This package provides a high level interface for PETSc, enabling the use of PETSc as an
A low level interface is also available in the submodule
The package supports 64-bit integers the
PetscInt type described in
the PETSc documentation, and
Complex128 for the
PetscScalar type. In a default build of the package, all types can be used
simultaneously, using multiple dispatch to determine which version of PETSc
This package requires the MPI.jl package be installed. Once it is installed you should be able to run both Julia and Petsc in parallel using MPI for all communication. The testing verifies that PETSc can be used both serially and in parallel.
To use the package, simply put
using PETSc at the top of your Julia source file. The module exports the names of all the functions, as well as the PETSc data type aliases and constants such as
In general, it is possible to run PETSc in parallel. To do so with 4 processors, do:
mpirun -np 4 julia ./name_of_file
Note that this launches 4 independent Julia processes. They are not aware of each other using Julia's built-in parallelism, and MPI is used for all communications.
To run in serial, do:
Even when running serially, the MPI.jl package must be installed.
An example of using a Krylov subspace method to solve a linear system is in
test/test_ksp.jl, which solves a simple system with a Krylov subspace method and compares the result with a direct solve using Julia's backslash operator. This works in serial and in parallel. It requires some variables declared at the top of
runtests.jl to work.
PetscVec is implemented. Some additional PETSc
BLAS functions are wrapped as well.
The AbstractArray interface for
PetscMat is implemented. Preallocation
is supported through optional keyword arguments to the matrix constructor or
setpreallocation function. It possible to set multiple values in the
matrix without intermediate assembly using the
assemble function or by
Mat object field
false and calling
Just enough KSP functions are implimented to do a GMRES solve. Adding more functionality is the current priority.
/src : source files. PETSc.jl is the main file containing initialization, with the functions for each type of Petsc object in its own file. All constants are declared in
/src/generated: auto generated wrappers from Clang.jl. Not directly useful, but easy to modify to make useful
/test : contains
runtest.jl, which does some setup and runs all tests on all three version of Petsc currently supported. Tests for each type of Petsc object (mirroring the files in
/src) are contained in separate files.
/deps : builds Petsc if needed. See description below
By default, building the package will build 3 versions of PETSc in the
directory, and writes the file
lib_locations.jl to the
directory to tell the package the location of the libraries. Note that
this builds the debug versions of PETSc, which are recommended to use for all
development. If you wish to do high performance computations, you should
build the optimized versions of the library. See the PETSc website for
If you wish to build fewer than 3 version of PETSc or to use your own build
of PETSc rather than having the package build it for you, there a several
environmental variables that control what the build system will do.
For all the variables listed below,
name is one of
ComplexDouble, and specifies which version of the library the variable
If the varibles
JULIA_PETSC_name_ARCH are set to
PETSC_ARCH of an existing PETSc installation, the build
system will use that PETSc installation for the version of PETSc specified by
If the variable
JULIA_PETSC_name_NOBUILD exists (the value does not matter),
then the package will not build a version the
named version of PETSc.
If the variable
JULIA_PETSC_OPT exists (the value does not matter), then
a set of default optimization flags are passed to the PETSc
If the variable
JULIA_PETSC_FLAGS exists and
JULIA_PETSC_OPT does not,
its value is used passed to the
PETSc configure script (for all builds). The user should never specify
--with-precision, because this
would break the build process for the different version of PETSc.
If neither of the above variables exist, a standard build is performed.
PETSc uses preprocessor variables to decide what code to include when compiling
the library. Clang does not know what preprocessor variables were defined at
compile time, so it does not correctly detect the typealiases
PetscReal, etc. To correctly autogenerate wrappers, the proper variables must be passed to Clang with the -D switch. Note that users will not need to generate their own wrappers because they have already been generated and commit to the repo.
2 months ago