After an epic battle the tips of each Android build are ready to be released.
Of special note is Bernhard's contribution which not only landed many patches to fix strict-aliasing issues in the code but pushed 3 changes:
skia: Add -ffast-math, -flto http://review.android.git.linaro.org/#change,773
opengl: Enable -flto and -ffast-math http://review.android.git.linaro.org/#change,774
and:
iMX53: Switch to kernel 3.1 http://review.android.git.linaro.org/#change,777
...after sorting a couple of config options that weren't correct:
<bero> -CONFIG_PAGE_OFFSET=0xC0000000 works! <bero> +CONFIG_PAGE_OFFSET=0x80000000 broken
<bero> -CONFIG_VMSPLIT_3G=y works! <bero> +CONFIG_VMSPLIT_2G=y broken
The first 2 changes have a fairly dramatic effect on the builds, making them work much better and smoothly. Although the new iMX53 build based on the 3.1 kernel needs some work, getting over this hump is a major achievement.
Here are all the tip builds. Each has been verified.
https://android-build.linaro.org/builds/~linaro-android/staging-imx53/#build... on iMX53 (3.1) https://android-build.linaro.org/builds/~linaro-android/staging-imx53/#build... on iMX53 (2.6.38) <- RC off of this https://android-build.linaro.org/builds/~linaro-android/panda/#build=351 on 4460 and 4430 https://android-build.linaro.org/builds/~linaro-android/staging-panda/#build... on 4460 and 4430 https://android-build.linaro.org/builds/~linaro-android/staging-origen/#buil... on origen https://android-build.linaro.org/builds/~linaro-android/landing-snowball/#bu... on snowball https://android-build.linaro.org/builds/~linaro-android/beagle/#build=319 on beagle_xM and beagle https://android-build.linaro.org/builds/~linaro-android/staging-snowball/#bu... snowball https://android-build.linaro.org/builds/~linaro-android/tracking-panda/#buil... on 4460 and 4430
On Sat, 22 Oct 2011, Zach Pfeffer wrote:
After an epic battle the tips of each Android build are ready to be released.
skia: Add -ffast-math, -flto http://review.android.git.linaro.org/#change,773
Good to see that ffast-math is making a difference at O3 but how does this affect your benchmarking activity ? Do we know why this difference occurs ? Is this because of extra vectorization for Floating point in skia or some other reason ?
Also, I'm sure you've considered this but I thought it better to check.
Given that the benchmarking pages plot skia results - https://wiki.linaro.org/Platform/Android/AndroidToolchainBenchmarking/2011-0...
a double check that this doesn't result in there being a spike in the results for 2011.10 would be good. If not, historical data needs to be measured for this change given that the said options weren't used while measuring the historical data.
Cheers, Ramana
---
Ramana Radhakrishnan Toolchain Working Group Linaro
On 10/22/2011 01:31 PM, Ramana Radhakrishnan wrote:
On Sat, 22 Oct 2011, Zach Pfeffer wrote:
After an epic battle the tips of each Android build are ready to be released.
skia: Add -ffast-math, -flto http://review.android.git.linaro.org/#change,773
Good to see that ffast-math is making a difference at O3 but how does this affect your benchmarking activity ? Do we know why this difference occurs ? Is this because of extra vectorization for Floating point in skia or some other reason ?
Also, I'm sure you've considered this but I thought it better to check.
Given that the benchmarking pages plot skia results - https://wiki.linaro.org/Platform/Android/AndroidToolchainBenchmarking/2011-0...
a double check that this doesn't result in there being a spike in the results for 2011.10 would be good. If not, historical data needs to be measured for this change given that the said options weren't used while measuring the historical data.
Neither the 2011.09 nor the soon to be published 2011.10 benchmarking was done using these flags. I'm not sure if I'll have the cycles to try this out this month or not. However, I'll add this to next month's testing.
On 23 October 2011 20:48, Andy Doan andy.doan@linaro.org wrote:
On 10/22/2011 01:31 PM, Ramana Radhakrishnan wrote:
On Sat, 22 Oct 2011, Zach Pfeffer wrote:
After an epic battle the tips of each Android build are ready to be released.
skia: Add -ffast-math, -flto http://review.android.git.linaro.org/#change,773
Good to see that ffast-math is making a difference at O3 but how does this affect your benchmarking activity ? Do we know why this difference occurs ? Is this because of extra vectorization for Floating point in skia or some other reason ?
Also, I'm sure you've considered this but I thought it better to check.
Given that the benchmarking pages plot skia results - https://wiki.linaro.org/Platform/Android/AndroidToolchainBenchmarking/2011-0...
a double check that this doesn't result in there being a spike in the results for 2011.10 would be good. If not, historical data needs to be measured for this change given that the said options weren't used while measuring the historical data.
Neither the 2011.09 nor the soon to be published 2011.10 benchmarking was done using these flags. I'm not sure if I'll have the cycles to try this out this month or not. However, I'll add this to next month's testing.
Cool. Thanks Andy!
a double check that this doesn't result in there being a spike in the results for 2011.10 would be good. If not, historical data needs to be measured for this change given that the said options weren't used while measuring the historical data.
Neither the 2011.09 nor the soon to be published 2011.10 benchmarking was done using these flags. I'm not sure if I'll have the cycles to try this out this month or not. However, I'll add this to next month's testing.
That is good news.
Err - I was actually suggesting not adding it or making sure that the new numbers you get with these options aren't used to plot the historical performance of the toolchain.
Ramana
On 10/24/2011 07:01 PM, Ramana Radhakrishnan wrote:
a double check that this doesn't result in there being a spike in the results for 2011.10 would be good. If not, historical data needs to be measured for this change given that the said options weren't used while measuring the historical data.
Neither the 2011.09 nor the soon to be published 2011.10 benchmarking was done using these flags. I'm not sure if I'll have the cycles to try this out this month or not. However, I'll add this to next month's testing.
That is good news.
Err - I was actually suggesting not adding it or making sure that the new numbers you get with these options aren't used to plot the historical performance of the toolchain.
agreed.
-
Andy Doan -
Ramana Radhakrishnan -
Zach Pfeffer