37 lines
1.5 KiB
Plaintext
37 lines
1.5 KiB
Plaintext
This project provides a build framework for TomsFastMath on Android.
|
|
|
|
Notes / TODO:
|
|
|
|
To use the faster ARM inline assembly, TFM_ARM must be defined at compile time. Unfortunately, is
|
|
incompatible with the armeabi build target because some of the instructions are not supported. The
|
|
Android SDK defines a few variables which may help with detection:
|
|
|
|
when compiled for armeabi-v7a:
|
|
__thumb2__
|
|
__ARM_ARCH_7A__
|
|
when compiled for armeabi:
|
|
__SOFTFP__
|
|
|
|
For now, the armeabi target will use the unoptimized C code. TFM_ARM is conditionally defined
|
|
for the armeabi-v7a target.
|
|
|
|
Don't forget to change FP_MAX_SIZE in src/headers/tfm.h to suit your needs, e.g.
|
|
|
|
#define FP_MAX_SIZE (16384+(8*DIGIT_BIT))
|
|
|
|
Build instructions:
|
|
|
|
(1) Clone this git repository to your project's JNI directory, naming the directory "tfm".
|
|
(2) Modify your JNI root's Android.mk and Application.mk using the files in this directory as a model.
|
|
(3) Run ndk-build (available from the Android NDK); see Gingerbread's NDK docs for more.
|
|
|
|
Known bugs / issues:
|
|
|
|
* armeabi builds are known to produce binary code that is over three times larger than armeabi-v7a.
|
|
This is probably due to either a bug in the Android NDK or limitations of the Thumb16 instruction set.
|
|
|
|
* NEON support can't really be turned on, because not all armeabi-v7a boards support it.
|
|
After some trials, it appears that gcc can auto-vectorize some loops when NEON support is enabled
|
|
and this leads to a performance increase. But doing so will cause the code to crash on phones
|
|
where NEON isn't supported...
|