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 "verify.h"
cannam@95: 
cannam@95: /* copy real A into real B, using output stride of A and input stride of B */
cannam@95: typedef struct {
cannam@95:      dotens2_closure k;
cannam@95:      R *ra;
cannam@95:      R *rb;
cannam@95: } cpyr_closure;
cannam@95: 
cannam@95: static void cpyr0(dotens2_closure *k_,
cannam@95:                   int indxa, int ondxa, int indxb, int ondxb)
cannam@95: {
cannam@95:      cpyr_closure *k = (cpyr_closure *)k_;
cannam@95:      k->rb[indxb] = k->ra[ondxa];
cannam@95:      UNUSED(indxa); UNUSED(ondxb);
cannam@95: }
cannam@95: 
cannam@95: static void cpyr(R *ra, const bench_tensor *sza, 
cannam@95: 		 R *rb, const bench_tensor *szb)
cannam@95: {
cannam@95:      cpyr_closure k;
cannam@95:      k.k.apply = cpyr0;
cannam@95:      k.ra = ra; k.rb = rb;
cannam@95:      bench_dotens2(sza, szb, &k.k);
cannam@95: }
cannam@95: 
cannam@95: /* copy unpacked halfcomplex A[n] into packed-complex B[n], using output stride
cannam@95:    of A and input stride of B.  Only copies non-redundant half; other
cannam@95:    half must be copied via mkhermitian. */
cannam@95: typedef struct {
cannam@95:      dotens2_closure k;
cannam@95:      int n;
cannam@95:      int as;
cannam@95:      int scalea;
cannam@95:      R *ra, *ia;
cannam@95:      R *rb, *ib;
cannam@95: } cpyhc2_closure;
cannam@95: 
cannam@95: static void cpyhc20(dotens2_closure *k_, 
cannam@95: 		    int indxa, int ondxa, int indxb, int ondxb)
cannam@95: {
cannam@95:      cpyhc2_closure *k = (cpyhc2_closure *)k_;
cannam@95:      int i, n = k->n;
cannam@95:      int scalea = k->scalea;
cannam@95:      int as = k->as * scalea;
cannam@95:      R *ra = k->ra + ondxa * scalea, *ia = k->ia + ondxa * scalea;
cannam@95:      R *rb = k->rb + indxb, *ib = k->ib + indxb;
cannam@95:      UNUSED(indxa); UNUSED(ondxb);
cannam@95: 
cannam@95:      for (i = 0; i < n/2 + 1; ++i) {
cannam@95: 	  rb[2*i] = ra[as*i];
cannam@95: 	  ib[2*i] = ia[as*i];
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: static void cpyhc2(R *ra, R *ia,
cannam@95: 		   const bench_tensor *sza, const bench_tensor *vecsza,
cannam@95: 		   int scalea,
cannam@95: 		   R *rb, R *ib, const bench_tensor *szb)
cannam@95: {
cannam@95:      cpyhc2_closure k;
cannam@95:      BENCH_ASSERT(sza->rnk <= 1);
cannam@95:      k.k.apply = cpyhc20;
cannam@95:      k.n = tensor_sz(sza);
cannam@95:      k.scalea = scalea;
cannam@95:      if (!FINITE_RNK(sza->rnk) || sza->rnk == 0)
cannam@95: 	  k.as = 0;
cannam@95:      else
cannam@95: 	  k.as = sza->dims[0].os;
cannam@95:      k.ra = ra; k.ia = ia; k.rb = rb; k.ib = ib;
cannam@95:      bench_dotens2(vecsza, szb, &k.k);
cannam@95: }
cannam@95: 
cannam@95: /* icpyhc2 is the inverse of cpyhc2 */
cannam@95: 
cannam@95: static void icpyhc20(dotens2_closure *k_, 
cannam@95: 		     int indxa, int ondxa, int indxb, int ondxb)
