On Mon, 6 Jun 2011 16:23:00 -0700, "Segovia, Benjamin" benjamin.segovia@intel.com wrote:
Hello all,
I saw at some point that per-process GTT (ppGTT) may be (or is already) implemented to handle paging. Right now, I am investigating some flat space addressing (ab)using surface states. The idea is to create a surface state (raw buffer only, this is GPGPU stuff) as big enough to cover the entire address space such that I will only manipulate offsets as pointers in this surface instead of dealing with both offsets and surface IDs (in other words, segmented address space).
My concern is relative to the way bo buffers are mapped. Basically, I must be sure that _all_ of them are either mapped using ppGTT or GTT. Otherwise, this will also bring another form of segmentation. If ppGTT is implemented or will be implemented, will there be anyway to know how a bo is mapped?
If an exec would be done through PPGTT, it would be something you ask for as part of the execbuffers call. It won't change out from under you.
However, we stopped looking into PPGTT because we've been told by hw people that it's hopelessly broken on all hardware in undocumented ways (and basically no other details than that).