Merge pull request #29557 from fengyuentau:4x/hal_rvv/split_opt

hal_rvv core: Cover more data types for split with native RVV intrinsics #29557

Covering 16-bit, 32-bit and 64-bit split.

### Pull Request Readiness Checklist

See details at https://github.com/opencv/opencv/wiki/How_to_contribute#making-a-good-pull-request

- [x] I agree to contribute to the project under Apache 2 License.
- [x] To the best of my knowledge, the proposed patch is not based on a code under GPL or another license that is incompatible with OpenCV
- [x] The PR is proposed to the proper branch
- [x] There is a reference to the original bug report and related work
- [x] There is accuracy test, performance test and test data in opencv_extra repository, if applicable
      Patch to opencv_extra has the same branch name.
- [x] The feature is well documented and sample code can be built with the project CMake

hal/riscv-rvv/include/core.hpp

@@ -99,9 +99,18 @@ int fast_atan_64(const double* y, const double* x, double* dst, size_t n, bool a
 /* ############ split ############ */
 
 int split8u(const uchar* src, uchar** dst, int len, int cn);
+int split16u(const ushort* src, ushort** dst, int len, int cn);
+int split32s(const int* src, int** dst, int len, int cn);
+int split64s(const int64* src, int64** dst, int len, int cn);
 
 #undef cv_hal_split8u
 #define cv_hal_split8u cv::rvv_hal::core::split8u
+#undef cv_hal_split16u
+#define cv_hal_split16u cv::rvv_hal::core::split16u
+#undef cv_hal_split32s
+#define cv_hal_split32s cv::rvv_hal::core::split32s
+#undef cv_hal_split64s
+#define cv_hal_split64s cv::rvv_hal::core::split64s
 
 /* ############ sqrt ############ */
 

hal/riscv-rvv/src/core/copy_mask.cpp

@@ -30,8 +30,27 @@ static int copyToMasked_e##X##c1(const uchar *src_data, size_t src_step, const u
     return CV_HAL_ERROR_OK; \
 }
 
-CV_HAL_RVV_COPY_MASK_eXc1(8,  8)
-CV_HAL_RVV_COPY_MASK_eXc1(16, 4)
+#define CV_HAL_RVV_COPY_MASK_MERGE_eXc1(X, data_lmul, mask_lmul) \
+static int copyToMasked_e##X##c1(const uchar *src_data, size_t src_step, const uchar *mask_data, size_t mask_step, \
+                                 uchar *dst_data, size_t dst_step, int width, int height) { \
+    for (; height--; mask_data += mask_step, src_data += src_step, dst_data += dst_step) { \
+        const uint##X##_t *src = (const uint##X##_t*)src_data; \
+        uint##X##_t *dst = (uint##X##_t*)dst_data; \
+        int vl; \
+        for (int i = 0; i < width; i += vl) { \
+            vl = __riscv_vsetvl_e##X##m##data_lmul(width - i); \
+            auto m = __riscv_vmsne(__riscv_vle8_v_u8m##mask_lmul(mask_data + i, vl), 0, vl); \
+            auto src_v = __riscv_vle##X##_v_u##X##m##data_lmul(src + i, vl); \
+            auto dst_v = __riscv_vle##X##_v_u##X##m##data_lmul(dst + i, vl); \
+            __riscv_vse##X##_v_u##X##m##data_lmul(dst + i, __riscv_vmerge(dst_v, src_v, m, vl), vl); \
+        } \
+    } \
+    return CV_HAL_ERROR_OK; \
+}
+
+CV_HAL_RVV_COPY_MASK_MERGE_eXc1(8,  8, 8)
+CV_HAL_RVV_COPY_MASK_MERGE_eXc1(16, 8, 4)
+
 CV_HAL_RVV_COPY_MASK_eXc1(32, 2)
 CV_HAL_RVV_COPY_MASK_eXc1(64, 1)
 

