7/29/2016

Scatter Plots

I tweaked my scatter plot generation per Fabian. They're still log-log but now labeled with the non-log axis values so it's easier to read off the actual ratio and speed without doing a pow2 in your head.

Silesia :

Game Test Set :

Seven, total :

Seven, all files :

Slow slow compressors

lzma is really too slow to decode.

total                : Kraken     : 2.914 to 1 : 1053.961 MB/s
total                : lzma       : 3.186 to 1 : 52.660 MB/s

(Win64 Core i7-3770 3.4 GHz)

Kraken is around 20X faster than lzma, but lzma compresses better (about 9%). That's already a tip that something is horribly wrong; you have a 2000% speed difference and a 9% size difference.

If we look at the total time to load compressed from disk + decompress, we can make these speedup factor curves :

At very low disk speeds, the higher compression of lzma provides a speedup over Kraken. But how slow does the disk have to be? You can see the intersection of the curves is between 0 and 1 on the log scale, that's 1-2 MB/s !!

For any disk faster than 2 MB/s , load+decomp is *way* faster with Kraken. At a disk speed of 16 MB/s or so (log scale 4) the full load+decomp for Kraken is around 2X faster than with lzma. And that's still a very slow disk (around optical speed).

Now, this is a speedup factor for load *then* decomp. If you are fully overlapping overlapping IO with decompression, then some of the decode time is hidden.

*But* that also assumes that you have a whole core to give to decompression. And it assumes you have no other CPU to work to do after loading.

The idea that you can hide decompressor time in IO time only works if you have enough independent loads so that there's lots to overlap (because if you don't, then the first IO and last decompress will never overlap anything), and it assumes you have no other CPU work to do.

In theory I absolutely love the idea that you just load pre-baked data which is all ready to go, and you just point at it, so there's no CPU work in loading other than decompression, but in practice that is almost never the case. eg. for loading compressed web pages, there's tons of CPU work that needs to be ton to parse the HTML or JS or whatever, so the idea that you can hide the decompressor time in the network latency is a lie - the decompressor time adds on to the later processing time and adds directly onto total load latency.

The other factor that people often ignore is the fact that loading these days is heterogeneous.

What you actually encounter is something like this :


Download from internet ~ 1 MB/s
Load from optimal disc ~ 20 MB/s
Load from slow HDD ~ 80 MB/s
Fast SSD ~ 500 MB/s
NVMe drive on PCIe ~ 1-2 GB/s
Load from cache in RAM ~ 8 GB/s

We have very heterogeneous loading - even for a single file loaded by the same application.

The first time you load it, maybe you download from the internet, and in that case a slow decompressor like lzma might be best. But the next time you load it's from the cache in RAM. And the time after that it's from HDD. In those cases, using lzma is a disaster (in the sense that the loading is now nearly instant, but you spend seconds decoding; or in the sense that just loading uncompressed would have been way faster).

One issue that I think is not considered is that making the right choice in the slow-disk zone is not that big of a deal. On a 1 MB/s disk, the difference in "speedup" between lzma and Kraken is maybe 2% in favor of lzma. But on a 100 MB/s it's something like 400% in favor of Kraken.

Now in theory maybe it would be nice to have different compressors for download & disk storage; like you use something like lzma for downloadable, and then decode and re-encode to ZStd for HDD loading. In practice nobody does that and the advantage over just using ZStd all the time is very marginal.

Also in theory it would be nice if the OS cache would cache the decompressed data rather than caching the compressed data.


TODO : time lzma on PS4. Usually PS4 is 2-4X slower than my PC, so that puts lzma somewhere in the 10-25 mb/s range, which is very borderline for keeping up with the optical disc.


DVD 16x is ~ 20 MB/s (max)
PS4 Blu-Ray is 6x ~ 27 MB/s (max)

PS4 transparently caches Blu-Ray to HDD

Of course because of the transparent caching to HDD, if you actually keep files in lzma on the disc, and they are cached to HDD, loading them from HDD is a huge mismatch and makes lzma the bottleneck.

That is, in practice on the PS4 when you load files from disc, they are sometimes actually coming from the HDD transparent cache, so you sometimes get 20 MB/s speeds, and sometimes 100 MB/s.


Now of course we'd love to have a higher-ratio compressor than Kraken which isn't so slow. Right now, we just don't have it. We have Kraken at 1000 MB/s , LZNA at 120 MB/s , lzma at 50 MB/s - it's too big of a step down in speed, even for LZNA.

In order for the size gain of lzma/LZNA to be worth it, it needs to run a *lot* faster, more like 400 mb/s. There needs to be a new algorithmic step in the high compress domain to get there.

At the moment the only reason to use the slower decoders than Kraken is if you simply must get smaller files and damn the speed; like if you have a downloadable app size hard limit and just have to fit in 100 MB, or if you are running out of room on an optical disc or whatever.

