diff --git a/modules/gapi/perf/cpu/gapi_core_perf_tests_fluid.cpp b/modules/gapi/perf/cpu/gapi_core_perf_tests_fluid.cpp
index a142109315..d91ce65fff 100644
--- a/modules/gapi/perf/cpu/gapi_core_perf_tests_fluid.cpp
+++ b/modules/gapi/perf/cpu/gapi_core_perf_tests_fluid.cpp
@@ -28,8 +28,8 @@ INSTANTIATE_TEST_CASE_P(SqrtPerfTestFluid, SqrtPerfTest,
INSTANTIATE_TEST_CASE_P(AddPerfTestFluid, AddPerfTest,
Combine(Values(Tolerance_FloatRel_IntAbs(1e-6, 1).to_compare_f()),
Values(szSmall128, szVGA, sz720p, sz1080p),
- Values(CV_8UC1, CV_8UC3, CV_16SC1, CV_32FC1),
- Values(-1, CV_8U, CV_32F),
+ Values(CV_8UC1, CV_8UC3, CV_16UC1, CV_16SC1, CV_32FC1),
+ Values(-1, CV_8U, CV_16U, CV_16S, CV_32F),
Values(cv::compile_args(CORE_FLUID))));
INSTANTIATE_TEST_CASE_P(AddCPerfTestFluid, AddCPerfTest,
diff --git a/modules/gapi/src/backends/fluid/gfluidcore.cpp b/modules/gapi/src/backends/fluid/gfluidcore.cpp
index bb33c45d85..866381f39b 100644
--- a/modules/gapi/src/backends/fluid/gfluidcore.cpp
+++ b/modules/gapi/src/backends/fluid/gfluidcore.cpp
@@ -379,136 +379,6 @@ CV_ALWAYS_INLINE int absdiff_simd(const T in1[], const T in2[], T out[], int len
return 0;
}
-template<typename T, typename VT>
-CV_ALWAYS_INLINE int add_simd_sametype(const T in1[], const T in2[], T out[], int length)
-{
- constexpr int nlanes = static_cast<int>(VT::nlanes);
-
- if (length < nlanes)
- return 0;
-
- int x = 0;
- for (;;)
- {
- for (; x <= length - nlanes; x += nlanes)
- {
- VT a = vx_load(&in1[x]);
- VT b = vx_load(&in2[x]);
- vx_store(&out[x], a + b);
- }
-
- if (x < length && (in1 != out) && (in2 != out))
- {
- x = length - nlanes;
- continue; // process one more time (unaligned tail)
- }
- break;
- }
-
- return x;
-}
-
-template<typename SRC, typename DST>
-CV_ALWAYS_INLINE int add_simd(const SRC in1[], const SRC in2[], DST out[], int length)
-{
- if (std::is_same<DST, float>::value && !std::is_same<SRC, float>::value)
- return 0;
-
- if (std::is_same<DST, SRC>::value)
- {
- if (std::is_same<DST, uchar>::value)
- {
- return add_simd_sametype<uchar, v_uint8>(reinterpret_cast<const uchar*>(in1),
- reinterpret_cast<const uchar*>(in2),
- reinterpret_cast<uchar*>(out), length);
- }
- else if (std::is_same<DST, short>::value)
- {
- return add_simd_sametype<short, v_int16>(reinterpret_cast<const short*>(in1),
- reinterpret_cast<const short*>(in2),
- reinterpret_cast<short*>(out), length);
- }
- else if (std::is_same<DST, float>::value)
- {
- return add_simd_sametype<float, v_float32>(reinterpret_cast<const float*>(in1),
- reinterpret_cast<const float*>(in2),
- reinterpret_cast<float*>(out), length);
- }
- }
- else if (std::is_same<SRC, short>::value && std::is_same<DST, uchar>::value)
- {
- constexpr int nlanes = static_cast<int>(v_uint8::nlanes);
-
- if (length < nlanes)
- return 0;
-
- int x = 0;
- for (;;)
- {
- for (; x <= length - nlanes; x += nlanes)
- {
- v_int16 a1 = vx_load(reinterpret_cast<const short*>(&in1[x]));
- v_int16 a2 = vx_load(reinterpret_cast<const short*>(&in1[x + nlanes / 2]));
- v_int16 b1 = vx_load(reinterpret_cast<const short*>(&in2[x]));
- v_int16 b2 = vx_load(reinterpret_cast<const short*>(&in2[x + nlanes / 2]));
