nmatrix 0.0.1

data/ext/nmatrix/nmatrix.h

@@ -0,0 +1,912 @@
+/////////////////////////////////////////////////////////////////////
+// = 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 - 2012, Ruby Science Foundation
+// NMatrix is Copyright (c) 2012, 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
+//
+// == nmatrix.h
+//
+
+#ifndef NMATRIX_H
+#define NMATRIX_H
+
+#include "nmatrix_config.h"
+
+#include <cblas.h>
+
+#include <math.h>
+
+#include <stdlib.h>
+#include <stdio.h>
+#include <string.h>
+#include <ruby.h>
+#define RUBY_ZERO INT2FIX(0)
+
+#ifdef BENCHMARK // SOURCE: http://stackoverflow.com/questions/2349776/how-can-i-benchmark-a-c-program-easily
+# include <sys/time.h>
+# include <sys/resource.h>
+#endif
+
+#include "dtypes.h"
+
+#include <stddef.h>
+#ifdef HAVE_STDINT_H
+# include <stdint.h>
+#endif
+
+/*
+  Data types used in NArray / NMatrix :
+  Please modify these types if your system has any different type.
+*/
+
+
+/* NM_BYTE : unsigned 8-bit integer */
+#ifndef HAVE_U_INT8_T
+# ifdef HAVE_UINT8_T
+typedef uint8_t			u_int8_t;
+# else
+typedef unsigned char		u_int8_t;
+# endif
+#endif
+
+//#ifndef HAVE_INT8_T
+//typedef char                   int8_t;
+//#endif
+
+#ifndef HAVE_INT16_T
+# if SIZEOF_SHORT == 2
+typedef short                  int16_t;
+# else
+---->> Please define int16_t manually because sizeof(short) != 2. <<----
+# endif
+#endif /* HAVE_INT16_T */
+
+#ifndef HAVE_INT32_T
+# if SIZEOF_LONG == 4
+typedef long                   int32_t;
+# else
+#  if SIZEOF_INT == 4
+typedef int                    int32_t;
+#  else
+---->> Please define int32_t manually because sizeof(long) != 4. <<----
+#  endif
+# endif
+#endif /* HAVE_INT32_T */
+
+/* unsigned 32-bit integer */
+#ifndef HAVE_U_INT32_T
+# ifdef HAVE_UINT32_T
+typedef uint32_t			u_int32_t;
+# else
+#  if SIZEOF_LONG == 4
+typedef unsigned long                   u_int32_t;
+#  else
+#   if SIZEOF_INT == 4
+typedef unsigned int                    u_int32_t;
+#   else
+---->> Please define u_int32_t manually because sizeof(long) != 4. <<----
+#   endif
+#  endif
+# endif
+#endif /* HAVE_U_INT32_T */
+
+#ifndef HAVE_INT64_T
+# if SIZEOF_QUAD == 8
+typedef quad                   int64_t;
+# else
+#  if SIZEOF_LONG == 8
+typedef long                   int64_t;
+#  else
+---->> Please define int64_t manually because sizeof(quad) != 8. <<----
+#  endif
+# endif
+#endif /* HAVE_INT64_T */
+
+/* unsigned 64-bit integer */
+#ifndef HAVE_U_INT64_T
+# ifdef HAVE_UINT64_T
+typedef uint64_t            u_int64_t;
+# else
+#  if SIZEOF_QUAD == 8
+typedef unsigned quad       u_int64_t;
+#  else
+#   if SIZEOF_LONG == 8
+typedef unsigned long       u_int64_t;
+#   else
+---->> Please define u_int64_t manually because sizeof(quad) != 8. <<----
+#   endif
+#  endif
+# endif
+#endif /* HAVE_U_INT64_T */
+
+
+#ifndef HAVE_SIZE_T /// If you modify this, make sure to modify the definition of y_size_t and Y_SIZE_T!
+typedef u_int64_t    size_t;
+# define NM_SIZE_T   NM_INT64
+#else
+# if SIZEOF_SIZE_T == 8
+#  define NM_SIZE_T  NM_INT64
+# else
+#  if SIZEOF_SIZE_T == 4
+#   define NM_SIZE_T NM_INT32
+#  else
+---->> Please define size_t and y_size_t manually because sizeof(size_t) is neither 8 nor 4. <<----
+#  endif
+# endif
+#endif
+
+// for when we need to return array indices.