7/27/2016

Improving the compression of block-compressed textures

ADD: the best solution for this now is our new product Oodle Texture

I'm leaving this post for historical reference, but it is now obsolete and you should look at the Oodle Texture results instead.


I'm often asked how to get more compression on block-compressed textures. (and was asked again today, hence writing this)

If you're working from existing BCn data (not RGB originals), there's not a lot you can do. One small thing you can do is to de-interleave the end points and indexes.

In BC1/DXT1 for example each block is 4 bytes of end points then 4 bytes of index data. You can take each alternating 4 bytes out and instead put all the end points together, then all the index data together. Sometimes this improves compression, sometimes it doesn't, it depends on the compressor and the data. When it does help it's in the 5-10% range.

If you're using mips, then you can convert the BCn endpoint colors into deltas from the parent mip. (that will only be a good idea if your BCn encoder is one that is *not* aggressive about finding endpoints outside of the original image colors)

If you have original RGB data, you can make new BCn data that's designed to be more compressible. This opens up a whole new world of powerful possibilities : R/D decisions in the BCn encoding.

There are some obvious basic ideas, like - instead of aggressive end point searches, only choose end points that occur or are near the colors in the block; try to reuse previous index dwords rather than make new ones; try to use completely flat color blocks with the whole index dword = 0, etc.

See for example Jon Olick's articles on doing this .

But unless you really need to do it manually for some reason you should probably just use Rich Geldreich's cool tool crunch .

Crunch can make its own "crn" compressed texture format, which you would need to load with the crn-lib. crn-lib would decode the crn file back to BC1 at load time. That may be an awesome thing to do, I really can't comment because I haven't looked into it in detail.

Let's assume for the moment that you don't want to use the "crn" custom compressed format. You just want DDS or raw BCn data that you will run through your normal compression pipeline (Oodle or whatever). Crunch can also make BCn data that's more compressible by reducing its complexity, choosing encodings that are lower quality but less complex.

You tell it to make BCn and output DDS and you can specify a "quality". Then when you run it through your back-end compressor you get smaller files :


lena 512x512 RGB (absolutely terrible as a representative game texture)

DXT1 DDS quality 255

 rmse : 7.6352 psnr : 30.5086

Oodle LZNA   :   131,200 ->   102,888 =  6.274 bpb =  1.275 to 1
Oodle Kraken :   131,200 ->   107,960 =  6.583 bpb =  1.215 to 1

DXT1 DDS quality 200

 rmse : 8.2322 psnr : 29.8547

Oodle LZNA   :   131,200 ->    80,264 =  4.894 bpb =  1.635 to 1
Oodle Kraken :   131,200 ->    85,268 =  5.199 bpb =  1.539 to 1

CRN quality 255

 rmse : 8.2699 psnr : 29.8150
   crunched.crn                74,294

DXT1 DDS quality 128

 rmse : 9.0698 psnr : 29.0131

Oodle LZNA   :   131,200 ->    62,960 =  3.839 bpb =  2.084 to 1
Oodle Kraken :   131,200 ->    66,628 =  4.063 bpb =  1.969 to 1

CRN quality 160

rmse : 9.0277 psnr : 29.0535

crunched.crn                53,574

DXT1 DDS quality 80

 rmse : 10.2521 psnr : 27.9488

Oodle LZNA   :   131,200 ->    50,983 =  3.109 bpb =  2.573 to 1
Oodle Kraken :   131,200 ->    54,096 =  3.299 bpb =  2.425 to 1

CRN quality 100

 rmse : 10.1770 psnr : 28.0126

 crunched.crn               41,533

So going from rmse 7.64 to 10.26 we reduced the Oodle-compressed DDS files to about half their size! Pretty cool.

The CRN format files are even smaller at equal error. (unfortunately the -quality setting is not a direct comparison, you have to hunt around to find qualities that give equal rmse to do a comparison).

For my reference :


@echo test.bat [file] [quality]
@echo quality in 1-255 optional 128 default
set file=%1
if "%file%"=="" end.bat 
set q=%2
if "%q%"=="" set q=128
crunch_x64.exe -DXT1 -fileformat dds -file %file% -maxmips 1 -quality %q% -out crunched.dds
@REM -forceprimaryencoding ??

@REM verified same :
@REM radbitmaptest64 copy crunched.dds crunched_dds_un.bmp
crunch_x64.exe -R8G8B8 -file crunched.dds -out crunched_dds_un.bmp -fileformat bmp

crunch_x64.exe -DXT1 -file %file% -maxmips 1 -quality %q% -out crunched.crn

crunch_x64.exe -R8G8B8 -file crunched.crn -out crunched_crn_un.bmp -fileformat bmp

call bmp mse %file% crunched_dds_un.bmp
@echo --------------------------

call ooz crunched.dds -zc7 -zl8
call ooz crunched.dds -zc8 -zl8

call bmp mse %file% crunched_crn_un.bmp
@echo --------------------------

call d crunched.crn

7/26/2016

Oodle 2.3.0 Thread-Phased Performance

Thread-Phased (two threads) decoding can be used with Mermaid just as it was with Kraken.

