CompuBench RS RenderScript Benchmark

Hmm the results are all over the place - just looking at S4 Pro devices there's a factor of 5 or so between slowest and fastest. Can you somehow tell from the results which ones had (even if just partial) gpu acceleration?
 
Hmm the results are all over the place - just looking at S4 Pro devices there's a factor of 5 or so between slowest and fastest. Can you somehow tell from the results which ones had (even if just partial) gpu acceleration?

We use the RenderScript support library which back-ports all API features from Android 4.3 to earlier OS versions, but runs always in CPU mode on 4.2 and earlier OS-es. So there can be large performance differences on the same device with different (eg 4.2 vs 4.3) OS versions.

Please also note, the GPU acceleration we are seeing is mostly done through the (sometimes hidden) OpenCL drivers which are still not very mature and not all test kernels compile to CL easily.

We also saw some test where certain kernels run on GPU, but others on CPU at the same time! So some numbers are actually mixed-mode.

As we cannot force either CPU or GPU on consumer devices without direct access, things are a bit unpredictable currently. This is life with a young API.
 
We also saw some test where certain kernels run on GPU, but others on CPU at the same time! So some numbers are actually mixed-mode.

As we cannot force either CPU or GPU on consumer devices without direct access, things are a bit unpredictable currently. This is life with a young API.
You have Google's dick-headedness to thank for that. A RenderScript benchmark is kind of an oxymoron.

As long as Google doesn't support proper OpenCL in Android, real compute will be useless.
 
You have Google's dick-headedness to thank for that. A RenderScript benchmark is kind of an oxymoron.

As long as Google doesn't support proper OpenCL in Android, real compute will be useless.

I feel we already have enough threads debating rs, ocl, and android. If you feel that this debate needs to be revisited, please do so in those threads.
 
Back
Top