added area fast mode to ocl::resize

This commit is contained in:
Ilya Lavrenov 2013-11-18 18:46:59 +04:00
parent f70d63e4c9
commit 198cd1a40d
5 changed files with 121 additions and 28 deletions

View File

@ -191,7 +191,7 @@ typedef TestBaseWithParam<resizeAreaParams> resizeAreaFixture;
PERF_TEST_P(resizeAreaFixture, resize,
::testing::Combine(OCL_TYPICAL_MAT_SIZES,
OCL_PERF_ENUM(CV_8UC1, CV_8UC4, CV_32FC1, CV_32FC4),
::testing::Values(0.6, 0.3)))
::testing::Values(0.3, 0.5, 0.6)))
{
const resizeAreaParams params = GetParam();
const Size srcSize = get<0>(params);

View File

@ -318,28 +318,46 @@ namespace cv
ofs_tab[dx] = k;
}
static void computeResizeAreaFastTabs(int * dmap_tab, int * smap_tab, int scale, int dcols, int scol)
{
for (int i = 0; i < dcols; ++i)
dmap_tab[i] = scale * i;
for (int i = 0, size = dcols * scale; i < size; ++i)
smap_tab[i] = std::min(scol - 1, i);
}
static void resize_gpu( const oclMat &src, oclMat &dst, double ifx, double ify, int interpolation)
{
float ifxf = (float)ifx, ifyf = (float)ify;
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 ocn = interpolation == INTER_LINEAR ? dst.oclchannels() : -1;
int depth = interpolation == INTER_LINEAR ? dst.depth() : -1;
int ocn = dst.oclchannels(), depth = dst.depth();
const char * const interMap[] = { "NN", "LN", "CUBIC", "AREA", "LAN4" };
std::string kernelName = std::string("resize") + interMap[interpolation];
const char * const typeMap[] = { "uchar", "char", "ushort", "short", "int", "float", "double" };
const char * const channelMap[] = { "" , "", "2", "4", "4" };
std::string buildOption = format("-D %s -D T=%s%s", interMap[interpolation], typeMap[dst.depth()], channelMap[dst.oclchannels()]);
std::string buildOption = format("-D %s -D T=%s%s", interMap[interpolation], typeMap[depth], channelMap[ocn]);
int wdepth = std::max(src.depth(), CV_32F);
// check if fx, fy is integer and then we have inter area fast mode
int iscale_x = saturate_cast<int>(ifx);
int iscale_y = saturate_cast<int>(ify);
bool is_area_fast = std::abs(ifx - iscale_x) < DBL_EPSILON &&
std::abs(ify - iscale_y) < DBL_EPSILON;
if (is_area_fast)
wdepth = std::max(src.depth(), CV_32S);
if (interpolation != INTER_NEAREST)
{
buildOption += format(" -D WT=%s -D WTV=%s%s -D convertToWTV=convert_%s%s -D convertToT=convert_%s%s%s",
typeMap[wdepth], typeMap[wdepth], channelMap[dst.oclchannels()],
typeMap[wdepth], channelMap[dst.oclchannels()],
typeMap[src.depth()], channelMap[dst.oclchannels()], src.depth() <= CV_32S ? "_sat_rte" : "");
typeMap[wdepth], typeMap[wdepth], channelMap[ocn],
typeMap[wdepth], channelMap[ocn],
typeMap[src.depth()], channelMap[ocn], src.depth() <= CV_32S ? "_sat_rte" : "");
}
size_t blkSizeX = 16, blkSizeY = 16;
@ -352,26 +370,48 @@ namespace cv
else
glbSizeX = dst.cols;
static oclMat alphaOcl, mapOcl, tabofsOcl;
oclMat alphaOcl, mapOcl, tabofsOcl;
if (interpolation == INTER_AREA)
{
Size ssize = src.size(), dsize = dst.size();
int xytab_size = (ssize.width + ssize.height) << 1;
int tabofs_size = dsize.height + dsize.width + 2;
if (is_area_fast)
{
kernelName += "_FAST";
int wdepth2 = std::max(CV_32F, src.depth());
