[FFmpeg-devel] [PATCH v1 1/2] lavf/vf_transpose: add exif orientation support

Jun Li junli1026 at gmail.com
Mon May 27 02:06:04 EEST 2019


On Sun, May 26, 2019 at 2:09 AM Paul B Mahol <onemda at gmail.com> wrote:

> On 5/26/19, Jun Li <junli1026 at gmail.com> wrote:
> > On Sun, May 26, 2019 at 1:16 AM Paul B Mahol <onemda at gmail.com> wrote:
> >
> >> On 5/25/19, Jun Li <junli1026 at gmail.com> wrote:
> >> > Add exif orientation support and expose an option.
> >> > ---
> >> >  libavfilter/vf_transpose.c | 258
> +++++++++++++++++++++++++++++--------
> >> >  1 file changed, 207 insertions(+), 51 deletions(-)
> >> >
> >> > diff --git a/libavfilter/vf_transpose.c b/libavfilter/vf_transpose.c
> >> > index dd54947bd9..4aebfb9ee4 100644
> >> > --- a/libavfilter/vf_transpose.c
> >> > +++ b/libavfilter/vf_transpose.c
> >> > @@ -46,6 +46,9 @@ typedef struct TransVtable {
> >> >      void (*transpose_block)(uint8_t *src, ptrdiff_t src_linesize,
> >> >                              uint8_t *dst, ptrdiff_t dst_linesize,
> >> >                              int w, int h);
> >> > +    void (*copyline_block)(uint8_t *src, ptrdiff_t src_linesize,
> >> > +                           uint8_t *dst, ptrdiff_t dst_linesize,
> >> > +                           int line_dir, int w, int h);
> >> >  } TransVtable;
> >> >
> >>
> >> This is slow, the operations should be in one loop.
> >>
> >
> > Thanks Paul for review.
> > I see transpose is using " ctx->internal->execute(ctx, filter_slice, &td,
> > NULL, FFMIN(outlink->h, ff_filter_get_nb_threads(ctx)));" .
> > which looks like leveraging multi-threading.
> > If we do that in one loop, I guess it would be hard to use
> multi-threading
> > to accelerate ? But correct me if I am wrong.
>
> Transpose filter does one thing and that is transposing.
>
> If I'm not mistaken you are adding vertical and horizontal flipping?
>

Thanks for review.
Yes, I am doing vflip, hflip and rotate180(vflip+hflip).


> If you really need to do that, do it efficiently in one loop, by that I
> mean
> you do not do cascading operations at all.
>

Are you suggesting not use function pointer, and call the function directly
like this:
            switch (s->pixsteps[plane]) {
                case 1:  copyline_block_8(src, srclinesize, dst,
dstlinesize, line_dir, outw, end - start); break;
                case 2:  copyline_block_16(src, srclinesize, dst,
dstlinesize, line_dir, outw, end - start); break;
                case 3:  copyline_block_24(src, srclinesize, dst,
dstlinesize, line_dir, outw, end - start); break;
                case 4:  copyline_block_32(src, srclinesize, dst,
dstlinesize, line_dir, outw, end - start); break;
                case 6:  copyline_block_48(src, srclinesize, dst,
dstlinesize, line_dir, outw, end - start); break;
                case 8:  copyline_block_64(src, srclinesize, dst,
dstlinesize, line_dir, outw, end - start); break;
            }

I did some perf comparsion, they are very close and no significant
difference, about 9ms for one frame.

Also I did some comparison between current hflip filter and this patch,
they take almost the same time, about 9ms too on my machine and I see 13
threads/jobs are used for filtering.
I just use the simple command "ffmpeg  -i test.bmp -vf
transpose=orientation=2  -y hello.jpg".
Current vflip wins a lot and I see it is not copying frame at all, just
like "orientation=1" or "passthrough" cases.

I hope I understand you correctly, especially about "cascading operations".
But correct me if I am wrong.
Thanks for your time Paul.

