the address calculation is performed as: (Y + _i * Y_MULTIPLIER) * STRIDE_Y 528 * @param[in] STRIDE_Y Stride Y (in bytes) used to load each row. 529 * @param[out] dst Output tile 530 */ 531 #define T_LOAD(DATA_TYPE, HEIGHT, WIDTH, TENSOR_TYPE, TENSOR, X, Y, YI_MULTIPLIER, STRIDE_Y, dst) \ 532 ({ \ 533 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \ 534 { \ 535 dst[_i].v = V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, ((Y) + _i * (int)(YI_MULTIPLIER)), STRIDE_Y); \ 536 }) \ 537 }) 538 539 /** Load a tile from global memory (tensor) using an indirect Y index tile 540 * 541 * @param[in] DATA_TYPE Data type 542 * @param[in] HEIGHT Number of dst rows 543 * @param[in] WIDTH Number of dst columns 544 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported 545 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16) 546 * @param[in] TENSOR Tensor basename 547 * @param[in] X Starting X position 548 * @param[in] STRIDE_Y Stride Y (in bytes) 549 * @param[in] indirect_y Indirect Y index tile 550 * @param[out] dst Output tile 551 */ 552 #define T_LOAD_INDIRECT(DATA_TYPE, HEIGHT, WIDTH, TENSOR_TYPE, TENSOR, X, STRIDE_Y, indirect_y, dst) \ 553 ({ \ 554 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \ 555 { \ 556 dst[_i].v = V_LOAD(DATA_TYPE, WIDTH, TENSOR_TYPE, TENSOR, X, (indirect_y[_i].v), STRIDE_Y); \ 557 }) \ 558 }) 559 560 /** Load a tile from global memory (tensor) using an indirect Y index tile and conditionally use a different length for the load 561 * 562 * @note If WIDTH1_CONDITION is true, the load will use the WIDTH1 length for the store 563 * @note The vectors are stored in reverse order so the invalid rows are overwritten by the valid ones 564 * 565 * @param[in] DATA_TYPE Data type 566 * @param[in] HEIGHT Number of dst rows 567 * @param[in] WIDTH0 Store width to use if WIDTH1_CONDITION = false 568 * @param[in] WIDTH1 Store width to use if WIDTH1_CONDITION = true 569 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). 570 * In case of cl_image, only WIDTH multiples of 4 are supported (4, 8, 16) 571 * @param[in] TENSOR Tensor basename 572 * @param[in] X Starting X position 573 * @param[in] STRIDE_Y Stride Y (in bytes) used to load each row. 574 * @param[in] WIDTH1_CONDITION Condition to select the WIDTH1 store 575 * @param[out] dst Output tile 576 * @param[out] indirect_y Indirect Y index tile 577 */ 578 #define T_LOAD_INDIRECT_WIDTH_SELECT(DATA_TYPE, HEIGHT, WIDTH0, WIDTH1, TENSOR_TYPE, TENSOR, X, STRIDE_Y, WIDTH1_CONDITION, dst, indirect_y) \ 579 ({ \ 580 if(WIDTH1_CONDITION) \ 581 { \ 582 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \ 583 { \ 584 VLOAD_PARTIAL(WIDTH0, WIDTH1) \ 585 (dst[HEIGHT - 1 - _i].v, 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \ 586 }) \ 587 } \ 588 else \ 589 { \ 590 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \ 591 { \ 592 dst[HEIGHT - 1 - _i].v = V_LOAD(DATA_TYPE, WIDTH0, TENSOR_TYPE, TENSOR, X, (indirect_y[HEIGHT - 1 - _i].v), STRIDE_Y); \ 593 }) \ 594 } \ 595 }) 596 /** Load a tile from global memory (tensor) when the tensor is stored using a NHWC layout 597 * 598 * @param[in] DATA_TYPE Data type 599 * @param[in] TILE_HEIGHT Number of elements to load from Y (height) dimension 600 * @param[in] TILE_WIDTH Number of elements to load from X (width) dimension 601 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension 602 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported 603 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16) 604 * @param[in] TENSOR Tensor basename 605 * @param[in] B Starting batch index 606 * @param[in] Y Starting Y index 607 * @param[in] X Starting X index 608 * @param[in] C Starting C index 609 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension 610 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension 611 * @param[in] STRIDE_Y Stride Y (in bytes) 612 * @param[out] dst Output tile 613 */ 614 #define T_LOAD_NHWC(DATA_TYPE, TILE_HEIGHT, TILE_WIDTH, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, STRIDE_Y, dst) \ 615 ({ \ 616 LOOP_UNROLLING(int, _yk, 0, 1, TILE_HEIGHT, \ 617 { \ 618 LOOP_UNROLLING(int, _xk, 0, 1, TILE_WIDTH, \ 619 { \ 620 int _src_y = (X) + _xk + ((Y) + _yk) * (TENSOR_WIDTH); \ 621 _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT); \ 622 int _src_valid_y = (((X) + _xk) >= 0 && ((X) + _xk) < (int)(TENSOR_WIDTH) && ((Y) + _yk) >= 0 && ((Y) + _yk) < (int)(TENSOR_HEIGHT)); \ 623 if(_src_valid_y != 0) \ 624 { \ 625 dst[_xk + _yk * (TILE_WIDTH)].