cannam@95: {
cannam@95:      cpyhc2_closure *k = (cpyhc2_closure *)k_;
cannam@95:      int i, n = k->n;
cannam@95:      int scalea = k->scalea;
cannam@95:      int as = k->as * scalea;
cannam@95:      R *ra = k->ra + indxa * scalea, *ia = k->ia + indxa * scalea;
cannam@95:      R *rb = k->rb + ondxb, *ib = k->ib + ondxb;
cannam@95:      UNUSED(ondxa); UNUSED(indxb);
cannam@95: 
cannam@95:      for (i = 0; i < n/2 + 1; ++i) {
cannam@95: 	  ra[as*i] = rb[2*i];
cannam@95: 	  ia[as*i] = ib[2*i];
cannam@95:      }
cannam@95: }
cannam@95: 
cannam@95: static void icpyhc2(R *ra, R *ia, 
cannam@95: 		    const bench_tensor *sza, const bench_tensor *vecsza,
cannam@95: 		    int scalea,
cannam@95: 		    R *rb, R *ib, const bench_tensor *szb)
cannam@95: {
cannam@95:      cpyhc2_closure k;
cannam@95:      BENCH_ASSERT(sza->rnk <= 1);
cannam@95:      k.k.apply = icpyhc20;
cannam@95:      k.n = tensor_sz(sza);
cannam@95:      k.scalea = scalea;
cannam@95:      if (!FINITE_RNK(sza->rnk) || sza->rnk == 0)
cannam@95: 	  k.as = 0;
cannam@95:      else
cannam@95: 	  k.as = sza->dims[0].is;
cannam@95:      k.ra = ra; k.ia = ia; k.rb = rb; k.ib = ib;
cannam@95:      bench_dotens2(vecsza, szb, &k.k);
cannam@95: }
cannam@95: 
cannam@95: typedef struct {
cannam@95:      dofft_closure k;
cannam@95:      bench_problem *p;
cannam@95: } dofft_rdft2_closure;
cannam@95: 
cannam@95: static void rdft2_apply(dofft_closure *k_, 
cannam@95: 			bench_complex *in, bench_complex *out)
cannam@95: {
cannam@95:      dofft_rdft2_closure *k = (dofft_rdft2_closure *)k_;
cannam@95:      bench_problem *p = k->p;
cannam@95:      bench_tensor *totalsz, *pckdsz, *totalsz_swap, *pckdsz_swap;
cannam@95:      bench_tensor *probsz2, *totalsz2, *pckdsz2;
cannam@95:      bench_tensor *probsz2_swap, *totalsz2_swap, *pckdsz2_swap;
cannam@95:      bench_real *ri, *ii, *ro, *io;
cannam@95:      int n2, totalscale;
cannam@95: 
cannam@95:      totalsz = tensor_append(p->vecsz, p->sz);
cannam@95:      pckdsz = verify_pack(totalsz, 2);
cannam@95:      n2 = tensor_sz(totalsz);
cannam@95:      if (FINITE_RNK(p->sz->rnk) && p->sz->rnk > 0)
cannam@95: 	  n2 = (n2 / p->sz->dims[p->sz->rnk - 1].n) * 
cannam@95: 	       (p->sz->dims[p->sz->rnk - 1].n / 2 + 1);
cannam@95:      ri = (bench_real *) p->in;
cannam@95:      ro = (bench_real *) p->out;
cannam@95: 
cannam@95:      if (FINITE_RNK(p->sz->rnk) && p->sz->rnk > 0 && n2 > 0) {
cannam@95: 	  probsz2 = tensor_copy_sub(p->sz, p->sz->rnk - 1, 1);
cannam@95: 	  totalsz2 = tensor_copy_sub(totalsz, 0, totalsz->rnk - 1);
cannam@95: 	  pckdsz2 = tensor_copy_sub(pckdsz, 0, pckdsz->rnk - 1);
cannam@95:      }
cannam@95:      else {
cannam@95: 	  probsz2 = mktensor(0);
cannam@95: 	  totalsz2 = tensor_copy(totalsz);
cannam@95: 	  pckdsz2 = tensor_copy(pckdsz);
cannam@95:      }
cannam@95: 
cannam@95:      totalsz_swap = tensor_copy_swapio(totalsz);
