| Message ID | 1556015482-19503-1-git-send-email-yejun.guo@intel.com |
|---|---|
| State | Superseded |
| Headers | show |
> print_in_columns() { > - cols=$(expr $ncols / 24) > - cat | tr ' ' '\n' | sort | pr -r "-$cols" -w $ncols -t > + col_width=24 > + cols=$(expr $ncols / $col_width) > + rows=$(expr $(expr $# + $cols - 1) / $cols) > + for row in $(seq $rows); do > + index=$row > + line="" > + fmt="" > + for col in $(seq $cols); do > + if [ $index -le $# ]; then > + eval line='"$line "${'$index'}' > + fmt="$fmt%-${col_width}s" > + fi > + index=$(expr $index + $rows) > + done > + printf "$fmt\n" $line > + done | sed 's/ *$//' > } The new code is relatively slow. On linux it adds ~ 1.5s (1500 ms) to the configure time in both bash and dash on my system, which is roughly additional ~20% to configure run time. On Windows this will easily add a minute or more (I didn't test). Few things to consider: - print_in_column iterates over a _lot_ of values - hundreds or more. - Subshells - `$(cmd ...)` are relatively very expensive, especially in hot (inner) loops, and especially on Windows. - $(expr ...) can typically be replaced with shell arithmetics - $((...)) . Your part 2 already uses it, and regardless it's been used in configure before (in pushvar and popvar), so you should use it where possible. - `for col in $(seq $cols)` need not invoke `seq` on each iterations to always produce the same output. You can capture its output once on startup. - All the places which use the new print_in_columns want the result sorted and duplicate pipe through `tr` and `sort`. The original version already did `cat | tr ' ' '\n' | sort` in one place, and you can simply replace it with `set -- $(cat | tr ' ' '\n' | sort)` to get the values as positional parameters. This will also keep the interface the same as before and will reduce the patch size. And finally, is there a good reason to sort the results in columns and not rows? As you rightfully mentioned, outputs can span several pages, and when reading it on screen it might be more convenient to read row by row than column by column. This could also simplify the new code a lot.
> -----Original Message----- > From: ffmpeg-devel [mailto:ffmpeg-devel-bounces@ffmpeg.org] On Behalf Of > avih > Sent: Tuesday, April 23, 2019 8:14 PM > To: FFmpeg development discussions and patches <ffmpeg-devel@ffmpeg.org> > Cc: Guo, Yejun <yejun.guo@intel.com> > Subject: Re: [FFmpeg-devel] [PATCH V3 1/2] configure: sort > decoder/encoder/filter/... names in alphabet order > > > print_in_columns() { > > - cols=$(expr $ncols / 24) > > - cat | tr ' ' '\n' | sort | pr -r "-$cols" -w $ncols -t > > + col_width=24 > > + cols=$(expr $ncols / $col_width) > > + rows=$(expr $(expr $# + $cols - 1) / $cols) > > + for row in $(seq $rows); do > > + index=$row > > + line="" > > + fmt="" > > + for col in $(seq $cols); do > > + if [ $index -le $# ]; then > > + eval line='"$line "${'$index'}' > > + fmt="$fmt%-${col_width}s" > > + fi > > + index=$(expr $index + $rows) > > + done > > + printf "$fmt\n" $line > > + done | sed 's/ *$//' > > } > > > The new code is relatively slow. > > On linux it adds ~ 1.5s (1500 ms) to the configure time in both bash and dash > on my system, which is roughly additional ~20% to configure run time. > On Windows this will easily add a minute or more (I didn't test). I tried with mingw on windows, the whole configure 'date;./configure;date' takes about 2:25 (minute:second) with master, it takes about 3:41 with this v3 patch, and takes about 2:15 with v4 patch with your suggestions, v4 will be sent out next. I also tried on ubuntu 16.04, the consumed times are: 8.5 seconds (master), 12 seconds (v3 patch), and 7.5 seconds (v4 patch). > > > Few things to consider: > > - print_in_column iterates over a _lot_ of values - hundreds or more. > > - Subshells - `$(cmd ...)` are relatively very expensive, especially in hot > (inner) loops, and especially on Windows. > > - $(expr ...) can typically be replaced with shell arithmetics - $((...)) . > Your part 2 already uses it, and regardless it's been used in configure > before (in pushvar and popvar), so you should use it where possible. I just thought they are similar and so choose it by random. Find the performance is better after change to $((...)). > > - `for col in $(seq $cols)` need not invoke `seq` on each iterations to always > produce the same output. You can capture its output once on startup. thanks, and it helps the performance too. > > - All the places which use the new print_in_columns want the result sorted and > duplicate pipe through `tr` and `sort`. The original version already did > `cat | tr ' ' '\n' | sort` in one place, and you can simply replace it with > `set -- $(cat | tr ' ' '\n' | sort)` to get the values as positional > parameters. This will also keep the interface the same as before and will > reduce the patch size. thanks, good point. > > And finally, is there a good reason to sort the results in columns and not rows? > As you rightfully mentioned, outputs can span several pages, and when reading > it on screen it might be more convenient to read row by row than column by > column. This could also simplify the new code a lot. To check a name with human eye, after we get the nearby area for the check: If row by row, we have to turn our eyes from left to right, from up to down, and then again left to right, up to down, maybe again and again. And there is still possibility that we missed the check. If column by column, we just turn from up to down (the same column) and can easily check if the name is there or not. > > _______________________________________________ > ffmpeg-devel mailing list > ffmpeg-devel@ffmpeg.org > https://ffmpeg.org/mailman/listinfo/ffmpeg-devel > > To unsubscribe, visit link above, or email > ffmpeg-devel-request@ffmpeg.org with subject "unsubscribe".
