don't pick threshold based on register_size

[chriselrod]

I ran the benchmarks on a system with register_size() == 64, and got the following GFLOPS on the master branch for sizes 4:8:500

old_gflops = [0.16210526315789472, 1.6781567157355763, 4.161433018404238, 7.197283531409167, 9.553030303030303, 11.572753482232796, 14.429034619809643, 16.746275715299124, 19.578947368421055, 21.99035679845709, 24.80128720836685, 27.395831851483035, 29.192711404646776, 31.158719856208496, 33.36904068827221, 35.04422745548535, 36.832800851970184, 38.45884561636294, 40.34930579920806, 41.74674634794157, 42.88618689412984, 43.41364958680523, 46.127380024496155, 48.31694140867722, 49.149166776013644, 50.6332078728454, 52.40313864176383, 53.361424069431436, 54.484825818962534, 55.665026474153144, 56.34745964051333, 57.06077387621847, 57.672247676957994, 58.81788096735113, 59.81234582000914, 61.201434991044714, 61.60337858639627, 62.26612862173807, 61.90732041204154, 63.83436121917612, 64.29267682226191, 64.66860695953035, 64.70087883095505, 63.02217143268487, 63.342450710412066, 62.40403463747168, 62.800555826020585, 64.99590271988143, 62.627694636203955, 63.44711043553702, 65.32155270694881, 65.44405587845723, 67.48039597885897, 67.21316766343575, 64.40518187418853, 65.53358371634931, 66.40292069907099, 65.77859620654205, 68.44577118422902, 69.95890248896575, 67.95310444896346, 69.75019630350812, 71.11808005326851]

vs on this PR

new_gflops = [0.14776471843857145, 1.594246329038058, 4.056428001064679, 7.003855050115651, 9.445973663089681, 11.798402175760666, 14.065660685154976, 16.81253462055868, 19.46212787338274, 21.86716463106894, 24.676946410515672, 26.936951428471517, 28.929145361577795, 31.068750933439542, 33.41226763391412, 35.264421821978594, 36.813198509845655, 38.66570163487739, 40.477121882542235, 41.85745830605519, 42.8681555068836, 43.31440606859706, 46.13178282241772, 48.02345401835342, 49.25548520293771, 50.49430103220947, 52.26192142577662, 53.20286698053545, 54.56292220552537, 55.71549705859817, 56.36502191487289, 57.02239791081617, 57.628588073820275, 58.548880604989264, 59.739683915748884, 60.97500460799949, 61.46329113924051, 62.184114689244645, 61.86037659703268, 63.43681821206945, 63.936741274300914, 64.10524762182544, 64.36412311533029, 62.51022439067927, 63.10966290369639, 60.69208268872103, 62.12551472946074, 63.450891278513666, 62.35895975844578, 63.878358004244625, 65.83827413941691, 64.16653728850349, 65.1365044394016, 67.48519372013227, 65.46371685339177, 66.80207753266224, 66.97536258325441, 65.8370965883286, 68.24679282093642, 69.15932131070008, 65.85227749579352, 69.26174482564507, 70.6934274926346]

These are quite similar.

julia> sum(log, old_gflops / new_gflops)
0.4199631234593538

This is a <1% regression on average

julia> exp(ans/63)
1.0066883490969607

which could also be noise.