potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-29 23:33:05 +04:00
Code Packages Releases Wiki Activity

ICV2017u3 package update;

Browse Source 
- Optimizations set change. Now IPP integrations will provide code for SSE42, AVX2 and AVX512 (SKX) CPUs only. For HW below SSE42 IPP code is disabled.
- Performance regressions fixes for IPP code paths;
- cv::boxFilter integration improvement;
- cv::filter2D integration improvement;
This commit is contained in:
Pavel Vlasov
2017-08-17 14:57:58 +03:00
parent 87c27a074d
commit a57718e1ac
30 changed files with 727 additions and 584 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/src/convert.cpp
+49 -32
View File
@@ -86,12 +86,6 @@ static MergeFunc getMergeFunc(int depth)
}
#ifdef HAVE_IPP
#ifdef HAVE_IPP_IW
extern "C" {
IW_DECL(IppStatus) llwiCopySplit(const void *pSrc, int srcStep, void* const pDstOrig[], int dstStep,
IppiSize size, int typeSize, int channels);
}
#endif
namespace cv {
static bool ipp_split(const Mat& src, Mat* mv, int channels)
@@ -114,7 +108,7 @@ static bool ipp_split(const Mat& src, Mat* mv, int channels)
return false;
}
return CV_INSTRUMENT_FUN_IPP(llwiCopySplit, src.ptr(), (int)src.step, dstPtrs, (int)dstStep, size, (int)src.elemSize1(), channels) >= 0;
return CV_INSTRUMENT_FUN_IPP(llwiCopySplit, src.ptr(), (int)src.step, dstPtrs, (int)dstStep, size, (int)src.elemSize1(), channels, 0) >= 0;
}
else
{
@@ -132,7 +126,7 @@ static bool ipp_split(const Mat& src, Mat* mv, int channels)
for( size_t i = 0; i < it.nplanes; i++, ++it )
{
if(CV_INSTRUMENT_FUN_IPP(llwiCopySplit, ptrs[0], 0, (void**)&ptrs[1], 0, size, (int)src.elemSize1(), channels) < 0)
if(CV_INSTRUMENT_FUN_IPP(llwiCopySplit, ptrs[0], 0, (void**)&ptrs[1], 0, size, (int)src.elemSize1(), channels, 0) < 0)
return false;
}
return true;
@@ -273,12 +267,6 @@ void cv::split(InputArray _m, OutputArrayOfArrays _mv)
}
#ifdef HAVE_IPP
#ifdef HAVE_IPP_IW
extern "C" {
IW_DECL(IppStatus) llwiCopyMerge(const void* const pSrc[], int srcStep, void *pDst, int dstStep,
IppiSize size, int typeSize, int channels);
}
#endif
namespace cv {
static bool ipp_merge(const Mat* mv, Mat& dst, int channels)
@@ -301,7 +289,7 @@ static bool ipp_merge(const Mat* mv, Mat& dst, int channels)
return false;
}
return CV_INSTRUMENT_FUN_IPP(llwiCopyMerge, srcPtrs, (int)srcStep, dst.ptr(), (int)dst.step, size, (int)mv[0].elemSize1(), channels) >= 0;
return CV_INSTRUMENT_FUN_IPP(llwiCopyMerge, srcPtrs, (int)srcStep, dst.ptr(), (int)dst.step, size, (int)mv[0].elemSize1(), channels, 0) >= 0;
}
else
{
@@ -319,7 +307,7 @@ static bool ipp_merge(const Mat* mv, Mat& dst, int channels)
for( size_t i = 0; i < it.nplanes; i++, ++it )
{
if(CV_INSTRUMENT_FUN_IPP(llwiCopyMerge, (const void**)&ptrs[1], 0, ptrs[0], 0, size, (int)mv[0].elemSize1(), channels) < 0)
if(CV_INSTRUMENT_FUN_IPP(llwiCopyMerge, (const void**)&ptrs[1], 0, ptrs[0], 0, size, (int)mv[0].elemSize1(), channels, 0) < 0)
return false;
