narray-bigmem 0.0.0

data/ext/narray/na_linalg.c

@@ -0,0 +1,635 @@
+/*
+ * na_linalg.c
+ * Numerical Array Extention for Ruby
+ *   (C) Copyright 2000-2008 by Masahiro TANAKA
+ */
+#include <ruby.h>
+#include "narray.h"
+#include "narray_local.h"
+#define ARRAY_BUF
+
+/*
+  a_ij == a[j,i]
+     j  - >
+   i 11 21 31
+   | 12 22 32
+   v 13 23 33
+*/
+
+#define SWAPMEM(a,b,tmp,sz) \
+{ memcpy(tmp,a,sz); memcpy(a,b,sz); memcpy(b,tmp,sz); }
+
+typedef struct NARRAY_FUNCSET {
+  int   elmsz;
+  char *zero;
+  char *one;
+  char *tiny;
+  void (*set)();
+  void (*neg)();
+  void (*rcp)();
+  void (*abs)();
+  void (*add)();
+  void (*sbt)();
+  void (*mul)();
+  void (*div)();
+  void (*mod)();
+  void (*muladd)();
+  void (*mulsbt)();
+  void (*cmp)();
+  int  (*sort)();
+  void (*min)();
+  void (*max)();
+} na_funcset_t;
+
+VALUE cNMatrix, cNVector, cNMatrixLU;
+static na_funcset_t na_funcset[NA_NTYPES];
+static ID id_lu, id_pivot;
+
+
+static void
+na_loop_linalg( int nd, char *p1, char *p2, char *p3,
+		struct slice *s1, struct slice *s2, struct slice *s3,
+		void (*func)(), na_shape_t *shape, int type )
+{
+  int i;
+  int ps1 = s1[0].pstep;
+  int ps2 = s2[0].pstep;
+  int ps3 = s3[0].pstep;
+  int *si;
+
+  if (nd==0) {
+    (*func)(1, p1, 0, p2, 0, p3, 0, shape, type);
+    return;
+  }
+
+  si = ALLOCA_N(int,nd);
+  i  = nd;
+  s1[i].p = p1;
+  s2[i].p = p2;
+  s3[i].p = p3;
+
+  for(;;) {
+    /* set pointers */
+    while (i > 0) {
+      --i;
+      s3[i].p = s3[i].pbeg + s3[i+1].p;
+      s2[i].p = s2[i].pbeg + s2[i+1].p;
+      s1[i].p = s1[i].pbeg + s1[i+1].p;
+      si[i] = s1[i].n;
+    }
+    /* rank 0 loop */
+    (*func)(s2[0].n, s1[0].p, ps1, s2[0].p, ps2, s3[0].p, ps3, shape, type);
+    /* rank up */
+    do {
+      if ( ++i >= nd ) return;
+    } while ( --si[i] == 0 );
+    /* next point */
+    s1[i].p += s1[i].pstep;
+    s2[i].p += s2[i].pstep;
+    s3[i].p += s3[i].pstep;
+  }
+}
+
+static na_shape_t
+na_shape_total( int n, na_shape_t *shape )
+{
+  na_shape_t total=1;
+
+  for (; n>0; --n)
+    total *= *(shape++);
+
+  return total;
+}
+
+static void
+na_exec_linalg( struct NARRAY *a1, struct NARRAY *a2, struct NARRAY *a3,
+		int ncd1, int ncd2, int ncd3, void (*func)() )
+{
+  int   ndim, ncd, nsz1, nsz2, nsz3;
+  na_shape_t  *itr, *shp1, *shp2, *shp3;
+  struct slice *s1, *s2, *s3;
+
+  ncd  = NA_MAX3(ncd1,ncd2,ncd3); /* class dim */
+  ndim = NA_MAX3(a1->rank-ncd1, a2->rank-ncd2, a3->rank-ncd3) + ncd;
+
+  NA_ALLOC_SLICE(s1,(ndim+1)*3,shp1,ndim*4);
+  shp2 = &shp1[ndim];
+  shp3 = &shp2[ndim];
+  itr  = &shp3[ndim];
+  s2   = &s1[ndim+1];
+  s3   = &s2[ndim+1];
+
+  na_shape_copy( ndim, shp1, a1 );
