Fix ocl::bruteforcematcher crash on Intel OCL

This commit is contained in:
yao 2013-02-28 14:31:08 +08:00
parent 620c699456
commit 504008dbe0
2 changed files with 730 additions and 679 deletions

View File

@ -51,7 +51,6 @@ using namespace cv;
using namespace cv::ocl;
using namespace std;
using namespace std;
namespace cv
{
namespace ocl
@ -62,7 +61,7 @@ namespace cv
}
template < int BLOCK_SIZE, int MAX_DESC_LEN, typename T/*, typename Mask*/ >
void matchUnrolledCached(const oclMat &query, const oclMat &train, const oclMat &mask,
void matchUnrolledCached(const oclMat &query, const oclMat &train, const oclMat &/*mask*/,
const oclMat &trainIdx, const oclMat &distance, int distType)
{
cv::ocl::Context *ctx = query.clCxt;
@ -77,7 +76,7 @@ void matchUnrolledCached(const oclMat &query, const oclMat &train, const oclMat
{
args.push_back( make_pair( sizeof(cl_mem), (void *)&query.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&train.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
//args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&trainIdx.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&distance.data ));
args.push_back( make_pair( smemSize, (void *)NULL));
@ -103,7 +102,7 @@ void matchUnrolledCached(const oclMat /*query*/, const oclMat * /*trains*/, int
}
template < int BLOCK_SIZE, typename T/*, typename Mask*/ >
void match(const oclMat &query, const oclMat &train, const oclMat &mask,
void match(const oclMat &query, const oclMat &train, const oclMat &/*mask*/,
const oclMat &trainIdx, const oclMat &distance, int distType)
{
cv::ocl::Context *ctx = query.clCxt;
@ -117,7 +116,7 @@ void match(const oclMat &query, const oclMat &train, const oclMat &mask,
{
args.push_back( make_pair( sizeof(cl_mem), (void *)&query.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&train.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
//args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&trainIdx.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&distance.data ));
args.push_back( make_pair( smemSize, (void *)NULL));
@ -143,7 +142,7 @@ void match(const oclMat /*query*/, const oclMat * /*trains*/, int /*n*/, const o
//radius_matchUnrolledCached
template < int BLOCK_SIZE, int MAX_DESC_LEN, typename T/*, typename Mask*/ >
void matchUnrolledCached(const oclMat &query, const oclMat &train, float maxDistance, const oclMat &mask,
void matchUnrolledCached(const oclMat &query, const oclMat &train, float maxDistance, const oclMat &/*mask*/,
const oclMat &trainIdx, const oclMat &distance, const oclMat &nMatches, int distType)
{
cv::ocl::Context *ctx = query.clCxt;
@ -159,7 +158,7 @@ void matchUnrolledCached(const oclMat &query, const oclMat &train, float maxDist
args.push_back( make_pair( sizeof(cl_mem), (void *)&query.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&train.data ));
args.push_back( make_pair( sizeof(cl_float), (void *)&maxDistance ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
//args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&trainIdx.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&distance.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&nMatches.data ));
@ -183,7 +182,7 @@ void matchUnrolledCached(const oclMat &query, const oclMat &train, float maxDist
//radius_match
template < int BLOCK_SIZE, typename T/*, typename Mask*/ >
void radius_match(const oclMat &query, const oclMat &train, float maxDistance, const oclMat &mask,
void radius_match(const oclMat &query, const oclMat &train, float maxDistance, const oclMat &/*mask*/,
const oclMat &trainIdx, const oclMat &distance, const oclMat &nMatches, int distType)
{
cv::ocl::Context *ctx = query.clCxt;
@ -198,7 +197,7 @@ void radius_match(const oclMat &query, const oclMat &train, float maxDistance, c
args.push_back( make_pair( sizeof(cl_mem), (void *)&query.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&train.data ));
args.push_back( make_pair( sizeof(cl_float), (void *)&maxDistance ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
//args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&trainIdx.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&distance.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&nMatches.data ));
@ -472,7 +471,7 @@ void matchDispatcher(const oclMat &query, const oclMat &train, int n, float maxD
//knn match Dispatcher
template < int BLOCK_SIZE, int MAX_DESC_LEN, typename T/*, typename Mask*/ >
void knn_matchUnrolledCached(const oclMat &query, const oclMat &train, const oclMat &mask,
void knn_matchUnrolledCached(const oclMat &query, const oclMat &train, const oclMat &/*mask*/,
const oclMat &trainIdx, const oclMat &distance, int distType)
{
cv::ocl::Context *ctx = query.clCxt;
@ -487,7 +486,7 @@ void knn_matchUnrolledCached(const oclMat &query, const oclMat &train, const ocl
{
