[FFmpeg-devel,v3] aacenc: add SIMD optimizations for abs_pow34 and quantization

Message ID CAE9qxYB4fyrierEMSG8=5Ei3XMkAa2uLFw1Az5nsiHG2TTzgZw@mail.gmail.com
State Accepted
Headers

Commit Message

Rostislav Pehlivanov Oct. 18, 2016, 4:33 p.m. UTC
On 18 October 2016 at 16:32, James Almer <jamrial@gmail.com> wrote:

> On 10/18/2016 12:07 PM, Rostislav Pehlivanov wrote:
> > diff --git a/libavcodec/aacenc.c b/libavcodec/aacenc.c
> > index ee3cbf8..622f0ba 100644
> > --- a/libavcodec/aacenc.c
> > +++ b/libavcodec/aacenc.c
> > @@ -1033,6 +1033,12 @@ static av_cold int aac_encode_init(AVCodecContext
> *avctx)
> >      ff_lpc_init(&s->lpc, 2*avctx->frame_size, TNS_MAX_ORDER,
> FF_LPC_TYPE_LEVINSON);
> >      s->random_state = 0x1f2e3d4c;
> >
> > +    s->abs_pow34   = &abs_pow34_v;
> > +    s->quant_bands = &quantize_bands;
>
> No need for & in these.
>
> > +
> > +    if (ARCH_X86)
> > +        ff_aac_dsp_init_x86(s);
> > +
> >      if (HAVE_MIPSDSP)
> >          ff_aac_coder_init_mips(s);
>
> [...]
>
> > diff --git a/libavcodec/x86/aacencdsp.asm b/libavcodec/x86/aacencdsp.asm
> > new file mode 100644
> > index 0000000..dd7b022
> > --- /dev/null
> > +++ b/libavcodec/x86/aacencdsp.asm
> > @@ -0,0 +1,88 @@
> > +;**********************************************************
> ********************
> > +;* SIMD optimized AAC encoder DSP functions
> > +;*
> > +;* Copyright (C) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
> > +;*
> > +;* This file is part of FFmpeg.
> > +;*
> > +;* FFmpeg is free software; you can redistribute it and/or
> > +;* modify it under the terms of the GNU Lesser General Public
> > +;* License as published by the Free Software Foundation; either
> > +;* version 2.1 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
> > +;* Lesser General Public License for more details.
> > +;*
> > +;* You should have received a copy of the GNU Lesser 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_RODATA
> > +
> > +float_abs_mask: times 4 dd 0x7fffffff
> > +
> > +SECTION .text
> > +
> > +;*******************************************************************
> > +;void ff_abs_pow34(float *out, const float *in, const int size);
> > +;*******************************************************************
> > +INIT_XMM sse
> > +cglobal abs_pow34, 3, 3, 3, out, in, size
> > +    mova   m2, [float_abs_mask]
> > +    shl    sizeq, 2
> > +    add    inq, sizeq
> > +    add    outq, sizeq
> > +    neg    sizeq
> > +.loop:
> > +    movaps m0, [inq+sizeq]
> > +    andps  m0, m2
>
> Remove the movaps and do
>
> andps  m0, m2, [inq+sizeq]
>
> Instead. Sorry i didn't notice this last time.
>
> > +    sqrtps m1, m0
> > +    mulps  m0, m1
> > +    sqrtps m0, m0
> > +    mova   [outq+sizeq], m0
> > +    add    sizeq, mmsize
> > +    jl    .loop
> > +    RET
> > +
> > +;*******************************************************************
> > +;void ff_aac_quantize_bands(int *out, const float *in, const float
> *scaled,
> > +;                           int size, int is_signed, int maxval, const
> float Q34,
> > +;                           const float rounding)
> > +;*******************************************************************
> > +INIT_XMM sse2
> > +cglobal aac_quantize_bands, 5, 5, 6, out, in, scaled, size, is_signed,
> maxval, Q34, rounding
> > +%if UNIX64 == 0
> > +    movss     m0, Q34m
> > +    movss     m1, roundingm
> > +    cvtsi2ss  m3, maxvald
> > +%else
> > +    cvtsi2ss  m3, dword maxvalm
> > +%endif
>
> The other way around. Unix64 is the one that has maxval on a reg regardless
> of how you init the function, whereas win64 and any x86_32 target have it
> on stack.
>
> > +    shufps    m0, m0, 0
> > +    shufps    m1, m1, 0
> > +    shufps    m3, m3, 0
> > +    shl       is_signedd, 31
> > +    movd      m4, is_signedd
> > +    shufps    m4, m4, 0
> > +    shl       sized,   2
> > +    add       inq, sizeq
> > +    add       outq, sizeq
> > +    add       scaledq, sizeq
> > +    neg       sizeq
> > +.loop:
> > +    mulps     m2, m0, [scaledq+sizeq]
> > +    addps     m2, m1
> > +    minps     m2, m3
>
> > +    movaps    m5, [inq+sizeq]
> > +    andps     m5, m4
>
> Same as in abs_pow34, remove movaps and do
>
> andps     m5, m4, [inq+sizeq]
>
> > +    orps      m2, m5
> > +    cvttps2dq m2, m2
> > +    mova      [outq+sizeq], m2
> > +    add       sizeq, mmsize
> > +    jl       .loop
> > +    RET
> > diff --git a/libavcodec/x86/aacencdsp_init.c b/libavcodec/x86/aacencdsp_
> init.c
> > new file mode 100644
> > index 0000000..aefaa15
> > --- /dev/null
> > +++ b/libavcodec/x86/aacencdsp_init.c
> > @@ -0,0 +1,43 @@
> > +/*
> > + * AAC encoder assembly optimizations
> > + * Copyright (C) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
> > + *
> > + * This file is part of FFmpeg.
> > + *
> > + * FFmpeg is free software; you can redistribute it and/or
> > + * modify it under the terms of the GNU Lesser General Public
> > + * License as published by the Free Software Foundation; either
> > + * version 2.1 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
> > + * Lesser General Public License for more details.
> > + *
> > + * You should have received a copy of the GNU Lesser 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 "config.h"
> > +
> > +#include "libavutil/float_dsp.h"
> > +#include "libavutil/x86/cpu.h"
> > +#include "libavcodec/aacenc.h"
> > +
> > +void ff_abs_pow34_sse(float *out, const float *in, const int size);
> > +
> > +void ff_aac_quantize_bands_sse2(int *out, const float *in, const float
> *scaled,
> > +                                int size, int is_signed, int maxval,
> const float Q34,
> > +                                const float rounding);
> > +
> > +av_cold void ff_aac_dsp_init_x86(AACEncContext *s)
> > +{
> > +    int cpu_flags = av_get_cpu_flags();
> > +
> > +    if (EXTERNAL_SSE(cpu_flags))
> > +        s->abs_pow34   = &ff_abs_pow34_sse;
> > +
> > +    if (EXTERNAL_SSE2(cpu_flags))
> > +        s->quant_bands = &ff_aac_quantize_bands_sse2;
>
> Same with these.
>
> Passes fate on mingw-w64 (x86_64) after fixing the maxval mixup above.
> Crashes otherwise.
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>

