Chris@10: /*
Chris@10:  * Copyright (c) 2003, 2007-11 Matteo Frigo
Chris@10:  * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
Chris@10:  *
Chris@10:  * This program is free software; you can redistribute it and/or modify
Chris@10:  * it under the terms of the GNU General Public License as published by
Chris@10:  * the Free Software Foundation; either version 2 of the License, or
Chris@10:  * (at your option) any later version.
Chris@10:  *
Chris@10:  * This program is distributed in the hope that it will be useful,
Chris@10:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
Chris@10:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
Chris@10:  * GNU General Public License for more details.
Chris@10:  *
Chris@10:  * You should have received a copy of the GNU General Public License
Chris@10:  * along with this program; if not, write to the Free Software
Chris@10:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
Chris@10:  *
Chris@10:  */
Chris@10: 
Chris@10: 
Chris@10: #include "ct.h"
Chris@10: 
Chris@10: typedef struct {
Chris@10:      ct_solver super;
Chris@10:      const ct_desc *desc;
Chris@10:      int bufferedp;
Chris@10:      kdftw k;
Chris@10: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_dftw super;
Chris@10:      kdftw k;
Chris@10:      INT r;
Chris@10:      stride rs;
Chris@10:      INT m, ms, v, vs, mb, me, extra_iter;
Chris@10:      stride brs;
Chris@10:      twid *td;
Chris@10:      const S *slv;
Chris@10: } P;
Chris@10: 
Chris@10: 
Chris@10: /*************************************************************
Chris@10:   Nonbuffered code
Chris@10:  *************************************************************/
Chris@10: static void apply(const plan *ego_, R *rio, R *iio)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      INT i;
Chris@10:      ASSERT_ALIGNED_DOUBLE;
Chris@10:      for (i = 0; i < ego->v; ++i, rio += ego->vs, iio += ego->vs) {
Chris@10: 	  INT  mb = ego->mb, ms = ego->ms;
Chris@10: 	  ego->k(rio + mb*ms, iio + mb*ms, ego->td->W, 
Chris@10: 		 ego->rs, mb, ego->me, ms);
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static void apply_extra_iter(const plan *ego_, R *rio, R *iio)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      INT i, v = ego->v, vs = ego->vs;
Chris@10:      INT mb = ego->mb, me = ego->me, mm = me - 1, ms = ego->ms;
Chris@10:      ASSERT_ALIGNED_DOUBLE;
Chris@10:      for (i = 0; i < v; ++i, rio += vs, iio += vs) {
Chris@10: 	  ego->k(rio + mb*ms, iio + mb*ms, ego->td->W, 
Chris@10: 		 ego->rs, mb, mm, ms);
Chris@10: 	  ego->k(rio + mm*ms, iio + mm*ms, ego->td->W, 
Chris@10: 		 ego->rs, mm, mm+2, 0);
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: /*************************************************************
Chris@10:   Buffered code
Chris@10:  *************************************************************/
Chris@10: static void dobatch(const P *ego, R *rA, R *iA, INT mb, INT me, R *buf)
Chris@10: {
Chris@10:      INT brs = WS(ego->brs, 1);
Chris@10:      INT rs = WS(ego->rs, 1);
Chris@10:      INT ms = ego->ms;
Chris@10: 
Chris@10:      X(cpy2d_pair_ci)(rA + mb*ms, iA + mb*ms, buf, buf + 1,
Chris@10: 		      ego->r, rs, brs,
Chris@10: 		      me - mb, ms, 2);
Chris@10:      ego->k(buf, buf + 1, ego->td->W, ego->brs, mb, me, 2);
Chris@10:      X(cpy2d_pair_co)(buf, buf + 1, rA + mb*ms, iA + mb*ms,
Chris@10: 		      ego->r, brs, rs,
