Matthias Braun via llvm-dev
2016-Oct-08 01:28 UTC
[llvm-dev] [test-suite] making the test-suite succeed with "-Ofast" and "-ffp-contract=on"
> On Oct 7, 2016, at 5:56 PM, Hal Finkel via llvm-dev <llvm-dev at lists.llvm.org> wrote: > > ----- Original Message ----- >> From: "Sebastian Pop" <sebpop.llvm at gmail.com <mailto:sebpop.llvm at gmail.com>> >> To: "Renato Golin" <renato.golin at linaro.org <mailto:renato.golin at linaro.org>> >> Cc: "Kristof Beyls" <Kristof.Beyls at arm.com <mailto:Kristof.Beyls at arm.com>>, "Sebastian Paul Pop" <s.pop at samsung.com <mailto:s.pop at samsung.com>>, "llvm-dev" >> <llvm-dev at lists.llvm.org <mailto:llvm-dev at lists.llvm.org>>, "nd" <nd at arm.com <mailto:nd at arm.com>>, "Abe Skolnik" <a.skolnik at samsung.com <mailto:a.skolnik at samsung.com>>, "Clang Dev" >> <cfe-dev at lists.llvm.org <mailto:cfe-dev at lists.llvm.org>>, "Hal Finkel" <hfinkel at anl.gov <mailto:hfinkel at anl.gov>>, "Stephen Canon" <scanon at apple.com <mailto:scanon at apple.com>>, "Matthias Braun" >> <matze at braunis.de <mailto:matze at braunis.de>> >> Sent: Friday, October 7, 2016 7:34:40 PM >> Subject: [test-suite] making the test-suite succeed with "-Ofast" and "-ffp-contract=on" >> >> Hi, >> >> I would like to provide a summary of the different proposals on how >> to >> fix the test-suite to make it succeed when specifying extra CFLAGS >> "-Ofast" and "-ffp-contract=on". I would like to expose the issue >> and >> proposed ways to fix it to other potential reviewers that could >> provide extra feedback. We also need to decide which proposal (or >> combination of) to implement and commit. >> >> Proposal 1: https://reviews.llvm.org/D25277 >> modify the CMakes to compile and run each of these benchmarks twice: >> once with added CFLAGS -ffp-contract=off. Record on disk the full >> output of both runs and compare with FP_TOLERANCE. Hash the output >> of >> the run with -ffp-contract=off and exact match against the reference >> output. >> >> The good for Proposal 1: >> - changes contained in the build system: no change to the code of the >> benchmarks >> - runs benchmarks under an extra configuration with CFLAGS +>> -ffp-contract=off >> >> The bad for Proposal 1: >> - compilation time will double >> - running time on the device will double >> - build system is more complex >> - the build directory goes from 300M to 1.2G due to the extra >> reference outputs recorded under -ffp-contract=off, >> - when running test-suite over small devices it will cost 1G more >> transfer over the network. > > I prefer proposal 1 (although, to be fair, it was something I suggested). Being the the business of trying to heavily modify every benchmark that does floating-point computation, as in proposal 2, does not seem to scale well, and can't always be done regardless. > > We can make some effort to reduce the size of the problems being computed by some of the benchmarks (e.g. pollybench); I think that is reasonable and will help with the extra space requirements. That having been said, functionally speaking, our test suite is at least an order of magnitude too small, and so my sympathy is somewhat limited. We're going to have to find a way to execute the test suite in stages on smaller devices to limit the peak usage, if not because of this then because we've added a lot more test applications and benchmarks in the future. > > -Hal > >> >> Proposal 2: https://reviews.llvm.org/D25346 >> like Proposal 1, except that there are no files written to disk >> (transferred over the network from the device to the host that does >> the fpcmp and hashing), the outputs of both normal compilation and >> the >> kernel compiled under "#pragma STDC FP_CONTRACT OFF" are computed and >> compared on the device running the benchmark. The output of >> -ffp-contract=off is written to disk, and as currently done in the >> test-suite, the output is hashed and exactly matched against the >> reference output. >> >> The good for Proposal 2: >> - no modifications to CMake and Makefiles >> - no extra space to store the extra reference output >> - tests both user CFLAGS specified mode and