Merge pull request #1418 from vpisarev:24_ipp_sepfilter_dft

modules/core/src/dxt.cpp

@@ -50,7 +50,7 @@ namespace cv
 # pragma warning(disable: 4748)
 #endif
 
-#if defined HAVE_IPP && IPP_VERSION_MAJOR >= 7
+#if defined HAVE_IPP && IPP_VERSION_MAJOR*100 + IPP_VERSION_MINOR >= 701
 #define USE_IPP_DFT 1
 #else
 #undef USE_IPP_DFT

modules/core/src/stat.cpp

@@ -1607,13 +1607,15 @@ double cv::norm( InputArray _src, int normType, InputArray _mask )
     int depth = src.depth(), cn = src.channels();
 
     normType &= 7;
-    CV_Assert( normType == NORM_INF || normType == NORM_L1 || normType == NORM_L2 || normType == NORM_L2SQR ||
+    CV_Assert( normType == NORM_INF || normType == NORM_L1 ||
+               normType == NORM_L2 || normType == NORM_L2SQR ||
                ((normType == NORM_HAMMING || normType == NORM_HAMMING2) && src.type() == CV_8U) );
 
 #if defined (HAVE_IPP) && (IPP_VERSION_MAJOR >= 7)
     size_t total_size = src.total();
     int rows = src.size[0], cols = (int)(total_size/rows);
-    if( src.dims == 2 || (src.isContinuous() && mask.isContinuous() && cols > 0 && (size_t)rows*cols == total_size)
+    if( (src.dims == 2 || (src.isContinuous() && mask.isContinuous()))
+        && cols > 0 && (size_t)rows*cols == total_size
         && (normType == NORM_INF || normType == NORM_L1 || normType == NORM_L2 || normType == NORM_L2SQR) )
     {
         IppiSize sz = { cols, rows };
@@ -1900,8 +1902,10 @@ double cv::norm( InputArray _src1, InputArray _src2, int normType, InputArray _m
                 ((normType == NORM_HAMMING || normType == NORM_HAMMING2) && src1.type() == CV_8U) );
         size_t total_size = src1.total();
         int rows = src1.size[0], cols = (int)(total_size/rows);
-        if( src1.dims == 2 || (src1.isContinuous() && src2.isContinuous() && mask.isContinuous() && cols > 0 && (size_t)rows*cols == total_size)
-            && (normType == NORM_INF || normType == NORM_L1 || normType == NORM_L2 || normType == NORM_L2SQR) )
+        if( (src1.dims == 2 || (src1.isContinuous() && src2.isContinuous() && mask.isContinuous()))
+            && cols > 0 && (size_t)rows*cols == total_size
+            && (normType == NORM_INF || normType == NORM_L1 ||
+                normType == NORM_L2 || normType == NORM_L2SQR) )
         {
             IppiSize sz = { cols, rows };
             int type = src1.type();
@@ -1974,13 +1978,15 @@ double cv::norm( InputArray _src1, InputArray _src2, int normType, InputArray _m
     CV_Assert( src1.size == src2.size && src1.type() == src2.type() );
 
     normType &= 7;
-    CV_Assert( normType == NORM_INF || normType == NORM_L1 || normType == NORM_L2 || normType == NORM_L2SQR ||
+    CV_Assert( normType == NORM_INF || normType == NORM_L1 ||
+               normType == NORM_L2 || normType == NORM_L2SQR ||
               ((normType == NORM_HAMMING || normType == NORM_HAMMING2) && src1.type() == CV_8U) );
 
 #if defined (HAVE_IPP) && (IPP_VERSION_MAJOR >= 7)
     size_t total_size = src1.total();
     int rows = src1.size[0], cols = (int)(total_size/rows);
-    if( src1.dims == 2 || (src1.isContinuous() && src2.isContinuous() && mask.isContinuous() && cols > 0 && (size_t)rows*cols == total_size)
+    if( (src1.dims == 2 || (src1.isContinuous() && src2.isContinuous() && mask.isContinuous()))
+        && cols > 0 && (size_t)rows*cols == total_size
         && (normType == NORM_INF || normType == NORM_L1 || normType == NORM_L2 || normType == NORM_L2SQR) )
     {
         IppiSize sz = { cols, rows };

modules/imgproc/src/filter.cpp

@@ -46,6 +46,12 @@
                                     Base Image Filter
 \****************************************************************************************/
 
