4/30/2013

04-30-13 - Packing Values in Bits - Flat Codes

One of the very simplest forms of packing values in bits is simply to store a value with non-power-of-2 range and all values of equal probability.

You have a value that's in [0,N). Ideally all code lengths would be the same ( log2(N) ) which is fractional for N not a power of 2. With just bit output, we can't write fractional bits, so we will lose some efficiency. But how much exactly?

You can of course trivially write a symbol in [0,N) by using log2ceil(N) bits. That's just going up to the next integer bit count. But you're wasting values in there, so you can take each wasted value and use it to reduce the length of a code that you need. eg. for N = 5 , start with log2ceil(N) bits :

0 : 000
1 : 001
2 : 010
3 : 011
4 : 100
x : 101
x : 110
x : 111
The first five codes are used for our values, and the last three are wasted. Rearrange to interleave the wasted codewords :
0 : 000
x : 001
1 : 010
x : 011
2 : 100
x : 101
3 : 110
4 : 111
now since we have adjacent codes where one is used and one is not used, we can reduce the length of those codes and still have a prefix code. That is, if we see the two bits "00" we know that it must always be a value of 0, because "001" is wasted. So simply don't send the third bit in that case :
0 : 00
1 : 01
2 : 10
3 : 110
4 : 111

(this is a general way of constructing shorter prefix codes when you have wasted values). You can see that the number of wasted values we had at the top is the number of codes that can be shortened by one bit.

A flat code is written thusly :


void OutputFlat(int sym, int N)
{
    ASSERT( N >= 2 && sym >= 0 && sym < N );

    int B = intlog2ceil(N);
    int T = (1<<B) - N;
    // T is the number of "wasted values"
    if ( sym < T )
    {
        // write in B-1 bits
        PutBits(sym, B-1);
    }
    else
    {
        // write in B bits
        // push value up by T
        PutBits(sym+T, B);
    }
}

int InputFlat(int sym,int N)
{
    ASSERT( N >= 2 && sym >= 0 && sym < N );

    int B = intlog2ceil(N);
    int T = (1<<B) - N;

    int sym = GetBits(B-1);
    if ( sym < T )
    {
        return sym;
    }
    else
    {
        // need one more bit :
        int ret = (sym<<1) - T + GetBits(1);        
        return ret;
    }
}

That is, we write (T) values in (B-1) bits, and (N-T) in (B) bits. The intlog2ceil can be slow, so in practice you would want to precompute that or pass it in as a parameter.

So, what is the loss vs. ideal, and where does it occur? Let's work it out :


H = log2(N)  is the ideal (fractional) entropy

N is in (2^(B-1),2^B]
so H is in (B-1,B]

The number of bits written by the flat code is :

L = ( T * (B-1) + (N-T) * B ) / N

with T = 2^B - N

Let's set

N = f * 2^B

with f in (0.5,1] our fractional position in the range.

so T = 2^B * (1 - f)

At f = 0.5 and 1.0 there's no loss, so there must be a maximum in that interval.

Doing some simplifying :

L = (T * (B-1) + (N-T) * B)/N
L = (T * B - T + N*B - T * B)/N
L = ( N*B - T)/N = B - T/N
T/N = (1-f)/f = (1/f) - 1
L = B - (1/f) + 1

The excess bits is :

E = L - H

H = log2(N) = log2( f * 2^B ) = B + log2(f)

E = (B - (1/f) + 1) - (B + log2(f))
E = 1 - (1/f) - log2(f)

so find the maximum of E by taking a derivative :

d/df(E) = 0
d/df(E) = 1/f^2 - (1/f)/ln2
1/f^2 = (1/f)/ln2
1/f = 1/ln(2)
f = ln(2)
f = 0.6931472...

and at that spot the excess is :

E = 1 - (1/ln2) - ln(ln2)/ln2
E = 0.08607133...

The worst case is 8.6% of a bit per symbol excess. The worst case appears periodically, once for each power of two.

