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: #include "rdft.h"
cannam@95: #include <stddef.h>
cannam@95: 
cannam@95: static void destroy(problem *ego_)
cannam@95: {
cannam@95:      problem_rdft *ego = (problem_rdft *) ego_;
cannam@95: #if !defined(STRUCT_HACK_C99) && !defined(STRUCT_HACK_KR)
cannam@95:      X(ifree0)(ego->kind);
cannam@95: #endif
cannam@95:      X(tensor_destroy2)(ego->vecsz, ego->sz);
cannam@95:      X(ifree)(ego_);
cannam@95: }
cannam@95: 
cannam@95: static void kind_hash(md5 *m, const rdft_kind *kind, int rnk)
cannam@95: {
cannam@95:      int i;
cannam@95:      for (i = 0; i < rnk; ++i)
cannam@95: 	  X(md5int)(m, kind[i]);
cannam@95: }
cannam@95: 
cannam@95: static void hash(const problem *p_, md5 *m)
cannam@95: {
cannam@95:      const problem_rdft *p = (const problem_rdft *) p_;
cannam@95:      X(md5puts)(m, "rdft");
cannam@95:      X(md5int)(m, p->I == p->O);
cannam@95:      kind_hash(m, p->kind, p->sz->rnk);
cannam@95:      X(md5int)(m, X(alignment_of)(p->I));
cannam@95:      X(md5int)(m, X(alignment_of)(p->O));
cannam@95:      X(tensor_md5)(m, p->sz);
cannam@95:      X(tensor_md5)(m, p->vecsz);
cannam@95: }
cannam@95: 
cannam@95: static void recur(const iodim *dims, int rnk, R *I)
cannam@95: {
cannam@95:      if (rnk == RNK_MINFTY)
cannam@95:           return;
cannam@95:      else if (rnk == 0)
cannam@95:           I[0] = K(0.0);
cannam@95:      else if (rnk > 0) {
cannam@95:           INT i, n = dims[0].n, is = dims[0].is;
cannam@95: 
cannam@95: 	  if (rnk == 1) {
cannam@95: 	       /* this case is redundant but faster */
cannam@95: 	       for (i = 0; i < n; ++i)
cannam@95: 		    I[i * is] = K(0.0);
cannam@95: 	  } else {
cannam@95: 	       for (i = 0; i < n; ++i)
cannam@95: 		    recur(dims + 1, rnk - 1, I + i * is);
cannam@95: 	  }
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: void X(rdft_zerotens)(tensor *sz, R *I)
cannam@95: {
cannam@95:      recur(sz->dims, sz->rnk, I);
cannam@95: }
cannam@95: 
cannam@95: #define KSTR_LEN 8
cannam@95: 
cannam@95: const char *X(rdft_kind_str)(rdft_kind kind)
cannam@95: {
cannam@95:      static const char kstr[][KSTR_LEN] = {
cannam@95: 	  "r2hc", "r2hc01", "r2hc10", "r2hc11",
cannam@95: 	  "hc2r", "hc2r01", "hc2r10", "hc2r11",
cannam@95: 	  "dht",
cannam@95: 	  "redft00", "redft01", "redft10", "redft11",
cannam@95: 	  "rodft00", "rodft01", "rodft10", "rodft11"
cannam@95:      };
cannam@95:      A(kind >= 0 && kind < sizeof(kstr) / KSTR_LEN);
cannam@95:      return kstr[kind];
cannam@95: }
cannam@95: 
cannam@95: static void print(const problem *ego_, printer *p)
cannam@95: {
cannam@95:      const problem_rdft *ego = (const problem_rdft *) ego_;
cannam@95:      int i;
cannam@95:      p->print(p, "(rdft %d %D %T %T", 
cannam@95: 	      X(alignment_of)(ego->I),
cannam@95: 	      (INT)(ego->O - ego->I), 
cannam@95: 	      ego->sz,
cannam@95: 	      ego->vecsz);
cannam@95:      for (i = 0; i < ego->sz->rnk; ++i)
cannam@95: 	  p->print(p, " %d", (int)ego->kind[i]);
cannam@95:      p->print(p, ")");
cannam@95: }
cannam@95: 
cannam@95: static void zero(const problem *ego_)
cannam@95: {
cannam@95:      const problem_rdft *ego = (const problem_rdft *) ego_;
cannam@95:      tensor *sz = X(tensor_append)(ego->vecsz, ego->sz);
cannam@95:      X(rdft_zerotens)(sz, UNTAINT(ego->I));
cannam@95:      X(tensor_destroy)(sz);
cannam@95: }
cannam@95: 
cannam@95: static const problem_adt padt =
cannam@95: {
cannam@95:      PROBLEM_RDFT,
cannam@95:      hash,
cannam@95:      zero,
cannam@95:      print,
cannam@95:      destroy
cannam@95: };
cannam@95: 
cannam@95: /* Dimensions of size 1 that are not REDFT/RODFT are no-ops and can be
cannam@95:    eliminated.  REDFT/RODFT unit dimensions often have factors of 2.0
cannam@95:    and suchlike from normalization and phases, although in principle
cannam@95:    these constant factors from different dimensions could be combined. */
cannam@95: static int nontrivial(const iodim *d, rdft_kind kind)
cannam@95: {
cannam@95:      return (d->n > 1 || kind == R2HC11 || kind == HC2R11
cannam@95: 	     || (REODFT_KINDP(kind) && kind != REDFT01 && kind != RODFT01));
cannam@95: }
cannam@95: 
