int EnterTaskManager ()
{
if (task_manager)
{
// no task manager started
return 0;
}
task_manager = new TaskManager();
cout << IM(3) << "task-based parallelization (C++11 threads) using "<< task_manager->GetNumThreads() << " threads" << endl;
#ifdef USE_NUMA
numa_run_on_node (0);
#endif
#ifndef WIN32
// master has maximal priority !
int policy;
struct sched_param param;
pthread_getschedparam(pthread_self(), &policy, ¶m);
param.sched_priority = sched_get_priority_max(policy);
pthread_setschedparam(pthread_self(), policy, ¶m);
#endif // WIN32
task_manager->StartWorkers();
ParallelFor (Range(100), [&] (int i) { ; }); // startup
return task_manager->GetNumThreads();
}
void ComputeBeamDiffusionBSSRDF(Float g, Float eta, BSSRDFTable *t) {
// Choose radius values of the diffusion profile discretization
t->radiusSamples[0] = 0;
t->radiusSamples[1] = 2.5e-3f;
for (int i = 2; i < t->nRadiusSamples; ++i)
t->radiusSamples[i] = t->radiusSamples[i - 1] * 1.2f;
// Choose albedo values of the diffusion profile discretization
for (int i = 0; i < t->nRhoSamples; ++i)
t->rhoSamples[i] =
(1 - std::exp(-8 * i / (Float)(t->nRhoSamples - 1))) /
(1 - std::exp(-8));
ParallelFor([&](int i) {
// Compute the diffusion profile for the _i_th albedo sample
// Compute scattering profile for chosen albedo $\rho$
for (int j = 0; j < t->nRadiusSamples; ++j) {
Float rho = t->rhoSamples[i], r = t->radiusSamples[j];
t->profile[i * t->nRadiusSamples + j] =
2 * Pi * r * (BeamDiffusionSS(rho, 1 - rho, g, eta, r) +
BeamDiffusionMS(rho, 1 - rho, g, eta, r));
}
// Compute effective albedo $\rho_{\roman{eff}}$ and CDF for importance
// sampling
t->rhoEff[i] =
IntegrateCatmullRom(t->nRadiusSamples, t->radiusSamples.get(),
&t->profile[i * t->nRadiusSamples],
&t->profileCDF[i * t->nRadiusSamples]);
}, t->nRhoSamples);
}
// RealisticCamera Method Definitions
RealisticCamera::RealisticCamera(const AnimatedTransform &CameraToWorld,
Float shutterOpen, Float shutterClose,
Float apertureDiameter, Float focusDistance,
bool simpleWeighting, const char *lensFile,
Film *film, const Medium *medium)
: Camera(CameraToWorld, shutterOpen, shutterClose, film, medium),
simpleWeighting(simpleWeighting) {
// Load element data from lens description file
std::vector<Float> lensData;
if (ReadFloatFile(lensFile, &lensData) == false) {
Error("Error reading lens specification file \"%s\".", lensFile);
return;
}
if ((lensData.size() % 4) != 0) {
Error(
"Excess values in lens specification file \"%s\"; "
"must be multiple-of-four values, read %d.",
lensFile, (int)lensData.size());
return;
}
for (int i = 0; i < (int)lensData.size(); i += 4) {
if (lensData[i] == 0) {
if (apertureDiameter > lensData[i + 3]) {
Warning(
"Specified aperture diameter %f is greater than maximum "
"possible %f. Clamping it.",
apertureDiameter, lensData[i + 3]);
} else {
lensData[i + 3] = apertureDiameter;
}
}
elementInterfaces.push_back((LensElementInterface) {
lensData[i] * (Float).001, lensData[i + 1] * (Float).001,
lensData[i + 2], lensData[i + 3] * Float(.001) / Float(2.)
