Merge pull request #2687 from ilya-lavrenov:tapi_norm_inf_234cn

modules/core/src/opencl/reduce.cl

@@ -50,6 +50,36 @@
 #endif
 #endif
 
+#if defined OP_NORM_INF_MASK || defined OP_MIN_MAX_LOC || defined OP_MIN_MAX_LOC_MASK
+
+#ifdef DEPTH_0
+#define MIN_VAL 0
+#define MAX_VAL 255
+#elif defined DEPTH_1
+#define MIN_VAL -128
+#define MAX_VAL 127
+#elif defined DEPTH_2
+#define MIN_VAL 0
+#define MAX_VAL 65535
+#elif defined DEPTH_3
+#define MIN_VAL -32768
+#define MAX_VAL 32767
+#elif defined DEPTH_4
+#define MIN_VAL INT_MIN
+#define MAX_VAL INT_MAX
+#elif defined DEPTH_5
+#define MIN_VAL (-FLT_MAX)
+#define MAX_VAL FLT_MAX
+#elif defined DEPTH_6
+#define MIN_VAL (-DBL_MAX)
+#define MAX_VAL DBL_MAX
+#endif
+
+#define dstT srcT
+#define dstT1 srcT1
+
+#endif // min/max stuff
+
 #define noconvert
 
 #if cn != 3
@@ -145,40 +175,31 @@
 #define CALC_RESULT \
     storepix(localmem[0], dstptr + dstTSIZE * gid)
 
-// minMaxLoc stuff
-#elif defined OP_MIN_MAX_LOC || defined OP_MIN_MAX_LOC_MASK
+// norm (NORM_INF) with cn > 1 and mask
+#elif defined OP_NORM_INF_MASK
 
-#ifdef DEPTH_0
-#define srcT uchar
-#define MIN_VAL 0
-#define MAX_VAL 255
-#elif defined DEPTH_1
-#define srcT char
-#define MIN_VAL -128
-#define MAX_VAL 127
-#elif defined DEPTH_2
-#define srcT ushort
-#define MIN_VAL 0
-#define MAX_VAL 65535
-#elif defined DEPTH_3
-#define srcT short
-#define MIN_VAL -32768
-#define MAX_VAL 32767
-#elif defined DEPTH_4
-#define srcT int
-#define MIN_VAL INT_MIN
-#define MAX_VAL INT_MAX
-#elif defined DEPTH_5
-#define srcT float
-#define MIN_VAL (-FLT_MAX)
-#define MAX_VAL FLT_MAX
-#elif defined DEPTH_6
-#define srcT double
-#define MIN_VAL (-DBL_MAX)
-#define MAX_VAL DBL_MAX
-#endif
+#define DECLARE_LOCAL_MEM \
+    __local srcT localmem_max[WGS2_ALIGNED]
+#define DEFINE_ACCUMULATOR \
+    srcT maxval = MIN_VAL, temp
+#define REDUCE_GLOBAL \
+    int mask_index = mad24(id / cols, mask_step, mask_offset + (id % cols)); \
+    if (mask[mask_index]) \
+    { \
+        temp = loadpix(srcptr + src_index); \
+        maxval = max(maxval, (srcT)(temp >= 0 ? temp : -temp)); \
+    }
+#define SET_LOCAL_1 \
+    localmem_max[lid] = maxval
+#define REDUCE_LOCAL_1 \
+    localmem_max[lid - WGS2_ALIGNED] = max(maxval, localmem_max[lid - WGS2_ALIGNED])
+#define REDUCE_LOCAL_2 \
+    localmem_max[lid] = max(localmem_max[lid], localmem_max[lid2])
+#define CALC_RESULT \
+    storepix(localmem_max[0], dstptr + dstTSIZE * gid)
 
-#define dstT srcT
+// minMaxLoc stuff
+#elif defined OP_MIN_MAX_LOC || defined OP_MIN_MAX_LOC_MASK
 
 #define DECLARE_LOCAL_MEM \
     __local srcT localmem_min[WGS2_ALIGNED]; \

modules/core/src/stat.cpp

@@ -41,10 +41,11 @@
 //M*/
 
 #include "precomp.hpp"
-#include "opencl_kernels.hpp"
 #include <climits>
 #include <limits>
 
