mirror of
https://github.com/tildearrow/furnace.git
synced 2024-11-24 13:35:11 +00:00
54e93db207
not reliable yet
215 lines
6.3 KiB
C
215 lines
6.3 KiB
C
/*
|
|
* Copyright (c) 2003, 2007-14 Matteo Frigo
|
|
* Copyright (c) 2003, 2007-14 Massachusetts Institute of Technology
|
|
*
|
|
* This program is free software; you can redistribute it and/or modify
|
|
* it under the terms of the GNU General Public License as published by
|
|
* the Free Software Foundation; either version 2 of the License, or
|
|
* (at your option) any later version.
|
|
*
|
|
* This program is distributed in the hope that it will be useful,
|
|
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
|
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
|
|
* GNU General Public License for more details.
|
|
*
|
|
* You should have received a copy of the GNU General Public License
|
|
* along with this program; if not, write to the Free Software
|
|
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
|
*
|
|
*/
|
|
|
|
|
|
#include "kernel/ifftw.h"
|
|
|
|
static int signof(INT x)
|
|
{
|
|
if (x < 0) return -1;
|
|
if (x == 0) return 0;
|
|
/* if (x > 0) */ return 1;
|
|
}
|
|
|
|
/* total order among iodim's */
|
|
int X(dimcmp)(const iodim *a, const iodim *b)
|
|
{
|
|
INT sai = X(iabs)(a->is), sbi = X(iabs)(b->is);
|
|
INT sao = X(iabs)(a->os), sbo = X(iabs)(b->os);
|
|
INT sam = X(imin)(sai, sao), sbm = X(imin)(sbi, sbo);
|
|
|
|
/* in descending order of min{istride, ostride} */
|
|
if (sam != sbm)
|
|
return signof(sbm - sam);
|
|
|
|
/* in case of a tie, in descending order of istride */
|
|
if (sbi != sai)
|
|
return signof(sbi - sai);
|
|
|
|
/* in case of a tie, in descending order of ostride */
|
|
if (sbo != sao)
|
|
return signof(sbo - sao);
|
|
|
|
/* in case of a tie, in ascending order of n */
|
|
return signof(a->n - b->n);
|
|
}
|
|
|
|
static void canonicalize(tensor *x)
|
|
{
|
|
if (x->rnk > 1) {
|
|
qsort(x->dims, (unsigned)x->rnk, sizeof(iodim),
|
|
(int (*)(const void *, const void *))X(dimcmp));
|
|
}
|
|
}
|
|
|
|
static int compare_by_istride(const iodim *a, const iodim *b)
|
|
{
|
|
INT sai = X(iabs)(a->is), sbi = X(iabs)(b->is);
|
|
|
|
/* in descending order of istride */
|
|
return signof(sbi - sai);
|
|
}
|
|
|
|
static tensor *really_compress(const tensor *sz)
|
|
{
|
|
int i, rnk;
|
|
tensor *x;
|
|
|
|
A(FINITE_RNK(sz->rnk));
|
|
for (i = rnk = 0; i < sz->rnk; ++i) {
|
|
A(sz->dims[i].n > 0);
|
|
if (sz->dims[i].n != 1)
|
|
++rnk;
|
|
}
|
|
|
|
x = X(mktensor)(rnk);
|
|
for (i = rnk = 0; i < sz->rnk; ++i) {
|
|
if (sz->dims[i].n != 1)
|
|
x->dims[rnk++] = sz->dims[i];
|
|
}
|
|
return x;
|
|
}
|
|
|
|
/* Like tensor_copy, but eliminate n == 1 dimensions, which
|
|
never affect any transform or transform vector.