Hopefully I didn't mess anything up and the attached patch is the last
revision of it.
  

Comments

Michael Niedermayer Oct. 18, 2016, 8:04 p.m. UTC | #1
On Tue, Oct 18, 2016 at 05:33:13PM +0100, Rostislav Pehlivanov wrote:
> On 18 October 2016 at 16:32, James Almer <jamrial@gmail.com> wrote:
> 
> > On 10/18/2016 12:07 PM, Rostislav Pehlivanov wrote:
> > > diff --git a/libavcodec/aacenc.c b/libavcodec/aacenc.c
> > > index ee3cbf8..622f0ba 100644
> > > --- a/libavcodec/aacenc.c
> > > +++ b/libavcodec/aacenc.c
> > > @@ -1033,6 +1033,12 @@ static av_cold int aac_encode_init(AVCodecContext
> > *avctx)
> > >      ff_lpc_init(&s->lpc, 2*avctx->frame_size, TNS_MAX_ORDER,
> > FF_LPC_TYPE_LEVINSON);
> > >      s->random_state = 0x1f2e3d4c;
> > >
> > > +    s->abs_pow34   = &abs_pow34_v;
> > > +    s->quant_bands = &quantize_bands;
> >
> > No need for & in these.
> >
> > > +
> > > +    if (ARCH_X86)
> > > +        ff_aac_dsp_init_x86(s);
> > > +
> > >      if (HAVE_MIPSDSP)
> > >          ff_aac_coder_init_mips(s);
> >
> > [...]
> >
> > > diff --git a/libavcodec/x86/aacencdsp.asm b/libavcodec/x86/aacencdsp.asm
> > > new file mode 100644
> > > index 0000000..dd7b022
> > > --- /dev/null
> > > +++ b/libavcodec/x86/aacencdsp.asm
> > > @@ -0,0 +1,88 @@
> > > +;**********************************************************
> > ********************
> > > +;* SIMD optimized AAC encoder DSP functions
> > > +;*
> > > +;* Copyright (C) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
> > > +;*
> > > +;* This file is part of FFmpeg.
> > > +;*
> > > +;* FFmpeg is free software; you can redistribute it and/or
> > > +;* modify it under the terms of the GNU Lesser General Public
> > > +;* License as published by the Free Software Foundation; either
> > > +;* version 2.1 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
> > > +;* Lesser General Public License for more details.
> > > +;*
> > > +;* You should have received a copy of the GNU Lesser 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_RODATA
> > > +
> > > +float_abs_mask: times 4 dd 0x7fffffff
> > > +
> > > +SECTION .text
> > > +
> > > +;*******************************************************************
> > > +;void ff_abs_pow34(float *out, const float *in, const int size);
> > > +;*******************************************************************
> > > +INIT_XMM sse
> > > +cglobal abs_pow34, 3, 3, 3, out, in, size
> > > +    mova   m2, [float_abs_mask]
> > > +    shl    sizeq, 2
> > > +    add    inq, sizeq
> > > +    add    outq, sizeq
> > > +    neg    sizeq
> > > +.loop:
> > > +    movaps m0, [inq+sizeq]
> > > +    andps  m0, m2
> >
> > Remove the movaps and do
> >
> > andps  m0, m2, [inq+sizeq]
> >
> > Instead. Sorry i didn't notice this last time.
> >
> > > +    sqrtps m1, m0
> > > +    mulps  m0, m1
> > > +    sqrtps m0, m0
> > > +    mova   [outq+sizeq], m0
> > > +    add    sizeq, mmsize
> > > +    jl    .loop
> > > +    RET
> > > +
> > > +;*******************************************************************
> > > +;void ff_aac_quantize_bands(int *out, const float *in, const float
> > *scaled,
> > > +;                           int size, int is_signed, int maxval, const
> > float Q34,
> > > +;                           const float rounding)
> > > +;*******************************************************************
> > > +INIT_XMM sse2
> > > +cglobal aac_quantize_bands, 5, 5, 6, out, in, scaled, size, is_signed,
> > maxval, Q34, rounding
> > > +%if UNIX64 == 0
> > > +    movss     m0, Q34m
> > > +    movss     m1, roundingm
> > > +    cvtsi2ss  m3, maxvald
> > > +%else
> > > +    cvtsi2ss  m3, dword maxvalm
> > > +%endif
> >
> > The other way around. Unix64 is the one that has maxval on a reg regardless
> > of how you init the function, whereas win64 and any x86_32 target have it
> > on stack.
> >
> > > +    shufps    m0, m0, 0
> > > +    shufps    m1, m1, 0
> > > +    shufps    m3, m3, 0
> > > +    shl       is_signedd, 31
> > > +    movd      m4, is_signedd
> > > +    shufps    m4, m4, 0
> > > +    shl       sized,   2
> > > +    add       inq, sizeq
> > > +    add       outq, sizeq
> > > +    add       scaledq, sizeq
> > > +    neg       sizeq
> > > +.loop:
> > > +    mulps     m2, m0, [scaledq+sizeq]
> > > +    addps     m2, m1
> > > +    minps     m2, m3
> >
> > > +    movaps    m5, [inq+sizeq]
> > > +    andps     m5, m4
> >
> > Same as in abs_pow34, remove movaps and do
> >
> > andps     m5, m4, [inq+sizeq]
> >
> > > +    orps      m2, m5
> > > +    cvttps2dq m2, m2
> > > +    mova      [outq+sizeq], m2
> > > +    add       sizeq, mmsize
> > > +    jl       .loop
> > > +    RET
> > > diff --git a/libavcodec/x86/aacencdsp_init.c b/libavcodec/x86/aacencdsp_
> > init.c
> > > new file mode 100644
> > > index 0000000..aefaa15
> > > --- /dev/null
> > > +++ b/libavcodec/x86/aacencdsp_init.c
> > > @@ -0,0 +1,43 @@
> > > +/*
> > > + * AAC encoder assembly optimizations
> > > + * Copyright (C) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
> > > + *
> > > + * This file is part of FFmpeg.
> > > + *
> > > + * FFmpeg is free software; you can redistribute it and/or
> > > + * modify it under the terms of the GNU Lesser General Public
> > > + * License as published by the Free Software Foundation; either
> > > + * version 2.1 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
> > > + * Lesser General Public License for more details.
> > > + *
> > > + * You should have received a copy of the GNU Lesser 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 "config.h"
> > > +
> > > +#include "libavutil/float_dsp.h"
> > > +#include "libavutil/x86/cpu.h"
> > > +#include "libavcodec/aacenc.h"
> > > +
> > > +void ff_abs_pow34_sse(float *out, const float *in, const int size);
> > > +
> > > +void ff_aac_quantize_bands_sse2(int *out, const float *in, const float
> > *scaled,
> > > +                                int size, int is_signed, int maxval,
> > const float Q34,
> > > +                                const float rounding);
> > > +
> > > +av_cold void ff_aac_dsp_init_x86(AACEncContext *s)
> > > +{
> > > +    int cpu_flags = av_get_cpu_flags();
> > > +
> > > +    if (EXTERNAL_SSE(cpu_flags))
> > > +        s->abs_pow34   = &ff_abs_pow34_sse;
> > > +
> > > +    if (EXTERNAL_SSE2(cpu_flags))
> > > +        s->quant_bands = &ff_aac_quantize_bands_sse2;
> >
> > Same with these.
> >
> > Passes fate on mingw-w64 (x86_64) after fixing the maxval mixup above.
> > Crashes otherwise.
> >
> > _______________________________________________
> > ffmpeg-devel mailing list
> > ffmpeg-devel@ffmpeg.org
> > http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> >
> 
> Hopefully I didn't mess anything up and the attached patch is the last
> revision of it.

