lemire
commented
5 years ago For correct analysis, I recommend running the instrumented benchmark with the -v flag for verbose outputs, you something like this...
pospopcnt_u16_avx512_csa instructions per cycle 1.93, cycles per 16-bit word: 0.154, instructions per 16-bit word 0.298
min: 7970 cycles, 15417 instructions, 1 branch mis., 540 cache ref., 0 cache mis.
avg: 8155.0 cycles, 15417.0 instructions, 1.2 branch mis., 644.1 cache ref., 1.6 cache mis.See that the average and min. cycles are different by about 2%? This gives you an estimate of your incertitude of that measure.
You can also run the test multiple times...
$ ./instrumented_benchmark | grep pospopcnt_u16_avx512_csa
pospopcnt_u16_avx512_csa cycles per 16-bit word: 0.152
$ ./instrumented_benchmark | grep pospopcnt_u16_avx512_csa
pospopcnt_u16_avx512_csa cycles per 16-bit word: 0.151
$ ./instrumented_benchmark | grep pospopcnt_u16_avx512_csa
pospopcnt_u16_avx512_csa cycles per 16-bit word: 0.152
$ ./instrumented_benchmark | grep pospopcnt_u16_avx512_csa
pospopcnt_u16_avx512_csa cycles per 16-bit word: 0.152
$ ./instrumented_benchmark | grep pospopcnt_u16_avx512_csa
pospopcnt_u16_avx512_csa cycles per 16-bit word: 0.155
$ ./instrumented_benchmark | grep pospopcnt_u16_avx512_csa
pospopcnt_u16_avx512_csa cycles per 16-bit word: 0.153
$ ./instrumented_benchmark | grep pospopcnt_u16_avx512_csa
pospopcnt_u16_avx512_csa cycles per 16-bit word: 0.156
$ ./instrumented_benchmark | grep pospopcnt_u16_avx512_csa
pospopcnt_u16_avx512_csa cycles per 16-bit word: 0.151So the values range between 0.151 and 0.156... so about a 3% incertitude.
You definitively do not want to get something that goes from 0.134 to 0.151 cycles per word for the same algorithm. Let me know if this happens and it is one of my test machines. Sometimes they get misconfigured.
mklarqvist
WojciechMula
This is slightly faster[1] code, but requires the AVX512VBMI extension. Can we assume this extension? I also checked four-fold sum, but couldn't measure any further improvement.
[1] Our tool
instrumented_benchmarkshows significant variations of measurements. On CNL I got 0.134 to 0.151 cycles per word, thus I'm not sure if it's faster. Current baseline is 0.148 cycles.