[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: [arm-gnu] cs 2010q1 toolchain generates slower code than 2007q3?


  • To: Maxim Kuvyrkov <maxim@xxxxxxxxxxxxxxxx>
  • Subject: Re: [arm-gnu] cs 2010q1 toolchain generates slower code than 2007q3?
  • From: Jauder Ho <codecrumb@xxxxxxxxx>
  • Date: Fri, 25 Jun 2010 13:34:19 -0700

Maxim,

I just tried compiling and testing as you requested and do see a variance.

Code is from http://www.netlib.org/benchmark/linpackc. Only changes are to
comment out getrusage() and change NTIMES to 100

This is with the 2007toolchain and compiling with -Os -mcpu=cortex-a8
-mtune=cortex-a8 -march=armv7-a -mfpu=neon -mfloat-abi=softfp just to
establish what I am seeing with 2007.

Unrolled Single  Precision 24700 Kflops ; 100 Reps


All following results are with the 2010q toolchain.

For performance:
* -O2 -mcpu=cortex-a8

Unrolled Single  Precision 12530 Kflops ; 100 Reps

* -O2 -mcpu=cortex-a8 -mfpu=neon -mfloat-abi=softfp

Unrolled Single  Precision 15431 Kflops ; 100 Reps


For size:
* -Os -mcpu=cortex-a8 -mthumb

Unrolled Single  Precision 11369 Kflops ; 100 Reps

* -Os -mcpu=cortex-a8 -mthumb -mfpu=neon -mfloat-abi=softfp

Unrolled Single  Precision 18070 Kflops ; 100 Reps



As you can see, there is quite a bit of variance.




On Wed, Jun 23, 2010 at 11:13 AM, Maxim Kuvyrkov <maxim@xxxxxxxxxxxxxxxx>wrote:

> On 6/22/10 11:20 AM, codecrumb wrote:
> ...
>
>  As a quick test, I decided to use c linpack. Since the Pre is a
>> cortex-a8, I ended up compiling with "-Os -mcpu=cortex-a8
>> -mtune=cortex-a8 -march=armv7-a -mfpu=neon -mfloat-abi=softfp".
>>
>
> Could you give a try to the following set of options:
>
> For performance:
> * -O2 -mcpu=cortex-a8
> * -O2 -mcpu=cortex-a8 -mfpu=neon -mfloat-abi=softfp
>
> For size:
> * -Os -mcpu=cortex-a8 -mthumb
> * -Os -mcpu=cortex-a8 -mthumb -mfpu=neon -mfloat-abi=softfp
>
> If there are significant (>5%) speed of size degradation on any of these
> options, please let us know.  As usual, initial analysis of the regressions
> are much appreciated.
>
> Thank you,
>
> --
> Maxim Kuvyrkov
> CodeSourcery
> maxim@xxxxxxxxxxxxxxxx
> (650) 331-3385 x724
>