});
}
// Compute lens--film distance for given focus distance
Float fb = FocusBinarySearch(focusDistance);
Info("Binary search focus: %f -> %f\n", fb, FocusDistance(fb));
elementInterfaces.back().thickness = FocusThickLens(focusDistance);
Info("Thick lens focus: %f -> %f\n", elementInterfaces.back().thickness,
FocusDistance(elementInterfaces.back().thickness));
// Compute exit pupil bounds at sampled points on the film
int nSamples = 64;
exitPupilBounds.resize(nSamples);
ParallelFor([&](int i) {
Float r0 = (Float)i / nSamples * film->diagonal / 2;
Float r1 = (Float)(i + 1) / nSamples * film->diagonal / 2;
exitPupilBounds[i] = BoundExitPupil(r0, r1);
}, nSamples);
}
// InfiniteAreaLight Method Definitions
InfiniteAreaLight::InfiniteAreaLight(const Transform &LightToWorld,
const Spectrum &L, int nSamples,
const std::string &texmap)
: Light((int)LightFlags::Infinite, LightToWorld, MediumInterface(),
nSamples) {
// Read texel data from _texmap_ and initialize _Lmap_
Point2i resolution;
std::unique_ptr<RGBSpectrum[]> texels(nullptr);
if (texmap != "") {
texels = ReadImage(texmap, &resolution);
if (texels)
for (int i = 0; i < resolution.x * resolution.y; ++i)
texels[i] *= L.ToRGBSpectrum();
}
if (!texels) {
resolution.x = resolution.y = 1;
texels = std::unique_ptr<RGBSpectrum[]>(new RGBSpectrum[1]);
texels[0] = L.ToRGBSpectrum();
}
Lmap.reset(new MIPMap<RGBSpectrum>(resolution, texels.get()));
// Initialize sampling PDFs for infinite area light
// Compute scalar-valued image _img_ from environment map
int width = 2 * Lmap->Width(), height = 2 * Lmap->Height();
std::unique_ptr<Float[]> img(new Float[width * height]);
float fwidth = 0.5f / std::min(width, height);
ParallelFor(
[&](int64_t v) {
Float vp = (v + .5f) / (Float)height;
Float sinTheta = std::sin(Pi * (v + .5f) / height);
for (int u = 0; u < width; ++u) {
Float up = (u + .5f) / (Float)width;
img[u + v * width] = Lmap->Lookup(Point2f(up, vp), fwidth).y();
img[u + v * width] *= sinTheta;
}
},
height, 32);
// Compute sampling distributions for rows and columns of image
distribution.reset(new Distribution2D(img.get(), width, height));
}
void FMeshElementCollector::ProcessTasks()
{
check(IsInRenderingThread());
check(!ParallelTasks.Num() || bUseAsyncTasks);
if (ParallelTasks.Num())
{
QUICK_SCOPE_CYCLE_COUNTER(STAT_FMeshElementCollector_ProcessTasks);
TArray<TFunction<void()>*, SceneRenderingAllocator>& LocalParallelTasks(ParallelTasks);
ParallelFor(ParallelTasks.Num(),
[&LocalParallelTasks](int32 Index)
{
TFunction<void()>* Func = LocalParallelTasks[Index];
(*Func)();
Func->~TFunction<void()>();
}
);
ParallelTasks.Empty();
}
}
void MLTIntegrator::Render(const Scene &scene) {
ProfilePhase p(Prof::IntegratorRender);
std::unique_ptr<Distribution1D> lightDistr =
ComputeLightPowerDistribution(scene);
// Generate bootstrap samples and compute normalization constant $b$
int nBootstrapSamples = nBootstrap * (maxDepth + 1);
std::vector<Float> bootstrapWeights(nBootstrapSamples, 0);
if (scene.lights.size() > 0) {
ProgressReporter progress(nBootstrap / 256,
"Generating bootstrap paths");
std::vector<MemoryArena> bootstrapThreadArenas(MaxThreadIndex());
int chunkSize = Clamp(nBootstrap / 128, 1, 8192);
ParallelFor([&](int i) {
// Generate _i_th bootstrap sample
MemoryArena &arena = bootstrapThreadArenas[threadIndex];
for (int depth = 0; depth <= maxDepth; ++depth) {
int rngIndex = i * (maxDepth + 1) + depth;
MLTSampler sampler(mutationsPerPixel, rngIndex, sigma,
largeStepProbability, nSampleStreams);
Point2f pRaster;
bootstrapWeights[rngIndex] =
L(scene, arena, lightDistr, sampler, depth, &pRaster).y();
arena.Reset();
}
if ((i + 1 % 256) == 0) progress.Update();