Best Regards,
-Jun

>
> > Best Regards,
> > -Jun
> >
> >
> >> >  typedef struct TransContext {
> >> > @@ -56,7 +59,7 @@ typedef struct TransContext {
> >> >
> >> >      int passthrough;    ///< PassthroughType, landscape passthrough
> >> > mode
> >> > enabled
> >> >      int dir;            ///< TransposeDir
> >> > -
> >> > +    int orientation;    ///< Orientation
> >> >      TransVtable vtables[4];
> >> >  } TransContext;
> >> >
> >> > @@ -182,6 +185,105 @@ static void transpose_8x8_64_c(uint8_t *src,
> >> ptrdiff_t
> >> > src_linesize,
> >> >      transpose_block_64_c(src, src_linesize, dst, dst_linesize, 8, 8);
> >> >  }
> >> >
> >> > +
> >> > +static void copyline_block_8(uint8_t *src, ptrdiff_t src_linesize,
> >> > +                             uint8_t *dst, ptrdiff_t dst_linesize,
> >> > +                             int line_dir, int w, int h)
> >> > +{
> >> > +    int x, y, i;
> >> > +    for (y = 0; y < h; y++, dst += dst_linesize, src +=
> src_linesize) {
> >> > +        for (x = 0; x < w; x ++) {
> >> > +            i = line_dir < 0 ? w-x-1 : x;
> >> > +            dst[i] = src[x];
> >> > +        }
> >> > +    }
> >> > +}
> >> > +
> >> > +static void copyline_block_16(uint8_t *src, ptrdiff_t src_linesize,
> >> > +                              uint8_t *dst, ptrdiff_t dst_linesize,
> >> > +                              int line_dir, int w, int h)
> >> > +{
> >> > +    int x, y, i;
> >> > +    for (y = 0; y < h; y++, dst += dst_linesize, src +=
> src_linesize) {
> >> > +        for (x = 0; x < w; x ++) {
> >> > +            i = line_dir < 0 ? w-x-1 : x;
> >> > +            *((uint16_t *)(dst + 2*i)) = *((uint16_t *)(src + 2*x));
> >> > +        }
> >> > +    }
> >> > +}
> >> > +
> >> > +static void copyline_block_24(uint8_t *src, ptrdiff_t src_linesize,
> >> > +                              uint8_t *dst, ptrdiff_t dst_linesize,
> >> > +                              int line_dir, int w, int h)
> >> > +{
> >> > +    int x, y, i;
> >> > +    for (y = 0; y < h; y++, dst += dst_linesize, src +=
> src_linesize) {
> >> > +        for (x = 0; x < w; x++) {
> >> > +            int32_t v = AV_RB24(src + 3*x);
> >> > +            i = line_dir < 0 ? w-x-1 : x;
> >> > +            AV_WB24(dst + 3*i, v);
> >> > +        }
> >> > +    }
> >> > +}
> >> > +
> >> > +static void copyline_block_32(uint8_t *src, ptrdiff_t src_linesize,
> >> > +                              uint8_t *dst, ptrdiff_t dst_linesize,
> >> > +                              int line_dir, int w, int h)
> >> > +{
> >> > +    int x, y, i;
> >> > +    for (y = 0; y < h; y++, dst += dst_linesize, src +=
> src_linesize) {
> >> > +        for (x = 0; x < w; x ++) {
> >> > +            i = line_dir < 0 ? w-x-1 : x;
> >> > +            *((uint32_t *)(dst + 4*i)) = *((uint32_t *)(src + 4*x));
> >> > +        }
> >> > +    }
> >> > +}
> >> > +
> >> > +static void copyline_block_48(uint8_t *src, ptrdiff_t src_linesize,
> >> > +                              uint8_t *dst, ptrdiff_t dst_linesize,
> >> > +                              int line_dir, int w, int h)
> >> > +{
> >> > +    int x, y, i;
> >> > +    for (y = 0; y < h; y++, dst += dst_linesize, src +=
> src_linesize) {
> >> > +        for (x = 0; x < w; x++) {
> >> > +            int64_t v = AV_RB48(src + 6*x);
> >> > +            i = line_dir < 0 ? w-x-1 : x;
> >> > +            AV_WB48(dst + 6*i, v);
> >> > +        }
> >> > +    }
> >> > +}
> >> > +
> >> > +static void copyline_block_64(uint8_t *src, ptrdiff_t src_linesize,
> >> > +                              uint8_t *dst, ptrdiff_t dst_linesize,
