third_party/opencv
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

Merge pull request #7233 from tomoaki0705:featureUniversalIntrinsicFp16

Browse Source
This commit is contained in:
Alexander Alekhin
2016-09-07 14:42:46 +00:00
parent 48af5e5579 903789f7af
commit c8d77fd961
4 changed files with 149 additions and 36 deletions
Download Patch File Download Diff File Expand all files Collapse all files
modules/core/include/opencv2/core/hal/intrin_neon.hpp
+51
View File
@@ -275,6 +275,39 @@ struct v_float64x2
};
#endif
#if defined (HAVE_FP16)
// Workaround for old comiplers
template <typename T> static inline int16x4_t vreinterpret_s16_f16(T a)
{ return (int16x4_t)a; }
template <typename T> static inline float16x4_t vreinterpret_f16_s16(T a)
{ return (float16x4_t)a; }
template <typename T> static inline float16x4_t vld1_f16(const T* ptr)
{ return vreinterpret_f16_s16(vld1_s16((const short*)ptr)); }
template <typename T> static inline void vst1_f16(T* ptr, float16x4_t a)
{ vst1_s16((short*)ptr, vreinterpret_s16_f16(a)); }
static inline short vget_lane_f16(float16x4_t a, int b)
{ return vget_lane_s16(vreinterpret_s16_f16(a), b); }
struct v_float16x4
{
typedef short lane_type;
enum { nlanes = 4 };
v_float16x4() {}
explicit v_float16x4(float16x4_t v) : val(v) {}
v_float16x4(short v0, short v1, short v2, short v3)
{
short v[] = {v0, v1, v2, v3};
val = vld1_f16(v);
}
short get0() const
{
return vget_lane_f16(val, 0);
}
float16x4_t val;
};
#endif
#define OPENCV_HAL_IMPL_NEON_INIT(_Tpv, _Tp, suffix) \
inline v_##_Tpv v_setzero_##suffix() { return v_##_Tpv(vdupq_n_##suffix((_Tp)0)); } \
inline v_##_Tpv v_setall_##suffix(_Tp v) { return v_##_Tpv(vdupq_n_##suffix(v)); } \
@@ -734,6 +767,14 @@ OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(v_float32x4, float, f32)
OPENCV_HAL_IMPL_NEON_LOADSTORE_OP(v_float64x2, double, f64)
#endif
#if defined (HAVE_FP16)
// Workaround for old comiplers
inline v_float16x4 v_load_f16(const short* ptr)
{ return v_float16x4(vld1_f16(ptr)); }
inline void v_store_f16(short* ptr, v_float16x4& a)
{ vst1_f16(ptr, a.val); }
#endif
#define OPENCV_HAL_IMPL_NEON_REDUCE_OP_4(_Tpvec, scalartype, func, scalar_func) \
inline scalartype v_reduce_##func(const _Tpvec& a) \
{ \
@@ -1146,7 +1187,17 @@ inline v_float64x2 v_cvt_f64_high(const v_float32x4& a)
}
#endif
#if defined (HAVE_FP16)
inline v_float32x4 v_cvt_f32(const v_float16x4& a)
{
return v_float32x4(vcvt_f32_f16(a.val));
}
inline v_float16x4 v_cvt_f16(const v_float32x4& a)
{
return v_float16x4(vcvt_f16_f32(a.val));
}
#endif
//! @endcond
modules/core/include/opencv2/core/hal/intrin_sse.hpp
+39
View File
@@ -252,6 +252,26 @@ struct v_float64x2
__m128d val;
};
#if defined(HAVE_FP16)
struct v_float16x4
{
typedef short lane_type;
enum { nlanes = 4 };
v_float16x4() {}
explicit v_float16x4(__m128i v) : val(v) {}
v_float16x4(short v0, short v1, short v2, short v3)
{
val = _mm_setr_epi16(v0, v1, v2, v3, 0, 0, 0, 0);
}
short get0() const
{
return (short)_mm_cvtsi128_si32(val);
}
__m128i val;
};
#endif
#define OPENCV_HAL_IMPL_SSE_INITVEC(_Tpvec, _Tp, suffix, zsuffix, ssuffix, _Tps, cast) \
inline _Tpvec v_setzero_##suffix() { return _Tpvec(_mm_setzero_##zsuffix()); } \
inline _Tpvec v_setall_##suffix(_Tp v) { return _Tpvec(_mm_set1_##ssuffix((_Tps)v)); } \
@@ -1021,6 +1041,13 @@ inline void v_store_high(_Tp* ptr, const _Tpvec& a) \
OPENCV_HAL_IMPL_SSE_LOADSTORE_FLT_OP(v_float32x4, float, ps)
OPENCV_HAL_IMPL_SSE_LOADSTORE_FLT_OP(v_float64x2, double, pd)
#if defined(HAVE_FP16)
inline v_float16x4 v_load_f16(const short* ptr)
{ return v_float16x4(_mm_loadl_epi64((const __m128i*)ptr)); }
inline void v_store_f16(short* ptr, v_float16x4& a)
{ _mm_storel_epi64((__m128i*)ptr, a.val); }
#endif
#define OPENCV_HAL_IMPL_SSE_REDUCE_OP_4(_Tpvec, scalartype, func, scalar_func) \
inline scalartype v_reduce_##func(const _Tpvec& a) \
{ \
@@ -1626,6 +1653,18 @@ inline v_float64x2 v_cvt_f64_high(const v_float32x4& a)
return v_float64x2(_mm_cvtps_pd(_mm_castsi128_ps(_mm_srli_si128(_mm_castps_si128(a.val),8))));
}
#if defined(HAVE_FP16)
inline v_float32x4 v_cvt_f32(const v_float16x4& a)
{
return v_float32x4(_mm_cvtph_ps(a.val));
}
inline v_float16x4 v_cvt_f16(const v_float32x4& a)
{
return v_float16x4(_mm_cvtps_ph(a.val, 0));
}
#endif
//! @endcond
}