It's not as big a benefit on Mermaid as it is for Kraken. (Kraken gets around 1.5X, Mermaid much less).

There is a little neat thing about thread-phased Mermaid though, and that's the ability to run Mermaid+ at almost the exact same speed as Mermaid.

Mermaid+ is a hybrid option that's hidden in Oodle 2.3.0 and will be exposed in the next release. It gets compression between Mermaid & Kraken, with a small speed hit vs Mermaid.


Seven :

ooSelkie    :  2.19:1 ,    3.0 enc MB/s , 3668.0 dec MB/s
ootp2Mermaid:  2.46:1 ,    2.3 enc MB/s , 3046.1 dec MB/s
lz4hc       :  2.00:1 ,   12.8 enc MB/s , 2532.4 dec MB/s
ooMermaid   :  2.46:1 ,    2.3 enc MB/s , 2364.4 dec MB/s
ootp2Kraken :  2.91:1 ,    2.6 enc MB/s , 1660.9 dec MB/s
ooKraken    :  2.91:1 ,    2.6 enc MB/s , 1049.6 dec MB/s
zlib9       :  2.33:1 ,    7.9 enc MB/s ,  315.1 dec MB/s

ootp2Merm+  :  2.64:1 ,    2.3 enc MB/s , 3042.4 dec MB/s
ooMermaid+  :  2.64:1 ,    2.3 enc MB/s , 2044.5 dec MB/s

Silesia :

ooSelkie    :  3.05:1 ,    1.3 enc MB/s , 2878.4 dec MB/s
ootp2Mermaid:  3.57:1 ,    1.1 enc MB/s , 2600.8 dec MB/s
lz4hc       :  2.72:1 ,   13.6 enc MB/s , 2273.5 dec MB/s
ooMermaid   :  3.57:1 ,    1.1 enc MB/s , 1994.2 dec MB/s
ootp2Kraken :  4.08:1 ,    1.2 enc MB/s , 1434.4 dec MB/s
ooKraken    :  4.08:1 ,    1.2 enc MB/s , 1000.5 dec MB/s
zlib9       :  3.13:1 ,    8.3 enc MB/s ,  358.4 dec MB/s

ootp2Merm+  :  3.58:1 ,    1.1 enc MB/s , 2583.9 dec MB/s
ooMermaid+  :  3.58:1 ,    1.1 enc MB/s , 1986.8 dec MB/s

Game Test Set :

ooSelkie    :  2.03:1 ,    3.3 enc MB/s , 4548.0 dec MB/s
lz4hc       :  1.78:1 ,   14.0 enc MB/s , 3171.1 dec MB/s
ootp2Mermaid:  2.28:1 ,    2.7 enc MB/s , 3099.4 dec MB/s
ooMermaid   :  2.28:1 ,    2.7 enc MB/s , 2622.8 dec MB/s
ootp2Kraken :  2.57:1 ,    2.9 enc MB/s , 1812.7 dec MB/s
ooKraken    :  2.57:1 ,    2.9 enc MB/s , 1335.9 dec MB/s
zlib9       :  1.99:1 ,    8.3 enc MB/s ,  337.2 dec MB/s

ootp2Merm+  :  2.35:1 ,    2.7 enc MB/s , 3034.9 dec MB/s
ooMermaid+  :  2.35:1 ,    2.7 enc MB/s , 2409.0 dec MB/s

Pulling out just the relevant numbers on Seven, you can see Mermaid+ is between Mermaid and Kraken, but thread-phased it runs at full Mermaid speed :

Seven :

ooMermaid   :  2.46:1 ,    2.3 enc MB/s , 2364.4 dec MB/s
ooMermaid+  :  2.64:1 ,    2.3 enc MB/s , 2044.5 dec MB/s
ooKraken    :  2.91:1 ,    2.6 enc MB/s , 1049.6 dec MB/s

ootp2Mermaid:  2.46:1 ,    2.3 enc MB/s , 3046.1 dec MB/s
ootp2Merm+  :  2.64:1 ,    2.3 enc MB/s , 3042.4 dec MB/s
ootp2Kraken :  2.91:1 ,    2.6 enc MB/s , 1660.9 dec MB/s

7/25/2016

Introducing Oodle Mermaid and Selkie

I'm pleased to announce the release of two new compressors in Oodle 2.3.0 : Mermaid and Selkie.

Mermaid and Selkie are the super-fast-to-decode distant relatives of Kraken. They use some of the same ideas and technology as Kraken, but are independent compressors targetted at even higher speed and lower compression. Mermaid & Selkie make huge strides in what's possible in compression in the high-speed domain, the same way that Kraken did in the high-compression domain.

Mermaid is about twice as fast as Kraken, but with compression around Zlib levels.

Selkie is one of the fastest decompressors in the world, and also gets much more compression than other very-high-speed compressors.