fast-math and >> fp-contraction=off. >> >> The bad for Proposal 2: >> - compilation time will double: e.g., Polly will optimize both >> kernels, >> - memory requirements on the device will almost double: added one >> extra output array, input arrays are not modified, so no need to >> duplicate them, >> - compute time on the device will more than double: running the >> kernel >> twice, plus an extra loop over both outputs to compare with >> FP_TOLERANCE. >> - requires modifications to the code of the benchmarks: some >> benchmarks may not be easily modified and will need to be only run >> under -ffp-contract=off (as in Proposal 3.) >> >> Proposal 3: https://reviews.llvm.org/D25351 >> modify the Makefiles and CMakes to explicitly specify the flags under >> which the results will match the recorded reference output. >> >> The good for Proposal 3: >> - no modifications to the benchmarks >> - minimal modifications to the build system >> >> The bad for Proposal 3: >> - these benchmarks will not be tested with -ffp-contract=on: exact >> matching of the reference output requires -ffp-contract=off >> - adding more tests (as in Proposals 1 and 2) is actually a good >> thing >> for the test-suite >> >> I would like to invite other people to review the above proposals and >> suggest a way forward on fixing the current state of the test-suite >> when running under CFLAGS="-Ofast" and "-ffp-contract=on." Once >> consensus is achieved, I am willing to implement and follow up with >> addressing all reviews necessary to commit the change to the >> test-suite. >> >> Thank you, >> Sebastian >>- First: I don't think we can find a 100% solution for the -ffp-contract=on differences; fpcmp with tolerances won't work on the output of oggenc. Luckily this seems to be the only problematic benchmark today. But at least for that one I see no better solution than adding the -ffp-contract=off switch. - We should consider Polybench to be the problem here! Benchmarks that just run for a few seconds and produce hundreds of megabytes output are useless as a compiler/CPU benchmarks (time is really spend in libc, the kernel and waiting for disks). In case of well behaving benchmarks Proposal 1 is unnecessary: We can just ship the reference results together with the benchmark and use fpcmp with tolerances, we do that with most other benchmarks today. We just don't really want to do that in the case of Polybench because the output is so huge, so instead we went for just shipping a md5sum of the output which now failed in combination with floating point accuracy swings, starting this whole discussion... - Because of the nature of Polybench I'd rather see Proposal 2 implemented. Compilation time of polybench is small compared to many of the other benchmarks, if we run into memory issues we can reduce the size of the arrays to normalize runtimes (so far I have no reason to believe we do though, looking at some random polybenchs it seemed by default they create a 1024*1024 array of doubles which should only be 8Meg per array). And hey if we modify the benchmarks anyway we could also add some checksumming to the code (maybe bitcasting the doubles to integers, adjusting for endianess and XOR'ing them together is enough?) and avoid all the I/O. - I personally could live with Proposal 3 on the grounds of just declaring polybench a problematic benchmark so -ffp-contract=off is fine as a stopgap measure and relying on the fact that we have several other benchmarks that have smaller references outputs and use fpcmp correctly. Of course Proposal 2 is the saner solution here. - Matthias -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20161007/af11fbcb/attachment.html>
Hal Finkel via llvm-dev
2016-Oct-08 01:42 UTC
[llvm-dev] [test-suite] making the test-suite succeed with "-Ofast" and "-ffp-contract=on"
