improved performance of moments (on 720p or larger images)

modules/imgproc/src/moments.cpp

@@ -365,7 +365,7 @@ Moments::Moments( double _m00, double _m10, double _m01, double _m20, double _m1
 
 static bool ocl_moments( InputArray _src, Moments& m)
 {
-    const int TILE_SIZE = 16;
+    const int TILE_SIZE = 32;
     const int K = 10;
     ocl::Kernel k("moments", ocl::imgproc::moments_oclsrc, format("-D TILE_SIZE=%d", TILE_SIZE));
     if( k.empty() )
@@ -378,10 +378,10 @@ static bool ocl_moments( InputArray _src, Moments& m)
     int ntiles = xtiles*ytiles;
     UMat umbuf(1, ntiles*K, CV_32S);
 
-    size_t globalsize[] = {xtiles, ytiles};
+    size_t globalsize[] = {xtiles, sz.height}, localsize[] = {1, TILE_SIZE};
     bool ok = k.args(ocl::KernelArg::ReadOnly(src),
                      ocl::KernelArg::PtrWriteOnly(umbuf),
-                     xtiles).run(2, globalsize, 0, true);
+                     xtiles).run(2, globalsize, localsize, true);
     if(!ok)
         return false;
     Mat mbuf = umbuf.getMat(ACCESS_READ);

modules/imgproc/src/opencl/moments.cl

 /* See LICENSE file in the root OpenCV directory */
 
-#if TILE_SIZE > 16
-#error "TILE SIZE should be <= 16"
+#if TILE_SIZE != 32
+#error "TILE SIZE should be 32"
 #endif
 
 __kernel void moments(__global const uchar* src, int src_step, int src_offset,
                       int src_rows, int src_cols, __global int* mom0, int xtiles)
 {
-    int x = get_global_id(0);
-    int y = get_global_id(1);
-    int x_min = x*TILE_SIZE;
-    int y_min = y*TILE_SIZE;
+    int x0 = get_global_id(0);
+    int y0 = get_group_id(1);
+    int x, y = get_local_id(1);
+    int x_min = x0*TILE_SIZE;
+    int ypix = y0*TILE_SIZE + y;
+    __local int mom[TILE_SIZE][10];
 
-    if( x_min < src_cols && y_min < src_rows )
+    if( x_min < src_cols && y0*TILE_SIZE < src_rows )
     {
-        int x_max = min(src_cols - x_min, TILE_SIZE);
-        int y_max = min(src_rows - y_min, TILE_SIZE);
-        int m00=0, m10=0, m01=0, m20=0, m11=0, m02=0, m30=0, m21=0, m12=0, m03=0;
-        __global const uchar* ptr = src + src_offset + y_min*src_step + x_min;
-        __global int* mom = mom0 + (xtiles*y + x)*10;
-        x = x_max & -4;
-
-        for( y = 0; y < y_max; y++, ptr += src_step )
+        if( ypix < src_rows )
         {
+            int x_max = min(src_cols - x_min, TILE_SIZE);
+            __global const uchar* ptr = src + src_offset + ypix*src_step + x_min;
             int4 S = (int4)(0,0,0,0), p;
 
             #define SUM_ELEM(elem, ofs) \
-                (int4)(1, (ofs), ((ofs)*(ofs)), ((ofs)*(ofs)*(ofs)))*elem
+                (int4)(1, (ofs), (ofs)*(ofs), (ofs)*(ofs)*(ofs))*elem
+
+            x = x_max & -4;
             if( x_max >= 4 )
             {
                 p = convert_int4(vload4(0, ptr));
@@ -51,6 +50,30 @@ __kernel void moments(__global const uchar* src, int src_step, int src_offset,
                 }
             }
 
