cv::accumulateSquare && cv::acumulateProduct

diff --git a/modules/imgproc/src/accum.cpp b/modules/imgproc/src/accum.cpp
index 7c78795..41d0807 100644 (file)
--- a/modules/imgproc/src/accum.cpp
+++ b/modules/imgproc/src/accum.cpp
@@ -457,6 +457,56 @@ void cv::accumulateSquare( InputArray _src, InputOutputArray _dst, InputArray _m
 
     Mat src = _src.getMat(), dst = _dst.getMat(), mask = _mask.getMat();
 
+#ifdef HAVE_IPP
+    if (src.dims <= 2 || (src.isContinuous() && dst.isContinuous() && (mask.empty() || mask.isContinuous())))
+    {
+        typedef IppStatus (CV_STDCALL * ippiAddSquare)(const void * pSrc, int srcStep, Ipp32f * pSrcDst, int srcdstStep, IppiSize roiSize);
+        typedef IppStatus (CV_STDCALL * ippiAddSquareMask)(const void * pSrc, int srcStep, const Ipp8u * pMask, int maskStep, Ipp32f * pSrcDst,
+                                                           int srcDstStep, IppiSize roiSize);
+        ippiAddSquare ippFunc = 0;
+        ippiAddSquareMask ippFuncMask = 0;
+
+        if (mask.empty())
+        {
+            ippFunc = sdepth == CV_8U && ddepth == CV_32F ? (ippiAddSquare)ippiAddSquare_8u32f_C1IR :
+                sdepth == CV_16U && ddepth == CV_32F ? (ippiAddSquare)ippiAddSquare_16u32f_C1IR :
+                sdepth == CV_32F && ddepth == CV_32F ? (ippiAddSquare)ippiAddSquare_32f_C1IR : 0;
+        }
+        else if (scn == 1)
+        {
+            ippFuncMask = sdepth == CV_8U && ddepth == CV_32F ? (ippiAddSquareMask)ippiAddSquare_8u32f_C1IMR :
+                sdepth == CV_16U && ddepth == CV_32F ? (ippiAddSquareMask)ippiAddSquare_16u32f_C1IMR :
+                sdepth == CV_32F && ddepth == CV_32F ? (ippiAddSquareMask)ippiAddSquare_32f_C1IMR : 0;
+        }
+
+        if (ippFunc || ippFuncMask)
+        {
+            IppStatus status = ippStsNoErr;
+
+            Size size = src.size();
+            int srcstep = (int)src.step, dststep = (int)dst.step, maskstep = (int)mask.step;
+            if (src.isContinuous() && dst.isContinuous() && mask.isContinuous())
+            {
+                srcstep = src.total() * src.elemSize();
+                dststep = dst.total() * dst.elemSize();
+                maskstep = mask.total() * mask.elemSize();
+                size.width = (int)src.total();
+                size.height = 1;
+            }
+            size.width *= scn;
+
+            if (mask.empty())
+                status = ippFunc(src.data, srcstep, (Ipp32f *)dst.data, dststep, ippiSize(size.width, size.height));
+            else
+                status = ippFuncMask(src.data, srcstep, (Ipp8u *)mask.data, maskstep,
+                                     (Ipp32f *)dst.data, dststep, ippiSize(size.width, size.height));
+
+            if (status == ippStsNoErr)
+                return;
+        }
+    }
+#endif
+
     int fidx = getAccTabIdx(sdepth, ddepth);
     AccFunc func = fidx >= 0 ? accSqrTab[fidx] : 0;
     CV_Assert( func != 0 );
@@ -485,6 +535,59 @@ void cv::accumulateProduct( InputArray _src1, InputArray _src2,
 
     Mat src1 = _src1.getMat(), src2 = _src2.getMat(), dst = _dst.getMat(), mask = _mask.getMat();
 
