Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Add support for verifying Enzyme Gradients with Clad Gradients #488

Open
wants to merge 1 commit into
base: master
Choose a base branch
from

Conversation

Nirhar
Copy link
Contributor

@Nirhar Nirhar commented Aug 25, 2022

This commit generates code that will verify the results of Enzyme Gradients with Clad Gradients.
For example, if previously the following code was generated for differentiating with enzyme for a function:

void f1_grad_enzyme(double arr[2], clad::array_ref<double> _d_arr) {
    double *d_arr = _d_arr.ptr();
    __enzyme_autodiff_f1(f1, arr, d_arr);
}

The above code will be appended with checks to verify the calculated gradients. Thus the newly generated code would be:

void f1_grad_enzyme(double arr[2], clad::array_ref<double> _d_arr) {
    double *d_arr = _d_arr.ptr();
    __enzyme_autodiff_f1(f1, arr, d_arr);
    double cladResult1[2];
    f1_grad(arr, cladResult1);
    EssentiallyEqualArrays(cladResult1, _d_arr.ptr(), 2UL);
}

EssentiallyEqualArrays and EssentiallyEqual are functions defined in Differentiator.h

Only functions with primitive type and ConstantArray type parameters can be verified in this manner.

To trigger this verification one must append the following flag to clang while compiling the function to be generated: -Xclang -plugin-arg-clad -Xclang -fcheck-enzyme-with-clad

@Nirhar Nirhar marked this pull request as draft August 25, 2022 14:08
@Nirhar
Copy link
Contributor Author

Nirhar commented Aug 25, 2022

This PR fails when I attempt to differentiate multiple functions with enzyme and verify it with clad. Then clad outputs the wrong gradients in some runs of the compiled code but not in others.
Here is a failing example:

double f1(double arr[2]) { return arr[0] * arr[1]; }
double f2(double x, double y, double z){
    return x * y * z;
}
int main(){
  auto f1_grad = clad::gradient<clad::opts::use_enzyme>(f1);
  double f1_v[2] = {3, 4};
  double f1_g[2] = {0};
  f1_grad.execute(f1_v, f1_g);
  printf("d_x = %.2f, d_y = %.2f\n", f1_g[0], f1_g[1]);

  auto f2_grad=clad::gradient<clad::opts::use_enzyme>(f2);
  double f2_res[3];
  double f2_x=3,f2_y=4,f2_z=5;
  f2_grad.execute(f2_x,f2_y,f2_z,&f2_res[0],&f2_res[1],&f2_res[2]);
}

If each function is individually differentiated and verified(in different files), this error does not occur. Only when they are jointly differentiated the error occurs.
The derivatives generated for the above functions are:

void f1_grad(double arr[2], clad::array_ref<double> _d_arr) {
    double _t0;
    double _t1;
    _t1 = arr[0];
    _t0 = arr[1];
    double f1_return = _t1 * _t0;
    goto _label0;
  _label0:
    {
        double _r0 = 1 * _t0;
        _d_arr[0] += _r0;
        double _r1 = _t1 * 1;
        _d_arr[1] += _r1;
    }
}
void f1_grad_enzyme(double arr[2], clad::array_ref<double> _d_arr) {
    double *d_arr = _d_arr.ptr();
    __enzyme_autodiff_f1(f1, arr, d_arr);
    //Verification part below
    double cladResult1[2];
    f1_grad(arr, cladResult1);
    EssentiallyEqualArrays(cladResult1, _d_arr.ptr(), 2UL);
}
void f2_grad(double x, double y, double z, clad::array_ref<double> _d_x, clad::array_ref<double> _d_y, clad::array_ref<double> _d_z) {
    double _t0;
    double _t1;
    double _t2;
    double _t3;
    _t2 = x;
    _t1 = y;
    _t3 = _t2 * _t1;
    _t0 = z;
    double f2_return = _t3 * _t0;
    goto _label0;
  _label0:
    {
        double _r0 = 1 * _t0;
        double _r1 = _r0 * _t1;
        * _d_x += _r1;
        double _r2 = _t2 * _r0;
        * _d_y += _r2;
        double _r3 = _t3 * 1;
        * _d_z += _r3;
    }
}
void f2_grad_enzyme(double x, double y, double z, clad::array_ref<double> _d_x, clad::array_ref<double> _d_y, clad::array_ref<double> _d_z) {
    clad::EnzymeGradient<3> grad = __enzyme_autodiff_f2(f2, x, y, z);
    * _d_x = grad.d_arr[0U];
    * _d_y = grad.d_arr[1U];
    * _d_z = grad.d_arr[2U];
    //Verification part below
    double cladResult1;
    double cladResult2;
    double cladResult3;
    f2_grad(x, y, z, &cladResult1, &cladResult2, &cladResult3);
    EssentiallyEqual(cladResult1, * _d_x);
    EssentiallyEqual(cladResult2, * _d_y);
    EssentiallyEqual(cladResult3, * _d_z);
}

@Nirhar
Copy link
Contributor Author

Nirhar commented Aug 25, 2022

cc @vgvassilev @parth-07 for review and some help

@Nirhar
Copy link
Contributor Author

Nirhar commented Aug 26, 2022

The above problem has been fixed. It was simply that I was not initializing declared variables in some places

@codecov
Copy link

codecov bot commented Aug 26, 2022

Codecov Report

Merging #488 (2d05fa0) into master (09e2ed0) will increase coverage by 0.10%.
The diff coverage is 100.00%.

