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: 
cannam@167: #include "rdft/hc2hc.h"
cannam@167: 
cannam@167: typedef struct {
cannam@167:      hc2hc_solver super;
cannam@167:      const hc2hc_desc *desc;
cannam@167:      khc2hc k;
cannam@167:      int bufferedp;
cannam@167: } S;
cannam@167: 
cannam@167: typedef struct {
cannam@167:      plan_hc2hc super;
cannam@167:      khc2hc k;
cannam@167:      plan *cld0, *cldm; /* children for 0th and middle butterflies */
cannam@167:      INT r, m, v;
cannam@167:      INT ms, vs, mb, me;
cannam@167:      stride rs, brs;
cannam@167:      twid *td;
cannam@167:      const S *slv;
cannam@167: } P;
cannam@167: 
cannam@167: /*************************************************************
cannam@167:   Nonbuffered code
cannam@167: *************************************************************/
cannam@167: static void apply(const plan *ego_, R *IO)
cannam@167: {
cannam@167:      const P *ego = (const P *) ego_;
cannam@167:      plan_rdft *cld0 = (plan_rdft *) ego->cld0;
cannam@167:      plan_rdft *cldm = (plan_rdft *) ego->cldm;
cannam@167:      INT i, m = ego->m, v = ego->v;
cannam@167:      INT mb = ego->mb, me = ego->me;
cannam@167:      INT ms = ego->ms, vs = ego->vs;
cannam@167: 
cannam@167:      for (i = 0; i < v; ++i, IO += vs) {
cannam@167: 	  cld0->apply((plan *) cld0, IO, IO);
cannam@167: 	  ego->k(IO + ms * mb, IO + (m - mb) * ms, 
cannam@167: 		 ego->td->W, ego->rs, mb, me, ms);
cannam@167: 	  cldm->apply((plan *) cldm, IO + (m/2) * ms, IO + (m/2) * ms);
cannam@167:      }
cannam@167: }
cannam@167: 
cannam@167: /*************************************************************
cannam@167:   Buffered code
cannam@167: *************************************************************/
cannam@167: 
cannam@167: /* should not be 2^k to avoid associativity conflicts */
cannam@167: static INT compute_batchsize(INT radix)
cannam@167: {
cannam@167:      /* round up to multiple of 4 */
cannam@167:      radix += 3;
cannam@167:      radix &= -4;
cannam@167: 
cannam@167:      return (radix + 2);
cannam@167: }
cannam@167: 
cannam@167: static void dobatch(const P *ego, R *IOp, R *IOm,
cannam@167: 		    INT mb, INT me, R *bufp)
cannam@167: {
cannam@167:      INT b = WS(ego->brs, 1);
cannam@167:      INT rs = WS(ego->rs, 1);
cannam@167:      INT r = ego->r;
cannam@167:      INT ms = ego->ms;
cannam@167:      R *bufm = bufp + b - 1;
cannam@167: 
cannam@167:      X(cpy2d_ci)(IOp + mb * ms, bufp, r, rs, b, me - mb,  ms,  1, 1);
cannam@167:      X(cpy2d_ci)(IOm - mb * ms, bufm, r, rs, b, me - mb, -ms, -1, 1);
cannam@167: 
cannam@167:      ego->k(bufp, bufm, ego->td->W, ego->brs, mb, me, 1);
cannam@167: 
cannam@167:      X(cpy2d_co)(bufp, IOp + mb * ms, r, b, rs, me - mb,  1,  ms, 1);
cannam@167:      X(cpy2d_co)(bufm, IOm - mb * ms, r, b, rs, me - mb, -1, -ms, 1);
cannam@167: }
cannam@167: 
cannam@167: static void apply_buf(const plan *ego_, R *IO)
cannam@167: {
cannam@167:      const P *ego = (const P *) ego_;
cannam@167:      plan_rdft *cld0 = (plan_rdft *) ego->cld0;
cannam@167:      plan_rdft *cldm = (plan_rdft *) ego->cldm;
cannam@167:      INT i, j, m = ego->m, v = ego->v, r = ego->r;
cannam@167:      INT mb = ego->mb, me = ego->me, ms = ego->ms;
cannam@167:      INT batchsz = compute_batchsize(r);
cannam@167:      R *buf;
cannam@167:      size_t bufsz = r * batchsz * 2 * sizeof(R);
cannam@167: 
cannam@167:      BUF_ALLOC(R *, buf, bufsz);
cannam@167: 
cannam@167:      for (i = 0; i < v; ++i, IO += ego->vs) {
cannam@167: 	  R *IOp = IO;