+#include "opencl_kernels.hpp"
+
 namespace cv
 {
 
@@ -1245,7 +1246,7 @@ void getMinMaxRes(const Mat &minv, const Mat &maxv, const Mat &minl, const Mat &
     T min = std::numeric_limits<T>::max();
     T max = std::numeric_limits<T>::min() > 0 ? -std::numeric_limits<T>::max() : std::numeric_limits<T>::min();
     int minloc = INT_MAX, maxloc = INT_MAX;
-    for( int i = 0; i < groupnum; i++)
+    for (int i = 0; i < groupnum; i++)
     {
         T current_min = minv.at<T>(0,i);
         T current_max = maxv.at<T>(0,i);
@@ -1262,16 +1263,16 @@ void getMinMaxRes(const Mat &minv, const Mat &maxv, const Mat &minl, const Mat &
         }
     }
     bool zero_mask = (maxloc == INT_MAX) || (minloc == INT_MAX);
-    if(minVal)
+    if (minVal)
         *minVal = zero_mask ? 0 : (double)min;
-    if(maxVal)
+    if (maxVal)
         *maxVal = zero_mask ? 0 : (double)max;
-    if(minLoc)
+    if (minLoc)
     {
         minLoc[0] = zero_mask ? -1 : minloc/cols;
         minLoc[1] = zero_mask ? -1 : minloc%cols;
     }
-    if(maxLoc)
+    if (maxLoc)
     {
         maxLoc[0] = zero_mask ? -1 : maxloc/cols;
         maxLoc[1] = zero_mask ? -1 : maxloc%cols;
@@ -1300,8 +1301,9 @@ static bool ocl_minMaxIdx( InputArray _src, double* minVal, double* maxVal, int*
         wgs2_aligned <<= 1;
     wgs2_aligned >>= 1;
 
-    String opts = format("-D DEPTH_%d -D OP_MIN_MAX_LOC%s -D WGS=%d -D WGS2_ALIGNED=%d%s",
-        depth, _mask.empty() ? "" : "_MASK", (int)wgs, wgs2_aligned, doubleSupport ? " -D DOUBLE_SUPPORT" : "");
+    String opts = format("-D DEPTH_%d -D srcT=%s -D OP_MIN_MAX_LOC%s -D WGS=%d -D WGS2_ALIGNED=%d%s",
+                         depth, ocl::typeToStr(depth), _mask.empty() ? "" : "_MASK", (int)wgs,
+                         wgs2_aligned, doubleSupport ? " -D DOUBLE_SUPPORT" : "");
 
     ocl::Kernel k("reduce", ocl::core::reduce_oclsrc, opts);
     if (k.empty())
@@ -1980,39 +1982,70 @@ static bool ocl_norm( InputArray _src, int normType, InputArray _mask, double &
             haveMask = _mask.kind() != _InputArray::NONE;
 
     if ( !(normType == NORM_INF || normType == NORM_L1 || normType == NORM_L2 || normType == NORM_L2SQR) ||
-         (!doubleSupport && depth == CV_64F) || (normType == NORM_INF && haveMask && cn != 1))
+         (!doubleSupport && depth == CV_64F))
         return false;
 
     UMat src = _src.getUMat();
 
     if (normType == NORM_INF)
     {
-        UMat abssrc;
+        if (cn == 1 || !haveMask)
+        {
+            UMat abssrc;
+
+            if (depth != CV_8U && depth != CV_16U)
+            {
+                int wdepth = std::max(CV_32S, depth);
+                char cvt[50];
+
+                ocl::Kernel kabs("KF", ocl::core::arithm_oclsrc,
+                                 format("-D UNARY_OP -D OP_ABS_NOSAT -D dstT=%s -D srcT1=%s -D convertToDT=%s%s",
+                                        ocl::typeToStr(wdepth), ocl::typeToStr(depth),
+                                        ocl::convertTypeStr(depth, wdepth, 1, cvt),
+                                        doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
+                if (kabs.empty())
+                    return false;
+
+                abssrc.create(src.size(), CV_MAKE_TYPE(wdepth, cn));
+                kabs.args(ocl::KernelArg::ReadOnlyNoSize(src), ocl::KernelArg::WriteOnly(abssrc, cn));
+
+                size_t globalsize[2] = { src.cols * cn, src.rows };
+                if (!kabs.run(2, globalsize, NULL, false))
+                    return false;
+            }
+            else
+                abssrc = src;
 
