[FFmpeg-devel] [PATCH v6 1/3] mips: improve detection of ISAs, FPU and ASEs (DSP, MSA)
Vicente Olivert Riera
Vincent.Riera at imgtec.com
Mon Jan 25 11:35:12 CET 2016
On 21/01/16 23:17, Michael Niedermayer wrote:
> On Wed, Jan 20, 2016 at 02:33:21PM +0000, Vicente Olivert Riera wrote:
>> On 20/01/16 13:15, Michael Niedermayer wrote:
>>> On Wed, Jan 20, 2016 at 12:21:37PM +0000, Vicente Olivert Riera wrote:
>>>> On 20/01/16 12:05, Michael Niedermayer wrote:
>>>>> On Wed, Jan 20, 2016 at 11:37:52AM +0000, Vicente Olivert Riera wrote:
>>>>>> Hello Michael,
>>>>>>
>>>>>> On 19/01/16 22:11, Michael Niedermayer wrote:
>>>>>>> On Thu, Jan 14, 2016 at 02:59:03PM +0000, Vicente Olivert Riera wrote:
>>>>>>>> Signed-off-by: Vicente Olivert Riera <Vincent.Riera at imgtec.com>
>>>>>>>> ---
>>>>>>>> Changes v5 -> v6:
>>>>>>>> - Multiple changes:
>>>>>>>> - Use separate if blocks when detecting the MIPS ISA, this way when a block
>>>>>>>> disables one ISA it will be detected by the next one. Before the elif was
>>>>>>>> preventing this to be done.
>>>>>>>> - Don't use check_cflags and use only check_ldflags instead, which will
>>>>>>>> check if that flag is valid for compiling and linking. The compiler may
>>>>>>>> accept some cflags for compiling but not for linking, so if we check first
>>>>>>>> for the cflags it will succeed, and the cflags will be added to the CFLAGS
>>>>>>>> variable despite of the check_ldflags failing after that. This can cause
>>>>>>>> problems because of having multiple incompatible cflags enabled at the same
>>>>>>>> time.
>>>>>>>> - Check if mipsfpu has been disabled, and if so, then use -msoft-float. This
>>>>>>>> way we override the compiler default behavious which can be building for
>>>>>>>> hard-float.
>>>>>>>> - Explicitly disable the features that fail any check by adding "|| disable
>>>>>>>> $feature" at the end of the checks chain.
>>>>>>>>
>>>>>>>> Changes v4 -> v5:
>>>>>>>> - Check for msa.h and disable msa if not found.
>>>>>>>>
>>>>>>>> Changes v3 -> v4:
>>>>>>>> - Nothing.
>>>>>>>>
>>>>>>>> Changes v2 -> v3:
>>>>>>>> - Nothing.
>>>>>>>>
>>>>>>>> Changes v1 -> v2:
>>>>>>>> - Add a block of code for mipsdspr2 which was previously placed in the
>>>>>>>> first patch of this series causing a syntax error in the configure
>>>>>>>> script.
>>>>>>>> https://lists.ffmpeg.org/pipermail/ffmpeg-devel/2015-November/183444.html
>>>>>>>
>>>>>>> make distclean ; ../configure --cross-prefix=/usr/mips-linux-gnu/bin/ --cc='ccache mips-linux-gnu-gcc-4.4' --arch=mips --target-os=linux --enable-cross-compile --disable-mipsfpu && make -j12
>>>>>>> results in pages of:
>>>>>>>
>>>>>>> /usr/lib/gcc/mips-linux-gnu/4.4.5/../../../../mips-linux-gnu/bin/ld: Warning: libavfilter/filtfmts-test uses hard float, libswresample/libswresample.a(dither.o) uses soft float
>>>>>>> /usr/lib/gcc/mips-linux-gnu/4.4.5/../../../../mips-linux-gnu/bin/ld: Warning: libavfilter/filtfmts-test uses hard float, libswresample/libswresample.a(resample.o) uses soft float
>>>>>>> /usr/lib/gcc/mips-linux-gnu/4.4.5/../../../../mips-linux-gnu/bin/ld: Warning: libavfilter/filtfmts-test uses hard float, libswresample/libswresample.a(resample_dsp.o) uses soft float
>>>>>>> /usr/lib/gcc/mips-linux-gnu/4.4.5/../../../../mips-linux-gnu/bin/ld: Warning: libavfilter/filtfmts-test uses hard float, libavutil/libavutil.a(adler32.o) uses soft float
>>>>>>> /usr/lib/gcc/mips-linux-gnu/4.4.5/../../../../mips-linux-gnu/bin/ld: Warning: libavfilter/filtfmts-test uses hard float, libavutil/libavutil.a(aes.o) uses soft float
>>>>>>> /usr/lib/gcc/mips-linux-gnu/4.4.5/../../../../mips-linux-gnu/bin/ld: Warning: libavfilter/filtfmts-test uses hard float, libavutil/libavutil.a(aes_ctr.o) uses soft float
>>>>>>> /usr/lib/gcc/mips-linux-gnu/4.4.5/../../../../mips-linux-gnu/bin/ld: Warning: libavfilter/filtfmts-test uses hard float, libavutil/libavutil.a(audio_fifo.o) uses soft float
>>>>>>>
>>>>>>>
>>>>>>> [...]
