fix : vx_spawn

[talubik]

Hello. I have encountered a bug in vx_spawn. When the number of groups (or tasks) does not divide evenly across active
cores, the old formula produced wrong offsets for cores with
core_id < remaining, causing some work items to be skipped and others
to be processed twice.

Cause of bug

Before computing the offset, total_groups_per_core is incremented for
cores that receive an extra group:

if (core_id < remaining_groups_per_core)
    ++total_groups_per_core;

The old formula then used this already-mutated variable:

group_offset = core_id * total_groups_per_core + MIN(core_id, remaining);

This double-counts the remainder for affected cores. The same bug was
present in the task distribution path (all_tasks_offset).

Example

10 groups, 4 cores → base = 2, remaining = 2

Core	old offset	correct offset	groups assigned (old)
0	0	0	[0, 1, 2]
1	4	3	[4, 5, 6] — group 3 skipped
2	6	6	[6, 7] — group 6 duplicate
3	8	8	[8, 9]

This pull request fixes this bug.

[davymillion]

Hi, I encountered the same issue so I agree with the bug.
Your logic in the else statement is hard to understand and I think this can be simplified with the already defined variable total_groups_per_core which is equal to your A) base_groups + 1 and B) base_groups in the A) if core_id < remaining_groups_per_core and B) else cases, respectively. Here is the patch i've done in my code (same should be applied for all_tasks_offset):

// calculate offsets for group distribution
uint32_t remaining_offset = 0;
if (core_id >= remaining_groups_per_core)
  // The remaining groups per core have been distributed (1 by 1) on the
  // previous cores (whose core_id are < remaining_groups_per_core), so we
  // now need to offset by the total remaining groups
  remaining_offset = remaining_groups_per_core;
uint32_t group_offset = core_id * total_groups_per_core + remaining_offset;

If I'm not mistaken, this is equivalent to your code. For the if case in your code, this is easy to show as remaining_offset stays at 0 and total_groups_per_core = base_groups + 1. For the else condition, due to the fact that total_groups_per_core = base_groups, your group_offset computation can be developed and simplified: group_offset = remaining_groups_per_core * (total_groups_per_core + 1) + (core_id - remaining_groups_per_core) * (total_groups_per_core) = remaining_groups_per_core + total_groups_per_core * core_id, which is what computes the patch I just shared.