nmatrix-atlas 0.2.0

data/ext/nmatrix/util/util.h

@@ -0,0 +1,78 @@
+/////////////////////////////////////////////////////////////////////
+// = NMatrix
+//
+// A linear algebra library for scientific computation in Ruby.
+// NMatrix is part of SciRuby.
+//
+// NMatrix was originally inspired by and derived from NArray, by
+// Masahiro Tanaka: http://narray.rubyforge.org
+//
+// == Copyright Information
+//
+// SciRuby is Copyright (c) 2010 - 2014, Ruby Science Foundation
+// NMatrix is Copyright (c) 2012 - 2014, John Woods and the Ruby Science Foundation
+//
+// Please see LICENSE.txt for additional copyright notices.
+//
+// == Contributing
+//
+// By contributing source code to SciRuby, you agree to be bound by
+// our Contributor Agreement:
+//
+// * https://github.com/SciRuby/sciruby/wiki/Contributor-Agreement
+//
+// == util.h
+//
+// Header file for utility functions and data.
+
+#ifndef UTIL_H
+#define UTIL_H
+
+/*
+ * Standard Includes
+ */
+
+/*
+ * Project Includes
+ */
+
+#include "types.h"
+
+/*
+ * Macros
+ */
+
+/*
+ * Types
+ */
+
+/*
+ * Data
+ */
+
+/*
+ * Functions
+ */
+namespace nm {
+  template <typename Type>
+  inline Type gcf(Type x, Type y) {
+    Type t;
+
+    if (x < 0) x = -x;
+    if (y < 0) y = -y;
+
+    if (x == 0) return y;
+    if (y == 0) return x;
+
+    while (x > 0) {
+      t = x;
+      x = y % x;
+      y = t;
+    }
+
+    return y;
+  }
+} // end of namespace nm
+
+
+#endif // UTIL_H

data/ext/nmatrix_atlas/extconf.rb

@@ -0,0 +1,250 @@
+# = NMatrix
+#
+# A linear algebra library for scientific computation in Ruby.
+# NMatrix is part of SciRuby.
+#
+# NMatrix was originally inspired by and derived from NArray, by
+# Masahiro Tanaka: http://narray.rubyforge.org
+#
+# == Copyright Information
+#
+# SciRuby is Copyright (c) 2010 - 2014, Ruby Science Foundation
+# NMatrix is Copyright (c) 2012 - 2014, John Woods and the Ruby Science Foundation
+#
+# Please see LICENSE.txt for additional copyright notices.
+#
+# == Contributing
+#
+# By contributing source code to SciRuby, you agree to be bound by
+# our Contributor Agreement:
+#
+# * https://github.com/SciRuby/sciruby/wiki/Contributor-Agreement
+#
+# == extconf.rb
+#
+# This file checks for ATLAS and other necessary headers, and
+# generates a Makefile for compiling NMatrix.
+
+require "mkmf"
+
+
+# Function derived from NArray's extconf.rb.
+def have_type(type, header=nil) #:nodoc:
+  printf "checking for %s... ", type
+  STDOUT.flush
+
+  src = <<"SRC"
+#include <ruby.h>
+SRC
+
+
+  src << <<"SRC" unless header.nil?
+#include <#{header}>
+SRC
+
+  r = try_link(src + <<"SRC")
+  int main() { return 0; }
+  int t() { #{type} a; return 0; }
+SRC
+
+  unless r
+    print "no\n"
+    return false
+  end
+
+  $defs.push(format("-DHAVE_%s", type.upcase))
+
+  print "yes\n"
+
+  return true
+end
+
+# Function derived from NArray's extconf.rb.
+def create_conf_h(file) #:nodoc:
+  print "creating #{file}\n"
+  File.open(file, 'w') do |hfile|
+    header_guard = file.upcase.sub(/\s|\./, '_')
+
+    hfile.puts "#ifndef #{header_guard}"
+    hfile.puts "#define #{header_guard}"
+    hfile.puts
+
+    # FIXME: Find a better way to do this:
+    hfile.puts "#define RUBY_2 1" if RUBY_VERSION >= '2.0'
+
+    for line in $defs
+      line =~ /^-D(.*)/
+      hfile.printf "#define %s 1\n", $1
+    end
+
+    hfile.puts
+    hfile.puts "#endif"
+  end
+end
+
+if RUBY_VERSION < '1.9'
+  raise(NotImplementedError, "Sorry, you need at least Ruby 1.9!")
+else
+  #$INSTALLFILES = [['nmatrix.h', '$(archdir)'], ['nmatrix.hpp', '$(archdir)'], ['nmatrix_config.h', '$(archdir)'], ['nm_memory.h', '$(archdir)']]
+  if /cygwin|mingw/ =~ RUBY_PLATFORM
+    #$INSTALLFILES << ['libnmatrix.a', '$(archdir)']
+  end
+end
+
+if /cygwin|mingw/ =~ RUBY_PLATFORM
+  CONFIG["DLDFLAGS"] << " --output-lib libnmatrix.a"
+end
+
+$DEBUG = true
+#not the right way to add this include directory
+$CFLAGS = ["-Wall -Werror=return-type -I$(srcdir)/../nmatrix",$CFLAGS].join(" ")
+$CXXFLAGS = ["-Wall -Werror=return-type -I$(srcdir)/../nmatrix",$CXXFLAGS].join(" ")
+$CPPFLAGS = ["-Wall -Werror=return-type -I$(srcdir)/../nmatrix",$CPPFLAGS].join(" ")
+
+# When adding objects here, make sure their directories are included in CLEANOBJS down at the bottom of extconf.rb.
+# Why not just autogenerate this list from all .c/.cpp files in directory?
+basenames = %w{nmatrix_atlas math_atlas}
+$objs = basenames.map { |b| "#{b}.o"   }
+$srcs = basenames.map { |b| "#{b}.cpp" }
+
+#CONFIG['CXX'] = 'clang++'
+CONFIG['CXX'] = 'g++'
+
+def find_newer_gplusplus #:nodoc:
+  print "checking for apparent GNU g++ binary with C++0x/C++11 support... "
+  [9,8,7,6,5,4,3].each do |minor|
+    ver = "4.#{minor}"
+    gpp = "g++-#{ver}"
+    result = `which #{gpp}`
+    next if result.empty?
+    CONFIG['CXX'] = gpp
+    puts ver
+    return CONFIG['CXX']
+  end
+  false
+end
+
+def gplusplus_version
+  cxxvar = proc { |n| `#{CONFIG['CXX']} -E -dM - </dev/null | grep #{n}`.chomp.split(' ')[2] }
+  major = cxxvar.call('__GNUC__')
+  minor = cxxvar.call('__GNUC_MINOR__')
+  patch = cxxvar.call('__GNUC_PATCHLEVEL__')
+
+  raise("unable to determine g++ version (match to get version was nil)") if major.nil? || minor.nil? || patch.nil?
+
+  "#{major}.#{minor}.#{patch}"
+end
+
+
+if CONFIG['CXX'] == 'clang++'
+  $CPP_STANDARD = 'c++11'
+
+else
+  version = gplusplus_version
+  if version < '4.3.0' && CONFIG['CXX'] == 'g++'  # see if we can find a newer G++, unless it's been overridden by user
+    if !find_newer_gplusplus
+      raise("You need a version of g++ which supports -std=c++0x or -std=c++11. If you're on a Mac and using Homebrew, we recommend using mac-brew-gcc.sh to install a more recent g++.")
+    end
+    version = gplusplus_version
+  end
+
+  if version < '4.7.0'
+    $CPP_STANDARD = 'c++0x'
+  else
+    $CPP_STANDARD = 'c++11'
+  end
+  puts "using C++ standard... #{$CPP_STANDARD}"
+  puts "g++ reports version... " + `#{CONFIG['CXX']} --version|head -n 1|cut -f 3 -d " "`
+end
+
+# add smmp in to get generic transp; remove smmp2 to eliminate funcptr transp
+
+# The next line allows the user to supply --with-atlas-dir=/usr/local/atlas,
+# --with-atlas-lib or --with-atlas-include and tell the compiler where to look
+# for ATLAS. The same for all the others
+#
+#dir_config("clapack", ["/usr/local/atlas/include"], [])
+#
+#
+
+# Is g++ having trouble finding your header files?
+# Try this:
+#   export C_INCLUDE_PATH=/usr/local/atlas/include
+#   export CPLUS_INCLUDE_PATH=/usr/local/atlas/include
+# (substituting in the path of your cblas.h and clapack.h for the path I used). -- JW 8/27/12
+
+idefaults = {lapack: ["/usr/include/atlas"],
+             cblas: ["/usr/local/atlas/include", "/usr/include/atlas"],
+             atlas: ["/usr/local/atlas/include", "/usr/include/atlas"]}
+
+# For some reason, if we try to look for /usr/lib64/atlas on a Mac OS X Mavericks system, and the directory does not
+# exist, it will give a linker error -- even if the lib dir is already correctly included with -L. So we need to check
+# that Dir.exists?(d) for each.
+ldefaults = {lapack: ["/usr/local/lib", "/usr/local/atlas/lib", "/usr/lib64/atlas"].delete_if { |d| !Dir.exists?(d) },
+             cblas: ["/usr/local/lib", "/usr/local/atlas/lib", "/usr/lib64/atlas"].delete_if { |d| !Dir.exists?(d) },
+             atlas: ["/usr/local/lib", "/usr/local/atlas/lib", "/usr/lib", "/usr/lib64/atlas"].delete_if { |d| !Dir.exists?(d) }}
+
+if have_library("clapack") # Usually only applies for Mac OS X
+  $libs += " -lclapack "
+end
+
+unless have_library("lapack")
+  dir_config("lapack", idefaults[:lapack], ldefaults[:lapack])
+end
+
+unless have_library("cblas")
+  dir_config("cblas", idefaults[:cblas], ldefaults[:cblas])
+end
+
+unless have_library("atlas")
+  dir_config("atlas", idefaults[:atlas], ldefaults[:atlas])
+end
+
+# If BLAS and LAPACK headers are in an atlas directory, prefer those. Otherwise,
+# we try our luck with the default location.