>>>>>>>
>>>>>>
>>>>>> I can't reproduce your issue. Look:
>>>>>>
>>>>>> $ make distclean ; ./configure --cross-prefix=mips-linux-gnu-
>>>>>> --arch=mips --target-os=linux --enable-cross-compile --disable-mipsfpu
>>>>>>
>>>>>> And then:
>>>>>>
>>>>>> $ make -j1 V=1 2>&1 | tee build.log
>>>>>>
>>>>>> And then:
>>>>>>
>>>>>> $ grep "uses soft float" build.log
>>>>>> $ grep "filtfmts-test" build.log
>>>>>>
>>>>>> It gives me nothing.
>>>>>>
>>>>>> Also I try to find for that filtfmts-test file which your logs say that
>>>>>> is using hard-float, and I cannot find it:
>>>>>>
>>>>>
>>>>>> $ find . -name "filtfmts-test*"
>>>>>> $
>>>>>
>>>>> see libavfilter/filtfmts.c
>>>>
>>>> But do you have a generated filtfmts-test binary or a filtfmts.o file?
>>>> Because when my build is finished I don't any anything of that.
>>>
>>> make libavfilter/filtfmts-test
>>> CC libavfilter/filtfmts-test.o
>>> libavfilter/filtfmts.c:141:9: warning: ignoring return value of function declared with warn_unused_result attribute [-Wunused-result]
>>> ff_default_query_formats(filter_ctx);
>>> ^~~~~~~~~~~~~~~~~~~~~~~~ ~~~~~~~~~~
>>> 1 warning generated.
>>> LD libavfilter/filtfmts-test
>>>
>>> the source and object files differ in their names
>>>
>>
>> I see. Ok, look at this:
>>
>> $ make libavfilter/filtfmts-test
>> CC libavfilter/filtfmts-test.o
>> libavfilter/filtfmts.c: In function 'main':
>> libavfilter/filtfmts.c:141:9: warning: ignoring return value of
>> 'ff_default_query_formats', declared with attribute warn_unused_result
>> [-Wunused-result]
>> ff_default_query_formats(filter_ctx);
>> ^
>> LD libavfilter/filtfmts-test
>>
>> $ mips-linux-gnu-objdump -x libavfilter/filtfmts-test | grep "FP ABI"
>> FP ABI: Soft float
>>
>> It shows to me that it has been for soft-float.
>>
>>>>
>>>>> our compilers differ
>>>>> the one i have here is a binary from emdebian
>>>>> are you droping support for that ?
>>>>> if so that should be clearly stated in the commit message, and the
>>>>> compiler cleanly detected and a clear warning given to the user why
>>>>> his/her compiler isnt supported and what is supported
>>>>
>>>> Well, my intention isn't dropping support for anything, although this is
>>>> what the emdebian website says:
>>>>
>>>> "Status of Emdebian distributions
>>>>
>>>> As of July 2014, updates to the Emdebian distributions ceased. There
>>>> will be no further updates and no further stable releases."
>>>
>>> i know, the problem is not me, i can update, IIRC you guys posted a
>>> binary already
>>> but others might use similarly build or old compilers
>>> so keeping this one here gives some wider coverage ... and its less
>>> work
>>
>> I also want to keep it.
