I wrote a simple test to prove that SVM fine-grained buffer atomics work on my system. The Intel OpenCL driver says that my GPU (HD Graphics 630) supports fine-grained buffers with atomics.
If I use any NEO driver version, including version 24.20.100.6136 currently listed on Intel’s website, the test fails. With NEO drivers, memory is not synchronizing between the CPU and GPU when I make an atomic access. If a GPU kernel makes an atomic modification to shared virtual memory, the CPU cannot see the change until the entire GPU kernel (not just the atomic access) has completed.
If I downgrade to a pre-NEO driver (confirmed with 22.20.16.4836 and 21.20.16.4590), I see that memory is synchronizing with each atomic access as expected. When the GPU makes an atomic modification to SVM, the CPU sees the change immediately while the GPU kernel is still running.
I have been trying to implement an approach like Intel’s presentation “GPU Daemon: Road to Zero Cost Submission,” but that is impossible without working shared atomics.
When will SVM atomics be supported by the NEO driver?
