Changeset 2a4a23 in git

Timestamp:

Apr 20, 2018, 3:03:49 PM (5 years ago)

Author:

Hans Schoenemann <hannes@…>

Branches:

(u'spielwiese', '8e0ad00ce244dfd0756200662572aef8402f13d5')

Children:

eb6e1d58eb9834a39221581ba56d2e26f61499cc

Parents:

65c1bd5b526a9447123bf85f32538aff3498501c

Message:

fix: Z/p-arithmetic for all architectures

Files:

• Singular/misc_ip.cc (modified) (1 diff)
• libpolys/coeffs/modulop.cc (modified) (6 diffs)
• libpolys/coeffs/modulop.h (modified) (2 diffs)
• m4/cpu-check.m4 (modified) (1 diff)

Singular/misc_ip.cc

@@ -895 +895 @@
               StringAppendS("AvoidBranching,");
 #endif
-#ifdef HAVE_MULT_MOD
+#ifdef HAVE_GENERIC_MULT
               StringAppendS("GenericMult,");
 #else
               StringAppendS("TableMult,");
+#endif
+#ifdef HAVE_INVTABLE
+              StringAppendS("invTable,");
+#else
+              StringAppendS("no invTable,");
 #endif
               StringAppendS("\n\t");

libpolys/coeffs/modulop.cc

@@ -207 +207 @@
   s += (s >> 31) & R->ch;
   #endif
+  return s;
 #endif
 }
@@ -213 +214 @@
 {
   n_Test(c, r);
-#if !defined(HAVE_MULT_MOD) && (!defined(HAVE_DIV_MOD))
+#ifndef HAVE_GENERIC_MULT
+  #ifndef HAVE_INVTABLE
   number d = (number)(long)r->npExpTable[r->npPminus1M - r->npLogTable[(long)c]];
-#else
+  #else
   long inv=(long)r->npInvTable[(long)c];
   if (inv==0)
   {
-    #ifndef HAVE_MULT_MOD
     inv = (long)r->npExpTable[r->npPminus1M - r->npLogTable[(long)c]];
-    #else
+    r->npInvTable[(long)c]=inv;
+  }
+  number d = (number)inv;
+  #endif
+#else
+  #ifdef HAVE_INVTABLE
+  long inv=(long)r->npInvTable[(long)c];
+  if (inv==0)
+  {
     inv=InvMod((long)c,r);
-    #endif
     r->npInvTable[(long)c]=inv;
   }
+  #else
+  long  inv=InvMod((long)c,r);
+  #endif
   number d = (number)inv;
 #endif
@@ -245 +256 @@
 
   number d;
-#ifndef HAVE_DIV_MOD
+#ifndef HAVE_GENERIC_MULT
   int s = r->npLogTable[(long)a] - r->npLogTable[(long)b];
+  #ifdef HAVE_GENERIC_ADD
   if (s < 0)
     s += r->npPminus1M;
+  #else
+    #if SIZEOF_LONG == 8
+    s += ((long)s >> 63) & r->npPminus1M;
+    #else
+    s += ((long)s >> 31) & r->npPminus1M;
+    #endif
+  #endif
   d = (number)(long)r->npExpTable[s];
 #else
@@ -395 +414 @@
 void npKillChar(coeffs r)
 {
-  #ifdef HAVE_DIV_MOD
+  #ifdef HAVE_INVTABLE
   if (r->npInvTable!=NULL)
-  omFreeSize( (void *)r->npInvTable, r->ch*sizeof(unsigned short) );
-  r->npInvTable=NULL;
-  #else
+  {
+    omFreeSize( (void *)r->npInvTable, r->ch*sizeof(unsigned short) );
+    r->npInvTable=NULL;
+  }
+  #endif
+  #ifndef HAVE_GENERIC_MULT
   if (r->npExpTable!=NULL)
   {
@@ -548 +570 @@
 #endif
   {
-#ifdef HAVE_DIV_MOD
+#ifdef HAVE_INVTABLE
     r->npInvTable=(unsigned short*)omAlloc0( r->ch*sizeof(unsigned short) );
-#elif (!defined(HAVE_MULT_MOD)||(!HAVE_DIV_MOD))
+#endif
+#ifndef HAVE_GENERIC_MULT
     r->npExpTable=(unsigned short *)omAlloc0( r->ch*sizeof(unsigned short) );
     r->npLogTable=(unsigned short *)omAlloc0( r->ch*sizeof(unsigned short) );
@@ -801 +824 @@
 }
 
