Performance

Every port is benchmarked on the same eight operations and the same 64 dates, defined once in the hub’s spec/BENCHMARK.md. Each port also measures every other Bikram Sambat library it could find in its language, on the same machine and the same dates.

In all seven languages, every port is faster than all 33 other libraries on both conversions. None of those libraries gives the same answer as the hub’s data on all 64 dates; each report lists where they differ.

Against other libraries

BS to AD (B1), time per call. “Next fastest” is the fastest other library in that language.

Language	NepDate	Next fastest	Its time	Measured
Rust benchmark report	2.9 ns	nepdate-cli	61 ns	4
C++ benchmark report	7.1 ns	nepdate-cli	33 ns	3
Go benchmark report	6.2 ns	go-nepali	173 ns	4
Swift benchmark report	21 ns	Nepali-Date-Picker-SPM	922 ns	2
Kotlin benchmark report	21 ns	adbs	403 ns	3
TypeScript benchmark report	31 ns	nepali-datetime	380 ns	9
Python benchmark report	724 ns	bikram-sambat	1.24 µs	8

“Measured” counts the other libraries timed in that language. Each language’s report, linked from its name, has AD to BS (B2) too, every library’s row, which dates each one disagrees on, and why.

Every operation

Time per call for B1 to B8. Lower is better. B8 builds a whole month grid with tithi and holidays, 29 to 32 days.

• B1 BS to AD
• B2 AD to BS
• B3 Add 1000 days
• B4 Add 13 months
• B5 Long Nepali date with weekday
• B6 Strict parse
• B7 Patro info for a day
• B8 Patro month grid

Language	B1	B2	B3	B4	B5	B6	B7	B8
Rust	2.9 ns	3.1 ns	3.2 ns	2.6 ns	16 ns	18 ns	3.5 ns	299 ns
C++	7.1 ns	8.9 ns	7.5 ns	5.6 ns	28 ns	12 ns	13 ns	517 ns
Go	6.2 ns	5.0 ns	4.7 ns	4.1 ns	31 ns	35 ns	12 ns	398 ns
Swift	21 ns	20 ns	14 ns	14 ns	391 ns	53 ns	40 ns	1.53 µs
Kotlin	21 ns	12 ns	15 ns	13 ns	192 ns	39 ns	4.5 ns	135 ns
TypeScript	31 ns	27 ns	51 ns	56 ns	175 ns	66 ns	5.6 ns	517 ns
Python	724 ns	822 ns	662 ns	598 ns	527 ns	1.32 µs	128 ns	21.6 µs
C# NepDate	13 ns	19 ns	18 ns	19 ns	153 ns	22 ns	48 ns	3.13 µs

C# NepDate 2.0.7 on .NET 10 is the reference (its report). The conversions, arithmetic and patro lookups allocate nothing in Rust, C++, Go and Swift.

Reading the numbers

• Compare within a row of the first table, not across languages. Each language was measured on its own runtime, and Swift on GitHub’s macOS runner (Apple M1), not the laptop the others used (Intel Core i5-12500H).
• Some runs were not on a quiet machine. Each report records the CPU load during its run; the stable release repeats them on an idle machine, pinned to one core, as the spec asks.
• Speed is not the reason to pick a date library; correct dates are. Every port is checked against the same 20,360 golden rows and 7,176 month boundaries first (Same answer everywhere), and the benchmarks run only after that.

The method, the 64 dates and the rules for timing are in spec/BENCHMARK.md.