-
- vx_store(reinterpret_cast<uchar*>(&out[x]), v_pack_u(a1 + b1, a2 + b2));
- }
-
- if (x < length)
- {
- CV_DbgAssert((reinterpret_cast<const short*>(in1) != reinterpret_cast<const short*>(out)) &&
- (reinterpret_cast<const short*>(in2) != reinterpret_cast<const short*>(out)));
- x = length - nlanes;
- continue; // process one more time (unaligned tail)
- }
- break;
- }
-
- return x;
- }
- else if (std::is_same<SRC, float>::value && std::is_same<DST, uchar>::value)
- {
- constexpr int nlanes = static_cast<int>(v_uint8::nlanes);
-
- if (length < nlanes)
- return 0;
-
- int x = 0;
- for (;;)
- {
- for (; x <= length - nlanes; x += nlanes)
- {
- v_float32 a1 = vx_load(reinterpret_cast<const float*>(&in1[x]));
- v_float32 a2 = vx_load(reinterpret_cast<const float*>(&in1[x + nlanes / 4]));
- v_float32 a3 = vx_load(reinterpret_cast<const float*>(&in1[x + 2 * nlanes / 4]));
- v_float32 a4 = vx_load(reinterpret_cast<const float*>(&in1[x + 3 * nlanes / 4]));
-
- v_float32 b1 = vx_load(reinterpret_cast<const float*>(&in2[x]));
- v_float32 b2 = vx_load(reinterpret_cast<const float*>(&in2[x + nlanes / 4]));
- v_float32 b3 = vx_load(reinterpret_cast<const float*>(&in2[x + 2 * nlanes / 4]));
- v_float32 b4 = vx_load(reinterpret_cast<const float*>(&in2[x + 3 * nlanes / 4]));
-
- vx_store(reinterpret_cast<uchar*>(&out[x]), v_pack_u(v_pack(v_round(a1 + b1), v_round(a2 + b2)),
- v_pack(v_round(a3 + b3), v_round(a4 + b4))));
- }
-
- if (x < length)
- {
- CV_DbgAssert((reinterpret_cast<const float*>(in1) != reinterpret_cast<const float*>(out)) &&
- (reinterpret_cast<const float*>(in2) != reinterpret_cast<const float*>(out)));
- x = length - nlanes;
- continue; // process one more time (unaligned tail)
- }
- break;
- }
-
- return x;
- }
-
- return 0;
-}
-
template<typename T, typename VT>
CV_ALWAYS_INLINE int sub_simd_sametype(const T in1[], const T in2[], T out[], int length)
{
@@ -641,7 +511,7 @@ CV_ALWAYS_INLINE int sub_simd(const SRC in1[], const SRC in2[], DST out[], int l
#endif // CV_SIMD
template<typename DST, typename SRC1, typename SRC2>
-static CV_ALWAYS_INLINE void run_arithm(Buffer &dst, const View &src1, const View &src2,
+CV_ALWAYS_INLINE void run_arithm(Buffer &dst, const View &src1, const View &src2,
Arithm arithm, double scale=1)
{
static_assert(std::is_same<SRC1, SRC2>::value, "wrong types");
@@ -652,7 +522,7 @@ static CV_ALWAYS_INLINE void run_arithm(Buffer &dst, const View &src1, const Vie
int width = dst.length();
int chan = dst.meta().chan;
- int length = width * chan;
+ const int length = width * chan;
// NB: assume in/out types are not 64-bits
float _scale = static_cast<float>( scale );
@@ -708,13 +578,22 @@ GAPI_FLUID_KERNEL(GFluidAdd, cv::gapi::core::GAdd, false)
static void run(const View &src1, const View &src2, int /*dtype*/, Buffer &dst)
{
// DST SRC1 SRC2 OP __VA_ARGS__
- BINARY_(uchar , uchar , uchar , run_arithm, dst, src1, src2, ARITHM_ADD);
- BINARY_(uchar , short, short, run_arithm, dst, src1, src2, ARITHM_ADD);
- BINARY_(uchar , float, float, run_arithm, dst, src1, src2, ARITHM_ADD);
