cascadedetect.cl 26.6 KB
Newer Older
wester committed
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182
///////////////////////////// OpenCL kernels for face detection //////////////////////////////
////////////////////////////// see the opencv/doc/license.txt ///////////////////////////////

//
// the code has been derived from the OpenCL Haar cascade kernel by
//
//    Niko Li, newlife20080214@gmail.com
//    Wang Weiyan, wangweiyanster@gmail.com
//    Jia Haipeng, jiahaipeng95@gmail.com
//    Nathan, liujun@multicorewareinc.com
//    Peng Xiao, pengxiao@outlook.com
//    Erping Pang, erping@multicorewareinc.com
//

#ifdef HAAR
typedef struct __attribute__((aligned(4))) OptHaarFeature
{
    int4 ofs[3] __attribute__((aligned (4)));
    float4 weight __attribute__((aligned (4)));
}
OptHaarFeature;
#endif

#ifdef LBP
typedef struct __attribute__((aligned(4))) OptLBPFeature
{
    int16 ofs __attribute__((aligned (4)));
}
OptLBPFeature;
#endif

typedef struct __attribute__((aligned(4))) Stump
{
    float4 st __attribute__((aligned (4)));
}
Stump;

typedef struct __attribute__((aligned(4))) Node
{
    int4 n __attribute__((aligned (4)));
}
Node;

typedef struct __attribute__((aligned (4))) Stage
{
    int first __attribute__((aligned (4)));
    int ntrees __attribute__((aligned (4)));
    float threshold __attribute__((aligned (4)));
}
Stage;

typedef struct __attribute__((aligned (4))) ScaleData
{
    float scale __attribute__((aligned (4)));
    int szi_width __attribute__((aligned (4)));
    int szi_height __attribute__((aligned (4)));
    int layer_ofs __attribute__((aligned (4)));
    int ystep __attribute__((aligned (4)));
}
ScaleData;

#ifndef SUM_BUF_SIZE
#define SUM_BUF_SIZE 0
#endif

#ifndef NODE_COUNT
#define NODE_COUNT 1
#endif

#ifdef HAAR
__kernel __attribute__((reqd_work_group_size(LOCAL_SIZE_X,LOCAL_SIZE_Y,1)))
void runHaarClassifier(
    int nscales, __global const ScaleData* scaleData,
    __global const int* sum,
    int _sumstep, int sumoffset,
    __global const OptHaarFeature* optfeatures,
    __global const Stage* stages,
    __global const Node* nodes,
    __global const float* leaves0,

    volatile __global int* facepos,
    int4 normrect, int sqofs, int2 windowsize)
{
    int lx = get_local_id(0);
    int ly = get_local_id(1);
    int groupIdx = get_group_id(0);
    int i, ngroups = get_global_size(0)/LOCAL_SIZE_X;
    int scaleIdx, tileIdx, stageIdx;
    int sumstep = (int)(_sumstep/sizeof(int));
    int4 nofs0 = (int4)(mad24(normrect.y, sumstep, normrect.x),
                        mad24(normrect.y, sumstep, normrect.x + normrect.z),
                        mad24(normrect.y + normrect.w, sumstep, normrect.x),
                        mad24(normrect.y + normrect.w, sumstep, normrect.x + normrect.z));
    int normarea = normrect.z * normrect.w;
    float invarea = 1.f/normarea;
    int lidx = ly*LOCAL_SIZE_X + lx;

    #if SUM_BUF_SIZE > 0
    int4 nofs = (int4)(mad24(normrect.y, SUM_BUF_STEP, normrect.x),
                       mad24(normrect.y, SUM_BUF_STEP, normrect.x + normrect.z),
                       mad24(normrect.y + normrect.w, SUM_BUF_STEP, normrect.x),
                       mad24(normrect.y + normrect.w, SUM_BUF_STEP, normrect.x + normrect.z));
    #else
    int4 nofs = nofs0;
    #endif
    #define LOCAL_SIZE (LOCAL_SIZE_X*LOCAL_SIZE_Y)
    __local int lstore[SUM_BUF_SIZE + LOCAL_SIZE*5/2+1];
    #if SUM_BUF_SIZE > 0
    __local int* ibuf = lstore;
    __local int* lcount = ibuf + SUM_BUF_SIZE;
    #else
    __local int* lcount = lstore;
    #endif
    __local float* lnf = (__local float*)(lcount + 1);
    __local float* lpartsum = lnf + LOCAL_SIZE;
    __local short* lbuf = (__local short*)(lpartsum + LOCAL_SIZE);