diff --git a/configure b/configure index 3b11ffe..f8032aa 100755 --- a/configure +++ b/configure @@ -3832,14 +3832,28 @@ die_unknown(){ } print_in_columns() { - cols=$(expr $ncols / 24) - cat | tr ' ' '\n' | sort | pr -r "-$cols" -w $ncols -t + col_width=24 + cols=$(expr $ncols / $col_width) + rows=$(expr $(expr $# + $cols - 1) / $cols) + for row in $(seq $rows); do + index=$row + line="" + fmt="" + for col in $(seq $cols); do + if [ $index -le $# ]; then + eval line='"$line "${'$index'}' + fmt="$fmt%-${col_width}s" + fi + index=$(expr $index + $rows) + done + printf "$fmt\n" $line + done | sed 's/ *$//' } show_list() { suffix=_$1 shift - echo $* | sed s/$suffix//g | print_in_columns + print_in_columns $(echo $* | sed s/$suffix//g | tr ' ' '\n' | sort) exit 0 } @@ -7121,32 +7135,32 @@ test -n "$random_seed" && echo echo "External libraries:" -print_enabled '' $EXTERNAL_LIBRARY_LIST $EXTERNAL_AUTODETECT_LIBRARY_LIST | print_in_columns +print_in_columns $(print_enabled '' $EXTERNAL_LIBRARY_LIST $EXTERNAL_AUTODETECT_LIBRARY_LIST | tr ' ' '\n' | sort) echo echo "External libraries providing hardware acceleration:" -print_enabled '' $HWACCEL_LIBRARY_LIST $HWACCEL_AUTODETECT_LIBRARY_LIST | print_in_columns +print_in_columns $(print_enabled '' $HWACCEL_LIBRARY_LIST $HWACCEL_AUTODETECT_LIBRARY_LIST | tr ' ' '\n' | sort) echo echo "Libraries:" -print_enabled '' $LIBRARY_LIST | print_in_columns +print_in_columns $(print_enabled '' $LIBRARY_LIST | tr ' ' '\n' | sort) echo echo "Programs:" -print_enabled '' $PROGRAM_LIST | print_in_columns +print_in_columns $(print_enabled '' $PROGRAM_LIST | tr ' ' '\n' | sort) echo for type in decoder encoder hwaccel parser demuxer muxer protocol filter bsf indev outdev; do echo "Enabled ${type}s:" eval list=\$$(toupper $type)_LIST - print_enabled '_*' $list | print_in_columns + print_in_columns $(print_enabled '_*' $list | tr ' ' '\n' | sort) echo done if test -n "$ignore_tests"; then ignore_tests=$(echo $ignore_tests | tr ',' ' ') echo "Ignored FATE tests:" - echo $ignore_tests | print_in_columns + print_in_columns $(echo $ignore_tests | tr ' ' '\n' | sort) echo fi
take decoder names an example, with the default page length, shell command 'pr' needs two pages for all the decoder names. The names are firstly printed in the first page, then in the second page. So, as a whole, the names are sorted neither in column order nor in row order. It's a little confused. One method is to calculate the proper page length, so all the names are printed in one page by 'pr -l', and so strictly in alphabet order, column by column. Another method is to use command printf instead of pr, because buybox doesn't have pr. This patch refines print_in_columns to print the names with printf in alphabet order, very similar with 'pr -l', except the case when the last column is not fully filled with names. The interface of print_in_columns changed, the input needs to be sorted first, and then pass into print_in_columns as function parameters. It gives the flexibility the input parameters can be considered as an array and can be indexed. contributor: Alexander Strasser <eclipse7@gmx.net> Signed-off-by: Guo, Yejun <yejun.guo@intel.com> --- configure | 32 +++++++++++++++++++++++--------- 1 file changed, 23 insertions(+), 9 deletions(-)