The Raspbian project is a rebuild of Debian for the Raspberry Pi. adama did some benchmarks that show the improvement in going from ARMv4T with soft float to ARMv6 with hard float: http://www.memetic.org/raspbian-benchmarking-armel-vs-armhf/
The page says that there's a 4-10 % improvement on integer programs and up to 40 % on floating point programs. It's hard to tell the new instruction vs pipeline influence and if the baseline is soft float or softfp/VFP based.
-- Michael
+++ Michael Hope [2012-07-18 14:37 +1200]:
The Raspbian project is a rebuild of Debian for the Raspberry Pi. adama did some benchmarks that show the improvement in going from ARMv4T with soft float to ARMv6 with hard float: http://www.memetic.org/raspbian-benchmarking-armel-vs-armhf/
The page says that there's a 4-10 % improvement on integer programs and up to 40 % on floating point programs. It's hard to tell the new instruction vs pipeline influence and if the baseline is soft float or softfp/VFP based.
He said he used debian armel which is soft-float by default, but packages can specify a softfp/vfp build if they wish (I don't know if any do). The tests also don't say anything about controlling for the version of gcc in use. Random binaries in debian armel could have been built with an older gcc than the one used to build rasbian, depending on how long ago they were last uploaded and exactly where he was getting his packages from.
So it's hard to know how much of the improvement is due to compiling for v6 over v4t, soft-float vs vfp, armhf ABI vs armel ABI, and possibly variation in gcc used. I suspect all of those are mixed in.
Interesting nevertheless.
Wookey
That sounds similar to what Konstantinos found compiling with and without hard float ABI for ARMv7 architecture.
Dave
On 18/07/12 14:52, Wookey wrote:
+++ Michael Hope [2012-07-18 14:37 +1200]:
The Raspbian project is a rebuild of Debian for the Raspberry Pi. adama did some benchmarks that show the improvement in going from ARMv4T with soft float to ARMv6 with hard float: http://www.memetic.org/raspbian-benchmarking-armel-vs-armhf/
The page says that there's a 4-10 % improvement on integer programs and up to 40 % on floating point programs. It's hard to tell the new instruction vs pipeline influence and if the baseline is soft float or softfp/VFP based.
He said he used debian armel which is soft-float by default, but packages can specify a softfp/vfp build if they wish (I don't know if any do). The tests also don't say anything about controlling for the version of gcc in use. Random binaries in debian armel could have been built with an older gcc than the one used to build rasbian, depending on how long ago they were last uploaded and exactly where he was getting his packages from.
So it's hard to know how much of the improvement is due to compiling for v6 over v4t, soft-float vs vfp, armhf ABI vs armel ABI, and possibly variation in gcc used. I suspect all of those are mixed in.
Interesting nevertheless.
Wookey
On Wed, Jul 18, 2012 at 03:00:23PM +0100, David Rusling wrote:
So it's hard to know how much of the improvement is due to compiling for v6 over v4t, soft-float vs vfp, armhf ABI vs armel ABI, and possibly variation in gcc used. I suspect all of those are mixed in.
Agreed, but wow, the MP3 encode differences of 5x+ are dramatic (is that using lameenc?). I assume the Ogg encoder is fixed-point and therefore doesn't get the hf benefit.
On 18 July 2012 19:21, Christian Robottom Reis kiko@linaro.org wrote:
On Wed, Jul 18, 2012 at 03:00:23PM +0100, David Rusling wrote:
So it's hard to know how much of the improvement is due to compiling for v6 over v4t, soft-float vs vfp, armhf ABI vs armel ABI, and possibly variation in gcc used. I suspect all of those are mixed in.
Agreed, but wow, the MP3 encode differences of 5x+ are dramatic (is that using lameenc?). I assume the Ogg encoder is fixed-point and therefore doesn't get the hf benefit.
For all we know, some of the tested apps have hand-written assembly optimisations for armv5/6 too. That would easily speed things up by a factor of 2 or more even if no floating-point is used.
There are so many variables here, most of them unknown, that saying anything beyond "it got a bit faster" is pure speculation at best.
+++ Christian Robottom Reis [2012-07-18 15:21 -0300]:
On Wed, Jul 18, 2012 at 03:00:23PM +0100, David Rusling wrote:
So it's hard to know how much of the improvement is due to compiling for v6 over v4t, soft-float vs vfp, armhf ABI vs armel ABI, and possibly variation in gcc used. I suspect all of those are mixed in.
Agreed, but wow, the MP3 encode differences of 5x+ are dramatic (is that using lameenc?).
That's probably largely due to softfp vs vfp. We'd expect that to make a pretty dramatic difference.
I assume the Ogg encoder is fixed-point and therefore doesn't get the hf benefit.
You need to link against libtremor to get integer ogg en/decoding and few apps do, so I'd expect it to be fp too. But without the details of what was tested we can't tell from here.
Wookey
-
Christian Robottom Reis -
David Rusling -
Mans Rullgard -
Michael Hope -
Wookey