----- Original Message -----> From: "Matthias Braun" <matze at braunis.de> > To: "Hal Finkel" <hfinkel at anl.gov> > Cc: "Sebastian Pop" <sebpop.llvm at gmail.com>, "Sebastian Paul Pop" > <s.pop at samsung.com>, "llvm-dev" <llvm-dev at lists.llvm.org>, "Clang > Dev" <cfe-dev at lists.llvm.org>, "nd" <nd at arm.com>, "Abe Skolnik" > <a.skolnik at samsung.com> > Sent: Friday, October 7, 2016 8:28:09 PM > Subject: Re: [llvm-dev] [test-suite] making the test-suite succeed > with "-Ofast" and "-ffp-contract=on"> > On Oct 7, 2016, at 5:56 PM, Hal Finkel via llvm-dev < > > llvm-dev at lists.llvm.org > wrote: >> > ----- Original Message ----- >> > > From: "Sebastian Pop" < sebpop.llvm at gmail.com > > > > > > > To: "Renato Golin" < renato.golin at linaro.org > > > > > > > Cc: "Kristof Beyls" < Kristof.Beyls at arm.com >, "Sebastian Paul > > > Pop" > > > < > > > s.pop at samsung.com >, "llvm-dev" > > > > > > < llvm-dev at lists.llvm.org >, "nd" < nd at arm.com >, "Abe Skolnik" < > > > a.skolnik at samsung.com >, "Clang Dev" > > > > > > < cfe-dev at lists.llvm.org >, "Hal Finkel" < hfinkel at anl.gov >, > > > "Stephen Canon" < scanon at apple.com >, "Matthias Braun" > > > > > > < matze at braunis.de > > > > > > > Sent: Friday, October 7, 2016 7:34:40 PM > > > > > > Subject: [test-suite] making the test-suite succeed with "-Ofast" > > > and > > > "-ffp-contract=on" > > >> > > Hi, > > >> > > I would like to provide a summary of the different proposals on > > > how > > > > > > to > > > > > > fix the test-suite to make it succeed when specifying extra > > > CFLAGS > > > > > > "-Ofast" and "-ffp-contract=on". I would like to expose the issue > > > > > > and > > > > > > proposed ways to fix it to other potential reviewers that could > > > > > > provide extra feedback. We also need to decide which proposal (or > > > > > > combination of) to implement and commit. > > >> > > Proposal 1: https://reviews.llvm.org/D25277 > > > > > > modify the CMakes to compile and run each of these benchmarks > > > twice: > > > > > > once with added CFLAGS -ffp-contract=off. Record on disk the full > > > > > > output of both runs and compare with FP_TOLERANCE. Hash the > > > output > > > > > > of > > > > > > the run with -ffp-contract=off and exact match against the > > > reference > > > > > > output. > > >> > > The good for Proposal 1: > > > > > > - changes contained in the build system: no change to the code of > > > the > > > > > > benchmarks > > > > > > - runs benchmarks under an extra configuration with CFLAGS +> > > > > > -ffp-contract=off > > >> > > The bad for Proposal 1: > > > > > > - compilation time will double > > > > > > - running time on the device will double > > > > > > - build system is more complex > > > > > > - the build directory goes from 300M to 1.2G due to the extra > > > > > > reference outputs recorded under -ffp-contract=off, > > > > > > - when running test-suite over small devices it will cost 1G more > > > > > > transfer over the network. > > >> > I prefer proposal 1 (although, to be fair, it was something I > > suggested). Being the the business of trying to heavily modify > > every > > benchmark that does floating-point computation, as in proposal 2, > > does not seem to scale well, and can't always be done regardless. >> > We can make some effort to reduce the size of the problems being > > computed by some of the benchmarks (e.g. pollybench); I think that > > is reasonable and will help with the extra space requirements. That > > having been said, functionally speaking, our test suite is at least > > an order of magnitude too small, and so my sympathy is somewhat > > limited. We're going to have to find a way to execute the test > > suite > > in stages on smaller devices to limit the peak usage, if not > > because > > of this then because we've added a lot more test applications and > > benchmarks in the future. >> > -Hal >> > > Proposal 2: https://reviews.llvm.org/D25346 > > > > > > like Proposal 1, except that there are no files written to disk > > > > > > (transferred over the network from the device to the host that > > > does > > > > > > the fpcmp and hashing), the outputs of both normal compilation > > > and > > > > > > the > > > > > > kernel compiled under "#pragma STDC FP_CONTRACT OFF" are computed > > > and > > > > > > compared on the device running the benchmark. The output of > > > > > > -ffp-contract=off is written to disk, and as currently