2 * Copyright (c) 2007-2012, Novell Inc.
4 * This program is licensed under the BSD license, read LICENSE.BSD
5 * for further information
11 * Paging and compression functions for the vertical repository data.
12 * Vertical data is grouped by key, normal data is grouped by solvable.
13 * This makes searching for a string in vertical data fast as there's
14 * no need to skip over data if keys we're not interested in.
16 * The vertical data is split into pages, each page is compressed with a fast
17 * compression algorithm. These pages are read in on demand, not recently used
18 * pages automatically get dropped.
21 #define _XOPEN_SOURCE 500
23 #include <sys/types.h>
38 #define BLOCK_SIZE (65536*1)
39 #if BLOCK_SIZE <= 65536
46 The format is tailored for fast decompression (i.e. only byte based),
47 and skewed to ASCII content (highest bit often not set):
50 - self-describing ASCII character hex L
51 b 100lllll <l+1 bytes>
52 - literal run of length l+1
54 - back ref of length l+2, at offset -(o+1) (o < 1 << 10)
56 - back ref of length l+2+8, at offset -(o+1) (o < 1 << 8)
58 - back ref of length l+3, at offset -(o+1) (o < 1 << 16)
59 f1 1111llll <8l> <8o> <8o>
60 - back ref, length l+19 (l < 1<<12), offset -(o+1) (o < 1<<16)
61 f2 11110lll <8l> <8o> <8o>
62 - back ref, length l+19 (l < 1<<11), offset -(o+1) (o < 1<<16)
63 g 11111lll <8l> <8o> <8o> <8o>
64 - back ref, length l+5 (l < 1<<11), offset -(o+1) (o < 1<<24)
66 Generally for a literal of length L we need L+1 bytes, hence it is
67 better to encode also very short backrefs (2 chars) as backrefs if
68 their offset is small, as that only needs two bytes. Except if we
69 already have a literal run, in that case it's better to append there,
70 instead of breaking it for a backref. So given a potential backref
71 at offset O, length L the strategy is as follows:
73 L < 2 : encode as 1-literal
74 L == 2, O > 1024 : encode as 1-literal
75 L == 2, have already literals: encode as 1-literal
77 L >= 2, L <= 9, O < 1024 : encode as c
78 L >= 10, L <= 41, O < 256 : encode as d
79 else we have either O >= 1024, or L >= 42:
80 L < 3 : encode as 1-literal
81 L >= 3, L <= 18, O < 65536 : encode as e
82 L >= 19, L <= 4095+18, O < 65536 : encode as f
83 else we have either L >= 4096+18 or O >= 65536.
84 O >= 65536: encode as 1-literal, too bad
85 (with the current block size this can't happen)
86 L >= 4096+18, so reduce to 4095+18 : encode as f
91 compress_buf(const unsigned char *in, unsigned int in_len,
92 unsigned char *out, unsigned int out_len)
94 unsigned int oo = 0; /* out-offset */
95 unsigned int io = 0; /* in-offset */
98 Ref hnext[BLOCK_SIZE];
99 unsigned int litofs = 0;
100 memset(htab, -1, sizeof (htab));
101 memset(hnext, -1, sizeof (hnext));
102 while (io + 2 < in_len)
104 /* Search for a match of the string starting at IN, we have at
105 least three characters. */
106 unsigned int hval = in[io] | in[io + 1] << 8 | in[io + 2] << 16;
107 unsigned int try, mlen, mofs, tries;
108 hval = (hval ^ (hval << 5) ^ (hval >> 5)) - hval * 5;
109 hval = hval & (HS - 1);
111 hnext[io] = htab[hval];
116 for (tries = 0; try != -1 && tries < 12; tries++)
119 && in[try] == in[io] && in[try + 1] == in[io + 1])
122 mofs = (io - try) - 1;
127 for (; try != -1 && tries < 12; tries++)
129 /* assert(mlen >= 2); */
130 /* assert(io + mlen < in_len); */
131 /* Try a match starting from [io] with the strings at [try].
