Improve the performance of fast_nlmeans_denoising_opencl.

1. Remove unnecessary barriers.
2. Adjust CTA_SIZE based on the following cases for Intel platform:
   a) OCL_Photo_DenoisingGrayscale.DenoisingGrayscale
   b) OCL_Photo_DenoisingColored.DenoisingColored

diff --git a/modules/photo/src/fast_nlmeans_denoising_opencl.hpp b/modules/photo/src/fast_nlmeans_denoising_opencl.hpp
index ae17390..1cdd8fa 100644 (file)
--- a/modules/photo/src/fast_nlmeans_denoising_opencl.hpp
+++ b/modules/photo/src/fast_nlmeans_denoising_opencl.hpp
@@ -19,7 +19,8 @@ enum
 {
     BLOCK_ROWS = 32,
     BLOCK_COLS = 32,
-    CTA_SIZE = 256
+    CTA_SIZE_INTEL = 64,
+    CTA_SIZE_DEFAULT = 256
 };
 
 static int divUp(int a, int b)
@@ -70,6 +71,7 @@ static bool ocl_fastNlMeansDenoising(InputArray _src, OutputArray _dst, float h,
                                      int templateWindowSize, int searchWindowSize)
 {
     int type = _src.type(), cn = CV_MAT_CN(type);
+    int ctaSize = ocl::Device::getDefault().isIntel() ? CTA_SIZE_INTEL : CTA_SIZE_DEFAULT;
     Size size = _src.size();
 
     if ( type != CV_8UC1 && type != CV_8UC2 && type != CV_8UC4 )
@@ -86,12 +88,12 @@ static bool ocl_fastNlMeansDenoising(InputArray _src, OutputArray _dst, float h,
     String opts = format("-D OP_CALC_FASTNLMEANS -D TEMPLATE_SIZE=%d -D SEARCH_SIZE=%d"
                          " -D uchar_t=%s -D int_t=%s -D BLOCK_COLS=%d -D BLOCK_ROWS=%d"
                          " -D CTA_SIZE=%d -D TEMPLATE_SIZE2=%d -D SEARCH_SIZE2=%d"
-                         " -D convert_int_t=%s -D cn=%d -D CTA_SIZE2=%d -D convert_uchar_t=%s",
+                         " -D convert_int_t=%s -D cn=%d -D convert_uchar_t=%s",
                          templateWindowSize, searchWindowSize, ocl::typeToStr(type),
-                         ocl::typeToStr(CV_32SC(cn)), BLOCK_COLS, BLOCK_ROWS, CTA_SIZE,
+                         ocl::typeToStr(CV_32SC(cn)), BLOCK_COLS, BLOCK_ROWS, ctaSize,
                          templateWindowHalfWize, searchWindowHalfSize,
                          ocl::convertTypeStr(CV_8U, CV_32S, cn, cvt[0]), cn,
-                         CTA_SIZE >> 1, ocl::convertTypeStr(CV_32S, CV_8U, cn, cvt[1]));
+                         ocl::convertTypeStr(CV_32S, CV_8U, cn, cvt[1]));
 
     ocl::Kernel k("fastNlMeansDenoising", ocl::photo::nlmeans_oclsrc, opts);
     if (k.empty())
@@ -120,7 +122,7 @@ static bool ocl_fastNlMeansDenoising(InputArray _src, OutputArray _dst, float h,
            ocl::KernelArg::PtrReadOnly(almostDist2Weight),
            ocl::KernelArg::PtrReadOnly(buffer), almostTemplateWindowSizeSqBinShift);
 
-    size_t globalsize[2] = { nblocksx * CTA_SIZE, nblocksy }, localsize[2] = { CTA_SIZE, 1 };
+    size_t globalsize[2] = { nblocksx * ctaSize, nblocksy }, localsize[2] = { ctaSize, 1 };
     return k.run(2, globalsize, localsize, false);
 }
 

diff --git a/modules/photo/src/opencl/nlmeans.cl b/modules/photo/src/opencl/nlmeans.cl
index 67a6ec6..152f4dd 100644 (file)
--- a/modules/photo/src/opencl/nlmeans.cl
+++ b/modules/photo/src/opencl/nlmeans.cl
@@ -206,22 +206,11 @@ inline void convolveWindow(__global const uchar * src, int src_step, int src_off
         weighted_sum += (int_t)(weight) * src_value;
     }
 
-    if (id >= CTA_SIZE2)
-    {
-        int id2 = id - CTA_SIZE2;
-        weights_local[id2] = weights;
-        weighted_sum_local[id2] = weighted_sum;
-    }
-    barrier(CLK_LOCAL_MEM_FENCE);
-
-    if (id < CTA_SIZE2)
-    {
-        weights_local[id] += weights;
-        weighted_sum_local[id] += weighted_sum;
-    }
+    weights_local[id] = weights;
+    weighted_sum_local[id] = weighted_sum;
     barrier(CLK_LOCAL_MEM_FENCE);
 
-    for (int lsize = CTA_SIZE2 >> 1; lsize > 2; lsize >>= 1)
+    for (int lsize = CTA_SIZE >> 1; lsize > 2; lsize >>= 1)
     {
         if (id < lsize)
         {
@@ -252,8 +241,8 @@ __kernel void fastNlMeansDenoising(__global const uchar * src, int src_step, int
     int block_y = get_group_id(1);
     int id = get_local_id(0), first;
 
-    __local int dists[SEARCH_SIZE_SQ], weights[CTA_SIZE2];
-    __local int_t weighted_sum[CTA_SIZE2];
+    __local int dists[SEARCH_SIZE_SQ], weights[CTA_SIZE];
+    __local int_t weighted_sum[CTA_SIZE];
 
     int x0 = block_x * BLOCK_COLS, x1 = min(x0 + BLOCK_COLS, dst_cols);
     int y0 = block_y * BLOCK_ROWS, y1 = min(y0 + BLOCK_ROWS, dst_rows);
@@ -281,7 +270,6 @@ __kernel void fastNlMeansDenoising(__global const uchar * src, int src_step, int
 
                 first = (first + 1) % TEMPLATE_SIZE;
             }
-            barrier(CLK_LOCAL_MEM_FENCE);
 
             convolveWindow(src, src_step, src_offset, dists, almostDist2Weight, dst, dst_step, dst_offset,
                 y, x, id, weights, weighted_sum, almostTemplateWindowSizeSqBinShift);