+// This must never be larger than size_t
+typedef uint32_t    y_size_t;
+#define Y_SIZE_T    NM_INT32
+
+
+#ifdef HAVE_STDBOOL_H
+# include <stdbool.h>
+#else
+typedef char    bool;
+# define true    1;
+# define false   0;
+#endif
+
+
+typedef struct { float r,i; } complex64;
+typedef struct { double r,i; } complex128;
+typedef struct { int16_t n,d; } rational32;
+typedef struct { int32_t n,d; } rational64;
+typedef struct { int64_t n,d; } rational128;
+
+
+#if SIZEOF_INT == 8
+# define DEFAULT_DTYPE  NM_INT64
+#else
+# if SIZEOF_INT == 4
+#  define DEFAULT_DTYPE NM_INT32
+# else
+#  define DEFAULT_DTYPE NM_INT16
+# endif
+#endif
+
+
+#define YALE_GROWTH_CONSTANT    1.5
+
+
+enum NMatrix_STypes {
+  S_DENSE,
+  S_LIST,
+  S_YALE,
+  S_TYPES
+};
+
+
+// Element-wise operations (see blas/elementwise.template.c)
+enum NMatrix_Ops {
+  NM_OP_ADD = '+',
+  NM_OP_SUB = '-',
+  NM_OP_MUL = '*',
+  NM_OP_DIV = '/',
+  NM_OP_MOD = '%',
+  NM_OP_BANG = '!',
+  NM_OP_NEG, // unary minus
+  NM_OP_EQEQ, // ==
+  NM_OP_NEQ, // !=
+  NM_OP_GT = '>', // >
+  NM_OP_LT = '<', // <
+  NM_OP_GTE = ',', // >=
+  NM_OP_LTE = '.', // <=
+  NM_OP_NOT = '~',
+  NM_OP_AND = '&',
+  NM_OP_OR = '|',
+  NM_OP_XOR = '^',
+  NM_OP_LSH, // <<
+  NM_OP_RSH // >>
+};
+
+
+/* Singly-linked ordered list
+ * - holds keys and values
+ * - no duplicate keys
+ * - keys are ordered
+ * - values may be lists themselves
+ */
+typedef struct l_node { /* Linked list node */
+  size_t key;
+  void*  val;
+  struct l_node * next; // next
+} NODE;
+
+typedef struct l_list {
+  NODE* first;
+} LIST;
+
+
+// two vectors and a capacity
+typedef struct y_vector {
+  void*  ija;
+  void*  a;
+  size_t capacity;
+} VECTOR;
+
+
+typedef struct common_s { // Common elements found in all _s types.
+  int8_t    dtype;
+  size_t    rank;
+  size_t*   shape;
+} STORAGE;
+
+
+typedef struct list_s {
+  int8_t    dtype;
+  size_t    rank;
+  size_t*   shape;
+  void*     default_val;
+  LIST*     rows;
+} LIST_STORAGE;
+
+
+typedef struct dense_s {
+  int8_t    dtype;
+  size_t    rank;
+  size_t*   shape;
+  void*     elements;
+} DENSE_STORAGE;
+
+
+typedef struct yale_s {
+  int8_t    dtype;
+  size_t    rank;
+  size_t*   shape;
+  size_t    ndnz; // strictly non-diagonal non-zero count!
+  size_t    capacity;
+  int8_t    index_dtype;
+  void*     ija;
+  void*     a;
+} YALE_STORAGE;
+
+
+typedef struct numeric_matrix {
+  int8_t   stype;             /* method of storage (csc, dense, etc) */
+  STORAGE* storage;           /* pointer to storage struct */
+} NMATRIX;
+
+
+/* Local */
+
+typedef union {
+  u_int8_t b[2];
+  int16_t s;
+} nm_size16_t;
+
+typedef union {
+  u_int8_t b[4];
+  int32_t  i;
+  float    f;
+} nm_size32_t;
+
+typedef union {
+  u_int8_t  b[8];
+  int64_t   q;
+  float     f[2];
+  double    d;
+  complex64 c;
+} nm_size64_t;
+
+typedef union {
+  u_int8_t   b[16];
+  int64_t    i[2];
+  double     d[2];
+  float      f[4];
+  complex64  c[2];
+  complex128 z;
+  rational32 r[4];
+  rational64 ra[2];
+  rational128 rat;
+  VALUE      v[2];
+} nm_size128_t;
+
+
+// For calling cblas_gemm functions (see cblas.c)
+typedef struct cblas_param_t {
+  int M, N, K, lda, ldb, ldc;
+  void *A, *B, *C;
+  nm_size128_t alpha, beta;
+} DENSE_PARAM;
+
+
+// Formerly in smmp.h:
+typedef struct smmp_param_t {
+  void *ia, *ja, *a;
+  bool diag;
+} YALE_PARAM;
+
+// Shouldn't be necessary, as they're defined in nmatrix.h:
+// (Oddly, though, these fix the error.)