args.push_back( make_pair( sizeof(cl_mem), (void *)&query.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&train.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
//args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&trainIdx.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&distance.data ));
args.push_back( make_pair( smemSize, (void *)NULL));
@ -507,7 +506,7 @@ void knn_matchUnrolledCached(const oclMat &query, const oclMat &train, const ocl
}
template < int BLOCK_SIZE, typename T/*, typename Mask*/ >
void knn_match(const oclMat &query, const oclMat &train, const oclMat &mask,
void knn_match(const oclMat &query, const oclMat &train, const oclMat &/*mask*/,
const oclMat &trainIdx, const oclMat &distance, int distType)
{
cv::ocl::Context *ctx = query.clCxt;
@ -521,7 +520,7 @@ void knn_match(const oclMat &query, const oclMat &train, const oclMat &mask,
{
args.push_back( make_pair( sizeof(cl_mem), (void *)&query.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&train.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
//args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&trainIdx.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&distance.data ));
args.push_back( make_pair( smemSize, (void *)NULL));
@ -540,7 +539,7 @@ void knn_match(const oclMat &query, const oclMat &train, const oclMat &mask,
}
template < int BLOCK_SIZE, int MAX_DESC_LEN, typename T/*, typename Mask*/ >
void calcDistanceUnrolled(const oclMat &query, const oclMat &train, const oclMat &mask, const oclMat &allDist, int distType)
void calcDistanceUnrolled(const oclMat &query, const oclMat &train, const oclMat &/*mask*/, const oclMat &allDist, int distType)
{
cv::ocl::Context *ctx = query.clCxt;
size_t globalSize[] = {(query.rows + BLOCK_SIZE - 1) / BLOCK_SIZE * BLOCK_SIZE, BLOCK_SIZE, 1};
@ -554,7 +553,7 @@ void calcDistanceUnrolled(const oclMat &query, const oclMat &train, const oclMat
{
args.push_back( make_pair( sizeof(cl_mem), (void *)&query.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&train.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
//args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&allDist.data ));
args.push_back( make_pair( smemSize, (void *)NULL));
args.push_back( make_pair( sizeof(cl_int), (void *)&block_size ));
@ -573,7 +572,7 @@ void calcDistanceUnrolled(const oclMat &query, const oclMat &train, const oclMat
}
template < int BLOCK_SIZE, typename T/*, typename Mask*/ >
void calcDistance(const oclMat &query, const oclMat &train, const oclMat &mask, const oclMat &allDist, int distType)
void calcDistance(const oclMat &query, const oclMat &train, const oclMat &/*mask*/, const oclMat &allDist, int distType)
{
cv::ocl::Context *ctx = query.clCxt;
size_t globalSize[] = {(query.rows + BLOCK_SIZE - 1) / BLOCK_SIZE * BLOCK_SIZE, BLOCK_SIZE, 1};
@ -586,7 +585,7 @@ void calcDistance(const oclMat &query, const oclMat &train, const oclMat &mask,
{
args.push_back( make_pair( sizeof(cl_mem), (void *)&query.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&train.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
//args.push_back( make_pair( sizeof(cl_mem), (void *)&mask.data ));
args.push_back( make_pair( sizeof(cl_mem), (void *)&allDist.data ));
args.push_back( make_pair( smemSize, (void *)NULL));
args.push_back( make_pair( sizeof(cl_int), (void *)&block_size ));
@ -691,7 +690,7 @@ void findKnnMatch(int k, const oclMat &trainIdx, const oclMat &distance, const o
}
}
static void findKnnMatchDispatcher(int k, const oclMat &trainIdx, const oclMat &distance, const oclMat &allDist, int distType)
void findKnnMatchDispatcher(int k, const oclMat &trainIdx, const oclMat &distance, const oclMat &allDist, int distType)
{
findKnnMatch<256>(k, trainIdx, distance, allDist, distType);
}
@ -1007,6 +1006,7 @@ void cv::ocl::BruteForceMatcher_OCL_base::matchConvert(const Mat &trainIdx, cons
void cv::ocl::BruteForceMatcher_OCL_base::match(const oclMat &query, const oclMat &train, vector<DMatch> &matches, const oclMat &mask)
{
assert(mask.empty()); // mask is not supported at the moment
oclMat trainIdx, distance;
matchSingle(query, train, trainIdx, distance, mask);
matchDownload(trainIdx, distance, matches);
@ -1697,3 +1697,5 @@ void cv::ocl::BruteForceMatcher_OCL_base::radiusMatch(const oclMat &query, vecto
radiusMatchCollection(query, trainIdx, imgIdx, distance, nMatches, maxDistance, masks);
radiusMatchDownload(trainIdx, imgIdx, distance, nMatches, matches, compactResult);
}

View File

@ -5,11 +5,13 @@ int bit1Count(float x)
{
int c = 0;
int ix = (int)x;
for (int i = 0 ; i < 32 ; i++)
{
c += ix & 0x1;
ix >>= 1;
}
return (float)c;
}
/* 2dim launch, global size: dim0 is (query rows + block_size - 1) / block_size * block_size, dim1 is block_size
@ -18,7 +20,7 @@ local size: dim0 is block_size, dim1 is block_size.
