Merge pull request #2876 from vbystricky:oclopt_integralsum

modules/imgproc/perf/opencl/perf_imgproc.cpp

@@ -231,7 +231,7 @@ OCL_PERF_TEST_P(IntegralFixture, Integral1, ::testing::Combine(OCL_TEST_SIZES, O
 
     OCL_TEST_CYCLE() cv::integral(src, dst, ddepth);
 
-    SANITY_CHECK(dst, 1e-6, ERROR_RELATIVE);
+    SANITY_CHECK(dst, 2e-6, ERROR_RELATIVE);
 }
 
 OCL_PERF_TEST_P(IntegralFixture, Integral2, ::testing::Combine(OCL_TEST_SIZES, OCL_PERF_ENUM(CV_32S, CV_32F)))
@@ -243,11 +243,11 @@ OCL_PERF_TEST_P(IntegralFixture, Integral2, ::testing::Combine(OCL_TEST_SIZES, O
     checkDeviceMaxMemoryAllocSize(srcSize, ddepth);
 
     UMat src(srcSize, CV_8UC1), sum(srcSize + Size(1, 1), ddepth), sqsum(srcSize + Size(1, 1), CV_32F);
-    declare.in(src, WARMUP_RNG).out(sum).out(sqsum);
+    declare.in(src, WARMUP_RNG).out(sum, sqsum);
 
     OCL_TEST_CYCLE() cv::integral(src, sum, sqsum, ddepth, CV_32F);
 
-    SANITY_CHECK(sum, 1e-6, ERROR_RELATIVE);
+    SANITY_CHECK(sum, 2e-4, ERROR_RELATIVE);
     SANITY_CHECK(sqsum, 5e-5, ERROR_RELATIVE);
 }
 

modules/imgproc/src/sumpixels.cpp

@@ -235,97 +235,87 @@ typedef void (*IntegralFunc)(const uchar* src, size_t srcstep, uchar* sum, size_
 
 #ifdef HAVE_OPENCL
 
-enum { vlen = 4 };
-
 static bool ocl_integral( InputArray _src, OutputArray _sum, int sdepth )
 {
-    if ( _src.type() != CV_8UC1 || _src.step() % vlen != 0 || _src.offset() % vlen != 0  ||
-         !(sdepth == CV_32S || sdepth == CV_32F) )
-        return false;
+    bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
 
-    ocl::Kernel k1("integral_sum_cols", ocl::imgproc::integral_sum_oclsrc,
-                   format("-D sdepth=%d", sdepth));
-    if (k1.empty())
+    if ( (_src.type() != CV_8UC1) ||
+        !(sdepth == CV_32S || sdepth == CV_32F || (doubleSupport && sdepth == CV_64F)))
         return false;
 
-    Size size = _src.size(), t_size = Size(((size.height + vlen - 1) / vlen) * vlen, size.width),
-            ssize(size.width + 1, size.height + 1);
-    _sum.create(ssize, sdepth);
-    UMat src = _src.getUMat(), t_sum(t_size, sdepth), sum = _sum.getUMat();
-    t_sum = t_sum(Range::all(), Range(0, size.height));
+    static const int tileSize = 16;
+
+    String build_opt = format("-D sumT=%s -D LOCAL_SUM_SIZE=%d%s",
+                                ocl::typeToStr(sdepth), tileSize,
+                                doubleSupport ? " -D DOUBLE_SUPPORT" : "");
+
+    ocl::Kernel kcols("integral_sum_cols", ocl::imgproc::integral_sum_oclsrc, build_opt);
+    if (kcols.empty())
+        return false;
 
-    int offset = (int)src.offset / vlen;
-    int vcols = (src.cols + vlen - 1) / vlen;
-    int sum_offset = (int)sum.offset / vlen;
+    UMat src = _src.getUMat();
+    Size src_size = src.size();
+    Size bufsize(((src_size.height + tileSize - 1) / tileSize) * tileSize, ((src_size.width + tileSize - 1) / tileSize) * tileSize);
+    UMat buf(bufsize, sdepth);
+    kcols.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::WriteOnlyNoSize(buf));
+    size_t gt = src.cols, lt = tileSize;
+    if (!kcols.run(1, &gt, &lt, false))
+        return false;
 
-    k1.args(ocl::KernelArg::PtrReadOnly(src), ocl::KernelArg::PtrWriteOnly(t_sum),
-            offset, src.rows, src.cols, (int)src.step, (int)t_sum.step);
-    size_t gt = ((vcols + 1) / 2) * 256, lt = 256;
-    if (!k1.run(1, &gt, &lt, false))
+    ocl::Kernel krows("integral_sum_rows", ocl::imgproc::integral_sum_oclsrc, build_opt);
+    if (krows.empty())
         return false;
 
-    ocl::Kernel k2("integral_sum_rows", ocl::imgproc::integral_sum_oclsrc,
-                   format("-D sdepth=%d", sdepth));
-    k2.args(ocl::KernelArg::PtrReadOnly(t_sum), ocl::KernelArg::PtrWriteOnly(sum),
-            t_sum.rows, t_sum.cols, (int)t_sum.step, (int)sum.step, sum_offset);
+    Size sumsize(src_size.width + 1, src_size.height + 1);
+    _sum.create(sumsize, sdepth);
+    UMat sum = _sum.getUMat();
 