( Oodle is my data compression library that we sell at RAD Game Tools , read more about it there )

Kraken, Mermaid, and Selkie all use an architecture that makes space-speed decisions in the encoder to give the best tradeoff of compressed size vs decoding speed. The three compressors have different performance targets and make decisions suited for each one's usage domain (Kraken favors more compression and will give up some speed, Selkie strongly favors speed, Mermaid is in between).

For detailed information about the new Mermaid and Selkie I've written a series of posts :

cbloom rants Introducing Oodle Mermaid and Selkie
cbloom rants Oodle 2.3.0 All Test Sets
cbloom rants Oodle 2.3.0 ARM Report
cbloom rants Oodle Mermaid and Selkie on PS4
cbloom rants Oodle Mermaid
cbloom rants Oodle Selkie
RAD Game Tools - Oodle Network and Data Compression

Here are some representative numbers on the Silesia test set : (sum of time and size on individual files)


Oodle 2.3.0 Silesia -z6

Kraken     : 4.082 to 1 : 999.389 MB/s
Mermaid    : 3.571 to 1 : 2022.038 MB/s
Selkie     : 3.053 to 1 : 2929.770 MB/s

zstdmax    : 4.013 to 1 : 468.497 MB/s
zlib9      : 3.128 to 1 : 358.681 MB/s
lz4hc      : 2.723 to 1 : 2267.021 MB/s

on Win64 (Core i7-3770 3.4 GHz)

On Silesia, Mermaid is 5.65X faster to decode than zlib, and gets 14% more compression. Selkie is 1.3X faster to decode than LZ4 and gets 12% more compression.

Charts on Silesia total : (charts show time and size - lower is better!)

And the speedup chart on Silesia, which demonstrates the space-speed efficiency of a compressor in different usage domains.

Kraken was a huge step in the Pareto frontier that pushed the achievable speedup factor way up beyond what other compressers were doing. There's a pre-Kraken curve where we thought the best possible tradeoff existed, that most other compressors in the world roughly lie on (or under). Kraken set a new frontier way up on its own with nobody to join it; Mermaid & Selkie are the partners on that new curve that have their peaks at higher speeds than Kraken.

You can also see this big jump of the new family very easily in scatter plots, which we'll see in later posts .

Oodle Mermaid and Selkie on PS4

The PS4 is a lovely platform to benchmark on because it's standard. It's also very easy to run tests on. Performance of these compressors on the Xbox One is extremely similar, small variatation due to clock rate (1.75 vs 1.6 GHz) and compiler (MSVC vs clang/llvm).

Everything is slow on the PS4 in absolute terms (it's a slow chip and difficult to optimize for). The Oodle compressors do very well, even better in relative terms on PS4 than on typical PC's.

Kraken is usually around ~2X faster than ZStd on PC's, but is 3X faster on PS4. Mermaid is usually just slightly slower than LZ4 on PC's, but is solidly faster than LZ4 on PS4.


lzt99 :

Kraken     : 2.477 to 1 : 390.582 MB/s
Mermaid    : 2.279 to 1 : 749.896 MB/s
Selkie     : 1.937 to 1 : 1159.064 MB/s

zstd       : 2.374 to 1 : 133.498 MB/s
miniz      : 1.883 to 1 : 85.654 MB/s
lz4hc-safe : 1.669 to 1 : 673.616 MB/s
LZSSE8     : 1.626 to 1 : 767.106 MB/s

Mermaid is faster than LZ4 on PS4 !! Wow! And the compression level is in a totally different domain than other super-fast decompressors like LZ4 or LZSSE.

lzt99 is a good case for Selkie & Mermaid. Selkie beats zlib compression ratio while being 75% faster than LZ4.

All compressors here are fuzz-safe, and run in safe mode if they have optional safe/unsafe modes.

Charts : (showing time and size - lower is better!)

lzt99 :

the raw data :


PS4 : Oodle 230 : (-z6)

inName : lzt:/lzt99

reference :

miniz      : 24,700,820 ->13,120,668 =  4.249 bpb =  1.883 to 1
miniz_decompress_time : 288.379 millis, 18.61 c/b, rate= 85.65 mb/s

zstd       : 24,700,820 ->10,403,228 =  3.369 bpb =  2.374 to 1
zstd_decompress_time : 185.028 millis, 11.94 c/b, rate= 133.50 mb/s

lz4hc      : 24,700,820 ->14,801,510 =  4.794 bpb =  1.669 to 1
LZ4_decompress_safe_time : 36.669 millis, 2.37 c/b, rate= 673.62 mb/s

LZSSE8     : 24,700,820 ->15,190,395 =  4.920 bpb =  1.626 to 1
decode_time      : 32.200 millis, 2.08 c/b, rate= 767.11 mb/s

Oodle :

Kraken     : 24,700,820 -> 9,970,882 =  3.229 bpb =  2.477 to 1
decode           : 63.241 millis, 4.08 c/b, rate= 390.58 mb/s