+  na_shape_copy( ndim, shp2, a2 );
+  na_shape_copy( ndim, shp3, a3 );
+  ndim -= ncd;
+  shp1 += ncd1;
+  shp2 += ncd2;
+  shp3 += ncd3;
+  na_shape_max3( ndim, itr, shp1, shp2, shp3 );
+
+  ndim = na_set_slice_3obj( ndim, s1, s2, s3, shp1, shp2, shp3, itr );
+
+  nsz1 = na_shape_total(a1->rank-ncd1,a1->shape+ncd1);
+  nsz2 = na_shape_total(a2->rank-ncd2,a2->shape+ncd2);
+  nsz3 = na_shape_total(a3->rank-ncd3,a3->shape+ncd3);
+
+  na_init_slice(s1, ndim, shp1, na_sizeof[a1->type]*nsz1 );
+  na_init_slice(s2, ndim, shp2, na_sizeof[a2->type]*nsz2 );
+  na_init_slice(s3, ndim, shp3, na_sizeof[a3->type]*nsz3 );
+
+  na_loop_linalg( ndim, a1->ptr, a2->ptr, a3->ptr,
+		  s1, s2, s3, func, a2->shape, a2->type );
+  xfree(s1);
+}
+
+
+static int
+na_lu_fact_func_body(int ni, char *a, char *idx, na_shape_t *shape, int type, char *buf)
+{
+  na_shape_t i, j, k;
+  int imax;
+
+  char *amax, *rtmp;
+  char *aa, *aii, *aij, *ai0, *a0i, *a0j;
+  char *v, *vi;
+
+  na_funcset_t *f = &na_funcset[type];
+  na_funcset_t *r = &na_funcset[na_cast_real[type]];
+
+  int status = 0;
+  na_shape_t n = shape[0];
+  na_shape_t relmsz = r->elmsz;
+  na_shape_t felmsz = f->elmsz;
+  na_shape_t rowsz  = felmsz*n;
+  na_shape_t matsz  = rowsz*n;
+  na_shape_t diagsz = rowsz + felmsz;
+
+  v    = buf + rowsz;
+  amax = v   + relmsz*n;
+
+  while (ni-->0) {
+
+    aa = a;
+    vi = v;
+
+    /* v[j] = 1/max( abs( a[i,j] ) ) */
+    for (j=0;j<n;++j) {
+      f->abs(n, buf, relmsz, aa, felmsz);
+
+      r->set(1, amax,0, r->zero,0);
+      rtmp = buf;
+      for (i=0;i<n;++i) {
+	if (r->sort(rtmp, amax) == 1)
+	  r->set(1, amax,0, rtmp,0);
+	rtmp += relmsz;
+      }
+
+      if (r->sort(amax,r->tiny) != 1)
+	status = 2; /* Singular Matrix */
+
+      r->rcp(1, vi,0, amax,0);
+      vi += relmsz;
+    }
+
+    ai0 = a0i = aii = a;
+    vi  = v;
+
+    for (i=0;i<n;++i) {
+
+      f->set(n, buf, felmsz, ai0, rowsz);
+
+      aij = buf;
+      a0j = a;
+      /* a[i,j(<i)]  -=  sum(k<j){ a[i,k]*a[k,j] } */
+      for (j=1;j<i;++j) {
+	aij += felmsz;
+	a0j += rowsz;
+	f->mulsbt(j, aij, 0, buf, felmsz, a0j, felmsz);
+      }
+      /* a[i,j(>=i)]  -=  sum(k<i){ a[i,k]*a[k,j] } */
+      for (  ;j<n;++j) {
+	aij += felmsz;
+	a0j += rowsz;
+	f->mulsbt(i, aij, 0, buf, felmsz, a0j, felmsz);
+      }
+      f->set(n, ai0, rowsz, buf, felmsz);
+
+      /* pivoting
+	 imax = max_idx( abs( a[i,j(>=i)] ) * v[j(>=i)] ) */
+      f->abs(n-i, buf, relmsz, aii, rowsz);
+      r->mul(n-i, buf, relmsz, vi, relmsz);
+
+      r->set(1, amax,0, r->zero,0);
+      rtmp = buf;
+      imax = 0;
+      for (j=i;j<n;++j) {
+	if (r->sort(rtmp,amax) == 1) {
+	  r->set(1, amax,0, rtmp,0);