-    size_t gt2 = t_sum.cols  * 32, lt2 = 256;
-    return k2.run(1, &gt2, &lt2, false);
+    krows.args(ocl::KernelArg::ReadOnlyNoSize(buf), ocl::KernelArg::WriteOnly(sum));
+    gt = src.rows;
+    return krows.run(1, &gt, &lt, false);
 }
 
 static bool ocl_integral( InputArray _src, OutputArray _sum, OutputArray _sqsum, int sdepth, int sqdepth )
 {
     bool doubleSupport = ocl::Device::getDefault().doubleFPConfig() > 0;
 
-    if ( _src.type() != CV_8UC1 || _src.step() % vlen != 0 || _src.offset() % vlen != 0 ||
-         (!doubleSupport && (sdepth == CV_64F || sqdepth == CV_64F)) )
-        return false;
-
-    char cvt[40];
-    String opts = format("-D sdepth=%d -D sqdepth=%d -D TYPE=%s -D TYPE4=%s4 -D convert_TYPE4=%s%s",
-                         sdepth, sqdepth, ocl::typeToStr(sqdepth), ocl::typeToStr(sqdepth),
-                         ocl::convertTypeStr(sdepth, sqdepth, 4, cvt),
-                         doubleSupport ? " -D DOUBLE_SUPPORT" : "");
-
-    ocl::Kernel k1("integral_cols", ocl::imgproc::integral_sqrsum_oclsrc, opts);
-    if (k1.empty())
+    if ( _src.type() != CV_8UC1 || (!doubleSupport && (sdepth == CV_64F || sqdepth == CV_64F)) )
         return false;
 
-    Size size = _src.size(), dsize = Size(size.width + 1, size.height + 1),
-            t_size = Size(((size.height + vlen - 1) / vlen) * vlen, size.width);
-    UMat src = _src.getUMat(), t_sum(t_size, sdepth), t_sqsum(t_size, sqdepth);
-    t_sum = t_sum(Range::all(), Range(0, size.height));
-    t_sqsum = t_sqsum(Range::all(), Range(0, size.height));
-
-    _sum.create(dsize, sdepth);
-    _sqsum.create(dsize, sqdepth);
-    UMat sum = _sum.getUMat(), sqsum = _sqsum.getUMat();
+    static const int tileSize = 16;
 
-    int offset = (int)src.offset / vlen;
-    int pre_invalid = src.offset % vlen;
-    int vcols = (pre_invalid + src.cols + vlen - 1) / vlen;
-    int sum_offset = (int)(sum.offset / sum.elemSize());
-    int sqsum_offset = (int)(sqsum.offset / sqsum.elemSize());
+    String build_opt = format("-D SUM_SQUARE -D sumT=%s -D sumSQT=%s -D LOCAL_SUM_SIZE=%d%s",
+                                ocl::typeToStr(sdepth), ocl::typeToStr(sqdepth),
+                                tileSize,
+                                doubleSupport ? " -D DOUBLE_SUPPORT" : "");
 
-    k1.args(ocl::KernelArg::PtrReadOnly(src), ocl::KernelArg::PtrWriteOnly(t_sum),
-            ocl::KernelArg::PtrWriteOnly(t_sqsum), offset, pre_invalid, src.rows,
-            src.cols, (int)src.step, (int)t_sum.step, (int)t_sqsum.step);
+    ocl::Kernel kcols("integral_sum_cols", ocl::imgproc::integral_sum_oclsrc, build_opt);
+    if (kcols.empty())
+        return false;
 
-    size_t gt = ((vcols + 1) / 2) * 256, lt = 256;
-    if (!k1.run(1, &gt, &lt, false))
+    UMat src = _src.getUMat();
+    Size src_size = src.size();
+    Size bufsize(((src_size.height + tileSize - 1) / tileSize) * tileSize, ((src_size.width + tileSize - 1) / tileSize) * tileSize);
+    UMat buf(bufsize, sdepth);
+    UMat buf_sq(bufsize, sqdepth);
+    kcols.args(ocl::KernelArg::ReadOnly(src), ocl::KernelArg::WriteOnlyNoSize(buf), ocl::KernelArg::WriteOnlyNoSize(buf_sq));
+    size_t gt = src.cols, lt = tileSize;
+    if (!kcols.run(1, &gt, &lt, false))
         return false;
 
-    ocl::Kernel k2("integral_rows", ocl::imgproc::integral_sqrsum_oclsrc, opts);
-    if (k2.empty())
+    ocl::Kernel krows("integral_sum_rows", ocl::imgproc::integral_sum_oclsrc, build_opt);
+    if (krows.empty())
         return false;
 
-    k2.args(ocl::KernelArg::PtrReadOnly(t_sum), ocl::KernelArg::PtrReadOnly(t_sqsum),
-            ocl::KernelArg::PtrWriteOnly(sum), ocl::KernelArg::PtrWriteOnly(sqsum),
-            t_sum.rows, t_sum.cols, (int)t_sum.step, (int)t_sqsum.step,
-            (int)sum.step, (int)sqsum.step, sum_offset, sqsum_offset);
+    Size sumsize(src_size.width + 1, src_size.height + 1);
+    _sum.create(sumsize, sdepth);
+    UMat sum = _sum.getUMat();
+    _sqsum.create(sumsize, sqdepth);
+    UMat sum_sq = _sqsum.getUMat();
 
-    size_t gt2 = t_sum.cols  * 32, lt2 = 256;
-    return k2.run(1, &gt2, &lt2, false);
+    krows.args(ocl::KernelArg::ReadOnlyNoSize(buf), ocl::KernelArg::ReadOnlyNoSize(buf_sq), ocl::KernelArg::WriteOnly(sum), ocl::KernelArg::WriteOnlyNoSize(sum_sq));
+    gt = src.rows;
+    return krows.run(1, &gt, &lt, false);
 }
 
 #endif