cannam@167: 	  R *IOm = IO + m * ms;
cannam@167: 
cannam@167: 	  cld0->apply((plan *) cld0, IO, IO);
cannam@167: 
cannam@167: 	  for (j = mb; j + batchsz < me; j += batchsz) 	       
cannam@167: 	       dobatch(ego, IOp, IOm, j, j + batchsz, buf);
cannam@167: 
cannam@167: 	  dobatch(ego, IOp, IOm, j, me, buf);
cannam@167: 
cannam@167: 	  cldm->apply((plan *) cldm, IO + ms * (m/2), IO + ms * (m/2));
cannam@167:      }
cannam@167: 
cannam@167:      BUF_FREE(buf, bufsz);
cannam@167: }
cannam@167: 
cannam@167: static void awake(plan *ego_, enum wakefulness wakefulness)
cannam@167: {
cannam@167:      P *ego = (P *) ego_;
cannam@167: 
cannam@167:      X(plan_awake)(ego->cld0, wakefulness);
cannam@167:      X(plan_awake)(ego->cldm, wakefulness);
cannam@167:      X(twiddle_awake)(wakefulness, &ego->td, ego->slv->desc->tw, 
cannam@167: 		      ego->r * ego->m, ego->r, (ego->m - 1) / 2);
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->cld0);
cannam@167:      X(plan_destroy_internal)(ego->cldm);
cannam@167:      X(stride_destroy)(ego->rs);
cannam@167:      X(stride_destroy)(ego->brs);
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:      const S *slv = ego->slv;
cannam@167:      const hc2hc_desc *e = slv->desc;
cannam@167:      INT batchsz = compute_batchsize(ego->r);
cannam@167: 
cannam@167:      if (slv->bufferedp)
cannam@167: 	  p->print(p, "(hc2hc-directbuf/%D-%D/%D%v \"%s\"%(%p%)%(%p%))",
cannam@167: 		   batchsz, ego->r, X(twiddle_length)(ego->r, e->tw), 
cannam@167: 		   ego->v, e->nam, ego->cld0, ego->cldm);
cannam@167:      else
cannam@167: 	  p->print(p, "(hc2hc-direct-%D/%D%v \"%s\"%(%p%)%(%p%))",
cannam@167: 		   ego->r, X(twiddle_length)(ego->r, e->tw), ego->v, e->nam,
cannam@167: 		   ego->cld0, ego->cldm);
cannam@167: }
cannam@167: 
cannam@167: static int applicable0(const S *ego, rdft_kind kind, INT r)
cannam@167: {
cannam@167:      const hc2hc_desc *e = ego->desc;
cannam@167: 
cannam@167:      return (1
cannam@167: 	     && r == e->radix
cannam@167: 	     && kind == e->genus->kind
cannam@167: 	  );
cannam@167: }
cannam@167: 
cannam@167: static int applicable(const S *ego, rdft_kind kind, INT r, INT m, INT v,
cannam@167: 		      const planner *plnr)
cannam@167: {
cannam@167:      if (!applicable0(ego, kind, r))
cannam@167:           return 0;
cannam@167: 
cannam@167:      if (NO_UGLYP(plnr) && X(ct_uglyp)((ego->bufferedp? (INT)512 : (INT)16),
cannam@167: 				       v, m * r, r)) 
cannam@167: 	  return 0;
cannam@167: 
cannam@167:      return 1;
cannam@167: }
cannam@167: 
cannam@167: #define CLDMP(m, mstart, mcount) (2 * ((mstart) + (mcount)) == (m) + 2)
cannam@167: #define CLD0P(mstart) ((mstart) == 0)
cannam@167: 
cannam@167: static plan *mkcldw(const hc2hc_solver *ego_, 
cannam@167: 		    rdft_kind kind, INT r, INT m, INT ms, INT v, INT vs, 
cannam@167: 		    INT mstart, INT mcount,
cannam@167: 		    R *IO, planner *plnr)
cannam@167: {
cannam@167:      const S *ego = (const S *) ego_;
cannam@167:      P *pln;
cannam@167:      const hc2hc_desc *e = ego->desc;
cannam@167:      plan *cld0 = 0, *cldm = 0;
cannam@167:      INT imid = (m / 2) * ms;
cannam@167:      INT rs = m * ms;
cannam@167: 
cannam@167:      static const plan_adt padt = {
cannam@167: 	  0, awake, print, destroy
cannam@167:      };
cannam@167: 
cannam@167:      if (!applicable(ego, kind, r, m, v, plnr))
cannam@167:           return (plan *)0;
cannam@167: 
cannam@167:      cld0 = X(mkplan_d)(
cannam@167: 	  plnr, 
cannam@167: 	  X(mkproblem_rdft_1_d)((CLD0P(mstart) ?