The actual excess bits output for some low N's :

The worst case actually occurs as N->large, because at higher N you can get f closer to that worst case fraction (ln(2)). At lower N, the integer steps mean you miss the worst case and so waste less. This is perhaps a bit surprising, you might think that the worst case would be at something like N = 3.

In fact for N = 3 :


H = l2(3) = 1.584962 ...

L = average length written by OutputFlat

L = (1+2+2)/3 = 1.66666...

E = L - H = 0.08170421...

(obviously if you measure the loss as a percentage of the output length, the worst case is at N=3, and there it's 5.155% of the entropy).

4/21/2013

04-21-13 - How to grow Linux disk under VMWare

There's a lot of these guides around the net, but I found them all a bit confusing to follow, so my experience :
  • 1. Power off the VM.

  • 2. Make a backup of the whole VM in case something goes wrong. Just find the dir containing it and copy the whole thing.

  • 3. VMWare Settings -> Hardware -> Hard Disk -> Utilities -> Expand change to whatever size you want.

  • 4. This has just expanded the virtual disk, now you must grow the partition on the disk. Linux does not have good tools to grow a partition that is running the OS, so don't boot your VM back into Linux.

    (there is some LVM stuff that lets you make multiple partitions and then treat them as a single one, but for a Unix newb like myself that looks too scary).

  • 5. Download GParted ISO. VMWare Settings -> Hardware -> CD/DVD -> Use ISO -> point it at GParted.

    Also make sure "Connect at Power On" is checked.

  • 6. Now you have to get the virtual machine to boot from CD. Getting into the BIOS interactively was impossible for me. Fortunately VMWare has a solution. Find your VM files and edit the .vmx config file. Add this line :
    bios.forceSetupOnce = "TRUE"
    

  • 7. Power on the VM and you should enter the BIOS. Go to "Boot" and put the CD first. Save and Exit and you should now boot into GParted.

  • 8. Using GParted is pretty self explanatory. It's a good tool. When you're done, shut down and Power Off the VM.

    My VM was set up with a swap partitition, so I had to move that to the end before I could grow the primary partition. I hear that you can set up Linux with a swap file instead of a swap partition; that would be preferable. A swap partition makes zero sense in a VM where the disks are virtualized anyway (so the advantage of keeping the swap thrashing off your main disk doesn't exist). Not something I want to change though.

  • 9. VMWare Settings -> Hardware -> CD/DVD -> turn off "Connect at Power On"
It seems to me it's a good idea to leave it this way - BIOS is set to boot first from CD, but the VM is set with no CD hardware enabled. This makes it easy to change the ISO and just check that box any time you want to boot from an ISO, rather than having to go into that BIOS nightmare again.


More generally, what have I learned about multi-platform development from working at RAD ?

That it's horrible, really horrible, and I pray that I never have to do it again in my life. Ugh.

Just writing cross-platform code is not the issue (though that's horrible enough, solely due to stupid non-fundamental issues like the fact that struct packing isn't standardized, adding signed ints isn't standardized, restrict/noalias isn't standardized, inline linkage varies greatly, etc. urg wtf etc etc). If you're just releasing some code on the net and offering it for many platforms (leaving it up to the downloaders to actually build it and test it), your life is easy. The horrible part is if you actually have to maintain machines and build systems for all those platforms, test them, be able to debug on them, keep all the sdk's up to date, etc. etc.

(in general coding is easy when you don't actually test your code and make sure it works well, which a surprising number of people think is "done"; hey it compiles, I'm done! umm, no...)

(I guess that's a more general life thing; I observe a lot of people who just do things and don't actually measure whether the "doing" was successful or done well, but they just move on and are generally happy. People who stress over whether what they're doing is actually a good job or not are massively less happy but also actually do good work.)

