Radeon Gallium3D Hierarchical-Z Updated (R600)

[pixo]

Enybody else has a problem that after some time the apps die with not enought free memry?

[whizse]

It looks like Glisse is working on making r600g a lot faster, as his project.
I sure hope 8.1 has both these included -- I wonder if 50% up is a reasonable expectation?
(based on the '20-30%' for 2d tiling, and a wag of similar improvements from HiZ)

Anyhow, is there a git branch somewhere with patches applied?

Testing out Doom3 it's actually somewhat slower with HiZ than without it, which was a bit surprising. On r300g HiZ in Doom3 made quite a big difference.

[RavFX]

Save, in EVE-Online after +- 20 minutes...

The GPU started to chain soft-lock until I magic-key reboot with the yesterday kernel.

[allquixotic]

So far, so pretty good... working well with Linux Mint's cinnamon fork of gnome-shell, and I had a stable experience running Star Trek Online with wine 1.4.0-rc2. Well, for about half an hour, and then I got GPU soft lockups and had to kill X, but amazingly didn't have to reboot.

Seems like maybe an instruction got emitted that didn't bounce off the CS checker and caused the GPU to lock up, but the driver was able to reset it:

[ 3558.663953] radeon 0000:04:00.0: GPU lockup CP stall for more than 10011msec
[ 3558.663956] GPU lockup (waiting for 0x000B8073 last fence id 0x000B806B)
[ 3558.665062] radeon 0000:04:00.0: GPU softreset
[ 3558.665065] radeon 0000:04:00.0: GRBM_STATUS=0xF5700828
[ 3558.665067] radeon 0000:04:00.0: GRBM_STATUS_SE0=0x00000001
[ 3558.665069] radeon 0000:04:00.0: GRBM_STATUS_SE1=0xFC000001
[ 3558.665071] radeon 0000:04:00.0: SRBM_STATUS=0x20000AC0
[ 3558.665088] radeon 0000:04:00.0: GRBM_SOFT_RESET=0x00007F6B
[ 3558.665192] radeon 0000:04:00.0: GRBM_STATUS=0x00003828
[ 3558.665194] radeon 0000:04:00.0: GRBM_STATUS_SE0=0x00000007
[ 3558.665196] radeon 0000:04:00.0: GRBM_STATUS_SE1=0x00000007
[ 3558.665198] radeon 0000:04:00.0: SRBM_STATUS=0x200000C0
[ 3558.666206] radeon 0000:04:00.0: GPU reset succeed
[ 3558.791800] [drm] PCIE GART of 512M enabled (table at 0x0000000000040000).
[ 3558.791945] radeon 0000:04:00.0: WB enabled
[ 3558.791948] [drm] fence driver on ring 0 use gpu addr 0x40000c00 and cpu addr 0xffff88041ed7ec00
[ 3558.808610] [drm] ring test on 0 succeeded in 1 usecs
[ 3558.808626] [drm] ib test on ring 0 succeeded in 1 usecs

[curaga]

Just the other day I was wondering why the Z reject speed on my r700 sucked. Looking forward to this being in master.

[elanthis]

I despise ZCULL (NV's name for HiZ), it keeps pretending to "work" yet doesn't improve performance at all (on nv50+ that is) ... probably needs more magic. Or it's all covered up by other bottlenecks.

Then something is wrong with the driver, or as you say, other bottlenecks. Or possibly the apps you're testing are just doing one of the several things that disables early-z and hence hierarchical z-buffer optimizations.

[Nedanfor]

Just side note, both of my patch are self contains, ie you don't need any other patch, just my 2 patch (one for kernel one for mesa, i don't need a patch for ddx).

First of all: thank you, you've made my day When do you expect to submit the Mesa patch to Master?

[calim]

ZCULL

I despise ZCULL (NV's name for HiZ), it keeps pretending to "work" yet doesn't improve performance at all (on nv50+ that is) ... probably needs more magic. Or it's all covered up by other bottlenecks.

[darkbasic]

Please be patient. There is long time for Linux kernel 3.4 and Mesa8.1 to come.

Nobody needs to wait 3.4 and 8.1 to test it, also Jerome never told it will be mainlined in the 3.4 merge window (which is unlikely considering patches are floating around since one year).

Even HiZ for R300g isn't still enabled by default (and it will probably never be).

[Drago]

Great! What is the ETA of this feature?
How is the current state?
Do we have any benchmarks?

Please be patient. There is long time for Linux kernel 3.4 and Mesa8.1 to come. Michael soon will show us the present Jerome will give us all