Merge pull request #23246 from mshabunin:rvv07-support

alalek · alalek · commit 426212785445 · 2023-02-20T18:06:30.000Z
diff --git a/modules/core/include/opencv2/core/hal/intrin_rvv.hpp b/modules/core/include/opencv2/core/hal/intrin_rvv.hpp
@@ -10,13 +10,27 @@
 
 #include <algorithm>
 
+// Building for T-Head C906 core with RVV 0.7.1 using toolchain
+// https://github.com/T-head-Semi/xuantie-gnu-toolchain
+// with option '-march=rv64gcv0p7'
+#ifdef __THEAD_VERSION__
+#   if __riscv_v == 7000
+#       include <fenv.h>
+#       define CV_RVV_THEAD_0_7
+#   endif
+#endif
+
 namespace cv
 {
 
 CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
 
 #define CV_SIMD128 1
-#define CV_SIMD128_64F 1
+#ifndef CV_RVV_THEAD_0_7
+#   define CV_SIMD128_64F 1
+#else
+#   define CV_SIMD128_64F 0
+#endif
 
 //////////// Unsupported native intrinsics in C++ ////////////
 // The following types have been defined in clang, but not in GCC yet.
@@ -1001,14 +1015,17 @@ OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint64x2, float32x4, u64, f32, u, f, 6
 OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(int8x16, float32x4, s8, f32, i, f, 8, 32)
 OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(int16x8, float32x4, s16, f32, i, f, 16, 32)
 OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(int64x2, float32x4, s64, f32, i, f, 64, 32)
+#if CV_SIMD128_64F
 OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint8x16, float64x2, u8, f64, u, f, 8, 64)
 OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint16x8, float64x2, u16, f64, u, f, 16, 64)
 OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(uint32x4, float64x2, u32, f64, u, f, 32, 64)
 OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(int8x16, float64x2, s8, f64, i, f, 8, 64)
 OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(int16x8, float64x2, s16, f64, i, f, 16, 64)
 OPENCV_HAL_IMPL_RVV_TWO_TIMES_REINTERPRET(int32x4, float64x2, s32, f64, i, f, 32, 64)
+#endif
 
 // Three times reinterpret
+#if CV_SIMD128_64F
 inline v_float32x4 v_reinterpret_as_f32(const v_float64x2& v) \
 { \
     return v_float32x4(vreinterpret_v_u32m1_f32m1(vreinterpret_v_u64m1_u32m1(vreinterpret_v_f64m1_u64m1(v))));\
@@ -1017,6 +1034,7 @@ inline v_float64x2 v_reinterpret_as_f64(const v_float32x4& v) \
 { \
     return v_float64x2(vreinterpret_v_u64m1_f64m1(vreinterpret_v_u32m1_u64m1(vreinterpret_v_f32m1_u32m1(v))));\
 }
+#endif
 
 ////////////// Extract //////////////
 
@@ -1920,13 +1938,15 @@ inline v_float64x2 v_muladd(const v_float64x2& a, const v_float64x2& b, const v_
 #define OPENCV_HAL_IMPL_RVV_CHECK_ALLANY(_Tpvec, suffix, shift, vl) \
 inline bool v_check_all(const _Tpvec& a) \
 { \
-    v_uint64x2 v = v_uint64x2(vreinterpret_v_##suffix##m1_u64m1(vsrl_vx_##suffix##m1(vnot_v_##suffix##m1(a, vl), shift, vl))); \
-    return (v.val[0] | v.val[1]) == 0; \
+    auto v0 = vsrl_vx_##suffix##m1(vnot_v_##suffix##m1(a, vl), shift, vl); \
+    v_uint32x4 v = v_uint32x4(v_reinterpret_as_u32(_Tpvec(v0))); \
+    return (v.val[0] | v.val[1] | v.val[2] | v.val[3]) == 0; \
 } \
 inline bool v_check_any(const _Tpvec& a) \
 { \
-    v_uint64x2 v = v_uint64x2(vreinterpret_v_##suffix##m1_u64m1(vsrl_vx_##suffix##m1(a, shift, vl))); \
-    return (v.val[0] | v.val[1]) != 0; \
+    auto v0 = vsrl_vx_##suffix##m1(a, shift, vl); \
+    v_uint32x4 v = v_uint32x4(v_reinterpret_as_u32(_Tpvec(v0))); \
+    return (v.val[0] | v.val[1] | v.val[2] | v.val[3]) != 0; \
 }
 