Chris@10: 		      me - mb, 2, ms);
Chris@10: }
Chris@10: 
Chris@10: /* must be even for SIMD alignment; should not be 2^k to avoid
Chris@10:    associativity conflicts */
Chris@10: static INT compute_batchsize(INT radix)
Chris@10: {
Chris@10:      /* round up to multiple of 4 */
Chris@10:      radix += 3;
Chris@10:      radix &= -4;
Chris@10: 
Chris@10:      return (radix + 2);
Chris@10: }
Chris@10: 
Chris@10: static void apply_buf(const plan *ego_, R *rio, R *iio)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      INT i, j, v = ego->v, r = ego->r;
Chris@10:      INT batchsz = compute_batchsize(r);
Chris@10:      R *buf;
Chris@10:      INT mb = ego->mb, me = ego->me;
Chris@10:      size_t bufsz = r * batchsz * 2 * sizeof(R);
Chris@10: 
Chris@10:      BUF_ALLOC(R *, buf, bufsz);
Chris@10: 
Chris@10:      for (i = 0; i < v; ++i, rio += ego->vs, iio += ego->vs) {
Chris@10: 	  for (j = mb; j + batchsz < me; j += batchsz) 
Chris@10: 	       dobatch(ego, rio, iio, j, j + batchsz, buf);
Chris@10: 
Chris@10: 	  dobatch(ego, rio, iio, j, me, buf);
Chris@10:      }
Chris@10: 
Chris@10:      BUF_FREE(buf, bufsz);
Chris@10: }
Chris@10: 
Chris@10: /*************************************************************
Chris@10:   common code
Chris@10:  *************************************************************/
Chris@10: static void awake(plan *ego_, enum wakefulness wakefulness)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10: 
Chris@10:      X(twiddle_awake)(wakefulness, &ego->td, ego->slv->desc->tw,
Chris@10: 		      ego->r * ego->m, ego->r, ego->m + ego->extra_iter);
Chris@10: }
Chris@10: 
Chris@10: static void destroy(plan *ego_)
Chris@10: {
Chris@10:      P *ego = (P *) ego_;
Chris@10:      X(stride_destroy)(ego->brs);
Chris@10:      X(stride_destroy)(ego->rs);
Chris@10: }
Chris@10: 
Chris@10: static void print(const plan *ego_, printer *p)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      const S *slv = ego->slv;
Chris@10:      const ct_desc *e = slv->desc;
Chris@10: 
Chris@10:      if (slv->bufferedp)
Chris@10: 	  p->print(p, "(dftw-directbuf/%D-%D/%D%v \"%s\")",
Chris@10: 		   compute_batchsize(ego->r), ego->r,
Chris@10: 		   X(twiddle_length)(ego->r, e->tw), ego->v, e->nam);
Chris@10:      else
Chris@10: 	  p->print(p, "(dftw-direct-%D/%D%v \"%s\")",
Chris@10: 		   ego->r, X(twiddle_length)(ego->r, e->tw), ego->v, e->nam);
Chris@10: }
Chris@10: 
Chris@10: static int applicable0(const S *ego,
Chris@10: 		       INT r, INT irs, INT ors,
Chris@10: 		       INT m, INT ms,
Chris@10: 		       INT v, INT ivs, INT ovs,
Chris@10: 		       INT mb, INT me,
Chris@10: 		       R *rio, R *iio,
Chris@10: 		       const planner *plnr, INT *extra_iter)
Chris@10: {
Chris@10:      const ct_desc *e = ego->desc;
Chris@10:      UNUSED(v);
Chris@10: 
Chris@10:      return (
Chris@10: 	  1
Chris@10: 	  && r == e->radix
Chris@10: 	  && irs == ors /* in-place along R */
Chris@10: 	  && ivs == ovs /* in-place along V */
Chris@10: 
Chris@10: 	  /* check for alignment/vector length restrictions */
Chris@10: 	  && ((*extra_iter = 0,
Chris@10: 	       e->genus->okp(e, rio, iio, irs, ivs, m, mb, me, ms, plnr))
Chris@10: 	      ||
Chris@10: 	      (*extra_iter = 1,
Chris@10: 	       (1
Chris@10: 		/* FIXME: require full array, otherwise some threads
Chris@10: 		   may be extra_iter and other threads won't be.
