Merge pull request #5493 from lupustr3:pvlasov/ipp9_fixes

modules/core/include/opencv2/core/private.hpp

@@ -241,6 +241,26 @@ static inline IppDataType ippiGetDataType(int depth)
         depth == CV_64F ? ipp64f : (IppDataType)-1;
 }
 
+// IPP temporary buffer hepler
+template<typename T>
+class IppAutoBuffer
+{
+public:
+    IppAutoBuffer() { m_pBuffer = NULL; }
+    IppAutoBuffer(int size) { Alloc(size); }
+    ~IppAutoBuffer() { Release(); }
+    T* Alloc(int size) { m_pBuffer = (T*)ippMalloc(size); return m_pBuffer; }
+    void Release() { if(m_pBuffer) ippFree(m_pBuffer); }
+    inline operator T* () { return (T*)m_pBuffer;}
+    inline operator const T* () const { return (const T*)m_pBuffer;}
+private:
+    // Disable copy operations
+    IppAutoBuffer(IppAutoBuffer &) {};
+    IppAutoBuffer& operator =(const IppAutoBuffer &) {return *this;};
+
+    T* m_pBuffer;
+};
+
 #else
 #define IPP_VERSION_X100 0
 #endif

modules/core/src/matmul.cpp

@@ -3131,7 +3131,7 @@ static double dotProd_16u(const ushort* src1, const ushort* src2, int len)
 
 static double dotProd_16s(const short* src1, const short* src2, int len)
 {
-#if (ARITHM_USE_IPP == 1)
+#if (ARITHM_USE_IPP == 1) && (IPP_VERSION_X100 != 900) // bug in IPP 9.0.0
     CV_IPP_CHECK()
     {
         double r = 0;

modules/core/src/system.cpp

@@ -1318,6 +1318,12 @@ public:
                 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")

modules/imgproc/src/filter.cpp

@@ -4579,7 +4579,11 @@ static bool ipp_filter2D( InputArray _src, OutputArray _dst, int ddepth,
     int stype = src.type(), sdepth = CV_MAT_DEPTH(stype), cn = CV_MAT_CN(stype),
         ktype = kernel.type(), kdepth = CV_MAT_DEPTH(ktype);
     bool isolated = (borderType & BORDER_ISOLATED) != 0;
+#if IPP_VERSION_X100 >= 900
+    Point ippAnchor((kernel.cols-1)/2, (kernel.rows-1)/2);
+#else
     Point ippAnchor(kernel.cols >> 1, kernel.rows >> 1);
+#endif
     int borderTypeNI = borderType & ~BORDER_ISOLATED;
     IppiBorderType ippBorderType = ippiGetBorderType(borderTypeNI);
 
@@ -4610,24 +4614,64 @@ static bool ipp_filter2D( InputArray _src, OutputArray _dst, int ddepth,
 
             if ((status = ippiFilterBorderGetSize(kernelSize, dstRoiSize, dataType, kernelType, cn, &specSize, &bufsize)) >= 0)
             {
-                IppiFilterBorderSpec * spec = (IppiFilterBorderSpec *)ippMalloc(specSize);
-                Ipp8u * buffer = ippsMalloc_8u(bufsize);
+                IppAutoBuffer<IppiFilterBorderSpec> spec(specSize);
+                IppAutoBuffer<Ipp8u> buffer(bufsize);
                 Ipp32f borderValue[4] = { 0, 0, 0, 0 };
 