+/*typedef uint8_t   u_int8_t;
+typedef uint16_t  u_int16_t;
+typedef uint32_t  u_int32_t;
+typedef uint64_t  u_int64_t; */
+
+
+// rational.c
+int64_t nmrb_gcd(int64_t x, int64_t y);
+
+// BLAS functions
+#define SMMP_MAX_THREE(a,b,c) ((a)>(b) ? ( (a)>(c) ? (a) : (c) ) : ( (b)>(c) ? (b) : (c) ))
+#define SMMP_MIN(a,b) ((a)>(b) ? (b) : (a))
+#define SMMP_MAX(a,b) ((a)>(b) ? (a) : (b))
+
+void transp(y_size_t n, y_size_t m, void* ia, void* ja, bool diaga, void* a, void* ib, void* jb, void* b, bool move, int8_t itype, int8_t dtype);
+
+void i8_symbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_symbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_symbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_symbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+
+void i8_b_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_b_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_b_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_b_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i8_i8_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_i8_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_i8_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_i8_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i8_i16_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_i16_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_i16_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_i16_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i8_i32_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_i32_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_i32_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_i32_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i8_i64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_i64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_i64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_i64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i8_f32_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_f32_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_f32_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_f32_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i8_f64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_f64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_f64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_f64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i8_c64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_c64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_c64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_c64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i8_c128_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_c128_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_c128_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_c128_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i8_r32_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_r32_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_r32_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_r32_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i8_r64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_r64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_r64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_r64_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i8_r128_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_r128_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_r128_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_r128_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i8_v_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i16_v_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i32_v_smmp_sort_columns_(y_size_t, YALE_PARAM);
+void i64_v_smmp_sort_columns_(y_size_t, YALE_PARAM);
+
+void i8_b_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_b_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_b_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_b_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_i8_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_i8_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_i8_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_i8_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_i16_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_i16_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_i16_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_i16_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_i32_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_i32_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_i32_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_i32_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_i64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_i64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_i64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_i64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_f32_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_f32_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_f32_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_f32_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_f64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_f64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_f64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_f64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_c64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_c64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_c64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_c64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_c128_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_c128_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_c128_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_c128_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_r32_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_r32_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_r32_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_r32_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_r64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_r64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_r64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_r64_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_r128_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_r128_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_r128_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_r128_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_v_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_v_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_v_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_v_smmp(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+
+void i8_b_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_b_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_b_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_b_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_i8_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_i8_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_i8_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_i8_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_i16_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_i16_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_i16_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_i16_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_i32_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_i32_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_i32_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_i32_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_i64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_i64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_i64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_i64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_f32_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_f32_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_f32_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_f32_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_f64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_f64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_f64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_f64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_c64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_c64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_c64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_c64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_c128_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_c128_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_c128_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_c128_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_r32_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_r32_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_r32_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_r32_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_r64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_r64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_r64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_r64_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_r128_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_r128_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_r128_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_r128_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i8_v_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i16_v_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i32_v_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+void