generic-library/opencv
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Improve ocl cvt_color performance for the following conversions: RGB<->BGR, RGB->Gray, RGB<->XYZ, RGB<->YCrCb, RGB<->YUV, and mRGBA<->RGBA.

Browse Source 
The improvement was done basically by processing more than 1 pixel by each work-item and using vector's operations.
new performance tests were added
This commit is contained in:
krodyush
2013-12-19 22:56:46 +04:00
parent 26f5d2d639
commit 27c1bd2762
3 changed files with 955 additions and 177 deletions
Download Patch File Download Diff File Expand all files Collapse all files

97
modules/ocl/perf/perf_color.cpp
View File

@@ -57,9 +57,39 @@ CV_ENUM(ConversionTypes, CV_RGB2GRAY, CV_RGB2BGR, CV_RGB2YUV, CV_YUV2RGB, CV_RGB
CV_HLS2RGB, CV_BGR5652BGR, CV_BGR2BGR565, CV_RGBA2mRGBA, CV_mRGBA2RGBA, CV_YUV2RGB_NV12)
typedef tuple<Size, tuple<ConversionTypes, int, int> > cvtColorParams;
typedef TestBaseWithParam<cvtColorParams> cvtColorFixture;
typedef TestBaseWithParam<cvtColorParams> cvtColorU8Fixture;
typedef TestBaseWithParam<cvtColorParams> cvtColorF32Fixture;
typedef TestBaseWithParam<cvtColorParams> cvtColorU16Fixture;
PERF_TEST_P(cvtColorFixture, cvtColor, testing::Combine(
#define RUN_CVT_PERF_TEST \
cvtColorParams params = GetParam();\
const Size srcSize = get<0>(params);\
const tuple<int, int, int> conversionParams = get<1>(params);\
const int code = get<0>(conversionParams), scn = get<1>(conversionParams),\
dcn = get<2>(conversionParams);\
\
Mat src(srcSize, CV_8UC(scn)), dst(srcSize, CV_8UC(scn));\
declare.in(src, WARMUP_RNG).out(dst);\
\
if (RUN_OCL_IMPL)\
{\
ocl::oclMat oclSrc(src), oclDst(src.size(), dst.type());\
\
OCL_TEST_CYCLE() ocl::cvtColor(oclSrc, oclDst, code, dcn);\
oclDst.download(dst);\
\
SANITY_CHECK(dst, 1);\
}\
else if (RUN_PLAIN_IMPL)\
{\
TEST_CYCLE() cv::cvtColor(src, dst, code, dcn);\
\
SANITY_CHECK(dst);\
}\
else\
OCL_PERF_ELSE\
PERF_TEST_P(cvtColorU8Fixture, cvtColor, testing::Combine(
testing::Values(Size(1000, 1002), Size(2000, 2004), Size(4000, 4008)),
testing::Values(
make_tuple(ConversionTypes(CV_RGB2GRAY), 3, 1),
@@ -81,30 +111,41 @@ PERF_TEST_P(cvtColorFixture, cvtColor, testing::Combine(
make_tuple(ConversionTypes(CV_YUV2RGB_NV12), 1, 3)
)))
{
cvtColorParams params = GetParam();
const Size srcSize = get<0>(params);
const tuple<int, int, int> conversionParams = get<1>(params);
const int code = get<0>(conversionParams), scn = get<1>(conversionParams),
dcn = get<2>(conversionParams);
Mat src(srcSize, CV_8UC(scn)), dst(srcSize, CV_8UC(scn));
declare.in(src, WARMUP_RNG).out(dst);
if (RUN_OCL_IMPL)
{
ocl::oclMat oclSrc(src), oclDst(src.size(), dst.type());
OCL_TEST_CYCLE() ocl::cvtColor(oclSrc, oclDst, code, dcn);
oclDst.download(dst);
SANITY_CHECK(dst, 1);
}
else if (RUN_PLAIN_IMPL)
{
TEST_CYCLE() cv::cvtColor(src, dst, code, dcn);
SANITY_CHECK(dst);
}
else
OCL_PERF_ELSE
RUN_CVT_PERF_TEST
}
PERF_TEST_P(cvtColorF32Fixture, cvtColor, testing::Combine(
testing::Values(Size(1000, 1002), Size(2000, 2004), Size(4000, 4008)),
testing::Values(
make_tuple(ConversionTypes(CV_RGB2GRAY), 3, 1),
make_tuple(ConversionTypes(CV_RGB2BGR), 3, 3),
make_tuple(ConversionTypes(CV_RGB2YUV), 3, 3),
make_tuple(ConversionTypes(CV_YUV2RGB), 3, 3),
make_tuple(ConversionTypes(CV_RGB2YCrCb), 3, 3),
make_tuple(ConversionTypes(CV_YCrCb2RGB), 3, 3),
make_tuple(ConversionTypes(CV_RGB2XYZ), 3, 3),
make_tuple(ConversionTypes(CV_XYZ2RGB), 3, 3),