132 That's only sensible if TRY actually is before IO (can happen
133 with uninit hash table). If we have a previous match already
134 we're only going to take the new one if it's longer, hence
135 check the potentially last character. */
136 if (try < io && in[try + mlen] == in[io + mlen])
138 unsigned int this_len, this_ofs;
139 if (memcmp(in + try, in + io, mlen))
142 /* Now try extending the match by more characters. */
144 io + this_len < in_len
145 && in[try + this_len] == in[io + this_len]; this_len++)
149 for (testi = 0; testi < this_len; testi++)
150 assert(in[try + testi] == in[io + testi]);
152 this_ofs = (io - try) - 1;
153 /*if (this_ofs > 65535)
156 assert(this_len >= 2);
157 assert(this_len >= mlen);
158 assert(this_len > mlen || (this_len == mlen && this_ofs > mofs));
160 mlen = this_len, mofs = this_ofs;
161 /* If our match extends up to the end of input, no next
162 match can become better. This is not just an
163 optimization, it establishes a loop invariant
164 (io + mlen < in_len). */
165 if (io + mlen >= in_len)
170 /*if (io - try - 1 >= 65536)
177 /*fprintf(stderr, "%d %d\n", mlen, mofs);*/
178 if (mlen == 2 && (litofs || mofs >= 1024))
180 /*else if (mofs >= 65536)
182 else if (mofs >= 65536)
184 if (mlen >= 2048 + 5)
191 /*else if (mlen >= 4096 + 19)
193 else if (mlen >= 2048 + 19)
195 /* Skip this match if the next character would deliver a better one,
196 but only do this if we have the chance to really extend the
197 length (i.e. our current length isn't yet the (conservative)
199 if (mlen && mlen < (2048 + 5) && io + 3 < in_len)
202 in[io + 1] | in[io + 2] << 8 | in[io + 3] << 16;
204 hval = (hval ^ (hval << 5) ^ (hval >> 5)) - hval * 5;
205 hval = hval & (HS - 1);
208 && in[try] == in[io + 1] && in[try + 1] == in[io + 2])
210 unsigned int this_len;
213 io + 1 + this_len < in_len
214 && in[try + this_len] == in[io + 1 + this_len];
217 if (this_len >= mlen)
234 litlen = io - litofs;
235 /* fprintf(stderr, "lit: %d\n", litlen); */
238 unsigned int easy_sz;
239 /* Emit everything we can as self-describers. As soon as
240 we hit a byte we can't emit as such we're going to emit
241 a length descriptor anyway, so we can as well include
242 bytes < 0x80 which might follow afterwards in that run. */
244 easy_sz < litlen && in[litofs + easy_sz] < 0x80;
249 if (oo + easy_sz >= out_len)
251 memcpy(out + oo, in + litofs, easy_sz);
260 if (oo + 1 + litlen >= out_len)
262 out[oo++] = 0x80 | (litlen - 1);
264 out[oo++] = in[litofs++];
269 /* Literal length > 32, so chunk it. */
270 if (oo + 1 + 32 >= out_len)
272 out[oo++] = 0x80 | 31;
273 memcpy(out + oo, in + litofs, 32);
282 /* fprintf(stderr, "ref: %d @ %d\n", mlen, mofs); */
284 if (mlen >= 2 && mlen <= 9 && mofs < 1024)
286 if (oo + 2 >= out_len)
288 out[oo++] = 0xa0 | ((mofs & 0x300) >> 5) | (mlen - 2);
289 out[oo++] = mofs & 0xff;
291 else if (mlen >= 10 && mlen <= 41 && mofs < 256)
293 if (oo + 2 >= out_len)
295 out[oo++] = 0xc0 | (mlen - 10);
298 else if (mofs >= 65536)
300 assert(mlen >= 5 && mlen < 2048 + 5);
301 if (oo + 5 >= out_len)
303 out[oo++] = 0xf8 | ((mlen - 5) >> 8);
304 out[oo++] = (mlen - 5) & 0xff;
305 out[oo++] = mofs & 0xff;
306 out[oo++] = (mofs >> 8) & 0xff;
307 out[oo++] = mofs >> 16;
309 else if (mlen >= 3 && mlen <= 18)
311 assert(mofs < 65536);
312 if (oo + 3 >= out_len)
314 out[oo++] = 0xe0 | (mlen - 3);
315 out[oo++] = mofs & 0xff;
316 out[oo++] = mofs >> 8;
320 assert(mlen >= 19 && mlen <= 4095 + 19 && mofs < 65536);
321 if (oo + 4 >= out_len)
323 out[oo++] = 0xf0 | ((mlen - 19) >> 8);
324 out[oo++] = (mlen - 19) & 0xff;
325 out[oo++] = mofs & 0xff;
326 out[oo++] = mofs >> 8;
328 /* Insert the hashes for the compressed run [io..io+mlen-1].
