static void RE_rayobject_blibvh_add(RayObject *o, RayObject *ob)
{
BVHObject *obj = (BVHObject*)o;
float min_max[6];
INIT_MINMAX(min_max, min_max+3);
RE_rayobject_merge_bb(ob, min_max, min_max+3);
DO_MIN(min_max, obj->bb[0]);
DO_MAX(min_max + 3, obj->bb[1]);
BLI_bvhtree_insert(obj->bvh, obj->next_leaf - obj->leafs, min_max, 2);
*(obj->next_leaf++) = ob;
}
/*
====================
CL_FinishTimeDemo
====================
*/
static void CL_FinishTimeDemo (void)
{
int frames;
int i;
double time, totalfpsavg;
double fpsmin, fpsavg, fpsmax; // report min/avg/max fps
static int benchmark_runs = 0;
char vabuf[1024];
cls.timedemo = false;
frames = cls.td_frames;
time = realtime - cls.td_starttime;
totalfpsavg = time > 0 ? frames / time : 0;
fpsmin = cls.td_onesecondminfps;
fpsavg = cls.td_onesecondavgcount ? cls.td_onesecondavgfps / cls.td_onesecondavgcount : 0;
fpsmax = cls.td_onesecondmaxfps;
// LordHavoc: timedemo now prints out 7 digits of fraction, and min/avg/max
Con_Printf("%i frames %5.7f seconds %5.7f fps, one-second fps min/avg/max: %.0f %.0f %.0f (%i seconds)\n", frames, time, totalfpsavg, fpsmin, fpsavg, fpsmax, cls.td_onesecondavgcount);
Log_Printf("benchmark.log", "date %s | enginedate %s | demo %s | commandline %s | run %d | result %i frames %5.7f seconds %5.7f fps, one-second fps min/avg/max: %.0f %.0f %.0f (%i seconds)\n", Sys_TimeString("%Y-%m-%d %H:%M:%S"), buildstring, cls.demoname, cmdline.string, benchmark_runs + 1, frames, time, totalfpsavg, fpsmin, fpsavg, fpsmax, cls.td_onesecondavgcount);
if (COM_CheckParm("-benchmark"))
{
++benchmark_runs;
i = COM_CheckParm("-benchmarkruns");
if(i && i + 1 < com_argc)
{
static benchmarkhistory_t *history = NULL;
if(!history)
history = (benchmarkhistory_t *)Z_Malloc(sizeof(*history) * atoi(com_argv[i + 1]));
history[benchmark_runs - 1].frames = frames;
history[benchmark_runs - 1].time = time;
history[benchmark_runs - 1].totalfpsavg = totalfpsavg;
history[benchmark_runs - 1].fpsmin = fpsmin;
history[benchmark_runs - 1].fpsavg = fpsavg;
history[benchmark_runs - 1].fpsmax = fpsmax;
if(atoi(com_argv[i + 1]) > benchmark_runs)
{
// restart the benchmark
Cbuf_AddText(va(vabuf, sizeof(vabuf), "timedemo %s\n", cls.demoname));
// cannot execute here
}
else
{
// print statistics
int first = COM_CheckParm("-benchmarkruns_skipfirst") ? 1 : 0;
if(benchmark_runs > first)
{
#define DO_MIN(f) \
for(i = first; i < benchmark_runs; ++i) if((i == first) || (history[i].f < f)) f = history[i].f
#define DO_MAX(f) \
for(i = first; i < benchmark_runs; ++i) if((i == first) || (history[i].f > f)) f = history[i].f
#define DO_MED(f) \
doublecmp_offset = (char *)&history->f - (char *)history; \
qsort(history + first, benchmark_runs - first, sizeof(*history), doublecmp_withoffset); \
if((first + benchmark_runs) & 1) \
f = history[(first + benchmark_runs - 1) / 2].f; \
else \
f = (history[(first + benchmark_runs - 2) / 2].f + history[(first + benchmark_runs) / 2].f) / 2
DO_MIN(frames);
DO_MAX(time);
DO_MIN(totalfpsavg);
DO_MIN(fpsmin);
DO_MIN(fpsavg);
DO_MIN(fpsmax);
Con_Printf("MIN: %i frames %5.7f seconds %5.7f fps, one-second fps min/avg/max: %.0f %.0f %.0f (%i seconds)\n", frames, time, totalfpsavg, fpsmin, fpsavg, fpsmax, cls.td_onesecondavgcount);
DO_MED(frames);
DO_MED(time);
DO_MED(totalfpsavg);
DO_MED(fpsmin);
DO_MED(fpsavg);
DO_MED(fpsmax);
Con_Printf("MED: %i frames %5.7f seconds %5.7f fps, one-second fps min/avg/max: %.0f %.0f %.0f (%i seconds)\n", frames, time, totalfpsavg, fpsmin, fpsavg, fpsmax, cls.td_onesecondavgcount);
DO_MAX(frames);
DO_MIN(time);
DO_MAX(totalfpsavg);
DO_MAX(fpsmin);
DO_MAX(fpsavg);
DO_MAX(fpsmax);
