vectorized ocl::threshold for single channel images

diff --git a/modules/ocl/perf/perf_imgproc.cpp b/modules/ocl/perf/perf_imgproc.cpp
index 5c89988..c57950f 100644 (file)
--- a/modules/ocl/perf/perf_imgproc.cpp
+++ b/modules/ocl/perf/perf_imgproc.cpp
@@ -366,21 +366,23 @@ PERF_TEST_P(resizeFixture, resize,
 
 ///////////// threshold////////////////////////
 
-CV_ENUM(ThreshType, THRESH_BINARY, THRESH_BINARY_INV, THRESH_TRUNC, THRESH_TOZERO, THRESH_TOZERO_INV)
+CV_ENUM(ThreshType, THRESH_BINARY, THRESH_TOZERO_INV)
 
-typedef tuple<Size, ThreshType> ThreshParams;
+typedef tuple<Size, MatType, ThreshType> ThreshParams;
 typedef TestBaseWithParam<ThreshParams> ThreshFixture;
 
 PERF_TEST_P(ThreshFixture, threshold,
             ::testing::Combine(OCL_TYPICAL_MAT_SIZES,
+                               OCL_PERF_ENUM(CV_8UC1, CV_8UC4, CV_16SC1, CV_16SC4, CV_32FC1),
                                ThreshType::all()))
 {
     const ThreshParams params = GetParam();
     const Size srcSize = get<0>(params);
-    const int threshType = get<1>(params);
+    const int srcType = get<1>(params);
+    const int threshType = get<2>(params);
     const double maxValue = 220.0, threshold = 50;
 
-    Mat src(srcSize, CV_8U), dst(srcSize, CV_8U);
+    Mat src(srcSize, srcType), dst(srcSize, srcType);
     randu(src, 0, 100);
     declare.in(src).out(dst);
 

diff --git a/modules/ocl/src/imgproc.cpp b/modules/ocl/src/imgproc.cpp
index 930865c..adfd88c 100644 (file)
--- a/modules/ocl/src/imgproc.cpp
+++ b/modules/ocl/src/imgproc.cpp
@@ -118,22 +118,20 @@ namespace cv
         static void threshold_runner(const oclMat &src, oclMat &dst, double thresh, double maxVal, int thresholdType)
         {
             bool ival = src.depth() < CV_32F;
+            int cn = src.channels(), vecSize = 4, depth = src.depth();
             std::vector<uchar> thresholdValue = scalarToVector(cv::Scalar::all(ival ? cvFloor(thresh) : thresh), dst.depth(),
                                                                dst.oclchannels(), dst.channels());
             std::vector<uchar> maxValue = scalarToVector(cv::Scalar::all(maxVal), dst.depth(), dst.oclchannels(), dst.channels());
 
-            size_t localThreads[3] = { 16, 16, 1 };
-            size_t globalThreads[3] = { dst.cols, dst.rows, 1 };
-
             const char * const thresholdMap[] = { "THRESH_BINARY", "THRESH_BINARY_INV", "THRESH_TRUNC",
                                                   "THRESH_TOZERO", "THRESH_TOZERO_INV" };
             const char * const channelMap[] = { "", "", "2", "4", "4" };
             const char * const typeMap[] = { "uchar", "char", "ushort", "short", "int", "float", "double" };
-            std::string buildOptions = format("-D T=%s%s -D %s", typeMap[src.depth()], channelMap[src.channels()],
-                                              thresholdMap[thresholdType]);
+            std::string buildOptions = format("-D T=%s%s -D %s", typeMap[depth], channelMap[cn], thresholdMap[thresholdType]);
 
-            int src_step = src.step / src.elemSize(), src_offset = src.offset / src.elemSize();
-            int dst_step = dst.step / dst.elemSize(), dst_offset = dst.offset / dst.elemSize();
+            int elemSize = src.elemSize();
+            int src_step = src.step / elemSize, src_offset = src.offset / elemSize;
+            int dst_step = dst.step / elemSize, dst_offset = dst.offset / elemSize;
 
             vector< pair<size_t, const void *> > args;
             args.push_back( make_pair(sizeof(cl_mem), (void *)&src.data));
@@ -142,11 +140,32 @@ namespace cv
             args.push_back( make_pair(sizeof(cl_mem), (void *)&dst.data));
             args.push_back( make_pair(sizeof(cl_int), (void *)&dst_offset));
             args.push_back( make_pair(sizeof(cl_int), (void *)&dst_step));
-            args.push_back( make_pair(sizeof(cl_int), (void *)&dst.rows));
-            args.push_back( make_pair(sizeof(cl_int), (void *)&dst.cols));
             args.push_back( make_pair(thresholdValue.size(), (void *)&thresholdValue[0]));
             args.push_back( make_pair(maxValue.size(), (void *)&maxValue[0]));
 