cannam@95:      pckdsz_swap = tensor_copy_swapio(pckdsz);
cannam@95:      totalsz2_swap = tensor_copy_swapio(totalsz2);
cannam@95:      pckdsz2_swap = tensor_copy_swapio(pckdsz2);
cannam@95:      probsz2_swap = tensor_copy_swapio(probsz2);
cannam@95: 
cannam@95:      /* confusion: the stride is the distance between complex elements
cannam@95: 	when using interleaved format, but it is the distance between
cannam@95: 	real elements when using split format */
cannam@95:      if (p->split) {
cannam@95: 	  ii = p->ini ? (bench_real *) p->ini : ri + n2;
cannam@95: 	  io = p->outi ? (bench_real *) p->outi : ro + n2;
cannam@95: 	  totalscale = 1;
cannam@95:      } else {
cannam@95: 	  ii = p->ini ? (bench_real *) p->ini : ri + 1;
cannam@95: 	  io = p->outi ? (bench_real *) p->outi : ro + 1;
cannam@95: 	  totalscale = 2;
cannam@95:      }
cannam@95: 
cannam@95:      if (p->sign < 0) { /* R2HC */
cannam@95: 	  int N, vN, i;
cannam@95: 	  cpyr(&c_re(in[0]), pckdsz, ri, totalsz);
cannam@95: 	  after_problem_rcopy_from(p, ri);
cannam@95: 	  doit(1, p);
cannam@95: 	  after_problem_hccopy_to(p, ro, io);
cannam@95: 	  if (k->k.recopy_input)
cannam@95: 	       cpyr(ri, totalsz_swap, &c_re(in[0]), pckdsz_swap);
cannam@95: 	  cpyhc2(ro, io, probsz2, totalsz2, totalscale,
cannam@95: 		 &c_re(out[0]), &c_im(out[0]), pckdsz2);
cannam@95: 	  N = tensor_sz(p->sz);
cannam@95: 	  vN = tensor_sz(p->vecsz);
cannam@95: 	  for (i = 0; i < vN; ++i)
cannam@95: 	       mkhermitian(out + i*N, p->sz->rnk, p->sz->dims, 1);
cannam@95:      }
cannam@95:      else { /* HC2R */
cannam@95: 	  icpyhc2(ri, ii, probsz2, totalsz2, totalscale,
cannam@95: 		  &c_re(in[0]), &c_im(in[0]), pckdsz2);
cannam@95: 	  after_problem_hccopy_from(p, ri, ii);
cannam@95: 	  doit(1, p);
cannam@95: 	  after_problem_rcopy_to(p, ro);
cannam@95: 	  if (k->k.recopy_input)
cannam@95: 	       cpyhc2(ri, ii, probsz2_swap, totalsz2_swap, totalscale,
cannam@95: 		      &c_re(in[0]), &c_im(in[0]), pckdsz2_swap);
cannam@95: 	  mkreal(out, tensor_sz(pckdsz));
cannam@95: 	  cpyr(ro, totalsz, &c_re(out[0]), pckdsz);
cannam@95:      }
cannam@95: 
cannam@95:      tensor_destroy(totalsz);
cannam@95:      tensor_destroy(pckdsz);
cannam@95:      tensor_destroy(totalsz_swap);
cannam@95:      tensor_destroy(pckdsz_swap);
cannam@95:      tensor_destroy(probsz2);
cannam@95:      tensor_destroy(totalsz2);
cannam@95:      tensor_destroy(pckdsz2);
cannam@95:      tensor_destroy(probsz2_swap);
cannam@95:      tensor_destroy(totalsz2_swap);
cannam@95:      tensor_destroy(pckdsz2_swap);
cannam@95: }
cannam@95: 
cannam@95: void verify_rdft2(bench_problem *p, int rounds, double tol, errors *e)
cannam@95: {
cannam@95:      C *inA, *inB, *inC, *outA, *outB, *outC, *tmp;
cannam@95:      int n, vecn, N;
cannam@95:      dofft_rdft2_closure k;
cannam@95: 
cannam@95:      BENCH_ASSERT(p->kind == PROBLEM_REAL);