+#ifdef HAVE_IPP
+    if (src1.dims <= 2 || (src1.isContinuous() && src2.isContinuous() && dst.isContinuous()))
+    {
+        typedef IppStatus (CV_STDCALL * ippiAddProduct)(const void * pSrc1, int src1Step, const void * pSrc2,
+                                                        int src2Step, Ipp32f * pSrcDst, int srcDstStep, IppiSize roiSize);
+        typedef IppStatus (CV_STDCALL * ippiAddProductMask)(const void * pSrc1, int src1Step, const void * pSrc2, int src2Step,
+                                                            const Ipp8u * pMask, int maskStep, Ipp32f * pSrcDst, int srcDstStep, IppiSize roiSize);
+        ippiAddProduct ippFunc = 0;
+        ippiAddProductMask ippFuncMask = 0;
+
+        if (mask.empty())
+        {
+            ippFunc = sdepth == CV_8U && ddepth == CV_32F ? (ippiAddProduct)ippiAddProduct_8u32f_C1IR :
+                sdepth == CV_16U && ddepth == CV_32F ? (ippiAddProduct)ippiAddProduct_16u32f_C1IR :
+                sdepth == CV_32F && ddepth == CV_32F ? (ippiAddProduct)ippiAddProduct_32f_C1IR : 0;
+        }
+        else if (scn == 1)
+        {
+            ippFuncMask = sdepth == CV_8U && ddepth == CV_32F ? (ippiAddProductMask)ippiAddProduct_8u32f_C1IMR :
+                sdepth == CV_16U && ddepth == CV_32F ? (ippiAddProductMask)ippiAddProduct_16u32f_C1IMR :
+                sdepth == CV_32F && ddepth == CV_32F ? (ippiAddProductMask)ippiAddProduct_32f_C1IMR : 0;
+        }
+
+        if (ippFunc || ippFuncMask)
+        {
+            IppStatus status = ippStsNoErr;
+
+            Size size = src1.size();
+            int src1step = (int)src1.step, src2step = (int)src2.step, dststep = (int)dst.step, maskstep = (int)mask.step;
+            if (src1.isContinuous() && src2.isContinuous() && dst.isContinuous() && mask.empty())
+            {
+                src1step = src1.total() * src1.elemSize();
+                src2step = src2.total() * src2.elemSize();
+                dststep = dst.total() * dst.elemSize();
+                maskstep = mask.total() * mask.elemSize();
+                size.width = (int)src1.total();
+                size.height = 1;
+            }
+            size.width *= scn;
+
+            if (mask.empty())
+                status = ippFunc(src1.data, src1step, src2.data, src2step, (Ipp32f *)dst.data,
+                                 dststep, ippiSize(size.width, size.height));
+            else
+                status = ippFuncMask(src1.data, src1step, src2.data, src2step, (Ipp8u *)mask.data, maskstep,
+                                     (Ipp32f *)dst.data, dststep, ippiSize(size.width, size.height));
+
+            if (status == ippStsNoErr)
+                return;
+        }
+    }
+#endif
+
     int fidx = getAccTabIdx(sdepth, ddepth);
     AccProdFunc func = fidx >= 0 ? accProdTab[fidx] : 0;
     CV_Assert( func != 0 );
@@ -513,35 +616,54 @@ void cv::accumulateWeighted( InputArray _src, InputOutputArray _dst,
     Mat src = _src.getMat(), dst = _dst.getMat(), mask = _mask.getMat();
 
