Improve obj and grad! #375
Conversation
amontoison
force-pushed
the
obj_grad_optimized
branch
from
8f530f5 to
3805079
Compare
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #375 +/- ##
===========================================
- Coverage 89.11% 77.49% -11.63%
===========================================
Files 5 8 +3
Lines 790 1524 +734
===========================================
+ Hits 704 1181 +477
- Misses 86 343 +257 ☔ View full report in Codecov by Sentry. |
|
Memory leak in Please submit a bug report with steps to reproduce this fault, and any error messages that follow (in their entirety). Thanks.
Exception: EXCEPTION_ACCESS_VIOLATION at 0x6c301825 -- .text at D:\a\CUTEst.jl\CUTEst.jl\deps\files\libHS6_double.DLL (unknown line)
in expression starting at D:\a\CUTEst.jl\CUTEst.jl\test\nlpmodelstest.jl:1
.text at D:\a\CUTEst.jl\CUTEst.jl\deps\files\libHS6_double.DLL (unknown line)
.text at D:\a\CUTEst.jl\CUTEst.jl\deps\files\libHS6_double.DLL (unknown line)
cutest_cigr_ at D:\a\CUTEst.jl\CUTEst.jl\deps\files\libHS6_double.DLL (unknown line)
cutest_cigr_ at D:\a\CUTEst.jl\CUTEst.jl\src\libcutest.jl:411
cigr at D:\a\CUTEst.jl\CUTEst.jl\src\core_interface.jl:3312 [inlined]
grad! at D:\a\CUTEst.jl\CUTEst.jl\src\julia_interface.jl:32 |
|
|
||
| function NLPModels.obj(nlp::CUTEstModel{T}, x::AbstractVector) where T | ||
| @lencheck nlp.meta.nvar x | ||
| x_ = Vector{T}(x) |
There was a problem hiding this comment.
Maybe this could be preallocated in the problem structure?
There was a problem hiding this comment.
Yes, we could but it's the same issue with all routines.
I'm wondering if we should not remove the dispatch for input vectors of wrong type.
If the CUTEstModel is in double precision, the user should provide input arrays in double precision.
What do you think?
Otherwise, we need to duplicate storage in the CUTEstModel.
There was a problem hiding this comment.
Most of Julia dispatches in that way. It’s useful because we routinely call problem functions with views.
There was a problem hiding this comment.
Views are handled by the other method with x::StrideOneVector{T}.
This kind of method is only used if x is a Vector{Float32} or a CuVector{Float64} and nlp is a CUTEstModel{Float64}.
But should we handle that?
| function NLPModels.grad!(nlp::CUTEstModel{T}, x::AbstractVector, g::AbstractVector) where T | ||
| @lencheck nlp.meta.nvar x g | ||
| x_ = Vector{T}(x) | ||
| g_ = Vector{T}(g) |
amontoison
force-pushed
the
obj_grad_optimized
branch
from
3805079 to
54c2cae
Compare
amontoison
merged commit 217a575
into
JuliaSmoothOptimizers:main
No description provided.