Skip to content

Commit 7915c11

Browse files
authored
Merge pull request #3114 from embg/fast_extdict_pipeline2
Software pipeline for ZSTD_compressBlock_fast_extDict
2 parents ea763f3 + 3be9a81 commit 7915c11

File tree

2 files changed

+302
-158
lines changed

2 files changed

+302
-158
lines changed

lib/compress/zstd_fast.c

Lines changed: 207 additions & 63 deletions
Original file line numberDiff line numberDiff line change
@@ -587,11 +587,10 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
587587
U32* const hashTable = ms->hashTable;
588588
U32 const hlog = cParams->hashLog;
589589
/* support stepSize of 0 */
590-
U32 const stepSize = cParams->targetLength + !(cParams->targetLength);
590+
size_t const stepSize = cParams->targetLength + !(cParams->targetLength) + 1;
591591
const BYTE* const base = ms->window.base;
592592
const BYTE* const dictBase = ms->window.dictBase;
593593
const BYTE* const istart = (const BYTE*)src;
594-
const BYTE* ip = istart;
595594
const BYTE* anchor = istart;
596595
const U32 endIndex = (U32)((size_t)(istart - base) + srcSize);
597596
const U32 lowLimit = ZSTD_getLowestMatchIndex(ms, endIndex, cParams->windowLog);
@@ -605,6 +604,27 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
605604
const BYTE* const ilimit = iend - 8;
606605
U32 offset_1=rep[0], offset_2=rep[1];
607606

607+
const BYTE* ip0 = istart;
608+
const BYTE* ip1;
609+
const BYTE* ip2;
610+
const BYTE* ip3;
611+
U32 current0;
612+
613+
614+
size_t hash0; /* hash for ip0 */
615+
size_t hash1; /* hash for ip1 */
616+
U32 idx; /* match idx for ip0 */
617+
const BYTE* idxBase; /* base pointer for idx */
618+
619+
U32 offcode;
620+
const BYTE* match0;
621+
size_t mLength;
622+
const BYTE* matchEnd = 0; /* initialize to avoid warning, assert != 0 later */
623+
624+
size_t step;
625+
const BYTE* nextStep;
626+
const size_t kStepIncr = (1 << (kSearchStrength - 1));
627+
608628
(void)hasStep; /* not currently specialized on whether it's accelerated */
609629

610630
DEBUGLOG(5, "ZSTD_compressBlock_fast_extDict_generic (offset_1=%u)", offset_1);
@@ -613,75 +633,198 @@ static size_t ZSTD_compressBlock_fast_extDict_generic(
613633
if (prefixStartIndex == dictStartIndex)
614634
return ZSTD_compressBlock_fast(ms, seqStore, rep, src, srcSize);
615635

616-
/* Search Loop */
617-
while (ip < ilimit) { /* < instead of <=, because (ip+1) */
618-
const size_t h = ZSTD_hashPtr(ip, hlog, mls);
619-
const U32 matchIndex = hashTable[h];
620-
const BYTE* const matchBase = matchIndex < prefixStartIndex ? dictBase : base;
621-
const BYTE* match = matchBase + matchIndex;
622-
const U32 curr = (U32)(ip-base);
623-
const U32 repIndex = curr + 1 - offset_1;
624-
const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
625-
const BYTE* const repMatch = repBase + repIndex;
626-
hashTable[h] = curr; /* update hash table */
627-
DEBUGLOG(7, "offset_1 = %u , curr = %u", offset_1, curr);
628-
629-
if ( ( ((U32)((prefixStartIndex-1) - repIndex) >= 3) /* intentional underflow */
630-
& (offset_1 <= curr+1 - dictStartIndex) ) /* note: we are searching at curr+1 */
631-
&& (MEM_read32(repMatch) == MEM_read32(ip+1)) ) {
632-
const BYTE* const repMatchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
633-
size_t const rLength = ZSTD_count_2segments(ip+1 +4, repMatch +4, iend, repMatchEnd, prefixStart) + 4;
634-
ip++;
635-
ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, REPCODE1_TO_OFFBASE, rLength);
636-
ip += rLength;
637-
anchor = ip;
638-
} else {
639-
if ( (matchIndex < dictStartIndex) ||
640-
(MEM_read32(match) != MEM_read32(ip)) ) {
641-
assert(stepSize >= 1);
642-
ip += ((ip-anchor) >> kSearchStrength) + stepSize;
643-
continue;
636+
{ U32 const curr = (U32)(ip0 - base);
637+
U32 const maxRep = curr - dictStartIndex;
638+
if (offset_2 >= maxRep) offset_2 = 0;
639+
if (offset_1 >= maxRep) offset_1 = 0;
640+
}
641+
642+
/* start each op */
643+
_start: /* Requires: ip0 */
644+
645+
step = stepSize;
646+
nextStep = ip0 + kStepIncr;
647+
648+
/* calculate positions, ip0 - anchor == 0, so we skip step calc */
649+
ip1 = ip0 + 1;
650+
ip2 = ip0 + step;
651+
ip3 = ip2 + 1;
652+
653+
if (ip3 >= ilimit) {
654+
goto _cleanup;
655+
}
656+
657+
hash0 = ZSTD_hashPtr(ip0, hlog, mls);
658+
hash1 = ZSTD_hashPtr(ip1, hlog, mls);
659+
660+
idx = hashTable[hash0];
661+
idxBase = idx < prefixStartIndex ? dictBase : base;
662+
663+
do {
664+
{ /* load repcode match for ip[2] */
665+
U32 const current2 = (U32)(ip2 - base);
666+
U32 const repIndex = current2 - offset_1;
667+
const BYTE* const repBase = repIndex < prefixStartIndex ? dictBase : base;
668+
U32 rval;
669+
if ( ((U32)(prefixStartIndex - repIndex) >= 4) /* intentional underflow */
670+
& (offset_1 > 0) ) {
671+
rval = MEM_read32(repBase + repIndex);
672+
} else {
673+
rval = MEM_read32(ip2) ^ 1; /* guaranteed to not match. */
644674
}
645-
{ const BYTE* const matchEnd = matchIndex < prefixStartIndex ? dictEnd : iend;
646-
const BYTE* const lowMatchPtr = matchIndex < prefixStartIndex ? dictStart : prefixStart;
647-
U32 const offset = curr - matchIndex;
648-
size_t mLength = ZSTD_count_2segments(ip+4, match+4, iend, matchEnd, prefixStart) + 4;
649-
while (((ip>anchor) & (match>lowMatchPtr)) && (ip[-1] == match[-1])) { ip--; match--; mLength++; } /* catch up */
650-
offset_2 = offset_1; offset_1 = offset; /* update offset history */
651-
ZSTD_storeSeq(seqStore, (size_t)(ip-anchor), anchor, iend, OFFSET_TO_OFFBASE(offset), mLength);
652-
ip += mLength;
653-
anchor = ip;
675+
676+
/* write back hash table entry */
677+
current0 = (U32)(ip0 - base);
678+
hashTable[hash0] = current0;
679+
680+
/* check repcode at ip[2] */
681+
if (MEM_read32(ip2) == rval) {
682+
ip0 = ip2;
683+
match0 = repBase + repIndex;
684+
matchEnd = repIndex < prefixStartIndex ? dictEnd : iend;
685+
assert((match0 != prefixStart) & (match0 != dictStart));
686+
mLength = ip0[-1] == match0[-1];
687+
ip0 -= mLength;
688+
match0 -= mLength;
689+
offcode = REPCODE1_TO_OFFBASE;
690+
mLength += 4;
691+
goto _match;
654692
} }
655693

656-
if (ip <= ilimit) {
657-
/* Fill Table */
658-
hashTable[ZSTD_hashPtr(base+curr+2, hlog, mls)] = curr+2;
659-
hashTable[ZSTD_hashPtr(ip-2, hlog, mls)] = (U32)(ip-2-base);
660-
/* check immediate repcode */
661-
while (ip <= ilimit) {
662-
U32 const current2 = (U32)(ip-base);
663-
U32 const repIndex2 = current2 - offset_2;
664-
const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
665-
if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (offset_2 <= curr - dictStartIndex)) /* intentional overflow */
666-
&& (MEM_read32(repMatch2) == MEM_read32(ip)) ) {
667-
const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
668-
size_t const repLength2 = ZSTD_count_2segments(ip+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
669-
{ U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; } /* swap offset_2 <=> offset_1 */
670-
ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
671-
hashTable[ZSTD_hashPtr(ip, hlog, mls)] = current2;
672-
ip += repLength2;
673-
anchor = ip;
674-
continue;
675-
}
676-
break;
677-
} } }
694+
{ /* load match for ip[0] */
695+
U32 const mval = idx >= dictStartIndex ?
696+
MEM_read32(idxBase + idx) :
697+
MEM_read32(ip0) ^ 1; /* guaranteed not to match */
698+
699+
/* check match at ip[0] */
700+
if (MEM_read32(ip0) == mval) {
701+
/* found a match! */
702+
goto _offset;
703+
} }
704+
705+
/* lookup ip[1] */
706+
idx = hashTable[hash1];
707+
idxBase = idx < prefixStartIndex ? dictBase : base;
708+
709+
/* hash ip[2] */
710+
hash0 = hash1;
711+
hash1 = ZSTD_hashPtr(ip2, hlog, mls);
712+
713+
/* advance to next positions */
714+
ip0 = ip1;
715+
ip1 = ip2;
716+
ip2 = ip3;
717+
718+
/* write back hash table entry */
719+
current0 = (U32)(ip0 - base);
720+
hashTable[hash0] = current0;
721+
722+
{ /* load match for ip[0] */
723+
U32 const mval = idx >= dictStartIndex ?
724+
MEM_read32(idxBase + idx) :
725+
MEM_read32(ip0) ^ 1; /* guaranteed not to match */
726+
727+
/* check match at ip[0] */
728+
if (MEM_read32(ip0) == mval) {
729+
/* found a match! */
730+
goto _offset;
731+
} }
732+
733+
/* lookup ip[1] */
734+
idx = hashTable[hash1];
735+
idxBase = idx < prefixStartIndex ? dictBase : base;
736+
737+
/* hash ip[2] */
738+
hash0 = hash1;
739+
hash1 = ZSTD_hashPtr(ip2, hlog, mls);
740+
741+
/* advance to next positions */
742+
ip0 = ip1;
743+
ip1 = ip2;
744+
ip2 = ip0 + step;
745+
ip3 = ip1 + step;
746+
747+
/* calculate step */
748+
if (ip2 >= nextStep) {
749+
step++;
750+
PREFETCH_L1(ip1 + 64);
751+
PREFETCH_L1(ip1 + 128);
752+
nextStep += kStepIncr;
753+
}
754+
} while (ip3 < ilimit);
755+
756+
_cleanup:
757+
/* Note that there are probably still a couple positions we could search.
758+
* However, it seems to be a meaningful performance hit to try to search
759+
* them. So let's not. */
678760