+	  imax = j;
+	}
+	rtmp += relmsz;
+      }
+
+      if (r->sort(amax,r->tiny)!=1)
+	status = 1; /* Singular Matrix */
+
+      if (i != imax) {
+	/* a[*,i] <=> a[*,imax] */
+	SWAPMEM(a+i*rowsz, a+imax*rowsz, buf, rowsz);
+	/* v[i]   <=> v[imax] */
+	SWAPMEM(vi, v+imax*relmsz, buf, relmsz);
+	NA_SWAP(((int32_t*)idx)[i],((int32_t*)idx)[imax],k);
+      }
+
+      /* a[i,j(>i)]  = a[i,j]/a[i,i] */
+      f->div(n-i-1, aii+rowsz, rowsz, aii, 0);
+
+      ai0 += felmsz;
+      a0i += rowsz;
+      aii += diagsz;
+      vi  += relmsz;
+    }
+
+    a   += matsz;
+    idx += sizeof(int32_t)*n;
+  }
+  return status;
+}
+
+
+
+static int
+ na_lu_fact_func(int ni, char *a, char *idx, na_shape_t *shape, int type)
+{
+  volatile VALUE val;
+  char *buf;
+  int status;
+  na_shape_t size, n=shape[0];
+
+  if (type==NA_ROBJ) {
+    VALUE *mem;
+    int i;
+    size = n*2+1;
+    mem = ALLOC_N(VALUE, size);
+    for (i=0; i<size; i++) mem[i] = Qnil;
+    val = rb_ary_new4(size, mem);
+    xfree(mem);
+    buf = (char*)((RARRAY_PTR(val)));
+    status = na_lu_fact_func_body( ni, a, idx, shape, type, buf );
+  } else {
+    size = na_sizeof[type]*n + na_sizeof[na_cast_real[type]]*(n+1);
+    buf = ALLOC_N(char, size);
+    status = na_lu_fact_func_body( ni, a, idx, shape, type, buf );
+    xfree(buf);
+  }
+  return status;
+}
+
+
+/* :nodoc: */
+static VALUE
+ na_lu_fact_bang(VALUE self)
+{ 
+  na_shape_t i, total, n;
+  int sz, stat;
+  struct NARRAY *ary;
+  VALUE piv;
+  char *ptr, *idx;
+  void (*func)();
+
+  GetNArray(self,ary);
+
+  /* shape & dimension check */
+  if (ary->rank<2)
+    rb_raise(rb_eTypeError,"dim(=%i) < 2", ary->rank);
+
+  n = ary->shape[0];
+  if (n != ary->shape[1])
+    rb_raise(rb_eTypeError,"not square matrix");
+
+  total=1;
+  for (i=2; i<ary->rank; ++i)
+    total *= ary->shape[i];
+
+  piv = na_make_object(NA_LINT, ary->rank-1, ary->shape+1, cNVector);
+
+  /* prepare pivot index */
+  func = IndGenFuncs[NA_LINT];
+  sz   = na_sizeof[NA_LINT];
+  ptr  = idx = ((struct NARRAY *)DATA_PTR(piv))->ptr;
+  for (i=0; i<total; ++i) {
+    func(n,ptr,sz,0,1);
+    ptr += n*sz;
+  }
+
+  stat = na_lu_fact_func(total, ary->ptr, idx, ary->shape, ary->type);
+
+  if (stat!=0)
+    rb_raise(rb_eZeroDivError,"singular matrix, status=%i",stat);
+
+  return rb_funcall(cNMatrixLU,na_id_new,2,self,piv);
+}
+
+
+/* :nodoc: */
+static VALUE
+ na_lu_fact(VALUE self)
+{
+  return na_lu_fact_bang( na_clone(self) );
+}
+
+
+static void
+na_lu_pivot_func( int ni,
+		  char *x, int ps1, char *y, int ps2, char *idx, int ps3,
+		  na_shape_t *shape, int type )
+{
+  int i, n, sz;
+  char *xi;
+  na_funcset_t *f = &na_funcset[type];
