In deepTools:

alignmentSieve --samFlagInclude 2 -b input.bam -o output.bam

You can use multiple threads if you like. Of course this uses pysam, which uses htslib, which is part of samtools, but most tools will use something related to samtools somewhere...it's a standard tool after all.

using samjdk http://lindenb.github.io/jvarkit/SamJdk.html

java -jar  dist/samjdk.jar -e 'return record.getProperPairFlag();' input.bam

Thanks for the above responses, but we also found Sambamba which is pretty fast sambamba view -f bam -F "proper_pair" -o output.bam input.bam