Integrating ODES on GPU?

Hi there: I am having a tough time integrating odes on my GPU in MATLAB. I have tried MATLAB ode solvers which, of course, didn't work. I tried a simple Euler solver to execute through 'arrayfun' which didn't worked either. It looks like a lot of functionality is not yet supported for computation on GPUs. Please help me out if I am not aware of something or if there is an easy way out. I am running MATLAB R2011b on Windows 8. Thanks, Shiv

[Matt J]

It's hard to know what to advise, since we have no idea what you tried.

I'm not sure how the GPU could help with ODE solving, since there isn't any obvious parallel structure to the operations that that requires.

[Shiv Tewari]

Thanks, Matt. When I said integrating odes on GPU I meant integrating several uncoupled odes. Ode integration doesn't have, obviously, a parallel structure as it a iterative procedure. Hope it makes sense. I am having problem integrating uncoupled odes on GPU? PS: I tried to integrate a single ode first using MATLAB ode solvers that didn't worked at the first place.

[Matt J]

If you only have "several" ODEs, it sounds like a parallel CPU implementation, e.g., using parfor, may be more suitable than a gpu driven approach.

[Carlos Ricagno]

I have the same problem, parfor is a good option, but i need more power, If somebody knows how to implement ode in gpu would be great,

[Tuan Pham]

I am also in need of an ode in gpu and was wondering if any resources are available in MATLAB?

Answer 1

[Geng]

Higher matlab versions includes more functions, but still too few to solve an ODE or a PDE. Arrayfun is a good tool with unexpected flaws. The best way to implement an ODE solver on GPU with matlab is matlab-with-CUDA. First you write the parallel parts in CUDA (.cu file) and compell out a '.ptx' file using whatever method. Then with the .cu and .ptx file, you can obtain a matlab parallel function. You will get more from the Parallel Computing Toolbox introduction document of matlab.

[Shiv Tewari]

Thanks Geng. I am also heading to the same solution. - Shiv

Answer 2

[majid]

Have you made any progress Shiv?! I could not find any material regarding CUDA and ODE for Matlab! Any help is appreciated.

[Vipul Singhal]

Just want to echo this question. Would love to see if someone has made progress in this direction. I think doing things like parameter estimation / sensitivity analysis would be greatly accelerated if we could figure out how to run an ODE on a GPU.

[Hamid Osooli]

Try this one

https://github.com/astroHaoPeng/ode45gpu

[Gonzalo Villegas Curulla]

Dear Hamid,

Please inspect the error message below when running Test_02_RTBP.m in your github repo:

"

Error using gpuArray/arrayfun

Unable to resolve the function handle.

Error in Test_02_RTBP (line 36)

[TGPU,YGPU1,YGPU2,YGPU3,YGPU4,YGPU5,YGPU6,FLAG] =

arrayfun(@ode45GPU_simple,tspan(1),tspan(end),y0GPU(:,1),y0GPU(:,2),y0GPU(:,3),y0GPU(:,4),y0GPU(:,5),y0GPU(:,6),AbsTol,RelTol,0.001,mu,e);

"