1 /*M/////////////////////////////////////////////////////////////////////////////////////// 2 // 3 // IMPORTANT: READ BEFORE DOWNLOADING, COPYING, INSTALLING OR USING. 4 // 5 // By downloading, copying, installing or using the software you agree to this license. 6 // If you do not agree to this license, do not download, install, 7 // copy or use the software. 8 // 9 // 10 // License Agreement 11 // For Open Source Computer Vision Library 12 // 13 // Copyright (C) 2000-2008, Intel Corporation, all rights reserved. 14 // Copyright (C) 2009, Willow Garage Inc., all rights reserved. 15 // Third party copyrights are property of their respective owners. 16 // 17 // Redistribution and use in source and binary forms, with or without modification, 18 // are permitted provided that the following conditions are met: 19 // 20 // * Redistribution's of source code must retain the above copyright notice, 21 // this list of conditions and the following disclaimer. 22 // 23 // * Redistribution's in binary form must reproduce the above copyright notice, 24 // this list of conditions and the following disclaimer in the documentation 25 // and/or other materials provided with the distribution. 26 // 27 // * The name of the copyright holders may not be used to endorse or promote products 28 // derived from this software without specific prior written permission. 29 // 30 // This software is provided by the copyright holders and contributors "as is" and 31 // any express or implied warranties, including, but not limited to, the implied 32 // warranties of merchantability and fitness for a particular purpose are disclaimed. 33 // In no event shall the Intel Corporation or contributors be liable for any direct, 34 // indirect, incidental, special, exemplary, or consequential damages 35 // (including, but not limited to, procurement of substitute goods or services; 36 // loss of use, data, or profits; or business interruption) however caused 37 // and on any theory of liability, whether in contract, strict liability, 38 // or tort (including negligence or otherwise) arising in any way out of 39 // the use of this software, even if advised of the possibility of such damage. 40 // 41 //M*/ 42 43 #ifndef __OPENCV_CUDA_PRED_VAL_REDUCE_DETAIL_HPP__ 44 #define __OPENCV_CUDA_PRED_VAL_REDUCE_DETAIL_HPP__ 45 46 #include <thrust/tuple.h> 47 #include "../warp.hpp" 48 #include "../warp_shuffle.hpp" 49 50 //! @cond IGNORED 51 52 namespace cv { namespace cuda { namespace device 53 { 54 namespace reduce_key_val_detail 55 { 56 template <typename T> struct GetType; 57 template <typename T> struct GetType<T*> 58 { 59 typedef T type; 60 }; 61 template <typename T> struct GetType<volatile T*> 62 { 63 typedef T type; 64 }; 65 template <typename T> struct GetType<T&> 66 { 67 typedef T type; 68 }; 69 70 template <unsigned int I, unsigned int N> 71 struct For 72 { 73 template <class PointerTuple, class ReferenceTuple> 74 static __device__ void loadToSmem(const PointerTuple& smem, const ReferenceTuple& data, unsigned int tid) 75 { 76 thrust::get<I>(smem)[tid] = thrust::get<I>(data); 77 78 For<I + 1, N>::loadToSmem(smem, data, tid); 79 } 80 template <class PointerTuple, class ReferenceTuple> 81 static __device__ void loadFromSmem(const PointerTuple& smem, const ReferenceTuple& data, unsigned int tid) 82 { 83 thrust::get<I>(data) = thrust::get<I>(smem)[tid]; 84 85 For<I + 1, N>::loadFromSmem(smem, data, tid); 86 } 87 88 template <class ReferenceTuple> 89 static __device__ void copyShfl(const ReferenceTuple& val, unsigned int delta, int width) 90 { 91 