+if have_header("atlas/cblas.h")
+  have_header("atlas/clapack.h")
+else
+  have_header("cblas.h")
+  have_header("clapack.h")
+end
+
+
+# Although have_func is supposed to take a list as its second argument, I find that it simply
+# applies a :to_s to the second arg and doesn't actually check each one. We may want to put
+# have_func calls inside an :each block which checks atlas/clapack.h, cblas.h, clapack.h, and
+# lastly lapack.h. On Ubuntu, it only works if I use atlas/clapack.h. --@mohawkjohn 8/20/14
+have_func("clapack_dgetrf", "atlas/clapack.h")
+have_func("clapack_dgetri", "atlas/clapack.h")
+have_func("dgesvd_", "clapack.h") # This may not do anything. dgesvd_ seems to be in LAPACK, not CLAPACK.
+
+have_func("cblas_dgemm", "cblas.h")
+
+#have_func("rb_scan_args", "ruby.h")
+
+#find_library("lapack", "clapack_dgetrf")
+#find_library("cblas", "cblas_dgemm")
+#find_library("atlas", "ATL_dgemmNN")
+# Order matters here: ATLAS has to go after LAPACK: http://mail.scipy.org/pipermail/scipy-user/2007-January/010717.html
+$libs += " -llapack -lcblas -latlas "
+#$libs += " -lprofiler "
+
+
+# For release, these next two should both be changed to -O3.
+$CFLAGS += " -O3" #" -O0 -g "
+#$CFLAGS += " -static -O0 -g "
+$CPPFLAGS += " -O3 -std=#{$CPP_STANDARD}" #" -O0 -g -std=#{$CPP_STANDARD} " #-fmax-errors=10 -save-temps
+#$CPPFLAGS += " -static -O0 -g -std=#{$CPP_STANDARD} "
+
+CONFIG['warnflags'].gsub!('-Wshorten-64-to-32', '') # doesn't work except in Mac-patched gcc (4.2)
+CONFIG['warnflags'].gsub!('-Wdeclaration-after-statement', '')
+CONFIG['warnflags'].gsub!('-Wimplicit-function-declaration', '')
+
+create_conf_h("nmatrix_atlas_config.h")
+create_makefile("nmatrix_atlas")
+
+# to clean up object files in subdirectories:
+open('Makefile', 'a') do |f|
+  clean_objs_paths = %w{ }.map { |d| "#{d}/*.#{CONFIG["OBJEXT"]}" }
+  f.write("CLEANOBJS := $(CLEANOBJS) #{clean_objs_paths.join(' ')}")
+end

data/ext/nmatrix_atlas/math_atlas.cpp

@@ -0,0 +1,1206 @@
+/////////////////////////////////////////////////////////////////////
+// = NMatrix
+//
+// A linear algebra library for scientific computation in Ruby.
+// NMatrix is part of SciRuby.
+//
+// NMatrix was originally inspired by and derived from NArray, by
+// Masahiro Tanaka: http://narray.rubyforge.org
+//
+// == Copyright Information
+//
+// SciRuby is Copyright (c) 2010 - 2014, Ruby Science Foundation
+// NMatrix is Copyright (c) 2012 - 2014, John Woods and the Ruby Science Foundation
+//
+// Please see LICENSE.txt for additional copyright notices.
+//
+// == Contributing
+//
+// By contributing source code to SciRuby, you agree to be bound by
+// our Contributor Agreement:
+//
+// * https://github.com/SciRuby/sciruby/wiki/Contributor-Agreement
+//
+// == math_atlas.cpp
+//
+// Ruby-exposed CBLAS and LAPACK functions that call ATLAS
+// functions.
+//
+
+/*
+ * Project Includes
+ */
+
+#include "data/data.h"
+
+#include "math_atlas/inc.h"
+
+#include "math/util.h"
+
+//BLAS
+#include "math_atlas/cblas_templates_atlas.h"
+
+//LAPACK
+#include "math/laswp.h"
+#include "math_atlas/clapack_templates.h"
+
+#include "math_atlas/gesvd.h"
+#include "math_atlas/gesdd.h"
+#include "math_atlas/geev.h"
+
+
+/*
+ * Forward Declarations
+ */
+
+extern "C" {
+  /* BLAS Level 1. */
+  static VALUE nm_atlas_cblas_scal(VALUE self, VALUE n, VALUE scale, VALUE vector, VALUE incx);
+  static VALUE nm_atlas_cblas_nrm2(VALUE self, VALUE n, VALUE x, VALUE incx);
+  static VALUE nm_atlas_cblas_asum(VALUE self, VALUE n, VALUE x, VALUE incx);
+  static VALUE nm_atlas_cblas_rot(VALUE self, VALUE n, VALUE x, VALUE incx, VALUE y, VALUE incy, VALUE c, VALUE s);
+  static VALUE nm_atlas_cblas_rotg(VALUE self, VALUE ab);
+  static VALUE nm_atlas_cblas_imax(VALUE self, VALUE n, VALUE x, VALUE incx);
+
+  /* BLAS Level 2. */
+  static VALUE nm_atlas_cblas_gemv(VALUE self, VALUE trans_a, VALUE m, VALUE n, VALUE vAlpha, VALUE a, VALUE lda,
+                             VALUE x, VALUE incx, VALUE vBeta, VALUE y, VALUE incy);
+
+  /* BLAS Level 3. */
+  static VALUE nm_atlas_cblas_gemm(VALUE self, VALUE order, VALUE trans_a, VALUE trans_b, VALUE m, VALUE n, VALUE k, VALUE vAlpha,
+                             VALUE a, VALUE lda, VALUE b, VALUE ldb, VALUE vBeta, VALUE c, VALUE ldc);
+  static VALUE nm_atlas_cblas_trsm(VALUE self, VALUE order, VALUE side, VALUE uplo, VALUE trans_a, VALUE diag, VALUE m, VALUE n,
+                             VALUE vAlpha, VALUE a, VALUE lda, VALUE b, VALUE ldb);
+  static VALUE nm_atlas_cblas_trmm(VALUE self, VALUE order, VALUE side, VALUE uplo, VALUE trans_a, VALUE diag, VALUE m, VALUE n,
+                             VALUE alpha, VALUE a, VALUE lda, VALUE b, VALUE ldb);
+  static VALUE nm_atlas_cblas_herk(VALUE self, VALUE order, VALUE uplo, VALUE trans, VALUE n, VALUE k, VALUE alpha, VALUE a,
+                             VALUE lda, VALUE beta, VALUE c, VALUE ldc);
+  static VALUE nm_atlas_cblas_syrk(VALUE self, VALUE order, VALUE uplo, VALUE trans, VALUE n, VALUE k, VALUE alpha, VALUE a,
+                             VALUE lda, VALUE beta, VALUE c, VALUE ldc);
+
+  /* LAPACK. */
+  static VALUE nm_atlas_has_clapack(VALUE self);
+  static VALUE nm_atlas_clapack_getrf(VALUE self, VALUE order, VALUE m, VALUE n, VALUE a, VALUE lda);
+  static VALUE nm_atlas_clapack_potrf(VALUE self, VALUE order, VALUE uplo, VALUE n, VALUE a, VALUE lda);
+  static VALUE nm_atlas_clapack_getrs(VALUE self, VALUE order, VALUE trans, VALUE n, VALUE nrhs, VALUE a, VALUE lda, VALUE ipiv, VALUE b, VALUE ldb);
+  static VALUE nm_atlas_clapack_potrs(VALUE self, VALUE order, VALUE uplo, VALUE n, VALUE nrhs, VALUE a, VALUE lda, VALUE b, VALUE ldb);
+  static VALUE nm_atlas_clapack_getri(VALUE self, VALUE order, VALUE n, VALUE a, VALUE lda, VALUE ipiv);
+  static VALUE nm_atlas_clapack_potri(VALUE self, VALUE order, VALUE uplo, VALUE n, VALUE a, VALUE lda);
+  static VALUE nm_atlas_clapack_laswp(VALUE self, VALUE n, VALUE a, VALUE lda, VALUE k1, VALUE k2, VALUE ipiv, VALUE incx);
+
+  static VALUE nm_atlas_lapack_gesvd(VALUE self, VALUE jobu, VALUE jobvt, VALUE m, VALUE n, VALUE a, VALUE lda, VALUE s, VALUE u, VALUE ldu, VALUE vt, VALUE ldvt, VALUE lworkspace_size);
+  static VALUE nm_atlas_lapack_gesdd(VALUE self, VALUE jobz, VALUE m, VALUE n, VALUE a, VALUE lda, VALUE s, VALUE u, VALUE ldu, VALUE vt, VALUE ldvt, VALUE lworkspace_size);
+  static VALUE nm_atlas_lapack_geev(VALUE self, VALUE compute_left, VALUE compute_right, VALUE n, VALUE a, VALUE lda, VALUE w, VALUE wi, VALUE vl, VALUE ldvl, VALUE vr, VALUE ldvr, VALUE lwork);
+}
+
+////////////////////
+// Math Functions //
+////////////////////
+
+namespace nm { 
+  namespace math {
+  namespace atlas {
+
+    /*
+     * Function signature conversion for calling CBLAS' gesvd functions as directly as possible.
+     */
+    template <typename DType, typename CType>
+    inline static int lapack_gesvd(char jobu, char jobvt, int m, int n, void* a, int lda, void* s, void* u, int ldu, void* vt, int ldvt, void* work, int lwork, void* rwork) {
+      return gesvd<DType,CType>(jobu, jobvt, m, n, reinterpret_cast<DType*>(a), lda, reinterpret_cast<CType*>(s), reinterpret_cast<DType*>(u), ldu, reinterpret_cast<DType*>(vt), ldvt, reinterpret_cast<DType*>(work), lwork, reinterpret_cast<CType*>(rwork));
+    }
+
+    /*
+     * Function signature conversion for calling CBLAS' gesdd functions as directly as possible.
