How to debug with numexpr?

[27rabbitlt]

Hello~

I'm curious about the implementation of Numexpr and really attracted by the 20x speed up (tested on my machine) compared with Numpy.

I used to dig deep in Numpy's C code to find out how np.sum(a, b) is implemented using gdb but it doesn't work here.

I guess there are two reasons:

1. I don't know how to build and use Numexpr in-place
2. I don't know how to add debugging options when compiling Numexpr (the compiling process or compiling flags seem not to be accessible, we could only run python setup.py build)

I believe there must be a method for you guys, developpers and maintainers, to debug with this complicated project. Are there any suggestions or instructions?

[robbmcleod]

We're using the new pyproject.toml based build system with some very minor modifications. To install in-place, just use Pip:

pip install -e .

I don't use gdb, NumExpr is very heavily multi-threaded so I don't see how much use it would be. I use valgrind for locating C++ bugs which is a very different approach.

Another thing that is advisable, because NumExpr 2.8 is full of deeply nested macros, is to use the -E flag with gcc on build once to generate all the expanded source files. This would be added into extra_cflags in setup.py on building the extension.

[27rabbitlt]

@robbmcleod So much thanks!
I will use pip install -e ., I didn't know this before.

I use gdb because I used it for Numpy, which is completely not multi-threaded. Now I will try to use valgrind instead.

[27rabbitlt]

So much thanks again, though I have another new question, I think I should close this one and open another since they are irrelevant.