Mermaid    : 24,700,820 ->10,838,455 =  3.510 bpb =  2.279 to 1
decode           : 32.939 millis, 2.13 c/b, rate= 749.90 mb/s

Selkie : 24,700,820 ->12,752,506 =  4.130 bpb =  1.937 to 1
decode           : 21.311 millis, 1.38 c/b, rate= 1159.06 mb/s

BTW for reference, the previous best compressor in Mermaid's domain was LZNIB. Before these new compressors, LZNIB was quite unique in that it got good decode speeds, much faster than the LZ-Huffs of the time (eg. 3X faster than ZStd) but with compression usually better than ZLib. Well, LZNIB is still quite good compared to other competition, but it's just clobbered by the new Oceanic Bestiary compressors. The new compressor in this domain is Mermaid and it creams LZNIB for both size and speed :


LZNIB -z6  : 24,700,820 ->12,015,591 =  3.892 bpb =  2.056 to 1 
decode           : 58.710 millis, 3.79 c/b, rate= 420.73 mb/s

Mermaid    : 24,700,820 ->10,838,455 =  3.510 bpb =  2.279 to 1
decode           : 32.939 millis, 2.13 c/b, rate= 749.90 mb/s

See the index of this series of posts for more information : Introducing Oodle Mermaid and Selkie .
For more about Oodle visit RAD Game Tools

Oodle Mermaid

Mermaid is a new compressor with a unique balance of space and speed. Mermaid is very close to LZ4 decode speeds, while usually beating Zlib compression ratios.

There's really nothing even close. It's way beyond what was previously thought possible.

Mermaid supports unbounded distance match references. This is part of how it gets such high compression. It does so in a new way which reduces the speed penalty normally incurred by large-window LZ's.

Mermaid almost always compresses better than ZLib. The only exception is on small files, less than 32k or so. The whole Oceanic Bestiary family is best suited to files over 64k. They work fine on smaller files, but they lose their huge advantage. It's always best to combine small files into larger units for compression, particularly so with these compressors.

There's not really any single compressor to compare Mermaid to. What we can do is compare vs. Zlib's compression ratio and LZ4's speed. A kind of mythological hybrid like a Chimera, the head of a Zlib and the legs of an LZ4.

Tests on Win64 (Core i7-3770 3.4 GHz) :

Silesia :

On Silesia, Mermaid is just slightly slower than LZ4 but compresses much more than Zlib !!

PD3D :

On PD3D, Mermaid gets almost exactly the compression level of ZLib but the decode speed of LZ4. Magic! It turns out you *can* have your cake and eat it too.

Game Test Set :

lzt99 :

Mermaid really compresses well on lzt99 ; not only does it kill Zlib, it gets close to high compression LZ-Huffs like RAR. (RAR gets 10826108 , Mermaid 10838455 bytes).

Seven :

Because of the space-speed optimizing nature of Mermaid, it will make decisions to be slower than LZ4 when it can find big compression gains. For example if you look at the individual files of the "Seven" test set below - Mermaid is typically right around the same speed as LZ4 or even faster (baby7,dds7,exe7,game7,wad7 - all same speed or faster than LZ4). On a few files it decides to take an encoding slower to decode than LZ4 - model7,enwik7, and records7. The biggest differences are enwik7 and records7, but if you look at the compression ratios - those are all the files where it found huge size differences over LZ4. It has an internal exchange rate for time vs. bytes that it must meet in order to take that encoding, trying to optimize for its space-speed target usage.

Seven files :


Silesia              : Mermaid    : 3.571 to 1 : 2022.038 MB/s
Silesia              : lz4hc      : 2.723 to 1 : 2267.021 MB/s
Silesia              : zlib9      : 3.128 to 1 : 358.681 MB/s

GameTestSet          : Mermaid    : 2.284 to 1 : 2718.095 MB/s
GameTestSet          : lz4hc      : 1.776 to 1 : 3226.887 MB/s
GameTestSet          : zlib9      : 1.992 to 1 : 337.986 MB/s

lzt99                : Mermaid    : 2.279 to 1 : 2283.278 MB/s
lzt99                : lz4hc      : 1.669 to 1 : 2605.366 MB/s
lzt99                : zlib9      : 1.883 to 1 : 309.304 MB/s

PD3D                 : Mermaid    : 2.875 to 1 : 2308.830 MB/s
PD3D                 : lz4hc      : 2.238 to 1 : 2369.666 MB/s
PD3D                 : zlib9      : 2.886 to 1 : 382.349 MB/s

Seven                : Mermaid    : 2.462 to 1 : 2374.212 MB/s
Seven                : lz4hc      : 2.000 to 1 : 2521.296 MB/s
Seven                : zlib9      : 2.329 to 1 : 315.370 MB/s

See the index of this series of posts for more information : Introducing Oodle Mermaid and Selkie .
For more about Oodle visit RAD Game Tools

Oodle Selkie

Selkie is the faster cousin of Mermaid, distant relative of Kraken.

