| 1 | // Copyright 2009-2021 Intel Corporation |
| 2 | // SPDX-License-Identifier: Apache-2.0 |
| 3 | |
| 4 | // TODO: |
| 5 | // - adjust parallel build thresholds |
| 6 | // - openNodesBasedOnExtend should consider max extended size |
| 7 | |
| 8 | #pragma once |
| 9 | |
| 10 | #include "heuristic_binning.h" |
| 11 | #include "heuristic_spatial.h" |
| 12 | |
| 13 | /* stop opening of all bref.geomIDs are the same */ |
| 14 | #define EQUAL_GEOMID_STOP_CRITERIA 1 |
| 15 | |
| 16 | /* 10% spatial extend threshold */ |
| 17 | #define MAX_EXTEND_THRESHOLD 0.1f |
| 18 | |
| 19 | /* maximum is 8 children */ |
| 20 | #define MAX_OPENED_CHILD_NODES 8 |
| 21 | |
| 22 | /* open until all build refs are below threshold size in one step */ |
| 23 | #define USE_LOOP_OPENING 0 |
| 24 | |
| 25 | namespace embree |
| 26 | { |
| 27 | namespace isa |
| 28 | { |
| 29 | /*! Performs standard object binning */ |
| 30 | template<typename NodeOpenerFunc, typename PrimRef, size_t OBJECT_BINS> |
| 31 | struct HeuristicArrayOpenMergeSAH |
| 32 | { |
| 33 | typedef BinSplit<OBJECT_BINS> Split; |
| 34 | typedef BinInfoT<OBJECT_BINS,PrimRef,BBox3fa> Binner; |
| 35 | |
| 36 | static const size_t PARALLEL_THRESHOLD = 1024; |
| 37 | static const size_t PARALLEL_FIND_BLOCK_SIZE = 512; |
| 38 | static const size_t PARALLEL_PARTITION_BLOCK_SIZE = 128; |
| 39 | |
| 40 | static const size_t MOVE_STEP_SIZE = 64; |
| 41 | static const size_t CREATE_SPLITS_STEP_SIZE = 128; |
| 42 | |
| 43 | __forceinline HeuristicArrayOpenMergeSAH () |
| 44 | : prims0(nullptr) {} |
| 45 | |
| 46 | /*! remember prim array */ |
| 47 | __forceinline HeuristicArrayOpenMergeSAH (const NodeOpenerFunc& nodeOpenerFunc, PrimRef* prims0, size_t max_open_size) |
| 48 | : prims0(prims0), nodeOpenerFunc(nodeOpenerFunc), max_open_size(max_open_size) |
| 49 | { |
| 50 | assert(max_open_size <= MAX_OPENED_CHILD_NODES); |
| 51 | } |
| 52 | |
| 53 | struct OpenHeuristic |
| 54 | { |
| 55 | __forceinline ( const PrimInfoExtRange& pinfo ) |
| 56 | { |
| 57 | const Vec3fa diag = pinfo.geomBounds.size(); |
| 58 | dim = maxDim(diag); |
| 59 | assert(diag[dim] > 0.0f); |
| 60 | inv_max_extend = 1.0f / diag[dim]; |
| 61 | } |
| 62 | |
| 63 | __forceinline bool operator () ( PrimRef& prim ) const { |
| 64 | return !prim.node.isLeaf() && prim.bounds().size()[dim] * inv_max_extend > MAX_EXTEND_THRESHOLD; |
| 65 | } |
| 66 | |
| 67 | private: |
| 68 | size_t dim; |
| 69 | float inv_max_extend; |
| 70 | }; |
| 71 | |
| 72 | /*! compute extended ranges */ |
| 73 | __forceinline void (const PrimInfoExtRange& set, PrimInfoExtRange& lset, PrimInfoExtRange& rset, const size_t lweight, const size_t rweight) |
| 74 | { |
| 75 | assert(set.ext_range_size() > 0); |
| 76 | const float left_factor = (float)lweight / (lweight + rweight); |
| 77 | const size_t ext_range_size = set.ext_range_size(); |
| 78 | const size_t left_ext_range_size = min((size_t)(floorf(left_factor * ext_range_size)),ext_range_size); |