cannam@167: 				 X(mktensor_1d)(r, rs, rs) : X(mktensor_0d)()),
cannam@167: 				X(mktensor_0d)(),
cannam@167: 				TAINT(IO, vs), TAINT(IO, vs), 
cannam@167: 				kind));
cannam@167:      if (!cld0) goto nada;
cannam@167: 
cannam@167:      cldm = X(mkplan_d)(
cannam@167: 	  plnr, 
cannam@167: 	  X(mkproblem_rdft_1_d)((CLDMP(m, mstart, mcount) ?
cannam@167: 				 X(mktensor_1d)(r, rs, rs) : X(mktensor_0d)()),
cannam@167: 				X(mktensor_0d)(),
cannam@167: 				TAINT(IO + imid, vs), TAINT(IO + imid, vs),
cannam@167: 				kind == R2HC ? R2HCII : HC2RIII));
cannam@167:      if (!cldm) goto nada;
cannam@167: 	  
cannam@167:      pln = MKPLAN_HC2HC(P, &padt, ego->bufferedp ? apply_buf : apply);
cannam@167: 
cannam@167:      pln->k = ego->k;
cannam@167:      pln->td = 0;
cannam@167:      pln->r = r; pln->rs = X(mkstride)(r, rs);
cannam@167:      pln->m = m; pln->ms = ms;
cannam@167:      pln->v = v; pln->vs = vs;
cannam@167:      pln->slv = ego;
cannam@167:      pln->brs = X(mkstride)(r, 2 * compute_batchsize(r));
cannam@167:      pln->cld0 = cld0;
cannam@167:      pln->cldm = cldm;
cannam@167:      pln->mb = mstart + CLD0P(mstart);
cannam@167:      pln->me = mstart + mcount - CLDMP(m, mstart, mcount);
cannam@167: 
cannam@167:      X(ops_zero)(&pln->super.super.ops);
cannam@167:      X(ops_madd2)(v * ((pln->me - pln->mb) / e->genus->vl),
cannam@167: 		  &e->ops, &pln->super.super.ops);
cannam@167:      X(ops_madd2)(v, &cld0->ops, &pln->super.super.ops);
cannam@167:      X(ops_madd2)(v, &cldm->ops, &pln->super.super.ops);
cannam@167: 
cannam@167:      if (ego->bufferedp) 
cannam@167: 	  pln->super.super.ops.other += 4 * r * (pln->me - pln->mb) * v;
cannam@167: 
cannam@167:      pln->super.super.could_prune_now_p =
cannam@167: 	  (!ego->bufferedp && r >= 5 && r < 64 && m >= r);
cannam@167: 
cannam@167:      return &(pln->super.super);
cannam@167: 
cannam@167:  nada:
cannam@167:      X(plan_destroy_internal)(cld0);
cannam@167:      X(plan_destroy_internal)(cldm);
cannam@167:      return 0;
cannam@167: }
cannam@167: 
cannam@167: static void regone(planner *plnr, khc2hc codelet, const hc2hc_desc *desc,
cannam@167: 		   int bufferedp)
cannam@167: {
cannam@167:      S *slv = (S *)X(mksolver_hc2hc)(sizeof(S), desc->radix, mkcldw);
cannam@167:      slv->k = codelet;
cannam@167:      slv->desc = desc;
cannam@167:      slv->bufferedp = bufferedp;
cannam@167:      REGISTER_SOLVER(plnr, &(slv->super.super));
cannam@167:      if (X(mksolver_hc2hc_hook)) {
cannam@167: 	  slv = (S *)X(mksolver_hc2hc_hook)(sizeof(S), desc->radix, mkcldw);
cannam@167: 	  slv->k = codelet;
cannam@167: 	  slv->desc = desc;
cannam@167: 	  slv->bufferedp = bufferedp;
cannam@167: 	  REGISTER_SOLVER(plnr, &(slv->super.super));
cannam@167:      }
cannam@167: }
cannam@167: 
cannam@167: void X(regsolver_hc2hc_direct)(planner *plnr, khc2hc codelet,
cannam@167: 			       const hc2hc_desc *desc)
cannam@167: {
cannam@167:      regone(plnr, codelet, desc, /* bufferedp */0);
cannam@167:      regone(plnr, codelet, desc, /* bufferedp */1);
cannam@167: }