https://github.com/utkarsh530/mixedprecisiondiffeq.jl

Mixed Precision ODE solvers in Julia
https://github.com/utkarsh530/mixedprecisiondiffeq.jl

Last synced: about 1 month ago
JSON representation

Mixed Precision ODE solvers in Julia

• Host: GitHub
• URL: https://github.com/utkarsh530/mixedprecisiondiffeq.jl
• Owner: utkarsh530
• License: mit
• Created: 2023-04-27T19:41:46.000Z (over 1 year ago)
• Default Branch: main
• Last Pushed: 2024-04-16T21:30:13.000Z (9 months ago)
• Last Synced: 2024-11-10T08:34:33.402Z (2 months ago)
• Language: Julia
• Size: 639 KB
• Stars: 2
• Watchers: 2
• Forks: 1
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
[![CI](https://github.com/utkarsh530/MixedPrecisionDiffEq.jl/actions/workflows/CI.yml/badge.svg)](https://github.com/utkarsh530/MixedPrecisionDiffEq.jl/actions/workflows/CI.yml) [![codecov](https://codecov.io/gh/utkarsh530/MixedPrecisionDiffEq.jl/branch/main/graph/badge.svg?token=30TQNN2BL6)](https://codecov.io/gh/utkarsh530/MixedPrecisionDiffEq.jl)

# MixedPrecisionDiffEq.jl

Mixed Precision ODE solvers in Julia compatible with SciML ecosystem support!

# Setup

For GPU usage, ensure that you have a NVIDIA GPU working. While in the directory, run:

```

$ julia --project=. --threads=auto

julia> using Pkg

julia> Pkg.instantiate()

julia> Pkg.precompile()

```

# Usage

## Mixed precision methods in linear systems

```julia

using LinearSolve, MixedPrecisionDiffEq

prob = LinearProblem(rand(100,100), rand(100))

sol = solve(prob, MixedPrecisionLinsolve())

```

The default matrix factorization choice is `RFLUFactorization`, which is a recursive LU factorization.

However, you can simply use any factorization used by `LinearSolve.jl` by simply changing the argument as:

```julia

sol = solve(prob, MixedPrecisionLinsolve(FastLUFactorization()))

```

## Using mixed precision methods in ODEs

```julia

using OrdinaryDiffEq, MixedPrecisionDiffEq

function f!(du, u, p, t)

  du[1] = -p[1]*u[1] + p[2]*u[2]*u[3]

  du[2] = p[1]*u[1] - p[2]*u[2]*u[3] - p[3]*u[2]*u[2]

  du[3] = p[3]*u[2]*u[2]

end

u0 = [1.0,0.0,0.0]

tspan = (0.0, 1e5)

p = [0.04, 1e4, 3e7]

prob = ODEProblem(f!,u0, tspan, p)

sol = solve(prob, TRBDF2(;linsolve = MixedPrecisionLinsolve()))

```

## GPU acceleration of linear solve with mixed precision

To use GPU acceleration, run:

```julia

sol = solve(prob, TRBDF2(;linsolve = MixedPrecisionCudaOffloadFactorization()))

```

The GPU method only supports LU factorization as of now.