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 "verify.h"
Chris@10: 
Chris@10: /* copy real A into real B, using output stride of A and input stride of B */
Chris@10: typedef struct {
Chris@10:      dotens2_closure k;
Chris@10:      R *ra;
Chris@10:      R *rb;
Chris@10: } cpyr_closure;
Chris@10: 
Chris@10: static void cpyr0(dotens2_closure *k_,
Chris@10:                   int indxa, int ondxa, int indxb, int ondxb)
Chris@10: {
Chris@10:      cpyr_closure *k = (cpyr_closure *)k_;
Chris@10:      k->rb[indxb] = k->ra[ondxa];
Chris@10:      UNUSED(indxa); UNUSED(ondxb);
Chris@10: }
Chris@10: 
Chris@10: static void cpyr(R *ra, const bench_tensor *sza, 
Chris@10: 		 R *rb, const bench_tensor *szb)
Chris@10: {
Chris@10:      cpyr_closure k;
Chris@10:      k.k.apply = cpyr0;
Chris@10:      k.ra = ra; k.rb = rb;
Chris@10:      bench_dotens2(sza, szb, &k.k);
Chris@10: }
Chris@10: 
Chris@10: /* copy unpacked halfcomplex A[n] into packed-complex B[n], using output stride
Chris@10:    of A and input stride of B.  Only copies non-redundant half; other
Chris@10:    half must be copied via mkhermitian. */
Chris@10: typedef struct {
Chris@10:      dotens2_closure k;
Chris@10:      int n;
Chris@10:      int as;
Chris@10:      int scalea;
Chris@10:      R *ra, *ia;
Chris@10:      R *rb, *ib;
Chris@10: } cpyhc2_closure;
Chris@10: 
Chris@10: static void cpyhc20(dotens2_closure *k_, 
Chris@10: 		    int indxa, int ondxa, int indxb, int ondxb)
Chris@10: {
Chris@10:      cpyhc2_closure *k = (cpyhc2_closure *)k_;
Chris@10:      int i, n = k->n;
Chris@10:      int scalea = k->scalea;
Chris@10:      int as = k->as * scalea;
Chris@10:      R *ra = k->ra + ondxa * scalea, *ia = k->ia + ondxa * scalea;
Chris@10:      R *rb = k->rb + indxb, *ib = k->ib + indxb;
Chris@10:      UNUSED(indxa); UNUSED(ondxb);
Chris@10: 
Chris@10:      for (i = 0; i < n/2 + 1; ++i) {
Chris@10: 	  rb[2*i] = ra[as*i];
Chris@10: 	  ib[2*i] = ia[as*i];
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static void cpyhc2(R *ra, R *ia,
Chris@10: 		   const bench_tensor *sza, const bench_tensor *vecsza,
Chris@10: 		   int scalea,
Chris@10: 		   R *rb, R *ib, const bench_tensor *szb)
Chris@10: {
Chris@10:      cpyhc2_closure k;
Chris@10:      BENCH_ASSERT(sza->rnk <= 1);
Chris@10:      k.k.apply = cpyhc20;
Chris@10:      k.n = tensor_sz(sza);
Chris@10:      k.scalea = scalea;
Chris@10:      if (!FINITE_RNK(sza->rnk) || sza->rnk == 0)
Chris@10: 	  k.as = 0;
Chris@10:      else
Chris@10: 	  k.as = sza->dims[0].os;
Chris@10:      k.ra = ra; k.ia = ia; k.rb = rb; k.ib = ib;
Chris@10:      bench_dotens2(vecsza, szb, &k.k);
Chris@10: }
Chris@10: 
Chris@10: /* icpyhc2 is the inverse of cpyhc2 */
Chris@10: 
Chris@10: static void icpyhc20(dotens2_closure *k_, 
Chris@10: 		     int indxa, int ondxa, int indxb, int ondxb)
