Merge pull request #2751 from akarsakov:ocl_thresh_opt

modules/imgproc/src/opencl/threshold.cl

@@ -53,29 +53,40 @@
 
 __kernel void threshold(__global const uchar * srcptr, int src_step, int src_offset,
                         __global uchar * dstptr, int dst_step, int dst_offset, int rows, int cols,
-                        T1 thresh, T1 max_val)
+                        T1 thresh, T1 max_val, T1 min_val)
 {
     int gx = get_global_id(0);
-    int gy = get_global_id(1);
+    int gy = get_global_id(1) * STRIDE_SIZE;
 
-    if (gx < cols && gy < rows)
+    if (gx < cols)
     {
         int src_index = mad24(gy, src_step, mad24(gx, (int)sizeof(T), src_offset));
         int dst_index = mad24(gy, dst_step, mad24(gx, (int)sizeof(T), dst_offset));
 
-        T sdata = *(__global const T *)(srcptr + src_index);
-        __global T * dst = (__global T *)(dstptr + dst_index);
+        #pragma unroll
+        for (int i = 0; i < STRIDE_SIZE; i++)
+        {
+            if (gy < rows)
+            {
+                T sdata = *(__global const T *)(srcptr + src_index);
+                __global T * dst = (__global T *)(dstptr + dst_index);
 
-#ifdef THRESH_BINARY
-        dst[0] = sdata > (T)(thresh) ? (T)(max_val) : (T)(0);
-#elif defined THRESH_BINARY_INV
-        dst[0] = sdata > (T)(thresh) ? (T)(0) : (T)(max_val);
-#elif defined THRESH_TRUNC
-        dst[0] = sdata > (T)(thresh) ? (T)(thresh) : sdata;
-#elif defined THRESH_TOZERO
-        dst[0] = sdata > (T)(thresh) ? sdata : (T)(0);
-#elif defined THRESH_TOZERO_INV
-        dst[0] = sdata > (T)(thresh) ? (T)(0) : sdata;
-#endif
+                #ifdef THRESH_BINARY
+                        dst[0] = sdata > (thresh) ? (T)(max_val) : (T)(0);
+                #elif defined THRESH_BINARY_INV
+                        dst[0] = sdata > (thresh) ? (T)(0) : (T)(max_val);
+                #elif defined THRESH_TRUNC
+                        dst[0] = clamp(sdata, (T)min_val, (T)(thresh));
+                #elif defined THRESH_TOZERO
+                        dst[0] = sdata > (thresh) ? sdata : (T)(0);
+                #elif defined THRESH_TOZERO_INV
+                        dst[0] = sdata > (thresh) ? (T)(0) : sdata;
+                #endif
+
+                gy++;
+                src_index += src_step;
+                dst_index += dst_step;
+            }
+        }
     }
 }

modules/imgproc/src/thresh.cpp

@@ -833,9 +833,12 @@ static bool ocl_threshold( InputArray _src, OutputArray _dst, double & thresh, d
 
     const char * const thresholdMap[] = { "THRESH_BINARY", "THRESH_BINARY_INV", "THRESH_TRUNC",
                                           "THRESH_TOZERO", "THRESH_TOZERO_INV" };
+    ocl::Device dev = ocl::Device::getDefault();
+    int stride_size = dev.isIntel() && (dev.type() & ocl::Device::TYPE_GPU) ? 4 : 1;
+
     ocl::Kernel k("threshold", ocl::imgproc::threshold_oclsrc,
-                  format("-D %s -D T=%s -D T1=%s%s", thresholdMap[thresh_type],
-                         ocl::typeToStr(ktype), ocl::typeToStr(depth),
+                  format("-D %s -D T=%s -D T1=%s -D STRIDE_SIZE=%d%s", thresholdMap[thresh_type],
+                         ocl::typeToStr(ktype), ocl::typeToStr(depth), stride_size,
                          doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
     if (k.empty())
         return false;
@@ -847,11 +850,16 @@ static bool ocl_threshold( InputArray _src, OutputArray _dst, double & thresh, d
     if (depth <= CV_32S)
         thresh = cvFloor(thresh);
 
+    const double min_vals[] = { 0, CHAR_MIN, 0, SHRT_MIN, INT_MIN, -FLT_MAX, -DBL_MAX, 0 };
+    double min_val = min_vals[depth];
+
     k.args(ocl::KernelArg::ReadOnlyNoSize(src), ocl::KernelArg::WriteOnly(dst, cn, kercn),
            ocl::KernelArg::Constant(Mat(1, 1, depth, Scalar::all(thresh))),
-           ocl::KernelArg::Constant(Mat(1, 1, depth, Scalar::all(maxval))));
+           ocl::KernelArg::Constant(Mat(1, 1, depth, Scalar::all(maxval))),
+           ocl::KernelArg::Constant(Mat(1, 1, depth, Scalar::all(min_val))));
 
     size_t globalsize[2] = { dst.cols * cn / kercn, dst.rows };
+    globalsize[1] = (globalsize[1] + stride_size - 1) / stride_size;
     return k.run(2, globalsize, NULL, false);
 }