Broadcast benchmarks #30
Conversation
|
@jrevels, it seems like there is a |
|
It's not necessary for you to update the JLD file. The failure is due to JuliaCI/BenchmarkTools.jl#23. That PR shouldn't have affected BaseBenchmarks yet, since I haven't tagged it in a release, but the Travis script in this repo clones BenchmarkTools directly instead of using the latest stable release (a holdover from before BenchmarkTools was registered). I'll open a PR to fix the Travis script, and rerun the tests here. |
|
The Travis fix is merged, if you'd like to rebase and rerun the tests here. |
|
Okay, thanks. |
| R = Array(Float64, length(x),length(y)) | ||
| r = similar(z) | ||
|
|
||
| g["fusion", "Float64", size(r), 1] = @benchmarkable perf_bcast!($r, $z) |
There was a problem hiding this comment.
As it is, the fully qualified ID for this benchmark will be ["broadcast", "fusion", ("fusion", "Float64", size(R), 1)]. Can we remove the extra "fusion" (same goes for the other benchmarks where the group name is duplicated in the local key)?
|
Tests are green; any other comments? |
| perf_op_bcast!(r, x) = @dotcompat r .= 3 .* x .- 4 .* x.^2 .+ x .* x .- x .^ 3 | ||
| perf_op_bcast!(R, x, y) = @dotcompat R .= 3 .* x .- 4 .* y.^2 .+ x .* y .- x .^ 3 | ||
|
|
||
| g["dotop", "Float64", size(r), 1] = @benchmarkable perf_op_bcast!($r, $z) |
There was a problem hiding this comment.
"dotop" is redundant here for the reasons previously mentioned.
BTW, apologies for the slow review. I've been hiking on vacation, but just got back, so things should move a bit more quickly now.
There was a problem hiding this comment.
It needs something to distinguish it from the other benchmarks in the fusion group.
There was a problem hiding this comment.
Unless I'm mistaken, it's already in a different group (the "dotop" group as defined on line 50). The difference in the fully qualified IDs is thus the group name.
There was a problem hiding this comment.
whoops, right, I missed that addgroup line
* feed quasiquote variables in as arguments interpolating these into the AST gives the optimizer too much flexibility for a benchmark, that can mean the compiler will be able to just constant-fold away the work * handle case where core isn't and Expr not sure what it means to benchmark a constant value, but need to handle this case anyways
This adds a new suite (
broadcast) of benchmarks testing the new broadcast-fusion functionality, the upcoming dot-operator=broadcast change, and broadcast on sparse vectors, for JuliaLang/julia#16285.