[alpaka] Should we define ALPAKA_DEBUG_OFFLOAD_ASSUME_HOST for the CPU backends?

[makortel]

#219 changed all uses of assert() to use ALPAKA_ASSERT_OFFLOAD, and this disables the assertions by default. This behavior corresponds what we do in cuda program, but in cudacompat we have asserts enabled. A ALPAKA_DEBUG_OFFLOAD_ASSUME_HOST macro can be used to enable assertions, so the question is, should we enable that macro for Alpaka CPU backends to be consistent with the behavior of cudacompat?

(or vice versa, should we disable asserts also in cudacompat?)

[fwyzard]

I think the rationale is that

• assert on the cpu is relatively cheap
• assert on the gpu is expensive

So, yes, to me it would make sense to enable the asserts by default on the cpu backends.

[fwyzard]

By the way, we could also (ask the Alpaka developers to) consider replacing

  assert(expression);

with something like

  [[noreturn]] __device__
  inline void printAssertMessage(const char* file, int line, const char* func, const char* expression, const char* message = 0) {
    printf("%s:%d:%s: block: [%d,%d,%d], thread: [%d,%d,%d] Assertion `%s` failed.\n",
        file, line, func,
        blockIdx.x, blockIdx.y, blockIdx.z,
        threadIdx.x, threadIdx.y, threadIdx.z,
        expression);
    if (message) {
      printf("%s\n", message);
    }
    __trap();
  }

  #define my_assert(expression) if (__builtin_expect(not static_cast<bool>(expression), false)) printAssertMessage(__FILE__, __LINE__, __PRETTY_FUNCTION__, #expression, nullptr)

if it is confirmed that it has better performance.

[fwyzard]

Looks like __trap() is not available on HIP/ROCm ?