+            int max_index = dst.cols, cols = dst.cols;
+            if (cn == 1 && vecSize > 1)
+            {
+                CV_Assert(((vecSize - 1) & vecSize) == 0 && vecSize <= 16);
+                cols = divUp(cols, vecSize);
+                buildOptions += format(" -D VECTORIZED -D VT=%s%d -D VLOADN=vload%d -D VECSIZE=%d -D VSTOREN=vstore%d",
+                                       typeMap[depth], vecSize, vecSize, vecSize, vecSize);
+
+                int vecSizeBytes = vecSize * dst.elemSize1();
+                if ((dst.offset % dst.step) % vecSizeBytes == 0 && dst.step % vecSizeBytes == 0)
+                    buildOptions += " -D DST_ALIGNED";
+                if ((src.offset % src.step) % vecSizeBytes == 0 && src.step % vecSizeBytes == 0)
+                    buildOptions += " -D SRC_ALIGNED";
+
+                args.push_back( make_pair(sizeof(cl_int), (void *)&max_index));
+            }
+
+            args.push_back( make_pair(sizeof(cl_int), (void *)&dst.rows));
+            args.push_back( make_pair(sizeof(cl_int), (void *)&cols));
+
+            size_t localThreads[3] = { 16, 16, 1 };
+            size_t globalThreads[3] = { cols, dst.rows, 1 };
+
             openCLExecuteKernel(src.clCxt, &imgproc_threshold, "threshold", globalThreads, localThreads, args,
                                 -1, -1, buildOptions.c_str());
         }

diff --git a/modules/ocl/src/opencl/imgproc_threshold.cl b/modules/ocl/src/opencl/imgproc_threshold.cl
index 81f2a74..6b847c8 100644 (file)
--- a/modules/ocl/src/opencl/imgproc_threshold.cl
+++ b/modules/ocl/src/opencl/imgproc_threshold.cl
@@ -51,9 +51,63 @@
 #endif
 #endif
 
+#ifdef VECTORIZED
+
+__kernel void threshold(__global const T * restrict src, int src_offset, int src_step,
+                        __global T * dst, int dst_offset, int dst_step,
+                        T thresh, T max_val, int max_index, int rows, int cols)
+{
+    int gx = get_global_id(0);
+    int gy = get_global_id(1);
+
+    if (gx < cols && gy < rows)
+    {
+        gx *= VECSIZE;
+        int src_index = mad24(gy, src_step, src_offset + gx);
+        int dst_index = mad24(gy, dst_step, dst_offset + gx);
+
+#ifdef SRC_ALIGNED
+        VT sdata = *((__global VT *)(src + src_index));
+#else
+        VT sdata = VLOADN(0, src + src_index);
+#endif
+        VT vthresh = (VT)(thresh), zero = (VT)(0);
+
+#ifdef THRESH_BINARY
+        VT vecValue = sdata > vthresh ? max_val : zero;
+#elif defined THRESH_BINARY_INV
+        VT vecValue = sdata > vthresh ? zero : max_val;
+#elif defined THRESH_TRUNC
+        VT vecValue = sdata > vthresh ? thresh : sdata;
+#elif defined THRESH_TOZERO
+        VT vecValue = sdata > vthresh ? sdata : zero;
+#elif defined THRESH_TOZERO_INV
+        VT vecValue = sdata > vthresh ? zero : sdata;
+#endif
+
+        if (gx + VECSIZE <= max_index)
+#ifdef DST_ALIGNED
+            *(__global VT*)(dst + dst_index) = vecValue;
+#else
+            VSTOREN(vecValue, 0, dst + dst_index);
+#endif
+        else
+        {
+            T array[VECSIZE];
+            VSTOREN(vecValue, 0, array);
+            #pragma unroll
+            for (int i = 0; i < VECSIZE; ++i)
+                if (gx + i < max_index)
+                    dst[dst_index + i] = array[i];
+        }
+    }
+}
+
+#else
+
 __kernel void threshold(__global const T * restrict src, int src_offset, int src_step,
                         __global T * dst, int dst_offset, int dst_step,
-                        int rows, int cols, T thresh, T max_val)
+                        T thresh, T max_val, int rows, int cols)
 {
     int gx = get_global_id(0);
     int gy = get_global_id(1);
@@ -78,3 +132,5 @@ __kernel void threshold(__global const T * restrict src, int src_offset, int src
 #endif
     }
 }
+
+#endif