329 For [io] we have it already done at the start of the loop.
330 So it's from [io+1..io+mlen-1], and we need three chars per
331 hash, so the accessed characters will be [io+1..io+mlen-1+2],
332 ergo io+mlen+1 < in_len. */
340 in[io] | in[io + 1] << 8 | in[io + 2] << 16;
341 hval = (hval ^ (hval << 5) ^ (hval >> 5)) - hval * 5;
342 hval = hval & (HS - 1);
343 hnext[io] = htab[hval];
350 /* We might have some characters left. */
351 if (io < in_len && !litofs)
358 litlen = io - litofs;
359 /* fprintf(stderr, "lit: %d\n", litlen); */
362 unsigned int easy_sz;
363 /* Emit everything we can as self-describers. As soon as we hit a
364 byte we can't emit as such we're going to emit a length
365 descriptor anyway, so we can as well include bytes < 0x80 which
366 might follow afterwards in that run. */
367 for (easy_sz = 0; easy_sz < litlen && in[litofs + easy_sz] < 0x80;
372 if (oo + easy_sz >= out_len)
374 memcpy(out + oo, in + litofs, easy_sz);
383 if (oo + 1 + litlen >= out_len)
385 out[oo++] = 0x80 | (litlen - 1);
387 out[oo++] = in[litofs++];
392 /* Literal length > 32, so chunk it. */
393 if (oo + 1 + 32 >= out_len)
395 out[oo++] = 0x80 | 31;
396 memcpy(out + oo, in + litofs, 32);
408 unchecked_decompress_buf(const unsigned char *in, unsigned int in_len,
410 unsigned int out_len __attribute__((unused)))
412 unsigned char *orig_out = out;
413 const unsigned char *in_end = in + in_len;
416 unsigned int first = *in++;
421 /* This default case can't happen, but GCCs VRP is not strong
422 enough to see this, so make this explicitely not fall to
423 the end of the switch, so that we don't have to initialize
431 /* fprintf (stderr, "lit: 1\n"); */
435 /* b 100lllll <l+1 bytes> */
437 unsigned int l = first & 31;
438 /* fprintf (stderr, "lit: %d\n", l); */
445 /* c 101oolll <8o> */
447 o = first & (3 << 3);
448 o = (o << 5) | *in++;
449 first = (first & 7) + 2;
453 /* d 110lllll <8o> */
456 first = (first & 31) + 10;
460 /* e 1110llll <8o> <8o> */
462 o = in[0] | (in[1] << 8);
469 /* f1 1111llll <8o> <8o> <8l> */
470 /* f2 11110lll <8o> <8o> <8l> */
471 /* g 11111lll <8o> <8o> <8o> <8l> */
476 first = (((first - 8) << 8) | in[0]) + 5;
477 o = in[1] | (in[2] << 8) | (in[3] << 16);
482 first = ((first << 8) | in[0]) + 19;
483 o = in[1] | (in[2] << 8);
489 /* fprintf(stderr, "ref: %d @ %d\n", first, o); */
493 /* We know that first will not be zero, and this loop structure is
494 better optimizable. */
504 case 18: *out = *(out + o); out++;
505 case 17: *out = *(out + o); out++;
506 case 16: *out = *(out + o); out++;
507 case 15: *out = *(out + o); out++;
508 case 14: *out = *(out + o); out++;
509 case 13: *out = *(out + o); out++;
510 case 12: *out = *(out + o); out++;
511 case 11: *out = *(out + o); out++;
512 case 10: *out = *(out + o); out++;
513 case 9: *out = *(out + o); out++;
514 case 8: *out = *(out + o); out++;
515 case 7: *out = *(out + o); out++;
516 case 6: *out = *(out + o); out++;
517 case 5: *out = *(out + o); out++;
518 case 4: *out = *(out + o); out++;
519 case 3: *out = *(out + o); out++;
520 case 2: *out = *(out + o); out++;
521 case 1: *out = *(out + o); out++;
529 case 0: *out = *(out + o); out++;
530 case 15: *out = *(out + o); out++;
531 case 14: *out = *(out + o); out++;
532 case 13: *out = *(out + o); out++;
533 case 12: *out = *(out + o); out++;
534 case 11: *out = *(out + o); out++;
535 case 10: *out = *(out + o); out++;
536 case 9: *out = *(out + o); out++;
537 case 8: *out = *(out + o); out++;
538 case 7: *out = *(out + o); out++;
539 case 6: *out = *(out + o); out++;
540 case 5: *out = *(out + o); out++;
