cannam@95: /*
cannam@95:  * Copyright (c) 2003, 2007-11 Matteo Frigo
cannam@95:  * Copyright (c) 2003, 2007-11 Massachusetts Institute of Technology
cannam@95:  *
cannam@95:  * This program is free software; you can redistribute it and/or modify
cannam@95:  * it under the terms of the GNU General Public License as published by
cannam@95:  * the Free Software Foundation; either version 2 of the License, or
cannam@95:  * (at your option) any later version.
cannam@95:  *
cannam@95:  * This program is distributed in the hope that it will be useful,
cannam@95:  * but WITHOUT ANY WARRANTY; without even the implied warranty of
cannam@95:  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the
cannam@95:  * GNU General Public License for more details.
cannam@95:  *
cannam@95:  * You should have received a copy of the GNU General Public License
cannam@95:  * along with this program; if not, write to the Free Software
cannam@95:  * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA  02110-1301  USA
cannam@95:  *
cannam@95:  */
cannam@95: 
cannam@95: 
cannam@95: /* direct RDFT solver, using r2c codelets */
cannam@95: 
cannam@95: #include "rdft.h"
cannam@95: 
cannam@95: typedef struct {
cannam@95:      solver super;
cannam@95:      const kr2c_desc *desc;
cannam@95:      kr2c k;
cannam@95:      int bufferedp;
cannam@95: } S;
cannam@95: 
cannam@95: typedef struct {
cannam@95:      plan_rdft super;
cannam@95: 
cannam@95:      stride rs, csr, csi;
cannam@95:      stride brs, bcsr, bcsi;
cannam@95:      INT n, vl, rs0, ivs, ovs, ioffset, bioffset;
cannam@95:      kr2c k;
cannam@95:      const S *slv;
cannam@95: } P;
cannam@95: 
cannam@95: /*************************************************************
cannam@95:   Nonbuffered code
cannam@95:  *************************************************************/
cannam@95: static void apply_r2hc(const plan *ego_, R *I, R *O)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      ASSERT_ALIGNED_DOUBLE;
cannam@95:      ego->k(I, I + ego->rs0, O, O + ego->ioffset, 
cannam@95: 	    ego->rs, ego->csr, ego->csi,
cannam@95: 	    ego->vl, ego->ivs, ego->ovs);
cannam@95: }
cannam@95: 
cannam@95: static void apply_hc2r(const plan *ego_, R *I, R *O)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      ASSERT_ALIGNED_DOUBLE;
cannam@95:      ego->k(O, O + ego->rs0, I, I + ego->ioffset, 
cannam@95: 	    ego->rs, ego->csr, ego->csi,
cannam@95: 	    ego->vl, ego->ivs, ego->ovs);
cannam@95: }
cannam@95: 
cannam@95: /*************************************************************
cannam@95:   Buffered code
cannam@95:  *************************************************************/
cannam@95: /* should not be 2^k to avoid associativity conflicts */
cannam@95: static INT compute_batchsize(INT radix)
cannam@95: {
cannam@95:      /* round up to multiple of 4 */
cannam@95:      radix += 3;
cannam@95:      radix &= -4;
cannam@95: 
cannam@95:      return (radix + 2);
cannam@95: }
cannam@95: 
cannam@95: static void dobatch_r2hc(const P *ego, R *I, R *O, R *buf, INT batchsz)
cannam@95: {
cannam@95:      X(cpy2d_ci)(I, buf,
cannam@95: 		 ego->n, ego->rs0, WS(ego->bcsr /* hack */, 1),
cannam@95: 		 batchsz, ego->ivs, 1, 1);