done in > > > the > > > > > > test-suite, the output is hashed and exactly matched against the > > > > > > reference output. > > >> > > The good for Proposal 2: > > > > > > - no modifications to CMake and Makefiles > > > > > > - no extra space to store the extra reference output > > > > > > - tests both user CFLAGS specified mode and fast-math and > > > > > > fp-contraction=off. > > >> > > The bad for Proposal 2: > > > > > > - compilation time will double: e.g., Polly will optimize both > > > > > > kernels, > > > > > > - memory requirements on the device will almost double: added one > > > > > > extra output array, input arrays are not modified, so no need to > > > > > > duplicate them, > > > > > > - compute time on the device will more than double: running the > > > > > > kernel > > > > > > twice, plus an extra loop over both outputs to compare with > > > > > > FP_TOLERANCE. > > > > > > - requires modifications to the code of the benchmarks: some > > > > > > benchmarks may not be easily modified and will need to be only > > > run > > > > > > under -ffp-contract=off (as in Proposal 3.) > > >> > > Proposal 3: https://reviews.llvm.org/D25351 > > > > > > modify the Makefiles and CMakes to explicitly specify the flags > > > under > > > > > > which the results will match the recorded reference output. > > >> > > The good for Proposal 3: > > > > > > - no modifications to the benchmarks > > > > > > - minimal modifications to the build system > > >> > > The bad for Proposal 3: > > > > > > - these benchmarks will not be tested with -ffp-contract=on: > > > exact > > > > > > matching of the reference output requires -ffp-contract=off > > > > > > - adding more tests (as in Proposals 1 and 2) is actually a good > > > > > > thing > > > > > > for the test-suite > > >> > > I would like to invite other people to review the above proposals > > > and > > > > > > suggest a way forward on fixing the current state of the > > > test-suite > > > > > > when running under CFLAGS="-Ofast" and "-ffp-contract=on." Once > > > > > > consensus is achieved, I am willing to implement and follow up > > > with > > > > > > addressing all reviews necessary to commit the change to the > > > > > > test-suite. > > >> > > Thank you, > > > > > > Sebastian > > >> - First: I don't think we can find a 100% solution for the > -ffp-contract=on differences; fpcmp with tolerances won't work on > the output of oggenc. Luckily this seems to be the only problematic > benchmark today. But at least for that one I see no better solution > than adding the -ffp-contract=off switch.I agree. With application benchmarks like oggenc, I don't see any better solution.> - We should consider Polybench to be the problem here! Benchmarks > that just run for a few seconds and produce hundreds of megabytes > output are useless as a compiler/CPU benchmarks (time is really > spend in libc, the kernel and waiting for disks). In case of well > behaving benchmarks Proposal 1 is unnecessary: We can just ship the > reference results together with the benchmark and use fpcmp with > tolerances, we do that with most other benchmarks today. We just > don't really want to do that in the case of Polybench because the > output is so huge, so instead we went for just shipping a md5sum of > the output which now failed in combination with floating point > accuracy swings, starting this whole discussion...> - Because of the nature of Polybench I'd rather see Proposal 2 > implemented. Compilation time of polybench is small compared to many > of the other benchmarks, if we run into memory issues we can reduce > the size of the arrays to normalize runtimes (so far I have no > reason to believe we do though, looking at some random polybenchs it > seemed by default they create a 1024*1024 array of doubles which > should only be 8Meg per array). And hey if we modify the benchmarks > anyway we could also add some checksumming to the code (maybe > bitcasting the doubles to integers, adjusting for endianess and > XOR'ing them together is enough?) and avoid all the I/O.I agree that, regardless, the polybench benchmarks should be modified to do less I/O.