Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

GAPI Fluid SIMD:Add support of new several types for the Merge3. #21797

Merged
merged 4 commits into from
May 31, 2023
Merged

GAPI Fluid SIMD:Add support of new several types for the Merge3. #21797

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
3601b29
GAPI Fluid SIMD:Add support of new several types for the Merge3.
Mar 30, 2022
f09e84c
Fixes for the Split/Merge and ConvertTo issues.
Mar 30, 2022
9d6b84f
Applied comments
Apr 1, 2022
42653f5
Merge branch '4.x' into ak/merge3_extend_supported_types
May 29, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Jump to
Jump to file
Failed to load files.
Diff view
Diff view
2 changes: 1 addition & 1 deletion modules/gapi/perf/common/gapi_core_perf_tests.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -62,7 +62,7 @@ namespace opencv_test
class InRangePerfTest : public TestPerfParams<tuple<compare_f, cv::Size, MatType, cv::GCompileArgs>> {};
class Split3PerfTest : public TestPerfParams<tuple<compare_f, cv::Size, cv::GCompileArgs>> {};
class Split4PerfTest : public TestPerfParams<tuple<compare_f, cv::Size, cv::GCompileArgs>> {};
class Merge3PerfTest : public TestPerfParams<tuple<compare_f, cv::Size, cv::GCompileArgs>> {};
class Merge3PerfTest : public TestPerfParams<tuple<compare_f, cv::Size, MatType, cv::GCompileArgs>> {};
class Merge4PerfTest : public TestPerfParams<tuple<compare_f, cv::Size, cv::GCompileArgs>> {};
class RemapPerfTest : public TestPerfParams<tuple<compare_f, cv::Size, MatType, cv::GCompileArgs>> {};
class FlipPerfTest : public TestPerfParams<tuple<compare_f, cv::Size, MatType, int, cv::GCompileArgs>> {};
Expand Down
7 changes: 4 additions & 3 deletions modules/gapi/perf/common/gapi_core_perf_tests_inl.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1577,11 +1577,12 @@ PERF_TEST_P_(Merge3PerfTest, TestPerformance)
{
compare_f cmpF;
cv::Size sz;
MatType type = -1;
cv::GCompileArgs compile_args;
std::tie(cmpF, sz, compile_args) = GetParam();
std::tie(cmpF, sz, type, compile_args) = GetParam();

initMatsRandU(CV_8UC1, sz, CV_8UC3);
cv::Mat in_mat3(sz, CV_8UC1);
initMatsRandU(type, sz, CV_MAKETYPE(type, 3));
cv::Mat in_mat3(sz, type);
cv::Scalar mean = cv::Scalar::all(127);
cv::Scalar stddev = cv::Scalar::all(40.f);
cv::randn(in_mat3, mean, stddev);
Expand Down
1 change: 1 addition & 0 deletions modules/gapi/perf/cpu/gapi_core_perf_tests_cpu.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -252,6 +252,7 @@ INSTANTIATE_TEST_CASE_P(Split4PerfTestCPU, Split4PerfTest,
INSTANTIATE_TEST_CASE_P(Merge3PerfTestCPU, Merge3PerfTest,
Combine(Values(AbsExact().to_compare_f()),
Values(szSmall128, szVGA, sz720p, sz1080p),
Values(CV_8U),
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Should other types be listed here?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

..and in GPU tests, too.

Values(cv::compile_args(CORE_CPU))));

INSTANTIATE_TEST_CASE_P(Merge4PerfTestCPU, Merge4PerfTest,
Expand Down
1 change: 1 addition & 0 deletions modules/gapi/perf/cpu/gapi_core_perf_tests_fluid.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -253,6 +253,7 @@ INSTANTIATE_TEST_CASE_P(Split4PerfTestFluid, Split4PerfTest,
INSTANTIATE_TEST_CASE_P(Merge3PerfTestFluid, Merge3PerfTest,
Combine(Values(AbsExact().to_compare_f()),
Values(szSmall128, szVGA, sz720p, sz1080p),
Values(CV_8U, CV_16S, CV_16U, CV_32F),
Values(cv::compile_args(CORE_FLUID))));