thrust::get<I>(val) = shfl_down(thrust::get<I>(val), delta, width); 92 93 For<I + 1, N>::copyShfl(val, delta, width); 94 } 95 template <class PointerTuple, class ReferenceTuple> 96 static __device__ void copy(const PointerTuple& svals, const ReferenceTuple& val, unsigned int tid, unsigned int delta) 97 { 98 thrust::get<I>(svals)[tid] = thrust::get<I>(val) = thrust::get<I>(svals)[tid + delta]; 99 100 For<I + 1, N>::copy(svals, val, tid, delta); 101 } 102 103 template <class KeyReferenceTuple, class ValReferenceTuple, class CmpTuple> 104 static __device__ void mergeShfl(const KeyReferenceTuple& key, const ValReferenceTuple& val, const CmpTuple& cmp, unsigned int delta, int width) 105 { 106 typename GetType<typename thrust::tuple_element<I, KeyReferenceTuple>::type>::type reg = shfl_down(thrust::get<I>(key), delta, width); 107 108 if (thrust::get<I>(cmp)(reg, thrust::get<I>(key))) 109 { 110 thrust::get<I>(key) = reg; 111 thrust::get<I>(val) = shfl_down(thrust::get<I>(val), delta, width); 112 } 113 114 For<I + 1, N>::mergeShfl(key, val, cmp, delta, width); 115 } 116 template <class KeyPointerTuple, class KeyReferenceTuple, class ValPointerTuple, class ValReferenceTuple, class CmpTuple> 117 static __device__ void merge(const KeyPointerTuple& skeys, const KeyReferenceTuple& key, 118 const ValPointerTuple& svals, const ValReferenceTuple& val, 119 const CmpTuple& cmp, 120 unsigned int tid, unsigned int delta) 121 { 122 typename GetType<typename thrust::tuple_element<I, KeyPointerTuple>::type>::type reg = thrust::get<I>(skeys)[tid + delta]; 123 124 if (thrust::get<I>(cmp)(reg, thrust::get<I>(key))) 125 { 126 thrust::get<I>(skeys)[tid] = thrust::get<I>(key) = reg; 127 thrust::get<I>(svals)[tid] = thrust::get<I>(val) = thrust::get<I>(svals)[tid + delta]; 128 } 129 130 For<I + 1, N>::merge(skeys, key, svals, val, cmp, tid, delta); 131 } 132 }; 133 template <unsigned int N> 134 struct For<N, N> 135 { 136 template <class PointerTuple, class ReferenceTuple> 137 static __device__ void loadToSmem(const PointerTuple&, const ReferenceTuple&, unsigned int) 138 { 139 } 140 template <class PointerTuple, class ReferenceTuple> 141 static __device__ void loadFromSmem(const PointerTuple&, const ReferenceTuple&, unsigned int) 142 { 143 } 144 145 template <class ReferenceTuple> 146 static __device__ void copyShfl(const ReferenceTuple&, unsigned int, int) 147 { 148 } 149 template <class PointerTuple, class ReferenceTuple> 150 static __device__ void copy(const PointerTuple&, const ReferenceTuple&, unsigned int, unsigned int) 151 { 152 } 153 154 template <class KeyReferenceTuple, class ValReferenceTuple, class CmpTuple> 155 static __device__ void mergeShfl(const KeyReferenceTuple&, const ValReferenceTuple&, const CmpTuple&, unsigned int, int) 156 { 157 } 158 template <class KeyPointerTuple, class KeyReferenceTuple, class ValPointerTuple, class ValReferenceTuple, class CmpTuple> 159 static __device__ void merge(const KeyPointerTuple&, const KeyReferenceTuple&, 160 const ValPointerTuple&, const ValReferenceTuple&, 161 const CmpTuple&, 162 unsigned int, unsigned int) 163 { 164 } 165 }; 166 167 ////////////////////////////////////////////////////// 168 // loadToSmem 169 170 template <typename T> 171 __device__ __forceinline__ void loadToSmem(volatile T* smem, T& data, unsigned int tid) 172 { 173 smem[tid] = data; 174 } 175 template <typename T> 176 __device__ __forceinline__ void loadFromSmem(volatile T* smem, T& data, unsigned int tid) 177 { 178 data = smem[tid]; 179 } 180 template <typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9, 181 typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9> 182 __device__ __forceinline__ void loadToSmem(const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& smem, 183 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& data, 184 unsigned int tid) 185 { 186 For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::loadToSmem(smem, data, tid); 187 } 188 template <typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9, 189 typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9> 190 __device__ __forceinline__ void loadFromSmem(const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& smem, 191 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& data, 192 unsigned int tid) 193 { 194 For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::loadFromSmem(smem, data, tid); 195 } 196 197 ////////////////////////////////////////////////////// 198 // copyVals 199 200 template <typename V> 201 __device__ __forceinline__ void copyValsShfl(V& val, unsigned int delta, int width) 202 { 203 val = shfl_down(val, delta, width); 204 } 205 template <typename V> 206 __device__ __forceinline__ void copyVals(volatile V* svals, V& val, unsigned int tid, unsigned int delta) 207 { 208 svals[tid] = val = svals[tid + delta]; 209 } 210 template <typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9> 211 __device__ __forceinline__ void copyValsShfl(const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val, 212 unsigned int delta, 213 int width) 214 { 215 For<0, thrust::tuple_size<thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9> >::value>::copyShfl(val, delta, width); 216 } 217 template <typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9, 218 typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9> 219 __device__ __forceinline__ void copyVals(const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals, 220 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val, 221 unsigned int tid, unsigned int delta) 222 { 223 For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::copy(svals, val, tid, delta); 224 } 225 226 ////////////////////////////////////////////////////// 227 // merge 228 229 template <typename K, typename V, class Cmp> 230 __device__ __forceinline__ void mergeShfl(K& key, V& val, const Cmp& cmp, unsigned int delta, int width) 231 { 232 K reg = shfl_down(key, delta, width); 233 234 if (cmp(reg, key)) 235 { 236 key = reg; 237 copyValsShfl(val, delta, width); 238 } 239 } 240 template <typename K, typename V, class Cmp> 241 __device__ __forceinline__ void merge(volatile K* skeys, K& key, volatile V* svals, V& val, const Cmp& cmp, unsigned int tid, unsigned int delta) 242 { 243 K reg = skeys[tid + delta]; 244 245 if (cmp(reg, key)) 246 { 247 skeys[tid] = key = reg; 248 copyVals(svals, val, tid, delta); 249 } 250 } 251 template <typename K, 252 typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9, 253 class Cmp> 254 __device__ __forceinline__ void mergeShfl(K& key, 255 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val, 256 const Cmp& cmp, 257 unsigned int delta, int width) 258 { 259 K reg = shfl_down(key, delta, width); 260 261 if (cmp(reg, key)) 262 { 263 key = reg; 264 copyValsShfl(val, delta, width); 265 } 266 } 267 template <typename K, 268 typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9, 269 typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9, 270 class Cmp> 271 __device__ __forceinline__ void merge(volatile K* skeys, K& key, 272 const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals, 273 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val, 