EDIT: I realized that I, unfortunately, overlooked a semicolon at the end of the while statement in the first example code and misinterpreted it myself. So there is in fact an empty loop for threads with threadIdx.x != s, a convergency point after that loop and a thread waiting at this point for all the others without incrementing the s variable. I am leaving the original (uncorrected) question below for anyone interested in it. Be aware, that there is a semicolon missing at the end of the second line in the first example and thus, s++ has nothing in common with the cycle body.
--
We were studying serialization in our CUDA lesson and our teacher told us that a code like this:
__shared__ int s = 0;
while (s != threadIdx.x)
s++; // serialized code
would end up with a HW deadlock because the nvcc compiler puts a reconvergence point between the while (s != threadIdx.x) and s++ statements. If I understand it correctly, this means that once the reconvergence point is reached by a thread, this thread stops execution and waits for the other threads until they reach the point too. In this example, however, this never happens, because thread #0 enters the body of the while loop, reaches the reconvergence point without incrementing the s variable and other threads get stuck in an endless loop.
A working solution should be the following:
__shared__ int s = 0;
while (s < blockDim.x)
if (threadIdx.x == s)
s++; // serialized code
Here, all threads within a block enter the body of the loop, all evaluate the condition and only thread #0 increments the s variable in the first iteration (and loop goes on).
My question is, why does the second example work if the first hangs? To be more specific, the if statement is just another point of divergence and in terms of the Assembler language should be compiled into the same conditional jump instruction as the condition in the loop. So why isn't there any reconvergence point before s++ in the second example and has it in fact gone immediately after the statement?
In other sources I have only found that a divergent code is computed independently for every branch - e.g. in an if/else statement, first the if branch is computed with all else-branched threads masked within the same warp and then the other threads compute the else branch while the first wait. There's a reconvergence point after the if/else statement. Why then does the first example freeze, not having the loop split into two branches (a true branch for one thread and a waiting false branch for all the others in a warp)?
Thank you.
It does not make sense to put the reconvergence point between the call to while (s != threadIdx.x) and s++;. It disrupts the program flow since the reconvergence point for a piece of code should be reachable by all threads at compile time. Below picture shows the flowchart of your first piece of code and possible and impossible points of reconvergence.
Regarding this answer about recording the convergence point via SSY instruction, I created below simple kernel resembling your first piece of code
__global__ void kernel_1() {
__shared__ int s;
if(threadIdx.x==0)
s = 0;
__syncthreads();
while (s == threadIdx.x)
s++; // serialized code
}
and compiled it for CC=3.5 with -O3. Below is the result of using cuobjdumbinary tool for the output to observe the CUDA assembly. The result is:
I'm not an expert in reading CUDA assembly but I can see while loop condition checks in lines 0038 and 00a0. At line 00a8, it branches to 0x80 if it satisfies the while loop condition and executes the code block again. The introduction of the reconvergence point is at line 0058 introducing line 0xb8 as the reconvergence point which is after the loop condition check near the exit.
Overall, it is not clear what you're trying to achieve with this piece of code. Also in the second piece of code, the reconvergence point should be again after while loop code block (I don't mean between while and if).
The reason why it "hangs" is neither a HW deadlock nor branching, at least not directly. You produce an endless loop for one or multiple threads (as already suspected).
In your example, there isn't really a convergence point. Since you do not use any synchronization, there aren't any threads that actually wait. What happens here with the while-loop is pretty much a busy-wait.
A kernel only finishes if all threads return. Since you have one (or multiple) endless loops (by accident maybe even none - this is unlikely however) the kernel will never finish.
You declared a shared variable s. This variable is known to all threads within a block.
With your while-statement you basically say (to each thread): increment s until it reaches the value of your (local) thread id. Since all threads are incrementing s in parallel, you introduce race conditions.
Example:
List item
Thread 5 is looping and checking for s to become 5
s is 4
Two threads increment s, it becomes 6
At the same time thread 5 only reached the end of its loop.
Now it reaches the next loop iteration and checks for s and it's not 5.
Thread 5 will never be able to finish since you check via == and the value of s already exceeded the value of the thread id.
Also your solution is quite confusing, because each thread executes the serialized code consecutively (which probably was the intention after all - even though that actually is strange):
Thread 0 will execute the serialized code
After that, thread 1 will execute the serialized code
and so on
Most examples show a program where each thread works on some code, then all threads are synchronized and only single thread executes some more code (maybe it needed the results of all threads).