Chris@10: 		   Generating the proper twiddle factors is a pain in
Chris@10: 		   this case */
Chris@10: 		&& mb == 0 && me == m
Chris@10: 		&& e->genus->okp(e, rio, iio, irs, ivs,
Chris@10: 				 m, mb, me - 1, ms, plnr)
Chris@10: 		&& e->genus->okp(e, rio, iio, irs, ivs,
Chris@10: 				 m, me - 1, me + 1, ms, plnr))))
Chris@10: 
Chris@10: 	  && (e->genus->okp(e, rio + ivs, iio + ivs, irs, ivs,
Chris@10: 			    m, mb, me - *extra_iter, ms, plnr))
Chris@10: 
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: static int applicable0_buf(const S *ego,
Chris@10: 			   INT r, INT irs, INT ors,
Chris@10: 			   INT m, INT ms,
Chris@10: 			   INT v, INT ivs, INT ovs,
Chris@10: 			   INT mb, INT me,
Chris@10: 			   R *rio, R *iio,
Chris@10: 			   const planner *plnr)
Chris@10: {
Chris@10:      const ct_desc *e = ego->desc;
Chris@10:      INT batchsz;
Chris@10:      UNUSED(v); UNUSED(ms); UNUSED(rio); UNUSED(iio);
Chris@10: 
Chris@10:      return (
Chris@10: 	  1
Chris@10: 	  && r == e->radix
Chris@10: 	  && irs == ors /* in-place along R */
Chris@10: 	  && ivs == ovs /* in-place along V */
Chris@10: 
Chris@10: 	  /* check for alignment/vector length restrictions, both for
Chris@10: 	     batchsize and for the remainder */
Chris@10: 	  && (batchsz = compute_batchsize(r), 1)
Chris@10: 	  && (e->genus->okp(e, 0, ((const R *)0) + 1, 2 * batchsz, 0,
Chris@10: 			    m, mb, mb + batchsz, 2, plnr))
Chris@10: 	  && (e->genus->okp(e, 0, ((const R *)0) + 1, 2 * batchsz, 0,
Chris@10: 			    m, mb, me, 2, plnr))
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: static int applicable(const S *ego,
Chris@10: 		      INT r, INT irs, INT ors,
Chris@10: 		      INT m, INT ms,
Chris@10: 		      INT v, INT ivs, INT ovs,
Chris@10: 		      INT mb, INT me,
Chris@10: 		      R *rio, R *iio,
Chris@10: 		      const planner *plnr, INT *extra_iter)
Chris@10: {
Chris@10:      if (ego->bufferedp) {
Chris@10: 	  *extra_iter = 0;
Chris@10: 	  if (!applicable0_buf(ego,
Chris@10: 			       r, irs, ors, m, ms, v, ivs, ovs, mb, me,
Chris@10: 			       rio, iio, plnr))
Chris@10: 	       return 0;
Chris@10:      } else {
Chris@10: 	  if (!applicable0(ego,
Chris@10: 			   r, irs, ors, m, ms, v, ivs, ovs, mb, me,
Chris@10: 			   rio, iio, plnr, extra_iter))
Chris@10: 	       return 0;
Chris@10:      }
Chris@10: 
Chris@10:      if (NO_UGLYP(plnr) && X(ct_uglyp)((ego->bufferedp? (INT)512 : (INT)16),
Chris@10: 				       v, m * r, r))
Chris@10: 	  return 0;
Chris@10: 
Chris@10:      if (m * r > 262144 && NO_FIXED_RADIX_LARGE_NP(plnr))
Chris@10: 	  return 0;
Chris@10: 
Chris@10:      return 1;
Chris@10: }
Chris@10: 
Chris@10: static plan *mkcldw(const ct_solver *ego_,
Chris@10: 		    INT r, INT irs, INT ors,
Chris@10: 		    INT m, INT ms,
Chris@10: 		    INT v, INT ivs, INT ovs,
Chris@10: 		    INT mstart, INT mcount,