Selkie is all about decode speed, it aims to be the fastest mainstream decompressor in the world, and still gets more compression than anything in the high-speed domain.

Selkie does not currently have a super fast encoder. It's got good optimal parse encoders that produce carefully tuned encoded file which offer excellent space-speed tradeoff.

The closest compressors to Selkie are the fast byte-wise small-window coders like LZ4 and LZSSE (and Oodle's LZB16). These are all just obsolete now (in terms of ratio vs decode speed), Selkie gets a lot more compression (sometimes close to Zlib compression levels!) and is also much faster.

Selkie will not compress tiny buffers, or files that only compress a little bit. For example if you give Selkie something like an mp3, it might be able to compress it to 95% of its original size, saving a few bytes. Selkie will refuse to do that and just give you the original uncompressed file. If you wanted that compression, that means you wanted to save only a few bytes at a large time cost, which means you don't actually want a fast compressor like Selkie. You in fact wanted a compressor that was more willing to trade time for bytes, such as Mermaid or Kraken. Selkie will not abide logical inconsistency.

Selkie generally beats LZ4 compression even on small files (under 64k) but really gets ahead on files larger than 64k where the unbounded match distances can find big wins.

As usual, I'm not picking on LZ4 here because it's bad; I'm comparing to it because it's the best of the rest, and it's widely known. Both decompressors are run fuzz-safe.


Tests on Win64 (Core i7-3770 3.4 GHz) : 
(for reference, this machine runs memcpy at roughly 8 GB/s)
(total of time & size on each test set)

gametestset : ooSelkie    : 143,579,361 ->70,716,380 =  3.940 bpb =  2.030 to 1 
gametestset : decode      : 29.239 millis, 0.69 c/b, rate= 4910.61 mb/s

gametestset : lz4hc       : 143,579,361 ->80,835,018 =  4.504 bpb =  1.776 to 1 
gametestset : decode      : 44.495 millis, 1.05 c/b, rate= 3226.89 mb/s

pd3d : ooSelkie    : 31,941,800 ->13,428,298 =  3.363 bpb =  2.379 to 1 
pd3d : decode      : 8.381 millis, 0.89 c/b, rate= 3811.29 mb/s

pd3d : lz4hc       : 31,941,800 ->14,273,195 =  3.575 bpb =  2.238 to 1 
pd3d : decode      : 13.479 millis, 1.44 c/b, rate= 2369.67 mb/s

seven : ooSelkie    : 80,000,000 ->36,460,084 =  3.646 bpb =  2.194 to 1 
seven : decode      : 21.458 millis, 0.91 c/b, rate= 3728.26 mb/s

seven : lz4hc       : 80,000,000 ->39,990,656 =  3.999 bpb =  2.000 to 1 
seven : decode      : 31.730 millis, 1.35 c/b, rate= 2521.30 mb/s

silesia : ooSelkie    : 211,938,580 ->69,430,966 =  2.621 bpb =  3.053 to 1 
silesia : decode      : 72.340 millis, 1.16 c/b, rate= 2929.77 mb/s

silesia : lz4hc       : 211,938,580 ->77,841,566 =  2.938 bpb =  2.723 to 1 
silesia : decode      : 93.488 millis, 1.50 c/b, rate= 2267.02 mb/s

The edge that Selkie has over LZ4 is even greater on more difficult platforms like the PS4.

To get a better idea of the magic of Selkie it's useful to look at the other Oodle compressors that are similar to Selkie.

LZB16 is Oodle's LZ4 variant; it gets slightly more compression and slightly more decode speed, but they're roughly equal. It's included here for comparison to LZBLW.

Oodle's LZBLW is perhaps the most similar compressor to Selkie. It's like LZB16 (LZ4) but adds large-window matches. That ability to do long-distance matches hurts speed a tiny bit (2873 mb/s -> 2596 mb/s), but helps compression a lot.

Oodle's LZNIB is nibble-wise, with unbounded offsets and a rep match. It gets good compression, generally better than Zlib, with speed much higher than any LZ-Huff. LZNIB is in a pretty unique space speed tradeoff zone without much competition outside of Oodle.


lz4hc     : 24,700,820 ->14,801,510 =  4.794 bpb =  1.669 to 1 
decode    : 9.481 millis, 1.31 c/b, rate= 2605.37 mb/s

ooLZB16   : 24,700,820 ->14,754,643 =  4.779 bpb =  1.674 to 1
decode    : 8.597 millis, 1.18 c/b, rate= 2873.17 mb/s

ooLZNIB   : 24,700,820 ->12,014,368 =  3.891 bpb =  2.056 to 1
decode    : 17.420 millis, 2.40 c/b, rate= 1417.93 mb/s

ooLZBLW   : 24,700,820 ->13,349,800 =  4.324 bpb =  1.850 to 1
decode    : 9.512 millis, 1.31 c/b, rate= 2596.80 mb/s

ooSelkie  : 24,700,820 ->12,752,506 =  4.130 bpb =  1.937 to 1 
decode    : 6.410 millis, 0.88 c/b, rate= 3853.57 mb/s

LZNIB and LZBLW were both pretty cool before Selkie, but now they're just obsolete.