Chris@10: {
Chris@10:      cpyhc2_closure *k = (cpyhc2_closure *)k_;
Chris@10:      int i, n = k->n;
Chris@10:      int scalea = k->scalea;
Chris@10:      int as = k->as * scalea;
Chris@10:      R *ra = k->ra + indxa * scalea, *ia = k->ia + indxa * scalea;
Chris@10:      R *rb = k->rb + ondxb, *ib = k->ib + ondxb;
Chris@10:      UNUSED(ondxa); UNUSED(indxb);
Chris@10: 
Chris@10:      for (i = 0; i < n/2 + 1; ++i) {
Chris@10: 	  ra[as*i] = rb[2*i];
Chris@10: 	  ia[as*i] = ib[2*i];
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static void icpyhc2(R *ra, R *ia, 
Chris@10: 		    const bench_tensor *sza, const bench_tensor *vecsza,
Chris@10: 		    int scalea,
Chris@10: 		    R *rb, R *ib, const bench_tensor *szb)
Chris@10: {
Chris@10:      cpyhc2_closure k;
Chris@10:      BENCH_ASSERT(sza->rnk <= 1);
Chris@10:      k.k.apply = icpyhc20;
Chris@10:      k.n = tensor_sz(sza);
Chris@10:      k.scalea = scalea;
Chris@10:      if (!FINITE_RNK(sza->rnk) || sza->rnk == 0)
Chris@10: 	  k.as = 0;
Chris@10:      else
Chris@10: 	  k.as = sza->dims[0].is;
Chris@10:      k.ra = ra; k.ia = ia; k.rb = rb; k.ib = ib;
Chris@10:      bench_dotens2(vecsza, szb, &k.k);
Chris@10: }
Chris@10: 
Chris@10: typedef struct {
Chris@10:      dofft_closure k;
Chris@10:      bench_problem *p;
Chris@10: } dofft_rdft2_closure;
Chris@10: 
Chris@10: static void rdft2_apply(dofft_closure *k_, 
Chris@10: 			bench_complex *in, bench_complex *out)
Chris@10: {
Chris@10:      dofft_rdft2_closure *k = (dofft_rdft2_closure *)k_;
Chris@10:      bench_problem *p = k->p;
Chris@10:      bench_tensor *totalsz, *pckdsz, *totalsz_swap, *pckdsz_swap;
Chris@10:      bench_tensor *probsz2, *totalsz2, *pckdsz2;
Chris@10:      bench_tensor *probsz2_swap, *totalsz2_swap, *pckdsz2_swap;
Chris@10:      bench_real *ri, *ii, *ro, *io;
Chris@10:      int n2, totalscale;
Chris@10: 
Chris@10:      totalsz = tensor_append(p->vecsz, p->sz);
Chris@10:      pckdsz = verify_pack(totalsz, 2);
Chris@10:      n2 = tensor_sz(totalsz);
Chris@10:      if (FINITE_RNK(p->sz->rnk) && p->sz->rnk > 0)
Chris@10: 	  n2 = (n2 / p->sz->dims[p->sz->rnk - 1].n) * 
Chris@10: 	       (p->sz->dims[p->sz->rnk - 1].n / 2 + 1);
Chris@10:      ri = (bench_real *) p->in;
Chris@10:      ro = (bench_real *) p->out;
Chris@10: 
Chris@10:      if (FINITE_RNK(p->sz->rnk) && p->sz->rnk > 0 && n2 > 0) {
Chris@10: 	  probsz2 = tensor_copy_sub(p->sz, p->sz->rnk - 1, 1);
Chris@10: 	  totalsz2 = tensor_copy_sub(totalsz, 0, totalsz->rnk - 1);
Chris@10: 	  pckdsz2 = tensor_copy_sub(pckdsz, 0, pckdsz->rnk - 1);
Chris@10:      }
Chris@10:      else {
Chris@10: 	  probsz2 = mktensor(0);
Chris@10: 	  totalsz2 = tensor_copy(totalsz);
Chris@10: 	  pckdsz2 = tensor_copy(pckdsz);
Chris@10:      }
Chris@10: 
Chris@10:      totalsz_swap = tensor_copy_swapio(totalsz);