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \ 626 } \ 627 }) \ 628 }) \ 629 }) 630 631 /** Load a tile from global memory (tensor) when the tensor is stored using a NHWC layout with dilation for the X and Y increments 632 * 633 * @param[in] DATA_TYPE Data type 634 * @param[in] TILE_HEIGHT Number of elements to load from Y (height) dimension 635 * @param[in] TILE_WIDTH Number of elements to load from X (width) dimension 636 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension 637 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported 638 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16) 639 * @param[in] TENSOR Tensor basename 640 * @param[in] B Starting batch index 641 * @param[in] Y Starting Y index 642 * @param[in] X Starting X index 643 * @param[in] C Starting C index 644 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension 645 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension 646 * @param[in] DILATION_X Dilation for the X increment 647 * @param[in] DILATION_Y Dilation for the Y increment 648 * @param[in] BOUNDARY_CHECK Boundary check flag. If true, it checks for any out-of-bound reads 649 * @param[out] dst Output tile 650 */ 651 #define T_LOAD_NHWC_WITH_DILATION(DATA_TYPE, TILE_HEIGHT, TILE_WIDTH, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, DILATION_X, DILATION_Y, BOUNDARY_CHECK, dst) \ 652 ({ \ 653 LOOP_UNROLLING(int, _yk, 0, 1, TILE_HEIGHT, \ 654 { \ 655 LOOP_UNROLLING(int, _xk, 0, 1, TILE_WIDTH, \ 656 { \ 657 int _src_y = (X) + _xk * (DILATION_X); \ 658 int _src_z = ((Y) + _yk * (DILATION_Y)); \ 659 int _src_w = (B); \ 660 bool _src_valid_y = (((X) + _xk * (DILATION_X)) >= 0) && (((X) + _xk * (DILATION_X)) < (int)(TENSOR_WIDTH)) && (((Y) + _yk * (DILATION_Y)) >= 0) && (((Y) + _yk * (DILATION_Y)) < (int)(TENSOR_HEIGHT)); \ 661 if(!(BOUNDARY_CHECK)) \ 662 { \ 663 dst[_xk + _yk * (TILE_WIDTH)].v = VLOAD(TILE_CHANNELS) \ 664 (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (C) * sizeof(DATA_TYPE) + (_src_y) * (TENSOR##_stride_y) + (_src_z) * (TENSOR##_stride_z) + (_src_w) * (TENSOR##_stride_w))); \ 665 } \ 666 else \ 667 { \ 668 if(_src_valid_y) \ 669 { \ 670 dst[_xk + _yk * (TILE_WIDTH)].v = VLOAD(TILE_CHANNELS) \ 671 (0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (C) * sizeof(DATA_TYPE) + (_src_y) * (TENSOR##_stride_y) + (_src_z) * (TENSOR##_stride_z) + (_src_w) * (TENSOR##_stride_w))); \ 672 } \ 673 } \ 674 }) \ 675 }) \ 676 }) 677 678 /** Load a tile from global memory (tensor) when the tensor is stored using a NHWC layout using indirect X and Y coordinates 679 * 680 * @param[in] DATA_TYPE Data type 681 * @param[in] TILE_AREA Number of elements to load from Y (height) dimension * Number of elements to load from X (width) dimension 682 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension 683 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported 684 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16) 685 * @param[in] TENSOR Tensor basename 686 * @param[in] B Starting batch index 687 * @param[in] Y Starting Y index 688 * @param[in] X Starting X index 689 * @param[in] C Starting C index 690 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension 691 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension 692 * @param[in] STRIDE_Y Stride Y (in bytes) 693 * @param[out] xi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect X coordinate 694 * @param[out] yi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect Y coordinate 695 * @param[out] dst Output tile 696 */ 697 #define T_LOAD_NHWC_INDIRECT(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, STRIDE_Y, xi, yi, dst) \ 698 ({ \ 699 LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \ 700 { \ 701 int _src_y = (X) + xi[_i].v + ((Y) + yi[_i].v) * (TENSOR_WIDTH); \ 702 _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT); \ 703 int _src_valid_y = (((X) + xi[_i].v) >= 0 && ((X) + xi[_i].v) < (int)(TENSOR_WIDTH) && ((Y) + yi[_i].v) >= 0 && ((Y) + yi[_i].v) < (int)(TENSOR_HEIGHT)); \ 704 if(_src_valid_y != 0) \ 705 { \ 706 dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \ 707 } \ 708 }) \ 709 }) 710 711 /** Load a tile from global memory (tensor) using an indirect buffer for the Y coordinates 712 * 713 * @param[in] DATA_TYPE Data type 714 * @param[in] TILE_AREA Number of elements to load from Y (height) dimension * Number of elements to load from X (width) dimension 715 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension 716 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). 