I feel like I spend 90% of my time on stupid fucking non-algorithmic issues like this Linux partition resizing shit (probably more like 20%, but that's still frustratingly high). The regression tests are failing on Linux, okay have to figure out why, oh it's because the VM disk is too small, okay how do I fix that; or the PS4 compiler has a bug I have to work around, or the system software on this system has a bug, or the Mac clang wants to spew pointless warnings about anonymous namespaces, or my tests aren't working on Xenon .. spend some time digging .. oh the console is just turned off, or the IP changed or it got reflashed and my SDK doesn't work anymore, and blah blah fucking blah. God dammit I just want to be able to write algorithms. I miss coding, I miss thinking about hard problems. Le sigh.

I've written before about how in my imagination I could hire some kid for $50k to do all this shit work for me and it would be a huge win for me overall. But I'm afraid it's not that easy in reality.

What really should exist is a "coder cloud" service. There should be a bunch of VMs of different OS'es with various compilers and SDKs installed, so I can just say "build my shit for X with Y". Of course you need to be able to run tests on that system as well, and if something goes wrong you need remote desktop for interactive debugging. It's got to have every platform, including things like game consoles where you need license agreements, which is probably a no-go in reality because corporations are jerks. There's got to be superb customer service, because if I can't rely on it for builds at every moment of every day then it's a no-go. Unfortunately, programmers are almost uniformly overly optimistic about this kind of thing (in that they massively overestimate their own ability to manage these things quickly) so wouldn't want to pay what it costs to run that service.

4/10/2013

04-10-13 - Waitset Resignal Notes

I've been redoing my low level waitset and want to make some notes. Some previous discussion of the same issues here :

cbloom rants 11-28-11 - Some lock-free rambling
cbloom rants 11-30-11 - Some more Waitset notes
cbloom rants 12-08-11 - Some Semaphores

In particular, two big things occurred to me :

1. I talked before about the "passing on the signal" issue. See the above posts for more in depth details, but in brief the issue is if you are trying to do NotifyOne (instead of NotifyAll), and you have a double-check waitset like this :


{
waiter = waitset.prepare_wait(condition);

if ( double check )
{
    waiter.cancel();
}
else
{
    waiter.wait();
    // possibly loop and re-check condition
}

}

then if you get a signal between prepare_wait and cancel, you didn't need that signal, so a wakeup of another thread that did need that signal can be missed.

Now, I talked about this before as an "ugly hack", but over time thinking about it, it doesn't seem so bad. In particular, if you put the resignal inside the cancel() , so that the client code looks just like the above, it doesn't need to know about the fact that the resignal mechanism is happening at all.

So, the new concept is that cancel atomically removes the waiter from the waitset and sees if it got a signal that it didn't consume. If so, it just passes on that signal. The fact that this is okay and not a hack came to me when I thought about under what conditions this actually happens. If you recall from the earlier posts, the need for resignal comes from situations like :


T0 posts sem , and signals no one
T1 posts sem , and signals T3
T2 tries to dec count and sees none, goes into wait()
T3 tries to dec count and gets one, goes into cancel(), but also got the signal - must resignal T2

the thing is this can only happen if all the threads are awake and racing against each other (it requires a very specific interleaving); that is, the T3 in particular that decs count and does the resignal had to be awake anyway (because its first check saw no count, but its double check did dec count, so it must have raced with the sem post). It's not like you wake up a thread you shouldn't have and then pass it on. The thread wakeup scheme is just changed from :

T0 sem.post --wakes--> T2 sem.wait
T1 sem.post --wakes--> T3 sem.wait

to :

T0 sem.post
T1 sem.post --wakes--> T3 sem.wait --wakes--> T2 sem.wait

that is, one of the consumer threads wakes its peer. This is a tiny performance loss, but it's a pretty rare race, so really not a bad thing.

The whole "double check" pathway in waitset only happens in a race case. It occurs when one thread sets the condition you want right at the same time that you check it, so your first check fails and after you prepare_wait, your second check passes. The resignal only occurs if you are in that race path, and also the setting thread sent you a signal between your prepare_wait and cancel, *and* there's another thread waiting on that same signal that should have gotten it. Basically this case is quite rare, we don't care too much about it being fast or elegant (as long as it's not disastrously slow), we just need behavior to be correct when it does happen - and the "pass on the signal" mechanism gives you that.

