Working on a problem that requires a GA. I have all of that working and spent quite a bit of time trimming the fat and optimizing the code before resorting to compiler optimization. Because the GA runs as a result of user input it has to find a solution within a reasonable period of time, otherwise the UI stalls and it just won't play well at all. I got this binary GA solving a 27 variable problem in about 0.1s on an iPhone 3GS.
In order to achieve this level of performance the entire GA was coded in C, not Objective-C.
In a search for a further reduction of run time I was considering the idea of using the "-O3" optimization switch only for the solver module. I tried it and it cut run time nearly in half.
Should I be concerned about any gotcha's by setting optimization to "-O3"? Keep in mind that I am doing this at the file level and not for the entire project.
-O3 flag will make the code work in the same way as before (only faster), as long as you don't do any tricky stuff that is unsafe or otherwise dependant on what the compiler does to it.
Also, as suggested in the comments, it might be a good idea to let the computation run in a separate thread to prevent the UI from locking up. That also gives you the flexibility to make the computation more expensive, or to display a progress bar, or whatever.
Tricky stuff
Optimisation will produce unexpected results if you try to access stuff on the stack directly, or move the stack pointer somewhere else, or if you do something inherently illegal, like forget to initialise a variable (some compilers (MinGW) will set them to 0).
E.g.
int main() {
int array[10];
array[-2] = 13; // some negative value might get the return address
}
Some other tricky stuff involves the optimiser stuffing up by itself. Here's an example of when -O3 completely breaks the code.
Related
has anyone experience the following issue?
A stack variable getting changed/corrupted after calling ne10 assembly function such as ne10_len_vec2f_neon?
e.g
float gain = 8.0;
ne10_len_vec2f_neon(src, dst, len);
after the call to ne10_len_vec2f_neon, the value of gain changes as its memory is getting corrupted.
1. Note this only happens when the project is compiled in release build but not debug build.
2. Does Ne10 assembly functions preserve registers?
3. Replacing the assembly function call to c equivalent such as ne10_len_vec2f_c and both release and debug build seem to work OK.
thanks for any help on this. Not sure if there's an inherent issue within the program or it is really the call to ne10_len_vec2f_neon causing the corruption with release build.enter code here
I had a quick rummage through the master NEON code here:
https://github.com/projectNe10/Ne10/blob/master/modules/math/NE10_len.neon.s
... and it doesn't really touch address-based stack at all, so not sure it's a stack problem in memory.
However based on what I remember of the NEON procedure call standard q4-q7 (alias d8-d15 or s16-s31) should be preserved by the callee, and as far as I can tell that code is clobbering q4-6 without the necessary save/restore, so it does indeed look like it's clobbering the stack in registers.
In the failed case do you know if gain is still stored in FPU registers, and if yes which ones? If it's stored in any of s16/17/18/19 then this looks like the problem. It also seems plausible that a compiler would choose to use s16 upwards for things it needs to keep across a function call, as it avoids the need to touch in-RAM stack memory.
In terms of a fix, if you perform the following replacements:
s/q4/q8/
s/q5/q9/
s/q6/q10/
in that file, then I think it should work; no means to test here, but those higher register blocks are not callee saved.
For example, if I implement some simple object caching, which method is faster?
1. return isset($cache[$cls]) ? $cache[$cls] : $cache[$cls] = new $cls;
2. return #$cache[$cls] ?: $cache[$cls] = new $cls;
I read somewhere # takes significant time to execute (and I wonder why), especially when warnings/notices are actually being issued and suppressed. isset() on the other hand means an extra hash lookup. So which is better and why?
I do want to keep E_NOTICE on globally, both on dev and production servers.
I wouldn't worry about which method is FASTER. That is a micro-optimization. I would worry more about which is more readable code and better coding practice.
I would certainly prefer your first option over the second, as your intent is much clearer. Also, best to keep away edge condition problems by always explicitly testing variables to make sure you are getting what you are expecting to get. For example, what if the class stored in $cache[$cls] is not of type $cls?
Personally, if I typically would not expect the index on $cache to be unset, then I would also put error handling in there rather than using ternary operations. If I could reasonably expect that that index would be unset on a regular basis, then I would make class $cls behave as a singleton and have your code be something like
return $cls::get_instance();
The isset() approach is better. It is code that explicitly states the index may be undefined. Suppressing the error is sloppy coding.
According to this article 10 Performance Tips to Speed Up PHP, warnings take additional execution time and also claims the # operator is "expensive."