    for( scaleIdx = nscales-1; scaleIdx >= 0; scaleIdx-- )
    {
        __global const ScaleData* s = scaleData + scaleIdx;
        int ystep = s->ystep;
        int2 worksize = (int2)(max(s->szi_width - windowsize.x, 0), max(s->szi_height - windowsize.y, 0));
        int2 ntiles = (int2)((worksize.x + LOCAL_SIZE_X-1)/LOCAL_SIZE_X,
                             (worksize.y + LOCAL_SIZE_Y-1)/LOCAL_SIZE_Y);
        int totalTiles = ntiles.x*ntiles.y;

        for( tileIdx = groupIdx; tileIdx < totalTiles; tileIdx += ngroups )
        {
            int ix0 = (tileIdx % ntiles.x)*LOCAL_SIZE_X;
            int iy0 = (tileIdx / ntiles.x)*LOCAL_SIZE_Y;
            int ix = lx, iy = ly;
            __global const int* psum0 = sum + mad24(iy0, sumstep, ix0) + s->layer_ofs;
            __global const int* psum1 = psum0 + mad24(iy, sumstep, ix);

            if( ix0 >= worksize.x || iy0 >= worksize.y )
                continue;
            #if SUM_BUF_SIZE > 0
            for( i = lidx*4; i < SUM_BUF_SIZE; i += LOCAL_SIZE_X*LOCAL_SIZE_Y*4 )
            {
                int dy = i/SUM_BUF_STEP, dx = i - dy*SUM_BUF_STEP;
                vstore4(vload4(0, psum0 + mad24(dy, sumstep, dx)), 0, ibuf+i);
            }
            #endif

            if( lidx == 0 )
                lcount[0] = 0;
            barrier(CLK_LOCAL_MEM_FENCE);

            if( ix0 + ix < worksize.x && iy0 + iy < worksize.y )
            {
                #if NODE_COUNT==1
                __global const Stump* stump = (__global const Stump*)nodes;
                #else
                __global const Node* node = nodes;
                __global const float* leaves = leaves0;
                #endif
                #if SUM_BUF_SIZE > 0
                __local const int* psum = ibuf + mad24(iy, SUM_BUF_STEP, ix);
                #else
                __global const int* psum = psum1;
                #endif

                __global const int* psqsum = (__global const int*)(psum1 + sqofs);
                float sval = (psum[nofs.x] - psum[nofs.y] - psum[nofs.z] + psum[nofs.w])*invarea;
                float sqval = (psqsum[nofs0.x] - psqsum[nofs0.y] - psqsum[nofs0.z] + psqsum[nofs0.w])*invarea;
                float nf = (float)normarea * sqrt(max(sqval - sval * sval, 0.f));
                nf = nf > 0 ? nf : 1.f;

                for( stageIdx = 0; stageIdx < SPLIT_STAGE; stageIdx++ )
                {
                    int ntrees = stages[stageIdx].ntrees;
                    float s = 0.f;
                    #if NODE_COUNT==1
                    for( i = 0; i < ntrees; i++ )
                    {
                        float4 st = stump[i].st;
                        __global const OptHaarFeature* f = optfeatures + as_int(st.x);
                        float4 weight = f->weight;

                        int4 ofs = f->ofs[0];
                        sval = (psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w])*weight.x;
                        ofs = f->ofs[1];
a  
Kai Westerkamp committed
183
                        sval = mad((psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.y, sval);
wester committed
184 185 186
                        if( weight.z > 0 )
                        {
                            ofs = f->ofs[2];
a  
Kai Westerkamp committed
187
                            sval = mad((psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.z, sval);
wester committed
188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206
                        }

                        s += (sval < st.y*nf) ? st.z : st.w;
                    }
                    stump += ntrees;
                    #else
                    for( i = 0; i < ntrees; i++, node += NODE_COUNT, leaves += NODE_COUNT+1 )
                    {
                        int idx = 0;
                        do
                        {
                            int4 n = node[idx].n;
                            __global const OptHaarFeature* f = optfeatures + n.x;
                            float4 weight = f->weight;