__kernel void BruteForceMatch_UnrollMatch(
__global float *query,
__global float *train,
__global float *mask,
//__global float *mask,
__global int *bestTrainIdx,
__global float *bestDistance,
__local float *sharebuffer,
@ -40,6 +42,7 @@ __kernel void BruteForceMatch_UnrollMatch(
__local float *s_train = sharebuffer + block_size * max_desc_len;
int queryIdx = groupidx * block_size + lidy;
// load the query into local memory.
for (int i = 0 ; i < max_desc_len / block_size; i ++)
{
@ -52,9 +55,11 @@ __kernel void BruteForceMatch_UnrollMatch(
// loopUnrolledCached to find the best trainIdx and best distance.
volatile int imgIdx = 0;
for (int t = 0 ; t < (train_rows + block_size - 1) / block_size ; t++)
{
float result = 0;
for (int i = 0 ; i < max_desc_len / block_size ; i++)
{
//load a block_size * block_size block into local train.
@ -70,24 +75,30 @@ __kernel void BruteForceMatch_UnrollMatch(
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; j++)
{
result += fabs(s_query[lidy * max_desc_len + i * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; j++)
{
float qr = s_query[lidy * max_desc_len + i * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; j++)
{
//result += popcount((uint)s_query[lidy * max_desc_len + i * block_size + j] ^ (uint)s_train[j * block_size + lidx]);
result += bit1Count((uint)s_query[lidy * max_desc_len + i * block_size + j] ^(uint)s_train[j * block_size + lidx]);
}
break;
}
@ -136,7 +147,7 @@ __kernel void BruteForceMatch_UnrollMatch(
__kernel void BruteForceMatch_Match(
__global float *query,
__global float *train,
__global float *mask,
//__global float *mask,
__global int *bestTrainIdx,
__global float *bestDistance,
__local float *sharebuffer,
@ -166,6 +177,7 @@ __kernel void BruteForceMatch_Match(
{
//Dist dist;
float result = 0;
for (int i = 0 ; i < (query_cols + block_size - 1) / block_size ; i++)
{
const int loadx = lidx + i * block_size;
@ -187,24 +199,30 @@ __kernel void BruteForceMatch_Match(
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; j++)
{
result += fabs(s_query[lidy * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; j++)
{
float qr = s_query[lidy * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; j++)
{
//result += popcount((uint)s_query[lidy * block_size + j] ^ (uint)s_train[j * block_size + lidx]);
result += bit1Count((uint)s_query[lidy * block_size + j] ^(uint)s_train[(uint)j * block_size + lidx]);
}
break;
}
@ -256,7 +274,7 @@ __kernel void BruteForceMatch_RadiusUnrollMatch(
__global float *query,
__global float *train,
float maxDistance,
__global float *mask,
//__global float *mask,
__global int *bestTrainIdx,
__global float *bestDistance,
__global int *nMatches,
@ -285,6 +303,7 @@ __kernel void BruteForceMatch_RadiusUnrollMatch(
__local float *s_train = sharebuffer + block_size * block_size;
float result = 0;
for (int i = 0 ; i < max_desc_len / block_size ; ++i)
{
//load a block_size * block_size block into local train.
@ -302,23 +321,29 @@ __kernel void BruteForceMatch_RadiusUnrollMatch(
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; ++j)
{
result += fabs(s_query[lidy * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; ++j)
{
float qr = s_query[lidy * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; ++j)
{
result += bit1Count((uint)s_query[lidy * block_size + j] ^(uint)s_train[j * block_size + lidx]);
}
break;
}
@ -343,7 +368,7 @@ __kernel void BruteForceMatch_RadiusMatch(
__global float *query,
__global float *train,
float maxDistance,
__global float *mask,
//__global float *mask,
__global int *bestTrainIdx,
__global float *bestDistance,
__global int *nMatches,
@ -371,6 +396,7 @@ __kernel void BruteForceMatch_RadiusMatch(
__local float *s_train = sharebuffer + block_size * block_size;
float result = 0;
for (int i = 0 ; i < (query_cols + block_size - 1) / block_size ; ++i)
{
//load a block_size * block_size block into local train.