|
|
|
|
Also, we sort the tensor into a canonical order of decreasing
|
|
strides (see X(dimcmp) for an exact definition). In general,
|
|
processing a loop/array in order of decreasing stride will improve
|
|
locality. Both forward and backwards traversal of the tensor are
|
|
considered e.g. by vrank-geq1, so sorting in increasing
|
|
vs. decreasing order is not really important. */
|
|
tensor *X(tensor_compress)(const tensor *sz)
|
|
{
|
|
tensor *x = really_compress(sz);
|
|
canonicalize(x);
|
|
return x;
|
|
}
|
|
|
|
/* Return whether the strides of a and b are such that they form an
|
|
effective contiguous 1d array. Assumes that a.is >= b.is. */
|
|
static int strides_contig(iodim *a, iodim *b)
|
|
{
|
|
return (a->is == b->is * b->n && a->os == b->os * b->n);
|
|
}
|
|
|
|
/* Like tensor_compress, but also compress into one dimension any
|
|
group of dimensions that form a contiguous block of indices with
|
|
some stride. (This can safely be done for transform vector sizes.) */
|
|
tensor *X(tensor_compress_contiguous)(const tensor *sz)
|
|
{
|
|
int i, rnk;
|
|
tensor *sz2, *x;
|
|
|
|
if (X(tensor_sz)(sz) == 0)
|
|
return X(mktensor)(RNK_MINFTY);
|
|
|
|
sz2 = really_compress(sz);
|
|
A(FINITE_RNK(sz2->rnk));
|
|
|
|
if (sz2->rnk <= 1) { /* nothing to compress. */
|
|
if (0) {
|
|
/* this call is redundant, because "sz->rnk <= 1" implies
|
|
that the tensor is already canonical, but I am writing
|
|
it explicitly because "logically" we need to canonicalize
|
|
the tensor before returning. */
|
|
canonicalize(sz2);
|
|
}
|
|
return sz2;
|
|
}
|
|
|
|
/* sort in descending order of |istride|, so that compressible
|
|
dimensions appear contigously */
|
|
qsort(sz2->dims, (unsigned)sz2->rnk, sizeof(iodim),
|
|
(int (*)(const void *, const void *))compare_by_istride);
|
|
|
|
/* compute what the rank will be after compression */
|
|
for (i = rnk = 1; i < sz2->rnk; ++i)
|
|
if (!strides_contig(sz2->dims + i - 1, sz2->dims + i))
|
|
++rnk;
|
|
|
|
/* merge adjacent dimensions whenever possible */
|
|
x = X(mktensor)(rnk);
|
|
x->dims[0] = sz2->dims[0];
|
|
for (i = rnk = 1; i < sz2->rnk; ++i) {
|
|
if (strides_contig(sz2->dims + i - 1, sz2->dims + i)) {
|
|
x->dims[rnk - 1].n *= sz2->dims[i].n;
|
|
x->dims[rnk - 1].is = sz2->dims[i].is;
|
|
x->dims[rnk - 1].os = sz2->dims[i].os;
|
|
} else {
|
|
A(rnk < x->rnk);
|
|
x->dims[rnk++] = sz2->dims[i];
|
|
}
|
|
}
|
|
|
|
X(tensor_destroy)(sz2);
|
|
|
|
/* reduce to canonical form */
|
|
canonicalize(x);
|
|
return x;
|
|
}
|
|
|
|
/* The inverse of X(tensor_append): splits the sz tensor into
|
|
tensor a followed by tensor b, where a's rank is arnk. */
|
|
void X(tensor_split)(const tensor *sz, tensor **a, int arnk, tensor **b)
|
|
{
|
|
A(FINITE_RNK(sz->rnk) && FINITE_RNK(arnk));
|
|
|
|
*a = X(tensor_copy_sub)(sz, 0, arnk);
|
|
*b = X(tensor_copy_sub)(sz, arnk, sz->rnk - arnk);
|
|
}
|
|
|
|
/* TRUE if the two tensors are equal */
|
|
int X(tensor_equal)(const tensor *a, const tensor *b)
|
|
{
|
|
if (a->rnk != b->rnk)
|
|
return 0;
|
|
|
|
if (FINITE_RNK(a->rnk)) {
|
|
int i;
|
|
for (i = 0; i < a->rnk; ++i)
|
|
if (0
|
|
|| a->dims[i].n != b->dims[i].n
|
|
|| a->dims[i].is != b->dims[i].is
|
|
|| a->dims[i].os != b->dims[i].os
|
|
)
|
|
return 0;
|
|
}
|
|
|
|
return 1;
|
|
}
|
|
|
|
/* TRUE if the sets of input and output locations described by
|
|
(append sz vecsz) are the same */
|
|
int X(tensor_inplace_locations)(const tensor *sz, const tensor *vecsz)
|
|
{
|
|
tensor *t = X(tensor_append)(sz, vecsz);
|
|
tensor *ti = X(tensor_copy_inplace)(t, INPLACE_IS);
|
|
tensor *to = X(tensor_copy_inplace)(t, INPLACE_OS);
|
|
tensor *tic = X(tensor_compress_contiguous)(ti);
|
|
tensor *toc = X(tensor_compress_contiguous)(to);
|
|
|
|
int retval = X(tensor_equal)(tic, toc);
|
|
|
|
X(tensor_destroy)(t);
|
|
X(tensor_destroy4)(ti, to, tic, toc);
|
|
|
|
return retval;
|
|
}
|