i64_v_numbmm_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, YALE_PARAM);
+
+void i8_b_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_b_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_b_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_b_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i8_i8_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_i8_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_i8_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_i8_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i8_i16_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_i16_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_i16_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_i16_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i8_i32_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_i32_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_i32_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_i32_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i8_i64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_i64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_i64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_i64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i8_f32_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_f32_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_f32_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_f32_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i8_f64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_f64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_f64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_f64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i8_c64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_c64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_c64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_c64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i8_c128_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_c128_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_c128_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_c128_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i8_r32_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_r32_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_r32_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_r32_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i8_r64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_r64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_r64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_r64_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i8_r128_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_r128_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_r128_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_r128_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i8_v_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i16_v_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i32_v_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+void i64_v_transp_(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool);
+
+
+
+// For binary operations involving matrices that need to be casted.
+typedef struct storage_pair_t {
+  STORAGE* left;
+  STORAGE* right;
+} STORAGE_PAIR;
+
+
+#ifndef NMATRIX_C
+extern VALUE cNMatrix;
+
+extern const int nm_sizeof[NM_TYPES+1];
+#endif
+
+
+#define NM_MAX_RANK 15
+
+#define UnwrapNMatrix(obj,var)  Data_Get_Struct(obj, struct numeric_matrix, var)
+#define IsNMatrix(obj)          (rb_obj_is_kind_of(obj, CNMatrix)==Qtrue)
+
+#define NM_STORAGE(val)         (((struct numeric_matrix*)DATA_PTR(val))->storage)
+//#define NM_PTR(a, p)            ((a)->ptr+(p)*nm_sizeof[(a)->type])
+#define NM_STRUCT(val)          ((struct numeric_matrix*)DATA_PTR(val))
+//#define NM_PTR_TYPE(val,type)   (type)(((struct numeric_matrix*)DATA_PTR(val))->ptr)
+#define NM_RANK(val)            (((STORAGE*)(NM_STORAGE(val)))->rank)
+#define NM_DTYPE(val)           (((STORAGE*)(NM_STORAGE(val)))->dtype)
+#define NM_STYPE(val)           (((struct numeric_matrix*)DATA_PTR(val))->stype)
+#define NM_SHAPE(val,i)         (((STORAGE*)(NM_STORAGE(val)))->shape[(i)])
+#define NM_SHAPE0(val)          (((struct numeric_matrix*)DATA_PTR(val))->shape[0])
+#define NM_SHAPE1(val)          (((struct numeric_matrix*)DATA_PTR(val))->shape[1])
+#define NM_SIZEOF_DTYPE(val)    (nm_sizeof[NM_DTYPE(val)])
+#define NM_REF(val,coords)      (RefFuncs[NM_STYPE(val)]( NM_STORAGE(val), coords, NM_SIZEOF_DTYPE(val) ))
+
+#define NM_IsNMatrix(obj) (rb_obj_is_kind_of(obj, cNMatrix)==Qtrue)
+#define NM_IsArray(obj)   (TYPE(obj)==T_ARRAY || rb_obj_is_kind_of(obj,cNMatrix)==Qtrue)
+#define NM_IsROBJ(d) ((d)->dtype==NM_ROBJ)
+#define NM_IsINTEGER(a) \
+    (NM_DTYPE(a)==NM_BYTE || NM_DTYPE(a)==NM_INT8 || NM_DTYPE(a)==NM_INT16 || NM_DTYPE(a)==NM_INT32 || NM_DTYPE(a)==NM_INT64)
+#define NM_IsCOMPLEX(a) \
+    (NM_DTYPE(a)==NM_COMPLEX32 || NM_DTYPE(a)==NM_COMPLEX64)
+#define NM_MAX(a,b) (((a)>(b))?(a):(b))
+#define NM_SWAP(a,b,tmp) {(tmp)=(a);(a)=(b);(b)=(tmp);}
+
+//#define NUM2REAL(v) NUM2DBL( rb_funcall((v),nm_id_real,0) ) // deprecated
+#define REAL2DBL(v) NUM2DBL( rb_funcall((v),nm_id_real,0) )
+//#define NUM2IMAG(v) NUM2DBL( rb_funcall((v),nm_id_imag,0) ) // deprecated
+#define IMAG2DBL(v) NUM2DBL( rb_funcall((v),nm_id_imag,0) )
+
+#define NUM2NUMER(v) NUM2INT( rb_funcall((v), nm_id_numer,0) ) // deprecated
+#define NUMER2INT(v) NUM2INT( rb_funcall((v), nm_id_numer,0) )
+#define NUM2DENOM(v) NUM2INT( rb_funcall((v), nm_id_denom,0) ) // deprecated
+#define DENOM2INT(v) NUM2INT( rb_funcall((v), nm_id_denom,0) )
+
+#define IS_NUMERIC(v)   (FIXNUM_P(v) || TYPE(v) == T_FLOAT || TYPE(v) == T_COMPLEX || TYPE(v) == T_RATIONAL)
+#define IS_STRING(v)    (TYPE(v) == T_STRING)
+
+#define CheckNMatrixType(v)   if (TYPE(v) != T_DATA || RDATA(v)->dfree != (RUBY_DATA_FUNC)nm_delete) rb_raise(rb_eTypeError, "expected NMatrix on left-hand side of operation");
+
+//#define YALE_JA_START(sptr)             (((YALE_STORAGE*)(sptr))->shape[0]+1)
+#define YALE_IJA(sptr,elem_size,i)          (void*)( (char*)(((YALE_STORAGE*)(sptr))->ija) + i * elem_size )
+//#define YALE_JA(sptr,dtype,j)           ((((dtype)*)((YALE_STORAGE*)(sptr))->ija)[(YALE_JA_START(sptr))+j])
+#define YALE_ROW_LENGTH(sptr,elem_size,i)   (*(size_t*)YALE_IA((sptr),(elem_size),(i)+1) - *(size_t*)YALE_IJA((sptr),(elem_size),(i)))
+#define YALE_A(sptr,elem_size,i)            (void*)((char*)(((YALE_STORAGE*)(sptr))->a) + elem_size * i)
+#define YALE_DIAG(sptr, elem_size, i)       ( YALE_A((sptr),(elem_size),(i)) )
+//#define YALE_LU(sptr,dtype,i,j)             (((dtype)*)(((YALE_STORAGE*)(sptr))->a)[ YALE_JA_START(sptr) +  ])
+#define YALE_MINIMUM(sptr)                  (((YALE_STORAGE*)(sptr))->shape[0]*2 + 1) // arbitrarily defined
+#define YALE_SIZE_PTR(sptr,elem_size)       (void*)((char*)((YALE_STORAGE*)(sptr))->ija + ((YALE_STORAGE*)(sptr))->shape[0]*elem_size )
+#define YALE_MAX_SIZE(sptr)                 (((YALE_STORAGE*)(sptr))->shape[0] * ((YALE_STORAGE*)(sptr))->shape[1] + 1)
+#define YALE_IA_SIZE(sptr)                  ((YALE_STORAGE*)(sptr))->shape[0]
+
+// None of these next three return anything. They set a reference directly.