buildOption += format(" -D WT2V=%s%s -D convertToWT2V=convert_%s%s -D AREA_FAST -D XSCALE=%d -D YSCALE=%d -D SCALE=%f",
typeMap[wdepth2], channelMap[ocn], typeMap[wdepth2], channelMap[ocn],
iscale_x, iscale_y, 1.0f / (iscale_x * iscale_y));
AutoBuffer<int> _xymap_tab(xytab_size), _xyofs_tab(tabofs_size);
AutoBuffer<float> _xyalpha_tab(xytab_size);
int * xmap_tab = _xymap_tab, * ymap_tab = _xymap_tab + (ssize.width << 1);
float * xalpha_tab = _xyalpha_tab, * yalpha_tab = _xyalpha_tab + (ssize.width << 1);
int * xofs_tab = _xyofs_tab, * yofs_tab = _xyofs_tab + dsize.width + 1;
int smap_tab_size = dst.cols * iscale_x + dst.rows * iscale_y;
AutoBuffer<int> dmap_tab(dst.cols + dst.rows), smap_tab(smap_tab_size);
int * dxmap_tab = dmap_tab, * dymap_tab = dxmap_tab + dst.cols;
int * sxmap_tab = smap_tab, * symap_tab = smap_tab + dst.cols * iscale_y;
computeResizeAreaTabs(ssize.width, dsize.width, ifx, xmap_tab, xalpha_tab, xofs_tab);
computeResizeAreaTabs(ssize.height, dsize.height, ify, ymap_tab, yalpha_tab, yofs_tab);
computeResizeAreaFastTabs(dxmap_tab, sxmap_tab, iscale_x, dst.cols, src.cols);
computeResizeAreaFastTabs(dymap_tab, symap_tab, iscale_y, dst.rows, src.rows);
// loading precomputed arrays to GPU
alphaOcl = oclMat(1, xytab_size, CV_32FC1, (void *)_xyalpha_tab);
mapOcl = oclMat(1, xytab_size, CV_32SC1, (void *)_xymap_tab);
tabofsOcl = oclMat(1, tabofs_size, CV_32SC1, (void *)_xyofs_tab);
tabofsOcl = oclMat(1, dst.cols + dst.rows, CV_32SC1, (void *)dmap_tab);
mapOcl = oclMat(1, smap_tab_size, CV_32SC1, (void *)smap_tab);
}
else
{
Size ssize = src.size(), dsize = dst.size();
int xytab_size = (ssize.width + ssize.height) << 1;
int tabofs_size = dsize.height + dsize.width + 2;
AutoBuffer<int> _xymap_tab(xytab_size), _xyofs_tab(tabofs_size);
AutoBuffer<float> _xyalpha_tab(xytab_size);
int * xmap_tab = _xymap_tab, * ymap_tab = _xymap_tab + (ssize.width << 1);
float * xalpha_tab = _xyalpha_tab, * yalpha_tab = _xyalpha_tab + (ssize.width << 1);
int * xofs_tab = _xyofs_tab, * yofs_tab = _xyofs_tab + dsize.width + 1;
computeResizeAreaTabs(ssize.width, dsize.width, ifx, xmap_tab, xalpha_tab, xofs_tab);
computeResizeAreaTabs(ssize.height, dsize.height, ify, ymap_tab, yalpha_tab, yofs_tab);
// loading precomputed arrays to GPU
alphaOcl = oclMat(1, xytab_size, CV_32FC1, (void *)_xyalpha_tab);
mapOcl = oclMat(1, xytab_size, CV_32SC1, (void *)_xymap_tab);
tabofsOcl = oclMat(1, tabofs_size, CV_32SC1, (void *)_xyofs_tab);
}
}
size_t globalThreads[3] = { glbSizeX, dst.rows, 1 };
@ -400,12 +440,18 @@ namespace cv
args.push_back( make_pair(sizeof(cl_float), (void *)&ifyf));
}
if (interpolation == INTER_AREA)
{
// precomputed tabs
if (!tabofsOcl.empty())
args.push_back( make_pair(sizeof(cl_mem), (void *)&tabofsOcl.data));
if (!mapOcl.empty())
args.push_back( make_pair(sizeof(cl_mem), (void *)&mapOcl.data));
if (!alphaOcl.empty())
args.push_back( make_pair(sizeof(cl_mem), (void *)&alphaOcl.data));
}
ocn = interpolation == INTER_LINEAR ? ocn : -1;
depth = interpolation == INTER_LINEAR ? depth : -1;
openCLExecuteKernel(src.clCxt, &imgproc_resize, kernelName, globalThreads, localThreads, args,
ocn, depth, buildOption.c_str());

