Skip to content

13rac1/fastmath

Repository files navigation

FastMath for Go

"I can make things very fast if they don’t have to be correct." — Russ Cox

8 and 16 bit math functions for when speed matters more than precision. Potential use cases include LED displays, 2D/3D graphics, and games.

Functions

  • Sin8() / Sin16()
  • Cos8() / Cos16()
  • Random8() / Random16()
  • Random8Limit() / Random16Limit()
  • Random8Range() / Random16Range()
  • Random16SetSeed() / Random16GetSeed() / Random16AddEntropy()
  • Scale8() / Scale8Video()
  • NScale8x3() / NScale8x3Video() - Scale three 8 bit integers at the same time.
  • Scale16() / Scale16By8()
  • QAdd8() / QSub8() / QMul8() - Saturating non-overflowing math functions.
  • Abs8()
  • Sqrt16()

Note: Functionality already well handled by the Go runtime has not be re-implemented.

Approximation Error

Computer-based math functions have an error delta verses the pure mathematical results. The Golang Standard Library's math functions are precise up to 64 bit floats. The math functions provided by this library sacrifice additional precision for speed by working with small integers.

  • Sin8() - Max Error: 1.63%, Average Error: 0.78%
  • Sin16() - Max Error: 0.34%, Average Error: 0.19%

Benchmarks

Run on a Intel(R) Core(TM) i7-7600U CPU @ 2.80GHz.

BenchmarkStdLibFallbackSqrt-4   20000000                 50.7 ns/op
BenchmarkStdLibDefaultSqrt-4    2000000000               0.30 ns/op
BenchmarkSqrt16-4               200000000                9.12 ns/op
BenchmarkStdLibRandom8-4        50000000                 25.6 ns/op
BenchmarkRandom8-4              1000000000               2.12 ns/op
BenchmarkStdLibSin8-4           50000000                 20.8 ns/op
BenchmarkSin8-4                 300000000                4.07 ns/op
BenchmarkStdLibSin16-4          50000000                 20.0 ns/op
BenchmarkSin16-4                2000000000               0.94 ns/op

Random8(), Sin8() and Sin16() are significantly faster than using the equivalent Go Standard Library's Math package functions.

Sqrt16() is compared against both the default compiled math.Sqrt() and a copy of the fallback Standard Library sqrt() function. The default version is optimized by the Go compiler into a single instruction on the AMD64 architecture, so the fallback version is used for a fair comparison.

TODO

  • Should Sin8() be a lookup table? Why is it 4ns/op vs Sin16()@0.9ns/op?
  • Add ARM assembly implementations from upstream, benchmark difference.

License

Licensed MIT

© 2019 Brad Erickson

Based on FastLED MIT-licensed code:

© 2013 FastLED

Parts of test-only BSD code:

© 2009 The Go Authors