}, nBootstrap, chunkSize);
progress.Done();
}
Distribution1D bootstrap(&bootstrapWeights[0], nBootstrapSamples);
Float b = bootstrap.funcInt * (maxDepth + 1);
// Run _nChains_ Markov chains in parallel
Film &film = *camera->film;
int64_t nTotalMutations =
(int64_t)mutationsPerPixel * (int64_t)film.GetSampleBounds().Area();
if (scene.lights.size() > 0) {
StatTimer timer(&renderingTime);
const int progressFrequency = 32768;
ProgressReporter progress(nTotalMutations / progressFrequency,
"Rendering");
ParallelFor([&](int i) {
int64_t nChainMutations =
std::min((i + 1) * nTotalMutations / nChains, nTotalMutations) -
i * nTotalMutations / nChains;
// Follow {i}th Markov chain for _nChainMutations_
MemoryArena arena;
// Select initial state from the set of bootstrap samples
RNG rng(i);
int bootstrapIndex = bootstrap.SampleDiscrete(rng.UniformFloat());
int depth = bootstrapIndex % (maxDepth + 1);
// Initialize local variables for selected state
MLTSampler sampler(mutationsPerPixel, bootstrapIndex, sigma,
largeStepProbability, nSampleStreams);
Point2f pCurrent;
Spectrum LCurrent =
L(scene, arena, lightDistr, sampler, depth, &pCurrent);
// Run the Markov chain for _nChainMutations_ steps
for (int64_t j = 0; j < nChainMutations; ++j) {
sampler.StartIteration();
Point2f pProposed;
Spectrum LProposed =
L(scene, arena, lightDistr, sampler, depth, &pProposed);
// Compute acceptance probability for proposed sample
Float accept = std::min((Float)1, LProposed.y() / LCurrent.y());
// Splat both current and proposed samples to _film_
if (accept > 0)
film.AddSplat(pProposed,
LProposed * accept / LProposed.y());
film.AddSplat(pCurrent, LCurrent * (1 - accept) / LCurrent.y());
// Accept or reject the proposal
if (rng.UniformFloat() < accept) {
pCurrent = pProposed;
LCurrent = LProposed;
sampler.Accept();
++acceptedMutations;
} else
sampler.Reject();
++totalMutations;
if ((i * nTotalMutations / nChains + j) % progressFrequency ==
0)
progress.Update();
arena.Reset();
}
}, nChains);
progress.Done();
}
// Store final image computed with MLT
camera->film->WriteImage(b / mutationsPerPixel);
}
BVHBuildNode *BVHAccel::HLBVHBuild(
MemoryArena &arena, const std::vector<BVHPrimitiveInfo> &primitiveInfo,
int *totalNodes,
std::vector<std::shared_ptr<Primitive>> &orderedPrims) const {
// Compute bounding box of all primitive centroids
Bounds3f bounds;
for (const BVHPrimitiveInfo &pi : primitiveInfo)
bounds = Union(bounds, pi.centroid);
// Compute Morton indices of primitives
std::vector<MortonPrimitive> mortonPrims(primitiveInfo.size());
ParallelFor([&](int i) {
// Initialize _mortionPrims[i]_ for _i_th primitive
constexpr int mortonBits = 10;
constexpr int mortonScale = 1 << mortonBits;
mortonPrims[i].primitiveIndex = primitiveInfo[i].primitiveNumber;
Vector3f centroidOffset = bounds.Offset(primitiveInfo[i].centroid);
mortonPrims[i].mortonCode = EncodeMorton3(centroidOffset * mortonScale);
}, primitiveInfo.size(), 512);
// Radix sort primitive Morton indices
RadixSort(&mortonPrims);
// Create LBVH treelets at bottom of BVH
// Find intervals of primitives for each treelet
std::vector<LBVHTreelet> treeletsToBuild;
for (int start = 0, end = 1; end <= (int)mortonPrims.size(); ++end) {
uint32_t mask = 0b00111111111111000000000000000000;
if (end == (int)mortonPrims.size() ||
((mortonPrims[start].mortonCode & mask) !=
(mortonPrims[end].mortonCode & mask))) {
// Add entry to _treeletsToBuild_ for this treelet
int nPrimitives = end - start;
int maxBVHNodes = 2 * nPrimitives;
BVHBuildNode *nodes = arena.Alloc<BVHBuildNode>(maxBVHNodes, false);
treeletsToBuild.push_back({start, nPrimitives, nodes});
start = end;
}
}
// Create LBVHs for treelets in parallel
