generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Fixed gpu::matchTemplate for correct handling of big templates. Added tests

Browse Source
This commit is contained in:
Alexey Spizhevoy
2012-03-26 09:19:33 +00:00
parent 098fc1a62e
commit 5434a9a5ec
4 changed files with 61 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
modules/gpu/src/cuda/match_template.cu
View File

@@ -216,7 +216,7 @@ namespace cv { namespace gpu { namespace device
// Prepared_SQDIFF
template <int cn>
__global__ void matchTemplatePreparedKernel_SQDIFF_8U(int w, int h, const PtrStep<unsigned long long> image_sqsum, unsigned int templ_sqsum, DevMem2Df result)
__global__ void matchTemplatePreparedKernel_SQDIFF_8U(int w, int h, const PtrStep<unsigned long long> image_sqsum, unsigned long long templ_sqsum, DevMem2Df result)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
@@ -232,7 +232,7 @@ namespace cv { namespace gpu { namespace device
}
template <int cn>
void matchTemplatePrepared_SQDIFF_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned int templ_sqsum, DevMem2Df result, cudaStream_t stream)
void matchTemplatePrepared_SQDIFF_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned long long templ_sqsum, DevMem2Df result, cudaStream_t stream)
{
const dim3 threads(32, 8);
const dim3 grid(divUp(result.cols, threads.x), divUp(result.rows, threads.y));
@@ -244,10 +244,10 @@ namespace cv { namespace gpu { namespace device
cudaSafeCall( cudaDeviceSynchronize() );
}
void matchTemplatePrepared_SQDIFF_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned int templ_sqsum, DevMem2Df result, int cn,
void matchTemplatePrepared_SQDIFF_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned long long templ_sqsum, DevMem2Df result, int cn,
cudaStream_t stream)
{
typedef void (*caller_t)(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned int templ_sqsum, DevMem2Df result, cudaStream_t stream);
typedef void (*caller_t)(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned long long templ_sqsum, DevMem2Df result, cudaStream_t stream);
static const caller_t callers[] =
{
@@ -284,7 +284,9 @@ namespace cv { namespace gpu { namespace device
template <int cn>
__global__ void matchTemplatePreparedKernel_SQDIFF_NORMED_8U(int w, int h, const PtrStep<unsigned long long> image_sqsum, unsigned int templ_sqsum, DevMem2Df result)
__global__ void matchTemplatePreparedKernel_SQDIFF_NORMED_8U(
int w, int h, const PtrStep<unsigned long long> image_sqsum,
unsigned long long templ_sqsum, DevMem2Df result)
{
const int x = blockIdx.x * blockDim.x + threadIdx.x;
const int y = blockIdx.y * blockDim.y + threadIdx.y;
@@ -301,7 +303,7 @@ namespace cv { namespace gpu { namespace device
}
template <int cn>
void matchTemplatePrepared_SQDIFF_NORMED_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned int templ_sqsum,
void matchTemplatePrepared_SQDIFF_NORMED_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned long long templ_sqsum,
DevMem2Df result, cudaStream_t stream)
{
const dim3 threads(32, 8);
@@ -315,10 +317,10 @@ namespace cv { namespace gpu { namespace device
}
void matchTemplatePrepared_SQDIFF_NORMED_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned int templ_sqsum,
void matchTemplatePrepared_SQDIFF_NORMED_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned long long templ_sqsum,
DevMem2Df result, int cn, cudaStream_t stream)
{
typedef void (*caller_t)(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned int templ_sqsum, DevMem2Df result, cudaStream_t stream);
typedef void (*caller_t)(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned long long templ_sqsum, DevMem2Df result, cudaStream_t stream);
static const caller_t callers[] =
{
0, matchTemplatePrepared_SQDIFF_NORMED_8U<1>, matchTemplatePrepared_SQDIFF_NORMED_8U<2>, matchTemplatePrepared_SQDIFF_NORMED_8U<3>, matchTemplatePrepared_SQDIFF_NORMED_8U<4>

2
modules/gpu/src/cuda/matrix_reductions.cu
View File

@@ -1739,7 +1739,7 @@ namespace cv { namespace gpu { namespace device
template <typename T>
void sqrSumCaller(const DevMem2Db src, PtrStepb buf, double* sum, int cn)
{
typedef typename SumType<T>::R R;
typedef double R;
dim3 threads, grid;
estimateThreadCfg(src.cols, src.rows, threads, grid);

8
modules/gpu/src/match_template.cpp
View File

@@ -62,10 +62,10 @@ namespace cv { namespace gpu { namespace device
void matchTemplateNaive_SQDIFF_8U(const DevMem2Db image, const DevMem2Db templ, DevMem2Df result, int cn, cudaStream_t stream);
void matchTemplateNaive_SQDIFF_32F(const DevMem2Db image, const DevMem2Db templ, DevMem2Df result, int cn, cudaStream_t stream);
void matchTemplatePrepared_SQDIFF_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned int templ_sqsum, DevMem2Df result,
void matchTemplatePrepared_SQDIFF_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned long long templ_sqsum, DevMem2Df result,
int cn, cudaStream_t stream);
void matchTemplatePrepared_SQDIFF_NORMED_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned int templ_sqsum, DevMem2Df result,
void matchTemplatePrepared_SQDIFF_NORMED_8U(int w, int h, const DevMem2D_<unsigned long long> image_sqsum, unsigned long long templ_sqsum, DevMem2Df result,
int cn, cudaStream_t stream);
void matchTemplatePrepared_CCOFF_8U(int w, int h, const DevMem2D_<unsigned int> image_sum, unsigned int templ_sum, DevMem2Df result, cudaStream_t stream);
@@ -248,7 +248,7 @@ namespace
GpuMat img_sqsum;
sqrIntegral(image.reshape(1), img_sqsum, stream);
unsigned int templ_sqsum = (unsigned int)sqrSum(templ.reshape(1))[0];
unsigned long long templ_sqsum = (unsigned long long)sqrSum(templ.reshape(1))[0];
matchTemplate_CCORR_8U(image, templ, result, stream);
matchTemplatePrepared_SQDIFF_8U(templ.cols, templ.rows, img_sqsum, templ_sqsum, result, image.channels(), StreamAccessor::getStream(stream));
@@ -260,7 +260,7 @@ namespace
GpuMat img_sqsum;
sqrIntegral(image.reshape(1), img_sqsum, stream);
unsigned int templ_sqsum = (unsigned int)sqrSum(templ.reshape(1))[0];
unsigned long long templ_sqsum = (unsigned long long)sqrSum(templ.reshape(1))[0];
matchTemplate_CCORR_8U(image, templ, result, stream);
matchTemplatePrepared_SQDIFF_NORMED_8U(templ.cols, templ.rows, img_sqsum, templ_sqsum, result, image.channels(), StreamAccessor::getStream(stream));