Con_Printf("MAX: %i frames %5.7f seconds %5.7f fps, one-second fps min/avg/max: %.0f %.0f %.0f (%i seconds)\n", frames, time, totalfpsavg, fpsmin, fpsavg, fpsmax, cls.td_onesecondavgcount);
}
Z_Free(history);
history = NULL;
Host_Quit_f();
}
}
else
Host_Quit_f();
}
}
static void RE_rayobject_blibvh_bb(RayObject *o, float *min, float *max)
{
BVHObject *obj = (BVHObject*)o;
DO_MIN(obj->bb[0], min);
DO_MAX(obj->bb[1], max);
}
/* Object Surface Area Heuristic splitter */
int rtbuild_heuristic_object_split(RTBuilder *b, int nchilds)
{
int size = rtbuild_size(b);
assert(nchilds == 2);
assert(size > 1);
int baxis = -1, boffset = 0;
if (size > nchilds) {
float bcost = FLT_MAX;
baxis = -1, boffset = size / 2;
SweepCost *sweep = (SweepCost *)MEM_mallocN(sizeof(SweepCost) * size, "RTBuilder.HeuristicSweep");
for (int axis = 0; axis < 3; axis++) {
SweepCost sweep_left;
RTBuilder::Object **obj = b->sorted_begin[axis];
// float right_cost = 0;
for (int i = size - 1; i >= 0; i--) {
if (i == size - 1) {
copy_v3_v3(sweep[i].bb, obj[i]->bb);
copy_v3_v3(sweep[i].bb + 3, obj[i]->bb + 3);
sweep[i].cost = obj[i]->cost;
}
else {
sweep[i].bb[0] = min_ff(obj[i]->bb[0], sweep[i + 1].bb[0]);
sweep[i].bb[1] = min_ff(obj[i]->bb[1], sweep[i + 1].bb[1]);
sweep[i].bb[2] = min_ff(obj[i]->bb[2], sweep[i + 1].bb[2]);
sweep[i].bb[3] = max_ff(obj[i]->bb[3], sweep[i + 1].bb[3]);
sweep[i].bb[4] = max_ff(obj[i]->bb[4], sweep[i + 1].bb[4]);
sweep[i].bb[5] = max_ff(obj[i]->bb[5], sweep[i + 1].bb[5]);
sweep[i].cost = obj[i]->cost + sweep[i + 1].cost;
}
// right_cost += obj[i]->cost;
}
sweep_left.bb[0] = obj[0]->bb[0];
sweep_left.bb[1] = obj[0]->bb[1];
sweep_left.bb[2] = obj[0]->bb[2];
sweep_left.bb[3] = obj[0]->bb[3];
sweep_left.bb[4] = obj[0]->bb[4];
sweep_left.bb[5] = obj[0]->bb[5];
sweep_left.cost = obj[0]->cost;
// right_cost -= obj[0]->cost; if (right_cost < 0) right_cost = 0;
for (int i = 1; i < size; i++) {
//Worst case heuristic (cost of each child is linear)
float hcost, left_side, right_side;
// not using log seems to have no impact on raytracing perf, but
// makes tree construction quicker, left out for now to test (brecht)
// left_side = bb_area(sweep_left.bb, sweep_left.bb + 3) * (sweep_left.cost + logf((float)i));
// right_side = bb_area(sweep[i].bb, sweep[i].bb + 3) * (sweep[i].cost + logf((float)size - i));
left_side = bb_area(sweep_left.bb, sweep_left.bb + 3) * (sweep_left.cost);
right_side = bb_area(sweep[i].bb, sweep[i].bb + 3) * (sweep[i].cost);
hcost = left_side + right_side;
assert(left_side >= 0);
assert(right_side >= 0);
if (left_side > bcost) break; //No way we can find a better heuristic in this axis
assert(hcost >= 0);
// this makes sure the tree built is the same whatever is the order of the sorting axis
if (hcost < bcost || (hcost == bcost && axis < baxis)) {
bcost = hcost;
baxis = axis;
boffset = i;
}
DO_MIN(obj[i]->bb, sweep_left.bb);
DO_MAX(obj[i]->bb + 3, sweep_left.bb + 3);
sweep_left.cost += obj[i]->cost;
// right_cost -= obj[i]->cost; if (right_cost < 0) right_cost = 0;
}
//assert(baxis >= 0 && baxis < 3);
if (!(baxis >= 0 && baxis < 3))
baxis = 0;
}
MEM_freeN(sweep);
}
else if (size == 2) {
baxis = 0;
boffset = 1;
}
else if (size == 1) {
b->child_offset[0] = 0;
b->child_offset[1] = 1;
return 1;
}
b->child_offset[0] = 0;
b->child_offset[1] = boffset;
b->child_offset[2] = size;
/* Adjust sorted arrays for childs */
for (int i = 0; i < boffset; i++) b->sorted_begin[baxis][i]->selected = true;
for (int i = boffset; i < size; i++) b->sorted_begin[baxis][i]->selected = false;
for (int i = 0; i < 3; i++)
std::stable_partition(b->sorted_begin[i], b->sorted_end[i], selected_node);
return nchilds;
}
void rtbuild_merge_bb(RTBuilder *b, float min[3], float max[3])
{
rtbuild_calc_bb(b);
DO_MIN(b->bb, min);
DO_MAX(b->bb + 3, max);
}