-                Mat reversedKernel;
-                flip(kernel, reversedKernel, -1);
-
-                if ((kdepth == CV_32F && (status = ippiFilterBorderInit_32f((const Ipp32f *)reversedKernel.data, kernelSize,
-                    dataType, cn, ippRndFinancial, spec)) >= 0 ) ||
-                    (kdepth == CV_16S && (status = ippiFilterBorderInit_16s((const Ipp16s *)reversedKernel.data,
-                    kernelSize, 0, dataType, cn, ippRndFinancial, spec)) >= 0))
+                if(kdepth == CV_32F)
                 {
-                    status = ippFunc(src.data, (int)src.step, dst.data, (int)dst.step, dstRoiSize,
-                        ippBorderType, borderValue, spec, buffer);
+                    Ipp32f *pKerBuffer = (Ipp32f*)kernel.data;
+                    IppAutoBuffer<Ipp32f> kerTmp;
+                    int kerStep = sizeof(Ipp32f)*kernelSize.width;
+#if IPP_VERSION_X100 >= 900
+                    if(kernel.step != kerStep)
+                    {
+                        kerTmp.Alloc(kerStep*kernelSize.height);
+                        if(ippiCopy_32f_C1R((Ipp32f*)kernel.data, (int)kernel.step, kerTmp, kerStep, kernelSize) < 0)
+                            return false;
+                        pKerBuffer = kerTmp;
+                    }
+#else
+                    kerTmp.Alloc(kerStep*kernelSize.height);
+                    Mat kerFlip(Size(kernelSize.width, kernelSize.height), CV_32FC1, kerTmp, kerStep);
+                    flip(kernel, kerFlip, -1);
+                    pKerBuffer = kerTmp;
+#endif
+
+                    if((status = ippiFilterBorderInit_32f(pKerBuffer, kernelSize,
+                        dataType, cn, ippRndFinancial, spec)) >= 0 )
+                    {
+                        status = ippFunc(src.data, (int)src.step, dst.data, (int)dst.step, dstRoiSize,
+                            ippBorderType, borderValue, spec, buffer);
+                    }
                 }
+                else if(kdepth == CV_16S)
+                {
+                    Ipp16s *pKerBuffer = (Ipp16s*)kernel.data;
+                    IppAutoBuffer<Ipp16s> kerTmp;
+                    int kerStep = sizeof(Ipp16s)*kernelSize.width;
+#if IPP_VERSION_X100 >= 900
+                    if(kernel.step != kerStep)
+                    {
+                        kerTmp.Alloc(kerStep*kernelSize.height);
+                        if(ippiCopy_16s_C1R((Ipp16s*)kernel.data, (int)kernel.step, kerTmp, kerStep, kernelSize) < 0)
+                            return false;
+                        pKerBuffer = kerTmp;
+                    }
+#else
+                    kerTmp.Alloc(kerStep*kernelSize.height);
+                    Mat kerFlip(Size(kernelSize.width, kernelSize.height), CV_16SC1, kerTmp, kerStep);
+                    flip(kernel, kerFlip, -1);
+                    pKerBuffer = kerTmp;
+#endif
 
-                ippsFree(buffer);
-                ippsFree(spec);
+                    if((status = ippiFilterBorderInit_16s(pKerBuffer, kernelSize,
+                        0, dataType, cn, ippRndFinancial, spec)) >= 0)
+                    {
+                        status = ippFunc(src.data, (int)src.step, dst.data, (int)dst.step, dstRoiSize,
+                            ippBorderType, borderValue, spec, buffer);
+                    }
+                }
             }
 
             if (status >= 0)

modules/imgproc/src/morph.cpp

@@ -1231,17 +1231,18 @@ static bool ipp_MorphReplicate(int op, const Mat &src, Mat &dst, const Mat &kern
     }
     else
     {
+#if IPP_VERSION_X100 != 900 // Problems with accuracy in 9.0.0
 #if IPP_VERSION_X100 >= 900
-        if (((kernel.cols - 1) / 2 != anchor.x) || ((kernel.rows - 1) / 2 != anchor.y)) // Arbitrary anchor is no longer supporeted since IPP 9.0.0
+        if (((kernelSize.width - 1) / 2 != anchor.x) || ((kernelSize.height - 1) / 2 != anchor.y)) // Arbitrary anchor is no longer supporeted since IPP 9.0.0
             return false;
 
