Towards shared memory consistency models for GPUs