hal/riscv-rvv/src/core/merge.cpp

@@ -28,6 +28,12 @@
 #define __riscv_vcreate_v_u16m4x2(v0, v1) OPENCV_HAL_IMPL_RVV_VCREATE_x2(u16, 4, v0, v1)
 #define __riscv_vcreate_v_u16m2x3(v0, v1, v2) OPENCV_HAL_IMPL_RVV_VCREATE_x3(u16, 2, v0, v1, v2)
 #define __riscv_vcreate_v_u16m2x4(v0, v1, v2, v3) OPENCV_HAL_IMPL_RVV_VCREATE_x4(u16, 2, v0, v1, v2, v3)
+#define __riscv_vcreate_v_i32m4x2(v0, v1) OPENCV_HAL_IMPL_RVV_VCREATE_x2(i32, 4, v0, v1)
+#define __riscv_vcreate_v_i32m2x3(v0, v1, v2) OPENCV_HAL_IMPL_RVV_VCREATE_x3(i32, 2, v0, v1, v2)
+#define __riscv_vcreate_v_i32m2x4(v0, v1, v2, v3) OPENCV_HAL_IMPL_RVV_VCREATE_x4(i32, 2, v0, v1, v2, v3)
+#define __riscv_vcreate_v_i64m4x2(v0, v1) OPENCV_HAL_IMPL_RVV_VCREATE_x2(i64, 4, v0, v1)
+#define __riscv_vcreate_v_i64m2x3(v0, v1, v2) OPENCV_HAL_IMPL_RVV_VCREATE_x3(i64, 2, v0, v1, v2)
+#define __riscv_vcreate_v_i64m2x4(v0, v1, v2, v3) OPENCV_HAL_IMPL_RVV_VCREATE_x4(i64, 2, v0, v1, v2, v3)
 #endif  // clang < 18
 
 namespace cv { namespace rvv_hal { namespace core {
@@ -204,146 +210,177 @@ int merge16u(const ushort** src, ushort* dst, int len, int cn ) {
     return CV_HAL_ERROR_OK;
 }
 