717 * When TENSOR_TYPE=IMAGE, the if condition for the out-of-bound check can be skipped 718 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16) 719 * @param[in] TENSOR Tensor basename 720 * @param[in] C Starting C index 721 * @param[in] STRIDE_Y Stride Y (in bytes) 722 * @param[out] yi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect Y coordinate 723 * 16 is the maximum indirect buffer size. 724 * @param[out] dst Output tile 725 */ 726 #define T_LOAD2D_INDIRECT(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) T_LOAD2D_INDIRECT_STR(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) 727 #define T_LOAD2D_INDIRECT_STR(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) T_LOAD2D_INDIRECT_##TENSOR_TYPE(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) 728 #define T_LOAD2D_INDIRECT_BUFFER(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) \ 729 ({ \ 730 LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \ 731 { \ 732 if(yi[0].s[_i] >= 0) \ 733 { \ 734 dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, yi[0].s[_i], STRIDE_Y); \ 735 } \ 736 }) \ 737 }) 738 739 #define T_LOAD2D_INDIRECT_IMAGE(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, STRIDE_Y, yi, dst) \ 740 ({ \ 741 LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \ 742 { \ 743 dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, yi[0].s[_i], STRIDE_Y); \ 744 }) \ 745 }) 746 747 /** Load a tile from global memory (tensor) when the tensor is stored using a NDHWC layout using indirect X, Y and Z coordinates 748 * 749 * @param[in] DATA_TYPE Data type 750 * @param[in] TILE_AREA Number of elements to load from Y (height) dimension * Number of elements to load from X (width) dimension 751 * @param[in] TILE_CHANNELS Number of elements to load from C (channel) dimension 752 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported 753 * In case of cl_image, only TILE_CHANNELS multiples of 4 are supported (4, 8, 16) 754 * @param[in] TENSOR Tensor basename 755 * @param[in] B Starting batch index 756 * @param[in] Z Starting Z index 757 * @param[in] Y Starting Y index 758 * @param[in] X Starting X index 759 * @param[in] C Starting C index 760 * @param[in] TENSOR_WIDTH Number of elements to load from X (width) dimension 761 * @param[in] TENSOR_HEIGHT Number of elements to load from Y (height) dimension 762 * @param[in] TENSOR_DEPTH Number of elements to load from Z (depth) dimension 763 * @param[in] STRIDE_Y Stride Y (in bytes) 764 * @param[out] xi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect X coordinate 765 * @param[out] yi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect Y coordinate 766 * @param[out] zi A tile with (TILE_WIDTH x TILE_HEIGHT) values with the indirect Z coordinate 767 * @param[out] dst Output tile 768 */ 769 #define T_LOAD_NDHWC_INDIRECT(DATA_TYPE, TILE_AREA, TILE_CHANNELS, TENSOR_TYPE, TENSOR, B, Z, Y, X, C, TENSOR_WIDTH, TENSOR_HEIGHT, TENSOR_DEPTH, STRIDE_Y, xi, yi, zi, dst) \ 770 ({ \ 771 LOOP_UNROLLING(int, _i, 0, 1, TILE_AREA, \ 772 { \ 773 int _src_y = (X) + xi[_i].v + ((Y) + yi[_i].v) * (TENSOR_WIDTH) + ((Z) + zi[_i].v) * (TENSOR_WIDTH * TENSOR_HEIGHT); \ 774 _src_y += (B) * (int)(TENSOR_WIDTH) * (int)(TENSOR_HEIGHT) * (int)(TENSOR_DEPTH); \ 775 int _src_valid_y = (((X) + xi[_i].v) >= 0 && ((X) + xi[_i].v) < (int)(TENSOR_WIDTH) && ((Y) + yi[_i].v) >= 0 && ((Y) + yi[_i].v) < (int)(TENSOR_HEIGHT) \ 776 && ((Z) + zi[_i].v) >= 0 && ((Z) + zi[_i].v) < (int)(TENSOR_DEPTH)); \ 777 if(_src_valid_y != 0) \ 778 { \ 779 dst[_i].v = V_LOAD(DATA_TYPE, TILE_CHANNELS, TENSOR_TYPE, TENSOR, C, _src_y, STRIDE_Y); \ 780 } \ 781 }) \ 782 }) 783 784 /** Store a tile to global memory (tensor) using an indirect Y index tile and conditionally use a different length for the store 785 * 786 * @note If WIDTH1_CONDITION is true, the store will use the WIDTH1 length for the store 787 * @note The vectors are stored in reverse order so the invalid rows are overwritten by the valid ones 788 * 789 * @param[in] DATA_TYPE Data type 790 * @param[in] HEIGHT Number of src rows 791 * @param[in] WIDTH0 Store width to use if WIDTH1_CONDITION = false 792 * @param[in] WIDTH1 Store width to use if WIDTH1_CONDITION = true 793 * @param[in] TENSOR_TYPE Type of cl_type used to store the tensor in global memory (BUFFER=cl_buffer, IMAGE=cl_image). Currently BUFFER only is supported 794 * cl_image is not supported. 