Cleaning up warnings and errors beforehand can also keep you from
using # error suppression, which is expensive.
Additionally, the # will not suppress the errors with respect to custom error handlers:
http://www.php.net/manual/en/language.operators.errorcontrol.php
If you have set a custom error handler function with
set_error_handler() then it will still get called, but this custom
error handler can (and should) call error_reporting() which will
return 0 when the call that triggered the error was preceded by an #.
If the track_errors feature is enabled, any error message generated by
the expression will be saved in the variable $php_errormsg. This
variable will be overwritten on each error, so check early if you want
to use it.
# temporarily changes the error_reporting state, that's why it is said to take time.
If you expect a certain value, the first thing to do to validate it, is to check that it is defined. If you have notices, it's probably because you're missing something. Using isset() is, in my opinion, a good practice.
I ran timing tests for both cases, using hash keys of various lengths, also using various hit/miss ratios for the hash table, plus with and without E_NOTICE.
The results were: with error_reporting(E_ALL) the isset() variant was faster than the # by some 20-30%. Platform used: command line PHP 5.4.7 on OS X 10.8.
However, with error_reporting(E_ALL & ~E_NOTICE) the difference was within 1-2% for short hash keys, and up 10% for longer ones (16 chars).
Note that the first variant executes 2 hash table lookups, whereas the variant with # does only one lookup.
Thus, # is inferior in all scenarios and I wonder if there are any plans to optimize it.
I think you have your priorities a little mixed up here.
First of all, if you want to get a real world test of which is faster - load test them. As stated though suppressing will probably be slower.
The problem here is if you have performance issues with regular code, you should be upgrading your hardware, or optimize the grand logic of your code rather than preventing proper execution and error checking.
Suppressing errors to steal the tiniest fraction of a speed gain won't do you any favours in the long run. Especially if you think that this error may keep happening time and time again, and cause your app to run more slowly than if the error was caught and fixed.
Do the risks caused by bypassing Perl safe signals for example like shown in the second timeout example in the DBI documentation concern only the code that uses such bypassing?
The code in that example works hard to localize the change to just that section of code, or any code called from it.
There is not 100% guarantee that no code will be effected outside the code that bypasses safe signals, because signals are no longer safe. In the example the call being timed out is a DBI->connect. For most DBD's this will be implemented mostly in C, unless the C code can handle being aborted and tried again you might find that some data structures internal to the DBD, or the libraries it uses, are left in a inconstant state.
The chances of the example code going wrong is probably incredibly tiny. My personal anecdote on the issues is that I had used the traditional Perl signal handling for years before safe signals were introduced and for a long time I had never had a problem. I hadn't even been very cautious about what I did in my signal handlers. Then we managed to hit a data set that actually did trigger memory corruptions in about 1 out of ever 100 runs. Just modifying the signal handlers to use better practices, similar to those in the example, eliminated our issues.
What does that even mean? By using unsafe signals, you can corrupt Perl's internals and Perl variables. It can also cause problem if a non-reentrant C library call is interrupted.
This can lead to SEGFAULTs and other problems, and those may only manifest themselves outside the block where the timeout is in effect.
I've recently learned Haskell, and am trying to carry the pure functional style over to my other code when possible. An important aspect of this is treating all variables as immutable, i.e. constants. In order to do so, many computations that would be implemented using loops in an imperative style have to be performed using recursion, which typically incurs a memory penalty due to the allocation a new stack frame for each function call. In the special case of a tail call (where the return value of a called function is immediately returned to the callee's caller), however, this penalty can be bypassed by a process called tail call optimization (in one method, this can be done by essentially replacing a call with a jmp after setting up the stack properly). Does MATLAB perform TCO by default, or is there a way to tell it to?
If I define a simple tail-recursive function:
function tailtest(n)
if n==0; feature memstats; return; end
tailtest(n-1);
end
and call it so that it will recurse quite deeply:
set(0,'RecursionLimit',10000);
tailtest(1000);
then it doesn't look as if stack frames are eating a lot of memory. However, if I make it recurse much deeper:
set(0,'RecursionLimit',10000);
tailtest(5000);
then (on my machine, today) MATLAB simply crashes: the process unceremoniously dies.
I don't think this is consistent with MATLAB doing any TCO; the case where a function tail-calls itself, only in one place, with no local variables other than a single argument, is just about as simple as anyone could hope for.
So: No, it appears that MATLAB does not do TCO at all, at least by default. I haven't (so far) looked for options that might enable it. I'd be surprised if there were any.
