/** Copyright (C) powturbo 2013-2019 GPL v2 License This program is free software; you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation; either version 2 of the License, or (at your option) any later version. This program is distributed in the hope that it will be useful, but WITHOUT ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for more details. You should have received a copy of the GNU General Public License along with this program; if not, write to the Free Software Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA. - homepage : https://sites.google.com/site/powturbo/ - github : https://github.com/powturbo - twitter : https://twitter.com/powturbo - email : powturbo [_AT_] gmail [_DOT_] com **/ // "Floating Point + Integer Compression (All integer compression functions can be used for float/double and vice versa)" #ifndef USIZE #pragma warning( disable : 4005) #pragma warning( disable : 4090) #pragma warning( disable : 4068) #define BITUTIL_IN #include "conf.h" #include "vp4.h" #include "bitutil.h" #include "fp.h" //---------------------- template generation -------------------------------------------- #define VSIZE 128 #define P4ENC p4enc #define P4DEC p4dec #define P4ENCV p4enc #define P4DECV p4dec #define NL 18 #define N4 17 // must be > 16 #define N_0 3 #define N_1 4 #define N2 3 #define N3 5 #define USIZE 8 #include "fp.c" #define P4ENCV p4enc128v #define P4DECV p4dec128v #define N_0 3 #define N_1 5 #define N2 6 #define N3 12 #define USIZE 16 #include "fp.c" #define N_0 4 #define N_1 6 #define N2 6 // for seconds time series #define N3 10 #define USIZE 32 #include "fp.c" #define N_1 7 #define N2 6 // for seconds/milliseconds,... time series #define N3 12 #define N4 20 // must be > 16 #define USIZE 64 #include "fp.c" #else //-------------------------------------- Template functions ------------------------------------------------------------ #define XORENC( _u_, _pu_, _usize_) ((_u_)^(_pu_)) // xor predictor #define XORDEC( _u_, _pu_, _usize_) ((_u_)^(_pu_)) #define ZZAGENC(_u_, _pu_, _usize_) TEMPLATE2(zigzagenc,_usize_)((_u_)-(_pu_)) //zigzag predictor #define ZZAGDEC(_u_, _pu_, _usize_) (TEMPLATE2(zigzagdec,_usize_)(_u_)+(_pu_)) #define uint_t TEMPLATE3(uint, USIZE, _t) #define int_t TEMPLATE3(int, USIZE, _t) //-------- TurboPFor Zigzag of zigzag for unsorted/sorted integer/floating point array --------------------------------------- size_t TEMPLATE2(p4nzzenc128v,USIZE)(uint_t *in, size_t n, unsigned char *out, uint_t start) { uint_t _p[VSIZE+32], *ip, *p, pd = 0; unsigned char *op = out; #define FE(i,_usize_) { TEMPLATE3(uint, USIZE, _t) u = ip[i]; start = u-start; p[i] = ZZAGENC(start,pd,_usize_); pd = start; start = u; } for(ip = in; ip != in + (n&~(VSIZE-1)); ) { for(p = _p; p != &_p[VSIZE]; p+=4,ip+=4) { FE(0,USIZE); FE(1,USIZE); FE(2,USIZE); FE(3,USIZE); } op = TEMPLATE2(P4ENCV,USIZE)(_p, VSIZE, op); PREFETCH(ip+512,0); } if(n = (in+n)-ip) { for(p = _p; p != &_p[n]; p++,ip++) FE(0,USIZE); op = TEMPLATE2(P4ENC,USIZE)(_p, n, op); } return op - out; } size_t TEMPLATE2(p4nzzdec128v,USIZE)(unsigned char *in, size_t n, uint_t *out, uint_t start) { uint_t _p[VSIZE+32],*p, *op, pd=0; unsigned char *ip = in; #define FD(i,_usize_) { TEMPLATE3(uint, USIZE, _t) u = ZZAGDEC(p[i],start+pd,_usize_); op[i] = u; pd = u - start; start = u; } for(op = out; op != out+(n&~(VSIZE-1)); ) { PREFETCH(ip+512,0); for(ip = TEMPLATE2(P4DECV,USIZE)(ip, VSIZE, _p), p = _p; p != &_p[VSIZE]; p+=4,op+=4) { FD(0,USIZE); FD(1,USIZE); FD(2,USIZE); FD(3,USIZE); } } if(n = (out+n) - op) for(ip = TEMPLATE2(P4DEC,USIZE)(ip, n, _p), p = _p; p != &_p[n]; p++,op++) FD(0,USIZE); return ip - in; } /*---------------- TurboFloat XOR: last value Predictor with TurboPFor --------------------------------------------------------- Compress significantly (115% - 160%) better than Facebook's Gorilla algorithm for values BEST results are obtained with LOSSY COMPRESSION (using fppad32/fppad64 in bitutil.c) 1: XOR value with previous value. We have now leading (for common sign/exponent bits) + mantissa trailing zero bits 2: Eliminate the common block leading zeros of sign/exponent by shifting all values in the block to left 3: reverse values to bring the mantissa trailing zero bits to left for better compression with TurboPFor */ size_t TEMPLATE2(fpxenc,USIZE)(uint_t *in, size_t n, unsigned char *out, uint_t start) { uint_t _p[VSIZE+32], *ip, *p; unsigned char *op = out; #if defined(__AVX2__) && USIZE >= 32 #define _mm256_set1_epi64(a) _mm256_set1_epi64x(a) __m256i sv = TEMPLATE2(_mm256_set1_epi, USIZE)(start); #elif (defined(__SSSE3__) || defined(__ARM_NEON)) && (USIZE == 16 || USIZE == 32) #define _mm_set1_epi64(a) _mm_set1_epi64x(a) __m128i sv = TEMPLATE2(_mm_set1_epi, USIZE)(start); #endif #define FE(i,_usize_) { TEMPLATE3(uint, _usize_, _t) u = ip[i]; p[i] = XORENC(u, start,_usize_); b |= p[i]; start = u; } for(ip = in; ip != in + (n&~(VSIZE-1)); ) { uint_t b = 0; #if defined(__AVX2__) && USIZE >= 32 __m256i bv = _mm256_setzero_si256(); for(p = _p; p != &_p[VSIZE]; p+=64/(USIZE/8),ip+=64/(USIZE/8)) { __m256i v0 = _mm256_loadu_si256((__m256i *) ip); __m256i v1 = _mm256_loadu_si256((__m256i *)(ip+32/(USIZE/8))); sv = TEMPLATE2(mm256_xore_epi, USIZE)(v0,sv); bv = _mm256_or_si256(bv, sv); _mm256_storeu_si256((__m256i *) p, sv); sv = v0; sv = TEMPLATE2(mm256_xore_epi, USIZE)(v1,sv); bv = _mm256_or_si256(bv, sv); _mm256_storeu_si256((__m256i *)(p+32/(USIZE/8)), sv); sv = v1; } start = (uint_t)TEMPLATE2(_mm256_extract_epi,USIZE)(sv, 256/USIZE-1); b = TEMPLATE2(mm256_hor_epi, USIZE)(bv); #elif (defined(__SSSE3__) || defined(__ARM_NEON)) && (USIZE == 16 || USIZE == 32) __m128i bv = _mm_setzero_si128(); for(p = _p; p != &_p[VSIZE]; p+=32/(USIZE/8),ip+=32/(USIZE/8)) { __m128i v0 = _mm_loadu_si128((__m128i *) ip); __m128i v1 = _mm_loadu_si128((__m128i *)(ip+16/(USIZE/8))); sv = TEMPLATE2(mm_xore_epi, USIZE)(v0,sv); bv = _mm_or_si128(bv, sv); _mm_storeu_si128((__m128i *) p, sv); sv = v0; sv = TEMPLATE2(mm_xore_epi, USIZE)(v1,sv); bv = _mm_or_si128(bv, sv); _mm_storeu_si128((__m128i *)(p+16/(USIZE/8)), sv); sv = v1; } start = (uint_t)TEMPLATE2(_mm_cvtsi128_si,USIZE)(_mm_srli_si128(sv,16-USIZE/8)); b = TEMPLATE2(mm_hor_epi, USIZE)(bv); #else for(p = _p; p != &_p[VSIZE]; p+=4,ip+=4) { FE(0,USIZE); FE(1,USIZE); FE(2,USIZE); FE(3,USIZE); } #endif *op++ = b = TEMPLATE2(clz,USIZE)(b); #define TR(i,_usize_) p[i] = TEMPLATE2(rbit,_usize_)(p[i]<= 32 for(p = _p; p != &_p[VSIZE]; p+=64/(USIZE/8)) { __m256i v0 = _mm256_loadu_si256((__m256i *)p); __m256i v1 = _mm256_loadu_si256((__m256i *)(p+32/(USIZE/8))); v0 = TEMPLATE2(_mm256_slli_epi, USIZE)(v0,b); v1 = TEMPLATE2(_mm256_slli_epi, USIZE)(v1,b); v0 = TEMPLATE2( mm256_rbit_epi, USIZE)(v0); v1 = TEMPLATE2( mm256_rbit_epi, USIZE)(v1); _mm256_storeu_si256((__m256i *) p, v0); _mm256_storeu_si256((__m256i *)(p+32/(USIZE/8)), v1); } #elif (defined(__SSSE3__) || defined(__ARM_NEON)) && (USIZE == 16 || USIZE == 32) for(p = _p; p != &_p[VSIZE]; p+=32/(USIZE/8)) { __m128i v0 = _mm_loadu_si128((__m128i *) p); __m128i v1 = _mm_loadu_si128((__m128i *)(p+16/(USIZE/8))); v0 = TEMPLATE2(_mm_slli_epi, USIZE)(v0,b); v0 = TEMPLATE2( mm_rbit_epi, USIZE)(v0); v1 = TEMPLATE2(_mm_slli_epi, USIZE)(v1,b); v1 = TEMPLATE2( mm_rbit_epi, USIZE)(v1); _mm_storeu_si128((__m128i *) p, v0); _mm_storeu_si128((__m128i *)(p+16/(USIZE/8)), v1); } #else for(p = _p; p != &_p[VSIZE]; p+=4) { TR(0,USIZE); TR(1,USIZE); TR(2,USIZE); TR(3,USIZE); } #endif op = TEMPLATE2(P4ENCV,USIZE)(_p, VSIZE, op); PREFETCH(ip+512,0); } if(n = (in+n)-ip) { uint_t b = 0; for(p = _p; p != &_p[n]; p++,ip++) FE(0,USIZE); b = TEMPLATE2(clz,USIZE)(b); *op++ = b; for(p = _p; p != &_p[n]; p++) TR(0,USIZE); op = TEMPLATE2(P4ENC,USIZE)(_p, n, op); } return op - out; } size_t TEMPLATE2(fpxdec,USIZE)(unsigned char *in, size_t n, uint_t *out, uint_t start) { uint_t *op, _p[VSIZE+32],*p; unsigned char *ip = in; #if defined(__AVX2__) && USIZE >= 32 #define _mm256_set1_epi64(a) _mm256_set1_epi64x(a) __m256i sv = TEMPLATE2(_mm256_set1_epi, USIZE)(start); #elif (defined(__SSSE3__) || defined(__ARM_NEON)) && (USIZE == 16 || USIZE == 32) #define _mm_set1_epi64(a) _mm_set1_epi64x(a) __m128i sv = TEMPLATE2(_mm_set1_epi, USIZE)(start); #endif #define FD(i,_usize_) { TEMPLATE3(uint, USIZE, _t) u = p[i]; u = TEMPLATE2(rbit,_usize_)(u)>>b; u = XORDEC(u, start,_usize_); op[i] = start = u; } for(op = out; op != out+(n&~(VSIZE-1)); ) { PREFETCH(ip+512,0); unsigned b = *ip++; ip = TEMPLATE2(P4DECV,USIZE)(ip, VSIZE, _p); #if defined(__AVX2__) && USIZE >= 32 for(p = _p; p != &_p[VSIZE]; p+=64/(USIZE/8),op+=64/(USIZE/8)) { __m256i v0 = _mm256_loadu_si256((__m256i *)p); __m256i v1 = _mm256_loadu_si256((__m256i *)(p+32/(USIZE/8))); v0 = TEMPLATE2( mm256_rbit_epi, USIZE)(v0); v1 = TEMPLATE2( mm256_rbit_epi, USIZE)(v1); v0 = TEMPLATE2(_mm256_srli_epi, USIZE)(v0,b); v1 = TEMPLATE2(_mm256_srli_epi, USIZE)(v1,b); v0 = TEMPLATE2( mm256_xord_epi, USIZE)(v0,sv); sv = TEMPLATE2( mm256_xord_epi, USIZE)(v1,v0); _mm256_storeu_si256((__m256i *)op, v0); _mm256_storeu_si256((__m256i *)(op+32/(USIZE/8)), sv); } start = (uint_t)TEMPLATE2(_mm256_extract_epi,USIZE)(sv, 256/USIZE-1); #elif (defined(__SSSE3__) || defined(__ARM_NEON)) && (USIZE == 16 || USIZE == 32) for(p = _p; p != &_p[VSIZE]; p+=32/(USIZE/8),op+=32/(USIZE/8)) { __m128i v0 = _mm_loadu_si128((__m128i *)p); __m128i v1 = _mm_loadu_si128((__m128i *)(p+16/(USIZE/8))); v0 = TEMPLATE2( mm_rbit_epi, USIZE)(v0); v0 = TEMPLATE2(_mm_srli_epi, USIZE)(v0,b); v0 = TEMPLATE2( mm_xord_epi, USIZE)(v0,sv); v1 = TEMPLATE2( mm_rbit_epi, USIZE)(v1); v1 = TEMPLATE2(_mm_srli_epi, USIZE)(v1,b); sv = TEMPLATE2( mm_xord_epi, USIZE)(v1,v0); _mm_storeu_si128((__m128i *) op, v0); _mm_storeu_si128((__m128i *)(op+16/(USIZE/8)), sv); } start = (uint_t)TEMPLATE2(_mm_cvtsi128_si,USIZE)(_mm_srli_si128(sv,16-USIZE/8)); #else for(p = _p; p != &_p[VSIZE]; p+=4,op+=4) { FD(0,USIZE); FD(1,USIZE); FD(2,USIZE); FD(3,USIZE); } #endif } if(n = (out+n) - op) { uint_t b = *ip++; for(ip = TEMPLATE2(P4DEC,USIZE)(ip, n, _p), p = _p; p < &_p[n]; p++,op++) FD(0,USIZE); } return ip - in; } //-------- TurboFloat FCM: Finite Context Method Predictor --------------------------------------------------------------- #define HBITS 13 //15 #define HASH64(_h_,_u_) (((_h_)<<5 ^ (_u_)>>50) & ((1u<>23) & ((1u<>12) & ((1u<> 5) & ((1u<= 32 #define _mm256_set1_epi64(a) _mm256_set1_epi64x(a) __m256i sv = TEMPLATE2(_mm256_set1_epi, USIZE)(start); #elif (defined(__SSSE3__) || defined(__ARM_NEON)) && (USIZE == 16 || USIZE == 32) #define _mm_set1_epi64(a) _mm_set1_epi64x(a) __m128i sv = TEMPLATE2(_mm_set1_epi, USIZE)(start); #endif for(ip = in; ip != in + (n&~(VSIZE-1)); ) { uint_t b = 0; #define FE(i,_usize_) { TEMPLATE3(uint, _usize_, _t) u = ip[i]; p[i] = XORENC(u, htab[h],_usize_); b |= p[i]; htab[h] = u; h = TEMPLATE2(HASH,_usize_)(h,u); } for(p = _p; p != &_p[VSIZE]; p+=4,ip+=4) { FE(0,USIZE); FE(1,USIZE); FE(2,USIZE); FE(3,USIZE); } *op++ = b = TEMPLATE2(clz,USIZE)(b); #if defined(__AVX2__) && USIZE >= 32 for(p = _p; p != &_p[VSIZE]; p+=64/(USIZE/8)) { __m256i v0 = _mm256_loadu_si256((__m256i *)p); __m256i v1 = _mm256_loadu_si256((__m256i *)(p+32/(USIZE/8))); v0 = TEMPLATE2(_mm256_slli_epi, USIZE)(v0,b); v1 = TEMPLATE2(_mm256_slli_epi, USIZE)(v1,b); v0 = TEMPLATE2( mm256_rbit_epi, USIZE)(v0); v1 = TEMPLATE2( mm256_rbit_epi, USIZE)(v1); _mm256_storeu_si256((__m256i *) p, v0); _mm256_storeu_si256((__m256i *)(p+32/(USIZE/8)), v1); } #elif (defined(__SSSE3__) || defined(__ARM_NEON)) && (USIZE == 16 || USIZE == 32) for(p = _p; p != &_p[VSIZE]; p+=32/(USIZE/8)) { __m128i v0 = _mm_loadu_si128((__m128i *) p); __m128i v1 = _mm_loadu_si128((__m128i *)(p+16/(USIZE/8))); v0 = TEMPLATE2(_mm_slli_epi, USIZE)(v0,b); v0 = TEMPLATE2( mm_rbit_epi, USIZE)(v0); v1 = TEMPLATE2(_mm_slli_epi, USIZE)(v1,b); v1 = TEMPLATE2( mm_rbit_epi, USIZE)(v1); _mm_storeu_si128((__m128i *) p, v0); _mm_storeu_si128((__m128i *)(p+16/(USIZE/8)), v1); } #else #define TR(i,_usize_) p[i] = TEMPLATE2(rbit,_usize_)(p[i]<>b;\ u = XORDEC(u, htab[h], _usize_); op[i] = u; htab[h] = u; h = TEMPLATE2(HASH,_usize_)(h,u);\ } for(op = (uint_t*)out; op != out+(n&~(VSIZE-1)); ) { PREFETCH(ip+512,0); unsigned b = *ip++; ip = TEMPLATE2(P4DECV,USIZE)(ip, VSIZE, _p); for(p = _p; p != &_p[VSIZE]; p+=4,op+=4) { FD(0,USIZE); FD(1,USIZE); FD(2,USIZE); FD(3,USIZE); } } if(n = ((uint_t *)out+n) - op) { unsigned b = *ip++; ip = TEMPLATE2(P4DEC,USIZE)(ip, n, _p); for(p = _p; p != &_p[n]; p++,op++) FD(0,USIZE); } return ip - in; } //-------- TurboFloat DFCM: Differential Finite Context Method Predictor ---------------------------------------------------------- size_t TEMPLATE2(fpdfcmenc,USIZE)(uint_t *in, size_t n, unsigned char *out, uint_t start) { uint_t *ip, _p[VSIZE+32], h = 0, *p, htab[1<>b; u = XORDEC(u, (htab[h]+start),_usize_); \ op[i] = u; htab[h] = start = u-start; h = TEMPLATE2(HASH,_usize_)(h,start); start = u;\ } for(op = (uint_t*)out; op != out+(n&~(VSIZE-1)); ) { PREFETCH(ip+512,0); uint_t b = *ip++; ip = TEMPLATE2(P4DECV,USIZE)(ip, VSIZE, _p); for(p = _p; p != &_p[VSIZE]; p+=4,op+=4) { FD(0,USIZE); FD(1,USIZE); FD(2,USIZE); FD(3,USIZE); } } if(n = ((uint_t *)out+n) - op) { uint_t b = *ip++; ip = TEMPLATE2(P4DEC,USIZE)(ip, n, _p); for(p = _p; p != &_p[n]; p++,op++) FD(0,USIZE); } return ip - in; } //-------- TurboFloat 2D DFCM: Differential Finite Context Method Predictor ---------------------------------------------------------- size_t TEMPLATE2(fp2dfcmenc,USIZE)(uint_t *in, size_t n, unsigned char *out, uint_t start) { uint_t *ip, _p[VSIZE+32], h = 0, *p, htab[1<>b; u = XORDEC(u, (htab[h]+start),_usize_);\ op[i] = u; htab[h] = start = u-start; h = TEMPLATE2(HASH,_usize_)(h,start); start = start0; start0 = u;\ } for(op = (uint_t*)out; op != out+(n&~(VSIZE-1)); ) { PREFETCH(ip+512,0); uint_t b = *ip++; ip = TEMPLATE2(P4DECV,USIZE)(ip, VSIZE, _p); for(p = _p; p != &_p[VSIZE]; p+=4,op+=4) { FD(0,USIZE); FD(1,USIZE); FD(2,USIZE); FD(3,USIZE); } } if(n = ((uint_t *)out+n) - op) { uint_t b = *ip++; ip = TEMPLATE2(P4DEC,USIZE)(ip, n, _p); for(p = _p; p != &_p[n]; p++,op++) FD(0,USIZE); } return ip - in; } //-------- TurboGorilla : Improved Gorilla style (see Facebook paper) Floating point compression with bitio ------------------------------------ #define bitput2(_bw_,_br_, _n1_, _n2_, _x_) {\ if(!_x_) bitput(_bw_,_br_, 1, 1);/*1*/\ else if( _x_ < (1<< (_n1_-1))) bitput(_bw_,_br_, _n1_+2,_x_<<2|2);/*10*/\ else bitput(_bw_,_br_, _n2_+2,_x_<<2 );/*00*/\ } #define bitget2(_bw_,_br_, _n1_, _n2_, _x_) { _x_ = bitbw(_bw_,_br_);\ if(_x_ & 1) bitrmv(_bw_,_br_, 0+1), _x_ = 0;\ else if(_x_ & 2) bitrmv(_bw_,_br_,_n1_+2), _x_ = BZHI32(_x_>>2, _n1_);\ else bitrmv(_bw_,_br_,_n2_+2), _x_ = BZHI32(_x_>>2, _n2_);\ } #define BSIZE(_usize_) (_usize_==64?6:(_usize_==32?5:(_usize_==16?4:3))) size_t