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

[James Almer]

On 10/17/2016 6:24 PM, Rostislav Pehlivanov wrote:
> diff --git a/libavcodec/aacenc_utils.h b/libavcodec/aacenc_utils.h
> index ff9188a..f5cf77d 100644
> --- a/libavcodec/aacenc_utils.h
> +++ b/libavcodec/aacenc_utils.h
> @@ -37,7 +37,7 @@
>  #define ROUND_TO_ZERO 0.1054f
>  #define C_QUANT 0.4054f
>  
> -static inline void abs_pow34_v(float *out, const float *in, const int size)
> +static inline void abs_pow34_v(float *out, const float *in, const int64_t size)

Why int64_t? There's no need for values that big...

>  {
>      int i;
>      for (i = 0; i < size; i++) {

...And it's certainly not correct, seeing this for loop here.

> diff --git a/libavcodec/x86/aacencdsp.asm b/libavcodec/x86/aacencdsp.asm
> new file mode 100644
> index 0000000..ff4019f
> --- /dev/null
> +++ b/libavcodec/x86/aacencdsp.asm
> @@ -0,0 +1,87 @@
> +;******************************************************************************
> +;* SIMD optimized AAC encoder DSP functions
> +;*
> +;* Copyright (C) 2016 Rostislav Pehlivanov <atomnuker at 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_sse(float *out, const float *in, const int64_t 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:
> +    mova   m0, [inq+sizeq]
> +    andps  m0, m2
> +    sqrtps m1, m0
> +    mulps  m0, m1
> +    sqrtps m0, m0
> +    mova   [outq+sizeq], m0
> +    add    sizeq, mmsize
> +    jl     .loop
> +    RET
> +
> +;*******************************************************************
> +;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)
> +;*******************************************************************
> +INIT_XMM sse2
> +cglobal aac_quantize_bands, 6, 6, 6, out, in, scaled, size, is_signed, maxval, Q34, rounding

5, 5, 6

No need to load maxval into a gpr on x86_32 and win64 when cvtsi2ss
can also load it directly from memory.

> +%if UNIX64 == 0
> +    movss     m0, Q34m
> +    movss     m1, roundingm
> +%endif
> +    SPLATD    m0
> +    SPLATD    m1

shufps m0, m0, 0
shufps m1, m1, 0

On sse2, SPLATD will expand to pshufd, so better stay in float domain.

If you add an AVX/FMA3 version of this function, you could instead use
vbroadcastss in them, and save yourself the movss.

> +    cvtsi2ss  m3, maxvald
> +    SPLATD    m3

%if UNIX64
    cvtsi2ss  m3, maxvald
%else
    cvtsi2ss  m3, dword maxvalm
%endif
    shufps    m3, m3, 0

After you made the function 5, 5, 6.

> +    shl       is_signedd, 31
> +    movd      m4, is_signedd
> +    SPLATD    m4

Use shufps here since the instruction using this register in the loop
should be a float one.

> +    shl       sizeq,   2

Even if it doesn't come from stack on win64, better use "sized" anyway.

> +    add       inq,     sizeq
> +    add       outq,    sizeq
> +    add       scaledq, sizeq
> +    neg       sizeq
> +.loop:
> +    mova      m2, [scaledq+sizeq]
> +    mulps     m2, m0

mulps m2, m0, [scaledq+sizeq]

> +    addps     m2, m1

You could combine the mulps and addps into a single fmaddps if you add
an FMA3 version. Something like

movaps  m2, [scaledq+sizeq]
fmaddps m2, m2, m0, m1

The movaps is needed because, unlike FMA4's fmaddps which is
non-destructive, FMA3 needs dst to be the same as one of the three src
registers.

> +    minps     m2, m3
> +    mova      m5, [inq+sizeq]

movaps m5, [inq+sizeq]

> +    pand      m5, m4

andps m5, m4

> +    orps      m2, m5
> +    cvttps2dq m2, m2
> +    mova      [outq+sizeq], m2
> +    add       sizeq, mmsize
> +    jl       .loop
> +    RET

Can't you make these function also process eight floats per loop?
YMM versions would be nice, seeing how much these functions affect
overall encoding speed.