 #ifdef HAVE_IPP
-    typedef IppStatus (CV_STDCALL * ippiAddWeighted)(const void * pSrc, int srcStep, Ipp32f * pSrcDst, int srcdstStep, IppiSize roiSize, Ipp32f alpha);
-    typedef IppStatus (CV_STDCALL * ippiAddWeightedMask)(const void * pSrc, int srcStep, const Ipp8u * pMask, int maskStep, Ipp32f * pSrcDst,
-                                                         int srcDstStep, IppiSize roiSize, Ipp32f alpha);
-    ippiAddWeighted ippFunc = 0;
-    ippiAddWeightedMask ippFuncMask = 0;
-
-    if (mask.empty())
-    {
-        ippFunc = sdepth == CV_8U && ddepth == CV_32F ? (ippiAddWeighted)ippiAddWeighted_8u32f_C1IR :
-            sdepth == CV_16U && ddepth == CV_32F ? (ippiAddWeighted)ippiAddWeighted_16u32f_C1IR :
-            sdepth == CV_32F && ddepth == CV_32F ? (ippiAddWeighted)ippiAddWeighted_32f_C1IR : 0;
-    }
-    else
-    {
-        ippFuncMask = sdepth == CV_8U && ddepth == CV_32F ? (ippiAddWeightedMask)ippiAddWeighted_8u32f_C1IMR :
-            sdepth == CV_16U && ddepth == CV_32F ? (ippiAddWeightedMask)ippiAddWeighted_16u32f_C1IMR :
-            sdepth == CV_32F && ddepth == CV_32F ? (ippiAddWeightedMask)ippiAddWeighted_32f_C1IMR : 0;
-    }
-    if (ippFunc || ippFuncMask)
+    if (src.dims <= 2 || (src.isContinuous() && dst.isContinuous() && mask.isContinuous()))
     {
-        IppStatus status = ippStsNoErr;
+        typedef IppStatus (CV_STDCALL * ippiAddWeighted)(const void * pSrc, int srcStep, Ipp32f * pSrcDst, int srcdstStep,
+                                                         IppiSize roiSize, Ipp32f alpha);
+        typedef IppStatus (CV_STDCALL * ippiAddWeightedMask)(const void * pSrc, int srcStep, const Ipp8u * pMask,
+                                                             int maskStep, Ipp32f * pSrcDst,
+                                                             int srcDstStep, IppiSize roiSize, Ipp32f alpha);
+        ippiAddWeighted ippFunc = 0;
+        ippiAddWeightedMask ippFuncMask = 0;
+
         if (mask.empty())
-            status = ippFunc(src.data, (int)src.step, (Ipp32f *)dst.data, (int)dst.step, ippiSize(src.cols * scn, src.rows), (Ipp32f)alpha);
-        else
-            status = ippFuncMask(src.data, (int)src.step, (Ipp8u *)mask.data, (int)mask.step,
-                                 (Ipp32f *)dst.data, (int)dst.step, ippiSize(src.cols * scn, src.rows), (Ipp32f)alpha);
+        {
+            ippFunc = sdepth == CV_8U && ddepth == CV_32F ? (ippiAddWeighted)ippiAddWeighted_8u32f_C1IR :
+                sdepth == CV_16U && ddepth == CV_32F ? (ippiAddWeighted)ippiAddWeighted_16u32f_C1IR :
+                sdepth == CV_32F && ddepth == CV_32F ? (ippiAddWeighted)ippiAddWeighted_32f_C1IR : 0;
+        }
+        else if (scn == 1)
+        {
+            ippFuncMask = sdepth == CV_8U && ddepth == CV_32F ? (ippiAddWeightedMask)ippiAddWeighted_8u32f_C1IMR :
+                sdepth == CV_16U && ddepth == CV_32F ? (ippiAddWeightedMask)ippiAddWeighted_16u32f_C1IMR :
+                sdepth == CV_32F && ddepth == CV_32F ? (ippiAddWeightedMask)ippiAddWeighted_32f_C1IMR : 0;
+        }
 
-        if (status == ippStsNoErr)
-            return;
+        if (ippFunc || ippFuncMask)
+        {
+            IppStatus status = ippStsNoErr;
+
+            Size size = src.size();
+            int srcstep = (int)src.step, dststep = (int)dst.step, maskstep = (int)mask.step;
+            if (src.isContinuous() && dst.isContinuous() && mask.empty())
+            {
+                srcstep = src.total() * src.elemSize();
+                dststep = dst.total() * dst.elemSize();
+                maskstep = mask.total() * mask.elemSize();
+                size.width = (int)src.total();
+                size.height = 1;
+            }
+            size.width *= scn;
+
+            if (mask.empty())
+                status = ippFunc(src.data, srcstep, (Ipp32f *)dst.data, dststep, ippiSize(size.width, size.height), (Ipp32f)alpha);
+            else
+                status = ippFuncMask(src.data, srcstep, (Ipp8u *)mask.data, maskstep,
+                                     (Ipp32f *)dst.data, dststep, ippiSize(size.width, size.height), (Ipp32f)alpha);
+
+            if (status == ippStsNoErr)
+                return;
+        }
     }
 #endif