> - I personally could live with Proposal 3 on the grounds of just > declaring polybench a problematic benchmark so -ffp-contract=off is > fine as a stopgap measure and relying on the fact that we have > several other benchmarks that have smaller references outputs and > use fpcmp correctly. Of course Proposal 2 is the saner solution > here.I agree with Renato that there is a danger in the method suggested by the third proposal. In the very circumstances where the FP-contraction logic is known to be active, we'd be disabling it. We've had bugs in this logic in the past, both in the frontend and in the backend, that have at least caused crashes. We don't want the test suite to lose sensitivity to these (coverage is already not great). If FP-contraction being on is the default compiler mode, then we're more likely to run into problems in less-convenient ways than test-suite failures. -Hal> - Matthias-- Hal Finkel Lead, Compiler Technology and Programming Languages Leadership Computing Facility Argonne National Laboratory -------------- next part -------------- An HTML attachment was scrubbed... URL: <http://lists.llvm.org/pipermail/llvm-dev/attachments/20161007/207ed394/attachment.html>
Renato Golin via llvm-dev
2016-Oct-08 11:17 UTC
[llvm-dev] [test-suite] making the test-suite succeed with "-Ofast" and "-ffp-contract=on"
Proposal 4: Investigate each problematic benchmark and apply the best solution for each one of them, independently. For oggenc we may need something different. While investigating povray on a similar case (very small FP differences over very few points of the output), I noticed we were emitting NEON instructions as if they were IEEE compliant (they're not). That lead me to fix a bad compiler bug. Are we sure all the FP=on differences are *just* due to fusions? If so, then lets look at the benchmarks and make them output less garbage without resorting to hashes. I've done that to a number of tests and benchmarks already. It's quite boring, yes, but it's necessary if we want them to be meaningful. The proposal 2 is actually good for Polybench, at least for the one case where Sebastian has implemented. Yes, it doubles run time, but it's validation run time, which is part of the test, and it doesn't bloat disk/memory. Benchmark bots nowadays only run for a few iterations anyway, and even the ARM bot (the slowest) is now only taking 2hs per build. My proposal is to go through all 50 cases and propose the lowest number of solutions possible for all of them. I'm guessing this will be between 2 and 4 different cases. cheers, --renato
Sebastian Pop via llvm-dev
2016-Oct-08 13:25 UTC
[llvm-dev] [cfe-dev] [test-suite] making the test-suite succeed with "-Ofast" and "-ffp-contract=on"
On Sat, Oct 8, 2016 at 6:17 AM, Renato Golin via cfe-dev <cfe-dev at lists.llvm.org> wrote:> Proposal 4: > > Investigate each problematic benchmark and apply the best solution for > each one of them, independently. For oggenc we may need something > different. >[...]> My proposal is to go through all 50 cases and propose the lowest > number of solutions possible for all of them. I'm guessing this will > be between 2 and 4 different cases. >I like Proposal 4: we need different patches to different problems. I am sure we do not have an understanding of all the problems in the 50 currently failing benchmarks, so we will need to analyze each problem.> The proposal 2 is actually good for Polybench, at least for the one > case where Sebastian has implemented. Yes, it doubles run time, but > it's validation run time, which is part of the test, and it doesn't > bloat disk/memory.I see that handling Polybench separately as in Proposal 2 also falls under Proposal 4, as handling that benchmark separately from the other ones that may have different problems. A separate follow-up patch can link a hashing algorithm in each test of Polybench and output the hashed result to reduce I/O. If everybody agrees on starting by fixing Polybench as described in Proposal 2, I will complete the implementation of that patch, and follow-up with the hashing of the output. Thanks, Sebastian