INSTANTIATE_TEST_CASE_P(Merge4PerfTestFluid, Merge4PerfTest,
Expand Down
1 change: 1 addition & 0 deletions modules/gapi/perf/gpu/gapi_core_perf_tests_gpu.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -242,6 +242,7 @@ INSTANTIATE_TEST_CASE_P(Split4PerfTestGPU, Split4PerfTest,
INSTANTIATE_TEST_CASE_P(Merge3PerfTestGPU, Merge3PerfTest,
Combine(Values(AbsExact().to_compare_f()),
Values( szSmall128, szVGA, sz720p, sz1080p ),
Values(CV_8U),
Values(cv::compile_args(CORE_GPU))));

INSTANTIATE_TEST_CASE_P(Merge4PerfTestGPU, Merge4PerfTest,
Expand Down
66 changes: 45 additions & 21 deletions modules/gapi/src/backends/fluid/gfluidcore.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -2320,12 +2320,15 @@ GAPI_FLUID_KERNEL(GFluidSplit3, cv::gapi::core::GSplit3, false)

static void run(const View &src, Buffer &dst1, Buffer &dst2, Buffer &dst3)
{
GAPI_Assert((src.meta().depth == CV_8U) && (dst1.meta().depth == CV_8U) &&
(dst2.meta().depth == CV_8U) && (dst3.meta().depth == CV_8U) &&
(3 == src.meta().chan));

const auto *in = src.InLine<uchar>(0);
auto *out1 = dst1.OutLine<uchar>();
auto *out2 = dst2.OutLine<uchar>();
auto *out3 = dst3.OutLine<uchar>();

GAPI_Assert(3 == src.meta().chan);
int width = src.length();
int w = 0;

Expand All @@ -2348,13 +2351,16 @@ GAPI_FLUID_KERNEL(GFluidSplit4, cv::gapi::core::GSplit4, false)

static void run(const View &src, Buffer &dst1, Buffer &dst2, Buffer &dst3, Buffer &dst4)
{
GAPI_Assert((src.meta().depth == CV_8U) && (dst1.meta().depth == CV_8U) &&
(dst2.meta().depth == CV_8U) && (dst3.meta().depth == CV_8U) &&
(dst4.meta().depth == CV_8U) && (4 == src.meta().chan));

const auto *in = src.InLine<uchar>(0);
auto *out1 = dst1.OutLine<uchar>();
auto *out2 = dst2.OutLine<uchar>();
auto *out3 = dst3.OutLine<uchar>();
auto *out4 = dst4.OutLine<uchar>();

GAPI_Assert(4 == src.meta().chan);
int width = src.length();
int w = 0;

Expand All @@ -2372,31 +2378,46 @@ GAPI_FLUID_KERNEL(GFluidSplit4, cv::gapi::core::GSplit4, false)
}
};

template<typename T>
CV_ALWAYS_INLINE void run_merge3(Buffer& dst, const View& src1, const View& src2, const View& src3)
{
const auto* in1 = src1.InLine<T>(0);
const auto* in2 = src2.InLine<T>(0);
const auto* in3 = src3.InLine<T>(0);
auto* out = dst.OutLine<T>();

int width = dst.length();
int w = 0;

#if CV_SIMD
w = merge3_simd(in1, in2, in3, out, width);
#endif

for (; w < width; w++)
{
out[3 * w] = in1[w];
out[3 * w + 1] = in2[w];
out[3 * w + 2] = in3[w];
}
}

GAPI_FLUID_KERNEL(GFluidMerge3, cv::gapi::core::GMerge3, false)
{
static const int Window = 1;

static void run(const View &src1, const View &src2, const View &src3, Buffer &dst)
static void run(const View& src1, const View& src2, const View& src3, Buffer& dst)
{
const auto *in1 = src1.InLine<uchar>(0);
const auto *in2 = src2.InLine<uchar>(0);
const auto *in3 = src3.InLine<uchar>(0);
auto *out = dst.OutLine<uchar>();

GAPI_Assert(3 == dst.meta().chan);
int width = dst.length();
int w = 0;
GAPI_Assert((src1.meta().depth == dst.meta().depth) &&
(src1.meta().depth == src2.meta().depth) &&
(src1.meta().depth == src3.meta().depth));