274 const Cmp& cmp, unsigned int tid, unsigned int delta) 275 { 276 K reg = skeys[tid + delta]; 277 278 if (cmp(reg, key)) 279 { 280 skeys[tid] = key = reg; 281 copyVals(svals, val, tid, delta); 282 } 283 } 284 template <typename KR0, typename KR1, typename KR2, typename KR3, typename KR4, typename KR5, typename KR6, typename KR7, typename KR8, typename KR9, 285 typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9, 286 class Cmp0, class Cmp1, class Cmp2, class Cmp3, class Cmp4, class Cmp5, class Cmp6, class Cmp7, class Cmp8, class Cmp9> 287 __device__ __forceinline__ void mergeShfl(const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>& key, 288 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val, 289 const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>& cmp, 290 unsigned int delta, int width) 291 { 292 For<0, thrust::tuple_size<thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9> >::value>::mergeShfl(key, val, cmp, delta, width); 293 } 294 template <typename KP0, typename KP1, typename KP2, typename KP3, typename KP4, typename KP5, typename KP6, typename KP7, typename KP8, typename KP9, 295 typename KR0, typename KR1, typename KR2, typename KR3, typename KR4, typename KR5, typename KR6, typename KR7, typename KR8, typename KR9, 296 typename VP0, typename VP1, typename VP2, typename VP3, typename VP4, typename VP5, typename VP6, typename VP7, typename VP8, typename VP9, 297 typename VR0, typename VR1, typename VR2, typename VR3, typename VR4, typename VR5, typename VR6, typename VR7, typename VR8, typename VR9, 298 class Cmp0, class Cmp1, class Cmp2, class Cmp3, class Cmp4, class Cmp5, class Cmp6, class Cmp7, class Cmp8, class Cmp9> 299 __device__ __forceinline__ void merge(const thrust::tuple<KP0, KP1, KP2, KP3, KP4, KP5, KP6, KP7, KP8, KP9>& skeys, 300 const thrust::tuple<KR0, KR1, KR2, KR3, KR4, KR5, KR6, KR7, KR8, KR9>& key, 301 const thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9>& svals, 302 const thrust::tuple<VR0, VR1, VR2, VR3, VR4, VR5, VR6, VR7, VR8, VR9>& val, 303 const thrust::tuple<Cmp0, Cmp1, Cmp2, Cmp3, Cmp4, Cmp5, Cmp6, Cmp7, Cmp8, Cmp9>& cmp, 304 unsigned int tid, unsigned int delta) 305 { 306 For<0, thrust::tuple_size<thrust::tuple<VP0, VP1, VP2, VP3, VP4, VP5, VP6, VP7, VP8, VP9> >::value>::merge(skeys, key, svals, val, cmp, tid, delta); 307 } 308 309 ////////////////////////////////////////////////////// 310 // Generic 311 312 template <unsigned int N> struct Generic 313 { 314 template <class KP, class KR, class VP, class VR, class Cmp> 315 static __device__ void reduce(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp) 316 { 317 loadToSmem(skeys, key, tid); 318 loadValsToSmem(svals, val, tid); 319 if (N >= 32) 320 __syncthreads(); 321 322 if (N >= 2048) 323 { 324 if (tid < 1024) 325 merge(skeys, key, svals, val, cmp, tid, 1024); 326 327 __syncthreads(); 328 } 329 if (N >= 1024) 330 { 331 if (tid < 512) 332 merge(skeys, key, svals, val, cmp, tid, 512); 333 334 __syncthreads(); 335 } 336 if (N >= 512) 337 { 338 if (tid < 256) 339 merge(skeys, key, svals, val, cmp, tid, 256); 340 341 __syncthreads(); 342 } 343 if (N >= 256) 344 { 345 if (tid < 128) 346 merge(skeys, key, svals, val, cmp, tid, 128); 347 348 __syncthreads(); 349 } 350 if (N >= 128) 351 { 352 if (tid < 64) 353 merge(skeys, key, svals, val, cmp, tid, 64); 354 355 __syncthreads(); 356 } 357 if (N >= 64) 358 { 359 if (tid < 32) 360 merge(skeys, key, svals, val, cmp, tid, 32); 361 } 362 363 if (tid < 16) 364 { 365 merge(skeys, key, svals, val, cmp, tid, 16); 366 merge(skeys, key, svals, val, cmp, tid, 