Marek Has Been Taking To AMDGPU LLVM Optimizations

[lem79]

Hmm, seems my other reply might have disappeared.

I was saying how I loved the GPUTach feature on my Vega 56 (Sapphire). I hope that feature continues to appear on future AMD cards. I can easily tell how loaded the GPU is by looking down at my case and seeing how many LEDs are lit up. Quite a few times I've been surprised that the GPU isn't that busy, despite the framerate not being as high as I expect.

[oleyska]

Hmm, seems my other reply might have disappeared.

I was saying how I loved the GPUTach feature on my Vega 56 (Sapphire). I hope that feature continues to appear on future AMD cards. I can easily tell how loaded the GPU is by looking down at my case and seeing how many LEDs are lit up. Quite a few times I've been surprised that the GPU isn't that busy, despite the framerate not being as high as I expect.

are you using windows sometimes when you notice it?
I run "Chill" in windows most of the time and my graphics card at 50%+ power limit at 1670\\1100 (watercooled) and it's on holiday most of the time with chill on.

[cb88]

From the exact link you posted: "So CPU bound for sure"
50 % CPU simply means that only half of the available cores are maxed out. Using one of 8 available CPU cores would give you 12.5 % CPU utilization.
Also in the thread it's mentioned, that the GPU does not even switch to maximum clock rate, indicating that it doesn't get new commands fast enough to warrant that.

You obviously didn't look at the screenshots. They should peak usages on each core of around 40-60 percent none of them were hitting 100% as they would if CPU bound.

[oleyska]

You obviously didn't look at the screenshots. They should peak usages on each core of around 40-60 percent none of them were hitting 100% as they would if CPU bound.

Wait times on Ryzen as memory latency may be an issue ?
Can anyone with Vega and not ryzen and some fancy Intel ringbus cpu (KL, SL, CFL etc) test?

[Azpegath]

Sounds like more good work from marek!

Another thing that I would really like AMD/Marek to look at was to "fix" the percentage difference between Nvidia and AMD for CPU bound scenarios. It feels like that difference could resonate to other, non cpu-bound, scenarios, causing a measurable decrease in performance. An overhead is always an overhead, even though it might get smaller and smaller, at least relatively, when changing use case.

[cb88]

Wait times on Ryzen as memory latency may be an issue ?
Can anyone with Vega and not Ryzen and some fancy Intel ringbus cpu (KL, SL, CFL etc) test?

Could be, in which case it is probably scheduler bound and not necessarily CPU bound. Either that or someone should post results of 2133 memory vs 3466 or 3600Mhz.

[shmerl]

Could be, in which case it is probably scheduler bound and not necessarily CPU bound. Either that or someone should post results of 2133 memory vs 3466 or 3600Mhz.

I don't have Vega, but it works around the same with RX 480 on i7 Haswell, and Ryzen 7 1700X (GPU is maxed out, around 40 fps on 1920 x 1200).

[VikingGe]

You obviously didn't look at the screenshots. They should peak usages on each core of around 40-60 percent none of them were hitting 100% as they would if CPU bound.

That's just not true. First of all, threads aren't necessarily pinned to cores and will be moved around by the scheduler, and with CSMT and possibly mesa_glthread around, there are numerous threads involved in rendering that may require synchronization. You simply cannot tell from the CPU load numbers alone whether or not a game is CPU limited.

In Witcher 3 on wine, the GPU load on my RX 480 is as low as 20-30% in crowded areas like Beauclair, with the game being basically unplayable at 20 or less FPS. Of course that's a CPU bottleneck. And it will probably still be a CPU bottleneck if you throw a 5.0 GHz Intel chip at it, wined3d just requires ridiculous amounts of processing power. Especially in games that make heavy use of D3D11 deferred contexts.

[baka0815]

Would be nice, if the lighting problem in HITMAN (2016) would be fixed in LLVM for radeonsi (AMD R9 280).
The bug exists since LLVM 4.0 and is still present in 5.x.

[shmerl]

In Witcher 3 on wine, the GPU load on my RX 480 is as low as 20-30% in crowded areas like Beauclair, with the game being basically unplayable at 20 or less FPS. Of course that's a CPU bottleneck.

Multiple threads shouldn't starve the CPU if you have something like eight core CPU with SMT. Did you observe CPU utilization? Can you please post your benchmarks here: https://www.gamingonlinux.com/forum/topic/2753