On Fri, May 20, 2016 at 08:02:16PM +0300, Marius Vlad wrote: > There's no need to multiply the number of batches with the number of > engines as intel_require_memory() already compares against the aperture > size (count * BATCH_SIZE). nengines was because I first planned to allocate an array for each engine, because I wanted each to run as independently as possible (since we overwrite the batch if it is active, we will stall). Then I realised the futility and overallocation since we only have the single address space and so allocated the single array and set each engine off on a random order to hopefully avoid them lockstepping. > This also removes the weird assertion messages where we need > bogus amounts of RAM. It won't. However you put it, we need 256 exabytes to run the test on bdw+. Unless we can constrain the per-context GTT (i.e. a constrction flag to limit the address space to 32bits).... Or we just don't use the allow-48bit flag and pretend that is equivalent. > Also tune down the timeout from from 10s to 2s to speed up BAT. It's not going to make any difference on the slowest machines, but meh. -Chris -- Chris Wilson, Intel Open Source Technology Centre _______________________________________________ Intel-gfx mailing list Intel-gfx@xxxxxxxxxxxxxxxxxxxxx https://lists.freedesktop.org/mailman/listinfo/intel-gfx
