fftw-3[-single] 3.2.1 : should not have --enable-fma for Intel builds

[eborisch@…]

Both the fftw-3 and fftw-3-single ports have --enable-fma in their configure.args. These should be removed, as they are incorrect for Intel builds, and replaced with the appropriate platform-specific args.

Suggested args:
PPC: --enable-fma --enable-altivec
Intel: --enable-sse OR --enable-sse2 (for fftw-3-single, and fftw-3, respectively)

It might also be nice to add '-O3 -fomit-frame-pointer -fstrict-aliasing' to the configure.cflags-append -- most folks using FFTW are going for speed; might as well build for it. (These flags were found on the FFTW benchmark site: ​http://www.fftw.org/speed/Pentium4-3.06GHz/ and ​http://www.fftw.org/speed/G5-2GHz-macosx/)

[skymoo (Adam Mercer)]

Cc Me!

[skymoo (Adam Mercer)]

On Intel Leopard fftw-3 and fftw-3-single build without error using these suggested flags. I've attached patches, can anyone test on PPC?

[skymoo (Adam Mercer)]

apply flags to fftw-3

[skymoo (Adam Mercer)]

apply flags to fftw-3-single

[skymoo (Adam Mercer)]

One thing I've just noticed is that the Portfiles already used the -fno-common cflag and the page you linked to didn't mention this cflag. Should this be used?

[skymoo (Adam Mercer)]

fftw-3 fails to build with the universal variant with the following error:

libtool: compile:  gcc -std=gnu99 -DHAVE_CONFIG_H -I. -I../.. -I../../kernel -I../../simd -msse2 -O2 -fno-common -O3 -fomit-frame-pointer -fstrict-aliasing -isysroot /Developer/SDKs/MacOSX10.5.sdk -arch ppc -D_THREAD_SAFE -MT sse2.lo -MD -MP -MF .deps/sse2.Tpo -c sse2.c  -fno-common -DPIC -o .libs/sse2.o
sse2.c:27: error: variable 'fftw_sse2_pm' has initializer but incomplete type
sse2.c:28: error: extra brace group at end of initializer
sse2.c:28: error: (near initialization for 'fftw_sse2_pm')
sse2.c:28: warning: excess elements in union initializer
sse2.c:28: warning: (near initialization for 'fftw_sse2_pm')
make[3]: *** [sse2.lo] Error 1
make[2]: *** [all-recursive] Error 1
make[1]: *** [all-recursive] Error 1
make: *** [all] Error 2

Also on Intel Leopard. Looks like the --enable-sse2 option is being passed to configure for the PPC build.

[MarcusCalhoun-Lopez (Marcus Calhoun-Lopez)]

Cc Me!

[eborisch@…]

Replying to ram@…:

One thing I've just noticed is that the Portfiles already used the -fno-common cflag and the page you linked to didn't mention this cflag. Should this be used?

The -fno-common originated while addressing Ticket #14240 -- I'd guess leave it in for now!

[eborisch@…]

How about this for fftw-3:

array set merger_configure_args {
        ppc    "--enable-fma --enable-altivec"
        ppc64  "--enable-fma --enable-altivec"
        i386    --enable-sse2
        x86_64  --enable-sse2
}

and this for fftw-3-single:

array set merger_configure_args {
        ppc    "--enable-fma --enable-altivec"
        ppc64  "--enable-fma --enable-altivec"
        i386    --enable-sse
        x86_64  --enable-sse
}

(Keeping your other changes to configure.args and configure.cflags-append)

[eborisch@…]

The configure argument "-enable-type-prefix" in fftw-3-single should go away as well (not a valid - or needed - argument anymore.)

[skymoo (Adam Mercer)]

Replying to eborisch@…:

How about this for fftw-3:

array set merger_configure_args {
        ppc    "--enable-fma --enable-altivec"
        ppc64  "--enable-fma --enable-altivec"
        i386    --enable-sse2
        x86_64  --enable-sse2
}

That does the job, although --enable-altivec is only appropriate for the single precision build.

Replying to eborisch@…:

The configure argument "-enable-type-prefix" in fftw-3-single should go away as well (not a valid - or needed - argument anymore.)

Done

Changes committed in r47232, thanks!

[skymoo (Adam Mercer)]

Reopening as the platform specific flags, --enable-sse2 for example, are not applied when the port is built for a single platform, i.e. not universally.

[skymoo (Adam Mercer)]

Platform specific flags applied to non-universal builds in r47243

[(none)]

Milestone Port Bugs deleted