Find a solution for sending data to device about gfx HOT 4 CLOSED

The holy grail of buffer sharing with the gpu these days is to have memory mapped on both the cpu and the gpu at the same time for the entire lifetime of the buffer/texture object and work out synchronization explicitly using fences. ARB_buffer_storage is pretty neat, is supported by Mesa 10.2, has been available with nvidia cards for a while but I am not sure how widely it is supported by AMD hardware). Gralloc textures, on recent versions of Android also have this kind of zero-copy-always-mapped model. Each platform has its own specific APIs to achieve this kind of stuff although sometimes you have a map/unmap-stype API instead of fences (unfortunately).
If you manage to design the API to be able to take advantage of this whenever possible, it'd be awesome since at this point you can't get much more efficient than that as the gpu doen't have to pay the cost of any synchronization and only needs to do a copy if the memory has a different representation inside the gpu (swizzled textures) or if the gpu has its own memory. This may be at the expense of a complex or unsafe API, but streaming data to the gpu is so often a bottleneck that a bit f "unsafety" here might be worth it.

from gfx.

@nical I find it hard to think about mapped memory across the task boundary that we have between the renderer and the device...

from gfx.

Well now we don't have the explicit task barrier, so time to revisit this issue.

from gfx.

Closing this in favor of #236. Now that we don't have the copy into the intermediate Vec, we have no more overhead than raw OpenGL calls. Buffer mapping would still be advantageous, though.

from gfx.