AMD RadeonSI Graphics Driver Still Troublesome On Linux

[Rakot]

Hello, Alex,

I have muxless laptop with intel hd 4000 + radeon 7750M. So physically there is no display connected to discrete graphic card. I have the following results of unigine tropics benchmark (1024x768 with all high settings):

intel hd 4000                   22.7 fps
radeon 7750M (fglrx)            44.9 fps
radeon 7750M (radeonsi auto)     9.6 fps
radeon 7750M (radeonsi high)    24.1 fps

Software:
kernel 3.11.1 with dpm enabled
The rest is from today's git

With auto my radeon GPU is always in low state:

uvd    vclk: 0 dclk: 0
power level 0    sclk: 30000 mclk: 15000 vddc: 850 vddci: 900 pcie gen: 2

Only using high performance level helps me. In bug report 69395 you attached a patch. Will it help in my laptop's case?

Also do you know what is the state of dynamic switching of discrete GPU? Dave merged nouveau patches for nvidia and his branch for radeons wasn't touched for three weeks. Is there a chance to have this support in 3.12 kernel?

[FourDMusic]

Don't panic

They've got Marek.

[baffledmollusc]

Kudos Michael, that was a good set of benchmarks. Relevant variables clearly seperated out and compared, with some demanding tests included (even if most of them failed!) rather than just open source games with ridiculous frame rates.

Thanks!

[Michael]

Micheal is there any way to do a single test (for example Xonotic) when benchmarking an openbenchmarking.org ID (1309171-SO-AMDRADEON66)?
Did you receive my last e-mail? I wasn't able to submit the results...

No I didn't get any email. Michael at phoronix.com.

Not without hand editing the XML, though would be a good feature to come up with a good way to add that support, it would be easy but would just be a matter of thinking of a good way to expose/input that request to only benchmark a select portion(s) of the comparison.

[agd5f]

For us with a lack of understanding. Just what pieces are missing/needs to be done in the free driver to make perform as well as the closed driver?

Mostly porting of additional hardware and software optimizations from r600g to radeonsi. Of the top of my head:
- hyperZ support
- fast color clears
- enabling 2D tiling by default on SI (can probably done now that mesa 9.2 has been released)
- using the sDMA engines for blits and uploads
- shader compiler improvements

[darkbasic]

Micheal is there any way to do a single test (for example Xonotic) when benchmarking an openbenchmarking.org ID (1309171-SO-AMDRADEON66)?
Did you receive my last e-mail? I wasn't able to submit the results...

[CrvenaZvezda]

For us with a lack of understanding. Just what pieces are missing/needs to be done in the free driver to make perform as well as the closed driver?

[agd5f]

You can see the effects of the clocks between the 3.8 and 3.11 and 3.12 tests. Prior to 3.11, the kernel left the clocks at their boot up levels which are usually very low. In 3.11, the default clocks are now programmed. The default clocks vary from board to board. Sometimes they are low like the boot up clocks, sometimes they are higher like on older dGPUs. So depending on your board, you are not likely to see much of a difference in performance if your particular board already has high default clocks. You will save power however since the clocks will be reduced when the GPU is not busy.

[MrCooper]

Tropics

What was the problem with Tropics? It works fine for me.

Also, which version of LLVM was used for these tests? 3.4 SVN is recommended for Mesa master.

[dffx]

As it seems that AMD is focusing more Linux-related efforts to the open source module, what is the implication for the future? If the FOSS module is the future of AMD drivers on Linux, does this mean that acceptable performance on the radeon module will always be 2 or 3 generations behind the latest and greatest?