Compilation problem with ftree-parallelize-loops

Message

Elv13 · Post by **Elv13** » Sat Jun 12, 2010 9:16 pm

Hi,

After sawing more and more use of loop parallisation on gentoo user make.conf, I tried too. But it does not work fine here, most package using it compile, but packages linking to those packages have errors like:

Code: Select all

crc32.o: In function `get_crc_table._loopfn.0':
crc32.c:(.text+0x25a): undefined reference to `omp_get_num_threads'
crc32.c:(.text+0x261): undefined reference to `omp_get_thread_num'
inflatef.o: In function `inflate_dynamic':
inflate.c:(.text+0x118a): undefined reference to `GOMP_parallel_start'
inflate.c:(.text+0x1197): undefined reference to `GOMP_parallel_end'
inflatef.o: In function `inflate_fixed':
inflate.c:(.text+0x18de): undefined reference to `GOMP_parallel_start'
inflate.c:(.text+0x18eb): undefined reference to `GOMP_parallel_end'
inflate.c:(.text+0x19de): undefined reference to `GOMP_parallel_start'
inflate.c:(.text+0x19eb): undefined reference to `GOMP_parallel_end'
inflate.c:(.text+0x1ade): undefined reference to `GOMP_parallel_start'
inflate.c:(.text+0x1aeb): undefined reference to `GOMP_parallel_end'
inflate.c:(.text+0x1c5e): undefined reference to `GOMP_parallel_start'
inflate.c:(.text+0x1c6b): undefined reference to `GOMP_parallel_end'
inflatef.o: In function `inflate_fixed._loopfn.7':
inflate.c:(.text+0x22e5): undefined reference to `omp_get_num_threads'
inflate.c:(.text+0x22ed): undefined reference to `omp_get_thread_num'
inflatef.o: In function `inflate_fixed._loopfn.6':
inflate.c:(.text+0x2365): undefined reference to `omp_get_num_threads'
inflate.c:(.text+0x236d): undefined reference to `omp_get_thread_num'
inflatef.o: In function `inflate_fixed._loopfn.5':
inflate.c:(.text+0x23e5): undefined reference to `omp_get_num_threads'
inflate.c:(.text+0x23ed): undefined reference to `omp_get_thread_num'
inflatef.o: In function `inflate_fixed._loopfn.4':
inflate.c:(.text+0x2465): undefined reference to `omp_get_num_threads'
inflate.c:(.text+0x246d): undefined reference to `omp_get_thread_num'
inflatef.o: In function `inflate_dynamic._loopfn.3':
inflate.c:(.text+0x24e7): undefined reference to `omp_get_num_threads'
inflate.c:(.text+0x24ef): undefined reference to `omp_get_thread_num'
inflatef.o: In function `inflate_dynamic._loopfn.2':
inflate.c:(.text+0x2575): undefined reference to `omp_get_num_threads'
inflate.c:(.text+0x257d): undefined reference to `omp_get_thread_num'
inflatef.o: In function `inflate_dynamic._loopfn.1':
inflate.c:(.text+0x25f6): undefined reference to `omp_get_num_threads'

I added -fopenmp, but it does not work. I tried disabling every options down to -O0. -O0 work fine, but anything more than that produce the error. Adding -fno-guess-branch-probability make -O1 work, but some package still fail with -O2. I wont post my make.conf or emerge --info, as it is the one I used to do my tests (I used an empty make.conf with only ftree-parallelize-loops and -O2). Any idea how to solve the problem?

Elv13 · Post by **Elv13** » Wed Jun 16, 2010 9:12 pm

Elv13 · Post by **Elv13** » Tue Jun 29, 2010 3:13 am

look like this does not help: (compiling glibc)

Code: Select all

r/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/elf:/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/dlfcn:/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/nss:/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/nis:/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/rt:/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/resolv:/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/crypt:/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/nptl -o /var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc.so -T /var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/shlib.lds /var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/csu/abi-note.o /var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/elf/soinit.os /var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os /var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/elf/sofini.os /var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/elf/interp.os /var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/elf/ld.so -lgcc
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `initialize_env':
unwind-pe.c:(.text+0xf3776): undefined reference to `pthread_attr_setstacksize'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_new_icv':
(.text+0xf7fd0): undefined reference to `pthread_setspecific'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `team_destructor':
unwind-pe.c:(.text+0xf800e): undefined reference to `pthread_key_delete'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `initialize_team':
unwind-pe.c:(.text+0xf8046): undefined reference to `pthread_key_create'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_team_start':
(.text+0xf85ae): undefined reference to `pthread_create'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_team_start':
(.text+0xf8818): undefined reference to `pthread_attr_getstacksize'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_team_start':
(.text+0xf8835): undefined reference to `pthread_attr_setstacksize'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_team_start':
(.text+0xf88ab): undefined reference to `pthread_setspecific'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `get_num_procs':
unwind-pe.c:(.text+0xf9193): undefined reference to `pthread_getaffinity_np'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_init_num_threads':
(.text+0xf9215): undefined reference to `pthread_getaffinity_np'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_ialias_omp_get_wtick':
(.text+0xf9aa6): undefined reference to `clock_getres'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_ialias_omp_get_wtick':
(.text+0xf9ad4): undefined reference to `clock_getres'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_ialias_omp_get_wtime':
(.text+0xf9b06): undefined reference to `clock_gettime'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_ialias_omp_get_wtime':
(.text+0xf9b34): undefined reference to `clock_gettime'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_init_thread_affinity':
(.text+0xf9ff4): undefined reference to `pthread_attr_setaffinity_np'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_init_affinity':
(.text+0xfa055): undefined reference to `pthread_getaffinity_np'
/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc_pic.os: In function `gomp_init_affinity':
(.text+0xfa1dc): undefined reference to `pthread_setaffinity_np'
collect2: ld returned 1 exit status
make[1]: *** [/var/tmp/portage/sys-libs/glibc-2.11.2/work/build-x86-x86_64-pc-linux-gnu-nptl/libc.so] Error 1
make[1]: Leaving directory `/var/tmp/portage/sys-libs/glibc-2.11.2/work/glibc-2.11.2'
make: *** [all] Error 2
 * ERROR: sys-libs/glibc-2.11.2 failed:
 *   make for x86 failed
 * 
 * Call stack:
 *           ebuild.sh, line   54:  Called src_compile
 *         environment, line 3892:  Called eblit-run 'src_compile'

