#include "config.h" #include "bench.h" #include #define DG unsigned short #define ACC unsigned long #define REAL bench_real #define BITS_IN_REAL 53 /* mantissa */ #define SHFT 16 #define RADIX 65536L #define IRADIX (1.0 / RADIX) #define LO(x) ((x) & (RADIX - 1)) #define HI(x) ((x) >> SHFT) #define HI_SIGNED(x) \ ((((x) + (ACC)(RADIX >> 1) * RADIX) >> SHFT) - (RADIX >> 1)) #define ZEROEXP (-32768) #define LEN 10 typedef struct { short sign; short expt; DG d[LEN]; } N[1]; #define EXA a->expt #define EXB b->expt #define EXC c->expt #define AD a->d #define BD b->d #define SGNA a->sign #define SGNB b->sign static const N zero = {{ 1, ZEROEXP, {0} }}; static void cpy(const N a, N b) { *b = *a; } static void fromreal(REAL x, N a) { int i, e; cpy(zero, a); if (x == 0.0) return; if (x >= 0) { SGNA = 1; } else { SGNA = -1; x = -x; } e = 0; while (x >= 1.0) { x *= IRADIX; ++e; } while (x < IRADIX) { x *= RADIX; --e; } EXA = e; for (i = LEN - 1; i >= 0 && x != 0.0; --i) { REAL y; x *= RADIX; y = (REAL) ((int) x); AD[i] = (DG)y; x -= y; } } static void fromshort(int x, N a) { cpy(zero, a); if (x < 0) { x = -x; SGNA = -1; } else { SGNA = 1; } EXA = 1; AD[LEN - 1] = x; } static void pack(DG *d, int e, int s, int l, N a) { int i, j; for (i = l - 1; i >= 0; --i, --e) if (d[i] != 0) break; if (i < 0) { /* number is zero */ cpy(zero, a); } else { EXA = e; SGNA = s; if (i >= LEN - 1) { for (j = LEN - 1; j >= 0; --i, --j) AD[j] = d[i]; } else { for (j = LEN - 1; i >= 0; --i, --j) AD[j] = d[i]; for ( ; j >= 0; --j) AD[j] = 0; } } } /* compare absolute values */ static int abscmp(const N a, const N b) { int i; if (EXA > EXB) return 1; if (EXA < EXB) return -1; for (i = LEN - 1; i >= 0; --i) { if (AD[i] > BD[i]) return 1; if (AD[i] < BD[i]) return -1; } return 0; } static int eq(const N a, const N b) { return (SGNA == SGNB) && (abscmp(a, b) == 0); } /* add magnitudes, for |a| >= |b| */ static void addmag0(int s, const N a, const N b, N c) { int ia, ib; ACC r = 0; DG d[LEN + 1]; for (ia = 0, ib = EXA - EXB; ib < LEN; ++ia, ++ib) { r += (ACC)AD[ia] + (ACC)BD[ib]; d[ia] = LO(r); r = HI(r); } for (; ia < LEN; ++ia) { r += (ACC)AD[ia]; d[ia] = LO(r); r = HI(r); } d[ia] = LO(r); pack(d, EXA + 1, s * SGNA, LEN + 1, c); } static void addmag(int s, const N a, const N b, N c) { if (abscmp(a, b) > 0) addmag0(1, a, b, c); else addmag0(s, b, a, c); } /* subtract magnitudes, for |a| >= |b| */ static void submag0(int s, const N a, const N b, N c) { int ia, ib; ACC r = 0; DG d[LEN]; for (ia = 0, ib = EXA - EXB; ib < LEN; ++ia, ++ib) { r += (ACC)AD[ia] - (ACC)BD[ib]; d[ia] = LO(r); r = HI_SIGNED(r); } for (; ia < LEN; ++ia) { r += (ACC)AD[ia]; d[ia] = LO(r); r = HI_SIGNED(r); } pack(d, EXA, s * SGNA, LEN, c); } static void submag(int s, const N a, const N b, N