So, your second example "works" because no thread is stuck in an endless loop, however I can't think of a reason why anyone would use such a code,
since it is confusing and, well, not parallel at all.
I'm trying to incorporate error checking within a pure procedure I am writing. I would like something like:
pure real function func1(output_unit,a)
implicit none
integer :: a, output_unit
if (a < 0) then
write(output_unit,*) 'Error in function func1: argument must be a nonnegative integer. It is ', a
else
func1 = a/3
endif
return
end function func1
However, pure functions are not allowed to have IO statements to external files, so I tried passing a unit number to the function, e.g. output_unit = 6, which is the default output. gfortran still regards this as illegal. Is there a way around this? Is it possible to make the function a derived type (instead of intrinsic type real here) which outputs a string when there is an error?
You are not the first person to have this problem, and I'm happy to say that this flaw in the standard will be remedied in Fortran 2015. As stated in this document (page 6, header "Approved changes to the standard"), "the restriction on the appearance of an error stop statement in a pure procedure should be removed".
The Fortran 2008 standard included the error stop statement in the context of some new parallel computing features. It signals an error and makes all processes stop as soon as is practicable. Currently, neither stop nor error stop statements are allowed in pure procedures, because they're obviously not thread-safe. In practice this is unnecessarily restrictive in cases where an internal error occurs.
Depending on your compiler, you may have to wait patiently for the implementation. I know that Intel has implemented it in their ifort compiler. ("F2015: Lift restriction on STOP and ERROR STOP in PURE/ELEMENTAL procedures")
alternative
For an alternative approach, you could have a look at this question, though in you case this is probably slightly trickier as you have to change the do concurrent keyword, not just pure.
(end of proper answer)
if getting dirty hands is an option ...
In the meantime you could do something brutal like
pure subroutine internal_error(error_msg)
! Try hard to produce a runtime error, regardless of compiler flags.
! This is useful in pure subprograms where you want to produce an error,
! preferably with a traceback.
!
! Though far from pretty, this solution contains all the ugliness in this
! single subprogram.
!
! TODO: replace with ERROR STOP when supported by compiler
implicit none
character(*), intent(in) :: error_msg
integer, dimension(:), allocatable :: molested
allocate(molested(2))
allocate(molested(2))
molested(3) = molested(4)
molested(1) = -10
molested(2) = sqrt(real(molested(1)))
deallocate(molested)
deallocate(molested)
molested(3) = molested(-10)
end subroutine internal_error
Should anyone ask, you didn't get this from me.
I've found an answer myself, detailed here. It uses what is considered "obsolescent", but still does the trick; it is called alternate return. Write the procedure as a subroutine as it doesn't work on functions.
pure real subroutine procA(arg1)
implicit none
integer :: arg1
if (arg < 0) then
return 1 ! exit the function and go to the first label supplied
! when function was called. Also return 2, 3 etc.
else
procA = ... ! whatever it should do under normal circumstances
endif
endsubroutine procA
....
! later on, procedure is called
num = procA(a, *220)
220 write(6,*) 'Error with func1: you've probably supplied a negative argument'
What would probably be better is what eriktous suggested--get the procedure to return a status, perhaps as a logical value or an integer, and get the program to check this value every time after it calls the procedure. If all's well, carry on. Otherwise, print a relevant error message.
Comments welcome.
I often use this code pattern:
while(true) {
//do something
if(<some condition>) {
break;
}
}
Another programmer told me that this was bad practice and that I should replace it with the more standard:
while(!<some condition>) {
//do something
}
His reasoning was that you could "forget the break" too easily and have an endless loop. I told him that in the second example you could just as easily put in a condition which never returned true and so just as easily have an endless loop, so both are equally valid practices.
Further, I often prefer the former as it makes the code easier to read when you have multiple break points, i.e. multiple conditions which get out of the loop.
Can anyone enrichen this argument by adding evidence for one side or the other?
There is a discrepancy between the two examples. The first will execute the "do something" at least once every time even if the statement is never true. The second will only "do something" when the statement evaluates to true.
I think what you are looking for is a do-while loop. I 100% agree that while (true) is not a good idea because it makes it hard to maintain this code and the way you are escaping the loop is very goto esque which is considered bad practice.
Try:
do {
//do something
} while (!something);
Check your individual language documentation for the exact syntax. But look at this code, it basically does what is in the do, then checks the while portion to see if it should do it again.
To quote that noted developer of days gone by, Wordsworth:
...
