AuroraMiddleware/zstd
1
0
Fork 0
Code Issues Pull Requests Releases Wiki Activity

Fixed speed regression

Browse Source
This commit is contained in:
Yann Collet 2015-11-01 12:40:22 +01:00
parent be2010ea1b
commit 9b11b46f8a
2 changed files with 131 additions and 114 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

216
lib/zstdhc.c
View File

@ -235,7 +235,7 @@ size_t ZSTD_HC_insertAndFindBestMatch (
size_t ml=0;
/* HC4 match finder */
matchIndex = ZSTD_HC_insertAndFindFirstIndex(zc, ip, matchLengthSearch);
matchIndex = ZSTD_HC_insertAndFindFirstIndex (zc, ip, matchLengthSearch);
while ((matchIndex>=lowLimit) && (nbAttempts))
{
@ -273,7 +273,7 @@ size_t ZSTD_HC_insertAndFindBestMatch (
}
static size_t ZSTD_HC_insertAndFindBestMatch_selectMLS (
FORCE_INLINE size_t ZSTD_HC_insertAndFindBestMatch_selectMLS (
ZSTD_HC_CCtx* zc, /* Index table will be updated */
const BYTE* ip, const BYTE* const iLimit,
const BYTE** matchpos,
@ -289,6 +289,107 @@ static size_t ZSTD_HC_insertAndFindBestMatch_selectMLS (
}
size_t ZSTD_HC_compressBlock_lazy(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
{
seqStore_t* seqStorePtr = &(ctx->seqStore);
const BYTE* const istart = (const BYTE*)src;
const BYTE* ip = istart;
const BYTE* anchor = istart;
const BYTE* const iend = istart + srcSize;
const BYTE* const ilimit = iend - 8;
const BYTE* match = istart;
size_t offset_2=REPCODE_STARTVALUE, offset_1=REPCODE_STARTVALUE;
const U32 maxSearches = 1 << ctx->params.searchLog;
const U32 mls = ctx->params.searchLength;
/* init */
ZSTD_resetSeqStore(seqStorePtr);
if (((ip-ctx->base) - ctx->dictLimit) < REPCODE_STARTVALUE) ip += REPCODE_STARTVALUE;
/* Match Loop */
while (ip <= ilimit)
{
size_t matchLength;
size_t offset;
const BYTE* start;
/* try to find a first match */
if (MEM_read32(ip) == MEM_read32(ip - offset_2))
{
/* repcode : we take it*/
size_t offtmp = offset_2;
size_t litLength = ip - anchor;
matchLength = ZSTD_count(ip+MINMATCH, ip+MINMATCH-offset_2, iend);
offset_2 = offset_1;
offset_1 = offtmp;
ZSTD_storeSeq(seqStorePtr, litLength, anchor, 0, matchLength);
ip += matchLength+MINMATCH;
anchor = ip;
continue;
}
offset_2 = offset_1;
matchLength = ZSTD_HC_insertAndFindBestMatch_selectMLS(ctx, ip, iend, &match, maxSearches, mls);
if (!matchLength) { ip++; continue; }
/* let's try to find a better solution */
offset = ip - match;
start = ip;
while (ip<ilimit)
{
ip ++;
if (MEM_read32(ip) == MEM_read32(ip - offset_1))
{
size_t ml2 = ZSTD_count(ip+MINMATCH, ip+MINMATCH-offset_1, iend) + MINMATCH;
int gain2 = (int)(ml2 * 4);
int gain1 = (int)(matchLength*4 - ZSTD_highbit((U32)offset));
if (gain2 > gain1)
{
matchLength = ml2, offset = 0, start = ip;
break;
}
}
{
size_t ml2 = ZSTD_HC_insertAndFindBestMatch_selectMLS(ctx, ip, iend, &match, maxSearches, mls);
size_t offset2 = ip - match;
int gain2 = (int)(ml2*5 - ZSTD_highbit((U32)offset2)); /* raw approx */
int gain1 = (int)(matchLength*5 - ZSTD_highbit((U32)offset));
if (gain2 > gain1)
{
matchLength = ml2, offset = offset2, start = ip;
continue; /* search a better one */
}
}
break; /* nothing found : store previous one */
}
/* store sequence */
{
size_t litLength = start - anchor;
if (offset) offset_1 = offset;
ZSTD_storeSeq(seqStorePtr, litLength, anchor, offset, matchLength-MINMATCH);
ip = start + matchLength;
anchor = ip;
}
}
/* Last Literals */
{
size_t lastLLSize = iend - anchor;
memcpy(seqStorePtr->lit, anchor, lastLLSize);
seqStorePtr->lit += lastLLSize;
}
/* Final compression stage */
return ZSTD_compressSequences((BYTE*)dst, maxDstSize,
seqStorePtr, srcSize);
}
size_t ZSTD_HC_compressBlock_greedy(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
{
seqStore_t* seqStorePtr = &(ctx->seqStore);
@ -362,112 +463,14 @@ size_t ZSTD_HC_compressBlock_greedy(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstS
