RenderScript和V8支持库之间的区别

Question

While working with V8 support library and the default RenderScript runtime in Android, I noticed that the same RenderScript code at the Java level can produce different results.

For example, there were slight differences in the resulting images, probably because of the different implementation of some intrinsics.

I also noticed that the support library version is "more asynchronous" than the native one. The native version would take some time to process "forEach()" methods, while the support library would return almost instantly after a "forEach()" call and do all the work when "copyTo" or "finish()" was called.

I also noticed that when " Allocation.createFromBitmap() " was called, the support library would use the bitmap memory and not allocate extra space outside the VM, while the native version would always allocate more space. I came to this conclusion using Qualcom's Trepn profiler .

Are the differences documented somewhere?

Answer 1

RenderScript support library is essentially the same as the native RenderScript in CPU path.

There are several differences though:

1. Certain features that depend on private Android APIs (non-NDK API) do not exist in the support lib.
2. The kernels (.rs) files are pre-compiled for the support lib, while the native RenderScript will do a JIT compilation on device.
3. The support lib is also able to execute in the native RenderScript path, if the the Android version on the device is higher than both "renderscriptTargetApi" and "targetSdkVersion".

The differences in memory usage and other behaviors may come from the GPU driver on the device. GPU may need to copy the bitmap to GPU memory with specified strides / alignments.