+#define YaleGetIJA(victim,s,i)              (SetFuncs[Y_SIZE_T][(s)->index_dtype](1, &(victim), 0, YALE_IJA((s), nm_sizeof[s->index_dtype], (i)), 0))
+#define YaleSetIJA(i,s,from)                (SetFuncs[s->index_dtype][Y_SIZE_T](1, YALE_IJA((s), nm_sizeof[s->index_dtype], (i)), 0, &(from), 0))
+#define YaleGetSize(sz,s)                   (SetFuncs[Y_SIZE_T][(s)->index_dtype](1, &sz, 0, (YALE_SIZE_PTR((s), nm_sizeof[(s)->index_dtype])), 0))
+//#define YALE_FIRST_NZ_ROW_ENTRY(sptr,elem_size,i)
+
+
+#if !defined RSTRING_LEN
+#define RSTRING_LEN(a) RSTRING(a)->len
+#endif
+#if !defined RSTRING_PTR
+#define RSTRING_PTR(a) RSTRING(a)->ptr
+#endif
+#if !defined RARRAY_LEN
+#define RARRAY_LEN(a) RARRAY(a)->len
+#endif
+#if !defined RARRAY_PTR
+#define RARRAY_PTR(a) RARRAY(a)->ptr
+#endif
+
+#define NM_INDEX_TYPES  NM_FLOAT32
+
+
+typedef void     (*nm_setfunc_t[NM_TYPES][NM_TYPES])(); // copy functions
+typedef void     (*nm_incfunc_t[NM_TYPES])();           // increment functions
+typedef void*    (*nm_stype_ref_t[S_TYPES])(STORAGE*, size_t*);        // get/ref
+typedef VALUE    (*nm_stype_ins_t[S_TYPES])(STORAGE*, size_t*, VALUE); // insert
+typedef STORAGE* (*nm_create_storage_t[S_TYPES])();
+typedef STORAGE* (*nm_cast_copy_storage_t[S_TYPES])();
+typedef STORAGE* (*nm_scast_copy_storage_t[S_TYPES][S_TYPES])();
+typedef NMATRIX* (*nm_matrix_multiply_op_t[S_TYPES])();
+typedef NMATRIX* (*nm_elementwise_binary_op_casted_t[S_TYPES])();
+typedef int      (*nm_d_elementwise_binary_op_t[NM_TYPES])();
+typedef int      (*nm_y_elementwise_binary_op_t[NM_TYPES][NM_INDEX_TYPES])();
+typedef bool     (*nm_compare_t[S_TYPES])();
+typedef void     (*nm_delete_t[S_TYPES])();
+typedef void     (*nm_mark_t[S_TYPES])(void*);
+typedef void     (*nm_gemm_t[NM_TYPES])();           // general matrix/matrix multiply
+typedef void     (*nm_dense_transpose_t[NM_TYPES])(); // dense transpose
+typedef void     (*nm_gemv_t[NM_TYPES])();           // general matrix/vector multiply
+typedef void     (*nm_smmp_t[NM_TYPES][NM_INDEX_TYPES])(); // sparse (yale) multiply
+typedef void     (*nm_smmp_transpose_t[NM_TYPES][NM_INDEX_TYPES])(y_size_t, y_size_t, YALE_PARAM, YALE_PARAM, bool); // sparse (yale) transpose
+//typedef void (*nm_setsf_t[S_TYPES][S_TYPES])();
+//typedef void (*nm_setdf_t[NM_DTYPES][NM_DTYPES])();
+
+extern nm_setfunc_t SetFuncs;
+extern nm_incfunc_t Increment;
+extern ID nm_id_real, nm_id_imag;
+extern ID nm_id_denom, nm_id_numer;
+extern ID nm_id_mult, nm_id_multeq, nm_id_add;
+
+/* blas.c */
+int r32gemm(enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, const int M, const int N, const int K, const rational32 alpha, const rational32* A, const int lda, const rational32* B, const int ldb, const rational32 beta, rational32* C, const int ldc);
+int r32gemv(enum CBLAS_TRANSPOSE Trans, const size_t M, const size_t N, const rational32 alpha, const rational32* A, const size_t lda, const rational32* X, const int incX, const rational32 beta, rational32* Y, const int incY);
+int r64gemm(enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, const int M, const int N, const int K, const rational64 alpha, const rational64* A, const int lda, const rational64* B, const int ldb, const rational64 beta, rational64* C, const int ldc);
+int r64gemv(enum CBLAS_TRANSPOSE Trans, const size_t M, const size_t N, const rational64 alpha, const rational64* A, const size_t lda, const rational64* X, const int incX, const rational64 beta, rational64* Y, const int incY);
+int r128gemm(enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, const int M, const int N, const int K, const rational128 alpha, const rational128* A, const int lda, const rational128* B, const int ldb, const rational128 beta, rational128* C, const int ldc);
+int r128gemv(enum CBLAS_TRANSPOSE Trans, const size_t M, const size_t N, const rational128 alpha, const rational128* A, const size_t lda, const rational128* X, const int incX, const rational128 beta, rational128* Y, const int incY);
+int bgemm(enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, const int M, const int N, const int K, const u_int8_t alpha, const u_int8_t* A, const int lda, const u_int8_t* B, const int ldb, const u_int8_t beta, u_int8_t* C, const int ldc);
+int bgemv(enum CBLAS_TRANSPOSE Trans, const size_t M, const size_t N, const u_int8_t alpha, const u_int8_t* A, const size_t lda, const u_int8_t* X, const int incX, const u_int8_t beta, u_int8_t* Y, const int incY);
+int i8gemm(enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, const int M, const int N, const int K, const int8_t alpha, const int8_t* A, const int lda, const int8_t* B, const int ldb, const int8_t beta, int8_t* C, const int ldc);
+int i8gemv(enum CBLAS_TRANSPOSE Trans, const size_t M, const size_t N, const int8_t alpha, const int8_t* A, const size_t lda, const int8_t* X, const int incX, const int8_t beta, int8_t* Y, const int incY);
+int i16gemm(enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, const int M, const int N, const int K, const int16_t alpha, const int16_t* A, const int lda, const int16_t* B, const int ldb, const int16_t beta, int16_t* C, const int ldc);
+int i16gemv(enum CBLAS_TRANSPOSE Trans, const size_t M, const size_t N, const int16_t alpha, const int16_t* A, const size_t lda, const int16_t* X, const int incX, const int16_t beta, int16_t* Y, const int incY);