>  aaccoder.c            |   27 ++++++++-------
>  aaccoder_trellis.h    |    2 -
>  aaccoder_twoloop.h    |    2 -
>  aacenc.c              |    6 +++
>  aacenc.h              |    6 +++
>  aacenc_is.c           |    6 +--
>  aacenc_ltp.c          |    4 +-
>  aacenc_pred.c         |    6 +--
>  aacenc_quantization.h |    4 +-
>  aacenc_utils.h        |    2 -
>  x86/Makefile          |    2 +
>  x86/aacencdsp.asm     |   86 ++++++++++++++++++++++++++++++++++++++++++++++++++
>  x86/aacencdsp_init.c  |   43 +++++++++++++++++++++++++
>  13 files changed, 170 insertions(+), 26 deletions(-)
> e9878e2480cd612e6852783003f47768f5ced052  0001-aacenc-add-SIMD-optimizations-for-abs_pow34-and-quan.patch
> From eb95cf5aeacdc7d191c5b7cf8df40a1e5306b733 Mon Sep 17 00:00:00 2001
> From: Rostislav Pehlivanov <atomnuker@gmail.com>
> Date: Sat, 8 Oct 2016 15:59:14 +0100
> Subject: [PATCH] aacenc: add SIMD optimizations for abs_pow34 and quantization

tested on x86 linux32/64 mingw32/64, arm qemu and mips qemu
all work now

[...]
  
Rostislav Pehlivanov Oct. 18, 2016, 8:42 p.m. UTC | #2
On 18 October 2016 at 21:04, Michael Niedermayer <michael@niedermayer.cc>
wrote:

