mirror/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2024-11-28 13:10:12 +08:00
Code Issues Actions Packages Projects Releases Wiki Activity

speedUp

Browse Source
This commit is contained in:
Konstantin Matskevich 2014-03-13 09:33:15 +04:00
parent 91e4334227
commit af1084eb58
3 changed files with 49 additions and 53 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

94
modules/calib3d/src/opencl/stereobm.cl
View File

@ -51,47 +51,27 @@
#define MAX_VAL 32767 #define MAX_VAL 32767
void calcDisp(__local short * cost, __global short * disp, int uniquenessRatio/*, int textureTreshold, short textsum*/, void calcDisp(__local short * cost, __global short * disp, int uniquenessRatio/*, int textureTreshold, short textsum*/,
int mindisp, int ndisp, int w, __local short * dispbuf, int d, int x, int y, int cols, int rows, int wsz2) int mindisp, int ndisp, int w, __local int * bestDisp, __local int * bestCost, int d, int x, int y, int cols, int rows, int wsz2)
{ {
short FILTERED = (mindisp - 1)<<4; short FILTERED = (mindisp - 1)<<4;
short best_disp, best_cost; int best_disp = *bestDisp, best_cost = *bestCost, best_disp_back = ndisp - best_disp - 1;
dispbuf[d] = d;
barrier(CLK_LOCAL_MEM_FENCE);
for(int lsize = tsize/2 >> 1; lsize > 0; lsize >>= 1)
{
short lid1 = dispbuf[d], lid2 = dispbuf[d+lsize],
cost1 = cost[lid1*w], cost2 = cost[lid2*w];
if (d < lsize)
{
dispbuf[d] = (cost1 < cost2) ? lid1 : (cost1==cost2) ? max(lid1, lid2) : lid2;
}
barrier(CLK_LOCAL_MEM_FENCE);
}
best_disp = ndisp - dispbuf[0] - 1;
best_cost = cost[(ndisp-best_disp-1)*w];
barrier(CLK_LOCAL_MEM_FENCE);
int thresh = best_cost + (best_cost * uniquenessRatio/100); int thresh = best_cost + (best_cost * uniquenessRatio/100);
bool notUniq = ( (cost[0] <= thresh) && (d < (best_disp_back - 1) || d > (best_disp_back + 1) ) );
bool notUniq = ( (cost[d*w] <= thresh) && (d < (ndisp - best_disp - 2) || d > (ndisp - best_disp) ) );
if(notUniq) if(notUniq)
dispbuf[0] = FILTERED; *bestCost = FILTERED;
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
// best_disp = (textsum < textureTreshold) ? FILTERED : best_disp; // best_disp = (textsum < textureTreshold) ? FILTERED : best_disp;
if( dispbuf[0] != FILTERED && x < cols-wsz2-mindisp && y < rows-wsz2) if( *bestCost != FILTERED && x < cols-wsz2-mindisp && y < rows-wsz2 && d == best_disp_back)
{ {
cost = &cost[0] + (ndisp - best_disp - 1)*w; int y3 = (best_disp_back > 0) ? cost[-w] : cost[w],
int y3 = ((ndisp - best_disp - 1) > 0) ? cost[-w] : cost[w],
y2 = cost[0], y2 = cost[0],
y1 = ((ndisp - best_disp - 1) < ndisp-1) ? cost[w] : cost[-w]; y1 = (best_disp_back < ndisp-1) ? cost[w] : cost[-w];
int d = y3+y1-2*y2 + abs(y3-y1); int d = y3+y1-2*y2 + abs(y3-y1);
disp[0] = (short)(((ndisp - best_disp - 1 + mindisp)*256 + (d != 0 ? (y3-y1)*256/d : 0) + 15) >> 4); disp[0] = (short)(((best_disp_back + mindisp)*256 + (d != 0 ? (y3-y1)*256/d : 0) + 15) >> 4);
} }
} }
@ -115,11 +95,11 @@ short calcCostBorder(__global const uchar * leftptr, __global const uchar * righ
int idx = mad24(y+wsz2*(2*nthread-1), cols, x+wsz2*(1-2*nthread)); int idx = mad24(y+wsz2*(2*nthread-1), cols, x+wsz2*(1-2*nthread));
left = leftptr + idx; left = leftptr + idx;
right = rightptr + (idx - d); right = rightptr + (idx - d);
int shift = 1*nthread + cols*(1-nthread);
short costdiff = 0; short costdiff = 0;
for(int i = 0; i < winsize; i++) for(int i = 0; i < winsize; i++)
{ {
int shift = 1*nthread + cols*(1-nthread);
costdiff += abs( left[0] - right[0] ); costdiff += abs( left[0] - right[0] );
left += shift; left += shift;
right += shift; right += shift;
@ -131,7 +111,7 @@ short calcCostBorder(__global const uchar * leftptr, __global const uchar * righ
} }
short calcCostInside(__global const uchar * leftptr, __global const uchar * rightptr, int x, int y, short calcCostInside(__global const uchar * leftptr, __global const uchar * rightptr, int x, int y,
int wsz2, int cols, int d, short cost_up_left, short cost_up, short cost_left) int wsz2, int cols, int d, short cost_up_left, short cost_up, short cost_left, int winsize)
{ {
__global const uchar * left, * right; __global const uchar * left, * right;
int idx = mad24(y-wsz2-1, cols, x-wsz2-1); int idx = mad24(y-wsz2-1, cols, x-wsz2-1);
@ -139,9 +119,9 @@ short calcCostInside(__global const uchar * leftptr, __global const uchar * righ
right = rightptr + (idx - d); right = rightptr + (idx - d);
uchar corrner1 = abs(left[0] - right[0]), uchar corrner1 = abs(left[0] - right[0]),
corrner2 = abs(left[wsz] - right[wsz]), corrner2 = abs(left[winsize] - right[winsize]),
corrner3 = abs(left[(wsz)*cols] - right[(wsz)*cols]), corrner3 = abs(left[(winsize)*cols] - right[(winsize)*cols]),
corrner4 = abs(left[(wsz)*cols + wsz] - right[(wsz)*cols + wsz]); corrner4 = abs(left[(winsize)*cols + winsize] - right[(winsize)*cols + winsize]);