In truth the prison, unto which we doom
Ourselves, no prison is; and hence for me,
In sundry moods, 'twas pastime to be bound
Within the Sonnet's scanty plot of ground;
Pleased if some souls (for such their needs must be)
Who have felt the weight of too much liberty,
Should find brief solace there, as I have found.
Wordsworth accepted the strict requirements of the sonnet as a liberating frame, rather than as a straightjacket. I'd suggest that the heart of "structured programming" is about giving up the freedom to build arbitrarily-complex flow graphs in favor of a liberating ease of understanding.
I freely agree that sometimes an early exit is the simplest way to express an action. However, my experience has been that when I force myself to use the simplest possible control structures (and really think about designing within those constraints), I most often find that the result is simpler, clearer code. The drawback with
while (true) {
action0;
if (test0) break;
action1;
}
is that it's easy to let action0 and action1 become larger and larger chunks of code, or to add "just one more" test-break-action sequence, until it becomes difficult to point to a specific line and answer the question, "What conditions do I know hold at this point?" So, without making rules for other programmers, I try to avoid the while (true) {...} idiom in my own code whenever possible.
When you can write your code in the form
while (condition) { ... }
or
while (!condition) { ... }
with no exits (break, continue, or goto) in the body, that form is preferred, because someone can read the code and understand the termination condition just by looking at the header. That's good.
But lots of loops don't fit this model, and the infinite loop with explicit exit(s) in the middle is an honorable model. (Loops with continue are usually harder to understand than loops with break.) If you want some evidence or authority to cite, look no further than Don Knuth's famous paper on Structured Programming with Goto Statements; you will find all the examples, arguments, and explanations you could want.
A minor point of idiom: writing while (true) { ... } brands you as an old Pascal programmer or perhaps these days a Java programmer. If you are writing in C or C++, the preferred idiom is
for (;;) { ... }
There's no good reason for this, but you should write it this way because this is the way C programmers expect to see it.
I prefer
while(!<some condition>) {
//do something
}
but I think it's more a matter of readability, rather than the potential to "forget the break." I think that forgetting the break is a rather weak argument, as that would be a bug and you'd find and fix it right away.
The argument I have against using a break to get out of an endless loop is that you're essentially using the break statement as a goto. I'm not religiously against using goto (if the language supports it, it's fair game), but I do try to replace it if there's a more readable alternative.
In the case of many break points I would replace them with
while( !<some condition> ||
!<some other condition> ||
!<something completely different> ) {
//do something
}
Consolidating all of the stop conditions this way makes it a lot easier to see what's going to end this loop. break statements could be sprinkled around, and that's anything but readable.
while (true) might make sense if you have many statements and you want to stop if any fail
while (true) {
if (!function1() ) return;
if (!function2() ) return;
if (!function3() ) return;
if (!function4() ) return;
}
is better than
while (!fail) {
if (!fail) {
fail = function1()
}
if (!fail) {
fail = function2()
}
........
}
Javier made an interesting comment on my earlier answer (the one quoting Wordsworth):
I think while(true){} is a more 'pure' construct than while(condition){}.
and I couldn't respond adequately in 300 characters (sorry!)
In my teaching and mentoring, I've informally defined "complexity" as "How much of the rest of the code I need to have in my head to be able to understand this single line or expression?" The more stuff I have to bear in mind, the more complex the code is. The more the code tells me explicitly, the less complex.
So, with the goal of reducing complexity, let me reply to Javier in terms of completeness and strength rather than purity.
I think of this code fragment:
while (c1) {
// p1
a1;
// p2
...
// pz
az;
}
as expressing two things simultaneously:
the (entire) body will be repeated as long as c1 remains true, and
at point 1, where a1 is performed, c1 is guaranteed to hold.
The difference is one of perspective; the first of these has to do with the outer, dynamic behavior of the entire loop in general, while the second is useful to understanding the inner, static guarantee which I can count on while thinking about a1 in particular. Of course the net effect of a1 may invalidate c1, requiring that I think harder about what I can count on at point 2, etc.
Let's put a specific (tiny) example in place to think about the condition and first action:
while (index < length(someString)) {
// p1
char c = someString.charAt(index++);
// p2
...
}
The "outer" issue is that the loop is clearly doing something within someString that can only be done as long as index is positioned in the someString. This sets up an expectation that we'll be modifying either index or someString within the body (at a location and manner not known until I examine the body) so that termination eventually occurs. That gives me both context and expectation for thinking about the body.
