Add fmodf128

[tgross35]

This function is significantly slower than all others so includes an override in EXTREMELY_SLOW_TESTS. Without it, PR CI takes ~1hour and the extensive tests in CI take ~1day.

[tgross35]

Based on the icount, it looks like fmodf128 should be about 28x slower than fmod, which seems reasonable given non-native integer sizes and a few paths with possible f128 multiplication. I might just need to turn down the test iteration here so it completes in a reasonable amount of time.

 icount::icount_bench_fmod_group::icount_bench_fmod logspace:setup_fmod()
  Baselines:                      softfloat|softfloat
  Instructions:                     1103106|1109057              (-0.53658%) [-1.00539x]
  L1 Hits:                          1105306|1111254              (-0.53525%) [-1.00538x]
  L2 Hits:                                0|1                    (-100.000%) [---inf---]
  RAM Hits:                               8|10                   (-20.0000%) [-1.25000x]
  Total read+write:                 1105314|1111265              (-0.53552%) [-1.00538x]
  Estimated Cycles:                 1105586|1111609              (-0.54183%) [-1.00545x]
icount::icount_bench_fmodf128_group::icount_bench_fmodf128 logspace:setup_fmodf128()
  Baselines:                      softfloat|softfloat
  Instructions:                    31327744|N/A                  (*********)
  L1 Hits:                         31361122|N/A                  (*********)
  L2 Hits:                                3|N/A                  (*********)
  RAM Hits:                              37|N/A                  (*********)
  Total read+write:                31361162|N/A                  (*********)
  Estimated Cycles:                31362432|N/A                  (*********)
icount::icount_bench_fmodf16_group::icount_bench_fmodf16 logspace:setup_fmodf16()
  Baselines:                      softfloat|softfloat
  Instructions:                       82388|N/A                  (*********)
  L1 Hits:                            97228|N/A                  (*********)
  L2 Hits:                                1|N/A                  (*********)
  RAM Hits:                              15|N/A                  (*********)
  Total read+write:                   97244|N/A                  (*********)
  Estimated Cycles:                   97758|N/A                  (*********)
icount::icount_bench_fmodf_group::icount_bench_fmodf logspace:setup_fmodf()
  Baselines:                      softfloat|softfloat
  Instructions:                      186830|188298               (-0.77962%) [-1.00786x]
  L1 Hits:                           189029|190498               (-0.77114%) [-1.00777x]
  L2 Hits:                                1|0                    (+++inf+++) [+++inf+++]
  RAM Hits:                               8|8                    (No change)
  Total read+write:                  189038|190506               (-0.77058%) [-1.00777x]
  Estimated Cycles:                  189314|190778               (-0.76738%) [-1.00773x]

[tgross35]

I am collecting a baseline for extensive test runtime at #471. fmod is doing slightly under 1M iterations/second. The extensive CI on this PR has fmodf128 at about 50k/s, which loosely correlates with the icount ratio.

Think I might just need to turn down the test iterations here unless there are some obvious perf wins.

Edit: the f64 baseline completed in about 30 minutes.

[tgross35]

Seems like this is more or less expected. I added a way to override the default number of iterations to make sure tests don't just run forever.

[tgross35]

With the override, extensive tests complete in 12 minutes and normal PR CI looks about the same. Seems reasonable 🎉