cannam@95: 
cannam@95:      if (IABS(WS(ego->csr, 1)) < IABS(ego->ovs)) {
cannam@95: 	  /* transform directly to output */
cannam@95: 	  ego->k(buf, buf + WS(ego->bcsr /* hack */, 1), 
cannam@95: 		 O, O + ego->ioffset, 
cannam@95: 		 ego->brs, ego->csr, ego->csi,
cannam@95: 		 batchsz, 1, ego->ovs);
cannam@95:      } else {
cannam@95: 	  /* transform to buffer and copy back */
cannam@95: 	  ego->k(buf, buf + WS(ego->bcsr /* hack */, 1), 
cannam@95: 		 buf, buf + ego->bioffset, 
cannam@95: 		 ego->brs, ego->bcsr, ego->bcsi,
cannam@95: 		 batchsz, 1, 1);
cannam@95: 	  X(cpy2d_co)(buf, O,
cannam@95: 		      ego->n, WS(ego->bcsr, 1), WS(ego->csr, 1),  
cannam@95: 		      batchsz, 1, ego->ovs, 1);
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: static void dobatch_hc2r(const P *ego, R *I, R *O, R *buf, INT batchsz)
cannam@95: {
cannam@95:      if (IABS(WS(ego->csr, 1)) < IABS(ego->ivs)) {
cannam@95: 	  /* transform directly from input */
cannam@95: 	  ego->k(buf, buf + WS(ego->bcsr /* hack */, 1),
cannam@95: 		 I, I + ego->ioffset, 
cannam@95: 		 ego->brs, ego->csr, ego->csi,
cannam@95: 		 batchsz, ego->ivs, 1);
cannam@95:      } else {
cannam@95: 	  /* copy into buffer and transform in place */
cannam@95: 	  X(cpy2d_ci)(I, buf,
cannam@95: 		      ego->n, WS(ego->csr, 1), WS(ego->bcsr, 1),
cannam@95: 		      batchsz, ego->ivs, 1, 1);
cannam@95: 	  ego->k(buf, buf + WS(ego->bcsr /* hack */, 1),
cannam@95: 		 buf, buf + ego->bioffset, 
cannam@95: 		 ego->brs, ego->bcsr, ego->bcsi,
cannam@95: 		 batchsz, 1, 1);
cannam@95:      }
cannam@95:      X(cpy2d_co)(buf, O,
cannam@95: 		 ego->n, WS(ego->bcsr /* hack */, 1), ego->rs0,
cannam@95: 		 batchsz, 1, ego->ovs, 1);
cannam@95: }
cannam@95: 
cannam@95: static void iterate(const P *ego, R *I, R *O,
cannam@95: 		    void (*dobatch)(const P *ego, R *I, R *O, 
cannam@95: 				    R *buf, INT batchsz))
cannam@95: {
cannam@95:      R *buf;
cannam@95:      INT vl = ego->vl;
cannam@95:      INT n = ego->n;
cannam@95:      INT i;
cannam@95:      INT batchsz = compute_batchsize(n);
cannam@95:      size_t bufsz = n * batchsz * sizeof(R);
cannam@95: 
cannam@95:      BUF_ALLOC(R *, buf, bufsz);
cannam@95: 
cannam@95:      for (i = 0; i < vl - batchsz; i += batchsz) {
cannam@95: 	  dobatch(ego, I, O, buf, batchsz);
cannam@95: 	  I += batchsz * ego->ivs;
cannam@95: 	  O += batchsz * ego->ovs;
cannam@95:      }
cannam@95:      dobatch(ego, I, O, buf, vl - i);
cannam@95: 
cannam@95:      BUF_FREE(buf, bufsz);
cannam@95: }
cannam@95: 
cannam@95: static void apply_buf_r2hc(const plan *ego_, R *I, R *O)
cannam@95: {
cannam@95:      iterate((const P *) ego_, I, O, dobatch_r2hc);
cannam@95: }
cannam@95: 
cannam@95: static void apply_buf_hc2r(const plan *ego_, R *I, R *O)
cannam@95: {
cannam@95:      iterate((const P *) ego_, I, O, dobatch_hc2r);
cannam@95: }
cannam@95: 
cannam@95: static void destroy(plan *ego_)
cannam@95: {
cannam@95:      P *ego = (P *) ego_;
cannam@95:      X(stride_destroy)(ego->rs);
cannam@95:      X(stride_destroy)(ego->csr);
cannam@95:      X(stride_destroy)(ego->csi);
cannam@95:      X(stride_destroy)(ego->brs);
cannam@95:      X(stride_destroy)(ego->bcsr);
