Merge pull request #2560 from akarsakov:gaussianblur_integer

modules/imgproc/src/opencl/filterSepCol.cl

@@ -3,6 +3,7 @@
 //
 // Copyright (C) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved.
 // Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved.
+// Copyright (C) 2014, Itseez, Inc, all rights reserved.
 // Third party copyrights are property of their respective owners.
 //
 // @Authors
@@ -60,7 +61,7 @@
 #endif
 
 #define DIG(a) a,
-__constant float mat_kernel[] = { COEFF };
+__constant srcT1 mat_kernel[] = { COEFF };
 
 __kernel void col_filter(__global const uchar * src, int src_step, int src_offset, int src_whole_rows, int src_whole_cols,
                          __global uchar * dst, int dst_step, int dst_offset, int dst_rows, int dst_cols, float delta)
@@ -96,9 +97,17 @@ __kernel void col_filter(__global const uchar * src, int src_step, int src_offse
     {
         temp[0] = LDS_DAT[l_y + RADIUSY - i][l_x];
         temp[1] = LDS_DAT[l_y + RADIUSY + i][l_x];
+#ifndef INTEGER_ARITHMETIC
         sum += mad(temp[0], mat_kernel[RADIUSY - i], temp[1] * mat_kernel[RADIUSY + i]);
+#else
+        sum += mad24(temp[0],mat_kernel[RADIUSY - i], temp[1] * mat_kernel[RADIUSY + i]);
+#endif
     }
 
+#ifdef INTEGER_ARITHMETIC
+    sum = (sum + (1 << (SHIFT_BITS-1))) >> SHIFT_BITS;
+#endif
+
     // write the result to dst
     if (x < dst_cols && y < dst_rows)
     {

modules/imgproc/src/opencl/filterSepRow.cl

@@ -3,6 +3,7 @@
 //
 // Copyright (C) 2010-2012, Institute Of Software Chinese Academy Of Science, all rights reserved.
 // Copyright (C) 2010-2012, Advanced Micro Devices, Inc., all rights reserved.
+// Copyright (C) 2014, Itseez, Inc, all rights reserved.
 // Third party copyrights are property of their respective owners.
 //
 // @Authors
@@ -138,7 +139,15 @@
 #endif
 
 #define DIG(a) a,
-__constant float mat_kernel[] = { COEFF };
+__constant dstT1 mat_kernel[] = { COEFF };
+
+#ifndef INTEGER_ARITHMETIC
+#define dstT4 float4
+#define convertDstVec convert_float4
+#else
+#define dstT4 int4
+#define convertDstVec convert_int4
+#endif
 
 __kernel void row_filter_C1_D0(__global const uchar * src, int src_step_in_pixel, int src_offset_x, int src_offset_y,
                                int src_cols, int src_rows, int src_whole_cols, int src_whole_rows,
@@ -155,7 +164,7 @@ __kernel void row_filter_C1_D0(__global const uchar * src, int src_step_in_pixel
     int start_y = y + src_offset_y - radiusy;
     int start_addr = mad24(start_y, src_step_in_pixel, start_x);
 
-    float4 sum;
+    dstT4 sum;
     uchar4 temp[READ_TIMES_ROW];
 
     __local uchar4 LDS_DAT[LSIZE1][READ_TIMES_ROW * LSIZE0 + 1];
@@ -249,19 +258,23 @@ __kernel void row_filter_C1_D0(__global const uchar * src, int src_step_in_pixel
     barrier(CLK_LOCAL_MEM_FENCE);
 
     // read pixels from lds and calculate the result
-    sum = convert_float4(vload4(0,(__local uchar *)&LDS_DAT[l_y][l_x]+RADIUSX+offset)) * mat_kernel[RADIUSX];
+    sum = convertDstVec(vload4(0,(__local uchar *)&LDS_DAT[l_y][l_x]+RADIUSX+offset)) * mat_kernel[RADIUSX];
     for (int i = 1; i <= RADIUSX; ++i)
     {
         temp[0] = vload4(0, (__local uchar*)&LDS_DAT[l_y][l_x] + RADIUSX + offset - i);
         temp[1] = vload4(0, (__local uchar*)&LDS_DAT[l_y][l_x] + RADIUSX + offset + i);
-        sum += mad(convert_float4(temp[0]), mat_kernel[RADIUSX-i], convert_float4(temp[1]) * mat_kernel[RADIUSX + i]);
+#ifndef INTEGER_ARITHMETIC
+        sum += mad(convertDstVec(temp[0]), mat_kernel[RADIUSX-i], convertDstVec(temp[1]) * mat_kernel[RADIUSX + i]);
+#else
+        sum += mad24(convertDstVec(temp[0]), mat_kernel[RADIUSX-i], convertDstVec(temp[1]) * mat_kernel[RADIUSX + i]);
+#endif
     }
 
     start_addr = mad24(y, dst_step_in_pixel, x);
 
