ddidderr/zstd-rs
1
0
Fork 0
Code Issues Pull Requests Packages Projects Releases Wiki Activity

improved 4reps in ZSTD_compressBlock_lazy_generic

Browse Source 
kSlotNew = 0
This commit is contained in:
inikep
2016-03-22 11:56:22 +01:00
parent f2fa0e1198
commit 2b942881ac
5 changed files with 26 additions and 19 deletions
Download Patch File Download Diff File Expand all files Collapse all files
lib/zstd_compress.c
+7 -7
View File
@@ -1617,8 +1617,8 @@ void ZSTD_compressBlock_lazy_generic(ZSTD_CCtx* ctx,
if (depth==0) goto _storeSequence; if (depth==0) goto _storeSequence;
} else { } else {
size_t mlRep = ZSTD_count(ip+MINMATCH, ip+MINMATCH-rep[i], iend) + MINMATCH; size_t mlRep = ZSTD_count(ip+MINMATCH, ip+MINMATCH-rep[i], iend) + MINMATCH;
int gain2 = (int)(mlRep * 3 /*- ZSTD_highbit((U32)i+1)*/); int gain2 = (int)(mlRep * 3 /*- ZSTD_highbit((U32)i+1)*/ + (i==1));
int gain1 = (int)(matchLength*3 - /*ZSTD_highbit((U32)offset+1)*/ + 1); int gain1 = (int)(matchLength*3 - /*ZSTD_highbit((U32)offset+1)*/ + 1 + (offset==1));
if (gain2 > gain1) if (gain2 > gain1)
matchLength = mlRep, offset = i; matchLength = mlRep, offset = i;
} }
@@ -1642,10 +1642,10 @@ void ZSTD_compressBlock_lazy_generic(ZSTD_CCtx* ctx,
while (ip<ilimit) { while (ip<ilimit) {
ip ++; ip ++;
for (int i=0; i<ZSTD_REP_NUM; i++) for (int i=0; i<ZSTD_REP_NUM; i++)
if ((offset >= ZSTD_REP_NUM) && (MEM_read32(ip) == MEM_read32(ip - rep[i]))) { if (MEM_read32(ip) == MEM_read32(ip - rep[i])) {
size_t mlRep = ZSTD_count(ip+MINMATCH, ip+MINMATCH-rep[i], iend) + MINMATCH; size_t mlRep = ZSTD_count(ip+MINMATCH, ip+MINMATCH-rep[i], iend) + MINMATCH;
int gain2 = (int)(mlRep * 3); int gain2 = (int)(mlRep * 3);
int gain1 = (int)(matchLength*3 - ZSTD_highbit((U32)offset+1) + 1); int gain1 = (int)(matchLength*3 - ZSTD_highbit((U32)offset+1) + 1 + (offset<ZSTD_REP_NUM));
if ((mlRep >= MINMATCH) && (gain2 > gain1)) if ((mlRep >= MINMATCH) && (gain2 > gain1))
matchLength = mlRep, offset = i, start = ip; matchLength = mlRep, offset = i, start = ip;
} }
@@ -1663,10 +1663,10 @@ void ZSTD_compressBlock_lazy_generic(ZSTD_CCtx* ctx,
if ((depth==2) && (ip<ilimit)) { if ((depth==2) && (ip<ilimit)) {
ip ++; ip ++;
for (int i=0; i<ZSTD_REP_NUM; i++) for (int i=0; i<ZSTD_REP_NUM; i++)
if ((offset >= ZSTD_REP_NUM) && (MEM_read32(ip) == MEM_read32(ip - rep[i]))) { if (MEM_read32(ip) == MEM_read32(ip - rep[i])) {
size_t ml2 = ZSTD_count(ip+MINMATCH, ip+MINMATCH-rep[i], iend) + MINMATCH; size_t ml2 = ZSTD_count(ip+MINMATCH, ip+MINMATCH-rep[i], iend) + MINMATCH;
int gain2 = (int)(ml2 * 4); int gain2 = (int)(ml2 * 4);
int gain1 = (int)(matchLength*4 - ZSTD_highbit((U32)offset+1) + 1); int gain1 = (int)(matchLength*4 - ZSTD_highbit((U32)offset+1) + 1 + (offset<ZSTD_REP_NUM));
if ((ml2 >= MINMATCH) && (gain2 > gain1)) if ((ml2 >= MINMATCH) && (gain2 > gain1))
