Chris@10: /*
Chris@10:  * Copyright (c) 2001 Matteo Frigo
Chris@10:  * Copyright (c) 2001 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: #include "bench.h"
Chris@10: #include <stdlib.h>
Chris@10: 
Chris@10: bench_tensor *mktensor(int rnk) 
Chris@10: {
Chris@10:      bench_tensor *x;
Chris@10: 
Chris@10:      BENCH_ASSERT(rnk >= 0);
Chris@10: 
Chris@10:      x = (bench_tensor *)bench_malloc(sizeof(bench_tensor));
Chris@10:      if (FINITE_RNK(rnk) && rnk > 0)
Chris@10:           x->dims = (bench_iodim *)bench_malloc(sizeof(bench_iodim) * rnk);
Chris@10:      else
Chris@10:           x->dims = 0;
Chris@10: 
Chris@10:      x->rnk = rnk;
Chris@10:      return x;
Chris@10: }
Chris@10: 
Chris@10: void tensor_destroy(bench_tensor *sz)
Chris@10: {
Chris@10:      bench_free0(sz->dims);
Chris@10:      bench_free(sz);
Chris@10: }
Chris@10: 
Chris@10: int tensor_sz(const bench_tensor *sz)
Chris@10: {
Chris@10:      int i, n = 1;
Chris@10: 
Chris@10:      if (!FINITE_RNK(sz->rnk))
Chris@10:           return 0;
Chris@10: 
Chris@10:      for (i = 0; i < sz->rnk; ++i)
Chris@10:           n *= sz->dims[i].n;
Chris@10:      return n;
Chris@10: }
Chris@10: 
Chris@10: 
Chris@10: /* total order among bench_iodim's */
Chris@10: static int dimcmp(const bench_iodim *a, const bench_iodim *b)
Chris@10: {
Chris@10:      if (b->is != a->is)
Chris@10:           return (b->is - a->is);	/* shorter strides go later */
Chris@10:      if (b->os != a->os)
Chris@10:           return (b->os - a->os);	/* shorter strides go later */
Chris@10:      return (int)(a->n - b->n);	        /* larger n's go later */
Chris@10: }
Chris@10: 
Chris@10: bench_tensor *tensor_compress(const bench_tensor *sz)
Chris@10: {
Chris@10:      int i, rnk;
Chris@10:      bench_tensor *x;
Chris@10: 
Chris@10:      BENCH_ASSERT(FINITE_RNK(sz->rnk));
Chris@10:      for (i = rnk = 0; i < sz->rnk; ++i) {
Chris@10:           BENCH_ASSERT(sz->dims[i].n > 0);
Chris@10:           if (sz->dims[i].n != 1)
Chris@10:                ++rnk;
Chris@10:      }
Chris@10: 
Chris@10:      x = mktensor(rnk);
Chris@10:      for (i = rnk = 0; i < sz->rnk; ++i) {
Chris@10:           if (sz->dims[i].n != 1)
Chris@10:                x->dims[rnk++] = sz->dims[i];
Chris@10:      }
Chris@10: 
Chris@10:      if (rnk) {
Chris@10: 	  /* God knows how qsort() behaves if n==0 */
Chris@10: 	  qsort(x->dims, (size_t)x->rnk, sizeof(bench_iodim),
Chris@10: 		(int (*)(const void *, const void *))dimcmp);
Chris@10:      }
Chris@10: 
Chris@10:      return x;
Chris@10: }
Chris@10: 
Chris@10: int tensor_unitstridep(bench_tensor *t)
Chris@10: {
Chris@10:      BENCH_ASSERT(FINITE_RNK(t->rnk));
Chris@10:      return (t->rnk == 0 ||
Chris@10: 	     (t->dims[t->rnk - 1].is == 1 && t->dims[t->rnk - 1].os == 1));
Chris@10: }
Chris@10: 
Chris@10: /* detect screwy real padded rowmajor... ugh */
Chris@10: int tensor_real_rowmajorp(bench_tensor *t, int sign, int in_place)
Chris@10: {
Chris@10:      int i;
Chris@10: 
Chris@10:      BENCH_ASSERT(FINITE_RNK(t->rnk));
Chris@10: 
Chris@10:      i = t->rnk - 1;
Chris@10: 
Chris@10:      if (--i >= 0) {
Chris@10:           bench_iodim *d = t->dims + i;
Chris@10: 	  if (sign < 0) {
Chris@10: 	       if (d[0].is != d[1].is * (in_place ? 2*(d[1].n/2 + 1) : d[1].n))
Chris@10: 		    return 0;
Chris@10: 	       if (d[0].os != d[1].os * (d[1].n/2 + 1))
Chris@10: 		    return 0;
Chris@10: 	  }
Chris@10: 	  else {
Chris@10: 	       if (d[0].is != d[1].is * (d[1].n/2 + 1))
Chris@10: 		    return 0;
Chris@10: 	       if (d[0].os != d[1].os * (in_place ? 2*(d[1].n/2 + 1) : d[1].n))
Chris@10: 		    return 0;
Chris@10: 	  }
Chris@10:      }
Chris@10: 
