03-31-13 - Market Equilibrium

I'm sure there's some standard economic theory of all this but hey let's reinvent the wheel without any background.

There's a fundamental principal of any healthy (*) market that the reward for some labor is equal across all fields - proportional only to standard factors like the risk factor, the scarcity of labor, the capital required for entry, etc. (* = more on "healthy" later). The point is that those factors have *nothing* to do with the details of the field.

The basic factor at play is that if some field changes and suddenly becomes much more profitable, then people will flood into that field, and the risk-equal-capital-return will keep going down until it becomes equal to other fields. Water flows downhill, you know.

When people like Alan Greenspan try to tell you that oh this new field is completely unlike anything we've seen in the past because of blah blah - it doesn't matter, they may have lots of great points that seem reasonable in isolation, but the equilibrium still applies. The pay of a computer programmer is set by the pay of a farmer, because if the difference were out of whack, the farmer would quit farming and start programming; the pay of programmers will go down and the wages of farmers will go up, then the price of lettuce will go up, and in the end a programmer won't be able to buy any more lettuce than anyone else in a similar job. ("similar" only in terms of risk, ease of entry, rarity of talent, etc.)

We went through a drive-through car wash yesterday and Tasha idly wondered how much the car wash operator makes from an operation like that. Well, I bet it's about the same as a quick-lube place makes, and that's about the same as a dry cleaner, and it's about the same as a pizza place (which has less capital outlay but more risk), because if one of them was much more profitable, there would be more competition until equilibrium was reached.

Specifically I've been thinking about this because of the current indie game boom on the PC, which seems to be a bit of a magic gold rush at the moment. That almost inevitably has to die out, it's just a question of when. (so hurry up and get your game out before it does!).

But of course that leads us into the issue of broken markets, since all current game sales avenues are deeply broken markets.

Equilibrium (like most naive economic theory) only applies to markets where there's fluidity, robust competition, no monopolistic control, free information, etc. And of course those don't happen in the real world.

Whenever a market is not healthy, it provides an opportunity for unbalanced reward, well out of equilibrium.

Lack of information can be particularly be a factor in small niches. There can be a company that does something random like make height-adjustable massage tables. If they're a private operation and nobody really pays attention to them, they can have super high profit levels for something that's not particularly difficult - way out of equilibrium. If other people knew how easy that business was, lots of others would enter, but due to lack of information they don't.

Patents and other such mechanisms that create legally enforced distortions of the market. Of course things like the cable and utility systems are even worse.

On a large scale, government distortion means that huge fields like health care, finance, insurance, oil, farming, etc. are all more profitable than they should be.

Perhaps the biggest issue in downloadable games is the oligopoly of App Store and Steam. This creates an unhealthy market distortion and it's hard to say exactly what the long term affect of that will be. (of course you don't see it as "unhealthy" if you are the one benefiting from the favor of the great powers; it's unhealthy in a market fluidity and fair competition sense, and may slow or prevent equilibrium)

Of course new fields are not yet in equilibrium, and one of the best ways to "get rich quick" is to chase new fields. Software has been out of equilibrium for the past 50 years, and is only recently settling down. Eventually software will be a very poorly paid field, because it requires very little capital to become a programmer, it's low risk, and there are lots of people who can do it.

Note that in *every* field the best will always rise to the top and be paid accordingly.

Games used to be a great field to work in because it was a new field. New fields are exciting, they offer great opportunities for innovation, and they attract the best people. Mature industries are well into equilibrium and the only chances for great success are through either big risk, big capital investment, or crookedness.

03-31-13 - Index - Game Threading Architecture

Gathering the series for an index post :

cbloom rants 08-01-11 - A game threading model
cbloom rants 12-03-11 - Worker Thread system with reverse dependencies
cbloom rants 03-05-12 - Oodle Handle Table
cbloom rants 03-08-12 - Oodle Coroutines
cbloom rants 06-21-12 - Two Alternative Oodles
cbloom rants 07-19-12 - Experimental Futures in Oodle
cbloom rants 10-26-12 - Oodle Rewrite Thoughts
cbloom rants 12-18-12 - Async-Await ; Microsoft's Coroutines
cbloom rants 12-21-12 - Coroutine-centric Architecture
cbloom rants 12-21-12 - Coroutines From Lambdas
cbloom rants 12-06-12 - Theoretical Oodle Rewrite Continued
cbloom rants 02-23-13 - Threading - Reasoning Behind Coroutine Centric Design

I believe this is a good architecture, using the techniques that we currently have available, without doing anything that I consider bananas like writing your own programming language (*). Of course if you are platform-specific or know you can use C++11 there are small ways to make things more convenient, but the fundamental architecture would be about the same (and assuming that you will never need to port to a broken platform is a mistake I know well).