- BINARY_( short, short, short, run_arithm, dst, src1, src2, ARITHM_ADD);
- BINARY_( float, uchar , uchar , run_arithm, dst, src1, src2, ARITHM_ADD);
- BINARY_( float, short, short, run_arithm, dst, src1, src2, ARITHM_ADD);
- BINARY_( float, float, float, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(uchar, uchar, uchar, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(uchar, ushort, ushort, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(uchar, short, short, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(uchar, float, float, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(short, short, short, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(short, uchar, uchar, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(short, ushort, ushort, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(short, float, float, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(ushort, ushort, ushort, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(ushort, uchar, uchar, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(ushort, short, short, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(ushort, float, float, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(float, uchar, uchar, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(float, ushort, ushort, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(float, short, short, run_arithm, dst, src1, src2, ARITHM_ADD);
+ BINARY_(float, float, float, run_arithm, dst, src1, src2, ARITHM_ADD);
CV_Error(cv::Error::StsBadArg, "unsupported combination of types");
}
diff --git a/modules/gapi/src/backends/fluid/gfluidcore_func.dispatch.cpp b/modules/gapi/src/backends/fluid/gfluidcore_func.dispatch.cpp
index a682fb7dbb..d80a6b29c0 100644
--- a/modules/gapi/src/backends/fluid/gfluidcore_func.dispatch.cpp
+++ b/modules/gapi/src/backends/fluid/gfluidcore_func.dispatch.cpp
@@ -290,6 +290,33 @@ int merge4_simd(const uchar in1[], const uchar in2[], const uchar in3[],
CV_CPU_DISPATCH_MODES_ALL);
}
+#define ADD_SIMD(SRC, DST) \
+int add_simd(const SRC in1[], const SRC in2[], DST out[], const int length) \
+{ \
+ \
+ CV_CPU_DISPATCH(add_simd, (in1, in2, out, length), \
+ CV_CPU_DISPATCH_MODES_ALL); \
+}
+
+ADD_SIMD(uchar, uchar)
+ADD_SIMD(ushort, uchar)
+ADD_SIMD(short, uchar)
+ADD_SIMD(float, uchar)
+ADD_SIMD(short, short)
+ADD_SIMD(ushort, short)
+ADD_SIMD(uchar, short)
+ADD_SIMD(float, short)
+ADD_SIMD(ushort, ushort)
+ADD_SIMD(uchar, ushort)
+ADD_SIMD(short, ushort)
+ADD_SIMD(float, ushort)
+ADD_SIMD(uchar, float)
+ADD_SIMD(ushort, float)
+ADD_SIMD(short, float)
+ADD_SIMD(float, float)
+
+#undef ADD_SIMD
+
} // namespace fluid
} // namespace gapi
} // namespace cv
diff --git a/modules/gapi/src/backends/fluid/gfluidcore_func.hpp b/modules/gapi/src/backends/fluid/gfluidcore_func.hpp
index 975383a8d9..052adbe2fd 100644
--- a/modules/gapi/src/backends/fluid/gfluidcore_func.hpp
+++ b/modules/gapi/src/backends/fluid/gfluidcore_func.hpp
@@ -222,6 +222,28 @@ int merge3_simd(const uchar in1[], const uchar in2[], const uchar in3[],
int merge4_simd(const uchar in1[], const uchar in2[], const uchar in3[],
const uchar in4[], uchar out[], const int width);
