Re: [eigen] 3.0.4 coming soon, testing appreciated |
[ Thread Index |
Date Index
| More lists.tuxfamily.org/eigen Archives
]
- To: eigen@xxxxxxxxxxxxxxxxxxx
- Subject: Re: [eigen] 3.0.4 coming soon, testing appreciated
- From: Rhys Ulerich <rhys.ulerich@xxxxxxxxx>
- Date: Wed, 23 Nov 2011 14:43:00 -0600
- Dkim-signature: v=1; a=rsa-sha256; c=relaxed/relaxed; d=gmail.com; s=gamma; h=mime-version:in-reply-to:references:from:date:message-id:subject:to :content-type:content-transfer-encoding; bh=6ATX8iaTQuM7jQmoHmcrNVwS6NrEWdyR96HYZgTtRoo=; b=fAP9HutERATuVcPUMuKU3kA/9Tj9iUav0LFnEQ8vxKFtst9+WcWRfTdqinHfchI/KR pqkA7qnrJzNrpH94A0tMdQfB6IHds21ajopZpmFn//eNr4oOP+bRG9YbL10ZvxPgM/4Q J2QtIWIExnp/t9Tmg/8kynxXKRAkmqHEYijSI=
>>> I see
>>> icpc 12.0.4 20110427
>>> icpc 12.1.0 20111011
>>> both build all the 'make check' targets. I noticed some terribly long
>>> individual test times for those two builds...
>> Can you profile this?
>> Is Eigen 3.0.3 performing better?
> I'm building the outer product of {Intel 11.1 latest,
> Intel 12.1 latest} and {3.0.3 release, 3.0 tip} ...
> to get better test case timing info.
Here's gross timing information from a release build with
-fno-omit-frame-pointer executing 'EIGEN_REPEAT=100 ./check.sh
product_trmm_.' on a quiet x86_64 Linux system. Notice the repeat
count of 100. I picked those tests as they popped up as slow when I
initially ran make check on a login node earlier in this discussion.
Versions are icpc (ICC) 11.1 20100806 and icpc (ICC) 12.1.0 20111011.
Test project /g/g10/ulerich1/Build/build-3.0.3-intel-11.1
1/4 Test #215: product_trmm_1 ................... Passed 14.28 sec
2/4 Test #216: product_trmm_2 ................... Passed 17.99 sec
3/4 Test #217: product_trmm_3 ................... Passed 75.14 sec
4/4 Test #218: product_trmm_4 ................... Passed 36.13 sec
Test project /g/g10/ulerich1/Build/build-3.0.4-intel-11.1
1/4 Test #215: product_trmm_1 ................... Passed 19.55 sec
2/4 Test #216: product_trmm_2 ................... Passed 24.53 sec
3/4 Test #217: product_trmm_3 ................... Passed 87.97 sec
4/4 Test #218: product_trmm_4 ................... Passed 32.09 sec
Test project /g/g10/ulerich1/Build/build-3.0.3-intel-12.1
1/4 Test #215: product_trmm_1 ................... Passed 10.07 sec
2/4 Test #216: product_trmm_2 ................... Passed 16.13 sec
3/4 Test #217: product_trmm_3 ................... Passed 79.36 sec
4/4 Test #218: product_trmm_4 ................... Passed 35.73 sec
Test project /g/g10/ulerich1/Build/build-3.0.4-intel-12.1
1/4 Test #215: product_trmm_1 ................... Passed 12.57 sec
2/4 Test #216: product_trmm_2 ................... Passed 19.36 sec
3/4 Test #217: product_trmm_3 ................... Passed 86.88 sec
4/4 Test #218: product_trmm_4 ................... Passed 27.58 sec
Good news: Across the board, Intel 12.1 does better than Intel 11.1 on
3.0 tip (what I called 3.0.4 above).
Bad news: On all of product_trmm_{1,2,3} the performance is worse on
3.0 tip than it was on 3.0.3 on both Intel 11.1 and 12.1. In
particular, compare product_trmm_{1,2} on Intel 11.1.
I can do some more detailed profiling on product_trmm_{1,2,3} using
Intel 11.1. Any preference for with which one I start?
- Rhys