|
- #ifndef __INC_LIB8TION_H
- #define __INC_LIB8TION_H
-
- #include "FastLED.h"
-
- #ifndef __INC_LED_SYSDEFS_H
- #error WTH? led_sysdefs needs to be included first
- #endif
-
- FASTLED_NAMESPACE_BEGIN
-
- /*
-
- Fast, efficient 8-bit math functions specifically
- designed for high-performance LED programming.
-
- Because of the AVR(Arduino) and ARM assembly language
- implementations provided, using these functions often
- results in smaller and faster code than the equivalent
- program using plain "C" arithmetic and logic.
-
-
- Included are:
-
-
- - Saturating unsigned 8-bit add and subtract.
- Instead of wrapping around if an overflow occurs,
- these routines just 'clamp' the output at a maxumum
- of 255, or a minimum of 0. Useful for adding pixel
- values. E.g., qadd8( 200, 100) = 255.
-
- qadd8( i, j) == MIN( (i + j), 0xFF )
- qsub8( i, j) == MAX( (i - j), 0 )
-
- - Saturating signed 8-bit ("7-bit") add.
- qadd7( i, j) == MIN( (i + j), 0x7F)
-
-
- - Scaling (down) of unsigned 8- and 16- bit values.
- Scaledown value is specified in 1/256ths.
- scale8( i, sc) == (i * sc) / 256
- scale16by8( i, sc) == (i * sc) / 256
-
- Example: scaling a 0-255 value down into a
- range from 0-99:
- downscaled = scale8( originalnumber, 100);
-
- A special version of scale8 is provided for scaling
- LED brightness values, to make sure that they don't
- accidentally scale down to total black at low
- dimming levels, since that would look wrong:
- scale8_video( i, sc) = ((i * sc) / 256) +? 1
-
- Example: reducing an LED brightness by a
- dimming factor:
- new_bright = scale8_video( orig_bright, dimming);
-
-
- - Fast 8- and 16- bit unsigned random numbers.
- Significantly faster than Arduino random(), but
- also somewhat less random. You can add entropy.
- random8() == random from 0..255
- random8( n) == random from 0..(N-1)
- random8( n, m) == random from N..(M-1)
-
- random16() == random from 0..65535
- random16( n) == random from 0..(N-1)
- random16( n, m) == random from N..(M-1)
-
- random16_set_seed( k) == seed = k
- random16_add_entropy( k) == seed += k
-
-
- - Absolute value of a signed 8-bit value.
- abs8( i) == abs( i)
-
-
- - 8-bit math operations which return 8-bit values.
- These are provided mostly for completeness,
- not particularly for performance.
- mul8( i, j) == (i * j) & 0xFF
- add8( i, j) == (i + j) & 0xFF
- sub8( i, j) == (i - j) & 0xFF
-
-
- - Fast 16-bit approximations of sin and cos.
- Input angle is a uint16_t from 0-65535.
- Output is a signed int16_t from -32767 to 32767.
- sin16( x) == sin( (x/32768.0) * pi) * 32767
- cos16( x) == cos( (x/32768.0) * pi) * 32767
- Accurate to more than 99% in all cases.
-
- - Fast 8-bit approximations of sin and cos.
- Input angle is a uint8_t from 0-255.
- Output is an UNsigned uint8_t from 0 to 255.
- sin8( x) == (sin( (x/128.0) * pi) * 128) + 128
- cos8( x) == (cos( (x/128.0) * pi) * 128) + 128
- Accurate to within about 2%.
-
-
- - Fast 8-bit "easing in/out" function.
- ease8InOutCubic(x) == 3(x^i) - 2(x^3)
- ease8InOutApprox(x) ==
- faster, rougher, approximation of cubic easing
- ease8InOutQuad(x) == quadratic (vs cubic) easing
-
- - Cubic, Quadratic, and Triangle wave functions.
- Input is a uint8_t representing phase withing the wave,
- similar to how sin8 takes an angle 'theta'.
- Output is a uint8_t representing the amplitude of
- the wave at that point.
- cubicwave8( x)
- quadwave8( x)
- triwave8( x)
-
- - Square root for 16-bit integers. About three times
- faster and five times smaller than Arduino's built-in
- generic 32-bit sqrt routine.
- sqrt16( uint16_t x ) == sqrt( x)
-
- - Dimming and brightening functions for 8-bit
- light values.
- dim8_video( x) == scale8_video( x, x)
- dim8_raw( x) == scale8( x, x)
- dim8_lin( x) == (x<128) ? ((x+1)/2) : scale8(x,x)
- brighten8_video( x) == 255 - dim8_video( 255 - x)
- brighten8_raw( x) == 255 - dim8_raw( 255 - x)
- brighten8_lin( x) == 255 - dim8_lin( 255 - x)
- The dimming functions in particular are suitable
- for making LED light output appear more 'linear'.
-
-
- - Linear interpolation between two values, with the
- fraction between them expressed as an 8- or 16-bit
- fixed point fraction (fract8 or fract16).
- lerp8by8( fromU8, toU8, fract8 )
- lerp16by8( fromU16, toU16, fract8 )
- lerp15by8( fromS16, toS16, fract8 )
- == from + (( to - from ) * fract8) / 256)
- lerp16by16( fromU16, toU16, fract16 )
- == from + (( to - from ) * fract16) / 65536)
- map8( in, rangeStart, rangeEnd)
- == map( in, 0, 255, rangeStart, rangeEnd);
-
- - Optimized memmove, memcpy, and memset, that are
- faster than standard avr-libc 1.8.
- memmove8( dest, src, bytecount)
- memcpy8( dest, src, bytecount)
- memset8( buf, value, bytecount)
-
- - Beat generators which return sine or sawtooth
- waves in a specified number of Beats Per Minute.