+#define ADD_SIMD(SRC, DST) \
+int add_simd(const SRC in1[], const SRC in2[], DST out[], const int length);
+
+ADD_SIMD(uchar, uchar)
+ADD_SIMD(ushort, uchar)
+ADD_SIMD(short, uchar)
+ADD_SIMD(float, uchar)
+ADD_SIMD(short, short)
+ADD_SIMD(ushort, short)
+ADD_SIMD(uchar, short)
+ADD_SIMD(float, short)
+ADD_SIMD(ushort, ushort)
+ADD_SIMD(uchar, ushort)
+ADD_SIMD(short, ushort)
+ADD_SIMD(float, ushort)
+ADD_SIMD(uchar, float)
+ADD_SIMD(ushort, float)
+ADD_SIMD(short, float)
+ADD_SIMD(float, float)
+
+#undef ADD_SIMD
+
} // namespace fluid
} // namespace gapi
} // namespace cv
diff --git a/modules/gapi/src/backends/fluid/gfluidcore_func.simd.hpp b/modules/gapi/src/backends/fluid/gfluidcore_func.simd.hpp
index b92d92d0cf..4c324daa25 100644
--- a/modules/gapi/src/backends/fluid/gfluidcore_func.simd.hpp
+++ b/modules/gapi/src/backends/fluid/gfluidcore_func.simd.hpp
@@ -231,6 +231,28 @@ DIVRC_SIMD(float, float)
#undef DIVRC_SIMD
+#define ADD_SIMD(SRC, DST) \
+int add_simd(const SRC in1[], const SRC in2[], DST out[], const int length);
+
+ADD_SIMD(uchar, uchar)
+ADD_SIMD(ushort, uchar)
+ADD_SIMD(short, uchar)
+ADD_SIMD(float, uchar)
+ADD_SIMD(short, short)
+ADD_SIMD(ushort, short)
+ADD_SIMD(uchar, short)
+ADD_SIMD(float, short)
+ADD_SIMD(ushort, ushort)
+ADD_SIMD(uchar, ushort)
+ADD_SIMD(short, ushort)
+ADD_SIMD(float, ushort)
+ADD_SIMD(uchar, float)
+ADD_SIMD(ushort, float)
+ADD_SIMD(short, float)
+ADD_SIMD(float, float)
+
+#undef ADD_SIMD
+
int split3_simd(const uchar in[], uchar out1[], uchar out2[],
uchar out3[], const int width);
@@ -2503,6 +2525,178 @@ int merge4_simd(const uchar in1[], const uchar in2[], const uchar in3[],
return x;
}
+//-------------------------
+//
+// Fluid kernels: Add
+//
+//-------------------------
+
+CV_ALWAYS_INLINE void add_uchar_store(uchar* outx, const v_uint16& c1, const v_uint16& c2)
+{
+ vx_store(outx, v_pack(c1, c2));
+}
+
+CV_ALWAYS_INLINE void add_uchar_store(uchar* outx, const v_int16& c1, const v_int16& c2)
+{
+ vx_store(outx, v_pack_u(c1, c2));
+}
+
+template<typename SRC, typename DST>
+CV_ALWAYS_INLINE
+typename std::enable_if<std::is_same<SRC, DST>::value, void>::type
+add_simd_impl(const SRC* in1x, const SRC* in2x, DST* outx)
+{
+ vector_type_of_t<SRC> a = vx_load(in1x);
+ vector_type_of_t<SRC> b = vx_load(in2x);
+ vx_store(outx, a + b);
+}
+
+template<typename SRC>
+CV_ALWAYS_INLINE
+typename std::enable_if<std::is_same<SRC, short>::value ||
+ std::is_same<SRC, ushort>::value, void>::type
+add_simd_impl(const SRC* in1x, const SRC* in2x, uchar* outx)
+{
+ constexpr int nlanes = v_uint8::nlanes;
+
+ vector_type_of_t<SRC> a1 = vx_load(in1x);
+ vector_type_of_t<SRC> a2 = vx_load(&in1x[nlanes / 2]);
+ vector_type_of_t<SRC> b1 = vx_load(in2x);
+ vector_type_of_t<SRC> b2 = vx_load(&in2x[nlanes / 2]);
+
+ add_uchar_store(outx, a1 + b1, a2 + b2);
+}
+
+CV_ALWAYS_INLINE void add_simd_impl(const float* in1x, const float* in2x, uchar* outx)
+{
+ constexpr int nlanes = v_uint8::nlanes;
+
+ v_float32 a1 = vx_load(in1x);
+ v_float32 a2 = vx_load(&in1x[nlanes / 4]);
+ v_float32 a3 = vx_load(&in1x[2 * nlanes / 4]);
+ v_float32 a4 = vx_load(&in1x[3 * nlanes / 4]);