> On Tue, Oct 18, 2016 at 05:33:13PM +0100, Rostislav Pehlivanov wrote:
> > On 18 October 2016 at 16:32, James Almer <jamrial@gmail.com> wrote:
> >
> > > On 10/18/2016 12:07 PM, Rostislav Pehlivanov wrote:
> > > > diff --git a/libavcodec/aacenc.c b/libavcodec/aacenc.c
> > > > index ee3cbf8..622f0ba 100644
> > > > --- a/libavcodec/aacenc.c
> > > > +++ b/libavcodec/aacenc.c
> > > > @@ -1033,6 +1033,12 @@ static av_cold int
> aac_encode_init(AVCodecContext
> > > *avctx)
> > > >      ff_lpc_init(&s->lpc, 2*avctx->frame_size, TNS_MAX_ORDER,
> > > FF_LPC_TYPE_LEVINSON);
> > > >      s->random_state = 0x1f2e3d4c;
> > > >
> > > > +    s->abs_pow34   = &abs_pow34_v;
> > > > +    s->quant_bands = &quantize_bands;
> > >
> > > No need for & in these.
> > >
> > > > +
> > > > +    if (ARCH_X86)
> > > > +        ff_aac_dsp_init_x86(s);
> > > > +
> > > >      if (HAVE_MIPSDSP)
> > > >          ff_aac_coder_init_mips(s);
> > >
> > > [...]
> > >
> > > > diff --git a/libavcodec/x86/aacencdsp.asm
> b/libavcodec/x86/aacencdsp.asm
> > > > new file mode 100644
> > > > index 0000000..dd7b022
> > > > --- /dev/null
> > > > +++ b/libavcodec/x86/aacencdsp.asm
> > > > @@ -0,0 +1,88 @@
> > > > +;**********************************************************
> > > ********************
> > > > +;* SIMD optimized AAC encoder DSP functions
> > > > +;*
> > > > +;* Copyright (C) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
> > > > +;*
> > > > +;* This file is part of FFmpeg.
> > > > +;*
> > > > +;* FFmpeg is free software; you can redistribute it and/or
> > > > +;* modify it under the terms of the GNU Lesser General Public
> > > > +;* License as published by the Free Software Foundation; either
> > > > +;* version 2.1 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
> > > > +;* Lesser General Public License for more details.
> > > > +;*
> > > > +;* You should have received a copy of the GNU Lesser 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_RODATA
> > > > +
> > > > +float_abs_mask: times 4 dd 0x7fffffff
> > > > +
> > > > +SECTION .text
> > > > +
> > > > +;**********************************************************
> *********
> > > > +;void ff_abs_pow34(float *out, const float *in, const int size);
> > > > +;**********************************************************
> *********
> > > > +INIT_XMM sse
> > > > +cglobal abs_pow34, 3, 3, 3, out, in, size
> > > > +    mova   m2, [float_abs_mask]
> > > > +    shl    sizeq, 2
> > > > +    add    inq, sizeq
> > > > +    add    outq, sizeq
> > > > +    neg    sizeq
> > > > +.loop:
> > > > +    movaps m0, [inq+sizeq]
> > > > +    andps  m0, m2
> > >
> > > Remove the movaps and do
> > >
> > > andps  m0, m2, [inq+sizeq]
> > >
> > > Instead. Sorry i didn't notice this last time.
> > >
> > > > +    sqrtps m1, m0
> > > > +    mulps  m0, m1
> > > > +    sqrtps m0, m0
> > > > +    mova   [outq+sizeq], m0
> > > > +    add    sizeq, mmsize
> > > > +    jl    .loop
> > > > +    RET
> > > > +
> > > > +;**********************************************************
> *********
> > > > +;void ff_aac_quantize_bands(int *out, const float *in, const float
> > > *scaled,
> > > > +;                           int size, int is_signed, int maxval,
> const
> > > float Q34,
> > > > +;                           const float rounding)
> > > > +;**********************************************************
> *********
> > > > +INIT_XMM sse2
> > > > +cglobal aac_quantize_bands, 5, 5, 6, out, in, scaled, size,
> is_signed,
> > > maxval, Q34, rounding
> > > > +%if UNIX64 == 0
> > > > +    movss     m0, Q34m
> > > > +    movss     m1, roundingm
> > > > +    cvtsi2ss  m3, maxvald
> > > > +%else
> > > > +    cvtsi2ss  m3, dword maxvalm
> > > > +%endif
> > >
> > > The other way around. Unix64 is the one that has maxval on a reg
> regardless
> > > of how you init the function, whereas win64 and any x86_32 target have
> it
> > > on stack.
> > >
> > > > +    shufps    m0, m0, 0
> > > > +    shufps    m1, m1, 0
> > > > +    shufps    m3, m3, 0
> > > > +    shl       is_signedd, 31
> > > > +    movd      m4, is_signedd
> > > > +    shufps    m4, m4, 0
> > > > +    shl       sized,   2
> > > > +    add       inq, sizeq
> > > > +    add       outq, sizeq
> > > > +    add       scaledq, sizeq
> > > > +    neg       sizeq
> > > > +.loop:
> > > > +    mulps     m2, m0, [scaledq+sizeq]
> > > > +    addps     m2, m1
> > > > +    minps     m2, m3
> > >
> > > > +    movaps    m5, [inq+sizeq]
> > > > +    andps     m5, m4
> > >
> > > Same as in abs_pow34, remove movaps and do
> > >
> > > andps     m5, m4, [inq+sizeq]
> > >
> > > > +    orps      m2, m5
> > > > +    cvttps2dq m2, m2
> > > > +    mova      [outq+sizeq], m2
> > > > +    add       sizeq, mmsize
> > > > +    jl       .loop
> > > > +    RET
> > > > diff --git a/libavcodec/x86/aacencdsp_init.c
> b/libavcodec/x86/aacencdsp_
> > > init.c
> > > > new file mode 100644
> > > > index 0000000..aefaa15
> > > > --- /dev/null
> > > > +++ b/libavcodec/x86/aacencdsp_init.c
> > > > @@ -0,0 +1,43 @@
> > > > +/*
> > > > + * AAC encoder assembly optimizations
> > > > + * Copyright (C) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
> > > > + *
> > > > + * This file is part of FFmpeg.
> > > > + *
> > > > + * FFmpeg is free software; you can redistribute it and/or
> > > > + * modify it under the terms of the GNU Lesser General Public
> > > > + * License as published by the Free Software Foundation; either
> > > > + * version 2.1 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
> > > > + * Lesser General Public License for more details.
> > > > + *
> > > > + * You should have received a copy of the GNU Lesser 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 "config.h"
> > > > +
> > > > +#include "libavutil/float_dsp.h"
> > > > +#include "libavutil/x86/cpu.h"
> > > > +#include "libavcodec/aacenc.h"
> > > > +
> > > > +void ff_abs_pow34_sse(float *out, const float *in, const int size);
> > > > +
> > > > +void ff_aac_quantize_bands_sse2(int *out, const float *in, const
> float
> > > *scaled,
> > > > +                                int size, int is_signed, int maxval,
> > > const float Q34,
> > > > +                                const float rounding);
> > > > +
> > > > +av_cold void ff_aac_dsp_init_x86(AACEncContext *s)
> > > > +{
> > > > +    int cpu_flags = av_get_cpu_flags();
> > > > +
> > > > +    if (EXTERNAL_SSE(cpu_flags))
> > > > +        s->abs_pow34   = &ff_abs_pow34_sse;
> > > > +
> > > > +    if (EXTERNAL_SSE2(cpu_flags))
> > > > +        s->quant_bands = &ff_aac_quantize_bands_sse2;
> > >
> > > Same with these.
> > >
> > > Passes fate on mingw-w64 (x86_64) after fixing the maxval mixup above.
> > > Crashes otherwise.
> > >
> > > _______________________________________________
> > > ffmpeg-devel mailing list
> > > ffmpeg-devel@ffmpeg.org
> > > http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> > >
> >
> > Hopefully I didn't mess anything up and the attached patch is the last
> > revision of it.
>
> >  aaccoder.c            |   27 ++++++++-------
> >  aaccoder_trellis.h    |    2 -
> >  aaccoder_twoloop.h    |    2 -
> >  aacenc.c              |    6 +++
> >  aacenc.h              |    6 +++
> >  aacenc_is.c           |    6 +--
> >  aacenc_ltp.c          |    4 +-
> >  aacenc_pred.c         |    6 +--
> >  aacenc_quantization.h |    4 +-
> >  aacenc_utils.h        |    2 -
> >  x86/Makefile          |    2 +
> >  x86/aacencdsp.asm     |   86 ++++++++++++++++++++++++++++++
> ++++++++++++++++++++
> >  x86/aacencdsp_init.c  |   43 +++++++++++++++++++++++++
> >  13 files changed, 170 insertions(+), 26 deletions(-)
> > e9878e2480cd612e6852783003f47768f5ced052  0001-aacenc-add-SIMD-
> optimizations-for-abs_pow34-and-quan.patch
> > From eb95cf5aeacdc7d191c5b7cf8df40a1e5306b733 Mon Sep 17 00:00:00 2001
> > From: Rostislav Pehlivanov <atomnuker@gmail.com>
> > Date: Sat, 8 Oct 2016 15:59:14 +0100
> > Subject: [PATCH] aacenc: add SIMD optimizations for abs_pow34 and
> quantization
>
> tested on x86 linux32/64 mingw32/64, arm qemu and mips qemu
> all work now
>
> [...]
>
> --
> Michael     GnuPG fingerprint: 9FF2128B147EF6730BADF133611EC787040B0FAB
>
> It is dangerous to be right in matters on which the established authorities
> are wrong. -- Voltaire
>
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel@ffmpeg.org
> http://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
>
Pushed
Thanks to both of you for the review and testing
  