(* = a lot of people that I consider usually smart seem to think that writing a custom language is a great solution for lots of problems. Whenever we're talking about "oh reflection in C is a mess" or "dependency analysis should be automatic", they'll throw out "well if you had the time you would just write a custom language that does all this better". Would you? I certainly wouldn't. I like using tools that actually work, that new hires are familiar with, etc. etc. I don't have to list the pros of sticking with standard languages. In my experience every clever custom language for games is a huge fucking disaster and I would never advocate that as a good solution for any problem. It's not a question of limitted dev times and budgets.)

03-31-13 - Endian-Independent Structs

I dunno, maybe this is common practice, but I've never seen it before.

The easy way to load many file formats (I'll use a BMP here to be concrete) is just to point a struct at it :

    U16 bfType; 
    U32 bfSize; 
    U16 bfReserved1; 
    U16 bfReserved2; 
    U32 bfOffBits; 
} __attribute__ ((__packed__));


if ( bmfh->bfType != 0x4D42 )
    ERROR_RETURN("not a BM",0);


but of course this doesn't work cross platform.

So people do all kinds of convoluted things (which I have usually done), like changing to a method like :

U16 bfType = Get16LE(&ptr);
U32 bfSize = Get32LE(&ptr);

or they'll do some crazy struct-parse fixup thing which I've always found to be bananas.

But there's a super trivial and convenient solution :

    U16LE bfType; 
    U32LE bfSize; 
    U16LE bfReserved1; 
    U16LE bfReserved2; 
    U32LE bfOffBits; 
} __attribute__ ((__packed__));

where U16LE is just U16 on little-endian platforms and is a class that does bswap on itself on big-endian platforms.

Then you can still just use the old struct-pointing method and everything just works. Duh, I can't believe I didn't think of this earlier.

Similarly, here's a WAV header :

struct WAV_header_LE
    U32LE FOURCC_RIFF; // RIFF Header 
    U32LE riffChunkSize; // RIFF Chunk Size 
    U32LE FOURCC_WAVE; // WAVE Header 
    U32LE FOURCC_FMT; // FMT header 
    U32LE fmtChunkSize; // Size of the fmt chunk 
    U16LE audioFormat; // Audio format 1=PCM,6=mulaw,7=alaw, 257=IBM Mu-Law, 258=IBM A-Law, 259=ADPCM 
    U16LE numChan; // Number of channels 1=Mono 2=Sterio 
    U32LE samplesPerSec; // Sampling Frequency in Hz 
    U32LE bytesPerSec; // bytes per second 
    U16LE blockAlign; // normall NumChan* bytes per sample
    U16LE bitsPerSample; // Number of bits per sample 
}  __attribute__ ((__packed__));;


For file-input type structs, you just do this and there's no penalty. For structs you keep in memory you wouldn't want to eat the bswap all the time, but even in that case this provides a simple way to get the swizzle into native structs by just copying all the members over.

Of course if you have the Reflection-Visitor system that I'm fond of, that's also a good way to go. (cursed C, give me a "do this macro on all members").


03-30-13 - Error codes

Some random rambling on the topic of returning error codes.

Recently I've been fixing up a bunch of code that does things like

void  MutexLock( Mutex * m )
    if ( ! m ) return;

yikes. Invalid argument and you just silently do nothing. No thank you.

We should all know that silently nopping in failure cases is pretty horrible. But I'm also dealing with a lot of error code returns, and it occurs to me that returning an error code in that situation is not much better.

Personally I want unexpected or unhandleable errors to just blow up my app. In my own code I would just assert; unfortunately that's not viable in OS code or perhaps even in a library.

The classic example is malloc. I hate mallocs that return null. If I run out of memory, there's no way I'm handling it cleanly and reducing my footprint and carrying on. Just blow up my app. Personally whenever I implement an allocator if it can't get memory from the OS it just prints a message and exits (*).

(* = aside : even better is "functions that don't fail" which I might write more about later; basically the idea is the function tries to handle the failure case itself and never returns it out to the larger app. So in the case of malloc it might print a message like "tried to alloc N bytes; (a)bort/(r)etry/return (n)ull?". Another common case is when you try to open a file for write and it fails for whatever reason, it should just handle that at the low level and say "couldn't open X for write; (a)bort/(r)etry/change (n)ame?" )

I think error code returns are okay for *expected* and *recoverable* errors.

On functions that you realistically expect to always succeed and will not check error codes for, they shouldn't return error codes at all. I wrote recently about wrapping system APIs for portable code ; an example of the style of level 2 wrapping that I like is to "fix" the error returns.

(obviously this is not something the OS should do, they just have to return every error; it requires app-specific knowledge about what kind of errors your app can encounter and successfully recover from and continue, vs. ones that just mean you have a catastrophic unexpected bug)

For example, functions like lock & unlock a mutex shouldn't fail (in my code). 99% of the user code in the world that locks and unlocks mutexes doesn't check the return value, they just call lock and then proceed assuming the lock succeeded - so don't return it :

void mypthread_mutex_lock(mypthread_mutex_t *mutex)
    int ret = pthread_mutex_lock(mutex);
    if ( ret != 0 )

When you get a crazy unexpected error like that, the app should just blow up right at the call site (rather than silently failing and then blowing up somewhere weird later on because the mutex wasn't actually locked).