                            int4 ofs = f->ofs[0];

                            sval = (psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w])*weight.x;
                            ofs = f->ofs[1];
a  
Kai Westerkamp committed
207
                            sval = mad((psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.y, sval);
wester committed
208 209 210
                            if( weight.z > 0 )
                            {
                                ofs = f->ofs[2];
a  
Kai Westerkamp committed
211
                                sval = mad((psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.z, sval);
wester committed
212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283
                            }

                            idx = (sval < as_float(n.y)*nf) ? n.z : n.w;
                        }
                        while(idx > 0);
                        s += leaves[-idx];
                    }
                    #endif

                    if( s < stages[stageIdx].threshold )
                        break;
                }

                if( stageIdx == SPLIT_STAGE && (ystep == 1 || ((ix | iy) & 1) == 0) )
                {
                    int count = atomic_inc(lcount);
                    lbuf[count] = (int)(ix | (iy << 8));
                    lnf[count] = nf;
                }
            }

            for( stageIdx = SPLIT_STAGE; stageIdx < N_STAGES; stageIdx++ )
            {
                barrier(CLK_LOCAL_MEM_FENCE);
                int nrects = lcount[0];

                if( nrects == 0 )
                    break;
                barrier(CLK_LOCAL_MEM_FENCE);
                if( lidx == 0 )
                    lcount[0] = 0;

                {
                    #if NODE_COUNT == 1
                    __global const Stump* stump = (__global const Stump*)nodes + stages[stageIdx].first;
                    #else
                    __global const Node* node = nodes + stages[stageIdx].first*NODE_COUNT;
                    __global const float* leaves = leaves0 + stages[stageIdx].first*(NODE_COUNT+1);
                    #endif
                    int nparts = LOCAL_SIZE / nrects;
                    int ntrees = stages[stageIdx].ntrees;
                    int ntrees_p = (ntrees + nparts - 1)/nparts;
                    int nr = lidx / nparts;
                    int partidx = -1, idxval = 0;
                    float partsum = 0.f, nf = 0.f;

                    if( nr < nrects )
                    {
                        partidx = lidx % nparts;
                        idxval = lbuf[nr];
                        nf = lnf[nr];

                        {
                        int ntrees0 = ntrees_p*partidx;
                        int ntrees1 = min(ntrees0 + ntrees_p, ntrees);
                        int ix1 = idxval & 255, iy1 = idxval >> 8;
                        #if SUM_BUF_SIZE > 0
                        __local const int* psum = ibuf + mad24(iy1, SUM_BUF_STEP, ix1);
                        #else
                        __global const int* psum = psum0 + mad24(iy1, sumstep, ix1);
                        #endif

                        #if NODE_COUNT == 1
                        for( i = ntrees0; i < ntrees1; i++ )
                        {
                            float4 st = stump[i].st;
                            __global const OptHaarFeature* f = optfeatures + as_int(st.x);
                            float4 weight = f->weight;

                            int4 ofs = f->ofs[0];
                            float sval = (psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w])*weight.x;
                            ofs = f->ofs[1];
a  
Kai Westerkamp committed
284
                            sval = mad((psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.y, sval);
wester committed
285 286 287 288
                            //if( weight.z > 0 )
                            if( fabs(weight.z) > 0 )
                            {
                                ofs = f->ofs[2];
a  
Kai Westerkamp committed
289
                                sval = mad((psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.z, sval);
wester committed
290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306
                            }

                            partsum += (sval < st.y*nf) ? st.z : st.w;
                        }
                        #else
                        for( i = ntrees0; i < ntrees1; i++ )
                        {
                            int idx = 0;
                            do
                            {
                                int4 n = node[i*2 + idx].n;
                                __global const OptHaarFeature* f = optfeatures + n.x;
                                float4 weight = f->weight;
                                int4 ofs = f->ofs[0];