cannam@95:      X(stride_destroy)(ego->bcsi);
cannam@95: }
cannam@95: 
cannam@95: static void print(const plan *ego_, printer *p)
cannam@95: {
cannam@95:      const P *ego = (const P *) ego_;
cannam@95:      const S *s = ego->slv;
cannam@95: 
cannam@95:      if (ego->slv->bufferedp)
cannam@95: 	  p->print(p, "(rdft-%s-directbuf/%D-r2c-%D%v \"%s\")", 
cannam@95: 		   X(rdft_kind_str)(s->desc->genus->kind), 
cannam@95: 		   /* hack */ WS(ego->bcsr, 1), ego->n, 
cannam@95: 		   ego->vl, s->desc->nam);
cannam@95: 
cannam@95:      else 
cannam@95: 	  p->print(p, "(rdft-%s-direct-r2c-%D%v \"%s\")", 
cannam@95: 		   X(rdft_kind_str)(s->desc->genus->kind), ego->n, 
cannam@95: 		   ego->vl, s->desc->nam);
cannam@95: }
cannam@95: 
cannam@95: static INT ioffset(rdft_kind kind, INT sz, INT s)
cannam@95: {
cannam@95:      return(s * ((kind == R2HC || kind == HC2R) ? sz : (sz - 1)));
cannam@95: }
cannam@95: 
cannam@95: static int applicable(const solver *ego_, const problem *p_)
cannam@95: {
cannam@95:      const S *ego = (const S *) ego_;
cannam@95:      const kr2c_desc *desc = ego->desc;
cannam@95:      const problem_rdft *p = (const problem_rdft *) p_;
cannam@95:      INT vl, ivs, ovs;
cannam@95: 
cannam@95:      return (
cannam@95: 	  1
cannam@95: 	  && p->sz->rnk == 1
cannam@95: 	  && p->vecsz->rnk <= 1
cannam@95: 	  && p->sz->dims[0].n == desc->n
cannam@95: 	  && p->kind[0] == desc->genus->kind
cannam@95: 
cannam@95: 	  /* check strides etc */
cannam@95: 	  && X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs)
cannam@95: 
cannam@95: 	  && (0
cannam@95: 	      /* can operate out-of-place */
cannam@95: 	      || p->I != p->O
cannam@95: 
cannam@95: 	      /* computing one transform */
cannam@95: 	      || vl == 1
cannam@95: 
cannam@95: 	      /* can operate in-place as long as strides are the same */
cannam@95: 	      || X(tensor_inplace_strides2)(p->sz, p->vecsz)
cannam@95: 	       )
cannam@95: 	  );
cannam@95: }
cannam@95: 
cannam@95: static int applicable_buf(const solver *ego_, const problem *p_)
cannam@95: {
cannam@95:      const S *ego = (const S *) ego_;
cannam@95:      const kr2c_desc *desc = ego->desc;
cannam@95:      const problem_rdft *p = (const problem_rdft *) p_;
cannam@95:      INT vl, ivs, ovs, batchsz;
cannam@95: 
cannam@95:      return (
cannam@95: 	  1
cannam@95: 	  && p->sz->rnk == 1
cannam@95: 	  && p->vecsz->rnk <= 1
cannam@95: 	  && p->sz->dims[0].n == desc->n
cannam@95: 	  && p->kind[0] == desc->genus->kind
cannam@95: 
cannam@95: 	  /* check strides etc */
cannam@95: 	  && X(tensor_tornk1)(p->vecsz, &vl, &ivs, &ovs)
cannam@95: 
cannam@95: 	  && (batchsz = compute_batchsize(desc->n), 1)
cannam@95: 
cannam@95: 	  && (0
cannam@95: 	      /* can operate out-of-place */
cannam@95: 	      || p->I != p->O
cannam@95: 
cannam@95: 	      /* can operate in-place as long as strides are the same */
cannam@95: 	      || X(tensor_inplace_strides2)(p->sz, p->vecsz)
cannam@95: 
cannam@95: 	      /* can do it if the problem fits in the buffer, no matter
cannam@95: 		 what the strides are */
cannam@95: 	      || vl <= batchsz
cannam@95: 	       )
cannam@95: 	  );
cannam@95: }
cannam@95: 
