Advertisements

Teensy 3.6 Double Precision Floats

Having spent a bit of effort wringing enough precision from an Arduino to make the 60 kHz quartz resonator tester, this came as a relief:

DDS frequency:  180000000.0000000 Hz
      epsilon:          0.0000001 Hz
         step:          0.0419095 Hz

Center frequency:  146520000.0000000 Hz
  146520000.0000001 Hz
  146520000.0000002 Hz
  146520000.0000003 Hz
  146520000.0000004 Hz
  146520000.0000004 Hz
  146520000.0000005 Hz
  146520000.0000006 Hz
  146520000.0000007 Hz
  146520000.0000008 Hz
  146520000.0000009 Hz
  146520000.0000010 Hz

... snippage ...

  146520000.0000099 Hz
  146520000.0000100 Hz
  146520000.0419195 Hz
  146520000.0838290 Hz
  146520000.1257386 Hz
  146520000.1676481 Hz
  146520000.2095576 Hz
  146520000.2514671 Hz
  146520000.2933766 Hz
  146520000.3352861 Hz
  146520000.3771957 Hz
  146520000.4191052 Hz
  146520000.4610147 Hz
  146520000.5029242 Hz
  146520000.5448337 Hz
  146520000.5867432 Hz
  146520000.6286528 Hz
  146520000.6705623 Hz
  146520000.7124718 Hz
  146520000.7543813 Hz
  146520000.7962908 Hz
  146520000.8382003 Hz
  146520000.8801098 Hz
  146520000.9220194 Hz
  146520000.9639289 Hz
  146520001.0058384 Hz
  146520001.0477479 Hz
  146520001.0896574 Hz
  146520001.1315669 Hz
  146520001.1734765 Hz

Which comes from a PJRC Teensy 3.6 running this code:

double DDSFreq, EpsilonFreq, DDSStepFreq;
double CenterFreq, TestFreq;

... in setup() ...

  DDSFreq = 180.0e6;
  EpsilonFreq = 1.0e-7;
  DDSStepFreq = DDSFreq / (1LL << 32);
  Serial.printf("DDS frequency: %18.7f Hz\n",DDSFreq);
  Serial.printf("      epsilon: %18.7f Hz\n",EpsilonFreq);
  Serial.printf("         step: %18.7f Hz\n\n",DDSStepFreq);

  CenterFreq = 146520000.0;
  TestFreq = CenterFreq;
  Serial.printf("Center frequency: %18.7f Hz\n",CenterFreq);

... in loop() ...

  if (TestFreq < (CenterFreq + 100*EpsilonFreq))
    TestFreq += EpsilonFreq;
  else
    TestFreq += DDSStepFreq;

  Serial.printf(" %18.7f Hz\n",TestFreq);

The IEEE-754 spec says a double floating-point variable carries about 15.9 decimal digits, which agrees with the 9 integer + 7 fraction digits. The highlight lowlight (gray bar) in the first figure shows the slight stumble where adding 1e-7 changes the sum, but not quite enough to affect the displayed fraction.

In round numbers, an increment of 1e-5 would work just fine:

  
  146520000.0000100 Hz
  146520000.0000200 Hz
  146520000.0000300 Hz
  146520000.0000401 Hz
  146520000.0000501 Hz
  146520000.0000601 Hz
  146520000.0000701 Hz
  146520000.0000801 Hz
  146520000.0000901 Hz
  146520000.0001001 Hz
  146520000.0001101 Hz
  146520000.0001202 Hz
  146520000.0001302 Hz
  146520000.0001402 Hz
  146520000.0001502 Hz
  146520000.0001602 Hz
  146520000.0001702 Hz
  146520000.0001802 Hz
  146520000.0001903 Hz
  146520000.0002003 Hz
  146520000.0002103 Hz
  146520000.0002203 Hz
  146520000.0002303 Hz

You’d use the “smallest of all” epsilon in a multiplied increment, perhaps to tick a value based on a knob or some such. Fine-tuning a VHF frequency with millihertz steps probably doesn’t make much practical sense.