                                float sval = (psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w])*weight.x;
                                ofs = f->ofs[1];
a  
Kai Westerkamp committed
307
                                sval = mad((psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.y, sval);
wester committed
308 309 310
                                if( weight.z > 0 )
                                {
                                    ofs = f->ofs[2];
a  
Kai Westerkamp committed
311
                                    sval = mad((psum[ofs.x] - psum[ofs.y] - psum[ofs.z] + psum[ofs.w]), weight.z, sval);
wester committed
312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564 565 566 567 568 569 570 571 572 573 574 575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630 631 632 633 634 635 636 637 638 639 640 641 642 643 644 645 646 647 648 649 650 651 652 653 654 655 656 657 658 659 660 661
                                }

                                idx = (sval < as_float(n.y)*nf) ? n.z : n.w;
                            }
                            while(idx > 0);
                            partsum += leaves[i*3-idx];
                        }
                        #endif
                        }
                    }
                    lpartsum[lidx] = partsum;
                    barrier(CLK_LOCAL_MEM_FENCE);

                    if( partidx == 0 )
                    {
                        float s = lpartsum[nr*nparts];
                        for( i = 1; i < nparts; i++ )
                            s += lpartsum[i + nr*nparts];
                        if( s >= stages[stageIdx].threshold )
                        {
                            int count = atomic_inc(lcount);
                            lbuf[count] = idxval;
                            lnf[count] = nf;
                        }
                    }
                }
            }

            barrier(CLK_LOCAL_MEM_FENCE);
            if( stageIdx == N_STAGES )
            {
                int nrects = lcount[0];
                if( lidx < nrects )
                {
                    int nfaces = atomic_inc(facepos);
                    if( nfaces < MAX_FACES )
                    {
                        volatile __global int* face = facepos + 1 + nfaces*3;
                        int val = lbuf[lidx];
                        face[0] = scaleIdx;
                        face[1] = ix0 + (val & 255);
                        face[2] = iy0 + (val >> 8);
                    }
                }
            }
        }
    }
}
#endif

#ifdef LBP
#undef CALC_SUM_OFS_
#define CALC_SUM_OFS_(p0, p1, p2, p3, ptr) \
    ((ptr)[p0] - (ptr)[p1] - (ptr)[p2] + (ptr)[p3])

__kernel void runLBPClassifierStumpSimple(
    int nscales, __global const ScaleData* scaleData,
    __global const int* sum,
    int _sumstep, int sumoffset,
    __global const OptLBPFeature* optfeatures,
    __global const Stage* stages,
    __global const Stump* stumps,
    __global const int* bitsets,
    int bitsetSize,

    volatile __global int* facepos,
    int2 windowsize)
{
    int lx = get_local_id(0);
    int ly = get_local_id(1);
    int local_size_x = get_local_size(0);
    int local_size_y = get_local_size(1);
    int groupIdx = get_group_id(1)*get_num_groups(0) + get_group_id(0);
    int ngroups = get_num_groups(0)*get_num_groups(1);
    int scaleIdx, tileIdx, stageIdx;
    int sumstep = (int)(_sumstep/sizeof(int));

    for( scaleIdx = nscales-1; scaleIdx >= 0; scaleIdx-- )
    {
        __global const ScaleData* s = scaleData + scaleIdx;
        int ystep = s->ystep;
        int2 worksize = (int2)(max(s->szi_width - windowsize.x, 0), max(s->szi_height - windowsize.y, 0));
        int2 ntiles = (int2)((worksize.x/ystep + local_size_x-1)/local_size_x,
                             (worksize.y/ystep + local_size_y-1)/local_size_y);
        int totalTiles = ntiles.x*ntiles.y;

        for( tileIdx = groupIdx; tileIdx < totalTiles; tileIdx += ngroups )
        {
            int iy = mad24((tileIdx / ntiles.x), local_size_y, ly) * ystep;
            int ix = mad24((tileIdx % ntiles.x), local_size_x, lx) * ystep;

            if( ix < worksize.x && iy < worksize.y )
            {
                __global const int* p = sum + mad24(iy, sumstep, ix) + s->layer_ofs;
                __global const Stump* stump = stumps;
                __global const int* bitset = bitsets;

                for( stageIdx = 0; stageIdx < N_STAGES; stageIdx++ )
                {
                    int i, ntrees = stages[stageIdx].ntrees;
                    float s = 0.f;
                    for( i = 0; i < ntrees; i++, stump++, bitset += bitsetSize )
                    {
                        float4 st = stump->st;
                        __global const OptLBPFeature* f = optfeatures + as_int(st.x);
                        int16 ofs = f->ofs;