make_tuple(ConversionTypes(CV_RGB2HSV), 3, 3),
make_tuple(ConversionTypes(CV_HSV2RGB), 3, 3),
make_tuple(ConversionTypes(CV_RGB2HLS), 3, 3),
make_tuple(ConversionTypes(CV_HLS2RGB), 3, 3)
)))
{
RUN_CVT_PERF_TEST
}
PERF_TEST_P(cvtColorU16Fixture, cvtColor, testing::Combine(
testing::Values(Size(1000, 1002), Size(2000, 2004), Size(4000, 4008)),
testing::Values(
make_tuple(ConversionTypes(CV_RGB2GRAY), 3, 1),
make_tuple(ConversionTypes(CV_RGB2BGR), 3, 3),
make_tuple(ConversionTypes(CV_RGB2YUV), 3, 3),
make_tuple(ConversionTypes(CV_YUV2RGB), 3, 3),
make_tuple(ConversionTypes(CV_RGB2YCrCb), 3, 3),
make_tuple(ConversionTypes(CV_YCrCb2RGB), 3, 3),
make_tuple(ConversionTypes(CV_RGB2XYZ), 3, 3),
make_tuple(ConversionTypes(CV_XYZ2RGB), 3, 3)
)))
{
RUN_CVT_PERF_TEST
}

186
modules/ocl/src/color.cpp
View File

@@ -56,8 +56,19 @@ static void fromRGB_caller(const oclMat &src, oclMat &dst, int bidx, const std::
{
int src_offset = src.offset / src.elemSize1(), src_step = src.step1();
int dst_offset = dst.offset / dst.elemSize1(), dst_step = dst.step1();
int pixels_per_work_item = 1;
std::string build_options = format("-D DEPTH_%d", src.depth());
if (Context::getContext()->supportsFeature(FEATURE_CL_INTEL_DEVICE))
{
if ((src.cols % 4 == 0) && (src.depth() == CV_8U))
pixels_per_work_item = 4;
else if (src.cols % 2 == 0)
pixels_per_work_item = 2;
else
pixels_per_work_item = 1;
}
std::string build_options = format("-D DEPTH_%d -D scn=%d -D bidx=%d -D pixels_per_work_item=%d", src.depth(), src.oclchannels(), bidx, pixels_per_work_item);
if (!additionalOptions.empty())
build_options += additionalOptions;
@@ -66,7 +77,6 @@ static void fromRGB_caller(const oclMat &src, oclMat &dst, int bidx, const std::
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.rows));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_step));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_step));
args.push_back( make_pair( sizeof(cl_int) , (void *)&bidx));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset ));
@@ -77,6 +87,73 @@ static void fromRGB_caller(const oclMat &src, oclMat &dst, int bidx, const std::
if (!data2.empty())
args.push_back( make_pair( sizeof(cl_mem) , (void *)&data2.data ));
size_t gt[3] = { dst.cols/pixels_per_work_item, dst.rows, 1 };
#ifdef ANDROID
size_t lt[3] = { 16, 10, 1 };
#else
size_t lt[3] = { 16, 16, 1 };
#endif
openCLExecuteKernel(src.clCxt, &cvt_color, kernelName.c_str(), gt, lt, args, -1, -1, build_options.c_str());
}
static void toHSV_caller(const oclMat &src, oclMat &dst, int bidx, const std::string & kernelName,
const std::string & additionalOptions = std::string(),
const oclMat & data1 = oclMat(), const oclMat & data2 = oclMat())
{
int src_offset = src.offset / src.elemSize1(), src_step = src.step1();
int dst_offset = dst.offset / dst.elemSize1(), dst_step = dst.step1();
std::string build_options = format("-D DEPTH_%d -D scn=%d -D bidx=%d", src.depth(), src.oclchannels(), bidx);
if (!additionalOptions.empty())
build_options += additionalOptions;
vector<pair<size_t , const void *> > args;
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.cols));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.rows));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_step));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_step));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_offset ));
if (!data1.empty())
