daiduo/TurboPFor-Integer-Compression
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

BitUtil: c/c++ header

Browse Source
This commit is contained in:
x
2019-12-19 18:45:13 +01:00
parent cfdfa1bfc0
commit cfdfe11189
1 changed files with 10 additions and 12 deletions
Download Patch File Download Diff File Expand all files Collapse all files

22
bitutil.h
View File

@ -115,6 +115,10 @@ static ALWAYS_INLINE __m128i mm_delta_epi32(__m128i v, __m128i sv) { return _mm_
static ALWAYS_INLINE __m128i mm_xore_epi16( __m128i v, __m128i sv) { return _mm_xor_si128(v, _mm_alignr_epi8(v, sv, 14)); } static ALWAYS_INLINE __m128i mm_xore_epi16( __m128i v, __m128i sv) { return _mm_xor_si128(v, _mm_alignr_epi8(v, sv, 14)); }
static ALWAYS_INLINE __m128i mm_xore_epi32( __m128i v, __m128i sv) { return _mm_xor_si128(v, _mm_alignr_epi8(v, sv, 12)); } static ALWAYS_INLINE __m128i mm_xore_epi32( __m128i v, __m128i sv) { return _mm_xor_si128(v, _mm_alignr_epi8(v, sv, 12)); }
#define MM_HDEC_EPI32(_v_,_sv_,_hop_) { _v_ = _hop_(_v_, _mm_slli_si128(_v_, 4)); _v_ = _hop_(mm_shuffle_nnnn_epi32(_sv_, 3), _hop_(_mm_slli_si128(_v_, 8), _v_)); }
static ALWAYS_INLINE __m128i mm_scan_epi32(__m128i v, __m128i sv) { MM_HDEC_EPI32(v,sv,_mm_add_epi32); return v; }
static ALWAYS_INLINE __m128i mm_xord_epi32(__m128i v, __m128i sv) { MM_HDEC_EPI32(v,sv,_mm_xor_si128); return v; }
#define MM_HDEC_EPI16(_v_,_sv_,_hop_) {\ #define MM_HDEC_EPI16(_v_,_sv_,_hop_) {\
_v_ = _hop_( _v_, _mm_slli_si128(_v_, 2));\ _v_ = _hop_( _v_, _mm_slli_si128(_v_, 2));\
_v_ = _hop_( _v_, _mm_slli_si128(_v_, 4));\ _v_ = _hop_( _v_, _mm_slli_si128(_v_, 4));\
@ -123,21 +127,15 @@ static ALWAYS_INLINE __m128i mm_xore_epi32( __m128i v, __m128i sv) { return _mm_
static ALWAYS_INLINE __m128i mm_scan_epi16(__m128i v, __m128i sv) { MM_HDEC_EPI16(v,sv,_mm_add_epi16); return v; } static ALWAYS_INLINE __m128i mm_scan_epi16(__m128i v, __m128i sv) { MM_HDEC_EPI16(v,sv,_mm_add_epi16); return v; }
static ALWAYS_INLINE __m128i mm_xord_epi16(__m128i v, __m128i sv) { MM_HDEC_EPI16(v,sv,_mm_xor_si128); return v; } static ALWAYS_INLINE __m128i mm_xord_epi16(__m128i v, __m128i sv) { MM_HDEC_EPI16(v,sv,_mm_xor_si128); return v; }
#elif defined(__SSE2__)
static ALWAYS_INLINE __m128i mm_delta_epi16(__m128i v, __m128i sv) { return _mm_sub_epi16(v, _mm_or_si128(_mm_srli_si128(sv, 14), _mm_slli_si128(v, 2))); }
static ALWAYS_INLINE __m128i mm_xore_epi16( __m128i v, __m128i sv) { return _mm_xor_epi16(v, _mm_or_si128(_mm_srli_si128(sv, 14), _mm_slli_si128(v, 2))); }
static ALWAYS_INLINE __m128i mm_delta_epi32(__m128i v, __m128i sv) { return _mm_sub_epi32(v, _mm_or_si128(_mm_srli_si128(sv, 12), _mm_slli_si128(v, 4))); }
static ALWAYS_INLINE __m128i mm_xore_epi32( __m128i v, __m128i sv) { return _mm_xor_epi32(v, _mm_or_si128(_mm_srli_si128(sv, 12), _mm_slli_si128(v, 4))); }
#endif
#if defined(__SSE2__) || defined(__ARM_NEON)
#define MM_HDEC_EPI32(_v_,_sv_,_hop_) { _v_ = _hop_(_v_, _mm_slli_si128(_v_, 4)); _v_ = _hop_(mm_shuffle_nnnn_epi32(_sv_, 3), _hop_(_mm_slli_si128(_v_, 8), _v_)); }
static ALWAYS_INLINE __m128i mm_scan_epi32(__m128i v, __m128i sv) { MM_HDEC_EPI32(v,sv,_mm_add_epi32); return v; }
static ALWAYS_INLINE __m128i mm_xord_epi32(__m128i v, __m128i sv) { MM_HDEC_EPI32(v,sv,_mm_xor_si128); return v; }
//-------- scan with vi delta > 0 ----------------------------- //-------- scan with vi delta > 0 -----------------------------
static ALWAYS_INLINE __m128i mm_scani_epi16(__m128i v, __m128i sv, __m128i vi) { return _mm_add_epi16(mm_scan_epi16(v, sv), vi); } static ALWAYS_INLINE __m128i mm_scani_epi16(__m128i v, __m128i sv, __m128i vi) { return _mm_add_epi16(mm_scan_epi16(v, sv), vi); }
static ALWAYS_INLINE __m128i mm_scani_epi32(__m128i v, __m128i sv, __m128i vi) { return _mm_add_epi32(mm_scan_epi32(v, sv), vi); } static ALWAYS_INLINE __m128i mm_scani_epi32(__m128i v, __m128i sv, __m128i vi) { return _mm_add_epi32(mm_scan_epi32(v, sv), vi); }
#elif defined(__SSE2__)
static ALWAYS_INLINE __m128i mm_delta_epi16(__m128i v, __m128i sv) { return _mm_sub_epi16(v, _mm_or_si128(_mm_srli_si128(sv, 14), _mm_slli_si128(v, 2))); }
static ALWAYS_INLINE __m128i mm_xore_epi16( __m128i v, __m128i sv) { return _mm_xor_si128(v, _mm_or_si128(_mm_srli_si128(sv, 14), _mm_slli_si128(v, 2))); }
static ALWAYS_INLINE __m128i mm_delta_epi32(__m128i v, __m128i sv) { return _mm_sub_epi32(v, _mm_or_si128(_mm_srli_si128(sv, 12), _mm_slli_si128(v, 4))); }
static ALWAYS_INLINE __m128i mm_xore_epi32( __m128i v, __m128i sv) { return _mm_xor_si128(v, _mm_or_si128(_mm_srli_si128(sv, 12), _mm_slli_si128(v, 4))); }
#endif #endif
//------------------ Horizontal OR ----------------------------------------------- //------------------ Horizontal OR -----------------------------------------------