                        int cval = CALC_SUM_OFS_( ofs.s5, ofs.s6, ofs.s9, ofs.sa, p );

                        int mask, idx = (CALC_SUM_OFS_( ofs.s0, ofs.s1, ofs.s4, ofs.s5, p ) >= cval ? 4 : 0); // 0
                        idx |= (CALC_SUM_OFS_( ofs.s1, ofs.s2, ofs.s5, ofs.s6, p ) >= cval ? 2 : 0); // 1
                        idx |= (CALC_SUM_OFS_( ofs.s2, ofs.s3, ofs.s6, ofs.s7, p ) >= cval ? 1 : 0); // 2

                        mask = (CALC_SUM_OFS_( ofs.s6, ofs.s7, ofs.sa, ofs.sb, p ) >= cval ? 16 : 0); // 5
                        mask |= (CALC_SUM_OFS_( ofs.sa, ofs.sb, ofs.se, ofs.sf, p ) >= cval ? 8 : 0);  // 8
                        mask |= (CALC_SUM_OFS_( ofs.s9, ofs.sa, ofs.sd, ofs.se, p ) >= cval ? 4 : 0);  // 7
                        mask |= (CALC_SUM_OFS_( ofs.s8, ofs.s9, ofs.sc, ofs.sd, p ) >= cval ? 2 : 0);  // 6
                        mask |= (CALC_SUM_OFS_( ofs.s4, ofs.s5, ofs.s8, ofs.s9, p ) >= cval ? 1 : 0);  // 7

                        s += (bitset[idx] & (1 << mask)) ? st.z : st.w;
                    }

                    if( s < stages[stageIdx].threshold )
                        break;
                }

                if( stageIdx == N_STAGES )
                {
                    int nfaces = atomic_inc(facepos);
                    if( nfaces < MAX_FACES )
                    {
                        volatile __global int* face = facepos + 1 + nfaces*3;
                        face[0] = scaleIdx;
                        face[1] = ix;
                        face[2] = iy;
                    }
                }
            }
        }
    }
}

__kernel __attribute__((reqd_work_group_size(LOCAL_SIZE_X,LOCAL_SIZE_Y,1)))
void runLBPClassifierStump(
    int nscales, __global const ScaleData* scaleData,
    __global const int* sum,
    int _sumstep, int sumoffset,
    __global const OptLBPFeature* optfeatures,
    __global const Stage* stages,
    __global const Stump* stumps,
    __global const int* bitsets,
    int bitsetSize,

    volatile __global int* facepos,
    int2 windowsize)
{
    int lx = get_local_id(0);
    int ly = get_local_id(1);
    int groupIdx = get_group_id(0);
    int i, ngroups = get_global_size(0)/LOCAL_SIZE_X;
    int scaleIdx, tileIdx, stageIdx;
    int sumstep = (int)(_sumstep/sizeof(int));
    int lidx = ly*LOCAL_SIZE_X + lx;

    #define LOCAL_SIZE (LOCAL_SIZE_X*LOCAL_SIZE_Y)
    __local int lstore[SUM_BUF_SIZE + LOCAL_SIZE*3/2+1];
    #if SUM_BUF_SIZE > 0
    __local int* ibuf = lstore;
    __local int* lcount = ibuf + SUM_BUF_SIZE;
    #else
    __local int* lcount = lstore;
    #endif
    __local float* lpartsum = (__local float*)(lcount + 1);
    __local short* lbuf = (__local short*)(lpartsum + LOCAL_SIZE);

    for( scaleIdx = nscales-1; scaleIdx >= 0; scaleIdx-- )
    {
        __global const ScaleData* s = scaleData + scaleIdx;
        int ystep = s->ystep;
        int2 worksize = (int2)(max(s->szi_width - windowsize.x, 0), max(s->szi_height - windowsize.y, 0));
        int2 ntiles = (int2)((worksize.x + LOCAL_SIZE_X-1)/LOCAL_SIZE_X,
                             (worksize.y + LOCAL_SIZE_Y-1)/LOCAL_SIZE_Y);
        int totalTiles = ntiles.x*ntiles.y;