- Sine wave beat generators can specify a low and
- high range for the output. Sawtooth wave beat
- generators always range 0-255 or 0-65535.
- beatsin8( BPM, low8, high8)
- = (sine(beatphase) * (high8-low8)) + low8
- beatsin16( BPM, low16, high16)
- = (sine(beatphase) * (high16-low16)) + low16
- beatsin88( BPM88, low16, high16)
- = (sine(beatphase) * (high16-low16)) + low16
- beat8( BPM) = 8-bit repeating sawtooth wave
- beat16( BPM) = 16-bit repeating sawtooth wave
- beat88( BPM88) = 16-bit repeating sawtooth wave
- BPM is beats per minute in either simple form
- e.g. 120, or Q8.8 fixed-point form.
- BPM88 is beats per minute in ONLY Q8.8 fixed-point
- form.
-
- Lib8tion is pronounced like 'libation': lie-BAY-shun
-
- */
-
-
-
- #include <stdint.h>
-
- #define LIB8STATIC __attribute__ ((unused)) static inline
- #define LIB8STATIC_ALWAYS_INLINE __attribute__ ((always_inline)) static inline
-
- #if !defined(__AVR__)
- #include <string.h>
- // for memmove, memcpy, and memset if not defined here
- #endif // end of !defined(__AVR__)
-
- #if defined(__arm__)
-
- #if defined(FASTLED_TEENSY3)
- // Can use Cortex M4 DSP instructions
- #define QADD8_C 0
- #define QADD7_C 0
- #define QADD8_ARM_DSP_ASM 1
- #define QADD7_ARM_DSP_ASM 1
- #else
- // Generic ARM
- #define QADD8_C 1
- #define QADD7_C 1
- #endif // end of defined(FASTLED_TEENSY3)
-
- #define QSUB8_C 1
- #define SCALE8_C 1
- #define SCALE16BY8_C 1
- #define SCALE16_C 1
- #define ABS8_C 1
- #define MUL8_C 1
- #define QMUL8_C 1
- #define ADD8_C 1
- #define SUB8_C 1
- #define EASE8_C 1
- #define AVG8_C 1
- #define AVG7_C 1
- #define AVG16_C 1
- #define AVG15_C 1
- #define BLEND8_C 1
-
- // end of #if defined(__arm__)
-
- #elif defined(ARDUINO_ARCH_APOLLO3)
-
- // Default to using the standard C functions for now
- #define QADD8_C 1
- #define QADD7_C 1
- #define QSUB8_C 1
- #define SCALE8_C 1
- #define SCALE16BY8_C 1
- #define SCALE16_C 1
- #define ABS8_C 1
- #define MUL8_C 1
- #define QMUL8_C 1
- #define ADD8_C 1
- #define SUB8_C 1
- #define EASE8_C 1
- #define AVG8_C 1
- #define AVG7_C 1
- #define AVG16_C 1
- #define AVG15_C 1
- #define BLEND8_C 1
-
- // end of #elif defined(ARDUINO_ARCH_APOLLO3)
-
- #elif defined(__AVR__)
-
- // AVR ATmega and friends Arduino
-
- #define QADD8_C 0
- #define QADD7_C 0
- #define QSUB8_C 0
- #define ABS8_C 0
- #define ADD8_C 0
- #define SUB8_C 0
- #define AVG8_C 0
- #define AVG7_C 0
- #define AVG16_C 0
- #define AVG15_C 0
-
- #define QADD8_AVRASM 1
- #define QADD7_AVRASM 1
- #define QSUB8_AVRASM 1
- #define ABS8_AVRASM 1
- #define ADD8_AVRASM 1
- #define SUB8_AVRASM 1
- #define AVG8_AVRASM 1
- #define AVG7_AVRASM 1
- #define AVG16_AVRASM 1
- #define AVG15_AVRASM 1
-
- // Note: these require hardware MUL instruction
- // -- sorry, ATtiny!
- #if !defined(LIB8_ATTINY)
- #define SCALE8_C 0
- #define SCALE16BY8_C 0
- #define SCALE16_C 0
- #define MUL8_C 0
- #define QMUL8_C 0
- #define EASE8_C 0
- #define BLEND8_C 0
- #define SCALE8_AVRASM 1
- #define SCALE16BY8_AVRASM 1
- #define SCALE16_AVRASM 1
- #define MUL8_AVRASM 1
- #define QMUL8_AVRASM 1
- #define EASE8_AVRASM 1
- #define CLEANUP_R1_AVRASM 1
- #define BLEND8_AVRASM 1
- #else
- // On ATtiny, we just use C implementations
- #define SCALE8_C 1
- #define SCALE16BY8_C 1
- #define SCALE16_C 1
- #define MUL8_C 1
- #define QMUL8_C 1
- #define EASE8_C 1
- #define BLEND8_C 1
- #define SCALE8_AVRASM 0
- #define SCALE16BY8_AVRASM 0
- #define SCALE16_AVRASM 0
- #define MUL8_AVRASM 0
- #define QMUL8_AVRASM 0
- #define EASE8_AVRASM 0
- #define BLEND8_AVRASM 0
- #endif // end of !defined(LIB8_ATTINY)
-
- // end of #elif defined(__AVR__)
-
- #else
-
- // unspecified architecture, so
- // no ASM, everything in C
- #define QADD8_C 1
- #define QADD7_C 1
- #define QSUB8_C 1
- #define SCALE8_C 1
- #define SCALE16BY8_C 1
- #define SCALE16_C 1
- #define ABS8_C 1
- #define MUL8_C 1
- #define QMUL8_C 1
- #define ADD8_C 1
- #define SUB8_C 1
- #define EASE8_C 1
- #define AVG8_C 1
- #define AVG7_C 1
- #define AVG16_C 1
- #define AVG15_C 1
- #define BLEND8_C 1
-
- #endif
-
- ///@defgroup lib8tion Fast math functions
- ///A variety of functions for working with numbers.
