improve ALPAKA_ASSERT_OFFLOAD

[fwyzard]

It's not clear to me what the meaning of ALPAKA_ASSERT_OFFLOAD is :-)

The current definition is

#define ALPAKA_ASSERT(...) assert(__VA_ARGS__)

#if defined(ALPAKA_DEBUG_OFFLOAD_ASSUME_HOST) || defined(SYCL_EXT_ONEAPI_ASSERT)
#    define ALPAKA_ASSERT_OFFLOAD(EXPRESSION) ALPAKA_ASSERT(EXPRESSION)
#elif defined __AMDGCN__ && (!defined NDEBUG)
#    define ALPAKA_ASSERT_OFFLOAD(EXPRESSION)                                                                         \
        do                                                                                                            \
        {                                                                                                             \
            if(!(EXPRESSION))                                                                                         \
                __builtin_trap();                                                                                     \
        } while(false)
#else
#    define ALPAKA_ASSERT_OFFLOAD(EXPRESSION)                                                                         \
        do                                                                                                            \
        {                                                                                                             \
        } while(false)
#endif

So, if I interpret it correctly:

• if ALPAKA_DEBUG_OFFLOAD_ASSUME_HOST is defined, ALPAKA_ASSERT_OFFLOAD expands to ALPAKA_ASSERT, which expands to assert, which expands to a runtime check only if NDEBUG is not defined;

• if ALPAKA_DEBUG_OFFLOAD_ASSUME_HOST is not defined:

  • for the CPU backends, ALPAKA_ASSERT_OFFLOAD always expands to a no-op;
  • for the CUDA backend, ALPAKA_ASSERT_OFFLOAD always expands to a no-op;
  • for the ROCm backend, ALPAKA_ASSERT_OFFLOAD expands to a runtime check if NDEBUG is not defined;
  • for the SYCL backend, ALPAKA_ASSERT_OFFLOAD always expands to assert, so it expands to a runtime check if NDEBUG is not defined.

It looks like the behaviour is different for the different backends, which seems counter intuitive.

So... what is the expected behaviour ?

1. always compile to a runtime check
2. compile to a runtime check if NDEBUG is not defined
3. compile to a runtime check if some other macro is not defined
4. always compile to a runtime check on the CPU, and compile to a runtime check on the other devices if NDEBUG is not defined
5. always compile to a runtime check on the CPU, and compile to a runtime check on the other devices if some other macro is not defined
6. something else
7. always compile to a no-op