Initial RGBA16F capabilities for vc4

[nkreeger]

Hi @anholt - here is a few things that I think are needed for https://github.com/anholt/mesa/issues/114

I am not a domain expert in drivers so I'm mostly piecing together bits that I see the v3d driver does already. Happy to iterate on this if you are open to some collaboration in helping me bridge some domain knowledge gap here.

[anholt]

Looks like a reasonable start. Note that src/gallium/driivers/vc4/kernel/is just copies of the kernel code used for the software simulator, and you'll need to modify your actual kernel drivers/gpu/drm/vc4/ code :) (I'll make sure that vc4/kernel is synced up when merging)

[nkreeger]

Oh ok - so the actual change needs to go in here? https://github.com/torvalds/linux/tree/master/drivers/gpu/drm/vc4

[nkreeger]

Or if I'm just targeting Raspberry Pi - here https://github.com/raspberrypi/linux/tree/rpi-4.14.y/drivers/gpu/drm/vc4 ?

[anholt]

Yeah, feel free to work against downstream rpi's kernel. The validation code hasn't changed in so long that we should be able to apply any patches that you send (to the dri-devel mailing list, since the kernel has an archaic contribution process) to master successfully.

[nkreeger]

Great thanks for the info - I just sent the vc4_validate.c change to dri-devel mailing list. I assume it is a good idea to keep these patches small - so I'll try tackling the HDR changes later in another patch.

[nkreeger]

@anholt PTAL when you have a chance.

I'm not sure we need the "is float texture" flag - can we just rely on the enum pipe_format format field of vc4_key? Do you just want to avoid a future if condition with a bunch of format == .... checks?

Also - it looks like we can rely on the nir_op-* commands for unpack and packing?

[anholt]

Yeah, I'd forgotten that we had the format in the key already.

Ideally we would have our texturing unpacks be in NIR, but I haven't done that yet. If we did, there would be a chance for NIR to do some useful optimizations ("oh, hey, you unpacked and then repacked the value, we can just skip that") You'll see at the end of ntq_emit_tex() that we do qir_UNPACK_8_F() for unorm8-to-float conversion, so that's the thing that would need a variant for unpacking 16f instead.

[nkreeger]

OK thanks that is helpful - to clarify: Texture packs are handled from NIR but texture unpacks are not?

[anholt]

You mean packing for TLB writes? Yeah, I moved that to NIR, and some sort of similar logic might work for texturing. (It's kind of weird because we don't get to have NIR have a different number of components returned from the texture instructions, so we'd end up having returned components that are just unused)

[anholt]

WIP moving the texture unpacking to NIR for v3d, should be straightforward to use in vc4 too:

https://gitlab.freedesktop.org/anholt/mesa/commits/v3d-lower-tex-results

[anholt]

The NIR unpacking support is now merged: https://gitlab.freedesktop.org/mesa/mesa/merge_requests/45

[nkreeger]

Thanks for the heads up -

Sorry been trying to stay low during the holiday break. I'll be back to work next week :-)

Hope you're having a nice new year so far. Nick

On Fri, Jan 4, 2019 at 4:01 PM Eric Anholt [email protected] wrote:

The NIR unpacking support is now merged: https://gitlab.freedesktop.org/mesa/mesa/merge_requests/45

— You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub https://github.com/anholt/mesa/pull/115#issuecomment-451605057, or mute the thread https://github.com/notifications/unsubscribe-auth/AASsZFxh-quyhzenT3VaDBQIs5MXvPrlks5u_-t3gaJpZM4X9QxK .