-        #define IPP_MORPH_CASE(cvtype, flavor, data_type) \
+        #define IPP_MORPH_CASE(cvtype, flavor, data_type, cn) \
         case cvtype: \
             {\
                 if (op == MORPH_ERODE)\
                 {\
                     int bufSize = 0;\
-                    if (0 > ippiFilterMinBorderGetBufferSize(roiSize, kernelSize, ipp##data_type, 1, &bufSize))\
+                    if (0 > ippiFilterMinBorderGetBufferSize(roiSize, kernelSize, ipp##data_type, cn, &bufSize))\
                         return false;\
                     AutoBuffer<uchar> buf(bufSize + 64);\
                     uchar* buffer = alignPtr((uchar*)buf, 32);\
@@ -1250,7 +1251,7 @@ static bool ipp_MorphReplicate(int op, const Mat &src, Mat &dst, const Mat &kern
                 else\
                 {\
                     int bufSize = 0;\
-                    if (0 > ippiFilterMaxBorderGetBufferSize(roiSize, kernelSize, ipp##data_type, 1, &bufSize))\
+                    if (0 > ippiFilterMaxBorderGetBufferSize(roiSize, kernelSize, ipp##data_type, cn, &bufSize))\
                         return false;\
                     AutoBuffer<uchar> buf(bufSize + 64);\
                     uchar* buffer = alignPtr((uchar*)buf, 32);\
@@ -1261,7 +1262,7 @@ static bool ipp_MorphReplicate(int op, const Mat &src, Mat &dst, const Mat &kern
 #else
         IppiPoint point = {anchor.x, anchor.y};
 
-        #define IPP_MORPH_CASE(cvtype, flavor, data_type) \
+        #define IPP_MORPH_CASE(cvtype, flavor, data_type, cn) \
         case cvtype: \
             {\
                 int bufSize = 0;\
@@ -1279,17 +1280,18 @@ static bool ipp_MorphReplicate(int op, const Mat &src, Mat &dst, const Mat &kern
         CV_SUPPRESS_DEPRECATED_START
         switch (type)
         {
-        IPP_MORPH_CASE(CV_8UC1, 8u_C1R, 8u);
-        IPP_MORPH_CASE(CV_8UC3, 8u_C3R, 8u);
-        IPP_MORPH_CASE(CV_8UC4, 8u_C4R, 8u);
-        IPP_MORPH_CASE(CV_32FC1, 32f_C1R, 32f);
-        IPP_MORPH_CASE(CV_32FC3, 32f_C3R, 32f);
-        IPP_MORPH_CASE(CV_32FC4, 32f_C4R, 32f);
+        IPP_MORPH_CASE(CV_8UC1, 8u_C1R, 8u, 1);
+        IPP_MORPH_CASE(CV_8UC3, 8u_C3R, 8u, 3);
+        IPP_MORPH_CASE(CV_8UC4, 8u_C4R, 8u, 4);
+        IPP_MORPH_CASE(CV_32FC1, 32f_C1R, 32f, 1);
+        IPP_MORPH_CASE(CV_32FC3, 32f_C3R, 32f, 3);
+        IPP_MORPH_CASE(CV_32FC4, 32f_C4R, 32f, 4);
         default:
             ;
         }
         CV_SUPPRESS_DEPRECATED_END
         #undef IPP_MORPH_CASE
+#endif
     }
 #else
     CV_UNUSED(op); CV_UNUSED(src); CV_UNUSED(dst); CV_UNUSED(kernel); CV_UNUSED(ksize); CV_UNUSED(anchor); CV_UNUSED(rectKernel);

modules/imgproc/src/smooth.cpp

@@ -1695,32 +1695,33 @@ static bool ipp_GaussianBlur( InputArray _src, OutputArray _dst, Size ksize,
 
             if (ippiFilterGaussianGetBufferSize(roiSize, (Ipp32u)ksize.width, dataType, cn, &specSize, &bufferSize) >= 0)
             {
-                IppFilterGaussianSpec * pSpec = (IppFilterGaussianSpec *)ippMalloc(specSize);
-                Ipp8u * pBuffer = (Ipp8u*)ippMalloc(bufferSize);
+                IppAutoBuffer<IppFilterGaussianSpec> spec(specSize);
+                IppAutoBuffer<Ipp8u> buffer(bufferSize);
 
