GAPI Fluid: SIMD optimization for sep filters 5x5 kernel size (gaussBlur)

modules/gapi/src/backends/fluid/gfluidimgproc.cpp

@@ -599,6 +599,7 @@ static void run_sepfilter(Buffer& dst, const View& src,
 {
     constexpr int kMax = 11;
     GAPI_Assert(kxLen <= kMax && kyLen <= kMax);
+    GAPI_Assert(kxLen == kyLen);
 
     const SRC *in[kMax];
           DST *out;
@@ -625,6 +626,13 @@ static void run_sepfilter(Buffer& dst, const View& src,
         int border = xborder;
         run_sepfilter3x3_impl(out, in, width, chan, kx, ky, border, scale, delta, buf, y, y0);
     }
+    else if (kxLen == 5 && kyLen == 5)
+    {
+        int y = dst.y();
+        int y0 = dst.priv().writeStart();
+
+        run_sepfilter5x5_impl(out, in, width, chan, kx, ky, xborder, scale, delta, buf, y, y0);
+    }
     else
     {
         int length = chan * width;
@@ -788,7 +796,9 @@ GAPI_FLUID_KERNEL(GFluidGaussBlur, cv::gapi::imgproc::GGaussBlur, true)
                               Buffer&    dst,
                               Buffer&    scratch)
     {
-        int kxsize = ksize.width;
+        GAPI_Assert(ksize.height == ksize.width);
+        GAPI_Assert((ksize.height == 3) || (ksize.height == 5));
+        const int kxsize = ksize.width;
         int kysize = ksize.height;
 
         auto *kx = scratch.OutLine<float>(); // cached kernX data
@@ -801,7 +811,7 @@ GAPI_FLUID_KERNEL(GFluidGaussBlur, cv::gapi::imgproc::GGaussBlur, true)
         constexpr int buffSize = 5;
         GAPI_Assert(ksize.height <= buffSize);
 
-        float *buf[buffSize]{};
+        float *buf[buffSize] = { nullptr };
 
         buf[0] = ky + kysize;
         for (int i = 1; i < ksize.height; ++i)

modules/gapi/src/backends/fluid/gfluidimgproc_func.dispatch.cpp

@@ -119,6 +119,28 @@ RUN_SEPFILTER3X3_IMPL( float,  float)
 
 #undef RUN_SEPFILTER3X3_IMPL
 
+#define RUN_SEPFILTER5x5_IMPL(DST, SRC)                                     \
+void run_sepfilter5x5_impl(DST out[], const SRC *in[], int width, int chan, \
+                           const float kx[], const float ky[], int border,  \
+                           float scale, float delta,                        \
+                           float *buf[], int y, int y0)                     \
+{                                                                           \
+    CV_CPU_DISPATCH(run_sepfilter5x5_impl,                                  \
+        (out, in, width, chan, kx, ky, border, scale, delta, buf,y, y0),    \
+        CV_CPU_DISPATCH_MODES_ALL);                                         \
+}
+
+RUN_SEPFILTER5x5_IMPL(uchar, uchar)
+RUN_SEPFILTER5x5_IMPL(short, uchar)
+RUN_SEPFILTER5x5_IMPL(float, uchar)
+RUN_SEPFILTER5x5_IMPL(ushort, ushort)
+RUN_SEPFILTER5x5_IMPL(short, ushort)
+RUN_SEPFILTER5x5_IMPL(float, ushort)
+RUN_SEPFILTER5x5_IMPL(short, short)
+RUN_SEPFILTER5x5_IMPL(float, short)
+RUN_SEPFILTER5x5_IMPL(float, float)
+
+#undef RUN_SEPFILTER5x5_IMPL
 //-------------------------
 //
 // Fluid kernels: Filter 2D

modules/gapi/src/backends/fluid/gfluidimgproc_func.hpp

@@ -78,6 +78,25 @@ RUN_SEPFILTER3X3_IMPL( float,  float)
 
 #undef RUN_SEPFILTER3X3_IMPL
 
+#define RUN_SEPFILTER5x5_IMPL(DST, SRC)                                     \
+void run_sepfilter5x5_impl(DST out[], const SRC *in[], int width, int chan, \
+                           const float kx[], const float ky[], int border,  \
+                           float scale, float delta,                        \
+                           float *buf[], int y, int y0);
+
+
+RUN_SEPFILTER5x5_IMPL(uchar, uchar)
+RUN_SEPFILTER5x5_IMPL(short, uchar)
+RUN_SEPFILTER5x5_IMPL(float, uchar)
+RUN_SEPFILTER5x5_IMPL(ushort, ushort)
+RUN_SEPFILTER5x5_IMPL(short, ushort)
+RUN_SEPFILTER5x5_IMPL(float, ushort)
+RUN_SEPFILTER5x5_IMPL(short, short)
+RUN_SEPFILTER5x5_IMPL(float, short)
+RUN_SEPFILTER5x5_IMPL(float, float)
+
+#undef RUN_SEPFILTER5x5_IMPL
+
 //-------------------------
 //
 // Fluid kernels: Filter 2D