Intel Publishes More Haswell Graphics Driver Code

[allquixotic]

The userspace bits for Trinity were published the same afternoon, as Michael hinted. Maybe look in the commit log before badmouthing us ?

http://cgit.freedesktop.org/xorg/dri...video-ati/log/
http://cgit.freedesktop.org/mesa/mesa/log/



That's because the foundation work required for GL 2.1 was being added at the same time, and the Gallium3D paths were brand new as well. That only has to happen once.

I wasn't badmouthing anyone; I was just saying that the article said the support wasn't there yet, and I figured that since Michael is so hyper about posting articles whenever someone makes a git commit, we'd hear about it in a new article if the support had actually landed. I guess I should have checked first. But I was explaining to the other guy that, going on what Michael said that it was only kernel support at that time, this wasn't enough to start using the driver.

I'll definitely take a look (or take a compile :P) of the stuff out there right now, and see if I can report some bugs... that said, I have a much less "unusual" part this time around, because I'm guessing that you guys pretty much put a lot of effort into supporting the 7970 "Tahiti" during your internal testing. I mean it would be very awkward if you spent less time on your flagship GPU than the others in the series. Last time around I had the dual GPU, which for a while was somewhat of a bastard child due to the variances in its hardware compared to the 5870. I've learned my lesson on that point.

Question: Does all the tiling work that Jerome Glisse did for Evergreen have to happen all over again for the Northern Islands series? Or will the existing tiling work provide at least a springboard with which to more rapidly bring up tiling on NI?

Just asking, because I saw the very visible performance improvements of 1D and then 2D tiling on Evergreen, and I'd wager that the NI will perform worse than Evergreen during the period where Evergreen has tiling and NI doesn't.

Then again, maybe you guys went all-out and NI already has tiling in the initial release?! I haven't trawled that far into the source yet to tell, so I figure it's easier to just ask here.

Or, heck, don't even answer me here on the forum. Instead, a constructive request! We like positive suggestions around here, don't we?

Please edit http://www.x.org/wiki/RadeonFeature (once x.org is up; the site is down for me at the moment) and update it with any changes which have occurred due to the public release of the driver.

EDIT: I erroneously called the HD7000 series "Northern Islands" in this post. It's Southern Islands. I knew that, but I'm still thinking in terms of last generation. Silly me!

[bridgman]

I wasn't badmouthing anyone; I was just saying that the article said the support wasn't there yet, and I figured that since Michael is so hyper about posting articles whenever someone makes a git commit, we'd hear about it in a new article if the support had actually landed.

Fair point

But I was explaining to the other guy that, going on what Michael said that it was only kernel support at that time, this wasn't enough to start using the driver.

Dave (aka airlied) was working on a generic modesetting driver for X which just needs KMS to operate... sort of like VESA but with native modesetting support. I don't know whether that X driver is working with the current SI KMS code but I imagine it either does work or is really close.

The generic modesetting driver is a really useful idea (something I only realized in the last couple of weeks), because it makes it worthwhile to release initial kernel driver support at the modesetting stage without waiting for accel support before you release... and in many cases modesetting code is the "safest" and therefore the best candidate for releasing *before* hardware launch.

I'll definitely take a look (or take a compile :P) of the stuff out there right now, and see if I can report some bugs... that said, I have a much less "unusual" part this time around, because I'm guessing that you guys pretty much put a lot of effort into supporting the 7970 "Tahiti" during your internal testing. I mean it would be very awkward if you spent less time on your flagship GPU than the others in the series.

It might sound funny, but we don't actually gate release of open source code on testing or bug fixing, just "does it include enough to be useful" and "have we jumped through the IP/legal hoops". Testing and bug fixing can be done better once the code is in public.

Strictly speaking the code doesn't even have to work before we make the initial release, but we normally find that making something actually work is a real good guide to identifying the information we need to review before release. We may end up pushing the userspace code for SI before it actually works, for example, depending on whether "making it work" or "being able to release" happens first.

In terms of which hardware we use for testing, the general rule is "use whatever you can lay hands on", which usually means a mix of whatever chip was first and whatever chip had the best early yields

Question: Does all the tiling work that Jerome Glisse did for Evergreen have to happen all over again for the Northern Islands series? Or will the existing tiling work provide at least a springboard with which to more rapidly bring up tiling on NI?

IIRC the tiling work was done for all generations supported by r600g at the same time, including NI, but each hardware generation tends to have some unique quirks and therefore some unique bugs. See below...

Then again, maybe you guys went all-out and NI already has tiling in the initial release?! I haven't trawled that far into the source yet to tell, so I figure it's easier to just ask here.

The general rule is that when adding new features you add to all the hardware generations supported by the driver when you start work, and when adding support for new hardware you add support for all the features & functionality that the driver supported when you start. You don't see a lot of "add this feature to only that generation" except in cases where all the earlier chips are too old to support the feature (when adding a feature) or there are radical changes in programming model which make it worth omitting a feature from the initial support in order to get the rest of the functionality into users hands more quickly (when adding new hardware support).

[allquixotic]

Looks like http://www.x.org/wiki/RadeonFeature is back online; it was down earlier.

Can someone very familiar with the recent HD7000 code, add a column to the features table for SI? And fill in the boxes to the best of your knowledge?

I'd like to know the details of which of the RadeonFeatures are supported in the latest SI code, versus NI and Evergreen.

[bridgman]

I started making changes, but quickly concluded we should wait until userspace is actually running.

[smitty3268]

I started making changes, but quickly concluded we should wait until userspace is actually running.

That's disappointing. I was hoping the code was at least up to legal review, but if it's not even running yet it sounds like we've got a while to wait for it.

[curaga]

Anyway, I'm looking forward to a fanless HD 7000-based card (7550?)

Sadly, IIRC they weren't planning any lower-end GCN cards, the 7750 is the lowest, anything below is/will be a rebranded HD 6000.
You can buy a passive 7750 though (Sapphire).