#if CV_SIMD
w = merge3_simd(in1, in2, in3, out, width);
#endif
// SRC/DST TYPE OP __VA_ARGS__
MERGE3_(uchar, run_merge3, dst, src1, src2, src3);
MERGE3_(ushort, run_merge3, dst, src1, src2, src3);
MERGE3_(short, run_merge3, dst, src1, src2, src3);
MERGE3_(float, run_merge3, dst, src1, src2, src3);

for (; w < width; w++)
{
out[3*w ] = in1[w];
out[3*w + 1] = in2[w];
out[3*w + 2] = in3[w];
}
CV_Error(cv::Error::StsBadArg, "unsupported combination of types");
}
};

Expand All @@ -2407,13 +2428,16 @@ GAPI_FLUID_KERNEL(GFluidMerge4, cv::gapi::core::GMerge4, false)
static void run(const View &src1, const View &src2, const View &src3, const View &src4,
Buffer &dst)
{
GAPI_Assert((dst.meta().depth == CV_8U) && (src1.meta().depth == CV_8U) &&
(src2.meta().depth == CV_8U) && (src3.meta().depth == CV_8U) &&
(4 == dst.meta().chan));

const auto *in1 = src1.InLine<uchar>(0);
const auto *in2 = src2.InLine<uchar>(0);
const auto *in3 = src3.InLine<uchar>(0);
const auto *in4 = src4.InLine<uchar>(0);
auto *out = dst.OutLine<uchar>();

GAPI_Assert(4 == dst.meta().chan);
int width = dst.length();

int w = 0; // cycle counter
Expand Down
18 changes: 13 additions & 5 deletions modules/gapi/src/backends/fluid/gfluidcore_func.dispatch.cpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -276,13 +276,21 @@ int split4_simd(const uchar in[], uchar out1[], uchar out2[],
CV_CPU_DISPATCH_MODES_ALL);
}

int merge3_simd(const uchar in1[], const uchar in2[], const uchar in3[],
uchar out[], const int width)
{
CV_CPU_DISPATCH(merge3_simd, (in1, in2, in3, out, width),
CV_CPU_DISPATCH_MODES_ALL);
#define MERGE3_SIMD(T) \
int merge3_simd(const T in1[], const T in2[], const T in3[], \
T out[], const int width) \
{ \
CV_CPU_DISPATCH(merge3_simd, (in1, in2, in3, out, width), \
CV_CPU_DISPATCH_MODES_ALL); \
}

MERGE3_SIMD(uchar)
MERGE3_SIMD(short)
MERGE3_SIMD(ushort)
MERGE3_SIMD(float)

#undef MERGE3_SIMD

int merge4_simd(const uchar in1[], const uchar in2[], const uchar in3[],
const uchar in4[], uchar out[], const int width)
{
Expand Down
12 changes: 10 additions & 2 deletions modules/gapi/src/backends/fluid/gfluidcore_func.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -216,8 +216,16 @@ int split3_simd(const uchar in[], uchar out1[], uchar out2[],
int split4_simd(const uchar in[], uchar out1[], uchar out2[],
uchar out3[], uchar out4[], const int width);

int merge3_simd(const uchar in1[], const uchar in2[], const uchar in3[],
uchar out[], const int width);
#define MERGE3_SIMD(T) \
int merge3_simd(const T in1[], const T in2[], const T in3[], \
T out[], const int width);

MERGE3_SIMD(uchar)
MERGE3_SIMD(short)
MERGE3_SIMD(ushort)
MERGE3_SIMD(float)

#undef MERGE3_SIMD

int merge4_simd(const uchar in1[], const uchar in2[], const uchar in3[],
const uchar in4[], uchar out[], const int width);
Expand Down
80 changes: 51 additions & 29 deletions modules/gapi/src/backends/fluid/gfluidcore_func.simd.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -322,12 +322,21 @@ int split3_simd(const uchar in[], uchar out1[], uchar out2[],
int split4_simd(const uchar in[], uchar out1[], uchar out2[],
uchar out3[], uchar out4[], const int width);

int merge3_simd(const uchar in1[], const uchar in2[], const uchar in3[],
uchar out[], const int width);
#define MERGE3_SIMD(T) \
int merge3_simd(const T in1[], const T in2[], const T in3[], \
T out[], const int width);

MERGE3_SIMD(uchar)
MERGE3_SIMD(short)
MERGE3_SIMD(ushort)
MERGE3_SIMD(float)

