Presentation by Roman Mazur, senior Android Developer at Stanfy at Android Developer Days in Ankara, June 2013. This presentation dedicated to techniques of optimization of big arrays processing using Android NDK and RenderScript.

1. Increasing performance
of big arrays processing
on Android
... cannot come up with a shorter title ...

2. The Story
Java (SDK)
C++ (NDK)
vs ?
OpenSL ES API
AudioTrack API

3. Application base concepts
short[]
track.write(...);

4. Main operation
for (int outIdx = start; outIdx < end; outIdx++)
{
// ...
output[outIdx] = source[srcIdx];
// optional: pitch shifting calculations
}

5. Implementation
● Think about avoiding GC
● Think more about avoiding GC
● Think a lot about avoiding GC
● Utilize RenderScript

6. RenderScript
● Can be used not only for images processing
● Our implementation switches from pure Java
to RenderScript when possible
● Reuse Allocations

7. And it worked fast enough! On tablets...
http://hdwpapers.com/ship_silversea_shadow_wallpaper-wallpapers.html

8. But then we ran it on budget phones...
http://www.theatlantic.com/infocus/2012/01/the-wreck-of-the-costa-concordia/100224/

9. Optimizations time!

10. http://ericlanke.blogspot.com/2012/11/too-big-to-measure.html
Measure

11. Traceview
1. Generate trace logs with
Debug.startMethodTracing("my")
or DDMS
2. View traces with
$traceview my.trace

12. Traceview

13. AspectJ
public aspect BenchmarkAspect {
pointcut benchmarkPointcuts() :
execution(* audio.Voice.fill(..));
Object around() : benchmarkPointcuts() {
storeStartTime();
try {
return proceed();
} finally {
storeEndTime(); // and write to log
}
}
}

14. AspectJ
● Allows to wrap execution of methods we are
interested in
● Works on Android with compile time
weavements

15. Our instruments
● Traceview
for detecting slowest parts of our code
● AspectJ
for more precise measuring of execution
time

16. http://www.allstuffpics.net/2011/04/wallpaper-bike-and-cars-collection-hd.html
Act

17. Main operation
for (int outIdx = start; outIdx < end; outIdx++)
{
// ...
output[outIdx] = source[srcIdx];
// optional: pitch shifting calculations
}

18. Using NDK
● Main point:
rewriting critical parts in C++ helps
● JNI calls overhead was negligible in our
case

19. Java implementation

20. C++ implementation
10 times!

21. Main operation
for (int outIdx = start; outIdx < end; outIdx++)
{
// ...
output[outIdx] = source[srcIdx];
// optional: pitch shifting calculations
}

22. Fixed point arithmetics
typedef int32_t ffloat;
#define FRAC_BITS 10
const ffloat ONE = 1 << FRAC_BITS;
static inline ffloat int_to_ffloat(jint v) {
return (ffloat)(v << FRAC_BITS);
}
static inline ffloat float_to_ffloat(jfloat v) {
register jint intpart = (jint) v;
return int_to_ffloat(intpart)
+ (ffloat)((v - intpart) * ONE);
}
25%
faster

23. After all optimizations...
GC: 14.7%

24. Solution
● Moving sample buffers to native heap
completely

25. Java side
public class Buffer {
private int nPointer;
public Buffer() { nPointer = nativeInit(); }
public void doSomething() { nativeDo(nPointer); }
protected void finalize() {
super.finalize();
nativeFinalize(nPointer);
}
private static native int nativeInit();
private static native void nativeFinalize(int np);
private static native void nativeDo(int np);
}

26. Native side
jint XXX_nativeInit() {
NativeBuffer* b = new NativeBuffer();
return (jint)b;
}
void XXX_nativeDo(jint np) {
NativeBuffer* b = (NativeBuffer*)np;
b->nativeMethodCall();
}
void XXX_finalizeNative(jint np) {
NativeBuffer* b = (NativeBuffer*)np;
delete b;
}

27. http://www.citizenship-aei.org/2012/04/americans-failing-citizenship-test-again/
Don't forget to test

28. Tests
● Robolectric based tests for major part of
audio engine components
● Android instrumentation tests for
RenderScript and native implementations.

29. Thank you!
Roman Mazur
Head of Android/Java Unit at Stanfy
mazur.roman@gmail.com
+Roman Mazur
@roman_mazur