Switch to the fast broadcast implementation #716
Merged
Conversation
This file contains bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
julia> using OrdinaryDiffEq, BenchmarkTools
julia> using OrdinaryDiffEq: perform_step!, initialize!
julia> prob10 = ODEProblem((du, u, p, t)->copyto!(du, u),ones(10),(0.0,10.0));
julia> prob100 = ODEProblem((du, u, p, t)->copyto!(du, u),ones(100),(0.0,10.0));
julia> prob1000 = ODEProblem((du, u, p, t)->copyto!(du, u),ones(1000),(0.0,10.0));
julia> integ10 = init(prob10, Tsit5()); integ100 = init(prob100, Tsit5()); integ1000 = init(prob1000, Tsit5());
julia> initialize!(integ10, integ10.cache); initialize!(integ100, integ100.cache); initialize!(integ1000, integ1000.cache);
julia> @btime perform_step!($integ10, $integ10.cache, false) # PR
251.051 ns (0 allocations: 0 bytes)
julia> @btime perform_step!($integ100, $integ100.cache, false) # PR
645.738 ns (0 allocations: 0 bytes)
julia> @btime perform_step!($integ1000, $integ1000.cache, false) # PR
5.743 μs (0 allocations: 0 bytes)
julia> @btime perform_step!($integ10, $integ10.cache, false) # Master
220.204 ns (0 allocations: 0 bytes)
julia> @btime perform_step!($integ100, $integ100.cache, false) # Master
620.471 ns (0 allocations: 0 bytes)
julia> @btime perform_step!($integ1000, $integ1000.cache, false) # Master
5.754 μs (0 allocations: 0 bytes)
|
Codecov Report
@@ Coverage Diff @@
## master #716 +/- ##
==========================================
- Coverage 72.31% 72.01% -0.31%
==========================================
Files 93 93
Lines 29182 28808 -374
==========================================
- Hits 21104 20747 -357
+ Misses 8078 8061 -17
Continue to review full report at Codecov.
|
|
With SciML/DiffEqBase.jl#204, I got using OrdinaryDiffEq, BenchmarkTools
prob = ODEProblem((du, u, p, t)->copyto!(du, u),ones(1000),(0.0,10.0))
integ = init(prob, Tsit5());
@btime OrdinaryDiffEq.perform_step!($integ, $(integ.cache)) # bc
@btime OrdinaryDiffEq.perform_step!($integ, $(integ.cache)) # loop
#=
julia> @btime OrdinaryDiffEq.perform_step!($integ, $(integ.cache)) # bc
4.122 μs (0 allocations: 0 bytes)
julia> @btime OrdinaryDiffEq.perform_step!($integ, $(integ.cache)) # loop
4.129 μs (0 allocations: 0 bytes)
=#
integ = init(prob, Vern9());
@btime OrdinaryDiffEq.perform_step!($integ, $(integ.cache)) # bc
@btime OrdinaryDiffEq.perform_step!($integ, $(integ.cache)) # loop
#=
julia> @btime OrdinaryDiffEq.perform_step!($integ, $(integ.cache)) # bc
11.411 μs (0 allocations: 0 bytes)
julia> @btime OrdinaryDiffEq.perform_step!($integ, $(integ.cache)) # loop
20.662 μs (0 allocations: 0 bytes)
=#. I think that it is safe to say that there is no regression :-) |
|
This PR needs SciML/DiffEqBase.jl#204 |
| @@ -50,7 +50,7 @@ function qradd!(Q::AbstractMatrix, R::AbstractMatrix, v::AbstractVector, k::Int) | |||
| @inbounds begin | |||
| d = norm(v) | |||
| R[k, k] = d | |||
| @. Q[:, k] = v / d | |||
| @.. @view(Q[:, k]) = v / d | |||
There was a problem hiding this comment.
oh lol. That might effect timings.
|
Nord fails. |
ChrisRackauckas
approved these changes
Apr 13, 2019
This was referenced Apr 13, 2019
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Macro
@..is the fast broadcast implementation which assumes that all arrays are non-extruded, so that LLVM can always do a runtime alias check and vectorize.