OSDev.org

I had a gtx750ti and spent some time on C99 extensions on parallel computing.
Setup CUDA8.0 at home and looked at some sample projects in VS15 as a baby step.
I looked at source for deviceQuery utility, built and ran which probes lot of GPU characteristics on the system if it is found.
Now digging deeper, tried to look into one of the api function, and yes the header file is available in the cuda installation directory. However actual source of function body in does not seem to be available anywhere (was looking for files with *.c extension). grep with cygwin revealed something in *.lib. According to google search, lib appears to be a several compiled obj put together for distribution.
This means actual cuda source is not distributed?

That's correct. These cards are also, generally, undocumented at the hardware level. I think nvidia stated they would provide documentation for the hardware, but I am unsure of that status of that.

An alternative to cuda is OpenCL. It's lesser used (especially in the nerual network/machine learning community) and I think the performance is generally worse too.

CUDA is for when you want to do efficient machine learning. OpenCL is for when you want to waste power and heat mining bitcoin or upscaling anime.

Seriously, though, CUDA's closed up pretty tightly. Nvidia knows what they have, so they can do that. You want similar performance out of OpenCL, you need to spend more money and basically double your power/heat budget on AMD cards, and good luck finding any while the third wave of cryptocurrency miners are busy buying them all up to launder triad money (no joke).

that is good to know. i have applied rigorously for several openings at Nvidia in hopes of getting job over next few months. Lol, if i can get in and hands-on with cuda low-level, that would be awesome.