return cost_up + cost_left - cost_up_left + corrner1 - return cost_up + cost_left - cost_up_left + corrner1 -
corrner2 - corrner3 + corrner4; corrner2 - corrner3 + corrner4;
@ -162,9 +142,11 @@ __kernel void stereoBM_opt(__global const uchar * leftptr, __global const uchar
__global short * disp; __global short * disp;
__global const uchar * left, * right; __global const uchar * left, * right;
__local short dispbuf[tsize];
__local short costFunc[csize]; __local short costFunc[csize];
__local short * cost; __local short * cost;
__local int best_disp[2];
__local int best_cost[2];
best_cost[nthread] = MAX_VAL;
short costbuf[wsz]; short costbuf[wsz];
int head = 0; int head = 0;
@ -180,18 +162,19 @@ __kernel void stereoBM_opt(__global const uchar * leftptr, __global const uchar
short tempcost = 0; short tempcost = 0;
if(x < cols-wsz2-mindisp && y < rows-wsz2) if(x < cols-wsz2-mindisp && y < rows-wsz2)
{ {
for(int i = 0; i < wsz; i++) int shift = 1*nthread + cols*(1-nthread);
for(int i = 0; i < winsize; i++)
{ {
int idx = mad24(y-wsz2+i*nthread, cols, x-wsz2+i*(1-nthread)); int idx = mad24(y-wsz2+i*nthread, cols, x-wsz2+i*(1-nthread));
left = leftptr + idx; left = leftptr + idx;
right = rightptr + (idx - d); right = rightptr + (idx - d);
short costdiff = 0; short costdiff = 0;
for(int j = 0; j < wsz; j++) for(int j = 0; j < winsize; j++)
{ {
costdiff += abs( left[0] - right[0] ); costdiff += abs( left[0] - right[0] );
left += 1*nthread + cols*(1-nthread); left += shift;
right += 1*nthread + cols*(1-nthread);// maybe use ? operator right += shift;
} }
if(nthread==1) if(nthread==1)
{ {
@ -203,42 +186,55 @@ __kernel void stereoBM_opt(__global const uchar * leftptr, __global const uchar
} }
if(nthread==1) if(nthread==1)
{ {
cost[0] = tempcost; cost[0] = tempcost;
atomic_min(best_cost+nthread, tempcost);
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if(best_cost[1] == tempcost)
best_disp[1] = ndisp - d - 1;
barrier(CLK_LOCAL_MEM_FENCE);
int dispIdx = mad24(gy, disp_step, disp_offset + gx*(int)sizeof(short)); int dispIdx = mad24(gy, disp_step, disp_offset + gx*(int)sizeof(short));
disp = (__global short *)(dispptr + dispIdx); disp = (__global short *)(dispptr + dispIdx);
calcDisp(&costFunc[sizeY - 1 + lx - ly], disp, uniquenessRatio, //textureTreshold, textsum, calcDisp(cost, disp, uniquenessRatio, //textureTreshold, textsum,
mindisp, ndisp, 2*sizeY, &dispbuf[nthread*tsize/2], d, x, y, cols, rows, wsz2); mindisp, ndisp, 2*sizeY, best_disp + 1, best_cost+1, d, x, y, cols, rows, wsz2);
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
lx = 1 - nthread; lx = 1 - nthread;
ly = nthread; ly = nthread;
for(int i = 0; i < iters; i++) for(int i = 0; i < sizeY*sizeX/2; i++)
{ {
x = (lx < sizeX) ? gx + shiftX + lx : cols; x = (lx < sizeX) ? gx + shiftX + lx : cols;
y = (ly < sizeY) ? gy + shiftY + ly : rows; y = (ly < sizeY) ? gy + shiftY + ly : rows;
best_cost[nthread] = MAX_VAL;
barrier(CLK_LOCAL_MEM_FENCE);
costIdx = calcLocalIdx(lx, ly, d, sizeY); costIdx = calcLocalIdx(lx, ly, d, sizeY);
cost = costFunc + costIdx; cost = costFunc + costIdx;
if(x < cols-wsz2-mindisp && y < rows-wsz2 ) if(x < cols-wsz2-mindisp && y < rows-wsz2 )
{ {
cost[0] = ( ly*(1-nthread) + lx*nthread == 0 ) ? tempcost = ( ly*(1-nthread) + lx*nthread == 0 ) ?
calcCostBorder(leftptr, rightptr, x, y, nthread, wsz2, costbuf, &head, cols, d, calcCostBorder(leftptr, rightptr, x, y, nthread, wsz2, costbuf, &head, cols, d,
costFunc[calcLocalIdx(lx-1*(1-nthread), ly-1*nthread, d, sizeY)], winsize) : cost[2*nthread-1], winsize) :
calcCostInside(leftptr, rightptr, x, y, wsz2, cols, d, calcCostInside(leftptr, rightptr, x, y, wsz2, cols, d,
costFunc[calcLocalIdx(lx-1, ly-1, d, sizeY)], cost[0], cost[1], cost[-1], winsize);
costFunc[calcLocalIdx(lx, ly-1, d, sizeY)], cost[0] = tempcost;
costFunc[calcLocalIdx(lx-1, ly, d, sizeY)]); atomic_min(best_cost + nthread, tempcost);
} }
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
if(best_cost[nthread] == tempcost)
best_disp[nthread] = ndisp - d - 1;
barrier(CLK_LOCAL_MEM_FENCE);
int dispIdx = mad24(gy+ly, disp_step, disp_offset + (gx+lx)*(int)sizeof(short)); int dispIdx = mad24(gy+ly, disp_step, disp_offset + (gx+lx)*(int)sizeof(short));
disp = (__global short *)(dispptr + dispIdx); disp = (__global short *)(dispptr + dispIdx);
calcDisp(&costFunc[sizeY - 1 - ly + lx], disp, uniquenessRatio, //textureTreshold, textsum, calcDisp(cost, disp, uniquenessRatio, //textureTreshold, textsum,
mindisp, ndisp, 2*sizeY, &dispbuf[nthread*tsize/2], d, x, y, cols, rows, wsz2); mindisp, ndisp, 2*sizeY, best_disp + nthread, best_cost + nthread, d, x, y, cols, rows, wsz2);
barrier(CLK_LOCAL_MEM_FENCE); barrier(CLK_LOCAL_MEM_FENCE);
calcNewCoordinates(&lx, &ly, nthread); calcNewCoordinates(&lx, &ly, nthread);