- ///@{
-
-
- ///////////////////////////////////////////////////////////////////////
- //
- // typdefs for fixed-point fractional types.
- //
- // sfract7 should be interpreted as signed 128ths.
- // fract8 should be interpreted as unsigned 256ths.
- // sfract15 should be interpreted as signed 32768ths.
- // fract16 should be interpreted as unsigned 65536ths.
- //
- // Example: if a fract8 has the value "64", that should be interpreted
- // as 64/256ths, or one-quarter.
- //
- //
- // fract8 range is 0 to 0.99609375
- // in steps of 0.00390625
- //
- // sfract7 range is -0.9921875 to 0.9921875
- // in steps of 0.0078125
- //
- // fract16 range is 0 to 0.99998474121
- // in steps of 0.00001525878
- //
- // sfract15 range is -0.99996948242 to 0.99996948242
- // in steps of 0.00003051757
- //
-
- /// ANSI unsigned short _Fract. range is 0 to 0.99609375
- /// in steps of 0.00390625
- typedef uint8_t fract8; ///< ANSI: unsigned short _Fract
-
- /// ANSI: signed short _Fract. range is -0.9921875 to 0.9921875
- /// in steps of 0.0078125
- typedef int8_t sfract7; ///< ANSI: signed short _Fract
-
- /// ANSI: unsigned _Fract. range is 0 to 0.99998474121
- /// in steps of 0.00001525878
- typedef uint16_t fract16; ///< ANSI: unsigned _Fract
-
- /// ANSI: signed _Fract. range is -0.99996948242 to 0.99996948242
- /// in steps of 0.00003051757
- typedef int16_t sfract15; ///< ANSI: signed _Fract
-
-
- // accumXY types should be interpreted as X bits of integer,
- // and Y bits of fraction.
- // E.g., accum88 has 8 bits of int, 8 bits of fraction
-
- typedef uint16_t accum88; ///< ANSI: unsigned short _Accum. 8 bits int, 8 bits fraction
- typedef int16_t saccum78; ///< ANSI: signed short _Accum. 7 bits int, 8 bits fraction
- typedef uint32_t accum1616;///< ANSI: signed _Accum. 16 bits int, 16 bits fraction
- typedef int32_t saccum1516;///< ANSI: signed _Accum. 15 bits int, 16 bits fraction
- typedef uint16_t accum124; ///< no direct ANSI counterpart. 12 bits int, 4 bits fraction
- typedef int32_t saccum114;///< no direct ANSI counterpart. 1 bit int, 14 bits fraction
-
-
- /// typedef for IEEE754 "binary32" float type internals
- typedef union {
- uint32_t i;
- float f;
- struct {
- uint32_t mantissa: 23;
- uint32_t exponent: 8;
- uint32_t signbit: 1;
- };
- struct {
- uint32_t mant7 : 7;
- uint32_t mant16: 16;
- uint32_t exp_ : 8;
- uint32_t sb_ : 1;
- };
- struct {
- uint32_t mant_lo8 : 8;
- uint32_t mant_hi16_exp_lo1 : 16;
- uint32_t sb_exphi7 : 8;
- };
- } IEEE754binary32_t;
-
- #include "lib8tion/math8.h"
- #include "lib8tion/scale8.h"
- #include "lib8tion/random8.h"
- #include "lib8tion/trig8.h"
-
- ///////////////////////////////////////////////////////////////////////
-
-
-
-
-
-
-
- ///////////////////////////////////////////////////////////////////////
- //
- // float-to-fixed and fixed-to-float conversions
- //
- // Note that anything involving a 'float' on AVR will be slower.
-
- /// sfract15ToFloat: conversion from sfract15 fixed point to
- /// IEEE754 32-bit float.
- LIB8STATIC float sfract15ToFloat( sfract15 y)
- {
- return y / 32768.0;
- }
-
- /// conversion from IEEE754 float in the range (-1,1)
- /// to 16-bit fixed point. Note that the extremes of
- /// one and negative one are NOT representable. The
- /// representable range is basically
- LIB8STATIC sfract15 floatToSfract15( float f)
- {
- return f * 32768.0;
- }
-
-
-
- ///////////////////////////////////////////////////////////////////////
- //
- // memmove8, memcpy8, and memset8:
- // alternatives to memmove, memcpy, and memset that are
- // faster on AVR than standard avr-libc 1.8
-
- #if defined(__AVR__)
- extern "C" {
- void * memmove8( void * dst, const void * src, uint16_t num );
- void * memcpy8 ( void * dst, const void * src, uint16_t num ) __attribute__ ((noinline));
- void * memset8 ( void * ptr, uint8_t value, uint16_t num ) __attribute__ ((noinline)) ;
- }
- #else
- // on non-AVR platforms, these names just call standard libc.
- #define memmove8 memmove
- #define memcpy8 memcpy
- #define memset8 memset
- #endif
-
-
- ///////////////////////////////////////////////////////////////////////
- //
- // linear interpolation, such as could be used for Perlin noise, etc.
- //
-
- // A note on the structure of the lerp functions:
- // The cases for b>a and b<=a are handled separately for
- // speed: without knowing the relative order of a and b,
- // the value (a-b) might be overflow the width of a or b,
- // and have to be promoted to a wider, slower type.
- // To avoid that, we separate the two cases, and are able
- // to do all the math in the same width as the arguments,
- // which is much faster and smaller on AVR.
