I have a section of a Fortran90 program that should be parallelized with OpenMP.
!$omp parallel num_threads(8) &
!$omp private(j, s, prop_states) &
!$omp firstprivate(targets, pulses)
! ... modify something in pulses. targets(s)%ham contains pointers to
! elements of pulses ...
do s = 1, n_systems
prop_states(s) = targets(s)%psi_i
call prop(prop_states(s), targets(s)%grid, targets(s)%ham, &
& targets(s)%work, para)
end do
!$omp end parallel
What I'm unsure about is whether complex data structures can be private to each thread (and how this should be done -- is firstprivate correct?). In the example code above, targets is of a somewhat complicated user-defined type, with equally complex sub-fields. For example, targets(s)%ham%op(1)%pulse is a pointer to some element of an array pulses. Also, targets(s)%work contains allocated space to be used as work arrays in Fast-Fourier-Transforms.
Obviously, every thread needs to maintain an independent copy both of targets and of pulses, and maintain the pointers between the two independently. It seems to me that this might be asking a little bit too much from the automatic memory management of OpenMP. Is this correct, or should this work out of the box?
The alternative of course is to create copies of the original data within each thread (stored in an array), and use this private copied data in the call to prop.
From my reading of the OpenMP 2.5 standard you can't use the targets of Fortran pointers in private (or firstprivate or threadprivate) clauses, which seems to rule out your code. Having said that, it's not something I've ever tried in OpenMP so if you bash ahead and get anywhere, do let us know.
And firstprivate is correct if your private variables are to be initialised, on entry into the parallel region, with the value of the variables of the same name at the entry to the parallel region.
I guess you will probably have to implement your plan B.
Related
Say that I have a construct like this:
for(int i=0;i<5000;i++){
const int upper_bound = f(i);
#pragma acc parallel loop
for(int j=0;j<upper_bound;j++){
//Do work...
}
}
Where f is a monotonically-decreasing function of i.
Since num_gangs, num_workers, and vector_length are not set, OpenACC chooses what it thinks is an appropriate scheduling.
But does it choose such a scheduling afresh each time it encounters the pragma, or only once the first time the pragma is encountered?
Looking at the output of PGI_ACC_TIME suggests that scheduling is only performed once.
The PGI compiler will choose how to decompose the work at compile-time, but will generally determine the number of gangs at runtime. Gangs are inherently scalable parallelism, so the decision on how many can be deferred until runtime. The vector length and number of workers affects how the underlying kernel gets generated, so they're generally selected at compile-time to maximize optimization opportunities. With loops like these, where the bounds aren't really known at compile-time, the compiler has to generate some extra code in the kernel to ensure exactly the correct number of iterations are performed.
According to OpenAcc 2.6 specification[1] Line 1357 and 1358:
A loop associated with a loop construct that does not have a seq clause must be written such that the loop iteration count is computable when entering the loop construct.
Which seems to be the case, so your code is valid.
However, note it is implementation defined how to distribute the work among the gangs and workers, and it may be that the PGI compiler is simply doing some simple partitioning of the iterations.
You could manually define values of gang/workers using num_gangs and num_workers, and the integer expression passed to those clauses can depend on the value of your function (See 2.5.7 and 2.5.8 on OpenACC specification).
[1] https://www.openacc.org/sites/default/files/inline-files/OpenACC.2.6.final.pdf
quite newbie with OpenACC here, so please be patient :-)
I'm trying to port some Fortran code to use OpenACC, and I'm finding a strange (at least to me) behaviour.
The code is given below, but as you can see is just some nested loops which ultimately update the variable zc, which I copyout. I have tried to make private copies where I think they are needed and for the moment specified that all loops are independent. Now, when I compile with and without OpenACC all is fine if I remove the line "if(mu2-mup2.ne.q2) cycle", but if that line is present, then the results for the zc calculated with OpenACC are very different to those calculated without OpenACC.
Any ideas why that line could be giving me trouble?
