Improve performance of logpdf for DiagNormal and IsoNormal
#1991
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #1991 +/- ##
==========================================
+ Coverage 86.28% 86.30% +0.02%
==========================================
Files 146 146
Lines 8787 8801 +14
==========================================
+ Hits 7582 7596 +14
Misses 1205 1205 ☔ View full report in Codecov by Sentry. 🚀 New features to boost your workflow:
|
| @@ -261,6 +261,25 @@ logdetcov(d::MvNormal) = logdet(d.Σ) | |||
|
|
|||
| sqmahal(d::MvNormal, x::AbstractVector) = invquad(d.Σ, x .- d.μ) | |||
|
|
|||
| function sqmahal(d::DiagNormal, x::AbstractVector) | |||
| # Faster than above as this avoids calculating (x .- d.µ) | |||
| T = promote_type(partype(d), eltype(x)) | |||
There was a problem hiding this comment.
This is not correct eg if the parameters and x are arrays of integers. One could either compute the first element outside of the loop or possibly use a functional approach (e.g. with mapreduce and Base.Broadcast.broadcasted(...)).
There was a problem hiding this comment.
The functional approach probably also works better with GPU arrays.
| end | ||
|
|
||
| function sqmahal(d::IsoNormal, x::AbstractVector) | ||
| T = promote_type(partype(d), eltype(x)) |
There was a problem hiding this comment.
Very artificial, but this is also not guaranteed to be correct e.g. for Boolean values.
| for i in eachindex(x) | ||
| @inbounds sum += abs2(x[i] - d.μ[i]) / d.Σ[i, i] |
There was a problem hiding this comment.
Indices might not be compatible.
| T = promote_type(partype(d), eltype(x)) | ||
| sum = zero(T) | ||
| for i in eachindex(x) | ||
| @inbounds sum += abs2(x[i] - d.μ[i]) |
There was a problem hiding this comment.
Indices might not be compatible
| for i in eachindex(x) | ||
| @inbounds sum += abs2(x[i] - d.μ[i]) | ||
| end | ||
| return sum / d.Σ[1, 1] |
There was a problem hiding this comment.
Might not use one-based indexing.
This PR (sort of) closes #1989. I didn't attempt to make
rand()faster because the performance drop there was smaller, but the changes here improve the performance oflogpdfonDiagNormalandIsoNormalby about 2–3x.The offending code that made the previous
logpdfslower was the calculation ofx .- d.µ:Distributions.jl/src/multivariate/mvnormal.jl
Line 262 in abb151c
The new methods avoid this and are thus both faster and avoid allocations.