The DDS frequency increment works out to 41.9095 mHz, slightly larger than with the Arduino, because it’s fot a cheap DDS eBay module with an AD9851 running a 180 MHz (6 × 30 MHz ) clock.

Advertisements

,

  1. #1 by Hexley Ball on 2018-04-19 - 12:37

    “The highlight lowlight (gray bar) in the first figure shows the slight stumble where adding 1e-7 changes the sum, but not quite enough to affect the displayed fraction.”

    For some reason this reminds me of a class I took many moons ago on the topic of computer methods for mathematical computation. This was back in the day when the universe of computing platforms was much more diverse than it is today, and some had odd word sizes. Not just 8, 16, or 32-bit architectures, but also 12, 18, and 30-bit, IIRC.

    So one of the tricks in writing portable code back then was to determine the native precision at run time. This meant running a loop that added an “epsilon” value to a variable and then checked the result. If the new value of the variable was larger than the starting value, then epsilon had not yet reached the limit of precision. So divide epsilon by 2 and repeat the the experiment until the difference was no longer detectable. Then you had found the limit of native precision. Kind of like your DDS routine :-)

    That was a fascinating class. It addressed numerical integration, solution of a set of linear equations, finding the zero of a function, and so on, All simple in theory, but not in practice. The problems were set in Fortran (!). Just for the heck of it, I ran my homework on an 8-bit Z80-based microcomputer that had a snazzy Fortran compiler, and was amazed when I got the same results as folks who used the big IBM timesharing system on campus. OK, it took me 10-20 minutes to run what they did in a few seconds. But I loved not being tied to Big Iron and being able to run code at home.

    • #2 by Ed on 2018-04-19 - 18:06

      Ewwww! Next, you’re going to be rhapsodizing over self-modifying code! [grin]

      Speaking of weird architectures, I thought the ones-complement 60-bit numbers in Lehigh’s CDC 6400 made perfect sense. Not that I quite understood what I was doing, though …

    • #3 by RCPete on 2018-04-19 - 22:24

      I was programming on Teradyne automatic test machines. The older ones used DEC processors and were 12 bits. Newer ones seemed to be Teradyne specials, and used an 18 bit word. This was in 1976-79.

      I was in Illinois when they were working on the Plato (computer aided learning) project. IIRC, they were on a CDC 7000, but I was working on a tool for an odd peripheral. (The project used a 19-ish inch monochrome plasma display that could be doubled up as a back projection screen. We were adding a “voice” with an addressable disc. Take a big chunk of 3M mag film (18-24″ diameter), and run read/record heads on actuators for tracking. The mag disc could be located independently from the main platen, and certain lengths of sound could be recorded. It might have been state of the art, but it was one hell of a kluge.)

      I was doing test boxes for the players; toggle switches for location, theta and time. Did a fair number of those boxes. I kept my sanity, but wasn’t sure about the designers…

  2. #4 by William Rutiser on 2018-04-20 - 11:55

    Thinking about the least significant bits of floating point numbers is much easier when they are displayed in hex instead of decimal. The crux of the matter is that the fraction “1/5” does not have a terminating binary representation. Conversion from binary floating point to decimal is an ugly process with conflicting objectives.

  1. Frequency Modulated DDS: SPI Mock 1 | The Smell of Molten Projects in the Morning

Spam comments vanish. Comment moderation may cause a delay.

Fill in your details below or click an icon to log in:

WordPress.com Logo

You are commenting using your WordPress.com account. Log Out /  Change )

Google+ photo

You are commenting using your Google+ account. Log Out /  Change )

Twitter picture

You are commenting using your Twitter account. Log Out /  Change )

Facebook photo

You are commenting using your Facebook account. Log Out /  Change )

Connecting to %s