TEMPLATE2(fpgenc,USIZE)(uint_t *in, size_t n, unsigned char *out, uint_t start) { uint_t *ip; unsigned ol = 0,ot = 0; unsigned char *op = out; bitdef(bw,br); if(start) { ol = TEMPLATE2(clz,USIZE)(start); ot = TEMPLATE2(ctz,USIZE)(start); } #define FE(i,_usize_) { TEMPLATE3(uint, _usize_, _t) z = XORENC(ip[i], start,_usize_); start = ip[i];\ if(likely(!z)) bitput( bw,br, 1, 1);\ else { unsigned t = TEMPLATE2(ctz,_usize_)(z), l = TEMPLATE2(clz,_usize_)(z);\ unsigned s = _usize_ - l - t, os = _usize_ - ol - ot;\ if(l >= ol && t >= ot && os < 6+5+s) { bitput( bw,br, 2, 2); TEMPLATE2(bitput,_usize_)(bw,br, os, z>>ot,op); }\ else { bitput( bw,br, 2+BSIZE(_usize_), l<<2); bitput2(bw,br, N_0, N_1, t); bitenorm(bw,br,op);TEMPLATE2(bitput,_usize_)(bw,br, s, z>>t,op); ol = l; ot = t; }\ } bitenorm(bw,br,op);\ } for(ip = in; ip != in + (n&~(4-1)); ip+=4) { PREFETCH(ip+512,0); FE(0,USIZE); FE(1,USIZE); FE(2,USIZE); FE(3,USIZE); } for( ; ip != in + n ; ip++) FE(0,USIZE); bitflush(bw,br,op); return op - out; } size_t TEMPLATE2(fpgdec,USIZE)(unsigned char *in, size_t n, uint_t *out, uint_t start) { if(!n) return 0; uint_t *op; unsigned ol = 0,ot = 0,x; unsigned char *ip = in; bitdef(bw,br); if(start) { ol = TEMPLATE2(clz,USIZE)(start); ot = TEMPLATE2(ctz,USIZE)(start); } #define FD(i,_usize_) { TEMPLATE3(uint, _usize_, _t) z=0; unsigned _x; BITGET32(bw,br,1,_x); \ if(likely(!_x)) { BITGET32(bw,br,1,_x);\ if(!_x) { BITGET32(bw,br,BSIZE(_usize_),ol); bitget2(bw,br, N_0, N_1, ot); bitdnorm(bw,br,ip); }\ TEMPLATE2(bitget,_usize_)(bw,br,_usize_ - ol - ot,z,ip);\ z<<=ot;\ } op[i] = start = XORDEC(z, start,_usize_); bitdnorm(bw,br,ip);\ } for(bitdnorm(bw,br,ip),op = out; op != out+(n&~(4-1)); op+=4) { FD(0,USIZE); FD(1,USIZE); FD(2,USIZE); FD(3,USIZE); PREFETCH(ip+512,0); } for( ; op != out+n; op++) FD(0,USIZE); bitalign(bw,br,ip); return ip - in; } //------ Zigzag of zigzag with bitio for timestamps with bitio ------------------------------------------------------------------------------------------ // Improved Gorilla style compression with sliding zigzag of delta + RLE + overflow handling for timestamps in time series. // More than 300 times better compression and several times faster #define OVERFLOW if(op >= out_) { *out++ = 1<<4; /*bitini(bw,br); bitput(bw,br,4+3,1<<4); bitflush(bw,br,out);*/ memcpy(out,in,n*sizeof(in[0])); return 1+n*sizeof(in[0]); } size_t TEMPLATE2(bvzzenc,USIZE)(uint_t *in, size_t n, unsigned char *out, uint_t start) { uint_t *ip = in, pd = 0, *pp = in,dd; unsigned char *op = out, *out_ = out+n*sizeof(in[0]); bitdef(bw,br); #define FE(_pp_, _ip_, _d_, _op_,_usize_) do {\ uint64_t _r = _ip_ - _pp_;\ if(_r > NL) { _r -= NL; unsigned _b = (bsr64(_r)+7)>>3; bitput(bw,br,4+3+3,(_b-1)<<(4+3)); bitput64(bw,br,_b<<3, _r, _op_); bitenorm(bw,br,_op_); }\ else while(_r--) { bitput(bw,br,1,1); bitenorm(bw,br,_op_); }\ _d_ = TEMPLATE2(zigzagenc,_usize_)(_d_);\ if(!