PathScale hackers reverse-engineer Nvidia GPUs to write a better compiler

[wmf]

http://www.hpcwire.com/features/PathScale-Looks-to-One-Up-CUDA-OpenCL-with-New-GPU-Compiler-97089024.html?viewAll=y

[daeken]

I was recently doing some research into writing compilers targeting NVidia GPUs, and I came upon some good info:

- http://blog.langly.org/2009/02/12/cuda-hacking-ptx-code/ -- Breakdown of some CUDA code compiled down to PTX assembly.

- http://wiki.github.com/laanwj/decuda/ -- Assembler and disassembler for CUDA binaries.

I'm really excited to see what people come up with compiler-wise here. GPGPU stuff is still very new; I can't wait to see what people come up with.

[manvsmachine]

I really hope that they're successful. I feel like GPGPU is too immature of a field for it to already get locked down two one or two implementations.

That said, they've got a long road ahead of them. GPU development iterates much more quickly that that of CPUs; there has been an entirely new hardware architecture (Fermi), three of four driver updates, and two major CUDA API revisions released by NVIDIA in the 6 months since I did my independent study on the topic.

[CTOPathScale → manvsmachine]

Oh we fully know the challenges we face in making a fully automatic solution.. Also keep in mind that if we're successful it would only be to Nvidia's benefit to help us. In many ways I'm glad the GPU is so volatile since it pushes things forward at a much faster pace. Legacy would only get in the way and while Fermi is a leap in technology, but it's not a complete rewrite. There's still enough familiar things on there so that we had a good start even from day 1. Most importantly is that HMPP directives are now an open standard which we'll be defining and improving along with other industry partners.

[sadiq]

I wonder if they're doing AMD GPUs too?

That should be easier, given the specs for those are freely available.

[CTOPathScale → sadiq]

We've not announced any plans for AMD at this moment. If someone who is invested heavily in AMD hardware is interested they should contact us. I'm certain we could make a great solution, but from our evaluations there's just not enough market demand.

Also your perception of what AMD has released isn't entirely correct. The low level details for things around launching a compute program on ATI Evergreen hasn't been made public yet. They've only released the graphics details and we're hoping they release the compute stuff soon. This isn't to say we couldn't figure it out, but we're focused on what customers want.