-#if defined __GNUC__ && !defined(__clang__)
-__attribute__((optimize("no-tree-vectorize")))
-#endif
 int merge32s(const int** src, int* dst, int len, int cn ) {
-    int k = cn % 4 ? cn % 4 : 4;
-    int i, j;
-    if( k == 1 )
+    int vl = 0;
+    if (cn == 1)
     {
         const int* src0 = src[0];
-        #if defined(__clang__)
-        #pragma clang loop vectorize(disable)
-        #endif
-        for( i = j = 0; i < len; i++, j += cn )
-            dst[j] = src0[i];
-    }
-    else if( k == 2 )
-    {
-        const int *src0 = src[0], *src1 = src[1];
-        i = j = 0;
-        #if defined(__clang__)
-        #pragma clang loop vectorize(disable)
-        #endif
-        for( ; i < len; i++, j += cn )
+        for (int i = 0; i < len; i += vl)
         {
-            dst[j] = src0[i];
-            dst[j+1] = src1[i];
+            vl = __riscv_vsetvl_e32m8(len - i);
+            __riscv_vse32_v_i32m8(dst + i, __riscv_vle32_v_i32m8(src0 + i, vl), vl);
         }
     }
-    else if( k == 3 )
+    else if (cn == 2)
+    {
+        const int *src0 = src[0], *src1 = src[1];
+        for (int i = 0; i < len; i += vl)
+        {
+            vl = __riscv_vsetvl_e32m4(len - i);
+            vint32m4x2_t seg = __riscv_vcreate_v_i32m4x2(
+                __riscv_vle32_v_i32m4(src0 + i, vl),
+                __riscv_vle32_v_i32m4(src1 + i, vl)
+            );
+            __riscv_vsseg2e32_v_i32m4x2(dst + i * cn, seg, vl);
+        }
+    }
+    else if (cn == 3)
     {
         const int *src0 = src[0], *src1 = src[1], *src2 = src[2];
-        i = j = 0;
-        #if defined(__clang__)
-        #pragma clang loop vectorize(disable)
-        #endif
-        for( ; i < len; i++, j += cn )
+        for (int i = 0; i < len; i += vl)
         {
-            dst[j] = src0[i];
-            dst[j+1] = src1[i];
-            dst[j+2] = src2[i];
+            vl = __riscv_vsetvl_e32m2(len - i);
+            vint32m2x3_t seg = __riscv_vcreate_v_i32m2x3(
+                __riscv_vle32_v_i32m2(src0 + i, vl),
+                __riscv_vle32_v_i32m2(src1 + i, vl),
+                __riscv_vle32_v_i32m2(src2 + i, vl)
+            );
+            __riscv_vsseg3e32_v_i32m2x3(dst + i * cn, seg, vl);
+        }
+    }
+    else if (cn == 4)
+    {
+        const int *src0 = src[0], *src1 = src[1], *src2 = src[2], *src3 = src[3];
+        for (int i = 0; i < len; i += vl)
+        {
+            vl = __riscv_vsetvl_e32m2(len - i);
+            vint32m2x4_t seg = __riscv_vcreate_v_i32m2x4(
+                __riscv_vle32_v_i32m2(src0 + i, vl),
+                __riscv_vle32_v_i32m2(src1 + i, vl),
+                __riscv_vle32_v_i32m2(src2 + i, vl),
+                __riscv_vle32_v_i32m2(src3 + i, vl)
+            );
+            __riscv_vsseg4e32_v_i32m2x4(dst + i * cn, seg, vl);
         }
     }
     else
     {
-        const int *src0 = src[0], *src1 = src[1], *src2 = src[2], *src3 = src[3];
-        i = j = 0;
-        #if defined(__clang__)
-        #pragma clang loop vectorize(disable)
-        #endif
-        for( ; i < len; i++, j += cn )
+        int k = 0;
+        for (; k <= cn - 4; k += 4)
         {
-            dst[j] = src0[i]; dst[j+1] = src1[i];
-            dst[j+2] = src2[i]; dst[j+3] = src3[i];
+            const int *src0 = src[k], *src1 = src[k + 1], *src2 = src[k + 2], *src3 = src[k + 3];
+            for (int i = 0; i < len; i += vl)
+            {
+                vl = __riscv_vsetvl_e32m2(len - i);
+                vint32m2x4_t seg = __riscv_vcreate_v_i32m2x4(
+                    __riscv_vle32_v_i32m2(src0 + i, vl),
+                    __riscv_vle32_v_i32m2(src1 + i, vl),
+                    __riscv_vle32_v_i32m2(src2 + i, vl),
+                    __riscv_vle32_v_i32m2(src3 + i, vl)
+                );
+                __riscv_vssseg4e32_v_i32m2x4(dst + k + i * cn, cn * sizeof(int), seg, vl);
+            }
         }
-    }
-    #if defined(__clang__)
-    #pragma clang loop vectorize(disable)
-    #endif
-    for( ; k < cn; k += 4 )
-    {
-        const int *src0 = src[k], *src1 = src[k+1], *src2 = src[k+2], *src3 = src[k+3];
-        for( i = 0, j = k; i < len; i++, j += cn )
+        for (; k < cn; ++k)
         {
-            dst[j] = src0[i]; dst[j+1] = src1[i];
-            dst[j+2] = src2[i]; dst[j+3] = src3[i];
+            const int* srcK = src[k];
+            for (int i = 0; i < len; i += vl)
+            {
+                vl = __riscv_vsetvl_e32m2(len - i);
+                vint32m2_t seg = __riscv_vle32_v_i32m2(srcK + i, vl);
+                __riscv_vsse32_v_i32m2(dst + k + i * cn, cn * sizeof(int), seg, vl);
+            }
         }
     }
     return CV_HAL_ERROR_OK;
 }
 
