Radeon HD 7000 Series Linux Driver Support

[calim]

no the hd7870 is the old VLIW 4D architecture

only the hd7950 and hd7970 are build in the new architecture.



the fermi only do have some features bot not all and not the key features.

the fermi do not have x86 64bit shared ram space.
the fermi do not have the new ram
the fermi is not directX 12.

and so one and so one

Fermi doesn't need weird x86 space, NV have nice working virtual memory (you can back pages with CPU and GPU memory transparently) and a 40 bit address space since GeForce 8.

And it's Direct3D 11.1, not 12 - and there's not really any great changes there ... well, they added back "logical operations in a render target", which OpenGL had since version 1.1.

[Qaridarium]

Fermi doesn't need weird x86 space,

bullshit the next nvidia gpu will have x86_64 64bit space support.
its useful because this allows the cpu to use vram as normal ram and the gpu can use the normal ram as vram.
without streaming coping data between spaces.

NV have nice working virtual memory (you can back pages with CPU and GPU memory transparently) and a 40 bit address space since GeForce 8.

yes nice "nice" but i talk about the future.

And it's Direct3D 11.1, not 12 - and there's not really any great changes there ... well, they added back "logical operations in a render target", which OpenGL had since version 1.1.

sure fermi is dx11.1 but the new amd architecture hd7970 is directX 12.

the workload in dx12 is different they drop shader based light effect and go for light raytracing light effect.
the ray-tracing will only be in the light effect all others are still raster rendering.
maybe they do object collision calculations also via ray-tracing.

directX 12 is an raster+light raytracing mix.

[glisse]

bullshit the next nvidia gpu will have x86_64 64bit space support.
its useful because this allows the cpu to use vram as normal ram and the gpu can use the normal ram as vram.
without streaming coping data between spaces.

You misunderstood the so call 64bit space support. It doesn't give anything to the cpu. For discret GPU it just means you can make the gpu address space equal to your cpu address space. But it's not completely true. NVidia have part of this since long time in form of virtual address space. But to go all the way down where GPU can access all system page of a process you need a lot more than a new GPU. You need new iommu and new pci protocol to support that.

And CPU access to VRAM is likely not what they had in mind.

Note also that most of marketing about all this is aimed at next generation APU where some of this does make lot more sense.

[curaga]

The ones that were just canceled?

[bridgman]

The ones that were just canceled?

No.

That's the turnaround time in doing work behind closed-doors first, then clearing AMD's legal review, publishing of the hardware enablement code and other features, and then repeating the process. Unlike Intel, AMD doesn't publicly work on their next-generation open-source driver support code. They push when it's ready and when the higher-ups tell them that it's okay.

Actually I was talking more about the fact that it usually takes as much or more work to get a new 3D engine up and running with any kind of acceleration than it does to go from "OK, I guess it runs but it looks awful" to "hey, it works pretty well".

[gpuhackr]

Just out of curiosity anyone here know when the ISA manual and 3D register reference will be released for the upcoming vector 7xxx series?

[Drago]

Just out of curiosity anyone here know when the ISA manual and 3D register reference will be released for the upcoming vector 7xxx series?

Most likely after the chips start shipping, and initial FOSS support lands in the kernel/mesa.

[dacresbu]

where do start with linux graphics driver dev? what kind of science (compiler theory?) or math (linear algebra?) do I need to know to help? I can't seem to find any documentation between register guides and diagrams.

[smitty3268]

where do start with linux graphics driver dev? what kind of science (compiler theory?) or math (linear algebra?) do I need to know to help? I can't seem to find any documentation between register guides and diagrams.

It depends on what parts you want to work on.

If you're looking to work on the OpenGL Mesa side of things, the devs have said you mainly need to be very familiar with OpenGL. Create some 3D apps, make sure you understand how that API works, and you should be able to dive right in. Compiler theory and matrix math is probably very helpful background, but if you can create some non-trivial OpenGL apps you probably know enough to work on Mesa.

If you're looking more at enabling features in the hardware, that's more likely to be inside the kernel and is tougher to get started in. You should probably look at a few of the simpler kernel drivers, maybe try submitting a patch or 2 for another driver first, something simpler than graphics. Once you get the hang of kernel development, working with the gpus isn't much different than any other piece of hardware.

[dacresbu]

thanks, this is enlightening. perhaps a usb driver? I had this idea to make a "target disk mode" for linux. that would require making a computer behave like a slave type usb device (usb is slave/master pattern protocol right?) Theres documentation for making usb drivers for interfacing with usb devices but I cant find docs for making usb linux devices to serve content via usb.