CUDA Math Libraries- Possible Underflow Exceptions?

Accelerated Computing GPU-Accelerated Libraries
1

Hello!
We are currently using the CUDA Math Library to experiment with the numerical stability of its Math APIs. To verify correctness, we compare CUDA Math APIs with the corresponding C programming math functions. We have encountered some issues, particularly with underflow errors, where the C versions identify the underflow exception, but the CUDA versions output -inf/0.0 values.

We understand that CUDA lacks floating-point exception detection mechanisms. However, we want to verify if these are actually underflow errors. Below are the functions and the inputs we used to identify the errors:

  1. asin- input: 1.4693679385278594e-39; C result: Underflow exception; CUDA result: -inf
  2. erf - input: -1.0101903176080569e-38; C result: Underflow exception; CUDA result: -inf
  3. erfc- input: 450.4062194824219; C result: Underflow exception; CUDA result: 0.0
  4. exp- input: -1099461492736.0; C result: Underflow exception; CUDA result: 0.0
  5. exp10- input: -63.3827476501; C result: Underflow exception; CUDA result: 0.0
  6. exp2- input: -7316.86279; C result: Underflow exception; CUDA result: 0.0
  7. tgamma- input- -128.25; C result: Underflow exception; CUDA result: 0.0
2

I suggest providing a simple test case, like I have done here.

3

Hi! Robert,
I have provided a template for test case we used here:

4

I’ve provided a response there. CUDA device code doesn’t throw exceptions of any kind.

5

Thank you for the response. We want to confirm such situations we identified during our investigation.

6

Is this the same reason for occurence of NaN results that "CUDA device code doesn’t throw exceptions of any kind " in the Math APIs ? in NaN results

7

I’ve provided an answer there.