Dolphin on the pi 4?

[JonnyH]

(06-25-2019, 03:30 AM)mstreurman Wrote: The issue isn't necessarily the CPU but the GPU... as the Shield TV uses "real" OpenGL and any other ARM device uses OpenGL ES which is a subset of OpenGL that miss extensions that Dolphin needs for better performance.

The reason why the shield drivers are "faster" isn't because they expose OpenGL or OpenGL|ES, but simply that it's build from a battle-tested codebase with hundreds of devs.

As for the rpi - yes, the A72 is a *lot* faster at dolphin, but there's already plenty of devices with fast CPUs that get let down by other things (like gpu drivers etc.)

And as for a 64-bit userspace - isn't it all upstream now? Even the VideoCore driver is in mesa? So it /should/ be possible for someone knowledgeable to put something together? Right? At least nowhere near the difficulty of trying to pry info out of phone vendors

[bomblord]

(06-25-2019, 04:11 AM)JonnyH Wrote: The reason why the shield drivers are "faster" isn't because they expose OpenGL or OpenGL|ES, but simply that it's build from a battle-tested codebase with hundreds of devs.

As for the rpi - yes, the A72 is a *lot* faster at dolphin, but there's already plenty of devices with fast CPUs that get let down by other things (like gpu drivers etc.)

And as for a 64-bit userspace - isn't it all upstream now? Even the VideoCore driver is in mesa? So it /should/ be possible for someone knowledgeable to put something together? Right? At least nowhere near the difficulty of trying to pry info out of phone vendors

Thanks for the additional info!

To the bolded:That's oddly specific I take it you have experience with that xD

[sirdaniel]

Those internet announcements are flawed. I dont get why commercial sites spread so much hype about RBpi4 while logical deduction cannot follow 3x performance with 0,1 Ghz increase.

Here are some real numbers and datasheets from the raspberry pi foundation>> https://www.raspberrypi.org/documentation/hardware/raspberrypi/bcm2711/

Cpu gets overall performance score about 50% more than previous generation. Linpack test measures only floating point unit.
Cpu is 64-bit arm v8. But why gpu is 32 bit?? Same as pevious model (not sure how is that meaningful)


Hevc and h264 decode - great. H264 only 1080@60fps - not sure if that equals to 4k@25fps yet.

USB 3.0 - great.

HDMI 2.0 >> Native 4:2:0 and 10-bit/HDR transmission (ITU-R BT.2020)

GPU opengles 3.0 only....   but 100 Mgz faster and much more bandwidth

RAM 4GB - great

the rest of specification is improved too but out of my interest.

Overall great move, but 4GB model costs almost 50% higher. More ram would be needed in graphic tasks and fluid 4k movie decoding.

[bomblord]

(06-25-2019, 07:17 AM)sirdaniel Wrote: Those internet announcements are flawed. I dont get why commercial sites spread so much hype about RBpi4 while logical deduction cannot follow 3x performance with 0,1 Ghz increase.

Here are some real numbers and datasheets from the raspberry pi foundation>> https://www.raspberrypi.org/documentation/hardware/raspberrypi/bcm2711/

Cpu gets overall performance score about 50% more than previous generation. Linpack test measures only floating point unit.
Cpu is 64-bit arm v8. But why gpu is 32 bit?? Same as pevious model (not sure how is that meaningful)


Hevc and h264 decode - great. H264 only 1080@60fps - not sure if that equals to 4k@25fps yet.

USB 3.0 - great.

HDMI 2.0 >> Native 4:2:0 and 10-bit/HDR transmission (ITU-R BT.2020)

GPU opengles 3.0 only....   but 100 Mgz faster and much more bandwidth

RAM 4GB - great

the rest of specification is improved too but out of my interest.

Overall great move, but 4GB model costs almost 50% higher. More ram would be needed in graphic tasks and fluid 4k movie decoding.

Plenty of benchmarks showing CPU performance gains well above 50% it depends on the task.

