ICODE-ADC/lammps
4
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge pull request #2613 from lammps/ssao_bugfix

Browse Source 
Fix bug in dump image SSAO depth shading
This commit is contained in:
Axel Kohlmeyer
2021-02-19 18:47:41 -05:00
committed by GitHub
parent 0f23036450 0a355c0194
commit 421eb5df0e
2 changed files with 131 additions and 91 deletions
Download Patch File Download Diff File Expand all files Collapse all files

218
src/image.cpp
View File

@ -113,6 +113,11 @@ Image::Image(LAMMPS *lmp, int nmap_caller) : Pointers(lmp)
backLightColor[2] = 0.9;
random = nullptr;
// MPI_Gatherv vectors
recvcounts = nullptr;
displs = nullptr;
}
/* ---------------------------------------------------------------------- */
@ -134,6 +139,9 @@ Image::~Image()
memory->destroy(rgbcopy);
if (random) delete random;
memory->destroy(recvcounts);
memory->destroy(displs);
}
/* ----------------------------------------------------------------------
@ -334,16 +342,37 @@ void Image::merge()
// extra SSAO enhancement
// bcast full image to all procs
// each works on subset of pixels
// gather result back to proc 0
// MPI_Gather() result back to proc 0
// use Gatherv() if subset of pixels is not the same size on every proc
if (ssao) {
MPI_Bcast(imageBuffer,npixels*3,MPI_BYTE,0,world);
MPI_Bcast(surfaceBuffer,npixels*2,MPI_DOUBLE,0,world);
MPI_Bcast(depthBuffer,npixels,MPI_DOUBLE,0,world);
compute_SSAO();
int pixelPart = height/nprocs * width*3;
MPI_Gather(imageBuffer+me*pixelPart,pixelPart,MPI_BYTE,
rgbcopy,pixelPart,MPI_BYTE,0,world);
int pixelstart = 3 * static_cast<int> (1.0*me/nprocs * npixels);
int pixelstop = 3 * static_cast<int> (1.0*(me+1)/nprocs * npixels);
int mypixels = pixelstop - pixelstart;
if (npixels % nprocs == 0) {
MPI_Gather(imageBuffer+pixelstart,mypixels,MPI_BYTE,
rgbcopy,mypixels,MPI_BYTE,0,world);
} else {
if (recvcounts == nullptr) {
memory->create(recvcounts,nprocs,"image:recvcounts");
memory->create(displs,nprocs,"image:displs");
MPI_Allgather(&mypixels,1,MPI_INT,recvcounts,1,MPI_INT,world);
displs[0] = 0;
for (int i = 1; i < nprocs; i++)
displs[i] = displs[i-1] + recvcounts[i-1];
}
MPI_Gatherv(imageBuffer+pixelstart,mypixels,MPI_BYTE,
rgbcopy,recvcounts,displs,MPI_BYTE,0,world);
}
writeBuffer = rgbcopy;
} else {
writeBuffer = imageBuffer;
@ -880,110 +909,117 @@ void Image::compute_SSAO()
-tanPerPixel / zoom;
int pixelRadius = (int) trunc (SSAORadius / pixelWidth + 0.5);
int x,y,s;
int hPart = height / nprocs;
int index = me * hPart * width;
for (y = me * hPart; y < (me + 1) * hPart; y ++) {
for (x = 0; x < width; x ++, index ++) {
double cdepth = depthBuffer[index];
if (cdepth < 0) { continue; }
// each proc is assigned a subset of contiguous pixels from the full image
// pixels are contiguous in x (columns within a row), then by row
// index = pixels from 0 to npixel-1
// x = column # from 0 to width-1
// y = row # from 0 to height-1
double sx = surfaceBuffer[index * 2 + 0];
double sy = surfaceBuffer[index * 2 + 1];
double sin_t = -sqrt(sx*sx + sy*sy);
int pixelstart = static_cast<int> (1.0*me/nprocs * npixels);
int pixelstop = static_cast<int> (1.0*(me+1)/nprocs * npixels);
double mytheta = random->uniform() * SSAOJitter;
double ao = 0.0;
for (int index = pixelstart; index < pixelstop; index++) {
int x = index % width;
int y = index / width;
for (s = 0; s < SSAOSamples; s ++) {
double hx = cos(mytheta);
double hy = sin(mytheta);
mytheta += delTheta;
double cdepth = depthBuffer[index];
if (cdepth < 0) { continue; }
// multiply by z cross surface tangent
// so that dot (aka cos) works here
double sx = surfaceBuffer[index * 2 + 0];
double sy = surfaceBuffer[index * 2 + 1];
double sin_t = -sqrt(sx*sx + sy*sy);
double scaled_sin_t = sin_t * (hx*sy + hy*sx);
double mytheta = random->uniform() * SSAOJitter;
double ao = 0.0;
// Bresenham's line algorithm to march over depthBuffer
for (int s = 0; s < SSAOSamples; s ++) {
double hx = cos(mytheta);
double hy = sin(mytheta);
mytheta += delTheta;
