Ensure checksize inlines #14609
Ensure checksize inlines #14609
Conversation
In JuliaLang#13612 I converted checksize from a generated function to a recursive lispy definition, but the methods are too complicated to be automatically inlined. Manually adding the inline annotation fixes this performance regression JuliaLang#14594. Master is now faster than 0.4.0 on most of the array perf tests.
|
On my computer, I still see a few perf regressions from 0.4.0: More than 2x perf regression: More than 1.2x perf regression: |
| @@ -247,18 +247,25 @@ end | |||
| end | |||
|
|
|||
| # checksize ensures the output array A is the correct size for the given indices | |||
| @noinline throw_checksize_error(arr, dim, idx) = throw(DimensionMismatch("index $d selects $(length(I[d])) elements, but size(A, $d) = $(size(A,d))")) | |||
| @noinline throw_checksize_error(A, dim, idx) = throw(DimensionMismatch("index $dim selects $(length(idx)) elements, but size(A, $dim) = $(size(A,dim))")) | |||
| @noinline throw_checksize_error(A, dim, idx::AbstractArray{Bool}) = throw(DimensionMismatch("index $dim selects $(sum(idx)) elements, but size(A, $dim) = $(size(A,dim))")) | |||
There was a problem hiding this comment.
isn't the size of the boolean array more important than how many elements are true?
There was a problem hiding this comment.
This is checking the size of the output array… it's essentially ensuring that similar returned an object of the correct size. It's of dubious value (looks like nobody hit the error since I had also introduced a bug in the error message), but it's cheap when done correctly. IIRC, it's what the builtin array types did before #10525, but my memory there is fuzzy.
There was a problem hiding this comment.
Oh my mistake, I thought it was checking the size of the input. I should read the comments. Too used to there not being any comments to read, or something.
|
Maybe you could run a benchmark against julia 0.4.2 as well? |
|
Regarding the regression in On master there is a significant time spent in I'm having a hard time digging through the generated functions coupled with the Cartesian macros so can't really pinpoint exactly where it happens.. |
|
Hmm, regarding my last comment, it could also be that we just have better inlining debug info for the profiler. |
|
The GC "regressions" that @nanosoldier detected here are erroneous, see JuliaCI/BenchmarkTrackers.jl#5. The raw time improvements are great! |
In #13612 I converted checksize from a generated function to a recursive lispy definition, but the methods are too complicated to be automatically inlined. Manually adding the inline annotation fixes this performance regression #14594. Master is now faster than 0.4.0 on most of the array perf tests.