     // write the result to dst
     if ((x+3<dst_cols) & (y<dst_rows))
-        *(__global float4*)&dst[start_addr] = sum;
+        *(__global dstT4*)&dst[start_addr] = sum;
     else if ((x+2<dst_cols) && (y<dst_rows))
     {
         dst[start_addr] = sum.x;
@@ -355,7 +368,11 @@ __kernel void row_filter(__global const uchar * src, int src_step, int src_offse
     {
         temp[0] = LDS_DAT[l_y][l_x + RADIUSX - i];
         temp[1] = LDS_DAT[l_y][l_x + RADIUSX + i];
+#ifndef INTEGER_ARITHMETIC
         sum += mad(convertToDstT(temp[0]), mat_kernel[RADIUSX - i], convertToDstT(temp[1]) * mat_kernel[RADIUSX + i]);
+#else
+        sum += mad24(convertToDstT(temp[0]), mat_kernel[RADIUSX - i], convertToDstT(temp[1]) * mat_kernel[RADIUSX + i]);
+#endif
     }
 
     // write the result to dst

modules/imgproc/src/opencl/filterSep_singlePass.cl

@@ -100,8 +100,8 @@
 // horizontal and vertical filter kernels
 // should be defined on host during compile time to avoid overhead
 #define DIG(a) a,
-__constant float mat_kernelX[] = { KERNEL_MATRIX_X };
-__constant float mat_kernelY[] = { KERNEL_MATRIX_Y };
+__constant WT1 mat_kernelX[] = { KERNEL_MATRIX_X };
+__constant WT1 mat_kernelY[] = { KERNEL_MATRIX_Y };
 
 __kernel void sep_filter(__global uchar* Src, int src_step, int srcOffsetX, int srcOffsetY, int height, int width,
                          __global uchar* Dst, int dst_step, int dst_offset, int dst_rows, int dst_cols, float delta)
@@ -124,8 +124,6 @@ __kernel void sep_filter(__global uchar* Src, int src_step, int srcOffsetX, int
     // calculate pixel position in source image taking image offset into account
     int srcX = x + srcOffsetX - RADIUSX;
     int srcY = y + srcOffsetY - RADIUSY;
-    int xb = srcX;
-    int yb = srcY;
 
     // extrapolate coordinates, if needed
     // and read my own source pixel into local memory
@@ -159,12 +157,16 @@ __kernel void sep_filter(__global uchar* Src, int src_step, int srcOffsetX, int
     // do vertical filter pass
     // and store intermediate results to second local memory array
     int i, clocX = lix;
-    WT sum = 0.0f;
+    WT sum = (WT) 0;
     do
     {
-        sum = 0.0f;
+        sum = (WT) 0;
         for (i=0; i<=2*RADIUSY; i++)
+#ifndef INTEGER_ARITHMETIC
             sum = mad(lsmem[liy+i][clocX], mat_kernelY[i], sum);
+#else
+            sum = mad24(lsmem[liy+i][clocX], mat_kernelY[i], sum);
+#endif
         lsmemDy[liy][clocX] = sum;
         clocX += BLK_X;
     }
@@ -180,7 +182,13 @@ __kernel void sep_filter(__global uchar* Src, int src_step, int srcOffsetX, int
     // and calculate final result
     sum = 0.0f;
     for (i=0; i<=2*RADIUSX; i++)
+#ifndef INTEGER_ARITHMETIC
         sum = mad(lsmemDy[liy][lix+i], mat_kernelX[i], sum);
+#else
+        sum = mad24(lsmemDy[liy][lix+i], mat_kernelX[i], sum);
+
+    sum = (sum + (1 << (SHIFT_BITS-1))) >> SHIFT_BITS;
+#endif
 
     // store result into destination image
     storepix(convertToDstT(sum + (WT)(delta)), Dst + mad24(y, dst_step, mad24(x, DSTSIZE, dst_offset)));

modules/imgproc/test/ocl/test_filters.cpp

@@ -219,7 +219,7 @@ OCL_TEST_P(GaussianBlurTest, Mat)
         OCL_OFF(cv::GaussianBlur(src_roi, dst_roi, Size(ksize, ksize), sigma1, sigma2, borderType));
         OCL_ON(cv::GaussianBlur(usrc_roi, udst_roi, Size(ksize, ksize), sigma1, sigma2, borderType));
 
-        Near(CV_MAT_DEPTH(type) == CV_8U ? 3 : 5e-5, false);
+        Near(CV_MAT_DEPTH(type) >= CV_32F ? 5e-5 : 1, false);
     }
 }