/* Final compression stage */
return ZSTD_compressSequences((BYTE*)dst, maxDstSize,
seqStorePtr, srcSize);
}
size_t ZSTD_HC_compressBlock_lazy(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
{
seqStore_t* seqStorePtr = &(ctx->seqStore);
const BYTE* const istart = (const BYTE*)src;
const BYTE* ip = istart;
const BYTE* anchor = istart;
const BYTE* const iend = istart + srcSize;
const BYTE* const ilimit = iend - 8;
const BYTE* match = istart;
size_t offset_2=REPCODE_STARTVALUE, offset_1=REPCODE_STARTVALUE;
const U32 maxSearches = 1 << ctx->params.searchLog;
const U32 mls = ctx->params.searchLength;
/* init */
ZSTD_resetSeqStore(seqStorePtr);
if (((ip-ctx->base) - ctx->dictLimit) < REPCODE_STARTVALUE) ip += REPCODE_STARTVALUE;
/* Match Loop */
while (ip <= ilimit)
{
size_t matchLength;
size_t offset;
const BYTE* start;
/* try to find a first match */
if (MEM_read32(ip) == MEM_read32(ip - offset_2))
{
/* repcode : we take it*/
size_t offtmp = offset_2;
size_t litLength = ip - anchor;
matchLength = ZSTD_count(ip+MINMATCH, ip+MINMATCH-offset_2, iend);
offset_2 = offset_1;
offset_1 = offtmp;
ZSTD_storeSeq(seqStorePtr, litLength, anchor, 0, matchLength);
ip += matchLength+MINMATCH;
anchor = ip;
continue;
}
offset_2 = offset_1;
matchLength = ZSTD_HC_insertAndFindBestMatch_selectMLS(ctx, ip, iend, &match, maxSearches, mls);
if (!matchLength) { ip++; continue; }
/* let's try to find a better solution */
offset = ip - match;
start = ip;
while (ip<ilimit)
{
ip ++;
if (MEM_read32(ip) == MEM_read32(ip - offset_1))
{
size_t ml2 = ZSTD_count(ip+MINMATCH, ip+MINMATCH-offset_1, iend) + MINMATCH;
int gain1 = matchLength - (ZSTD_highbit((U32)offset) / 3);
if ((int)ml2 > gain1)
{
matchLength = ml2, offset = 0, start = ip;
break;
}
}
{
size_t ml2 = ZSTD_HC_insertAndFindBestMatch_selectMLS(ctx, ip, iend, &match, maxSearches, mls);
size_t offset2 = ip - match;
int gain2 = ml2 - (ZSTD_highbit((U32)offset2) / 4); /* raw approx */
int gain1 = matchLength - (ZSTD_highbit((U32)offset) / 4);
if (gain2 > gain1)
{
matchLength = ml2, offset = offset2, start = ip;
continue; /* search a better one */
}
}
break; /* nothing found : store previous one */
}
/* store sequence */
{
size_t litLength = start - anchor;
if (offset) offset_1 = offset;
ZSTD_storeSeq(seqStorePtr, litLength, anchor, offset, matchLength-MINMATCH);
ip = start + matchLength;
anchor = ip;
}
}
/* Last Literals */
{
size_t lastLLSize = iend - anchor;
memcpy(seqStorePtr->lit, anchor, lastLLSize);
seqStorePtr->lit += lastLLSize;
}
/* Final compression stage */
return ZSTD_compressSequences((BYTE*)dst, maxDstSize,
seqStorePtr, srcSize);
}
size_t ZSTD_HC_compressBlock(ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize)
{
if (ctx->params.strategy==ZSTD_HC_greedy)
if (ctx->params.strategy == ZSTD_HC_greedy)
return ZSTD_HC_compressBlock_greedy(ctx, dst, maxDstSize, src, srcSize);
return ZSTD_HC_compressBlock_lazy(ctx, dst, maxDstSize, src, srcSize);
}
@ -483,10 +486,17 @@ static size_t ZSTD_HC_compress_generic (ZSTD_HC_CCtx* ctxPtr,
BYTE* const ostart = (BYTE*)dst;
BYTE* op = ostart;
BYTE* const oend = op + maxDstSize;
size_t (*blockCompressor) (ZSTD_HC_CCtx* ctx, void* dst, size_t maxDstSize, const void* src, size_t srcSize);
if (ctxPtr->params.strategy == ZSTD_HC_greedy)
blockCompressor = ZSTD_HC_compressBlock_greedy;
else
blockCompressor = ZSTD_HC_compressBlock_lazy;
while (remaining > blockSize)
{
size_t cSize = ZSTD_HC_compressBlock(ctxPtr, op+3, oend-op, ip, blockSize);
size_t cSize = blockCompressor(ctxPtr, op+3, oend-op, ip, blockSize);
if (cSize == 0)
{
@ -509,7 +519,7 @@ static size_t ZSTD_HC_compress_generic (ZSTD_HC_CCtx* ctxPtr,
/* last block */
{
size_t cSize = ZSTD_HC_compressBlock(ctxPtr, op+3, oend-op, ip, remaining);
size_t cSize = blockCompressor(ctxPtr, op+3, oend-op, ip, remaining);
if (cSize == 0)
{