cannam@95: 
cannam@95:      if (!FINITE_RNK(p->sz->rnk) || !FINITE_RNK(p->vecsz->rnk))
cannam@95: 	  return;      /* give up */
cannam@95: 
cannam@95:      k.k.apply = rdft2_apply;
cannam@95:      k.k.recopy_input = 0;
cannam@95:      k.p = p;
cannam@95: 
cannam@95:      if (rounds == 0)
cannam@95: 	  rounds = 20;  /* default value */
cannam@95: 
cannam@95:      n = tensor_sz(p->sz);
cannam@95:      vecn = tensor_sz(p->vecsz);
cannam@95:      N = n * vecn;
cannam@95: 
cannam@95:      inA = (C *) bench_malloc(N * sizeof(C));
cannam@95:      inB = (C *) bench_malloc(N * sizeof(C));
cannam@95:      inC = (C *) bench_malloc(N * sizeof(C));
cannam@95:      outA = (C *) bench_malloc(N * sizeof(C));
cannam@95:      outB = (C *) bench_malloc(N * sizeof(C));
cannam@95:      outC = (C *) bench_malloc(N * sizeof(C));
cannam@95:      tmp = (C *) bench_malloc(N * sizeof(C));
cannam@95: 
cannam@95:      e->i = impulse(&k.k, n, vecn, inA, inB, inC, outA, outB, outC, 
cannam@95: 		    tmp, rounds, tol);
cannam@95:      e->l = linear(&k.k, 1, N, inA, inB, inC, outA, outB, outC,
cannam@95: 		   tmp, rounds, tol);
cannam@95: 
cannam@95:      e->s = 0.0;
cannam@95:      if (p->sign < 0)
cannam@95: 	  e->s = dmax(e->s, tf_shift(&k.k, 1, p->sz, n, vecn, p->sign,
cannam@95: 				     inA, inB, outA, outB, 
cannam@95: 				     tmp, rounds, tol, TIME_SHIFT));
cannam@95:      else
cannam@95: 	  e->s = dmax(e->s, tf_shift(&k.k, 1, p->sz, n, vecn, p->sign,
cannam@95: 				     inA, inB, outA, outB, 
cannam@95: 				     tmp, rounds, tol, FREQ_SHIFT));
cannam@95:      
cannam@95:      if (!p->in_place && !p->destroy_input)
cannam@95: 	  preserves_input(&k.k, p->sign < 0 ? mkreal : mkhermitian1,
cannam@95: 			  N, inA, inB, outB, rounds);
cannam@95: 
cannam@95:      bench_free(tmp);
cannam@95:      bench_free(outC);
cannam@95:      bench_free(outB);
cannam@95:      bench_free(outA);
cannam@95:      bench_free(inC);
cannam@95:      bench_free(inB);
cannam@95:      bench_free(inA);
cannam@95: }
cannam@95: 
cannam@95: void accuracy_rdft2(bench_problem *p, int rounds, int impulse_rounds,
cannam@95: 		    double t[6])
cannam@95: {
cannam@95:      dofft_rdft2_closure k;
cannam@95:      int n;
cannam@95:      C *a, *b;
cannam@95: 
cannam@95:      BENCH_ASSERT(p->kind == PROBLEM_REAL);
cannam@95:      BENCH_ASSERT(p->sz->rnk == 1);
cannam@95:      BENCH_ASSERT(p->vecsz->rnk == 0);
cannam@95: 
cannam@95:      k.k.apply = rdft2_apply;
cannam@95:      k.k.recopy_input = 0;
cannam@95:      k.p = p;
cannam@95:      n = tensor_sz(p->sz);
cannam@95: 
cannam@95:      a = (C *) bench_malloc(n * sizeof(C));
cannam@95:      b = (C *) bench_malloc(n * sizeof(C));
cannam@95:      accuracy_test(&k.k, p->sign < 0 ? mkreal : mkhermitian1, p->sign, 
cannam@95: 		   n, a, b, rounds, impulse_rounds, t);
cannam@95:      bench_free(b);
cannam@95:      bench_free(a);
cannam@95: }