| 79 | const size_t right_ext_range_size = ext_range_size - left_ext_range_size; |
| 80 | lset.set_ext_range(lset.end() + left_ext_range_size); |
| 81 | rset.set_ext_range(rset.end() + right_ext_range_size); |
| 82 | } |
| 83 | |
| 84 | /*! move ranges */ |
| 85 | __forceinline void (const PrimInfoExtRange& set, const PrimInfoExtRange& lset, PrimInfoExtRange& rset) |
| 86 | { |
| 87 | const size_t left_ext_range_size = lset.ext_range_size(); |
| 88 | const size_t right_size = rset.size(); |
| 89 | |
| 90 | /* has the left child an extended range? */ |
| 91 | if (left_ext_range_size > 0) |
| 92 | { |
| 93 | /* left extended range smaller than right range ? */ |
| 94 | if (left_ext_range_size < right_size) |
| 95 | { |
| 96 | /* only move a small part of the beginning of the right range to the end */ |
| 97 | parallel_for( rset.begin(), rset.begin()+left_ext_range_size, MOVE_STEP_SIZE, [&](const range<size_t>& r) { |
| 98 | for (size_t i=r.begin(); i<r.end(); i++) |
| 99 | prims0[i+right_size] = prims0[i]; |
| 100 | }); |
| 101 | } |
| 102 | else |
| 103 | { |
| 104 | /* no overlap, move entire right range to new location, can be made fully parallel */ |
| 105 | parallel_for( rset.begin(), rset.end(), MOVE_STEP_SIZE, [&](const range<size_t>& r) { |
| 106 | for (size_t i=r.begin(); i<r.end(); i++) |
| 107 | prims0[i+left_ext_range_size] = prims0[i]; |
| 108 | }); |
| 109 | } |
| 110 | /* update right range */ |
| 111 | assert(rset.ext_end() + left_ext_range_size == set.ext_end()); |
| 112 | rset.move_right(left_ext_range_size); |
| 113 | } |
| 114 | } |
| 115 | |
| 116 | /* estimates the extra space required when opening, and checks if all primitives are from same geometry */ |
| 117 | __noinline std::pair<size_t,bool> (const PrimInfoExtRange& set) |
| 118 | { |
| 119 | const OpenHeuristic heuristic(set); |
| 120 | const unsigned int geomID = prims0[set.begin()].geomID(); |
| 121 | |
| 122 | auto body = [&] (const range<size_t>& r) -> std::pair<size_t,bool> { |
| 123 | bool commonGeomID = true; |
| 124 | size_t opens = 0; |
| 125 | for (size_t i=r.begin(); i<r.end(); i++) { |
| 126 | commonGeomID &= prims0[i].geomID() == geomID; |
| 127 | if (heuristic(prims0[i])) |
| 128 | opens += prims0[i].node.getN()-1; // coarse approximation |
| 129 | } |
| 130 | return std::pair<size_t,bool>(opens,commonGeomID); |
| 131 | }; |
| 132 | auto reduction = [&] (const std::pair<size_t,bool>& b0, const std::pair<size_t,bool>& b1) -> std::pair<size_t,bool> { |
| 133 | return std::pair<size_t,bool>(b0.first+b1.first,b0.second && b1.second); |
| 134 | }; |
| 135 | return parallel_reduce(set.begin(),set.end(),PARALLEL_FIND_BLOCK_SIZE,PARALLEL_THRESHOLD,std::pair<size_t,bool>(0,true),body,reduction); |
| 136 | } |
| 137 | |
| 138 | // FIXME: should consider maximum available extended size |
| 139 | __noinline void (PrimInfoExtRange& set) |
| 140 | { |
| 141 | const OpenHeuristic heuristic(set); |
