T-API: optimized ocl_flip

diff --git a/modules/core/src/copy.cpp b/modules/core/src/copy.cpp
index b007f3c..7b9c038 100644 (file)
--- a/modules/core/src/copy.cpp
+++ b/modules/core/src/copy.cpp
@@ -610,6 +610,7 @@ flipVert( const uchar* src0, size_t sstep, uchar* dst0, size_t dstep, Size size,
 
 #ifdef HAVE_OPENCL
 
+#define DIVUP(total, grain) (((total) + (grain) - 1) / (grain))
 enum { FLIP_COLS = 1 << 0, FLIP_ROWS = 1 << 1, FLIP_BOTH = FLIP_ROWS | FLIP_COLS };
 
 static bool ocl_flip(InputArray _src, OutputArray _dst, int flipCode )
@@ -628,9 +629,12 @@ static bool ocl_flip(InputArray _src, OutputArray _dst, int flipCode )
     else
         kernelName = "arithm_flip_rows_cols", flipType = FLIP_BOTH;
 
+    ocl::Device dev = ocl::Device::getDefault();
+    int pxPerWIy = (dev.isIntel() && (dev.type() & ocl::Device::TYPE_GPU)) ? 4 : 1;
+
     ocl::Kernel k(kernelName, ocl::core::flip_oclsrc,
-        format( "-D T=%s -D T1=%s -D cn=%d", ocl::memopTypeToStr(type),
-                ocl::memopTypeToStr(depth), cn));
+        format( "-D T=%s -D T1=%s -D cn=%d -D PIX_PER_WI_Y=%d", ocl::memopTypeToStr(type),
+                ocl::memopTypeToStr(depth), cn, pxPerWIy));
     if (k.empty())
         return false;
 
@@ -645,10 +649,13 @@ static bool ocl_flip(InputArray _src, OutputArray _dst, int flipCode )
     k.args(ocl::KernelArg::ReadOnlyNoSize(src),
            ocl::KernelArg::WriteOnly(dst), rows, cols);
 
-    size_t maxWorkGroupSize = ocl::Device::getDefault().maxWorkGroupSize();
+    size_t maxWorkGroupSize = dev.maxWorkGroupSize();
     CV_Assert(maxWorkGroupSize % 4 == 0);
+
     size_t globalsize[2] = { cols, rows }, localsize[2] = { maxWorkGroupSize / 4, 4 };
-    return k.run(2, globalsize, flipType == FLIP_COLS ? localsize : NULL, false);
+    globalsize[1] = DIVUP(globalsize[1], pxPerWIy);
+
+    return k.run(2, globalsize, (flipType == FLIP_COLS) && (!dev.isIntel()) ? localsize : NULL, false);
 }
 
 #endif

diff --git a/modules/core/src/opencl/flip.cl b/modules/core/src/opencl/flip.cl
index bacfe7a..c81dd43 100644 (file)
--- a/modules/core/src/opencl/flip.cl
+++ b/modules/core/src/opencl/flip.cl
@@ -54,15 +54,19 @@ __kernel void arithm_flip_rows(__global const uchar * srcptr, int src_step, int
                                int rows, int cols, int thread_rows, int thread_cols)
 {
     int x = get_global_id(0);
-    int y = get_global_id(1);
+    int y = get_global_id(1)*PIX_PER_WI_Y;
 
-    if (x < cols && y < thread_rows)
+    if (x < cols)
     {
-        T src0 = loadpix(srcptr + mad24(y, src_step, mad24(x, TSIZE, src_offset)));
-        T src1 = loadpix(srcptr + mad24(rows - y - 1, src_step, mad24(x, TSIZE, src_offset)));
+        #pragma unroll
+        for (int cy = 0; cy < PIX_PER_WI_Y && y < thread_rows; ++cy, ++y)
+        {
+            T src0 = loadpix(srcptr + mad24(y, src_step, mad24(x, TSIZE, src_offset)));
+            T src1 = loadpix(srcptr + mad24(rows - y - 1, src_step, mad24(x, TSIZE, src_offset)));
 
