Skip to content

O
opencv
Commit 0904f10a authored by Konstantin Matskevich's avatar Konstantin Matskevich
Browse files
Options

optimizations

parent bfc843a5
Show whitespace changes
Inline Side-by-side
Showing with 134 additions and 91 deletions
modules/calib3d/perf/opencl/perf_stereobm.cpp
View file @ 0904f10a
......@@ -51,7 +51,7 @@ namespace ocl {
typedef std::tr1::tuple<int, int> StereoBMFixture_t;
typedef TestBaseWithParam<StereoBMFixture_t> StereoBMFixture;
OCL_PERF_TEST_P(StereoBMFixture, StereoBM, ::testing::Combine(OCL_PERF_ENUM(32, 64, 128), OCL_PERF_ENUM(11,21) ) )
OCL_PERF_TEST_P(StereoBMFixture, StereoBM, ::testing::Combine(OCL_PERF_ENUM(32, 64), OCL_PERF_ENUM(11,21) ) )
{
const int n_disp = get<0>(GetParam()), winSize = get<1>(GetParam());
UMat left, right, disp;
......@@ -64,11 +64,11 @@ OCL_PERF_TEST_P(StereoBMFixture, StereoBM, ::testing::Combine(OCL_PERF_ENUM(32,
declare.in(left, right);
Ptr<StereoBM> bm = createStereoBM( n_disp, winSize );
bm->setPreFilterType(bm->PREFILTER_NORMALIZED_RESPONSE);
bm->setPreFilterType(bm->PREFILTER_XSOBEL);
OCL_TEST_CYCLE() bm->compute(left, right, disp);
SANITY_CHECK(disp, 0.05, ERROR_RELATIVE);
SANITY_CHECK(disp, 1e-3, ERROR_RELATIVE);
}
}//ocl
......
modules/calib3d/src/opencl/stereobm.cl
View file @ 0904f10a
......@@ -47,90 +47,119 @@
#ifdef csize
__kernel void stereoBM_opt(__global const uchar * left, __global const uchar * right, __global uchar * dispptr,
#define MAX_VAL 32767
__kernel void stereoBM_opt(__global const uchar * leftptr, __global const uchar * rightptr, __global uchar * dispptr,
int disp_step, int disp_offset, int rows, int cols, int mindisp, int ndisp,
int preFilterCap, int winsize, int textureTreshold, int uniquenessRatio)
int preFilterCap, int nthreads, int textureTreshold, int uniquenessRatio)
{
int total_x = get_global_id(0);
int gx = get_group_id(0), x = gx*ndisp;
int y = get_global_id(1);
int d = get_local_id(0) + mindisp;
int wsz2 = winsize/2;
int x = get_global_id(0);
int total_y = get_global_id(1);
int z = get_local_id(2);
int d = get_local_id(1);
int gy = get_group_id(1), y = gy*ndisp + z*ndisp/nthreads;
int wsz2 = wsz/2;
short FILTERED = (mindisp - 1)<<4;
__local int cost[csize];
int textsum[tsize];
if( total_x<cols && y<rows && d<ndisp)
{
int dispIdx = mad24(y, disp_step, disp_offset + total_x*(int)sizeof(short) );
__local short costFunc[csize];
short textsum[tsize];
__local short * cost = &costFunc[0] + d + ndisp*ndisp/nthreads*z;
__global uchar * left, * right;
int dispIdx = mad24(total_y, disp_step, disp_offset + x*(int)sizeof(short) );
__global short * disp = (__global short*)(dispptr + dispIdx);
disp[0] = FILTERED;
if( (total_x > ndisp-1) && (y > wsz2-1) && (total_x < cols + ndisp - cols%ndisp) && (y < rows - wsz2))
{
for(; (x <= ndisp+mindisp+wsz2-2); x++)
if( x < cols && total_y < rows)
{
cost[(d-mindisp)+ndisp*(x%(gx*ndisp))] = INT_MAX;
textsum[x%(gx*ndisp)] = INT_MAX;
disp[0] = FILTERED;
}
cost[(d-mindisp)+ndisp*(x%(gx*ndisp))] = 0;
textsum[x%(gx*ndisp)] = 0;
for(int i = -wsz2; i < wsz2+1; i++)
for(int j = -wsz2; j < wsz2+1; j++)
short costbuf[wsz];
short textbuf[wsz];
int head = 0;
if( (x > ndisp+mindisp+wsz2-2) && (x < cols - wsz2 - mindisp) )
{
cost[(d-mindisp)+ndisp*(x%(gx*ndisp))] += abs( left[min( y+i, rows-1 ) * cols + min( x+j, cols-1 )]