Chris@10: 		    R *rio, R *iio,
Chris@10: 		    planner *plnr)
Chris@10: {
Chris@10:      const S *ego = (const S *) ego_;
Chris@10:      P *pln;
Chris@10:      const ct_desc *e = ego->desc;
Chris@10:      INT extra_iter;
Chris@10: 
Chris@10:      static const plan_adt padt = {
Chris@10: 	  0, awake, print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      A(mstart >= 0 && mstart + mcount <= m);
Chris@10:      if (!applicable(ego,
Chris@10: 		     r, irs, ors, m, ms, v, ivs, ovs, mstart, mstart + mcount,
Chris@10: 		     rio, iio, plnr, &extra_iter))
Chris@10:           return (plan *)0;
Chris@10: 
Chris@10:      if (ego->bufferedp) {
Chris@10: 	  pln = MKPLAN_DFTW(P, &padt, apply_buf);
Chris@10:      } else {
Chris@10: 	  pln = MKPLAN_DFTW(P, &padt, extra_iter ? apply_extra_iter : apply);
Chris@10:      }
Chris@10: 
Chris@10:      pln->k = ego->k;
Chris@10:      pln->rs = X(mkstride)(r, irs);
Chris@10:      pln->td = 0;
Chris@10:      pln->r = r;
Chris@10:      pln->m = m;
Chris@10:      pln->ms = ms;
Chris@10:      pln->v = v;
Chris@10:      pln->vs = ivs;
Chris@10:      pln->mb = mstart;
Chris@10:      pln->me = mstart + mcount;
Chris@10:      pln->slv = ego;
Chris@10:      pln->brs = X(mkstride)(r, 2 * compute_batchsize(r));
Chris@10:      pln->extra_iter = extra_iter;
Chris@10: 
Chris@10:      X(ops_zero)(&pln->super.super.ops);
Chris@10:      X(ops_madd2)(v * (mcount/e->genus->vl), &e->ops, &pln->super.super.ops);
Chris@10: 
Chris@10:      if (ego->bufferedp) {
Chris@10: 	  /* 8 load/stores * N * V */
Chris@10: 	  pln->super.super.ops.other += 8 * r * mcount * v;
Chris@10:      }
Chris@10: 
Chris@10:      pln->super.super.could_prune_now_p =
Chris@10: 	  (!ego->bufferedp && r >= 5 && r < 64 && m >= r);
Chris@10:      return &(pln->super.super);
Chris@10: }
Chris@10: 
Chris@10: static void regone(planner *plnr, kdftw codelet,
Chris@10: 		   const ct_desc *desc, int dec, int bufferedp)
Chris@10: {
Chris@10:      S *slv = (S *)X(mksolver_ct)(sizeof(S), desc->radix, dec, mkcldw, 0);
Chris@10:      slv->k = codelet;
Chris@10:      slv->desc = desc;
Chris@10:      slv->bufferedp = bufferedp;
Chris@10:      REGISTER_SOLVER(plnr, &(slv->super.super));
Chris@10:      if (X(mksolver_ct_hook)) {
Chris@10: 	  slv = (S *)X(mksolver_ct_hook)(sizeof(S), desc->radix,
Chris@10: 					 dec, mkcldw, 0);
Chris@10: 	  slv->k = codelet;
Chris@10: 	  slv->desc = desc;
Chris@10: 	  slv->bufferedp = bufferedp;
Chris@10: 	  REGISTER_SOLVER(plnr, &(slv->super.super));
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: void X(regsolver_ct_directw)(planner *plnr, kdftw codelet,
Chris@10: 			     const ct_desc *desc, int dec)
Chris@10: {
Chris@10:      regone(plnr, codelet, desc, dec, /* bufferedp */ 0);
Chris@10:      regone(plnr, codelet, desc, dec, /* bufferedp */ 1);
Chris@10: }