mklarqvist / positional-popcount Goto Github PK

If this project wants to collect all possible methods, here is another one:
https://gist.github.com/aqrit/c729815b0165c139d0bac642ab7ee104

Expected to be 37.5% slower than sse2_sad.

I'm thinking about doing a rewrite, here are some random thoughts:

1. pospopcnt_csa_sse() implies two unnecessary temp variables. It could be rewritten as:
   b ^= a;
   a ^= c;
   c ^= b; // xor sum
   b |= a;
   b ^= c; // carry

2. The throughput of a CSA is better than its latency. So it is probably best to keep two independent CSAs executing in parallel. Specifically, passing v1 as a [in/out] parameter
   to every bottom tier CSA may create a serial dependency.

3. At least two passes through the CSA Unit are required for the CSA Unit to pay for itself.
   So we may be able to better space loads inside the loop body...
   If we process half a loop, before before initially dropping into the loop body, then we can spend the back half of the loop pre-loading for the next pass.

4. Assuming two loads per cycle on AVX2 machines… It may be better to trade out instructions for more i/o? Maybe spill the output of the CSA Unit into a buffer, then we could go back and do another --or several more-- passes of 16:1 reductions before accumulating.

For reference, so we can appreciate the speeds.

@WojciechMula

The latest commit broke build on MinGW, but it seems to be some compiler-related problem: "Error: invalid register for .seh_savexmm".