}
return true;
@@ -820,16 +808,10 @@ void cv::mixChannels(InputArrayOfArrays src, InputOutputArrayOfArrays dst,
}
#ifdef HAVE_IPP
#ifdef HAVE_IPP_IW
extern "C" {
IW_DECL(IppStatus) llwiCopyMixed(const void *pSrc, int srcStep, int srcChannels, void *pDst, int dstStep, int dstChannels,
IppiSize size, int typeSize, int channelsShift);
}
#endif
namespace cv
{
static bool ipp_extractInsertChannel(const Mat &src, Mat &dst, int channel)
static bool ipp_extractChannel(const Mat &src, Mat &dst, int channel)
{
#ifdef HAVE_IPP_IW
CV_INSTRUMENT_REGION_IPP()
@@ -840,14 +822,11 @@ static bool ipp_extractInsertChannel(const Mat &src, Mat &dst, int channel)
if(src.dims != dst.dims)
return false;
if(srcChannels == dstChannels || (srcChannels != 1 && dstChannels != 1))
return false;
if(src.dims <= 2)
{
IppiSize size = ippiSize(src.size());
return CV_INSTRUMENT_FUN_IPP(llwiCopyMixed, src.ptr(), (int)src.step, srcChannels, dst.ptr(), (int)dst.step, dstChannels, size, (int)src.elemSize1(), channel) >= 0;
return CV_INSTRUMENT_FUN_IPP(llwiCopyChannel, src.ptr(), (int)src.step, srcChannels, channel, dst.ptr(), (int)dst.step, dstChannels, 0, size, (int)src.elemSize1()) >= 0;
}
else
{
@@ -859,7 +838,45 @@ static bool ipp_extractInsertChannel(const Mat &src, Mat &dst, int channel)
for( size_t i = 0; i < it.nplanes; i++, ++it )
{
if(CV_INSTRUMENT_FUN_IPP(llwiCopyMixed, ptrs[0], 0, srcChannels, ptrs[1], 0, dstChannels, size, (int)src.elemSize1(), channel) < 0)
if(CV_INSTRUMENT_FUN_IPP(llwiCopyChannel, ptrs[0], 0, srcChannels, channel, ptrs[1], 0, dstChannels, 0, size, (int)src.elemSize1()) < 0)
return false;
}
return true;
}
#else
CV_UNUSED(src); CV_UNUSED(dst); CV_UNUSED(channel);
return false;
#endif
}
static bool ipp_insertChannel(const Mat &src, Mat &dst, int channel)
{
#ifdef HAVE_IPP_IW
CV_INSTRUMENT_REGION_IPP()
int srcChannels = src.channels();
int dstChannels = dst.channels();
if(src.dims != dst.dims)
return false;
if(src.dims <= 2)
{
IppiSize size = ippiSize(src.size());
return CV_INSTRUMENT_FUN_IPP(llwiCopyChannel, src.ptr(), (int)src.step, srcChannels, 0, dst.ptr(), (int)dst.step, dstChannels, channel, size, (int)src.elemSize1()) >= 0;
}
else
{
const Mat *arrays[] = {&dst, NULL};
uchar *ptrs[2] = {NULL};
NAryMatIterator it(arrays, ptrs);
IppiSize size = {(int)it.size, 1};
for( size_t i = 0; i < it.nplanes; i++, ++it )
{
if(CV_INSTRUMENT_FUN_IPP(llwiCopyChannel, ptrs[0], 0, srcChannels, 0, ptrs[1], 0, dstChannels, channel, size, (int)src.elemSize1()) < 0)
return false;
}
return true;
@@ -893,7 +910,7 @@ void cv::extractChannel(InputArray _src, OutputArray _dst, int coi)
_dst.create(src.dims, &src.size[0], depth);
Mat dst = _dst.getMat();
CV_IPP_RUN_FAST(ipp_extractInsertChannel(src, dst, coi))
CV_IPP_RUN_FAST(ipp_extractChannel(src, dst, coi))
mixChannels(&src, 1, &dst, 1, ch, 1);
}
@@ -917,7 +934,7 @@ void cv::insertChannel(InputArray _src, InputOutputArray _dst, int coi)