The advantage of being able to do just a NotifyOne instead of a NotifyAll is so huge that it's worth adopting this as standard practice in waitset.

2. It then occurred to me that the waitset PrepareWait and Cancel could be made lock-free pretty trivially.

Conceptually, they are made lock free by turning them into messages. "Notify" is now the receiver of messages and the scheme is now :


{
waiter w;
waitset : send message { prepare_wait , &w, condition };

if ( double check )
{
    waitset : send message { cancel , &w };
    return;
}

w.wait();
}

-------

{
waitset Notify(condition) :
first consume all messages
do prepare_wait and cancel actions
then do the normal notify
eg. see if there are any waiters that want to know about "condition"
}

The result is that the entire wait-side operation is lock free. The notify-side still uses a lock to ensure the consistency of the wait list.

This greatly reduces contention in the most common usage patterns :


Mutex :

only the mutex owner does Notify
 - so contention of the waitset lock is non-existant
many threads may try to lock a mutex
 - they do not have any waitset-lock contention

Semaphore :

the common case of one producer and many consumers (lots of threads do wait() )
 - zero contention of the waitset lock

the less common case of many producers and few consumers is slow

Another way to look at it is instead of doing little bits of waitlist maintenance in three different places (prepare_wait,notify,cancel) which each have to take a lock on the list, all the maintenance is moved to one spot.

Now there are some subtleties.

If you used a fresh "waiter" every time, things would be simple. But for efficiency you don't want to do that. In fact I use one unique waiter per thread. There's only one OS waitable handle needed per thread and you can use that to implement every threading primitive. But now you have to be able to recycle the waiter. Note that you don't have to worry about other threads using your waiter; the waiter is per-thread so you just have to worry about when you come around and use it again yourself.

If you didn't try to do the lock-free wait-side, recycling would be easy. But with the lock-free wait side there are some issues.

First is that when you do a prepare-then-cancel , your cancel might not actually be done for a long time (it was just a request). So if you come back around on the same thread and call prepare() again, prepare has to check if that earlier cancel has been processed or not. If it has not, then you just have to force the Notify-side list maintenance to be done immediately.

The second related issue is that the lock-free wait-side can give you spurious signals to your waiter. Normally prepare_wait could clear the OS handle, so that when you wait on it you know that you got the signal you wanted. But because prepare_wait is just a message and doesn't take the lock on the waitlist, you might actually still be in the waitlist from the previous time you used your waiter. Thus you can get a signal that you didn't want. There are a few solutions to this; one is to allow spurious signals (I don't love that); another is to detect that the signal is spurious and wait again (I do this). Another is to always just grab the waitlist lock (and do nothing) in either cancel or prepare_wait.


Ok, so we now have a clean waitset that can do NotifyOne and gaurantee no spurious signals. Let's use it.

You may recall we've looked at a simple waitset-based mutex before :


U32 thinlock;

Lock :
{
    // first check :
    while( Exchange(&thinlock,1) != 0 )
    {
        waiter w; // get from TLS
        waitset.PrepareWait( &w, &thinlock );

        // double check and put in waiter flag :
        if ( Exchange(&thinlock,2) == 0 )
        {
            // got it
            w.Cancel();
            return;
        }

        w.Wait();
    }
}

Unlock :
{
    if ( Exchange(&thinlock,0) == 2 )
    {
        waitset.NotifyAll( &thinlock );
    }
}
This mutex is non-recursive, and of course you should spin doing some TryLocks before going into the wait loop for efficiency.

This was an okay way to build a mutex on waitset when all you had was NotifyAll. It only does the notify if there are waiters, but the big problem with it is if you have multiple waiters, it wakes them all and then they all run in to try to grab the mutex, and all but one fail and go back to sleep. This is a common type of unnecessary-wakeup thread-thrashing pattern that sucks really bad.