+int i32gemm(enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, const int M, const int N, const int K, const int32_t alpha, const int32_t* A, const int lda, const int32_t* B, const int ldb, const int32_t beta, int32_t* C, const int ldc);
+int i32gemv(enum CBLAS_TRANSPOSE Trans, const size_t M, const size_t N, const int32_t alpha, const int32_t* A, const size_t lda, const int32_t* X, const int incX, const int32_t beta, int32_t* Y, const int incY);
+int i64gemm(enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, const int M, const int N, const int K, const int64_t alpha, const int64_t* A, const int lda, const int64_t* B, const int ldb, const int64_t beta, int64_t* C, const int ldc);
+int i64gemv(enum CBLAS_TRANSPOSE Trans, const size_t M, const size_t N, const int64_t alpha, const int64_t* A, const size_t lda, const int64_t* X, const int incX, const int64_t beta, int64_t* Y, const int incY);
+int vgemm(enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, const int M, const int N, const int K, const VALUE alpha, const VALUE* A, const int lda, const VALUE* B, const int ldb, const VALUE beta, VALUE* C, const int ldc);
+
+// void %%TYPE_ABBREV%%transp(unsigned int M, unsigned int N, const %%TYPE%%* A, int lda, %%TYPE%%* B, int ldb);
+void btransp(const unsigned int, const unsigned int, const u_int8_t*, const int, u_int8_t*, const int);
+void i8transp(const unsigned int, const unsigned int, const int8_t*, const int, int8_t*, const int);
+void i16transp(const unsigned int, const unsigned int, const int16_t*, const int, int16_t*, const int);
+void i32transp(const unsigned int, const unsigned int, const int32_t*, const int, int32_t*, const int);
+void i64transp(const unsigned int, const unsigned int, const int64_t*, const int, int64_t*, const int);
+void f32transp(const unsigned int, const unsigned int, const float*, const int, float*, const int);
+void f64transp(const unsigned int, const unsigned int, const double*, const int, double*, const int);
+void c64transp(const unsigned int, const unsigned int, const complex64*, const int, complex64*, const int);
+void c128transp(const unsigned int, const unsigned int, const complex128*, const int, complex128*, const int);
+void r32transp(const unsigned int, const unsigned int, const rational32*, const int, rational32*, const int);
+void r64transp(const unsigned int, const unsigned int, const rational64*, const int, rational64*, const int);
+void r128transp(const unsigned int, const unsigned int, const rational128*, const int, rational128*, const int);
+void vtransp(const unsigned int, const unsigned int, const VALUE*, const int, VALUE*, const int);
+
+int nm_d_b_elementwise(const u_int8_t* A, const u_int8_t* B, u_int8_t* C, size_t n, enum NMatrix_Ops op);
+int nm_d_i8_elementwise(const int8_t* A, const int8_t* B, int8_t* C, size_t n, enum NMatrix_Ops op);
+int nm_d_i16_elementwise(const int16_t* A, const int16_t* B, int16_t* C, size_t n, enum NMatrix_Ops op);
+int nm_d_i32_elementwise(const int32_t* A, const int32_t* B, int32_t* C, size_t n, enum NMatrix_Ops op);
+int nm_d_i64_elementwise(const int64_t* A, const int64_t* B, int64_t* C, size_t n, enum NMatrix_Ops op);
+int nm_d_f32_elementwise(const float* A, const float* B, float* C, size_t n, enum NMatrix_Ops op);
+int nm_d_f64_elementwise(const double* A, const double* B, double* C, size_t n, enum NMatrix_Ops op);
+int nm_d_c64_elementwise(const complex64* A, const complex64* B, complex64* C, size_t n, enum NMatrix_Ops op);
+int nm_d_c128_elementwise(const complex128* A, const complex128* B, complex128* C, size_t n, enum NMatrix_Ops op);
+int nm_d_r32_elementwise(const rational32* A, const rational32* B, rational32* C, size_t n, enum NMatrix_Ops op);
+int nm_d_r64_elementwise(const rational64* A, const rational64* B, rational64* C, size_t n, enum NMatrix_Ops op);
+int nm_d_r128_elementwise(const rational128* A, const rational128* B, rational128* C, size_t n, enum NMatrix_Ops op);
+int nm_d_v_elementwise(const VALUE* A, const VALUE* B, VALUE* C, size_t n, enum NMatrix_Ops op);
+
+// These are in blas.c but are needed by smmp2.c (the smmp template stuff)
+rational128 r128_muldiv(int64_t anum, int64_t aden, int64_t bnum, int64_t bden, char k);
+rational128 r128_addsub(int64_t anum, int64_t aden, int64_t bnum, int64_t bden, char k);
+rational128 r128_mod(int64_t anum, int64_t aden, int64_t bnum, int64_t bden);
+rational128 r128_bang(int64_t, int64_t);
+rational128 r128_negate(int64_t, int64_t);
+rational64 r64_muldiv(int64_t anum, int64_t aden, int64_t bnum, int64_t bden, char k);
+rational64 r64_addsub(int64_t anum, int64_t aden, int64_t bnum, int64_t bden, char k);
+rational64 r64_mod(int32_t anum, int32_t aden, int32_t bnum, int32_t bden);
+rational64 r64_bang(int32_t, int32_t);
+rational64 r64_negate(int32_t, int32_t);
+rational32 r32_muldiv(int64_t anum, int64_t aden, int64_t bnum, int64_t bden, char k);
+rational32 r32_addsub(int64_t anum, int64_t aden, int64_t bnum, int64_t bden, char k);
+rational32 r32_mod(int16_t anum, int16_t aden, int16_t bnum, int16_t bden);
+rational32 r32_bang(int16_t, int16_t);
+rational32 r32_negate(int16_t, int16_t);
+
+rational32 BOOL2R32(bool);
+rational64 BOOL2R64(bool);
+rational128 BOOL2R128(bool);
+
+
+/* cblas.c */
+DENSE_PARAM init_cblas_params_for_nm_multiply_matrix(int8_t dtype);
+void cblas_bgemm_(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+void cblas_bgemv_(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, DENSE_PARAM p);