+     */
+    template <typename DType, typename CType>
+    inline static int lapack_gesdd(char jobz, int m, int n, void* a, int lda, void* s, void* u, int ldu, void* vt, int ldvt, void* work, int lwork, int* iwork, void* rwork) {
+      return gesdd<DType,CType>(jobz, m, n, reinterpret_cast<DType*>(a), lda, reinterpret_cast<CType*>(s), reinterpret_cast<DType*>(u), ldu, reinterpret_cast<DType*>(vt), ldvt, reinterpret_cast<DType*>(work), lwork, iwork, reinterpret_cast<CType*>(rwork));
+    }
+
+
+  }
+  }
+}
+
+extern "C" {
+
+///////////////////
+// Ruby Bindings //
+///////////////////
+
+void nm_math_init_atlas() {
+  VALUE cNMatrix_ATLAS = rb_define_module_under(cNMatrix, "ATLAS");
+
+  rb_define_singleton_method(cNMatrix, "has_clapack?", (METHOD)nm_atlas_has_clapack, 0);
+
+  VALUE cNMatrix_ATLAS_LAPACK = rb_define_module_under(cNMatrix_ATLAS, "LAPACK");
+
+  /* ATLAS-CLAPACK Functions */
+  rb_define_singleton_method(cNMatrix_ATLAS_LAPACK, "clapack_getrf", (METHOD)nm_atlas_clapack_getrf, 5);
+  rb_define_singleton_method(cNMatrix_ATLAS_LAPACK, "clapack_potrf", (METHOD)nm_atlas_clapack_potrf, 5);
+  rb_define_singleton_method(cNMatrix_ATLAS_LAPACK, "clapack_getrs", (METHOD)nm_atlas_clapack_getrs, 9);
+  rb_define_singleton_method(cNMatrix_ATLAS_LAPACK, "clapack_potrs", (METHOD)nm_atlas_clapack_potrs, 8);
+  rb_define_singleton_method(cNMatrix_ATLAS_LAPACK, "clapack_getri", (METHOD)nm_atlas_clapack_getri, 5);
+  rb_define_singleton_method(cNMatrix_ATLAS_LAPACK, "clapack_potri", (METHOD)nm_atlas_clapack_potri, 5);
+  rb_define_singleton_method(cNMatrix_ATLAS_LAPACK, "clapack_laswp", (METHOD)nm_atlas_clapack_laswp, 7);
+
+  /* Non-ATLAS regular LAPACK Functions called via Fortran interface */
+  rb_define_singleton_method(cNMatrix_ATLAS_LAPACK, "lapack_gesvd", (METHOD)nm_atlas_lapack_gesvd, 12);
+  rb_define_singleton_method(cNMatrix_ATLAS_LAPACK, "lapack_gesdd", (METHOD)nm_atlas_lapack_gesdd, 11);
+  rb_define_singleton_method(cNMatrix_ATLAS_LAPACK, "lapack_geev",  (METHOD)nm_atlas_lapack_geev,  12);
+
+  VALUE cNMatrix_ATLAS_BLAS = rb_define_module_under(cNMatrix_ATLAS, "BLAS");
+
+  //BLAS Level 1
+  rb_define_singleton_method(cNMatrix_ATLAS_BLAS, "cblas_scal", (METHOD)nm_atlas_cblas_scal, 4);
+  rb_define_singleton_method(cNMatrix_ATLAS_BLAS, "cblas_nrm2", (METHOD)nm_atlas_cblas_nrm2, 3);
+  rb_define_singleton_method(cNMatrix_ATLAS_BLAS, "cblas_asum", (METHOD)nm_atlas_cblas_asum, 3);
+  rb_define_singleton_method(cNMatrix_ATLAS_BLAS, "cblas_rot",  (METHOD)nm_atlas_cblas_rot,  7);
+  rb_define_singleton_method(cNMatrix_ATLAS_BLAS, "cblas_rotg", (METHOD)nm_atlas_cblas_rotg, 1);
+  rb_define_singleton_method(cNMatrix_ATLAS_BLAS, "cblas_imax", (METHOD)nm_atlas_cblas_imax, 3);
+
+  //BLAS Level 2
+  rb_define_singleton_method(cNMatrix_ATLAS_BLAS, "cblas_gemv", (METHOD)nm_atlas_cblas_gemv, 11);
+
+  //BLAS Level 3
+  rb_define_singleton_method(cNMatrix_ATLAS_BLAS, "cblas_gemm", (METHOD)nm_atlas_cblas_gemm, 14);
+  rb_define_singleton_method(cNMatrix_ATLAS_BLAS, "cblas_trsm", (METHOD)nm_atlas_cblas_trsm, 12);
+  rb_define_singleton_method(cNMatrix_ATLAS_BLAS, "cblas_trmm", (METHOD)nm_atlas_cblas_trmm, 12);
+  rb_define_singleton_method(cNMatrix_ATLAS_BLAS, "cblas_syrk", (METHOD)nm_atlas_cblas_syrk, 11);
+  rb_define_singleton_method(cNMatrix_ATLAS_BLAS, "cblas_herk", (METHOD)nm_atlas_cblas_herk, 11);
+
+}
+
+/*
+ * Simple way to check from within Ruby code if clapack functions are available, without
+ * having to wait around for an exception to be thrown.
+ */
+static VALUE nm_atlas_has_clapack(VALUE self) {
+#if defined (HAVE_CLAPACK_H) || defined (HAVE_ATLAS_CLAPACK_H)
+  return Qtrue;
+#else
+  return Qfalse;
+#endif
+}
+
+/*
+ * call-seq:
+ *     NMatrix::BLAS.cblas_scal(n, alpha, vector, inc) -> NMatrix
+ *
+ * BLAS level 1 function +scal+. Works with all dtypes.
+ *
+ * Scale +vector+ in-place by +alpha+ and also return it. The operation is as
+ * follows:
+ *  x <- alpha * x
+ *
+ * - +n+ -> Number of elements of +vector+.
+ * - +alpha+ -> Scalar value used in the operation.
+ * - +vector+ -> NMatrix of shape [n,1] or [1,n]. Modified in-place.
+ * - +inc+ -> Increment used in the scaling function. Should generally be 1.
+ */
+static VALUE nm_atlas_cblas_scal(VALUE self, VALUE n, VALUE alpha, VALUE vector, VALUE incx) {
+  nm::dtype_t dtype = NM_DTYPE(vector);
+
+  void* scalar = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]);
+  rubyval_to_cval(alpha, dtype, scalar);
+
+  NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::math::atlas::cblas_scal, void, const int n,
+      const void* scalar, void* x, const int incx);
+
+  ttable[dtype](FIX2INT(n), scalar, NM_STORAGE_DENSE(vector)->elements,
+      FIX2INT(incx));
+
+  return vector;
+}
+
+/*
+ * Call any of the cblas_xrotg functions as directly as possible.
+ *
+ * xROTG computes the elements of a Givens plane rotation matrix such that:
+ *
+ *  |  c s |   | a |   | r |
+ *  | -s c | * | b | = | 0 |
+ *
+ * where r = +- sqrt( a**2 + b**2 ) and c**2 + s**2 = 1.
+ *
+ * The Givens plane rotation can be used to introduce zero elements into a matrix selectively.
+ *
+ * This function differs from most of the other raw BLAS accessors. Instead of
+ * providing a, b, c, s as arguments, you should only provide a and b (the
+ * inputs), and you should provide them as the first two elements of any dense
+ * NMatrix type.
+ *
+ * The outputs [c,s] will be returned in a Ruby Array at the end; the input
+ * NMatrix will also be modified in-place.
+ *
+ * This function, like the other cblas_ functions, does minimal type-checking.
+ */
+static VALUE nm_atlas_cblas_rotg(VALUE self, VALUE ab) {
+  static void (*ttable[nm::NUM_DTYPES])(void* a, void* b, void* c, void* s) = {
+      NULL, NULL, NULL, NULL, NULL, // can't represent c and s as integers, so no point in having integer operations.
+      nm::math::atlas::cblas_rotg<float>,
+      nm::math::atlas::cblas_rotg<double>,
+      nm::math::atlas::cblas_rotg<nm::Complex64>,
+      nm::math::atlas::cblas_rotg<nm::Complex128>,
+      NULL //nm::math::atlas::cblas_rotg<nm::RubyObject>
+  };
+
+  nm::dtype_t dtype = NM_DTYPE(ab);
+
+  if (!ttable[dtype]) {
+    rb_raise(nm_eDataTypeError, "this operation undefined for integer vectors");
+    return Qnil;
+
+  } else {
+    NM_CONSERVATIVE(nm_register_value(&self));
+    NM_CONSERVATIVE(nm_register_value(&ab));
+    void *pC = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]),
+         *pS = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]);
+
+    // extract A and B from the NVector (first two elements)
+    void* pA = NM_STORAGE_DENSE(ab)->elements;
+    void* pB = (char*)(NM_STORAGE_DENSE(ab)->elements) + DTYPE_SIZES[dtype];
+    // c and s are output
+
+    ttable[dtype](pA, pB, pC, pS);
+
+    VALUE result = rb_ary_new2(2);
+
+    if (dtype == nm::RUBYOBJ) {
+      rb_ary_store(result, 0, *reinterpret_cast<VALUE*>(pC));
+      rb_ary_store(result, 1, *reinterpret_cast<VALUE*>(pS));
+    } else {
+      rb_ary_store(result, 0, rubyobj_from_cval(pC, dtype).rval);
+      rb_ary_store(result, 1, rubyobj_from_cval(pS, dtype).rval);
+    }
+    NM_CONSERVATIVE(nm_unregister_value(&ab));
+    NM_CONSERVATIVE(nm_unregister_value(&self));
+    return result;
+  }
+}
+
+
+/*
+ * Call any of the cblas_xrot functions as directly as possible.
+ *
+ * xROT is a BLAS level 1 routine (taking two vectors) which applies a plane rotation.