(any time you write threading code where the wakeup means "hey wakeup and see if you can grab an atomic" (as opposed to "wakeup you got it"), you should be concerned (particularly when the wake is a broadcast))

Now that we have NotifyOne we can fix that mutex :


U32 thinlock;

Lock :
{
    // first check :
    while( Exchange(&thinlock,2) != 0 ) // (*1)
    {
        waiter w; // get from TLS
        waitset.PrepareWait( &w, &thinlock );

        // double check and put in waiter flag :
        if ( Exchange(&thinlock,2) == 0 )
        {
            // got it
            w.Cancel(waitset_resignal_no); // (*2)
            return;
        }

        w.Wait();
    }
}

Unlock :
{
    if ( Exchange(&thinlock,0) == 2 ) // (*3)
    {
        waitset.NotifyOne( &thinlock );
    }
}
We changed the NotifyAll to NotifyOne , but two funny bits are worth noting : (*1) - we must now immediately exchange in the waiter-flag here; in the NotifyAll case it worked to put a 1 in there for funny reasons (see cbloom rants 07-15-11 - Review of many Mutex implementations , where this type of mutex is discussed as "Three-state mutex using Event" ), but it doesn't work with the NotifyOne. (*2) - with a mutex you do not need to pass on the signal when you stole it and cancelled. The reason is just that there can't possibly be any more mutex for another thread to consume. A mutex is a lot like a semaphore with a maximum count of 1 (actually it's exactly like it for non-recursive mutexes); you only need to pass on the signal when it's possible that some other thread needs to know about it. (*3) - you might think the check for == 2 here is dumb because we always put in a 2, but there's code you're not seeing. TryLock should still put in a 1, so in the uncontended cases the thinlock will have a value of 1 and no Notify is done. The thinlock only goes to a 2 if there is some contention, and then the value stays at 2 until the last unlock of that contended sequence.

Okay, so that works, but it's kind of silly. With the mechanism we have now we can do a much neater mutex :


U32 thinlock; // = 0 initializes thinlock

Lock :
{
    waiter w; // get from TLS
    waitset.PrepareWait( &w, &thinlock );

    if ( Fetch_Add(&thinlock,1) == 0 )
    {
        // got the lock (no need to resignal)
        w.Cancel(waitset_resignal_no);
        return;
    }

    w.Wait();
    // woke up - I have the lock !
}

Unlock :
{
    if ( Fetch_Add(&thinlock,-1) > 1 )
    {
        // there were waiters
        waitset.NotifyOne( &thinlock );
    }
}
The mutex is just a wait-count now. (as usual you should TryLock a few times before jumping in to the PrepareWait). This mutex is more elegant; it also has a small performance advantage in that it only calls NotifyOne when it really needs to; because its gate is also a wait-count it knows if it needs to Notify or not. The previous Mutex posted will always Notify on the last unlock whether or not it needs to (eg. it will always do one Notify too many).

This last mutex is also really just a semaphore. We can see it by writing a semaphore with our waitset :


U32 thinsem; // = 0 initializes thinsem

Wait :
{
    waiter w; // get from TLS
    waitset.PrepareWait( &w, &thinsem );

    if ( Fetch_Add(&thinsem,-1) > 0 )
    {
        // got a dec on count
        w.Cancel(waitset_resignal_yes); // (*1)
        return;
    }

    w.Wait();
    // woke up - I got the sem !
}

Post :
{
    if ( Fetch_add(&thinsem,1) < 0 )
    {
        waitset.NotifyOne( &thinsem );
    }
}

which is obviously the same. The only subtle change is at (*1) - with a semaphore we must do the resignal, because there may have been several posts to the sem (contrast with mutex where there can only be one Unlock at a time; and the mutex itself serializes the unlocks).