+
+  n = shape[1];
+  sz = f->elmsz * shape[0];
+
+  for (; ni>0; --ni) {
+    xi = x;
+    for (i=0; i<n; ++i) {
+      memcpy(xi, y+((int32_t*)idx)[i]*sz, sz);
+      xi += sz;
+    }
+    x   += ps1;
+    y   += ps2;
+    idx += ps3;
+  }
+}
+
+
+
+static void
+na_lu_solve_func_body( int ni,
+		       char *x, int ps1,  char *a, int ps2,
+		       na_shape_t *shape, int type, char *buf )
+{
+  char *aii, *a0i, *xx, *xi;
+  na_shape_t i,k;
+  na_funcset_t *f = &na_funcset[type];
+  na_shape_t n = shape[1];
+  int sz = na_sizeof[type];
+  na_shape_t xsz = shape[0] * sz;
+  na_shape_t rowsz = sz * n;
+  na_shape_t matsz = rowsz * n;
+  na_shape_t diagsz = rowsz + sz;
+
+  for (; ni>0; --ni) {
+
+    xx = x;
+
+    for (k=shape[0]; k>0; --k) { /* once if x is vector */
+
+      f->set(n, buf,sz, xx,xsz);
+
+      xi  = buf;
+      a0i = a;
+
+      /* solve Lx' = y' */
+      for (i=1; i<n; ++i) {
+	/* x[i] -= a[j(<i),i] * x[j(<i)] */
+	xi  += sz;
+	a0i += rowsz;
+	f->mulsbt(i, xi, 0, a0i, sz, buf, sz);
+      }
+
+      xi  = buf + sz*(n-1);
+      aii = a + (matsz-sz);
+
+      /* solve Ux = x' */
+      f->div(1, xi,0, aii,0);
+      for (i=n-1; i>0; --i) {
+	xi  -= sz;
+	aii -= diagsz;
+	/* x[i] -= a[j(>i),i] * x[j(>i)] */
+	f->mulsbt(n-i, xi,0, aii+sz, sz, xi+sz, sz);
+	/* x[i] /= a[i,i] */
+	f->div(1, xi,0, aii,0);
+      }
+
+      f->set(n, xx,xsz, buf,sz);
+
+      xx += sz;
+    }
+
+    x += ps1;
+    a += ps2;
+  }
+}
+
+
+static void
+na_lu_solve_func( int ni,
+		  char *z, int ps,  char *x, int ps1,  char *a, int ps2,
+		  na_shape_t *shape, int type )
+{
+  volatile VALUE val;
+  char *buf;
+  na_shape_t size;
+
+  if (type==NA_ROBJ) {
+    VALUE *mem;
+    na_shape_t i;
+    size = shape[1];
+    mem = ALLOC_N(VALUE, size);
+    for (i=0; i<size; i++) mem[i] = Qnil;
+    val = rb_ary_new4(size, mem);
+    xfree(mem);
+    buf = (char*)((RARRAY_PTR(val)));
+    na_lu_solve_func_body( ni, x, ps1, a, ps2, shape, type, buf );
+  } else {
+    size = shape[1] * na_sizeof[type];
+    buf = ALLOC_N(char, size);
+    na_lu_solve_func_body( ni, x, ps1, a, ps2, shape, type, buf );
+    xfree(buf);
+  }
+}
+
+
+static void
+na_shape_max2(int ndim, na_shape_t *shape, int n1, na_shape_t *shape1, int n2, na_shape_t *shape2)
+{
+  na_shape_t *tmp;
+  int  i;
+
+  if (n1 < n2) {
+    NA_SWAP(shape1,shape2,tmp);
+  }
+
+  for (i=0; i<n2; ++i) {
+    shape[i] = NA_MAX(shape1[i],shape2[i]);
+  }
+  for (   ; i<n1; ++i) {
+    shape[i] = shape1[i];
+  }
+  for (   ; i<ndim; ++i) {
+    shape[i] = 1;
+  }
+}
+
+
+
+/*
+ *  call-seq:
+ *     lu.solve(arg)  -> result
+ *
+ *  Solve with the result of LU factorization.