In other cases there are a mix of expected failures and unexpected ones, and the level-2 wrapper should differentiate between them :

bool mysem_trywait(mysem * sem)
        int res = sem_trywait( sem );
        if ( res == 0 ) return true; // got it

        int err = errno;
        if ( err == EINTR )
            // UNIX is such balls
        else if ( err == EAGAIN )
            // expected failure, no count in sem to dec :
            return false;
            // crazy failure; blow up :

(BTW best practice these days is always to copy "errno" out to an int, because errno may actually be #defined to a function call in the multithreaded world)

And since I just stumbled into it by accident, I may as well talk about EINTR. Now I understand that there may be legitimate reasons why you *want* an OS API that's interrupted by signals - we're going to ignore that, because that's not what the EINTR debate is about. So for purposes of discussion pretend that you never have a use case where you want EINTR and it's just a question of whether the API should put that trouble on the user or not.

I ranted about EINTR at RAD a while ago and was informed (reminded) this was an ancient argument that I was on the wrong side of.

Mmm. One thing certainly is true : if you want to write an operating system (or any piece of software) such that it is easy to port to lots of platforms and maintain for a long time, then it should be absolutely as simple as possible (meaning simple to implement, not simple in the API or simple to use), even at the cost of "rightness" and pain to the user. That I certainly agree with; UNIX has succeeded at being easy to port (and also succeeded at being a pain to the user).

But most people who argue on the pro-EINTR side of the argument are just wrong; they are confused about what the advantage of the pro-EINTR argument is (for example Jeff Atwood takes off on a general rant against complexity ; I think we all should know by now that huge complex APIs are bad; that's not interesting, and that's not what "Worse is Better" is about; or Jeff's example of INI files vs the registry - INI files are just massively better in every way, it's not related at all, there's no pro-con there).

(to be clear and simple : the pro-EINTR argument is entirely about simplicity of implementation and porting of the API; it's about requiring the minimum from the system)

The EINTR-returning API is not simpler (than one that doesn't force you to loop). Consider an API like this :

U64 system( U64 code );

doc :

if the top 32 bits of code are 77 this is a file open and the bottom 32 bits specify a device; the
return values then are 0 = call the same function again with the first 8 chars of the file name ...
if it returns 7 then you must sleep at least 1 milli and then call again with code = 44 ...
etc.. docs for 100 pages ...

what you should now realize is that *the docs are part of the API*. (that is not a "simple" API)

An API that requires you to carefully read about the weird special cases and understand what is going on inside the system is NOT a simple API. It might look simple, but it's in disguise. A simple API does what you expect it to. You should be able to just look at the function signature and guess what it does and be right 99% of the time.

Aside from the issue of simplicity, any API that requires you to write the exact same boiler-plate every time you use it is just a broken fucking API.

Also, I strongly believe that any API which returns error codes should be usable if you don't check the error code at all. Yeah yeah in real production code of course you check the error code, but for little test apps you should be able to do :

int fd = open("blah");



and that should work okay in my hack test app. Nope, not in UNIX it doesn't. Thanks to its wonderful "simplicity" you have to call "read" in a loop because it might decide to return before the whole read is done.

Another example that occurs to me is the reuse of keywords and syntax in C. Things like making "static" mean something completely different depending on how you use it makes the number of special keywords smaller. But I believe it actually makes the "API" of the language much *more* complex. Instead of having intuitive and obvious separate clear keywords for each meaning that you could perhaps figure out just by looking at them, you instead have to read a bunch of docs and have very technical knowledge of the internals of what the keywords mean in each usage. (there are legitimate advantages to minimizing the number of keywords, of course, like leaving as many names available to users as possible). Knowledge required to use an API is part of the API. Simplicity is determined by the amount of knowledge required to do things correctly.


03-26-13 - Simulating Deep Yield with a Wait

I'm becoming increasingly annoyed at my lack of "deep yield" for coroutines.

Any time you are in a work item, if you decide that you can get some more parallelism by doing a branch-merge inside that item, you need deep yield.

Remember you should never ever do an OS wait on a coroutine thread (with normal threads anyway; on a WinRT threadpool thread you can). The reason is the OS wait disables that worker thread, so you have one less. In the worst case, it leads to deadlock, because all your worker threads can be asleep waiting on work items, and with no worker threads they will never get done.

Anyway, I've cooked up a temporary work-around, it looks like this :

I'm in some function and I want to branch-merge

If I'm not on on a worker thread
  -> just do a normal branch-merge, send the work off and use a Wait for completion

If I am on a worker thread :

inc target worker thread count
if # currently live worker threads is < target count
  start a new worker thread (either create or wake from pool)

now do the branch-merge and use OS Wait
dec the target worker thread count

on each worker thread, after completing a work item and before popping more work :
if target worker thread count < currently live count
  stop self (go back into a sleeping state in the pool)

this is basically using OS threads to implement stack-saving deep yield. It's not awesome, but it is okay if deep yield is rare.