-        storepix(src1, dstptr + mad24(y, dst_step, mad24(x, TSIZE, dst_offset)));
-        storepix(src0, dstptr + mad24(rows - y - 1, dst_step, mad24(x, TSIZE, dst_offset)));
+            storepix(src1, dstptr + mad24(y, dst_step, mad24(x, TSIZE, dst_offset)));
+            storepix(src0, dstptr + mad24(rows - y - 1, dst_step, mad24(x, TSIZE, dst_offset)));
+        }
     }
 }
 
@@ -71,16 +75,20 @@ __kernel void arithm_flip_rows_cols(__global const uchar * srcptr, int src_step,
                                     int rows, int cols, int thread_rows, int thread_cols)
 {
     int x = get_global_id(0);
-    int y = get_global_id(1);
+    int y = get_global_id(1)*PIX_PER_WI_Y;
 
-    if (x < cols && y < thread_rows)
+    if (x < cols)
     {
         int x1 = cols - x - 1;
-        T src0 = loadpix(srcptr + mad24(y, src_step, mad24(x, TSIZE, src_offset)));
-        T src1 = loadpix(srcptr + mad24(rows - y - 1, src_step, mad24(x1, TSIZE, src_offset)));
+        #pragma unroll
+        for (int cy = 0; cy < PIX_PER_WI_Y && y < thread_rows; ++cy, ++y)
+        {
+            T src0 = loadpix(srcptr + mad24(y, src_step, mad24(x, TSIZE, src_offset)));
+            T src1 = loadpix(srcptr + mad24(rows - y - 1, src_step, mad24(x1, TSIZE, src_offset)));
 
-        storepix(src0, dstptr + mad24(rows - y - 1, dst_step, mad24(x1, TSIZE, dst_offset)));
-        storepix(src1, dstptr + mad24(y, dst_step, mad24(x, TSIZE, dst_offset)));
+            storepix(src0, dstptr + mad24(rows - y - 1, dst_step, mad24(x1, TSIZE, dst_offset)));
+            storepix(src1, dstptr + mad24(y, dst_step, mad24(x, TSIZE, dst_offset)));
+        }
     }
 }
 
@@ -89,15 +97,19 @@ __kernel void arithm_flip_cols(__global const uchar * srcptr, int src_step, int
                                int rows, int cols, int thread_rows, int thread_cols)
 {
     int x = get_global_id(0);
-    int y = get_global_id(1);
+    int y = get_global_id(1)*PIX_PER_WI_Y;
 
-    if (x < thread_cols && y < rows)
+    if (x < thread_cols)
     {
         int x1 = cols - x - 1;
-        T src0 = loadpix(srcptr + mad24(y, src_step, mad24(x, TSIZE, src_offset)));
-        T src1 = loadpix(srcptr + mad24(y, src_step, mad24(x1, TSIZE, src_offset)));
+        #pragma unroll
+        for (int cy = 0; cy < PIX_PER_WI_Y && y < rows; ++cy, ++y)
+        {
+            T src0 = loadpix(srcptr + mad24(y, src_step, mad24(x, TSIZE, src_offset)));
+            T src1 = loadpix(srcptr + mad24(y, src_step, mad24(x1, TSIZE, src_offset)));
 
-        storepix(src0, dstptr + mad24(y, dst_step, mad24(x1, TSIZE, dst_offset)));
-        storepix(src1, dstptr + mad24(y, dst_step, mad24(x, TSIZE, dst_offset)));
+            storepix(src0, dstptr + mad24(y, dst_step, mad24(x1, TSIZE, dst_offset)));
+            storepix(src1, dstptr + mad24(y, dst_step, mad24(x, TSIZE, dst_offset)));
+        }
     }
 }