+void cblas_i8gemm_(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+void cblas_i8gemv_(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, DENSE_PARAM p);
+void cblas_i16gemm_(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+void cblas_i16gemv_(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, DENSE_PARAM p);
+void cblas_i32gemm_(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+void cblas_i32gemv_(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, DENSE_PARAM p);
+void cblas_i64gemm_(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+void cblas_i64gemv_(enum CBLAS_ORDER Order, enum CBLAS_TRANSPOSE TransA, DENSE_PARAM p);
+void cblas_sgemm_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+void cblas_sgemv_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, DENSE_PARAM p);
+void cblas_dgemm_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+void cblas_dgemv_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, DENSE_PARAM p);
+void cblas_cgemm_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+void cblas_cgemv_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, DENSE_PARAM p);
+void cblas_zgemm_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+void cblas_zgemv_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, DENSE_PARAM p);
+void cblas_r32gemm_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+void cblas_r32gemv_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, DENSE_PARAM p);
+void cblas_r64gemm_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+void cblas_r64gemv_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, DENSE_PARAM p);
+void cblas_r128gemm_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+void cblas_r128gemv_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, DENSE_PARAM p);
+void cblas_vgemm_(const enum CBLAS_ORDER Order, const enum CBLAS_TRANSPOSE TransA, const enum CBLAS_TRANSPOSE TransB, DENSE_PARAM p);
+
+
+/* smmp2.c */
+int i8_b_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, u_int8_t* a, u_int8_t* b, u_int8_t* c);
+int i16_b_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, u_int8_t* a, u_int8_t* b, u_int8_t* c);
+int i32_b_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, u_int8_t* a, u_int8_t* b, u_int8_t* c);
+int i64_b_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, u_int8_t* a, u_int8_t* b, u_int8_t* c);
+int i8_i8_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, int8_t* a, int8_t* b, int8_t* c);
+int i16_i8_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, int8_t* a, int8_t* b, int8_t* c);
+int i32_i8_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, int8_t* a, int8_t* b, int8_t* c);
+int i64_i8_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, int8_t* a, int8_t* b, int8_t* c);
+int i8_i16_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, int16_t* a, int16_t* b, int16_t* c);
+int i16_i16_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, int16_t* a, int16_t* b, int16_t* c);
+int i32_i16_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, int16_t* a, int16_t* b, int16_t* c);
+int i64_i16_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, int16_t* a, int16_t* b, int16_t* c);
+int i8_i32_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, int32_t* a, int32_t* b, int32_t* c);
+int i16_i32_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, int32_t* a, int32_t* b, int32_t* c);
+int i32_i32_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, int32_t* a, int32_t* b, int32_t* c);
+int i64_i32_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, int32_t* a, int32_t* b, int32_t* c);
+int i8_i64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, int64_t* a, int64_t* b, int64_t* c);
+int i16_i64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, int64_t* a, int64_t* b, int64_t* c);
+int i32_i64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, int64_t* a, int64_t* b, int64_t* c);
+int i64_i64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, int64_t* a, int64_t* b, int64_t* c);
+int i8_f32_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, float* a, float* b, float* c);
+int i16_f32_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, float* a, float* b, float* c);
+int i32_f32_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, float* a, float* b, float* c);
+int i64_f32_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, float* a, float* b, float* c);
+int i8_f64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, double* a, double* b, double* c);
+int i16_f64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, double* a, double* b, double* c);
+int i32_f64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, double* a, double* b, double* c);
+int i64_f64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, double* a, double* b, double* c);
+int i8_c64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, complex64* a, complex64* b, complex64* c);
+int i16_c64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, complex64* a, complex64* b, complex64* c);
+int i32_c64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, complex64* a, complex64* b, complex64* c);
+int i64_c64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, complex64* a, complex64* b, complex64* c);
+int i8_c128_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, complex128* a, complex128* b, complex128* c);