+ *
+ * It's tough to find documentation on xROT. Here are what we think the arguments are for:
+ *  * n     :: number of elements to consider in x and y
+ *  * x     :: a vector (expects an NVector)
+ *  * incx  :: stride of x
+ *  * y     :: a vector (expects an NVector)
+ *  * incy  :: stride of y
+ *  * c     :: cosine of the angle of rotation
+ *  * s     :: sine of the angle of rotation
+ *
+ * Note that c and s will be the same dtype as x and y, except when x and y are complex. If x and y are complex, c and s
+ * will be float for Complex64 or double for Complex128.
+ *
+ * You probably don't want to call this function. Instead, why don't you try rot, which is more flexible
+ * with its arguments?
+ *
+ * This function does almost no type checking. Seriously, be really careful when you call it! There's no exception
+ * handling, so you can easily crash Ruby!
+ */
+static VALUE nm_atlas_cblas_rot(VALUE self, VALUE n, VALUE x, VALUE incx, VALUE y, VALUE incy, VALUE c, VALUE s) {
+  static void (*ttable[nm::NUM_DTYPES])(const int N, void*, const int, void*, const int, const void*, const void*) = {
+      NULL, NULL, NULL, NULL, NULL, // can't represent c and s as integers, so no point in having integer operations.
+      nm::math::atlas::cblas_rot<float,float>,
+      nm::math::atlas::cblas_rot<double,double>,
+      nm::math::atlas::cblas_rot<nm::Complex64,float>,
+      nm::math::atlas::cblas_rot<nm::Complex128,double>,
+      nm::math::atlas::cblas_rot<nm::RubyObject,nm::RubyObject>
+  };
+
+  nm::dtype_t dtype = NM_DTYPE(x);
+
+
+  if (!ttable[dtype]) {
+    rb_raise(nm_eDataTypeError, "this operation undefined for integer vectors");
+    return Qfalse;
+  } else {
+    void *pC, *pS;
+
+    // We need to ensure the cosine and sine arguments are the correct dtype -- which may differ from the actual dtype.
+    if (dtype == nm::COMPLEX64) {
+      pC = NM_ALLOCA_N(float,1);
+      pS = NM_ALLOCA_N(float,1);
+      rubyval_to_cval(c, nm::FLOAT32, pC);
+      rubyval_to_cval(s, nm::FLOAT32, pS);
+    } else if (dtype == nm::COMPLEX128) {
+      pC = NM_ALLOCA_N(double,1);
+      pS = NM_ALLOCA_N(double,1);
+      rubyval_to_cval(c, nm::FLOAT64, pC);
+      rubyval_to_cval(s, nm::FLOAT64, pS);
+    } else {
+      pC = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]);
+      pS = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]);
+      rubyval_to_cval(c, dtype, pC);
+      rubyval_to_cval(s, dtype, pS);
+    }
+
+
+    ttable[dtype](FIX2INT(n), NM_STORAGE_DENSE(x)->elements, FIX2INT(incx), NM_STORAGE_DENSE(y)->elements, FIX2INT(incy), pC, pS);
+
+    return Qtrue;
+  }
+}
+
+
+/*
+ * Call any of the cblas_xnrm2 functions as directly as possible.
+ *
+ * xNRM2 is a BLAS level 1 routine which calculates the 2-norm of an n-vector x.
+ *
+ * Arguments:
+ *  * n     :: length of x, must be at least 0
+ *  * x     :: pointer to first entry of input vector
+ *  * incx  :: stride of x, must be POSITIVE (ATLAS says non-zero, but 3.8.4 code only allows positive)
+ *
+ * You probably don't want to call this function. Instead, why don't you try nrm2, which is more flexible
+ * with its arguments?
+ *
+ * This function does almost no type checking. Seriously, be really careful when you call it! There's no exception
+ * handling, so you can easily crash Ruby!
+ */
+static VALUE nm_atlas_cblas_nrm2(VALUE self, VALUE n, VALUE x, VALUE incx) {
+
+  static void (*ttable[nm::NUM_DTYPES])(const int N, const void* X, const int incX, void* sum) = {
+      NULL, NULL, NULL, NULL, NULL, // no help for integers
+      nm::math::atlas::cblas_nrm2<float32_t,float32_t>,
+      nm::math::atlas::cblas_nrm2<float64_t,float64_t>,
+      nm::math::atlas::cblas_nrm2<float32_t,nm::Complex64>,
+      nm::math::atlas::cblas_nrm2<float64_t,nm::Complex128>,
+      nm::math::atlas::cblas_nrm2<nm::RubyObject,nm::RubyObject>
+  };
+
+  nm::dtype_t dtype  = NM_DTYPE(x);
+
+  if (!ttable[dtype]) {
+    rb_raise(nm_eDataTypeError, "this operation undefined for integer vectors");
+    return Qnil;
+
+  } else {
+    // Determine the return dtype and allocate it
+    nm::dtype_t rdtype = dtype;
+    if      (dtype == nm::COMPLEX64)  rdtype = nm::FLOAT32;
+    else if (dtype == nm::COMPLEX128) rdtype = nm::FLOAT64;
+
+    void *Result = NM_ALLOCA_N(char, DTYPE_SIZES[rdtype]);
+
+    ttable[dtype](FIX2INT(n), NM_STORAGE_DENSE(x)->elements, FIX2INT(incx), Result);
+
+    return rubyobj_from_cval(Result, rdtype).rval;
+  }
+}
+
+
+
+/*
+ * Call any of the cblas_xasum functions as directly as possible.
+ *
+ * xASUM is a BLAS level 1 routine which calculates the sum of absolute values of the entries
+ * of a vector x.
+ *
+ * Arguments:
+ *  * n     :: length of x, must be at least 0
+ *  * x     :: pointer to first entry of input vector
+ *  * incx  :: stride of x, must be POSITIVE (ATLAS says non-zero, but 3.8.4 code only allows positive)
+ *
+ * You probably don't want to call this function. Instead, why don't you try asum, which is more flexible
+ * with its arguments?
+ *
+ * This function does almost no type checking. Seriously, be really careful when you call it! There's no exception
+ * handling, so you can easily crash Ruby!
+ */
+static VALUE nm_atlas_cblas_asum(VALUE self, VALUE n, VALUE x, VALUE incx) {
+
+  static void (*ttable[nm::NUM_DTYPES])(const int N, const void* X, const int incX, void* sum) = {
+      nm::math::atlas::cblas_asum<uint8_t,uint8_t>,
+      nm::math::atlas::cblas_asum<int8_t,int8_t>,
+      nm::math::atlas::cblas_asum<int16_t,int16_t>,
+      nm::math::atlas::cblas_asum<int32_t,int32_t>,
+      nm::math::atlas::cblas_asum<int64_t,int64_t>,
+      nm::math::atlas::cblas_asum<float32_t,float32_t>,
+      nm::math::atlas::cblas_asum<float64_t,float64_t>,
+      nm::math::atlas::cblas_asum<float32_t,nm::Complex64>,
+      nm::math::atlas::cblas_asum<float64_t,nm::Complex128>,
+      nm::math::atlas::cblas_asum<nm::RubyObject,nm::RubyObject>
+  };
+
+  nm::dtype_t dtype  = NM_DTYPE(x);
+
+  // Determine the return dtype and allocate it
+  nm::dtype_t rdtype = dtype;
+  if      (dtype == nm::COMPLEX64)  rdtype = nm::FLOAT32;
+  else if (dtype == nm::COMPLEX128) rdtype = nm::FLOAT64;
+
+  void *Result = NM_ALLOCA_N(char, DTYPE_SIZES[rdtype]);
+
+  ttable[dtype](FIX2INT(n), NM_STORAGE_DENSE(x)->elements, FIX2INT(incx), Result);
+
+  return rubyobj_from_cval(Result, rdtype).rval;
+}
+
+/*
+ * call-seq:
+ *    NMatrix::BLAS.cblas_imax(n, vector, inc) -> Fixnum
+ *
+ * BLAS level 1 routine.
+ *
+ * Return the index of the largest element of +vector+.
+ *
+ * - +n+ -> Vector's size. Generally, you can use NMatrix#rows or NMatrix#cols.
+ * - +vector+ -> A NMatrix of shape [n,1] or [1,n] with any dtype.
+ * - +inc+ -> It's the increment used when searching. Use 1 except if you know
+ *   what you're doing.
+ */
+static VALUE nm_atlas_cblas_imax(VALUE self, VALUE n, VALUE x, VALUE incx) {
+  NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::math::atlas::cblas_imax, int, const int n, const void* x, const int incx);
+
+  nm::dtype_t dtype = NM_DTYPE(x);
+
+  int index = ttable[dtype](FIX2INT(n), NM_STORAGE_DENSE(x)->elements, FIX2INT(incx));
+
+  // Convert to Ruby's Int value.
+  return INT2FIX(index);
+}
+
+/* Call any of the cblas_xgemv functions as directly as possible.
+ *
+ * The cblas_xgemv functions (dgemv, sgemv, cgemv, and zgemv) define the following operation:
+ *
+ *    y = alpha*op(A)*x + beta*y
+ *
+ * where op(A) is one of <tt>op(A) = A</tt>, <tt>op(A) = A**T</tt>, or the complex conjugate of A.
+ *
+ * Note that this will only work for dense matrices that are of types :float32, :float64, :complex64, and :complex128.
+ * Other types are not implemented in BLAS, and while they exist in NMatrix, this method is intended only to
+ * expose the ultra-optimized ATLAS versions.
+ *
+ * == Arguments
+ * See: http://www.netlib.org/blas/dgemm.f
+ *
+ * You probably don't want to call this function. Instead, why don't you try cblas_gemv, which is more flexible
+ * with its arguments?
+ *
+ * This function does almost no type checking. Seriously, be really careful when you call it! There's no exception
+ * handling, so you can easily crash Ruby!