541 case 4: *out = *(out + o); out++;
542 case 3: *out = *(out + o); out++;
543 case 2: *out = *(out + o); out++;
544 case 1: *out = *(out + o); out++;
546 while ((int)(first -= 16) > 0);
552 return out - orig_out;
555 /**********************************************************************/
557 void repopagestore_init(Repopagestore *store)
559 memset(store, 0, sizeof(*store));
563 void repopagestore_free(Repopagestore *store)
565 solv_free(store->blob_store);
566 solv_free(store->pages);
567 solv_free(store->mapped);
568 if (store->pagefd != -1)
569 close(store->pagefd);
573 /**********************************************************************/
576 repopagestore_load_page_range(Repopagestore *store, unsigned int pstart, unsigned int pend)
578 /* Make sure all pages from PSTART to PEND (inclusive) are loaded,
579 and are consecutive. Return a pointer to the mapping of PSTART. */
580 unsigned char buf[REPOPAGE_BLOBSIZE];
581 unsigned int i, best;
585 /* Quick check in case the requested page is already mapped */
586 if (store->pages[pstart].mapped_at != -1)
587 return store->blob_store + store->pages[pstart].mapped_at;
591 /* Quick check in case all pages are already mapped and consecutive. */
592 for (i = pstart; i <= pend; i++)
593 if (store->pages[i].mapped_at == -1
595 && store->pages[i].mapped_at
596 != store->pages[i-1].mapped_at + REPOPAGE_BLOBSIZE))
599 return store->blob_store + store->pages[pstart].mapped_at;
602 if (store->pagefd == -1)
606 fprintf(stderr, "PAGE: want %d pages starting at %d\n", pend - pstart + 1, pstart);
609 /* Ensure that we can map the numbers of pages we need at all. */
610 if (pend - pstart + 1 > store->nmapped)
612 unsigned int oldcan = store->nmapped;
613 store->nmapped = pend - pstart + 1;
614 if (store->nmapped < 4)
616 store->mapped = solv_realloc2(store->mapped, store->nmapped, sizeof(store->mapped[0]));
617 memset(store->mapped + oldcan, 0, (store->nmapped - oldcan) * sizeof (store->mapped[0]));
618 store->blob_store = solv_realloc2(store->blob_store, store->nmapped, REPOPAGE_BLOBSIZE);
620 fprintf(stderr, "PAGE: can map %d pages\n", store->nmapped);
624 if (store->pages[pstart].mapped_at != -1)
626 /* assume forward search */
627 best = store->pages[pstart].mapped_at / REPOPAGE_BLOBSIZE;
628 if (best + (pend - pstart) >= store->nmapped)
631 else if (store->pages[pend].mapped_at != -1)
633 /* assume backward search */
634 best = store->pages[pend].mapped_at / REPOPAGE_BLOBSIZE;
635 if (best < pend - pstart)
636 best = store->nmapped - 1;
637 best -= pend - pstart;
641 /* choose some "random" location to avoid thrashing */
642 best = (pstart + store->rr_counter++) % (store->nmapped - pend + pstart);
645 /* So we want to map our pages from [best] to [best+pend-pstart].
646 Use a very simple strategy, which doesn't make the best use of
647 our resources, but works. Throw away all pages in that range
648 (even ours) then copy around ours or read them in. */
649 for (i = best; i < best + pend - pstart + 1; i++)
651 unsigned int pnum = store->mapped[i];
653 /* If this page is exactly at the right place already,
654 no need to evict it. */
655 && pnum != pstart + i - best)
658 /* Evict this page. */
660 fprintf(stderr, "PAGE: evict page %d from %d\n", pnum, i);
662 store->mapped[i] = 0;
663 store->pages[pnum].mapped_at = -1;
664 /* check if we can copy the correct content */
665 p = store->pages + (pstart + i - best);
666 if (p->mapped_at != -1 && p->mapped_at != i * REPOPAGE_BLOBSIZE)
668 void *dest = store->blob_store + i * REPOPAGE_BLOBSIZE;
