Message ID | 20170530173429.10885-1-kmark937@gmail.com |
---|---|
State | Accepted |
Commit | 3385989b98be7940044e4f0a6b431a0a00abf2fa |
Headers | show |
On Tue, May 30, 2017 at 01:34:29PM -0400, Kevin Mark wrote: > Variables pertaining to the main video are now available when > using the scale2ref filter. This allows, as an example, scaling a > video with another as a reference point while maintaining the > original aspect ratio of the primary/non-reference video. > > Consider the following graph: scale2ref=iw/6:-1 [main][ref] > This will scale [main] to 1/6 the width of [ref] while maintaining > the aspect ratio. This works well when the AR of [ref] is equal to > the AR of [main] only. What the above filter really does is > maintain the AR of [ref] when scaling [main]. So in all non-same-AR > situations [main] will appear stretched or compressed to conform to > the same AR of the reference video. Without doing this calculation > externally there is no way to scale in reference to another input > while maintaining AR in libavfilter. > > To make this possible, we introduce eight new constants to be used > in the w and h expressions only in the scale2ref filter: > > * main_w/main_h: width/height of the main input video > * main_a: aspect ratio of the main input video > * main_sar: sample aspect ratio of the main input video > * main_dar: display aspect ratio of the main input video > * main_hsub/main_vsub: horiz/vert chroma subsample vals of main > * mdar: a shorthand alias of main_dar > > Of course, not all of these constants are needed for maintaining the > AR, but adding additional constants in line of what is available for > in/out allows for other scaling possibilities I have not imagined. > > So to now scale a video to 1/6 the size of another video using the > width and maintaining its own aspect ratio you can do this: > > scale2ref=iw/6:ow/mdar [main][ref] > > This is ideal for picture-in-picture configurations where you could > have a square or 4:3 video overlaid on a corner of a larger 16:9 > feed all while keeping the scaled video in the corner at its correct > aspect ratio and always the same size relative to the larger video. > > I've tried to re-use as much code as possible. I could not find a way > to avoid duplication of the var_names array. It must now be kept in > sync with the other (the normal one and the scale2ref one) for > everything to work which does not seem ideal. For every new variable > introduced/removed into/from the normal scale filter one must be > added/removed to/from the scale2ref version. Suggestions on how to > avoid var_names duplication are welcome. > > var_values has been increased to always be large enough for the > additional scale2ref variables. I do not forsee this being a problem > as the names variable will always be the correct size. From my > understanding of av_expr_parse_and_eval it will stop processing > variables when it runs out of names even though there may be > additional (potentially uninitialized) entries in the values array. > The ideal solution here would be using a variable-length array but > that is unsupported in C90. > > This patch does not remove any functionality and is strictly a > feature patch. There are no API changes. Behavior does not change for > any previously valid inputs. > > The applicable documentation has also been updated. > > Signed-off-by: Kevin Mark <kmark937@gmail.com> > --- > doc/filters.texi | 26 ++++++++++++++++++- > libavfilter/scale.c | 72 ++++++++++++++++++++++++++++++++++++++++++++++++++--- > 2 files changed, 93 insertions(+), 5 deletions(-) applied can you add a fate test that uses the new identifers ? Thanks [...]