8); 367 merge(skeys, key, svals, val, cmp, tid, 4); 368 merge(skeys, key, svals, val, cmp, tid, 2); 369 merge(skeys, key, svals, val, cmp, tid, 1); 370 } 371 } 372 }; 373 374 template <unsigned int I, class KP, class KR, class VP, class VR, class Cmp> 375 struct Unroll 376 { 377 static __device__ void loopShfl(KR key, VR val, Cmp cmp, unsigned int N) 378 { 379 mergeShfl(key, val, cmp, I, N); 380 Unroll<I / 2, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, N); 381 } 382 static __device__ void loop(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp) 383 { 384 merge(skeys, key, svals, val, cmp, tid, I); 385 Unroll<I / 2, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp); 386 } 387 }; 388 template <class KP, class KR, class VP, class VR, class Cmp> 389 struct Unroll<0, KP, KR, VP, VR, Cmp> 390 { 391 static __device__ void loopShfl(KR, VR, Cmp, unsigned int) 392 { 393 } 394 static __device__ void loop(KP, KR, VP, VR, unsigned int, Cmp) 395 { 396 } 397 }; 398 399 template <unsigned int N> struct WarpOptimized 400 { 401 template <class KP, class KR, class VP, class VR, class Cmp> 402 static __device__ void reduce(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp) 403 { 404 #if 0 // __CUDA_ARCH__ >= 300 405 (void) skeys; 406 (void) svals; 407 (void) tid; 408 409 Unroll<N / 2, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, N); 410 #else 411 loadToSmem(skeys, key, tid); 412 loadToSmem(svals, val, tid); 413 414 if (tid < N / 2) 415 Unroll<N / 2, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp); 416 #endif 417 } 418 }; 419 420 template <unsigned int N> struct GenericOptimized32 421 { 422 enum { M = N / 32 }; 423 424 template <class KP, class KR, class VP, class VR, class Cmp> 425 static __device__ void reduce(KP skeys, KR key, VP svals, VR val, unsigned int tid, Cmp cmp) 426 { 427 const unsigned int laneId = Warp::laneId(); 428 429 #if 0 // __CUDA_ARCH__ >= 300 430 Unroll<16, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, warpSize); 431 432 if (laneId == 0) 433 { 434 loadToSmem(skeys, key, tid / 32); 435 loadToSmem(svals, val, tid / 32); 436 } 437 #else 438 loadToSmem(skeys, key, tid); 439 loadToSmem(svals, val, tid); 440 441 if (laneId < 16) 442 Unroll<16, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp); 443 444 __syncthreads(); 445 446 if (laneId == 0) 447 { 448 loadToSmem(skeys, key, tid / 32); 449 loadToSmem(svals, val, tid / 32); 450 } 451 #endif 452 453 __syncthreads(); 454 455 loadFromSmem(skeys, key, tid); 456 457 if (tid < 32) 458 { 459 #if 0 // __CUDA_ARCH__ >= 300 460 loadFromSmem(svals, val, tid); 461 462 Unroll<M / 2, KP, KR, VP, VR, Cmp>::loopShfl(key, val, cmp, M); 463 #else 464 Unroll<M / 2, KP, KR, VP, VR, Cmp>::loop(skeys, key, svals, val, tid, cmp); 465 #endif 466 } 467 } 468 }; 469 470 template <bool val, class T1, class T2> struct StaticIf; 471 template <class T1, class T2> struct StaticIf<true, T1, T2> 472 { 473 typedef T1 type; 474 }; 475 template <class T1, class T2> struct StaticIf<false, T1, T2> 476 { 477 typedef T2 type; 478 }; 479 480 template <unsigned int N> struct IsPowerOf2 481 { 482 enum { value = ((N != 0) && !(N & (N - 1))) }; 483 }; 484 485 template <unsigned int N> struct Dispatcher 486 { 487 typedef typename StaticIf< 488 (N <= 32) && IsPowerOf2<N>::value, 489 WarpOptimized<N>, 490 typename StaticIf< 491 (N <= 1024) && IsPowerOf2<N>::value, 492 GenericOptimized32<N>, 493 Generic<N> 494 >::type 495 >::type reductor; 496 }; 497 } 498 }}} 499 500 //! @endcond 501 502 #endif // __OPENCV_CUDA_PRED_VAL_REDUCE_DETAIL_HPP__ 503