Mat src = _src.getMat(), dst = _dst.getMat();
CV_IPP_RUN_FAST(ipp_extractInsertChannel(src, dst, coi))
CV_IPP_RUN_FAST(ipp_insertChannel(src, dst, coi))
mixChannels(&src, 1, &dst, 1, ch, 1);
}
@@ -5152,7 +5169,7 @@ static bool ipp_convertTo(Mat &src, Mat &dst, double alpha, double beta)
iwSrc.Init(ippiSize(sz), srcDepth, 1, NULL, (void*)src.ptr(), src.step);
iwDst.Init(ippiSize(sz), dstDepth, 1, NULL, (void*)dst.ptr(), dst.step);
CV_INSTRUMENT_FUN_IPP(::ipp::iwiScale, &iwSrc, &iwDst, alpha, beta, mode);
CV_INSTRUMENT_FUN_IPP(::ipp::iwiScale, iwSrc, iwDst, alpha, beta, ::ipp::IwiScaleParams(mode));
}
else
{
@@ -5168,7 +5185,7 @@ static bool ipp_convertTo(Mat &src, Mat &dst, double alpha, double beta)
iwSrc.m_ptr = ptrs[0];
iwDst.m_ptr = ptrs[1];
CV_INSTRUMENT_FUN_IPP(::ipp::iwiScale, &iwSrc, &iwDst, alpha, beta, mode);
CV_INSTRUMENT_FUN_IPP(::ipp::iwiScale, iwSrc, iwDst, alpha, beta, ::ipp::IwiScaleParams(mode));
}
}
}
modules/core/src/copy.cpp
+11 -23
View File
@@ -49,18 +49,6 @@
#include "precomp.hpp"
#include "opencl_kernels_core.hpp"
#ifdef HAVE_IPP_IW
extern "C" {
IW_DECL(IppStatus) llwiCopyMask(const void *pSrc, int srcStep, void *pDst, int dstStep,
IppiSize size, int typeSize, int channels, const Ipp8u *pMask, int maskStep);
IW_DECL(IppStatus) llwiSet(const double *pValue, void *pDst, int dstStep,
IppiSize size, IppDataType dataType, int channels);
IW_DECL(IppStatus) llwiSetMask(const double *pValue, void *pDst, int dstStep,
IppiSize size, IppDataType dataType, int channels, const Ipp8u *pMask, int maskStep);
IW_DECL(IppStatus) llwiCopyMakeBorder(const void *pSrc, IppSizeL srcStep, void *pDst, IppSizeL dstStep,
IppiSizeL size, IppDataType dataType, int channels, IppiBorderSize *pBorderSize, IppiBorderType border, const Ipp64f *pBorderVal);
}
#endif
namespace cv
{
@@ -480,9 +468,9 @@ static bool ipp_Mat_setTo_Mat(Mat &dst, Mat &_val, Mat &mask)
if(dst.dims <= 2)
{
IppiSize size = ippiSize(dst.size());
IppDataType dataType = ippiGetDataType(dst.depth());
::ipp::IwValue s;
IppiSize size = ippiSize(dst.size());
IppDataType dataType = ippiGetDataType(dst.depth());
::ipp::IwValueFloat s;
convertAndUnrollScalar(_val, CV_MAKETYPE(CV_64F, dst.channels()), (uchar*)((Ipp64f*)s), 1);
return CV_INSTRUMENT_FUN_IPP(llwiSetMask, s, dst.ptr(), (int)dst.step, size, dataType, dst.channels(), mask.ptr(), (int)mask.step) >= 0;
@@ -493,9 +481,9 @@ static bool ipp_Mat_setTo_Mat(Mat &dst, Mat &_val, Mat &mask)
uchar *ptrs[2] = {NULL};
NAryMatIterator it(arrays, ptrs);
IppiSize size = {(int)it.size, 1};
IppDataType dataType = ippiGetDataType(dst.depth());
::ipp::IwValue s;
IppiSize size = {(int)it.size, 1};
IppDataType dataType = ippiGetDataType(dst.depth());
::ipp::IwValueFloat s;
convertAndUnrollScalar(_val, CV_MAKETYPE(CV_64F, dst.channels()), (uchar*)((Ipp64f*)s), 1);
for( size_t i = 0; i < it.nplanes; i++, ++it)
@@ -717,7 +705,7 @@ static bool ipp_flip(Mat &src, Mat &dst, int flip_mode)
::ipp::IwiImage iwSrc = ippiGetImage(src);
::ipp::IwiImage iwDst = ippiGetImage(dst);
CV_INSTRUMENT_FUN_IPP(::ipp::iwiMirror, &iwSrc, &iwDst, ippMode);
CV_INSTRUMENT_FUN_IPP(::ipp::iwiMirror, iwSrc, iwDst, ippMode);
}
catch(::ipp::IwException)
{
@@ -1155,13 +1143,13 @@ static bool ipp_copyMakeBorder( Mat &_src, Mat &_dst, int top, int bottom,
if(_src.dims > 2)
return false;
Rect dstRect(borderSize.borderLeft, borderSize.borderTop,
_dst.cols - borderSize.borderRight - borderSize.borderLeft,
_dst.rows - borderSize.borderBottom - borderSize.borderTop);
Rect dstRect(borderSize.left, borderSize.top,
_dst.cols - borderSize.right - borderSize.left,
_dst.rows - borderSize.bottom - borderSize.top);
Mat subDst = Mat(_dst, dstRect);
Mat *pSrc = &_src;
return CV_INSTRUMENT_FUN_IPP(llwiCopyMakeBorder, pSrc->ptr(), pSrc->step, subDst.ptr(), subDst.step, size, dataType, _src.channels(), &borderSize, borderType, &value[0]) >= 0;
return CV_INSTRUMENT_FUN_IPP(llwiCopyMakeBorder, pSrc->ptr(), pSrc->step, subDst.ptr(), subDst.step, size, dataType, _src.channels(), borderSize, borderType, &value[0]) >= 0;
#else
CV_UNUSED(_src); CV_UNUSED(_dst); CV_UNUSED(top); CV_UNUSED(bottom); CV_UNUSED(left); CV_UNUSED(right);
CV_UNUSED(_borderType); CV_UNUSED(value);
modules/core/src/mathfuncs_core.dispatch.cpp
+4 -4
View File
@@ -44,7 +44,8 @@ void magnitude32f(const float* x, const float* y, float* mag, int len)
CV_INSTRUMENT_REGION()
CALL_HAL(magnitude32f, cv_hal_magnitude32f, x, y, mag, len);
CV_IPP_RUN(!IPP_DISABLE_PERF_MAG_SSE42 || (ipp::getIppFeatures()&ippCPUID_AVX), CV_INSTRUMENT_FUN_IPP(ippsMagnitude_32f, x, y, mag, len) >= 0);
// SSE42 performance issues
CV_IPP_RUN(IPP_VERSION_X100 > 201800 || cv::ipp::getIppTopFeatures() != ippCPUID_SSE42, CV_INSTRUMENT_FUN_IPP(ippsMagnitude_32f, x, y, mag, len) >= 0);
CV_CPU_DISPATCH(magnitude32f, (x, y, mag, len),
CV_CPU_DISPATCH_MODES_ALL);
@@ -55,7 +56,8 @@ void magnitude64f(const double* x, const double* y, double* mag, int len)
CV_INSTRUMENT_REGION()
CALL_HAL(magnitude64f, cv_hal_magnitude64f, x, y, mag, len);
CV_IPP_RUN(!IPP_DISABLE_PERF_MAG_SSE42 || (ipp::getIppFeatures()&ippCPUID_AVX), CV_INSTRUMENT_FUN_IPP(ippsMagnitude_64f, x, y, mag, len) >= 0);
// SSE42 performance issues
CV_IPP_RUN(IPP_VERSION_X100 > 201800 || cv::ipp::getIppTopFeatures() != ippCPUID_SSE42, CV_INSTRUMENT_FUN_IPP(ippsMagnitude_64f, x, y, mag, len) >= 0);
CV_CPU_DISPATCH(magnitude64f, (x, y, mag, len),
CV_CPU_DISPATCH_MODES_ALL);
@@ -91,7 +93,6 @@ void sqrt32f(const float* src, float* dst, int len)
CV_INSTRUMENT_REGION()
CALL_HAL(sqrt32f, cv_hal_sqrt32f, src, dst, len);
CV_IPP_RUN_FAST(CV_INSTRUMENT_FUN_IPP(ippsSqrt_32f_A21, src, dst, len) >= 0);
CV_CPU_DISPATCH(sqrt32f, (src, dst, len),
CV_CPU_DISPATCH_MODES_ALL);
@@ -103,7 +104,6 @@ void sqrt64f(const double* src, double* dst, int len)
CV_INSTRUMENT_REGION()
CALL_HAL(sqrt64f, cv_hal_sqrt64f, src, dst, len);