On 6/1/2017 6:37 PM, Michael Niedermayer wrote: > On Tue, May 30, 2017 at 01:34:29PM -0400, Kevin Mark wrote: >> Variables pertaining to the main video are now available when >> using the scale2ref filter. This allows, as an example, scaling a >> video with another as a reference point while maintaining the >> original aspect ratio of the primary/non-reference video. >> >> Consider the following graph: scale2ref=iw/6:-1 [main][ref] >> This will scale [main] to 1/6 the width of [ref] while maintaining >> the aspect ratio. This works well when the AR of [ref] is equal to >> the AR of [main] only. What the above filter really does is >> maintain the AR of [ref] when scaling [main]. So in all non-same-AR >> situations [main] will appear stretched or compressed to conform to >> the same AR of the reference video. Without doing this calculation >> externally there is no way to scale in reference to another input >> while maintaining AR in libavfilter. >> >> To make this possible, we introduce eight new constants to be used >> in the w and h expressions only in the scale2ref filter: >> >> * main_w/main_h: width/height of the main input video >> * main_a: aspect ratio of the main input video >> * main_sar: sample aspect ratio of the main input video >> * main_dar: display aspect ratio of the main input video >> * main_hsub/main_vsub: horiz/vert chroma subsample vals of main >> * mdar: a shorthand alias of main_dar >> >> Of course, not all of these constants are needed for maintaining the >> AR, but adding additional constants in line of what is available for >> in/out allows for other scaling possibilities I have not imagined. >> >> So to now scale a video to 1/6 the size of another video using the >> width and maintaining its own aspect ratio you can do this: >> >> scale2ref=iw/6:ow/mdar [main][ref] >> >> This is ideal for picture-in-picture configurations where you could >> have a square or 4:3 video overlaid on a corner of a larger 16:9 >> feed all while keeping the scaled video in the corner at its correct >> aspect ratio and always the same size relative to the larger video. >> >> I've tried to re-use as much code as possible. I could not find a way >> to avoid duplication of the var_names array. It must now be kept in >> sync with the other (the normal one and the scale2ref one) for >> everything to work which does not seem ideal. For every new variable >> introduced/removed into/from the normal scale filter one must be >> added/removed to/from the scale2ref version. Suggestions on how to >> avoid var_names duplication are welcome. >> >> var_values has been increased to always be large enough for the >> additional scale2ref variables. I do not forsee this being a problem >> as the names variable will always be the correct size. From my >> understanding of av_expr_parse_and_eval it will stop processing >> variables when it runs out of names even though there may be >> additional (potentially uninitialized) entries in the values array. >> The ideal solution here would be using a variable-length array but >> that is unsupported in C90. >> >> This patch does not remove any functionality and is strictly a >> feature patch. There are no API changes. Behavior does not change for >> any previously valid inputs. >> >> The applicable documentation has also been updated. >> >> Signed-off-by: Kevin Mark <kmark937@gmail.com> >> --- >> doc/filters.texi | 26 ++++++++++++++++++- >> libavfilter/scale.c | 72 ++++++++++++++++++++++++++++++++++++++++++++++++++--- >> 2 files changed, 93 insertions(+), 5 deletions(-) > > applied > > can you add a fate test that uses the new identifers ? > > Thanks This broke 820 tests with Valgrind. All of them seem to point to ==1253== Invalid read of size 8 ==1253== at 0x697ECA: ff_scale_eval_dimensions (scale.c:118)
On Fri, Jun 2, 2017 at 9:16 PM, James Almer <jamrial@gmail.com> wrote: > This broke 820 tests with Valgrind. All of them seem to point to > > ==1253== Invalid read of size 8 > ==1253== at 0x697ECA: ff_scale_eval_dimensions (scale.c:118) scale.c:118: const char scale2ref = outlink->src->inputs[1] == inlink; I believe the issue must be that ff_scale_eval_dimensions blindly assumes a second input exists. A potential fix may be: const char scale2ref = outlink->src->nb_inputs == 2 && outlink->src->inputs[1] == inlink; As long as I am correct in assuming that nb_inputs will equal the number of input links. I'll have to reproduce the valgrind tests myself and see if this fixes it. Thanks James!
On Fri, Jun 2, 2017 at 9:37 PM, Kevin Mark <kmark937@gmail.com> wrote: > On Fri, Jun 2, 2017 at 9:16 PM, James Almer <jamrial@gmail.com> wrote: >> This broke 820 tests with Valgrind. Patch has been submitted. Thanks again for the report!
