Whenever we try to enter a copy fast-path, there is a certain cost in checking
that all the preconditions are in place, but it's normally offset by the fact
that we can usually take the cheaper path. However, in a certain path we've
already established that "avail < literal_length", which usually means that
either the available space is small, or the literal is big. Both will disqualify
us from taking the fast path, and thus we take the hit from the precondition
checking without gaining much from having a fast path. Thus, simply don't try
the fast path in this situation -- we're already on a slow path anyway
(one where we need to refill more data from the reader).
I'm a bit surprised at how much this gained; it could be that this path is
more common than I thought, or that the simpler structure somehow makes the
compiler happier. I haven't looked at the assembler, but it's a win across
the board on both Core 2, Core i7 and Opteron, at least for the cases we
typically care about. The gains seem to be the largest on Core i7, though.
Results from my Core i7 workstation:
Benchmark Time(ns) CPU(ns) Iterations
---------------------------------------------------
BM_UFlat/0 73337 73091 190996 1.3GB/s html [ +1.7%]
BM_UFlat/1 696379 693501 20173 965.5MB/s urls [ +2.7%]
BM_UFlat/2 9765 9734 1472135 12.1GB/s jpg [ +0.7%]
BM_UFlat/3 29720 29621 472973 3.0GB/s pdf [ +1.8%]
BM_UFlat/4 294636 293834 47782 1.3GB/s html4 [ +2.3%]
BM_UFlat/5 28399 28320 494700 828.5MB/s cp [ +3.5%]
BM_UFlat/6 12795 12760 1000000 833.3MB/s c [ +1.2%]
BM_UFlat/7 3984 3973 3526448 893.2MB/s lsp [ +5.7%]
BM_UFlat/8 991996 989322 14141 992.6MB/s xls [ +3.3%]
BM_UFlat/9 228620 227835 61404 636.6MB/s txt1 [ +4.0%]
BM_UFlat/10 197114 196494 72165 607.5MB/s txt2 [ +3.5%]
BM_UFlat/11 605240 603437 23217 674.4MB/s txt3 [ +3.7%]
BM_UFlat/12 804157 802016 17456 573.0MB/s txt4 [ +3.9%]
BM_UFlat/13 347860 346998 40346 1.4GB/s bin [ +1.2%]
BM_UFlat/14 44684 44559 315315 818.4MB/s sum [ +2.3%]
BM_UFlat/15 5120 5106 2739726 789.4MB/s man [ +3.3%]
BM_UFlat/16 76591 76355 183486 1.4GB/s pb [ +2.8%]
BM_UFlat/17 238564 237828 58824 739.1MB/s gaviota [ +1.6%]
BM_UValidate/0 42194 42060 333333 2.3GB/s html [ -0.1%]
BM_UValidate/1 433182 432005 32407 1.5GB/s urls [ -0.1%]
BM_UValidate/2 197 196 71428571 603.3GB/s jpg [ +0.5%]
BM_UValidate/3 14494 14462 972222 6.1GB/s pdf [ +0.5%]
BM_UValidate/4 168444 167836 83832 2.3GB/s html4 [ +0.1%]
R=jeff
Revision created by MOE tool push_codebase.
git-svn-id: https://snappy.googlecode.com/svn/trunk@42 03e5f5b5-db94-4691-08a0-1a8bf15f6143
Looking up into and decoding the values from char_table has long shown up as a
hotspot in the decompressor. While it turns out that it's hard to make a more
efficient decoder for the copy ops, the literals are simple enough that we can
decode them without needing a table lookup. (This means that 1/4 of the table
is now unused, although that in itself doesn't buy us anything.)
The gains are small, but definitely present; some tests win as much as 10%,
but 1-4% is more typical. These results are from Core i7, in 64-bit mode;
Core 2 and Opteron show similar results. (I've run with more iterations
than unusual to make sure the smaller gains don't drown entirely in noise.)
Benchmark Time(ns) CPU(ns) Iterations
---------------------------------------------------
BM_UFlat/0 74665 74428 182055 1.3GB/s html [ +3.1%]
BM_UFlat/1 714106 711997 19663 940.4MB/s urls [ +4.4%]
BM_UFlat/2 9820 9789 1427115 12.1GB/s jpg [ -1.2%]
BM_UFlat/3 30461 30380 465116 2.9GB/s pdf [ +0.8%]
BM_UFlat/4 301445 300568 46512 1.3GB/s html4 [ +2.2%]
BM_UFlat/5 29338 29263 479452 801.8MB/s cp [ +1.6%]
BM_UFlat/6 13004 12970 1000000 819.9MB/s c [ +2.1%]
BM_UFlat/7 4180 4168 3349282 851.4MB/s lsp [ +1.3%]
BM_UFlat/8 1026149 1024000 10000 959.0MB/s xls [+10.7%]
BM_UFlat/9 237441 236830 59072 612.4MB/s txt1 [ +0.3%]
BM_UFlat/10 203966 203298 69307 587.2MB/s txt2 [ +0.8%]
BM_UFlat/11 627230 625000 22400 651.2MB/s txt3 [ +0.7%]
BM_UFlat/12 836188 833979 16787 551.0MB/s txt4 [ +1.3%]
BM_UFlat/13 351904 350750 39886 1.4GB/s bin [ +3.8%]
BM_UFlat/14 45685 45562 308370 800.4MB/s sum [ +5.9%]
BM_UFlat/15 5286 5270 2656546 764.9MB/s man [ +1.5%]
BM_UFlat/16 78774 78544 178117 1.4GB/s pb [ +4.3%]
BM_UFlat/17 242270 241345 58091 728.3MB/s gaviota [ +1.2%]
BM_UValidate/0 42149 42000 333333 2.3GB/s html [ -3.0%]
BM_UValidate/1 432741 431303 32483 1.5GB/s urls [ +7.8%]
BM_UValidate/2 198 197 71428571 600.7GB/s jpg [+16.8%]
BM_UValidate/3 14560 14521 965517 6.1GB/s pdf [ -4.1%]
BM_UValidate/4 169065 168671 83832 2.3GB/s html4 [ -2.9%]
R=jeff
Revision created by MOE tool push_codebase.
