[FFmpeg-devel] [PATCH] SSE2 version of vf_idet's filter_line()

Pascal Massimino pascal.massimino at gmail.com
Wed Sep 3 23:04:43 CEST 2014


Clément

On Wed, Sep 3, 2014 at 12:37 PM, Clément Bœsch <u at pkh.me> wrote:

> On Wed, Sep 03, 2014 at 07:05:48PM +0200, Pascal Massimino wrote:
> [...]
> > > > +    punpcklbw m3, m_zero
> > > > +    punpckhbw m4, m_zero
> > > > +
> > > > +    paddsw    m0, m3
> > > > +    paddsw    m1, m4
> > > > +
> > > > +    movq      m3, [bq+indexq*1]
> > > > +    movq      m4, m3
> > > > +    punpcklbw m3, m_zero
> > > > +    punpckhbw m4, m_zero
> > > > +
> > > > +    paddw     m3, m3
> > > > +    paddw     m4, m4
> > > > +    psubsw    m0, m3
> > > > +    psubsw    m1, m4
> > > > +
> > >
> > > > +    ABS1      m0, m5
> > > > +    ABS1      m1, m5
> > >
> > > ABS2?
> > >
> >
> > ABS2 requires the two tmp registers to be different (can't use m5 for
> both).
> >
>
> Aren't m3 and m4 available at that point?
>

Indeed. We can actually get rid of m5 use totally.

>
> (ABS2 has the benefit of doing some pairing so is faster than doing two
> ABS1)
>
> >
> > >
> > > > +    paddw     m0, m1
> > > > +    movq      m1, m0
> > > > +    punpcklwd m0, m_zero
> > > > +    punpckhwd m1, m_zero
> > > > +    paddd     m0, m1
> > > > +    paddd     m_sum, m0
> > > > +
> > > > +    add       indexq, 0x8
> > >
> > > > +    CMP       widthq, indexq
> > >
> > > Someone needs to confirm this, but I think you'll need to make width a
> > > ptrdiff_t and not an int
> > >
> >
> > changed to widthd/indexd, that's enough.
> >
>
> Hopefully...
>
> >
> > >
> > > Also... stupid question but what's CMP?
> > >
> >
> > it's equivalent to 'cmp DWORD' here iirc.
> >
>
> I believe you can keep it lowercase. I thought it was a macro but didn't
> see anything like this.
>

In any case, it doesn't compile here with lowercase here (yasm 1.3.0).
Not sure why. Any idea welcome...


> >
> >
> [...]
> > diff --git a/libavfilter/x86/vf_idet.asm b/libavfilter/x86/vf_idet.asm
> > new file mode 100644
> > index 0000000..19b7f3b
> > --- /dev/null
> > +++ b/libavfilter/x86/vf_idet.asm
> > @@ -0,0 +1,116 @@
> > +;;
> *****************************************************************************
> > +;; * x86-optimized functions for idet filter
> > +;; *
> > +;; * This file is part of FFmpeg.
> > +;; *
> > +;; * FFmpeg is free software; you can redistribute it and/or modify
> > +;; * it under the terms of the GNU General Public License as published
> by
> > +;; * the Free Software Foundation; either version 2 of the License, or
> > +;; * (at your option) any later version.
> > +;; *
> > +;; * FFmpeg is distributed in the hope that it will be useful,
> > +;; * but WITHOUT ANY WARRANTY; without even the implied warranty of
> > +;; * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
> > +;; * GNU General Public License for more details.
> > +;; *
> > +;; * You should have received a copy of the GNU General Public License
> along
> > +;; * with FFmpeg; if not, write to the Free Software Foundation, Inc.,
> > +;; * 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA.
> > +;;
> ******************************************************************************
> > +
> > +%include "libavutil/x86/x86util.asm"
> > +
> > +SECTION_TEXT
> > +
> > +; Implementation that does 8-bytes at a time using single-word
> operations.
> > +%macro IDET_FILTER_LINE 0
>
> > +cglobal idet_filter_line, 4, 8, 6, a, b, c, width, index
>
> Shouldn't this be 4,5,8? 4 args, 5 regs (4 args + 1 local var), 8 xmm
> (m0..m7)


> And BTW, you don't seem to use m2, so you can have 4,5,7 ; you probably
> want to define m_zero on m2 to avoid any confusion.
>

Actually, m5 m6 and m7 are now unused, so it's 4,5,6 now, eventually.