+ *  arg should be NMatrix or NVector instance.
+ *  Returns an instance of same class with arg.
+ */
+static VALUE
+na_lu_solve(VALUE self, volatile VALUE other)
+{
+  int  ndim;
+  na_shape_t n, *shape;
+  struct NARRAY *a1, *a2, *l, *p;
+  VALUE pv, obj, klass;
+  volatile VALUE lu;
+
+  klass = CLASS_OF(other);
+  if (klass==cNVector)
+    other = na_newdim_ref(1,(VALUE*)na_funcset[NA_ROBJ].zero,other);
+  else if (klass!=cNMatrix)
+    rb_raise(rb_eTypeError,"neither NMatrix or NVector");
+
+  lu = rb_ivar_get(self, id_lu);
+  pv = rb_ivar_get(self, id_pivot);
+
+  GetNArray(lu,l);
+
+  other = na_upcast_object(other,l->type);
+  GetNArray(other,a1);
+
+  lu = na_upcast_type(lu,a1->type);
+  GetNArray(lu,l);
+  GetNArray(pv,p);
+
+  n = l->shape[0];
+  if (n != a1->shape[1])
+    rb_raise(rb_eTypeError,"size mismatch (%zd!=%zd)",n,a1->shape[1]);
+
+  ndim  = NA_MAX(l->rank, a1->rank);
+  shape = ALLOCA_N(na_shape_t, ndim);
+
+  shape[0] = a1->shape[0];
+  na_shape_max2( ndim-1, shape+1, a1->rank-1, a1->shape+1,
+		 l->rank-1, l->shape+1 );
+  obj = na_make_object( a1->type, ndim, shape, klass );
+
+  GetNArray(obj,a2);
+
+  na_exec_linalg( a2, a1, p, 2, 2, 1, na_lu_pivot_func );
+  na_exec_linalg( a2, a2, l, 2, 2, 2, na_lu_solve_func );
+
+  if (klass==cNVector) {
+    shape = ALLOC_N(na_shape_t, ndim-1);
+    memcpy(shape,a2->shape+1,sizeof(na_shape_t)*(ndim-1));
+    xfree(a2->shape);
+    a2->shape = shape;
+    --(a2->rank);
+  }
+  return obj;
+}
+
+
+/* :nodoc: */
+static VALUE
+na_lu_init(VALUE self, VALUE lu, VALUE piv)
+{
+  int i;
+  struct NARRAY *l, *p;
+
+  if (CLASS_OF(lu)!=cNMatrix)
+    rb_raise(rb_eTypeError,"LU should be NMatrix");
+  if (CLASS_OF(piv)!=cNVector)
+    rb_raise(rb_eTypeError,"pivot should be NVector");
+
+  GetNArray(lu,l);
+  GetNArray(piv,p);
+
+  if (p->type != NA_LINT)
+    rb_raise(rb_eRuntimeError,"pivot type must be Integer");
+
+  if (l->rank != p->rank+1)
+    rb_raise(rb_eRuntimeError,"array dimension mismatch %i!=%i+1",
+	     l->rank, p->rank);
+
+  if (l->shape[0] != l->shape[1])
+    rb_raise(rb_eRuntimeError,"LU matrix (%zd,%zd) is not square",
+	     l->shape[0], l->shape[1]);
+
+  for (i=1; i<l->rank; ++i)
