FR_Math
A C language fixed-point math library for embedded systems.
FR_Math is a compact, integer-only fixed-point math library built for systems where floating point is too slow, too big, or unavailable. Designed for embedded targets ranging from legacy 16 MHz 68k processors to modern Cortex-M and RISC-V cores, it provides a full suite of math primitives — trigonometry, logarithms, roots, transforms, and signal generators — while remaining deterministic, portable, and small. Unlike traditional fixed-point libraries, FR_Math lets the caller choose the binary point per operation, trading precision and range explicitly instead of locking into a single format.
- Pure C (C99/C11/C17) with an optional C++ 2D-transform wrapper. Tested on gcc, clang, MSVC, IAR, Keil, sdcc, AVR-gcc, MSP430-gcc, RISC-V toolchains, and Arduino.
- Zero dependencies beyond
<stdint.h>. - Parameterized radix: every function takes the binary point as an argument, so you choose how many fractional bits you need per call.
- Deterministic, bounded error — every public symbol has a documented worst case in the API reference.
Measured accuracy
Errors below are measured at Q16.16 (s15.16). All functions accept any
radix — Q16.16 is just the reference point for the table.
Run make test-tdd to generate the TDD report
(build/test_tdd_report.md) with sweeps at radixes 8, 12, 16, and 24.
| Function | Max err (%)* | Avg err (%) | Note |
|---|---|---|---|
| sin/cos (BAM) | 0.1526 | 0.0030 | very fast binary angle trig |
| sin/cos (deg) | 0.1526 | 0.0029 | degree input trig fns |
| sin/cos (rad) | 0.1828 | 0.0033 | radian (traditional) trig |
| tan (BAM) | 0.5823 | 0.0008 | binary angle tangent; ±maxint at poles |
| tan (deg) | 0.5311 | 0.0008 | degree input tangent; saturated at poles |
| tan (rad) | 0.0386 | 0.0001 | radian (traditional) tangent |
| asin / acos | 0.7771 | 0.0280 | reverse trig, radian output |
| atan2 | 0.2564 | 0.0237 | reverse tangent, always safe |
| atan | 0.2425 | 0.0155 | reverse tangent, accepts up to maxint |
| sqrt | 0.0000 | 0.0000 | Round-to-nearest |
| log2 | 0.0116 | 0.0016 | shift/add only for speed |
| pow2 | 0.0018 | 0.0004 | shift/add only for speed |
| ln, log10 | 0.0004 | 0.0000 | shift/add only for speed |
| exp | 0.0003 | 0.0000 | shift/add only for speed |
| exp_fast | 0.0009 | 0.0001 | Shift-only scaling |
| pow10 | 0.0005 | 0.0000 | shift/add only for speed |
| pow10_fast | 0.0022 | 0.0002 | Shift-only scaling |
| hypot (exact) | 0.0000 | 0.0000 | Uses 64-bit intermediate |
| hypot_fast8 (8-seg) | 0.0915 | 0.0320 | Shift-only, no multiply |
*Relative error; reference clamped to 1% of full-scale output.
What’s in the box
| Area | Functions |
|---|---|
| Arithmetic | FR_ADD, FR_SUB, FR_DIV, FR_DIV32, FR_MOD, FR_FixMuls, FR_FixMulSat, FR_CHRDX |
| Utility | FR_MIN, FR_MAX, FR_CLAMP, FR_ABS, FR_SGN |
| Trig (integer deg) | fr_sin_deg, fr_cos_deg, fr_tan_deg, FR_SinI, FR_CosI, FR_TanI |
| Trig (radian/BAM) | fr_sin, fr_cos, fr_tan, fr_sin_bam, fr_cos_bam, fr_tan_bam |
| Inverse trig | FR_atan, FR_atan2, FR_asin, FR_acos |
| Log / exp | FR_log2, FR_ln, FR_log10, FR_pow2, FR_EXP, FR_POW10, FR_EXP_FAST, FR_POW10_FAST, FR_MULK28 |
| Roots | FR_sqrt, FR_hypot, FR_hypot_fast8 |
| Wave generators | fr_wave_sqr, fr_wave_pwm, fr_wave_tri, fr_wave_saw, fr_wave_tri_morph, fr_wave_noise |
| Envelope | fr_adsr_init, fr_adsr_trigger, fr_adsr_release, fr_adsr_step |
| 2D transforms | FR_Matrix2D_CPT (mul, add, sub, det, inv, setrotate, XFormPtI, XFormPtI16) |
| Formatted output | FR_printNumD, FR_printNumF, FR_printNumH, FR_numstr |
Every function is covered by the TDD characterization suite in the repo.