> Now, this macro is only used for MMX, so you can probably do 4,5,0 (no XMM
> reg). If you do 4,5,0, please move the INIT_MMX into the macro:
>
>     %macro IDET_FILTER_LINE_MMX 1
>     INIT_MMX %1
>     ...
>     %end macro
>

all done.


>
>     ...
>
>     INIT_MMX mmx
>     INIT_MMX mmxext
>
>
> > +    xor       indexq, indexq
> > +%define   m_zero m7
> > +%define   m_sum  m6
> > +    pxor      m_sum, m_sum
> > +    pxor      m_zero, m_zero
> > +
> > +.loop:
> > +    movu      m0, [aq+indexq*1]
> > +    punpckhbw m1, m0, m_zero
> > +    punpcklbw m0, m_zero
> > +
> > +    movu      m3, [cq+indexq*1]
> > +    punpckhbw m4, m3, m_zero
> > +    punpcklbw m3, m_zero
> > +
> > +    paddsw    m1, m4
> > +    paddsw    m0, m3
> > +
> > +    movu      m3, [bq+indexq*1]
> > +    punpckhbw m4, m3, m_zero
> > +    punpcklbw m3, m_zero
> > +
> > +    paddw     m4, m4
> > +    paddw     m3, m3
> > +    psubsw    m1, m4
> > +    psubsw    m0, m3
> > +
> > +    ABS1      m1, m5
> > +    ABS1      m0, m5
> > +
> > +    paddw     m0, m1
> > +    punpckhwd m1, m0, m_zero
> > +    punpcklwd m0, m_zero
> > +
> > +    paddd     m0, m1
> > +    paddd     m_sum, m0
> > +
> > +    add       indexq, 0x8
> > +    CMP       widthd, indexd
> > +    jg        .loop
> > +
> > +    mova      m0, m_sum
> > +    psrlq     m_sum, 0x20
> > +    paddq     m0, m_sum
> > +    movd      eax, m0
> > +    RET
> > +%endmacro
> > +
>
> > +%if ARCH_X86_32
> > +INIT_MMX mmxext
> > +IDET_FILTER_LINE
> > +
> > +INIT_MMX mmx
> > +IDET_FILTER_LINE
> > +%endif
> > +
> > +;; SSE2 8-bit implementation that does 16-bytes at a time:
> > +INIT_XMM sse2
>
> > +cglobal idet_filter_line, 4, 8, 6, a, b, c, width, index, total
>
> 4,6,7, AFAICT
>

done


>
> > +    xor       indexq, indexq
> > +    pxor      m0, m0
> > +    pxor      m1, m1
> > +
> > +.sse2_loop:
> > +    movu      m2, [bq+indexq*1]  ; B
> > +    movu      m3, [aq+indexq*1]  ; A
> > +    mova      m6, m2
> > +    mova      m4, m3
> > +    psubusb   m5, m2, m3         ; ba
> > +
> > +    movu      m3, [cq+indexq*1]  ; C
> > +    add       indexq, 0x10
> > +    psubusb   m4, m2             ; ab
> > +    CMP       indexd, widthd
> > +
> > +    psubusb   m6, m3             ; bc
> > +    psubusb   m3, m2             ; cb
> > +
> > +    psadbw    m4, m6             ; |ab - bc|
> > +    paddq     m0, m4
> > +    psadbw    m5, m3             ; |ba - cb|
> > +    paddq     m1, m5
> > +    jl       .sse2_loop
> > +
> > +    paddq     m0, m1
> > +    movhlps   m1, m0
> > +    paddq     m0, m1
> > +    movd      eax, m0
> > +    RET
> [...]
>
> --
> Clément B.
>
-------------- next part --------------
A non-text attachment was scrubbed...
Name: 0001-MMX-MMXEXT-SSE2-implementation-of-idet-s-filter_line.patch
Type: application/octet-stream
Size: 11929 bytes
Desc: not available
URL: <https://ffmpeg.org/pipermail/ffmpeg-devel/attachments/20140903/8a315dc7/attachment.obj>


More information about the ffmpeg-devel mailing list