-
- /// linear interpolation between two unsigned 8-bit values,
- /// with 8-bit fraction
- LIB8STATIC uint8_t lerp8by8( uint8_t a, uint8_t b, fract8 frac)
- {
- uint8_t result;
- if( b > a) {
- uint8_t delta = b - a;
- uint8_t scaled = scale8( delta, frac);
- result = a + scaled;
- } else {
- uint8_t delta = a - b;
- uint8_t scaled = scale8( delta, frac);
- result = a - scaled;
- }
- return result;
- }
-
- /// linear interpolation between two unsigned 16-bit values,
- /// with 16-bit fraction
- LIB8STATIC uint16_t lerp16by16( uint16_t a, uint16_t b, fract16 frac)
- {
- uint16_t result;
- if( b > a ) {
- uint16_t delta = b - a;
- uint16_t scaled = scale16(delta, frac);
- result = a + scaled;
- } else {
- uint16_t delta = a - b;
- uint16_t scaled = scale16( delta, frac);
- result = a - scaled;
- }
- return result;
- }
-
- /// linear interpolation between two unsigned 16-bit values,
- /// with 8-bit fraction
- LIB8STATIC uint16_t lerp16by8( uint16_t a, uint16_t b, fract8 frac)
- {
- uint16_t result;
- if( b > a) {
- uint16_t delta = b - a;
- uint16_t scaled = scale16by8( delta, frac);
- result = a + scaled;
- } else {
- uint16_t delta = a - b;
- uint16_t scaled = scale16by8( delta, frac);
- result = a - scaled;
- }
- return result;
- }
-
- /// linear interpolation between two signed 15-bit values,
- /// with 8-bit fraction
- LIB8STATIC int16_t lerp15by8( int16_t a, int16_t b, fract8 frac)
- {
- int16_t result;
- if( b > a) {
- uint16_t delta = b - a;
- uint16_t scaled = scale16by8( delta, frac);
- result = a + scaled;
- } else {
- uint16_t delta = a - b;
- uint16_t scaled = scale16by8( delta, frac);
- result = a - scaled;
- }
- return result;
- }
-
- /// linear interpolation between two signed 15-bit values,
- /// with 8-bit fraction
- LIB8STATIC int16_t lerp15by16( int16_t a, int16_t b, fract16 frac)
- {
- int16_t result;
- if( b > a) {
- uint16_t delta = b - a;
- uint16_t scaled = scale16( delta, frac);
- result = a + scaled;
- } else {
- uint16_t delta = a - b;
- uint16_t scaled = scale16( delta, frac);
- result = a - scaled;
- }
- return result;
- }
-
- /// map8: map from one full-range 8-bit value into a narrower
- /// range of 8-bit values, possibly a range of hues.
- ///
- /// E.g. map myValue into a hue in the range blue..purple..pink..red
- /// hue = map8( myValue, HUE_BLUE, HUE_RED);
- ///
- /// Combines nicely with the waveform functions (like sin8, etc)
- /// to produce continuous hue gradients back and forth:
- ///
- /// hue = map8( sin8( myValue), HUE_BLUE, HUE_RED);
- ///
- /// Mathematically simiar to lerp8by8, but arguments are more
- /// like Arduino's "map"; this function is similar to
- ///
- /// map( in, 0, 255, rangeStart, rangeEnd)
- ///
- /// but faster and specifically designed for 8-bit values.
- LIB8STATIC uint8_t map8( uint8_t in, uint8_t rangeStart, uint8_t rangeEnd)
- {
- uint8_t rangeWidth = rangeEnd - rangeStart;
- uint8_t out = scale8( in, rangeWidth);
- out += rangeStart;
- return out;
- }
-
-
- ///////////////////////////////////////////////////////////////////////
- //
- // easing functions; see http://easings.net
- //
-
- /// ease8InOutQuad: 8-bit quadratic ease-in / ease-out function
- /// Takes around 13 cycles on AVR
- #if EASE8_C == 1
- LIB8STATIC uint8_t ease8InOutQuad( uint8_t i)
- {
- uint8_t j = i;
- if( j & 0x80 ) {
- j = 255 - j;
- }
- uint8_t jj = scale8( j, j);
- uint8_t jj2 = jj << 1;
- if( i & 0x80 ) {
- jj2 = 255 - jj2;
- }
- return jj2;
- }
-
- #elif EASE8_AVRASM == 1
- // This AVR asm version of ease8InOutQuad preserves one more
- // low-bit of precision than the C version, and is also slightly
- // smaller and faster.
- LIB8STATIC uint8_t ease8InOutQuad(uint8_t val) {
- uint8_t j=val;
- asm volatile (
- "sbrc %[val], 7 \n"
- "com %[j] \n"
- "mul %[j], %[j] \n"
- "add r0, %[j] \n"
- "ldi %[j], 0 \n"
- "adc %[j], r1 \n"
- "lsl r0 \n" // carry = high bit of low byte of mul product
- "rol %[j] \n" // j = (j * 2) + carry // preserve add'l bit of precision
- "sbrc %[val], 7 \n"
- "com %[j] \n"
- "clr __zero_reg__ \n"
- : [j] "+&a" (j)
- : [val] "a" (val)
- : "r0", "r1"
- );
- return j;
- }
-
- #else
- #error "No implementation for ease8InOutQuad available."