03-26-13 - Oodle 1.1 and GDC

Hey it's GDC time again, so if you're here come on by the RAD booth and say "hi" (or "fuck you", or whatever).

The Oodle web site just went live a few days ago.

Sometimes I feel embarassed (ashamed? humiliated?) that it's taken me five years to write a file IO and data compression library. Other times I think I've basically written an entire OS by myself (and all the docs, and marketing materials, and a video compressor, and aborted paging engine, and a bunch of other crap) and that doesn't sound so bad. I suppose the truth is somewhere in the middle. (perhaps with Oodle finally being officially released and selling, I might write a little post-mortem about how it's gone, try to honestly look back at it a bit. (because lord knows what I need is more introspection in my life)).

Oodle 1.1 will be out any day now. Main new features :

Lots more platforms.  Almost everything except mobile platforms now.

LZNIB!  I think LZNIB is pretty great.  8X faster to decode than ZLIB and usually
makes smaller files.

Other junk :
All the compressors can run parallel encode & decode now.
Long-range-matcher for LZ matching on huge files (still only in-memory though).
Incremental compressors for online transmission, and faster resets.

Personally I'm excited the core architecture is finally settling down, and we have a more focused direction to go forward, which is mainly the compressors. I hope to be able to work on some new compressors for 1.2 (like a very-high-compression option, which I currently don't have), and then eventually move on to some image compression stuff.


03-19-13 - Windows Sleep Variation

Hmm, I've discovered that Sleep(n) behaves very differently on my three Windows boxes.

(Also remember there are a lot of other issues with Sleep(n) ; the times are only reliable here because this is in a no-op test app)

This actually started because I was looking into Linux thread sleep timing, so I wrote a little test to just Sleep(n) a bunch of times and measure the observed duration of the sleep.

(Of course on Windows I do timeBeginPeriod(1) and bump my thread to very high priority (and timeGetDevCaps says the minp is 1)).

Anyway, what I'm seeing is this :

Win7 :
sleep(1) : average = 0.999 , sdev = 0.035 ,min = 0.175 , max = 1.568
sleep(2) : average = 2.000 , sdev = 0.041 ,min = 1.344 , max = 2.660
sleep(3) : average = 3.000 , sdev = 0.040 ,min = 2.200 , max = 3.774

Sleep(n) averages n
duration in [n-1,n+1]

WinXP :
sleep(1) : average = 1.952 , sdev = 0.001 ,min = 1.902 , max = 1.966
sleep(2) : average = 2.929 , sdev = 0.004 ,min = 2.665 , max = 2.961
sleep(3) : average = 3.905 , sdev = 0.004 ,min = 3.640 , max = 3.927

Sleep(n) averages (n+1)
duration very close to (n+1) every time (tiny sdev)

Win8 :
sleep(1) : average = 2.002 , sdev = 0.111 ,min = 1.015 , max = 2.101
sleep(2) : average = 2.703 , sdev = 0.439 ,min = 2.017 , max = 3.085
sleep(3) : average = 3.630 , sdev = 0.452 ,min = 3.003 , max = 4.130

average no good
Sleep(n) minimum very precisely n
duration in [n,n+1] (+ a little error)
rather larger sdev

it's like completely different logic on each of my 3 machines. XP is the most precise, but it's sleeping for (n+1) millis instead of (n) ! Win8 has a very precise min of n, but the average and max is quite sloppy (sdev of almost half a milli, very high variation even with nothing happening on the system). Win7 hits the average really nicely but has a large range, and is the only one that will go well below the requested duration.

As noted before, I had a look at this because I'm running Linux in a VM and seeing very poor performance from my threading code under Linux-VM. So I ran this experiment :

Sleep(1) on Linux :

native : average = 1.094 , sdev = 0.015 , min = 1.054 , max = 1.224
in VM  : average = 3.270 , sdev =14.748 , min = 1.058 , max = 656.297

in VM2 : average = 1.308 , sdev = 2.757 , min = 1.052 , max = 154.025

obviously being inside a VM on Windows is not being very kind to Linux's threading system. On the native box, Linux's sleep time is way more reliable than Windows (small min-max range) (and this is just with default priority threads and SCHED_OTHER, not even using a high priority trick like with the Windows tests above).

added "in VM2". So the VM threading seems to be much better if you let it see many fewer cores than you have. I'm running on a 4 core (8 hypercore) machine; the base "in VM" numbers are with the VM set to see 4 cores. "in VM2" is with the VM set to 2 cores. Still a really bad max in there, but much better overall.


03-16-13 - Writing Portable Code Rambles

Some thoughts after spending some time on this (still a newbie). How I would do it differently if I started from scratch.

1. Obviously you all know the best practice of using your own data types (S32 or whatever) and making macros for any kind of common operation that the standards don't handle well (like use a SEL macro instead of ?: , make a macro for ROT, etc). Never use bit-fields, make your own macros for manipulating bits within words. You also have to make your own whole macro meta-language for things not quite in the language, like data alignment, restrict/alias, etc. etc. (god damn C standard people, spend some time on the actual problems that real coders face every day. Thanks mkay). That's background and it's the way to go.

