Why are we applying a 4096 alignment to our memory buffers?
In Benchmark::benchmark()
buffers are allocated with a hard-coded 4096 alignment. Why? This does not seem to be very future proof. Optimal memory alignment for the CPU RAM depends on the memory controller, motherboard memory traces and used memory modules. For GPU memory, or other device memory this relationship is further obfuscated. Anybody have an idea how we can do this more intelligently?