- #endif
-
- /// ease16InOutQuad: 16-bit quadratic ease-in / ease-out function
- // C implementation at this point
- LIB8STATIC uint16_t ease16InOutQuad( uint16_t i)
- {
- uint16_t j = i;
- if( j & 0x8000 ) {
- j = 65535 - j;
- }
- uint16_t jj = scale16( j, j);
- uint16_t jj2 = jj << 1;
- if( i & 0x8000 ) {
- jj2 = 65535 - jj2;
- }
- return jj2;
- }
-
-
- /// ease8InOutCubic: 8-bit cubic ease-in / ease-out function
- /// Takes around 18 cycles on AVR
- LIB8STATIC fract8 ease8InOutCubic( fract8 i)
- {
- uint8_t ii = scale8_LEAVING_R1_DIRTY( i, i);
- uint8_t iii = scale8_LEAVING_R1_DIRTY( ii, i);
-
- uint16_t r1 = (3 * (uint16_t)(ii)) - ( 2 * (uint16_t)(iii));
-
- /* the code generated for the above *'s automatically
- cleans up R1, so there's no need to explicitily call
- cleanup_R1(); */
-
- uint8_t result = r1;
-
- // if we got "256", return 255:
- if( r1 & 0x100 ) {
- result = 255;
- }
- return result;
- }
-
- /// ease8InOutApprox: fast, rough 8-bit ease-in/ease-out function
- /// shaped approximately like 'ease8InOutCubic',
- /// it's never off by more than a couple of percent
- /// from the actual cubic S-curve, and it executes
- /// more than twice as fast. Use when the cycles
- /// are more important than visual smoothness.
- /// Asm version takes around 7 cycles on AVR.
-
- #if EASE8_C == 1
- LIB8STATIC fract8 ease8InOutApprox( fract8 i)
- {
- if( i < 64) {
- // start with slope 0.5
- i /= 2;
- } else if( i > (255 - 64)) {
- // end with slope 0.5
- i = 255 - i;
- i /= 2;
- i = 255 - i;
- } else {
- // in the middle, use slope 192/128 = 1.5
- i -= 64;
- i += (i / 2);
- i += 32;
- }
-
- return i;
- }
-
- #elif EASE8_AVRASM == 1
- LIB8STATIC uint8_t ease8InOutApprox( fract8 i)
- {
- // takes around 7 cycles on AVR
- asm volatile (
- " subi %[i], 64 \n\t"
- " cpi %[i], 128 \n\t"
- " brcc Lshift_%= \n\t"
-
- // middle case
- " mov __tmp_reg__, %[i] \n\t"
- " lsr __tmp_reg__ \n\t"
- " add %[i], __tmp_reg__ \n\t"
- " subi %[i], 224 \n\t"
- " rjmp Ldone_%= \n\t"
-
- // start or end case
- "Lshift_%=: \n\t"
- " lsr %[i] \n\t"
- " subi %[i], 96 \n\t"
-
- "Ldone_%=: \n\t"
-
- : [i] "+&a" (i)
- :
- : "r0", "r1"
- );
- return i;
- }
- #else
- #error "No implementation for ease8 available."
- #endif
-
-
-
- /// triwave8: triangle (sawtooth) wave generator. Useful for
- /// turning a one-byte ever-increasing value into a
- /// one-byte value that oscillates up and down.
- ///
- /// input output
- /// 0..127 0..254 (positive slope)
- /// 128..255 254..0 (negative slope)
- ///
- /// On AVR this function takes just three cycles.
- ///
- LIB8STATIC uint8_t triwave8(uint8_t in)
- {
- if( in & 0x80) {
- in = 255 - in;
- }
- uint8_t out = in << 1;
- return out;
- }
-
-
- // quadwave8 and cubicwave8: S-shaped wave generators (like 'sine').
- // Useful for turning a one-byte 'counter' value into a
- // one-byte oscillating value that moves smoothly up and down,
- // with an 'acceleration' and 'deceleration' curve.
- //
- // These are even faster than 'sin8', and have
- // slightly different curve shapes.
- //
-
- /// quadwave8: quadratic waveform generator. Spends just a little more
- /// time at the limits than 'sine' does.
- LIB8STATIC uint8_t quadwave8(uint8_t in)
- {
- return ease8InOutQuad( triwave8( in));
- }
-
- /// cubicwave8: cubic waveform generator. Spends visibly more time
- /// at the limits than 'sine' does.
- LIB8STATIC uint8_t cubicwave8(uint8_t in)
- {
- return ease8InOutCubic( triwave8( in));
- }
-
- /// squarewave8: square wave generator. Useful for
- /// turning a one-byte ever-increasing value
- /// into a one-byte value that is either 0 or 255.
- /// The width of the output 'pulse' is
- /// determined by the pulsewidth argument:
- ///
- ///~~~
- /// If pulsewidth is 255, output is always 255.
- /// If pulsewidth < 255, then
- /// if input < pulsewidth then output is 255
- /// if input >= pulsewidth then output is 0
- ///~~~
- ///
- /// the output looking like:
- ///
- ///~~~
- /// 255 +--pulsewidth--+
- /// . | |
- /// 0 0 +--------(256-pulsewidth)--------
- ///~~~
- ///
- /// @param in
- /// @param pulsewidth
- /// @returns square wave output
- LIB8STATIC uint8_t squarewave8( uint8_t in, uint8_t pulsewidth=128)
- {
- if( in < pulsewidth || (pulsewidth == 255)) {
- return 255;
- } else {
- return 0;
- }
- }
-
-
-
-
- /// Template class for represneting fractional ints.
