4 * Copyright 2004, 2008 Develer S.r.l. (http://www.develer.com/)
8 * \brief Compute, save and load ramps for stepper motors (implementation)
12 * \author Simone Zinanni <s.zinanni@develer.com>
13 * \author Bernie Innocenti <bernie@codewiz.org>
14 * \author Giovanni Bajo <rasky@develer.com>
15 * \author Daniele Basile <asterix@develer.com>
18 * The formula used by the ramp is the following:
22 * f(t) = -------------
26 * Where <code>a</code> and <code>b</code> are the maximum and minimum speed
27 * respectively (minimum and maximum wavelength respectively), and <code>lerp</code>
28 * is a linear interpolation with a factor:
31 * lerp(a,b,t) = a + t * (b - a) = (a * (1 - t)) + (b * t)
34 * <code>t</code> must be in the [0,1] interval. It is easy to see that the
35 * following holds true:
41 * And that the function is monotonic. So, the function effectively interpolates
42 * between the maximum and minimum speed through its domain ([0,1] -> [b,a]).
44 * The curve drawn by this function is similar to 1 / (sqrt(n)), so it is slower
45 * than a linear acceleration (which would be 1/n).
47 * The floating point version uses a slightly modified function which accepts
48 * the parameter in the domain [0, MT] (where MT is maxTime, the length of the
49 * ramp, which is a setup parameter for the ramp). This is done to reduce the
50 * number of operations per step. The formula looks like this:
54 * g(t) = ----------------------------
55 * (a * MT) + t * (b - a)
58 * It can be shown that this <code>g(t) = f(t * MT)</code>. The denominator
59 * is a linear interpolation in the range [b*MT, a*MT], as t moves in the
60 * interval [0, MT]. So the interpolation interval of the function is again
61 * [b, a]. The implementation caches the value of the numerator and parts
62 * of the denominator, so that the formula becomes:
70 * g(t) = ----------------------
74 * and <code>t</code> is exactly the parameter that ramp_evaluate() gets,
75 * that is the current time (in range [0, MT]). The operations performed
76 * for each step are just an addition, a multiplication and a division.
78 * The fixed point version of the formula instead transforms the original
79 * function as follows:
83 * f(t) = ------------------------- = --------------------
85 * b * ( - * (1 - t) + t ) - * (1 - t) + t
89 * <code>t</code> must be computed by dividing the current time (24 bit integer)
90 * by the maximum time (24 bit integer). This is done by precomputing the
91 * reciprocal of the maximum time as a 0.32 fixed point number, and multiplying
92 * it to the current time. Multiplication is performed 8-bits a time by
93 * FIX_MULT32(), so that we end up with a 0.16 fixed point number for
94 * <code>t</code> (and <code>1-t</code> is just its twos-complement negation).
95 * <code>a/b</code> is in the range [0,1] (because a is always less than b,
96 * being the minimum wavelength), so it is precomputed as a 0.16 fixed point.
97 * The final step is then computing the denominator and executing the division
98 * (32 cycles using the 1-step division instruction in the DSP).
100 * The assembly implementation is needed for efficiency, but a C version of it
101 * can be easily written, in case it is needed in the future.
106 #include <cfg/debug.h>
108 #include <string.h> // memcpy()
110 void ramp_compute(struct Ramp *ramp, uint32_t clocksRamp, uint16_t clocksMinWL, uint16_t clocksMaxWL)
112 ASSERT(clocksMaxWL >= clocksMinWL);
114 // Save values in ramp struct
115 ramp->clocksRamp = clocksRamp;
116 ramp->clocksMinWL = clocksMinWL;
117 ramp->clocksMaxWL = clocksMaxWL;
119 #if RAMP_USE_FLOATING_POINT
120 ramp->precalc.gamma = ramp->clocksMaxWL - ramp->clocksMinWL;
121 ramp->precalc.beta = (float)ramp->clocksMinWL * (float)ramp->clocksRamp;
122 ramp->precalc.alpha = ramp->precalc.beta * (float)ramp->clocksMaxWL;
125 ramp->precalc.max_div_min = ((uint32_t)clocksMinWL << 16) / (uint32_t)clocksMaxWL;
127 /* Calcola 1/total_time in fixed point .32. Assumiamo che la rampa possa al
128 * massimo avere 25 bit (cioƩ valore in tick fino a 2^25, che con il
129 * prescaler=3 sono circa 7 secondi). Inoltre, togliamo qualche bit di precisione
130 * da destra (secondo quanto specificato in RAMP_CLOCK_SHIFT_PRECISION).
132 ASSERT(ramp->clocksRamp < (1UL << (24 + RAMP_CLOCK_SHIFT_PRECISION)));
133 ramp->precalc.inv_total_time = 0xFFFFFFFFUL / (ramp->clocksRamp >> RAMP_CLOCK_SHIFT_PRECISION);
134 ASSERT(ramp->precalc.inv_total_time < 0x1000000UL);
140 void ramp_setup(struct Ramp* ramp, uint32_t length, uint32_t minFreq, uint32_t maxFreq)
142 uint32_t minWL, maxWL;
144 minWL = TIME2CLOCKS(FREQ2MICROS(maxFreq));
145 maxWL = TIME2CLOCKS(FREQ2MICROS(minFreq));
147 ASSERT2(minWL < 65536UL, "Maximum frequency too high");
148 ASSERT2(maxWL < 65536UL, "Minimum frequency too high");
149 ASSERT(maxFreq > minFreq);
154 TIME2CLOCKS(FREQ2MICROS(maxFreq)),
155 TIME2CLOCKS(FREQ2MICROS(minFreq))
159 void ramp_default(struct Ramp *ramp)
161 ramp_setup(ramp, RAMP_DEF_TIME, RAMP_DEF_MINFREQ, RAMP_DEF_MAXFREQ);
164 #if RAMP_USE_FLOATING_POINT
166 float ramp_evaluate(const struct Ramp* ramp, float curClock)
168 return ramp->precalc.alpha / (curClock * ramp->precalc.gamma + ramp->precalc.beta);
173 INLINE uint32_t fix_mult32(uint32_t m1, uint32_t m2)
176 accum += m1 * ((m2 >> 0) & 0xFF);
178 accum += m1 * ((m2 >> 8) & 0xFF);
180 accum += m1 * ((m2 >> 16) & 0xFF);
185 INLINE uint16_t fix_mult16(uint16_t a, uint32_t b)
187 return (b*(uint32_t)a) >> 16;
190 uint16_t FAST_FUNC ramp_evaluate(const struct Ramp* ramp, uint32_t curClock)
192 uint16_t t = FIX_MULT32(curClock >> RAMP_CLOCK_SHIFT_PRECISION, ramp->precalc.inv_total_time);
193 uint16_t denom = fix_mult16((uint16_t)~t + 1, ramp->precalc.max_div_min) + t;
194 uint16_t cur_delta = ((uint32_t)ramp->clocksMinWL << 16) / denom;