Mercurial > octave
view scripts/sparse/svds.m @ 29949:f254c302bb9c
remove JIT compiler from Octave sources
As stated in the NEWS file entry added with this changeset, no one
has ever seriously taken on further development of the JIT compiler in
Octave since it was first added as part of a Google Summer of Code
project in 2012 and it still does nothing significant. It is out of
date with the default interpreter that walks the parse tree. Even
though we have fixed the configure script to disable it by default,
people still ask questions about how to build it, but it doesn’t seem
that they are doing that to work on it but because they think it will
make Octave code run faster (it never did, except for some extremely
simple bits of code as examples for demonstration purposes only).
* NEWS: Note change.
* configure.ac, acinclude.m4: Eliminate checks and macros related to
the JIT compiler and LLVM.
* basics.txi, install.txi, octave.texi, vectorize.txi: Remove mention
of JIT compiler and LLVM.
* jit-ir.cc, jit-ir.h, jit-typeinfo.cc, jit-typeinfo.h, jit-util.cc,
jit-util.h, pt-jit.cc, pt-jit.h: Delete.
* libinterp/parse-tree/module.mk: Update.
* Array-jit.cc: Delete.
* libinterp/template-inst/module.mk: Update.
* test/jit.tst: Delete.
* test/module.mk: Update.
* interpreter.cc (interpreter::interpreter): Don't check options for
debug_jit or jit_compiler.
* toplev.cc (F__octave_config_info__): Remove JIT compiler and LLVM
info from struct.
* ov-base.h (octave_base_value::grab, octave_base_value::release):
Delete.
* ov-builtin.h, ov-builtin.cc (octave_builtin::to_jit,
octave_builtin::stash_jit): Delete.
(octave_builtin::m_jtype): Delete data member and all uses.
* ov-usr-fcn.h, ov-usr-fcn.cc (octave_user_function::m_jit_info):
Delete data member and all uses.
(octave_user_function::get_info, octave_user_function::stash_info): Delete.
* options.h (DEBUG_JIT_OPTION, JIT_COMPILER_OPTION): Delete macro
definitions and all uses.
* octave.h, octave.cc (cmdline_options::cmdline_options): Don't handle
DEBUG_JIT_OPTION, JIT_COMPILER_OPTION): Delete.
(cmdline_options::debug_jit, cmdline_options::jit_compiler): Delete
functions and all uses.
(cmdline_options::m_debug_jit, cmdline_options::m_jit_compiler): Delete
data members and all uses.
(octave_getopt_options long_opts): Remove "debug-jit" and
"jit-compiler" from the list.
* pt-eval.cc (tree_evaluator::visit_simple_for_command,
tree_evaluator::visit_complex_for_command,
tree_evaluator::visit_while_command,
tree_evaluator::execute_user_function): Eliminate JIT compiler code.
* pt-loop.h, pt-loop.cc (tree_while_command::get_info,
tree_while_command::stash_info, tree_simple_for_command::get_info,
tree_simple_for_command::stash_info): Delete functions and all uses.
(tree_while_command::m_compiled, tree_simple_for_command::m_compiled):
Delete member variable and all uses.
* usage.h (usage_string, octave_print_verbose_usage_and_exit): Remove
[--debug-jit] and [--jit-compiler] from the message.
* Array.h (Array<T>::Array): Remove constructor that was only intended
to be used by the JIT compiler.
(Array<T>::jit_ref_count, Array<T>::jit_slice_data,
Array<T>::jit_dimensions, Array<T>::jit_array_rep): Delete.
* Marray.h (MArray<T>::MArray): Remove constructor that was only
intended to be used by the JIT compiler.
* NDArray.h (NDArray::NDarray): Remove constructor that was only
intended to be used by the JIT compiler.
* dim-vector.h (dim_vector::to_jit): Delete.
(dim_vector::dim_vector): Remove constructor that was only intended to
be used by the JIT compiler.
* codeql-analysis.yaml, make.yaml: Don't require llvm-dev.
* subst-config-vals.in.sh, subst-cross-config-vals.in.sh: Don't
substitute OCTAVE_CONF_LLVM_CPPFLAGS, OCTAVE_CONF_LLVM_LDFLAGS, or
OCTAVE_CONF_LLVM_LIBS.
* Doxyfile.in: Don't define HAVE_LLVM.
* aspell-octave.en.pws: Eliminate jit, JIT, and LLVM from the list of
spelling exceptions.
* build-env.h, build-env.in.cc (LLVM_CPPFLAGS, LLVM_LDFLAGS,
LLVM_LIBS): Delete variables and all uses.
* libinterp/corefcn/module.mk (%canon_reldir%_libcorefcn_la_CPPFLAGS):
Remove $(LLVM_CPPFLAGS) from the list.