- right[min( y+i, rows-1 ) * cols + min( x+j-d, cols-1 )] );
textsum[x%(gx*ndisp)] += abs( left[min( y+i, rows-1 ) * cols + min( x+j, cols-1 )] - preFilterCap );
}
x++;
for(; (x < gx*ndisp + ndisp) && (x < cols-wsz2-mindisp); x++)
cost += (y < wsz2) ? ndisp*wsz2 : 0;
y = (y<wsz2) ? wsz2 : y;
cost[0] = 0;
textsum[y-(gy*ndisp)] = 0;
for(int i = -wsz2; (i < wsz2+1) && (y < rows-wsz2); i++)
{
cost[(d-mindisp)+ndisp*(x%(gx*ndisp))] = cost[(d-mindisp)+ndisp*((x-1)%(gx*ndisp))];
textsum[x%(gx*ndisp)] = textsum[(x-1)%(gx*ndisp)];
for(int i = -wsz2; i < wsz2+1; i++)
left = leftptr + mad24(y+i, cols, x-wsz2);
right = rightptr + mad24(y+i, cols, x-wsz2-d-mindisp);
int costdiff = 0, textdiff = 0;
#pragma unroll
for(int j = 0; j < wsz; j++)
{
cost[(d-mindisp)+ndisp*(x%(gx*ndisp))] += -abs( left[min( y+i, rows-1 ) * cols + min( x-wsz2-1, cols-1 )]
- right[min( y+i, rows-1 ) * cols + min( x-wsz2-1-d, cols-1 )] ) +
abs( left[min( y+i, rows-1 ) * cols + min( x+wsz2, cols-1 )]
- right[min( y+i, rows-1 ) * cols + min( x+wsz2-d, cols-1 )] );
textsum[x%(gx*ndisp)] += -abs( left[min( y+i, rows-1 ) * cols + min( x-wsz2-1, cols-1 )] - preFilterCap ) +
abs( left[min( y+i, rows-1 ) * cols + min( x+wsz2, cols-1 )] - preFilterCap );
costdiff += abs( left[0] - right[0] );
textdiff += abs( left[0] - preFilterCap );
left++; right++;
}
cost[0] += costdiff;
textsum[y-(gy*ndisp)] += textdiff;
costbuf[head] = costdiff;
textbuf[head] = textdiff;
head++;
}
y++;
for(; (y < gy*ndisp + ndisp/nthreads + z*ndisp/nthreads) && (y < rows-wsz2); y++)
{
head = head%wsz;
cost += ndisp;
cost[0] = cost[-ndisp];
textsum[y-(gy*ndisp)] = textsum[(y-1)-(gy*ndisp)];
left = leftptr + mad24(y-wsz2-1, cols, x - wsz2);
right = rightptr + mad24(y-wsz2-1, cols, x - wsz2 - d - mindisp);
for(; (x > cols - (cols-1)%ndisp - 1) && (x < cols + ndisp - (cols-1)%ndisp - 1); x++)
int costdiff = 0, textdiff = 0;
#pragma unroll
for(int i = 0; i < wsz; i++)
{
cost[(d-mindisp)+ndisp*(x%(gx*ndisp))] = INT_MAX;
textsum[x%(gx*ndisp)] = INT_MAX;
costdiff +=
abs( left[wsz*cols] - right[wsz*cols] );
textdiff += abs( left[wsz*cols] - preFilterCap );
left++; right++;
}
cost[0] += costdiff - costbuf[head];
textsum[y-(gy*ndisp)] += textdiff - textbuf[head];
costbuf[head] = costdiff;
textbuf[head] = textdiff;
head++;
}
barrier(CLK_LOCAL_MEM_FENCE);
int best_disp = FILTERED, best_cost = INT_MAX-1;
for(int i = 0; (i < ndisp); i++)
cost = &costFunc[0] + d*ndisp;
short best_disp = FILTERED, best_cost = MAX_VAL-1;
#pragma unroll
for(int i = 0; i < tsize; i++)
{
best_cost = (cost[i + ndisp*(d-mindisp)] < best_cost) ? cost[i + ndisp*(d-mindisp)] : best_cost;
best_disp = (best_cost == cost[i + ndisp*(d-mindisp)]) ? i+mindisp : best_disp;
short c = cost[0];
best_cost = (c < best_cost) ? c : best_cost;
best_disp = (best_cost == c) ? ndisp - i - 1 : best_disp;
cost++;
}
cost = &costFunc[0] + d*ndisp;
int thresh = best_cost + (best_cost * uniquenessRatio/100);
for(int i = 0; (i < ndisp) && (uniquenessRatio > 0); i++)
#pragma unroll
for(int i = 0; (i < tsize) && (uniquenessRatio > 0); i++)
{
best_disp = ( (cost[i + ndisp*(d-mindisp)] <= thresh) && (i < best_disp - mindisp - 1 || i > best_disp - mindisp + 1) ) ?
best_disp = ( (cost[0] <= thresh) && (i < (ndisp - best_disp - 2) || i > (ndisp - best_disp) ) ) ?