cannam@95: problem *X(mkproblem_rdft)(const tensor *sz, const tensor *vecsz,
cannam@95: 			   R *I, R *O, const rdft_kind *kind)
cannam@95: {
cannam@95:      problem_rdft *ego;
cannam@95:      int rnk = sz->rnk;
cannam@95:      int i;
cannam@95: 
cannam@95:      A(X(tensor_kosherp)(sz));
cannam@95:      A(X(tensor_kosherp)(vecsz));
cannam@95:      A(FINITE_RNK(sz->rnk));
cannam@95: 
cannam@95:      if (UNTAINT(I) == UNTAINT(O))
cannam@95: 	  I = O = JOIN_TAINT(I, O);
cannam@95: 
cannam@95:      if (I == O && !X(tensor_inplace_locations)(sz, vecsz))
cannam@95: 	  return X(mkproblem_unsolvable)();
cannam@95: 
cannam@95:      for (i = rnk = 0; i < sz->rnk; ++i) {
cannam@95:           A(sz->dims[i].n > 0);
cannam@95:           if (nontrivial(sz->dims + i, kind[i]))
cannam@95:                ++rnk;
cannam@95:      }
cannam@95: 
cannam@95: #if defined(STRUCT_HACK_KR)
cannam@95:      ego = (problem_rdft *) X(mkproblem)(sizeof(problem_rdft)
cannam@95: 					 + sizeof(rdft_kind)
cannam@95: 					 * (rnk > 0 ? rnk - 1 : 0), &padt);
cannam@95: #elif defined(STRUCT_HACK_C99)
cannam@95:      ego = (problem_rdft *) X(mkproblem)(sizeof(problem_rdft)
cannam@95: 					 + sizeof(rdft_kind) * rnk, &padt);
cannam@95: #else
cannam@95:      ego = (problem_rdft *) X(mkproblem)(sizeof(problem_rdft), &padt);
cannam@95:      ego->kind = (rdft_kind *) MALLOC(sizeof(rdft_kind) * rnk, PROBLEMS);
cannam@95: #endif
cannam@95: 
cannam@95:      /* do compression and sorting as in X(tensor_compress), but take
cannam@95: 	transform kind into account (sigh) */
cannam@95:      ego->sz = X(mktensor)(rnk);
cannam@95:      for (i = rnk = 0; i < sz->rnk; ++i) {
cannam@95:           if (nontrivial(sz->dims + i, kind[i])) {
cannam@95: 	       ego->kind[rnk] = kind[i];
cannam@95:                ego->sz->dims[rnk++] = sz->dims[i];
cannam@95: 	  }
cannam@95:      }
cannam@95:      for (i = 0; i + 1 < rnk; ++i) {
cannam@95: 	  int j;
cannam@95: 	  for (j = i + 1; j < rnk; ++j)
cannam@95: 	       if (X(dimcmp)(ego->sz->dims + i, ego->sz->dims + j) > 0) {
cannam@95: 		    iodim dswap;
cannam@95: 		    rdft_kind kswap;
cannam@95: 		    dswap = ego->sz->dims[i];
cannam@95: 		    ego->sz->dims[i] = ego->sz->dims[j];
cannam@95: 		    ego->sz->dims[j] = dswap;
cannam@95: 		    kswap = ego->kind[i];
cannam@95: 		    ego->kind[i] = ego->kind[j];
cannam@95: 		    ego->kind[j] = kswap;
cannam@95: 	       }
cannam@95:      }
cannam@95: 
cannam@95:      for (i = 0; i < rnk; ++i)
cannam@95: 	  if (ego->sz->dims[i].n == 2 && (ego->kind[i] == REDFT00
cannam@95: 					  || ego->kind[i] == DHT
cannam@95: 					  || ego->kind[i] == HC2R))
cannam@95: 	       ego->kind[i] = R2HC; /* size-2 transforms are equivalent */
cannam@95: 
cannam@95:      ego->vecsz = X(tensor_compress_contiguous)(vecsz);
cannam@95:      ego->I = I;
cannam@95:      ego->O = O;
cannam@95: 
cannam@95:      A(FINITE_RNK(ego->sz->rnk));
cannam@95: 
cannam@95:      return &(ego->super);
cannam@95: }
cannam@95: 
cannam@95: /* Same as X(mkproblem_rdft), but also destroy input tensors. */
cannam@95: problem *X(mkproblem_rdft_d)(tensor *sz, tensor *vecsz,
cannam@95: 			     R *I, R *O, const rdft_kind *kind)
cannam@95: {
cannam@95:      problem *p = X(mkproblem_rdft)(sz, vecsz, I, O, kind);
cannam@95:      X(tensor_destroy2)(vecsz, sz);
cannam@95:      return p;
cannam@95: }
cannam@95: 
cannam@95: /* As above, but for rnk <= 1 only and takes a scalar kind parameter */
cannam@95: problem *X(mkproblem_rdft_1)(const tensor *sz, const tensor *vecsz,
cannam@95: 			     R *I, R *O, rdft_kind kind)
cannam@95: {
cannam@95:      A(sz->rnk <= 1);
cannam@95:      return X(mkproblem_rdft)(sz, vecsz, I, O, &kind);
cannam@95: }
cannam@95: 
cannam@95: problem *X(mkproblem_rdft_1_d)(tensor *sz, tensor *vecsz,
cannam@95: 			       R *I, R *O, rdft_kind kind)
cannam@95: {
cannam@95:      A(sz->rnk <= 1);
cannam@95:      return X(mkproblem_rdft_d)(sz, vecsz, I, O, &kind);
cannam@95: }
cannam@95: 
cannam@95: /* create a zero-dimensional problem */
cannam@95: problem *X(mkproblem_rdft_0_d)(tensor *vecsz, R *I, R *O)
cannam@95: {
cannam@95:      return X(mkproblem_rdft_d)(X(mktensor_0d)(), vecsz, I, O, 
cannam@95: 				(const rdft_kind *)0);
cannam@95: }