4
modules/calib3d/src/stereobm.cpp
View File

@ -743,9 +743,9 @@ static bool ocl_stereobm_opt( InputArray _left, InputArray _right,
int wsz = state->SADWindowSize; int wsz = state->SADWindowSize;
int wsz2 = wsz/2; int wsz2 = wsz/2;
int sizeX = 13, sizeY = sizeX-1, N = ndisp*2; int sizeX = std::max(11, 27 - ocl::Device::getDefault().maxComputeUnits() ), sizeY = sizeX-1, N = ndisp*2;
ocl::Kernel k("stereoBM_opt", ocl::calib3d::stereobm_oclsrc, cv::format("-D csize=%d -D tsize=%d -D wsz=%d -D iters=%d", (2*sizeY)*ndisp, 2*ndisp, wsz, sizeX*sizeY/2) ); ocl::Kernel k("stereoBM_opt", ocl::calib3d::stereobm_oclsrc, cv::format("-D csize=%d -D wsz=%d", (2*sizeY)*ndisp, wsz) );
if(k.empty()) if(k.empty())
return false; return false;

4
modules/calib3d/test/opencl/test_stereobm.cpp
View File

@ -94,8 +94,8 @@ OCL_TEST_P(StereoBMFixture, StereoBM)
Mat t; absdiff(disp, udisp, t); Mat t; absdiff(disp, udisp, t);
for(int i = 0; i<t.rows; i++) for(int i = 0; i<t.rows; i++)
for(int j = 0; j< t.cols; j++) for(int j = 0; j< t.cols; j++)
if(t.at<short>(i,j) > 0) // if(t.at<short>(i,j) > 0)
// if(i == 5 && j == 68) if(i == 5 && j == 36)
printf("%d %d cv: %d ocl: %d\n", i, j, disp.at<short>(i,j), udisp.getMat(ACCESS_READ).at<short>(i,j) ); printf("%d %d cv: %d ocl: %d\n", i, j, disp.at<short>(i,j), udisp.getMat(ACCESS_READ).at<short>(i,j) );
/* imshow("diff.png", t*100); /* imshow("diff.png", t*100);
imshow("cv.png", disp*100); imshow("cv.png", disp*100);