-#if defined __GNUC__ && !defined(__clang__)
-__attribute__((optimize("no-tree-vectorize")))
-#endif
 int merge64s(const int64** src, int64* dst, int len, int cn ) {
-    int k = cn % 4 ? cn % 4 : 4;
-    int i, j;
-    if( k == 1 )
+    int vl = 0;
+    if (cn == 1)
     {
         const int64* src0 = src[0];
-        #if defined(__clang__)
-        #pragma clang loop vectorize(disable)
-        #endif
-        for( i = j = 0; i < len; i++, j += cn )
-            dst[j] = src0[i];
-    }
-    else if( k == 2 )
-    {
-        const int64 *src0 = src[0], *src1 = src[1];
-        i = j = 0;
-        #if defined(__clang__)
-        #pragma clang loop vectorize(disable)
-        #endif
-        for( ; i < len; i++, j += cn )
+        for (int i = 0; i < len; i += vl)
         {
-            dst[j] = src0[i];
-            dst[j+1] = src1[i];
+            vl = __riscv_vsetvl_e64m8(len - i);
+            __riscv_vse64_v_i64m8(dst + i, __riscv_vle64_v_i64m8(src0 + i, vl), vl);
         }
     }
-    else if( k == 3 )
+    else if (cn == 2)
+    {
+        const int64 *src0 = src[0], *src1 = src[1];
+        for (int i = 0; i < len; i += vl)
+        {
+            vl = __riscv_vsetvl_e64m4(len - i);
+            vint64m4x2_t seg = __riscv_vcreate_v_i64m4x2(
+                __riscv_vle64_v_i64m4(src0 + i, vl),
+                __riscv_vle64_v_i64m4(src1 + i, vl)
+            );
+            __riscv_vsseg2e64_v_i64m4x2(dst + i * cn, seg, vl);
+        }
+    }
+    else if (cn == 3)
     {
         const int64 *src0 = src[0], *src1 = src[1], *src2 = src[2];
-        i = j = 0;
-        #if defined(__clang__)
-        #pragma clang loop vectorize(disable)
-        #endif
-        for( ; i < len; i++, j += cn )
+        for (int i = 0; i < len; i += vl)
         {
-            dst[j] = src0[i];
-            dst[j+1] = src1[i];
-            dst[j+2] = src2[i];
+            vl = __riscv_vsetvl_e64m2(len - i);
+            vint64m2x3_t seg = __riscv_vcreate_v_i64m2x3(
+                __riscv_vle64_v_i64m2(src0 + i, vl),
+                __riscv_vle64_v_i64m2(src1 + i, vl),
+                __riscv_vle64_v_i64m2(src2 + i, vl)
+            );
+            __riscv_vsseg3e64_v_i64m2x3(dst + i * cn, seg, vl);
+        }
+    }
+    else if (cn == 4)
+    {
+        const int64 *src0 = src[0], *src1 = src[1], *src2 = src[2], *src3 = src[3];
+        for (int i = 0; i < len; i += vl)
+        {
+            vl = __riscv_vsetvl_e64m2(len - i);
+            vint64m2x4_t seg = __riscv_vcreate_v_i64m2x4(
+                __riscv_vle64_v_i64m2(src0 + i, vl),
+                __riscv_vle64_v_i64m2(src1 + i, vl),
+                __riscv_vle64_v_i64m2(src2 + i, vl),
+                __riscv_vle64_v_i64m2(src3 + i, vl)
+            );
+            __riscv_vsseg4e64_v_i64m2x4(dst + i * cn, seg, vl);
         }
     }
     else
     {
-        const int64 *src0 = src[0], *src1 = src[1], *src2 = src[2], *src3 = src[3];
-        i = j = 0;
-        #if defined(__clang__)
-        #pragma clang loop vectorize(disable)
-        #endif
-        for( ; i < len; i++, j += cn )
+        int k = 0;
+        for (; k <= cn - 4; k += 4)
         {
-            dst[j] = src0[i]; dst[j+1] = src1[i];
-            dst[j+2] = src2[i]; dst[j+3] = src3[i];
+            const int64 *src0 = src[k], *src1 = src[k + 1], *src2 = src[k + 2], *src3 = src[k + 3];
+            for (int i = 0; i < len; i += vl)
+            {
+                vl = __riscv_vsetvl_e64m2(len - i);
+                vint64m2x4_t seg = __riscv_vcreate_v_i64m2x4(
+                    __riscv_vle64_v_i64m2(src0 + i, vl),
+                    __riscv_vle64_v_i64m2(src1 + i, vl),
+                    __riscv_vle64_v_i64m2(src2 + i, vl),
+                    __riscv_vle64_v_i64m2(src3 + i, vl)
+                );
+                __riscv_vssseg4e64_v_i64m2x4(dst + k + i * cn, cn * sizeof(int64), seg, vl);
+            }
         }
-    }
-    #if defined(__clang__)
-    #pragma clang loop vectorize(disable)
-    #endif
-    for( ; k < cn; k += 4 )
-    {
-        const int64 *src0 = src[k], *src1 = src[k+1], *src2 = src[k+2], *src3 = src[k+3];
-        for( i = 0, j = k; i < len; i++, j += cn )
+        for (; k < cn; ++k)
         {
-            dst[j] = src0[i]; dst[j+1] = src1[i];
-            dst[j+2] = src2[i]; dst[j+3] = src3[i];
+            const int64* srcK = src[k];
+            for (int i = 0; i < len; i += vl)
+            {
+                vl = __riscv_vsetvl_e64m2(len - i);
+                vint64m2_t seg = __riscv_vle64_v_i64m2(srcK + i, vl);
+                __riscv_vsse64_v_i64m2(dst + k + i * cn, cn * sizeof(int64), seg, vl);
+            }
         }
     }
     return CV_HAL_ERROR_OK;
 }
 
