5 ms·

Accelerating Conway's Game of Life Using CUDA

58 points by brendanrayw 5y ago

brendanrayw 5y ago
After having attended a few CUDA workshops at NVIDA’s latest GTC, I was inspired to continue learning CUDA on my own. To do so I decided to build John Conway’s famous “Game of Life” and use CUDA to accelerate the program. I explore multiple different CUDA techniques including managed memory, pinned memory, multiple streams, and asynchronous memory transfers.
- gtn42 5y ago
  Nice, thanks for sharing your experience!
- IdiocyInAction 5y ago
  Nice. I took a CUDA course at uni where I built a neural network and a physics simulation. Optimizing them was very involved, but ultimately very cool; I learned a ton of stuff. I'd love to work with CUDA in practice, but there's not that many jobs around.
- rrss 5y ago
  this was a fun read, thanks for sharing. FYI, the transfers from pageable memory almost certainly do not go to the storage device in your system, unless you have high memory pressure. "pageable" (as a cuda-ism) does mean that the buffer may be paged out to storage, but as a result it means that (more importantly) even if the buffer is in RAM, the GPU cannot access it directly. so for pageable copies the flow is probably not: storage → buffer in RAM → device, but rather: original buffer in RAM (inaccessible to the device) → intermediate buffer in RAM (accessible to the device) → device. also, in several places you use the term 'stack' where I think it should just be 'RAM' / main memory.
  - brendanrayw 5y ago
    Thanks for reading! I appreciate the feedback and the info, I'll keep that in mind.
- joe_the_user 5y ago
  Thanks for your effort! I really like the idea, it's similar to a more ambitious project I'm thinking of. And I do have questions Is your board a giant two-dimensional array in memory? Are your threads/kernels reading from this array and then writing back to it? Do you do synchronization to make sure reads happen before the later rights? Do you do any verification that your transition happen correctly? Do you have an estimate for time spend in - transfer from global GPU memory to each kernel, calculations in the kernel, and time spent idling through synchronization (assuming you do it).
iseanstevens 5y ago
Need to re-read in detail, but seems like great learning material. I especially liked the line including “with less power comes greater simplicity” :)
- brendanrayw 5y ago
  Thanks for reading :D
jacquesm 5y ago
Interesting. I'm assuming you are familiar with Hashlife? If not check it out, it is absolutely amazing how fast it is, and as a study in memoization maybe it will inspire you on how you can get some more mileage out of your CUDA version. https://en.wikipedia.org/wiki/Hashlife https://en.wikipedia.org/wiki/Hashlife
- buescher 5y ago
  Agreed! The CUDA implementation is nice and I was going to say "now do Hashlife" myself. Here's the original paper https://www.lri.fr/~filliatr/m1/gol/gosper-84.pdf https://www.lri.fr/~filliatr/m1/gol/gosper-84.pdf

58 points by brendanrayw 5y ago

After having attended a few CUDA workshops at NVIDA’s latest GTC, I was inspired to continue learning CUDA on my own. To do so I decided to build John Conway’s famous “Game of Life” and use CUDA to accelerate the program. I explore multiple different CUDA techniques including managed memory, pinned memory, multiple streams, and asynchronous memory transfers.

gtn42 5y ago
Nice, thanks for sharing your experience!
IdiocyInAction 5y ago
Nice. I took a CUDA course at uni where I built a neural network and a physics simulation. Optimizing them was very involved, but ultimately very cool; I learned a ton of stuff. I'd love to work with CUDA in practice, but there's not that many jobs around.
rrss 5y ago
this was a fun read, thanks for sharing. FYI, the transfers from pageable memory almost certainly do not go to the storage device in your system, unless you have high memory pressure. "pageable" (as a cuda-ism) does mean that the buffer may be paged out to storage, but as a result it means that (more importantly) even if the buffer is in RAM, the GPU cannot access it directly. so for pageable copies the flow is probably not: storage → buffer in RAM → device, but rather: original buffer in RAM (inaccessible to the device) → intermediate buffer in RAM (accessible to the device) → device. also, in several places you use the term 'stack' where I think it should just be 'RAM' / main memory.
- brendanrayw 5y ago
  Thanks for reading! I appreciate the feedback and the info, I'll keep that in mind.
joe_the_user 5y ago
Thanks for your effort! I really like the idea, it's similar to a more ambitious project I'm thinking of. And I do have questions Is your board a giant two-dimensional array in memory? Are your threads/kernels reading from this array and then writing back to it? Do you do synchronization to make sure reads happen before the later rights? Do you do any verification that your transition happen correctly? Do you have an estimate for time spend in - transfer from global GPU memory to each kernel, calculations in the kernel, and time spent idling through synchronization (assuming you do it).

iseanstevens 5y ago

Need to re-read in detail, but seems like great learning material. I especially liked the line including “with less power comes greater simplicity” :)

brendanrayw 5y ago
Thanks for reading :D

jacquesm 5y ago

Interesting. I'm assuming you are familiar with Hashlife? If not check it out, it is absolutely amazing how fast it is, and as a study in memoization maybe it will inspire you on how you can get some more mileage out of your CUDA version. https://en.wikipedia.org/wiki/Hashlife https://en.wikipedia.org/wiki/Hashlife

buescher 5y ago
Agreed! The CUDA implementation is nice and I was going to say "now do Hashlife" myself. Here's the original paper https://www.lri.fr/~filliatr/m1/gol/gosper-84.pdf https://www.lri.fr/~filliatr/m1/gol/gosper-84.pdf