[FFmpeg-devel] [PATCH] Move MLP's dot product to DSPContext

[Michael Niedermayer]

On Fri, May 15, 2009 at 09:41:30PM -0300, Ramiro Polla wrote:
> On Fri, May 15, 2009 at 4:43 PM, Michael Niedermayer <michaelni at gmx.at> wrote:
> > On Fri, May 15, 2009 at 04:33:12PM -0300, Ramiro Polla wrote:
> >> On Fri, May 15, 2009 at 3:43 PM, Michael Niedermayer <michaelni at gmx.at> wrote:
> >> > On Fri, May 15, 2009 at 01:30:17PM -0300, Ramiro Polla wrote:
> >> >> On Fri, May 15, 2009 at 12:15 PM, Michael Niedermayer <michaelni at gmx.at> wrote:
> >> >> > On Fri, May 15, 2009 at 10:38:01AM -0300, Ramiro Polla wrote:
> >> >> >> On Fri, May 15, 2009 at 5:49 AM, Diego Biurrun <diego at biurrun.de> wrote:
> >> >> >> > On Thu, May 14, 2009 at 07:44:36PM -0300, Ramiro Polla wrote:
> >> >> >> >> #if ARCH_X86_64 || (ARCH_X86_32 && HAVE_7REGS)
> >> >> >> >>
> >> >> >> >>
> >> >> >> >> #if ARCH_X86_64
> >> >> >> >>
> >> >> >> >> #else /* ARCH_X86_32 && HAVE_7REGS */
> >> >> >> >>
> >> >> >> >> #endif
> >> >> >> >>
> >> >> >> >> #endif /* ARCH_X86_64 || (ARCH_X86_32 && HAVE_7REGS) */
> >> >> >> >
> >> >> >> > This had me confused for a moment, please comment the #endif.
> >> >> >>
> >> >> >> I used Loren's suggestion and added some comments to the endifs.
> >> >> >>
> >> >> >> I also fixed the indentation of the function declaration and
> >> >> >> simplified the constraints ifdef bit more. And also removed some
> >> >> >> unnecessarily exported symbols and renamed them.
> >> >> > [...]
> >> >> >> ? ? ? ? : [s]"r"(state),
> >> >> >> ? ? ? ? ? [c]"r"(coeff),
> >> >> >> ? ? ? ? ? [b]"r"(sample_buffer),
> >> >> >> ? ? ? ? ? [i]CONSTRAINT(blocksize),
> >> >> >> ? ? ? ? ? [mask ? ]CONSTRAINT((x86_reg)mask),
> >> >> >> ? ? ? ? ? [firjump]CONSTRAINT(firjump),
> >> >> >> ? ? ? ? ? [iirjump]CONSTRAINT(iirjump),
> >> >> >> ? ? ? ? ? [shift ?]COUNTER(filter_shift),
> >> >> >> ? ? ? ? ? [binc]"i"(sizeof(int32_t)* MAX_CHANNELS),
> >> >> >> ? ? ? ? ? [is ?]"i"(sizeof(int32_t)*(MAX_FIR_ORDER + MAX_BLOCKSIZE)),
> >> >> >> ? ? ? ? ? [ic ?]"i"(sizeof(int32_t)* MAX_FIR_ORDER)
> >> >> >
> >> >> > this would break gcc 2.95 (fate)
> >> >>
> >> >> It takes 11 operands on x86_32, so it doesn't compile on gcc-2.95
> >> >> anyways. Suggestions:
> >> >> - enclose it in HAVE_TEN_OPERANDS and leave the code with []s.
> >> >> - remove []s and hardcode binc, is and ic values.
> >> >>
> >> >> Is there a way to expand sizeof(int32_t)*foobar into the asm without
> >> >> using the "i" constraint?
> >> >
> >> > sizeof(int32_t) is 4
> >>
> >> - Removed []s
> >> - Replaced "i"s by AVSTRINGIFYed defines.
> >> - Added const to [fi]irjump to avoid a warning.
> > [...]
> >
> >> ? ? ? ? "sub ? $4 ? ? ? , ?%0 ? ? ? ? \n\t"
> >> ? ? ? ? "mov ?"RESULT32", (%0) ? ? ? ?\n\t"
> >> ? ? ? ? "mov ?"RESULT32", (%2) ? ? ? ?\n\t"
> >> ? ? ? ? "add $"BINC" ? ?, ?%2 ? ? ? ? \n\t"
> >> ? ? ? ? "sub ?"ACCUM" ? ,"RESULT" ? ? \n\t"
> >> ? ? ? ? "mov ?"RESULT32","IOFFS"(%0) ?\n\t"
> >> ? ? ? ? "incl ? ? ? ? ? ? ?%3 ? ? ? ? \n\t"
> >> ? ? ? ? "js 1b ? ? ? ? ? ? ? ? ? ? ? ?\n\t"
> >> ? ? ? ? :
> >> ? ? ? ? : /* 0*/"r"(state),
> >
> > you are changing pure input operands, this leads to undefined behavior.
> > also mixed input+output count as 2 operands in relation to the 10 limit
> > in at least some gcc versions IIRC (so be carefull how you solve this)
> 
> #if HAVE_TEN_OPERANDS ?

hmm, well ok
and patch ok, less macro would be better though ...

[...]
-- 
Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB

Those who are too smart to engage in politics are punished by being
governed by those who are dumber. -- Plato 
-------------- next part --------------
A non-text attachment was scrubbed...
Name: not available
Type: application/pgp-signature
Size: 189 bytes
Desc: Digital signature
URL: <http://lists.mplayerhq.hu/pipermail/ffmpeg-devel/attachments/20090520/28050d34/attachment.pgp>