docs menu

Module Rig_nv_abi.Local_memory

Kernels' local memory.

Each thread of a launch has local memory of its own, for its stack and spills, which kernels see through the local memory window (Method.local_memory_window). A GPU takes it from one allocation, given with Method.local_memory, which holds as much for every thread every multiprocessor may run concurrently. The driver holds it and grows it on demand (Gpu.t.local), sized by make. Each launch says how much one thread has (Qmd.t.local).

#type t = {

#per_thread : int;
(*
The bytes of each thread, a multiple of 32.
*)
#per_tpc : int;
(*
The bytes of each texture processing cluster, a multiple of 32 KiB, as Method.local_memory takes them.
*)
#bytes : int;
(*
The allocation's size, a multiple of 128 KiB.
*)

}

The type for local memory.

#val make : Gpu.t -> int -> t

make g n is the local memory of g for kernels whose threads need n bytes each (Launch.local_bytes):

per_thread is n rounded up to a multiple of 32;
per_tpc is 32 * per_thread * g.warps_per_sm * g.sms_per_tpc, a warp's share for every warp of the cluster, rounded up to a multiple of 32 KiB;
bytes is per_tpc * g.tpcs_per_gpc * g.gpcs rounded up to a multiple of 128 KiB.

It is all 0 for n = 0.

Raises Invalid_argument if n is negative, or so large that bytes would be past max_int.