Skip to main content
Signal Processing

You are not logged in. Your edit will be placed in a queue until it is peer reviewed.

We welcome edits that make the post easier to understand and more valuable for readers. Because community members review edits, please try to make the post substantially better than how you found it, for example, by fixing grammar or adding additional resources and hyperlinks.

Required fields*

4
  • 5
    $\begingroup$ (+1) Very informative answer, though it was scary when you dropped the other F word at the end of the first paragraph! $\endgroup$ Commented Nov 21, 2021 at 0:59
  • $\begingroup$ //" then learn how to do it with reduced data widths and with fixed point..."// ---- they might not need to, they could have an ARM or a SHArC or something. ---- //" ...Then learn how to do it so that it runs fast."// That's usually necessary for embedded processing because sometimes they throw the least MIPS power necessary to accomplish a job. But this answer is spot on. Brings me back to 1979 to 1990 . $\endgroup$ Commented Nov 23, 2021 at 0:49
  • $\begingroup$ @robertbristow-johnson: I'm currently working on a cost-sensitive video processing app. So, yes -- speed is essential, and part of that is going to be half-precision floats or integers or something. $\endgroup$ Commented Nov 23, 2021 at 1:48
  • 1
    $\begingroup$ If you get to influence the design of the "half-precision floats ...or something", and if you're not doing Block Floating-Point, I really recommend a float format that is like the DEC PDP-10 but put in denormals at the bottom of the range. Then no bit pattern is not a number. And fixed-point compare works with floating-point numbers. And you may not need 8 bits for the exponent, but do the constant-offset exponent like IEEE. 00000 is your smallest exponent. $\endgroup$ Commented Nov 23, 2021 at 1:59