This discussion has been locked.

You can no longer post new replies to this discussion. If you have a question you can start a new discussion

Why my 3d game is slow only on ARM Mali GPU ?

Hi all,

I'm new on this forum. I'm here cuz my game (minetest) is slow on only on ARM Mali-400 MP4 GPU (on this smart console http://www.jxdofficial.com/jxd-s7800b-7-inch-quad-core-game-console.html). This 3d project uses the open source engine Irrlicht-gles. I test this game on several Android smartphones/tablets but I have only a low framerate (5-8 FPS) on this GPU. A lot of users have this problem : https://forum.minetest.net/viewtopic.php?f=3&t=9389

I fixed a slow framerate in my 3d GLES2 demo with my own engine: I only replaced glBufferSubData(...) by glBufferData(...) and it works. But this fix does nothing on Irrlicht-gles.

This is the source code of the OpenGLES 2 Driver/Renderer for Irrlicht-gles : https://github.com/zaki/irrlicht/blob/ogl-es/source/Irrlicht/COGLES2Driver.cpp.

Can someone take a look on it?

ps: I'm not here to blame someone and ARM products are very awesome ;)

Julia

Parents

0 Peter Harris over 7 years ago in reply to Myy

> Im surprised you got better performances with glBufferData then SubData, as that tends to force the GPU to recreate yet another buffer instead of reusing the existing one.

If you are only doing partial replacement of a resource which still referenced by queued rendering, using glBufferSubData can force the driver to both allocate a "ghost" resource and a memcpy of the parts which have not been replaced. See:

community.arm.com/.../mali-performance-6-efficiently-updating-dynamic-resources

... for more details.
Cancel
Up 0 Down

Cancel

Reply

0 Peter Harris over 7 years ago in reply to Myy

> Im surprised you got better performances with glBufferData then SubData, as that tends to force the GPU to recreate yet another buffer instead of reusing the existing one.

If you are only doing partial replacement of a resource which still referenced by queued rendering, using glBufferSubData can force the driver to both allocate a "ghost" resource and a memcpy of the parts which have not been replaced. See:

community.arm.com/.../mali-performance-6-efficiently-updating-dynamic-resources

... for more details.
Cancel
Up 0 Down

Cancel

Children

0 Myy over 7 years ago in reply to Peter Harris

Ah ! So, while it doesn't flush the entire draw pipeline, glBufferSubData still has its drawbacks. Sometimes I wish there would be some documented GLES 2 headers with mentions of these drawbacks, so that IDE would show these informations while calling these functions.

Anyway, I guess that for better performances, small buffers and triple buffering is kind of a norm with OpenGL ES 2.x apps on Android.

While glancing at the provided code, there's some mentions of generating shadows using Stencil operations, which also might be an issue if I understand correctly.
Try to disable such operations and see if you get better FPS.

github.com/.../COGLES2Driver.cpp

community.arm.com/.../the-mali-gpu-an-abstract-machine-part-2---tile-based-rendering
Cancel
Up 0 Down

Cancel