FILTERED : best_disp;
cost++;
}
disp[0] = textsum[d-mindisp] < textureTreshold ? (FILTERED) : (best_disp == FILTERED) ? (short)(best_disp) : (short)(best_disp);
best_disp = (total_y >= rows-wsz2) || (total_y < wsz2) || (textsum[d] < textureTreshold) ? FILTERED : best_disp;
if( best_disp != FILTERED )
{
int y1 = (best_disp > mindisp) ? cost[(best_disp-mindisp-1) + ndisp*(d-mindisp)] :
cost[(best_disp-mindisp+1) + ndisp*(d-mindisp)],
y2 = cost[(best_disp-mindisp) + ndisp*(d-mindisp)],
y3 = (best_disp < mindisp+ndisp-1) ? cost[(best_disp-mindisp+1) + ndisp*(d-mindisp)] :
cost[(best_disp-mindisp-1) + ndisp*(d-mindisp)];
float a = (y3 - ((best_disp+1)*(y2-y1) + best_disp*y1 - (best_disp-1)*y2)/(best_disp - (best_disp-1)) )/
((best_disp+1)*((best_disp+1) - (best_disp-1) - best_disp) + (best_disp-1)*best_disp);
float b = (y2 - y1)/(best_disp - (best_disp-1)) - a*((best_disp-1)+best_disp);
disp[0] = (y1 == y2 || y3 == y2) ? (short)(best_disp*16) :(short)(-b/(2*a)*16);
cost = &costFunc[0] + (ndisp - best_disp - 1) + ndisp*d;
int y3 = ((ndisp - best_disp - 1) > 0) ? cost[-1] : cost[1],
y2 = cost[0],
y1 = ((ndisp - best_disp - 1) < ndisp-1) ? cost[1] : cost[-1];
d = y3+y1-2*y2 + abs(y3-y1);
if( x < cols && total_y < rows)
{
disp[0] = (short)(((ndisp - best_disp - 1 + mindisp)*256 + (d != 0 ? (y3-y1)*256/d : 0) + 15) >> 4);
}
}
}
......@@ -161,21 +190,20 @@ __kernel void stereoBM_BF(__global const uchar * left, __global const uchar * ri
for(int d = mindisp; d < ndisp+mindisp; d++)
{
cost[d-mindisp] = 0;
cost[(ndisp-1) - (d - mindisp)] = 0;
for(int i = -wsz2; i < wsz2+1; i++)
for(int j = -wsz2; j < wsz2+1; j++)
{
textsum += abs( left[min( y+i, rows-1 ) * cols + min( x+j, cols-1 )] - preFilterCap );
cost[d-mindisp] += abs( left[min( y+i, rows-1 ) * cols + min( x+j, cols-1 )]
- right[min( y+i, rows-1 ) * cols + min( x+j-d, cols-1 )] );
textsum += (d == mindisp) ? abs( left[ (y+i) * cols + x + j] - preFilterCap ) : 0;
cost[(ndisp-1) - (d - mindisp)] += abs(left[(y+i) * cols + x+j] - right[(y+i) * cols + x+j-d] );
}
}
int best_disp = mindisp, best_cost = cost[0];
for(int d = mindisp; d < ndisp+mindisp; d++)
int best_disp = -1, best_cost = INT_MAX;
for(int d = ndisp + mindisp - 1; d > mindisp-1; d--)
{
best_cost = (cost[d-mindisp] < best_cost) ? cost[d-mindisp] : best_cost;
best_disp = (best_cost == cost[d-mindisp]) ? d : best_disp;
best_disp = (best_cost == cost[d-mindisp]) ? (d) : best_disp;
}
int thresh = best_cost + (best_cost * uniquenessRatio/100);
......@@ -191,10 +219,8 @@ __kernel void stereoBM_BF(__global const uchar * left, __global const uchar * ri
int y1 = (best_disp > mindisp) ? cost[best_disp-mindisp-1] : cost[best_disp-mindisp+1],
y2 = cost[best_disp-mindisp],
y3 = (best_disp < mindisp+ndisp-1) ? cost[best_disp-mindisp+1] : cost[best_disp-mindisp-1];
float a = (y3 - ((best_disp+1)*(y2-y1) + best_disp*y1 - (best_disp-1)*y2)/(best_disp - (best_disp-1)) )/
((best_disp+1)*((best_disp+1) - (best_disp-1) - best_disp) + (best_disp-1)*best_disp);
float b = (y2 - y1)/(best_disp - (best_disp-1)) - a*((best_disp-1)+best_disp);
disp[0] = (y1 == y2 || y2 == y3) ? (short)(best_disp*16) : (short)(-b/(2*a)*16);
int _d = y3+y1-2*y2 + abs(y3-y1);