args.push_back( make_pair( sizeof(cl_mem) , (void *)&data1.data ));
if (!data2.empty())
args.push_back( make_pair( sizeof(cl_mem) , (void *)&data2.data ));
size_t gt[3] = { dst.cols, dst.rows, 1 };
#ifdef ANDROID
size_t lt[3] = { 16, 10, 1 };
#else
size_t lt[3] = { 16, 16, 1 };
#endif
openCLExecuteKernel(src.clCxt, &cvt_color, kernelName.c_str(), gt, lt, args, -1, -1, build_options.c_str());
}
static void fromGray_caller(const oclMat &src, oclMat &dst, int bidx, const std::string & kernelName,
const std::string & additionalOptions = std::string(), const oclMat & data = oclMat())
{
std::string build_options = format("-D DEPTH_%d -D dcn=%d -D bidx=%d", src.depth(), dst.channels(), bidx);
if (!additionalOptions.empty())
build_options += additionalOptions;
int src_offset = src.offset / src.elemSize1(), src_step = src.step1();
int dst_offset = dst.offset / dst.elemSize1(), dst_step = dst.step1();
vector<pair<size_t , const void *> > args;
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.cols));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.rows));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_step));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_step));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_offset ));
if (!data.empty())
args.push_back( make_pair( sizeof(cl_mem) , (void *)&data.data ));
size_t gt[3] = { dst.cols, dst.rows, 1 };
#ifdef ANDROID
size_t lt[3] = { 16, 10, 1 };
@@ -89,7 +166,50 @@ static void fromRGB_caller(const oclMat &src, oclMat &dst, int bidx, const std::
static void toRGB_caller(const oclMat &src, oclMat &dst, int bidx, const std::string & kernelName,
const std::string & additionalOptions = std::string(), const oclMat & data = oclMat())
{
std::string build_options = format("-D DEPTH_%d -D dcn=%d", src.depth(), dst.channels());
int src_offset = src.offset / src.elemSize1(), src_step = src.step1();
int dst_offset = dst.offset / dst.elemSize1(), dst_step = dst.step1();
int pixels_per_work_item = 1;
if (Context::getContext()->supportsFeature(FEATURE_CL_INTEL_DEVICE))
{
if ((src.cols % 4 == 0) && (src.depth() == CV_8U))
pixels_per_work_item = 4;
else if (src.cols % 2 == 0)
pixels_per_work_item = 2;
else
pixels_per_work_item = 1;
}
std::string build_options = format("-D DEPTH_%d -D dcn=%d -D bidx=%d -D pixels_per_work_item=%d", src.depth(), dst.channels(), bidx, pixels_per_work_item);
if (!additionalOptions.empty())
build_options += additionalOptions;
vector<pair<size_t , const void *> > args;
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.cols));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.rows));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_step));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_step));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_offset ));
if (!data.empty())
args.push_back( make_pair( sizeof(cl_mem) , (void *)&data.data ));
size_t gt[3] = { dst.cols/pixels_per_work_item, dst.rows, 1 };
#ifdef ANDROID
size_t lt[3] = { 16, 10, 1 };
#else
size_t lt[3] = { 16, 16, 1 };
#endif
openCLExecuteKernel(src.clCxt, &cvt_color, kernelName.c_str(), gt, lt, args, -1, -1, build_options.c_str());
}
static void toRGB_NV12_caller(const oclMat &src, oclMat &dst, int bidx, const std::string & kernelName,
const std::string & additionalOptions = std::string(), const oclMat & data = oclMat())
{
std::string build_options = format("-D DEPTH_%d -D dcn=%d -D bidx=%d", src.depth(), dst.channels(), bidx);
