Accessor Requirements for an Implementation (Data Consistency)

Buffer Memory Layout

typical metadata like size, type, etc.
must maintain a point-of-truth for the data, but can have multiple copies of the data in various memory spaces. the rules for accessors below ensure that there are only read copies or a single write copy of the data at any given time. This maintains data consistency for the point-of-truth.
the point-of-truth can and usually is spread across memory spaces.
buffers should provide an api for setting the following hints:
- read-mostly: used in UVM runtimes to telling the underlying driver how to manage the unified memory space
- preferred memory space(s): used to force the point-of-truth to be in a specific memory space(s)
- access via hint: can sometimes help the runtime to optimize the location of read-only copies of the data

Only one host exists for a given system, even one with multiple CPUs.
If const, assumed read-only. Accesses will not signal to the owning buffer that it must update the point-of-truth data. Reads must be blocked if any device task is writing to the buffer until the task completes.
If non-const, assumed read-write. Accesses will signal to the owning buffer that it must update the point-of-truth data as all accesses are assumed to be a write operation. This makes it important for users to use const accessors when possible. Accesses will be blocked while any device task using the buffer is running, even if the device is only reading from the buffer. The accessor must be updated with point-of-truth data after any device task that writes to the buffer completes.

Must implement open which notifies the owning buffer that it is to be used by a device task. The open call must wait until the buffer signals that it is safe to use the accessor. To allow overlapping open requests for different buffers used in a task, the open call must return a future.
Must implement close which notifies the owning buffer that the device task is complete. It must also return a future. The buffer will use the close call to determine if any other accessors to the data are now safe to be used. It will also update the point-of-truth and any host accessors if the accessor was non-const

only one host
everything else is either a node or a device
devices are attached to nodes
buffer memory management should be deterministic so that layout metadata is the same on all nodes

create buffer, and runtime puts point-of-truth in a memory space(s)
construct a gpu_task:
- register inputs, each register call takes an optional read stencil to reduce the amount of data copied, empty access stencil indicates write-only operation
- final setup: queries point-of-truth locations for the buffers associated with each input to determine the best memory space(s) to use for the task. this can be overridden by the user. sets up stencil mapping for each input if necessary (e.g. hash map from global index to local index)
run the task when ready:
- prepare: request access from buffers for each input
- execute: run the task
- finalize: each input is closed which notifies the buffer that it is no longer in use