+
 #endif // CV_HAL_RVV_1P0_ENABLED
 
 }}} // cv::rvv_hal::core

hal/riscv-rvv/src/core/split.cpp

@@ -8,84 +8,92 @@ namespace cv { namespace rvv_hal { namespace core {
 
 #if CV_HAL_RVV_1P0_ENABLED
 
-int split8u(const uchar* src, uchar** dst, int len, int cn)
-{
-    int vl = 0;
-    if (cn == 1)
-    {
-        uchar* dst0 = dst[0];
-        for (int i = 0; i < len; i += vl)
-        {
-            vl = __riscv_vsetvl_e8m8(len - i);
-            __riscv_vse8_v_u8m8(dst0 + i, __riscv_vle8_v_u8m8(src + i, vl), vl);
-        }
-    }
-    else if (cn == 2)
-    {
-        uchar *dst0 = dst[0], *dst1 = dst[1];
-        for (int i = 0; i < len; i += vl)
-        {
-            vl = __riscv_vsetvl_e8m4(len - i);
-            vuint8m4x2_t seg = __riscv_vlseg2e8_v_u8m4x2(src + i * cn, vl);
-            __riscv_vse8_v_u8m4(dst0 + i, __riscv_vget_v_u8m4x2_u8m4(seg, 0), vl);
-            __riscv_vse8_v_u8m4(dst1 + i, __riscv_vget_v_u8m4x2_u8m4(seg, 1), vl);
-        }
-    }
-    else if (cn == 3)
-    {
-        uchar *dst0 = dst[0], *dst1 = dst[1], *dst2 = dst[2];
-        for (int i = 0; i < len; i += vl)
-        {
-            vl = __riscv_vsetvl_e8m2(len - i);
-            vuint8m2x3_t seg = __riscv_vlseg3e8_v_u8m2x3(src + i * cn, vl);
-            __riscv_vse8_v_u8m2(dst0 + i, __riscv_vget_v_u8m2x3_u8m2(seg, 0), vl);
-            __riscv_vse8_v_u8m2(dst1 + i, __riscv_vget_v_u8m2x3_u8m2(seg, 1), vl);
-            __riscv_vse8_v_u8m2(dst2 + i, __riscv_vget_v_u8m2x3_u8m2(seg, 2), vl);
-        }
-    }
-    else if (cn == 4)
-    {
-        uchar *dst0 = dst[0], *dst1 = dst[1], *dst2 = dst[2], *dst3 = dst[3];
-        for (int i = 0; i < len; i += vl)
-        {
-            vl = __riscv_vsetvl_e8m2(len - i);
-            vuint8m2x4_t seg = __riscv_vlseg4e8_v_u8m2x4(src + i * cn, vl);
-            __riscv_vse8_v_u8m2(dst0 + i, __riscv_vget_v_u8m2x4_u8m2(seg, 0), vl);
-            __riscv_vse8_v_u8m2(dst1 + i, __riscv_vget_v_u8m2x4_u8m2(seg, 1), vl);
-            __riscv_vse8_v_u8m2(dst2 + i, __riscv_vget_v_u8m2x4_u8m2(seg, 2), vl);
-            __riscv_vse8_v_u8m2(dst3 + i, __riscv_vget_v_u8m2x4_u8m2(seg, 3), vl);
-        }
-    }
-    else
-    {
-        int k = 0;
-        for (; k <= cn - 4; k += 4)
-        {
-            uchar *dst0 = dst[k], *dst1 = dst[k + 1], *dst2 = dst[k + 2], *dst3 = dst[k + 3];
-            for (int i = 0; i < len; i += vl)
-            {
-                vl = __riscv_vsetvl_e8m2(len - i);
-                vuint8m2x4_t seg = __riscv_vlsseg4e8_v_u8m2x4(src + k + i * cn, cn, vl);
-                __riscv_vse8_v_u8m2(dst0 + i, __riscv_vget_v_u8m2x4_u8m2(seg, 0), vl);
