Memory corruption bug

[MikeCernea]

Summary

I’m seeing intermittent SEGV_ACCERR crashes when using the mapmap optimizer to select the best camera for each face during mesh texturing. The optimization itself works, but the process occasionally ends with a memory access violation.

The crashes also occur when forcing a single thread, so it doesn’t look like a straightforward data race.

---

Details

• Component: mapmap optimizer (used for per-face best camera selection in mesh texturing)
• Repro: on my dataset, the crash happens roughly 1 in 5 runs
• Failure mode: GeneralDPNodeSolver::optimize_node() inside an AVX intrinsic (avxintrin.h:911), with SEGV_ACCERR (access error - permission violation, not unmapped memory) / memory corruption–style crash

I also noticed this allocator pattern in a couple of places:

m_value_allocator(
    (tbb::tbb_allocator<_s_t<COSTTYPE, SIMDWIDTH>>*)(
        new tbb::cache_aligned_allocator<_s_t<COSTTYPE, SIMDWIDTH>>
    )
);

I don’t understand the purpose of allocating a tbb::cache_aligned_allocator and then casting it to tbb::tbb_allocator. I tried changing it to:

m_value_allocator(
    new tbb::cache_aligned_allocator<_s_t<COSTTYPE, SIMDWIDTH>>()
);

but the intermittent crash still occurs.

---

Questions

1. Has anyone seen similar intermittent SEGV_ACCERR / memory corruption issues with mapmap?
2. What is the intended rationale for the allocator recast above?
   • Is this pattern actually safe, or could it be undefined behavior and related to the crash?
3. Is there a recommended way to configure allocators for mapmap (especially if cache-aligned allocation is desired)?

I can provide stack traces, build flags, and environment details if needed.

[MikeCernea]

After replacing all the tbb::tbb_allocator and tbb::cache_aligned_allocator throughout the code and switching to openMP parallel processing, I no longer have the memory corruption issues. What version of TBB was used to develop and test the latest mapmap cpu version?

[dthuerck]

Hi @MikeCernea , thanks for the investigative work and your issue! What version of AVX are you using? Could you please share your system and setup? Imho (see https://www.aussieai.com/book/ch17-avx-memory-alignment-issues) AVX1 and AVX2's alignments are different.

The tbb::cache_aligned_allocator is still created, just sored as its interface pointer, that should not result in errors. The issue looks like an out-of-bounds access through AVX.