fftsg_h.c

This is a package to calculate Discrete Fourier/Cosine/Sine Transforms of 1-dimensional sequences of

#endif
#include <pthread.h>
#include <stdio.h>
#include <stdlib.h>
#define cdft_thread_t pthread_t
#define cdft_thread_create(thp,func,argp) { \
    if (pthread_create(thp, NULL, func, (void *) argp) != 0) { \
        fprintf(stderr, "cdft thread error\n"); \
        exit(1); \
    } \
}
#define cdft_thread_wait(th) { \
    if (pthread_join(th, NULL) != 0) { \
        fprintf(stderr, "cdft thread error\n"); \
        exit(1); \
    } \
}
#endif /* USE_CDFT_PTHREADS */


#ifdef USE_CDFT_WINTHREADS
#define USE_CDFT_THREADS
#ifndef CDFT_THREADS_BEGIN_N
#define CDFT_THREADS_BEGIN_N 32768
#endif
#ifndef CDFT_4THREADS_BEGIN_N
#define CDFT_4THREADS_BEGIN_N 524288
#endif
#include <windows.h>
#include <stdio.h>
#include <stdlib.h>
#define cdft_thread_t HANDLE
#define cdft_thread_create(thp,func,argp) { \
    DWORD thid; \
    *(thp) = CreateThread(NULL, 0, (LPTHREAD_START_ROUTINE) func, (LPVOID) argp, 0, &thid); \
    if (*(thp) == 0) { \
        fprintf(stderr, "cdft thread error\n"); \
        exit(1); \
    } \
}
#define cdft_thread_wait(th) { \
    WaitForSingleObject(th, INFINITE); \
    CloseHandle(th); \
}
#endif /* USE_CDFT_WINTHREADS */


#ifndef CDFT_LOOP_DIV  /* control of the CDFT's speed & tolerance */
#define CDFT_LOOP_DIV 32
#endif

#ifndef RDFT_LOOP_DIV  /* control of the RDFT's speed & tolerance */
#define RDFT_LOOP_DIV 64
#endif

#ifndef DCST_LOOP_DIV  /* control of the DCT,DST's speed & tolerance */
#define DCST_LOOP_DIV 64
#endif


void cftfsub(int n, double *a)
{
    void bitrv2(int n, double *a);
    void bitrv216(double *a);
    void bitrv208(double *a);
    void cftmdl1(int n, double *a);
    void cftrec4(int n, double *a);
    void cftleaf(int n, int isplt, double *a);
    void cftfx41(int n, double *a);
    void cftf161(double *a);
    void cftf081(double *a);
    void cftf040(double *a);
    void cftx020(double *a);
#ifdef USE_CDFT_THREADS
    void cftrec4_th(int n, double *a);
#endif /* USE_CDFT_THREADS */
    
    if (n > 8) {
        if (n > 32) {
            cftmdl1(n, a);
#ifdef USE_CDFT_THREADS
            if (n > CDFT_THREADS_BEGIN_N) {
                cftrec4_th(n, a);
            } else 
#endif /* USE_CDFT_THREADS */
            if (n > 512) {
                cftrec4(n, a);
            } else if (n > 128) {
                cftleaf(n, 1, a);
            } else {
                cftfx41(n, a);
            }
            bitrv2(n, a);
        } else if (n == 32) {
            cftf161(a);
            bitrv216(a);
        } else {
            cftf081(a);
            bitrv208(a);
        }
    } else if (n == 8) {
        cftf040(a);
    } else if (n == 4) {
        cftx020(a);
    }
}


void cftbsub(int n, double *a)
{
    void bitrv2conj(int n, double *a);
    void bitrv216neg(double *a);
    void bitrv208neg(double *a);
    void cftb1st(int n, double *a);
    void cftrec4(int n, double *a);
    void cftleaf(int n, int isplt, double *a);
    void cftfx41(int n, double *a);
    void cftf161(double *a);
    void cftf081(double *a);
    void cftb040(double *a);
    void cftx020(double *a);
#ifdef USE_CDFT_THREADS
    void cftrec4_th(int n, double *a);
#endif /* USE_CDFT_THREADS */
    
    if (n > 8) {
        if (n > 32) {
            cftb1st(n, a);
#ifdef USE_CDFT_THREADS
            if (n > CDFT_THREADS_BEGIN_N) {
                cftrec4_th(n, a);
            } else 
#endif /* USE_CDFT_THREADS */
            if (n > 512) {
                cftrec4(n, a);
            } else if (n > 128) {
                cftleaf(n, 1, a);
            } else {
                cftfx41(n, a);
            }
            bitrv2conj(n, a);
        } else if (n == 32) {
            cftf161(a);
            bitrv216neg(a);
        } else {
            cftf081(a);
            bitrv208neg(a);
        }
    } else if (n == 8) {
        cftb040(a);
    } else if (n == 4) {
        cftx020(a);
    }
}


