Chris@82: /* Chris@82: * Copyright (c) 2003, 2007-14 Matteo Frigo Chris@82: * Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology Chris@82: * Chris@82: * This program is free software; you can redistribute it and/or modify Chris@82: * it under the terms of the GNU General Public License as published by Chris@82: * the Free Software Foundation; either version 2 of the License, or Chris@82: * (at your option) any later version. Chris@82: * Chris@82: * This program is distributed in the hope that it will be useful, Chris@82: * but WITHOUT ANY WARRANTY; without even the implied warranty of Chris@82: * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the Chris@82: * GNU General Public License for more details. Chris@82: * Chris@82: * You should have received a copy of the GNU General Public License Chris@82: * along with this program; if not, write to the Free Software Chris@82: * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA Chris@82: * Chris@82: */ Chris@82: Chris@82: Chris@82: #include "dft/ct.h" Chris@82: Chris@82: typedef struct { Chris@82: ct_solver super; Chris@82: const ct_desc *desc; Chris@82: int bufferedp; Chris@82: kdftw k; Chris@82: } S; Chris@82: Chris@82: typedef struct { Chris@82: plan_dftw super; Chris@82: kdftw k; Chris@82: INT r; Chris@82: stride rs; Chris@82: INT m, ms, v, vs, mb, me, extra_iter; Chris@82: stride brs; Chris@82: twid *td; Chris@82: const S *slv; Chris@82: } P; Chris@82: Chris@82: Chris@82: /************************************************************* Chris@82: Nonbuffered code Chris@82: *************************************************************/ Chris@82: static void apply(const plan *ego_, R *rio, R *iio) Chris@82: { Chris@82: const P *ego = (const P *) ego_; Chris@82: INT i; Chris@82: ASSERT_ALIGNED_DOUBLE; Chris@82: for (i = 0; i < ego->v; ++i, rio += ego->vs, iio += ego->vs) { Chris@82: INT mb = ego->mb, ms = ego->ms; Chris@82: ego->k(rio + mb*ms, iio + mb*ms, ego->td->W, Chris@82: ego->rs, mb, ego->me, ms); Chris@82: } Chris@82: } Chris@82: Chris@82: static void apply_extra_iter(const plan *ego_, R *rio, R *iio) Chris@82: { Chris@82: const P *ego = (const P *) ego_; Chris@82: INT i, v = ego->v, vs = ego->vs; Chris@82: INT mb = ego->mb, me = ego->me, mm = me - 1, ms = ego->ms; Chris@82: ASSERT_ALIGNED_DOUBLE; Chris@82: for (i = 0; i < v; ++i, rio += vs, iio += vs) { Chris@82: ego->k(rio + mb*ms, iio + mb*ms, ego->td->W, Chris@82: ego->rs, mb, mm, ms); Chris@82: ego->k(rio + mm*ms, iio + mm*ms, ego->td->W, Chris@82: ego->rs, mm, mm+2, 0); Chris@82: } Chris@82: } Chris@82: Chris@82: /************************************************************* Chris@82: Buffered code Chris@82: *************************************************************/ Chris@82: static void dobatch(const P *ego, R *rA, R *iA, INT mb, INT me, R *buf) Chris@82: { Chris@82: INT brs = WS(ego->brs, 1); Chris@82: INT rs = WS(ego->rs, 1); Chris@82: INT ms = ego->ms; Chris@82: Chris@82: X(cpy2d_pair_ci)(rA + mb*ms, iA + mb*ms, buf, buf + 1, Chris@82: ego->r, rs, brs, Chris@82: me - mb, ms, 2); Chris@82: ego->k(buf, buf + 1, ego->td->W, ego->brs, mb, me, 2); Chris@82: X(cpy2d_pair_co)(buf, buf + 1, rA + mb*ms, iA + mb*ms, Chris@82: ego->r, brs, rs, Chris@82: me - mb, 2, ms); Chris@82: } Chris@82: Chris@82: /* must be even for SIMD alignment; should not be 2^k to avoid Chris@82: associativity conflicts */ Chris@82: static INT compute_batchsize(INT radix) Chris@82: { Chris@82: /* round up to multiple of 4 */ Chris@82: radix += 3; Chris@82: radix &= -4; Chris@82: Chris@82: return (radix + 2); Chris@82: } Chris@82: Chris@82: static void apply_buf(const plan *ego_, R *rio, R *iio) Chris@82: { Chris@82: const P *ego = (const P *) ego_; Chris@82: INT i, j, v = ego->v, r = ego->r; Chris@82: INT batchsz = compute_batchsize(r); Chris@82: R *buf; Chris@82: INT mb = ego->mb, me = ego->me; Chris@82: size_t bufsz = r * batchsz * 2 * sizeof(R); Chris@82: Chris@82: BUF_ALLOC(R *, buf, bufsz); Chris@82: Chris@82: for (i = 0; i < v; ++i, rio += ego->vs, iio += ego->vs) { Chris@82: for (j = mb; j + batchsz < me; j += batchsz) Chris@82: dobatch(ego, rio, iio, j, j + batchsz, buf); Chris@82: Chris@82: dobatch(ego, rio, iio, j, me, buf); Chris@82: } Chris@82: Chris@82: BUF_FREE(buf, bufsz); Chris@82: } Chris@82: Chris@82: /************************************************************* Chris@82: common code Chris@82: *************************************************************/ Chris@82: static void awake(plan *ego_, enum wakefulness wakefulness) Chris@82: { Chris@82: P *ego = (P *) ego_; Chris@82: Chris@82: X(twiddle_awake)(wakefulness, &ego->td, ego->slv->desc->tw, Chris@82: ego->r * ego->m, ego->r, ego->m + ego->extra_iter); Chris@82: } Chris@82: Chris@82: static void destroy(plan *ego_) Chris@82: { Chris@82: P *ego = (P *) ego_; Chris@82: X(stride_destroy)(ego->brs); Chris@82: X(stride_destroy)(ego->rs); Chris@82: } Chris@82: Chris@82: static void print(const plan *ego_, printer *p) Chris@82: { Chris@82: const P *ego = (const P *) ego_; Chris@82: const S *slv = ego->slv; Chris@82: const ct_desc *e = slv->desc; Chris@82: Chris@82: if (slv->bufferedp) Chris@82: p->print(p, "(dftw-directbuf/%D-%D/%D%v \"%s\")", Chris@82: compute_batchsize(ego->r), ego->r, Chris@82: X(twiddle_length)(ego->r, e->tw), ego->v, e->nam); Chris@82: else Chris@82: p->print(p, "(dftw-direct-%D/%D%v \"%s\")", Chris@82: ego->r, X(twiddle_length)(ego->r, e->tw), ego->v, e->nam); Chris@82: } Chris@82: Chris@82: static int applicable0(const S *ego, Chris@82: INT r, INT irs, INT ors, Chris@82: INT m, INT ms, Chris@82: INT v, INT ivs, INT ovs, Chris@82: INT mb, INT me, Chris@82: R *rio, R *iio, Chris@82: const planner *plnr, INT *extra_iter) Chris@82: { Chris@82: const ct_desc *e = ego->desc; Chris@82: UNUSED(v); Chris@82: Chris@82: return ( Chris@82: 1 Chris@82: && r == e->radix Chris@82: && irs == ors /* in-place along R */ Chris@82: && ivs == ovs /* in-place along V */ Chris@82: Chris@82: /* check for alignment/vector length restrictions */ Chris@82: && ((*extra_iter = 0, Chris@82: e->genus->okp(e, rio, iio, irs, ivs, m, mb, me, ms, plnr)) Chris@82: || Chris@82: (*extra_iter = 1, Chris@82: (1 Chris@82: /* FIXME: require full array, otherwise some threads Chris@82: may be extra_iter and other threads won't be. Chris@82: Generating the proper twiddle factors is a pain in Chris@82: this case */ Chris@82: && mb == 0 && me == m Chris@82: && e->genus->okp(e, rio, iio, irs, ivs, Chris@82: m, mb, me - 1, ms, plnr) Chris@82: && e->genus->okp(e, rio, iio, irs, ivs, Chris@82: m, me - 1, me + 1, ms, plnr)))) Chris@82: Chris@82: && (e->genus->okp(e, rio + ivs, iio + ivs, irs, ivs, Chris@82: m, mb, me - *extra_iter, ms, plnr)) Chris@82: Chris@82: ); Chris@82: } Chris@82: Chris@82: static int applicable0_buf(const S *ego, Chris@82: INT r, INT irs, INT ors, Chris@82: INT m, INT ms, Chris@82: INT v, INT ivs, INT ovs, Chris@82: INT mb, INT me, Chris@82: R *rio, R *iio, Chris@82: const planner *plnr) Chris@82: { Chris@82: const ct_desc *e = ego->desc; Chris@82: INT batchsz; Chris@82: UNUSED(v); UNUSED(ms); UNUSED(rio); UNUSED(iio); Chris@82: Chris@82: return ( Chris@82: 1 Chris@82: && r == e->radix Chris@82: && irs == ors /* in-place along R */ Chris@82: && ivs == ovs /* in-place along V */ Chris@82: Chris@82: /* check for alignment/vector length restrictions, both for Chris@82: batchsize and for the remainder */ Chris@82: && (batchsz = compute_batchsize(r), 1) Chris@82: && (e->genus->okp(e, 0, ((const R *)0) + 1, 2 * batchsz, 0, Chris@82: m, mb, mb + batchsz, 2, plnr)) Chris@82: && (e->genus->okp(e, 0, ((const R *)0) + 1, 2 * batchsz, 0, Chris@82: m, mb, me, 2, plnr)) Chris@82: ); Chris@82: } Chris@82: Chris@82: static int applicable(const S *ego, Chris@82: INT r, INT irs, INT ors, Chris@82: INT m, INT ms, Chris@82: INT v, INT ivs, INT ovs, Chris@82: INT mb, INT me, Chris@82: R *rio, R *iio, Chris@82: const planner *plnr, INT *extra_iter) Chris@82: { Chris@82: if (ego->bufferedp) { Chris@82: *extra_iter = 0; Chris@82: if (!applicable0_buf(ego, Chris@82: r, irs, ors, m, ms, v, ivs, ovs, mb, me, Chris@82: rio, iio, plnr)) Chris@82: return 0; Chris@82: } else { Chris@82: if (!applicable0(ego, Chris@82: r, irs, ors, m, ms, v, ivs, ovs, mb, me, Chris@82: rio, iio, plnr, extra_iter)) Chris@82: return 0; Chris@82: } Chris@82: Chris@82: if (NO_UGLYP(plnr) && X(ct_uglyp)((ego->bufferedp? (INT)512 : (INT)16), Chris@82: v, m * r, r)) Chris@82: return 0; Chris@82: Chris@82: if (m * r > 262144 && NO_FIXED_RADIX_LARGE_NP(plnr)) Chris@82: return 0; Chris@82: Chris@82: return 1; Chris@82: } Chris@82: Chris@82: static plan *mkcldw(const ct_solver *ego_, Chris@82: INT r, INT irs, INT ors, Chris@82: INT m, INT ms, Chris@82: INT v, INT ivs, INT ovs, Chris@82: INT mstart, INT mcount, Chris@82: R *rio, R *iio, Chris@82: planner *plnr) Chris@82: { Chris@82: const S *ego = (const S *) ego_; Chris@82: P *pln; Chris@82: const ct_desc *e = ego->desc; Chris@82: INT extra_iter; Chris@82: Chris@82: static const plan_adt padt = { Chris@82: 0, awake, print, destroy Chris@82: }; Chris@82: Chris@82: A(mstart >= 0 && mstart + mcount <= m); Chris@82: if (!applicable(ego, Chris@82: r, irs, ors, m, ms, v, ivs, ovs, mstart, mstart + mcount, Chris@82: rio, iio, plnr, &extra_iter)) Chris@82: return (plan *)0; Chris@82: Chris@82: if (ego->bufferedp) { Chris@82: pln = MKPLAN_DFTW(P, &padt, apply_buf); Chris@82: } else { Chris@82: pln = MKPLAN_DFTW(P, &padt, extra_iter ? apply_extra_iter : apply); Chris@82: } Chris@82: Chris@82: pln->k = ego->k; Chris@82: pln->rs = X(mkstride)(r, irs); Chris@82: pln->td = 0; Chris@82: pln->r = r; Chris@82: pln->m = m; Chris@82: pln->ms = ms; Chris@82: pln->v = v; Chris@82: pln->vs = ivs; Chris@82: pln->mb = mstart; Chris@82: pln->me = mstart + mcount; Chris@82: pln->slv = ego; Chris@82: pln->brs = X(mkstride)(r, 2 * compute_batchsize(r)); Chris@82: pln->extra_iter = extra_iter; Chris@82: Chris@82: X(ops_zero)(&pln->super.super.ops); Chris@82: X(ops_madd2)(v * (mcount/e->genus->vl), &e->ops, &pln->super.super.ops); Chris@82: Chris@82: if (ego->bufferedp) { Chris@82: /* 8 load/stores * N * V */ Chris@82: pln->super.super.ops.other += 8 * r * mcount * v; Chris@82: } Chris@82: Chris@82: pln->super.super.could_prune_now_p = Chris@82: (!ego->bufferedp && r >= 5 && r < 64 && m >= r); Chris@82: return &(pln->super.super); Chris@82: } Chris@82: Chris@82: static void regone(planner *plnr, kdftw codelet, Chris@82: const ct_desc *desc, int dec, int bufferedp) Chris@82: { Chris@82: S *slv = (S *)X(mksolver_ct)(sizeof(S), desc->radix, dec, mkcldw, 0); Chris@82: slv->k = codelet; Chris@82: slv->desc = desc; Chris@82: slv->bufferedp = bufferedp; Chris@82: REGISTER_SOLVER(plnr, &(slv->super.super)); Chris@82: if (X(mksolver_ct_hook)) { Chris@82: slv = (S *)X(mksolver_ct_hook)(sizeof(S), desc->radix, Chris@82: dec, mkcldw, 0); Chris@82: slv->k = codelet; Chris@82: slv->desc = desc; Chris@82: slv->bufferedp = bufferedp; Chris@82: REGISTER_SOLVER(plnr, &(slv->super.super)); Chris@82: } Chris@82: } Chris@82: Chris@82: void X(regsolver_ct_directw)(planner *plnr, kdftw codelet, Chris@82: const ct_desc *desc, int dec) Chris@82: { Chris@82: regone(plnr, codelet, desc, dec, /* bufferedp */ 0); Chris@82: regone(plnr, codelet, desc, dec, /* bufferedp */ 1); Chris@82: }