Oh, one very subtle issue that I only discovered due to relacy :

waitset.Notify requires a #StoreLoad between the condition check and the notify call. That is, the standard pattern for any kind of "Publish" is something like :


Publish
{
    change shared variable
    if ( any waiters )
    {
        #StoreLoad
        waitset.Notify()
    }
}

Now, in most cases, such as the Sem and Mutex posted above, the Publish uses an atomic RMW op. If that is the case, then you don't need to add any more barriers - the RMW synchronizes for you. But if you do some kind of more weakly ordered primitive, then you must force a barrier there.

This is the exact same issue that I've run into before and forgot about again :

cbloom rants 07-31-11 - An example that needs seq_cst -
cbloom rants 08-09-11 - Threading Links (see threads on eventcount)
cbloom rants 06-01-12 - On C++ Atomic Fences Part 3

4/04/2013

04-04-13 - Tabdir

I made a fresh build of "tabdir" my old ( old ) utility that does a recursive dirlisting in tabbed-text format for "tabview".

Download : tabdir 320k zip

tabdir -?
usage : tabdir [opts] [dir]
options:
 -v  : view output after writing
 -p  : show progress of dir enumeration (with interactive keys)
 -w# : set # of worker threads
 -oN : output to name N [r:\tabdir.tab]

This new tabdir is built on Oodle so it has a multi-threaded dir lister for much greater speed. (*)

Also note to self : I fixed tabview so it works as a shell file association. I hit this all the time and always forget it : if something works on the command line but not as a shell association, it's probably because the shell passes you quotes around file names, so you need a little code to strip quotes from args.

Someday I'd like to write an even faster tabdir that reads the NTFS volume directory information directly, but chances are that will never happen.

One odd thing I've spotted with this tabdir is that the Windows SxS Assembly dirs take a ton of time to enumerate on my machine. I dunno if they're compressed or WTF the deal is with them (I pushed it on the todo stack to investigate), but they're like 10X slower than any other dir. (could just be the larger number of files in there; but I mean it's slow *per file*)

I never did this before because I didn't expect multi-threaded dir enumeration to be a big win; I thought it would just cause seek thrashing, and if you're IO bound anyway then multi-threading can't help, can it? Well, it turns out the Win32 dir enum functions have quite a lot of CPU overhead, so multi-threading does in fact help a bit :

nworkers| elapsed time
1       | 12.327
2       | 10.450
3       | 9.710
4       | 9.130

(* = actually the big speed win was not multi-threading, it's that the old tabdir did something rather dumb in the file enum. It would enum all files, and then do GetInfo() on each one to get the file sizes. The new one just uses the file infos that are returned as part of the Win32 enumeration, which is massively faster).

04-04-13 - Worker Thread Forward Permit Delay-Kicks

I've had this small annoying issue for a while, and finally thought of a pretty simple solution.

You may recall, I use a worker thread system with forward "permits" (reversed dependencies) . When any handle completes it sees if that completion should trigger any followup handles, and if so those are then launched. Handles may be SPU jobs or IOs or CPU jobs or whatever. The problem I will talk about occurred when the predessor and the followup were both CPU jobs.

I'll talk about a specific case to be concrete : decoding compressed data while reading it from disk.

To decode each chunk of LZ data, a chunk-decompress job is made. That job depends on the IO(s) that read in the compressed data for that chunk. It also depends on the previous chunk if the chunk is not a seek-reset point. So in the case of a non-reset chunk, you have a dependency on an IO and a previous CPU job. Your job will be started by one or the other, whichever finishes last.

Now, when decompression was IO bound, then the IO completions were kicking off the decompress jobs, and everything was fine.

In these timelines, the second line is IO and the bottom four are workers. (click images for higher res)

LZ Read and Decompress without seek-resets, IO bound :

You can see the funny fans of lines that show the dependency on the previous decompress job and also the IO. Yellow is a thread that's sleeping.

