first implementation of delayed update for btlazy2

This is a pretty nice speed win.

The new strategy consists in stacking new candidates as if it was a hash chain.
Then, only if there is a need to actually consult the chain, they are batch-updated,
before starting the match search itself.
This is supposed to be beneficial when skipping positions,
which happens a lot when using lazy strategy.

The baseline performance for btlazy2 on my laptop is :
15#calgary.tar       :   3265536 ->    955985 (3.416),  7.06 MB/s , 618.0 MB/s
15#enwik7            :  10000000 ->   3067341 (3.260),  4.65 MB/s , 521.2 MB/s
15#silesia.tar       : 211984896 ->  58095131 (3.649),  6.20 MB/s , 682.4 MB/s
(only level 15 remains for btlazy2, as this strategy is squeezed between lazy2 and btopt)

After this patch, and keeping all parameters identical,
speed is increased by a pretty good margin (+30-50%),
but compression ratio suffers a bit :
15#calgary.tar       :   3265536 ->    958060 (3.408),  9.12 MB/s , 621.1 MB/s
15#enwik7            :  10000000 ->   3078318 (3.249),  6.37 MB/s , 525.1 MB/s
15#silesia.tar       : 211984896 ->  58444111 (3.627),  9.89 MB/s , 680.4 MB/s

That's because I kept `1<<searchLog` as a maximum number of candidates to update.
But for a hash chain, this represents the total number of candidates in the chain,
while for the binary, it represents the maximum depth of searches.
Keep in mind that a lot of candidates won't even be visited in the btree,
since they are filtered out by the binary sort.

As a consequence, in the new implementation,
the effective depth of the binary tree is substantially shorter.

To compensate, it's enough to increase `searchLog` value.
Here is the result after adding just +1 to searchLog (level 15 setting in this patch):
15#calgary.tar       :   3265536 ->    956311 (3.415),  8.32 MB/s , 611.4 MB/s
15#enwik7            :  10000000 ->   3067655 (3.260),  5.43 MB/s , 535.5 MB/s
15#silesia.tar       : 211984896 ->  58113144 (3.648),  8.35 MB/s , 679.3 MB/s

aka, almost the same compression ratio as before,
but with a noticeable speed increase (+20-30%).