| 142 | const size_t ext_range_start = set.end(); |
| 143 | |
| 144 | if (false && set.size() < PARALLEL_THRESHOLD) |
| 145 | { |
| 146 | size_t = 0; |
| 147 | for (size_t i=set.begin(); i<set.end(); i++) |
| 148 | { |
| 149 | if (heuristic(prims0[i])) |
| 150 | { |
| 151 | PrimRef tmp[MAX_OPENED_CHILD_NODES]; |
| 152 | const size_t n = nodeOpenerFunc(prims0[i],tmp); |
| 153 | assert(extra_elements + n-1 <= set.ext_range_size()); |
| 154 | for (size_t j=0; j<n; j++) |
| 155 | set.extend_center2(tmp[j]); |
| 156 | |
| 157 | prims0[i] = tmp[0]; |
| 158 | for (size_t j=1; j<n; j++) |
| 159 | prims0[ext_range_start+extra_elements+j-1] = tmp[j]; |
| 160 | extra_elements += n-1; |
| 161 | } |
| 162 | } |
| 163 | set._end += extra_elements; |
| 164 | } |
| 165 | else |
| 166 | { |
| 167 | std::atomic<size_t> ext_elements; |
| 168 | ext_elements.store(0); |
| 169 | PrimInfo info = parallel_reduce( set.begin(), set.end(), CREATE_SPLITS_STEP_SIZE, PrimInfo(empty), [&](const range<size_t>& r) -> PrimInfo { |
| 170 | PrimInfo info(empty); |
| 171 | for (size_t i=r.begin(); i<r.end(); i++) |
| 172 | if (heuristic(prims0[i])) |
| 173 | { |
| 174 | PrimRef tmp[MAX_OPENED_CHILD_NODES]; |
| 175 | const size_t n = nodeOpenerFunc(prims0[i],tmp); |
| 176 | const size_t ID = ext_elements.fetch_add(n-1); |
| 177 | assert(ID + n-1 <= set.ext_range_size()); |
| 178 | |
| 179 | for (size_t j=0; j<n; j++) |
| 180 | info.extend_center2(tmp[j]); |
| 181 | |
| 182 | prims0[i] = tmp[0]; |
| 183 | for (size_t j=1; j<n; j++) |
| 184 | prims0[ext_range_start+ID+j-1] = tmp[j]; |
| 185 | } |
| 186 | return info; |
| 187 | }, [] (const PrimInfo& a, const PrimInfo& b) { return PrimInfo::merge(a,b); }); |
| 188 | set.centBounds.extend(info.centBounds); |
| 189 | assert(ext_elements.load() <= set.ext_range_size()); |
| 190 | set._end += ext_elements.load(); |
| 191 | } |
| 192 | } |
| 193 | |
| 194 | __noinline void (PrimInfoExtRange& set, const size_t est_new_elements) |
| 195 | { |
| 196 | const OpenHeuristic heuristic(set); |
| 197 | size_t = est_new_elements; |
| 198 | |
| 199 | while (next_iteration_extra_elements <= set.ext_range_size()) |
| 200 | { |
| 201 | next_iteration_extra_elements = 0; |
| 202 | size_t = 0; |
| 203 | const size_t ext_range_start = set.end(); |
| 204 | |
| 205 | for (size_t i=set.begin(); i<set.end(); i++) |
| 206 | { |
| 207 | if (heuristic(prims0[i])) |
| 208 | { |
| 209 | PrimRef tmp[MAX_OPENED_CHILD_NODES]; |
| 210 | const size_t n = nodeOpenerFunc(prims0[i],tmp); |
| 211 | assert(extra_elements + n-1 <= set.ext_range_size()); |
| 212 | for (size_t j=0;j<n;j++) |
| 213 | set.extend_center2(tmp[j]); |
| 214 | |
| 215 | prims0[i] = tmp[0]; |
| 216 | for (size_t j=1;j<n;j++) |
| 217 | prims0[ext_range_start+extra_elements+j-1] = tmp[j]; |
| 218 | extra_elements += n-1; |
| 219 | |
| 220 | for (size_t j=0; j<n; j++) |
| 221 | if (heuristic(tmp[j])) |
| 222 | next_iteration_extra_elements += tmp[j].node.getN()-1; // coarse approximation |