Many thanks in advance,
Ángel de Vicente
!$acc data copyout(zc)
!$acc update device(fact)
!$acc kernels
!$acc loop independent private(k2)
do k=kmin,kmax
k2=2*k
!$acc loop independent private(km,kp2,z0)
do kp=kmin,kmax
km = MIN(k,kp)
kp2=2*kp
z0=3.d0*dble(ju2+1)*dsqrt(dble(k2+1))*dsqrt(dble(kp2+1))
!$acc loop independent private(q2)
do q=-km,km
q2=2*q
!$acc loop independent
do mu2=-ju2,ju2,2
!$acc loop independent private(p2,z1)
do ml2=-jl2,jl2,2
p2=mu2-ml2
if(iabs(p2).gt.2) cycle
z1=w3js(ju2,jl2,2,mu2,-ml2,-p2)
!$acc loop independent private(pp2,z2)
do mup2=-ju2,ju2,2
if(mu2-mup2.ne.q2) cycle
pp2=mup2-ml2
if(iabs(pp2).gt.2) cycle
z2=w3js(ju2,jl2,2,mup2,-ml2,-pp2)
!$acc loop independent
do mlp2=-jl2,jl2,2
zc(ml2,mlp2,mu2,mup2,k,kp,q) = z2
enddo
enddo
enddo
enddo
end do
end do
end do
!$acc end kernels
!$acc end data
Without a reproducing example it's difficult to give a complete answer, but I'll do my best.
First, there are only three parallel dimensions in OpenACC: gang, worker, and vector. Hence, the compiler will need to ignore 4 of the 7 loop directives. Most likely the middle 4 (if using PGI, you can see which loops the compiler is parallelizing from the compiler feedback messages, i.e. -Minfo=accel). Not that you can't parallelize all the loops, but you'd need to make them tightly nested and then use the collapse clause to collapse them into a single parallel loop.
Also since scalars are private by default, there's no need to put them into a private clause (except for a few cases). While putting them in a private clause shouldn't impact correctness, it can cause performance slow downs since you'd be fetching the private copy from global memory rather than having the potential of the scalar being put into a register.
My guess is that the inner loops are not that large so may not be beneficial to parallelize. Hence, I would first try removing all the inner "loop" directives, and only parallelize the "k" and "kp" loops. Depending of the values of "kmin" and "kmax", I'd try collapsing them as well. Something like:
!$acc loop independent collapse(2)
do k=kmin,kmax
do kp=kmin,kmax
k2=2*k
km = MIN(k,kp)
Assuming that gets you the correct answers but not as much parallelism as you want, you can then try collapsing the middle two loops:
!$acc loop independent collapse(2)
do q=-km,km
do mu2=-ju2,ju2,2
q2=2*q
do ml2=-jl2,jl2,2
I wouldn't recommend parallelizing loops with cycles in them. Not that you can't, but doing so would hurt performance due to thread divergence.
If none of this helps, please post a full reproducing example.
This is a pretty simple question. I'm practicing using MPI in Fortran, and as far as I can tell, the sending and receiving sources must be arrays, like this:
do 20 i = 1, 25
final_sum(1) = final_sum(1) + random_numbers(i)
20 continue
print *, 'process final sum: ', final_sum(1)
call MPI_GATHER(final_sum,1,MPI_INTEGER,sum_recv_buff,1,MPI_INTEGER,0,MPI_COMM_WORLD,ierror)
Here, I'm using the array final_sum to hold a single value. Is there a better way to do this? This is my first time using Fortran, and since I've already done some practice in C, I was trying out Fortran to see the differences and similarities.
High Performance Mark is right. There is actually normally no type checking when calling MPI routines that work with buffers. The MPI routines make use of the Fortran feature that enables calling procedures with implicit interface. On machine language level just a pointer is passed (it may be a pointer to a temporary copy!). That means you can use scalars or arrays without any problem. Just use count 1 and the correct MPI type and you can pass and receive scalars.
The MPI derived types is a feature that will enable you to work with even more complicated pieces of data.
I'm faced with parallelizing an algorithm which in its serial implementation examines the six faces of a cube of array locations within a much larger three dimensional array. (That is, select an array element, and then define a cube or cuboid around that element 'n' elements distant in x, y, and z, bounded by the bounds of the array.
Each work unit looks something like this (Fortran pseudocode; the serial algorithm is in Fortran):
do n1=nlo,nhi
do o1=olo,ohi
if (somecondition(n1,o1) .eq. .TRUE.) then
retval =.TRUE.
RETURN
endif
end do
end do
Or C pseudocode:
for (n1=nlo,n1<=nhi,n++) {
for (o1=olo,o1<=ohi,o++) {
if(somecondition(n1,o1)!=0) {
return (bool)true;
}
}
}
There are six work units like this in the total algorithm, where the 'lo' and 'hi' values generally range between 10 and 300.
What I think would be best would be to schedule six or more threads of execution, round-robin if there aren't that many CPU cores, ideally with the loops executing in parallel, with the goal the same as the serial algorithm: somecondition() becomes True, execution among all the threads must immediately stop and a value of True set in a shared location.
