PyTorch master@92a0f78

Our digamma single-precision floating point accuracy is bad near the poles. This is not caught by our tests because we only test on a few values:

input = torch.tensor([-1.99999994], dtype=torch.float32)
fp32 = input.digamma()
fp64 = input.double().digamma()
print(((fp32.double() - fp64) / fp64).item())  # 0.24012388522631392

This is a relative error of 24%

We also are returning real numbers at the poles (for both float32 and float64) when we should be returning inf.