Module Rig_nv_abi

NVIDIA GPUs' formats, as values.

What compiled code and the driver of an NVIDIA GPU share: the words a channel runs, the entries of its ring, the descriptors its launches read, the cubins they run, and the record that describes the GPU. Nothing here acts on a GPU. Each value describes bytes, and its caller writes them.

The library describes bytes with two kinds of value:

A template (Template.t) is bytes around the caller's values: each value is an integer known now, written in place, or a slot, a value the caller numbers and fills later through a patch. Method and Gpfifo give each operation's words as a template, and Qmd a launch's descriptor and parameters.
A GPU (Gpu.t) is what the formats depend on: its classes, its geometry and the windows its driver chose, declared by the driver (declare). A cubin's kernel (Cubin) set up for launch on a GPU is a Launch.t, from which its descriptor is built; Local_memory sizes the local memory its threads take.

cubin bytes --Cubin.of_string--> Cubin.t --kernel--> Cubin.kernel
                                   |                      |
                          size, elf, patches       Launch.make Gpu.t
                                   |                      |
                      a loader writes the image      Launch.t
                                                          |
                                             Qmd.make, Qmd.parameters
                                                          |
Method.*, Gpfifo.entry --------> Template.t <-------------+
                                     |
                   bytes as they are, or filled at each use

Conventions

Addresses are the GPU's virtual addresses. A slot cannot be checked here: the caller keeps the alignment and range each function states for the value that fills it.

Misuse raises Invalid_argument: an integer outside the range its function states, a class or version no launch descriptor holds. A cubin that is not one, or a kernel that takes more than a launch can, is Error msg, msg saying why.

The library has no state: any domain may use any of its values.

References

Qmd and Launch name the documents of NVIDIA's that they follow.

Channel words

#module Template : sig ... end

Bytes around the caller's values.

#module Method : sig ... end

Channel methods.

#module Gpfifo : sig ... end

Ring entries.

GPUs

#module Gpu : sig ... end

NVIDIA GPUs, as their formats depend on them.

#val key : Gpu.t Stdlib.Type.Id.t

key is the key of an NVIDIA GPU's record. Compiled code finds the record of a device d with Rig.capability d key, which is None if d is not an NVIDIA GPU.

#val declare : Gpu.t -> Rig_edge.capability

declare c is c under key, as the facts of an NVIDIA GPU carry it (Rig_edge.facts).

#module Local_memory : sig ... end

Kernels' local memory.

Launches

#module Cubin : sig ... end

Cubins: the ELF objects NVIDIA's compilers make for a GPU architecture.

#module Launch : sig ... end

Kernels set up for launch on a GPU.

#module Qmd : sig ... end

Launch descriptors.