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: /* direct RDFT solver, using r2c codelets */
Chris@10: 
Chris@10: #include "rdft.h"
Chris@10: 
Chris@10: typedef struct {
Chris@10:      solver super;
Chris@10:      const kr2c_desc *desc;
Chris@10:      kr2c k;
Chris@10:      int bufferedp;
Chris@10: } S;
Chris@10: 
Chris@10: typedef struct {
Chris@10:      plan_rdft super;
Chris@10: 
Chris@10:      stride rs, csr, csi;
Chris@10:      stride brs, bcsr, bcsi;
Chris@10:      INT n, vl, rs0, ivs, ovs, ioffset, bioffset;
Chris@10:      kr2c k;
Chris@10:      const S *slv;
Chris@10: } P;
Chris@10: 
Chris@10: /*************************************************************
Chris@10:   Nonbuffered code
Chris@10:  *************************************************************/
Chris@10: static void apply_r2hc(const plan *ego_, R *I, R *O)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      ASSERT_ALIGNED_DOUBLE;
Chris@10:      ego->k(I, I + ego->rs0, O, O + ego->ioffset, 
Chris@10: 	    ego->rs, ego->csr, ego->csi,
Chris@10: 	    ego->vl, ego->ivs, ego->ovs);
Chris@10: }
Chris@10: 
Chris@10: static void apply_hc2r(const plan *ego_, R *I, R *O)
Chris@10: {
Chris@10:      const P *ego = (const P *) ego_;
Chris@10:      ASSERT_ALIGNED_DOUBLE;
Chris@10:      ego->k(O, O + ego->rs0, I, I + ego->ioffset, 
Chris@10: 	    ego->rs, ego->csr, ego->csi,
Chris@10: 	    ego->vl, ego->ivs, ego->ovs);
Chris@10: }
Chris@10: 
Chris@10: /*************************************************************
Chris@10:   Buffered code
Chris@10:  *************************************************************/
Chris@10: /* should not be 2^k to avoid 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 dobatch_r2hc(const P *ego, R *I, R *O, R *buf, INT batchsz)
Chris@10: {
Chris@10:      X(cpy2d_ci)(I, buf,
Chris@10: 		 ego->n, ego->rs0, WS(ego->bcsr /* hack */, 1),
Chris@10: 		 batchsz, ego->ivs, 1, 1);
Chris@10: 
Chris@10:      if (IABS(WS(ego->csr, 1)) < IABS(ego->ovs)) {
Chris@10: 	  /* transform directly to output */
Chris@10: 	  ego->k(buf, buf + WS(ego->bcsr /* hack */, 1), 
Chris@10: 		 O, O + ego->ioffset, 
Chris@10: 		 ego->brs, ego->csr, ego->csi,
Chris@10: 		 batchsz, 1, ego->ovs);
Chris@10:      } else {
Chris@10: 	  /* transform to buffer and copy back */
Chris@10: 	  ego->k(buf, buf + WS(ego->bcsr /* hack */, 1), 
Chris@10: 		 buf, buf + ego->bioffset, 
Chris@10: 		 ego->brs, ego->bcsr, ego->bcsi,
Chris@10: 		 batchsz, 1, 1);
Chris@10: 	  X(cpy2d_co)(buf, O,
Chris@10: 		      ego->n, WS(ego->bcsr, 1), WS(ego->csr, 1),  
Chris@10: 		      batchsz, 1, ego->ovs, 1);
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static void dobatch_hc2r(const P *ego, R *I, R *O, R *buf, INT batchsz)
Chris@10: {
Chris@10:      if (IABS(WS(ego->csr, 1)) < IABS(ego->ivs)) {
Chris@10: 	  /* transform directly from input */
Chris@10: 	  ego->k(buf, buf + WS(ego->bcsr /* hack */, 1),
Chris@10: 		 I, I + ego->ioffset, 
Chris@10: 		 ego->brs, ego->csr, ego->csi,
Chris@10: 		 batchsz, ego->ivs, 1);
Chris@10:      } else {
Chris@10: 	  /* copy into buffer and transform in place */
Chris@10: 	  X(cpy2d_ci)(I, buf,
Chris@10: 		      ego->n, WS(ego->csr, 1), WS(ego->bcsr, 1),
Chris@10: 		      batchsz, ego->ivs, 1, 1);
Chris@10: 	  ego->k(buf, buf + WS(ego->bcsr /* hack */, 1),
Chris@10: 		 buf, buf + ego->bioffset, 
Chris@10: 		 ego->brs, ego->bcsr, ego->bcsi,
Chris@10: 		 batchsz, 1, 1);
Chris@10:      }
Chris@10:      X(cpy2d_co)(buf, O,
Chris@10: 		 ego->n, WS(ego->bcsr /* hack */, 1), ego->rs0,
Chris@10: 		 batchsz, 1, ego->ovs, 1);
Chris@10: }
Chris@10: 
Chris@10: static void iterate(const P *ego, R *I, R *O,