| 223 | |
| 224 | } |
| 225 | } |
| 226 | assert( extra_elements <= set.ext_range_size()); |
| 227 | set._end += extra_elements; |
| 228 | |
| 229 | for (size_t i=set.begin();i<set.end();i++) |
| 230 | assert(prims0[i].numPrimitives() > 0); |
| 231 | |
| 232 | if (unlikely(next_iteration_extra_elements == 0)) break; |
| 233 | } |
| 234 | } |
| 235 | |
| 236 | __noinline const Split (PrimInfoExtRange& set, const size_t logBlockSize) |
| 237 | { |
| 238 | /* single element */ |
| 239 | if (set.size() <= 1) |
| 240 | return Split(); |
| 241 | |
| 242 | /* disable opening if there is no overlap */ |
| 243 | const size_t D = 4; |
| 244 | if (unlikely(set.has_ext_range() && set.size() <= D)) |
| 245 | { |
| 246 | bool disjoint = true; |
| 247 | for (size_t j=set.begin(); j<set.end()-1; j++) { |
| 248 | for (size_t i=set.begin()+1; i<set.end(); i++) { |
| 249 | if (conjoint(prims0[j].bounds(),prims0[i].bounds())) { |
| 250 | disjoint = false; break; |
| 251 | } |
| 252 | } |
| 253 | } |
| 254 | if (disjoint) set.set_ext_range(set.end()); /* disables opening */ |
| 255 | } |
| 256 | |
| 257 | std::pair<size_t,bool> p(0,false); |
| 258 | |
| 259 | /* disable opening when all primitives are from same geometry */ |
| 260 | if (unlikely(set.has_ext_range())) |
| 261 | { |
| 262 | p = getProperties(set); |
| 263 | #if EQUAL_GEOMID_STOP_CRITERIA == 1 |
| 264 | if (p.second) set.set_ext_range(set.end()); /* disable opening */ |
| 265 | #endif |
| 266 | } |
| 267 | |
| 268 | /* open nodes when we have sufficient space available */ |
| 269 | if (unlikely(set.has_ext_range())) |
| 270 | { |
| 271 | #if USE_LOOP_OPENING == 1 |
| 272 | openNodesBasedOnExtendLoop(set,p.first); |
| 273 | #else |
| 274 | if (p.first <= set.ext_range_size()) |
| 275 | openNodesBasedOnExtend(set); |
| 276 | #endif |
| 277 | |
| 278 | /* disable opening when insufficient space for opening a node available */ |
| 279 | if (set.ext_range_size() < max_open_size-1) |
| 280 | set.set_ext_range(set.end()); /* disable opening */ |
| 281 | } |
| 282 | |
| 283 | /* find best split */ |
| 284 | return object_find(set,logBlockSize); |
| 285 | } |
| 286 | |
| 287 | |
| 288 | /*! finds the best object split */ |
| 289 | __forceinline const Split (const PrimInfoExtRange& set,const size_t logBlockSize) |
| 290 | { |
| 291 | if (set.size() < PARALLEL_THRESHOLD) return sequential_object_find(set,logBlockSize); |
| 292 | else return parallel_object_find (set,logBlockSize); |
| 293 | } |
| 294 | |
| 295 | /*! finds the best object split */ |
| 296 | __noinline const Split (const PrimInfoExtRange& set, const size_t logBlockSize) |
| 297 | { |
| 298 | Binner binner(empty); |
| 299 | const BinMapping<OBJECT_BINS> mapping(set.centBounds); |
| 300 | binner.bin(prims0,set.begin(),set.end(),mapping); |
| 301 | return binner.best(mapping,logBlockSize); |
| 302 | } |
| 303 | |
| 304 | /*! finds the best split */ |