void bitrv2(int n, double *a)
{
    int j0, k0, j1, k1, l, m, i, j, k, nh;
    double xr, xi, yr, yi;
    
    m = 4;
    for (l = n >> 2; l > 8; l >>= 2) {
        m <<= 1;
    }
    nh = n >> 1;
    if (l == 8) {
        j0 = 0;
        for (k0 = 0; k0 < m; k0 += 4) {
            k = k0;
            for (j = j0; j < j0 + k0; j += 4) {
                xr = a[j];
                xi = a[j + 1];
                yr = a[k];
                yi = a[k + 1];
                a[j] = yr;
                a[j + 1] = yi;
                a[k] = xr;
                a[k + 1] = xi;
                j1 = j + m;
                k1 = k + 2 * m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 += m;
                k1 -= m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 += m;
                k1 += 2 * m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 += nh;
                k1 += 2;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 -= m;
                k1 -= 2 * m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 -= m;
                k1 += m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 -= m;
                k1 -= 2 * m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 += 2;
                k1 += nh;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 += m;
                k1 += 2 * m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 += m;
                k1 -= m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 += m;
                k1 += 2 * m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 -= nh;
                k1 -= 2;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 -= m;
                k1 -= 2 * m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 -= m;
                k1 += m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 -= m;
                k1 -= 2 * m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                for (i = nh >> 1; i > (k ^= i); i >>= 1);
            }
            k1 = j0 + k0;
            j1 = k1 + 2;
            k1 += nh;
            xr = a[j1];
            xi = a[j1 + 1];
            yr = a[k1];
            yi = a[k1 + 1];
            a[j1] = yr;
            a[j1 + 1] = yi;
            a[k1] = xr;
            a[k1 + 1] = xi;
            j1 += m;
            k1 += 2 * m;
            xr = a[j1];
            xi = a[j1 + 1];
            yr = a[k1];
            yi = a[k1 + 1];
            a[j1] = yr;
            a[j1 + 1] = yi;
            a[k1] = xr;
            a[k1 + 1] = xi;
            j1 += m;
            k1 -= m;
            xr = a[j1];
            xi = a[j1 + 1];
            yr = a[k1];
            yi = a[k1 + 1];
            a[j1] = yr;
            a[j1 + 1] = yi;
            a[k1] = xr;
            a[k1 + 1] = xi;
            j1 -= 2;
            k1 -= nh;
            xr = a[j1];
            xi = a[j1 + 1];
            yr = a[k1];
            yi = a[k1 + 1];
            a[j1] = yr;
            a[j1 + 1] = yi;
            a[k1] = xr;
            a[k1 + 1] = xi;
            j1 += nh + 2;
            k1 += nh + 2;
            xr = a[j1];
            xi = a[j1 + 1];
            yr = a[k1];
            yi = a[k1 + 1];
            a[j1] = yr;
            a[j1 + 1] = yi;
            a[k1] = xr;
            a[k1 + 1] = xi;
            j1 -= nh - m;
            k1 += 2 * m - 2;
            xr = a[j1];
            xi = a[j1 + 1];
            yr = a[k1];
            yi = a[k1 + 1];
            a[j1] = yr;
            a[j1 + 1] = yi;
            a[k1] = xr;
            a[k1 + 1] = xi;
            for (i = nh >> 1; i > (j0 ^= i); i >>= 1);
        }
    } else {
        j0 = 0;
        for (k0 = 0; k0 < m; k0 += 4) {
            k = k0;
            for (j = j0; j < j0 + k0; j += 4) {
                xr = a[j];
                xi = a[j + 1];
                yr = a[k];
                yi = a[k + 1];
                a[j] = yr;
                a[j + 1] = yi;
                a[k] = xr;
                a[k + 1] = xi;
                j1 = j + m;
                k1 = k + m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 += nh;
                k1 += 2;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 -= m;
                k1 -= m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 += 2;
                k1 += nh;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 += m;
                k1 += m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 -= nh;
                k1 -= 2;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                j1 -= m;
                k1 -= m;
                xr = a[j1];
                xi = a[j1 + 1];
                yr = a[k1];
                yi = a[k1 + 1];
                a[j1] = yr;
                a[j1 + 1] = yi;
                a[k1] = xr;
                a[k1 + 1] = xi;
                for (i = nh >> 1; i > (k ^= i); i >>= 1);
            }
            k1 = j0 + k0;
            j1 = k1 + 2;
            k1 += nh;
            xr = a[j1];
            xi = a[j1 + 1];
            yr = a[k1];
            yi = a[k1 + 1];
            a[j1] = yr;
            a[j1 + 1] = yi;
            a[k1] = xr;
            a[k1 + 1] = xi;
            j1 += m;
            k1 += m;
            xr = a[j1];
            xi = a[j1 + 1];