+            if( x_max >= 20 )
+            {
+                p = convert_int4(vload4(0, ptr+16));
+                S += SUM_ELEM(p.s0, 16) + SUM_ELEM(p.s1, 17) + SUM_ELEM(p.s2, 18) + SUM_ELEM(p.s3, 19);
+
+                if( x_max >= 24 )
+                {
+                    p = convert_int4(vload4(0, ptr+20));
+                    S += SUM_ELEM(p.s0, 20) + SUM_ELEM(p.s1, 21) + SUM_ELEM(p.s2, 22) + SUM_ELEM(p.s3, 23);
+
+                    if( x_max >= 28 )
+                    {
+                        p = convert_int4(vload4(0, ptr+24));
+                        S += SUM_ELEM(p.s0, 24) + SUM_ELEM(p.s1, 25) + SUM_ELEM(p.s2, 26) + SUM_ELEM(p.s3, 27);
+
+                        if( x_max >= 32 )
+                        {
+                            p = convert_int4(vload4(0, ptr+28));
+                            S += SUM_ELEM(p.s0, 28) + SUM_ELEM(p.s1, 29) + SUM_ELEM(p.s2, 30) + SUM_ELEM(p.s3, 31);
+                        }
+                    }
+                }
+            }
+
             if( x < x_max )
             {
                 int ps = ptr[x];
@@ -68,27 +91,57 @@ __kernel void moments(__global const uchar* src, int src_step, int src_offset,
             }
 
             int sy = y*y;
-            m00 += S.s0;
-            m10 += S.s1;
-            m01 += y*S.s0;
-            m20 += S.s2;
-            m11 += y*S.s1;
-            m02 += sy*S.s0;
-            m30 += S.s3;
-            m21 += y*S.s2;
-            m12 += sy*S.s1;
-            m03 += y*sy*S.s0;
+
+            mom[y][0] = S.s0;
+            mom[y][1] = S.s1;
+            mom[y][2] = y*S.s0;
+            mom[y][3] = S.s2;
+            mom[y][4] = y*S.s1;
+            mom[y][5] = sy*S.s0;
+            mom[y][6] = S.s3;
+            mom[y][7] = y*S.s2;
+            mom[y][8] = sy*S.s1;
+            mom[y][9] = y*sy*S.s0;
         }
+        else
+            mom[y][0] = mom[y][1] = mom[y][2] = mom[y][3] = mom[y][4] =
+                mom[y][5] = mom[y][6] = mom[y][7] = mom[y][8] = mom[y][9] = 0;
+        barrier(CLK_LOCAL_MEM_FENCE);
 
-        mom[0] = m00;
-        mom[1] = m10;
-        mom[2] = m01;
-        mom[3] = m20;
-        mom[4] = m11;
-        mom[5] = m02;
-        mom[6] = m30;
-        mom[7] = m21;
-        mom[8] = m12;
-        mom[9] = m03;
+        #define REDUCE(d) \
+        if( y < d ) \
+        { \
+            mom[y][0] += mom[y+d][0]; \
+            mom[y][1] += mom[y+d][1]; \
+            mom[y][2] += mom[y+d][2]; \
+            mom[y][3] += mom[y+d][3]; \
+            mom[y][4] += mom[y+d][4]; \
+            mom[y][5] += mom[y+d][5]; \
+            mom[y][6] += mom[y+d][6]; \
+            mom[y][7] += mom[y+d][7]; \
+            mom[y][8] += mom[y+d][8]; \
+            mom[y][9] += mom[y+d][9]; \
+        } \
+        barrier(CLK_LOCAL_MEM_FENCE)
+
+        REDUCE(16);
+        REDUCE(8);
+        REDUCE(4);
+        REDUCE(2);
+
+        if( y == 0 )
+        {
+            __global int* momout = mom0 + (y0*xtiles + x0)*10;
+            momout[0] = mom[0][0] + mom[1][0];
+            momout[1] = mom[0][1] + mom[1][1];
+            momout[2] = mom[0][2] + mom[1][2];
+            momout[3] = mom[0][3] + mom[1][3];
+            momout[4] = mom[0][4] + mom[1][4];
+            momout[5] = mom[0][5] + mom[1][5];
+            momout[6] = mom[0][6] + mom[1][6];
+            momout[7] = mom[0][7] + mom[1][7];
+            momout[8] = mom[0][8] + mom[1][8];
+            momout[9] = mom[0][9] + mom[1][9];
+        }
     }
 }