+ */
+static VALUE nm_atlas_cblas_gemv(VALUE self,
+                           VALUE trans_a,
+                           VALUE m, VALUE n,
+                           VALUE alpha,
+                           VALUE a, VALUE lda,
+                           VALUE x, VALUE incx,
+                           VALUE beta,
+                           VALUE y, VALUE incy)
+{
+  NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::math::atlas::cblas_gemv, bool, const enum CBLAS_TRANSPOSE, const int, const int, const void*, const void*, const int, const void*, const int, const void*, void*, const int)
+
+  nm::dtype_t dtype = NM_DTYPE(a);
+
+  void *pAlpha = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]),
+       *pBeta  = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]);
+  rubyval_to_cval(alpha, dtype, pAlpha);
+  rubyval_to_cval(beta, dtype, pBeta);
+
+  return ttable[dtype](blas_transpose_sym(trans_a), FIX2INT(m), FIX2INT(n), pAlpha, NM_STORAGE_DENSE(a)->elements, FIX2INT(lda), NM_STORAGE_DENSE(x)->elements, FIX2INT(incx), pBeta, NM_STORAGE_DENSE(y)->elements, FIX2INT(incy)) ? Qtrue : Qfalse;
+}
+
+/* Call any of the cblas_xgemm functions as directly as possible.
+ *
+ * The cblas_xgemm functions (dgemm, sgemm, cgemm, and zgemm) define the following operation:
+ *
+ *    C = alpha*op(A)*op(B) + beta*C
+ *
+ * where op(X) is one of <tt>op(X) = X</tt>, <tt>op(X) = X**T</tt>, or the complex conjugate of X.
+ *
+ * Note that this will only work for dense matrices that are of types :float32, :float64, :complex64, and :complex128.
+ * Other types are not implemented in BLAS, and while they exist in NMatrix, this method is intended only to
+ * expose the ultra-optimized ATLAS versions.
+ *
+ * == Arguments
+ * See: http://www.netlib.org/blas/dgemm.f
+ *
+ * You probably don't want to call this function. Instead, why don't you try gemm, which is more flexible
+ * with its arguments?
+ *
+ * This function does almost no type checking. Seriously, be really careful when you call it! There's no exception
+ * handling, so you can easily crash Ruby!
+ */
+static VALUE nm_atlas_cblas_gemm(VALUE self,
+                           VALUE order,
+                           VALUE trans_a, VALUE trans_b,
+                           VALUE m, VALUE n, VALUE k,
+                           VALUE alpha,
+                           VALUE a, VALUE lda,
+                           VALUE b, VALUE ldb,
+                           VALUE beta,
+                           VALUE c, VALUE ldc)
+{
+  NAMED_DTYPE_TEMPLATE_TABLE(ttable, nm::math::atlas::cblas_gemm, void, const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE trans_a, const enum CBLAS_TRANSPOSE trans_b, int m, int n, int k, void* alpha, void* a, int lda, void* b, int ldb, void* beta, void* c, int ldc);
+
+  nm::dtype_t dtype = NM_DTYPE(a);
+
+  void *pAlpha = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]),
+       *pBeta  = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]);
+  rubyval_to_cval(alpha, dtype, pAlpha);
+  rubyval_to_cval(beta, dtype, pBeta);
+
+  ttable[dtype](blas_order_sym(order), blas_transpose_sym(trans_a), blas_transpose_sym(trans_b), FIX2INT(m), FIX2INT(n), FIX2INT(k), pAlpha, NM_STORAGE_DENSE(a)->elements, FIX2INT(lda), NM_STORAGE_DENSE(b)->elements, FIX2INT(ldb), pBeta, NM_STORAGE_DENSE(c)->elements, FIX2INT(ldc));
+
+  return c;
+}
+
+
+static VALUE nm_atlas_cblas_trsm(VALUE self,
+                           VALUE order,
+                           VALUE side, VALUE uplo,
+                           VALUE trans_a, VALUE diag,
+                           VALUE m, VALUE n,
+                           VALUE alpha,
+                           VALUE a, VALUE lda,
+                           VALUE b, VALUE ldb)
+{
+  static void (*ttable[nm::NUM_DTYPES])(const enum CBLAS_ORDER, const enum CBLAS_SIDE, const enum CBLAS_UPLO,
+                                        const enum CBLAS_TRANSPOSE, const enum CBLAS_DIAG,
+                                        const int m, const int n, const void* alpha, const void* a,
+                                        const int lda, void* b, const int ldb) = {
+      NULL, NULL, NULL, NULL, NULL, // integers not allowed due to division
+      nm::math::atlas::cblas_trsm<float>,
+      nm::math::atlas::cblas_trsm<double>,
+      cblas_ctrsm, cblas_ztrsm, // call directly, same function signature!
+      nm::math::atlas::cblas_trsm<nm::RubyObject>
+  };
+
+  nm::dtype_t dtype = NM_DTYPE(a);
+
+  if (!ttable[dtype]) {
+    rb_raise(nm_eDataTypeError, "this matrix operation undefined for integer matrices");
+  } else {
+    void *pAlpha = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]);
+    rubyval_to_cval(alpha, dtype, pAlpha);
+
+    ttable[dtype](blas_order_sym(order), blas_side_sym(side), blas_uplo_sym(uplo), blas_transpose_sym(trans_a), blas_diag_sym(diag), FIX2INT(m), FIX2INT(n), pAlpha, NM_STORAGE_DENSE(a)->elements, FIX2INT(lda), NM_STORAGE_DENSE(b)->elements, FIX2INT(ldb));
+  }
+
+  return Qtrue;
+}
+
+static VALUE nm_atlas_cblas_trmm(VALUE self,
+                           VALUE order,
+                           VALUE side, VALUE uplo,
+                           VALUE trans_a, VALUE diag,
+                           VALUE m, VALUE n,
+                           VALUE alpha,
+                           VALUE a, VALUE lda,
+                           VALUE b, VALUE ldb)
+{
+  static void (*ttable[nm::NUM_DTYPES])(const enum CBLAS_ORDER,
+                                        const enum CBLAS_SIDE, const enum CBLAS_UPLO,
+                                        const enum CBLAS_TRANSPOSE, const enum CBLAS_DIAG,
+                                        const int m, const int n, const void* alpha, const void* a,
+                                        const int lda, void* b, const int ldb) = {
+      NULL, NULL, NULL, NULL, NULL, // integers not allowed due to division
+      nm::math::atlas::cblas_trmm<float>,
+      nm::math::atlas::cblas_trmm<double>,
+      cblas_ctrmm, cblas_ztrmm, // call directly, same function signature!
+      NULL
+  };
+
+  nm::dtype_t dtype = NM_DTYPE(a);
+
+  if (!ttable[dtype]) {
+    rb_raise(nm_eDataTypeError, "this matrix operation not yet defined for non-BLAS dtypes");
+  } else {
+    void *pAlpha = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]);
+    rubyval_to_cval(alpha, dtype, pAlpha);
+
+    ttable[dtype](blas_order_sym(order), blas_side_sym(side), blas_uplo_sym(uplo), blas_transpose_sym(trans_a), blas_diag_sym(diag), FIX2INT(m), FIX2INT(n), pAlpha, NM_STORAGE_DENSE(a)->elements, FIX2INT(lda), NM_STORAGE_DENSE(b)->elements, FIX2INT(ldb));
+  }
+
+  return b;
+}
+
+static VALUE nm_atlas_cblas_syrk(VALUE self,
+                           VALUE order,
+                           VALUE uplo,
+                           VALUE trans,
+                           VALUE n, VALUE k,
+                           VALUE alpha,
+                           VALUE a, VALUE lda,
+                           VALUE beta,
+                           VALUE c, VALUE ldc)
+{
+  static void (*ttable[nm::NUM_DTYPES])(const enum CBLAS_ORDER, const enum CBLAS_UPLO, const enum CBLAS_TRANSPOSE,
+                                        const int n, const int k, const void* alpha, const void* a,
+                                        const int lda, const void* beta, void* c, const int ldc) = {
+      NULL, NULL, NULL, NULL, NULL, // integers not allowed due to division
+      nm::math::atlas::cblas_syrk<float>,
+      nm::math::atlas::cblas_syrk<double>,
+      cblas_csyrk, cblas_zsyrk, // call directly, same function signature!
+      NULL
+  };
+
+  nm::dtype_t dtype = NM_DTYPE(a);
+
+  if (!ttable[dtype]) {
+    rb_raise(nm_eDataTypeError, "this matrix operation undefined for integer matrices");
+  } else {
+    void *pAlpha = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]),
+         *pBeta = NM_ALLOCA_N(char, DTYPE_SIZES[dtype]);
+    rubyval_to_cval(alpha, dtype, pAlpha);
+    rubyval_to_cval(beta, dtype, pBeta);
+
+    ttable[dtype](blas_order_sym(order), blas_uplo_sym(uplo), blas_transpose_sym(trans), FIX2INT(n), FIX2INT(k), pAlpha, NM_STORAGE_DENSE(a)->elements, FIX2INT(lda), pBeta, NM_STORAGE_DENSE(c)->elements, FIX2INT(ldc));
+  }
+
+  return Qtrue;
+}
+
+static VALUE nm_atlas_cblas_herk(VALUE self,
+                           VALUE order,
+                           VALUE uplo,
+                           VALUE trans,
+                           VALUE n, VALUE k,
+                           VALUE alpha,
+                           VALUE a, VALUE lda,
+                           VALUE beta,
+                           VALUE c, VALUE ldc)
+{
+
+  nm::dtype_t dtype = NM_DTYPE(a);
+
+  if (dtype == nm::COMPLEX64) {
+    cblas_cherk(blas_order_sym(order), blas_uplo_sym(uplo), blas_transpose_sym(trans), FIX2INT(n), FIX2INT(k), NUM2DBL(alpha), NM_STORAGE_DENSE(a)->elements, FIX2INT(lda), NUM2DBL(beta), NM_STORAGE_DENSE(c)->elements, FIX2INT(ldc));
+  } else if (dtype == nm::COMPLEX128) {
+    cblas_zherk(blas_order_sym(order), blas_uplo_sym(uplo), blas_transpose_sym(trans), FIX2INT(n), FIX2INT(k), NUM2DBL(alpha), NM_STORAGE_DENSE(a)->elements, FIX2INT(lda), NUM2DBL(beta), NM_STORAGE_DENSE(c)->elements, FIX2INT(ldc));
+  } else
+    rb_raise(rb_eNotImpError, "this matrix operation undefined for non-complex dtypes");
+  return Qtrue;
+}
+
+/*
+ * Function signature conversion for calling CBLAS' gesvd functions as directly as possible.