- template<class T, int F, int I> class q {
- T i:I;
- T f:F;
- public:
- q(float fx) { i = fx; f = (fx-i) * (1<<F); }
- q(uint8_t _i, uint8_t _f) {i=_i; f=_f; }
- uint32_t operator*(uint32_t v) { return (v*i) + ((v*f)>>F); }
- uint16_t operator*(uint16_t v) { return (v*i) + ((v*f)>>F); }
- int32_t operator*(int32_t v) { return (v*i) + ((v*f)>>F); }
- int16_t operator*(int16_t v) { return (v*i) + ((v*f)>>F); }
- #ifdef FASTLED_ARM
- int operator*(int v) { return (v*i) + ((v*f)>>F); }
- #endif
- #ifdef FASTLED_APOLLO3
- int operator*(int v) { return (v*i) + ((v*f)>>F); }
- #endif
- };
-
- template<class T, int F, int I> static uint32_t operator*(uint32_t v, q<T,F,I> & q) { return q * v; }
- template<class T, int F, int I> static uint16_t operator*(uint16_t v, q<T,F,I> & q) { return q * v; }
- template<class T, int F, int I> static int32_t operator*(int32_t v, q<T,F,I> & q) { return q * v; }
- template<class T, int F, int I> static int16_t operator*(int16_t v, q<T,F,I> & q) { return q * v; }
- #ifdef FASTLED_ARM
- template<class T, int F, int I> static int operator*(int v, q<T,F,I> & q) { return q * v; }
- #endif
- #ifdef FASTLED_APOLLO3
- template<class T, int F, int I> static int operator*(int v, q<T,F,I> & q) { return q * v; }
- #endif
-
- /// A 4.4 integer (4 bits integer, 4 bits fraction)
- typedef q<uint8_t, 4,4> q44;
- /// A 6.2 integer (6 bits integer, 2 bits fraction)
- typedef q<uint8_t, 6,2> q62;
- /// A 8.8 integer (8 bits integer, 8 bits fraction)
- typedef q<uint16_t, 8,8> q88;
- /// A 12.4 integer (12 bits integer, 4 bits fraction)
- typedef q<uint16_t, 12,4> q124;
-
-
-
- // Beat generators - These functions produce waves at a given
- // number of 'beats per minute'. Internally, they use
- // the Arduino function 'millis' to track elapsed time.
- // Accuracy is a bit better than one part in a thousand.
- //
- // beat8( BPM ) returns an 8-bit value that cycles 'BPM' times
- // per minute, rising from 0 to 255, resetting to zero,
- // rising up again, etc.. The output of this function
- // is suitable for feeding directly into sin8, and cos8,
- // triwave8, quadwave8, and cubicwave8.
- // beat16( BPM ) returns a 16-bit value that cycles 'BPM' times
- // per minute, rising from 0 to 65535, resetting to zero,
- // rising up again, etc. The output of this function is
- // suitable for feeding directly into sin16 and cos16.
- // beat88( BPM88) is the same as beat16, except that the BPM88 argument
- // MUST be in Q8.8 fixed point format, e.g. 120BPM must
- // be specified as 120*256 = 30720.
- // beatsin8( BPM, uint8_t low, uint8_t high) returns an 8-bit value that
- // rises and falls in a sine wave, 'BPM' times per minute,
- // between the values of 'low' and 'high'.
- // beatsin16( BPM, uint16_t low, uint16_t high) returns a 16-bit value
- // that rises and falls in a sine wave, 'BPM' times per
- // minute, between the values of 'low' and 'high'.
- // beatsin88( BPM88, ...) is the same as beatsin16, except that the
- // BPM88 argument MUST be in Q8.8 fixed point format,
- // e.g. 120BPM must be specified as 120*256 = 30720.
- //
- // BPM can be supplied two ways. The simpler way of specifying BPM is as
- // a simple 8-bit integer from 1-255, (e.g., "120").
- // The more sophisticated way of specifying BPM allows for fractional
- // "Q8.8" fixed point number (an 'accum88') with an 8-bit integer part and
- // an 8-bit fractional part. The easiest way to construct this is to multiply
- // a floating point BPM value (e.g. 120.3) by 256, (e.g. resulting in 30796
- // in this case), and pass that as the 16-bit BPM argument.
- // "BPM88" MUST always be specified in Q8.8 format.
- //
- // Originally designed to make an entire animation project pulse with brightness.
- // For that effect, add this line just above your existing call to "FastLED.show()":
- //
- // uint8_t bright = beatsin8( 60 /*BPM*/, 192 /*dimmest*/, 255 /*brightest*/ ));
- // FastLED.setBrightness( bright );
- // FastLED.show();
- //
- // The entire animation will now pulse between brightness 192 and 255 once per second.
-
-
- // The beat generators need access to a millisecond counter.
- // On Arduino, this is "millis()". On other platforms, you'll
- // need to provide a function with this signature:
- // uint32_t get_millisecond_timer();
- // that provides similar functionality.
- // You can also force use of the get_millisecond_timer function
- // by #defining USE_GET_MILLISECOND_TIMER.
- #if (defined(ARDUINO) || defined(SPARK) || defined(FASTLED_HAS_MILLIS)) && !defined(USE_GET_MILLISECOND_TIMER)
- // Forward declaration of Arduino function 'millis'.
- //uint32_t millis();
- #define GET_MILLIS millis
- #else
- uint32_t get_millisecond_timer();
- #define GET_MILLIS get_millisecond_timer
- #endif
-
- // beat16 generates a 16-bit 'sawtooth' wave at a given BPM,
- /// with BPM specified in Q8.8 fixed-point format; e.g.
- /// for this function, 120 BPM MUST BE specified as
- /// 120*256 = 30720.
- /// If you just want to specify "120", use beat16 or beat8.
- LIB8STATIC uint16_t beat88( accum88 beats_per_minute_88, uint32_t timebase = 0)
- {
- // BPM is 'beats per minute', or 'beats per 60000ms'.
- // To avoid using the (slower) division operator, we
- // want to convert 'beats per 60000ms' to 'beats per 65536ms',
- // and then use a simple, fast bit-shift to divide by 65536.
- //
- // The ratio 65536:60000 is 279.620266667:256; we'll call it 280:256.