        for( tileIdx = groupIdx; tileIdx < totalTiles; tileIdx += ngroups )
        {
            int ix0 = (tileIdx % ntiles.x)*LOCAL_SIZE_X;
            int iy0 = (tileIdx / ntiles.x)*LOCAL_SIZE_Y;
            int ix = lx, iy = ly;
            __global const int* psum0 = sum + mad24(iy0, sumstep, ix0) + s->layer_ofs;

            if( ix0 >= worksize.x || iy0 >= worksize.y )
                continue;
            #if SUM_BUF_SIZE > 0
            for( i = lidx*4; i < SUM_BUF_SIZE; i += LOCAL_SIZE_X*LOCAL_SIZE_Y*4 )
            {
                int dy = i/SUM_BUF_STEP, dx = i - dy*SUM_BUF_STEP;
                vstore4(vload4(0, psum0 + mad24(dy, sumstep, dx)), 0, ibuf+i);
            }
            barrier(CLK_LOCAL_MEM_FENCE);
            #endif

            if( lidx == 0 )
                lcount[0] = 0;
            barrier(CLK_LOCAL_MEM_FENCE);

            if( ix0 + ix < worksize.x && iy0 + iy < worksize.y )
            {
                __global const Stump* stump = stumps;
                __global const int* bitset = bitsets;
                #if SUM_BUF_SIZE > 0
                __local const int* p = ibuf + mad24(iy, SUM_BUF_STEP, ix);
                #else
                __global const int* p = psum0 + mad24(iy, sumstep, ix);
                #endif

                for( stageIdx = 0; stageIdx < SPLIT_STAGE; stageIdx++ )
                {
                    int ntrees = stages[stageIdx].ntrees;
                    float s = 0.f;
                    for( i = 0; i < ntrees; i++, stump++, bitset += bitsetSize )
                    {
                        float4 st = stump->st;
                        __global const OptLBPFeature* f = optfeatures + as_int(st.x);
                        int16 ofs = f->ofs;

                        int cval = CALC_SUM_OFS_( ofs.s5, ofs.s6, ofs.s9, ofs.sa, p );

                        int mask, idx = (CALC_SUM_OFS_( ofs.s0, ofs.s1, ofs.s4, ofs.s5, p ) >= cval ? 4 : 0); // 0
                        idx |= (CALC_SUM_OFS_( ofs.s1, ofs.s2, ofs.s5, ofs.s6, p ) >= cval ? 2 : 0); // 1
                        idx |= (CALC_SUM_OFS_( ofs.s2, ofs.s3, ofs.s6, ofs.s7, p ) >= cval ? 1 : 0); // 2

                        mask = (CALC_SUM_OFS_( ofs.s6, ofs.s7, ofs.sa, ofs.sb, p ) >= cval ? 16 : 0); // 5
                        mask |= (CALC_SUM_OFS_( ofs.sa, ofs.sb, ofs.se, ofs.sf, p ) >= cval ? 8 : 0);  // 8
                        mask |= (CALC_SUM_OFS_( ofs.s9, ofs.sa, ofs.sd, ofs.se, p ) >= cval ? 4 : 0);  // 7
                        mask |= (CALC_SUM_OFS_( ofs.s8, ofs.s9, ofs.sc, ofs.sd, p ) >= cval ? 2 : 0);  // 6
                        mask |= (CALC_SUM_OFS_( ofs.s4, ofs.s5, ofs.s8, ofs.s9, p ) >= cval ? 1 : 0);  // 7

                        s += (bitset[idx] & (1 << mask)) ? st.z : st.w;
                    }

                    if( s < stages[stageIdx].threshold )
                        break;
                }

                if( stageIdx == SPLIT_STAGE && (ystep == 1 || ((ix | iy) & 1) == 0) )
                {
                    int count = atomic_inc(lcount);
                    lbuf[count] = (int)(ix | (iy << 8));
                }
            }

            for( stageIdx = SPLIT_STAGE; stageIdx < N_STAGES; stageIdx++ )
            {
                int nrects = lcount[0];

                barrier(CLK_LOCAL_MEM_FENCE);
                if( nrects == 0 )
                    break;
                if( lidx == 0 )
                    lcount[0] = 0;

                {
                    __global const Stump* stump = stumps + stages[stageIdx].first;
                    __global const int* bitset = bitsets + stages[stageIdx].first*bitsetSize;
                    int nparts = LOCAL_SIZE / nrects;
                    int ntrees = stages[stageIdx].ntrees;
                    int ntrees_p = (ntrees + nparts - 1)/nparts;
                    int nr = lidx / nparts;
                    int partidx = -1, idxval = 0;
                    float partsum = 0.f, nf = 0.f;

                    if( nr < nrects )
                    {
                        partidx = lidx % nparts;
                        idxval = lbuf[nr];

                        {
                            int ntrees0 = ntrees_p*partidx;
                            int ntrees1 = min(ntrees0 + ntrees_p, ntrees);
                            int ix1 = idxval & 255, iy1 = idxval >> 8;
                            #if SUM_BUF_SIZE > 0
                            __local const int* p = ibuf + mad24(iy1, SUM_BUF_STEP, ix1);
                            #else
                            __global const int* p = psum0 + mad24(iy1, sumstep, ix1);
                            #endif

                            for( i = ntrees0; i < ntrees1; i++ )
                            {
                                float4 st = stump[i].st;
                                __global const OptLBPFeature* f = optfeatures + as_int(st.x);
                                int16 ofs = f->ofs;

                                #define CALC_SUM_OFS_(p0, p1, p2, p3, ptr) \
                                    ((ptr)[p0] - (ptr)[p1] - (ptr)[p2] + (ptr)[p3])

                                int cval = CALC_SUM_OFS_( ofs.s5, ofs.s6, ofs.s9, ofs.sa, p );

                                int mask, idx = (CALC_SUM_OFS_( ofs.s0, ofs.s1, ofs.s4, ofs.s5, p ) >= cval ? 4 : 0); // 0
                                idx |= (CALC_SUM_OFS_( ofs.s1, ofs.s2, ofs.s5, ofs.s6, p ) >= cval ? 2 : 0); // 1
                                idx |= (CALC_SUM_OFS_( ofs.s2, ofs.s3, ofs.s6, ofs.s7, p ) >= cval ? 1 : 0); // 2

                                mask = (CALC_SUM_OFS_( ofs.s6, ofs.s7, ofs.sa, ofs.sb, p ) >= cval ? 16 : 0); // 5
                                mask |= (CALC_SUM_OFS_( ofs.sa, ofs.sb, ofs.se, ofs.sf, p ) >= cval ? 8 : 0);  // 8
                                mask |= (CALC_SUM_OFS_( ofs.s9, ofs.sa, ofs.sd, ofs.se, p ) >= cval ? 4 : 0);  // 7
                                mask |= (CALC_SUM_OFS_( ofs.s8, ofs.s9, ofs.sc, ofs.sd, p ) >= cval ? 2 : 0);  // 6
                                mask |= (CALC_SUM_OFS_( ofs.s4, ofs.s5, ofs.s8, ofs.s9, p ) >= cval ? 1 : 0);  // 7

                                partsum += (bitset[i*bitsetSize + idx] & (1 << mask)) ? st.z : st.w;
                            }
                        }
                    }
                    lpartsum[lidx] = partsum;
                    barrier(CLK_LOCAL_MEM_FENCE);

                    if( partidx == 0 )
                    {
                        float s = lpartsum[nr*nparts];
                        for( i = 1; i < nparts; i++ )
                            s += lpartsum[i + nr*nparts];
                        if( s >= stages[stageIdx].threshold )
                        {
                            int count = atomic_inc(lcount);
                            lbuf[count] = idxval;
                        }
                    }
                }
            }

            barrier(CLK_LOCAL_MEM_FENCE);
            if( stageIdx == N_STAGES )
            {
                int nrects = lcount[0];
                if( lidx < nrects )
                {
                    int nfaces = atomic_inc(facepos);
                    if( nfaces < MAX_FACES )
                    {
                        volatile __global int* face = facepos + 1 + nfaces*3;
                        int val = lbuf[lidx];
                        face[0] = scaleIdx;
                        face[1] = ix0 + (val & 255);
                        face[2] = iy0 + (val >> 8);
                    }
                }
            }
        }
    }
}
#endif