drm MSI support

[Matt Dawson]

On Friday 10 October 2008 19:30:56 Robert Noland wrote:
> On Fri, 2008-10-10 at 18:53 +0100, Matt Dawson wrote:
> > On Saturday 04 October 2008 16:39:21 Robert Noland wrote:
> > > When drm loads it will also report that it has enabled MSI.
> > >
> > > Please send me reports of what chips do/don't work.
> >
> > Yep, looking good on an X850XT:
> >
> > drm0: <ATI Radeon R480 X850 XT> on vgapci0
> > info: [drm] MSI enabled 1 message(s)
> > info: [drm] Setting GART location based on new memory map
> > info: [drm] Loading R400 Microcode
> > info: [drm] Num pipes: 4
> > info: [drm] writeback test succeeded in 1 usecs
> > drm0: [ITHREAD]
> >
> > Pre-MSI
> > 8800 FPS in texcyl demo
> > 4800 FPS in glxgears
> > 602 FPS in terrain demo
> > glxs completed OK
> >
> > With MSI
> > 7450 FPS in texcyl demo
> > 4450 FPS in glxgears
> > 598 FPS in terrain demo
> > glxs completed OK
>
> I assume that you are using drm-msi-3.patch?

The current patch from your people.FreeBSD.org page, this one:
http://people.freebsd.org/~rnoland/drm-msi.patch

> I'm a little curious why performance seems slightly lower with msi.  We
> do have to re-arm the interrupt on radeons.  Is the interrupt shared in
> the non-msi case?

No, it's on IRQ 16 on its own as far as I can see:

workstation2 /home/matt # devinfo -u                                                   
Interrupt request lines:                                                                
    0 (root0)                                                                           
    1 (atkbd0)                                                                          
    3 (root0)                                                                           
    4 (sio0)                                                                            
    5 (root0)                                                                           
    6 (fdc0)                                                                            
    7 (ppc0)                                                                            
    8 (root0)                                                                           
    9 (acpi0)                                                                           
    10-11 (root0)                                                                       
    12 (psm0)                                                                           
    13 (root0)                                                                          
    14 (ata0)                                                                           
    15 (ata1)                                                                           
    16 (vgapci0)                                                                        
    17-19 (root0)                                                                       
    20 (atapci2)                                                                        
    21 (ohci0)                                                                          
    22 (ehci0)                                                                          
    23 (atapci1)

The board is an Asus A8N-VM-CSM.

> Yes, MSI seems to only be available on PCI-E radeons, so the only point
> of testing on these cards is to ensure nothing is broken.

It doesn't seem to break anything. Don't forget we're not supposed to be using 
glxgears or any of the demos as benchmarks - case in point: I can remember 
getting framerates of 12,000+ in glxgears on a 9000 Pro a while ago on 5.x, 
which suggests to me that this number is almost meaningless. Sadly, the Unix 
variant of GLExcess does not include benchmark functionality as does it's 
Win32 counterpart, so I am unable to extrapolate performance data by running 
it. Even timing it to completion doesn't help as it doesn't run with the 
fastest speed possible (it's controllable with the A and Z keys). All I run it 
for is to ensure that GL doesn't bomb or fall back to software rendering on 
the range of operations glxs requires. Perhaps I ought to install the 
linuxulator and slam Doom3 on there to run a timedemo to test this stuff?

> > Sorry for the delay. I had to set up -CURRENT on this box as it looks
> > like it will be handy to test these Radeons from time to time.
>
> Yes, particularly for newer chips being on -CURRENT is going to be
> helpful.  I can make patches for STABLE in most cases, but I'm already
> working with several different repos / code branches, so the quickest
> best way to get the new bling is going to be on -CURRENT.

I was actually thinking more long term. All my production boxen are 7-STABLE, 
albeit with your drm patches added. It seems there's only a couple of us with 
Radeon kit who are willing/daft enough to mess around with the source, so the 
more of us with -CURRENT installed, the better chance we have of seeing this 
stuff MFC'd into releases. That, and the fact that FBSD is a doddle to multi-
boot into however many environments you wish to run means there's no excuse 
not to ;)

I'm going to try to get my hands on an R500 soon (probably something low-end 
like an X1650, although I have seen an X1950Pro going for reasonable money), 
so I should be able to at least test R200-500 based cards.
-- 
Matt Dawson

matt at chronos.org.uk
MTD15-RIPE