git-svn-id: https://snappy.googlecode.com/svn/trunk@41 03e5f5b5-db94-4691-08a0-1a8bf15f6143
state of ip_ after decompression (or attempted decompresion) is
completely irrelevant, so we don't need the trailer.
Performance is, as expected, mostly flat -- there's a curious ~3-5%
loss in the "lsp" test, but that test case is so short it is hard to say
anything definitive about why (most likely, it's some sort of
unrelated effect).
R=jeff
git-svn-id: https://snappy.googlecode.com/svn/trunk@39 03e5f5b5-db94-4691-08a0-1a8bf15f6143
It is seemingly hard for the compiler to understand that ip_, the current input
pointer into the compressed data stream, can not alias on anything else, and
thus using it directly will incur memory traffic as it cannot be kept in a
register. The code already knew about this and cached it into a local
variable, but since Step() only decoded one tag, it had to move ip_ back into
place between every tag. This seems to have cost us a significant amount of
performance, so changing Step() into a function that decodes as much as it can
before it saves ip_ back and returns. (Note that Step() was already inlined,
so it is not the manual inlining that buys the performance here.)
The wins are about 3-6% for Core 2, 6-13% on Core i7 and 5-12% on Opteron
(for plain array-to-array decompression, in 64-bit opt mode).
There is a tiny difference in the behavior here; if an invalid literal is
encountered (ie., the writer refuses the Append() operation), ip_ will now
point to the byte past the tag byte, instead of where the literal was
originally thought to end. However, we don't use ip_ for anything after
DecompressAllTags() has returned, so this should not change external behavior
in any way.
Microbenchmark results for Core i7, 64-bit (Opteron results are similar):
Benchmark Time(ns) CPU(ns) Iterations
---------------------------------------------------
BM_UFlat/0 79134 79110 8835 1.2GB/s html [ +6.2%]
BM_UFlat/1 786126 786096 891 851.8MB/s urls [+10.0%]
BM_UFlat/2 9948 9948 69125 11.9GB/s jpg [ -1.3%]
BM_UFlat/3 31999 31998 21898 2.7GB/s pdf [ +6.5%]
BM_UFlat/4 318909 318829 2204 1.2GB/s html4 [ +6.5%]
BM_UFlat/5 31384 31390 22363 747.5MB/s cp [ +9.2%]
BM_UFlat/6 14037 14034 49858 757.7MB/s c [+10.6%]
BM_UFlat/7 4612 4612 151395 769.5MB/s lsp [ +9.5%]
BM_UFlat/8 1203174 1203007 582 816.3MB/s xls [+19.3%]
BM_UFlat/9 253869 253955 2757 571.1MB/s txt1 [+11.4%]
BM_UFlat/10 219292 219290 3194 544.4MB/s txt2 [+12.1%]
BM_UFlat/11 672135 672131 1000 605.5MB/s txt3 [+11.2%]
BM_UFlat/12 902512 902492 776 509.2MB/s txt4 [+12.5%]
BM_UFlat/13 372110 371998 1881 1.3GB/s bin [ +5.8%]
BM_UFlat/14 50407 50407 10000 723.5MB/s sum [+13.5%]
BM_UFlat/15 5699 5701 100000 707.2MB/s man [+12.4%]
BM_UFlat/16 83448 83424 8383 1.3GB/s pb [ +5.7%]
BM_UFlat/17 256958 256963 2723 684.1MB/s gaviota [ +7.9%]
BM_UValidate/0 42795 42796 16351 2.2GB/s html [+25.8%]
BM_UValidate/1 490672 490622 1427 1.3GB/s urls [+22.7%]
BM_UValidate/2 237 237 2950297 499.0GB/s jpg [+24.9%]
BM_UValidate/3 14610 14611 47901 6.0GB/s pdf [+26.8%]
BM_UValidate/4 171973 171990 4071 2.2GB/s html4 [+25.7%]
git-svn-id: https://snappy.googlecode.com/svn/trunk@38 03e5f5b5-db94-4691-08a0-1a8bf15f6143
Replace the Apache 2.0 license header by the BSD-type license header;
somehow a lot of the files were missed in the last round.
R=dannyb,csilvers
DELTA=147 (74 added, 2 deleted, 71 changed)
Change on 2011-03-25 19:25:07-07:00 by sesse
Unbreak the build; the relicensing removed a bit too much (only comments
were intended, but I also accidentially removed some of the top lines of
the actual source).
Revision created by MOE tool push_codebase.
MOE_MIGRATION=1072
git-svn-id: https://snappy.googlecode.com/svn/trunk@21 03e5f5b5-db94-4691-08a0-1a8bf15f6143
