Merge pull request #26789 from amane-ame:minmax_hal_rvv

amane-ame · web-flow · commit 13b2caffe09f · 2025-01-31T14:26:49.000+03:00
Add RISC-V HAL implementation for minMaxIdx #26789 On the RISC-V platform, `minMaxIdx` cannot benefit from Universal Intrinsics because the UI-optimized `minMaxIdx` only supports `CV_SIMD128` (and does not accept `CV_SIMD_SCALABLE` for RVV). https://github.com/opencv/opencv/blob/1d701d1690b8cc9aa6b86744bffd5d9841ac6fd3/modules/core/src/minmax.cpp#L209-L214 This patch implements `minMaxIdx` function in RVV_HAL using native intrinsic, optimizing the performance for all data types with one channel. Tested on MUSE-PI for both gcc 14.2 and clang 20.0. ``` $ opencv_test_core --gtest_filter="*MinMaxLoc*" $ opencv_perf_core --gtest_filter="*minMaxLoc*" ``` <img width="1122" alt="Untitled" src="https://github.com/user-attachments/assets/6a246852-87af-42c5-a50b-c349c2765f3f" /> ### 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 - [ ] The PR is proposed to the proper branch - [ ] There is a reference to the original bug report and related work - [ ] There is accuracy test, performance test and test data in opencv_extra repository, if applicable Patch to opencv_extra has the same branch name. - [ ] The feature is well documented and sample code can be built with the project CMake
diff --git a/3rdparty/hal_rvv/hal_rvv.hpp b/3rdparty/hal_rvv/hal_rvv.hpp
@@ -22,6 +22,7 @@
 #if defined(__riscv_v) && __riscv_v == 1000000
 #include "hal_rvv_1p0/merge.hpp" // core
 #include "hal_rvv_1p0/mean.hpp" // core
+#include "hal_rvv_1p0/minmax.hpp" // core
 #include "hal_rvv_1p0/atan.hpp" // core
 #endif
 
diff --git a/3rdparty/hal_rvv/hal_rvv_1p0/minmax.hpp b/3rdparty/hal_rvv/hal_rvv_1p0/minmax.hpp
@@ -0,0 +1,335 @@
+// This file is part of OpenCV project.
+// It is subject to the license terms in the LICENSE file found in the top-level directory
+// of this distribution and at https://opencv.org/license.html.
+#ifndef OPENCV_HAL_RVV_MINMAXIDX_HPP_INCLUDED
+#define OPENCV_HAL_RVV_MINMAXIDX_HPP_INCLUDED
+
+#include <riscv_vector.h>
+
+namespace cv { namespace cv_hal_rvv {
+
+#undef cv_hal_minMaxIdx
+#define cv_hal_minMaxIdx cv::cv_hal_rvv::minMaxIdx
+#undef cv_hal_minMaxIdxMaskStep
+#define cv_hal_minMaxIdxMaskStep cv::cv_hal_rvv::minMaxIdx
+
+namespace
+{
+    template<typename T> struct rvv;
+
+    #define HAL_RVV_GENERATOR(T, EEW, TYPE, IS_U, EMUL, M_EMUL, B_LEN) \
+    template<> struct rvv<T> \
+    { \
+        using vec_t = v##IS_U##int##EEW##EMUL##_t; \
+        using bool_t = vbool##B_LEN##_t; \
+        static inline size_t vsetvlmax() { return __riscv_vsetvlmax_e##EEW##EMUL(); } \
+        static inline size_t vsetvl(size_t a) { return __riscv_vsetvl_e##EEW##EMUL(a); } \
+        static inline vec_t vmv_v_x(T a, size_t b) { return __riscv_vmv_v_x_##TYPE##EMUL(a, b); } \
+        static inline vec_t vle(const T* a, size_t b) { return __riscv_vle##EEW##_v_##TYPE##EMUL(a, b); } \
+        static inline vuint8##M_EMUL##_t vle_mask(const uchar* a, size_t b) { return __riscv_vle8_v_u8##M_EMUL(a, b); } \
+        static inline vec_t vmin_tu(vec_t a, vec_t b, vec_t c, size_t d) { return __riscv_vmin##IS_U##_tu(a, b, c, d); } \
+        static inline vec_t vmax_tu(vec_t a, vec_t b, vec_t c, size_t d) { return __riscv_vmax##IS_U##_tu(a, b, c, d); } \