+    if (l->shape[i] != p->shape[i-1])
+      rb_raise(rb_eRuntimeError,"array size mismatch %zd!=%zd at %i",
+	       l->shape[i], p->shape[i-1], i);
+
+  rb_ivar_set(self, id_lu, lu);
+  rb_ivar_set(self, id_pivot, piv);
+  return Qnil;
+}
+
+
+
+void Init_na_linalg()
+{
+  static double tiny_d = 1e-15;
+  static float  tiny_f = (float)1e-7;
+  int i, sz;
+  int32_t one=1, zero=0;
+  static VALUE zerov = INT2FIX(0);
+  static VALUE onev = INT2FIX(1);
+  char *a = malloc(NA_NTYPES*sizeof(dcomplex)*2);
+
+  for (i=1;i<NA_NTYPES;++i) {
+    sz = na_funcset[i].elmsz = na_sizeof[i];
+    sz = (sz>((int)sizeof(int))) ? sz : (int)sizeof(int);
+    SetFuncs[i][NA_LINT](1, a,0, &one, 0);
+    na_funcset[i].one = a;
+    a += sz;
+    SetFuncs[i][NA_LINT](1, a,0, &zero,0);
+    na_funcset[i].zero = a;
+    na_funcset[i].tiny = a;
+    a += sz;
+    na_funcset[i].set = SetFuncs[i][i];
+    na_funcset[i].neg = NegFuncs[i];
+    na_funcset[i].rcp = RcpFuncs[i];
+    na_funcset[i].abs = AbsFuncs[i];
+    na_funcset[i].add = AddUFuncs[i];
+    na_funcset[i].sbt = SbtUFuncs[i];
+    na_funcset[i].mul = MulUFuncs[i];
+    na_funcset[i].div = DivUFuncs[i];
+    na_funcset[i].mod = ModUFuncs[i];
+    na_funcset[i].muladd = MulAddFuncs[i];
+    na_funcset[i].mulsbt = MulSbtFuncs[i];
+    na_funcset[i].cmp = CmpFuncs[i];
+    na_funcset[i].min = MinFuncs[i];
+    na_funcset[i].max = MaxFuncs[i];
+    na_funcset[i].sort = SortFuncs[i];
+  }
+  na_funcset[NA_SFLOAT].tiny = (char*)&tiny_f;
+  na_funcset[NA_DFLOAT].tiny = (char*)&tiny_d;
+  na_funcset[NA_ROBJ].zero = (char*)&zerov;
+  na_funcset[NA_ROBJ].one  = (char*)&onev;
+
+  cNVector = rb_define_class("NVector",cNArray);
+  cNMatrix = rb_define_class("NMatrix",cNArray);
+  cNMatrixLU = rb_define_class("NMatrixLU",rb_cObject);
+
+  rb_define_method(cNMatrix, "lu_fact!", na_lu_fact_bang, 0);
+  rb_define_alias(cNMatrix,  "lu!","lu_fact!");
+  rb_define_method(cNMatrix, "lu_fact",  na_lu_fact, 0);
+  rb_define_alias(cNMatrix,  "lu","lu_fact");
+
+  rb_define_method(cNMatrixLU, "initialize", na_lu_init, 2);
+  rb_define_method(cNMatrixLU, "solve", na_lu_solve, 1);
+
+  id_lu    = rb_intern("@lu");
+  id_pivot = rb_intern("@pivot");
+}