Merge pull request #19534 from tomoaki0705:fixCudaFiltersRace

pull/19582/head
Alexander Alekhin 4 years ago
commit 0421c3991d
  1. 89
      modules/cudafilters/src/cuda/median_filter.cu

@ -50,9 +50,6 @@
namespace cv { namespace cuda { namespace device
{
// // namespace imgproc
// {
__device__ void histogramAddAndSub8(int* H, const int * hist_colAdd,const int * hist_colSub){
int tx = threadIdx.x;
if (tx<8){
@ -120,6 +117,25 @@ namespace cv { namespace cuda { namespace device
luc[tx]=0;
}
#define scanNeighbor(array, range, index, threadIndex) \
{ \
int v = 0; \
if (index <= threadIndex && threadIndex < range) \
v = array[threadIndex] + array[threadIndex-index]; \
__syncthreads(); \
if (index <= threadIndex && threadIndex < range) \
array[threadIndex] = v; \
}
#define findMedian(array, range, threadIndex, result, count, position) \
if (threadIndex < range) \
{ \
if (array[threadIndex+1] > position && array[threadIndex] <= position) \
{ \
*result = threadIndex+1; \
*count = array[threadIndex]; \
} \
}
__device__ void histogramMedianPar8LookupOnly(int* H,int* Hscan, const int medPos,int* retval, int* countAtMed){
int tx=threadIdx.x;
*retval=*countAtMed=0;
@ -127,28 +143,14 @@ namespace cv { namespace cuda { namespace device
Hscan[tx]=H[tx];
}
__syncthreads();
if (1 <= tx && tx < 8 )
Hscan[tx]+=Hscan[tx-1];
scanNeighbor(Hscan, 8, 1, tx);
__syncthreads();
if (2 <= tx && tx < 8 )
Hscan[tx]+=Hscan[tx-2];
scanNeighbor(Hscan, 8, 2, tx);
__syncthreads();
if (4 <= tx && tx < 8 )
Hscan[tx]+=Hscan[tx-4];
scanNeighbor(Hscan, 8, 4, tx);
__syncthreads();
if(tx<7){
if(Hscan[tx+1] > medPos && Hscan[tx] < medPos){
*retval=tx+1;
*countAtMed=Hscan[tx];
}
else if(Hscan[tx]==medPos){
if(Hscan[tx+1]>medPos){
*retval=tx+1;
*countAtMed=Hscan[tx];
}
}
}
findMedian(Hscan, 7, tx, retval, countAtMed, medPos);
}
__device__ void histogramMedianPar32LookupOnly(int* H,int* Hscan, const int medPos,int* retval, int* countAtMed){
@ -158,33 +160,18 @@ namespace cv { namespace cuda { namespace device
Hscan[tx]=H[tx];
}
__syncthreads();
if ( 1 <= tx && tx < 32 )
Hscan[tx]+=Hscan[tx-1];
scanNeighbor(Hscan, 32, 1, tx);
__syncthreads();
if ( 2 <= tx && tx < 32 )
Hscan[tx]+=Hscan[tx-2];
scanNeighbor(Hscan, 32, 2, tx);
__syncthreads();
if ( 4 <= tx && tx < 32 )
Hscan[tx]+=Hscan[tx-4];
scanNeighbor(Hscan, 32, 4, tx);
__syncthreads();
if ( 8 <= tx && tx < 32 )
Hscan[tx]+=Hscan[tx-8];
scanNeighbor(Hscan, 32, 8, tx);
__syncthreads();
if ( 16 <= tx && tx < 32 )
Hscan[tx]+=Hscan[tx-16];
scanNeighbor(Hscan, 32, 16, tx);
__syncthreads();
if(tx<31){
if(Hscan[tx+1] > medPos && Hscan[tx] < medPos){
*retval=tx+1;
*countAtMed=Hscan[tx];
}
else if(Hscan[tx]==medPos){
if(Hscan[tx+1]>medPos){
*retval=tx+1;
*countAtMed=Hscan[tx];
}
}
}
findMedian(Hscan, 31, tx, retval, countAtMed, medPos);
}
__global__ void cuMedianFilterMultiBlock(PtrStepSzb src, PtrStepSzb dest, PtrStepSzi histPar, PtrStepSzi coarseHistGrid,int r, int medPos_)
@ -283,7 +270,6 @@ namespace cv { namespace cuda { namespace device
__syncthreads();
histogramMultipleAdd8(HCoarse,histCoarse, 2*r+1);
// __syncthreads();
int cols_m_1=cols-1;
for(int j=r;j<cols-r;j++){
@ -295,23 +281,24 @@ namespace cv { namespace cuda { namespace device
histogramMedianPar8LookupOnly(HCoarse,HCoarseScan,medPos, &firstBin,&countAtMed);
__syncthreads();
if ( luc[firstBin] <= (j-r))
int loopIndex = luc[firstBin];
if (loopIndex <= (j-r))
{
histogramClear32(HFine[firstBin]);
for ( luc[firstBin] = j-r; luc[firstBin] < ::min(j+r+1,cols); luc[firstBin]++ ){
histogramAdd32(HFine[firstBin], hist+(luc[firstBin]*256+(firstBin<<5) ) );
for ( loopIndex = j-r; loopIndex < ::min(j+r+1,cols); loopIndex++ ){
histogramAdd32(HFine[firstBin], hist+(loopIndex*256+(firstBin<<5) ) );
}
}
else{
for ( ; luc[firstBin] < (j+r+1);luc[firstBin]++ ) {
for ( ; loopIndex < (j+r+1);loopIndex++ ) {
histogramAddAndSub32(HFine[firstBin],
hist+(::min(luc[firstBin],cols_m_1)*256+(firstBin<<5) ),
hist+(::max(luc[firstBin]-2*r-1,0)*256+(firstBin<<5) ) );
hist+(::min(loopIndex,cols_m_1)*256+(firstBin<<5) ),
hist+(::max(loopIndex-2*r-1,0)*256+(firstBin<<5) ) );
__syncthreads();
}
}
__syncthreads();
luc[firstBin] = loopIndex;
int leftOver=medPos-countAtMed;
if(leftOver>=0){

Loading…
Cancel
Save