❗ Current head 2d05fa0 differs from pull request most recent head a882316. Consider uploading reports for the commit a882316 to get more accurate results

Impacted file tree graph

@@            Coverage Diff             @@
##           master     #488      +/-   ##
==========================================
+ Coverage   92.56%   92.67%   +0.10%     
==========================================
  Files          37       37              
  Lines        5528     5610      +82     
==========================================
+ Hits         5117     5199      +82     
  Misses        411      411              
Impacted Files Coverage Δ
include/clad/Differentiator/DiffPlanner.h 100.00% <ø> (ø)
include/clad/Differentiator/ReverseModeVisitor.h 98.70% <ø> (ø)
include/clad/Differentiator/VisitorBase.h 100.00% <ø> (ø)
tools/DerivedFnInfo.h 100.00% <ø> (ø)
lib/Differentiator/ReverseModeVisitor.cpp 96.16% <100.00%> (+0.14%) ⬆️
lib/Differentiator/VisitorBase.cpp 97.67% <100.00%> (+0.05%) ⬆️
tools/ClangPlugin.cpp 90.32% <100.00%> (+0.10%) ⬆️
tools/ClangPlugin.h 74.57% <100.00%> (+0.89%) ⬆️
tools/DerivedFnInfo.cpp 100.00% <100.00%> (ø)
Impacted Files Coverage Δ
include/clad/Differentiator/DiffPlanner.h 100.00% <ø> (ø)
include/clad/Differentiator/ReverseModeVisitor.h 98.70% <ø> (ø)
include/clad/Differentiator/VisitorBase.h 100.00% <ø> (ø)
tools/DerivedFnInfo.h 100.00% <ø> (ø)
lib/Differentiator/ReverseModeVisitor.cpp 96.16% <100.00%> (+0.14%) ⬆️
lib/Differentiator/VisitorBase.cpp 97.67% <100.00%> (+0.05%) ⬆️
tools/ClangPlugin.cpp 90.32% <100.00%> (+0.10%) ⬆️
tools/ClangPlugin.h 74.57% <100.00%> (+0.89%) ⬆️
tools/DerivedFnInfo.cpp 100.00% <100.00%> (ø)

This commit generates code that will verify the results of Enzyme Gradients with Clad Gradients.
For example, if previously the following code was generated for differentiating with enzyme for a function:
```cpp
void f1_grad_enzyme(double arr[2], clad::array_ref<double> _d_arr) {
    double *d_arr = _d_arr.ptr();
    __enzyme_autodiff_f1(f1, arr, d_arr);
}

```

The above code will be appended with checks to verify the calculated gradients. Thus the newly generated code would be:
```cpp
void f1_grad_enzyme(double arr[2], clad::array_ref<double> _d_arr) {
    double *d_arr = _d_arr.ptr();
    __enzyme_autodiff_f1(f1, arr, d_arr);
    double cladResult1[2];
    f1_grad(arr, cladResult1);
    EssentiallyEqualArrays(cladResult1, _d_arr.ptr(), 2UL);
}
```

`EssentiallyEqualArrays` and `EssentiallyEqual` are functions defined in Differentiator.h

Only functions with primitive type and ConstantArray type parameters can be verified in this manner.

To trigger this verification one must append the following flag to clang while compiling the function to be generated: `-Xclang -plugin-arg-clad -Xclang -fcheck-enzyme-with-clad`
@Nirhar Nirhar marked this pull request as ready for review August 26, 2022 15:12
@@ -450,6 +450,22 @@ namespace clad {
code);
}

void EssentiallyEqual(double a, double b) {
// FIXME: We should select epsilon value in a more robust way.
Copy link
Owner

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

@grimmmyshini, can you take a look?

bool ans = std::fabs(a - b) <=
((std::fabs(a > b) ? std::fabs(b) : std::fabs(a)) * epsilon);

assert(ans && "Clad Gradient is not equal to Enzyme Gradient");
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Instead of an assert, can we give a non-fatal error here?

@@ -413,6 +417,16 @@ namespace clad {
else
DifferentiateWithEnzyme();

if (use_enzyme && checkEnzymeWithClad) {
DiffRequest newRequest = const_cast<DiffRequest&>(request);
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why do you need const_cast here?

auto paramType = paramsRef[i]->getOriginalType();
llvm::SmallVector<Expr*, 2> equalityCheckArguments;
equalityCheckArguments.push_back(BuildDeclRef(cladResultDecls[i]));
if (paramType->isFloatingType()) {
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What if paramType is neither floating type nor an array type?

} else {
resultVar = BuildVarDecl(paramType, finalVarName, nullptr, true);
}
addToCurrentBlock(BuildDeclStmt(resultVar), direction::forward);
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We can move this statement outside the if-else chain.

}
addToCurrentBlock(BuildDeclStmt(resultVar), direction::forward);
cladGradArgs.push_back(BuildOp(UO_AddrOf, BuildDeclRef(resultVar)));
cladResultDecls.push_back(resultVar);
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We can move this outside if-else chain as well.

BuildDeclRef(paramsRef[i + numParams]), "ptr", {}));
ConstantArrayType* t = dyn_cast<ConstantArrayType>(
const_cast<Type*>(paramType.getTypePtr()));
int sizeOfArray = (int)(t->getSize().roundToDouble(false));
Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Why do you need C-style cast here?

// REQUIRES: Enzyme

#include "clad/Differentiator/Differentiator.h"

Copy link
Collaborator

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Can you please add a test containing nested function calls?

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

Successfully merging this pull request may close these issues.

3 participants