[BUG+FIX] SIGSEGV using allclose with user-owned exec with profiling activated

[osayamenja]

Describe the Bug
Hey folks! Wonderful work with this library, I am a huge fan!

Currently, using allclose with an external executor causes a SIGSEGV error which crashes the running program.

Here's a summary:

• Error occurs when a user passes an exec object with profiling activated to allclose
• allclose receives exec by value, so its destructor is called at the end of the function.

template <typename OutType, typename InType1, typename InType2>
void __MATX_INLINE__ allclose(OutType dest, const InType1 &in1, const InType2 &in2, double rtol, double atol, cudaExecutor exec = 0)

• The destructor destroys cuda events.
• However, this destructor is called again when the exec object goes out of scope in the user's program.
• This triggers a SIGSEGV as the events have already been destroyed in the previous call to the destructor.
• FIX: Changing allclose to receive exec by reference like so: (...cudaExecutor& exec) fixes the issue because it ensures the destructor is only called once: when the referred exec object goes out of scope.

To Reproduce

#include <matx.h>
int main() {
    cudaSetDevice(0);
    cudaStream_t stream;
    cudaStreamCreate(&stream);
    matx::cudaExecutor exec{stream, true}; // necessary for profiling to be on.
    auto tA = matx::make_tensor<float>({2, 2});
    (tA = matx::ones<float>(tA.Shape())).run(exec);
    auto tB = matx::make_tensor<float>({2, 2});
    (tB = matx::ones<float>(tB.Shape())).run(exec);
    auto result = matx::make_tensor<int>({});
    constexpr auto rtol = 1e-3;
    constexpr auto atol = 1e-4;
    matx::allclose(result, tA, tB, rtol, atol, exec); // exec destructor called within allclose
    printf("Is Close? %s\n", result() ? "Yes" : "No");
    cudaStreamDestroy(stream);
} // SIGSEGV as internal state (events) has already been destroyed

Expected Behavior
No program crash.

Code Snippets
See above.

System Details (please complete the following information):

• OS: Ubuntu 24.04.3 LTS (GNU/Linux 6.14.0-1021-gcp x86_64)
• CUDA version: CUDA 13.0
• g++ version: 13.3

Additional Context
N/A

[cliffburdick]

Thanks for reporting this @osayamenja! I will test your fix and see if this problem occurs anywhere else.