Patch

From eb95cf5aeacdc7d191c5b7cf8df40a1e5306b733 Mon Sep 17 00:00:00 2001
From: Rostislav Pehlivanov <atomnuker@gmail.com>
Date: Sat, 8 Oct 2016 15:59:14 +0100
Subject: [PATCH] aacenc: add SIMD optimizations for abs_pow34 and quantization

Performance improvements:

quant_bands:
with:     681 decicycles in quant_bands, 8388453 runs,    155 skips
without: 1190 decicycles in quant_bands, 8388386 runs,    222 skips
Around 42% for the function

Twoloop coder:

abs_pow34:
with/without: 7.82s/8.17s
Around 4% for the entire encoder

Both:
with/without: 7.15s/8.17s
Around 12% for the entire encoder

Fast coder:

abs_pow34:
with/without: 3.40s/3.77s
Around 10% for the entire encoder

Both:
with/without: 3.02s/3.77s
Around 20% faster for the entire encoder

Signed-off-by: Rostislav Pehlivanov <atomnuker@gmail.com>
---
 libavcodec/aaccoder.c            | 27 +++++++------
 libavcodec/aaccoder_trellis.h    |  2 +-
 libavcodec/aaccoder_twoloop.h    |  2 +-
 libavcodec/aacenc.c              |  6 +++
 libavcodec/aacenc.h              |  6 +++
 libavcodec/aacenc_is.c           |  6 +--
 libavcodec/aacenc_ltp.c          |  4 +-
 libavcodec/aacenc_pred.c         |  6 +--
 libavcodec/aacenc_quantization.h |  4 +-
 libavcodec/aacenc_utils.h        |  2 +-
 libavcodec/x86/Makefile          |  2 +
 libavcodec/x86/aacencdsp.asm     | 86 ++++++++++++++++++++++++++++++++++++++++
 libavcodec/x86/aacencdsp_init.c  | 43 ++++++++++++++++++++
 13 files changed, 170 insertions(+), 26 deletions(-)
 create mode 100644 libavcodec/x86/aacencdsp.asm
 create mode 100644 libavcodec/x86/aacencdsp_init.c

diff --git a/libavcodec/aaccoder.c b/libavcodec/aaccoder.c
index 35787e8..9f3b4ed 100644
--- a/libavcodec/aaccoder.c
+++ b/libavcodec/aaccoder.c
@@ -88,7 +88,7 @@  static void encode_window_bands_info(AACEncContext *s, SingleChannelElement *sce
     float next_minrd = INFINITY;
     int next_mincb = 0;
 
-    abs_pow34_v(s->scoefs, sce->coeffs, 1024);
+    s->abs_pow34(s->scoefs, sce->coeffs, 1024);
     start = win*128;
     for (cb = 0; cb < CB_TOT_ALL; cb++) {
         path[0][cb].cost     = 0.0f;
@@ -299,7 +299,7 @@  static void search_for_quantizers_anmr(AVCodecContext *avctx, AACEncContext *s,
         }
     }
     idx = 1;
-    abs_pow34_v(s->scoefs, sce->coeffs, 1024);
+    s->abs_pow34(s->scoefs, sce->coeffs, 1024);
     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
         start = w*128;
         for (g = 0; g < sce->ics.num_swb; g++) {
@@ -446,7 +446,7 @@  static void search_for_quantizers_fast(AVCodecContext *avctx, AACEncContext *s,
 
     if (!allz)
         return;
-    abs_pow34_v(s->scoefs, sce->coeffs, 1024);
+    s->abs_pow34(s->scoefs, sce->coeffs, 1024);
     ff_quantize_band_cost_cache_init(s);
 