The "inner" issue is that we're guaranteed that the action following point 1 will be legal, so while reading the code at point 2 I can think about what is being done with a char value I know has been legally obtained. (We can't even evaluate the condition if someString is a null ref, but I'm also assuming we've guarded against that in the context around this example!)
In contrast, a loop of the form:
while (true) {
// p1
a1;
// p2
...
}
lets me down on both issues. At the outer level, I am left wondering whether this means that I really should expect this loop to cycle forever (e.g. the main event dispatch loop of an operating system), or whether there's something else going on. This gives me neither an explicit context for reading the body, nor an expectation of what constitutes progress toward (uncertain) termination.
At the inner level, I have absolutely no explicit guarantee about any circumstances that may hold at point 1. The condition true, which is of course true everywhere, is the weakest possible statement about what we can know at any point in the program. Understanding the preconditions of an action are very valuable information when trying to think about what the action accomplishes!
So, I suggest that the while (true) ... idiom is much more incomplete and weak, and therefore more complex, than while (c1) ... according to the logic I've described above.
The problem is that not every algorithm sticks to the "while(cond){action}" model.
The general loop model is like this :
loop_prepare
loop:
action_A
if(cond) exit_loop
action_B
goto loop
after_loop_code
When there is no action_A you can replace it by :
loop_prepare
while(cond)
action_B
after_loop_code
When there is no action_B you can replace it by :
loop_prepare
do action_A
while(cond)
after_loop_code
In the general case, action_A will be executed n times and action_B will be executed (n-1) times.
A real life example is : print all the elements of a table separated by commas.
We want all the n elements with (n-1) commas.
You always can do some tricks to stick to the while-loop model, but this will always repeat code or check twice the same condition (for every loops) or add a new variable. So you will always be less efficient and less readable than the while-true-break loop model.
Example of (bad) "trick" : add variable and condition
loop_prepare
b=true // one more local variable : more complex code
while(b): // one more condition on every loop : less efficient
action_A
if(cond) b=false // the real condition is here
else action_B
after_loop_code
Example of (bad) "trick" : repeat the code. The repeated code must not be forgotten while modifying one of the two sections.
loop_prepare
action_A
while(cond):
action_B
action_A
after_loop_code
Note : in the last example, the programmer can obfuscate (willingly or not) the code by mixing the "loop_prepare" with the first "action_A", and action_B with the second action_A. So he can have the feeling he is not doing this.
The first is OK if there are many ways to break from the loop, or if the break condition cannot be expressed easily at the top of the loop (for example, the content of the loop needs to run halfway but the other half must not run, on the last iteration).
But if you can avoid it, you should, because programming should be about writing very complex things in the most obvious way possible, while also implementing features correctly and performantly. That's why your friend is, in the general case, correct. Your friend's way of writing loop constructs is much more obvious (assuming the conditions described in the preceding paragraph do not obtain).
There's a substantially identical question already in SO at Is WHILE TRUE…BREAK…END WHILE a good design?. #Glomek answered (in an underrated post):
Sometimes it's very good design. See Structured Programing With Goto Statements by Donald Knuth for some examples. I use this basic idea often for loops that run "n and a half times," especially read/process loops. However, I generally try to have only one break statement. This makes it easier to reason about the state of the program after the loop terminates.
Somewhat later, I responded with the related, and also woefully underrated, comment (in part because I didn't notice Glomek's the first time round, I think):
One fascinating article is Knuth's "Structured Programming with go to Statements" from 1974 (available in his book 'Literate Programming', and probably elsewhere too). It discusses, amongst other things, controlled ways of breaking out of loops, and (not using the term) the loop-and-a-half statement.
Ada also provides looping constructs, including
loopname:
loop
...
exit loopname when ...condition...;
...
end loop loopname;
The original question's code is similar to this in intent.
One difference between the referenced SO item and this is the 'final break'; that is a single-shot loop which uses break to exit the loop early. There have been questions on whether that is a good style too - I don't have the cross-reference at hand.
Sometime you need infinite loop, for example listening on port or waiting for connection.
So while(true)... should not categorized as good or bad, let situation decide what to use
It depends on what you’re trying to do, but in general I prefer putting the conditional in the while.
It’s simpler, since you don't need another test in the code.
It’s easier to read, since you don’t have to go hunting for a break inside the loop.
You’re reinventing the wheel. The whole point of while is to do something as long as a test is true. Why subvert that by putting the break condition somewhere else?