-        if (depth != CV_8U && depth != CV_16U)
+            cv::minMaxIdx(haveMask ? abssrc : abssrc.reshape(1), NULL, &result, NULL, NULL, _mask);
+        }
+        else
         {
-            int wdepth = std::max(CV_32S, depth);
-            char cvt[50];
-
-            ocl::Kernel kabs("KF", ocl::core::arithm_oclsrc,
-                             format("-D UNARY_OP -D OP_ABS_NOSAT -D dstT=%s -D srcT1=%s -D convertToDT=%s%s",
-                                    ocl::typeToStr(wdepth), ocl::typeToStr(depth),
-                                    ocl::convertTypeStr(depth, wdepth, 1, cvt),
-                                    doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
-            if (kabs.empty())
+            int dbsize = ocl::Device::getDefault().maxComputeUnits();
+            size_t wgs = ocl::Device::getDefault().maxWorkGroupSize();
+
+            int wgs2_aligned = 1;
+            while (wgs2_aligned < (int)wgs)
+                wgs2_aligned <<= 1;
+            wgs2_aligned >>= 1;
+
+            ocl::Kernel k("reduce", ocl::core::reduce_oclsrc,
+                          format("-D OP_NORM_INF_MASK -D HAVE_MASK -D DEPTH_%d"
+                                 " -D srcT=%s -D srcT1=%s -D WGS=%d -D cn=%d -D WGS2_ALIGNED=%d%s",
+                                 depth, ocl::typeToStr(type), ocl::typeToStr(depth),
+                                 wgs, cn, wgs2_aligned, doubleSupport ? " -D DOUBLE_SUPPORT" : ""));
+            if (k.empty())
                 return false;
 
-            abssrc.create(src.size(), CV_MAKE_TYPE(wdepth, cn));
-            kabs.args(ocl::KernelArg::ReadOnlyNoSize(src), ocl::KernelArg::WriteOnly(abssrc, cn));
+            UMat db(1, dbsize, type), mask = _mask.getUMat();
+            k.args(ocl::KernelArg::ReadOnlyNoSize(src), src.cols, (int)src.total(),
+                   dbsize, ocl::KernelArg::PtrWriteOnly(db), ocl::KernelArg::ReadOnlyNoSize(mask));
 
-            size_t globalsize[2] = { src.cols * cn, src.rows };
-            if (!kabs.run(2, globalsize, NULL, false))
+            size_t globalsize = dbsize * wgs;
+            if (!k.run(1, &globalsize, &wgs, true))
                 return false;
-        }
-        else
-            abssrc = src;
 
-        cv::minMaxIdx(haveMask ? abssrc : abssrc.reshape(1), NULL, &result, NULL, NULL, _mask);
+            minMaxIdx(db.getMat(ACCESS_READ), NULL, &result, NULL, NULL, noArray());
+        }
     }
     else if (normType == NORM_L1 || normType == NORM_L2 || normType == NORM_L2SQR)
     {

modules/core/test/ocl/test_arithm.cpp

@@ -1149,7 +1149,7 @@ OCL_TEST_P(MinMaxIdx, Mat)
         int p1[2], p2[2], up1[2], up2[2];
         double minv, maxv, uminv, umaxv;
 
-        if(src1_roi.channels() > 1)
+        if (cn > 1)
         {
             OCL_OFF(cv::minMaxIdx(src2_roi, &minv, &maxv) );
             OCL_ON(cv::minMaxIdx(usrc2_roi, &uminv, &umaxv));
@@ -1164,7 +1164,8 @@ OCL_TEST_P(MinMaxIdx, Mat)
 
             EXPECT_DOUBLE_EQ(minv, uminv);
             EXPECT_DOUBLE_EQ(maxv, umaxv);
-            for( int i = 0; i < 2; i++)
+
+            for (int i = 0; i < 2; i++)
             {
                 EXPECT_EQ(p1[i], up1[i]);
                 EXPECT_EQ(p2[i], up2[i]);