On Thu, Jun 1, 2017 at 5:37 PM, Michael Niedermayer <michael@niedermayer.cc> wrote: > applied > > can you add a fate test that uses the new identifers ? > > Thanks Certainly. Submitted a patch for a scale2ref-specific test. Thanks, Kevin
diff --git a/doc/filters.texi b/doc/filters.texi index 107fe61447..2cea6b74e6 100644 --- a/doc/filters.texi +++ b/doc/filters.texi @@ -12429,7 +12429,31 @@ Supersampling Scale (resize) the input video, based on a reference video. See the scale filter for available options, scale2ref supports the same but -uses the reference video instead of the main input as basis. +uses the reference video instead of the main input as basis. scale2ref also +supports the following additional constants for the @option{w} and +@option{h} options: + +@table @var +@item main_w +@item main_h +The main input video's width and height + +@item main_a +The same as @var{main_w} / @var{main_h} + +@item main_sar +The main input video's sample aspect ratio + +@item main_dar, mdar +The main input video's display aspect ratio. Calculated from +@code{(main_w / main_h) * main_sar}. + +@item main_hsub +@item main_vsub +The main input video's horizontal and vertical chroma subsample values. +For example for the pixel format "yuv422p" @var{hsub} is 2 and @var{vsub} +is 1. +@end table @subsection Examples diff --git a/libavfilter/scale.c b/libavfilter/scale.c index 50cd442849..e3a2fb5923 100644 --- a/libavfilter/scale.c +++ b/libavfilter/scale.c @@ -60,6 +60,49 @@ enum var_name { VARS_NB }; +/** + * This must be kept in sync with var_names so that it is always a + * complete list of var_names with the scale2ref specific names + * appended. scale2ref values must appear in the order they appear + * in the var_name_scale2ref enum but also be below all of the + * non-scale2ref specific values. + */ +static const char *const var_names_scale2ref[] = { + "PI", + "PHI", + "E", + "in_w", "iw", + "in_h", "ih", + "out_w", "ow", + "out_h", "oh", + "a", + "sar", + "dar", + "hsub", + "vsub", + "ohsub", + "ovsub", + "main_w", + "main_h", + "main_a", + "main_sar", + "main_dar", "mdar", + "main_hsub", + "main_vsub", + NULL +}; + +enum var_name_scale2ref { + VAR_S2R_MAIN_W, + VAR_S2R_MAIN_H, + VAR_S2R_MAIN_A, + VAR_S2R_MAIN_SAR, + VAR_S2R_MAIN_DAR, VAR_S2R_MDAR, + VAR_S2R_MAIN_HSUB, + VAR_S2R_MAIN_VSUB, + VARS_S2R_NB +}; + int ff_scale_eval_dimensions(void *log_ctx, const char *w_expr, const char *h_expr, AVFilterLink *inlink, AVFilterLink *outlink, @@ -72,7 +115,16 @@ int ff_scale_eval_dimensions(void *log_ctx, int factor_w, factor_h; int eval_w, eval_h; int ret; - double var_values[VARS_NB], res; + const char scale2ref = outlink->src->inputs[1] == inlink; + double var_values[VARS_NB + VARS_S2R_NB], res; + const AVPixFmtDescriptor *main_desc; + const AVFilterLink *main_link; + const char *const *names = scale2ref ? var_names_scale2ref : var_names; + + if (scale2ref) { + main_link = outlink->src->inputs[0]; + main_desc = av_pix_fmt_desc_get(main_link->format); + } var_values[VAR_PI] = M_PI; var_values[VAR_PHI] = M_PHI; @@ -90,20 +142,32 @@ int ff_scale_eval_dimensions(void *log_ctx, var_values[VAR_OHSUB] = 1 << out_desc->log2_chroma_w; var_values[VAR_OVSUB] = 1 << out_desc->log2_chroma_h; + if (scale2ref) { + var_values[VARS_NB + VAR_S2R_MAIN_W] = main_link->w; + var_values[VARS_NB + VAR_S2R_MAIN_H] = main_link->h; + var_values[VARS_NB + VAR_S2R_MAIN_A] = (double) main_link->w / main_link->h; + var_values[VARS_NB + VAR_S2R_MAIN_SAR] = main_link->sample_aspect_ratio.num ? + (double) main_link->sample_aspect_ratio.num / main_link->sample_aspect_ratio.den : 1; + var_values[VARS_NB + VAR_S2R_MAIN_DAR] = var_values[VARS_NB + VAR_S2R_MDAR] = + var_values[VARS_NB + VAR_S2R_MAIN_A] * var_values[VARS_NB + VAR_S2R_MAIN_SAR]; + var_values[VARS_NB + VAR_S2R_MAIN_HSUB] = 1 << main_desc->log2_chroma_w; + var_values[VARS_NB + VAR_S2R_MAIN_VSUB] = 1 << main_desc->log2_chroma_h; + } + /* evaluate