[FFmpeg-devel] [PATCH] avfilter/vf_transpose: add x86 SIMD

James Almer jamrial at gmail.com
Mon Oct 21 20:29:36 EEST 2019


On 10/21/2019 1:36 PM, James Almer wrote:
> On 10/21/2019 12:45 PM, Paul B Mahol wrote:
>> Signed-off-by: Paul B Mahol <onemda at gmail.com>
>> ---
>>  libavfilter/transpose.h             |  10 +++
>>  libavfilter/vf_transpose.c          |  18 ++---
>>  libavfilter/x86/Makefile            |   2 +
>>  libavfilter/x86/vf_transpose.asm    | 104 ++++++++++++++++++++++++++++
>>  libavfilter/x86/vf_transpose_init.c |  49 +++++++++++++
>>  5 files changed, 174 insertions(+), 9 deletions(-)
>>  create mode 100644 libavfilter/x86/vf_transpose.asm
>>  create mode 100644 libavfilter/x86/vf_transpose_init.c
>>
>> diff --git a/libavfilter/transpose.h b/libavfilter/transpose.h
>> index aa262b9487..f73a42864f 100644
>> --- a/libavfilter/transpose.h
>> +++ b/libavfilter/transpose.h
>> @@ -34,4 +34,14 @@ enum TransposeDir {
>>      TRANSPOSE_VFLIP,
>>  };
>>  
>> +typedef struct TransVtable {
>> +    void (*transpose_8x8)(uint8_t *src, ptrdiff_t src_linesize,
>> +                          uint8_t *dst, ptrdiff_t dst_linesize);
>> +    void (*transpose_block)(uint8_t *src, ptrdiff_t src_linesize,
>> +                            uint8_t *dst, ptrdiff_t dst_linesize,
>> +                            int w, int h);
>> +} TransVtable;
>> +
>> +void ff_transpose_init_x86(TransVtable *v, int pixstep);
>> +
>>  #endif
>> diff --git a/libavfilter/vf_transpose.c b/libavfilter/vf_transpose.c
>> index dd54947bd9..16ac6c311a 100644
>> --- a/libavfilter/vf_transpose.c
>> +++ b/libavfilter/vf_transpose.c
>> @@ -40,14 +40,6 @@
>>  #include "video.h"
>>  #include "transpose.h"
>>  
>> -typedef struct TransVtable {
>> -    void (*transpose_8x8)(uint8_t *src, ptrdiff_t src_linesize,
>> -                          uint8_t *dst, ptrdiff_t dst_linesize);
>> -    void (*transpose_block)(uint8_t *src, ptrdiff_t src_linesize,
>> -                            uint8_t *dst, ptrdiff_t dst_linesize,
>> -                            int w, int h);
>> -} TransVtable;
>> -
>>  typedef struct TransContext {
>>      const AVClass *class;
>>      int hsub, vsub;
>> @@ -243,7 +235,15 @@ static int config_props_output(AVFilterLink *outlink)
>>          }
>>      }
>>  
>> -    av_log(ctx, AV_LOG_VERBOSE,
>> +    if (ARCH_X86) {
>> +        for (int i = 0; i < 4; i++) {
>> +            TransVtable *v = &s->vtables[i];
>> +
>> +            ff_transpose_init_x86(v, s->pixsteps[i]);
>> +        }
>> +    }
>> +
>> +     av_log(ctx, AV_LOG_VERBOSE,
>>             "w:%d h:%d dir:%d -> w:%d h:%d rotation:%s vflip:%d\n",
>>             inlink->w, inlink->h, s->dir, outlink->w, outlink->h,
>>             s->dir == 1 || s->dir == 3 ? "clockwise" : "counterclockwise",
>> diff --git a/libavfilter/x86/Makefile b/libavfilter/x86/Makefile
>> index 06f832e36c..8d97e46c3f 100644
>> --- a/libavfilter/x86/Makefile
>> +++ b/libavfilter/x86/Makefile
>> @@ -31,6 +31,7 @@ OBJS-$(CONFIG_STEREO3D_FILTER)               += x86/vf_stereo3d_init.o
>>  OBJS-$(CONFIG_TBLEND_FILTER)                 += x86/vf_blend_init.o
>>  OBJS-$(CONFIG_THRESHOLD_FILTER)              += x86/vf_threshold_init.o
>>  OBJS-$(CONFIG_TINTERLACE_FILTER)             += x86/vf_tinterlace_init.o
>> +OBJS-$(CONFIG_TRANSPOSE_FILTER)              += x86/vf_transpose_init.o
>>  OBJS-$(CONFIG_VOLUME_FILTER)                 += x86/af_volume_init.o
>>  OBJS-$(CONFIG_V360_FILTER)                   += x86/vf_v360_init.o
>>  OBJS-$(CONFIG_W3FDIF_FILTER)                 += x86/vf_w3fdif_init.o
>> @@ -69,6 +70,7 @@ X86ASM-OBJS-$(CONFIG_STEREO3D_FILTER)        += x86/vf_stereo3d.o
>>  X86ASM-OBJS-$(CONFIG_TBLEND_FILTER)          += x86/vf_blend.o
>>  X86ASM-OBJS-$(CONFIG_THRESHOLD_FILTER)       += x86/vf_threshold.o
>>  X86ASM-OBJS-$(CONFIG_TINTERLACE_FILTER)      += x86/vf_interlace.o
>> +X86ASM-OBJS-$(CONFIG_TRANSPOSE_FILTER)       += x86/vf_transpose.o
>>  X86ASM-OBJS-$(CONFIG_VOLUME_FILTER)          += x86/af_volume.o
>>  X86ASM-OBJS-$(CONFIG_V360_FILTER)            += x86/vf_v360.o
>>  X86ASM-OBJS-$(CONFIG_W3FDIF_FILTER)          += x86/vf_w3fdif.o
>> diff --git a/libavfilter/x86/vf_transpose.asm b/libavfilter/x86/vf_transpose.asm
>> new file mode 100644
>> index 0000000000..6d925d5d97
>> --- /dev/null
>> +++ b/libavfilter/x86/vf_transpose.asm
>> @@ -0,0 +1,104 @@
>> +;*****************************************************************************
>> +;* x86-optimized functions for transpose filter
>> +;*
>> +;* Copyright (C) 2019 Paul B Mahol
>> +;*
>> +;* 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 .text
>> +
>> +;------------------------------------------------------------------------------
>> +; void ff_transpose_8x8(uint8_t *src, ptrdiff_t src_linesize,
>> +;                       uint8_t *dst, ptrdiff_t dst_linesize)
>> +;------------------------------------------------------------------------------
>> +
>> +INIT_XMM sse4
>> +cglobal transpose_8x8_8, 4,4,8, src, src_linesize, dst, dst_linesize
>> +    movu    m0, [srcq]
>> +    add    srcq, src_linesizeq
>> +    movu    m1, [srcq]
>> +    add    srcq, src_linesizeq
>> +    movu    m2, [srcq]
>> +    add    srcq, src_linesizeq
>> +    movu    m3, [srcq]
>> +    add    srcq, src_linesizeq
>> +    movu    m4, [srcq]
>> +    add    srcq, src_linesizeq
>> +    movu    m5, [srcq]
>> +    add    srcq, src_linesizeq
>> +    movu    m6, [srcq]
>> +    add    srcq, src_linesizeq
>> +    movu    m7, [srcq]
>> +
>> +    TRANSPOSE_8X8B 0, 1, 2, 3, 4, 5, 6, 7