 OPENCV_HAL_IMPL_RVV_CHECK_ALLANY(v_uint8x16, u8, 7, 16)
@@ -2042,28 +2062,18 @@ OPENCV_HAL_IMPL_RVV_ABSDIFF(v_float64x2, absdiff)
 OPENCV_HAL_IMPL_RVV_ABSDIFF(v_int8x16, absdiffs)
 OPENCV_HAL_IMPL_RVV_ABSDIFF(v_int16x8, absdiffs)
 
-// use reinterpret instead of c-style casting.
-#ifndef __clang__
-#define OPENCV_HAL_IMPL_RVV_ABSDIFF_S(_Tpvec, _rTpvec, _nwTpvec, sub, rshr, width, vl) \
-inline _rTpvec v_absdiff(const _Tpvec& a, const _Tpvec& b) \
+#define OPENCV_HAL_IMPL_RVV_ABSDIFF_S(ivec, uvec, itype, utype, isuf, usuf, vlen) \
+inline uvec v_absdiff(const ivec& a, const ivec& b) \
 { \
-    return _rTpvec(rshr(vreinterpret_v_i##width##m2_u##width##m2(sub(v_max(a, b), v_min(a, b), vl)), 0, vl)); \
+    itype max = vmax_vv_##isuf(a, b, vlen); \
+    itype min = vmin_vv_##isuf(a, b, vlen); \
+    return uvec(vreinterpret_v_##isuf##_##usuf(vsub_vv_##isuf(max, min, vlen))); \
 }
 
-OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int8x16, v_uint8x16, vuint16m2_t, vwsub_vv_i16m2, vnclipu_wx_u8m1, 16, 16)
-OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int16x8, v_uint16x8, vuint32m2_t, vwsub_vv_i32m2, vnclipu_wx_u16m1, 32, 8)
-OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int32x4, v_uint32x4, vuint64m2_t, vwsub_vv_i64m2, vnclipu_wx_u32m1, 64, 4)
-#else
-#define OPENCV_HAL_IMPL_RVV_ABSDIFF_S(_Tpvec, _rTpvec, _nwTpvec, sub, rshr, width, vl) \
-inline _rTpvec v_absdiff(const _Tpvec& a, const _Tpvec& b) \
-{ \
-    return _rTpvec(rshr(vreinterpret_u##width##m2(sub(v_max(a, b), v_min(a, b), vl)), 0, vl)); \
-}
+OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int8x16, v_uint8x16, vint8m1_t, vuint8m1_t, i8m1, u8m1, 16)
+OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int16x8, v_uint16x8, vint16m1_t, vuint16m1_t, i16m1, u16m1, 8)
+OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int32x4, v_uint32x4, vint32m1_t, vuint32m1_t, i32m1, u32m1, 4)
 
-OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int8x16, v_uint8x16, vuint16m2_t, vwsub_vv_i16m2, vnclipu_wx_u8m1, 16, 16)
-OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int16x8, v_uint16x8, vuint32m2_t, vwsub_vv_i32m2, vnclipu_wx_u16m1, 32, 8)
-OPENCV_HAL_IMPL_RVV_ABSDIFF_S(v_int32x4, v_uint32x4, vuint64m2_t, vwsub_vv_i64m2, vnclipu_wx_u32m1, 64, 4)
-#endif
 #define OPENCV_HAL_IMPL_RVV_ABS(_Tprvec, _Tpvec, suffix) \
 inline _Tprvec v_abs(const _Tpvec& a) \
 { \
@@ -2902,7 +2912,14 @@ inline v_int32x4 v_ceil(const v_float32x4& a)
 
