On modern versions of Android running in AOT mode
FloatMath is slower than Math. Calls to Math.sqrt(),
etc. are replaced by intrinsics which can be as small
as a single CPU opcode.
When running in interpreted mode the new
implementation is unfortunately slower, but I'm
judging this acceptable and likely to be improved
over time. This change saves a small amount of native
code.
Example timings:
Mako AOSP AOT:
Method: Original / New / Direct call to Math
ceil: 596ns / 146.ns / 111ns
sqrt: 694ns / 56ns / 25ns
Mako AOSP interpreted:
Method: Original / New / Direct call to Math
ceil: 1900ns / 2307ns / 1485ns
sqrt: 1998ns / 2603ns / 1788ns
Other calls Mako AOT:
Method: Original / New
cos: 635ns / 270ns
exp: 566ns / 324ns
floor: 604ns / 150ns
hypot: 631ns / 232ns
pow: 936ns / 643ns
sin: 641ns / 299ns
The advice to use Math directly, in preference to
FloatMath, is still good. FloatMath will be deprecated
separately.
Bug: https://code.google.com/p/android/issues/detail?id=36199
Change-Id: If07fcbd78543d13bc6d75f9743f999860e8d58d7
On modern versions of Android running in AOT mode
FloatMath is slower than Math. Calls to Math.sqrt(),
etc. are replaced by intrinsics which can be as small
as a single CPU opcode.
When running in interpreted mode the new
implementation is unfortunately slower, but I'm
judging this acceptable and likely to be improved
over time. This change saves a small amount of native
code.
Example timings:
Mako AOSP AOT:
Method: Original / New / Direct call to Math
ceil: 596ns / 146.ns / 111ns
sqrt: 694ns / 56ns / 25ns
Mako AOSP interpreted:
Method: Original / New / Direct call to Math
ceil: 1900ns / 2307ns / 1485ns
sqrt: 1998ns / 2603ns / 1788ns
Other calls Mako AOT:
Method: Original / New
cos: 635ns / 270ns
exp: 566ns / 324ns
floor: 604ns / 150ns
hypot: 631ns / 232ns
pow: 936ns / 643ns
sin: 641ns / 299ns
The advice to use Math directly, in preference to
FloatMath, is still good. FloatMath will be deprecated
separately.
Bug: https://code.google.com/p/android/issues/detail?id=36199
Change-Id: I8d1947d88b3c576643138b1df589fb9da7c1ab88
The HAVE_(BSD|GNU)_QSORT_R in AndroidConfig.h has never worked, and we
should just be using something standard anyway.
Change-Id: I784d6212f6c890a11c4af2f83d1ce2d279708652
The motivation is an API change: FloatMath is going to be
deprecated and/or removed. Performance is not the goal of
this change.
That said...
Math is faster than FloatMath with AOT compilation.
While making the change, occurances of:
{Float}Math.sqrt(x * x + y * y) and
{Float}Math.sqrt({Float}Math.pow(x, 2) + {Float}Math.pow(y, 2))
have been replaced with:
{(float)} Math.hypot(x, y)
Right now there is no runtime intrinsic for hypot so is not faster
in all cases for AOT compilation:
Math.sqrt(x * x + y * y) is faster than Math.hypot(x, y) with
AOT, but all other combinations of FloatMath, use of pow() etc.
are slower than hypot().
hypot() has the advantage of being self documenting and
could be optimized in future. None of the behavior differences
around NaN and rounding appear to be important for the cases
looked at: they all assume results and arguments are in range
and usually the results are cast to float.
Different implementations measured on hammerhead / L:
AOT compiled:
[FloatMath.hypot(x, y)]
benchmark=Hypot_FloatMathHypot} 633.85 ns; σ=0.32 ns @ 3 trials
[FloatMath.sqrt(x*x + y*y)]
benchmark=Hypot_FloatMathSqrtMult} 684.17 ns; σ=4.83 ns @ 3 trials
[FloatMath.sqrt(FloatMath.pow(x, 2) + FloatMath.pow(y, 2))]
benchmark=Hypot_FloatMathSqrtPow} 1270.65 ns; σ=12.20 ns @ 6 trials
[(float) Math.hypot(x, y)]
benchmark=Hypot_MathHypot} 96.80 ns; σ=0.05 ns @ 3 trials
[(float) Math.sqrt(x*x + y*y)]
benchmark=Hypot_MathSqrtMult} 23.97 ns; σ=0.01 ns @ 3 trials
[(float) Math.sqrt(Math.pow(x, 2) + Math.pow(y, 2))]
benchmark=Hypot_MathSqrtPow} 156.19 ns; σ=0.12 ns @ 3 trials
Interpreter:
benchmark=Hypot_FloatMathHypot} 1180.54 ns; σ=5.13 ns @ 3 trials
benchmark=Hypot_FloatMathSqrtMult} 1121.05 ns; σ=3.80 ns @ 3 trials
benchmark=Hypot_FloatMathSqrtPow} 3327.14 ns; σ=7.33 ns @ 3 trials
benchmark=Hypot_MathHypot} 856.57 ns; σ=1.41 ns @ 3 trials
benchmark=Hypot_MathSqrtMult} 1028.92 ns; σ=9.11 ns @ 3 trials
benchmark=Hypot_MathSqrtPow} 2539.47 ns; σ=24.44 ns @ 3 trials
Bug: https://code.google.com/p/android/issues/detail?id=36199
Change-Id: I06c91f682095e627cb547d60d936ef87941be692
HAVE_OOM_ADJ is really __linux__, and this code should never be run on
another platform.
Bug: 17641445
Change-Id: Ib8174106b6d03c9c2429794a4f582c09cb308646
Current implmntation only sends rotation updates on orientation
changes, so does not handle direct 0<->180 or 90<->270 transitions.
Update rotation based on an OrientationEventListener instead of
Intent.ACTION_CONFIGURATION_CHANGED
Bug 17606902
Change-Id: I01dfcd1c587f5b2e8a96365c2389782ad77936ef
Add the app directory to the arguments for starting a process.
Add a check for NeedsNativeBridge and a call to PreInitializeBridge
in the native fork code.
Change-Id: I0b93da93251c6b4638de786bf98cf99df07c3fc2
All supported locales use only U+2025 and U+2026 to represent
ellipses, and it will unlikely change in future. Given translated
resources are inconsistent and often use three dots it is safer
to use constants instead of resources.
Change-Id: I51a6cb903f62f739fbadd6b78e5765c0028d641a
The build system takes care of linking the correct STL and gtest
libraries for you, and specifying them manually confuses the build
system when using libc++.
Change-Id: I9f76ab26a63ace51293614cfb5ca002f37438e02