width and height */ av_expr_parse_and_eval(&res, (expr = w_expr), - var_names, var_values, + names, var_values, NULL, NULL, NULL, NULL, NULL, 0, log_ctx); eval_w = var_values[VAR_OUT_W] = var_values[VAR_OW] = res; if ((ret = av_expr_parse_and_eval(&res, (expr = h_expr), - var_names, var_values, + names, var_values, NULL, NULL, NULL, NULL, NULL, 0, log_ctx)) < 0) goto fail; eval_h = var_values[VAR_OUT_H] = var_values[VAR_OH] = res; /* evaluate again the width, as it may depend on the output height */ if ((ret = av_expr_parse_and_eval(&res, (expr = w_expr), - var_names, var_values, + names, var_values, NULL, NULL, NULL, NULL, NULL, 0, log_ctx)) < 0) goto fail; eval_w = res;
Variables pertaining to the main video are now available when using the scale2ref filter. This allows, as an example, scaling a video with another as a reference point while maintaining the original aspect ratio of the primary/non-reference video. Consider the following graph: scale2ref=iw/6:-1 [main][ref] This will scale [main] to 1/6 the width of [ref] while maintaining the aspect ratio. This works well when the AR of [ref] is equal to the AR of [main] only. What the above filter really does is maintain the AR of [ref] when scaling [main]. So in all non-same-AR situations [main] will appear stretched or compressed to conform to the same AR of the reference video. Without doing this calculation externally there is no way to scale in reference to another input while maintaining AR in libavfilter. To make this possible, we introduce eight new constants to be used in the w and h expressions only in the scale2ref filter: * main_w/main_h: width/height of the main input video * main_a: aspect ratio of the main input video * main_sar: sample aspect ratio of the main input video * main_dar: display aspect ratio of the main input video * main_hsub/main_vsub: horiz/vert chroma subsample vals of main * mdar: a shorthand alias of main_dar Of course, not all of these constants are needed for maintaining the AR, but adding additional constants in line of what is available for in/out allows for other scaling possibilities I have not imagined. So to now scale a video to 1/6 the size of another video using the width and maintaining its own aspect ratio you can do this: scale2ref=iw/6:ow/mdar [main][ref] This is ideal for picture-in-picture configurations where you could have a square or 4:3 video overlaid on a corner of a larger 16:9 feed all while keeping the scaled video in the corner at its correct aspect ratio and always the same size relative to the larger video. I've tried to re-use as much code as possible. I could not find a way to avoid duplication of the var_names array. It must now be kept in sync with the other (the normal one and the scale2ref one) for everything to work which does not seem ideal. For every new variable introduced/removed into/from the normal scale filter one must be added/removed to/from the scale2ref version. Suggestions on how to avoid var_names duplication are welcome. var_values has been increased to always be large enough for the additional scale2ref variables. I do not forsee this being a problem as the names variable will always be the correct size. From my understanding of av_expr_parse_and_eval it will stop processing variables when it runs out of names even though there may be additional (potentially uninitialized) entries in the values array. The ideal solution here would be using a variable-length array but that is unsupported in C90. This patch does not remove any functionality and is strictly a feature patch. There are no API changes. Behavior does not change for any previously valid inputs. The applicable documentation has also been updated. Signed-off-by: Kevin Mark <kmark937@gmail.com> --- doc/filters.texi | 26 ++++++++++++++++++- libavfilter/scale.c | 72 ++++++++++++++++++++++++++++++++++++++++++++++++++--- 2 files changed, 93 insertions(+), 5 deletions(-)