cannam@127: /* cannam@127: * Copyright (c) 2003, 2007-14 Matteo Frigo cannam@127: * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology cannam@127: * cannam@127: * This program is free software; you can redistribute it and/or modify cannam@127: * it under the terms of the GNU General Public License as published by cannam@127: * the Free Software Foundation; either version 2 of the License, or cannam@127: * (at your option) any later version. cannam@127: * cannam@127: * This program is distributed in the hope that it will be useful, cannam@127: * but WITHOUT ANY WARRANTY; without even the implied warranty of cannam@127: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the cannam@127: * GNU General Public License for more details. cannam@127: * cannam@127: * You should have received a copy of the GNU General Public License cannam@127: * along with this program; if not, write to the Free Software cannam@127: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA cannam@127: * cannam@127: */ cannam@127: cannam@127: cannam@127: #include "ct.h" cannam@127: cannam@127: typedef struct { cannam@127: ct_solver super; cannam@127: const ct_desc *desc; cannam@127: int bufferedp; cannam@127: kdftw k; cannam@127: } S; cannam@127: cannam@127: typedef struct { cannam@127: plan_dftw super; cannam@127: kdftw k; cannam@127: INT r; cannam@127: stride rs; cannam@127: INT m, ms, v, vs, mb, me, extra_iter; cannam@127: stride brs; cannam@127: twid *td; cannam@127: const S *slv; cannam@127: } P; cannam@127: cannam@127: cannam@127: /************************************************************* cannam@127: Nonbuffered code cannam@127: *************************************************************/ cannam@127: static void apply(const plan *ego_, R *rio, R *iio) cannam@127: { cannam@127: const P *ego = (const P *) ego_; cannam@127: INT i; cannam@127: ASSERT_ALIGNED_DOUBLE; cannam@127: for (i = 0; i < ego->v; ++i, rio += ego->vs, iio += ego->vs) { cannam@127: INT mb = ego->mb, ms = ego->ms; cannam@127: ego->k(rio + mb*ms, iio + mb*ms, ego->td->W, cannam@127: ego->rs, mb, ego->me, ms); cannam@127: } cannam@127: } cannam@127: cannam@127: static void apply_extra_iter(const plan *ego_, R *rio, R *iio) cannam@127: { cannam@127: const P *ego = (const P *) ego_; cannam@127: INT i, v = ego->v, vs = ego->vs; cannam@127: INT mb = ego->mb, me = ego->me, mm = me - 1, ms = ego->ms; cannam@127: ASSERT_ALIGNED_DOUBLE; cannam@127: for (i = 0; i < v; ++i, rio += vs, iio += vs) { cannam@127: ego->k(rio + mb*ms, iio + mb*ms, ego->td->W, cannam@127: ego->rs, mb, mm, ms); cannam@127: ego->k(rio + mm*ms, iio + mm*ms, ego->td->W, cannam@127: ego->rs, mm, mm+2, 0); cannam@127: } cannam@127: } cannam@127: cannam@127: /************************************************************* cannam@127: Buffered code cannam@127: *************************************************************/ cannam@127: static void dobatch(const P *ego, R *rA, R *iA, INT mb, INT me, R *buf) cannam@127: { cannam@127: INT brs = WS(ego->brs, 1); cannam@127: INT rs = WS(ego->rs, 1); cannam@127: INT ms = ego->ms; cannam@127: cannam@127: X(cpy2d_pair_ci)(rA + mb*ms, iA + mb*ms, buf, buf + 1, cannam@127: ego->r, rs, brs, cannam@127: me - mb, ms, 2); cannam@127: ego->k(buf, buf + 1, ego->td->W, ego->brs, mb, me, 2); cannam@127: X(cpy2d_pair_co)(buf, buf + 1, rA + mb*ms, iA + mb*ms, cannam@127: ego->r, brs, rs, cannam@127: me - mb, 2, ms); cannam@127: } cannam@127: cannam@127: /* must be even for SIMD alignment; should not be 2^k to avoid cannam@127: associativity conflicts */ cannam@127: static INT compute_batchsize(INT radix) cannam@127: { cannam@127: /* round up to multiple of 4 */ cannam@127: radix += 3; cannam@127: radix &= -4; cannam@127: cannam@127: return (radix + 2); cannam@127: } cannam@127: cannam@127: static void apply_buf(const plan *ego_, R *rio, R *iio) cannam@127: { cannam@127: const P *ego = (const P *) ego_; cannam@127: INT i, j, v = ego->v, r = ego->r; cannam@127: INT batchsz = compute_batchsize(r); cannam@127: R *buf; cannam@127: INT mb = ego->mb, me = ego->me; cannam@127: size_t bufsz = r * batchsz * 2 * sizeof(R); cannam@127: cannam@127: BUF_ALLOC(R *, buf, bufsz); cannam@127: cannam@127: for (i = 0; i < v; ++i, rio += ego->vs, iio += ego->vs) { cannam@127: for (j = mb; j + batchsz < me; j += batchsz) cannam@127: dobatch(ego, rio, iio, j, j + batchsz, buf); cannam@127: cannam@127: dobatch(ego, rio, iio, j, me, buf); cannam@127: } cannam@127: cannam@127: BUF_FREE(buf, bufsz); cannam@127: } cannam@127: cannam@127: /************************************************************* cannam@127: common code cannam@127: *************************************************************/ cannam@127: static void awake(plan *ego_, enum wakefulness wakefulness) cannam@127: { cannam@127: P *ego = (P *) ego_; cannam@127: cannam@127: X(twiddle_awake)(wakefulness, &ego->td, ego->slv->desc->tw, cannam@127: ego->r * ego->m, ego->r, ego->m + ego->extra_iter); cannam@127: } cannam@127: cannam@127: static void destroy(plan *ego_) cannam@127: { cannam@127: P *ego = (P *) ego_; cannam@127: X(stride_destroy)(ego->brs); cannam@127: X(stride_destroy)(ego->rs); cannam@127: } cannam@127: cannam@127: static void print(const plan *ego_, printer *p) cannam@127: { cannam@127: const P *ego = (const P *) ego_; cannam@127: const S *slv = ego->slv; cannam@127: const ct_desc *e = slv->desc; cannam@127: cannam@127: if (slv->bufferedp) cannam@127: p->print(p, "(dftw-directbuf/%D-%D/%D%v \"%s\")", cannam@127: compute_batchsize(ego->r), ego->r, cannam@127: X(twiddle_length)(ego->r, e->tw), ego->v, e->nam); cannam@127: else cannam@127: p->print(p, "(dftw-direct-%D/%D%v \"%s\")", cannam@127: ego->r, X(twiddle_length)(ego->r, e->tw), ego->v, e->nam); cannam@127: } cannam@127: cannam@127: static int applicable0(const S *ego, cannam@127: INT r, INT irs, INT ors, cannam@127: INT m, INT ms, cannam@127: INT v, INT ivs, INT ovs, cannam@127: INT mb, INT me, cannam@127: R *rio, R *iio, cannam@127: const planner *plnr, INT *extra_iter) cannam@127: { cannam@127: const ct_desc *e = ego->desc; cannam@127: UNUSED(v); cannam@127: cannam@127: return ( cannam@127: 1 cannam@127: && r == e->radix cannam@127: && irs == ors /* in-place along R */ cannam@127: && ivs == ovs /* in-place along V */ cannam@127: cannam@127: /* check for alignment/vector length restrictions */ cannam@127: && ((*extra_iter = 0, cannam@127: e->genus->okp(e, rio, iio, irs, ivs, m, mb, me, ms, plnr)) cannam@127: || cannam@127: (*extra_iter = 1, cannam@127: (1 cannam@127: /* FIXME: require full array, otherwise some threads cannam@127: may be extra_iter and other threads won't be. cannam@127: Generating the proper twiddle factors is a pain in cannam@127: this case */ cannam@127: && mb == 0 && me == m cannam@127: && e->genus->okp(e, rio, iio, irs, ivs, cannam@127: m, mb, me - 1, ms, plnr) cannam@127: && e->genus->okp(e, rio, iio, irs, ivs, cannam@127: m, me - 1, me + 1, ms, plnr)))) cannam@127: cannam@127: && (e->genus->okp(e, rio + ivs, iio + ivs, irs, ivs, cannam@127: m, mb, me - *extra_iter, ms, plnr)) cannam@127: cannam@127: ); cannam@127: } cannam@127: cannam@127: static int applicable0_buf(const S *ego, cannam@127: INT r, INT irs, INT ors, cannam@127: INT m, INT ms, cannam@127: INT v, INT ivs, INT ovs, cannam@127: INT mb, INT me, cannam@127: R *rio, R *iio, cannam@127: const planner *plnr) cannam@127: { cannam@127: const ct_desc *e = ego->desc; cannam@127: INT batchsz; cannam@127: UNUSED(v); UNUSED(ms); UNUSED(rio); UNUSED(iio); cannam@127: cannam@127: return ( cannam@127: 1 cannam@127: && r == e->radix cannam@127: && irs == ors /* in-place along R */ cannam@127: && ivs == ovs /* in-place along V */ cannam@127: cannam@127: /* check for alignment/vector length restrictions, both for cannam@127: batchsize and for the remainder */ cannam@127: && (batchsz = compute_batchsize(r), 1) cannam@127: && (e->genus->okp(e, 0, ((const R *)0) + 1, 2 * batchsz, 0, cannam@127: m, mb, mb + batchsz, 2, plnr)) cannam@127: && (e->genus->okp(e, 0, ((const R *)0) + 1, 2 * batchsz, 0, cannam@127: m, mb, me, 2, plnr)) cannam@127: ); cannam@127: } cannam@127: cannam@127: static int applicable(const S *ego, cannam@127: INT r, INT irs, INT ors, cannam@127: INT m, INT ms, cannam@127: INT v, INT ivs, INT ovs, cannam@127: INT mb, INT me, cannam@127: R *rio, R *iio, cannam@127: const planner *plnr, INT *extra_iter) cannam@127: { cannam@127: if (ego->bufferedp) { cannam@127: *extra_iter = 0; cannam@127: if (!applicable0_buf(ego, cannam@127: r, irs, ors, m, ms, v, ivs, ovs, mb, me, cannam@127: rio, iio, plnr)) cannam@127: return 0; cannam@127: } else { cannam@127: if (!applicable0(ego, cannam@127: r, irs, ors, m, ms, v, ivs, ovs, mb, me, cannam@127: rio, iio, plnr, extra_iter)) cannam@127: return 0; cannam@127: } cannam@127: cannam@127: if (NO_UGLYP(plnr) && X(ct_uglyp)((ego->bufferedp? (INT)512 : (INT)16), cannam@127: v, m * r, r)) cannam@127: return 0; cannam@127: cannam@127: if (m * r > 262144 && NO_FIXED_RADIX_LARGE_NP(plnr)) cannam@127: return 0; cannam@127: cannam@127: return 1; cannam@127: } cannam@127: cannam@127: static plan *mkcldw(const ct_solver *ego_, cannam@127: INT r, INT irs, INT ors, cannam@127: INT m, INT ms, cannam@127: INT v, INT ivs, INT ovs, cannam@127: INT mstart, INT mcount, cannam@127: R *rio, R *iio, cannam@127: planner *plnr) cannam@127: { cannam@127: const S *ego = (const S *) ego_; cannam@127: P *pln; cannam@127: const ct_desc *e = ego->desc; cannam@127: INT extra_iter; cannam@127: cannam@127: static const plan_adt padt = { cannam@127: 0, awake, print, destroy cannam@127: }; cannam@127: cannam@127: A(mstart >= 0 && mstart + mcount <= m); cannam@127: if (!applicable(ego, cannam@127: r, irs, ors, m, ms, v, ivs, ovs, mstart, mstart + mcount, cannam@127: rio, iio, plnr, &extra_iter)) cannam@127: return (plan *)0; cannam@127: cannam@127: if (ego->bufferedp) { cannam@127: pln = MKPLAN_DFTW(P, &padt, apply_buf); cannam@127: } else { cannam@127: pln = MKPLAN_DFTW(P, &padt, extra_iter ? apply_extra_iter : apply); cannam@127: } cannam@127: cannam@127: pln->k = ego->k; cannam@127: pln->rs = X(mkstride)(r, irs); cannam@127: pln->td = 0; cannam@127: pln->r = r; cannam@127: pln->m = m; cannam@127: pln->ms = ms; cannam@127: pln->v = v; cannam@127: pln->vs = ivs; cannam@127: pln->mb = mstart; cannam@127: pln->me = mstart + mcount; cannam@127: pln->slv = ego; cannam@127: pln->brs = X(mkstride)(r, 2 * compute_batchsize(r)); cannam@127: pln->extra_iter = extra_iter; cannam@127: cannam@127: X(ops_zero)(&pln->super.super.ops); cannam@127: X(ops_madd2)(v * (mcount/e->genus->vl), &e->ops, &pln->super.super.ops); cannam@127: cannam@127: if (ego->bufferedp) { cannam@127: /* 8 load/stores * N * V */ cannam@127: pln->super.super.ops.other += 8 * r * mcount * v; cannam@127: } cannam@127: cannam@127: pln->super.super.could_prune_now_p = cannam@127: (!ego->bufferedp && r >= 5 && r < 64 && m >= r); cannam@127: return &(pln->super.super); cannam@127: } cannam@127: cannam@127: static void regone(planner *plnr, kdftw codelet, cannam@127: const ct_desc *desc, int dec, int bufferedp) cannam@127: { cannam@127: S *slv = (S *)X(mksolver_ct)(sizeof(S), desc->radix, dec, mkcldw, 0); cannam@127: slv->k = codelet; cannam@127: slv->desc = desc; cannam@127: slv->bufferedp = bufferedp; cannam@127: REGISTER_SOLVER(plnr, &(slv->super.super)); cannam@127: if (X(mksolver_ct_hook)) { cannam@127: slv = (S *)X(mksolver_ct_hook)(sizeof(S), desc->radix, cannam@127: dec, mkcldw, 0); cannam@127: slv->k = codelet; cannam@127: slv->desc = desc; cannam@127: slv->bufferedp = bufferedp; cannam@127: REGISTER_SOLVER(plnr, &(slv->super.super)); cannam@127: } cannam@127: } cannam@127: cannam@127: void X(regsolver_ct_directw)(planner *plnr, kdftw codelet, cannam@127: const ct_desc *desc, int dec) cannam@127: { cannam@127: regone(plnr, codelet, desc, dec, /* bufferedp */ 0); cannam@127: regone(plnr, codelet, desc, dec, /* bufferedp */ 1); cannam@127: }