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

Added parallel version for DCT_ROWS

Browse Source
This commit is contained in:
Elena Gvozdeva 2014-04-21 12:43:26 +04:00
parent 6880dbfd95
commit 16629bff4c
2 changed files with 130 additions and 100 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

228
modules/core/src/dxt.cpp
View File

@ -2882,125 +2882,153 @@ static void IDCT_64f(const double* src, int src_step, double* dft_src, double* d
namespace cv namespace cv
{ {
#if defined HAVE_IPP && IPP_VERSION_MAJOR >= 7 #if defined HAVE_IPP && IPP_VERSION_MAJOR >= 7 && !defined HAVE_IPP_ICV_ONLY
typedef IppStatus (CV_STDCALL * ippiDCTFwdFunc)(const Ipp32f*, int, Ipp32f*, int, const IppiDCTFwdSpec_32f*, Ipp8u*); typedef IppStatus (CV_STDCALL * ippiDCTFunc)(const Ipp32f*, int, Ipp32f*, int, const void*, Ipp8u*);
typedef IppStatus (CV_STDCALL * ippiDCTInvFunc)(const Ipp32f*, int, Ipp32f*, int, const IppiDCTInvSpec_32f*, Ipp8u*); typedef IppStatus (CV_STDCALL * ippiDCTInitAlloc)(void**, IppiSize, IppHintAlgorithm);
typedef IppStatus (CV_STDCALL * ippiDCTFree)(void* pDCTSpec);
typedef IppStatus (CV_STDCALL * ippiDCTGetBufSize)(const void*, int*);
static bool ippi_DCT_Fwd(const Mat& src, Mat& dst, bool row) template <typename Dct>
class DctIPPLoop_Invoker : public ParallelLoopBody
{ {
if (src.type() != CV_32F) public:
return false;
IppStatus status; DctIPPLoop_Invoker(const Mat& _src, Mat& _dst, const Dct& _ippidct, bool _inv, bool *_ok) :
IppiDCTFwdSpec_32f* pDCTSpec; ParallelLoopBody(), src(_src), dst(_dst), ippidct(_ippidct), inv(_inv), ok(_ok)
Ipp8u *pBuffer;
int bufSize=0;
ippiDCTFwdFunc ippFunc = (ippiDCTFwdFunc)ippiDCTFwd_32f_C1R;
if (ippFunc==0)
return false;
IppiSize srcRoiSize = {src.cols, row ? 1 : src.rows};
CV_SUPPRESS_DEPRECATED_START
status = ippiDCTFwdInitAlloc_32f (&pDCTSpec, srcRoiSize, ippAlgHintNone);
if ( status < 0 )
{ {
ippiDCTFwdFree_32f(pDCTSpec); *ok = true;
return false;
} }
status = ippiDCTFwdGetBufSize_32f (pDCTSpec, &bufSize); virtual void operator()(const Range& range) const
if ( status < 0 )
{ {
ippiDCTFwdFree_32f(pDCTSpec); IppStatus status;
return false; void* pDCTSpec;
} AutoBuffer<uchar> buf;
uchar* pBuffer = 0;
int bufSize=0;
pBuffer = ippsMalloc_8u( bufSize ); IppiSize srcRoiSize = {src.cols, 1};
if (row) CV_SUPPRESS_DEPRECATED_START
{
for (int i=0; i<src.rows; i++) ippiDCTInitAlloc ippInitAlloc = inv ? (ippiDCTInitAlloc)ippiDCTInvInitAlloc_32f : (ippiDCTInitAlloc)ippiDCTFwdInitAlloc_32f;
ippiDCTFree ippFree = inv ? (ippiDCTFree)ippiDCTInvFree_32f : (ippiDCTFree)ippiDCTFwdFree_32f;
ippiDCTGetBufSize ippGetBufSize = inv ? (ippiDCTGetBufSize)ippiDCTInvGetBufSize_32f : (ippiDCTGetBufSize)ippiDCTFwdGetBufSize_32f;
status = ippInitAlloc(&pDCTSpec, srcRoiSize, ippAlgHintNone);
if ( status < 0 )
{ {