LZBLW gets a nice compression gain over LZB16, but Selkie gets even more, and is way faster!

LZNIB beats Selkie compression, but is way slower, around 3X slower, in fact it's slower than Mermaid (2283.28 mb/s and compresses to 10,838,455 = 3.510 bpb = 2.279 to 1).

You can see from the curves that Selkie just completely covers the curves of LZB16,LZBLW, and LZ4. When a curve is completely covered like that, it means that it was beaten for both space and speed, so there is no domain where that compressor is ever better. LZNIB just peeks out of the Selkie curve because it gets higher compression (albeit at lower speed), so there is a domain where it is the better choice - but in that domain Mermaid just completely dominates LZNIB, so it too is obsolete.

See the index of this series of posts for more information : Introducing Oodle Mermaid and Selkie .
For more about Oodle visit RAD Game Tools

Oodle 2.3.0 ARM Report

I prepared a detailed report of Oodle's performance on ARM mobile devices (Android and iOS).

The full report is here :

oodle_arm_report on cbloom.com

It's a thorough test on many devices and several corpora. See the full details there.

Cliff notes is : Oodle's great on ARM.

For example on the iPadAir2 64-bit , on Silesia :

We found that the iOS devices are generally very good and easy to program for. They're more like desktop Intel chips; they don't have any terrible performance cliffs. The Android ARM devices we tested on were rather more difficult. For one thing they have horrible thermal saturation problems that makes testing on them very difficult. They also have some odd performance quirks.

I'm sure we could get a lot more speed on ARM, but it's rather nasty to optimize for. For one thing the thermal problems mean that iterating and getting good numbers is a huge pain. It's hard to tell if a change helped or not. For another, there's a wide variety of devices and it's hard to tell which to optimize for, and they have different performance shortfalls. So there's definitely a lot left on the table here.

Mermaid & Selkie are quite special on ARM. Many of these devices have small caches (as small as 512k L2) and very slow main memory (slow wrst latency; they often have huge bandwidth, but latency is what I need). Mermaid & Selkie are able to use unbounded windows for LZ without suffering a huge speed hit, due to the unique way they are structured. Kraken doesn't have the same magic trick so it benefits from a limited window, as demonstrated in the report.

See the index of this series of posts for more information : Introducing Oodle Mermaid and Selkie .
For more about Oodle visit RAD Game Tools

Oodle 2.3.0 All Test Sets

Putting the total performance on various testsets together in one place. Tests on Win64 Core i7-3770 3.4 GHz as usual.

Showing speed & ratio here, higher is better.

As usual the total on a test set is total size of all individually compressed files, and total time.

I think the scatter plot most clearly shows the way Kraken, Mermaid & Selkie are just on a whole new Pareto Frontier than the older compressors. You can connect the dots of K-M-S performance for each test set and they form a very consistent space-speed tradeoff curve that's way above the previous best.

The raw numbers :

gametestset          : Kraken     : 2.566 to 1 : 1363.283 MB/s
gametestset          : Mermaid    : 2.284 to 1 : 2711.458 MB/s
gametestset          : Selkie     : 2.030 to 1 : 4870.413 MB/s
gametestset          : lz4hc      : 1.776 to 1 : 3223.279 MB/s
gametestset          : zlib9      : 1.992 to 1 : 338.063 MB/s
gametestset          : lzma       : 2.756 to 1 : 43.782 MB/s

pd3d                 : Kraken     : 3.647 to 1 : 1072.833 MB/s
pd3d                 : Mermaid    : 2.875 to 1 : 2299.860 MB/s
pd3d                 : Selkie     : 2.379 to 1 : 3784.850 MB/s
pd3d                 : lz4hc      : 2.238 to 1 : 2370.193 MB/s
pd3d                 : zlib9      : 2.886 to 1 : 382.226 MB/s
pd3d                 : lzma       : 4.044 to 1 : 63.878 MB/s

seven                : Kraken     : 2.914 to 1 : 1053.961 MB/s
seven                : Mermaid    : 2.462 to 1 : 2374.796 MB/s
seven                : Selkie     : 2.194 to 1 : 3717.074 MB/s
seven                : lz4hc      : 2.000 to 1 : 2522.824 MB/s
seven                : zlib9      : 2.329 to 1 : 315.344 MB/s
seven                : lzma       : 3.186 to 1 : 52.660 MB/s

silesia              : Kraken     : 4.082 to 1 : 1004.014 MB/s
silesia              : Mermaid    : 3.571 to 1 : 2002.079 MB/s
silesia              : Selkie     : 3.053 to 1 : 2889.536 MB/s
silesia              : lz4hc      : 2.723 to 1 : 2269.788 MB/s
silesia              : zlib9      : 3.128 to 1 : 358.593 MB/s
silesia              : lzma       : 4.369 to 1 : 78.655 MB/s

See the index of this series of posts for more information : Introducing Oodle Mermaid and Selkie .
For more about Oodle visit RAD Game Tools

7/24/2016

Oodle 2.3.0 : Kraken Improvement

Oodle 2.3.0 includes some pretty solid improvements to Kraken. The result is around a 10% gain in decode speed.