Chris@10: 		    void (*dobatch)(const P *ego, R *I, R *O, 
Chris@10: 				    R *buf, INT batchsz))
Chris@10: {
Chris@10:      R *buf;
Chris@10:      INT vl = ego->vl;
Chris@10:      INT n = ego->n;
Chris@10:      INT i;
Chris@10:      INT batchsz = compute_batchsize(n);
Chris@10:      size_t bufsz = n * batchsz * sizeof(R);
Chris@10: 
Chris@10:      BUF_ALLOC(R *, buf, bufsz);
Chris@10: 
Chris@10:      for (i = 0; i < vl - batchsz; i += batchsz) {
Chris@10: 	  dobatch(ego, I, O, buf, batchsz);
Chris@10: 	  I += batchsz * ego->ivs;
Chris@10: 	  O += batchsz * ego->ovs;
Chris@10:      }
Chris@10:      dobatch(ego, I, O, buf, vl - i);
Chris@10: 
Chris@10:      BUF_FREE(buf, bufsz);
Chris@10: }
Chris@10: 
Chris@10: static void apply_buf_r2hc(const plan *ego_, R *I, R *O)
Chris@10: {
Chris@10:      iterate((const P *) ego_, I, O, dobatch_r2hc);
Chris@10: }
Chris@10: 
Chris@10: static void apply_buf_hc2r(const plan *ego_, R *I, R *O)
Chris@10: {
Chris@10:      iterate((const P *) ego_, I, O, dobatch_hc2r);
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->rs);
Chris@10:      X(stride_destroy)(ego->csr);
Chris@10:      X(stride_destroy)(ego->csi);
Chris@10:      X(stride_destroy)(ego->brs);
Chris@10:      X(stride_destroy)(ego->bcsr);
Chris@10:      X(stride_destroy)(ego->bcsi);
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 *s = ego->slv;
Chris@10: 
Chris@10:      if (ego->slv->bufferedp)
Chris@10: 	  p->print(p, "(rdft-%s-directbuf/%D-r2c-%D%v \"%s\")", 
Chris@10: 		   X(rdft_kind_str)(s->desc->genus->kind), 
Chris@10: 		   /* hack */ WS(ego->bcsr, 1), ego->n, 
Chris@10: 		   ego->vl, s->desc->nam);
Chris@10: 
Chris@10:      else 
Chris@10: 	  p->print(p, "(rdft-%s-direct-r2c-%D%v \"%s\")", 
Chris@10: 		   X(rdft_kind_str)(s->desc->genus->kind), ego->n, 
Chris@10: 		   ego->vl, s->desc->nam);
Chris@10: }
Chris@10: 
Chris@10: static INT ioffset(rdft_kind kind, INT sz, INT s)
Chris@10: {
Chris@10:      return(s * ((kind == R2HC || kind == HC2R) ? sz : (sz - 1)));
Chris@10: }
Chris@10: 
Chris@10: static int applicable(const solver *ego_, const problem *p_)
Chris@10: {
Chris@10:      const S *ego = (const S *) ego_;
Chris@10:      const kr2c_desc *desc = ego->desc;
Chris@10:      const problem_rdft *p = (const problem_rdft *) p_;
Chris@10:      INT vl, ivs, ovs;
Chris@10: 
Chris@10:      return (
Chris@10: 	  1
Chris@10: 	  && p->sz->rnk == 1
Chris@10: 	  && p->vecsz->rnk <= 1
Chris@10: 	  && p->sz->dims[0].n == desc->n
Chris@10: 	  && p->kind[0] == desc->genus->kind
Chris@10: 
Chris@10: 	  /* check strides etc */
Chris@10: 	  && X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs)
Chris@10: 
Chris@10: 	  && (0
Chris@10: 	      /* can operate out-of-place */
Chris@10: 	      || p->I != p->O
Chris@10: 
Chris@10: 	      /* computing one transform */
Chris@10: 	      || vl == 1
Chris@10: 
Chris@10: 	      /* can operate in-place as long as strides are the same */
Chris@10: 	      || X(tensor_inplace_strides2)(p->sz, p->vecsz)
Chris@10: 	       )
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: static int applicable_buf(const solver *ego_, const problem *p_)
Chris@10: {
Chris@10:      const S *ego = (const S *) ego_;
Chris@10:      const kr2c_desc *desc = ego->desc;
Chris@10:      const problem_rdft *p = (const problem_rdft *) p_;
Chris@10:      INT vl, ivs, ovs, batchsz;
Chris@10: 
Chris@10:      return (
Chris@10: 	  1
Chris@10: 	  && p->sz->rnk == 1
Chris@10: 	  && p->vecsz->rnk <= 1
Chris@10: 	  && p->sz->dims[0].n == desc->n
Chris@10: 	  && p->kind[0] == desc->genus->kind
Chris@10: 
Chris@10: 	  /* check strides etc */
Chris@10: 	  && X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs)