CV_IPP_RUN_FAST(CV_INSTRUMENT_FUN_IPP(ippsSqrt_64f_A50, src, dst, len) >= 0);
CV_CPU_DISPATCH(sqrt64f, (src, dst, len),
CV_CPU_DISPATCH_MODES_ALL);
modules/core/src/matmul.cpp
+1 -1
View File
@@ -3101,7 +3101,7 @@ static double dotProd_8u(const uchar* src1, const uchar* src2, int len)
{
double r = 0;
#if ARITHM_USE_IPP
CV_IPP_RUN_FAST(CV_INSTRUMENT_FUN_IPP(ippiDotProd_8u64f_C1R, src1, len*sizeof(uchar), src2, len*sizeof(uchar), ippiSize(len, 1), &r) >= 0, r);
CV_IPP_RUN(IPP_VERSION_X100 > 201800 || cv::ipp::getIppTopFeatures() != ippCPUID_SSE42, CV_INSTRUMENT_FUN_IPP(ippiDotProd_8u64f_C1R, src1, len*sizeof(uchar), src2, len*sizeof(uchar), ippiSize(len, 1), &r) >= 0, r);
#endif
int i = 0;
modules/core/src/matrix.cpp
+1 -4
View File
@@ -4356,7 +4356,6 @@ template<typename T> static void sortIdx_( const Mat& src, Mat& dst, int flags )
}
#ifdef HAVE_IPP
#if !IPP_DISABLE_SORT_IDX
typedef IppStatus (CV_STDCALL *IppSortIndexFunc)(const void* pSrc, Ipp32s srcStrideBytes, Ipp32s *pDstIndx, int len, Ipp8u *pBuffer);
static IppSortIndexFunc getSortIndexFunc(int depth, bool sortDescending)
@@ -4435,7 +4434,6 @@ static bool ipp_sortIdx( const Mat& src, Mat& dst, int flags )
return true;
}
#endif
#endif
typedef void (*SortFunc)(const Mat& src, Mat& dst, int flags);
}
@@ -4472,9 +4470,8 @@ void cv::sortIdx( InputArray _src, OutputArray _dst, int flags )
_dst.release();
_dst.create( src.size(), CV_32S );
dst = _dst.getMat();
#if !IPP_DISABLE_SORT_IDX
CV_IPP_RUN_FAST(ipp_sortIdx(src, dst, flags));
#endif
static SortFunc tab[] =
{
modules/core/src/precomp.hpp
+4 -2
View File
@@ -257,7 +257,8 @@ struct CoreTLSData
//#ifdef HAVE_OPENCL
device(0), useOpenCL(-1),
//#endif
useIPP(-1)
useIPP(-1),
useIPP_NE(-1)
#ifdef HAVE_TEGRA_OPTIMIZATION
,useTegra(-1)
#endif
@@ -272,7 +273,8 @@ struct CoreTLSData
ocl::Queue oclQueue; // the queue used for running a kernel, see also getQueue, Kernel::run
int useOpenCL; // 1 - use, 0 - do not use, -1 - auto/not initialized
//#endif
int useIPP; // 1 - use, 0 - do not use, -1 - auto/not initialized
int useIPP; // 1 - use, 0 - do not use, -1 - auto/not initialized
int useIPP_NE; // 1 - use, 0 - do not use, -1 - auto/not initialized
#ifdef HAVE_TEGRA_OPTIMIZATION
int useTegra; // 1 - use, 0 - do not use, -1 - auto/not initialized
#endif
modules/core/src/stat.cpp
+23 -6
View File
@@ -1294,6 +1294,12 @@ static bool ipp_countNonZero( Mat &src, int &res )
{
CV_INSTRUMENT_REGION_IPP()
#if IPP_VERSION_X100 < 201801
// Poor performance of SSE42
if(cv::ipp::getIppTopFeatures() == ippCPUID_SSE42)
return false;
#endif
Ipp32s count = 0;
int depth = src.depth();
@@ -2531,15 +2537,16 @@ static bool ipp_minMaxIdx(Mat &src, double* _minVal, double* _maxVal, int* _minI
#if IPP_VERSION_X100 >= 700
CV_INSTRUMENT_REGION_IPP()
#if IPP_DISABLE_MINMAX_NAN_SSE42
#if IPP_VERSION_X100 < 201800
// cv::minMaxIdx problem with NaN input
// Disable 32F processing only
if(src.depth() == CV_32F && !(ipp::getIppFeatures()&ippCPUID_AVX))