Lean build options
Compile-time #define guards let you strip optional subsystems
for ROM-constrained targets. Define them before including
FR_math.h (or pass -D on the compiler command line):
| Define | What it removes | Typical savings |
|---|---|---|
FR_LEAN | Degree trig, BAM tan, angle converters, FR_log10, FR_hypot, waves + ADSR | ~3.7 KB |
FR_NO_PRINT | FR_printNumF, FR_printNumD, FR_printNumH, FR_numstr | ~1.3 KB |
FR_NO_WAVES | fr_wave_* (6 shapes), fr_adsr_* (ADSR envelope), FR_HZ2BAM_INC | ~0.6 KB |
FR_LEAN keeps only radian trig (sin, cos, tan), inverse trig,
sqrt, log2, ln, exp, pow2, and arithmetic — comparable to libfixmath’s
API but at 4.7 KB text vs libfixmath’s 4.9 KB + 112 KB BSS.
With FR_LEAN + FR_NO_PRINT the library compiles to
~4.7 KB on x86-64 / clang -Os.
/* Example: headless sensor node — math only, no print, no audio */
#define FR_NO_PRINT
#define FR_NO_WAVES
#include "FR_math.h"With -ffunction-sections and linker --gc-sections,
the linker will also strip any unused functions automatically, so these guards
are most useful when you include the library as a single .c file
or static archive without section-level dead-code elimination.
Why fixed-point?
Many modern microcontrollers have an FPU and can use float
freely. Older and low-cost MCUs remain common. Fixed-point is often faster and
more deterministic than float, and it excels in situations like:
- 8- and 16-bit MCUs (AVR, MSP430, 8051, SDCC) where the FPU does not exist and even software float is too slow or too large.
- Hot inner loops on any CPU where a
parameterized-radix integer multiply is faster and more
deterministic than a
float. Think DSP taps, PID loops, coordinate transforms inside a scanline renderer. - Bit-exact reproducibility across compilers, architectures, and hosts — something IEEE float does not give you in the general case.
- ROM-tight builds where linking
libmorlibgcc_spulls in more code than the whole application logic.
FR_Math is engineered for these use cases. It does not try to be a
generic float replacement.
Quick taste
#include "FR_math.h"
#define R 16 /* work at radix 16 (s15.16) throughout */
/* ---- Creating fixed-point values ----
*
* FR_NUM(integer, frac_digits, num_digits, radix) encodes a decimal
* literal at compile time. The fractional part is the digits AFTER
* the decimal point, and num_digits says how many digits that is.
* Think: FR_NUM(3, 14159, 5, 16) means "3.14159" at radix 16.
*/
s32 pi = FR_NUM(3, 14159, 5, R); /* 3.14159 → raw 205886 at r16 */
s32 half = FR_NUM(0, 5, 1, R); /* 0.5 → raw 32768 */
s32 neg = FR_NUM(-2, 75, 2, R); /* -2.75 → raw -180224 */
/* Or parse from a string at runtime (no floats, no strtod): */
s32 pi2 = FR_numstr("3.14159", R); /* same result as FR_NUM above */
/* Integer-to-fixed: I2FR(n, radix) just shifts left */
s32 two = I2FR(2, R); /* 2.0 → raw 131072 */
/* ---- Naming convention: macros vs functions ----
*
* UPPERCASE FR_ names are macros — they expand inline with no call
* overhead, and the compiler can constant-fold them. Use these for
* conversions and simple arithmetic:
* I2FR, FR2I, FR_NUM, FR_ADD, FR_DIV, FR_ABS, FR_CHRDX, FR_EXP ...