| 305 | __noinline const Split (const PrimInfoExtRange& set, const size_t logBlockSize) |
| 306 | { |
| 307 | Binner binner(empty); |
| 308 | const BinMapping<OBJECT_BINS> mapping(set.centBounds); |
| 309 | const BinMapping<OBJECT_BINS>& _mapping = mapping; // CLANG 3.4 parser bug workaround |
| 310 | auto body = [&] (const range<size_t>& r) -> Binner { |
| 311 | Binner binner(empty); binner.bin(prims0+r.begin(),r.size(),_mapping); return binner; |
| 312 | }; |
| 313 | auto reduction = [&] (const Binner& b0, const Binner& b1) -> Binner { |
| 314 | Binner r = b0; r.merge(b1,_mapping.size()); return r; |
| 315 | }; |
| 316 | binner = parallel_reduce(set.begin(),set.end(),PARALLEL_FIND_BLOCK_SIZE,binner,body,reduction); |
| 317 | return binner.best(mapping,logBlockSize); |
| 318 | } |
| 319 | |
| 320 | /*! array partitioning */ |
| 321 | __noinline void (const Split& split, const PrimInfoExtRange& set_i, PrimInfoExtRange& lset, PrimInfoExtRange& rset) |
| 322 | { |
| 323 | PrimInfoExtRange set = set_i; |
| 324 | |
| 325 | /* valid split */ |
| 326 | if (unlikely(!split.valid())) { |
| 327 | deterministic_order(set); |
| 328 | splitFallback(set,lset,rset); |
| 329 | return; |
| 330 | } |
| 331 | |
| 332 | std::pair<size_t,size_t> ext_weights(0,0); |
| 333 | |
| 334 | /* object split */ |
| 335 | if (likely(set.size() < PARALLEL_THRESHOLD)) |
| 336 | ext_weights = sequential_object_split(split,set,lset,rset); |
| 337 | else |
| 338 | ext_weights = parallel_object_split(split,set,lset,rset); |
| 339 | |
| 340 | /* if we have an extended range, set extended child ranges and move right split range */ |
| 341 | if (unlikely(set.has_ext_range())) |
| 342 | { |
| 343 | setExtentedRanges(set,lset,rset,ext_weights.first,ext_weights.second); |
| 344 | moveExtentedRange(set,lset,rset); |
| 345 | } |
| 346 | } |
| 347 | |
| 348 | /*! array partitioning */ |
| 349 | std::pair<size_t,size_t> (const Split& split, const PrimInfoExtRange& set, PrimInfoExtRange& lset, PrimInfoExtRange& rset) |
| 350 | { |
| 351 | const size_t begin = set.begin(); |
| 352 | const size_t end = set.end(); |
| 353 | PrimInfo local_left(empty); |
| 354 | PrimInfo local_right(empty); |
| 355 | const unsigned int splitPos = split.pos; |
| 356 | const unsigned int splitDim = split.dim; |
| 357 | const unsigned int splitDimMask = (unsigned int)1 << splitDim; |
| 358 | |
| 359 | const vint4 vSplitPos(splitPos); |
| 360 | const vbool4 vSplitMask( (int)splitDimMask ); |
| 361 | |
| 362 | size_t center = serial_partitioning(prims0, |
| 363 | begin,end,local_left,local_right, |
| 364 | [&] (const PrimRef& ref) { return split.mapping.bin_unsafe(ref,vSplitPos,vSplitMask); }, |
| 365 | [] (PrimInfo& pinfo,const PrimRef& ref) { pinfo.add_center2(ref); }); |
| 366 | |
| 367 | new (&lset) PrimInfoExtRange(begin,center,center,local_left); |
| 368 | new (&rset) PrimInfoExtRange(center,end,end,local_right); |
| 369 | assert(area(lset.geomBounds) >= 0.0f); |
