GPU-able Radiative Transfer code (GRTCODE)

Fork of GRTCODE originally written by Garret Wright ().

Overview

GRTCODE consists of a column-based line-by-line radiative transfer model capable of running on NVIDIA GPUs via CUDA-c and many-core CPUs via OpenMP. This package consists of several general purpose libraries and a set of binaries that specifically target popular radiative transfer benchmarks.

Libraries

libgrtcode_utilities.a

A utility library containing helper functions, error handlers, etc.

libgas_optics.a

A library that calculates the optical properties (i.e. optical depths) of an atmospheric column assumed to be made up of a series of "pressure levels". This library uses the HITRAN database to calculate spectra for gases including water vapor, ozone, carbon dioxide, etc. and several species of CFCs, HCFCs, and HFCs.

liblongwave.a

A four-stream longwave solver that uses a 2-term Pade approximation.

libshortwave.a

A two-stream solver that uses the delta-Eddington and Adding methods.

libgrtcode_fortran.a

FORTRAN bindings to the above libraries.

Binaries

circ

Program that calculates longwave and shortwave radiative fluxes for the CIRC cases.

rfmip-irf

Program that calculates longwave and shortwave radiative fluxes for the RFMIP-IRF offline benchmark cases.

Requirements

This repository is setup to use the GNU Build System. The libraries and binaries are written in c (ISO/IEC 9899:1999 standard) and thus requires a c compiler, such as the freely available gcc. In order to run on NVIDIA GPUs, a c++ compiler (such as g++), CUDA, and the NVCC compiler are also required. The optional FORTRAN front-end library obviously also requires a FORTRAN compiler (such as gfortran). Some of the binaries also require the netCDF library. A tarball with example input data can be downloaded from GFDL's FTP server. To run the tests, download the example data, untar in the base of the repository, and then run make check after building.

Building

CPU-only

To build using default settings, run the normal GNU Build System commands:

$ autoreconf --install
$ ./configure [--prefix <where_to_install>] [--enable-single-precision]
$ make
$ make check (this is optional but recommended)
$ make install

As usual, the default compilers and flags can be overridden by specifying CC, CPPFLAGS, CFLAGS, LDFLAGS) when running configure in the usual fashion:

$ ./configure CC=gcc CFLAGS='-O3 -fopenmp' LDFLAGS='-fopenmp'

This library can take advantage of thread-level parallelism through the use of OpenMP (the default settings run single-threaded). If your compiler supports OpenMP, make sure you activate it by setting the corresponding compiler option.

With CUDA

To build a CUDA-enabled version, run:

$ autoreconf --install
$ ./configure --enable-cuda
$ make
$ make check (this is optional, but recommended and requires downloading/untar-ing the example input data.)
$ make install

As stated above, CUDA-enable builds require NVCC and a c++ compiler. If not included in your $PATH, these (as well as NVCCFLAGS and NVCCLDFLAGS) can be set when running configure:

$ ./configure --enable-cuda CXX=g++ NVCC=/usr/local/cuda/bin/nvcc NVCCFLAGS='-g -O3' \
              NVCCLDFLAGS=-L/usr/local/cuda/lib

FORTRAN bindings

A FORTRAN front-end can be built for either the CPU-only or CUDA-enabled versions by running:

$ autoreconf --install
$ ./configure --enable-fortran
$ make
$ make check (this is optional, but recommended and requires downloading/untar-ing the example input data.)
$ make install

The default FORTRAN compiler and flags can also be overridden in the usual fashion:

./configure --enable-fortran FC="gfortran" FCFLAGS="-g -O3"