cannam@167: /* cannam@167: * Copyright (c) 2003, 2007-14 Matteo Frigo cannam@167: * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology cannam@167: * cannam@167: * This program is free software; you can redistribute it and/or modify cannam@167: * it under the terms of the GNU General Public License as published by cannam@167: * the Free Software Foundation; either version 2 of the License, or cannam@167: * (at your option) any later version. cannam@167: * cannam@167: * This program is distributed in the hope that it will be useful, cannam@167: * but WITHOUT ANY WARRANTY; without even the implied warranty of cannam@167: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the cannam@167: * GNU General Public License for more details. cannam@167: * cannam@167: * You should have received a copy of the GNU General Public License cannam@167: * along with this program; if not, write to the Free Software cannam@167: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA cannam@167: * cannam@167: */ cannam@167: cannam@167: /* express a hc2hc problem in terms of rdft + multiplication by cannam@167: twiddle factors */ cannam@167: cannam@167: #include "rdft/hc2hc.h" cannam@167: cannam@167: typedef hc2hc_solver S; cannam@167: cannam@167: typedef struct { cannam@167: plan_hc2hc super; cannam@167: cannam@167: INT r, m, s, vl, vs, mstart1, mcount1; cannam@167: plan *cld0; cannam@167: plan *cld; cannam@167: twid *td; cannam@167: } P; cannam@167: cannam@167: cannam@167: /**************************************************************/ cannam@167: static void mktwiddle(P *ego, enum wakefulness wakefulness) cannam@167: { cannam@167: static const tw_instr tw[] = { { TW_HALF, 0, 0 }, { TW_NEXT, 1, 0 } }; cannam@167: cannam@167: /* note that R and M are swapped, to allow for sequential cannam@167: access both to data and twiddles */ cannam@167: X(twiddle_awake)(wakefulness, &ego->td, tw, cannam@167: ego->r * ego->m, ego->m, ego->r); cannam@167: } cannam@167: cannam@167: static void bytwiddle(const P *ego, R *IO, R sign) cannam@167: { cannam@167: INT i, j, k; cannam@167: INT r = ego->r, m = ego->m, s = ego->s, vl = ego->vl, vs = ego->vs; cannam@167: INT ms = m * s; cannam@167: INT mstart1 = ego->mstart1, mcount1 = ego->mcount1; cannam@167: INT wrem = 2 * ((m-1)/2 - mcount1); cannam@167: cannam@167: for (i = 0; i < vl; ++i, IO += vs) { cannam@167: const R *W = ego->td->W; cannam@167: cannam@167: A(m % 2 == 1); cannam@167: for (k = 1, W += (m - 1) + 2*(mstart1-1); k < r; ++k) { cannam@167: /* pr := IO + (j + mstart1) * s + k * ms */ cannam@167: R *pr = IO + mstart1 * s + k * ms; cannam@167: cannam@167: /* pi := IO + (m - j - mstart1) * s + k * ms */ cannam@167: R *pi = IO - mstart1 * s + (k + 1) * ms; cannam@167: cannam@167: for (j = 0; j < mcount1; ++j, pr += s, pi -= s) { cannam@167: E xr = *pr; cannam@167: E xi = *pi; cannam@167: E wr = W[0]; cannam@167: E wi = sign * W[1]; cannam@167: *pr = xr * wr - xi * wi; cannam@167: *pi = xi * wr + xr * wi; cannam@167: W += 2; cannam@167: } cannam@167: W += wrem; cannam@167: } cannam@167: } cannam@167: } cannam@167: cannam@167: static void swapri(R *IO, INT r, INT m, INT s, INT jstart, INT jend) cannam@167: { cannam@167: INT k; cannam@167: INT ms = m * s; cannam@167: INT js = jstart * s; cannam@167: for (k = 0; k + k < r; ++k) { cannam@167: /* pr := IO + (m - j) * s + k * ms */ cannam@167: R *pr = IO + (k + 1) * ms - js; cannam@167: /* pi := IO + (m - j) * s + (r - 1 - k) * ms */ cannam@167: R *pi = IO + (r - k) * ms - js; cannam@167: INT j; cannam@167: for (j = jstart; j < jend; j += 1, pr -= s, pi -= s) { cannam@167: R t = *pr; cannam@167: *pr = *pi; cannam@167: *pi = t; cannam@167: } cannam@167: } cannam@167: } cannam@167: cannam@167: static void reorder_dit(const P *ego, R *IO) cannam@167: { cannam@167: INT i, k; cannam@167: INT r = ego->r, m = ego->m, s = ego->s, vl = ego->vl, vs = ego->vs; cannam@167: INT ms = m * s; cannam@167: INT mstart1 = ego->mstart1, mend1 = mstart1 + ego->mcount1; cannam@167: cannam@167: for (i = 0; i < vl; ++i, IO += vs) { cannam@167: for (k = 1; k + k < r; ++k) { cannam@167: R *p0 = IO + k * ms; cannam@167: R *p1 = IO + (r - k) * ms; cannam@167: INT j; cannam@167: cannam@167: for (j = mstart1; j < mend1; ++j) { cannam@167: E rp, ip, im, rm; cannam@167: rp = p0[j * s]; cannam@167: im = p1[ms - j * s]; cannam@167: rm = p1[j * s]; cannam@167: ip = p0[ms - j * s]; cannam@167: p0[j * s] = rp - im; cannam@167: p1[ms - j * s] = rp + im; cannam@167: p1[j * s] = rm - ip; cannam@167: p0[ms - j * s] = ip + rm; cannam@167: } cannam@167: } cannam@167: cannam@167: swapri(IO, r, m, s, mstart1, mend1); cannam@167: } cannam@167: } cannam@167: cannam@167: static void reorder_dif(const P *ego, R *IO) cannam@167: { cannam@167: INT i, k; cannam@167: INT r = ego->r, m = ego->m, s = ego->s, vl = ego->vl, vs = ego->vs; cannam@167: INT ms = m * s; cannam@167: INT mstart1 = ego->mstart1, mend1 = mstart1 + ego->mcount1; cannam@167: cannam@167: for (i = 0; i < vl; ++i, IO += vs) { cannam@167: swapri(IO, r, m, s, mstart1, mend1); cannam@167: cannam@167: for (k = 1; k + k < r; ++k) { cannam@167: R *p0 = IO + k * ms; cannam@167: R *p1 = IO + (r - k) * ms; cannam@167: const R half = K(0.5); cannam@167: INT j; cannam@167: cannam@167: for (j = mstart1; j < mend1; ++j) { cannam@167: E rp, ip, im, rm; cannam@167: rp = half * p0[j * s]; cannam@167: im = half * p1[ms - j * s]; cannam@167: rm = half * p1[j * s]; cannam@167: ip = half * p0[ms - j * s]; cannam@167: p0[j * s] = rp + im; cannam@167: p1[ms - j * s] = im - rp; cannam@167: p1[j * s] = rm + ip; cannam@167: p0[ms - j * s] = ip - rm; cannam@167: } cannam@167: } cannam@167: } cannam@167: } cannam@167: cannam@167: static int applicable(rdft_kind kind, INT r, INT m, const planner *plnr) cannam@167: { cannam@167: return (1 cannam@167: && (kind == R2HC || kind == HC2R) cannam@167: && (m % 2) cannam@167: && (r % 2) cannam@167: && !NO_SLOWP(plnr) cannam@167: ); cannam@167: } cannam@167: cannam@167: /**************************************************************/ cannam@167: cannam@167: static void apply_dit(const plan *ego_, R *IO) cannam@167: { cannam@167: const P *ego = (const P *) ego_; cannam@167: INT start; cannam@167: plan_rdft *cld, *cld0; cannam@167: cannam@167: bytwiddle(ego, IO, K(-1.0)); cannam@167: cannam@167: cld0 = (plan_rdft *) ego->cld0; cannam@167: cld0->apply(ego->cld0, IO, IO); cannam@167: cannam@167: start = ego->mstart1 * ego->s; cannam@167: cld = (plan_rdft *) ego->cld; cannam@167: cld->apply(ego->cld, IO + start, IO + start); cannam@167: cannam@167: reorder_dit(ego, IO); cannam@167: } cannam@167: cannam@167: static void apply_dif(const plan *ego_, R *IO) cannam@167: { cannam@167: const P *ego = (const P *) ego_; cannam@167: INT start; cannam@167: plan_rdft *cld, *cld0; cannam@167: cannam@167: reorder_dif(ego, IO); cannam@167: cannam@167: cld0 = (plan_rdft *) ego->cld0; cannam@167: cld0->apply(ego->cld0, IO, IO); cannam@167: cannam@167: start = ego->mstart1 * ego->s; cannam@167: cld = (plan_rdft *) ego->cld; cannam@167: cld->apply(ego->cld, IO + start, IO + start); cannam@167: cannam@167: bytwiddle(ego, IO, K(1.0)); cannam@167: } cannam@167: cannam@167: cannam@167: static void awake(plan *ego_, enum wakefulness wakefulness) cannam@167: { cannam@167: P *ego = (P *) ego_; cannam@167: X(plan_awake)(ego->cld0, wakefulness); cannam@167: X(plan_awake)(ego->cld, wakefulness); cannam@167: mktwiddle(ego, wakefulness); cannam@167: } cannam@167: cannam@167: static void destroy(plan *ego_) cannam@167: { cannam@167: P *ego = (P *) ego_; cannam@167: X(plan_destroy_internal)(ego->cld); cannam@167: X(plan_destroy_internal)(ego->cld0); cannam@167: } cannam@167: cannam@167: static void print(const plan *ego_, printer *p) cannam@167: { cannam@167: const P *ego = (const P *) ego_; cannam@167: p->print(p, "(hc2hc-generic-%s-%D-%D%v%(%p%)%(%p%))", cannam@167: ego->super.apply == apply_dit ? "dit" : "dif", cannam@167: ego->r, ego->m, ego->vl, ego->cld0, ego->cld); cannam@167: } cannam@167: cannam@167: static plan *mkcldw(const hc2hc_solver *ego_, cannam@167: rdft_kind kind, INT r, INT m, INT s, INT vl, INT vs, cannam@167: INT mstart, INT mcount, cannam@167: R *IO, planner *plnr) cannam@167: { cannam@167: P *pln; cannam@167: plan *cld0 = 0, *cld = 0; cannam@167: INT mstart1, mcount1, mstride; cannam@167: cannam@167: static const plan_adt padt = { cannam@167: 0, awake, print, destroy cannam@167: }; cannam@167: cannam@167: UNUSED(ego_); cannam@167: cannam@167: A(mstart >= 0 && mcount > 0 && mstart + mcount <= (m+2)/2); cannam@167: cannam@167: if (!applicable(kind, r, m, plnr)) cannam@167: return (plan *)0; cannam@167: cannam@167: A(m % 2); cannam@167: mstart1 = mstart + (mstart == 0); cannam@167: mcount1 = mcount - (mstart == 0); cannam@167: mstride = m - (mstart + mcount - 1) - mstart1; cannam@167: cannam@167: /* 0th (DC) transform (vl of these), if mstart == 0 */ cannam@167: cld0 = X(mkplan_d)(plnr, cannam@167: X(mkproblem_rdft_1_d)( cannam@167: mstart == 0 ? X(mktensor_1d)(r, m * s, m * s) cannam@167: : X(mktensor_0d)(), cannam@167: X(mktensor_1d)(vl, vs, vs), cannam@167: IO, IO, kind) cannam@167: ); cannam@167: if (!cld0) goto nada; cannam@167: cannam@167: /* twiddle transforms: there are 2 x mcount1 x vl of these cannam@167: (where 2 corresponds to the real and imaginary parts) ... cannam@167: the 2 x mcount1 loops are combined if mstart=0 and mcount=(m+2)/2. */ cannam@167: cld = X(mkplan_d)(plnr, cannam@167: X(mkproblem_rdft_1_d)( cannam@167: X(mktensor_1d)(r, m * s, m * s), cannam@167: X(mktensor_3d)(2, mstride * s, mstride * s, cannam@167: mcount1, s, s, cannam@167: vl, vs, vs), cannam@167: IO + s * mstart1, IO + s * mstart1, kind) cannam@167: ); cannam@167: if (!cld) goto nada; cannam@167: cannam@167: pln = MKPLAN_HC2HC(P, &padt, (kind == R2HC) ? apply_dit : apply_dif); cannam@167: pln->cld = cld; cannam@167: pln->cld0 = cld0; cannam@167: pln->r = r; cannam@167: pln->m = m; cannam@167: pln->s = s; cannam@167: pln->vl = vl; cannam@167: pln->vs = vs; cannam@167: pln->td = 0; cannam@167: pln->mstart1 = mstart1; cannam@167: pln->mcount1 = mcount1; cannam@167: cannam@167: { cannam@167: double n0 = 0.5 * (r - 1) * (2 * mcount1) * vl; cannam@167: pln->super.super.ops = cld->ops; cannam@167: pln->super.super.ops.mul += (kind == R2HC ? 5.0 : 7.0) * n0; cannam@167: pln->super.super.ops.add += 4.0 * n0; cannam@167: pln->super.super.ops.other += 11.0 * n0; cannam@167: } cannam@167: return &(pln->super.super); cannam@167: cannam@167: nada: cannam@167: X(plan_destroy_internal)(cld); cannam@167: X(plan_destroy_internal)(cld0); cannam@167: return (plan *) 0; cannam@167: } cannam@167: cannam@167: static void regsolver(planner *plnr, INT r) cannam@167: { cannam@167: S *slv = (S *)X(mksolver_hc2hc)(sizeof(S), r, mkcldw); cannam@167: REGISTER_SOLVER(plnr, &(slv->super)); cannam@167: if (X(mksolver_hc2hc_hook)) { cannam@167: slv = (S *)X(mksolver_hc2hc_hook)(sizeof(S), r, mkcldw); cannam@167: REGISTER_SOLVER(plnr, &(slv->super)); cannam@167: } cannam@167: } cannam@167: cannam@167: void X(hc2hc_generic_register)(planner *p) cannam@167: { cannam@167: regsolver(p, 0); cannam@167: }