- // The conversion is accurate to about 0.05%, more or less,
- // e.g. if you ask for "120 BPM", you'll get about "119.93".
- return (((GET_MILLIS()) - timebase) * beats_per_minute_88 * 280) >> 16;
- }
-
- /// beat16 generates a 16-bit 'sawtooth' wave at a given BPM
- LIB8STATIC uint16_t beat16( accum88 beats_per_minute, uint32_t timebase = 0)
- {
- // Convert simple 8-bit BPM's to full Q8.8 accum88's if needed
- if( beats_per_minute < 256) beats_per_minute <<= 8;
- return beat88(beats_per_minute, timebase);
- }
-
- /// beat8 generates an 8-bit 'sawtooth' wave at a given BPM
- LIB8STATIC uint8_t beat8( accum88 beats_per_minute, uint32_t timebase = 0)
- {
- return beat16( beats_per_minute, timebase) >> 8;
- }
-
- /// beatsin88 generates a 16-bit sine wave at a given BPM,
- /// that oscillates within a given range.
- /// For this function, BPM MUST BE SPECIFIED as
- /// a Q8.8 fixed-point value; e.g. 120BPM must be
- /// specified as 120*256 = 30720.
- /// If you just want to specify "120", use beatsin16 or beatsin8.
- LIB8STATIC uint16_t beatsin88( accum88 beats_per_minute_88, uint16_t lowest = 0, uint16_t highest = 65535,
- uint32_t timebase = 0, uint16_t phase_offset = 0)
- {
- uint16_t beat = beat88( beats_per_minute_88, timebase);
- uint16_t beatsin = (sin16( beat + phase_offset) + 32768);
- uint16_t rangewidth = highest - lowest;
- uint16_t scaledbeat = scale16( beatsin, rangewidth);
- uint16_t result = lowest + scaledbeat;
- return result;
- }
-
- /// beatsin16 generates a 16-bit sine wave at a given BPM,
- /// that oscillates within a given range.
- LIB8STATIC uint16_t beatsin16( accum88 beats_per_minute, uint16_t lowest = 0, uint16_t highest = 65535,
- uint32_t timebase = 0, uint16_t phase_offset = 0)
- {
- uint16_t beat = beat16( beats_per_minute, timebase);
- uint16_t beatsin = (sin16( beat + phase_offset) + 32768);
- uint16_t rangewidth = highest - lowest;
- uint16_t scaledbeat = scale16( beatsin, rangewidth);
- uint16_t result = lowest + scaledbeat;
- return result;
- }
-
- /// beatsin8 generates an 8-bit sine wave at a given BPM,
- /// that oscillates within a given range.
- LIB8STATIC uint8_t beatsin8( accum88 beats_per_minute, uint8_t lowest = 0, uint8_t highest = 255,
- uint32_t timebase = 0, uint8_t phase_offset = 0)
- {
- uint8_t beat = beat8( beats_per_minute, timebase);
- uint8_t beatsin = sin8( beat + phase_offset);
- uint8_t rangewidth = highest - lowest;
- uint8_t scaledbeat = scale8( beatsin, rangewidth);
- uint8_t result = lowest + scaledbeat;
- return result;
- }
-
-
- /// Return the current seconds since boot in a 16-bit value. Used as part of the
- /// "every N time-periods" mechanism
- LIB8STATIC uint16_t seconds16()
- {
- uint32_t ms = GET_MILLIS();
- uint16_t s16;
- s16 = ms / 1000;
- return s16;
- }
-
- /// Return the current minutes since boot in a 16-bit value. Used as part of the
- /// "every N time-periods" mechanism
- LIB8STATIC uint16_t minutes16()
- {
- uint32_t ms = GET_MILLIS();
- uint16_t m16;
- m16 = (ms / (60000L)) & 0xFFFF;
- return m16;
- }
-
- /// Return the current hours since boot in an 8-bit value. Used as part of the
- /// "every N time-periods" mechanism
- LIB8STATIC uint8_t hours8()
- {
- uint32_t ms = GET_MILLIS();
- uint8_t h8;
- h8 = (ms / (3600000L)) & 0xFF;
- return h8;
- }
-
-
- /// Helper routine to divide a 32-bit value by 1024, returning
- /// only the low 16 bits. You'd think this would be just
- /// result = (in32 >> 10) & 0xFFFF;
- /// and on ARM, that's what you want and all is well.
- /// But on AVR that code turns into a loop that executes
- /// a four-byte shift ten times: 40 shifts in all, plus loop
- /// overhead. This routine gets exactly the same result with
- /// just six shifts (vs 40), and no loop overhead.
- /// Used to convert millis to 'binary seconds' aka bseconds:
- /// one bsecond == 1024 millis.
- LIB8STATIC uint16_t div1024_32_16( uint32_t in32)
- {
- uint16_t out16;
- #if defined(__AVR__)
- asm volatile (
- " lsr %D[in] \n\t"
- " ror %C[in] \n\t"
- " ror %B[in] \n\t"
- " lsr %D[in] \n\t"
- " ror %C[in] \n\t"
- " ror %B[in] \n\t"
- " mov %B[out],%C[in] \n\t"
- " mov %A[out],%B[in] \n\t"
- : [in] "+r" (in32),
- [out] "=r" (out16)
- );
- #else
- out16 = (in32 >> 10) & 0xFFFF;
- #endif
- return out16;
- }
-
- /// bseconds16 returns the current time-since-boot in
- /// "binary seconds", which are actually 1024/1000 of a
- /// second long.
- LIB8STATIC uint16_t bseconds16()
- {
- uint32_t ms = GET_MILLIS();
- uint16_t s16;
- s16 = div1024_32_16( ms);
- return s16;
- }
-
-
- // Classes to implement "Every N Milliseconds", "Every N Seconds",
- // "Every N Minutes", "Every N Hours", and "Every N BSeconds".
