[clad] Bump clad to v1.10.

[guitargeek]

This release introduces significant updates, including support for clang versions 10 through 20 and improved diagnostics for unsupported features. Forward mode now handles variadic functions, while reverse mode sees major enhancements like marking const methods as non-differentiable, enabling differentiation of global variables and static member functions, generating constructor pullbacks automatically, and reducing tape usage through optimizations. CUDA support has been expanded with better handling of device pullbacks and kernel variables. For details, please see the release notes.

[guitargeek]

As one can see from the CladDericatorTests, the gradient generation with Clad doesn't seem to work anymore. There is no function call happening then calling exec() on the gradient, and dump() also doesn't give you any code output.

Here an example macro:

#include <Math/CladDerivator.h>

double g(double *variables) { return variables[0] * variables[0] * variables[1]; }

void simpleclad() {

    // Call clad to generate the gradient of g.
    auto g_grad = clad::gradient(g, "variables");

    // Execute the generated gradient function.
    double variables[]{3., 4.};
    double grad_output[]{0., 0.};
    g_grad.execute(variables, grad_output);
    std::cout << "grad_output[0]: " << grad_output[0] << std::endl;
    std::cout << "grad_output[1]: " << grad_output[1] << std::endl;

    // Dump the generated gradient code to standard output.
    g_grad.dump();

}

The output with ROOT compiled with Clad master is empty:

root [0]
Processing simpleclad.C...
grad_output[0]: 0
grad_output[1]: 0
The code is:

root [1]

[guitargeek]

The regression originated from this commit, @vgvassilev:

vgvassilev/clad@612eeb8

[github-actions]

Test Results

    18 files      18 suites   3d 22h 56m 23s ⏱️
 2 728 tests  2 728 ✅ 0 💤 0 ❌
47 707 runs  47 707 ✅ 0 💤 0 ❌

Results for commit 9b1aa3b.

♻️ This comment has been updated with latest results.

[vgvassilev]

The regression originated from this commit, @vgvassilev:

vgvassilev/clad@612eeb8

Can you try this patch:

diff --git a/tools/ClangPlugin.cpp b/tools/ClangPlugin.cpp
index 8f5a2550..919ccb79 100644
--- a/tools/ClangPlugin.cpp
+++ b/tools/ClangPlugin.cpp
@@ -138,7 +138,7 @@ namespace clad {
         if (!isa<FunctionDecl>(D))
           continue;
         FunctionDecl* FD = cast<FunctionDecl>(D);
-        if (FD->isConstexpr()) {
+        if (FD->isConstexpr() || !m_Multiplexer) {
           DiffCollector collector(DGR, CladEnabledRange, m_DiffRequestGraph, S,
                                   opts);
           break;

[guitargeek]

Indeed, that fixes it! Thanks for your quick help

[vgvassilev]

Indeed, that fixes it! Thanks for your quick help

Ok, in that case I can land that on the other side and release.

[vgvassilev]

@guitargeek, do we need somehow a clean build or there is something else to be fixed?

[guitargeek]

The problem will be fixed by this PR:
#18478

[vgvassilev]

@dpiparo this seems that it can go forward for the 6.36 release.

[guitargeek]

Hi @vgvassilev yes I think we can do that. Can you approve this PR so it can me merged? Then we'll see about the 6.36 backport

[vgvassilev]

Lgtm.