     for (w = 0; w < sce->ics.num_windows; w += sce->ics.group_len[w]) {
@@ -652,8 +652,8 @@  static void search_for_pns(AACEncContext *s, AVCodecContext *avctx, SingleChanne
                 s->fdsp->vector_fmul_scalar(PNS, PNS, scale, sce->ics.swb_sizes[g]);
                 pns_senergy = s->fdsp->scalarproduct_float(PNS, PNS, sce->ics.swb_sizes[g]);
                 pns_energy += pns_senergy;
-                abs_pow34_v(NOR34, &sce->coeffs[start_c], sce->ics.swb_sizes[g]);
-                abs_pow34_v(PNS34, PNS, sce->ics.swb_sizes[g]);
+                s->abs_pow34(NOR34, &sce->coeffs[start_c], sce->ics.swb_sizes[g]);
+                s->abs_pow34(PNS34, PNS, sce->ics.swb_sizes[g]);
                 dist1 += quantize_band_cost(s, &sce->coeffs[start_c],
                                             NOR34,
                                             sce->ics.swb_sizes[g],
@@ -757,8 +757,9 @@  static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
 {
     int start = 0, i, w, w2, g, sid_sf_boost, prev_mid, prev_side;
     uint8_t nextband0[128], nextband1[128];
-    float M[128], S[128];
-    float *L34 = s->scoefs, *R34 = s->scoefs + 128, *M34 = s->scoefs + 128*2, *S34 = s->scoefs + 128*3;
+    float *M   = s->scoefs + 128*0, *S   = s->scoefs + 128*1;
+    float *L34 = s->scoefs + 128*2, *R34 = s->scoefs + 128*3;
+    float *M34 = s->scoefs + 128*4, *S34 = s->scoefs + 128*5;
     const float lambda = s->lambda;
     const float mslambda = FFMIN(1.0f, lambda / 120.f);
     SingleChannelElement *sce0 = &cpe->ch[0];
@@ -789,8 +790,8 @@  static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
                         S[i] =  M[i]
                               - sce1->coeffs[start+(w+w2)*128+i];
                     }
-                    abs_pow34_v(M34, M, sce0->ics.swb_sizes[g]);
-                    abs_pow34_v(S34, S, sce0->ics.swb_sizes[g]);
+                    s->abs_pow34(M34, M, sce0->ics.swb_sizes[g]);
+                    s->abs_pow34(S34, S, sce0->ics.swb_sizes[g]);
                     for (i = 0; i < sce0->ics.swb_sizes[g]; i++ ) {
                         Mmax = FFMAX(Mmax, M34[i]);
                         Smax = FFMAX(Smax, S34[i]);
@@ -833,10 +834,10 @@  static void search_for_ms(AACEncContext *s, ChannelElement *cpe)
                                   - sce1->coeffs[start+(w+w2)*128+i];
                         }
 
-                        abs_pow34_v(L34, sce0->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
-                        abs_pow34_v(R34, sce1->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
-                        abs_pow34_v(M34, M,                         sce0->ics.swb_sizes[g]);
-                        abs_pow34_v(S34, S,                         sce0->ics.swb_sizes[g]);
+                        s->abs_pow34(L34, sce0->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
+                        s->abs_pow34(R34, sce1->coeffs+start+(w+w2)*128, sce0->ics.swb_sizes[g]);
+                        s->abs_pow34(M34, M,                         sce0->ics.swb_sizes[g]);
+                        s->abs_pow34(S34, S,                         sce0->ics.swb_sizes[g]);
                         dist1 += quantize_band_cost(s, &sce0->coeffs[start + (w+w2)*128],
                                                     L34,
                                                     sce0->ics.swb_sizes[g],
diff --git a/libavcodec/aaccoder_trellis.h b/libavcodec/aaccoder_trellis.h
index 0230052..940ebf0 100644
--- a/libavcodec/aaccoder_trellis.h
+++ b/libavcodec/aaccoder_trellis.h
@@ -70,7 +70,7 @@  static void codebook_trellis_rate(AACEncContext *s, SingleChannelElement *sce,
     float next_minbits = INFINITY;
     int next_mincb = 0;
 
-    abs_pow34_v(s->scoefs, sce->coeffs, 1024);
+    s->abs_pow34(s->scoefs, sce->coeffs, 1024);
     start = win*128;
     for (cb = 0; cb < CB_TOT_ALL; cb++) {
         path[0][cb].cost     = run_bits+4;
diff --git a/libavcodec/aaccoder_twoloop.h b/libavcodec/aaccoder_twoloop.h
index 42aea52..fb9849e 100644
--- a/libavcodec/aaccoder_twoloop.h
+++ b/libavcodec/aaccoder_twoloop.h
@@ -291,7 +291,7 @@  static void search_for_quantizers_twoloop(AVCodecContext *avctx,
 
     if (!allz)
         return;
-    abs_pow34_v(s->scoefs, sce->coeffs, 1024);
+    s->abs_pow34(s->scoefs, sce->coeffs, 1024);
     ff_quantize_band_cost_cache_init(s);
 
     for (i = 0; i < sizeof(minsf) / sizeof(minsf[0]); ++i)
diff --git a/libavcodec/aacenc.c b/libavcodec/aacenc.c
index ee3cbf8..363ed05 100644
--- a/libavcodec/aacenc.c
+++ b/libavcodec/aacenc.c
@@ -1033,6 +1033,12 @@  static av_cold int aac_encode_init(AVCodecContext *avctx)
     ff_lpc_init(&s->lpc, 2*avctx->frame_size, TNS_MAX_ORDER, FF_LPC_TYPE_LEVINSON);
     s->random_state = 0x1f2e3d4c;
 
+    s->abs_pow34   = abs_pow34_v;
+    s->quant_bands = quantize_bands;
+
+    if (ARCH_X86)
+        ff_aac_dsp_init_x86(s);
+
     if (HAVE_MIPSDSP)
         ff_aac_coder_init_mips(s);
 
diff --git a/libavcodec/aacenc.h b/libavcodec/aacenc.h
index 1ace00d..38a9734 100644
--- a/libavcodec/aacenc.h
+++ b/libavcodec/aacenc.h
@@ -127,11 +127,17 @@  typedef struct AACEncContext {
     uint16_t quantize_band_cost_cache_generation;
     AACQuantizeBandCostCacheEntry quantize_band_cost_cache[256][128]; ///< memoization area for quantize_band_cost
 
+    void (*abs_pow34)(float *out, const float *in, const int size);
+    void (*quant_bands)(int *out, const float *in, const float *scaled,
+                        int size, int is_signed, int maxval, const float Q34,
+                        const float rounding);
+
     struct {
         float *samples;
     } buffer;
 } AACEncContext;
 