 inline v_int32x4 v_trunc(const v_float32x4& a)
 {
+#ifndef CV_RVV_THEAD_0_7
     return v_int32x4(vfcvt_rtz_x_f_v_i32m1(a, 4));
+#else
+    const int old_round = fesetround(FE_TOWARDZERO);
+    vint32m1_t val = vfcvt_x_f_v_i32m1(a, 4);
+    fesetround(old_round);
+    return v_int32x4(val);
+#endif
 }
 #if CV_SIMD128_64F
 #ifndef __clang__
@@ -2938,7 +2955,14 @@ inline v_int32x4 v_trunc(const v_float64x2& a)
 {
     double arr[4] = {a.val[0], a.val[1], 0, 0};
     vfloat64m2_t tmp = vle64_v_f64m2(arr, 4);
+#ifndef CV_RVV_THEAD_0_7
     return v_int32x4(vfncvt_rtz_x_f_w_i32m1(tmp, 4));
+#else
+    const int old_round = fesetround(FE_TOWARDZERO);
+    vint32m1_t val = vfncvt_x_f_w_i32m1(tmp, 4);
+    fesetround(old_round);
+    return v_int32x4(val);
+#endif
 }
 
 #else
diff --git a/modules/core/include/opencv2/core/hal/intrin_rvv071.hpp b/modules/core/include/opencv2/core/hal/intrin_rvv071.hpp
@@ -19,7 +19,7 @@ namespace cv
 CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
 
 #define CV_SIMD128 1
-#define CV_SIMD128_64F 1
+#define CV_SIMD128_64F 0
 //////////// Types ////////////
 struct v_uint8x16
 {
@@ -2021,23 +2021,18 @@ inline v_int32x4 v_load_expand_q(const schar* ptr)
     c = vwadd_vv_i32m2(vget_i16m2_i16m1(b, 0), vmv_v_x_i16m1(0, 4), 4);    \
     return v_int32x4(vget_i32m2_i32m1(c, 0));
 }
-#define VITL_16 (vuint64m2_t){0x1303120211011000, 0x1707160615051404, 0x1B0B1A0A19091808, 0x1F0F1E0E1D0D1C0C}
-#define VITL_8 (vuint64m2_t){0x0009000100080000, 0x000B0003000A0002, 0x000D0005000C0004, 0x000F0007000E0006}
-#define VITL_4 (vuint64m2_t){0x0000000400000000, 0x0000000500000001, 0x0000000600000002, 0x0000000700000003}
-#define VITL_2 (vuint64m2_t){0, 2, 1, 3}
-#define LOW_4  0x0000000100000000, 0x0000000500000004
-#define LOW_8  0x0003000200010000, 0x000B000A00090008
-#define LOW_16 0x0706050403020100, 0x1716151413121110
-#define HIGH_4  0x0000000300000002, 0x0000000700000006
-#define HIGH_8  0x0007000600050004, 0x000F000E000D000C
-#define HIGH_16 0x0F0E0D0C0B0A0908,  0x1F1E1D1C1B1A1918
+#define VITL_16 (vuint32m2_t){0x11011000, 0x13031202, 0x15051404, 0x17071606, 0x19091808, 0x1B0B1A0A, 0x1D0D1C0C, 0x1F0F1E0E}
+#define VITL_8 (vuint32m2_t){0x00080000, 0x00090001, 0x000A0002, 0x000B0003, 0x000C0004, 0x000D0005, 0x000E0006, 0x000F0007}
+#define VITL_4 (vuint32m2_t){0x00000000, 0x00000004, 0x00000001, 0x00000005, 0x00000002, 0x00000006, 0x00000003, 0x00000007}
+#define VITL_2 (vuint32m2_t){0, 0, 2, 0, 1, 0, 3, 0}
+
 #define OPENCV_HAL_IMPL_RISCVV_UNPACKS(_Tpvec, _Tp, _T, _UTp, _UT, num, num2, len, numh) \
 inline void v_zip(const v_##_Tpvec& a0, const v_##_Tpvec& a1, v_##_Tpvec& b0, v_##_Tpvec& b1) \
 { \
     v##_Tp##m2_t tmp = vundefined_##_T##m2();\
     tmp = vset_##_T##m2(tmp, 0, a0.val); \
     tmp = vset_##_T##m2(tmp, 1, a1.val); \
-    vuint64m2_t mask = VITL_##num;    \
+    vuint32m2_t mask = VITL_##num;    \
     tmp = (v##_Tp##m2_t)vrgather_vv_##_T##m2((v##_Tp##m2_t)tmp, (v##_UTp##m2_t)mask, num2);    \
     b0.val = vget_##_T##m2_##_T##m1(tmp, 0); \
     b1.val = vget_##_T##m2_##_T##m1(tmp, 1); \
