The straight 256-bit AVX solution (no pext) goes a bit faster out of memory: AVX... 0.013 (speed up: 1.80) And even better out of L1: nate

The straight 256-bit AVX solution (no pext) goes a bit faster out of memory: AVX... 0.013 (speed up: 1.80) And even better out of L1: nate@skylake:~/git/WojciechMula-toys/base64/decode/sse$ ./speed input size: 8192 improved scalar... 2.481 scalar... 4.839 (speed up: 0.51) SSE... 1.918 (speed up: 1.29) SSE & BMI2... 1.668 (speed up: 1.49) AVX... 1.127 (speed up: 2.20) AVX & BMI2... 1.290 (speed up: 1.92) (Absolute times aren't comparable between Mem and L1, just the ratios within each)