This modification makes btlazy2 more competitive.
A new round of paramgrill will be necessary to determine which levels are impacted and could adopt the new strategy.
This commit is contained in:
Yann Collet
2017-12-28 16:58:57 +01:00
parent d55aea3c3b
commit 5235d8d6ba
5 changed files with 322 additions and 150 deletions
+141
View File
@@ -265,6 +265,147 @@ static U32 ZSTD_insertAndFindFirstIndexHash3 (ZSTD_CCtx* const cctx, const BYTE*
/*-*************************************
* Binary Tree search
***************************************/
/** ZSTD_insertBt1() : add one or multiple positions to tree.
* ip : assumed <= iend-8 .
* @return : nb of positions added */
static U32 ZSTD_insertBt1(ZSTD_CCtx* zc,
const BYTE* const ip, const BYTE* const iend,
U32 nbCompares, U32 const mls, U32 const extDict)
{
U32* const hashTable = zc->hashTable;
U32 const hashLog = zc->appliedParams.cParams.hashLog;
size_t const h = ZSTD_hashPtr(ip, hashLog, mls);
U32* const bt = zc->chainTable;
U32 const btLog = zc->appliedParams.cParams.chainLog - 1;
U32 const btMask = (1 << btLog) - 1;
U32 matchIndex = hashTable[h];
size_t commonLengthSmaller=0, commonLengthLarger=0;
const BYTE* const base = zc->base;
const BYTE* const dictBase = zc->dictBase;
const U32 dictLimit = zc->dictLimit;
const BYTE* const dictEnd = dictBase + dictLimit;
const BYTE* const prefixStart = base + dictLimit;
const BYTE* match;
const U32 current = (U32)(ip-base);
const U32 btLow = btMask >= current ? 0 : current - btMask;
U32* smallerPtr = bt + 2*(current&btMask);
U32* largerPtr = smallerPtr + 1;
U32 dummy32; /* to be nullified at the end */
U32 const windowLow = zc->lowLimit;
U32 matchEndIdx = current+8+1;
size_t bestLength = 8;
#ifdef ZSTD_C_PREDICT
U32 predictedSmall = *(bt + 2*((current-1)&btMask) + 0);
U32 predictedLarge = *(bt + 2*((current-1)&btMask) + 1);
predictedSmall += (predictedSmall>0);
predictedLarge += (predictedLarge>0);
#endif /* ZSTD_C_PREDICT */
DEBUGLOG(8, "ZSTD_insertBt1 (%u)", current);
assert(ip <= iend-8); /* required for h calculation */
hashTable[h] = current; /* Update Hash Table */
while (nbCompares-- && (matchIndex > windowLow)) {
U32* const nextPtr = bt + 2*(matchIndex & btMask);
size_t matchLength = MIN(commonLengthSmaller, commonLengthLarger); /* guaranteed minimum nb of common bytes */
assert(matchIndex < current);
#ifdef ZSTD_C_PREDICT /* note : can create issues when hlog small <= 11 */
const U32* predictPtr = bt + 2*((matchIndex-1) & btMask); /* written this way, as bt is a roll buffer */
if (matchIndex == predictedSmall) {
/* no need to check length, result known */
*smallerPtr = matchIndex;
if (matchIndex <= btLow) { smallerPtr=&dummy32; break; } /* beyond tree size, stop the search */
smallerPtr = nextPtr+1; /* new "smaller" => larger of match */
matchIndex = nextPtr[1]; /* new matchIndex larger than previous (closer to current) */
predictedSmall = predictPtr[1] + (predictPtr[1]>0);
continue;
}
if (matchIndex == predictedLarge) {
*largerPtr = matchIndex;
if (matchIndex <= btLow) { largerPtr=&dummy32; break; } /* beyond tree size, stop the search */
largerPtr = nextPtr;
matchIndex = nextPtr[0];
predictedLarge = predictPtr[0] + (predictPtr[0]>0);
continue;
}
#endif
if ((!extDict) || (matchIndex+matchLength >= dictLimit)) {
assert(matchIndex+matchLength >= dictLimit); /* might be wrong if extDict is incorrectly set to 0 */
match = base + matchIndex;
matchLength += ZSTD_count(ip+matchLength, match+matchLength, iend);
} else {
match = dictBase + matchIndex;
matchLength += ZSTD_count_2segments(ip+matchLength, match+matchLength, iend, dictEnd, prefixStart);
if (matchIndex+matchLength >= dictLimit)
match = base + matchIndex; /* to prepare for next usage of match[matchLength] */
}
if (matchLength > bestLength) {
bestLength = matchLength;
if (matchLength > matchEndIdx - matchIndex)
matchEndIdx = matchIndex + (U32)matchLength;
}
if (ip+matchLength == iend) { /* equal : no way to know if inf or sup */
break; /* drop , to guarantee consistency ; miss a bit of compression, but other solutions can corrupt tree */
}
if (match[matchLength] < ip[matchLength]) { /* necessarily within buffer */
/* match is smaller than current */
*smallerPtr = matchIndex; /* update smaller idx */
commonLengthSmaller = matchLength; /* all smaller will now have at least this guaranteed common length */
if (matchIndex <= btLow) { smallerPtr=&dummy32; break; } /* beyond tree size, stop searching */
smallerPtr = nextPtr+1; /* new "candidate" => larger than match, which was smaller than target */
matchIndex = nextPtr[1]; /* new matchIndex, larger than previous and closer to current */
} else {
/* match is larger than current */
*largerPtr = matchIndex;
commonLengthLarger = matchLength;
if (matchIndex <= btLow) { largerPtr=&dummy32; break; } /* beyond tree size, stop searching */
largerPtr = nextPtr;
matchIndex = nextPtr[0];
} }
*smallerPtr = *largerPtr = 0;
if (bestLength > 384) return MIN(192, (U32)(bestLength - 384)); /* speed optimization */
assert(matchEndIdx > current + 8);
return matchEndIdx - (current + 8);
}
FORCE_INLINE_TEMPLATE
void ZSTD_updateTree_internal(ZSTD_CCtx* zc,
const BYTE* const ip, const BYTE* const iend,
const U32 nbCompares, const U32 mls, const U32 extDict)
{
const BYTE* const base = zc->base;
U32 const target = (U32)(ip - base);
U32 idx = zc->nextToUpdate;
DEBUGLOG(7, "ZSTD_updateTree_internal, from %u to %u (extDict:%u)",
idx, target, extDict);
while(idx < target)
idx += ZSTD_insertBt1(zc, base+idx, iend, nbCompares, mls, extDict);
zc->nextToUpdate = target;
}
void ZSTD_updateTree(ZSTD_CCtx* zc,
const BYTE* const ip, const BYTE* const iend,
const U32 nbCompares, const U32 mls)
{
ZSTD_updateTree_internal(zc, ip, iend, nbCompares, mls, 0 /*extDict*/);
}
void ZSTD_updateTree_extDict(ZSTD_CCtx* zc,
const BYTE* const ip, const BYTE* const iend,
const U32 nbCompares, const U32 mls)
{
ZSTD_updateTree_internal(zc, ip, iend, nbCompares, mls, 1 /*extDict*/);
}
FORCE_INLINE_TEMPLATE
U32 ZSTD_insertBtAndGetAllMatches (
ZSTD_CCtx* zc,