std::atomic<int> atomicTotal(0), orderedPrimsOffset(0);
orderedPrims.resize(primitives.size());
ParallelFor([&](int i) {
// Generate _i_th LBVH treelet
int nodesCreated = 0;
const int firstBitIndex = 29 - 12;
LBVHTreelet &tr = treeletsToBuild[i];
tr.buildNodes =
emitLBVH(tr.buildNodes, primitiveInfo, &mortonPrims[tr.startIndex],
tr.nPrimitives, &nodesCreated, orderedPrims,
&orderedPrimsOffset, firstBitIndex);
atomicTotal += nodesCreated;
}, treeletsToBuild.size());
*totalNodes = atomicTotal;
// Create and return SAH BVH from LBVH treelets
std::vector<BVHBuildNode *> finishedTreelets;
finishedTreelets.reserve(treeletsToBuild.size());
for (LBVHTreelet &treelet : treeletsToBuild)
finishedTreelets.push_back(treelet.buildNodes);
return buildUpperSAH(arena, finishedTreelets, 0, finishedTreelets.size(),
totalNodes);
}
void BDPTIntegrator::Render(const Scene &scene) {
ProfilePhase p(Prof::IntegratorRender);
// Compute _lightDistr_ for sampling lights proportional to power
std::unique_ptr<Distribution1D> lightDistr =
ComputeLightPowerDistribution(scene);
// Partition the image into tiles
Film *film = camera->film;
const Bounds2i sampleBounds = film->GetSampleBounds();
const Vector2i sampleExtent = sampleBounds.Diagonal();
const int tileSize = 16;
const int nXTiles = (sampleExtent.x + tileSize - 1) / tileSize;
const int nYTiles = (sampleExtent.y + tileSize - 1) / tileSize;
ProgressReporter reporter(nXTiles * nYTiles, "Rendering");
// Allocate buffers for debug visualization
const int bufferCount = (1 + maxDepth) * (6 + maxDepth) / 2;
std::vector<std::unique_ptr<Film>> weightFilms(bufferCount);
if (visualizeStrategies || visualizeWeights) {
for (int depth = 0; depth <= maxDepth; ++depth) {
for (int s = 0; s <= depth + 2; ++s) {
int t = depth + 2 - s;
if (t == 0 || (s == 1 && t == 1)) continue;
char filename[32];
snprintf(filename, sizeof(filename),
"bdpt_d%02i_s%02i_t%02i.exr", depth, s, t);
weightFilms[BufferIndex(s, t)] = std::unique_ptr<Film>(new Film(
film->fullResolution,
Bounds2f(Point2f(0, 0), Point2f(1, 1)),
std::unique_ptr<Filter>(CreateBoxFilter(ParamSet())),
film->diagonal * 1000, filename, 1.f));
}
}
}
// Render and write the output image to disk
if (scene.lights.size() > 0) {
StatTimer timer(&renderingTime);
ParallelFor([&](const Point2i tile) {
// Render a single tile using BDPT
MemoryArena arena;
int seed = tile.y * nXTiles + tile.x;
std::unique_ptr<Sampler> tileSampler = sampler->Clone(seed);
int x0 = sampleBounds.pMin.x + tile.x * tileSize;
int x1 = std::min(x0 + tileSize, sampleBounds.pMax.x);
int y0 = sampleBounds.pMin.y + tile.y * tileSize;
int y1 = std::min(y0 + tileSize, sampleBounds.pMax.y);
Bounds2i tileBounds(Point2i(x0, y0), Point2i(x1, y1));
std::unique_ptr<FilmTile> filmTile =
camera->film->GetFilmTile(tileBounds);
for (Point2i pPixel : tileBounds) {
tileSampler->StartPixel(pPixel);
do {
// Generate a single sample using BDPT
Point2f pFilm = (Point2f)pPixel + tileSampler->Get2D();
// Trace the camera and light subpaths
Vertex *cameraVertices = arena.Alloc<Vertex>(maxDepth + 2);
Vertex *lightVertices = arena.Alloc<Vertex>(maxDepth + 1);
int nCamera = GenerateCameraSubpath(
scene, *tileSampler, arena, maxDepth + 2, *camera,
pFilm, cameraVertices);
int nLight = GenerateLightSubpath(
scene, *tileSampler, arena, maxDepth + 1,
cameraVertices[0].time(), *lightDistr, lightVertices);
// Execute all BDPT connection strategies
Spectrum L(0.f);
for (int t = 1; t <= nCamera; ++t) {
for (int s = 0; s <= nLight; ++s) {
int depth = t + s - 2;
if ((s == 1 && t == 1) || depth < 0 ||
depth > maxDepth)
continue;
// Execute the $(s, t)$ connection strategy and
// update _L_
Point2f pFilmNew = pFilm;
Float misWeight = 0.f;