> >> > +                              int line_dir, int w, int h)
> >> > +{
> >> > +    int x, y, i;
> >> > +    for (y = 0; y < h; y++, dst += dst_linesize, src +=
> src_linesize) {
> >> > +        for (x = 0; x < w; x ++) {
> >> > +            i = line_dir < 0 ? w-x-1 : x;
> >> > +            *((uint64_t *)(dst + 8*i)) = *((uint64_t *)(src + 8*x));
> >> > +        }
> >> > +    }
> >> > +}
> >> > +
> >> > +static void set_outlink_width_height(AVFilterLink *inlink,
> AVFilterLink
> >> > *outlink, int transpose)
> >> > +{
> >> > +    if (transpose) {
> >> > +        outlink->w = inlink->h;
> >> > +        outlink->h = inlink->w;
> >> > +
> >> > +        if (inlink->sample_aspect_ratio.num)
> >> > +            outlink->sample_aspect_ratio = av_div_q((AVRational) {
> 1, 1
> >> },
> >> > +
> >> > inlink->sample_aspect_ratio);
> >> > +        else
> >> > +            outlink->sample_aspect_ratio =
> inlink->sample_aspect_ratio;
> >> > +    } else {
> >> > +        outlink->w = inlink->w;
> >> > +        outlink->h = inlink->h;
> >> > +        outlink->sample_aspect_ratio = inlink->sample_aspect_ratio;
> >> > +    }
> >> > +}
> >> > +
> >> >  static int config_props_output(AVFilterLink *outlink)
> >> >  {
> >> >      AVFilterContext *ctx = outlink->src;
> >> > @@ -213,33 +315,44 @@ static int config_props_output(AVFilterLink
> >> *outlink)
> >> >
> >> >      av_assert0(desc_in->nb_components == desc_out->nb_components);
> >> >
> >> > -
> >> >      av_image_fill_max_pixsteps(s->pixsteps, NULL, desc_out);
> >> > +
> >> > +    if (s->orientation && s->orientation < 5) {
> >> > +        outlink->h = inlink->h;
> >> > +        outlink->w = inlink->w;
> >> > +        outlink->sample_aspect_ratio = inlink->sample_aspect_ratio;
> >> > +    } else {
> >> > +        outlink->w = inlink->h;
> >> > +        outlink->h = inlink->w;
> >> >
> >> > -    outlink->w = inlink->h;
> >> > -    outlink->h = inlink->w;
> >> > -
> >> > -    if (inlink->sample_aspect_ratio.num)
> >> > -        outlink->sample_aspect_ratio = av_div_q((AVRational) { 1, 1
> },
> >> > -
> >> > inlink->sample_aspect_ratio);
> >> > -    else
> >> > -        outlink->sample_aspect_ratio = inlink->sample_aspect_ratio;
> >> > +        if (inlink->sample_aspect_ratio.num)
> >> > +            outlink->sample_aspect_ratio = av_div_q((AVRational) {
> 1, 1
> >> },
> >> > +
> >> > inlink->sample_aspect_ratio);
> >> > +        else
> >> > +            outlink->sample_aspect_ratio =
> inlink->sample_aspect_ratio;
> >> > +    }
> >> >
> >> >      for (int i = 0; i < 4; i++) {
> >> >          TransVtable *v = &s->vtables[i];
> >> >          switch (s->pixsteps[i]) {
> >> >          case 1: v->transpose_block = transpose_block_8_c;
> >> > -                v->transpose_8x8   = transpose_8x8_8_c;  break;
> >> > +                v->transpose_8x8   = transpose_8x8_8_c;
> >> > +                v->copyline_block  = copyline_block_8; break;
> >> >          case 2: v->transpose_block = transpose_block_16_c;
> >> > -                v->transpose_8x8   = transpose_8x8_16_c; break;
> >> > +                v->transpose_8x8   = transpose_8x8_16_c;
> >> > +                v->copyline_block  = copyline_block_16; break;
> >> >          case 3: v->transpose_block = transpose_block_24_c;
> >> > -                v->transpose_8x8   = transpose_8x8_24_c; break;
> >> > +                v->transpose_8x8   = transpose_8x8_24_c;
> >> > +                v->copyline_block  = copyline_block_24; break;
> >> >          case 4: v->transpose_block = transpose_block_32_c;