if(src.depth() == CV_32F && cv::ipp::getIppTopFeatures() == ippCPUID_SSE42)
return false;
#endif
#if IPP_VERSION_X100 < 201801
// cv::minMaxIdx problem with index positions on AVX
#if IPP_VERSION_X100 < 201810
if(!mask.empty() && _maxIdx && ipp::getIppFeatures()&ippCPUID_AVX)
if(!mask.empty() && _maxIdx && cv::ipp::getIppTopFeatures() != ippCPUID_SSE42)
return false;
#endif
@@ -2550,8 +2557,8 @@ static bool ipp_minMaxIdx(Mat &src, double* _minVal, double* _maxVal, int* _minI
IppiPoint minIdx = {-1, -1};
IppiPoint maxIdx = {-1, -1};
float *pMinVal = (_minVal)?&minVal:NULL;
float *pMaxVal = (_maxVal)?&maxVal:NULL;
float *pMinVal = (_minVal || _minIdx)?&minVal:NULL;
float *pMaxVal = (_maxVal || _maxIdx)?&maxVal:NULL;
IppiPoint *pMinIdx = (_minIdx)?&minIdx:NULL;
IppiPoint *pMaxIdx = (_maxIdx)?&maxIdx:NULL;
@@ -2564,6 +2571,8 @@ static bool ipp_minMaxIdx(Mat &src, double* _minVal, double* _maxVal, int* _minI
ippMinMaxFun = ipp_minIdx_wrap;
else if(_maxVal && !_maxIdx && _minVal && !_minIdx)
ippMinMaxFun = ipp_minMax_wrap;
else if(!_maxVal && !_maxIdx && !_minVal && !_minIdx)
return false;
else
ippMinMaxFun = ipp_minMaxIndex_wrap;
}
@@ -2582,8 +2591,12 @@ static bool ipp_minMaxIdx(Mat &src, double* _minVal, double* _maxVal, int* _minI
*_maxVal = maxVal;
if(_minIdx)
{
#if IPP_VERSION_X100 < 201801
// Should be just ippStsNoOperation check, but there is a bug in the function so we need additional checks
if(status == ippStsNoOperation && !mask.empty() && !pMinIdx->x && !pMinIdx->y)
#else
if(status == ippStsNoOperation)
#endif
{
_minIdx[0] = -1;
_minIdx[1] = -1;
@@ -2596,8 +2609,12 @@ static bool ipp_minMaxIdx(Mat &src, double* _minVal, double* _maxVal, int* _minI
}
if(_maxIdx)
{
#if IPP_VERSION_X100 < 201801
// Should be just ippStsNoOperation check, but there is a bug in the function so we need additional checks
if(status == ippStsNoOperation && !mask.empty() && !pMaxIdx->x && !pMaxIdx->y)
#else
if(status == ippStsNoOperation)
#endif
{
_maxIdx[0] = -1;
_maxIdx[1] = -1;
modules/core/src/system.cpp
+164 -24
View File
@@ -1909,55 +1909,146 @@ struct IPPInitSingleton
public:
IPPInitSingleton()
{
useIPP = true;
ippStatus = 0;
funcname = NULL;
filename = NULL;
linen = 0;
ippFeatures = 0;
useIPP = true;
useIPP_NE = false;
ippStatus = 0;
funcname = NULL;
filename = NULL;
linen = 0;
cpuFeatures = 0;
ippFeatures = 0;
ippTopFeatures = 0;
pIppLibInfo = NULL;
ippStatus = ippGetCpuFeatures(&cpuFeatures, NULL);
if(ippStatus < 0)
{
std::cerr << "ERROR: IPP cannot detect CPU features, IPP was disabled " << std::endl;
useIPP = false;
return;
}
ippFeatures = cpuFeatures;
bool unsupported = false;
const char* pIppEnv = getenv("OPENCV_IPP");
cv::String env = pIppEnv;
if(env.size())
{
env = env.toLowerCase();
if(env.substr(0, 2) == "ne")
{
useIPP_NE = true;
env = env.substr(3, env.size());
}
if(env == "disabled")
{
std::cerr << "WARNING: IPP was disabled by OPENCV_IPP environment variable" << std::endl;