- #if 1
- #define INSTANTIATE_EVERY_N_TIME_PERIODS(NAME,TIMETYPE,TIMEGETTER) \
- class NAME { \
- public: \
- TIMETYPE mPrevTrigger; \
- TIMETYPE mPeriod; \
- \
- NAME() { reset(); mPeriod = 1; }; \
- NAME(TIMETYPE period) { reset(); setPeriod(period); }; \
- void setPeriod( TIMETYPE period) { mPeriod = period; }; \
- TIMETYPE getTime() { return (TIMETYPE)(TIMEGETTER()); }; \
- TIMETYPE getPeriod() { return mPeriod; }; \
- TIMETYPE getElapsed() { return getTime() - mPrevTrigger; } \
- TIMETYPE getRemaining() { return mPeriod - getElapsed(); } \
- TIMETYPE getLastTriggerTime() { return mPrevTrigger; } \
- bool ready() { \
- bool isReady = (getElapsed() >= mPeriod); \
- if( isReady ) { reset(); } \
- return isReady; \
- } \
- void reset() { mPrevTrigger = getTime(); }; \
- void trigger() { mPrevTrigger = getTime() - mPeriod; }; \
- \
- operator bool() { return ready(); } \
- };
- INSTANTIATE_EVERY_N_TIME_PERIODS(CEveryNMillis,uint32_t,GET_MILLIS);
- INSTANTIATE_EVERY_N_TIME_PERIODS(CEveryNSeconds,uint16_t,seconds16);
- INSTANTIATE_EVERY_N_TIME_PERIODS(CEveryNBSeconds,uint16_t,bseconds16);
- INSTANTIATE_EVERY_N_TIME_PERIODS(CEveryNMinutes,uint16_t,minutes16);
- INSTANTIATE_EVERY_N_TIME_PERIODS(CEveryNHours,uint8_t,hours8);
- #else
-
- // Under C++11 rules, we would be allowed to use not-external
- // -linkage-type symbols as template arguments,
- // e.g., LIB8STATIC seconds16, and we'd be able to use these
- // templates as shown below.
- // However, under C++03 rules, we cannot do that, and thus we
- // have to resort to the preprocessor to 'instantiate' 'templates',
- // as handled above.
- template<typename timeType,timeType (*timeGetter)()>
- class CEveryNTimePeriods {
- public:
- timeType mPrevTrigger;
- timeType mPeriod;
-
- CEveryNTimePeriods() { reset(); mPeriod = 1; };
- CEveryNTimePeriods(timeType period) { reset(); setPeriod(period); };
- void setPeriod( timeType period) { mPeriod = period; };
- timeType getTime() { return (timeType)(timeGetter()); };
- timeType getPeriod() { return mPeriod; };
- timeType getElapsed() { return getTime() - mPrevTrigger; }
- timeType getRemaining() { return mPeriod - getElapsed(); }
- timeType getLastTriggerTime() { return mPrevTrigger; }
- bool ready() {
- bool isReady = (getElapsed() >= mPeriod);
- if( isReady ) { reset(); }
- return isReady;
- }
- void reset() { mPrevTrigger = getTime(); };
- void trigger() { mPrevTrigger = getTime() - mPeriod; };
-
- operator bool() { return ready(); }
- };
- typedef CEveryNTimePeriods<uint16_t,seconds16> CEveryNSeconds;
- typedef CEveryNTimePeriods<uint16_t,bseconds16> CEveryNBSeconds;
- typedef CEveryNTimePeriods<uint32_t,millis> CEveryNMillis;
- typedef CEveryNTimePeriods<uint16_t,minutes16> CEveryNMinutes;
- typedef CEveryNTimePeriods<uint8_t,hours8> CEveryNHours;
- #endif
-
-
- #define CONCAT_HELPER( x, y ) x##y
- #define CONCAT_MACRO( x, y ) CONCAT_HELPER( x, y )
- #define EVERY_N_MILLIS(N) EVERY_N_MILLIS_I(CONCAT_MACRO(PER, __COUNTER__ ),N)
- #define EVERY_N_MILLIS_I(NAME,N) static CEveryNMillis NAME(N); if( NAME )
- #define EVERY_N_SECONDS(N) EVERY_N_SECONDS_I(CONCAT_MACRO(PER, __COUNTER__ ),N)
- #define EVERY_N_SECONDS_I(NAME,N) static CEveryNSeconds NAME(N); if( NAME )
- #define EVERY_N_BSECONDS(N) EVERY_N_BSECONDS_I(CONCAT_MACRO(PER, __COUNTER__ ),N)
- #define EVERY_N_BSECONDS_I(NAME,N) static CEveryNBSeconds NAME(N); if( NAME )
- #define EVERY_N_MINUTES(N) EVERY_N_MINUTES_I(CONCAT_MACRO(PER, __COUNTER__ ),N)
- #define EVERY_N_MINUTES_I(NAME,N) static CEveryNMinutes NAME(N); if( NAME )
- #define EVERY_N_HOURS(N) EVERY_N_HOURS_I(CONCAT_MACRO(PER, __COUNTER__ ),N)
- #define EVERY_N_HOURS_I(NAME,N) static CEveryNHours NAME(N); if( NAME )
-
- #define CEveryNMilliseconds CEveryNMillis
- #define EVERY_N_MILLISECONDS(N) EVERY_N_MILLIS(N)
- #define EVERY_N_MILLISECONDS_I(NAME,N) EVERY_N_MILLIS_I(NAME,N)
-
- FASTLED_NAMESPACE_END
- ///@}
-
- #endif
|