potok/opencv
1
0
Fork 0
mirror of https://github.com/opencv/opencv.git synced 2026-07-30 15:53:03 +04:00
Code Packages Releases Wiki Activity

Merge remote-tracking branch 'upstream/3.4' into merge-3.4

Browse Source
This commit is contained in:
Alexander Alekhin
2021-02-12 21:20:54 +00:00
parent ba8d20e9ae fba70f7991
commit e5d78960c6
20 changed files with 286 additions and 218 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/include/opencv2/core/hal/intrin_neon.hpp
+37 -1
View File
@@ -62,6 +62,22 @@ CV_CPU_OPTIMIZATION_HAL_NAMESPACE_BEGIN
#define CV_SIMD128_64F 0
#endif
// The following macro checks if the code is being compiled for the
// AArch64 execution state of Armv8, to enable the 128-bit
// intrinsics. The macro `__ARM_64BIT_STATE` is the one recommended by
// the Arm C Language Extension (ACLE) specifications [1] to check the
// availability of 128-bit intrinsics, and it is supporrted by clang
// and gcc. The macro `_M_ARM64` is the equivalent one for Microsoft
// Visual Studio [2] .
//
// [1] https://developer.arm.com/documentation/101028/0012/13--Advanced-SIMD--Neon--intrinsics
// [2] https://docs.microsoft.com/en-us/cpp/preprocessor/predefined-macros
#if defined(__ARM_64BIT_STATE) || defined(_M_ARM64)
#define CV_NEON_AARCH64 1
#else
#define CV_NEON_AARCH64 0
#endif
// TODO
#define CV_NEON_DOT 0
@@ -726,41 +742,61 @@ inline v_float64x2 v_dotprod_expand(const v_int32x4& a, const v_int32x4& b,
// 16 >> 32
inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b)
{
#if CV_NEON_AARCH64
int32x4_t p = vmull_s16(vget_low_s16(a.val), vget_low_s16(b.val));
return v_int32x4(vmlal_high_s16(p, a.val, b.val));
#else
int16x4_t a0 = vget_low_s16(a.val);
int16x4_t a1 = vget_high_s16(a.val);
int16x4_t b0 = vget_low_s16(b.val);
int16x4_t b1 = vget_high_s16(b.val);
int32x4_t p = vmull_s16(a0, b0);
return v_int32x4(vmlal_s16(p, a1, b1));
#endif
}
inline v_int32x4 v_dotprod_fast(const v_int16x8& a, const v_int16x8& b, const v_int32x4& c)
{
#if CV_NEON_AARCH64
int32x4_t p = vmlal_s16(c.val, vget_low_s16(a.val), vget_low_s16(b.val));
return v_int32x4(vmlal_high_s16(p, a.val, b.val));
#else
int16x4_t a0 = vget_low_s16(a.val);
int16x4_t a1 = vget_high_s16(a.val);
int16x4_t b0 = vget_low_s16(b.val);
int16x4_t b1 = vget_high_s16(b.val);
int32x4_t p = vmlal_s16(c.val, a0, b0);
return v_int32x4(vmlal_s16(p, a1, b1));
#endif
}
// 32 >> 64
inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b)
{
#if CV_NEON_AARCH64
int64x2_t p = vmull_s32(vget_low_s32(a.val), vget_low_s32(b.val));
return v_int64x2(vmlal_high_s32(p, a.val, b.val));
#else
int32x2_t a0 = vget_low_s32(a.val);
int32x2_t a1 = vget_high_s32(a.val);
int32x2_t b0 = vget_low_s32(b.val);
int32x2_t b1 = vget_high_s32(b.val);
int64x2_t p = vmull_s32(a0, b0);
return v_int64x2(vmlal_s32(p, a1, b1));
#endif
}
inline v_int64x2 v_dotprod_fast(const v_int32x4& a, const v_int32x4& b, const v_int64x2& c)
{
#if CV_NEON_AARCH64
int64x2_t p = vmlal_s32(c.val, vget_low_s32(a.val), vget_low_s32(b.val));
return v_int64x2(vmlal_high_s32(p, a.val, b.val));
#else
int32x2_t a0 = vget_low_s32(a.val);
int32x2_t a1 = vget_high_s32(a.val);
int32x2_t b0 = vget_low_s32(b.val);
int32x2_t b1 = vget_high_s32(b.val);
int64x2_t p = vmlal_s32(c.val, a0, b0);
return v_int64x2(vmlal_s32(p, a1, b1));
#endif
}
// 8 >> 32
@@ -1292,7 +1328,7 @@ inline int64 v_reduce_sum(const v_int64x2& a)
#if CV_SIMD128_64F
inline double v_reduce_sum(const v_float64x2& a)
{
return vgetq_lane_f64(a.val, 0) + vgetq_lane_f64(a.val, 1);
return vaddvq_f64(a.val);
}
#endif