Chris@10:      pckdsz_swap = tensor_copy_swapio(pckdsz);
Chris@10:      totalsz2_swap = tensor_copy_swapio(totalsz2);
Chris@10:      pckdsz2_swap = tensor_copy_swapio(pckdsz2);
Chris@10:      probsz2_swap = tensor_copy_swapio(probsz2);
Chris@10: 
Chris@10:      /* confusion: the stride is the distance between complex elements
Chris@10: 	when using interleaved format, but it is the distance between
Chris@10: 	real elements when using split format */
Chris@10:      if (p->split) {
Chris@10: 	  ii = p->ini ? (bench_real *) p->ini : ri + n2;
Chris@10: 	  io = p->outi ? (bench_real *) p->outi : ro + n2;
Chris@10: 	  totalscale = 1;
Chris@10:      } else {
Chris@10: 	  ii = p->ini ? (bench_real *) p->ini : ri + 1;
Chris@10: 	  io = p->outi ? (bench_real *) p->outi : ro + 1;
Chris@10: 	  totalscale = 2;
Chris@10:      }
Chris@10: 
Chris@10:      if (p->sign < 0) { /* R2HC */
Chris@10: 	  int N, vN, i;
Chris@10: 	  cpyr(&c_re(in[0]), pckdsz, ri, totalsz);
Chris@10: 	  after_problem_rcopy_from(p, ri);
Chris@10: 	  doit(1, p);
Chris@10: 	  after_problem_hccopy_to(p, ro, io);
Chris@10: 	  if (k->k.recopy_input)
Chris@10: 	       cpyr(ri, totalsz_swap, &c_re(in[0]), pckdsz_swap);
Chris@10: 	  cpyhc2(ro, io, probsz2, totalsz2, totalscale,
Chris@10: 		 &c_re(out[0]), &c_im(out[0]), pckdsz2);
Chris@10: 	  N = tensor_sz(p->sz);
Chris@10: 	  vN = tensor_sz(p->vecsz);
Chris@10: 	  for (i = 0; i < vN; ++i)
Chris@10: 	       mkhermitian(out + i*N, p->sz->rnk, p->sz->dims, 1);
Chris@10:      }
Chris@10:      else { /* HC2R */
Chris@10: 	  icpyhc2(ri, ii, probsz2, totalsz2, totalscale,
Chris@10: 		  &c_re(in[0]), &c_im(in[0]), pckdsz2);
Chris@10: 	  after_problem_hccopy_from(p, ri, ii);
Chris@10: 	  doit(1, p);
Chris@10: 	  after_problem_rcopy_to(p, ro);
Chris@10: 	  if (k->k.recopy_input)
Chris@10: 	       cpyhc2(ri, ii, probsz2_swap, totalsz2_swap, totalscale,
Chris@10: 		      &c_re(in[0]), &c_im(in[0]), pckdsz2_swap);
Chris@10: 	  mkreal(out, tensor_sz(pckdsz));
Chris@10: 	  cpyr(ro, totalsz, &c_re(out[0]), pckdsz);
Chris@10:      }
Chris@10: 
Chris@10:      tensor_destroy(totalsz);
Chris@10:      tensor_destroy(pckdsz);
Chris@10:      tensor_destroy(totalsz_swap);
Chris@10:      tensor_destroy(pckdsz_swap);
Chris@10:      tensor_destroy(probsz2);
Chris@10:      tensor_destroy(totalsz2);
Chris@10:      tensor_destroy(pckdsz2);
Chris@10:      tensor_destroy(probsz2_swap);
Chris@10:      tensor_destroy(totalsz2_swap);
Chris@10:      tensor_destroy(pckdsz2_swap);
Chris@10: }
Chris@10: 
Chris@10: void verify_rdft2(bench_problem *p, int rounds, double tol, errors *e)
Chris@10: {
Chris@10:      C *inA, *inB, *inC, *outA, *outB, *outC, *tmp;
Chris@10:      int n, vecn, N;
Chris@10:      dofft_rdft2_closure k;
Chris@10: 
Chris@10:      BENCH_ASSERT(p->kind == PROBLEM_REAL);