There were two major factors in the gains. One was just some more time optimizing some inner loops (including some new super-tight pathways from Fabian).

The other was more rigorous analysis of the space-speed tradeoff decisions inside Kraken. One of the fundamental things that makes Kraken work is the fact that it consider space-speed when making its internal decisions, but before 230 those decisions were made in a rather ad-hoc way. Making those decisions better means that even with the same decoder, the new encoder is able to create files that are the same size but decode faster.

The tradeoff point (technically, the lagrange lambda, or the exchange rate from time to bytes) that's used by Oodle to make space-speed decisions is exposed to the client in the OodleLZ_CompressOptions so you can adjust it to bias for compression or decode speed. Each compressor sets what I believe to be a reasonable default for its usage domain, so adjustments to this value should typically be small (you can't massively change behavior with it; Kraken won't start arithmetic coding things if you set the tradeoff really small, for example, there's a small window where the compressor works well and you can just bias sightly within that window).

Some dry numbers for reference :

On PS4 :


Oodle 230 Kraken -zl4 : 24,700,820 ->10,377,556 =  3.361 bpb =  2.380 to 1
decode only      : 65.547 millis, 4.23 c/b, rate= 376.84 mb/s

Oodle 230 Kraken -zl6 : 24,700,820 -> 9,970,882 =  3.229 bpb =  2.477 to 1
decode           : 63.453 millis, 4.09 c/b, rate= 389.28 mb/s

Oodle 230 Kraken -zl7 : 24,700,820 -> 9,734,771 =  3.153 bpb =  2.537 to 1
decode           : 67.915 millis, 4.38 c/b, rate= 363.70 mb/s

Oodle 220 Kraken -zl4 : 24,700,820 ->10,326,584 =  3.345 bpb =  2.392 to 1
decode only      : 0.073 seconds, 211.30 b/kc, rate= 336.76 mb/s

Oodle 220 Kraken -zl6 : 24,700,820 ->10,011,486 =  3.242 bpb =  2.467 to 1
decode           : 0.074 seconds, 208.83 b/kc, rate= 332.82 mb/s

Oodle 220 Kraken -zl7 : 24,700,820 -> 9,773,112 =  3.165 bpb =  2.527 to 1
decode           : 0.079 seconds, 196.70 b/kc, rate= 313.49 mb/s

On Win64 (Core i7-3770 3.4 GHz) :


Oodle 2.3.0 :

Silesia Kraken -z6

total   : 211,938,580 ->51,918,269 =  1.960 bpb =  4.082 to 1
decode           : 210.685 millis, 3.38 c/b, rate= 1005.95 mb/s
Weissman 1-256 : [8.575]

mozilla : 51,220,480 ->14,410,181 =  2.251 bpb =  3.554 to 1
decode only      : 51.280 millis, 3.41 c/b, rate= 998.83 mb/s

lzt99 : 24,700,820 -> 9,970,882 =  3.229 bpb =  2.477 to 1
decode only      : 20.943 millis, 2.89 c/b, rate= 1179.44 mb/s

win81 : 104,857,600 ->38,222,311 =  2.916 bpb =  2.743 to 1
decode only      : 108.344 millis, 3.52 c/b, rate= 967.82 mb/s

Oodle 2.2.0 :

Silesia Kraken -z6

total : 211,938,580 ->51,857,427 =  1.957 bpb =  4.087 to 1
decode   : 0.232 seconds, 268.43 b/kc, rate= 913.46 M/s
Weissman 1-256 : [8.431]

"silesia_mozilla"

Kraken 230  :  3.55:1 ,   998.8 dec mb/s
Kraken 220  :  3.60:1 ,   896.5 dec mb/s
Kraken 215  :  3.51:1 ,   928.0 dec mb/s

"lzt99"

Kraken 230  :  2.48:1 ,   998.8 dec mb/s
Kraken 220  :  2.53:1 ,   912.0 dec mb/s
Kraken 215  :  2.46:1 ,   957.1 dec mb/s

"win81"

Kraken 230  :  2.74:1 ,   967.8 dec mb/s
Kraken 220  :  2.77:1 ,   818.0 dec mb/s
Kraken 215  :  2.70:1 ,   877.0 dec mb/s

NOTE : Oodle 2.3.0 Kraken data cannot be read by Oodle 2.2.0 or earlier. Oodle 230 can load all old Oodle data (new versions of Oodle can always load all data created by older versions). If you need to make data that be loaded with an older version using Oodle 230, then you can set the minimum decoder version to something lower (by default it's the current version). Contact Oodle support for details.

Some of the biggest gains were found on ARM, which I'll post about more in the future.

old rants