+#if defined HAVE_IPP && IPP_VERSION_MAJOR*100 + IPP_VERSION_MINOR >= 701
+#define USE_IPP_SEP_FILTERS 1
+#else
+#undef USE_IPP_SEP_FILTERS
+#endif
+
 /*
  Various border types, image boundaries are denoted with '|'
 
@@ -1445,21 +1451,53 @@ struct RowVec_32f
     RowVec_32f( const Mat& _kernel )
     {
         kernel = _kernel;
+        haveSSE = checkHardwareSupport(CV_CPU_SSE);
+#ifdef USE_IPP_SEP_FILTERS
+        bufsz = -1;
+#endif
     }
 
     int operator()(const uchar* _src, uchar* _dst, int width, int cn) const
     {
-        if( !checkHardwareSupport(CV_CPU_SSE) )
-            return 0;
-
-        int i = 0, k, _ksize = kernel.rows + kernel.cols - 1;
+        int _ksize = kernel.rows + kernel.cols - 1;
+        const float* src0 = (const float*)_src;
         float* dst = (float*)_dst;
         const float* _kx = (const float*)kernel.data;
+
+#ifdef USE_IPP_SEP_FILTERS
+        IppiSize roisz = { width, 1 };
+        if( (cn == 1 || cn == 3) && width >= _ksize*8 )
+        {
+            if( bufsz < 0 )
+            {
+                if( (cn == 1 && ippiFilterRowBorderPipelineGetBufferSize_32f_C1R(roisz, _ksize, &bufsz) < 0) ||
+                    (cn == 3 && ippiFilterRowBorderPipelineGetBufferSize_32f_C3R(roisz, _ksize, &bufsz) < 0))
+                    return 0;
+            }
+            AutoBuffer<uchar> buf(bufsz + 64);
+            uchar* bufptr = alignPtr((uchar*)buf, 32);
+            int step = (int)(width*sizeof(dst[0])*cn);
+            float borderValue[] = {0.f, 0.f, 0.f};
+            // here is the trick. IPP needs border type and extrapolates the row. We did it already.
+            // So we pass anchor=0 and ignore the right tail of results since they are incorrect there.
+            if( (cn == 1 && ippiFilterRowBorderPipeline_32f_C1R(src0, step, &dst, roisz, _kx, _ksize, 0,
+                                                                ippBorderRepl, borderValue[0], bufptr) < 0) ||
+                (cn == 3 && ippiFilterRowBorderPipeline_32f_C3R(src0, step, &dst, roisz, _kx, _ksize, 0,
+                                                                ippBorderRepl, borderValue, bufptr) < 0))
+                return 0;
+            return width - _ksize + 1;
+        }
+#endif
+
+        if( !haveSSE )
+            return 0;
+
+        int i = 0, k;
         width *= cn;
 
         for( ; i <= width - 8; i += 8 )
         {
-            const float* src = (const float*)_src + i;
+            const float* src = src0 + i;
             __m128 f, s0 = _mm_setzero_ps(), s1 = s0, x0, x1;
             for( k = 0; k < _ksize; k++, src += cn )
             {
@@ -1478,6 +1516,10 @@ struct RowVec_32f
     }
 
     Mat kernel;
+    bool haveSSE;
+#ifdef USE_IPP_SEP_FILTERS
+    mutable int bufsz;
+#endif
 };
 
 

modules/imgproc/src/imgwarp.cpp

@@ -1689,12 +1689,10 @@ public:
           IppiRect dstroi = { 0, dsty, dstwidth, dstheight - dsty };
           int bufsize;
           ippiResizeGetBufSize( srcroi, dstroi, cn, mode, &bufsize );
-          Ipp8u *buf;
-          buf = ippsMalloc_8u( bufsize );
-          IppStatus sts;
-          if( func( src.data, ippiSize(src.cols, src.rows), (int)src.step[0], srcroi, dst.data, (int)dst.step[0], dstroi, inv_scale_x, inv_scale_y, 0, 0, mode, buf ) < 0 )
+          AutoBuffer<uchar> buf(bufsize + 64);
+          uchar* bufptr = alignPtr((uchar*)buf, 32);
+          if( func( src.data, ippiSize(src.cols, src.rows), (int)src.step[0], srcroi, dst.data, (int)dst.step[0], dstroi, inv_scale_x, inv_scale_y, 0, 0, mode, bufptr ) < 0 )
               *ok = false;
-          ippsFree(buf);
       }
 private:
     Mat &src;