Enabled tgamma test for bruteforce suite

[shajder]

Fixes #507 according to work plan from issue description.

[lakshmih]

Still reviewing/testing this. Thanks.

[lakshmih]

We would like to run this test since it is a new builtin being tested but could you rebase to tip first?

[shajder]

We would like to run this test since it is a new builtin being tested but could you rebase to tip first?

@lakshmih done

[AhmedAmraniAkdi]

Hello, we would like more time to review this if possible, thank you.

[lakshmih]

Need more time before this gets merged, still evaluating.

[paulfradgley]

Our implementation won't pass this test. We'll need some time to put some compiler engineers onto fixing it.

[lakshmih]

Can we update the reference algorithm? It produces wrong output for this result:

Input -> -1.92592994438723585306e-34 Reference value - Wolfram -> -5.19229685853482762853e+33 Reference value - CTS -> -1.03845937170696552571e+34 (Wrong)

CTS uses openlibm's tgammal. We suggest using openlibm's tgamma, which calculates the result correctly.

tgamma -> -5.19229685853482762853e+33 tgammal -> -1.03845937170696552571e+34 (Wrong)

openlibm also defaults to tgamma on systems where long double = double

[shajder]

Can we update the reference algorithm? It produces wrong output for this result:

Input -> -1.92592994438723585306e-34 Reference value - Wolfram -> -5.19229685853482762853e+33 Reference value - CTS -> -1.03845937170696552571e+34 (Wrong)

@lakshmih hmm, I just verified your case. Are you sure you are using latest commit for this PR ? I get a proper result (aligned with Wolfram output), am I missing something?

[lakshmih]

We’re now able to compute the correct reference values. However, we had a question regarding the required precision for the host reference:

Is 64-bit precision sufficient for this use case, or is 128-bit precision necessary? We’ve noticed that using 128-bit significantly increases execution time, so we’d like to confirm whether that level of precision is truly required.

[shajder]

We’re now able to compute the correct reference values. However, we had a question regarding the required precision for the host reference:

Is 64-bit precision sufficient for this use case, or is 128-bit precision necessary? We’ve noticed that using 128-bit significantly increases execution time, so we’d like to confirm whether that level of precision is truly required.

My perspective:

1. The 128-bit version of the reference function was significantly easier to port - 230 vs. 878 lines of code.

2. The assumption was that the 128-bit version offers higher precision, which seems desirable.

[shajder]

We’ve noticed that using 128-bit significantly increases execution time, so we’d like to confirm whether that level of precision is truly required.

@lakshmih as discussed in teleconference here is draft PR for comparison purpose