+ *
+ * xGESVD computes the singular value decomposition (SVD) of a real
+ * M-by-N matrix A, optionally computing the left and/or right singular
+ * vectors. The SVD is written
+ *
+ *      A = U * SIGMA * transpose(V)
+ *
+ * where SIGMA is an M-by-N matrix which is zero except for its
+ * min(m,n) diagonal elements, U is an M-by-M orthogonal matrix, and
+ * V is an N-by-N orthogonal matrix.  The diagonal elements of SIGMA
+ * are the singular values of A; they are real and non-negative, and
+ * are returned in descending order.  The first min(m,n) columns of
+ * U and V are the left and right singular vectors of A.
+ *
+ * Note that the routine returns V**T, not V.
+ */
+static VALUE nm_atlas_lapack_gesvd(VALUE self, VALUE jobu, VALUE jobvt, VALUE m, VALUE n, VALUE a, VALUE lda, VALUE s, VALUE u, VALUE ldu, VALUE vt, VALUE ldvt, VALUE lwork) {
+  static int (*gesvd_table[nm::NUM_DTYPES])(char, char, int, int, void* a, int, void* s, void* u, int, void* vt, int, void* work, int, void* rwork) = {
+    NULL, NULL, NULL, NULL, NULL, // no integer ops
+    nm::math::atlas::lapack_gesvd<float,float>,
+    nm::math::atlas::lapack_gesvd<double,double>,
+    nm::math::atlas::lapack_gesvd<nm::Complex64,float>,
+    nm::math::atlas::lapack_gesvd<nm::Complex128,double>,
+    NULL // no Ruby objects
+  };
+
+  nm::dtype_t dtype = NM_DTYPE(a);
+
+
+  if (!gesvd_table[dtype]) {
+    rb_raise(rb_eNotImpError, "this operation not yet implemented for non-BLAS dtypes");
+    return Qfalse;
+  } else {
+    int M = FIX2INT(m),
+        N = FIX2INT(n);
+
+    int min_mn  = NM_MIN(M,N);
+    int max_mn  = NM_MAX(M,N);
+
+    char JOBU = lapack_svd_job_sym(jobu),
+         JOBVT = lapack_svd_job_sym(jobvt);
+
+    // only need rwork for complex matrices
+    int rwork_size  = (dtype == nm::COMPLEX64 || dtype == nm::COMPLEX128) ? 5 * min_mn : 0;
+    void* rwork     = rwork_size > 0 ? NM_ALLOCA_N(char, DTYPE_SIZES[dtype] * rwork_size) : NULL;
+    int work_size   = FIX2INT(lwork);
+
+    // ignore user argument for lwork if it's too small.
+    work_size       = NM_MAX((dtype == nm::COMPLEX64 || dtype == nm::COMPLEX128 ? 2 * min_mn + max_mn : NM_MAX(3*min_mn + max_mn, 5*min_mn)), work_size);
+    void* work      = NM_ALLOCA_N(char, DTYPE_SIZES[dtype] * work_size);
+
+    int info = gesvd_table[dtype](JOBU, JOBVT, M, N, NM_STORAGE_DENSE(a)->elements, FIX2INT(lda),
+      NM_STORAGE_DENSE(s)->elements, NM_STORAGE_DENSE(u)->elements, FIX2INT(ldu), NM_STORAGE_DENSE(vt)->elements, FIX2INT(ldvt),
+      work, work_size, rwork);
+    return INT2FIX(info);
+  }
+}
+
+/*
+ * Function signature conversion for calling CBLAS' gesdd functions as directly as possible.
+ *
+ * xGESDD uses a divide-and-conquer strategy to compute the singular value decomposition (SVD) of a real
+ * M-by-N matrix A, optionally computing the left and/or right singular
+ * vectors. The SVD is written
+ *
+ *      A = U * SIGMA * transpose(V)
+ *
+ * where SIGMA is an M-by-N matrix which is zero except for its
+ * min(m,n) diagonal elements, U is an M-by-M orthogonal matrix, and
+ * V is an N-by-N orthogonal matrix.  The diagonal elements of SIGMA
+ * are the singular values of A; they are real and non-negative, and
+ * are returned in descending order.  The first min(m,n) columns of
+ * U and V are the left and right singular vectors of A.
+ *
+ * Note that the routine returns V**T, not V.
+ */
+static VALUE nm_atlas_lapack_gesdd(VALUE self, VALUE jobz, VALUE m, VALUE n, VALUE a, VALUE lda, VALUE s, VALUE u, VALUE ldu, VALUE vt, VALUE ldvt, VALUE lwork) {
+  static int (*gesdd_table[nm::NUM_DTYPES])(char, int, int, void* a, int, void* s, void* u, int, void* vt, int, void* work, int, int* iwork, void* rwork) = {
+    NULL, NULL, NULL, NULL, NULL, // no integer ops
+    nm::math::atlas::lapack_gesdd<float,float>,
+    nm::math::atlas::lapack_gesdd<double,double>,
+    nm::math::atlas::lapack_gesdd<nm::Complex64,float>,
+    nm::math::atlas::lapack_gesdd<nm::Complex128,double>,
+    NULL // no Ruby objects
+  };
+
+  nm::dtype_t dtype = NM_DTYPE(a);
+
+  if (!gesdd_table[dtype]) {
+    rb_raise(rb_eNotImpError, "this operation not yet implemented for non-BLAS dtypes");
+    return Qfalse;
+  } else {
+    int M = FIX2INT(m),
+        N = FIX2INT(n);
+
+    int min_mn  = NM_MIN(M,N);
+    int max_mn  = NM_MAX(M,N);
+
+    char JOBZ = lapack_svd_job_sym(jobz);
+
+    // only need rwork for complex matrices
+    void* rwork = NULL;
+
+    int work_size = FIX2INT(lwork); // Make sure we allocate enough work, regardless of the user request.
+    if (dtype == nm::COMPLEX64 || dtype == nm::COMPLEX128) {
+      int rwork_size = min_mn * (JOBZ == 'N' ? 5 : NM_MAX(5*min_mn + 7, 2*max_mn + 2*min_mn + 1));
+      rwork = NM_ALLOCA_N(char, DTYPE_SIZES[dtype] * rwork_size);
+
+      if (JOBZ == 'N')      work_size = NM_MAX(work_size, 3*min_mn + NM_MAX(max_mn, 6*min_mn));
+      else if (JOBZ == 'O') work_size = NM_MAX(work_size, 3*min_mn*min_mn + NM_MAX(max_mn, 5*min_mn*min_mn + 4*min_mn));
+      else                  work_size = NM_MAX(work_size, 3*min_mn*min_mn + NM_MAX(max_mn, 4*min_mn*min_mn + 4*min_mn));
+    } else {
+      if (JOBZ == 'N')      work_size = NM_MAX(work_size, 2*min_mn + max_mn);
+      else if (JOBZ == 'O') work_size = NM_MAX(work_size, 2*min_mn*min_mn + max_mn + 2*min_mn);
+      else                  work_size = NM_MAX(work_size, min_mn*min_mn + max_mn + 2*min_mn);
+    }
+    void* work  = NM_ALLOCA_N(char, DTYPE_SIZES[dtype] * work_size);
+    int* iwork  = NM_ALLOCA_N(int, 8*min_mn);
+
+    int info = gesdd_table[dtype](JOBZ, M, N, NM_STORAGE_DENSE(a)->elements, FIX2INT(lda),
+      NM_STORAGE_DENSE(s)->elements, NM_STORAGE_DENSE(u)->elements, FIX2INT(ldu), NM_STORAGE_DENSE(vt)->elements, FIX2INT(ldvt),
+      work, work_size, iwork, rwork);
+    return INT2FIX(info);
+  }
+}
+
+/*
+ * Function signature conversion for calling CBLAS' geev functions as directly as possible.
+ *
+ * GEEV computes for an N-by-N real nonsymmetric matrix A, the
+ * eigenvalues and, optionally, the left and/or right eigenvectors.
+ *
+ * The right eigenvector v(j) of A satisfies
+ *                    A * v(j) = lambda(j) * v(j)
+ * where lambda(j) is its eigenvalue.
+ *
+ * The left eigenvector u(j) of A satisfies
+ *                 u(j)**H * A = lambda(j) * u(j)**H
+ * where u(j)**H denotes the conjugate transpose of u(j).
+ *
+ * The computed eigenvectors are normalized to have Euclidean norm
+ * equal to 1 and largest component real.
