how is the speed compared to cuda based diff-rasterization?

[Erhushenshou]

Hi, thank you fr you wonderful work! I am curious about the difference in efficiency using tachi and the cuda? Do you have any starting tutorials in using tachi to do the parallelization?

[oliver-batchelor]

Hi Erhushenshou, Cuda and Taichi can be more or less the same speed for this kind of thing. You can use it more or less like a convenience wrapper over cuda, which I did for forward and backward passes. There's quite a bit of unwritten information, especially around how you use it for autodiff with torch integration, etc., and some pitfalls, too; it's rough around the edges and sometimes painful working around incomplete parts or bugs. There are a few other examples that I used for example, https://github.com/taichi-dev/taichi-nerfs (though not actively developed anymore, I think) The main strength I think Taichi has is templating with Python code (i.e. being able to recompile and use constants, lists etc. from python). Cheers, Oliver

…

On Mon, Feb 10, 2025 at 9:57 AM Erhushenshou ***@***.***> wrote: Hi, thank you fr you wonderful work! I am curious about the difference in efficiency using tachi and the cuda? Do you have any starting tutorials in using tachi to do the parallelization? — Reply to this email directly, view it on GitHub <#24>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAITRZL77LIBRV32VA7O3PL2O66LNAVCNFSM6AAAAABWZGO5GCVHI2DSMVQWIX3LMV43ASLTON2WKOZSHA2DAOJVGEZDINA> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>