Spoiler: (Show Spoiler)

But also some showing closer to 50%

Spoiler: (Show Spoiler)

It really depends on the load but It's running a much more efficient processor so looking at it from the perspective of it "only" being .1ghz increase seems strange to me.

[TKSilver]

(06-25-2019, 07:17 AM)sirdaniel Wrote: Those internet announcements are flawed. I dont get why commercial sites spread so much hype about RBpi4 while logical deduction cannot follow 3x performance with 0,1 Ghz increase.

Here are some real numbers and datasheets from the raspberry pi foundation>> https://www.raspberrypi.org/documentation/hardware/raspberrypi/bcm2711/

Cpu gets overall performance score about 50% more than previous generation. Linpack test measures only floating point unit.

It is moving from the energy efficient Cortex-A53 cores to the more performance oriented A72 ones.  This means changes like a 3 way super scaler vs a 2 way super scaler and more efficient (performance wise) out if order execution.  While it will probably not get a 3x increase in every workload there may be some that could see that due to the very slightly higher clockspeed and better IPC.  To be fair though I expect the majority of workloads to see a much more modest increase in performance.
Judging by other A72 based chipsets, personally I expect some minor improvements in emulation at the limits of what the 3b can handle (so better speed\compatibility for ps1/n64/psp) and for the emulators that are currently way outside it's ability to run to probably stay that way.  That being said I have been pleasantly surprised before by the amount of optimization the RPi community has done before getting sysytems to run better on it then they would on similarly specced hardware before, so maybe I will be wrong.

[sirdaniel]

Yea some benches are astonishing. Going to better arm v72 core made more performance. Anyway, file compression test might be also influenced by better Ram speed, what tom's hardware suggested. I would point out file reading/writing speed too as influencing factor on that particular test.

Let me explain, perspective i saw was triple performance of pi4 cpu. That was misleading what i tried to point out. Performance benefit comes from whole system.

[JonnyH]

(06-25-2019, 08:41 AM)TKSilver Wrote: This means changes like a 3 way super scaler vs a 2 way super scaler and more efficient (performance wise) out if order execution.

A53 are actually dual issue in-order cores, a72 being out-of-order - that's probably the biggest difference. I think everything these days is superscaler, it just means it's pipelined.
I seem to remember in dolphin the a57 could have a ~2x performance benefit at the same clockspeed over the a53 when running the JIT - as it doesn't generate code that really works well in-order.

An a72 isn't a lot faster, being more of a area/power/frequency update to the a57 than performance-per-clock, but it should be a significant boost compared to previous raspberry PIs.

[Helios]

So basically, you can probably look at this device as being a little slower than the Shield TV on CPU, but not by much. And massively slower than the Shield TV on GPU. As well as missing important GLES extensions we need to go fast.

The outlook does not look good. However, it's a $35 computer, and plenty of other things you all are probably into probably run nicely on it. It wouldn't be a bad purchase. The next step up that would make sense would be the Jetson Nano for $100. And that's just a mildly slower Shield TV.

[bomblord]

(06-25-2019, 03:02 PM)Helios Wrote: So basically, you can probably look at this device as being a little slower than the Shield TV on CPU, but not by much. And massively slower than the Shield TV on GPU. As well as missing important GLES extensions we need to go fast.

The outlook does not look good. However, it's a $35 computer, and plenty of other things you all are probably into probably run nicely on it. It wouldn't be a bad purchase. The next step up that would make sense would be the Jetson Nano for $100. And that's just a mildly slower Shield TV.

Can we at least expect to be able to try? I want to attempt to run some easy to run games if Dolphin does manage to at least compile and run on the device. I'm willing to pitch in if you need a device for testing

[namanix]

I've received my RP4 and will be installing it tonight. Has anyone tried to compile the dolphin source to test performance? Or is this impossible?
If no one has tried it. I will do that tonight and post some results

A lot of people in this forum think it won't be a great performance increase but there is a youtuber that actually did openGL benchmarks with performance boosts of 50% (See link with timestamp: https://youtu.be/EG5n-e7wDOQ?t=410)
I'm really looking forward to see how this device will perform with Dolphin.