+ */
+static VALUE nm_atlas_lapack_geev(VALUE self, VALUE compute_left, VALUE compute_right, VALUE n, VALUE a, VALUE lda, VALUE w, VALUE wi, VALUE vl, VALUE ldvl, VALUE vr, VALUE ldvr, VALUE lwork) {
+  static int (*geev_table[nm::NUM_DTYPES])(char, char, int, void* a, int, void* w, void* wi, void* vl, int, void* vr, int, void* work, int, void* rwork) = {
+    NULL, NULL, NULL, NULL, NULL, // no integer ops
+    nm::math::atlas::lapack_geev<float,float>,
+    nm::math::atlas::lapack_geev<double,double>,
+    nm::math::atlas::lapack_geev<nm::Complex64,float>,
+    nm::math::atlas::lapack_geev<nm::Complex128,double>,
+    NULL // no Ruby objects
+  };
+
+  nm::dtype_t dtype = NM_DTYPE(a);
+
+
+  if (!geev_table[dtype]) {
+    rb_raise(rb_eNotImpError, "this operation not yet implemented for non-BLAS dtypes");
+    return Qfalse;
+  } else {
+    int N = FIX2INT(n);
+
+    char JOBVL = lapack_evd_job_sym(compute_left),
+         JOBVR = lapack_evd_job_sym(compute_right);
+
+    void* A  = NM_STORAGE_DENSE(a)->elements;
+    void* WR = NM_STORAGE_DENSE(w)->elements;
+    void* WI = wi == Qnil ? NULL : NM_STORAGE_DENSE(wi)->elements;
+    void* VL = JOBVL == 'V' ? NM_STORAGE_DENSE(vl)->elements : NULL;
+    void* VR = JOBVR == 'V' ? NM_STORAGE_DENSE(vr)->elements : NULL;
+
+    // only need rwork for complex matrices (wi == Qnil for complex)
+    int rwork_size  = dtype == nm::COMPLEX64 || dtype == nm::COMPLEX128 ? N * DTYPE_SIZES[dtype] : 0; // 2*N*floattype for complex only, otherwise 0
+    void* rwork     = rwork_size > 0 ? NM_ALLOCA_N(char, rwork_size) : NULL;
+    int work_size   = FIX2INT(lwork);
+    void* work;
+
+    int info;
+
+    // if work size is 0 or -1, query.
+    if (work_size <= 0) {
+      work_size = -1;
+      work = NM_ALLOC_N(char, DTYPE_SIZES[dtype]); //2*N * DTYPE_SIZES[dtype]);
+      info = geev_table[dtype](JOBVL, JOBVR, N, A, FIX2INT(lda), WR, WI, VL, FIX2INT(ldvl), VR, FIX2INT(ldvr), work, work_size, rwork);
+      work_size = (int)(dtype == nm::COMPLEX64 || dtype == nm::FLOAT32 ? reinterpret_cast<float*>(work)[0] : reinterpret_cast<double*>(work)[0]);
+      // line above is basically: work_size = (int)(work[0]); // now have new work_size
+      NM_FREE(work);
+      if (info == 0)
+        rb_warn("geev: calculated optimal lwork of %d; to eliminate this message, use a positive value for lwork (at least 2*shape[i])", work_size);
+      else return INT2FIX(info); // error of some kind on query!
+    }
+
+    // if work size is < 2*N, just set it to 2*N
+    if (work_size < 2*N) work_size = 2*N;
+    if (work_size < 3*N && (dtype == nm::FLOAT32 || dtype == nm::FLOAT64)) {
+      work_size = JOBVL == 'V' || JOBVR == 'V' ? 4*N : 3*N;
+    }
+
+    // Allocate work array for actual run
+    work = NM_ALLOCA_N(char, work_size * DTYPE_SIZES[dtype]);
+
+    // Perform the actual calculation.
+    info = geev_table[dtype](JOBVL, JOBVR, N, A, FIX2INT(lda), WR, WI, VL, FIX2INT(ldvl), VR, FIX2INT(ldvr), work, work_size, rwork);
+
+    return INT2FIX(info);
+  }
+}
+
+/* Call any of the clapack_xgetrf functions as directly as possible.
+ *
+ * The clapack_getrf functions (dgetrf, sgetrf, cgetrf, and zgetrf) compute an LU factorization of a general M-by-N
+ * matrix A using partial pivoting with row interchanges.
+ *
+ * The factorization has the form:
+ *    A = P * L * U
+ * where P is a permutation matrix, L is lower triangular with unit diagonal elements (lower trapezoidal if m > n),
+ * and U is upper triangular (upper trapezoidal if m < n).
+ *
+ * This is the right-looking level 3 BLAS version of the algorithm.
+ *
+ * == Arguments
+ * See: http://www.netlib.org/lapack/double/dgetrf.f
+ * (You don't need argument 5; this is the value returned by this function.)
+ *
+ * You probably don't want to call this function. Instead, why don't you try clapack_getrf, which is more flexible
+ * with its arguments?
+ *
+ * This function does almost no type checking. Seriously, be really careful when you call it! There's no exception
+ * handling, so you can easily crash Ruby!
+ *
+ * Returns an array giving the pivot indices (normally these are argument #5).
+ */
+static VALUE nm_atlas_clapack_getrf(VALUE self, VALUE order, VALUE m, VALUE n, VALUE a, VALUE lda) {
+  static int (*ttable[nm::NUM_DTYPES])(const enum CBLAS_ORDER, const int m, const int n, void* a, const int lda, int* ipiv) = {
+      NULL, NULL, NULL, NULL, NULL, // integers not allowed due to division
+      nm::math::atlas::clapack_getrf<float>,
+      nm::math::atlas::clapack_getrf<double>,
+#if defined (HAVE_CLAPACK_H) || defined (HAVE_ATLAS_CLAPACK_H)
+      clapack_cgetrf, clapack_zgetrf, // call directly, same function signature!
+#else // Especially important for Mac OS, which doesn't seem to include the ATLAS clapack interface.
+      nm::math::atlas::clapack_getrf<nm::Complex64>,
+      nm::math::atlas::clapack_getrf<nm::Complex128>,
+#endif
+      nm::math::atlas::clapack_getrf<nm::RubyObject>
+  };
+
+  int M = FIX2INT(m),
+      N = FIX2INT(n);
+
+  // Allocate the pivot index array, which is of size MIN(M, N).
+  size_t ipiv_size = std::min(M,N);
+  int* ipiv = NM_ALLOCA_N(int, ipiv_size);
+
+  if (!ttable[NM_DTYPE(a)]) {
+    rb_raise(nm_eDataTypeError, "this matrix operation undefined for integer matrices");
+  } else {
+    // Call either our version of getrf or the LAPACK version.
+    ttable[NM_DTYPE(a)](blas_order_sym(order), M, N, NM_STORAGE_DENSE(a)->elements, FIX2INT(lda), ipiv);
+  }
+
+  // Result will be stored in a. We return ipiv as an array.
+  VALUE ipiv_array = rb_ary_new2(ipiv_size);
+  for (size_t i = 0; i < ipiv_size; ++i) {
+    rb_ary_store(ipiv_array, i, INT2FIX(ipiv[i]));
+  }
+
+  return ipiv_array;
+}
+
+
+/* Call any of the clapack_xpotrf functions as directly as possible.
+ *
+ * You probably don't want to call this function. Instead, why don't you try clapack_potrf, which is more flexible
+ * with its arguments?
+ *
+ * This function does almost no type checking. Seriously, be really careful when you call it! There's no exception
+ * handling, so you can easily crash Ruby!
+ */
+static VALUE nm_atlas_clapack_potrf(VALUE self, VALUE order, VALUE uplo, VALUE n, VALUE a, VALUE lda) {
+#if !defined(HAVE_CLAPACK_H) && !defined(HAVE_ATLAS_CLAPACK_H)
+  rb_raise(rb_eNotImpError, "potrf currently requires CLAPACK");
+#endif
+
+  static int (*ttable[nm::NUM_DTYPES])(const enum CBLAS_ORDER, const enum CBLAS_UPLO, const int n, void* a, const int lda) = {
+      NULL, NULL, NULL, NULL, NULL, // integers not allowed due to division
+      nm::math::atlas::clapack_potrf<float>,
+      nm::math::atlas::clapack_potrf<double>,
+#if defined (HAVE_CLAPACK_H) || defined (HAVE_ATLAS_CLAPACK_H)
+      clapack_cpotrf, clapack_zpotrf, // call directly, same function signature!
+#else // Especially important for Mac OS, which doesn't seem to include the ATLAS clapack interface.
+      nm::math::atlas::clapack_potrf<nm::Complex64>,
+      nm::math::atlas::clapack_potrf<nm::Complex128>,
+#endif
+      NULL
+  };
+
+  if (!ttable[NM_DTYPE(a)]) {
+    rb_raise(rb_eNotImpError, "this operation not yet implemented for non-BLAS dtypes");
+    // FIXME: Once BLAS dtypes are implemented, replace error above with the error below.
+    //rb_raise(nm_eDataTypeError, "this matrix operation undefined for integer matrices");
+  } else {
+    // Call either our version of potrf or the LAPACK version.
+    ttable[NM_DTYPE(a)](blas_order_sym(order), blas_uplo_sym(uplo), FIX2INT(n), NM_STORAGE_DENSE(a)->elements, FIX2INT(lda));
+  }
+
+  return a;
+}
+
+
+/*
+ * Call any of the clapack_xgetrs functions as directly as possible.
+ */
+static VALUE nm_atlas_clapack_getrs(VALUE self, VALUE order, VALUE trans, VALUE n, VALUE nrhs, VALUE a, VALUE lda, VALUE ipiv, VALUE b, VALUE ldb) {
+  static int (*ttable[nm::NUM_DTYPES])(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE Trans, const int N,
+                                       const int NRHS, const void* A, const int lda, const int* ipiv, void* B,
+                                       const int ldb) = {
+      NULL, NULL, NULL, NULL, NULL, // integers not allowed due to division
+      nm::math::atlas::clapack_getrs<float>,
+      nm::math::atlas::clapack_getrs<double>,
+#if defined (HAVE_CLAPACK_H) || defined (HAVE_ATLAS_CLAPACK_H)
+      clapack_cgetrs, clapack_zgetrs, // call directly, same function signature!
+#else // Especially important for Mac OS, which doesn't seem to include the ATLAS clapack interface.
+      nm::math::atlas::clapack_getrs<nm::Complex64>,
+      nm::math::atlas::clapack_getrs<nm::Complex128>,
+#endif
+      nm::math::atlas::clapack_getrs<nm::RubyObject>
+  };
+
+  // Allocate the C version of the pivot index array
+  int* ipiv_;
+  if (TYPE(ipiv) != T_ARRAY) {
+    rb_raise(rb_eArgError, "ipiv must be of type Array");
+  } else {
+    ipiv_ = NM_ALLOCA_N(int, RARRAY_LEN(ipiv));
+    for (int index = 0; index < RARRAY_LEN(ipiv); ++index) {
+      ipiv_[index] = FIX2INT( RARRAY_PTR(ipiv)[index] );
+    }
+  }
+
+  if (!ttable[NM_DTYPE(a)]) {
+    rb_raise(nm_eDataTypeError, "this matrix operation undefined for integer matrices");
+  } else {
+
+    // Call either our version of getrs or the LAPACK version.
