More AMD GPU LLVM Back-End Improvements For RadeonSI Coming

[phoronix]

Phoronix: More AMD GPU LLVM Back-End Improvements For RadeonSI Coming

Tom Stellard has more improvements coming for AMD's GPU LLVM back-end that can yield performance improvements...

http://www.phoronix.com/vr.php?view=MTgzMjE

[dungeon]

Great just tought what i can test this weekend

[schmidtbag]

haha "LLVM machine", a little redundant eh?

But anyway cool to see more progress on this. I might be getting a R9 290 soon and it seems like now (unlike a year ago) is a good time to consider it.

[mannerov]

For those not understanding what this is about:

When writing to a register it can take some time before you can read it again.
For example when you load something from memory, it can take long time.

You can still execute other instructions while the register is filled.
I think for memory reads, you have an instructions that waits the read is completed

So the goal is to change instruction order (or use more registers), to do the same thing, except making writes long before reads.

The reason it only improves performance slightly, is that SI cards have clever chips that do execute another job when a job is stalled because it is waiting for a load.
However switching job is not free, and potentially all jobs could be waiting, thus optimizing execution order can still improve performance.

I had tested old version of the patches (I'm happy work on them has resumed), and I had 10% increase for unigine heaven and 60% for dolphin emulator. I didn't test other apps.

[dungeon]

I had tested old version of the patches (I'm happy work on them has resumed), and I had 10% increase for unigine heaven and 60% for dolphin emulator. I didn't test other apps.

Testing it now... openarena/xonotic low feels the same (so it does not help in raw cases it seems), maybe slightly faster 1-2% on ultra... i wanna try xonotic high/ultra to made lockup but it does not lockup for me there .

But yeah Unigines seems all goes much up, for example Valley goes up by +11% here... nice improvment

[profoundWHALE]

haha "LLVM machine", a little redundant eh?

But anyway cool to see more progress on this. I might be getting a R9 290 soon and it seems like now (unlike a year ago) is a good time to consider it.

I got mine for half the price on ebay. Also, try to go for the non-reference cooler if possible. Otherwise you'll most likely have to get a aftermarket cooler and install it. Seriously, the fan is like vacuum cleaner when it's at 100%. Just search youtube for a sound comparison between the two. It's ridiculous.

[schmidtbag]

I got mine for half the price on ebay. Also, try to go for the non-reference cooler if possible. Otherwise you'll most likely have to get a aftermarket cooler and install it. Seriously, the fan is like vacuum cleaner when it's at 100%. Just search youtube for a sound comparison between the two. It's ridiculous.

I found 2 models for roughly $250, one in newegg (sapphire vapor-x with a rebate) and one in NCIXUS (gigabyte windforce with a promo code). Both come with a free game too. I prefer the sapphire model, since that has a better cooler, a much higher rating, it is quieter, and is overclocked a little. But yeah the reference cooler is a joke - AMD really hurt their reputation with that thing.

[dungeon]

Testing it now... openarena/xonotic low feels the same (so it does not help in raw cases it seems), maybe slightly faster 1-2% on ultra... i wanna try xonotic high/ultra to made lockup but it does not lockup for me there .

But yeah Unigines seems all goes much up, for example Valley goes up by +11% here... nice improvment

Further testing shows +1-4% in other idtech 3 engine games, i also have +5% in dhewm3 timedemo benchamrk idtech 4 game, etc.

For what is worth performance never goes down (i can't find single case yet ) only up , so very nice and steady improvment this is, kudos to Tom

[tstellar]

Testing it now... openarena/xonotic low feels the same (so it does not help in raw cases it seems), maybe slightly faster 1-2% on ultra... i wanna try xonotic high/ultra to made lockup but it does not lockup for me there .

But yeah Unigines seems all goes much up, for example Valley goes up by +11% here... nice improvment

Thanks for testing, what version of Unigine are you using, and what GPU do you have?

[dungeon]

Thanks for testing, what version of Unigine are you using, and what GPU do you have?

It is small Kabini GPU on Athlon 5350 APU . Free Basic Edition version of Unigine Valley, this https://unigine.com/products/valley/download/