Make your own defines for SIZEOF_POINTER since stupid C doesn't give you any way to check sizeof() in a macro. You probably also want SIZEOF_REGISTER. You need your own equivalent of ptrdiff_t and intptr_t. Best practice is to use pointer-sized ints for all indexing of arrays and buffer sizes.

(one annoying complication is that there are platforms with 64 bit pointers on which 64-bit int math is very slow; for example they might not have a 64-bit multiply at all and have to emulate it. In that case you will want to use 32-bit ints for array access when possible; bleh)

Avoid using "wchar_t" because it is not always the same size. Try to explicitly use UTF16 or UTF32 in your code. You could make your own SIZEOF_WCHAR and select one or the other on the appropriate platform. (really try to avoid using wchar at all; just use U16 or U32 and do your own UTF encoding).

One thing I would add to the macro meta-language next time is to wrap every single function (and class) in my code. That is, instead of :

int myfunc( int args );


FUNC1 int FUNC2 myfunc(int args );

or even better :

FUNC( int , myfunc , (int args) );

this gives you lots of power to add attributes and other munging as may be needed later on some platforms. If I was doing this again I would use the last style, and I would have two of them, a FUNC_PUBLIC and FUNC_PRIVATE to control linkage. Probably should have separate wrapper macros for the proto and the body.

While you're at it you may as well have a preamble in every func too :

FUNC_PUBLIC_BODY( int , myfunc , (int args) )


which lets you add automatic func tracing, profiling, logging, and so on.

I wish I had made several different layers of platform Id #defines. The first one you want is the lowest level, which explicitly Id's the current platform. These should be exclusive (no overlaps), something like OODLE_PLATFORM_X86X64_WIN32 or OODLE_PLATFORM_PS3_PPU.

Then I'd like another layer that's platform *groups*. For me the groups would probably be OODLE_PLATFORM_GROUP_PC , GROUP_CONSOLE, and GROUP_EMBEDDED. Those let you make gross characterizations like on "GROUP_PC" you use more memory and have more debug systems and such. With these mutually exclusive platform checks, you should never use an #else. That is, don't do :

.. some code ..
.. fallback ..
it's much better to explicitly enumerate which platforms you want to go to which code block, and then have an
#error new platform
at the end of every check. That way when you try building on new platforms that you haven't thought carefully about yet, you get nice compiler notification about all the places where you need to think "should it use this code path or should I write a new one". Fallbacks are evil! I hate fallbacks, give me errors.

Aside from the explicit platforms and groups I would have platform flags or caps which are non-mutually exclusive. Things like PLATFORM_FLAG_STDIN_CONSOLE.

While you want the raw platform checks, in end code I wish I had avoided using them explicitly, and instead converted them into logical queries about the platform. What I mean is, when you just have an "#if some platform" in the code, it doesn't make it clear why you care that's the platform, and it doesn't make it reusable. For example I have things like :

// .. do string matching by U64 and xor/cntlz
// unaligned U64 read may be slow
// do string match byte by byte
what I should have done is to introduce an abstraction layer in the #if that makes it clear what I am checking for, like :

#error classify me

// .. do string matching by U64 and xor/cntlz
// unaligned U64 read may be slow
// do string match byte by byte

then it's really clear what you want to know and how to classify new platforms. It also lets you reuse that toggle in lots of places without code duping the fiddly bit, which is the platform classification.

Note that when doing this, it's best to make high level usage-specific switches. You might be tempted to try to use platform attributes there. Like instead of "PLATFORM_SWITCH_DO_STRING_MATCH_BIGWORDS" you might want to use "PLATFORM_SWITCH_UNALIGNED_READ_PENALTY" . But that's not actually what you want to know, you want to know if on my particular application (LZ string match) it's better to use big words or not, and that might not match the low level attribute of the CPU.

It's really tempting to skip all this and abuse the switches you can see (lord knows I do it); I see (and write) lots of code that does evil things like using "#ifdef _MSC_VER" to mean something totally different like "is this x86 or x64" ? Of course that screws you when you move to another x86 platform and you aren't detecting it correctly (or when you use MSVC to make PPC or ARM compiles).

Okay, that's all pretty standard, now for the new bit :

2. I would opaque out the system APIs in two levels. I haven't actually ever done this, so grains of salt, but I'm pretty convinced it's the right way to go after working with a more standard system.

(for the record : the standard way is to make a set of wrappers that tries to behave the same on all systems, eg. that tries to hide what system you are on as much as possible. Then if you need to do platform-specific stuff you would just include the platform system headers and talk to them directly. That's what I'm saying is not good.)

In the proposed alternative, the first level would just be a wrapper on the system APIs with minimal or no behavior change. That is, it's just passing them through and standardizing naming and behavior.

At this level you are doing a few things :

2.A. Hiding the system includes from the rest of your app. System includes are often in different places, and often turn on compiler flags in nasty ways. You want to remove that variation from the rest of your code so that your main codebase only sees your own wrapper header.