+void ff_aac_dsp_init_x86(AACEncContext *s);
 void ff_aac_coder_init_mips(AACEncContext *c);
 void ff_quantize_band_cost_cache_init(struct AACEncContext *s);
 
diff --git a/libavcodec/aacenc_is.c b/libavcodec/aacenc_is.c
index 473897b..2f5b7eb 100644
--- a/libavcodec/aacenc_is.c
+++ b/libavcodec/aacenc_is.c
@@ -59,9 +59,9 @@  struct AACISError ff_aac_is_encoding_err(AACEncContext *s, ChannelElement *cpe,
         float minthr = FFMIN(band0->threshold, band1->threshold);
         for (i = 0; i < sce0->ics.swb_sizes[g]; i++)
             IS[i] = (L[start+(w+w2)*128+i] + phase*R[start+(w+w2)*128+i])*sqrt(ener0/ener01);
-        abs_pow34_v(L34, &L[start+(w+w2)*128], sce0->ics.swb_sizes[g]);
-        abs_pow34_v(R34, &R[start+(w+w2)*128], sce0->ics.swb_sizes[g]);
-        abs_pow34_v(I34, IS,                   sce0->ics.swb_sizes[g]);
+        s->abs_pow34(L34, &L[start+(w+w2)*128], sce0->ics.swb_sizes[g]);
+        s->abs_pow34(R34, &R[start+(w+w2)*128], sce0->ics.swb_sizes[g]);
+        s->abs_pow34(I34, IS,                   sce0->ics.swb_sizes[g]);
         maxval = find_max_val(1, sce0->ics.swb_sizes[g], I34);
         is_band_type = find_min_book(maxval, is_sf_idx);
         dist1 += quantize_band_cost(s, &L[start + (w+w2)*128], L34,
diff --git a/libavcodec/aacenc_ltp.c b/libavcodec/aacenc_ltp.c
index b9d43b4..1bec85b 100644
--- a/libavcodec/aacenc_ltp.c
+++ b/libavcodec/aacenc_ltp.c
@@ -190,8 +190,8 @@  void ff_aac_search_for_ltp(AACEncContext *s, SingleChannelElement *sce,
                 FFPsyBand *band = &s->psy.ch[s->cur_channel].psy_bands[(w+w2)*16+g];
                 for (i = 0; i < sce->ics.swb_sizes[g]; i++)
                     PCD[i] = sce->coeffs[start+(w+w2)*128+i] - sce->lcoeffs[start+(w+w2)*128+i];
-                abs_pow34_v(C34,  &sce->coeffs[start+(w+w2)*128],  sce->ics.swb_sizes[g]);
-                abs_pow34_v(PCD34, PCD, sce->ics.swb_sizes[g]);
+                s->abs_pow34(C34,  &sce->coeffs[start+(w+w2)*128],  sce->ics.swb_sizes[g]);
+                s->abs_pow34(PCD34, PCD, sce->ics.swb_sizes[g]);
                 dist1 += quantize_band_cost(s, &sce->coeffs[start+(w+w2)*128], C34, sce->ics.swb_sizes[g],
                                             sce->sf_idx[(w+w2)*16+g], sce->band_type[(w+w2)*16+g],
                                             s->lambda/band->threshold, INFINITY, &bits_tmp1, NULL, 0);
diff --git a/libavcodec/aacenc_pred.c b/libavcodec/aacenc_pred.c
index e77a3de..d111192 100644
--- a/libavcodec/aacenc_pred.c
+++ b/libavcodec/aacenc_pred.c
@@ -270,7 +270,7 @@  void ff_aac_search_for_pred(AACEncContext *s, SingleChannelElement *sce)
             continue;
 
         /* Normal coefficients */
-        abs_pow34_v(O34, &sce->coeffs[start_coef], num_coeffs);
+        s->abs_pow34(O34, &sce->coeffs[start_coef], num_coeffs);
         dist1 = quantize_and_encode_band_cost(s, NULL, &sce->coeffs[start_coef], NULL,
                                               O34, num_coeffs, sce->sf_idx[sfb],
                                               cb_n, s->lambda / band->threshold, INFINITY, &cost1, NULL, 0);
@@ -279,7 +279,7 @@  void ff_aac_search_for_pred(AACEncContext *s, SingleChannelElement *sce)
         /* Encoded coefficients - needed for #bits, band type and quant. error */
         for (i = 0; i < num_coeffs; i++)
             SENT[i] = sce->coeffs[start_coef + i] - sce->prcoeffs[start_coef + i];
-        abs_pow34_v(S34, SENT, num_coeffs);
+        s->abs_pow34(S34, SENT, num_coeffs);
         if (cb_n < RESERVED_BT)
             cb_p = av_clip(find_min_book(find_max_val(1, num_coeffs, S34), sce->sf_idx[sfb]), cb_min, cb_max);
         else
@@ -291,7 +291,7 @@  void ff_aac_search_for_pred(AACEncContext *s, SingleChannelElement *sce)
         /* Reconstructed coefficients - needed for distortion measurements */
         for (i = 0; i < num_coeffs; i++)
             sce->prcoeffs[start_coef + i] += QERR[i] != 0.0f ? (sce->prcoeffs[start_coef + i] - QERR[i]) : 0.0f;
-        abs_pow34_v(P34, &sce->prcoeffs[start_coef], num_coeffs);
+        s->abs_pow34(P34, &sce->prcoeffs[start_coef], num_coeffs);
         if (cb_n < RESERVED_BT)
             cb_p = av_clip(find_min_book(find_max_val(1, num_coeffs, P34), sce->sf_idx[sfb]), cb_min, cb_max);
         else
diff --git a/libavcodec/aacenc_quantization.h b/libavcodec/aacenc_quantization.h
index 4250407..fc5a46b 100644
--- a/libavcodec/aacenc_quantization.h
+++ b/libavcodec/aacenc_quantization.h
@@ -74,10 +74,10 @@  static av_always_inline float quantize_and_encode_band_cost_template(
         return cost * lambda;
     }
     if (!scaled) {
-        abs_pow34_v(s->scoefs, in, size);
+        s->abs_pow34(s->scoefs, in, size);
         scaled = s->scoefs;
     }
-    quantize_bands(s->qcoefs, in, scaled, size, Q34, !BT_UNSIGNED, aac_cb_maxval[cb], ROUNDING);
+    s->quant_bands(s->qcoefs, in, scaled, size, !BT_UNSIGNED, aac_cb_maxval[cb], Q34, ROUNDING);
     if (BT_UNSIGNED) {
         off = 0;
     } else {
diff --git a/libavcodec/aacenc_utils.h b/libavcodec/aacenc_utils.h
index ff9188a..bef4c10 100644
--- a/libavcodec/aacenc_utils.h
+++ b/libavcodec/aacenc_utils.h
@@ -63,7 +63,7 @@  static inline int quant(float coef, const float Q, const float rounding)
 }
 