+        static inline vec_t vmin_tumu(bool_t a, vec_t b, vec_t c, vec_t d, size_t e) { return __riscv_vmin##IS_U##_tumu(a, b, c, d, e); } \
+        static inline vec_t vmax_tumu(bool_t a, vec_t b, vec_t c, vec_t d, size_t e) { return __riscv_vmax##IS_U##_tumu(a, b, c, d, e); } \
+        static inline vec_t vredmin(vec_t a, vec_t b, size_t c) { return __riscv_vredmin##IS_U(a, b, c); } \
+        static inline vec_t vredmax(vec_t a, vec_t b, size_t c) { return __riscv_vredmax##IS_U(a, b, c); } \
+    };
+    HAL_RVV_GENERATOR(uchar , 8 , u8 , u, m1, m1 , 8 )
+    HAL_RVV_GENERATOR(schar , 8 , i8 ,  , m1, m1 , 8 )
+    HAL_RVV_GENERATOR(ushort, 16, u16, u, m1, mf2, 16)
+    HAL_RVV_GENERATOR(short , 16, i16,  , m1, mf2, 16)
+    #undef HAL_RVV_GENERATOR
+
+    #define HAL_RVV_GENERATOR(T, NAME, EEW, TYPE, IS_F, F_OR_S, F_OR_X, EMUL, M_EMUL, P_EMUL, B_LEN) \
+    template<> struct rvv<T> \
+    { \
+        using vec_t = v##NAME##EEW##EMUL##_t; \
+        using bool_t = vbool##B_LEN##_t; \
+        static inline size_t vsetvlmax() { return __riscv_vsetvlmax_e##EEW##EMUL(); } \
+        static inline size_t vsetvl(size_t a) { return __riscv_vsetvl_e##EEW##EMUL(a); } \
+        static inline vec_t vmv_v_x(T a, size_t b) { return __riscv_v##IS_F##mv_v_##F_OR_X##_##TYPE##EMUL(a, b); } \
+        static inline vuint32##P_EMUL##_t vid(size_t a) { return __riscv_vid_v_u32##P_EMUL(a); } \
+        static inline vuint32##P_EMUL##_t vundefined() { return __riscv_vundefined_u32##P_EMUL(); } \
+        static inline vec_t vle(const T* a, size_t b) { return __riscv_vle##EEW##_v_##TYPE##EMUL(a, b); } \
+        static inline vuint8##M_EMUL##_t vle_mask(const uchar* a, size_t b) { return __riscv_vle8_v_u8##M_EMUL(a, b); } \
+        static inline bool_t vmlt(vec_t a, vec_t b, size_t c) { return __riscv_vm##F_OR_S##lt(a, b, c); } \
+        static inline bool_t vmgt(vec_t a, vec_t b, size_t c) { return __riscv_vm##F_OR_S##gt(a, b, c); } \
+        static inline bool_t vmlt_mu(bool_t a, bool_t b, vec_t c, vec_t d, size_t e) { return __riscv_vm##F_OR_S##lt##_mu(a, b, c, d, e); } \
+        static inline bool_t vmgt_mu(bool_t a, bool_t b, vec_t c, vec_t d, size_t e) { return __riscv_vm##F_OR_S##gt##_mu(a, b, c, d, e); } \
+        static inline T vmv_x_s(vec_t a) { return __riscv_v##IS_F##mv_##F_OR_X(a); } \
+    };
+    HAL_RVV_GENERATOR(int   , int  , 32, i32,  , s, x, m4, m1 , m4, 8 )
+    HAL_RVV_GENERATOR(float , float, 32, f32, f, f, f, m4, m1 , m4, 8 )
+    HAL_RVV_GENERATOR(double, float, 64, f64, f, f, f, m4, mf2, m2, 16)
+    #undef HAL_RVV_GENERATOR
+}
+
+template<typename T>
+inline int minMaxIdxReadTwice(const uchar* src_data, size_t src_step, int width, int height, double* minVal, double* maxVal,
+                              int* minIdx, int* maxIdx, uchar* mask, size_t mask_step)
+{
+    int vlmax = rvv<T>::vsetvlmax();
+    auto vec_min = rvv<T>::vmv_v_x(std::numeric_limits<T>::max(), vlmax);
+    auto vec_max = rvv<T>::vmv_v_x(std::numeric_limits<T>::lowest(), vlmax);
+    T val_min, val_max;
+
+    if (mask)
+    {
+        for (int i = 0; i < height; i++)
+        {
+            const T* src_row = reinterpret_cast<const T*>(src_data + i * src_step);