useIPP = false;
}
#if IPP_VERSION_X100 >= 900
else if(env == "sse")
ippFeatures = ippCPUID_SSE;
else if(env == "sse2")
ippFeatures = ippCPUID_SSE2;
else if(env == "sse3")
ippFeatures = ippCPUID_SSE3;
else if(env == "ssse3")
ippFeatures = ippCPUID_SSSE3;
else if(env == "sse41")
ippFeatures = ippCPUID_SSE41;
else if(env == "sse42")
ippFeatures = ippCPUID_SSE42;
else if(env == "avx")
ippFeatures = ippCPUID_AVX;
{
if(!(cpuFeatures&ippCPUID_SSE42))
unsupported = true;
ippFeatures = ippCPUID_MMX|ippCPUID_SSE|ippCPUID_SSE2|ippCPUID_SSE3|ippCPUID_SSSE3|ippCPUID_SSE41|ippCPUID_SSE42;
ippFeatures |= (cpuFeatures&ippCPUID_AES);
ippFeatures |= (cpuFeatures&ippCPUID_CLMUL);
ippFeatures |= (cpuFeatures&ippCPUID_SHA);
}
else if(env == "avx2")
ippFeatures = ippCPUID_AVX2;
{
if(!(cpuFeatures&ippCPUID_AVX2))
unsupported = true;
ippFeatures = ippCPUID_MMX|ippCPUID_SSE|ippCPUID_SSE2|ippCPUID_SSE3|ippCPUID_SSSE3|ippCPUID_SSE41|ippCPUID_SSE42|ippCPUID_AVX|ippCPUID_AVX2;
ippFeatures |= (cpuFeatures&ippCPUID_AES);
ippFeatures |= (cpuFeatures&ippCPUID_CLMUL);
ippFeatures |= (cpuFeatures&ippCPUID_F16C);
ippFeatures |= (cpuFeatures&ippCPUID_ADCOX);
ippFeatures |= (cpuFeatures&ippCPUID_RDSEED);
ippFeatures |= (cpuFeatures&ippCPUID_PREFETCHW);
ippFeatures |= (cpuFeatures&ippCPUID_MPX);
}
#if defined (_M_AMD64) || defined (__x86_64__)
else if(env == "avx512")
{
if(!(cpuFeatures&ippCPUID_AVX512F))
unsupported = true;
ippFeatures = ippCPUID_MMX|ippCPUID_SSE|ippCPUID_SSE2|ippCPUID_SSE3|ippCPUID_SSSE3|ippCPUID_SSE41|ippCPUID_SSE42|ippCPUID_AVX|ippCPUID_AVX2|ippCPUID_AVX512F;
ippFeatures |= (cpuFeatures&ippCPUID_AES);
ippFeatures |= (cpuFeatures&ippCPUID_CLMUL);
ippFeatures |= (cpuFeatures&ippCPUID_F16C);
ippFeatures |= (cpuFeatures&ippCPUID_ADCOX);
ippFeatures |= (cpuFeatures&ippCPUID_RDSEED);
ippFeatures |= (cpuFeatures&ippCPUID_PREFETCHW);
ippFeatures |= (cpuFeatures&ippCPUID_MPX);
ippFeatures |= (cpuFeatures&ippCPUID_AVX512CD);
ippFeatures |= (cpuFeatures&ippCPUID_AVX512VL);
ippFeatures |= (cpuFeatures&ippCPUID_AVX512BW);
ippFeatures |= (cpuFeatures&ippCPUID_AVX512DQ);
ippFeatures |= (cpuFeatures&ippCPUID_AVX512ER);
ippFeatures |= (cpuFeatures&ippCPUID_AVX512PF);
ippFeatures |= (cpuFeatures&ippCPUID_AVX512VBMI);
}
#endif
else
std::cerr << "ERROR: Improper value of OPENCV_IPP: " << env.c_str() << std::endl;
std::cerr << "ERROR: Improper value of OPENCV_IPP: " << env.c_str() << ". Correct values are: disabled, sse42, avx2, avx512 (Intel64 only)" << std::endl;
}
if(unsupported)
{
std::cerr << "WARNING: selected IPP features are not supported by CPU. IPP was initialized with default features" << std::endl;
ippFeatures = cpuFeatures;
}
// Disable AVX1 since we don't track regressions for it. SSE42 will be used instead
if(cpuFeatures&ippCPUID_AVX && !(cpuFeatures&ippCPUID_AVX2))
ippFeatures &= ~ippCPUID_AVX;
// IPP integrations in OpenCV support only SSE4.2, AVX2 and AVX-512 optimizations.