You may notice that the worker threads are cycling around. That's not really ideal, but it's not related to the problem I'm talking about today. (that cycling is caused by the fact that the OS semaphore is FIFO. For something like worker threads, we'd actually rather have a LIFO semaphore, because it makes it more likely that you get a thread with something useful still hot in cache. Someday I'll replace my OS semaphore with my own LIFO one, but for now this is a minor performance bug). (Win32 docs say that they don't gaurantee any particular order, but in my experience threads of equal priority are always FIFO in Win32 semaphores)

Okay, now for the problem. When the IO was going fast, so we were CPU bound, it's the prior decompress job that triggers the followup work.

But something bad happened due to the forward permit system. The control flow was something like this :


On worker thread 0

wake from semaphore
do on an LZ decompress job
mark job done
completion change causes a permits check
permits check sees that there is a pending job triggered by this completion
  -> fire off that handle
   handle is pushed to worker thread system
   no worker is available to do it, so wake a new worker and give him the job
finalize (usually delete) job I just finished
look for more work to do
   there is none because it was already handed to a new worker

And it looked like this :

LZ Read and Decompress without seek-resets, CPU bound, naive permits :

You can see each subsequent decompress job is moving to another worker thread. Yuck, bad.

So the fix in Oodle is to use the "delay-kick" mechanism, which I'd already been using for coroutine refires (which had a similar problem; the problem occurred when you yielded a coroutine on something like an IO, and the IO was done almost immediately; the coroutine would get moved to another worker thread instead of just staying on the same one and continuing from the yield as if it wasn't there).

The scheme is something like this :


On each worker thread :

Try to pop a work item from the "delay kick queue"
  if there is more than one item in the DKQ,
    take one for myself and "kick" the remainder
    (kick means wake worker threads to do the jobs)

If nothing on DKQ, pop from the main queue
  if nothing on main queue, wait on work semaphore

Do your job

Set "delay kick" = true
  ("delay kick" has to be in TLS of course)
Mark job as done
Permit system checks for successor handles that can now run 
  if they exist, they are put in the DKQ instead of immediately firing
Set "delay kick" = false

Repeat

In brief : work that is made runnable by the completion of work is not fired until the worker thread that did the completion gets its own shot at grabbing that new work. If the completion made 4 jobs runnable, the worker will grab 1 for itself and kick the other 3. The kick is no longer in the completion phase, it's in the pop phase.

And the result is :

LZ Read and Decompress without seek-resets, CPU bound, delay-kick permits :

Molto superiore.

These last two timelines are on the same time scale, so you can see just from the visual that eliminating the unnecessary thread switching is about a 10% speedup.

Anyway, this particular issue may not apply to your worker thread system, or you may have other solutions. I think the main take-away is that while worker thread systems seem very simple to write at first, there's actually a huge amount of careful fiddling required to make them really run well. You have to be constantly vigilant about doing test runs and checking threadprofile views like this to ensure that what's actually happening matches what you think is happening. Err, oops, I think I just accidentally wrote an advertisement for Telemetry .

04-04-13 - Oodle Compression on BC1 and WAV

I put some stupid filters in, so here are some results for the record and my reference.

BC1 (DXT1/S3TC) DDS textures :

All compressors run in max-compress mode. Note that it's not entirely fair because Oodle has the BC1 swizzle and the others don't.

Some day I'd like to do a BC1-specific encoder. Various ideas and possibilities there. Also RD-DXTC.

I also did a WAV filter. This one is particularly ridiculous because nobody uses WAV, and if you want to compress audio you should use a domain-specific compressor, not just OodleLZ with a simple delta filter. I did it because I was annoyed that RAR beat me on WAVs (due to its having a multimedia filter), and RAR should never beat me.

WAV compression :

See also : same chart with 7z (not really fair cuz 7z doesn't have a WAV filter)

Happy to see that Oodle-filt handily beats RAR-filt. I'm using just a trivial linear gradient predictor :


out[i] = in[i] - 2*in[i-1] + in[i-2]

this could surely be better, but whatever, WAV filtering is not important.