I tried adding -pthread to my cflags, but no luck.

I installed a new Gentoo on my laptop using the same make.conf (with loop parallelization) and it work without issue, so there is something wrong here, but what

EndrjuX · Post by **EndrjuX** » Fri Jul 02, 2010 2:59 pm

Same problem here.

Some other apps that die with similar errors:
dev-lang/perl
media-plugins/live
sys-boot/grub
media-libs/freetype
media-gfx/splashutils
dev-java/icedtea
dev-python/PyQt4
app-office/openoffice

From emerge --info:

Code: Select all

System uname: Linux-2.6.34-gentoo-x86_64-AMD_Phenom-tm-_II_X4_955_Processor-with-gentoo-2.0.1
sys-devel/gcc:       4.5.0
CFLAGS="-O2 -march=native -pipe -floop-interchange -floop-strip-mine -floop-block -ftree-vectorize -ftree-loop-distribution -floop-parallelize-all -ftree-parallelize-loops=4"
LDFLAGS="-Wl,-O1 -Wl,--sort-common -Wl,--as-needed -fopenmp -lgomp"

As you can see I'm using graphite things, vectorization and parallelization of loops. And only parallelization is problematic. (Elv13 is using parallelization only, so it's not combination of those features) Adding "-lpthread" doesn't change anything for glibc, but some other apps die without "-fopenmp -lgomp". (Apps listed above die anyway)

I'm wondering if that whole loop parallelization thing really improves apps performance so much...

Elv13 · Post by **Elv13** » Sat Jul 03, 2010 8:05 pm

-floop-parallelize-all will probably make things worst, spliting a int i =0; i < 3 loop in 4 will be about 20x slower than just running it, if not more. Let the compiler decide if it is good or not. -lgomp is useless by the way, it will not improve things and if it override --as-needed, it will make things slower. I am also not sure about -ftree-loop-distribution, it may make things slower too. Loops have to be very complex for this to have any value, I am not sure when GCC decide to split loop, but if it is not very careful, when taking the choice, performances will degrade badly and eat CPU cycle for nothing.

I see "simple" parallelization of big loop as a free gain when the thread init time is smaller than the number of instances it will have to run, but for all other, deeper, loop optimization, I don't think it worth compiling all packages with them turned on.

EndrjuX · Post by **EndrjuX** » Sat Jul 03, 2010 8:52 pm

Have a look at http://gcc.gnu.org/wiki/Graphite/Parallelization

You can trigger it by 2 flags -floop-parallelize-all -ftree-parallelize-loops=4. Both of them is needed

Are you sure that -lgomp is useless? As I said some packages don't compile without those two in LDFLAGS, certainly one of them is required.

And after your post and some more reading in gcc manual i think i'll drop ftree-loop-distribution, thanks.

Elv13 · Post by **Elv13** » Tue Jul 06, 2010 5:08 pm

-lgomp is part of -fopenmp (at least, gcc should detect it need to be, ass gomp stand for Gnu/GCC OpenMP)

and read this page http://gcc.gnu.org/onlinedocs/gcc-4.5.0 ... tions.html
not the other one, it clearly say that -floop-parallelize-all does not check if it is profitable, and in most case, it is not.

Neskweek · Post by **Neskweek** » Tue Jul 20, 2010 2:11 pm

You can trigger it by 2 flags -floop-parallelize-all -ftree-parallelize-loops=4. Both of them is needed

Indeed

But since :

Calculating dependencies... done!

>>> Verifying ebuild manifests

>>> Starting parallel fetch

>>> Emerging (1 of 1171) sys-libs/zlib-1.2.5-r2
* zlib-1.2.5.tar.bz2 RMD160 SHA1 SHA256 size ... [ ok ]
* checking ebuild checksums ... [ ok ]
* checking auxfile checksums ... [ ok ]
* checking miscfile checksums ... [ ok ]
* Filtering out invalid CFLAG "-floop-parallelize-all"
* Filtering out invalid CXXFLAG "-floop-parallelize-all"

this instructions can't work properly

I tried with gcc (Gentoo 4.4.4-r1 p1.0, pie-0.4.5) 4.4.4
and this CFLAG :
-march=k8-sse3 -O2 -pipe -ftree-loop-linear -floop-block -floop-interchange -floop-strip-mine -fgraphite-identity -finline-functions -fira-algorithm=CB -fira-region=mixed -ftree-loop-distribution -floop-parallelize-all -ftree-parallelize-loops=4