- Built per-tensor-node stream-to-stream synchronization into copying functions.

- Re-introduced whole-device blocking synchronization, which now is just a slight optimization as it also cleans up event book-keeping.

- Simplifications: no more explicit compilation postponing; no more hard-coded pointers (all non-local arrays are passed by parameter).

{

dev;

ordinal;

cross_stream_candidates = Hashtbl.create (module Tnode);

owner_stream = Hashtbl.create (module Tnode);

shared_writer_streams = Hashtbl.create (module Tnode);

type ('buffer_ptr, 'dev, 'runner, 'event) device_ref = {

dev : 'dev;

ordinal : int;

cross_stream_candidates : 'buffer_ptr Hashtbl.M(Tnode).t;

owner_stream : ('buffer_ptr, 'dev, 'runner, 'event) stream_ref Hashtbl.M(Tnode).t;

shared_writer_streams :

('buffer_ptr, 'dev, 'runner, 'event) device_ref = {

dev : 'dev;

ordinal : int;

cross_stream_candidates : 'buffer_ptr Hashtbl.M(Tnode).t;

(** Freshly created arrays that might be shared across streams. The map can both grow and

Option.iter Backend.free_buffer ~f:(fun mem_free ->

if

Atomic.compare_and_set ctx.finalized false true

then (

Backend.await ctx.stream;

Map.iteri ctx.ctx_arrays ~f:(fun ~key ~data ->

Stdlib.Gc.full_major ();

(* TODO: is running again needed to give time to weak arrays to become empty? *)

Stdlib.Gc.full_major ();

match

Option.value_or_thunk backend_name ~default:(fun () ->

Utils.get_global_arg ~arg_name:"backend" ~default:"cc")

| "sync_cc" -> (module Make_device_backend_from_lowered (Schedulers.Sync) (Cc_backend) : Backend)

| "sync_gccjit" ->

(module Make_device_backend_from_lowered (Schedulers.Sync) (Gcc_backend) : Backend)

| backend -> invalid_arg [%string "Backends.fresh_backend: unknown backend %{backend}"]