I also did a simple BMP delta filter and EXE (BCJ/relative-call) transform. I don't really want to get into the business of offering all kinds of special case filters the way some of the more insane modern archivers do (like undoing ZLIB compression so you can recompress it, or WRT), but anyhoo there's a few.


ADDED : I will say something perhaps useful about the WAV filter.

There's a bit of a funny issue because the WAV data is 16 bit (or 24 or 32), and the back-end entropy coder in a simple LZ is 8 bit.

If you just take a 16-bit delta and put it into bytes, then most of your values will be around zero, and you'll make a stream like :

[00 00] [00 01] [FF FF] [FF F8] [00 04] ...
The bad thing you should notice here are the high bytes are switching between 00 and FF even though the values have quite a small range. (Note that the common thing of centering the values with +32768 doesn't change this at all).

You can make this much better just by doing a bias of +128. That makes it so the most important range of values (around zero (specifically [-128,127])) all have the same top byte.

I think it might be even slightly better to do a "folded" signed->unsigned map, like

{ 0,-1,1,-2,2,-3,...,32767,-32768 }
The main difference being that values like -129 and +128 get the same high byte in this mapping, rather than two different high bytes in the simple +128 bias scheme.

Of course you really want a separate 8-bit huffman for alternating pairs of bytes. One way to get that is to use a few bottom bits of position as part of the literal context. Also, the high byte should really be used as context for the low byte. But both of those are beyond the capabilities of my simple LZ-huffs so I just deinterleave the high and low bytes to two streams.

4/02/2013

04-02-13 - The Method of Random Holdouts

Quick and dirty game-programming style hacky way to make fitting model parameters somewhat better.

You have some testset {T} of many items, and you wish to fit some heuristic model M over T which has some parameters. There may be multiple forms of the model and you aren't sure which is best, so you wish to compare models against each other.

For concreteness, you might imagine that T is a bunch of images, and you are trying to make a perceptual DXTC coder; you measure block error in the encoder as something like (SSD + a * SATD ^ b + c * SSIM_8x8 ) , and the goal is to minimize the total image error in the outer loop, measured using something complex like IW-MS-SSIM or "MyDCTDelta" or whatever. So you are trying to fit the parameters {a,b,c} to minimize an error.

For reference, the naive training method is : run the model on all data in {T}, optimize parameters to minimize error over {T}.

The method of random holdouts goes like this :


Run many trials

On each trial, take the testset T and randomly separate it into a training set and a verification set.
Typically training set is something like 75% of the data and verification is 25%.

Optimize the model parameters on the {training set} to minimize the error measure over {training set}.

Now run the optimized model on the {verification set} and measure the error there.
This is the error that will be used to rate the model.

When you make the average error, compensate for the size of the model thusly :
average_error = sum_error / ( [num in {verification set}] - [dof in model] )

Record the optimal parameters and the error for that trial

Now you have optimized parameters for each trial, and an error for each trial. You can take the average over all trials, but you can also take the sdev. The sdev tells you how well your model is really working - if it's not close to zero then you are missing something important in your model. A term with a large sdev might just be a random factor that's not useful in the model, and you should try again without it.

The method of random holdouts reduces over-training risk, because in each run you are measuring error only on data samples that were not used in training.

The method of random holdouts gives you a decent way to compare models which may have different numbers of DOF. If you just use the naive method of training, then models with more DOF will always appear better, because they are just fitting your data set.

That is, in our example say that (SSD + a * SATD ^ b) is actually the ideal model and the extra term ( + c * SSIM_8x8 ) is not useful. As long as it's not just a linear combo of other terms, then naive training will find a "c" such that that term is used to compensate for variations in your particular testset. And in fact that incorrect "c" can be quite a large value (along with a negative "a").

This kind of method can also be used for fancier stuff like building complex models from ensembles of simple models, "boosting" models, etc. But it's useful even in this case where we wind up just using a simple linear model, because you can see how it varies over the random holdouts.