cannam@95: static plan *mkplan(const solver *ego_, const problem *p_, planner *plnr)
cannam@95: {
cannam@95:      const S *ego = (const S *) ego_;
cannam@95:      P *pln;
cannam@95:      const problem_rdft *p;
cannam@95:      iodim *d;
cannam@95:      INT rs, cs, b, n;
cannam@95: 
cannam@95:      static const plan_adt padt = {
cannam@95: 	  X(rdft_solve), X(null_awake), print, destroy
cannam@95:      };
cannam@95: 
cannam@95:      UNUSED(plnr);
cannam@95: 
cannam@95:      if (ego->bufferedp) {
cannam@95: 	  if (!applicable_buf(ego_, p_))
cannam@95: 	       return (plan *)0;
cannam@95:      } else {
cannam@95: 	  if (!applicable(ego_, p_))
cannam@95: 	       return (plan *)0;
cannam@95:      }
cannam@95: 
cannam@95:      p = (const problem_rdft *) p_;
cannam@95: 
cannam@95:      if (R2HC_KINDP(p->kind[0])) {
cannam@95: 	  rs = p->sz->dims[0].is; cs = p->sz->dims[0].os;
cannam@95: 	  pln = MKPLAN_RDFT(P, &padt, 
cannam@95: 			    ego->bufferedp ? apply_buf_r2hc : apply_r2hc);
cannam@95:      } else {
cannam@95: 	  rs = p->sz->dims[0].os; cs = p->sz->dims[0].is;
cannam@95: 	  pln = MKPLAN_RDFT(P, &padt, 
cannam@95: 			    ego->bufferedp ? apply_buf_hc2r : apply_hc2r);
cannam@95:      }
cannam@95: 
cannam@95:      d = p->sz->dims;
cannam@95:      n = d[0].n;
cannam@95: 
cannam@95:      pln->k = ego->k;
cannam@95:      pln->n = n;
cannam@95: 
cannam@95:      pln->rs0 = rs;
cannam@95:      pln->rs = X(mkstride)(n, 2 * rs);
cannam@95:      pln->csr = X(mkstride)(n, cs);
cannam@95:      pln->csi = X(mkstride)(n, -cs);
cannam@95:      pln->ioffset = ioffset(p->kind[0], n, cs);
cannam@95: 
cannam@95:      b = compute_batchsize(n);
cannam@95:      pln->brs = X(mkstride)(n, 2 * b);
cannam@95:      pln->bcsr = X(mkstride)(n, b);
cannam@95:      pln->bcsi = X(mkstride)(n, -b);
cannam@95:      pln->bioffset = ioffset(p->kind[0], n, b);
cannam@95: 
cannam@95:      X(tensor_tornk1)(p->vecsz, &pln->vl, &pln->ivs, &pln->ovs);
cannam@95: 
cannam@95:      pln->slv = ego;
cannam@95:      X(ops_zero)(&pln->super.super.ops);
cannam@95: 
cannam@95:      X(ops_madd2)(pln->vl / ego->desc->genus->vl,
cannam@95: 		  &ego->desc->ops,
cannam@95: 		  &pln->super.super.ops);
cannam@95: 
cannam@95:      if (ego->bufferedp) 
cannam@95: 	  pln->super.super.ops.other += 2 * n * pln->vl;
cannam@95: 
cannam@95:      pln->super.super.could_prune_now_p = !ego->bufferedp;
cannam@95: 
cannam@95:      return &(pln->super.super);
cannam@95: }
cannam@95: 
cannam@95: /* constructor */
cannam@95: static solver *mksolver(kr2c k, const kr2c_desc *desc, int bufferedp)
cannam@95: {
cannam@95:      static const solver_adt sadt = { PROBLEM_RDFT, mkplan, 0 };
cannam@95:      S *slv = MKSOLVER(S, &sadt);
cannam@95:      slv->k = k;
cannam@95:      slv->desc = desc;
cannam@95:      slv->bufferedp = bufferedp;
cannam@95:      return &(slv->super);
cannam@95: }
cannam@95: 
cannam@95: solver *X(mksolver_rdft_r2c_direct)(kr2c k, const kr2c_desc *desc)
cannam@95: {
cannam@95:      return mksolver(k, desc, 0);
cannam@95: }
cannam@95: 
cannam@95: solver *X(mksolver_rdft_r2c_directbuf)(kr2c k, const kr2c_desc *desc)
cannam@95: {
cannam@95:      return mksolver(k, desc, 1);
cannam@95: }