GameOptizationforMobileGPU

FPS

In Android, most of the phones set display refresh rate to 60. the v-Sync signal comes at every 16.67ms. Theoretically, FPS=60/n(n=1, 2, 3)could have stable visual experience. Other FPS has short-long frame(快慢帧) issue. User feels better at stable 30 FPS than 40.

简单来说，稳定的帧率比高帧的视觉体验会好很多，应该尽量避免长短帧

• 那么长短帧又该如何度量？

Jank

It hurts the user experience, but cannot be measured from the FPS or average FPS Many reasons

• CPU bound
• GPU bound
• GC
• Loading resource
• Long task

Tips

• DCVS and how to Measure Rendering Time
  • GPU adjust its frequency based on the load, so measuring the gpu rendering time should always know what frequency it is running
• Compute Shader
  • Compute shader on mobile is not as efficient as desktop. So we always prefer fragment shader than compute shader
  • If you do need compute shader, the group size need to be tuned carefully for different tiers of GPUs
• KHR_create_context_no_error
  • This extension allows the creation of an OpenGL or Open GL ES context that doesn’t generate errors if the context supports a no error mode
  • It could save about 20% rendering thread cpu time
• Direct fallback
  • Generally, direct rendering is not power efficient as bin rendering. You need to check carefully if your main surface is fallback to direct rendering This can be found in SDP
• Load/Store
  • In binning mode use sdp to check the load/store of each surface. Avoid unnecessary load/store
  • glcear
  • gllnvalidateFramebuffer
  • Vulkan load op clear/do not care
• Shadow
  • Check and use reasonable depth texture size
• MSAA
  • 4X MSAA has performance impact, try to use 2X if your game performance is not good in 4x Mode
• UI Size
  • Limit your UI resolution to 1080p, even the screen resolution is 2K on some phones

Shader

PCF

• Adreno support HW PCF in all tiers since A3xx, the gPU can do linear filter on the 2x 2 shadow depth texture
• To use HW pCF in OpenGL:
  • glTexParameterI(GL__TEXTURE_ 2D, GL_TEXTURE_ MIN_ FILTER, GL_LINEAR);
  • glTexParameteri(GL_TEXTURE_2D, GL_TEXTURE_ MAG_FILTER, GL LINEAR);
  • glTexParameteri(GL_TEXTURE_2D, GL- TEXTURE_ COMPARE_ MODE, GL_COMPARE_REF_TO. TEXTURE)
  • glTexParameteri(GL_ TEXTURE_ 2D, GL__TEXTURE_COMPARE_FUNC, GL_LEQUAL;

Half vs. Float

The Adreno GPU could run 2x performance when calculating 16bit ALU. So the rule is:

• If your shader could use a lot of midp instead of highp use midp could improve the performance significantly
• If just several midp in your shader, and these midp need to mixed calculated with highp, just use all as high As the conversion from midp to highp also cost time
• If you are not sure, use SDP to check the compiled instruction numbers

Discard and late Z

There are some cases that prevent the GPU from using early Z, which we should avoid

• Use discard instruction in the fragment shader
• Fragment shader output depth value or sample coverage
• Depth/stencil framebuffer fetch is enabled
• Others

Texture Sampling

Desktop renderer always uses a lot of texture samplings, But this needs to be very carefully on mobile

• Please be very careful when your use texture in Vs. t is hard to determine the performance impact but likely to be slow
• In the fragment shader, the proper number of texture samplings is varying from GPU to GPU. Try your numbers and test the gPU frequency and utilization
• Always avoid generating the UV on the fly

Out of Memory

• Check the kernel log to confirm if your issue is caused by out of vss. Need root permission to run below commands on android:

  adb shell dmesg

• Below log means driver failed to allocate 8M virtual memory

  kgsl kgsl-3d0: kgs_get_unmapped_areal_get svm_area: pid 27003 mmap_ base eaed9000 addr 0 pgoff 7fa len 8388608 failed error-12

• It could happen in many functions, but most likely in draw. The reason is very complex, it could be vss fragmentation, and also could be other non-graphics module use too many VSS and etc.
• How to debug/check
  • showmap -a [pid]
  • cat /proc/[pid]/maps