Installation

This page covers getting AtmosTransport.jl onto a machine and verifying the install with the synthetic-fixture test suite. No external met data is required to complete this page.

Prerequisites

Requirement	Notes
Julia	1.10 or newer; tested on 1.10 and 1.12. juliaup is the recommended installer.
Git	Standard.
Disk	~2 GB for source + dependencies + precompiled artifacts. Real meteorological datasets are separate (tens to hundreds of GB).
GPU backend (optional)	NVIDIA CUDA or Apple Silicon Metal can enable GPU acceleration. The CPU backend works without any GPU drivers.

Metal runs require Float32 numerics. The runtime rejects Metal with float_type = "Float64" before model construction.

Get the code

git clone https://github.com/RemoteSensingTools/AtmosTransport.git
cd AtmosTransport

Install Julia dependencies

The repository pins its dependencies via Project.toml and Manifest.toml. Instantiate them once:

julia --project=. -e 'using Pkg; Pkg.instantiate()'

Dependencies include KernelAbstractions, NCDatasets, FFTW, StaticArrays, and (as weak deps) CUDA and Metal. Plan on a ~5–10 minute first-time precompile.

Verify the install

The core test suite runs end-to-end on synthetic fixtures — no external met data needed. It exercises the preprocessor, the runtime, and the diagnostic tools across all three grid topologies.

julia --project=. -e 'using Pkg; Pkg.test()'

Expected output (abbreviated):

Test Summary:                | Pass  Total  Time
Met source trait             |   18     18  0.2s
IdentityRegrid passthrough   |   15     15  3.7s
GEOS reader                  |   48     48  8.3s
GEOS → CS passthrough        | 3467   3467 11.4s
GEOS convection wiring       |   26     26 12.8s
…

If Pkg.test() finishes green, the install is sound and you can proceed to First run.

Optional: GPU stack

If you have a GPU and want runtime acceleration, use the backend selector in your run TOML:

[architecture]
use_gpu = true
backend = "auto"   # or "cuda" / "metal"

[numerics]
float_type = "Float32"

The driver script (scripts/run_transport.jl) preloads CUDA or Metal before AtmosTransport when needed.

To verify the GPU is visible to Julia:

julia --project=. -e 'using CUDA; CUDA.versioninfo()'
julia --project=. -e 'using Metal; Metal.versioninfo()'

CUDA can run either Float32 or Float64 depending on the hardware and validation goal:

# in your run config TOML
[architecture]
use_gpu    = true
backend    = "cuda"

[numerics]
float_type = "Float32"   # or "Float64" on CUDA / CPU

Note that the L40S and most consumer / data-center GPUs lack first-class F64 throughput; F32 is the recommended default for production runs.

Optional: multi-threaded I/O

For double-buffered I/O overlap (disk reads in parallel with GPU compute), start Julia with multiple threads:

julia --threads=2 --project=. scripts/run_transport.jl <config.toml>

The runtime detects available threads automatically.

What's next

• First run — invoke the runtime on an existing config.

• Inspecting output — verify a transport binary or snapshot NetCDF.