if (!additionalOptions.empty())
build_options += additionalOptions;
@@ -101,7 +221,6 @@ static void toRGB_caller(const oclMat &src, oclMat &dst, int bidx, const std::st
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.rows));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_step));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_step));
args.push_back( make_pair( sizeof(cl_int) , (void *)&bidx));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset ));
@@ -119,10 +238,13 @@ static void toRGB_caller(const oclMat &src, oclMat &dst, int bidx, const std::st
openCLExecuteKernel(src.clCxt, &cvt_color, kernelName.c_str(), gt, lt, args, -1, -1, build_options.c_str());
}
static void RGB_caller(const oclMat &src, oclMat &dst, bool reverse)
static void fromHSV_caller(const oclMat &src, oclMat &dst, int bidx, const std::string & kernelName,
const std::string & additionalOptions = std::string(), const oclMat & data = oclMat())
{
std::string build_options = format("-D DEPTH_%d -D dcn=%d -D scn=%d -D %s", src.depth(),
dst.channels(), src.channels(), reverse ? "REVERSE" : "ORDER");
std::string build_options = format("-D DEPTH_%d -D dcn=%d -D bidx=%d", src.depth(), dst.channels(), bidx);
if (!additionalOptions.empty())
build_options += additionalOptions;
int src_offset = src.offset / src.elemSize1(), src_step = src.step1();
int dst_offset = dst.offset / dst.elemSize1(), dst_step = dst.step1();
@@ -136,6 +258,36 @@ static void RGB_caller(const oclMat &src, oclMat &dst, bool reverse)
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_offset ));
if (!data.empty())
args.push_back( make_pair( sizeof(cl_mem) , (void *)&data.data ));
size_t gt[3] = { dst.cols, dst.rows, 1 };
#ifdef ANDROID
size_t lt[3] = { 16, 10, 1 };
#else
size_t lt[3] = { 16, 16, 1 };
#endif
openCLExecuteKernel(src.clCxt, &cvt_color, kernelName.c_str(), gt, lt, args, -1, -1, build_options.c_str());
}
static void RGB_caller(const oclMat &src, oclMat &dst, bool reverse)
{
int src_offset = src.offset / src.elemSize1(), src_step = src.step1();
int dst_offset = dst.offset / dst.elemSize1(), dst_step = dst.step1();
std::string build_options = format("-D DEPTH_%d -D dcn=%d -D scn=%d -D %s",
src.depth(), dst.channels(), src.channels(), reverse ? "REVERSE" : "ORDER");
vector<pair<size_t , const void *> > args;
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.cols));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.rows));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_step));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_step));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset ));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_offset ));
size_t gt[3] = { dst.cols, dst.rows, 1 };
#ifdef ANDROID
size_t lt[3] = { 16, 10, 1 };
@@ -147,8 +299,8 @@ static void RGB_caller(const oclMat &src, oclMat &dst, bool reverse)
static void fromRGB5x5_caller(const oclMat &src, oclMat &dst, int bidx, int greenbits, const std::string & kernelName)
{
std::string build_options = format("-D DEPTH_%d -D greenbits=%d -D dcn=%d",
src.depth(), greenbits, dst.channels());
std::string build_options = format("-D DEPTH_%d -D greenbits=%d -D dcn=%d -D bidx=%d",
src.depth(), greenbits, dst.channels(), bidx);
int src_offset = src.offset >> 1, src_step = src.step >> 1;
int dst_offset = dst.offset / dst.elemSize1(), dst_step = dst.step / dst.elemSize1();