View File

@ -315,7 +315,44 @@ __kernel void resizeNN(__global T * dst, __global T * src,
}
}
#elif AREA
#elif defined AREA
#ifdef AREA_FAST
__kernel void resizeAREA_FAST(__global T * dst, __global T * src,
int dst_offset, int src_offset, int dst_step, int src_step,
int src_cols, int src_rows, int dst_cols, int dst_rows, WT ifx, WT ify,
__global const int * dmap_tab, __global const int * smap_tab)
{
int dx = get_global_id(0);
int dy = get_global_id(1);
if (dx < dst_cols && dy < dst_rows)
{
int dst_index = mad24(dy, dst_step, dst_offset + dx);
__global const int * xmap_tab = dmap_tab;
__global const int * ymap_tab = dmap_tab + dst_cols;
__global const int * sxmap_tab = smap_tab;
__global const int * symap_tab = smap_tab + XSCALE * dst_cols;
int sx = xmap_tab[dx], sy = ymap_tab[dy];
WTV sum = (WTV)(0);
#pragma unroll
for (int y = 0; y < YSCALE; ++y)
{
int src_index = mad24(symap_tab[y + sy], src_step, src_offset);
#pragma unroll
for (int x = 0; x < XSCALE; ++x)
sum += convertToWTV(src[src_index + sxmap_tab[sx + x]]);
}
dst[dst_index] = convertToT(convertToWT2V(sum) * (WT2V)(SCALE));
}
}
#else
__kernel void resizeAREA(__global T * dst, __global T * src,
int dst_offset, int src_offset, int dst_step, int src_step,
@ -364,3 +401,5 @@ __kernel void resizeAREA(__global T * dst, __global T * src,
}
#endif
#endif

View File

@ -231,7 +231,7 @@ double checkRectSimilarity(Size sz, std::vector<Rect>& ob1, std::vector<Rect>& o
return final_test_result;
}
void showDiff(const Mat& gold, const Mat& actual, double eps, bool alwaysShow)
void showDiff(const Mat& src, const Mat& gold, const Mat& actual, double eps, bool alwaysShow)
{
Mat diff, diff_thresh;
absdiff(gold, actual, diff);
@ -240,10 +240,18 @@ void showDiff(const Mat& gold, const Mat& actual, double eps, bool alwaysShow)
if (alwaysShow || cv::countNonZero(diff_thresh.reshape(1)) > 0)
{
#if 0
std::cout << "Src: " << std::endl << src << std::endl;
std::cout << "Reference: " << std::endl << gold << std::endl;
std::cout << "OpenCL: " << std::endl << actual << std::endl;
#endif
namedWindow("src", WINDOW_NORMAL);
namedWindow("gold", WINDOW_NORMAL);
namedWindow("actual", WINDOW_NORMAL);
namedWindow("diff", WINDOW_NORMAL);
imshow("src", src);
imshow("gold", gold);
imshow("actual", actual);
imshow("diff", diff);

View File

@ -52,7 +52,7 @@ extern int LOOP_TIMES;
namespace cvtest {
void showDiff(const Mat& gold, const Mat& actual, double eps, bool alwaysShow = false);
void showDiff(const Mat& src, const Mat& gold, const Mat& actual, double eps, bool alwaysShow = false);
cv::ocl::oclMat createMat_ocl(cv::RNG& rng, Size size, int type, bool useRoi);
cv::ocl::oclMat loadMat_ocl(cv::RNG& rng, const Mat& m, bool useRoi);