In cases where we don't blow out the stack, how much does recursion cost? See my comment to Bill Cheatham's answer: it looks like the time overhead is nontrivial but not insane.
... Except that Bill Cheatham deleted his answer after I left that comment. OK. So, I took a simple iterative implementation of the Fibonacci function and a simple tail-recursive one, doing essentially the same computation in both, and timed them both on fib(60). The recursive implementation took about 2.5 times longer to run than the iterative one. Of course the relative overhead will be smaller for functions that do more work than one addition and one subtraction per iteration.
(I also agree with delnan's sentiment: highly-recursive code of the sort that feels natural in Haskell is typically likely to be unidiomatic in MATLAB.)
There is a simple way to check this. Create this function tail_recursion_check:
function r = tail_recursion_check(n)
if n > 1
r = tail_recursion_check(n - 1);
else
error('error');
end
end
and run tail_recursion_check(10), for example. You are going to see a very long stack trace with 10 items that says error at line 3. If there were tail call optimization, you would only see one.
I've been making some progress with audio programming for iPhone. Now I'm doing some performance tuning, trying to see if I can squeeze more out of this little machine. Running Shark, I see that a significant part of my cpu power (16%) is getting eaten up by objc_msgSend. I understand I can speed this up somewhat by storing pointers to functions (IMP) rather than calling them using [object message] notation. But if I'm going to go through all this trouble, I wonder if I might just be better off using C++.
Any thoughts on this?
Objective C is absolutely fast enough for DSP/audio programming, because Objective C is a superset of C. You don't need to (and shouldn't) make everything a message. Where performance is critical, use plain C function calls (or use inline assembly, if there are hardware features you can leverage that way). Where performance isn't critical, and your application can benefit from the features of message indirection, use the square brackets.
The Accelerate framework on OS X, for example, is a great high-performance Objective C library. It only uses standard C99 function calls, and you can call them from Objective C code without any wrapping or indirection.
The problem with Objective-C and functions like DSP is not speed per se but rather the uncertainty of when the inevitable bottlenecks will occur.
All languages have bottlenecks but in static linked languages like C++ you can better predict when and where in the code they will occur. In the case of Objective-C's runtime coupling, the time it takes to find the appropriate object, the time it takes to send a message is not necessary slow but it is variable and unpredictable. Objective-C's flexibility in UI, data management and reuse work against it in the case of tightly timed task.
Most audio processing in the Apple API is done in C or C++ because of the need to nail down the time it takes code to execute. However, its easy to mix Objective-C, C and C++ in the same app. This allows you to pick the best language for the immediate task at hand.
Is Objective C fast enough for DSP/audio programming
Real Time Rendering
Definitely Not. The Objective-C runtime and its libraries are simply not designed for the demands of real time audio rendering. The fact is, it's virtually impossible to guarantee that using ObjC runtime or libraries such as Foundation (or even CoreFoundation) will not result your renderer missing its deadline.
The common case is a lock -- even a simple heap allocation (malloc, new/new[], [[NSObject alloc] init]) will likely require a lock.
To use ObjC is to utilize libraries and a runtime which assume locks are acceptable at any point within their execution. The lock can suspend execution of your render thread (e.g. during your render callback) while waiting to acquire the lock. Then you can miss your render deadline because your render thread is held up, ultimately resulting in dropouts/glitches.
Ask a pro audio plugin developer: they will tell you that blocking within the realtime render domain is forbidden. You cannot e.g. run to the filesystem or create heap allocations because you have no practical upper bound regarding the time it will take to finish.
Here's a nice introduction: http://www.rossbencina.com/code/real-time-audio-programming-101-time-waits-for-nothing
Offline Rendering
Yes, it would be acceptably fast in most scenarios for high level messaging. At the lower levels, I recommend against using ObjC because it would be wasteful -- it could take many, many times longer to render if ObjC messaging used at that level (compared to a C or C++ implementation).
See also: Will my iPhone app take a performance hit if I use Objective-C for low level code?
objc_msgSend is just a utility.
The cost of sending a message is not just the cost of sending the message.
It is the cost of doing everything that the message initiates.
(Just like the true cost of a function call is its inclusive cost, including I/O if there is any.)
What you need to know is where are the time-dominant messages coming from and going to and why.
Stack samples will tell you which routines / methods are being called so often that you should figure out how to call them more efficiently.
You may find that you're calling them more than you have to.
Especially if you find that many of the calls are for creating and deleting data structure, you can probably find better ways to do that.