| 370 | assert(area(rset.geomBounds) >= 0.0f); |
| 371 | return std::pair<size_t,size_t>(local_left.size(),local_right.size()); |
| 372 | } |
| 373 | |
| 374 | /*! array partitioning */ |
| 375 | __noinline std::pair<size_t,size_t> (const Split& split, const PrimInfoExtRange& set, PrimInfoExtRange& lset, PrimInfoExtRange& rset) |
| 376 | { |
| 377 | const size_t begin = set.begin(); |
| 378 | const size_t end = set.end(); |
| 379 | PrimInfo left(empty); |
| 380 | PrimInfo right(empty); |
| 381 | const unsigned int splitPos = split.pos; |
| 382 | const unsigned int splitDim = split.dim; |
| 383 | const unsigned int splitDimMask = (unsigned int)1 << splitDim; |
| 384 | |
| 385 | const vint4 vSplitPos(splitPos); |
| 386 | const vbool4 vSplitMask( (int)splitDimMask ); |
| 387 | auto isLeft = [&] (const PrimRef& ref) { return split.mapping.bin_unsafe(ref,vSplitPos,vSplitMask); }; |
| 388 | |
| 389 | const size_t center = parallel_partitioning( |
| 390 | prims0,begin,end,EmptyTy(),left,right,isLeft, |
| 391 | [] (PrimInfo& pinfo,const PrimRef& ref) { pinfo.add_center2(ref); }, |
| 392 | [] (PrimInfo& pinfo0,const PrimInfo& pinfo1) { pinfo0.merge(pinfo1); }, |
| 393 | PARALLEL_PARTITION_BLOCK_SIZE); |
| 394 | |
| 395 | new (&lset) PrimInfoExtRange(begin,center,center,left); |
| 396 | new (&rset) PrimInfoExtRange(center,end,end,right); |
| 397 | assert(area(lset.geomBounds) >= 0.0f); |
| 398 | assert(area(rset.geomBounds) >= 0.0f); |
| 399 | |
| 400 | return std::pair<size_t,size_t>(left.size(),right.size()); |
| 401 | } |
| 402 | |
| 403 | void deterministic_order(const extended_range<size_t>& set) |
| 404 | { |
| 405 | /* required as parallel partition destroys original primitive order */ |
| 406 | std::sort(&prims0[set.begin()],&prims0[set.end()]); |
| 407 | } |
| 408 | |
| 409 | __forceinline void (const PrimInfoExtRange& set, PrimInfoExtRange& lset, PrimInfoExtRange& rset) |
| 410 | { |
| 411 | const size_t begin = set.begin(); |
| 412 | const size_t end = set.end(); |
| 413 | const size_t center = (begin + end)/2; |
| 414 | |
| 415 | PrimInfo left(empty); |
| 416 | for (size_t i=begin; i<center; i++) |
| 417 | left.add_center2(prims0[i]); |
| 418 | |
| 419 | const size_t lweight = left.end; |
| 420 | |
| 421 | PrimInfo right(empty); |
| 422 | for (size_t i=center; i<end; i++) |
| 423 | right.add_center2(prims0[i]); |
| 424 | |
| 425 | const size_t rweight = right.end; |
| 426 | new (&lset) PrimInfoExtRange(begin,center,center,left); |
| 427 | new (&rset) PrimInfoExtRange(center,end,end,right); |
| 428 | |
| 429 | /* if we have an extended range */ |
| 430 | if (set.has_ext_range()) |
| 431 | { |
| 432 | setExtentedRanges(set,lset,rset,lweight,rweight); |
| 433 | moveExtentedRange(set,lset,rset); |
| 434 | } |
| 435 | } |
| 436 | |
| 437 | private: |
| 438 | PrimRef* const prims0; |
| 439 | const NodeOpenerFunc& nodeOpenerFunc; |
| 440 | size_t max_open_size; |
| 441 | }; |
| 442 | } |
| 443 | } |
| 444 | |