> >> > -                v->transpose_8x8   = transpose_8x8_32_c; break;
> >> > +                v->transpose_8x8   = transpose_8x8_32_c;
> >> > +                v->copyline_block  = copyline_block_32; break;
> >> >          case 6: v->transpose_block = transpose_block_48_c;
> >> > -                v->transpose_8x8   = transpose_8x8_48_c; break;
> >> > +                v->transpose_8x8   = transpose_8x8_48_c;
> >> > +                v->copyline_block  = copyline_block_48; break;
> >> >          case 8: v->transpose_block = transpose_block_64_c;
> >> > -                v->transpose_8x8   = transpose_8x8_64_c; break;
> >> > +                v->transpose_8x8   = transpose_8x8_64_c;
> >> > +                v->copyline_block  = copyline_block_64; break;
> >> >          }
> >> >      }
> >> >
> >> > @@ -285,42 +398,80 @@ static int filter_slice(AVFilterContext *ctx,
> void
> >> > *arg, int jobnr,
> >> >          uint8_t *dst, *src;
> >> >          int dstlinesize, srclinesize;
> >> >          int x, y;
> >> > +        int dir = s->dir;
> >> >          TransVtable *v = &s->vtables[plane];
> >> > +
> >> > +        if (s->orientation > 0 && s->orientation < 5) {
> >> > +            int line_dir = 1;
> >> > +            dstlinesize = out->linesize[plane];
> >> > +            dst         = out->data[plane] + start * dstlinesize;
> >> > +            srclinesize = in->linesize[plane];
> >> > +            src         = in->data[plane] + start * srclinesize;
> >> > +
> >> > +            switch (s->orientation) {
> >> > +                case 2: // mirror horizontal
> >> > +                    line_dir     = -1;
> >> > +                    break;
> >> > +                case 3: // rotate 180
> >> > +                    dst          = out->data[plane] + (outh - start -
> >> 1) *
> >> > dstlinesize;
> >> > +                    line_dir     = -1;
> >> > +                    dstlinesize *= -1;
> >> > +                    break;
> >> > +                case 4: // mirror vertical
> >> > +                    dst          = out->data[plane] + (outh - start -
> >> 1) *
> >> > dstlinesize;
> >> > +                    dstlinesize *= -1;
> >> > +                    break;
> >> > +                default:
> >> > +                    break;
> >> > +            }
> >> > +            v->copyline_block(src, srclinesize, dst, dstlinesize,
> >> line_dir,
> >> > outw, end - start);
> >> > +        } else {
> >> > +            dstlinesize = out->linesize[plane];
> >> > +            dst         = out->data[plane] + start * dstlinesize;
> >> > +            src         = in->data[plane];
> >> > +            srclinesize = in->linesize[plane];
> >> > +            switch (s->orientation) {
> >> > +                case 5: // mirror horizontal and rotate 270 CW
> >> > +                    dir = 0; break;
> >> > +                case 6: // rotate 90 CW
> >> > +                    dir = 1; break;
> >> > +                case 7: // mirror horizontal and rotate 90 CW
> >> > +                    dir = 3; break;
> >> > +                case 8: // rotate 270 CW
> >> > +                    dir = 2; break;
> >> > +                default: break;
> >> > +            }
> >> >
> >> > -        dstlinesize = out->linesize[plane];
> >> > -        dst         = out->data[plane] + start * dstlinesize;
> >> > -        src         = in->data[plane];
> >> > -        srclinesize = in->linesize[plane];
> >> > -
> >> > -        if (s->dir & 1) {
> >> > -            src         += in->linesize[plane] * (inh - 1);