2.B. Standardizing naming. For example the MSVC POSIX funcs are all named wrong; at this level you can patch that all up.

2.C. Fixing things that are slightly different or don't work on various platforms where they really should be the same. For example things like pthreads are not actually all the same on all the pthreads platforms, and that can catch you out in nasty ways. (eg. things like sem_init always failing on Mac).

Note this is *not* trying to make non-POSIX platforms look like POSIX. It's not hiding the system you're on, just wrapping it in a standard way.

2.D. I would also go ahead and add my own asserts for args and returns in this layer, because I hate functions that just return error codes when there's a catastrophic failure like a null arg or an EHEAPCORRUPT or whatever.

So once you have this wrapper you no longer call any system funcs directly from your main codebase, but you still would be doing things like :


    HANDLE h = platform_CreateFile( ... )


    int fd = platform_open( name , flags )

    #error unknown platform

that is, you're not hiding what platform you're on, you're still letting the larger codebase get to the low level calls, it's just the mess of how fucked they are that's hidden a bit.

3. You then have a second level of wrapping which tries to make same-action interfaces that dont require you to know what platform you're on. Second level is written on the first level.

The second level wrappers should be as high level as necessary to opaque out the operation. For example rather than having "make temp file name" and "open file" you might have "open file with temp name", because on some platforms that can be more efficient when you know it is a high-level combined op. You don't just have "GetTime" you have "GetTimeMonotonic" , because on some platforms they have an efficient monotonic clock for you, and on other platforms/hardwares you may have to do a lot of complicated work to ensure a reliable clock (that you don't want to do in the low level timer).

When a platform can't provide a high-level function efficiently, rather than emulate it in a complex way I'd rather just not have it - not a stub that fails, but no definition at all. That way I get a compile error and in those spots I can do something different, using the level 1 APIs.

The first level wrappers are very independent of the large code base's usage, but the second level wrappers are very much specifically designed for their usage.

To be clear about the problem of making platform-hiding second layer wrappers, consider something like OpenFile(). What are the args to that? What can it do? It's hopeless to make something that works on all platforms without greatly reducing the capabilities of some platforms. And the meaning of various options (like async, temporary, buffered, etc.) all changes with platform.

If you wanted to really make a general purpose multi-platform OpenFile you would have to use some kind of "caps" query system, where you first do something like OpenFile_QueryCaps( OF_DOES_UNBUFFERED_MEAN_ALIGNMENT_IS_REQUIRED ) and it would be an ugly disaster. (and it's obviously wrong on the face of it, because really what you're doing there is saying "is this win32" ?). The alternative to the crazy caps system is to just make the high level wrappers very limited and specific to your usage. So you could make a platform-agnostic wrapper like OpenFile_ForReadShared_StandardFlagsAndPermissions(). Then the platforms can all do slightly different things and satisfy the high level goal of the imperative in the best way for that platform.

A good second level has as few functions as possible, and they are as high level as possible. Making them very high level allows you to do different compound ops on the platform in a way that's hidden from the larger codebase.


03-10-13 - Two LZ Notes

Note 1 : on rep matches.

"Rep matches" are a little weird. They help a lot, but the reason why they help depends on the file you are compressing. (rep match = repeat match, gap match, aka "last offset")

On text files, they work as interrupted matches, or "gap matches". They let you generate something like :

stand on the floor
stand in the door

stand in the door
[stand ][i][n the ][d][oor]

[off 19, len 6][1 lit][rep len 6][1 lit][off 18, len 3]

that is, you have a long match of [stand on the ] but with a gap at the 'o'.

Now, something I observed was that more than one last offset continues to help. On text the main benefit from having two last offsets is that it lets you use a match for the gap. When the gap is not just one character but a word, you might want to use a match to put that word in, in which case the continuation after the gap is no longer the first last-offset, it's the second one. eg.

how to work with animals
how to cope with animals

[how to ][cope][ with animals]
[off 25 ][off 32][off 25 (rep2)]

You could imagine alternative coding structures that don't require keeping some number of "last offsets". (oddly, the last offset maintenance can be a large part of decode time, because maintaining an MTF list is something that CPUs do incredibly poorly). For example you could code with a scheme where you just send the entire long match, and then any time you send a long match you send a flag for "are there any gaps", and if so you then code some gaps inside the match.

The funny thing is, on binary files "last offsets" do something else which can be more important. They become the most common offsets. In particular, on highly structured binary data, they will generally be some factor of the structure size. eg. on a file that has a struct size of 36, and that struct has dwords and such in it, the last offsets will generally be things like 4,8,16,36, or 72. They provide a sort of dictionary of the most common offsets so that you can code those smaller. You are still getting the gap-match effect, but the common-offset benefit is much bigger on these files.

(aside : word-replacing transform on text really helps LZ (and everything) by removing the length variance of tokens. In particular for LZ77, word length variance breaks rep matches. There are lots of common occurances of a single replaced word in a phrase, like : "I want some stuff" -> "I want the stuff". You can't get a rep match here of [ stuff] because the offset changed because the substituted word was different length. If you do WRT first, then gap matches get these.)