+int i16_c128_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, complex128* a, complex128* b, complex128* c);
+int i32_c128_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, complex128* a, complex128* b, complex128* c);
+int i64_c128_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, complex128* a, complex128* b, complex128* c);
+int i8_r32_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, rational32* a, rational32* b, rational32* c);
+int i16_r32_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, rational32* a, rational32* b, rational32* c);
+int i32_r32_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, rational32* a, rational32* b, rational32* c);
+int i64_r32_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, rational32* a, rational32* b, rational32* c);
+int i8_r64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, rational64* a, rational64* b, rational64* c);
+int i16_r64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, rational64* a, rational64* b, rational64* c);
+int i32_r64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, rational64* a, rational64* b, rational64* c);
+int i64_r64_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, rational64* a, rational64* b, rational64* c);
+int i8_r128_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, rational128* a, rational128* b, rational128* c);
+int i16_r128_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, rational128* a, rational128* b, rational128* c);
+int i32_r128_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, rational128* a, rational128* b, rational128* c);
+int i64_r128_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, rational128* a, rational128* b, rational128* c);
+int i8_v_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int8_t* ija, const u_int8_t* ijb, const u_int8_t* ijc, VALUE* a, VALUE* b, VALUE* c);
+int i16_v_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int16_t* ija, const u_int16_t* ijb, const u_int16_t* ijc, VALUE* a, VALUE* b, VALUE* c);
+int i32_v_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int32_t* ija, const u_int32_t* ijb, const u_int32_t* ijc, VALUE* a, VALUE* b, VALUE* c);
+int i64_v_ew(y_size_t n, y_size_t m, enum NMatrix_Ops op, const u_int64_t* ija, const u_int64_t* ijb, const u_int64_t* ijc, VALUE* a, VALUE* b, VALUE* c);
+
+
+/* dense.c */
+DENSE_STORAGE*  create_dense_storage(int8_t dtype, size_t* shape, size_t rank, void* elements, size_t elements_length);
+void            delete_dense_storage(DENSE_STORAGE* s);
+void            mark_dense_storage(void* s);
+DENSE_STORAGE*  cast_copy_dense_storage(DENSE_STORAGE* rhs, int8_t new_dtype);
+
+size_t          count_dense_storage_elements(const DENSE_STORAGE* s);
+bool            dense_storage_eqeq(const DENSE_STORAGE*, const DENSE_STORAGE*);
+
+size_t          dense_storage_pos(DENSE_STORAGE* s, size_t* coords);
+void*           dense_storage_get(DENSE_STORAGE* s, size_t* coords);
+void            dense_storage_set(DENSE_STORAGE* s, size_t* coords, void* val);
+
+/* list.c */
+LIST_STORAGE*   create_list_storage(int8_t dtype, size_t* shape, size_t rank, void* init_val);
+void            delete_list_storage(LIST_STORAGE* s);
+void            mark_list_storage(void* s);
+LIST_STORAGE*   cast_copy_list_storage(LIST_STORAGE* rhs, int8_t new_dtype);
+size_t          count_storage_max_elements(const STORAGE*);
+
+void*           list_storage_get(LIST_STORAGE* s, size_t* coords);
+void*           list_storage_insert(LIST_STORAGE* s, size_t* coords, void* val);
+void*           list_storage_remove(LIST_STORAGE* s, size_t* coords);
+bool            list_storage_eqeq(const LIST_STORAGE*, const LIST_STORAGE*);
+
+/* yale.c */
+void print_vectors(YALE_STORAGE* s);
+YALE_STORAGE*   create_yale_storage(int8_t dtype, size_t* shape, size_t rank, size_t init_capacity);
+void            init_yale_storage(YALE_STORAGE* s);
+void            delete_yale_storage(YALE_STORAGE* s);
+void            mark_yale_storage(void* s);
+YALE_STORAGE*   cast_copy_yale_storage(YALE_STORAGE* rhs, int8_t new_dtype);
+bool            yale_storage_eqeq(const YALE_STORAGE*, const YALE_STORAGE*);
+
+void*           yale_storage_ref(YALE_STORAGE* s, size_t* coords);
+char            yale_storage_set(YALE_STORAGE* s, size_t* coords, void* v);
+
+YALE_STORAGE*   create_merged_yale_storage(const YALE_STORAGE*, const YALE_STORAGE*);
+
+size_t          count_list_storage_nd_elements(const LIST_STORAGE*);
+size_t          count_list_storage_elements(const LIST_STORAGE*);
+
+
+/* stype casts */
+DENSE_STORAGE* scast_copy_dense_yale(const YALE_STORAGE* rhs, int8_t l_dtype);
+DENSE_STORAGE* scast_copy_dense_list(const LIST_STORAGE* rhs, int8_t l_dtype);
+YALE_STORAGE* scast_copy_yale_dense(const DENSE_STORAGE* rhs, int8_t l_dtype);
+YALE_STORAGE* scast_copy_yale_list(const LIST_STORAGE* rhs, int8_t l_dtype);
+LIST_STORAGE* scast_copy_list_yale(const YALE_STORAGE* rhs, int8_t l_dtype);
+LIST_STORAGE* scast_copy_list_dense(const DENSE_STORAGE* rhs, int8_t l_dtype);
+
+/* nmatrix.c */
+void cast_copy_value_single(void* to, const void* from, int8_t l_dtype, int8_t r_dtype);
+int8_t nm_dtypestring_to_dtype(VALUE str);
+int8_t nm_dtypesymbol_to_dtype(VALUE sym);
+int8_t nm_stypestring_to_stype(VALUE str);
+int8_t nm_stypesymbol_to_stype(VALUE sym);
+int8_t nm_guess_dtype(VALUE v);
+size_t* nm_interpret_shape_arg(VALUE arg, size_t* rank);
+NMATRIX* nm_create(int8_t stype, void* storage);
+void Init_nmatrix();
+
+#endif