matchLength = ml2, offset = i, start = ip; matchLength = ml2, offset = i, start = ip;
} }
@@ -1693,7 +1693,7 @@ _storeSequence:
{ {
#if ZSTD_REP_NUM == 4 #if ZSTD_REP_NUM == 4
if (offset >= ZSTD_REP_NUM) { if (offset >= ZSTD_REP_NUM) {
#if 0 #if 1
rep[3] = rep[2]; rep[3] = rep[2];
rep[2] = rep[1]; rep[2] = rep[1];
rep[1] = rep[0]; rep[1] = rep[0];
lib/zstd_decompress.c
+9 -2
View File
@@ -691,19 +691,26 @@ static void ZSTD_decodeSequence(seq_t* seq, seqState_t* seqState, const U32 mls)
if (offset == 3) seqState->prevOffset[3] = seqState->prevOffset[2]; if (offset == 3) seqState->prevOffset[3] = seqState->prevOffset[2];
seqState->prevOffset[2] = seqState->prevOffset[1]; seqState->prevOffset[2] = seqState->prevOffset[1];
} }
offset = temp;
seqState->prevOffset[1] = seqState->prevOffset[0]; seqState->prevOffset[1] = seqState->prevOffset[0];
seqState->prevOffset[0] = offset; seqState->prevOffset[0] = offset = temp;
} else { } else {
offset = seqState->prevOffset[0]; offset = seqState->prevOffset[0];
} }
} else { } else {
offset -= ZSTD_REP_MOVE; offset -= ZSTD_REP_MOVE;
#if 1 // faster without kSlotNew
seqState->prevOffset[3] = seqState->prevOffset[2];
seqState->prevOffset[2] = seqState->prevOffset[1];
seqState->prevOffset[1] = seqState->prevOffset[0];
seqState->prevOffset[0] = offset;
#else
if (kSlotNew < 3) seqState->prevOffset[3] = seqState->prevOffset[2]; if (kSlotNew < 3) seqState->prevOffset[3] = seqState->prevOffset[2];
if (kSlotNew < 2) seqState->prevOffset[2] = seqState->prevOffset[1]; if (kSlotNew < 2) seqState->prevOffset[2] = seqState->prevOffset[1];
if (kSlotNew < 1) seqState->prevOffset[1] = seqState->prevOffset[0]; if (kSlotNew < 1) seqState->prevOffset[1] = seqState->prevOffset[0];
seqState->prevOffset[kSlotNew] = offset; seqState->prevOffset[kSlotNew] = offset;
#endif
} }
#else // ZSTD_REP_NUM == 1 #else // ZSTD_REP_NUM == 1
#if 0 #if 0
lib/zstd_internal.h
+1 -1
View File
@@ -71,7 +71,7 @@
#else #else
#define ZSTD_REP_NUM 1 #define ZSTD_REP_NUM 1
#define ZSTD_REP_INIT 2 #define ZSTD_REP_INIT 2
#define ZSTD_REP_MOVE 0//(ZSTD_REP_NUM-1) #define ZSTD_REP_MOVE 0 //(ZSTD_REP_NUM-1)
#endif #endif
#define KB *(1 <<10) #define KB *(1 <<10)
programs/Makefile
+1 -1
View File
@@ -46,7 +46,7 @@ DESTDIR?=
PREFIX ?= /usr/local PREFIX ?= /usr/local
CPPFLAGS= -I../lib -DZSTD_VERSION=\"$(VERSION)\" CPPFLAGS= -I../lib -DZSTD_VERSION=\"$(VERSION)\"
CFLAGS ?= -O3 # -falign-loops=32 # not always beneficial CFLAGS ?= -O3 # -falign-loops=32 # not always beneficial
CFLAGS += -std=c99 -Wall -Wextra -Wcast-qual -Wcast-align -Wshadow -Wstrict-aliasing=1 -Wswitch-enum -Wstrict-prototypes -Wundef CFLAGS += -std=gnu99 -Wall -Wextra -Wcast-qual -Wcast-align -Wshadow -Wstrict-aliasing=1 -Wswitch-enum -Wstrict-prototypes -Wundef
