I just updated the Rand-related dependencies and ran the benchmark:

Test	type	Time (`rand_distr`)	Time (`etf-rs`)
central_normal	`f32`	5.0159 ns	3.8457 ns
central_normal	`f64`	3.9653 ns	3.9148 ns
normal	`f64`	6.2197 ns	4.2361 ns
cauchy	`f32`	26.118 ns	4.0772 ns
cauchy	`f64`	21.983 ns	3.8984 ns
chi_squared (k=0.5)	`f32`	29.742 ns	7.8440 ns
chi_squared (k=0.5)	`f64`	41.105 ns	10.746 ns
chi_squared (k=2)	`f32`	7.7833 ns	4.2518 ns
chi_squared (k=2)	`f64`	6.3638 ns	3.7355 ns
chi_squared (k=2)	`f32`	7.7833 ns	4.2518 ns
chi_squared (k=2)	`f64`	6.3638 ns	3.7355 ns
chi_squared (k=5)	`f32`	16.573 ns	4.2282 ns
chi_squared (k=5)	`f64`	15.904 ns	3.9329 ns
chi_squared (k=1000)	`f32`	16.355 ns	4.0622 ns
chi_squared (k=1000)	`f64`	15.495 ns	3.9862 ns

Cauchy seems to be the obvious candidate for an ETF implementation since it is about 5-6 times faster than today's rand_distr::Cauchy and it can be generated from a single pre-computed table, just like rand_distr::Normal is today (it will be in such case slightly slower than in the above benchmark, but the relative difference should be the same as rand_distr::Normal vs rand_distr::StandardNormal).

Chi-squared is much faster than rand_distr::ChiSquared, but AFAIR it cannot be generated from a single pre-computed table.

Uh oh!

Investigate replacing Ziggurat sampling with ETF #9

Description

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions