On my Kaveri system, streaming data from CPU to GPU via VRAM is faster than via GTT both with the integrated GPU and with discrete GPUs. The following kernel patches make this safe by always flushing the HDP cache before submitting a command stream to the GPU, and bump the radeon DRM minor version. The following Mesa patches check for the bumped radeon DRM minor version, and if it's satisfied, they prefer CPU -> GPU streaming via VRAM and relax the restrictions for persistent mappings. [PATCH 1/2] drm/radeon: s/ioctl_wait_idle/mmio_hpd_flush/ [PATCH 2/2] drm/radeon: Always flush the HDP cache before submitting [PATCH 1/2] r600g/radeonsi: Reduce or even drop special treatment of [PATCH 2/2] r600g/radeonsi: Prefer VRAM for CPU -> GPU streaming _______________________________________________ dri-devel mailing list dri-devel@xxxxxxxxxxxxxxxxxxxxx http://lists.freedesktop.org/mailman/listinfo/dri-devel