670 fprintf(stderr, "PAGECOPY: %d from %d to %d\n", pstart + i - best, p->mapped_at / REPOPAGE_BLOBSIZE, i);
672 memcpy(dest, store->blob_store + p->mapped_at, REPOPAGE_BLOBSIZE);
673 store->mapped[p->mapped_at / REPOPAGE_BLOBSIZE] = 0;
674 p->mapped_at = i * REPOPAGE_BLOBSIZE;
675 store->mapped[i] = pstart + i - best + 1;
680 /* Everything is free now. Read in the pages we want. */
681 for (i = pstart; i <= pend; i++)
683 Attrblobpage *p = store->pages + i;
684 unsigned int pnum = i - pstart + best;
685 void *dest = store->blob_store + pnum * REPOPAGE_BLOBSIZE;
686 if (p->mapped_at != -1)
688 if (p->mapped_at != pnum * REPOPAGE_BLOBSIZE)
691 fprintf(stderr, "PAGECOPY: %d from %d to %d\n", i, p->mapped_at / REPOPAGE_BLOBSIZE, pnum);
693 /* Still mapped somewhere else, so just copy it from there. */
694 memcpy(dest, store->blob_store + p->mapped_at, REPOPAGE_BLOBSIZE);
695 store->mapped[p->mapped_at / REPOPAGE_BLOBSIZE] = 0;
700 unsigned int in_len = p->page_size;
701 unsigned int compressed = in_len & 1;
704 fprintf(stderr, "PAGEIN: %d to %d", i, pnum);
706 if (pread(store->pagefd, compressed ? buf : dest, in_len, store->file_offset + p->page_offset) != in_len)
708 perror("mapping pread");
713 unsigned int out_len;
714 out_len = unchecked_decompress_buf(buf, in_len, dest, REPOPAGE_BLOBSIZE);
715 if (out_len != REPOPAGE_BLOBSIZE && i < store->num_pages - 1)
718 fprintf(stderr, "can't decompress\n");
723 fprintf(stderr, " (expand %d to %d)", in_len, out_len);
727 fprintf(stderr, "\n");
730 p->mapped_at = pnum * REPOPAGE_BLOBSIZE;
731 store->mapped[pnum] = i + 1;
733 return store->blob_store + best * REPOPAGE_BLOBSIZE;
737 repopagestore_compress_page(unsigned char *page, unsigned int len, unsigned char *cpage, unsigned int max)
739 return compress_buf(page, len, cpage, max);
742 #define SOLV_ERROR_EOF 3
743 #define SOLV_ERROR_CORRUPT 6
745 static inline unsigned int
751 for (i = 0; i < 4; i++)
761 /* Try to either setup on-demand paging (using FP as backing
762 file), or in case that doesn't work (FP not seekable) slurps in
763 all pages and deactivates paging. */
765 repopagestore_read_or_setup_pages(Repopagestore *store, FILE *fp, unsigned int pagesz, unsigned int blobsz)
769 unsigned int can_seek;
770 unsigned int cur_page_ofs;
771 unsigned char buf[REPOPAGE_BLOBSIZE];
773 if (pagesz != REPOPAGE_BLOBSIZE)
775 /* We could handle this by slurping in everything. */
776 return SOLV_ERROR_CORRUPT;
779 if ((store->file_offset = ftell(fp)) < 0)
783 store->pagefd = dup(fileno(fp));
784 if (store->pagefd == -1)
787 fcntl(store->pagefd, F_SETFD, FD_CLOEXEC);
790 fprintf(stderr, "can %sseek\n", can_seek ? "" : "NOT ");
792 npages = (blobsz + REPOPAGE_BLOBSIZE - 1) / REPOPAGE_BLOBSIZE;
794 store->num_pages = npages;
795 store->pages = solv_malloc2(npages, sizeof(store->pages[0]));
797 /* If we can't seek on our input we have to slurp in everything. */
799 store->blob_store = solv_malloc2(npages, REPOPAGE_BLOBSIZE);
801 for (i = 0; i < npages; i++)
803 Attrblobpage *p = store->pages + i;
804 unsigned int in_len = read_u32(fp);
805 unsigned int compressed = in_len & 1;
808 fprintf(stderr, "page %d: len %d (%scompressed)\n",
809 i, in_len, compressed ? "" : "not ");
815 p->page_offset = cur_page_ofs;
816 p->page_size = in_len * 2 + compressed;
817 if (fseek(fp, in_len, SEEK_CUR) < 0)
819 /* We can't fall back to non-seeking behaviour as we already
820 read over some data pages without storing them away. */
821 close(store->pagefd);
823 return SOLV_ERROR_EOF;
825 cur_page_ofs += in_len;
829 unsigned int out_len;
