Add Varied analysis to the reverse mode #1084
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## master #1084 +/- ##
==========================================
+ Coverage 94.25% 94.34% +0.08%
==========================================
Files 55 50 -5
Lines 8445 8326 -119
==========================================
- Hits 7960 7855 -105
+ Misses 485 471 -14
... and 37 files with indirect coverage changes
|
ovdiiuv
force-pushed
the
activity
branch
2 times, most recently
from
65075a3 to
df82648
Compare
ovdiiuv
force-pushed
the
activity
branch
2 times, most recently
from
f5236aa to
c2ca79b
Compare
There was a problem hiding this comment.
Here is a partial review. Please address the clang-tidy complaints as most of them are good. I'd also rely on @PetroZarytskyi here for a review.
There was a problem hiding this comment.
Thank you for your work. I didn't mention pullbacks and call exprs because we already made huge progress there yesterday. Overall, the PR looks awesome. Its simplicity makes me feel that TBR was overcomplicated. I think we made the right decision to start with a simplified implementation.
|
|
||
| bool VariedAnalyzer::VisitDeclStmt(DeclStmt* DS) { | ||
| for (Decl* D : DS->decls()) { | ||
| if (!isa<VarDecl>(D)) |
There was a problem hiding this comment.
A test here could be class C{} c; that produces a DeclStmt whose Stmt is not a VarDecl.
test/Gradient/FunctionCalls.C
Outdated
| @@ -441,14 +441,14 @@ double do_nothing(double* u, double* v, double* w) { | |||
| return u[0]; | |||
| } | |||
|
|
|||
| // CHECK: void do_nothing_pullback(double *u, double *v, double *w, double _d_y, double *_d_u, double *_d_v, double *_d_w); | |||
There was a problem hiding this comment.
We used to remove non varied parameters from the adjoint, but it'd crash cms example, so I removed it for now.
There was a problem hiding this comment.
I am not sure I follow, Does not that change the pullback signatures randomly? @PetroZarytskyi.
| bool m_Varied = false; | ||
| bool m_Marking = false; | ||
|
|
||
| std::set<const clang::VarDecl*>& m_VariedDecls; |
There was a problem hiding this comment.
| std::set<const clang::VarDecl*>& m_VariedDecls; | |
| VarsData& m_VariedDecls; |
|
LGTM |
test/Gradient/FunctionCalls.C
Outdated
| @@ -441,14 +441,14 @@ double do_nothing(double* u, double* v, double* w) { | |||
| return u[0]; | |||
| } | |||
|
|
|||
| // CHECK: void do_nothing_pullback(double *u, double *v, double *w, double _d_y, double *_d_u, double *_d_v, double *_d_w); | |||
There was a problem hiding this comment.
I am not sure I follow, Does not that change the pullback signatures randomly? @PetroZarytskyi.
|
|
||
| bool VariedAnalyzer::VisitDeclStmt(DeclStmt* DS) { | ||
| for (Decl* D : DS->decls()) { | ||
| if (Expr* init = cast<VarDecl>(D)->getInit()) { |
There was a problem hiding this comment.
Here we will crash if D is not a VarDecl.
ovdiiuv
force-pushed
the
activity
branch
2 times, most recently
from
47ec589 to
03255f3
Compare
![github-actions[bot]](https://avatars.githubusercontent.com/in/15368?s=60&v=4){kind=small}
| enable_va_in_req = | ||
| clad::HasOption(bitmasked_opts_value, clad::opts::enable_va); | ||
| disable_va_in_req = | ||
| clad::HasOption(bitmasked_opts_value, clad::opts::disable_va); |
There was a problem hiding this comment.
warning: no member named 'disable_va' in 'clad::opts'; did you mean 'disable_aa'? [clang-diagnostic-error]
| clad::HasOption(bitmasked_opts_value, clad::opts::disable_va); | |
| clad::HasOption(bitmasked_opts_value, clad::opts::disable_aa); |
Additional context
include/clad/Differentiator/CladConfig.h:32: 'disable_aa' declared here
disable_aa = 1 << (ORDER_BITS + 6),
^
This PR adds Varied analysis support in the reverse mode, fixes wrong jacobian results and updates tests.