Merge pull request #21746 from anna-khakimova:ak/dynamic_dispatch_sub

alalek · alalek · commit 63c865ace64f · 2022-03-21T17:10:40.000Z
diff --git a/modules/gapi/perf/common/gapi_core_perf_tests_inl.hpp b/modules/gapi/perf/common/gapi_core_perf_tests_inl.hpp
@@ -436,7 +436,7 @@ PERF_TEST_P_(DivPerfTest, TestPerformance)
     // FIXIT Unstable input data for divide
     initMatsRandU(type, sz, dtype, false);
 
-    //This condition need to workaround issue in the OpenCV.
+    //This condition need to workaround the #21044 issue in the OpenCV.
     //It reinitializes divider matrix without zero values for CV_16S DST type.
     if (dtype == CV_16S && dtype != type)
         cv::randu(in_mat2, cv::Scalar::all(1), cv::Scalar::all(255));
@@ -482,7 +482,7 @@ PERF_TEST_P_(DivCPerfTest, TestPerformance)
     // FIXIT Unstable input data for divide
     initMatsRandU(type, sz, dtype, false);
 
-    //This condition need as workaround the issue in the OpenCV.
+    //This condition need to workaround the #21044 issue in the OpenCV.
     //It reinitializes divider scalar without zero values for CV_16S DST type.
     if (dtype == CV_16S || (type == CV_16S && dtype == -1))
         cv::randu(sc, cv::Scalar::all(1), cv::Scalar::all(SHRT_MAX));
@@ -528,7 +528,7 @@ PERF_TEST_P_(DivRCPerfTest, TestPerformance)
 
     // FIXIT Unstable input data for divide
     initMatsRandU(type, sz, dtype, false);
-    //This condition need as workaround the bug in the OpenCV.
+    //This condition need to workaround the #21044 issue in the OpenCV.
     //It reinitializes divider matrix without zero values for CV_16S DST type.
     if (dtype == CV_16S || (type == CV_16S && dtype == -1))
         cv::randu(in_mat1, cv::Scalar::all(1), cv::Scalar::all(255));
diff --git a/modules/gapi/perf/cpu/gapi_core_perf_tests_fluid.cpp b/modules/gapi/perf/cpu/gapi_core_perf_tests_fluid.cpp
@@ -40,10 +40,10 @@ INSTANTIATE_TEST_CASE_P(AddCPerfTestFluid, AddCPerfTest,
             Values(cv::compile_args(CORE_FLUID))));
 
 INSTANTIATE_TEST_CASE_P(SubPerfTestFluid, SubPerfTest,
-    Combine(Values(AbsExact().to_compare_f()),
+    Combine(Values(Tolerance_FloatRel_IntAbs(1e-6, 0).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(SubCPerfTestFluid, SubCPerfTest,
diff --git a/modules/gapi/src/backends/fluid/gfluidcore.cpp b/modules/gapi/src/backends/fluid/gfluidcore.cpp
@@ -378,141 +378,11 @@ 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 sub_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 sub_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 sub_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 sub_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 sub_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;
-}
 #endif // CV_SIMD
 
 template<typename DST, typename SRC1, typename SRC2>
 CV_ALWAYS_INLINE void run_arithm(Buffer &dst, const View &src1, const View &src2,
-                                        Arithm arithm, double scale=1)
+                                 Arithm arithm, double scale=1)
 {
     static_assert(std::is_same<SRC1, SRC2>::value, "wrong types");
 
@@ -607,10 +477,19 @@ GAPI_FLUID_KERNEL(GFluidSub, cv::gapi::core::GSub, false)
     {
         //      DST     SRC1    SRC2    OP          __VA_ARGS__
         BINARY_(uchar , uchar , uchar , run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
-        BINARY_(uchar ,  short,  short, run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
-        BINARY_(uchar ,  float,  float, run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
-        BINARY_( short,  short,  short, run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
-        BINARY_( float, uchar , uchar , run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(uchar,  ushort, ushort, run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(uchar,  short,  short,  run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(uchar,  float,  float,  run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(short,  short,  short,  run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(short,  uchar,  uchar,  run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(short,  ushort, ushort, run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(short,  float,  float,  run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(ushort, ushort, ushort, run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(ushort, uchar,  uchar,  run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(ushort, short,  short,  run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(ushort, float,  float,  run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(float,  uchar,  uchar,  run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
+        BINARY_(float,  ushort, ushort, run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
         BINARY_( float,  short,  short, run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
         BINARY_( float,  float,  float, run_arithm, dst, src1, src2, ARITHM_SUBTRACT);
 
diff --git a/modules/gapi/src/backends/fluid/gfluidcore_func.dispatch.cpp b/modules/gapi/src/backends/fluid/gfluidcore_func.dispatch.cpp
@@ -317,6 +317,33 @@ ADD_SIMD(float, float)
 
 #undef ADD_SIMD
 
+#define SUB_SIMD(SRC, DST)                                                    \
+int sub_simd(const SRC in1[], const SRC in2[], DST out[], const int length)   \
+{                                                                             \
+                                                                              \
+        CV_CPU_DISPATCH(sub_simd, (in1, in2, out, length),                    \
+                        CV_CPU_DISPATCH_MODES_ALL);                           \
+}
+
+SUB_SIMD(uchar, uchar)
+SUB_SIMD(ushort, uchar)
+SUB_SIMD(short, uchar)
+SUB_SIMD(float, uchar)
+SUB_SIMD(short, short)
+SUB_SIMD(ushort, short)
+SUB_SIMD(uchar, short)
+SUB_SIMD(float, short)
+SUB_SIMD(ushort, ushort)
+SUB_SIMD(uchar, ushort)
+SUB_SIMD(short, ushort)
+SUB_SIMD(float, ushort)
+SUB_SIMD(uchar, float)
+SUB_SIMD(ushort, float)
+SUB_SIMD(short, float)
+SUB_SIMD(float, float)
+
+#undef SUB_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
@@ -244,6 +244,28 @@ ADD_SIMD(float, float)
 
 #undef ADD_SIMD
 
+#define SUB_SIMD(SRC, DST)                                                     \
+int sub_simd(const SRC in1[], const SRC in2[], DST out[], const int length);
+
+SUB_SIMD(uchar, uchar)
+SUB_SIMD(ushort, uchar)
+SUB_SIMD(short, uchar)
+SUB_SIMD(float, uchar)
+SUB_SIMD(short, short)
+SUB_SIMD(ushort, short)
+SUB_SIMD(uchar, short)
+SUB_SIMD(float, short)
+SUB_SIMD(ushort, ushort)
+SUB_SIMD(uchar, ushort)
+SUB_SIMD(short, ushort)
+SUB_SIMD(float, ushort)
+SUB_SIMD(uchar, float)
+SUB_SIMD(ushort, float)
+SUB_SIMD(short, float)
+SUB_SIMD(float, float)
+
+#undef SUB_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