+            const uchar* mask_row = mask + i * mask_step;
+            int vl;
+            for (int j = 0; j < width; j += vl)
+            {
+                vl = rvv<T>::vsetvl(width - j);
+                auto vec_src = rvv<T>::vle(src_row + j, vl);
+                auto vec_mask = rvv<T>::vle_mask(mask_row + j, vl);
+                auto bool_mask = __riscv_vmsne(vec_mask, 0, vl);
+                vec_min = rvv<T>::vmin_tumu(bool_mask, vec_min, vec_min, vec_src, vl);
+                vec_max = rvv<T>::vmax_tumu(bool_mask, vec_max, vec_max, vec_src, vl);
+            }
+        }
+
+        auto sc_minval = rvv<T>::vmv_v_x(std::numeric_limits<T>::max(), vlmax);
+        auto sc_maxval = rvv<T>::vmv_v_x(std::numeric_limits<T>::lowest(), vlmax);
+        sc_minval = rvv<T>::vredmin(vec_min, sc_minval, vlmax);
+        sc_maxval = rvv<T>::vredmax(vec_max, sc_maxval, vlmax);
+        val_min = __riscv_vmv_x(sc_minval);
+        val_max = __riscv_vmv_x(sc_maxval);
+
+        bool found_min = !minIdx, found_max = !maxIdx;
+        for (int i = 0; i < height && (!found_min || !found_max); i++)
+        {
+            const T* src_row = reinterpret_cast<const T*>(src_data + i * src_step);
+            const uchar* mask_row = mask + i * mask_step;
+            int vl;
+            for (int j = 0; j < width && (!found_min || !found_max); j += vl)
+            {
+                vl = rvv<T>::vsetvl(width - j);
+                auto vec_src = rvv<T>::vle(src_row + j, vl);
+                auto vec_mask = rvv<T>::vle_mask(mask_row + j, vl);
+                auto bool_mask = __riscv_vmsne(vec_mask, 0, vl);
+                auto bool_zero = __riscv_vmxor(bool_mask, bool_mask, vl);
+                if (!found_min)
+                {
+                    auto bool_minpos = __riscv_vmseq_mu(bool_mask, bool_zero, vec_src, val_min, vl);
+                    int index = __riscv_vfirst(bool_minpos, vl);
+                    if (index != -1)
+                    {
+                        found_min = true;
+                        minIdx[0] = i;
+                        minIdx[1] = j + index;
+                    }
+                }
+                if (!found_max)
+                {
+                    auto bool_maxpos = __riscv_vmseq_mu(bool_mask, bool_zero, vec_src, val_max, vl);
+                    int index = __riscv_vfirst(bool_maxpos, vl);
+                    if (index != -1)
+                    {
+                        found_max = true;
+                        maxIdx[0] = i;
+                        maxIdx[1] = j + index;
+                    }
+                }
+            }
+        }
+    }
+    else
+    {
+        for (int i = 0; i < height; i++)
+        {
+            const T* src_row = reinterpret_cast<const T*>(src_data + i * src_step);
+            int vl;
+            for (int j = 0; j < width; j += vl)
+            {
+                vl = rvv<T>::vsetvl(width - j);
+                auto vec_src = rvv<T>::vle(src_row + j, vl);
+                vec_min = rvv<T>::vmin_tu(vec_min, vec_min, vec_src, vl);
+                vec_max = rvv<T>::vmax_tu(vec_max, vec_max, vec_src, vl);
+            }
+        }
+
+        auto sc_minval = rvv<T>::vmv_v_x(std::numeric_limits<T>::max(), vlmax);
