Rack/plugins/community/repos/squinkylabs-plug1/dsp/utils/LookupTable.h

#pragma once

#include "AudioMath.h"

#include <algorithm>
#include <assert.h>
#include <cmath>
#include <memory>
#include <emmintrin.h>
#include <functional>

template <typename T> class LookupTableParams;
/* Lookup table with evenly spaced lookup "bins"
 * Uses linear interpolation
 */

 // TODO: templatize on size?
template <typename T>
class LookupTable
{

public:
    LookupTable() = delete;       // we are only static

    /**
     * lookup does the table lookup.
     * input must be in the range specified at table creation time, otherwise
     * it will be limited to the range.
     * If allowOutsideDomain, will assert when input must be limited.
     */
    static T lookup(const LookupTableParams<T>& params, T input, bool allowOutsideDomain = false);

    /**
     * init will create the entries in the lookup table
     * bins is the number of entries desired in the lookup table.
     *      more bins means greater accuracy, but greater memory usage also.
     * xMin is the minimum input that will be passed to lookup()
     * xMax is the maximum input that will be passed to lookup().
     *      xMin..xMax is the domain of the function.
     * f is a continuous function that the lookup table will approximate.
     */
    static void init(LookupTableParams<T>& params, int bins, T xMin, T xMax, std::function<double(double)> f);

    /**
     * initDiscrete will make a table that interpolates between discrete values.
     * numEntries is the number of "points" that will be interpolated.
     * entries are the discrete y values to be interpolated.
     * Very important: x values are assumed to be 0..numEntries-1. That's because
     * this lookup table only works with uniform x value.
     */
    static void initDiscrete(LookupTableParams<T>& params, int numEntries, const T * yEntries);
private:
    static int cvtt(T *);

#ifdef _DEBUG
    static void checkInput(const LookupTableParams<T>& params, const T *in, int sampleFrames)
    {
        for (int i = 0; i < sampleFrames; ++i) {
            f_t input = in[i];
            assert(input >= params.xMin && input <= params.xMax);
        }
    }
#else
#define checkInput __noop
#endif

};

template<typename T>
inline T LookupTable<T>::lookup(const LookupTableParams<T>& params, T input, bool allowOutsideDomain)
{
    assert(allowOutsideDomain || (input >= params.xMin && input <= params.xMax));
                                                            // won't happen in the field,
                                                            // as assertions are disabled for release.
    input = std::min(input, params.xMax);
    input = std::max(input, params.xMin);
    assert(params.isValid());
    assert(input >= params.xMin && input <= params.xMax);

    // need to scale by bins
    T scaledInput = input * params.a + params.b;
    assert(params.a != 0);

    int input_int = cvtt(&scaledInput);
    T input_float = scaledInput - input_int;

    // Unfortunately, when we use float instead of doubles,
    // numeric precision issues get us "out of range". So
    // here we clamp to range. It would be more efficient if we didn't do this.
    // Perhaps the calculation of a and b could be done so this can't happen?
    if (input_float < 0) {
        input_float = 0;
    } else if (input_float > 1) {
        input_float = 1;
    }

    assert(input_float >= 0 && input_float <= 1);
    assert(input_int >= 0 && input_int <= params.numBins_i);

    T * entry = params.entries + (2 * input_int);
    T x = entry[0];
    x += input_float * entry[1];
    return x;
}

template<typename T>
inline void LookupTable<T>::init(LookupTableParams<T>& params,
    int bins, T x0In, T x1In, std::function<double(double)> f)
{
    params.alloc(bins);

    // f(x0 = ax + 0 == index
    // f(x0) = 0
    // f(x1) = bins
    params.a = (T) bins / (x1In - x0In);
    params.b = -params.a * x0In;

    if (x0In == 0) assert(params.b == 0);

    {
        assert(AudioMath::closeTo((params.a * x0In + params.b), 0, .0001));
        assert(AudioMath::closeTo((params.a * x1In + params.b), bins, .0001));
    }

    for (int i = 0; i <= bins; ++i) {
        double x0 = (i - params.b) / params.a;
        double x1 = ((i + 1) - params.b) / params.a;

        double y0 = f(x0);
        double y1 = f(x1);
        double slope = y1 - y0;
        T * entry = params.entries + (2 * i);
        entry[0] = (T) y0;
        entry[1] = (T) slope;
    }
    params.xMin = x0In;
    params.xMax = x1In;
}

template<typename T>
inline void LookupTable<T>::initDiscrete(LookupTableParams<T>& params, int numEntries, const T * entries)
{
    params.alloc(numEntries);
    // Since this version assumes x = 0, 1, 2 ....
    // We don't need an interpolation to find which bin we are in
    params.a = 1;
    params.b = 0;

    for (int i = 0; i < numEntries; ++i) {
        int x0 = i;
        int x1 = i + 1;

        // Ugh - this will need to get resolved when we "officially" make
        // all table support x values outside their range. But for now we 
        // have a problem: if x == Xmax (numEntries), we need an extra "bin"
        // at the end to look up that value, so here we make a flat bit of xMax.
        if (i == (numEntries - 1)) {
            --x1;
        }

        double y0 = entries[x0];
        double y1 = entries[x1];
        double slope = y1 - y0;

        T * entry = params.entries + (2 * i);
        entry[0] = (T) y0;
        entry[1] = (T) slope;
    }
    params.xMin = 0;
    params.xMax = T(numEntries - 1);
}

template<>
inline int LookupTable<double>::cvtt(double* input)
{
    auto x = _mm_load_sd(input);
    return _mm_cvttsd_si32(x);
}

template<>
inline int LookupTable<float>::cvtt(float* input)
{
    auto x = _mm_load_ss(input);
    return _mm_cvttss_si32(x);
}

/***************************************************************************/
extern int  _numLookupParams;

template <typename T>
class LookupTableParams
{
public:
    int numBins_i = 0;		// numBins will be number of entries (pairs of values)
    T a = 0, b = 0;			// lookup index = a * x + b, so domain of x can be whatever we want

    T * entries = 0;		// each entry is value, slope
    T xMin = 0;				// minimum x value we will accept as input
    T xMax = 0;				// max x value we will accept as input

    LookupTableParams()
    {
        ++_numLookupParams;
    }
    LookupTableParams(const LookupTableParams&) = delete;
    LookupTableParams& operator=(const LookupTableParams&) = delete;

    ~LookupTableParams()
    {
        free(entries);
        --_numLookupParams;
    }

    bool isValid() const
    {
        return ((entries != 0) &&
            (xMax > xMin) &&
            (numBins_i > 0)
            );
    }

    void alloc(int bins)
    {
        if (entries) free(entries);
        // allocate one extra, so we can index all the way to the end...
        entries = (T *) malloc((bins + 1) * 2 * sizeof(T));
        numBins_i = bins;
        a = 0;
        b = 0;
    }

};