21
programs/paramgrill.c
View File

@ -122,7 +122,8 @@ static U32 g_rand = 1;
static U32 g_singleRun = 0;
static U32 g_target = 0;
static U32 g_noSeed = 0;
static ZSTD_HC_parameters g_params = { 0, 0, 0, 0, 0, 0 };
static const ZSTD_HC_parameters* g_seedParams = ZSTD_HC_defaultParameters;
static ZSTD_HC_parameters g_params = { 0, 0, 0, 0, 0, ZSTD_HC_greedy };
void BMK_SetNbIterations(int nbLoops)
{
@ -623,9 +624,9 @@ static void playAround(FILE* f, winnerInfo_t* winners,
case 9:
p.searchLength--; break;
case 10:
p.strategy++; break;
p.strategy = ZSTD_HC_lazy; break;
case 11:
p.strategy--; break;
p.strategy = ZSTD_HC_greedy; break;
}
}
@ -674,7 +675,7 @@ static void BMK_selectRandomStart(
p.searchLog = FUZ_rand(&g_rand) % (ZSTD_HC_SEARCHLOG_MAX+1 - ZSTD_HC_SEARCHLOG_MIN) + ZSTD_HC_SEARCHLOG_MIN;
p.windowLog = FUZ_rand(&g_rand) % (ZSTD_HC_WINDOWLOG_MAX+1 - ZSTD_HC_WINDOWLOG_MIN) + ZSTD_HC_WINDOWLOG_MIN;
p.searchLength=FUZ_rand(&g_rand) % (ZSTD_HC_SEARCHLENGTH_MAX+1 - ZSTD_HC_SEARCHLENGTH_MIN) + ZSTD_HC_SEARCHLENGTH_MIN;
p.strategy = FUZ_rand(&g_rand) & 1;
p.strategy = (ZSTD_HC_strategy) (FUZ_rand(&g_rand) & 1);
playAround(f, winners, p, srcBuffer, srcSize, ctx);
}
else
@ -682,8 +683,6 @@ static void BMK_selectRandomStart(
}
static const ZSTD_HC_parameters* g_seedParams = ZSTD_HC_defaultParameters;
static void BMK_benchMem(void* srcBuffer, size_t srcSize)
{
ZSTD_HC_CCtx* ctx = ZSTD_HC_createCCtx();
@ -968,12 +967,20 @@ int main(int argc, char** argv)
while ((*argument>= '0') && (*argument<='9'))
g_params.searchLog *= 10, g_params.searchLog += *argument++ - '0';
continue;
case 'l':
case 'l': /* search length */
g_params.searchLength = 0;
argument++;
while ((*argument>= '0') && (*argument<='9'))
g_params.searchLength *= 10, g_params.searchLength += *argument++ - '0';
continue;
case 't': /* strategy */
g_params.strategy = ZSTD_HC_greedy;
argument++;
while ((*argument>= '0') && (*argument<='9'))
{
if (*argument++) g_params.strategy = ZSTD_HC_lazy;
}
continue;
case 'L':
{
int cLevel = 0;