It looks like our vectorizer doesn't know how to deal with complex numbers, unfortunately. Happy to take the promotion logic patch separately.

Discussed on slack. We will allow promotion between char and complex for now (this is not the same with numpy).

@Roger-luo This needs a rebase to trigger CI again. Also, I wonder if we can have some tests for this.

@ssnl thanks!

I'll need to fix the CI error I got with AT_DISPATCH_ALL_TYPES_AND_COMPLEX first... It works fine on my Mac but got error on Ubuntu :-(

Then I'll be able to add some tests.

(For the record: I helped Roger repro the issue offline)

On the AT_DISPATCH_ALL_TYPES issue: This might actually be related to a bug in GCC, where the visibility of a lambda and its captured objects list is treated differently by GCC: https://gcc.gnu.org/bugzilla/show_bug.cgi?id=80947 . Adding a __attribute__ ((visibility("hidden"))) to the tensor_cpu declaration in TensorFactories.cpp:643 is a suggested workaround that actually does the trick for me. It has some pitfalls, however. More on Slack.

Okay... this is related to different gcc versions, MWE:

#include <iostream>
#include <complex>

template <class T>
struct Vec256 {
public:
 T values[32 / sizeof(T)] = {0};
  static constexpr int size = 32 / sizeof(T);
  Vec256() {}
  Vec256(T val) {
    for (int i = 0; i != size; i++) {
      values[i] = val;
    }
  }
};

int main() {
  Vec256<std::complex<double>> r(1);
  std::cout << r.values[1] << std::endl;
  return 0;
}

This works fine with clang and later version of g++, but won't built with g++ 4.8. specialize the class manually for complex will work.

It seems that something breaks cuda tests:

22:56:08 ERROR: test_mul_intertype_scalar (__main__.TestCuda)
22:56:08 ----------------------------------------------------------------------
22:56:08 Traceback (most recent call last):
22:56:08   File "/var/lib/jenkins/workspace/test/common.py", line 275, in wrapper
22:56:08     method(*args, **kwargs)
22:56:08   File "test_cuda.py", line 925, in test_mul_intertype_scalar
22:56:08     test_mul(torch.float16)
22:56:08   File "test_cuda.py", line 922, in test_mul
22:56:08     x *= y
22:56:08 RuntimeError: TensorIterator expected type torch.cuda.DoubleTensor but got torch.cuda.HalfTensor[]

Is it a type promotion problem?

Looks like a promotion problem to me!

By the way, @gchanan brought to my attention that we have a bunch of latent bugs in PyTorch, owing to the fact that in a number of places we do if (scalar.is_float()) { /* assume it's float */ } else { /* assume it's int */ }. We'll need to fix these. Maybe this is causing the error you see above.

@Roger-luo Do you need any help with this patch?

Hi @Roger-luo!

Thank you for your pull request and welcome to our community.

Action Required

In order to merge any pull request (code, docs, etc.), we require contributors to sign our Contributor License Agreement, and we don't seem to have one on file for you.

Process

In order for us to review and merge your suggested changes, please sign at https://code.facebook.com/cla. If you are contributing on behalf of someone else (eg your employer), the individual CLA may not be sufficient and your employer may need to sign the corporate CLA.

Once the CLA is signed, our tooling will perform checks and validations. Afterwards, the pull request will be tagged with CLA signed. The tagging process may take up to 1 hour after signing. Please give it that time before contacting us about it.

If you have received this in error or have any questions, please contact us at [email protected]. Thanks!

Ah I think I should just close this PR since there's complex number now!