NVIDIA Has Someone Working On Nouveau CUDA Support

[brent]

Same here.
Although (thought experiment to follow). nouveau being gallium based, how would they go about implementing CUDA for nouveau? AFAIK the answer is "State tracker"; which is shared between drivers. CUDA could become cross vendor in OSS world...
Oh, and hell is apparently freezing over...

They probably won't do that. Nvidia just wants an open source kernel driver to simplify things for people working with embedded Linux.

[Ancurio]

Same here.
Although (thought experiment to follow). nouveau being gallium based, how would they go about implementing CUDA for nouveau? AFAIK the answer is "State tracker"; which is shared between drivers. CUDA could become cross vendor in OSS world...
Oh, and hell is apparently freezing over...

You're getting confused by the ambiguous naming schemes of FOSS graphics components. "noveau" is the name of both the user space gallium driver and the kernel driver. This work by the Nvidia engineer is about the latter. It's imaginable that if they can beat noveau into shape so it gives their CUDA blobs the functionality they require, one could run the same binaries on Linux simply with a different kernel backend.

[Dukenukemx]

NVIDIAs newest Tegra chips aren't only for mobile platforms, but also targeted at the automotive market for computer vision, self driving cars etc. For that you need CUDA.

http://www.nvidia.com/object/tegra-x1-processor.html

If Nvidia hardware was in cars the car would be stuck at 30 Mph and would require special Nvidia gas that no other car could use. I was thinking of putting a Raspberry Pi 2 in my car cause Carputer, and last I checked there's no Nvidia graphics in that. The Jetson Tk1 is $200. For that much I'd rather use a real computer.

[ZeDestructor]

If Nvidia hardware was in cars the car would be stuck at 30 Mph and would require special Nvidia gas that no other car could use. I was thinking of putting a Raspberry Pi 2 in my car cause Carputer, and last I checked there's no Nvidia graphics in that. The Jetson Tk1 is $200. For that much I'd rather use a real computer.

It's already in cars, and has been for a while: see Tesla cars and the new Audi TT, with more Audi cars planned to be Tegra-powered. They work just fine, and quite invisibly from the end user.

[SXX⁣]

You're getting confused by the ambiguous naming schemes of FOSS graphics components. "noveau" is the name of both the user space gallium driver and the kernel driver. This work by the Nvidia engineer is about the latter. It's imaginable that if they can beat noveau into shape so it gives their CUDA blobs the functionality they require, one could run the same binaries on Linux simply with a different kernel backend.

Fortunately there is one huge problem with it for Nvidia. Code that can be only used by blob in userspace will never be merged into mainline. So they'll have to improve open source user space driver so it's can use this work too.

[Ancurio]

Fortunately there is one huge problem with it for Nvidia. Code that can be only used by blob in userspace will never be merged into mainline. So they'll have to improve open source user space driver so it's can use this work too.

Is the granularity for that rule really at the ioctl level? I thought it was "an entire kernel driver with no FOSS userspace part is not welcome", I'd have assumed a few more ioctls that the official FOSS driver doesn't immediately use would be fine.

[MuPuF]

Is the granularity for that rule really at the ioctl level? I thought it was "an entire kernel driver with no FOSS userspace part is not welcome", I'd have assumed a few more ioctls that the official FOSS driver doesn't immediately use would be fine.

No, it is not OK at all. All the ioctls need to have an open source user that can be use to track down regressions.

[efikkan]

As a FOSS proponent you sould rather ask for proper OpenCL support for NVidia GPUs than promoting CUDA.

The CUDA compiler infrastructure is actually open. Back in the days CUDA 2.0 was actually the inspiration for OpenCL, and since then CUDA have progressed very rapidly. I wish we could have a new OpenCL version similar to CUDA 6 or 7, but I don't see it coming any time soon.

If Nvidia hardware was in cars the car would be stuck at 30 Mph and would require special Nvidia gas that no other car could use. I was thinking of putting a Raspberry Pi 2 in my car cause Carputer, and last I checked there's no Nvidia graphics in that. The Jetson Tk1 is $200. For that much I'd rather use a real computer.

You really don't understand what devkits are for, do you?

My 2014 BMW runs a Nvidia Tegra, I can asure you it goes way faster than 30 MPH.

[jntesteves]

My 2014 BMW runs a Nvidia Tegra, I can asure you it goes way faster than 30 MPH.

I think you missed his point, that Nvidia cripples it's own products to make them limited and non-interoperable with products from competitors.

[efikkan]

I think you missed his point, that Nvidia cripples it's own products to make them limited and non-interoperable with products from competitors.

No, I got it. It's the same bashing of Nvidia over and over again. The simple truth is noone offers better support for open standards than Nvidia.