@@ -157,7 +309,6 @@ static void fromRGB5x5_caller(const oclMat &src, oclMat &dst, int bidx, int gree
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.rows));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_step));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_step));
args.push_back( make_pair( sizeof(cl_int) , (void *)&bidx));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset ));
@@ -174,8 +325,8 @@ static void fromRGB5x5_caller(const oclMat &src, oclMat &dst, int bidx, int gree
static void toRGB5x5_caller(const oclMat &src, oclMat &dst, int bidx, int greenbits, const std::string & kernelName)
{
std::string build_options = format("-D DEPTH_%d -D greenbits=%d -D scn=%d",
src.depth(), greenbits, src.channels());
std::string build_options = format("-D DEPTH_%d -D greenbits=%d -D scn=%d -D bidx=%d",
src.depth(), greenbits, src.channels(), bidx);
int src_offset = (int)src.offset, src_step = (int)src.step;
int dst_offset = dst.offset >> 1, dst_step = dst.step >> 1;
@@ -184,7 +335,6 @@ static void toRGB5x5_caller(const oclMat &src, oclMat &dst, int bidx, int greenb
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst.rows));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_step));
args.push_back( make_pair( sizeof(cl_int) , (void *)&dst_step));
args.push_back( make_pair( sizeof(cl_int) , (void *)&bidx));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&src.data));
args.push_back( make_pair( sizeof(cl_mem) , (void *)&dst.data));
args.push_back( make_pair( sizeof(cl_int) , (void *)&src_offset ));
@@ -272,7 +422,7 @@ static void cvtColor_caller(const oclMat &src, oclMat &dst, int code, int dcn)
CV_Assert(scn == 1);
dcn = code == CV_GRAY2BGRA ? 4 : 3;
dst.create(sz, CV_MAKETYPE(depth, dcn));
toRGB_caller(src, dst, 0, "Gray2RGB");
fromGray_caller(src, dst, 0, "Gray2RGB");
break;
}
case CV_BGR2YUV: case CV_RGB2YUV:
@@ -303,7 +453,7 @@ static void cvtColor_caller(const oclMat &src, oclMat &dst, int code, int dcn)
Size dstSz(sz.width, sz.height * 2 / 3);
dst.create(dstSz, CV_MAKETYPE(depth, dcn));
toRGB_caller(src, dst, bidx, "YUV2RGBA_NV12");
toRGB_NV12_caller(src, dst, bidx, "YUV2RGBA_NV12");
break;
}
case CV_BGR2YCrCb: case CV_RGB2YCrCb:
@@ -460,11 +610,11 @@ static void cvtColor_caller(const oclMat &src, oclMat &dst, int code, int dcn)
initialized = true;
}
fromRGB_caller(src, dst, bidx, kernelName, format(" -D hrange=%d", hrange), sdiv_data, hrange == 256 ? hdiv_data256 : hdiv_data180);
toHSV_caller(src, dst, bidx, kernelName, format(" -D hrange=%d", hrange), sdiv_data, hrange == 256 ? hdiv_data256 : hdiv_data180);
return;
}
fromRGB_caller(src, dst, bidx, kernelName, format(" -D hscale=%f", hrange*(1.f/360.f)));
toHSV_caller(src, dst, bidx, kernelName, format(" -D hscale=%f", hrange*(1.f/360.f)));
break;
}
case CV_HSV2BGR: case CV_HSV2RGB: case CV_HSV2BGR_FULL: case CV_HSV2RGB_FULL:
@@ -483,7 +633,7 @@ static void cvtColor_caller(const oclMat &src, oclMat &dst, int code, int dcn)
dst.create(sz, CV_MAKETYPE(depth, dcn));
std::string kernelName = std::string(is_hsv ? "HSV" : "HLS") + "2RGB";
toRGB_caller(src, dst, bidx, kernelName, format(" -D hrange=%d -D hscale=%f", hrange, 6.f/hrange));
fromHSV_caller(src, dst, bidx, kernelName, format(" -D hrange=%d -D hscale=%f", hrange, 6.f/hrange));
break;
}
case CV_RGBA2mRGBA: case CV_mRGBA2RGBA:

849
modules/ocl/src/opencl/cvt_color.cl
View File

File diff suppressed because it is too large Load Diff