* libinterp/parse-tree/module.mk (%canon_reldir%_libparse_tree_la_CPPFLAGS):
Remove $(LLVM_CPPFLAGS) from the list.
author | John W. Eaton <jwe@octave.org> |
---|---|
date | Tue, 10 Aug 2021 16:42:29 -0400 |
parents | 7854d5752dd2 |
children | 796f54d4ddbf |
line wrap: on
line source
######################################################################## ## ## Copyright (C) 2006-2021 The Octave Project Developers ## ## See the file COPYRIGHT.md in the top-level directory of this ## distribution or <https://octave.org/copyright/>. ## ## This file is part of Octave. ## ## Octave 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 3 of the License, or ## (at your option) any later version. ## ## Octave 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 Octave; see the file COPYING. If not, see ## <https://www.gnu.org/licenses/>. ## ######################################################################## ## -*- texinfo -*- ## @deftypefn {} {@var{s} =} svds (@var{A}) ## @deftypefnx {} {@var{s} =} svds (@var{A}, @var{k}) ## @deftypefnx {} {@var{s} =} svds (@var{A}, @var{k}, @var{sigma}) ## @deftypefnx {} {@var{s} =} svds (@var{A}, @var{k}, @var{sigma}, @var{opts}) ## @deftypefnx {} {[@var{u}, @var{s}, @var{v}] =} svds (@dots{}) ## @deftypefnx {} {[@var{u}, @var{s}, @var{v}, @var{flag}] =} svds (@dots{}) ## ## Find a few singular values of the matrix @var{A}. ## ## The singular values are calculated using ## ## @example ## @group ## [@var{m}, @var{n}] = size (@var{A}); ## @var{s} = eigs ([sparse(@var{m}, @var{m}), @var{A}; ## @var{A}', sparse(@var{n}, @var{n})]) ## @end group ## @end example ## ## The eigenvalues returned by @code{eigs} correspond to the singular values ## of @var{A}. The number of singular values to calculate is given by @var{k} ## and defaults to 6. ## ## The argument @var{sigma} specifies which singular values to find. When ## @var{sigma} is the string @qcode{'L'}, the default, the largest singular ## values of @var{A} are found. Otherwise, @var{sigma} must be a real scalar ## and the singular values closest to @var{sigma} are found. As a corollary, ## @code{@var{sigma} = 0} finds the smallest singular values. Note that for ## relatively small values of @var{sigma}, there is a chance that the ## requested number of singular values will not be found. In that case ## @var{sigma} should be increased. ## ## @var{opts} is a structure defining options that @code{svds} will pass ## to @code{eigs}. The possible fields of this structure are documented in ## @code{eigs}. By default, @code{svds} sets the following three fields: ## ## @table @code ## @item tol ## The required convergence tolerance for the singular values. The default ## value is 1e-10. @code{eigs} is passed @code{@var{tol} / sqrt (2)}. ## ## @item maxit ## The maximum number of iterations. The default is 300. ## ## @item disp ## The level of diagnostic printout (0|1|2). If @code{disp} is 0 then ## diagnostics are disabled. The default value is 0. ## @end table ## ## If more than one output is requested then @code{svds} will return an ## approximation of the singular value decomposition of @var{A} ## ## @example ## @var{A}_approx = @var{u}*@var{s}*@var{v}' ## @end example ## ## @noindent ## where @var{A}_approx is a matrix of size @var{A} but only rank @var{k}. ## ## @var{flag} returns 0 if the algorithm has successfully converged, and 1 ## otherwise. The test for convergence is ## ## @example ## @group ## norm (@var{A}*@var{v} - @var{u}*@var{s}, 1) <= @var{tol} * norm (@var{A}, 1) ## @end group ## @end example ## ## @code{svds} is best for finding only a few singular values from a large ## sparse matrix. Otherwise, @code{svd (full (@var{A}))} will likely be more ## efficient. ## @seealso{svd, eigs} ## @end deftypefn function [u, s, v, flag] = svds (A, k, sigma, opts) persistent root2 = sqrt (2); if (nargin < 1) print_usage (); endif if (ndims (A) > 2) error ("svds: A must be a 2-D matrix"); endif if (nargin < 4) opts.tol = 0; # use ARPACK default opts.disp = 0; opts.maxit = 300; else if (! isstruct (opts)) error ("svds: OPTS must be a structure"); endif if (! isfield (opts, "tol")) opts.tol = 0; # use ARPACK default else opts.tol = opts.tol / root2; endif if (isfield (opts, "v0")) if (! isvector (opts.v0) || (length (opts.v0) != sum (size (A)))) error ("svds: OPTS.v0 must be a vector with rows (A) + columns (A) entries"); endif endif endif if (nargin < 3 || strcmp (sigma, "L")) if (isreal (A)) sigma = "LA"; else sigma = "LR"; endif elseif (isscalar (sigma) && isnumeric (sigma) && isreal (sigma)) if (sigma < 0) error ("svds: SIGMA must be a positive real value"); endif else error ("svds: SIGMA must be a positive real value or the