map_async for generated data?

[erlendp]

Should I favour BufferSlice::map_async when I need to update a buffer regularly with data generated on the fly, as presumably this will allow me to avoid allocating a temporary CPU buffer, or would I be better to just use Queue::write_buffer?

[cwfitzgerald]

map_async carries with it quite a bit of complexity, and write_buffer is designed with being a fast-path in mind (that also happens to be very easy to use)

[erlendp]

So it's safe to say that writing my data to an in memory data structure first, and then calling write_buffer will generally outperform map_async, followed by writing directly to a device buffer?

P.S. If the answer is "probably, but you should profile", then that's good enough for me!

[kvark]

A well-written map_async is able to be faster than a well-written write_buffer (on native only). It's especially visible in cases like video uploads - you have a concrete known size each frame, and re-using buffers of this size is straightforward. For random-sized updates it's more of a tie, there are small pros/cons. Generally, don't bother with map_async until you are doing an optimization and figuring out the hot spots.