> >> > -            srclinesize *= -1;
> >> > -        }
> >> > +            if (dir & 1) {
> >> > +                src         += in->linesize[plane] * (inh - 1);
> >> > +                srclinesize *= -1;
> >> > +            }
> >> >
> >> > -        if (s->dir & 2) {
> >> > -            dst          = out->data[plane] + dstlinesize * (outh -
> >> start -
> >> > 1);
> >> > -            dstlinesize *= -1;
> >> > -        }
> >> > +            if (dir & 2) {
> >> > +                dst          = out->data[plane] + dstlinesize *
> (outh -
> >> > start - 1);
> >> > +                dstlinesize *= -1;
> >> > +            }
> >> >
> >> > -        for (y = start; y < end - 7; y += 8) {
> >> > -            for (x = 0; x < outw - 7; x += 8) {
> >> > -                v->transpose_8x8(src + x * srclinesize + y * pixstep,
> >> > -                                 srclinesize,
> >> > -                                 dst + (y - start) * dstlinesize + x
> *
> >> > pixstep,
> >> > -                                 dstlinesize);
> >> > +            for (y = start; y < end - 7; y += 8) {
> >> > +                for (x = 0; x < outw - 7; x += 8) {
> >> > +                    v->transpose_8x8(src + x * srclinesize + y *
> >> pixstep,
> >> > +                                    srclinesize,
> >> > +                                    dst + (y - start) * dstlinesize
> + x
> >> *
> >> > pixstep,
> >> > +                                    dstlinesize);
> >> > +                }
> >> > +                if (outw - x > 0 && end - y > 0)
> >> > +                    v->transpose_block(src + x * srclinesize + y *
> >> pixstep,
> >> > +                                    srclinesize,
> >> > +                                    dst + (y - start) * dstlinesize
> + x
> >> *
> >> > pixstep,
> >> > +                                    dstlinesize, outw - x, end - y);
> >> >              }
> >> > -            if (outw - x > 0 && end - y > 0)
> >> > -                v->transpose_block(src + x * srclinesize + y *
> pixstep,
> >> > -                                   srclinesize,
> >> > -                                   dst + (y - start) * dstlinesize +
> x
> >> > *
> >> > pixstep,
> >> > -                                   dstlinesize, outw - x, end - y);
> >> > -        }
> >> >
> >> > -        if (end - y > 0)
> >> > -            v->transpose_block(src + 0 * srclinesize + y * pixstep,
> >> > -                               srclinesize,
> >> > -                               dst + (y - start) * dstlinesize + 0 *
> >> > pixstep,
> >> > -                               dstlinesize, outw, end - y);
> >> > +            if (end - y > 0)
> >> > +                v->transpose_block(src + 0 * srclinesize + y *
> pixstep,
> >> > +                                srclinesize,
> >> > +                                dst + (y - start) * dstlinesize + 0 *
> >> > pixstep,
> >> > +                                dstlinesize, outw, end - y);
> >> > +        }
> >> >      }
> >> >
> >> >      return 0;
> >> > @@ -334,7 +485,7 @@ static int filter_frame(AVFilterLink *inlink,
> >> > AVFrame
> >> > *in)
> >> >      ThreadData td;
> >> >      AVFrame *out;
> >> >
> >> > -    if (s->passthrough)
> >> > +    if (s->passthrough || s->orientation == 1)
> >> >          return ff_filter_frame(outlink, in);
> >> >
> >> >      out = ff_get_video_buffer(outlink, outlink->w, outlink->h);
> >> > @@ -347,8 +498,13 @@ static int filter_frame(AVFilterLink *inlink,
> >> AVFrame
> >> > *in)
> >> >      if (in->sample_aspect_ratio.num == 0) {
