postsd much slower than postmean

[william-denault]

Hello,

I was profiling some code, and I noticed that postsd was much slower than postmean (for normal mixture), which seems strange, given how these quantities are computed.
Please find a minimal example below:

`library(ashr)
library(microbenchmark)

Bhat <- rnorm(10000)
Shat <- runif(10000)
out <- ash(Bhat,Shat, mixcompdist="normal")

Bhat <- rnorm(10000)
Shat <- runif(10000)
m <- set_data(Bhat, Shat)

microbenchmark(
postmean = postmean( get_fitted_g(out),m),
postsd = postsd( get_fitted_g(out),m)
)
`

[pcarbo]

@william-denault Could you please share your benchmarking results?

[william-denault]

expr min lq mean median uq max neval
postmean 102.7708 113.6098 134.5917 117.8156 130.2786 294.1863 100
postsd 225.2152 238.8804 277.1865 252.2738 325.9447 420.9524 100

[pcarbo]

Yes, indeed it is about 2–3x slower. I don't see that as being a big problem, do you? Were there other settings where the difference in runtime was much larger?

[william-denault]

I was profiling fsusie, which uses these two functions a lot at every loop (2^S x number of covariates) so I just thought that could be place to gain speed (I will have a look at the counterpart of postsd and postmean in ebnm as Matthew suggested to me).

[pcarbo]

Okay, let us know what you find out.

[william-denault]

I have run some benchmark comparison and ebnm is actually 1.5 slower than ash when computing posterior quantities. I see the same pattern (posterior sd computation being slower than posterior mean). Furthermore it is seems that running ash with outputlevel= 0 is almost as fast as ebnm using prior_family_family="normal_scale_mixture". To be fair ebnm uses more component than ash.

[pcarbo]

That's helpful to know, thanks William!