795 * @param[in] TENSOR Tensor basename 796 * @param[in] X Starting X position 797 * @param[in] STRIDE_Y Stride Y (in bytes) 798 * @param[in] WIDTH1_CONDITION Condition to select the WIDTH1 store 799 * @param[in] src Input tile 800 * @param[in] indirect_y Indirect Y index tile 801 */ 802 #define T_STORE_INDIRECT_WIDTH_SELECT(DATA_TYPE, HEIGHT, WIDTH0, WIDTH1, TENSOR_TYPE, TENSOR, X, STRIDE_Y, WIDTH1_CONDITION, src, indirect_y) \ 803 ({ \ 804 if(WIDTH1_CONDITION) \ 805 { \ 806 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \ 807 { \ 808 VSTORE_PARTIAL(WIDTH0, WIDTH1) \ 809 (CONVERT(src[HEIGHT - 1 - _i].v, VEC_DATA_TYPE(DATA_TYPE, WIDTH0)), 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \ 810 }) \ 811 } \ 812 else \ 813 { \ 814 LOOP_UNROLLING(int, _i, 0, 1, HEIGHT, \ 815 { \ 816 VSTORE(WIDTH0) \ 817 (CONVERT(src[HEIGHT - 1 - _i].v, VEC_DATA_TYPE(DATA_TYPE, WIDTH0)), 0, (__global DATA_TYPE *)(TENSOR##_ptr + TENSOR##_offset_first_element_in_bytes + (X) * sizeof(DATA_TYPE) + (indirect_y[HEIGHT - 1 - _i].v) * STRIDE_Y)); \ 818 }) \ 819 } \ 820 }) 821 822 /** Offset correction for the QASYMM8 computation 823 * 824 * @param[in] ACC_DATA_TYPE Accumulator data type 825 * @param[in] M0 Number of src/dst rows 826 * @param[in] N0 Number of src/dst columns 827 * @param[in] K0 Number of src columns 828 * @param[in] SRC_OFFSET Source quantization offset 829 * @param[in] WEI_OFFSET Weights quantization shift 830 * @param[in] lhs LHS tile 831 * @param[in] rhs RHS tile 832 * @param[out] dst DST tile 833 */ 834 #define T_OFFSET_CORRECTION(ACC_DATA_TYPE, M0, N0, K0, SRC_OFFSET, WEI_OFFSET, lhs, rhs, dst) \ 835 ({ \ 836 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 837 { \ 838 ACC_DATA_TYPE _tm = 0; \ 839 LOOP_UNROLLING(int, _k0, 0, 1, K0, \ 840 { \ 841 _tm += ((ACC_DATA_TYPE)lhs[_m0].s[_k0] * (ACC_DATA_TYPE)WEI_OFFSET); \ 842 }) \ 843 LOOP_UNROLLING(int, _n0, 0, 1, N0, \ 844 { \ 845 dst[_m0].s[_n0] += _tm; \ 846 LOOP_UNROLLING(int, _k0, 0, 1, K0, \ 847 { \ 848 dst[_m0].s[_n0] += ((ACC_DATA_TYPE)rhs[_n0].s[_k0] * (ACC_DATA_TYPE)SRC_OFFSET); \ 849 }) \ 850 }) \ 851 }) \ 852 }) 853 854 /** 8-bit quantization with fixed-point scale 855 * 856 * @param[in] SRC_DATA_TYPE SRC data type 857 * @param[in] DST_DATA_TYPE DST data type 858 * @param[in] QUANTIZATION_TYPE Quantization type (PER_TENSOR or PER_CHANNEL) 859 * @param[in] M0 Number of src/dst rows 860 * @param[in] N0 Number of src/dst columns 861 * @param[in] DST_OFFSET Quantization offset used for both the per-tensor and per-channel quantization 862 * @param[in] DST_SHIFT Quantization shift for the per-tensor quantization 863 * @param[in] DST_MULTIPLIER Quantization multiplier for the per-tensor quantization 864 * @param[in] src Input tile 865 * @param[in] dst_multipliers Output multipliers tile for the per-channel quantization 866 * @param[in] dst_shifts Output shift tile for the per-channel quantization 867 * @param[out] dst Output tile 868 */ 869 #define T_QUANTIZE8(SRC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) T_QUANTIZE8_STR(SRC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) 870 #define T_QUANTIZE8_STR(SRC_DATA_TYPE, DST_DATA_TYPE, QUANTIZATION_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) T_QUANTIZE8_##QUANTIZATION_TYPE(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) 871 872 /** 8-bit per-tensor quantization with fixed-point scale 873 * 874 * @param[in] SRC_DATA_TYPE SRC data type 875 * @param[in] DST_DATA_TYPE DST data type 876 * @param[in] M0 Number of src/dst rows 877 * @param[in] N0 Number of src/dst columns 878 * @param[in] DST_OFFSET Quantization offset 879 * @param[in] DST_SHIFT Quantization shift for the per-tensor quantization 880 * @param[in] DST_MULTIPLIER Quantization multiplier for the per-tensor quantization 881 * @param[in] src Input tile 882 * @param[in] dst_multipliers (unused) 883 * @param[in] dst_shifts (unused) 884 * @param[out] dst Output tile 885 */ 886 #define T_QUANTIZE8_PER_TENSOR(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) \ 887 ({ \ 888 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 889 { \ 890 LOOP_UNROLLING(int, _n0, 0, 1, N0, \ 891 { \ 892 SRC_DATA_TYPE _tmp = 0; \ 893 SRC_DATA_TYPE _src = src[_m0].s[_n0]; \ 894 _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-DST_SHIFT)), ((SRC_DATA_TYPE)DST_SHIFT < (SRC_DATA_TYPE)0)); \ 895 SRC_DATA_TYPE overflow = _src == DST_MULTIPLIER && _src == INT_MIN; \ 896 long a_64 = (long)(_src); \ 897 long b_64 = (long)(DST_MULTIPLIER); \ 898 long ab_64 = a_64 * b_64; \ 899 long mask1 = 1 << 30; \ 900 long mask2 = 1 - (1 << 30); \ 901 long is_positive_or_zero = ab_64 >= 0; \ 902 long nudge = select(mask2, mask1, is_positive_or_zero); \ 903 SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \ 904 _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \ 905 if(DST_SHIFT >= 0) \ 906 { \ 907 long mask = ((((int)1) << DST_SHIFT) - (long)1); \ 908 long threshold = _tmp < (int)0 ? (mask >> 1) + (long)1 : (mask >> 1) + 0; \ 909 _tmp = (_tmp & mask) > threshold ? (_tmp >> DST_SHIFT) + (int)1 : (_tmp >> DST_SHIFT); \ 910 } \ 911 _tmp += DST_OFFSET; \ 912 dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \ 913 }) \ 914 }) \ 915 }) 916 917 /** 8-bit per-channel quantization with fixed-point scale 918 * 919 * @param[in] SRC_DATA_TYPE SRC data type 920 * @param[in] DST_DATA_TYPE DST data type 921 * @param[in] M0 Number of src/dst rows 922 * @param[in] N0 Number of src/dst columns 923 * @param[in] DST_OFFSET Quantization offset 924 * @param[in] DST_SHIFT (unused) 925 * @param[in] DST_MULTIPLIER (unused) 926 * @param[in] src Input tile 927 * @param[in] dst_multipliers Output multipliers tile for the per-channel quantization 928 * @param[in] dst_shifts Output shift tile for the per-channel quantization 929 * @param[out] dst Output tile 930 */ 931 #define T_QUANTIZE8_PER_CHANNEL(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst_multipliers, dst_shifts, dst) \ 932 ({ \ 933 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 934 { \ 935 LOOP_UNROLLING(int, _n0, 0, 1, N0, \ 936 { \ 937 SRC_DATA_TYPE _tmp = 0; \ 938 SRC_DATA_TYPE _tmp2 = 0; \ 939 SRC_DATA_TYPE _src = src[_m0].s[_n0]; \ 940 SRC_DATA_TYPE _dst_multiplier = dst_multipliers[0].s[_n0]; \ 941 SRC_DATA_TYPE _dst_shift = dst_shifts[0].s[_n0]; \ 942 _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-_dst_shift)), ((SRC_DATA_TYPE)_dst_shift < (SRC_DATA_TYPE)0)); \ 943 SRC_DATA_TYPE overflow = _src == _dst_multiplier && _src == INT_MIN; \ 944 long a_64 = (long)(_src); \ 945 long b_64 = (long)(_dst_multiplier); \ 946 long ab_64 = a_64 * b_64; \ 947 long mask1 = 1 << 30; \ 948 long mask2 = 1 - (1 << 30); \ 949 long is_positive_or_zero = ab_64 >= 0; \ 950 long nudge = select(mask2, mask1, is_positive_or_zero); \ 951 SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \ 952 _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \ 953 long mask = ((((int)1) << _dst_shift) - (int)1); \ 954 long threshold = (mask >> 1) + any(_tmp); \ 955 _tmp2 = _tmp >> _dst_shift; \ 956 _tmp2 += select(0, 1, (_tmp & mask) > threshold); \ 957 _tmp = select(_tmp, _tmp2, _dst_shift >= 0); \ 958 _tmp += DST_OFFSET; \ 959 dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \ 960 }) \ 961 }) \ 962 }) 963 964 /** Quantized the 8-bit tile with fixed-point scale for asymmetric 965 * 966 * @param[in] SRC_DATA_TYPE SRC data type 967 * @param[in] DST_DATA_TYPE DST data type 968 * @param[in] M0 Number of src/dst rows 969 * @param[in] N0 Number of src/dst columns 970 * @param[in] DST_OFFSET Quantization offset used for both the per-tensor and per-channel quantization 971 * @param[in] DST_SHIFT Quantization shift for the per-tensor quantization 972 * @param[in] DST_MULTIPLIER Quantization multiplier for the per-tensor quantization 973 * @param[in] src Input tile 974 * @param[out] dst Output tile 975 */ 976 #define T_QUANTIZE8_ASYMMETRIC(SRC_DATA_TYPE, DST_DATA_TYPE, M0, N0, DST_OFFSET, DST_SHIFT, DST_MULTIPLIER, src, dst) \ 977 ({ \ 978 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 979 { \ 980 LOOP_UNROLLING(int, _n0, 0, 1, N0, \ 981 { \ 982 SRC_DATA_TYPE _tmp = 0; \ 983 SRC_DATA_TYPE _src = src[_m0].s[_n0]; \ 984 _src *= select((SRC_DATA_TYPE)1, ((SRC_DATA_TYPE)1 << (SRC_DATA_TYPE)(-DST_SHIFT)), ((SRC_DATA_TYPE)DST_SHIFT < (SRC_DATA_TYPE)0)); \ 985 SRC_DATA_TYPE overflow = _src == DST_MULTIPLIER && _src == INT_MIN; \ 986 long a_64 = (long)(_src); \ 987 long b_64 = (long)(DST_MULTIPLIER); \ 988 long ab_64 = a_64 * b_64; \ 989 long mask1 = 1 << 30; \ 990 long mask2 = 1 - (1 << 30); \ 991 long is_positive_or_zero = ab_64 >= 0; \ 992 long nudge = select(mask2, mask1, is_positive_or_zero); \ 993 SRC_DATA_TYPE ab_x2_high32 = CONVERT((ab_64 + nudge) / (long)(1ll << 31), SRC_DATA_TYPE); \ 994 _tmp = select(ab_x2_high32, (SRC_DATA_TYPE)INT_MAX, overflow); \ 995 if(DST_SHIFT >= 0) \ 996 { \ 997 long mask = ((((int)1) << DST_SHIFT) - (int)1); \ 998 long threshold = _tmp < (int)0 ? (mask >> 1) + (long)1 : (mask >> 1) + 0; \ 999 _tmp = (_tmp & mask) > threshold ? (_tmp >> DST_SHIFT) + (int)1 : (_tmp >> DST_SHIFT); \ 1000 } \ 1001 _tmp += DST_OFFSET; \ 1002 dst[_m0].s[_n0] = CONVERT_SAT(_tmp, DST_DATA_TYPE); \ 1003 }) \ 1004 }) \ 1005 }) 1006 1007 /** Conditional rowset (memset by row) 1008 * 1009 * @note Set the row to VALUE_TO_SET if the corresponding mask == 0 1010 * 1011 * @param[in] DATA_TYPE Data type 1012 * @param[in] M0 Number of LHS rows 1013 * @param[in] N0 Number of LHS columns 1014 * @param[in] VALUE_TO_SET Value to set the row 1015 * @param[in, out] a Input/output tile 1016 * @param[out] mask Mask to check for setting the row to VALUE_TO_SET 1017 */ 1018 #define T_ROWSET_MASK(DATA_TYPE, M0, N0, VALUE_TO_SET, a, mask) \ 1019 ({ \ 1020 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 1021 { \ 1022 LOOP_UNROLLING(int, _n0, 0, 1, N0, \ 1023 { \ 1024 a[_m0].s[_n0] = select((DATA_TYPE)(a[_m0].s[_n0]), (DATA_TYPE)(VALUE_TO_SET), (SELECT_DATA_TYPE(DATA_TYPE))(mask[_m0].v == (DATA_TYPE)0)); \ 1025 }) \ 1026 }) \ 1027 }) 1028 1029 /** Element-wise activation for floating point types 1030 * 1031 * @note Performs: activation(LHS) = DST 1032 * 1033 * @param[in] DATA_TYPE SRC/DST data type 1034 * @param[in] M0 Number of SRC/DST rows 1035 * @param[in] N0 Number of SRC/DST columns 1036 * @param[in] ACTIVATION_TYPE Activation type 1037 * @param[in] A_VAL A value used for the activation (e.g. tanh_op, brelu,..) 1038 * @param[in] B_VAL B value used for the activation (e.g. tanh_op, brelu,..) 1039 * @param[out] src SRC tile 1040 * @param[out] dst DST tile 1041 */ 1042 #define T_ACTIVATION(DATA_TYPE, M0, N0, ACTIVATION_TYPE, A_VAL, B_VAL, src, dst) \ 1043 ({ \ 1044 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 1045 { \ 1046 dst[_m0].v = ACTIVATION(ACTIVATION_TYPE, DATA_TYPE, N0, src[_m0].v, A_VAL, B_VAL); \ 1047 }) \ 1048 }) 1049 1050 // RELU Activation 1051 #define relu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (max((DATA_TYPE)ZERO_VALUE, x)) 1052 // Bounded RELU Activation 1053 #define brelu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (min((DATA_TYPE)A_VAL, max((DATA_TYPE)ZERO_VALUE, x))) 1054 // Lower Upper Bounded RELU Activation 1055 #define lu_brelu_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (min(max(x, (DATA_TYPE)B_VAL), (DATA_TYPE)A_VAL)) 1056 // Hard Swish Activation 1057 #define hard_swish_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (x * ((min(max((DATA_TYPE)(x + (DATA_TYPE)3.f), (DATA_TYPE)0.f), (DATA_TYPE)6.f)) * (DATA_TYPE)0.166666667f)) 1058 // Identity Activation 1059 #define identity_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) (x) 1060 1061 #define ACT_OP_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) op##_op_quantized(DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) 1062 #define ACTIVATION_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) ACT_OP_QUANTIZED(op, DATA_TYPE, VEC_SIZE, ZERO_VALUE, A_VAL, B_VAL, x) 1063 1064 #define V_ADD(A_VAL, B_VAL) ((A_VAL) + (B_VAL)) 1065 #define V_SUB(A_VAL, B_VAL) ((A_VAL) - (B_VAL)) 1066 #define V_DIV(A_VAL, B_VAL) ((A_VAL) / (B_VAL)) 1067 #define V_MUL(A_VAL, B_VAL) ((A_VAL) * (B_VAL)) 1068 1069 /** Element-wise activation for quantized types 1070 * 1071 * @note Performs: activation(LHS) = DST 1072 * 1073 * @param[in] DATA_TYPE SRC/DST data type 1074 * @param[in] M0 Number of SRC/DST rows 1075 * @param[in] N0 Number of SRC/DST columns 1076 * @param[in] ACTIVATION_TYPE Activation type 1077 * @param[in] ZERO_VALUE The zero value to consider in the computation 1078 * @param[in] A_VAL A value used for the activation (e.g. tanh_op, brelu,..) 1079 * @param[in] B_VAL B value used for the activation (e.g. tanh_op, brelu,..) 1080 * @param[out] src SRC tile 1081 * @param[out] dst DST tile 1082 */ 1083 #define T_ACTIVATION_QUANTIZED(DATA_TYPE, M0, N0, ACTIVATION_TYPE, ZERO_VALUE, A_VAL, B_VAL, src, dst) \ 1084 ({ \ 1085 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 1086 { \ 1087 dst[_m0].v = ACTIVATION_QUANTIZED(ACTIVATION_TYPE, DATA_TYPE, N0, ZERO_VALUE, A_VAL, B_VAL, src[_m0].v); \ 1088 }) \ 1089 }) 1090 1091 /** Element-wise addition between two tiles 1092 * 1093 * @note Performs: LHS + RHS = DST 1094 * 1095 * @param[in] DATA_TYPE LHS/RHS/DST data type 1096 * @param[in] M0 Number of LHS rows 1097 * @param[in] N0 Number of LHS columns 1098 * @param[in] lhs LHS tile 1099 * @param[in] rhs Constant RHS tile 1100 * @param[out] dst DST tile 1101 */ 1102 #define T_ADD(DATA_TYPE, M0, N0, lhs, rhs, dst) \ 1103 ({ \ 1104 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 1105 { \ 1106 dst[_m0].v = lhs[_m0].v + rhs[_m0].v; \ 1107 }) \ 1108 }) 1109 1110 /** Element-wise addition with a constant value 1111 * 1112 * @note Performs: LHS + constant = DST 1113 * 1114 * @param[in] DATA_TYPE LHS/RHS/DST data type 1115 * @param[in] M0 Number of LHS rows 1116 * @param[in] N0 Number of LHS columns 1117 * @param[in] lhs LHS tile 1118 * @param[in] rhs_constant Constant value 1119 * @param[out] dst DST tile 1120 */ 1121 #define T_ADD_CONSTANT(DATA_TYPE, M0, N0, lhs, rhs_constant, dst) \ 1122 ({ \ 1123 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 1124 { \ 1125 dst[_m0].v = lhs[_m0].v + (DATA_TYPE)rhs_constant; \ 1126 }) \ 1127 }) 1128 1129 #define T_ELTWISE_BROADCAST_ADD_X(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) 1130 #define T_ELTWISE_BROADCAST_LHS_X_ADD(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_LHS_X(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) 1131 #define T_ELTWISE_BROADCAST_RHS_X_ADD(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) 1132 1133 #define T_ELTWISE_BROADCAST_LHS_X_SUB(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_LHS_X(V_SUB, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) 1134 #define T_ELTWISE_BROADCAST_RHS_X_SUB(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_SUB, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) 1135 1136 #define T_ELTWISE_BROADCAST_DIV_X(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_DIV, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) 1137 1138 #define T_ELTWISE_BROADCAST_LHS_X_MUL(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_LHS_X(V_MUL, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) 1139 #define T_ELTWISE_BROADCAST_RHS_X_MUL(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE_BROADCAST_X(V_MUL, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) 1140 1141 /** Element-wise scale with a constant value 1142 * 1143 * @note Performs: LHS * constant = DST 1144 * 1145 * @param[in] DATA_TYPE LHS/RHS/DST data type 1146 * @param[in] M0 Number of LHS rows 1147 * @param[in] N0 Number of LHS columns 1148 * @param[in] lhs LHS tile 1149 * @param[in] rhs_constant Constant value 1150 * @param[out] dst DST tile 1151 */ 1152 #define T_SCALE_CONSTANT(DATA_TYPE, M0, N0, lhs, rhs_constant, dst) \ 1153 ({ \ 1154 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 1155 { \ 1156 dst[_m0].v = lhs[_m0].v * (DATA_TYPE)rhs_constant; \ 1157 }) \ 1158 }) 1159 1160 /** Element-wise operation with RHS broadcasted (RHS has the X dimension only) 1161 * 1162 * @note Performs: LHS OP RHS[broadcasted] = DST 1163 * @note Both tiles must have same data type 1164 * 1165 * @param[in] T_ELWISE_OP Elementwise operator to perform 1166 * @param[in] DST_DATA_TYPE DST data type 1167 * @param[in] M0 Number of LHS rows 1168 * @param[in] N0 Number of LHS columns 1169 * @param[in] lhs LHS tile 1170 * @param[in] rhs RHS tile 1171 * @param[out] dst DST tile 1172 */ 1173 #define T_ELTWISE_BROADCAST_X(T_ELWISE_OP, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) \ 1174 ({ \ 1175 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 1176 { \ 1177 dst[_m0].v = T_ELWISE_OP(CONVERT(lhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)), CONVERT(rhs[0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0))); \ 1178 }) \ 1179 }) 1180 1181 /** Element-wise operation with LHS broadcasted (LHS has the X dimension only) 1182 * 1183 * @note Performs: LHS[broadcasted] OP RHS = DST 1184 * @note Both tiles must have same data type 1185 * 1186 * @param[in] T_ELWISE_OP Elementwise operator to perform 1187 * @param[in] DST_DATA_TYPE DST data type 1188 * @param[in] M0 Number of RHS rows 1189 * @param[in] N0 Number of RHS columns 1190 * @param[in] lhs LHS tile 1191 * @param[in] rhs RHS tile 1192 * @param[out] dst DST tile 1193 */ 1194 #define T_ELTWISE_BROADCAST_LHS_X(T_ELWISE_OP, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) \ 1195 ({ \ 1196 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 1197 { \ 1198 dst[_m0].v = T_ELWISE_OP(CONVERT(lhs[0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)), CONVERT(rhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0))); \ 