Note 2 : on offset structure.

I've had it in the back of my head for quite some time now to do an LZ compressor specifically designed for structured data.

One idea I had was to use "2d" match offsets. That is, send a {dx,dy} where dx is within the record and dy is different records. Like imagine the data is in a table, dy is going back rows, dx is an offset on the row. You probably want to mod dx around the row so its range is always the same, and special case dy=0 (matches within your own record).

It occurred to me that the standard way of sending LZ offsets these days actually already does this. The normal way that good LZ's send offsets these days is to break it into low and high parts :

low = offset & 7F;
high = offset >> 7;
or similar, then you send "high" using some kind of "NoSB" scheme (Number of Significant Bits is entropy coded, and the bits themselves are sent raw), and you send "low" with an order-0 entropy coder.

But this is just a 2d structured record offset for a particular power-of-2 record size. It's why when I've experimented with 2d offsets I haven't seen huge wins - because I'm already doing it.

There is some win to be had from custom 2d-offsets (vs. the standard low/high bits scheme) when the record size is not a power of two.


03-06-13 - Sympathy for the Library Writer

Over the years of being a coder who was a library-consumer and not a library-writer, I've done my share of griping about annoying API's or what I saw as pointless complication or ineffiency. Man, I've been humbled by my own experience trying to write a public library. It is *hard*.

The big problem with libraries is that you don't control how they're used. This is in contrast to game engines. Game engines are not libraries. I've worked on many game engines over the years, including ones that went out to large free user bases (Genesis 3d and Wild Tangent), and they are much much easier than libraries.

The difference is that game engines generally impose an architecture on the user. They force you to use it in a certain way. (this is of course why more advanced developers despise them so much; it sucks to have some 3rd party telling you your code architecture). It's totally acceptable if a game engine only works well when you use it in the approved way, and is really slow if you abuse it, or it could even crash if you use it oddly.

A library has to be flexible about how it's used; it can't impose a system on the user, like a certain threading model, or a certain memory management model, or even an error-handling style.

Personally when I do IO for games, I make a "tool path" that just uses stdio and is very simple and flexible, does streaming IO and text parsing and so on, but isn't shipped with the game, and I make a "game path" that only does large-block async IO that's pre-baked so you can just point at it. I find that system is powerful enough for my use, it's easy to write and use. It means that the "tool path" doesn't have to be particularly fast, and the fast game path doesn't need to support buffered character IO or anything other than big block reads.

But I can't force that model on clients, so I have to support all the permutations and I have to make them all decently fast.

A lot of times in the past I've complained about over-complicated APIs that have tons of crazy options that nobody ever needs (look at the IJG jpeg code for example). Well, now I see that often those complicated APIs were made because somebody (probably somebody important) needed those options. Of course as the library provider you can offer the complex interface and also simpler alternatives, but that has its own pitfalls of making the API bigger and more redundant (like if you offer OpenFileSimple and OpenFileComplex); in some ways it's better to only offer the complex API and make the user wrap it and reduce the parameter set to what they actually use.

There's also a sort of "liability" issue in libraries. Not exactly legal liability, but program bad behavior liability. Lots of things that would make the library easier to use and faster are naughty to do automatically. For example Oodle under Vista+ can run faster with elevated priviledge, to get access to some of the unsecure file APIs (like extending without zeroing), but it would be naughty for me to do that automatically, so instead I have to add an extra step to make the client specifically ask for that.

Optimization for me has really become a nightmare. At first I was trying to make every function fast, but it's impossible, there are just too many entry points and too many usage patterns. Now my philosophy is to make certain core functions fast, and then address problems in the bigger high level API as customers see issues. I remember as a game developer always being so pissed that all the GL drivers were specially optimized for Id. I would want to use the API in a slightly different style, and my way would be super slow, not for any good reason but just because it hadn't gotten the optimization loving of the important customer's use case.

I used to also rail about the "unnecessary" argument checking that all the 3d APIs do. It massively slows them down, and I would complain that I had ensured the arguments were good so just fucking pass them through, stop slowing me down with all your validation! But now I see that if you really do that, you will just constantly be crashing people as they pass in broken args. In fact arg validation is often the way that people figure out the API, either because they don't read the docs or because the docs are no good.

(this is not even getting into the issue of API design which is another area where I have been suitably humbled)

ADDENDUM : I guess I should mention the really obvious points that I didn't make.

1. One of the things that makes a public library so hard after release is that you can't refactor. The normal way I make APIs for myself (and for internal teams) is to sort of make an effort at a good API the first time, but it usually sucks, and you rip it out and go through big scourges of find-rep. That only works when you control all the code, the library and the consumer. It's only after several iterations that the API becomes really nice (and even then it's only nice for that specific use case, it might still suck in the wild).

2. APIs without users almost always suck. When someone goes away in a cave and works on a big new fancy library and then shows it to the world, it's probably terrible. This a problem that I think everyone at RAD faces. The code of mine that I really like is stuff that I use over and over, so that I see the flaws and when I want it to be easier to use I just go fix it.