status = ippFunc((float*)(src.data+i*src.step), (int)src.step,(float*)(dst.data+i*dst.step), (int)dst.step, pDCTSpec, pBuffer); ippFree(pDCTSpec);
if ( status < 0 ) *ok = false;
{ return;
ippsFree( pBuffer );
ippiDCTFwdFree_32f(pDCTSpec);
return false;
}
} }
status = ippGetBufSize(pDCTSpec, &bufSize);
if ( status < 0 )
{
ippFree(pDCTSpec);
*ok = false;
return;
}
buf.allocate( bufSize );
pBuffer = (uchar*)buf;
for( int i = range.start; i < range.end; ++i)
if(!ippidct((float*)(src.data+i*src.step), (int)src.step,(float*)(dst.data+i*dst.step), (int)dst.step, pDCTSpec, (Ipp8u*)pBuffer))
{
*ok = false;
}
ippFree( pDCTSpec);
CV_SUPPRESS_DEPRECATED_END
} }
else
status = ippFunc((float*)src.data, (int)src.step, (float*)dst.data, (int)dst.step, pDCTSpec, pBuffer);
ippsFree( pBuffer ); private:
ippiDCTFwdFree_32f(pDCTSpec); const Mat& src;
CV_SUPPRESS_DEPRECATED_END Mat& dst;
const Dct& ippidct;
bool inv;
bool *ok;
return status >= 0; const DctIPPLoop_Invoker& operator= (const DctIPPLoop_Invoker&);
};
template <typename Dct>
bool DctIPPLoop(const Mat& src, Mat& dst, const Dct& ippidct, bool inv)
{
bool ok;
parallel_for_(Range(0, src.rows), DctIPPLoop_Invoker<Dct>(src, dst, ippidct, inv, &ok), src.total()/(double)(1<<16) );
return ok;
} }
static bool ippi_DCT_Inv(const Mat& src, Mat& dst, bool row) struct IPPDCTFunctor
{
IPPDCTFunctor(ippiDCTFunc _func) : func(_func){}
bool operator()(const Ipp32f* src, int srcStep, Ipp32f* dst, int dstStep, const void* pDCTSpec, Ipp8u* pBuffer) const
{
return func ? func(src, srcStep, dst, dstStep, pDCTSpec, pBuffer) >= 0 : false;
}
private:
ippiDCTFunc func;
};
static bool ippi_DCT(const Mat& src, Mat& dst, bool inv, bool row)
{ {
if (src.type() != CV_32F) if (src.type() != CV_32F)
return false; return false;
IppStatus status; ippiDCTFunc ippFunc = inv ? (ippiDCTFunc)ippiDCTInv_32f_C1R : (ippiDCTFunc)ippiDCTFwd_32f_C1R ;
IppiDCTInvSpec_32f* pDCTSpec;
Ipp8u *pBuffer;
int bufSize=0;
ippiDCTInvFunc ippFunc = (ippiDCTInvFunc)ippiDCTInv_32f_C1R;
if (ippFunc==0)
return false;
IppiSize srcRoiSize = {src.cols, row ? 1 : src.rows};
CV_SUPPRESS_DEPRECATED_START
status = ippiDCTInvInitAlloc_32f (&pDCTSpec, srcRoiSize, ippAlgHintNone);
if ( status < 0 )
{
ippiDCTInvFree_32f(pDCTSpec);
return false;
}
status = ippiDCTInvGetBufSize_32f (pDCTSpec, &bufSize);
if ( status < 0 )
{
ippiDCTInvFree_32f(pDCTSpec);
return false;
}
pBuffer = ippsMalloc_8u( bufSize );
if (row) if (row)
{ if(DctIPPLoop(src,dst,IPPDCTFunctor(ippFunc),inv))
for (int i=0; i<src.rows; i++) return true;
{ else
status = ippFunc((float*)(src.data+i*src.step), (int)src.step,(float*)(dst.data+i*dst.step), (int)dst.step, pDCTSpec, pBuffer); return false;
if ( status < 0 )
{
ippsFree( pBuffer );
ippiDCTInvFree_32f(pDCTSpec);
return false;
}
}
}
else else
status = ippFunc((float*)src.data, (int)src.step, (float*)dst.data, (int)dst.step, pDCTSpec, pBuffer); {
IppStatus status;
void* pDCTSpec;
AutoBuffer<uchar> buf;