-                if (ippiFilterGaussianInit(roiSize, (Ipp32u)ksize.width, (Ipp32f)sigma1, ippBorder, dataType, 1, pSpec, pBuffer) >= 0)
+                if (ippiFilterGaussianInit(roiSize, (Ipp32u)ksize.width, (Ipp32f)sigma1, ippBorder, dataType, cn, spec, buffer) >= 0)
                 {
 #define IPP_FILTER_GAUSS_C1(ippfavor) \
                     { \
-                        typedef Ipp##ippfavor ippType; \
-                        ippType borderValues = 0; \
-                        status = ippiFilterGaussianBorder_##ippfavor##_C1R(src.ptr<ippType>(), (int)src.step, \
-                                dst.ptr<ippType>(), (int)dst.step, roiSize, borderValues, pSpec, pBuffer); \
+                        Ipp##ippfavor borderValues = 0; \
+                        status = ippiFilterGaussianBorder_##ippfavor##_C1R(src.ptr<Ipp##ippfavor>(), (int)src.step, \
+                                dst.ptr<Ipp##ippfavor>(), (int)dst.step, roiSize, borderValues, spec, buffer); \
                     }
 
 #define IPP_FILTER_GAUSS_CN(ippfavor, ippcn) \
                     { \
-                        typedef Ipp##ippfavor ippType; \
-                        ippType borderValues[] = { 0, 0, 0 }; \
-                        status = ippiFilterGaussianBorder_##ippfavor##_C##ippcn##R(src.ptr<ippType>(), (int)src.step, \
-                                dst.ptr<ippType>(), (int)dst.step, roiSize, borderValues, pSpec, pBuffer); \
+                        Ipp##ippfavor borderValues[] = { 0, 0, 0 }; \
+                        status = ippiFilterGaussianBorder_##ippfavor##_C##ippcn##R(src.ptr<Ipp##ippfavor>(), (int)src.step, \
+                                dst.ptr<Ipp##ippfavor>(), (int)dst.step, roiSize, borderValues, spec, buffer); \
                     }
 
                     IppStatus status = ippStsErr;
 #if !HAVE_ICV
+#if IPP_VERSION_X100 > 901 // Buffer overflow in IPP
                     if (type == CV_8UC1)
                         IPP_FILTER_GAUSS_C1(8u)
-                    else if (type == CV_8UC3)
+                    else
+#endif
+                    if (type == CV_8UC3)
                         IPP_FILTER_GAUSS_CN(8u, 3)
                     else if (type == CV_16UC1)
                         IPP_FILTER_GAUSS_C1(16u)
@@ -1737,11 +1738,6 @@ static bool ipp_GaussianBlur( InputArray _src, OutputArray _dst, Size ksize,
                     if (type == CV_32FC1)
                         IPP_FILTER_GAUSS_C1(32f)
 
-                    if (pSpec)
-                        ippFree(pSpec);
-                    if (pBuffer)
-                        ippFree(pBuffer);
-
                     if(status >= 0)
                         return true;
 

modules/imgproc/src/sumpixels.cpp

@@ -425,7 +425,7 @@ namespace cv
 {
 static bool ipp_integral(InputArray _src, OutputArray _sum, OutputArray _sqsum, OutputArray _tilted, int sdepth, int sqdepth)
 {
-#if !defined(HAVE_IPP_ICV_ONLY) // Disabled on ICV due invalid results
+#if !defined(HAVE_IPP_ICV_ONLY) && (IPP_VERSION_X100 != 900) // Disabled on ICV due invalid results
     int type = _src.type(), depth = CV_MAT_DEPTH(type), cn = CV_MAT_CN(type);
     if( sdepth <= 0 )
         sdepth = depth == CV_8U ? CV_32S : CV_64F;