manover

In other words, OpenCL supports this very limited buffer interface due to compatibility issues, i.e. this kind of MMIO is the lowest common denominator that has to be implemented by any GPU claiming OpenCL compatibility. Although, this does not preclude most desktop discrete GPUs from mapping their whole internal VRAM onto host's memory address space through PCI bus. It seems to be the common mechanism for a host to access VRAM in the modern ATI and NVidia GPUs from what I understood after skimming through several technical documents. It is, by the way, as far as I can tell, the main reason behind the infamous 'memory hole' in 32-bit Windows OS's (inability to use more than 2.5-3G of RAM). So, I guess, the correct answer to my initial question as to why it's not possible to use tmpfs with VRAM would be because that will require special memory allocation made in VRAM. Meaning, a patch to tmpfs code that can properly allocate memory in VRAM buffer would suffice if we are willing to limit compatibility to 64-bit x86 architecture with AMD/NVidia GPUs.

Yes, it's owned by the GPU, but you can map it to the regular space. In fact, this is exactly how textures and other data gets loaded to the video card. See http://en.wikipedia.org/wiki/Memory-mapped_I/O

Could anybody please explain to me why there is a need for a special treatment of VRAM compared to a regular system RAM in this use case? Assuming, we can perform an allocation in VRAM (probably using OpenCL API), why can't we use tmpfs/ramfs code? Do I understand correctly that PCI maps VRAM to a certain memory region and it is accessible via regular CPU instructions? Is it because CPU caching is different or VRAM is uncacheable? Or is it something else?