+        auto sc_maxval = rvv<T>::vmv_v_x(std::numeric_limits<T>::lowest(), vlmax);
+        sc_minval = rvv<T>::vredmin(vec_min, sc_minval, vlmax);
+        sc_maxval = rvv<T>::vredmax(vec_max, sc_maxval, vlmax);
+        val_min = __riscv_vmv_x(sc_minval);
+        val_max = __riscv_vmv_x(sc_maxval);
+
+        bool found_min = !minIdx, found_max = !maxIdx;
+        for (int i = 0; i < height && (!found_min || !found_max); i++)
+        {
+            const T* src_row = reinterpret_cast<const T*>(src_data + i * src_step);
+            int vl;
+            for (int j = 0; j < width && (!found_min || !found_max); j += vl)
+            {
+                vl = rvv<T>::vsetvl(width - j);
+                auto vec_src = rvv<T>::vle(src_row + j, vl);
+                if (!found_min)
+                {
+                    auto bool_minpos = __riscv_vmseq(vec_src, val_min, vl);
+                    int index = __riscv_vfirst(bool_minpos, vl);
+                    if (index != -1)
+                    {
+                        found_min = true;
+                        minIdx[0] = i;
+                        minIdx[1] = j + index;
+                    }
+                }
+                if (!found_max)
+                {
+                    auto bool_maxpos = __riscv_vmseq(vec_src, val_max, vl);
+                    int index = __riscv_vfirst(bool_maxpos, vl);
+                    if (index != -1)
+                    {
+                        found_max = true;
+                        maxIdx[0] = i;
+                        maxIdx[1] = j + index;
+                    }
+                }
+            }
+        }
+    }
+    if (minVal)
+    {
+        *minVal = val_min;
+    }
+    if (maxVal)
+    {
+        *maxVal = val_max;
+    }
+
+    return CV_HAL_ERROR_OK;
+}
+
+template<typename T>
+inline int minMaxIdxReadOnce(const uchar* src_data, size_t src_step, int width, int height, double* minVal, double* maxVal,
+                             int* minIdx, int* maxIdx, uchar* mask, size_t mask_step)
+{
+    int vlmax = rvv<T>::vsetvlmax();
+    auto vec_min = rvv<T>::vmv_v_x(std::numeric_limits<T>::max(), vlmax);
+    auto vec_max = rvv<T>::vmv_v_x(std::numeric_limits<T>::lowest(), vlmax);
+    auto vec_pos = rvv<T>::vid(vlmax);
+    auto vec_minpos = rvv<T>::vundefined(), vec_maxpos = rvv<T>::vundefined();
+    T val_min, val_max;
+
+    if (mask)
+    {
+        for (int i = 0; i < height; i++)
+        {
+            const T* src_row = reinterpret_cast<const T*>(src_data + i * src_step);
+            const uchar* mask_row = mask + i * mask_step;
+            int vl;
+            for (int j = 0; j < width; j += vl)
+            {
+                vl = rvv<T>::vsetvl(width - j);
+                auto vec_src = rvv<T>::vle(src_row + j, vl);
+                auto vec_mask = rvv<T>::vle_mask(mask_row + j, vl);
+                auto bool_mask = __riscv_vmsne(vec_mask, 0, vl);
+                auto bool_zero = __riscv_vmxor(bool_mask, bool_mask, vl);
+
+                auto bool_minpos = rvv<T>::vmlt_mu(bool_mask, bool_zero, vec_src, vec_min, vl);
+                auto bool_maxpos = rvv<T>::vmgt_mu(bool_mask, bool_zero, vec_src, vec_max, vl);
+                vec_minpos = __riscv_vmerge_tu(vec_minpos, vec_minpos, vec_pos, bool_minpos, vl);
