See proposal: remove complex numbers.
This repository:
- Implements (quite naively) most
complex64/complex128operations with user-defined typesComplex64andComplex128. - Measures performance of
complex64VSComplex64andcomplex128VSComplex128.
Tests are included.
# old is builtin complex64
# new is user-defined Complex64
Logical-4 2.89µs ± 1% 2.70µs ± 3% -6.47% (p=0.000 n=9+10)
Add64-4 1.50µs ± 0% 1.60µs ± 6% +6.74% (p=0.000 n=8+10)
Sub64-4 1.51µs ± 2% 1.70µs ± 0% +12.48% (p=0.000 n=10+8)
Mul64-4 6.56µs ± 2% 6.91µs ± 6% +5.45% (p=0.002 n=10+10)
Div64-4 53.2µs ± 3% 44.6µs ± 0% -16.06% (p=0.000 n=10+9)
go test -bench=.
BenchmarkLogicalBuiltin-4 500000 3087 ns/op
BenchmarkLogical-4 500000 2698 ns/op
BenchmarkAdd64Builtin-4 1000000 1496 ns/op
BenchmarkAdd64-4 1000000 1501 ns/op
BenchmarkSub64Builtin-4 1000000 1495 ns/op
BenchmarkSub64-4 1000000 1511 ns/op
BenchmarkMul64Builtin-4 200000 6466 ns/op
BenchmarkMul64-4 200000 6451 ns/op
BenchmarkDiv64Builtin-4 30000 52530 ns/op
BenchmarkDiv64-4 30000 44477 ns/op
Results stored in "benchmark-results/". Each result has "benchmark-results/$arch/$machine_id/" path and contains:
lscpu.txt, the result oflscpuinvocation result;builtin.out& `lib.out - results of the benchmark runs;uname.txt, optionaluname -ainvocation result;benchstat.txtresult ofbenchstat builtin.out lib.out;
Benchstat is required to aggregate benchmark statistics.
See run-benchmarks bash script for detailed instruction.
Most code looks the same, but some code compiled to different instruction sequences.
Look inside disasm.go to inspect objdump output.