GitHub

AMGX.jl

AMGX.jl wraps NVIDIA's AMGX, a GPU-accelerated algebraic multigrid solver library, for use from Julia.

Reading the official AMGX reference manual is recommended — the configuration parameters in particular are AMGX's own, and this package does not redocument them.

Installation

using Pkg; Pkg.add("AMGX")

Prebuilt binaries are available for Linux. On other systems you need a local AMGX build; point JULIA_AMGX_PATH at the shared library before loading the package:

ENV["JULIA_AMGX_PATH"] = "/path/to/libamgxsh.so"
using AMGX

A complete example

Solving a small system on the GPU:

using AMGX, CUDA, SparseArrays

AMGX.initialize()

config    = AMGX.Config(Dict("monitor_residual" => 1, "max_iters" => 100))
resources = AMGX.Resources(config)

A = sparse([1,1,2,2,3,3], [1,2,2,3,1,3], [4.0,1.0,4.0,1.0,1.0,4.0], 3, 3)
b = [1.0, 2.0, 3.0]

matrix = AMGX.AMGXMatrix(resources, AMGX.dDDI)
AMGX.upload!(matrix, CUDA.CUSPARSE.CuSparseMatrixCSR(A))

rhs = AMGX.AMGXVector(resources, AMGX.dDDI)
AMGX.upload!(rhs, b)

x = AMGX.AMGXVector(resources, AMGX.dDDI)
AMGX.set_zero!(x, 3)

solver = AMGX.Solver(resources, AMGX.dDDI, config)
AMGX.setup!(solver, matrix)
AMGX.solve!(x, solver, rhs)

@show AMGX.get_status(solver)
@show Vector(x)

# AMGX objects are not garbage collected: close them, children before parents
for obj in (solver, x, rhs, matrix, resources, config)
    close(obj)
end

AMGX.finalize()

See Memory management for why the close calls are necessary and how Defer.jl makes them less tedious.

Not implemented

These C API functions are not yet wrapped:

  • AMGX_read_system, AMGX_read_system_distributed
  • AMGX_write_system, AMGX_write_system_distributed
  • AMGX_config_create_from_file
  • AMGX_config_get_default_number_of_rings
  • AMGX_matrix_upload_all_global
  • AMGX_matrix_comm_from_maps, AMGX_matrix_comm_from_maps_one_ring
  • AMGX_vector_bind

The multi-GPU and distributed APIs are tracked in #8, which depends on MPI support in the JLL (#7).