*
* MixedCase FR_ names are legacy functions — they still work but
* map to the current lowercase names:
* FR_Cos → fr_cos_deg, FR_Sin → fr_sin_deg, FR_Tan → fr_tan_deg
*
* lowercase fr_ names are the current API (degree wrappers, radian
* trig, BAM trig, wave generators, ADSR envelopes):
* fr_cos_deg, fr_sin_deg, fr_tan_deg, fr_sin, fr_cos, fr_tan,
* fr_wave_tri, fr_adsr_step ...
*
* Other MixedCase / lowercase FR_ names are functions with loops,
* tables, or multi-step algorithms:
* FR_sqrt, FR_atan2, FR_log2, FR_pow2, FR_printNumF ...
*
* Some macros wrap functions: FR_EXP(x,r) scales x then calls
* FR_pow2 — one-liner convenience, heavy lifting in the function.
*/
/* ---- Math functions ---- */
s32 c45 = fr_cos_deg(45, 0); /* cos(45°) = 0.7071 */
s32 s30 = fr_sin(FR_numstr("0.5236", R), R); /* sin(0.5236 rad) */
s32 root2 = FR_sqrt(two, R); /* sqrt(2) = 1.4142 */
s32 angle = FR_atan2(I2FR(1,R), I2FR(1,R), R); /* atan2(1,1) rad */
s32 lg = FR_log2(I2FR(1000, R), R, R); /* log2(1000) ~ 9.97 */
s32 ex = FR_EXP(I2FR(1, R), R); /* macro: scales then calls
* FR_pow2 internally */
/* ---- Printing (serial / UART / file friendly) ----
*
* FR_printNumF takes a per-character output function — works with
* putchar, Serial.write, UART_putc, or any int(*)(char). No
* sprintf, no floats, no heap. Ideal for bare-metal targets.
*/
int my_putchar(char c) { return putchar(c); } /* or your UART func */
FR_printNumF(my_putchar, pi, R, 8, 5); /* prints " 3.14159" */
FR_printNumF(my_putchar, neg, R, 8, 2); /* prints " -2.75" */
FR_printNumD(my_putchar, FR2I(lg, R), 4); /* prints " 9" (integer)*/See Getting Started for a complete walkthrough, or jump straight to the Fixed-Point Primer if you want to understand how the radix notation works first.
Comparison
| Feature | libfixmath | CMSIS-DSP | FR_Math |
|---|---|---|---|
| Fixed format | Q16.16 only | Q31 / Q15 | Any radix |
| Angle input | Radians (Q16.16) | Radians (float) | BAM (u16), degrees, or radians |
| Exact cardinal angles | No | N/A | Yes |
| Multiply-free option | No | No | Yes (e.g. FR_EXP_FAST, FR_hypot_fast8) |
| Wave generators | No | No | 6 shapes + ADSR |
| Dependencies | None | ARM only | None |
| Code size (Cortex-M0, -Os) | 2.4 KB | ~40 KB+ | 3.4 KB lean / 5.7 KB full |
Sizes measured with arm-none-eabi-gcc -mcpu=cortex-m0
-mthumb -Os. libfixmath covers trig/sqrt/exp in Q16.16 only;
FR_Math includes log/ln/log10, wave generators, ADSR, print helpers,
and variable radix. CMSIS-DSP estimate is for the math function subset
only. See
scripts/crossbuild_sizes.sh
for the build script.
History
FR_Math has been in service since 2000, originally built for graphics transforms on 16 MHz 68k Palm Pilots (it shipped inside Trumpetsoft’s Inkstorm), then ported forward to ARM, x86, MIPS, RISC-V, and various 8/16-bit embedded targets. The current release has a full test suite, bit-exact numerical specification, and CI on every push.
License
BSD-2-Clause. Use it freely in open source or commercial projects.