disp[0] = (short)(((ndisp - (best_disp-mindisp) - 1 + mindisp)*256 + (_d != 0 ? (y3-y1)*256/_d : 0) + 15) >> 4);
}
}
}
......@@ -221,10 +247,10 @@ __kernel void prefilter_norm(__global unsigned char *input, __global unsigned ch
int cov2 = 0;
for(int i = -wsz2; i < wsz2+1; i++)
for(int j = -wsz2; j < wsz2+1; j++)
cov2 += input[min( max( (y+i),0 ),rows-1 ) * cols + min( max( (x+j),0 ),cols-1 )];
cov2 += input[clamp(y+i, 0, rows-1) * cols + clamp(x+j, 0, cols-1)];
int res = (cov1*scale_g - cov2*scale_s)>>10;
res = min(min(max(-prefilterCap, res), prefilterCap) + prefilterCap, 255);
res = min(clamp(res, -prefilterCap, prefilterCap) + prefilterCap, 255);
output[y * cols + x] = res & 0xFF;
}
}
......@@ -240,13 +266,13 @@ __kernel void prefilter_xsobel(__global unsigned char *input, __global unsigned
int x = get_global_id(0);
int y = get_global_id(1);
output[y * cols + x] = min(prefilterCap, 255) & 0xFF;
if(x < cols && y < rows-1 && x > 0)
if(x < cols && y < rows && x > 0 && !((y == rows-1)&(rows%2==1) ) )
{
int cov = input[((y > 0) ? y-1 : y+1) * cols + (x-1)] * (-1) + input[((y > 0) ? y-1 : y+1) * cols + ((x<cols-1) ? x+1 : x-1)] * (1) +
int cov = input[ ((y > 0) ? y-1 : y+1) * cols + (x-1)] * (-1) + input[ ((y > 0) ? y-1 : y+1) * cols + ((x<cols-1) ? x+1 : x-1)] * (1) +
input[ (y) * cols + (x-1)] * (-2) + input[ (y) * cols + ((x<cols-1) ? x+1 : x-1)] * (2) +
input[ (y+1) * cols + (x-1)] * (-1) + input[ (y+1) * cols + ((x<cols-1) ? x+1 : x-1)] * (1);
input[((y<rows-1)?(y+1):(y-1))* cols + (x-1)] * (-1) + input[((y<rows-1)?(y+1):(y-1))* cols + ((x<cols-1) ? x+1 : x-1)] * (1);
cov = min(min(max(-prefilterCap, cov), prefilterCap) + prefilterCap, 255);
cov = min(clamp(cov, -prefilterCap, prefilterCap) + prefilterCap, 255);
output[y * cols + x] = cov & 0xFF;
}
}
modules/calib3d/src/stereobm.cpp
View file @ 0904f10a
......@@ -180,13 +180,11 @@ static bool ocl_prefilter_xsobel(InputArray _input, OutputArray _output, int pre
_output.create(input.size(), input.type());
output = _output.getUMat();
size_t blockSize = 1;
size_t globalThreads[3] = { input.cols, input.rows, 1 };
size_t localThreads[3] = { blockSize, blockSize, 1 };
k.args(ocl::KernelArg::PtrReadOnly(input), ocl::KernelArg::PtrWriteOnly(output), input.rows, input.cols, prefilterCap);
return k.run(2, globalThreads, localThreads, false);
return k.run(2, globalThreads, NULL, false);
}
static void
......@@ -655,6 +653,7 @@ findStereoCorrespondenceBM( const Mat& left, const Mat& right,
mind = d;
}
}
tsum += htext[y + wsz2] - htext[y - wsz2 - 1];
if( tsum < textureThreshold )
{
......@@ -738,9 +737,9 @@ struct PrefilterInvoker : public ParallelLoopBody
static bool ocl_stereobm_opt( InputArray _left, InputArray _right,
OutputArray _disp, StereoBMParams* state)
{
{//printf("opt\n");
int ndisp = state->numDisparities;
ocl::Kernel k("stereoBM_opt", ocl::calib3d::stereobm_oclsrc, cv::format("-D csize=%d -D tsize=%d", ndisp*ndisp, ndisp) );
ocl::Kernel k("stereoBM_opt", ocl::calib3d::stereobm_oclsrc, cv::format("-D csize=%d -D tsize=%d -D wsz=%d", ndisp*ndisp, ndisp, state->SADWindowSize) );
if(k.empty())
return false;
