So you are probably working on radio telescopes, as you can do phase based corrections or telescope arrays?
(I’ve been doing lots of 3D realtime ultrasound transmit and receive beamforming on GPUs. If there is a commercial opportunity, we can work together.)
Yes, computation hardware usually beats memory accesses, especially on GPUs.
Your throughput rate most likely is limited by how fast you can read angles from global memory instead of by __sincosf performance.
Except if you reuse the angles and they are cached in L2 or L1.
(The 32x for shared memory was specifying 32 copies to avoid bank conflicts, but 32 coarse angles could be fine depending on how many rotations you would store.)
If you apply the phase rotations on several data points or combine it with filters or interpolation, the Tensor Cores could be a powerful solution for further processing in your case.