>>
>>>>
>>>> Anyway, how can I obtain the same compiler as you have in order to try
>>>> reproducing the bug?
>>>
>>> i dont know, if they arent on emdian anymore
>>> i dont seem to have the deb files in var/cache/apt/archive anymore
>>>
>>> Package: gcc-4.4-mips-linux-gnu
>>> Status: install ok installed
>>> Priority: extra
>>> Section: devel
>>> Installed-Size: 3088
>>> Maintainer: Debian GCC Maintainers <debian-gcc at lists.debian.org>
>>> Architecture: amd64
>>> Source: gcc-4.4
>>> Version: 4.4.5-8
>>> Provides: c-compiler-mips-linux-gnu
>>> Depends: gcc-4.4-mips-linux-gnu-base (= 4.4.5-8), cpp-4.4-mips-linux-gnu (= 4.4.5-8), binutils-mips-linux-gnu (>= 2.20.1-15~), libgcc1-mips-cross (>= 1:4.4.5-8), libgomp1-mips-cross (>= 4.4.5-8), libc6 (>= 2.11)
>>> Recommends: libc6-dev-mips-cross (>= 2.5)
>>> Suggests: gcc-4.4-multilib-mips-linux-gnu, libmudflap0-4.4-dev-mips-cross (>= 4.4.5-8), gcc-4.4-doc (>= 4.4.4-17), gcc-4.4-locales (>= 4.4.4-17), libgcc1-dbg-mips-cross, libgomp1-dbg-mips-cross, libmudflap0-dbg-mips-cross, libcloog-ppl0 (>= 0.15.9-2~), libppl-c2, libppl7
>>> Description: The GNU C compiler
>>> This is the GNU C compiler, a fairly portable optimizing compiler for C.
>>> .
>>> This package contains C cross-compiler for mips architecture.
>>> Homepage: http://gcc.gnu.org/
>>>
>>
>> Ok then.
>>
>>>>
>>>>> make distclean ; ../configure --cross-prefix=/usr/mips-linux-gnu/bin/ --cc='mips-linux-gnu-gcc-4.4' --arch=mips --target-os=linux --enable-cross-compile --disable-mipsfpu && make -j12 >& build
>>>>>
>>>>> grep 'uses soft float' build | wc
>>>>> 2327 24879 389225
>>>>>
>>>>> echo 'int main(void){ return 0; }' | mips-linux-gnu-gcc-4.4 -x c - -o out.o
>>>>> echo 'int main(void){ return 0; }' | mips-linux-gnu-gcc-4.4 -x c - -o out.o -msoft-float
>>>>> /usr/lib/gcc/mips-linux-gnu/4.4.5/../../../../mips-linux-gnu/bin/ld: Warning: out.o uses hard float, /tmp/ccFlutMG.o uses soft float
>>>>>
>>>>>
>>>>
>>>> But I don't understand how your compiler built something for hard-float
>>>> when the -msoft-float option should be used instead. This is what I have
>>>> (note that I passed V=1 to the make command):
>>>>
>>>> $ grep '\-mhard-float' build.log | wc -l
>>>> 0
>>>> $ grep '\-msoft-float' build.log | wc -l
>>>> 1411
>>>> $ grep "mips-linux-gnu-gcc" build.log | wc -l
>>>> 1411
>>>>
>>>> So the -mhard-float option wasn't used at all, and the -msoft-float one
>>>> was used every time the compiler was called.
>>>>
>>>> I'm curious, can you show me your output for those same commands, please?
>>>
>>> before the patch:
>>> grep '\-mhard-float' build.log-V1-disable-mipsfpu | wc -l
>>> 0
>>>
>>> grep '\-msoft-float' build.log-V1-disable-mipsfpu | wc -l
>>> 0
>>
>> So, before the patch when you pass the --disable-mipsfpu option, what
>> configure does is not adding the -mhard-float option to the CFLAGS. But
>> also it doesn't add -msoft-float. So no float option is passed to the
>> compiler at all.
>>
>> Given that you don't pass any of those two options to the compiler, the
>> default one will be used, which I think it will be -mhard-float.
