Probably by reading and analyzing the CIGAR strings, finding all operations before "H" and "S" and calculating the clipping locations based on mapping locations and operation lengths before clipping.

For example, the CIGAR for one read is 120M31S, the mapping position for the read is 12345, then the clipping location might be in the area close to 12345 + 120 = 12465.

using Bioalcidae http://lindenb.github.io/jvarkit/BioAlcidaeJdk.html : stream the bam, for each read extract the unclipped/start|end and clipped start/end. If there is a clip, print the BED record(s).

$ java -jar dist/bioalcidaejdk.jar -e 'stream().filter(R->!R.getReadUnmappedFlag()).forEach(R->{int x1=R.getUnclippedStart(),x2=R.getStart();if(x1<x2) println(R.getContig()+"\t"+(x1-1)+"\t"+(x2)+"\t"+R.getReadName());x1=R.getEnd();x2=R.getUnclippedEnd();if(x1<x2) println(R.getContig()+"\t"+(x1-1)+"\t"+(x2)+"\t"+R.getReadName());});'  src/test/resources/S1.bam 

RF01    567 576 RF01_508_1107_4:0:0_0:0:0_af
RF01    1679    1682    RF01_1679_2192_3:0:0_3:0:0_a8
RF01    1886    1889    RF01_1821_2315_3:0:0_2:0:0_98
RF01    1989    1993    RF01_1445_1994_3:0:0_3:0:0_5e
RF02    407 410 RF02_98_411_1:0:0_3:0:0_2e
RF02    446 450 RF02_382_836_3:0:0_0:0:0_82
RF02    1811    1821    RF02_1811_2313_5:0:0_2:0:0_74
RF02    2220    2224    RF02_1737_2289_3:0:0_4:0:0_5e
RF03    739 741 RF03_327_808_1:0:0_3:0:0_1e
RF03    813 822 RF03_309_823_1:0:0_4:0:0_82
(...)

In case you're really only interested in the start position, this might do the trick

perl -nwe 'next if /^@/;@a=split;if($a[5]~=/^\d*H*(\d+)S/{$a[3]+$1; print "$a[0]\t$a[3]"}' <samtools view my.bam >out.txt