FLAGS = $(CPPFLAGS) $(CFLAGS) $(LDFLAGS) $(MOREFLAGS) FLAGS = $(CPPFLAGS) $(CFLAGS) $(LDFLAGS) $(MOREFLAGS)
BINDIR = $(PREFIX)/bin BINDIR = $(PREFIX)/bin
programs/bench.c
+6 -6
View File
@@ -52,12 +52,12 @@
#include <stdio.h> /* fprintf, fopen, ftello64 */ #include <stdio.h> /* fprintf, fopen, ftello64 */
#include <sys/types.h> /* stat64 */ #include <sys/types.h> /* stat64 */
#include <sys/stat.h> /* stat64 */ #include <sys/stat.h> /* stat64 */
#include <time.h> /* clock_t, clock, CLOCKS_PER_SEC */ #include <time.h> /* clock_t, clock, nanosleep, CLOCKS_PER_SEC */
#ifdef WINDOWS #ifdef WINDOWS
#define mili_sleep(mili) Sleep(mili) #define mili_sleep(mili) Sleep(mili)
#else #else
#define mili_sleep(mili) usleep(mili*1000)
#include <sys/resource.h> /* setpriority */ #include <sys/resource.h> /* setpriority */
#define mili_sleep(mili) { struct timespec t; t.tv_sec=0; t.tv_nsec=mili*1000000L; nanosleep(&t, NULL); }
#endif #endif
/* sleep : posix - windows - others */ /* sleep : posix - windows - others */
@@ -379,11 +379,11 @@ _findError:
if (crcOrig == crcCheck) if (crcOrig == crcCheck)
{ {
DISPLAYLEVEL(2, "%2i-%-17.17s :%10i ->%10i (%5.3f),%6.1f MB/s ,%6.1f MB/s \n", cLevel, displayName, (int)srcSize, (int)cSize, ratio, (double)srcSize / fastestC / 1000., (double)srcSize / fastestD / 1000.);
result->ratio = ratio; result->ratio = ratio;
result->cSize = cSize; result->cSize = cSize;
result->cSpeed = (double)srcSize / fastestC / 1000.; result->cSpeed = (double)srcSize / 1000000. / (fastestC / CLOCKS_PER_SEC);
result->dSpeed = (double)srcSize / fastestD / 1000.; result->dSpeed = (double)srcSize / 1000000. / (fastestD / CLOCKS_PER_SEC);
DISPLAYLEVEL(2, "%2i-%-17.17s :%10i ->%10i (%5.3f),%6.1f MB/s ,%6.1f MB/s \n", cLevel, displayName, (int)srcSize, (int)cSize, ratio, result->cSpeed, result->dSpeed);
} }
else else
DISPLAYLEVEL(2, "%2i-\n", cLevel); DISPLAYLEVEL(2, "%2i-\n", cLevel);
@@ -449,7 +449,7 @@ static void BMK_benchCLevel(void* srcBuffer, size_t benchedSize,
dictBuffer, dictBufferSize, &result); dictBuffer, dictBufferSize, &result);
if (g_displayLevel == 1) { if (g_displayLevel == 1) {
if (additionalParam) if (additionalParam)
DISPLAY("%-3i%11i (%5.3f) %6.1f MB/s %6.1f MB/s %s (p=%d)\n", -l, (int)result.cSize, result.ratio, result.cSpeed, result.dSpeed, displayName, additionalParam); DISPLAY("%-3i%11i (%5.3f) %6.1f MB/s %6.1f MB/s %s (param=%d)\n", -l, (int)result.cSize, result.ratio, result.cSpeed, result.dSpeed, displayName, additionalParam);
else else
DISPLAY("%-3i%11i (%5.3f) %6.1f MB/s %6.1f MB/s %s\n", -l, (int)result.cSize, result.ratio, result.cSpeed, result.dSpeed, displayName); DISPLAY("%-3i%11i (%5.3f) %6.1f MB/s %6.1f MB/s %s\n", -l, (int)result.cSize, result.ratio, result.cSpeed, result.dSpeed, displayName);
total.cSize += result.cSize; total.cSize += result.cSize;