-                __riscv_vse8_v_u8m2(dst1 + i, __riscv_vget_v_u8m2x4_u8m2(seg, 1), vl);
-                __riscv_vse8_v_u8m2(dst2 + i, __riscv_vget_v_u8m2x4_u8m2(seg, 2), vl);
-                __riscv_vse8_v_u8m2(dst3 + i, __riscv_vget_v_u8m2x4_u8m2(seg, 3), vl);
-            }
-        }
-        for (; k < cn; ++k)
-        {
-            uchar* dstK = dst[k];
-            for (int i = 0; i < len; i += vl)
-            {
-                vl = __riscv_vsetvl_e8m2(len - i);
-                vuint8m2_t seg = __riscv_vlse8_v_u8m2(src + k + i * cn, cn, vl);
-                __riscv_vse8_v_u8m2(dstK + i, seg, vl);
-            }
-        }
-    }
-    return CV_HAL_ERROR_OK;
+#define OPENCV_HAL_IMPL_RVV_SPLIT(func, T, vtype, suffix, width) \
+int func(const T* src, T** dst, int len, int cn) \
+{ \
+    int vl = 0; \
+    if (cn == 1) \
+    { \
+        T* dst0 = dst[0]; \
+        for (int i = 0; i < len; i += vl) \
+        { \
+            vl = __riscv_vsetvl_e##width##m8(len - i); \
+            __riscv_vse##width##_v_##suffix##m8(dst0 + i, __riscv_vle##width##_v_##suffix##m8(src + i, vl), vl); \
+        } \
+    } \
+    else if (cn == 2) \
+    { \
+        T *dst0 = dst[0], *dst1 = dst[1]; \
+        for (int i = 0; i < len; i += vl) \
+        { \
+            vl = __riscv_vsetvl_e##width##m4(len - i); \
+            v##vtype##m4x2_t seg = __riscv_vlseg2e##width##_v_##suffix##m4x2(src + i * cn, vl); \
+            __riscv_vse##width##_v_##suffix##m4(dst0 + i, __riscv_vget_v_##suffix##m4x2_##suffix##m4(seg, 0), vl); \
+            __riscv_vse##width##_v_##suffix##m4(dst1 + i, __riscv_vget_v_##suffix##m4x2_##suffix##m4(seg, 1), vl); \
+        } \
+    } \
+    else if (cn == 3) \
+    { \
+        T *dst0 = dst[0], *dst1 = dst[1], *dst2 = dst[2]; \
+        for (int i = 0; i < len; i += vl) \
+        { \
+            vl = __riscv_vsetvl_e##width##m2(len - i); \
+            v##vtype##m2x3_t seg = __riscv_vlseg3e##width##_v_##suffix##m2x3(src + i * cn, vl); \
+            __riscv_vse##width##_v_##suffix##m2(dst0 + i, __riscv_vget_v_##suffix##m2x3_##suffix##m2(seg, 0), vl); \
+            __riscv_vse##width##_v_##suffix##m2(dst1 + i, __riscv_vget_v_##suffix##m2x3_##suffix##m2(seg, 1), vl); \
+            __riscv_vse##width##_v_##suffix##m2(dst2 + i, __riscv_vget_v_##suffix##m2x3_##suffix##m2(seg, 2), vl); \
+        } \
+    } \
+    else if (cn == 4) \
+    { \
+        T *dst0 = dst[0], *dst1 = dst[1], *dst2 = dst[2], *dst3 = dst[3]; \
+        for (int i = 0; i < len; i += vl) \
+        { \
+            vl = __riscv_vsetvl_e##width##m2(len - i); \
+            v##vtype##m2x4_t seg = __riscv_vlseg4e##width##_v_##suffix##m2x4(src + i * cn, vl); \