@ -388,23 +414,29 @@ __kernel void BruteForceMatch_RadiusMatch(
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; ++j)
{
result += fabs(s_query[lidy * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; ++j)
{
float qr = s_query[lidy * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; ++j)
{
result += bit1Count((uint)s_query[lidy * block_size + j] ^(uint)s_train[j * block_size + lidx]);
}
break;
}
@ -428,7 +460,7 @@ __kernel void BruteForceMatch_RadiusMatch(
__kernel void BruteForceMatch_knnUnrollMatch(
__global float *query,
__global float *train,
__global float *mask,
//__global float *mask,
__global int2 *bestTrainIdx,
__global float2 *bestDistance,
__local float *sharebuffer,
@ -464,9 +496,11 @@ __kernel void BruteForceMatch_knnUnrollMatch(
//loopUnrolledCached
volatile int imgIdx = 0;
for (int t = 0 ; t < (train_rows + block_size - 1) / block_size ; t++)
{
float result = 0;
for (int i = 0 ; i < max_desc_len / block_size ; i++)
{
const int loadX = lidx + i * block_size;
@ -483,24 +517,30 @@ __kernel void BruteForceMatch_knnUnrollMatch(
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; j++)
{
result += fabs(s_query[lidy * max_desc_len + i * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; j++)
{
float qr = s_query[lidy * max_desc_len + i * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; j++)
{
//result += popcount((uint)s_query[lidy * max_desc_len + i * block_size + j] ^ (uint)s_train[j * block_size + lidx]);
result += bit1Count((uint)s_query[lidy * max_desc_len + i * block_size + j] ^(uint)s_train[j * block_size + lidx]);
}
break;
}
@ -549,6 +589,7 @@ __kernel void BruteForceMatch_knnUnrollMatch(
for (int i = 0 ; i < block_size ; i++)
{
float val = s_distance[i];
if (val < bestDistance1)
{
bestDistance2 = bestDistance1;
@ -602,7 +643,7 @@ __kernel void BruteForceMatch_knnUnrollMatch(
__kernel void BruteForceMatch_knnMatch(
__global float *query,
__global float *train,
__global float *mask,
//__global float *mask,
__global int2 *bestTrainIdx,
__global float2 *bestDistance,
__local float *sharebuffer,
@ -632,6 +673,7 @@ __kernel void BruteForceMatch_knnMatch(
for (int t = 0 ; t < (train_rows + block_size - 1) / block_size ; t++)
{
float result = 0.0f;
for (int i = 0 ; i < (query_cols + block_size - 1) / block_size ; i++)
{
const int loadx = lidx + i * block_size;
@ -653,24 +695,30 @@ __kernel void BruteForceMatch_knnMatch(
switch (distType)
{
case 0:
for (int j = 0 ; j < block_size ; j++)
{
result += fabs(s_query[lidy * block_size + j] - s_train[j * block_size + lidx]);
}
break;
case 1:
for (int j = 0 ; j < block_size ; j++)
{
float qr = s_query[lidy * block_size + j] - s_train[j * block_size + lidx];
result += qr * qr;
}
break;
case 2:
for (int j = 0 ; j < block_size ; j++)
{
//result += popcount((uint)s_query[lidy * block_size + j] ^ (uint)s_train[j * block_size + lidx]);
result += bit1Count((uint)s_query[lidy * block_size + j] ^(uint)s_train[(uint)j * block_size + lidx]);
}
break;
}
@ -719,6 +767,7 @@ __kernel void BruteForceMatch_knnMatch(
for (int i = 0 ; i < block_size ; i++)
{
float val = s_distance[i];
if (val < bestDistance1)
{
bestDistance2 = bestDistance1;
@ -772,7 +821,7 @@ __kernel void BruteForceMatch_knnMatch(
kernel void BruteForceMatch_calcDistanceUnrolled(
__global float *query,
__global float *train,
__global float *mask,
//__global float *mask,
__global float *allDist,
__local float *sharebuffer,
int block_size,
@ -790,7 +839,7 @@ kernel void BruteForceMatch_calcDistanceUnrolled(
kernel void BruteForceMatch_calcDistance(
__global float *query,
__global float *train,
__global float *mask,
//__global float *mask,
__global float *allDist,
__local float *sharebuffer,
int block_size,