Spectrum Lpath = ConnectBDPT(
scene, lightVertices, cameraVertices, s, t,
*lightDistr, *camera, *tileSampler, &pFilmNew,
&misWeight);
if (visualizeStrategies || visualizeWeights) {
Spectrum value;
if (visualizeStrategies)
value =
misWeight == 0 ? 0 : Lpath / misWeight;
if (visualizeWeights) value = Lpath;
weightFilms[BufferIndex(s, t)]->AddSplat(
pFilmNew, value);
}
if (t != 1)
L += Lpath;
else
film->AddSplat(pFilmNew, Lpath);
}
}
filmTile->AddSample(pFilm, L);
arena.Reset();
} while (tileSampler->StartNextSample());
}
film->MergeFilmTile(std::move(filmTile));
reporter.Update();
}, Point2i(nXTiles, nYTiles));
reporter.Done();
}
film->WriteImage(1.0f / sampler->samplesPerPixel);
// Write buffers for debug visualization
if (visualizeStrategies || visualizeWeights) {
const Float invSampleCount = 1.0f / sampler->samplesPerPixel;
for (size_t i = 0; i < weightFilms.size(); ++i)
if (weightFilms[i]) weightFilms[i]->WriteImage(invSampleCount);
}
}
void MLTIntegrator::Render(const Scene &scene) {
lightDistr =
std::unique_ptr<Distribution1D>(ComputeLightSamplingCDF(scene));
Film &film = *camera->film;
// Generate bootstrap samples and compute $b$
int bootstrapSamples = nBootstrap * (maxDepth + 1);
std::unique_ptr<Float[]> bootstrapWeights(new Float[bootstrapSamples]);
{
ProgressReporter progress(nBootstrap, "Generating bootstrap paths");
ParallelFor([&](int k) {
// Generate a single bootstrap sample
MemoryArena arena;
for (int depth = 0; depth <= maxDepth; ++depth) {
uint32_t uIndex = k * (maxDepth + 1) + depth;
MLTSampler sampler(mutationsPerPixel, uIndex, sigma,
largeStepProb);
Point2f samplePos;
bootstrapWeights[uIndex] =
L(scene, arena, sampler, depth, &samplePos).y();
}
progress.Update();
}, nBootstrap);
progress.Done();
}
Distribution1D bootstrap(bootstrapWeights.get(), bootstrapSamples);
Float b = bootstrap.funcInt * (maxDepth + 1);
// Run _nChains_ Markov Chains in parallel
int64_t nTotalMutations =
mutationsPerPixel * (int64_t)film.GetSampleBounds().Area();
{
StatTimer timer(&renderingTime);
ProgressReporter progress(nTotalMutations / 100, "Rendering");
ParallelFor([&](int k) {
int64_t nChainMutations =
std::min((k + 1) * nTotalMutations / nChains, nTotalMutations) -
k * nTotalMutations / nChains;
MemoryArena arena;
std::unique_ptr<FilmTile> filmTile = film.GetFilmTile(Bounds2i(
film.croppedPixelBounds.pMin, film.croppedPixelBounds.pMin));
// Select initial state from the set of bootstrap samples
RNG rng(PCG32_DEFAULT_STATE, k);
int bootstrapIndex = bootstrap.SampleDiscrete(rng.UniformFloat());
int depth = bootstrapIndex % (maxDepth + 1);
// Initialize local variables for selected state
MLTSampler sampler(mutationsPerPixel, bootstrapIndex, sigma,
largeStepProb);
Point2f currentPos, proposalPos;
Spectrum currentL, proposalL;
currentL = L(scene, arena, sampler, depth, ¤tPos);
// Run the Markov Chain for _nChainMutations_ steps
for (int64_t i = 0; i != nChainMutations; ++i) {
sampler.Begin();
proposalL = L(scene, arena, sampler, depth, &proposalPos);
// Compute the acceptance rate
Float accept = std::min((Float)1, proposalL.y() / currentL.y());
// Splat both current and proposed samples to _FilmTile_
if (accept > 0)
filmTile->AddSplat(proposalPos,
proposalL * accept / proposalL.y());
filmTile->AddSplat(currentPos,
currentL * (1 - accept) / currentL.y());
// Accept or reject the proposal
if (rng.UniformFloat() < accept) {
currentPos = proposalPos;
currentL = proposalL;
sampler.Accept();
++acceptedMutations;
} else {
sampler.Reject();
}
++totalMutations;
if (i % 100 == 0) progress.Update();
}
film.MergeFilmTile(std::move(filmTile));
}, nChains);
progress.Done();
}
film.WriteImage(b / mutationsPerPixel);
}