compute Directory Reference

Compute back ends.

More...

Directory dependency graph for compute:

Files
file	compute.cpp
file	compute.h [code]
file	lua_compute.cpp
file	opencl.cpp
file	opencl.h [code]
file	pseudo.cpp

Detailed Description

Compute back ends.

Compute back ends for task execution in heterogeneous compute devices.

• OpenCLBackend: main/default OpenCL 1/2 back end.
• Pseudocomp: fake back end for testing.
• An absent back end is also supported, in which case native CPU code is used.

The following area of the screenshots page shows some of the compute capabilities:

https://bbguimaraes.com/nngn/screenshots/compute.html

Lua

The compute back end is exposed to Lua via the nngn.compute variable. Compiling a program and executing a kernel is done similarly to the C++ API:

prog = nngn.compute:create_prog(io.open("prog.cl"):read("a"))
nngn.compute:execute(prog, "fn", Compute.BLOCKING, {256, 256}, {16, 16}, {
    Compute.FLOAT, 3.1415,
    Compute.BYTEV, {0, 1, 2, 3},
})

Compute.BLOCKING makes the execution synchronous. Kernels can also be executed concurrently (requires a device with an out-of-order queue) and synchronized using events:

events = {}
-- Last two arguments are wait list and output events.
nngn.compute:execute(prog, "k0", 0, {1}, {1}, {}, nil, events)
-- Reuse output.
nngn.compute:execute(prog, "k1", 0, {1}, {1}, {}, nil, events)
-- Wait for all previous events, reuse output.
nngn.compute:execute(prog, "k2", 0, {1}, {1}, {}, events, events)
-- Wait for all previous events and release them.
nngn.compute:wait(events)
nngn.compute:release_events(events)

Separate tables can be used to construct more complex task graphs.

Buffers

To create and populate a device buffer:

b = assert(nngn.compute:create_buffer(Compute.READ_WRITE, size))
v = Compute.create_vector(size)
Compute.fill_rnd_vector(v)
assert(nngn.compute:write_buffer(b, v))

Common operations on buffers are creation:

rw = Compute.READ_WRITE
assert(nngn.compute:create_buffer(rw, Compute.FLOATV, size))
-- From existing data.
assert(nngn.compute:create_buffer(
    rw, Compute.FLOATV, size,
    Compute.create_vector(1024)))

write:

assert(nngn.compute:write_buffer(b, offset, n, Compute.FLOATV, v))

and read:

assert(nngn.compute:read_buffer(b, Compute.FLOATV, n, v))