>>
>> Look at this simple test:
>>
>> $ cat test.c
>> int main(){return 0;}
>> $ # USING -mhard-float
>> $ mips-linux-gnu-gcc -mhard-float test.c
>> $ mips-linux-gnu-objdump -x a.out | grep "FP ABI"
>> FP ABI: Hard float (double precision)
>> $ # USING -msoft-float
>> $ mips-linux-gnu-gcc -msoft-float test.c
>> $ mips-linux-gnu-objdump -x a.out | grep "FP ABI"
>> FP ABI: Soft float
>> $ # USING COMPILER'S DEFAULT
>> $ mips-linux-gnu-gcc test.c
>> $ mips-linux-gnu-objdump -x a.out | grep "FP ABI"
>> FP ABI: Hard float (double precision)
>>
>> So, as you can see, before that patch, when we were passing
>> --disable-mipsfpu to the configure script, we weren't building ffmpeg
>> for soft-float. We were actually building it for the compiler's default
>> floatness, which could be either soft-float or (more likely) hard-float.
>>
>> That's why you didn't see those warnings, because everything was being
>> built for hard-float.
>>
>>> after the patch:
>>> grep '\-mhard-float' build.log-V1-disable-mipsfpu-with-patch | wc -l
>>> 0
>>> grep '\-msoft-float' build.log-V1-disable-mipsfpu-with-patch | wc -l
>>> 1428
>>> grep "mips-linux-gnu-gcc" build.log-V1-disable-mipsfpu-with-patch | wc -l
>>> 1428
>>
>> This is what doesn't make sense to me. Every time your compiler is
>> called it has the -msoft-float option with it, so everything it has
>> built was built for soft-float.
>>
>>> also forget about filtfmts, its not specifici to that
>>> /usr/lib/gcc/mips-linux-gnu/4.4.5/../../../../mips-linux-gnu/bin/ld: Warning: ffprobe_g uses hard float, libavutil/libavutil.a(samplefmt.o) uses soft float
>>> /usr/lib/gcc/mips-linux-gnu/4.4.5/../../../../mips-linux-gnu/bin/ld: Warning: ffprobe_g uses hard float, libavutil/libavutil.a(sha.o) uses soft float
>>> /usr/lib/gcc/mips-linux-gnu/4.4.5/../../../../mips-linux-gnu/bin/ld: Warning: ffprobe_g uses hard float, libavutil/libavutil.a(sha512.o) uses soft float
>>> /usr/lib/gcc/mips-linux-gnu/4.4.5/../../../../mips-linux-gnu/bin/ld: Warning: ffprobe_g uses hard float, libavutil/libavutil.a(stereo3d.o) uses soft float
>>
>> Can you grep the output of "mips-linux-gnu-objdump -x ffprobe_g" and
>> search for "FP ABI" in order to see if it actually uses hard float?
>
> before and after the patch there is no match, greping for ABI
> 1 .note.ABI-tag 00000020 004001a4 004001a4 000001a4 2**2
> 004001a4 l d .note.ABI-tag 00000000 .note.ABI-tag
>
> mips-linux-gnu-objdump -x ffprobe_g | grep 'ABI'
> 1 .note.ABI-tag 00000020 004001a4 004001a4 000001a4 2**2
> 004001a4 l d .note.ABI-tag 00000000 .note.ABI-tag
>
> also both mips-linux-gnu-objdump and the normal hosts objdump give the
> same result
>
> [...]
>
Hello Michael,
I don't know what else to do. The think is that those are warning
messages, so it seems your build doesn't fail.
Before, when you tried to build ffmpeg for MIPS soft-float (passing the
--disable-mipsfpu option) you were actually building it for hard-float,
but you didn't noticed.
Now when you use the --disable-mipsfpu option, your ffmpeg (using that
compiler) is built for soft-float, although some hard-float bits are
included (I don't know why). With all the compilers I have tried the
whole ffmpeg is build for soft-float. There are no warnings about that.
So, what should we do here in order to get that patch applied? Should I
remove the part where -msoft-float is added to the flags in order to
have the useless behavior that we had before? I mean, building it always
for hard-float even if you use the --disable-mipsfpu option. Or can we
live with this small issue which happens only (as far as we know) with
your compiler?
Regards,
Vincent.
More information about the ffmpeg-devel
mailing list