+                vec_maxpos = __riscv_vmerge_tu(vec_maxpos, vec_maxpos, vec_pos, bool_maxpos, vl);
+
+                vec_min = __riscv_vmerge_tu(vec_min, vec_min, vec_src, bool_minpos, vl);
+                vec_max = __riscv_vmerge_tu(vec_max, vec_max, vec_src, bool_maxpos, vl);
+                vec_pos = __riscv_vadd(vec_pos, vl, vlmax);
+            }
+        }
+    }
+    else
+    {
+        for (int i = 0; i < height; i++)
+        {
+            const T* src_row = reinterpret_cast<const T*>(src_data + i * src_step);
+            int vl;
+            for (int j = 0; j < width; j += vl)
+            {
+                vl = rvv<T>::vsetvl(width - j);
+                auto vec_src = rvv<T>::vle(src_row + j, vl);
+
+                auto bool_minpos = rvv<T>::vmlt(vec_src, vec_min, vl);
+                auto bool_maxpos = rvv<T>::vmgt(vec_src, vec_max, vl);
+                vec_minpos = __riscv_vmerge_tu(vec_minpos, vec_minpos, vec_pos, bool_minpos, vl);
+                vec_maxpos = __riscv_vmerge_tu(vec_maxpos, vec_maxpos, vec_pos, bool_maxpos, vl);
+
+                vec_min = __riscv_vmerge_tu(vec_min, vec_min, vec_src, bool_minpos, vl);
+                vec_max = __riscv_vmerge_tu(vec_max, vec_max, vec_src, bool_maxpos, vl);
+                vec_pos = __riscv_vadd(vec_pos, vl, vlmax);
+            }
+        }
+    }
+
+    val_min = std::numeric_limits<T>::max();
+    val_max = std::numeric_limits<T>::lowest();
+    for (int i = 0; i < vlmax; i++)
+    {
+        if (val_min > rvv<T>::vmv_x_s(vec_min))
+        {
+            val_min = rvv<T>::vmv_x_s(vec_min);
+            if (minIdx)
+            {
+                minIdx[0] = __riscv_vmv_x(vec_minpos) / width;
+                minIdx[1] = __riscv_vmv_x(vec_minpos) % width;
+            }
+        }
+        if (val_max < rvv<T>::vmv_x_s(vec_max))
+        {
+            val_max = rvv<T>::vmv_x_s(vec_max);
+            if (maxIdx)
+            {
+                maxIdx[0] = __riscv_vmv_x(vec_maxpos) / width;
+                maxIdx[1] = __riscv_vmv_x(vec_maxpos) % width;
+            }
+        }
+        vec_min = __riscv_vslidedown(vec_min, 1, vlmax);
+        vec_max = __riscv_vslidedown(vec_max, 1, vlmax);
+        vec_minpos = __riscv_vslidedown(vec_minpos, 1, vlmax);
+        vec_maxpos = __riscv_vslidedown(vec_maxpos, 1, vlmax);
+    }
+    if (minVal)
+    {
+        *minVal = val_min;
+    }
+    if (maxVal)
+    {
+        *maxVal = val_max;
+    }
+
+    return CV_HAL_ERROR_OK;
+}
+
+inline int minMaxIdx(const uchar* src_data, size_t src_step, int width, int height, int depth, double* minVal, double* maxVal,
+                     int* minIdx, int* maxIdx, uchar* mask, size_t mask_step = 0)
+{
+    if (!mask_step)
+        mask_step = src_step;
+
+    switch (depth)
+    {
+    case CV_8UC1:
+        return minMaxIdxReadTwice<uchar>(src_data, src_step, width, height, minVal, maxVal, minIdx, maxIdx, mask, mask_step);
+    case CV_8SC1:
+        return minMaxIdxReadTwice<schar>(src_data, src_step, width, height, minVal, maxVal, minIdx, maxIdx, mask, mask_step);
+    case CV_16UC1:
+        return minMaxIdxReadTwice<ushort>(src_data, src_step, width, height, minVal, maxVal, minIdx, maxIdx, mask, mask_step);