Chris@10:      while (--i >= 0) {
Chris@10:           bench_iodim *d = t->dims + i;
Chris@10:           if (d[0].is != d[1].is * d[1].n)
Chris@10:                return 0;
Chris@10:           if (d[0].os != d[1].os * d[1].n)
Chris@10:                return 0;
Chris@10:      }
Chris@10:      return 1;
Chris@10: }
Chris@10: 
Chris@10: int tensor_rowmajorp(bench_tensor *t)
Chris@10: {
Chris@10:      int i;
Chris@10: 
Chris@10:      BENCH_ASSERT(FINITE_RNK(t->rnk));
Chris@10: 
Chris@10:      i = t->rnk - 1;
Chris@10:      while (--i >= 0) {
Chris@10: 	  bench_iodim *d = t->dims + i;
Chris@10: 	  if (d[0].is != d[1].is * d[1].n)
Chris@10: 	       return 0;
Chris@10: 	  if (d[0].os != d[1].os * d[1].n)
Chris@10: 	       return 0;
Chris@10:      }
Chris@10:      return 1;
Chris@10: }
Chris@10: 
Chris@10: static void dimcpy(bench_iodim *dst, const bench_iodim *src, int rnk)
Chris@10: {
Chris@10:      int i;
Chris@10:      if (FINITE_RNK(rnk))
Chris@10:           for (i = 0; i < rnk; ++i)
Chris@10:                dst[i] = src[i];
Chris@10: }
Chris@10: 
Chris@10: bench_tensor *tensor_append(const bench_tensor *a, const bench_tensor *b)
Chris@10: {
Chris@10:      if (!FINITE_RNK(a->rnk) || !FINITE_RNK(b->rnk)) {
Chris@10:           return mktensor(RNK_MINFTY);
Chris@10:      } else {
Chris@10: 	  bench_tensor *x = mktensor(a->rnk + b->rnk);
Chris@10:           dimcpy(x->dims, a->dims, a->rnk);
Chris@10:           dimcpy(x->dims + a->rnk, b->dims, b->rnk);
Chris@10: 	  return x;
Chris@10:      }
Chris@10: }
Chris@10: 
Chris@10: static int imax(int a, int b)
Chris@10: {
Chris@10:      return (a > b) ? a : b;
Chris@10: }
Chris@10: 
Chris@10: static int imin(int a, int b)
Chris@10: {
Chris@10:      return (a < b) ? a : b;
Chris@10: }
Chris@10: 
Chris@10: #define DEFBOUNDS(name, xs)			\
Chris@10: void name(bench_tensor *t, int *lbp, int *ubp)	\
Chris@10: {						\
Chris@10:      int lb = 0;				\
Chris@10:      int ub = 1;				\
Chris@10:      int i;					\
Chris@10: 						\
Chris@10:      BENCH_ASSERT(FINITE_RNK(t->rnk));		\
Chris@10: 						\
Chris@10:      for (i = 0; i < t->rnk; ++i) {		\
Chris@10: 	  bench_iodim *d = t->dims + i;		\
Chris@10: 	  int n = d->n;				\
Chris@10: 	  int s = d->xs;			\
Chris@10: 	  lb = imin(lb, lb + s * (n - 1));	\
Chris@10: 	  ub = imax(ub, ub + s * (n - 1));	\
Chris@10:      }						\
Chris@10: 						\
Chris@10:      *lbp = lb;					\
Chris@10:      *ubp = ub;					\
Chris@10: }
Chris@10: 
Chris@10: DEFBOUNDS(tensor_ibounds, is)
Chris@10: DEFBOUNDS(tensor_obounds, os)
Chris@10: 
Chris@10: bench_tensor *tensor_copy(const bench_tensor *sz)
Chris@10: {
Chris@10:      bench_tensor *x = mktensor(sz->rnk);
Chris@10:      dimcpy(x->dims, sz->dims, sz->rnk);
Chris@10:      return x;
Chris@10: }
Chris@10: 
Chris@10: /* Like tensor_copy, but copy only rnk dimensions starting with start_dim. */
Chris@10: bench_tensor *tensor_copy_sub(const bench_tensor *sz, int start_dim, int rnk)
Chris@10: {
Chris@10:      bench_tensor *x;
Chris@10: 
Chris@10:      BENCH_ASSERT(FINITE_RNK(sz->rnk) && start_dim + rnk <= sz->rnk);
Chris@10:      x = mktensor(rnk);
Chris@10:      dimcpy(x->dims, sz->dims + start_dim, rnk);
Chris@10:      return x;
Chris@10: }
Chris@10: 
Chris@10: bench_tensor *tensor_copy_swapio(const bench_tensor *sz)
Chris@10: {
Chris@10:      bench_tensor *x = tensor_copy(sz);
Chris@10:      int i;
Chris@10:      if (FINITE_RNK(x->rnk))
Chris@10: 	  for (i = 0; i < x->rnk; ++i) {
Chris@10: 	       int s;
Chris@10: 	       s = x->dims[i].is;
Chris@10: 	       x->dims[i].is = x->dims[i].os;
Chris@10: 	       x->dims[i].os = s;
Chris@10: 	  }
Chris@10:      return x;
Chris@10: }