+    ttable[NM_DTYPE(a)](blas_order_sym(order), blas_transpose_sym(trans), FIX2INT(n), FIX2INT(nrhs), NM_STORAGE_DENSE(a)->elements, FIX2INT(lda),
+                        ipiv_, NM_STORAGE_DENSE(b)->elements, FIX2INT(ldb));
+  }
+
+  // b is both returned and modified directly in the argument list.
+  return b;
+}
+
+
+/*
+ * Call any of the clapack_xpotrs functions as directly as possible.
+ */
+static VALUE nm_atlas_clapack_potrs(VALUE self, VALUE order, VALUE uplo, VALUE n, VALUE nrhs, VALUE a, VALUE lda, VALUE b, VALUE ldb) {
+  static int (*ttable[nm::NUM_DTYPES])(const enum CBLAS_ORDER Order, const enum CBLAS_UPLO Uplo, const int N,
+                                       const int NRHS, const void* A, const int lda, void* B, const int ldb) = {
+      NULL, NULL, NULL, NULL, NULL, // integers not allowed due to division
+      nm::math::atlas::clapack_potrs<float>,
+      nm::math::atlas::clapack_potrs<double>,
+#if defined (HAVE_CLAPACK_H) || defined (HAVE_ATLAS_CLAPACK_H)
+      clapack_cpotrs, clapack_zpotrs, // call directly, same function signature!
+#else // Especially important for Mac OS, which doesn't seem to include the ATLAS clapack interface.
+      nm::math::atlas::clapack_potrs<nm::Complex64>,
+      nm::math::atlas::clapack_potrs<nm::Complex128>,
+#endif
+      nm::math::atlas::clapack_potrs<nm::RubyObject>
+  };
+
+
+  if (!ttable[NM_DTYPE(a)]) {
+    rb_raise(nm_eDataTypeError, "this matrix operation undefined for integer matrices");
+  } else {
+
+    // Call either our version of potrs or the LAPACK version.
+    ttable[NM_DTYPE(a)](blas_order_sym(order), blas_uplo_sym(uplo), FIX2INT(n), FIX2INT(nrhs), NM_STORAGE_DENSE(a)->elements, FIX2INT(lda),
+                        NM_STORAGE_DENSE(b)->elements, FIX2INT(ldb));
+  }
+
+  // b is both returned and modified directly in the argument list.
+  return b;
+}
+
+/* Call any of the clapack_xgetri functions as directly as possible.
+ *
+ * You probably don't want to call this function. Instead, why don't you try clapack_getri, which is more flexible
+ * with its arguments?
+ *
+ * This function does almost no type checking. Seriously, be really careful when you call it! There's no exception
+ * handling, so you can easily crash Ruby!
+ *
+ * Returns an array giving the pivot indices (normally these are argument #5).
+ */
+static VALUE nm_atlas_clapack_getri(VALUE self, VALUE order, VALUE n, VALUE a, VALUE lda, VALUE ipiv) {
+#if !defined (HAVE_CLAPACK_H) && !defined (HAVE_ATLAS_CLAPACK_H)
+  rb_raise(rb_eNotImpError, "getri currently requires CLAPACK");
+#endif
+
+  static int (*ttable[nm::NUM_DTYPES])(const enum CBLAS_ORDER, const int n, void* a, const int lda, const int* ipiv) = {
+      NULL, NULL, NULL, NULL, NULL, // integers not allowed due to division
+      nm::math::atlas::clapack_getri<float>,
+      nm::math::atlas::clapack_getri<double>,
+#if defined (HAVE_CLAPACK_H) || defined (HAVE_ATLAS_CLAPACK_H)
+      clapack_cgetri, clapack_zgetri, // call directly, same function signature!
+#else // Especially important for Mac OS, which doesn't seem to include the ATLAS clapack interface.
+      nm::math::atlas::clapack_getri<nm::Complex64>,
+      nm::math::atlas::clapack_getri<nm::Complex128>,
+#endif
+      NULL
+  };
+
+  // Allocate the C version of the pivot index array
+  int* ipiv_;
+  if (TYPE(ipiv) != T_ARRAY) {
+    rb_raise(rb_eArgError, "ipiv must be of type Array");
+  } else {
+    ipiv_ = NM_ALLOCA_N(int, RARRAY_LEN(ipiv));
+    for (int index = 0; index < RARRAY_LEN(ipiv); ++index) {
+      ipiv_[index] = FIX2INT( RARRAY_PTR(ipiv)[index] );
+    }
+  }
+
+  if (!ttable[NM_DTYPE(a)]) {
+    rb_raise(rb_eNotImpError, "this operation not yet implemented for non-BLAS dtypes");
+    // FIXME: Once non-BLAS dtypes are implemented, replace error above with the error below.
+    //rb_raise(nm_eDataTypeError, "this matrix operation undefined for integer matrices");
+  } else {
+    // Call either our version of getri or the LAPACK version.
+    ttable[NM_DTYPE(a)](blas_order_sym(order), FIX2INT(n), NM_STORAGE_DENSE(a)->elements, FIX2INT(lda), ipiv_);
+  }
+
+  return a;
+}
+
+
+/* Call any of the clapack_xpotri functions as directly as possible.
+ *
+ * You probably don't want to call this function. Instead, why don't you try clapack_potri, which is more flexible
+ * with its arguments?
+ *
+ * This function does almost no type checking. Seriously, be really careful when you call it! There's no exception
+ * handling, so you can easily crash Ruby!
+ */
+static VALUE nm_atlas_clapack_potri(VALUE self, VALUE order, VALUE uplo, VALUE n, VALUE a, VALUE lda) {
+#if !defined (HAVE_CLAPACK_H) && !defined (HAVE_ATLAS_CLAPACK_H)
+  rb_raise(rb_eNotImpError, "getri currently requires CLAPACK");
+#endif
+
+  static int (*ttable[nm::NUM_DTYPES])(const enum CBLAS_ORDER, const enum CBLAS_UPLO, const int n, void* a, const int lda) = {
+      NULL, NULL, NULL, NULL, NULL, // integers not allowed due to division
+      nm::math::atlas::clapack_potri<float>,
+      nm::math::atlas::clapack_potri<double>,
+#if defined (HAVE_CLAPACK_H) || defined (HAVE_ATLAS_CLAPACK_H)
+      clapack_cpotri, clapack_zpotri, // call directly, same function signature!
+#else // Especially important for Mac OS, which doesn't seem to include the ATLAS clapack interface.
+      nm::math::atlas::clapack_potri<nm::Complex64>,
+      nm::math::atlas::clapack_potri<nm::Complex128>,
+#endif
+      NULL
+  };
+
+  if (!ttable[NM_DTYPE(a)]) {
+    rb_raise(rb_eNotImpError, "this operation not yet implemented for non-BLAS dtypes");
+    // FIXME: Once BLAS dtypes are implemented, replace error above with the error below.
+    //rb_raise(nm_eDataTypeError, "this matrix operation undefined for integer matrices");
+  } else {
+    // Call either our version of getri or the LAPACK version.
+    ttable[NM_DTYPE(a)](blas_order_sym(order), blas_uplo_sym(uplo), FIX2INT(n), NM_STORAGE_DENSE(a)->elements, FIX2INT(lda));
+  }
+
+  return a;
+}
+
+
+/*
+ * Call any of the clapack_xlaswp functions as directly as possible.
+ *
+ * Note that LAPACK's xlaswp functions accept a column-order matrix, but NMatrix uses row-order. Thus, n should be the
+ * number of rows and lda should be the number of columns, no matter what it says in the documentation for dlaswp.f.
+ */
+static VALUE nm_atlas_clapack_laswp(VALUE self, VALUE n, VALUE a, VALUE lda, VALUE k1, VALUE k2, VALUE ipiv, VALUE incx) {
+  //We have actually never used the ATLAS version of laswp. For the time being
+  //I will leave it like that and just always call the internal implementation.
+  //I don't know if there is a good reason for this or not.
+  //Maybe because our internal version swaps columns instead of rows.
+  static void (*ttable[nm::NUM_DTYPES])(const int n, void* a, const int lda, const int k1, const int k2, const int* ipiv, const int incx) = {
+      nm::math::clapack_laswp<uint8_t>,
+      nm::math::clapack_laswp<int8_t>,
+      nm::math::clapack_laswp<int16_t>,
+      nm::math::clapack_laswp<int32_t>,
+      nm::math::clapack_laswp<int64_t>,
+      nm::math::clapack_laswp<float>,
+      nm::math::clapack_laswp<double>,
+      nm::math::clapack_laswp<nm::Complex64>,
+      nm::math::clapack_laswp<nm::Complex128>,
+      nm::math::clapack_laswp<nm::RubyObject>
+  };
+
+  // Allocate the C version of the pivot index array
+  int* ipiv_;
+  if (TYPE(ipiv) != T_ARRAY) {
+    rb_raise(rb_eArgError, "ipiv must be of type Array");
+  } else {
+    ipiv_ = NM_ALLOCA_N(int, RARRAY_LEN(ipiv));
+    for (int index = 0; index < RARRAY_LEN(ipiv); ++index) {
+      ipiv_[index] = FIX2INT( RARRAY_PTR(ipiv)[index] );
+    }
+  }
+
+  // Call either our version of laswp or the LAPACK version.
+  ttable[NM_DTYPE(a)](FIX2INT(n), NM_STORAGE_DENSE(a)->elements, FIX2INT(lda), FIX2INT(k1), FIX2INT(k2), ipiv_, FIX2INT(incx));
+
+  // a is both returned and modified directly in the argument list.
+  return a;
+}
+
+
+}