通过找到将以60fps和30fps运行的最佳设置来摆脱性能设置旋钮

Question

Intro

I am developing a game where the quality of it is largley dependent on two factors.

1. How fancy the post processing effect, the fancier the better but also more gpu straining
2. How many entities can be simulated, which puts strain on the cpu.

The problem is it is a perfectly good game even without post processing which allows even pre 2010 devices to run it. And I have about 48 different levels of post processing, while most devices can run say level 10, other new devices can run level 48 processing and it looks significantly cooler.

My dilemma is I do not want my user to have to slog through an interface where the select how many entities, and tweak the two settings that effect post processing quality. This is for two reasons.

1. The main game already has plenty of interface elements on it, and I really do not want to add another. It would take away from the minimalist, and magical look of the game.
2. I had a thought to give the user a 30fps and a 60fps option (best settings that would run at that fps) that has really become a goal for me.

Question

However I am having issues implimenting my dream of a 30fps option and a 60fps option. Mostly because of two reasons that I am hoping you can help me out with.

1. Profiling takes way to long! Too many tests cause this process to take too long. Also the shorter I make the test the less accurate the results.
2. Profiling is not accurate, even though I cut out the data from the first and last 20% of the test it still gives slightly different results each time.
3. Since this is going to be an ios and android app there are SO many devices to profile, it would be difficult to program in predictions.

How would you test devices in order to figure out the best settings for 30fps and 60fps? No need to show code (articles would be nice) just explain the process.

If you want to read some more here is my current method.

My Current Method

Here is how I do it currently, and my test (although taking 3 minutes and being slightly inaccurate) has some assumptions that are making it faster.

1. Each new entity does not add significant time to GPU rendering time, meaning that the CPU and GPU tests can be done seperatly, rather then testing combinations.
2. Post processing is handled by two settings
   1. Downs-sample size: Basically the post processing effect is rendered at a lower resolution and fit onto a higher one. So if you down sample to 30% of the device resolution that means 70% less pixels do expensive calculations. The quality is also lower. I have simplified this by making 6 different levels of downsampling.
   2. Texture samples: All the processing that is done is convolution, and so a lot of texture samples need to happen for each pixel in the post processing stage. Currently for simplicity I came up with 8 different levels of this level 1 samples the texture 3 times per pixel, level 2 samples the texture 7 times per pixel etc. More samples = a more accurate and dramatic effect

However my method has some issues clouding its data, these are problems mostly because I can not predict how long these spikes will last and throw off data:

1. Every time I change one of the GPU settings there is a lag spike, so for changing texture samples the uniform causes the code to branch differently and the GPU recompiles, and for changing downsampling I have to regenerate Frame buffers.
2. There is also a lag spike at the beginning of each new CPU test because of creating new particles.
3. Profiling starts at the beginning of the launch, even though I wait two seconds some times it still isnt fully steady yet, and I will get really low results on the first test.
4. Some devices can go ridiculously high on the particle count, ironically the better the device the longer my cpu test takes.

Right now in order to test I am simply running the game as usual, and each new test I tweak the particle ammount, and gpu Ammount.

I first test the CPU, and turn post processing off. I start at 0 entities then go up by 100 entities. When the fps seems to drop below the target level (60fps then 30fps) then I test the one before it again. If this new test passed (ex: 500) but the last test failed (ex: 600) then I know 500 is the max amount of entities at that fps.

Then I test the GPU basically the same thing as the last test except I give the GPU more work each time. I broke it up into 4 tests. Luxury: Seeing if it is high end and finding the best setting, Level 6: Does level 6 texture sampling, and finds the lowest downsampling possible, Level 5: Does level 5 texture sampling and finds the lowest downsampling possible, Remedial: Does its best to find a glow setting that will work at maximum downsampling.

The texture sampling levels 5 and 6 are my goal. I dont really care how down-sampled the texture is. I would choose level 6 with down-sampling by 70% over level 5 with no down-sampling. Texture sampling levels 1-4 are remedial for the really low end devices, I only ever test these at the highest level of down-sampling. If I get something that works then that is great, if not then no glow is rendered. Texture sampling levels 7 and 8 are only for high end devices, and they only get tested with no down-sampling.

To speed up the GPU test I do it in the order Luxury-> Level 6-> Level 5->Remedial and if one of those passes it skips the other tests.

Answer 1

Your big problem is going to be thermal stability - most modern mid-end/high-end phones can generate more heat than they can dissipate if CPU and GPU are both running at maximum. They may sustain a measurably faster performance point for the initial 5-10 minutes after a cold start, but eventually will warm up enough so they need to down-clock. So your profiling problems become time variant, and you risk picking settings which work well for the first 10 minutes and then hit problems.

If you want a consistent user-experience you probably want to leave quite a bit of headroom rather than dialing things up as far as they can possibly go, and if you are leaving headroom the lack of 100% stability in the results matters less.

Personally as a user I don't find this type of setting too much of a problem - in your case it sounds like all you need is two sliders, one for CPU and GPU. I often turn down graphics deliberately below what my device is capable if the game allows to ensure my battery lasts longer (I play when I'm travelling so don't always have easy access to a charger).