+    case CV_16SC1:
+        return minMaxIdxReadTwice<short>(src_data, src_step, width, height, minVal, maxVal, minIdx, maxIdx, mask, mask_step);
+    case CV_32SC1:
+        return minMaxIdxReadOnce<int>(src_data, src_step, width, height, minVal, maxVal, minIdx, maxIdx, mask, mask_step);
+    case CV_32FC1:
+        return minMaxIdxReadOnce<float>(src_data, src_step, width, height, minVal, maxVal, minIdx, maxIdx, mask, mask_step);
+    case CV_64FC1:
+        return minMaxIdxReadOnce<double>(src_data, src_step, width, height, minVal, maxVal, minIdx, maxIdx, mask, mask_step);
+    }
+
+    return CV_HAL_ERROR_NOT_IMPLEMENTED;
+}
+
+}}
+
+#endif
diff --git a/modules/core/src/hal_replacement.hpp b/modules/core/src/hal_replacement.hpp
@@ -911,8 +911,26 @@ inline int hal_ni_gemm64fc(const double* src1, size_t src1_step, const double* s
 inline int hal_ni_minMaxIdx(const uchar* src_data, size_t src_step, int width, int height, int depth, double* minVal, double* maxVal,
                             int* minIdx, int* maxIdx, uchar* mask) { return CV_HAL_ERROR_NOT_IMPLEMENTED; }
 
+/**
+   @brief Finds the global minimum and maximum in an array.
+   @param src_data Source image
+   @param src_step Source image
+   @param width Source image dimensions
+   @param height Source image dimensions
+   @param depth Depth of source image
+   @param minVal Pointer to the returned global minimum and maximum in an array.
+   @param maxVal Pointer to the returned global minimum and maximum in an array.
+   @param minIdx Pointer to the returned minimum and maximum location.
+   @param maxIdx Pointer to the returned minimum and maximum location.
+   @param mask Specified array region.
+   @param mask_step Mask array step.
+*/
+inline int hal_ni_minMaxIdxMaskStep(const uchar* src_data, size_t src_step, int width, int height, int depth, double* minVal, double* maxVal,
+                                    int* minIdx, int* maxIdx, uchar* mask, size_t mask_step) { return CV_HAL_ERROR_NOT_IMPLEMENTED; }
+
 //! @cond IGNORED
 #define cv_hal_minMaxIdx hal_ni_minMaxIdx
+#define cv_hal_minMaxIdxMaskStep hal_ni_minMaxIdxMaskStep
 //! @endcond
 
 /**
diff --git a/modules/core/src/minmax.cpp b/modules/core/src/minmax.cpp
@@ -1514,10 +1514,19 @@ void cv::minMaxIdx(InputArray _src, double* minVal,
 
     if (src.dims <= 2)
     {
-        CALL_HAL(minMaxIdx, cv_hal_minMaxIdx, src.data, src.step, src.cols*cn, src.rows,
-                 src.depth(), minVal, maxVal, minIdx, maxIdx, mask.data);
+        if ((size_t)src.step == (size_t)mask.step)
+        {
+            CALL_HAL(minMaxIdx, cv_hal_minMaxIdx, src.data, src.step, src.cols*cn, src.rows,
+                     src.depth(), minVal, maxVal, minIdx, maxIdx, mask.data);
+        }
+        else
+        {
+            CALL_HAL(minMaxIdxMaskStep, cv_hal_minMaxIdxMaskStep, src.data, src.step, src.cols*cn, src.rows,
+                     src.depth(), minVal, maxVal, minIdx, maxIdx, mask.data, mask.step);
+        }
+
     }
-    else if (src.isContinuous())
+    else if (src.isContinuous() && mask.isContinuous())
     {
         int res = cv_hal_minMaxIdx(src.data, 0, (int)src.total()*cn, 1, src.depth(),
                                    minVal, maxVal, minIdx, maxIdx, mask.data);
