IWR1443: Radar Cube Memory

Storage requirement depends on the implementation choices. If 2D output is not written back to radar cube, then the storage requirement will be 216 KB because you use the number of chirp loops, not number of doppler bins which is the next power of two of the chirp loops. If 2D is written back then it is 256 KB because you use the number of doppler bins. There are trade-offs in doing either way, not necessarily related to whether or not number of chirp loops is a power of 2, the trade-off are mostly MIPS (and memory if num chirp loops is not power of 2) and to some extent quality related. If 2D output is not written back to L3, then doppler FFT will have to be recomputed during angle processing although this computation (and the reading of the cube) is only required on the detected points (versus the entire cube write out of 2D) so it depends on max detected points allowed in an implementation. The trade-off analysis is further complicated by whether you write higher precision (e.g 24-bit or 32-bit) output of 2D or lower precision 16-bit output (same as 1D output to not expand the size) and also the fact that memory I/O and compute operate in parallel in the typical implementation so it depends on whether compute or I/O dominates. Such trade-offs have been discussed in Section 4 of for 16xx but they could be seen in a similar light for 14xx (a similar document for 14xx does not exist). For example, in the current out of box demo on 14xx, an option is provided to select lower (16-bit) or higher (24-bit) precision for 2D output, the former case does not require recompute of 2D FFT during but sacrifices quality whereas latter requires more MIPS during angle processing (latter method is the default), this is described in the oob demo's doxygen documentation in section "Data Path - Higher precision 2nd Dimension Processing"