+            __riscv_vse##width##_v_##suffix##m2(dst0 + i, __riscv_vget_v_##suffix##m2x4_##suffix##m2(seg, 0), vl); \
+            __riscv_vse##width##_v_##suffix##m2(dst1 + i, __riscv_vget_v_##suffix##m2x4_##suffix##m2(seg, 1), vl); \
+            __riscv_vse##width##_v_##suffix##m2(dst2 + i, __riscv_vget_v_##suffix##m2x4_##suffix##m2(seg, 2), vl); \
+            __riscv_vse##width##_v_##suffix##m2(dst3 + i, __riscv_vget_v_##suffix##m2x4_##suffix##m2(seg, 3), vl); \
+        } \
+    } \
+    else \
+    { \
+        int k = 0; \
+        for (; k <= cn - 4; k += 4) \
+        { \
+            T *dst0 = dst[k], *dst1 = dst[k + 1], *dst2 = dst[k + 2], *dst3 = dst[k + 3]; \
+            for (int i = 0; i < len; i += vl) \
+            { \
+                vl = __riscv_vsetvl_e##width##m2(len - i); \
+                v##vtype##m2x4_t seg = __riscv_vlsseg4e##width##_v_##suffix##m2x4(src + k + i * cn, cn * sizeof(T), vl); \
+                __riscv_vse##width##_v_##suffix##m2(dst0 + i, __riscv_vget_v_##suffix##m2x4_##suffix##m2(seg, 0), vl); \
+                __riscv_vse##width##_v_##suffix##m2(dst1 + i, __riscv_vget_v_##suffix##m2x4_##suffix##m2(seg, 1), vl); \
+                __riscv_vse##width##_v_##suffix##m2(dst2 + i, __riscv_vget_v_##suffix##m2x4_##suffix##m2(seg, 2), vl); \
+                __riscv_vse##width##_v_##suffix##m2(dst3 + i, __riscv_vget_v_##suffix##m2x4_##suffix##m2(seg, 3), vl); \
+            } \
+        } \
+        for (; k < cn; ++k) \
+        { \
+            T* dstK = dst[k]; \
+            for (int i = 0; i < len; i += vl) \
+            { \
+                vl = __riscv_vsetvl_e##width##m2(len - i); \
+                v##vtype##m2_t seg = __riscv_vlse##width##_v_##suffix##m2(src + k + i * cn, cn * sizeof(T), vl); \
+                __riscv_vse##width##_v_##suffix##m2(dstK + i, seg, vl); \
+            } \
+        } \
+    } \
+    return CV_HAL_ERROR_OK; \
 }
 
+OPENCV_HAL_IMPL_RVV_SPLIT(split8u, uchar, uint8, u8, 8)
+OPENCV_HAL_IMPL_RVV_SPLIT(split16u, ushort, uint16, u16, 16)
+OPENCV_HAL_IMPL_RVV_SPLIT(split32s, int, int32, i32, 32)
+OPENCV_HAL_IMPL_RVV_SPLIT(split64s, int64, int64, i64, 64)
+
+#undef OPENCV_HAL_IMPL_RVV_SPLIT
+
 #endif // CV_HAL_RVV_1P0_ENABLED
 
 }}} // cv::rvv_hal::core

modules/core/include/opencv2/core/version.hpp

@@ -6,9 +6,9 @@
 #define OPENCV_VERSION_HPP
 
 #define CV_VERSION_MAJOR    4
-#define CV_VERSION_MINOR    14
+#define CV_VERSION_MINOR    15
 #define CV_VERSION_REVISION 0
-#define CV_VERSION_STATUS   "-pre"
+#define CV_VERSION_STATUS   "-dev"
 
 #define CVAUX_STR_EXP(__A)  #__A
 #define CVAUX_STR(__A)      CVAUX_STR_EXP(__A)