> >> >          out->sample_aspect_ratio = in->sample_aspect_ratio;
> >> >      } else {
> >> > -        out->sample_aspect_ratio.num = in->sample_aspect_ratio.den;
> >> > -        out->sample_aspect_ratio.den = in->sample_aspect_ratio.num;
> >> > +        if (s->orientation > 0 && s->orientation < 5) {
> >> > +            out->sample_aspect_ratio.num =
> in->sample_aspect_ratio.num;
> >> > +            out->sample_aspect_ratio.den =
> in->sample_aspect_ratio.den;
> >> > +        } else {
> >> > +            out->sample_aspect_ratio.num =
> in->sample_aspect_ratio.den;
> >> > +            out->sample_aspect_ratio.den =
> in->sample_aspect_ratio.num;
> >> > +        }
> >> >      }
> >> >
> >> >      td.in = in, td.out = out;
> >> > @@ -366,13 +522,13 @@ static const AVOption transpose_options[] = {
> >> >          { "clock",       "rotate clockwise",
> >> 0,
> >> > AV_OPT_TYPE_CONST, { .i64 = TRANSPOSE_CLOCK       }, .flags=FLAGS,
> .unit
> >> =
> >> > "dir" },
> >> >          { "cclock",      "rotate counter-clockwise",
> >> 0,
> >> > AV_OPT_TYPE_CONST, { .i64 = TRANSPOSE_CCLOCK      }, .flags=FLAGS,
> .unit
> >> =
> >> > "dir" },
> >> >          { "clock_flip",  "rotate clockwise with vertical flip",
> >>  0,
> >> > AV_OPT_TYPE_CONST, { .i64 = TRANSPOSE_CLOCK_FLIP  }, .flags=FLAGS,
> .unit
> >> =
> >> > "dir" },
> >> > -
> >> > +
> >> >      { "passthrough", "do not apply transposition if the input matches
> >> the
> >> > specified geometry",
> >> >        OFFSET(passthrough), AV_OPT_TYPE_INT,
> >> {.i64=TRANSPOSE_PT_TYPE_NONE},
> >> > 0, INT_MAX, FLAGS, "passthrough" },
> >> >          { "none",      "always apply transposition",   0,
> >> > AV_OPT_TYPE_CONST, {.i64=TRANSPOSE_PT_TYPE_NONE},      INT_MIN,
> INT_MAX,
> >> > FLAGS, "passthrough" },
> >> >          { "portrait",  "preserve portrait geometry",   0,
> >> > AV_OPT_TYPE_CONST, {.i64=TRANSPOSE_PT_TYPE_PORTRAIT},  INT_MIN,
> INT_MAX,
> >> > FLAGS, "passthrough" },
> >> >          { "landscape", "preserve landscape geometry",  0,
> >> > AV_OPT_TYPE_CONST, {.i64=TRANSPOSE_PT_TYPE_LANDSCAPE}, INT_MIN,
> INT_MAX,
> >> > FLAGS, "passthrough" },
> >> > -
> >> > +    { "orientation", "set exif orientation", OFFSET(orientation),
> >> > AV_OPT_TYPE_INT, {.i64 = 0 },  0, 8, FLAGS, "orientation" },
> >> >      { NULL }
> >> >  };
> >> >
> >> > --
> >> > 2.17.1
> >> >
> >> > _______________________________________________
> >> > ffmpeg-devel mailing list
> >> > ffmpeg-devel at ffmpeg.org
> >> > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> >> >
> >> > To unsubscribe, visit link above, or email
> >> > ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
> >> _______________________________________________
> >> ffmpeg-devel mailing list
> >> ffmpeg-devel at ffmpeg.org
> >> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> >>
> >> To unsubscribe, visit link above, or email
> >> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
> > _______________________________________________
> > ffmpeg-devel mailing list
> > ffmpeg-devel at ffmpeg.org
> > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
> >
> > To unsubscribe, visit link above, or email
> > ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".
> _______________________________________________
> ffmpeg-devel mailing list
> ffmpeg-devel at ffmpeg.org
> https://ffmpeg.org/mailman/listinfo/ffmpeg-devel
>
> To unsubscribe, visit link above, or email
> ffmpeg-devel-request at ffmpeg.org with subject "unsubscribe".


More information about the ffmpeg-devel mailing list