string 'L'"); endif [m, n] = size (A); max_a = max (abs (nonzeros (A))); if (isempty (max_a)) max_a = 0; endif ## Must initialize variable value, otherwise it may appear to interpreter ## that code is trying to call flag() colormap function. flag = 0; if (max_a == 0) s = zeros (k, 1); # special case of zero matrix else if (nargin < 2) k = min ([6, m, n]); else k = min ([k, m, n]); endif ## Scale everything by the 1-norm to make things more stable. b = A / max_a; b_opts = opts; ## Call to eigs is always a symmetric matrix by construction b_opts.issym = true; b_sigma = sigma; if (! ischar (b_sigma)) b_sigma /= max_a; endif if (b_sigma == 0) ## Find the smallest eigenvalues ## The eigenvalues returns by eigs for sigma=0 are symmetric about 0. ## As we are only interested in the positive eigenvalues, we have to ## double k and then throw out the k negative eigenvalues. ## Separately, if sigma is nonzero, but smaller than the smallest ## singular value, ARPACK may not return k eigenvalues. However, as ## computation scales with k we'd like to avoid doubling k for all ## scalar values of sigma. b_k = 2 * k; else b_k = k; # Normal case, find just the k largest eigenvalues endif if (nargout > 1) [V, s, flag] = eigs ([sparse(m,m), b; b', sparse(n,n)], b_k, b_sigma, b_opts); s = diag (s); else s = eigs ([sparse(m,m), b; b', sparse(n,n)], b_k, b_sigma, b_opts); endif if (ischar (sigma)) norma = max (s); else norma = normest (A); endif ## We wish to exclude all eigenvalues that are less than zero as these ## are artifacts of the way the matrix passed to eigs is formed. There ## is also the possibility that the value of sigma chosen is exactly ## a singular value, and in that case we're dead!! So have to rely on ## the warning from eigs. We exclude the singular values which are ## less than or equal to zero to within some tolerance scaled by the ## norm since if we don't we might end up with too many singular ## values. if (b_sigma == 0) if (sum (s>0) < k) ## It may happen that the number of positive s is less than k. ## In this case, take -s (if s in an eigenvalue, so is -s), ## flipped upside-down. s = flipud (-s); endif endif tol = norma * opts.tol; ind = find (s > tol); if (length (ind) < k) ## Too few eigenvalues returned. Add in any zero eigenvalues of B, ## including the nominally negative ones. zind = find (abs (s) <= tol); p = min (length (zind), k - length (ind)); ind = [ind; zind(1:p)]; elseif (length (ind) > k) ## Too many eigenvalues returned. Select according to criterion. if (b_sigma == 0) ind = ind(end+1-k:end); # smallest eigenvalues else ind = ind(1:k); # largest eigenvalues endif endif s = s(ind); if (length (s) < k) warning ("svds: returning fewer singular values than requested"); if (! ischar (sigma)) warning ("svds: try increasing the value of sigma"); endif endif s *= max_a; endif if (nargout < 2) u = s; else if (max_a == 0) u = eye (m, k); s = diag (s); v = eye (n, k); else u = root2 * V(1:m,ind); s = diag (s); v = root2 * V(m+1:end,ind); endif if (nargout > 3) flag = (flag != 0); endif endif endfunction %!shared n, k, A, u, s, v, opts, rand_state, randn_state, tol %! n = 100; %! k = 7; %! A = sparse ([3:n,1:n,1:(n-2)],[1:(n-2),1:n,3:n],[ones(1,n-2),0.4*n*ones(1,n),ones(1,n-2)]); %! [u,s,v] = svd (full (A)); %! s = diag (s); %! [~, idx] = sort (abs (s)); %! s = s(idx); %! u = u(:, idx); %! v = v(:, idx); %! rand_state = rand ("state"); %! rand ("state", 42); %! opts.v0 = rand (2*n,1); # Initialize eigs ARPACK starting vector %! # to guarantee reproducible results %!testif HAVE_ARPACK %! [u2,s2,v2,flag] = svds (A,k); %! s2 = diag (s2); %! assert (flag, ! 1); %! tol = 15 * eps * norm (s2, 1); %! assert (s2, s(end:-1:end-k+1), tol); %!testif HAVE_ARPACK, HAVE_UMFPACK %! [u2,s2,v2,flag] = svds (A,k,0,opts); %! s2 = diag (s2); %! assert (flag, ! 1); %! tol = 15 * eps * norm (s2, 1); %! assert (s2, s(k:-1:1), tol); %!testif HAVE_ARPACK, HAVE_UMFPACK %! idx = floor (n/2); %! % Don't put sigma right on a singular value or there are convergence issues %! sigma = 0.99*s(idx) + 0.01*s(idx+1); %! [u2,s2,v2,flag] = svds (A,k,sigma,opts); %! s2 = diag (s2); %! assert (flag, ! 1); %! tol = 15 * eps * norm (s2, 1); %! assert (s2, s((idx+floor (k/2)):-1:(idx-floor (k/2))), tol); %!testif HAVE_ARPACK %! [u2,s2,v2,flag] = svds (zeros (10), k); %! assert (u2, eye (10, k)); %! assert (s2, zeros (k)); %! assert (v2, eye (10, 7)); %! %!testif HAVE_ARPACK %! s = svds (speye (10)); %! assert (s, ones (6, 1), 8*eps); %!testif HAVE_ARPACK <57185> %! z = complex (ones (10), ones (10)); %! s = svds (z); %! assert (isreal (s)); %!test %! ## Restore random number generator seed at end of tests %! rand ("state", rand_state);