+
+ v_float32 b1 = vx_load(in2x);
+ v_float32 b2 = vx_load(&in2x[nlanes / 4]);
+ v_float32 b3 = vx_load(&in2x[2 * nlanes / 4]);
+ v_float32 b4 = vx_load(&in2x[3 * nlanes / 4]);
+
+ vx_store(outx, v_pack_u(v_pack(v_round(a1 + b1), v_round(a2 + b2)),
+ v_pack(v_round(a3 + b3), v_round(a4 + b4))));
+}
+
+CV_ALWAYS_INLINE void add_simd_impl(const uchar* in1x, const uchar* in2x, short* outx)
+{
+ v_int16 a = v_reinterpret_as_s16(vx_load_expand(in1x));
+ v_int16 b = v_reinterpret_as_s16(vx_load_expand(in2x));
+
+ vx_store(outx, a + b);
+}
+
+CV_ALWAYS_INLINE void add_simd_impl(const uchar* in1x, const uchar* in2x, ushort* outx)
+{
+ v_uint16 a = vx_load_expand(in1x);
+ v_uint16 b = vx_load_expand(in2x);
+
+ vx_store(outx, a + b);
+}
+
+template<typename DST>
+CV_ALWAYS_INLINE
+typename std::enable_if<std::is_same<DST, short>::value ||
+ std::is_same<DST, ushort>::value, void>::type
+add_simd_impl(const float* in1x, const float* in2x, DST* outx)
+{
+ constexpr int nlanes = vector_type_of_t<DST>::nlanes;
+ v_float32 a1 = vx_load(in1x);
+ v_float32 a2 = vx_load(&in1x[nlanes/2]);
+ v_float32 b1 = vx_load(in2x);
+ v_float32 b2 = vx_load(&in2x[nlanes/2]);
+
+ v_store_i16(outx, v_round(a1 + b1), v_round(a2 + b2));
+}
+
+CV_ALWAYS_INLINE void add_simd_impl(const short* in1x, const short* in2x, ushort* outx)
+{
+ v_int16 a = vx_load(in1x);
+ v_int32 a1 = v_expand_low(a);
+ v_int32 a2 = v_expand_high(a);
+
+ v_int16 b = vx_load(in2x);
+ v_int32 b1 = v_expand_low(b);
+ v_int32 b2 = v_expand_high(b);
+
+ vx_store(outx, v_pack_u(a1 + b1, a2 + b2));
+}
+
+CV_ALWAYS_INLINE void add_simd_impl(const ushort* in1x, const ushort* in2x, short* outx)
+{
+ v_uint16 a = vx_load(in1x);
+ v_uint32 a1 = v_expand_low(a);
+ v_uint32 a2 = v_expand_high(a);
+
+ v_uint16 b = vx_load(in2x);
+ v_uint32 b1 = v_expand_low(b);
+ v_uint32 b2 = v_expand_high(b);
+
+ vx_store(outx, v_reinterpret_as_s16(v_pack(a1 + b1, a2 + b2)));
+}
+
+template<typename SRC>
+CV_ALWAYS_INLINE void add_simd_impl(const SRC* in1x, const SRC* in2x, float* outx)
+{
+ v_float32 a = vg_load_f32(in1x);
+ v_float32 b = vg_load_f32(in2x);
+
+ vx_store(outx, a + b);
+}
+
+#define ADD_SIMD(SRC, DST) \
+int add_simd(const SRC in1[], const SRC in2[], DST out[], const int length) \
+{ \
+ constexpr int nlanes = vector_type_of_t<DST>::nlanes; \
+ \
+ if (length < nlanes) \
+ return 0; \
+ \
+ int x = 0; \
+ for (;;) \
+ { \
+ for (; x <= length - nlanes; x += nlanes) \
+ { \
+ add_simd_impl(&in1[x], &in2[x], &out[x]); \
+ } \
+ \
+ if (x < length) \
+ { \
+ x = length - nlanes; \
+ continue; \
+ } \
+ break; \
+ } \
+ \
+ return x; \
+}
+
+ADD_SIMD(uchar, uchar)
+ADD_SIMD(ushort, uchar)
+ADD_SIMD(short, uchar)
+ADD_SIMD(float, uchar)
+ADD_SIMD(short, short)
+ADD_SIMD(ushort, short)
+ADD_SIMD(uchar, short)
+ADD_SIMD(float, short)
+ADD_SIMD(ushort, ushort)
+ADD_SIMD(uchar, ushort)
+ADD_SIMD(short, ushort)
+ADD_SIMD(float, ushort)
+ADD_SIMD(uchar, float)
+ADD_SIMD(ushort, float)
+ADD_SIMD(short, float)
+ADD_SIMD(float, float)
+
+#undef ADD_SIMD
+
#endif // CV_CPU_OPTIMIZATION_DECLARATIONS_ONLY
CV_CPU_OPTIMIZATION_NAMESPACE_END