Chris@10: 
Chris@10: 	  && (batchsz = compute_batchsize(desc->n), 1)
Chris@10: 
Chris@10: 	  && (0
Chris@10: 	      /* can operate out-of-place */
Chris@10: 	      || p->I != p->O
Chris@10: 
Chris@10: 	      /* can operate in-place as long as strides are the same */
Chris@10: 	      || X(tensor_inplace_strides2)(p->sz, p->vecsz)
Chris@10: 
Chris@10: 	      /* can do it if the problem fits in the buffer, no matter
Chris@10: 		 what the strides are */
Chris@10: 	      || vl <= batchsz
Chris@10: 	       )
Chris@10: 	  );
Chris@10: }
Chris@10: 
Chris@10: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
Chris@10: {
Chris@10:      const S *ego = (const S *) ego_;
Chris@10:      P *pln;
Chris@10:      const problem_rdft *p;
Chris@10:      iodim *d;
Chris@10:      INT rs, cs, b, n;
Chris@10: 
Chris@10:      static const plan_adt padt = {
Chris@10: 	  X(rdft_solve), X(null_awake), print, destroy
Chris@10:      };
Chris@10: 
Chris@10:      UNUSED(plnr);
Chris@10: 
Chris@10:      if (ego->bufferedp) {
Chris@10: 	  if (!applicable_buf(ego_, p_))
Chris@10: 	       return (plan *)0;
Chris@10:      } else {
Chris@10: 	  if (!applicable(ego_, p_))
Chris@10: 	       return (plan *)0;
Chris@10:      }
Chris@10: 
Chris@10:      p = (const problem_rdft *) p_;
Chris@10: 
Chris@10:      if (R2HC_KINDP(p->kind[0])) {
Chris@10: 	  rs = p->sz->dims[0].is; cs = p->sz->dims[0].os;
Chris@10: 	  pln = MKPLAN_RDFT(P, &padt, 
Chris@10: 			    ego->bufferedp ? apply_buf_r2hc : apply_r2hc);
Chris@10:      } else {
Chris@10: 	  rs = p->sz->dims[0].os; cs = p->sz->dims[0].is;
Chris@10: 	  pln = MKPLAN_RDFT(P, &padt, 
Chris@10: 			    ego->bufferedp ? apply_buf_hc2r : apply_hc2r);
Chris@10:      }
Chris@10: 
Chris@10:      d = p->sz->dims;
Chris@10:      n = d[0].n;
Chris@10: 
Chris@10:      pln->k = ego->k;
Chris@10:      pln->n = n;
Chris@10: 
Chris@10:      pln->rs0 = rs;
Chris@10:      pln->rs = X(mkstride)(n, 2 * rs);
Chris@10:      pln->csr = X(mkstride)(n, cs);
Chris@10:      pln->csi = X(mkstride)(n, -cs);
Chris@10:      pln->ioffset = ioffset(p->kind[0], n, cs);
Chris@10: 
Chris@10:      b = compute_batchsize(n);
Chris@10:      pln->brs = X(mkstride)(n, 2 * b);
Chris@10:      pln->bcsr = X(mkstride)(n, b);
Chris@10:      pln->bcsi = X(mkstride)(n, -b);
Chris@10:      pln->bioffset = ioffset(p->kind[0], n, b);
Chris@10: 
Chris@10:      X(tensor_tornk1)(p->vecsz, &pln->vl, &pln->ivs, &pln->ovs);
Chris@10: 
Chris@10:      pln->slv = ego;
Chris@10:      X(ops_zero)(&pln->super.super.ops);
Chris@10: 
Chris@10:      X(ops_madd2)(pln->vl / ego->desc->genus->vl,
Chris@10: 		  &ego->desc->ops,
Chris@10: 		  &pln->super.super.ops);
Chris@10: 
Chris@10:      if (ego->bufferedp) 
Chris@10: 	  pln->super.super.ops.other += 2 * n * pln->vl;
Chris@10: 
Chris@10:      pln->super.super.could_prune_now_p = !ego->bufferedp;
Chris@10: 
Chris@10:      return &(pln->super.super);
Chris@10: }
Chris@10: 
Chris@10: /* constructor */
Chris@10: static solver *mksolver(kr2c k, const kr2c_desc *desc, int bufferedp)
Chris@10: {
Chris@10:      static const solver_adt sadt = { PROBLEM_RDFT, mkplan, 0 };
Chris@10:      S *slv = MKSOLVER(S, &sadt);
Chris@10:      slv->k = k;
Chris@10:      slv->desc = desc;
Chris@10:      slv->bufferedp = bufferedp;
Chris@10:      return &(slv->super);
Chris@10: }
Chris@10: 
Chris@10: solver *X(mksolver_rdft_r2c_direct)(kr2c k, const kr2c_desc *desc)
Chris@10: {
Chris@10:      return mksolver(k, desc, 0);
Chris@10: }
Chris@10: 
Chris@10: solver *X(mksolver_rdft_r2c_directbuf)(kr2c k, const kr2c_desc *desc)
Chris@10: {
Chris@10:      return mksolver(k, desc, 1);
Chris@10: }