Chris@10: 
Chris@10:      if (!FINITE_RNK(p->sz->rnk) || !FINITE_RNK(p->vecsz->rnk))
Chris@10: 	  return;      /* give up */
Chris@10: 
Chris@10:      k.k.apply = rdft2_apply;
Chris@10:      k.k.recopy_input = 0;
Chris@10:      k.p = p;
Chris@10: 
Chris@10:      if (rounds == 0)
Chris@10: 	  rounds = 20;  /* default value */
Chris@10: 
Chris@10:      n = tensor_sz(p->sz);
Chris@10:      vecn = tensor_sz(p->vecsz);
Chris@10:      N = n * vecn;
Chris@10: 
Chris@10:      inA = (C *) bench_malloc(N * sizeof(C));
Chris@10:      inB = (C *) bench_malloc(N * sizeof(C));
Chris@10:      inC = (C *) bench_malloc(N * sizeof(C));
Chris@10:      outA = (C *) bench_malloc(N * sizeof(C));
Chris@10:      outB = (C *) bench_malloc(N * sizeof(C));
Chris@10:      outC = (C *) bench_malloc(N * sizeof(C));
Chris@10:      tmp = (C *) bench_malloc(N * sizeof(C));
Chris@10: 
Chris@10:      e->i = impulse(&k.k, n, vecn, inA, inB, inC, outA, outB, outC, 
Chris@10: 		    tmp, rounds, tol);
Chris@10:      e->l = linear(&k.k, 1, N, inA, inB, inC, outA, outB, outC,
Chris@10: 		   tmp, rounds, tol);
Chris@10: 
Chris@10:      e->s = 0.0;
Chris@10:      if (p->sign < 0)
Chris@10: 	  e->s = dmax(e->s, tf_shift(&k.k, 1, p->sz, n, vecn, p->sign,
Chris@10: 				     inA, inB, outA, outB, 
Chris@10: 				     tmp, rounds, tol, TIME_SHIFT));
Chris@10:      else
Chris@10: 	  e->s = dmax(e->s, tf_shift(&k.k, 1, p->sz, n, vecn, p->sign,
Chris@10: 				     inA, inB, outA, outB, 
Chris@10: 				     tmp, rounds, tol, FREQ_SHIFT));
Chris@10:      
Chris@10:      if (!p->in_place && !p->destroy_input)
Chris@10: 	  preserves_input(&k.k, p->sign < 0 ? mkreal : mkhermitian1,
Chris@10: 			  N, inA, inB, outB, rounds);
Chris@10: 
Chris@10:      bench_free(tmp);
Chris@10:      bench_free(outC);
Chris@10:      bench_free(outB);
Chris@10:      bench_free(outA);
Chris@10:      bench_free(inC);
Chris@10:      bench_free(inB);
Chris@10:      bench_free(inA);
Chris@10: }
Chris@10: 
Chris@10: void accuracy_rdft2(bench_problem *p, int rounds, int impulse_rounds,
Chris@10: 		    double t[6])
Chris@10: {
Chris@10:      dofft_rdft2_closure k;
Chris@10:      int n;
Chris@10:      C *a, *b;
Chris@10: 
Chris@10:      BENCH_ASSERT(p->kind == PROBLEM_REAL);
Chris@10:      BENCH_ASSERT(p->sz->rnk == 1);
Chris@10:      BENCH_ASSERT(p->vecsz->rnk == 0);
Chris@10: 
Chris@10:      k.k.apply = rdft2_apply;
Chris@10:      k.k.recopy_input = 0;
Chris@10:      k.p = p;
Chris@10:      n = tensor_sz(p->sz);
Chris@10: 
Chris@10:      a = (C *) bench_malloc(n * sizeof(C));
Chris@10:      b = (C *) bench_malloc(n * sizeof(C));
Chris@10:      accuracy_test(&k.k, p->sign < 0 ? mkreal : mkhermitian1, p->sign, 
Chris@10: 		   n, a, b, rounds, impulse_rounds, t);
Chris@10:      bench_free(b);
Chris@10:      bench_free(a);
Chris@10: }