What techniques exist in a Windows compiler to facilitate parallelizing tasks like this? Obviously, I need a master thread which waits on a semaphore or the completion of the worker threads, so there is a need for nesting and signaling, but my experience with OpenMP is introductory at this point.
Are there message passing mechanisms in OpenMP?
EDIT: If the highest difference between "nlo" and "nhi" or "olo" and "ohi" is eight to ten, that would imply no more than 64 to 100 iterations for this nested loop, and no more than 384 to 600 iterations for the six work units together. Based on that, is it worth parallelizing at all?
Would it be better to parallelize the loop over the array elements and leave this algorithm serial, with multiple threads running the algorithm on different array elements? I'm thinking this from your comment "The time consumption comes from the fact that every element in the array must be tested like this. The arrays commonly have between four million and twenty million elements." The design of implementing the parallelelization of the array elements is also flexible in terms of the number threads. Unless there is a reason that the array elements have to be checked in some order?
It seems that the portion that you are showing us doesn't take that long to execute so making it take less clock time by making it parallel might not be easy ... there is always some overhead to multiple threads, and if there is not much time to gain, parallel code might not be faster.
One possibility is to use OpenMP to parallelize over the 6 loops -- declare logical :: array(6), allow each loop to run to completion, and then retval = any(array). Then you can check this value and return outside the parallelized loop. Add a schedule(dynamic) to the parallel do statement if you do this. Or, have a separate !$omp parallel and then put !$omp do schedule(dynamic) ... !$omp end do nowait around each of the 6 loops.
Or, you can follow the good advice by #M.S.B. and parallelize the outermost loop over the whole array. The problem here is that you cannot have a RETURN inside a parallel loop -- so label the second outermost loop (the largest one within the parallel part), and EXIT that loop -- smth like
retval = .FALSE.
!$omp parallel do default(private) shared(BIGARRAY,retval) schedule(dynamic,1)
do k=1,NN
if(.not. retval) then
outer2: do j=1,NN
do i=1,NN
! --- your loop #1
do n1=nlo,nhi
do o1=olo,ohi
if (somecondition(BIGARRAY(i,j,k),n1,o1)) then
retval =.TRUE.
exit outer2
endif
end do
end do
! --- your loops #2 ... #6 go here
end do
end do outer2
end if
end do
!$omp end parallel do
[edit: the if statement is there presuming that you need to find out if there is at least one element like that in the big array. If you need to figure the condition for every element, you can similarly either add a dummy loop exit or goto, skipping the rest of the processing for that element. Again, use schedule(dynamic) or schedule(guided).]
As a separate point, you might also want to check if it may be a good idea to go through the innermost loop by some larger step (depending on float size), compute a vector of logicals on each iteration and then aggregate the results, eg. smth like if(count(somecondition(x(o1:o1+step,n1,k)))>0); in this case the compiler may be able to vectorize somecondition.
I believe you can do what you want with the task construct introduced in OpenMP 3; Intel Fortran supports tasking in OpenMP. I don't use tasks often so I won't offer you any wonky pseudocode.
You already mentioned the obvious way to stop all threads as soon as any thread finds the ending condition: have each check some shared variable which gives the status of the ending condition, thereby determining whether to break out of the loops. Obviously this is an overhead, so if you decide to take this approach I would suggest a few things:
Use atomics to check the ending condition, this avoids expensive memory flushing as just the variable in question is flushed. Move to OpenMP 3.1, there are some new atomic operations supported.
Check infrequently, maybe like once per outer iteration. You should only be parallelizing large cases to overcome the overhead of multithreading.
This one is optional, but you can try adding compiler hints, e.g. if you expect a certain condition to be false most of the time, the compiler will optimize the code accordingly.
Another (somewhat dirty) approach is to use shared variables for the loop ranges for each thread, maybe use a shared array where index n is for thread n. When one thread finds the ending condition, it changes the loop ranges of all the other threads so that they stop. You'll need the appropriate memory synchronization. Basically the overhead has now moved from checking a dummy variable to synchronizing/checking loop conditions. Again probably not so good to do this frequently, so maybe use shared outer loop variables and private inner loop variables.
On another note, this reminds me of the classic polling versus interrupt problem. Unfortunately I don't think OpenMP supports interrupts where you can send some kind of kill signal to each thread.
There are hacking work-arounds like using a child process for just this parallel work and invoking the operating system scheduler to emulate interrupts, however this is rather tricky to get correct and would make your code extremely unportable.