_d_) bitput(bw,br, 1, 1);\ else if(_d_ < (1<< (N2-1))) bitput(bw,br, N2+2,_d_<<2|2);\ else if(_d_ < (1<< (N3-1))) bitput(bw,br, N3+3,_d_<<3|4);\ else if(_d_ < (1<< (N4-1))) bitput(bw,br, N4+4,_d_<<4|8);\ else { unsigned _b = (TEMPLATE2(bsr,_usize_)(_d_)+7)>>3; bitput(bw,br,4+3,(_b-1)<<4); TEMPLATE2(bitput,_usize_)(bw,br, _b<<3, _d_,_op_); }\ bitenorm(bw,br,_op_);\ } while(0) if(n > 4) for(; ip < in+(n-1-4);) { start = ip[0] - start; dd = start-pd; pd = start; start = ip[0]; if(dd) goto a; ip++; start = ip[0] - start; dd = start-pd; pd = start; start = ip[0]; if(dd) goto a; ip++; start = ip[0] - start; dd = start-pd; pd = start; start = ip[0]; if(dd) goto a; ip++; start = ip[0] - start; dd = start-pd; pd = start; start = ip[0]; if(dd) goto a; ip++; PREFETCH(ip+256,0); continue; a:; FE(pp,ip, dd, op,USIZE); pp = ++ip; OVERFLOW; } for(;ip < in+n;) { start = ip[0] - start; dd = start-pd; pd = start; start = ip[0]; if(dd) goto b; ip++; continue; b:; FE(pp,ip, dd, op,USIZE); pp = ++ip; OVERFLOW; } if(ip > pp) { start = ip[0] - start; dd = start-pd; FE(pp, ip, dd, op, USIZE); OVERFLOW; } bitflush(bw,br,op); return op - out; } size_t TEMPLATE2(bvzzdec,USIZE)(unsigned char *in, size_t n, uint_t *out, uint_t start) { if(!n) return 0; uint_t *op = out, pd = 0; unsigned char *ip = in; bitdef(bw,br); for(bitdnorm(bw,br,ip); op < out+n; ) { PREFETCH(ip+384,0); #if USIZE == 64 uint_t dd = bitbw(bw,br); #else uint32_t dd = bitbw(bw,br); #endif if(dd & 1) bitrmv(bw,br, 0+1), dd = 0; else if(dd & 2) bitrmv(bw,br,N2+2), dd = BZHI32(dd>>2, N2); else if(dd & 4) bitrmv(bw,br,N3+3), dd = BZHI32(dd>>3, N3); else if(dd & 8) bitrmv(bw,br,N4+4), dd = BZHI32(dd>>4, N4); else { unsigned b; uint_t *_op; uint64_t r; BITGET32(bw,br, 4+3, b); if((b>>=4) <= 1) { if(b==1) { // No compression, because of overflow memcpy(out,in+1, n*sizeof(out[0])); return 1+n*sizeof(out[0]); } BITGET32(bw,br,3,b); bitget32(bw,br,(b+1)<<3,r,ip); bitdnorm(bw,br,ip);//RLE //r+=NL; while(r--) *op++=(start+=pd); #if (defined(__SSE2__) /*|| defined(__ARM_NEON)*/) && USIZE == 32 __m128i sv = _mm_set1_epi32(start), cv = _mm_set_epi32(4*pd,3*pd,2*pd,1*pd); for(r += NL, _op = op; op != _op+(r&~7);) { sv = _mm_add_epi32(sv,cv); _mm_storeu_si128(op, sv); sv = mm_shuffle_nnnn_epi32(sv, 3); op += 4; //_mm_shuffle_epi32(sv, _MM_SHUFFLE(3, 3, 3, 3))->mm_shuffle_nnnn_epi32(sv, 3) sv = _mm_add_epi32(sv,cv); _mm_storeu_si128(op, sv); sv = mm_shuffle_nnnn_epi32(sv, 3); op += 4; } start = (unsigned)_mm_cvtsi128_si32(_mm_srli_si128(sv,12)); #else for(r+=NL, _op = op; op != _op+(r&~7); op += 8) op[0]=(start+=pd), op[1]=(start+=pd), op[2]=(start+=pd), op[3]=(start+=pd), op[4]=(start+=pd), op[5]=(start+=pd), op[6]=(start+=pd), op[7]=(start+=pd); #endif for(; op != _op+r; op++) *op = (start+=pd); continue; } TEMPLATE2(bitget,USIZE)(bw,br,(b+1)<<3,dd,ip); } pd += TEMPLATE2(zigzagdec,USIZE)(dd); *op++ = (start += pd); bitdnorm(bw,br,ip); } bitalign(bw,br,ip); return ip - in; } //-------- Zigzag with bit/io + RLE -------------------------------------------------------------------------- size_t TEMPLATE2(bvzenc,USIZE)(uint_t *in, size_t n, unsigned char *out, uint_t start) { uint_t *ip = in, *pp = in,dd; unsigned char *op = out, *out_ = out+n*sizeof(in[0]); bitdef(bw,br); #define FE(_pp_, _ip_, _d_, _op_,_usize_) do {\ uint64_t _r = _ip_ - _pp_;\ if(_r > NL) { _r -= NL; unsigned _b = (bsr64(_r)+7)>>3; bitput(bw,br,4+3+3,(_b-1)<<(4+3)); bitput64(bw,br,_b<<3, _r, _op_); bitenorm(bw,br,_op_); }\ else while(_r--) { bitput(bw,br,1,1); bitenorm(bw,br,_op_); }\ _d_ = TEMPLATE2(zigzagenc,_usize_)(_d_);\ if(!_d_) bitput(bw,br, 1, 1);\ else if(_d_ < (1<< (N2-1))) bitput(bw,br, N2+2,_d_<<2|2);\ else if(_d_ < (1<< (N3-1))) bitput(bw,br, N3+3,_d_<<3|4);\ else if(_d_ < (1<< (N4-1))) bitput(bw,br, N4+4,_d_<<4|8);\ else { unsigned _b = (TEMPLATE2(bsr,_usize_)(_d_)+7)>>3; bitput(bw,br,4+3,(_b-1)<<4); TEMPLATE2(bitput,_usize_)(bw,br, _b<<3, _d_,_op_); }\ bitenorm(bw,br,_op_);\ } while(0) if(n > 4) for(; ip < in+(n-1-4);) { dd = ip[0] - start; start = ip[0]; if(dd) goto a; ip++; dd = ip[0] - start; start = ip[0]; if(dd) goto a; ip++; dd = ip[0] - start; start = ip[0]; if(dd) goto a; ip++; dd = ip[0] - start; start = ip[0]; if(dd) goto a; ip++; PREFETCH(ip+256,0); continue; a:; FE(pp,ip, dd, op,USIZE); pp = ++ip; OVERFLOW; } for(;ip < in+n;) { dd = ip[0] - start; start = ip[0]; if(dd) goto b; ip++; continue; b:; FE(pp,ip, dd, op,USIZE); pp = ++ip; OVERFLOW; } if(ip > pp) { dd = ip[0] - start; start = ip[0]; FE(pp, ip, dd, op, USIZE); OVERFLOW; } bitflush(bw,br,op); return op - out; } size_t TEMPLATE2(bvzdec,USIZE)(unsigned char *in, size_t n, uint_t *out, uint_t start) { if(!n) return 0; uint_t *op = out; unsigned char *ip = in; bitdef(bw,br); for(bitdnorm(bw,br,ip); op < out+n; ) { PREFETCH(ip+384,0); #if USIZE == 64 uint_t dd = bitbw(bw,br); #else uint32_t dd = bitbw(bw,br); #endif if(dd & 1) bitrmv(bw,br, 0+1), dd = 0; else if(dd & 2) bitrmv(bw,br,N2+2), dd = BZHI32(dd>>2, N2); else if(dd & 4) bitrmv(bw,br,N3+3), dd = BZHI32(dd>>3, N3); else if(dd & 8) bitrmv(bw,br,N4+4), dd = BZHI32(dd>>4, N4); else { unsigned b; uint_t *_op; uint64_t r; BITGET32(bw,br, 4+3, b); if((b>>=4) <= 1) { if(b==1) { // No compression, because of overflow memcpy(out,in+1, n*sizeof(out[0])); return 1+n*sizeof(out[0]); } BITGET32(bw,br,3,b); bitget32(bw,br,(b+1)<<3,r,ip); bitdnorm(bw,br,ip);//RLE //r+=NL; while(r--) *op++=(start+=pd); #if (defined(__SSE2__) || defined(__ARM_NEON)) && USIZE == 32 __m128i sv = _mm_set1_epi32(start); for(r += NL, _op = op; op != _op+(r&~7);) { _mm_storeu_si128(op, sv); op += 4; _mm_storeu_si128(op, sv); op += 4; } #else for(r+=NL, _op = op; op != _op+(r&~7); op += 8) op[0]=op[1]=op[2]=op[3]=op[4]=op[5]=op[6]=op[7]=start; #endif for(; op != _op+r; op++) *op = start; continue; } TEMPLATE2(bitget,USIZE)(bw,br,(b+1)<<3,dd,ip); } dd = TEMPLATE2(zigzagdec,USIZE)(dd); *op++ = (start += dd); bitdnorm(bw,br,ip); } bitalign(bw,br,ip); return ip - in; } #undef USIZE #endif