Mesh File Version Check is Not Hit When Compiled with Kokkos OpenMP Backend

[Fuad-HH]

omega_h/src/Omega_h_file.cpp

Line 450 in 82824b2

OMEGA_H_CHECK(version <= latest_version);

This line gets hit when I want to read a mesh of version 11 without kokkos with an older Omega_h (it tried commit 09a5171) in release build.

assertion version <= latest_version failed at /users/hasanm4/lore/wsources/omega_h_scorec/src/Omega_h_file.cpp +441
Aborted (core dumped)

But when I build the same Omega_h with Kokkos (with OpenMP backend), it does not hit the check. It terminates with a std::bad_alloc error.

terminate called after throwing an instance of 'std::bad_alloc'
  what():  std::bad_alloc

Although if I build it in debug mode, it hits that check as expected:

describe: omega_h_scorec/src/Omega_h_file.cpp:441: void Omega_h::binary::read(std::istream&, Omega_h::Mesh*, Omega_h::I32): Assertion `version <= latest_version' failed.

[cwsmith]

Thanks @Fuad-HH . What you described is the 'expected' behavior of the OMEGA_H_CHECK(...) macro:

omega_h/src/Omega_h_fail.hpp

Lines 41 to 46 in 82824b2

	#if defined(OMEGA_H_USE_KOKKOS)
	# if defined(NDEBUG)
	# define OMEGA_H_CHECK(cond) static_cast<void>(cond)
	# else
	# define OMEGA_H_CHECK(cond) assert(cond)
	# endif

omega_h/src/Omega_h_fail.hpp

Lines 65 to 69 in 82824b2

	#else
	# define OMEGA_H_CHECK(cond) \
	((cond) ? ((void)0) \
	: Omega_h::fail("assertion %s failed at %s +%d\n", #cond, __FILE__, __LINE__))
	#endif

Looking back, to when those checks were added to file.cpp, the macro was much simpler:

https://github.com/sandialabs/omega_h/blob/398c8abafaee16df3e11ef49286a3f590235ee43/src/Omega_h_c.h#L78-L85

but I think you'd have a similar experience:

• CUDA Release build - assert disabled (same as current)
• CPU only build - always fails (same as current)

For conditionals like this (host code, typically ran once per rank), I think we should always run them as they are essentially 'free'. In PUMI we added an 'ALWAYS_ASSERT' macro that.... always ran regardless of the build configuration. I think adding something like that here for host only use makes sense.

[cwsmith]

for the record, here is the PCU_ALWAYS_ASSERT(...) macro:
https://github.com/SCOREC/core/blob/6a972af6daceec626e589b4d4f63acc22e27317f/pcu/pcu_util.h#L50-L60
(note, PCU is the 'parallel communication utility' where we happen to put a handful of utility functions there in PUMI).