#undef MERGE3_SIMD

int merge4_simd(const uchar in1[], const uchar in2[], const uchar in3[],
const uchar in4[], uchar out[], const int width);


#ifndef CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY

#define SRC_SHORT_OR_USHORT std::is_same<SRC, short>::value || std::is_same<SRC, ushort>::value
Expand Down Expand Up @@ -2530,33 +2539,41 @@ int split4_simd(const uchar in[], uchar out1[], uchar out2[],
//
//-------------------------

int merge3_simd(const uchar in1[], const uchar in2[], const uchar in3[],
uchar out[], const int width)
{
constexpr int nlanes = v_uint8::nlanes;
if (width < nlanes)
return 0;

int x = 0;
for (;;)
{
for (; x <= width - nlanes; x += nlanes)
{
v_uint8 a, b, c;
a = vx_load(&in1[x]);
b = vx_load(&in2[x]);
c = vx_load(&in3[x]);
v_store_interleave(&out[3 * x], a, b, c);
}
if (x < width)
{
x = width - nlanes;
continue;
}
break;
}
return x;
}
#define MERGE3_SIMD(T) \
int merge3_simd(const T in1[], const T in2[], const T in3[], \
T out[], const int width) \
{ \
constexpr int nlanes = vector_type_of_t<T>::nlanes; \
if (width < nlanes) \
return 0; \
\
int x = 0; \
for (;;) \
{ \
for (; x <= width - nlanes; x += nlanes) \
{ \
vector_type_of_t<T> a, b, c; \
a = vx_load(&in1[x]); \
b = vx_load(&in2[x]); \
c = vx_load(&in3[x]); \
v_store_interleave(&out[3 * x], a, b, c); \
} \
if (x < width) \
{ \
x = width - nlanes; \
continue; \
} \
break; \
} \
return x; \
}

MERGE3_SIMD(uchar)
MERGE3_SIMD(short)
MERGE3_SIMD(ushort)
MERGE3_SIMD(float)

#undef MERGE3_SIMD

//-------------------------
//
Expand Down Expand Up @@ -2926,6 +2943,8 @@ CV_ALWAYS_INLINE void convertto_simd_nocoeff_impl(const SRC* inx, float* outx)
int convertto_simd(const SRC in[], DST out[], const int length) \
{ \
constexpr int nlanes = vector_type_of_t<DST>::nlanes; \
if (length < nlanes) \
return 0; \
\
int x = 0; \
for (;;) \
Expand Down Expand Up @@ -3093,6 +3112,9 @@ int convertto_scaled_simd(const SRC in[], DST out[], const float alpha, \
const float beta, const int length) \
{ \
constexpr int nlanes = vector_type_of_t<DST>::nlanes; \
if (length < nlanes) \
return 0; \
\
v_float32 v_alpha = vx_setall_f32(alpha); \
v_float32 v_beta = vx_setall_f32(beta); \
\
Expand Down
17 changes: 17 additions & 0 deletions modules/gapi/src/backends/fluid/gfluidutils.hpp
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -86,6 +86,23 @@ using cv::gapi::own::rintd;
return; \
}

#define MERGE3_(T, OP, ...) \
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Cannot find place where is it used, could you point me out?

Copy link
Member Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes. It is used here.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

thanks

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

discussed offline:
to move out repeated conditions into beginning of function to simplify condition calculation complexity in actual type dispatching routing

if (cv::DataType<T>::depth == dst.meta().depth && \
cv::DataType<T>::depth == src1.meta().depth) \
{ \
GAPI_DbgAssert(dst.length() == src1.length()); \
GAPI_DbgAssert(dst.length() == src2.length()); \
GAPI_DbgAssert(dst.length() == src3.length()); \
Comment on lines +93 to +95
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

what is the reason to verify this?
normally in the framework we rely on our output buffers which are allocated per our needs.
it is not coming somewhere from the outside.

\
GAPI_DbgAssert(1 == src1.meta().chan); \
GAPI_DbgAssert(1 == src2.meta().chan); \
GAPI_DbgAssert(1 == src3.meta().chan); \
GAPI_DbgAssert(3 == dst.meta().chan); \
\
OP<T>(__VA_ARGS__); \
return; \
}
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Macro function bodies considered harmful, is there any reasonable issues with rewriting it to templates?


} // namespace fluid
} // namespace gapi
} // namespace cv
Expand Down