multiply and xor mix function, aka MUM

read functions. WARNING: we don't care about endianness, so results are different on big endian!

iterator functions These rely on device_array having contiguous memory layout -- let's hope this is the case despite the registerization optimizations

Both libc++ and libstdc++ define std::malloc as ::malloc and similarly * for std::calloc, std::aligned_alloc, and std::free, so it is enough to * i

This is a hack: If we just directly access num_outstanding_operations for testing, we trigger a load for this global variable, which would cause a bar