3. There are two separate issues about what makes an API "good". One is "is it good for the user?" and one is "is it good for the library maintainer?". Often they are the same but not always.

Anyway, the main point of this post is supposed to be : the next time you complain about a bad library design, there may well be valid reasons why it is the way it is; they have to balance a lot of competing goals. And even if they got it wrong, hey it's hard.


03-01-13 - Zopfli

zopfli seems to make small zips. There's no description of the algorithm so I can't comment on it. But hey, if you want small zips it seems to be the current leader.

(update : I've had a little look, and it seems to be pretty straightforward, it's an optimal parser + huff reset searcher. There are various other prior tools to do this (kzip,deflopt,defluff,etc). It's missing some of the things that I've written about before here, such as methods of dealing with the huff-parse feedback; the code looks pretty clean, so if you want a good zip-encoder code it looks like a good place to start.)

I've written these things before, but I will summarize here how to make small zips :

1. Use an exact (windowed) string matcher.

cbloom rants 09-24-12 - LZ String Matcher Decision Tree

2. Optimal parser. Optimal parsing zip is super easy because it has no "repeat match", so you can use plain old backwards scan. You do have the huffman code costs, so you have to consider at least one match candidate for each codeword length.

cbloom rants 10-10-08 - 7 - On LZ Optimal Parsing
cbloom rants 09-04-12 - LZ4 Optimal Parse

3. Deep huffman reset search. You can do this pretty easily by using some heuristics to set candidate points and then building a bottom-up tree. Zopfli seems to use a top-down greedy search. More huffman resets makes decode slower, so a good encoder should expose some kind space-speed tradeoff parameter (and/or a maximum number of resets).

cbloom rants 06-15-10 - Top down - Bottom up
cbloom rants 10-02-12 - Small note on Adaptive vs Static Modeling

4. Multi-parse. The optimal parser needs to be seeded in some way, with either initial code costs or some kind of heuristic parse. There may be multiple local minima, so the right way to do it is to run 4 seeds (or so) simultaneously with different strategies.

cbloom rants 09-11-12 - LZ MinMatchLen and Parse Strategies

5. The only unsolved bit : huffman - parse feedback. The only solution I know to this is iteration. You should use some tricks like smoothing and better handling of the zero-frequency symbols, but it's just heuristics and iteration.

One cool thing to have would be a cheap way to compute incremental huffman cost.

That is, say you have some array of symbols. The symbols have a corresponding histogram and huffman code. The full huffman cost is :

fullcost(symbol set) = cost{ transmit code lengths } + sum[n] { codelen[n] * count[n] }
that is, the cost to send the code lengths + the cost of sending all the symbols with those code lengths.

You'd like to be able to do an incremental update of fullcost. That is, if you add one more symbol to the set, what is the delta of fullcost ?

*if* the huffman code lengths don't change, then the delta is just +codelen[symbol].

But, the addition of the symbol might change the code lengths, which causes fullcost to change in several ways.

I'm not sure if there's some clever fast way to do incremental updates; like when adding the symbol pushes you over the threshold to change the huffman tree, it often only changes some small local part of the tree, so you don't have to re-sum your whole histogram, just the changed part. Then you could slide your partition point across an array and find the optimal point quite quickly.

Now some ranting.

How sad is it that we're still using zip?

I've been thinking about writing my own super-zipper for many years, but I always stop myself because WTF is the point? I don't mean for the world, I guess I see that it is useful for some people, but it does nothing for *me*. Hey I could write some thing and probably no one would use it and I wouldn't get any reward from it and it would just be another depressing waste of some great effort like so many other things in my life.

It's weird to me that the best code in the world tends to be the type of code that's given away for free. The little nuggets of pure genius, the code that really has something special in it - that tends to be the free code. I'm thinking of compressors, hashers, data structures, the real gems. Now, I'm all for free code and sharing knowledge and so on, but it's not equal. We (the producers of those gems) are getting fucked on the deal. Apple and the financial service industry are gouging me in every possible immoral way, and I'm giving away the best work of my life for nothing. It's a sucker move, but it's too late. The only sensible play in a realpolitik sense of your own life optimization is to not work in algorithms.

Obviously anyone who claims that patents provide money to inventors is either a liar (Myhrvold etc.) or just has no familiarity with actual engineering. I often think about LZ77 as a case in point. The people who made money off LZ77 patents were PK and Stac, both of whom contributed *absolutely nothing*. Their variants were completely trivial obvious extensions of the original idea. Of course the real inventors (L&Z, and the modern variant is really due to S&S) didn't patent and got nothing. Same thing with GIF and LZW, etc. etc. perhaps v42 goes in there somewhere too; not a single one of the compression-patent money makers was an innovator. (and this is even igoring the larger anti-patent argument, which is that things like LZ77 would have been obvious to any number of researchers in the field at the time; it's almost always impossible to attribute scientific invention/discovery to an individual)

old rants