Update in response to comment:
Try something like this:
char shared_var = 0;
#pragma omp parallel
{
//you should have some method for setting loop ranges for each thread
for (n1=nlo; n1<=nhi; n1++) {
for (o1=olo; o1<=ohi; o1++) {
if (somecondition(n1,o1)!=0) {
#pragma omp atomic write
shared_var = 1; //done marker, this will also trigger the other break below
break; //could instead use goto to break out of both loops in 1 go
}
}
#pragma omp atomic read
private_var = shared_var;
if (private_var!=0) break;
}
}
A suitable parallel approach might be, to let each worker examine a part of the overall problem, exactly as in the serial case and use a local (non-shared) variable for the result (retval). Finally do a reduction over all workers on these local variables into a shared overall result.
There is an mpi-version of a program which uses COMMON blocks to store arrays that are used everywhere through the code. Unfortunately, there is no way to declare arrays in COMMON block size of which would be known only run-time. So, as a workaround I decided to move that arrays in modules which accept ALLOCATABLE arrays inside. That is, all arrays in COMMON blocks were vanished, instead ALLOCATE was used. So, this was the only thing I changed in my program. Unfortunately, performance of the program was awful (when compared to COMMON blocks realization). As to mpi-settings, there is a single mpi-process on each computational node and each mpi-process has a single thread.
I found similar question asked here but don't think (don't understand :) ) how it could be applied to my case (where each process has a single thread). I appreciate any help.
Here is a simple example which illustrates what I was talking about (below is a pseudocode):
"SOURCE FILE":
SUBROUTINE ZEROSET()
INCLUDE 'FILE_1.INC'
INCLUDE 'FILE_2.INC'
INCLUDE 'FILE_3.INC'
....
INCLUDE 'FILE_N.INC'
ARRAY_1 = 0.0
ARRAY_2 = 0.0
ARRAY_3 = 0.0
ARRAY_4 = 0.0
...
ARRAY_N = 0.0
END SUBROUTINE
As you may see, ZEROSET() has no parallel or MPI stuff. FILE_1.INC, FILE_2, ... , FILE_N.INC are files where ARRAY_1, ARRAY_2 ... ARRAY_N are defined in COMMON blocks. Something like that
REAL ARRAY_1
COMMON /ARRAY_1/ ARRAY_1(NX, NY, NZ)
Where NX, NY, NZ are well defined parameters described with help of PARAMETER directive.
When I use modules, I just destroyed all COMMON blocks, so FILE_I.INC looks like
REAL, ALLOCATABLE:: ARRAY_I(:,:,:)
And then just changed "INCLUDE 'FILE_I.INC'" statement above to "USE FILE_I". Actually, when parallel program is executed, one particular process does not need a whole (NX, NY, NZ) domain, so I calculate parameters and then allocate ARRAY_I (only ONCE!).
Subroutine ZEROSET() is executed 0.18 seconds with COMMON blocks and 0.36 with modules (when array's dimensions are calculated runtime). So, the performance worsened by two times.
I hope that everything is clear now. I appreciate you help very much.
Using allocatable arrays in modules can often hurt performance because the compiler has no idea about sizes at compile time. You will get much better performance with many compilers with this code:
subroutine X
use Y ! Has allocatable array A(N,N) in it
call Z(A,N)
end subroutine
subroutine Z(A,N)
Integer N
real A(N,N)
do stuff here
end
Then this code:
subroutine X
use Y ! Has allocatable array A(N,N) in it
do stuff here
end subroutine
The compiler will know that the array is NxN and the do loops are over N and be able to take advantage of that fact (most codes work that way on arrays). Also, after any subroutine calls in "do stuff here", the compiler will have to assume that array "A" might have changed sizes or moved locations in memory and recheck. That kills optimization.
This should get you most of your performance back.
Common blocks are located in a specific place in memory also, and that allows optimizations also.
Actually I guess, your problem here is, in combination with stack vs. heap memory, indeed compiler optimization based. Depending on the compiler you're using, it might do some more efficient memory blanking, and for a fixed chunk of memory it does not even need to check the extent and location of it within the subroutine. Thus, in the fixed sized arrays there won't be nearly no overhead involved.
Is this routine called very often, or why do you care about these 0.18 s?
If it is indeed relevant, the best option would be to get rid of the 0 setting at all, and instead for example separate the first iteration loop and use it for the initialization, this way you do not have to introduce additional memory accesses, just for initialization with 0. However it would duplicate some code...
I could think of just these reasons when it comes to fortran performance using arrays:
arrays on the stack VS heap, but I doubt this could have a huge performance impact.
passing arrays to a subroutine, because the best way to do that depends on the array, see this page on using arrays efficiently