......@@ -748,8 +747,9 @@ static bool ocl_stereobm_opt( InputArray _left, InputArray _right,
_disp.create(_left.size(), CV_16S);
UMat disp = _disp.getUMat();
size_t globalThreads[3] = { left.cols, left.rows, 1 };
size_t localThreads[3] = {ndisp, 1, 1};
int nthreads = (ndisp <= 64) ? 2 : 4;
size_t globalThreads[3] = { left.cols, (left.rows - left.rows%ndisp + ndisp), nthreads};
size_t localThreads[3] = {1, ndisp, nthreads};
int idx = 0;
idx = k.set(idx, ocl::KernelArg::PtrReadOnly(left));
......@@ -758,11 +758,11 @@ static bool ocl_stereobm_opt( InputArray _left, InputArray _right,
idx = k.set(idx, state->minDisparity);
idx = k.set(idx, ndisp);
idx = k.set(idx, state->preFilterCap);
idx = k.set(idx, state->SADWindowSize);
idx = k.set(idx, nthreads);
idx = k.set(idx, state->textureThreshold);
idx = k.set(idx, state->uniquenessRatio);
return k.run(2, globalThreads, localThreads, false);
return k.run(3, globalThreads, localThreads, false);
}
static bool ocl_stereobm_bf(InputArray _left, InputArray _right,
......@@ -790,15 +790,16 @@ static bool ocl_stereobm_bf(InputArray _left, InputArray _right,
idx = k.set(idx, state->uniquenessRatio);
return k.run(2, globalThreads, NULL, false);
return false;
}
static bool ocl_stereo(InputArray _left, InputArray _right,
OutputArray _disp, StereoBMParams* state)
{
if(ocl::Device::getDefault().localMemSize() > state->numDisparities * state->numDisparities * sizeof(int) )
if(ocl::Device::getDefault().localMemSize() > state->numDisparities * state->numDisparities * sizeof(short) )
return ocl_stereobm_opt(_left, _right, _disp, state);
else
return ocl_stereobm_bf(_left, _right, _disp, state);
return false;//ocl_stereobm_bf(_left, _right, _disp, state);
}
struct FindStereoCorrespInvoker : public ParallelLoopBody
......@@ -992,7 +993,7 @@ public:
bufSize2 = width*height*(sizeof(Point_<short>) + sizeof(int) + sizeof(uchar));
#if CV_SSE2
bool useShorts = false;//params.preFilterCap <= 31 && params.SADWindowSize <= 21 && checkHardwareSupport(CV_CPU_SSE2);
bool useShorts = params.preFilterCap <= 31 && params.SADWindowSize <= 21 && checkHardwareSupport(CV_CPU_SSE2);
#else
const bool useShorts = false;
#endif
......
modules/calib3d/test/opencl/test_stereobm.cpp
View file @ 0904f10a
......@@ -81,11 +81,27 @@ OCL_TEST_P(StereoBMFixture, StereoBM)
{
Ptr<StereoBM> bm = createStereoBM( n_disp, winSize);
bm->setPreFilterType(bm->PREFILTER_XSOBEL);
// bm->setMinDisparity(15);
long t1 = clock();
OCL_OFF(bm->compute(left, right, disp));
long t2 = clock();
OCL_ON(bm->compute(uleft, uright, udisp));
Near(0.05);
cv::ocl::finish();
long t3 = clock();
std::cout << (double)(t2-t1)/CLOCKS_PER_SEC << " " << (double)(t3-t2)/CLOCKS_PER_SEC << std::endl;
/*
Mat t; absdiff(disp, udisp, t);
/* for(int i = 0; i<t.rows; i++)
for(int j = 0; j< t.cols; j++)
// if(t.at<short>(i,j) > 0)
if(i>=5 && i <=16 && j == 36+15)
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("cv.png", disp*100);
imshow("ocl.png", udisp.getMat(ACCESS_READ)*100);
waitKey(0);*/
Near(1e-3);
}
OCL_INSTANTIATE_TEST_CASE_P(StereoMatcher, StereoBMFixture, testing::Combine(testing::Values(32, 64, 128),
......
Markdown is supported
0% or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment