Compress LZSA2 a bit faster again

This commit is contained in:
Emmanuel Marty 2021-10-10 07:52:03 +02:00 committed by GitHub
parent c6a93601cf
commit bb1b4fda14
No known key found for this signature in database
GPG Key ID: 4AEE18F83AFDEB23

View File

@ -194,10 +194,12 @@ static void lzsa_insert_forward_match_v2(lzsa_compressor *pCompressor, const uns
for (j = 0; j < NARRIVALS_PER_POSITION_V2_BIG && arrival[j].from_slot; j++) { for (j = 0; j < NARRIVALS_PER_POSITION_V2_BIG && arrival[j].from_slot; j++) {
const int nRepOffset = arrival[j].rep_offset; const int nRepOffset = arrival[j].rep_offset;
if (nMatchOffset != nRepOffset && nRepOffset && arrival[j].rep_len >= MIN_MATCH_SIZE_V2) { if (nMatchOffset != nRepOffset && nRepOffset) {
const int nRepPos = arrival[j].rep_pos;
const int nRepLen = arrival[j].rep_len; const int nRepLen = arrival[j].rep_len;
if (nRepLen >= MIN_MATCH_SIZE_V2) {
const int nRepPos = arrival[j].rep_pos;
if (nRepPos > nMatchOffset && if (nRepPos > nMatchOffset &&
(nRepPos + nRepLen) <= nEndOffset) { (nRepPos + nRepLen) <= nEndOffset) {
@ -256,6 +258,7 @@ static void lzsa_insert_forward_match_v2(lzsa_compressor *pCompressor, const uns
} }
} }
} }
}
/** /**
* Attempt to pick optimal matches using a forward arrivals parser, so as to produce the smallest possible output that decompresses to the same input * Attempt to pick optimal matches using a forward arrivals parser, so as to produce the smallest possible output that decompresses to the same input
@ -525,16 +528,26 @@ static void lzsa_optimize_forward_v2(lzsa_compressor *pCompressor, const unsigne
if (n < nArrivalsPerPosition - 1) { if (n < nArrivalsPerPosition - 1) {
int nn; int nn;
if (!nInsertForwardReps || pDestSlots[nArrivalsPerPosition - 1].from_slot) {
for (nn = n; for (nn = n;
nn < nArrivalsPerPosition && pDestSlots[nn].cost == nCodingChoiceCost; nn < nArrivalsPerPosition && pDestSlots[nn].cost == nCodingChoiceCost;
nn++) { nn++) {
if (pDestSlots[nn].rep_offset == nMatchOffset && if (pDestSlots[nn].rep_offset == nMatchOffset) {
(!nInsertForwardReps || pDestSlots[nn].rep_pos >= i ||
pDestSlots[nArrivalsPerPosition - 1].from_slot)) {
exists = 1; exists = 1;
break; break;
} }
} }
}
else {
for (nn = n;
nn < nArrivalsPerPosition && pDestSlots[nn].cost == nCodingChoiceCost;
nn++) {
if (pDestSlots[nn].rep_offset == nMatchOffset && pDestSlots[nn].rep_pos >= i) {
exists = 1;
break;
}
}
}
if (!exists) { if (!exists) {
int z; int z;
@ -561,7 +574,7 @@ static void lzsa_optimize_forward_v2(lzsa_compressor *pCompressor, const unsigne
pDestArrival->rep_offset = nMatchOffset; pDestArrival->rep_offset = nMatchOffset;
pDestArrival->rep_pos = i; pDestArrival->rep_pos = i;
pDestArrival->rep_len = k; pDestArrival->rep_len = k;
nRepLenHandledMask[k >> 3] &= ~(1 << (k & 7)); nRepLenHandledMask[k >> 3] &= ~((1 ^ nReduce) << (k & 7));
} }
} }
} }
@ -577,8 +590,10 @@ static void lzsa_optimize_forward_v2(lzsa_compressor *pCompressor, const unsigne
nRepLenHandledMask[k >> 3] |= 1 << (k & 7); nRepLenHandledMask[k >> 3] |= 1 << (k & 7);
for (nCurRepMatchArrival = 0; (j = nRepMatchArrivalIdxAndLen[nCurRepMatchArrival]) >= 0; nCurRepMatchArrival += 2) { for (nCurRepMatchArrival = 0; (j = nRepMatchArrivalIdxAndLen[nCurRepMatchArrival]) >= 0; nCurRepMatchArrival += 2) {
if (nRepMatchArrivalIdxAndLen[nCurRepMatchArrival + 1] >= k) {
int nMaskOffset = (j << 7) + (k >> 3); int nMaskOffset = (j << 7) + (k >> 3);
if (nRepMatchArrivalIdxAndLen[nCurRepMatchArrival + 1] >= k && (nReduce || !(nRepSlotHandledMask[nMaskOffset] & (1 << (k & 7))))) {
if (nReduce || !(nRepSlotHandledMask[nMaskOffset] & (1 << (k & 7)))) {
const int nPrevCost = cur_arrival[j].cost & 0x3fffffff; const int nPrevCost = cur_arrival[j].cost & 0x3fffffff;
int nRepCodingChoiceCost = nPrevCost /* the actual cost of the literals themselves accumulates up the chain */ + nMatchLenCost; int nRepCodingChoiceCost = nPrevCost /* the actual cost of the literals themselves accumulates up the chain */ + nMatchLenCost;
int nScore = cur_arrival[j].score + 2; int nScore = cur_arrival[j].score + 2;
@ -644,7 +659,7 @@ static void lzsa_optimize_forward_v2(lzsa_compressor *pCompressor, const unsigne
pDestArrival->rep_offset = nRepOffset; pDestArrival->rep_offset = nRepOffset;
pDestArrival->rep_pos = i; pDestArrival->rep_pos = i;
pDestArrival->rep_len = k; pDestArrival->rep_len = k;
nRepLenHandledMask[k >> 3] &= ~(1 << (k & 7)); nRepLenHandledMask[k >> 3] &= ~((1 ^ nReduce) << (k & 7));
} }
} }
} }
@ -655,6 +670,7 @@ static void lzsa_optimize_forward_v2(lzsa_compressor *pCompressor, const unsigne
} }
} }
} }
}
if (k < nMaxRepInsertedLen) if (k < nMaxRepInsertedLen)
nMinOverallRepLen = k; nMinOverallRepLen = k;