This macro seems to only use the lower eight bytes of each register, for
both input and output. So i think you can replace the movu's with movq's.

>> +
>> +    movq [dstq], m0
>> +    add    dstq, dst_linesizeq
>> +    movq [dstq], m1
>> +    add    dstq, dst_linesizeq
>> +    movq [dstq], m2
>> +    add    dstq, dst_linesizeq
>> +    movq [dstq], m3
>> +    add    dstq, dst_linesizeq
>> +    movq [dstq], m4
>> +    add    dstq, dst_linesizeq
>> +    movq [dstq], m5
>> +    add    dstq, dst_linesizeq
>> +    movq [dstq], m6
>> +    add    dstq, dst_linesizeq
>> +    movq [dstq], m7
>> +    RET
> 
>     lea linesize3q, [src_linesizeq*3]
>     movu            m0, [srcq+src_linesizeq*0]
>     movu            m1, [srcq+src_linesizeq*1]
>     movu            m2, [srcq+src_linesizeq*2]
>     movu            m3, [srcq+linesize3q]
>     lea           srcq, [srcq+src_linesizeq*4]
>     movu            m4, [srcq+src_linesizeq*0]
>     movu            m5, [srcq+src_linesizeq*1]
>     movu            m6, [srcq+src_linesizeq*2]
>     movu            m7, [srcq+linesize3q]
> 
>     TRANSPOSE_8X8B 0, 1, 2, 3, 4, 5, 6, 7
> 
>     lea linesize3q, [dst_linesizeq*3]
>     movu [dstq+dst_linesizeq*0], m0
>     movu [dstq+dst_linesizeq*1], m1
>     movu [dstq+dst_linesizeq*2], m2
>     movu      [dstq+linesize3q], m3
>     lea                    dstq, [dstq+dst_linesizeq*4]
>     movu [dstq+dst_linesizeq*0], m4
>     movu [dstq+dst_linesizeq*1], m5
>     movu [dstq+dst_linesizeq*2], m6
>     movu      [dstq+linesize3q], m7
>     RET

Also, i obviously screwed up by using movu here.


More information about the ffmpeg-devel mailing list