(We had this previously, but I didn't see actual use for it, until now:)
We add an option to the benchmark executable to allow users to require that OpenCL state vector buffers are allocated in general heap. I'm not so lucky, with my personal devices, but this could sometimes allow the use of greater general heap capacities than GPU internal VRAM, potentially including swap space, dependent on DMA capabilities of the GPU/system.
(Rather, the ability to use swap disk, or more than GPU VRAM width, might properly depend on just the "DMA controller," but that's about as much as I know. For our part, we just CL_MEM_USE_HOST_PTR.)
(We had this previously, but I didn't see actual use for it, until now:)
We add an option to the benchmark executable to allow users to require that OpenCL state vector buffers are allocated in general heap. I'm not so lucky, with my personal devices, but this could sometimes allow the use of greater general heap capacities than GPU internal VRAM, potentially including swap space, dependent on DMA capabilities of the GPU/system.