It was a decent presentation. Popcount has a long history, going back to the very first programming text book. (So says Knuth.) Someone interested in it might also like the chapter in "Beautiful Code".
My own code does a lot of popcounts, of very wide (~1024 bit) values. A run might take several hours, so for me a 1% performance boost is still a couple of minutes of savings, and worth my development time to investigate.
On the other hand, I also use the POPCNT instruction if available, and fall back to the nibble LUT version using the shuffle instruction if available, before using the generic C versions. There are also some techniques that work for very large sets of bits that aren't covered in this talk, which focused only on 32 bit/word popcounts.
My own code does a lot of popcounts, of very wide (~1024 bit) values. A run might take several hours, so for me a 1% performance boost is still a couple of minutes of savings, and worth my development time to investigate.
On the other hand, I also use the POPCNT instruction if available, and fall back to the nibble LUT version using the shuffle instruction if available, before using the generic C versions. There are also some techniques that work for very large sets of bits that aren't covered in this talk, which focused only on 32 bit/word popcounts.