1199 }) \ 1200 }) 1201 1202 #define T_ELTWISE_ADD(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_ADD, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) 1203 #define T_ELTWISE_SUB(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_SUB, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) 1204 #define T_ELTWISE_DIV(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_DIV, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) 1205 #define T_ELTWISE_MUL(DST_DATA_TYPE, M0, N0, lhs, rhs, dst) T_ELTWISE(V_MUL, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) 1206 1207 /** Element-wise operation between two tiles (LHS and RHS) 1208 * 1209 * @note Performs: LHS OP RHS = DST 1210 * @note Both tiles must have same data type 1211 * 1212 * @param[in] T_ELWISE_OP Elementwise operator to perform 1213 * @param[in] DST_DATA_TYPE DST data type 1214 * @param[in] M0 Number of LHS rows 1215 * @param[in] N0 Number of LHS columns 1216 * @param[in] lhs LHS tile 1217 * @param[in] rhs RHS tile 1218 * @param[out] dst DST tile 1219 */ 1220 #define T_ELTWISE(T_ELWISE_OP, DST_DATA_TYPE, M0, N0, lhs, rhs, dst) \ 1221 ({ \ 1222 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 1223 { \ 1224 dst[_m0].v = T_ELWISE_OP(CONVERT(lhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0)), CONVERT(rhs[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0))); \ 1225 }) \ 1226 }) 1227 1228 /** Floor operation on a tile 1229 * 1230 * @note Performs: floor(SRC) = DST 1231 * @note Both tiles must have same data type 1232 * 1233 * @param[in] DST_DATA_TYPE DST data type 1234 * @param[in] M0 Number of SRC rows 1235 * @param[in] N0 Number of SRC columns 1236 * @param[in] src LHS tile 1237 * @param[out] dst DST tile 1238 */ 1239 #define T_FLOOR(DST_DATA_TYPE, M0, N0, src, dst) \ 1240 ({ \ 1241 LOOP_UNROLLING(int, _m0, 0, 1, M0, \ 1242 { \ 1243 dst[_m0].v = floor(CONVERT(src[_m0].v, VEC_DATA_TYPE(DST_DATA_TYPE, N0))); \ 1244 }) \ 1245 }) 1246 1247 /** Matrix multiplication 1248 * 1249 * @note Performs: LHS X RHS + DST = DST 1250 * 1251 * @param[in] LHS_DATA_TYPE LHS tile data type 1252 * @param[in] RHS_DATA_TYPE RHS tile data type 1253 * @param[in] DST_DATA_TYPE RHS tile data type 1254 * @param[in] M0 Number of LHS rows 1255 * @param[in] N0 Number of RHS columns 1256 * @param[in] K0 Number of LHS columns 1257 * @param[in] LHS_LAYOUT LHS layout (T= transposed, NT= not transposed) 1258 * @param[in] RHS_LAYOUT RHS layout (T= transposed, NT= not transposed) 1259 * @param[in] lhs LHS tile 1260 * @param[in] rhs RHS tile 1261 * @param[in, out] dst DST tile 1262 */ 1263 #define T_MMUL(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, LHS_LAYOUT, RHS_LAYOUT, lhs, rhs, dst) T_MMUL_##LHS_LAYOUT##_##RHS_LAYOUT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) 1264 #define T_MMUL_NT_T(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_##LHS_DATA_TYPE##_##RHS_DATA_TYPE##_##DST_DATA_TYPE(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) 1265 #define T_MMUL_NT_T_float_float_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) 1266 #define T_MMUL_NT_T_half_half_float(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) 1267 #define T_MMUL_NT_T_half_half_half(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) 1268 #define T_MMUL_NT_T_char_char_int(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) 1269 #define T_MMUL_NT_T_uchar_uchar_uint(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) 1270 #define T_MMUL_NT_T_uchar_uchar_int(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) 1271 #define T_MMUL_NT_T_FLOAT(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) \ 1272 { \ 1273 LOOP_UNROLLING(int, _m, 0, 1, M0, \ 1274 { \ 1275 LOOP_UNROLLING(int, _n, 0, 1, N0, \ 1276 { \ 1277 LOOP_UNROLLING(int, _k, 0, 1, K0, \ 1278 { \ 1279 dst[_m].s[_n] = fma((DST_DATA_TYPE)(lhs[_m].s[_k]), (DST_DATA_TYPE)(rhs[_n].s[_k]), dst[_m].s[_n]); \ 1280 }) \ 1281 }) \ 1282 }) \ 1283 } 1284 1285 #define T_MMUL_NT_T_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, M0, N0, K0, lhs, rhs, dst) \ 1286 ({ \ 1287 LOOP_UNROLLING(int, _m, 0, 1, M0, \ 1288 { \ 1289 LOOP_UNROLLING(int, _n, 0, 1, N0, \ 1290 { \ 1291 DOT_PRODUCT_INTEGER8(LHS_DATA_TYPE, RHS_DATA_TYPE, DST_DATA_TYPE, K0, (lhs[_m].v), (rhs[_n].v), dst[_m].s[_n]); \ 1292 }) \ 1293 }) \ 1294 }) 1295 1296 #endif /* SRC_CORE_CL_CL_KERNELS_TILE_HELPERS */ 1297