uchar* pBuffer = 0;
int bufSize=0;
ippFree( pBuffer ); IppiSize srcRoiSize = {src.cols, src.rows};
ippiDCTInvFree_32f(pDCTSpec);
CV_SUPPRESS_DEPRECATED_END
return status >= 0; CV_SUPPRESS_DEPRECATED_START
ippiDCTInitAlloc ippInitAlloc = inv ? (ippiDCTInitAlloc)ippiDCTInvInitAlloc_32f : (ippiDCTInitAlloc)ippiDCTFwdInitAlloc_32f;
ippiDCTFree ippFree = inv ? (ippiDCTFree)ippiDCTInvFree_32f : (ippiDCTFree)ippiDCTFwdFree_32f;
ippiDCTGetBufSize ippGetBufSize = inv ? (ippiDCTGetBufSize)ippiDCTInvGetBufSize_32f : (ippiDCTGetBufSize)ippiDCTFwdGetBufSize_32f;
status = ippInitAlloc(&pDCTSpec, srcRoiSize, ippAlgHintNone);
if ( status < 0 )
{
ippFree(pDCTSpec);
return false;
}
status = ippGetBufSize(pDCTSpec, &bufSize);
if ( status < 0 )
{
ippFree(pDCTSpec);
return false;
}
buf.allocate( bufSize );
pBuffer = (uchar*)buf;
status = ippFunc((float*)src.data, (int)src.step, (float*)dst.data, (int)dst.step, pDCTSpec, (Ipp8u*)pBuffer);
ippFree(pDCTSpec);
CV_SUPPRESS_DEPRECATED_END
return status >= 0;
}
} }
#endif #endif
@ -3019,7 +3047,7 @@ void cv::dct( InputArray _src0, OutputArray _dst, int flags )
bool inv = (flags & DCT_INVERSE) != 0; bool inv = (flags & DCT_INVERSE) != 0;
Mat src0 = _src0.getMat(), src = src0; Mat src0 = _src0.getMat(), src = src0;
int type = src.type(), depth = src.depth(); int type = src.type(), depth = src.depth();
void /* *spec_dft = 0, */ *spec = 0; void *spec = 0;
double scale = 1.; double scale = 1.;
int prev_len = 0, nf = 0, stage, end_stage; int prev_len = 0, nf = 0, stage, end_stage;
@ -3036,12 +3064,14 @@ void cv::dct( InputArray _src0, OutputArray _dst, int flags )
_dst.create( src.rows, src.cols, type ); _dst.create( src.rows, src.cols, type );
Mat dst = _dst.getMat(); Mat dst = _dst.getMat();
#if defined HAVE_IPP && IPP_VERSION_MAJOR >= 7 #if defined (HAVE_IPP) && (IPP_VERSION_MAJOR >= 7) && !defined HAVE_IPP_ICV_ONLY
bool row = (flags & DCT_ROWS) != 0; bool row = (flags & DCT_ROWS) != 0;
if (inv && ippi_DCT_Inv(src,dst,row)) if(!row || src.rows>(int)(1<<8))
return; {
if(ippi_DCT_Fwd(src,dst,row)) if(ippi_DCT(src,dst,inv, row))
return; return;
setIppErrorStatus();
}
#endif #endif
DCTFunc dct_func = dct_tbl[(int)inv + (depth == CV_64F)*2]; DCTFunc dct_func = dct_tbl[(int)inv + (depth == CV_64F)*2];

2
modules/imgproc/src/distransform.cpp
View File

@ -577,7 +577,7 @@ trueDistTrans( const Mat& src, Mat& dst )
for( ; i <= m*3; i++ ) for( ; i <= m*3; i++ )
sat_tab[i] = i - shift; sat_tab[i] = i - shift;
cv::parallel_for_(cv::Range(0, n), cv::DTColumnInvoker(&src, &dst, sat_tab, sqr_tab)); cv::parallel_for_(cv::Range(0, n), cv::DTColumnInvoker(&src, &dst, sat_tab, sqr_tab), src.total()/(double)(1<<16));
// stage 2: compute modified distance transform for each row // stage 2: compute modified distance transform for each row
float* inv_tab = sqr_tab + n; float* inv_tab = sqr_tab + n;