if(!(
cpuFeatures&ippCPUID_AVX512F ||
cpuFeatures&ippCPUID_AVX2 ||
cpuFeatures&ippCPUID_SSE42
))
{
useIPP = false;
return;
}
IPP_INITIALIZER(ippFeatures)
ippFeatures = ippGetEnabledCpuFeatures();
// Detect top level optimizations to make comparison easier for optimizations dependent conditions
if(ippFeatures&ippCPUID_AVX512F)
{
if((ippFeatures&ippCPUID_AVX512_SKX) == ippCPUID_AVX512_SKX)
ippTopFeatures = ippCPUID_AVX512_SKX;
else if((ippFeatures&ippCPUID_AVX512_KNL) == ippCPUID_AVX512_KNL)
ippTopFeatures = ippCPUID_AVX512_KNL;
else
ippTopFeatures = ippCPUID_AVX512F; // Unknown AVX512 configuration
}
else if(ippFeatures&ippCPUID_AVX2)
ippTopFeatures = ippCPUID_AVX2;
else if(ippFeatures&ippCPUID_SSE42)
ippTopFeatures = ippCPUID_SSE42;
pIppLibInfo = ippiGetLibVersion();
}
bool useIPP;
public:
bool useIPP;
bool useIPP_NE;
int ippStatus; // 0 - all is ok, -1 - IPP functions failed
int ippStatus; // 0 - all is ok, -1 - IPP functions failed
const char *funcname;
const char *filename;
int linen;
Ipp64u ippFeatures;
Ipp64u cpuFeatures;
Ipp64u ippTopFeatures;
const IppLibraryVersion *pIppLibInfo;
};
static IPPInitSingleton& getIPPSingleton()
@@ -1983,6 +2074,17 @@ int getIppFeatures()
#endif
}
unsigned long long getIppTopFeatures();
unsigned long long getIppTopFeatures()
{
#ifdef HAVE_IPP
return getIPPSingleton().ippTopFeatures;
#else
return 0;
#endif
}
void setIppStatus(int status, const char * const _funcname, const char * const _filename, int _line)
{
#ifdef HAVE_IPP
@@ -2013,6 +2115,19 @@ String getIppErrorLocation()
#endif
}
String getIppVersion()
{
#ifdef HAVE_IPP
const IppLibraryVersion *pInfo = getIPPSingleton().pIppLibInfo;
if(pInfo)
return format("%s %s %s", pInfo->Name, pInfo->Version, pInfo->BuildDate);
else
return String("error");
#else
return String("disabled");
#endif
}
bool useIPP()
{
#ifdef HAVE_IPP
@@ -2038,6 +2153,31 @@ void setUseIPP(bool flag)
#endif
}
bool useIPP_NE()
{
#ifdef HAVE_IPP
CoreTLSData* data = getCoreTlsData().get();
if(data->useIPP_NE < 0)
{
data->useIPP_NE = getIPPSingleton().useIPP_NE;
}
return (data->useIPP_NE > 0);
#else
return false;
#endif
}
void setUseIPP_NE(bool flag)
{
CoreTLSData* data = getCoreTlsData().get();
#ifdef HAVE_IPP
data->useIPP_NE = (getIPPSingleton().useIPP_NE)?flag:false;
#else
(void)flag;
data->useIPP_NE = false;
#endif
}
} // namespace ipp
} // namespace cv