I’d use a while(true) loop if I was writing a daemon or other process that should run until it gets killed.
If there's one (and only one) non-exceptional break condition, putting that condition directly into the control-flow construct (the while) is preferable. Seeing while(true) { ... } makes me as a code-reader think that there's no simple way to enumerate the break conditions and makes me think "look carefully at this and think about carefully about the break conditions (what is set before them in the current loop and what might have been set in the previous loop)"
In short, I'm with your colleague in the simplest case, but while(true){ ... } is not uncommon.
The perfect consultant's answer: it depends. Most cases, the right thing to do is either use a while loop
while (condition is true ) {
// do something
}
or a "repeat until" which is done in a C-like language with
do {
// do something
} while ( condition is true);
If either of these cases works, use them.
Sometimes, like in the inner loop of a server, you really mean that a program should keep going until something external interrupts it. (Consider, eg, an httpd daemon -- it isn't going to stop unless it crashes or it's stopped by a shutdown.)
THEN AND ONLY THEN use a while(1):
while(1) {
accept connection
fork child process
}
Final case is the rare occasion where you want to do some part of the function before terminating. In that case, use:
while(1) { // or for(;;)
// do some stuff
if (condition met) break;
// otherwise do more stuff.
}
I think the benefit of using "while(true)" is probably to let multiple exit condition easier to write especially if these exit condition has to appear in different location within the code block. However, for me, it could be chaotic when I have to dry-run the code to see how the code interacts.
Personally I will try to avoid while(true). The reason is that whenever I look back at the code written previously, I usually find that I need to figure out when it runs/terminates more than what it actually does. Therefore, having to locate the "breaks" first is a bit troublesome for me.
If there is a need for multiple exit condition, I tend to refactor the condition determining logic into a separate function so that the loop block looks clean and easier to understand.
No, that's not bad since you may not always know the exit condition when you setup the loop or may have multiple exit conditions. However it does require more care to prevent an infinite loop.
He is probably correct.
Functionally the two can be identical.
However, for readability and understanding program flow, the while(condition) is better. The break smacks more of a goto of sorts. The while (condition) is very clear on the conditions which continue the loop, etc. That doesn't mean break is wrong, just can be less readable.
A few advantages of using the latter construct that come to my mind:
it's easier to understand what the loop is doing without looking for breaks in the loop's code.
if you don't use other breaks in the loop code, there's only one exit point in your loop and that's the while() condition.
generally ends up being less code, which adds to readability.
I prefer the while(!) approach because it more clearly and immediately conveys the intent of the loop.
There has been much talk about readability here and its very well constructed but as with all loops that are not fixed in size (ie. do while and while) you run at a risk.
His reasoning was that you could "forget the break" too easily and have an endless loop.
Within a while loop you are in fact asking for a process that runs indefinitely unless something happens, and if that something does not happen within a certain parameter, you will get exactly what you wanted... an endless loop.
What your friend recommend is different from what you did. Your own code is more akin to
do{
// do something
}while(!<some condition>);
which always run the loop at least once, regardless of the condition.
But there are times breaks are perfectly okay, as mentioned by others. In response to your friend's worry of "forget the break", I often write in the following form:
while(true){
// do something
if(<some condition>) break;
// continue do something
}
By good indentation, the break point is clear to first time reader of the code, look as structural as codes which break at the beginning or bottom of a loop.
It's not so much the while(true) part that's bad, but the fact that you have to break or goto out of it that is the problem. break and goto are not really acceptable methods of flow control.
I also don't really see the point. Even in something that loops through the entire duration of a program, you can at least have like a boolean called Quit or something that you set to true to get out of the loop properly in a loop like while(!Quit)... Not just calling break at some arbitrary point and jumping out,
using loops like
while(1) { do stuff }
is necessary in some situations. If you do any embedded systems programming (think microcontrollers like PICs, MSP430, and DSP programming) then almost all your code will be in a while(1) loop. When coding for DSPs sometimes you just need a while(1){} and the rest of the code is an interrupt service routine (ISR).
If you loop over an external condition (not being changed inside the loop), you use while(t), where t is the condition. However, if the loop stops when the condition changes inside the loop, it's more convenient to have the exit point explicitly marked with break, instead of waiting for it to happen on the next iteration of the loop:
while (true) {
...
a := a + 1;
if (a > 10) break; // right here!
...
}
As was already mentioned in a few other answers, the less code you have to keep in your head while reading a particular line, the better.