679761
/* save reps for next block */
680-
rep[0] = offset_1;
681-
rep[1] = offset_2;
762+
rep[0] = offset_1 ? offset_1 : rep[0];
763+
rep[1] = offset_2 ? offset_2 : rep[1];
682764

683765
/* Return the last literals size */
684766
return (size_t)(iend - anchor);
767+
768+
_offset: /* Requires: ip0, idx, idxBase */
769+
770+
/* Compute the offset code. */
771+
{ U32 const offset = current0 - idx;
772+
const BYTE* const lowMatchPtr = idx < prefixStartIndex ? dictStart : prefixStart;
773+
matchEnd = idx < prefixStartIndex ? dictEnd : iend;
774+
match0 = idxBase + idx;
775+
offset_2 = offset_1;
776+
offset_1 = offset;
777+
offcode = OFFSET_TO_OFFBASE(offset);
778+
mLength = 4;
779+
780+
/* Count the backwards match length. */
781+
while (((ip0>anchor) & (match0>lowMatchPtr)) && (ip0[-1] == match0[-1])) {
782+
ip0--;
783+
match0--;
784+
mLength++;
785+
} }
786+
787+
_match: /* Requires: ip0, match0, offcode, matchEnd */
788+
789+
/* Count the forward length. */
790+
assert(matchEnd != 0);
791+
mLength += ZSTD_count_2segments(ip0 + mLength, match0 + mLength, iend, matchEnd, prefixStart);
792+
793+
ZSTD_storeSeq(seqStore, (size_t)(ip0 - anchor), anchor, iend, offcode, mLength);
794+
795+
ip0 += mLength;
796+
anchor = ip0;
797+
798+
/* write next hash table entry */
799+
if (ip1 < ip0) {
800+
hashTable[hash1] = (U32)(ip1 - base);
801+
}
802+
803+
/* Fill table and check for immediate repcode. */
804+
if (ip0 <= ilimit) {
805+
/* Fill Table */
806+
assert(base+current0+2 > istart); /* check base overflow */
807+
hashTable[ZSTD_hashPtr(base+current0+2, hlog, mls)] = current0+2; /* here because current+2 could be > iend-8 */
808+
hashTable[ZSTD_hashPtr(ip0-2, hlog, mls)] = (U32)(ip0-2-base);
809+
810+
while (ip0 <= ilimit) {
811+
U32 const repIndex2 = (U32)(ip0-base) - offset_2;
812+
const BYTE* const repMatch2 = repIndex2 < prefixStartIndex ? dictBase + repIndex2 : base + repIndex2;
813+
if ( (((U32)((prefixStartIndex-1) - repIndex2) >= 3) & (offset_2 > 0)) /* intentional underflow */
814+
&& (MEM_read32(repMatch2) == MEM_read32(ip0)) ) {
815+
const BYTE* const repEnd2 = repIndex2 < prefixStartIndex ? dictEnd : iend;
816+
size_t const repLength2 = ZSTD_count_2segments(ip0+4, repMatch2+4, iend, repEnd2, prefixStart) + 4;
817+
{ U32 const tmpOffset = offset_2; offset_2 = offset_1; offset_1 = tmpOffset; } /* swap offset_2 <=> offset_1 */
818+
ZSTD_storeSeq(seqStore, 0 /*litlen*/, anchor, iend, REPCODE1_TO_OFFBASE, repLength2);
819+
hashTable[ZSTD_hashPtr(ip0, hlog, mls)] = (U32)(ip0-base);
820+
ip0 += repLength2;
821+
anchor = ip0;
822+
continue;
823+
}
824+
break;
825+
} }
826+
827+
goto _start;
685828
}
686829

687830
ZSTD_GEN_FAST_FN(extDict, 4, 0)
@@ -694,6 +837,7 @@ size_t ZSTD_compressBlock_fast_extDict(
694837
void const* src, size_t srcSize)
695838
{
696839
U32 const mls = ms->cParams.minMatch;
840+
assert(ms->dictMatchState == NULL);
697841
switch(mls)
698842
{
699843
default: /* includes case 3 */

0 commit comments

Comments
 (0)