 static inline void quantize_bands(int *out, const float *in, const float *scaled,
-                                  int size, float Q34, int is_signed, int maxval,
+                                  int size, int is_signed, int maxval, const float Q34,
                                   const float rounding)
 {
     int i;
diff --git a/libavcodec/x86/Makefile b/libavcodec/x86/Makefile
index 522b6c2..1db1137 100644
--- a/libavcodec/x86/Makefile
+++ b/libavcodec/x86/Makefile
@@ -42,6 +42,7 @@  OBJS-$(CONFIG_XMM_CLOBBER_TEST)        += x86/w64xmmtest.o
 # decoders/encoders
 OBJS-$(CONFIG_AAC_DECODER)             += x86/aacpsdsp_init.o          \
                                           x86/sbrdsp_init.o
+OBJS-$(CONFIG_AAC_ENCODER)             += x86/aacencdsp_init.o
 OBJS-$(CONFIG_ADPCM_G722_DECODER)      += x86/g722dsp_init.o
 OBJS-$(CONFIG_ADPCM_G722_ENCODER)      += x86/g722dsp_init.o
 OBJS-$(CONFIG_ALAC_DECODER)            += x86/alacdsp_init.o
@@ -132,6 +133,7 @@  YASM-OBJS-$(CONFIG_VP8DSP)             += x86/vp8dsp.o                  \
 # decoders/encoders
 YASM-OBJS-$(CONFIG_AAC_DECODER)        += x86/aacpsdsp.o                \
                                           x86/sbrdsp.o
+YASM-OBJS-$(CONFIG_AAC_ENCODER)        += x86/aacencdsp.o
 YASM-OBJS-$(CONFIG_ADPCM_G722_DECODER) += x86/g722dsp.o
 YASM-OBJS-$(CONFIG_ADPCM_G722_ENCODER) += x86/g722dsp.o
 YASM-OBJS-$(CONFIG_ALAC_DECODER)       += x86/alacdsp.o
diff --git a/libavcodec/x86/aacencdsp.asm b/libavcodec/x86/aacencdsp.asm
new file mode 100644
index 0000000..97af571
--- /dev/null
+++ b/libavcodec/x86/aacencdsp.asm
@@ -0,0 +1,86 @@ 
+;******************************************************************************
+;* SIMD optimized AAC encoder DSP functions
+;*
+;* Copyright (C) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
+;*
+;* This file is part of FFmpeg.
+;*
+;* FFmpeg is free software; you can redistribute it and/or
+;* modify it under the terms of the GNU Lesser General Public
+;* License as published by the Free Software Foundation; either
+;* version 2.1 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
+;* Lesser General Public License for more details.
+;*
+;* You should have received a copy of the GNU Lesser 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_RODATA
+
+float_abs_mask: times 4 dd 0x7fffffff
+
+SECTION .text
+
+;*******************************************************************
+;void ff_abs_pow34(float *out, const float *in, const int size);
+;*******************************************************************
+INIT_XMM sse
+cglobal abs_pow34, 3, 3, 3, out, in, size
+    mova   m2, [float_abs_mask]
+    shl    sizeq, 2
+    add    inq, sizeq
+    add    outq, sizeq
+    neg    sizeq
+.loop:
+    andps  m0, m2, [inq+sizeq]
+    sqrtps m1, m0
+    mulps  m0, m1
+    sqrtps m0, m0
+    mova   [outq+sizeq], m0
+    add    sizeq, mmsize
+    jl    .loop
+    RET
+
+;*******************************************************************
+;void ff_aac_quantize_bands(int *out, const float *in, const float *scaled,
+;                           int size, int is_signed, int maxval, const float Q34,
+;                           const float rounding)
+;*******************************************************************
+INIT_XMM sse2
+cglobal aac_quantize_bands, 5, 5, 6, out, in, scaled, size, is_signed, maxval, Q34, rounding
+%if UNIX64 == 0
+    movss     m0, Q34m
+    movss     m1, roundingm
+    cvtsi2ss  m3, dword maxvalm
+%else
+    cvtsi2ss  m3, maxvald
+%endif
+    shufps    m0, m0, 0
+    shufps    m1, m1, 0
+    shufps    m3, m3, 0
+    shl       is_signedd, 31
+    movd      m4, is_signedd
+    shufps    m4, m4, 0
+    shl       sized,   2
+    add       inq, sizeq
+    add       outq, sizeq
+    add       scaledq, sizeq
+    neg       sizeq
+.loop:
+    mulps     m2, m0, [scaledq+sizeq]
+    addps     m2, m1
+    minps     m2, m3
+    andps     m5, m4, [inq+sizeq]
+    orps      m2, m5
+    cvttps2dq m2, m2
+    mova      [outq+sizeq], m2
+    add       sizeq, mmsize
+    jl       .loop
+    RET
diff --git a/libavcodec/x86/aacencdsp_init.c b/libavcodec/x86/aacencdsp_init.c
new file mode 100644
index 0000000..d761c3c
--- /dev/null
+++ b/libavcodec/x86/aacencdsp_init.c
@@ -0,0 +1,43 @@ 
+/*
+ * AAC encoder assembly optimizations
+ * Copyright (C) 2016 Rostislav Pehlivanov <atomnuker@gmail.com>
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 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
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser 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 "config.h"
+
+#include "libavutil/float_dsp.h"
+#include "libavutil/x86/cpu.h"
+#include "libavcodec/aacenc.h"
+
+void ff_abs_pow34_sse(float *out, const float *in, const int size);
+
+void ff_aac_quantize_bands_sse2(int *out, const float *in, const float *scaled,
+                                int size, int is_signed, int maxval, const float Q34,
+                                const float rounding);
+
+av_cold void ff_aac_dsp_init_x86(AACEncContext *s)
+{
+    int cpu_flags = av_get_cpu_flags();
+
+    if (EXTERNAL_SSE(cpu_flags))
+        s->abs_pow34   = ff_abs_pow34_sse;
+
+    if (EXTERNAL_SSE2(cpu_flags))
+        s->quant_bands = ff_aac_quantize_bands_sse2;
+}
-- 
2.9.3