int dx = static_cast<int> (hx * pixelRadius);
int dy = static_cast<int> (hy * pixelRadius);
int ex = x + dx;
if (ex < 0) { ex = 0; } if (ex >= width) { ex = width - 1; }
int ey = y + dy;
if (ey < 0) { ey = 0; } if (ey >= height) { ey = height - 1; }
double delta;
int small, large;
double lenIncr;
if (fabs(hx) > fabs(hy)) {
small = (hx > 0) ? 1 : -1;
large = (hy > 0) ? width : -width;
delta = fabs(hy / hx);
} else {
small = (hy > 0) ? width : -width;
large = (hx > 0) ? 1 : -1;
delta = fabs(hx / hy);
// multiply by z cross surface tangent
// so that dot (aka cos) works here
double scaled_sin_t = sin_t * (hx*sy + hy*sx);
// Bresenham's line algorithm to march over depthBuffer
int dx = static_cast<int> (hx * pixelRadius);
int dy = static_cast<int> (hy * pixelRadius);
int ex = x + dx;
if (ex < 0) { ex = 0; } if (ex >= width) { ex = width - 1; }
int ey = y + dy;
if (ey < 0) { ey = 0; } if (ey >= height) { ey = height - 1; }
double delta;
int small, large;
double lenIncr;
if (fabs(hx) > fabs(hy)) {
small = (hx > 0) ? 1 : -1;
large = (hy > 0) ? width : -width;
delta = fabs(hy / hx);
} else {
small = (hy > 0) ? width : -width;
large = (hx > 0) ? 1 : -1;
delta = fabs(hx / hy);
}
lenIncr = sqrt (1 + delta * delta) * pixelWidth;
// initialize with one step
// because the center point doesn't need testing
int end = ex + ey * width;
int ind = index + small;
double len = lenIncr;
double err = delta;
if (err >= 1.0) {
ind += large;
err -= 1.0;
}
double minPeak = -1;
double peakLen = 0.0;
int stepsTaken = 1;
while ((small > 0 && ind <= end) || (small < 0 && ind >= end)) {
if (ind < 0 || ind >= (width*height)) {
break;
}
lenIncr = sqrt (1 + delta * delta) * pixelWidth;
// initialize with one step
// because the center point doesn't need testing
// cdepth - depthBuffer B/C we want it in the negative z direction
int end = ex + ey * width;
int ind = index + small;
double len = lenIncr;
double err = delta;
if (minPeak < 0 || (depthBuffer[ind] >= 0 &&
depthBuffer[ind] < minPeak)) {
minPeak = depthBuffer[ind];
peakLen = len;
}
ind += small;
len += lenIncr;
err += delta;
if (err >= 1.0) {
ind += large;
err -= 1.0;
}
double minPeak = -1;
double peakLen = 0.0;
int stepsTaken = 1;
while ((small > 0 && ind <= end) || (small < 0 && ind >= end)) {
if (ind < 0 || ind >= (width*height)) {
break;
}
// cdepth - depthBuffer B/C we want it in the negative z direction
if (minPeak < 0 || (depthBuffer[ind] >= 0 &&
depthBuffer[ind] < minPeak)) {
minPeak = depthBuffer[ind];
peakLen = len;
}
ind += small;
len += lenIncr;
err += delta;
if (err >= 1.0) {
ind += large;
err -= 1.0;
}
stepsTaken ++;
}
if (peakLen > 0) {
double h = atan ((cdepth - minPeak) / peakLen);
ao += saturate(sin (h) - scaled_sin_t);
} else {
ao += saturate(-scaled_sin_t);
}
stepsTaken ++;
}
ao /= (double)SSAOSamples;
double c[3];
c[0] = (double) (*(unsigned char *) &imageBuffer[index * 3 + 0]);
c[1] = (double) (*(unsigned char *) &imageBuffer[index * 3 + 1]);
c[2] = (double) (*(unsigned char *) &imageBuffer[index * 3 + 2]);
c[0] *= (1.0 - ao);
c[1] *= (1.0 - ao);
c[2] *= (1.0 - ao);
imageBuffer[index * 3 + 0] = (int) c[0];
imageBuffer[index * 3 + 1] = (int) c[1];
imageBuffer[index * 3 + 2] = (int) c[2];
if (peakLen > 0) {
double h = atan ((cdepth - minPeak) / peakLen);
ao += saturate(sin (h) - scaled_sin_t);
} else {
ao += saturate(-scaled_sin_t);
}
}
ao /= (double)SSAOSamples;
double c[3];
c[0] = (double) (*(unsigned char *) &imageBuffer[index * 3 + 0]);
c[1] = (double) (*(unsigned char *) &imageBuffer[index * 3 + 1]);
c[2] = (double) (*(unsigned char *) &imageBuffer[index * 3 + 2]);
c[0] *= (1.0 - ao);
c[1] *= (1.0 - ao);
c[2] *= (1.0 - ao);
imageBuffer[index * 3 + 0] = (int) c[0];
imageBuffer[index * 3 + 1] = (int) c[1];
imageBuffer[index * 3 + 2] = (int) c[2];
}
}

4
src/image.h
View File

@ -73,6 +73,10 @@ class Image : protected Pointers {
double *depthcopy,*surfacecopy;
unsigned char *imageBuffer,*rgbcopy,*writeBuffer;
// MPI_Gatherv
int *recvcounts,*displs;
// constant view params
double FOV;