c) { if (abscmp(a, b) > 0) submag0(1, a, b, c); else submag0(s, b, a, c); } /* c = a + b */ static void add(const N a, const N b, N c) { if (SGNA == SGNB) addmag(1, a, b, c); else submag(1, a, b, c); } static void sub(const N a, const N b, N c) { if (SGNA == SGNB) submag(-1, a, b, c); else addmag(-1, a, b, c); } static void mul(const N a, const N b, N c) { DG d[2 * LEN]; int i, j, k; ACC r; for (i = 0; i < LEN; ++i) d[2 * i] = d[2 * i + 1] = 0; for (i = 0; i < LEN; ++i) { ACC ai = AD[i]; if (ai) { r = 0; for (j = 0, k = i; j < LEN; ++j, ++k) { r += ai * (ACC)BD[j] + (ACC)d[k]; d[k] = LO(r); r = HI(r); } d[k] = LO(r); } } pack(d, EXA + EXB, SGNA * SGNB, 2 * LEN, c); } static REAL toreal(const N a) { REAL h, l, f; int i, bits; ACC r; DG sticky; if (EXA != ZEROEXP) { f = IRADIX; i = LEN; bits = 0; h = (r = AD[--i]) * f; f *= IRADIX; for (bits = 0; r > 0; ++bits) r >>= 1; /* first digit */ while (bits + SHFT <= BITS_IN_REAL) { h += AD[--i] * f; f *= IRADIX; bits += SHFT; } /* guard digit (leave one bit for sticky bit, hence `<' instead of `<=') */ bits = 0; l = 0.0; while (bits + SHFT < BITS_IN_REAL) { l += AD[--i] * f; f *= IRADIX; bits += SHFT; } /* sticky bit */ sticky = 0; while (i > 0) sticky |= AD[--i]; if (sticky) l += (RADIX / 2) * f; h += l; for (i = 0; i < EXA; ++i) h *= (REAL)RADIX; for (i = 0; i > EXA; --i) h *= IRADIX; if (SGNA == -1) h = -h; return h; } else { return 0.0; } } static void neg(N a) { SGNA = -SGNA; } static void inv(const N a, N x) { N w, z, one, two; fromreal(1.0 / toreal(a), x); /* initial guess */ fromshort(1, one); fromshort(2, two); for (;;) { /* Newton */ mul(a, x, w); sub(two, w, z); if (eq(one, z)) break; mul(x, z, x); } } /* 2 pi */ static const N n2pi = {{ 1, 1, {18450, 59017, 1760, 5212, 9779, 4518, 2886, 54545, 18558, 6} }}; /* 1 / 31! */ static const N i31fac = {{ 1, -7, {28087, 45433, 51357, 24545, 14291, 3954, 57879, 8109, 38716, 41382} }}; /* 1 / 32! */ static const N i32fac = {{ 1, -7, {52078, 60811, 3652, 39679, 37310, 47227, 28432, 57597, 13497, 1293} }}; static void msin(const N a, N b) { N a2, g, k; int i; cpy(i31fac, g); cpy(g, b); mul(a, a, a2); /* Taylor */ for (i = 31; i > 1; i -= 2) { fromshort(i * (i - 1), k); mul(k, g, g); mul(a2, b, k); sub(g, k, b); } mul(a, b, b); } static void mcos(const N a, N b) { N a2, g, k; int i; cpy(i32fac, g); cpy(g, b); mul(a, a, a2); /* Taylor */ for (i = 32; i > 0; i -= 2) { fromshort(i * (i - 1), k); mul(k, g, g); mul(a2, b, k); sub(g, k, b); } } static void by2pi(REAL m, REAL n, N a) { N b; fromreal(n, b); inv(b, a); fromreal(m, b); mul(a, b, a); mul(n2pi, a, a); } static void sin2pi(REAL m, REAL n, N a); static void cos2pi(REAL m, REAL n, N a) { N b; if (m < 0) cos2pi(-m, n, a); else if (m > n * 0.5) cos2pi(n - m, n, a); else if (m > n * 0.25) {sin2pi(m - n * 0.25, n, a); neg(a);} else if (m > n * 0.125) sin2pi(n * 0.25 - m, n, a); else { by2pi(m, n, b); mcos(b, a); } } static void sin2pi(REAL m, REAL n, N a) { N b; if (m < 0) {sin2pi(-m, n, a); neg(a);} else if (m > n * 0.5) {sin2pi(n - m, n, a); neg(a);} else if (m > n * 0.25) {cos2pi(m - n * 0.25, n, a);} else if (m > n * 0.125) {cos2pi(n * 0.25 - m, n, a);} else {by2pi(m, n, b); msin(b, a);} } /*----------------------------------------------------------------------*/ /* FFT stuff */ /* (r0 + i i0)(r1 + i i1) */ static void cmul(N r0, N i0, N r1, N i1, N r2, N i2) { N s, t, q; mul(r0, r1, s); mul(i0, i1, t); sub(s, t, q); mul(r0, i1, s); mul(i0, r1, t); add(s, t, i2); cpy(q, r2); } /* (r0 - i i0)(r1 + i i1) */ static void cmulj(N r0, N i0, N r1, N i1, N r2, N i2) { N s, t, q; mul(r0, r1, s); mul(i0, i1, t); add(s, t, q); mul(r0, i1, s); mul(i0, r1, t); sub(s, t, i2); cpy(q, r2); } static void mcexp(int m, int n, N r, N i) { static int cached_n = -1; static N w[64][2]; int k, j; if (n != cached_n) { for (j = 1, k = 0; j < n; j += j, ++k) { cos2pi(j, n, w[k][0]); sin2pi(j, n, w[k][1]); } cached_n = n; } fromshort(1, r); fromshort(0, i); if (m > 0) { for (k = 0; m; ++k, m >>= 1) if (m & 1) cmul(w[k][0], w[k][1], r, i, r, i); } else { m = -m; for (k = 0; m; ++k, m >>= 1) if (m & 1) cmulj(w[k][0], w[k][1], r, i, r, i); } } static void bitrev(int n, N *a) { int i, j, m; for (i = j = 0; i < n - 1; ++i) { if (i < j) { N t; cpy(a[2*i], t); cpy(a[2*j], a[2*i]); cpy(t, a[2*j]); cpy(a[2*i+1], t); cpy(a[2*j+1], a[2*i+1]); cpy(t, a[2*j+1]); } /* bit reversed counter */ m = n; do { m >>= 1; j ^= m; } while (!(j & m)); } } static void fft0(int n, N *a, int sign) { int i, j, k; bitrev(n, a); for (i = 1; i < n; i = 2 * i) { for (j = 0; j < i; ++j) { N wr, wi; mcexp(sign * (int)j, 2 * i, wr, wi); for (k = j; k < n; k += 2 * i) { N *a0 = a + 2 * k; N *a1 = a0 + 2 * i; N r0, i0, r1, i1, t0, t1, xr, xi; cpy(a0[0], r0); cpy(a0[1], i0); cpy(a1[0], r1); cpy(a1[1], i1); mul(r1, wr, t0); mul(i1, wi, t1); sub(t0, t1, xr); mul(r1, wi, t0); mul(i1, wr, t1); add(t0, t1, xi); add(r0, xr, a0[0]); add(i0, xi, a0[1]); sub(r0, xr, a1[0]); sub(i0, xi, a1[1]); } } } } /* a[2*k]+i*a[2*k+1] = exp(2*pi*i*k^2/(2*n)) */ static void bluestein_sequence(int n, N *a) { int k, ksq, n2 = 2 * n; ksq = 1; /* (-1)^2 */ for (k = 0; k < n; ++k) { /* careful with overflow */ ksq = ksq + 2*k - 1; while (ksq > n2) ksq -= n2; mcexp(ksq, n2, a[2*k], a[2*k+1]); } } static int pow2_atleast(int x) { int h; for (h = 1; h < x; h = 2 * h) ; return h; } static N *cached_bluestein_w = 0; static N *cached_bluestein_y = 0; static int cached_bluestein_n = -1; static void bluestein(int n, N *a) { int nb = pow2_atleast(2 * n); N *b = (N *)bench_malloc(2 * nb * sizeof(N)); N *w = cached_bluestein_w; N *y = cached_bluestein_y; N nbinv; int i; fromreal(1.0 / nb, nbinv); /* exact because nb = 2^k */ if (cached_bluestein_n != n) { if (w) bench_free(w); if (y) bench_free(y); w = (N *)bench_malloc(2 * n * sizeof(N)); y = (N *)bench_malloc(2 * nb * sizeof(N)); cached_bluestein_n = n; cached_bluestein_w = w; cached_bluestein_y = y; bluestein_sequence(n, w); for (i = 0; i < 2*nb; ++i) cpy(zero, y[i]); for (i = 0; i < n; ++i) { cpy(w[2*i], y[2*i]); cpy(w[2*i+1], y[2*i+1]); } for (i = 1; i < n; ++i) { cpy(w[2*i], y[2*(nb-i)]); cpy(w[2*i+1], y[2*(nb-i)+1]); } fft0(nb, y, -1); } for (i = 0; i < 2*nb; ++i) cpy(zero, b[i]); for (i = 0; i < n; ++i) cmulj(w[2*i], w[2*i+1], a[2*i], a[2*i+1], b[2*i], b[2*i+1]); /* scaled convolution b * y */ fft0(nb, b, -1); for (i = 0; i < nb; ++i) cmul(b[2*i], b[2*i+1], y[2*i], y[2*i+1], b[2*i], b[2*i+1]); fft0(nb, b, 1); for (i = 0; i < n; ++i) { cmulj(w[2*i], w[2*i+1], b[2*i], b[2*i+1], a[2*i], a[2*i+1]); mul(nbinv, a[2*i], a[2*i]); mul(nbinv, a[2*i+1], a[2*i+1]); } bench_free(b); } static void swapri(int n, N *a) { int i; for (i = 0; i < n; ++i) { N t; cpy(a[2 * i], t); cpy(a[2 * i + 1], a[2 * i]); cpy(t, a[2 * i + 1]); } } static void fft1(int n, N *a, int sign) { if (power_of_two(n)) { fft0(n, a, sign); } else { if (sign == 1) swapri(n, a); bluestein(n, a); if (sign == 1) swapri(n, a); } } static void fromrealv(int n, bench_complex *a, N *b) { int i; for (i = 0; i < n; ++i) { fromreal(c_re(a[i]), b[2 * i]); fromreal(c_im(a[i]), b[2 * i + 1]); } } static void compare(int n, N *a, N *b, double *err) { int i; double e1, e2, einf; double n1, n2, ninf; e1 = e2 = einf = 0.0; n1 = n2 = ninf = 0.0; # define DO(x1, x2, xinf, var) { \ double d = var; \ if (d < 0) d = -d; \ x1 += d; x2 += d * d; if (d > xinf) xinf = d; \ } for (i = 0; i < 2 * n; ++i) { N dd; sub(a[i], b[i], dd); DO(n1, n2, ninf, toreal(a[i])); DO(e1, e2, einf, toreal(dd)); } # undef DO err[0] = e1 / n1; err[1] = sqrt(e2 / n2); err[2] = einf / ninf; } void fftaccuracy(int n, bench_complex *a, bench_complex *ffta, int sign, double err[6]) { N *b = (N *)bench_malloc(2 * n * sizeof(N)); N *fftb = (N *)bench_malloc(2 * n * sizeof(N)); N mn, ninv; int i; fromreal(n, mn); inv(mn, ninv); /* forward error */ fromrealv(n, a, b); fromrealv(n, ffta, fftb); fft1(n, b, sign); compare(n, b, fftb, err); /* backward error */ fromrealv(n, a, b); fromrealv(n, ffta, fftb); for (i = 0; i < 2 * n; ++i) mul(fftb[i], ninv, fftb[i]); fft1(n, fftb, -sign); compare(n, b, fftb, err + 3); bench_free(fftb); bench_free(b); } void fftaccuracy_done(void) { if (cached_bluestein_w) bench_free(cached_bluestein_w); if (cached_bluestein_y) bench_free(cached_bluestein_y); cached_bluestein_w = 0; cached_bluestein_y = 0; cached_bluestein_n = -1; }