Both Genrich and MACS2 can call peaks on paired-end ATAC-seq data and remove duplicates. Multi-mapped reads are typically removed/disregarded during the alignment step (i.e. bowtie2). Removing mitochondrial reads can be accomplished with multiple methods of which I prefer using sed but I can see how having this wrapped into the peak caller is convenient. I have tried both methods with ATAC-seq data.
An important difference is that MACS2 will call peaks on the paired end fragments and Genrich will call peaks on the cut sites (5' end extensions, set with -d, of the paired end fragments). With MACS2 you can use some suggested options to recapitulate the Genrich approach and calling peaks with Genrich 'not' in ATAC mode may give you similar MACS2 results as above. Overall each method above on the same bam file yields different peaks (as one would expect)
Downstream we supplied the peaks files to DAStk and were more satisfied with the MACS2 results but a bit unnerved by the differences.
I just learnt of Genrich this week - we installed it (which went smoothly), ran it and got results that didn't seem off. So, generally, I would recommend it although you may not want to use it for the super-high impact paper that you are submitting next week because it does not seem to be published yet and reviewers may complain about that.