improved 4reps in ZSTD_compressBlock_lazy_generic

kSlotNew = 0

diff --git a/lib/zstd_compress.c b/lib/zstd_compress.c
index 397e761892225815386e2b7c17a93357dc16a0b7..e9b9d4a6e6e75d637bd7f44560775b9e5b15eadd 100644 (file)
--- a/lib/zstd_compress.c
+++ b/lib/zstd_compress.c
@@ -1617,8 +1617,8 @@ void ZSTD_compressBlock_lazy_generic(ZSTD_CCtx* ctx,
                 if (depth==0) goto _storeSequence;
             } else {
                 size_t mlRep = ZSTD_count(ip+MINMATCH, ip+MINMATCH-rep[i], iend) + MINMATCH;
-                int gain2 = (int)(mlRep * 3 /*- ZSTD_highbit((U32)i+1)*/);
-                int gain1 = (int)(matchLength*3 - /*ZSTD_highbit((U32)offset+1)*/ + 1);
+                int gain2 = (int)(mlRep * 3 /*- ZSTD_highbit((U32)i+1)*/ + (i==1));
+                int gain1 = (int)(matchLength*3 - /*ZSTD_highbit((U32)offset+1)*/ + 1 + (offset==1));
                 if (gain2 > gain1)
                     matchLength = mlRep, offset = i;
             }
@@ -1642,10 +1642,10 @@ void ZSTD_compressBlock_lazy_generic(ZSTD_CCtx* ctx,
         while (ip<ilimit) {
             ip ++;
             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;
                 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))
                     matchLength = mlRep, offset = i, start = ip;
             }
@@ -1663,10 +1663,10 @@ void ZSTD_compressBlock_lazy_generic(ZSTD_CCtx* ctx,
             if ((depth==2) && (ip<ilimit)) {
                 ip ++;
                 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;
                     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))
                         matchLength = ml2, offset = i, start = ip;
                 }
@@ -1693,7 +1693,7 @@ _storeSequence:
         {
 #if ZSTD_REP_NUM == 4
             if (offset >= ZSTD_REP_NUM) {
-#if 0
+#if 1
                 rep[3] = rep[2];
                 rep[2] = rep[1];
                 rep[1] = rep[0];

diff --git a/lib/zstd_decompress.c b/lib/zstd_decompress.c
index f0475bf25a7aafeffbcb69ab7460b688ef8d155d..651087c4ded40341a408f7fd4064c2d4fe644a32 100644 (file)
--- a/lib/zstd_decompress.c
+++ b/lib/zstd_decompress.c
@@ -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];
                     seqState->prevOffset[2] = seqState->prevOffset[1];
                 }
-                offset = temp;
                 seqState->prevOffset[1] = seqState->prevOffset[0];
-                seqState->prevOffset[0] = offset;
+                seqState->prevOffset[0] = offset = temp;
+
             } else {
                 offset = seqState->prevOffset[0];
             }
 
         } else {
             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 < 2) seqState->prevOffset[2] = seqState->prevOffset[1];
             if (kSlotNew < 1) seqState->prevOffset[1] = seqState->prevOffset[0];               
             seqState->prevOffset[kSlotNew] = offset;
+#endif
         }
 #else // ZSTD_REP_NUM == 1
     #if 0

diff --git a/lib/zstd_internal.h b/lib/zstd_internal.h
index 0abdcdfa92d99f89185b52ad7048a062336fda52..5809f875493ced66f97de1e73a6b392d5477d83b 100644 (file)
--- a/lib/zstd_internal.h
+++ b/lib/zstd_internal.h
@@ -71,7 +71,7 @@
 #else
     #define ZSTD_REP_NUM    1
     #define ZSTD_REP_INIT   2
-    #define ZSTD_REP_MOVE  0//(ZSTD_REP_NUM-1)
+    #define ZSTD_REP_MOVE   0 //(ZSTD_REP_NUM-1)
 #endif
 
 #define KB *(1 <<10)

diff --git a/programs/Makefile b/programs/Makefile
index ca2ba99fc7ede9e20c6d38cfd145fa4eb5344f33..cd432f79049322edf1fad163d1157b334d695a74 100644 (file)
--- a/programs/Makefile
+++ b/programs/Makefile
@@ -46,7 +46,7 @@ DESTDIR?=
 PREFIX ?= /usr/local
 CPPFLAGS= -I../lib -DZSTD_VERSION=\"$(VERSION)\"
 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)
 
 BINDIR  = $(PREFIX)/bin

diff --git a/programs/bench.c b/programs/bench.c
index 5808c2ecc84c8de791857b9d0d441e8218279442..20f2ee0ab0397009e5a3530774949e3798d48fbd 100644 (file)
--- a/programs/bench.c
+++ b/programs/bench.c
@@ -52,14 +52,14 @@
 #include <stdio.h>       /* fprintf, fopen, ftello64 */
 #include <sys/types.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
-       #define mili_sleep(mili) Sleep(mili)
+    #define mili_sleep(mili) Sleep(mili)
 #else
-       #define mili_sleep(mili) usleep(mili*1000)
     #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
-    
+
 /* sleep : posix - windows - others */
 #if !defined(_WIN32) && (defined(__unix__) || defined(__unix) || (defined(__APPLE__) && defined(__MACH__)))
 #  include <unistd.h>
@@ -379,11 +379,11 @@ _findError:
 
         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->cSize = cSize;
-            result->cSpeed = (double)srcSize / fastestC / 1000.; 
-            result->dSpeed = (double)srcSize / fastestD / 1000.;
+            result->cSpeed = (double)srcSize / 1000000. / (fastestC / CLOCKS_PER_SEC); 
+            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
             DISPLAYLEVEL(2, "%2i-\n", cLevel);
@@ -449,7 +449,7 @@ static void BMK_benchCLevel(void* srcBuffer, size_t benchedSize,
                      dictBuffer, dictBufferSize, &result);
         if (g_displayLevel == 1) {
             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
                 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;