atomic for tasks

[pojeda]

Hi,

I wonder if there is support for atomic operations, especially in the context of tasks? I have extended a 2d integration from
one of our courses (https://uppmax.github.io/HPC-python/day4/parallel.html) using PyOMP:

import math
from time import perf_counter
from numba.openmp import njit
from numba.openmp import openmp_context as openmp

grid size

n = 100000

Task granularity - number of iterations per task

TASK_SIZE = 1000

@njit
def integration2d_omp(n):
h = math.pi / float(n)
total_sum = 0.0

with openmp("parallel"):
    with openmp("single"):
        # Create tasks for chunks of the outer loop
        num_tasks = (n + TASK_SIZE - 1) // TASK_SIZE
        
        for task_id in range(num_tasks):
            with openmp("task"):
                i_start = task_id * TASK_SIZE
                i_end = min(i_start + TASK_SIZE, n)
                local_sum = 0.0
                
                for i in range(i_start, i_end):
                    x = h * (i + 0.5)
                    for j in range(n):
                        y = h * (j + 0.5)
                        local_sum += math.sin(x + y)
                
                # Atomic update of the shared sum
                with openmp("atomic"):
                    total_sum += local_sum
        
        # Wait for all tasks to complete
        with openmp("taskwait"):
            pass

return h**2 * total_sum

if name == "main":
start = perf_counter()
integral = integration2d_omp(n)
end = perf_counter()
print(f"Integral value is {integral:e}, Error is {abs(integral - 0.0):e}")
print(f"Time spent: {end - start:.2f} sec")

However, the atomic seems not to be supported yet. Do you have any suggestions for using tasks for this problem?

[tgmattso]

Pedro, Support for atomic will need to be addressed by someone else on the team. I am confused. Why do you want to solve your problem with tasks? A much cleaner solution is to use a parallel loop with a reduction. The code would look something like this.

@njit <https://github.com/njit> def integration2d_omp(n): h = math.pi / float(n) total_sum = 0.0 with openmp(“parallel for reduction(+:total_sum)"): for i in range(0,n): x = h * (i + 0.5) for j in range(n): y = h * (j + 0.5) total_sum += math.sin(x + y) return h**2 * total_sum

—Tim

…

On Nov 25, 2025, at 7:41 AM, Pedro Ojeda ***@***.***> wrote: pojeda created an issue (Python-for-HPC/PyOMP#33) <#33> Hi, I wonder if there is support for atomic operations, especially in the context of tasks? I have extended a 2d integration from one of our courses (https://uppmax.github.io/HPC-python/day4/parallel.html) using PyOMP: import math from time import perf_counter from numba.openmp import njit from numba.openmp import openmp_context as openmp grid size n = 100000 Task granularity - number of iterations per task TASK_SIZE = 1000 @njit <https://github.com/njit> def integration2d_omp(n): h = math.pi / float(n) total_sum = 0.0 with openmp("parallel"): with openmp("single"): # Create tasks for chunks of the outer loop num_tasks = (n + TASK_SIZE - 1) // TASK_SIZE for task_id in range(num_tasks): with openmp("task"): i_start = task_id * TASK_SIZE i_end = min(i_start + TASK_SIZE, n) local_sum = 0.0 for i in range(i_start, i_end): x = h * (i + 0.5) for j in range(n): y = h * (j + 0.5) local_sum += math.sin(x + y) # Atomic update of the shared sum with openmp("atomic"): total_sum += local_sum # Wait for all tasks to complete with openmp("taskwait"): pass return h**2 * total_sum if name == "main": start = perf_counter() integral = integration2d_omp(n) end = perf_counter() print(f"Integral value is {integral:e}, Error is {abs(integral - 0.0):e}") print(f"Time spent: {end - start:.2f} sec") However, the atomic seems not to be supported yet. Do you have any suggestions for using tasks for this problem? — Reply to this email directly, view it on GitHub <#33>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AATVME5J4AFT2EDT44HQ6OD36R2BZAVCNFSM6AAAAACNFBXUO2VHI2DSMVQWIX3LMV43ASLTON2WKOZTGY3DGNJZGY2TQMA>. You are receiving this because you are subscribed to this thread.

[pojeda]

Hi,

I wanted to see PyOMP's capabilities for tasking, for teaching purposes. Atomic and taskloop are topics that, I think, can be very useful for our HPC users. I already have an example with the reduction operation; it works well. Thanks for your reply.

/Pedro

[tgmattso]

If you use a critical construct in place of your atomic directive, your program should work. As for taskloop, that construct is covered nicely by a for construct with the “schedule dynamic” clause. I may be telling you what you already know (since I know nothing of your background in parallel computing) but when teaching tasking, I like to focus on patterns that cannot be addressed through versions of the loop construct. The first is basically a version of what you did. A control structure executed by a single thread fills a task queue which the threads waiting at the implied barrier on the single execute. The point is to show students that a vast range of control structures beyond loops can be addressed thereby greatly expanding the range of algorithms that can be addressed with OpenMP The second is a parallel version of the divide and conquer pattern. I often show this with a program to do the numerical integration to estimate pi. I haven’t done so yet, but I need to create a DGEMM algorithm following Leiserson's work on cache oblivious algorithms. Those are by far the most commonly used patterns for task-level parallelism in OpenMP. —Tim

…

On Nov 25, 2025, at 10:38 PM, Pedro Ojeda ***@***.***> wrote: pojeda left a comment (Python-for-HPC/PyOMP#33) <#33 (comment)> Hi, I wanted to see PyOMP's capabilities for tasking, for teaching purposes. Atomic and taskloop are topics that, I think, can be very useful for our HPC users. I already have an example with the reduction operation; it works well. Thanks for your reply. /Pedro — Reply to this email directly, view it on GitHub <#33 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AATVME26XZ5GOSIOERRHW5D36VDF5AVCNFSM6AAAAACNFBXUO2VHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZTKNZZGUYTAMBQHE>. You are receiving this because you commented.