830 void *dest = store->blob_store + i * REPOPAGE_BLOBSIZE;
831 p->mapped_at = i * REPOPAGE_BLOBSIZE;
834 /* We can't seek, so suck everything in. */
835 if (fread(compressed ? buf : dest, in_len, 1, fp) != 1)
838 return SOLV_ERROR_EOF;
842 out_len = unchecked_decompress_buf(buf, in_len, dest, REPOPAGE_BLOBSIZE);
843 if (out_len != REPOPAGE_BLOBSIZE && i < npages - 1)
845 return SOLV_ERROR_CORRUPT;
854 repopagestore_disable_paging(Repopagestore *store)
856 if (store->num_pages)
857 repopagestore_load_page_range(store, 0, store->num_pages - 1);
863 transfer_file(FILE * from, FILE * to, int compress)
865 unsigned char inb[BLOCK_SIZE];
866 unsigned char outb[BLOCK_SIZE];
867 while (!feof (from) && !ferror (from))
869 unsigned int in_len, out_len;
872 in_len = fread(inb, 1, BLOCK_SIZE, from);
875 unsigned char *b = outb;
876 out_len = compress_buf(inb, in_len, outb, sizeof (outb));
878 b = inb, out_len = in_len;
879 if (fwrite(&out_len, sizeof (out_len), 1, to) != 1)
881 perror("write size");
884 if (fwrite(b, out_len, 1, to) != 1)
886 perror("write data");
893 if (fread(&in_len, sizeof(in_len), 1, from) != 1)
897 perror("can't read size");
900 if (fread(inb, in_len, 1, from) != 1)
902 perror("can't read data");
906 unchecked_decompress_buf(inb, in_len, outb, sizeof(outb));
907 if (fwrite(outb, out_len, 1, to) != 1)
909 perror("can't write output");
916 /* Just for benchmarking purposes. */
918 dumb_memcpy(void *dest, const void *src, unsigned int len)
927 benchmark(FILE * from)
929 unsigned char inb[BLOCK_SIZE];
930 unsigned char outb[BLOCK_SIZE];
931 unsigned int in_len = fread(inb, 1, BLOCK_SIZE, from);
932 unsigned int out_len;
935 perror("can't read from input");
939 unsigned int calib_loop;
940 unsigned int per_loop;
949 while ((clock() - start) < CLOCKS_PER_SEC / 4)
952 for (i = 0; i < calib_loop; i++)
953 dumb_memcpy(outb, inb, in_len);
954 per_loop += calib_loop;
957 fprintf(stderr, "memcpy:\nCalibrated to %d iterations per loop\n",
961 for (i = 0; i < 10; i++)
962 for (j = 0; j < per_loop; j++)
963 dumb_memcpy(outb, inb, in_len);
965 seconds = (end - start) / (float) CLOCKS_PER_SEC;
966 fprintf(stderr, "%.2f seconds == %.2f MB/s\n", seconds,
967 ((long long) in_len * per_loop * 10) / (1024 * 1024 * seconds));
973 while ((clock() - start) < CLOCKS_PER_SEC / 4)
976 for (i = 0; i < calib_loop; i++)
977 compress_buf(inb, in_len, outb, sizeof(outb));
978 per_loop += calib_loop;
981 fprintf(stderr, "compression:\nCalibrated to %d iterations per loop\n",
985 for (i = 0; i < 10; i++)
986 for (j = 0; j < per_loop; j++)
987 compress_buf(inb, in_len, outb, sizeof(outb));
989 seconds = (end - start) / (float) CLOCKS_PER_SEC;
990 fprintf(stderr, "%.2f seconds == %.2f MB/s\n", seconds,
991 ((long long) in_len * per_loop * 10) / (1024 * 1024 * seconds));
993 out_len = compress_buf(inb, in_len, outb, sizeof(outb));
998 while ((clock() - start) < CLOCKS_PER_SEC / 4)
1001 for (i = 0; i < calib_loop; i++)
1002 unchecked_decompress_buf(outb, out_len, inb, sizeof(inb));
1003 per_loop += calib_loop;
1006 fprintf(stderr, "decompression:\nCalibrated to %d iterations per loop\n",
1010 for (i = 0; i < 10; i++)
1011 for (j = 0; j < per_loop; j++)
1012 unchecked_decompress_buf(outb, out_len, inb, sizeof(inb));
1014 seconds = (end - start) / (float) CLOCKS_PER_SEC;
1015 fprintf(stderr, "%.2f seconds == %.2f MB/s\n", seconds,
1016 ((long long) in_len * per_loop * 10) / (1024 * 1024 * seconds));
1020 main(int argc, char *argv[])
1023 if (argc > 1 && !strcmp(argv[1], "-d"))
1025 if (argc > 1 && !strcmp(argv[1], "-b"))
1028 transfer_file(stdin, stdout, compress);