Unintended heap allocations

[fennel-labs]

ProxQP does a lot of work to avoid heap allocations. However, I found that some heap allocations are still performed after the initialization. Here is the call graph with the functions that allocate from the heap during solve():

Steps to reproduce:

• Take any QP (e.g. from testcases)
• Compile with:
  • Eigen 3.3.4
  • -O0 -g -DNDEBUG -std=gnu++17 -DPROXSUITE_VECTORIZE
• valgrind --tool=massif --xtree-memory=full <name_to_executable_to_examine>
• kcachegrind <name_of_output>

Desired behavior

ProxQP does not generate any heap allocations during solve. This is essential for running ProxQP on resource-constrained devices or under real-time constraints.

[ahoarau]

Seems to be eigen stuff. It has the nice Eigen::internal::set_is_malloc_allowed(false); function to detect runtime allocations.

With https://gitlab.com/libeigen/eigen/-/merge_requests/1055 it even supports reallocs.

[fennel-labs]

It is definitely Eigen, as the lowest node of the call graph above already reveals. This means, that Eigen is used wrong or has some unknown version incompatibility with ProxQP.

[ahoarau]

Then you can try to surround the solve() call with

Eigen::internal::set_is_malloc_allowed(false);
solver->solver();
Eigen::internal::set_is_malloc_allowed(true);

NOTE: Don't forget to compile in Debug with EIGEN_RUNTIME_NO_MALLOC defined somewhere.

[jcarpent]

@fennel-labs Thanks for raising this issue.
Could you try with Eigen 3.4.0?

[fabinsch]

With @Bambade we checked the dense backend and we could find that we actually have some allocations in the dense factorize function. It fails if we set_malloc_allowed to false here.

[fennel-labs]

I tried with 3.4.0, allocations still happen. Here is the corresponding call graph:

@fabinsch You can find all relevant function names with allocations in the figure.

[jcarpent]

@fennel-labs We have fixed the heap allocations in the code in #92.
Hope it will solve your issue.

[fennel-labs]

@jcarpent I can confirm, that this fixes the issue under Eigen 3.4.0. With 3.3.4 heap allocations are still observed.

[jcarpent]

Could provide the call graph for 3.3.4 after the fix?

[jcarpent]

@fennel-labs I've just updated the PR with further fixes. Should be fine hopefully.

[fennel-labs]

You were a minute faster than me, here is the call graph. I will try the new changes immediately.

[fennel-labs]

@jcarpent Some allocations are still left:

[jcarpent]

@fennel-labs 2ec747d should fix the issue on Eigen 3.3.4

[fennel-labs]

@jcarpent Yes, allocations are gone now for both, 3.3.4 and 3.4.0. Thank you so much!

[jcarpent]

Thanks @fennel-labs for the quick feedback. We really appreciate it.