C6748 Benchmarking: Run-Time Support Libraries

Clayton. I provided some information to you on a different thread also. Let me try and answer some of your questions here:

1) Your understanding is correct. See here: C:\Program Files\Texas Instruments\C6000 Code Generation Tools 6.1.15\lib\rtssrc\MATH\divf.c

2) Yes, the fastRTS implementation is what you should consider for achieving good performance. I provided you with the performance comparison with the last post.
Single precision divide performance:
* RTS = 540 clocks
* fastRTS = 37 clocks
* fastRTS with inlining (or vector invocation) = 3 clocks

 3) You won't suffer any detectable precision loss. We have verified the accuracy to match upto 0.000001 percent. I mentioned in my previous post about the updated fastRTS library that we are close to releasing. The updated release comes with test bench that you can use to do such analysis yourself.

4) Per my other comments, fastRTS is definitely an actively supported product and we have completed significant feature update to it that we will post in Oct. If you need an early prod, please let us know and we will provide

5) We have avoided special checks in the code. BUT, we do provide the entire fastRTS SW (in the updated release) in optimized C. So such special checks if needed are extremely easy to add. However I do recommend that such checks be added during the development phase of the project and removed for production

6) As I said there are some constraints are added. But in majority cases, these are don't care. For ex, the divsp function has the below constraint:
Special Cases:
* If | y| < 1.1755e-38 = 2- 126, then the return value is NaN = Not-a-Number (exponent and mantissa are all ones) > +/- 3.402823e+38 = +/- 1 * 2+128 (largest single-precision floating-point number) with the sign of x.

Regards,
Gagan