-static inline long nvInvMod(long a, const coeffs R)
-{
-#ifdef HAVE_DIV_MOD
-  return InvMod(a, R);
-#else
-/// TODO: use "long InvMod(long a, const coeffs R)"?!
-
-   long  s;
-
-   long  u, u0, u1, u2, q, r; // v0, v1, v2,
-
-   u1=1; // v1=0;
-   u2=0; // v2=1;
-   u = a;
-
-   long v = R->ch;
-
-   while (v != 0)
-   {
-      q = u / v;
-      r = u % v;
-      u = v;
-      v = r;
-      u0 = u2;
-//      v0 = v2;
-      u2 = u1 - q*u2;
-//      v2 = v1 - q*v2;
-      u1 = u0;
-//      v1 = v0;
-   }
-
-   s = u1;
-   //t = v1;
-   if (s < 0)
-      return s + R->ch;
-   else
-     return s;
-#endif
-}
-
 static inline number nvInversM (number c, const coeffs r)
 {
-  long inv=nvInvMod((long)c,r);
+  long inv=InvMod((long)c,r);
   return (number)inv;
 }

libpolys/coeffs/modulop.h

@@ -9 +9 @@
 #include "misc/auxiliary.h"
 
-// defines are in struct.h
+
 // define if a*b is with mod instead of tables
-//#define HAVE_MULT_MOD
-// define if a/b is with mod instead of tables
-//#define HAVE_DIV_MOD
+//#define HAVE_GENERIC_MULT
+// define if 1/b is from  tables
+//#define HAVE_INVTABLE
 // define if an if should be used
 //#define HAVE_GENERIC_ADD
+
+//#undef HAVE_GENERIC_ADD
+//#undef HAVE_GENERIC_MULT
+//#undef HAVE_INVTABLE
+
+//#define HAVE_GENERIC_ADD 1
+//#define HAVE_GENERIC_MULT 1
+//#define HAVE_INVTABLE 1
 
 // enable large primes (32749 < p < 2^31-)
@@ -41 +49 @@
 //    return (number)res;
 // }
-#ifdef HAVE_MULT_MOD
+#ifdef HAVE_GENERIC_MULT
 static inline number npMultM(number a, number b, const coeffs r)
 {

m4/cpu-check.m4

@@ -49 +49 @@
 
 AS_CASE([$host_cpu],
+dnl the following settings seems to be better on itanium processors
+  [ia64*], [
+      AC_DEFINE(HAVE_GENERIC_ADD,1,use branch for addition in Z/p otherwise it uses a generic add)],
 dnl the following settings seems to be better on i386 and x86_64 processors
-  [i*86*|x86_64*], [AC_DEFINE(HAVE_MULT_MOD,1,multiplication is fast on the cpu: a*b is with mod otherwise using tables of logartihms)],
-dnl the following settings seems to be better on itanium processors
-dnl AC_DEFINE(HAVE_MULT_MOD,1,)
-  [ia64*], [AC_DEFINE(HAVE_GENERIC_ADD,1,use branch for addition in Z/p otherwise it uses a generic add)],
+dnl AC_DEFINE(HAVE_GENERIC_MULT,1,)
+  [i*86*|x86_64*], [
+            AC_DEFINE(HAVE_GENERIC_ADD,1,use branch for addition in Z/p otherwise it uses a generic add)
+            AC_DEFINE(HAVE_GENERIC_MULT,1,multiplication is fast on the cpu: a*b is with mod otherwise using tables of logartihms)
+            ],
 dnl the following settings seems to be better on sparc processors
   [sparc*], [
-            AC_DEFINE(HAVE_MULT_MOD,1,multiplication is fast on the cpu: a*b is with mod otherwise using tables of logartihms)
-            AC_DEFINE(HAVE_DIV_MOD,1,division using extend euclidian algorithm otherwise using tables of logartihms)
+            AC_DEFINE(HAVE_GENERIC_ADD,1,use branch for addition in Z/p otherwise it uses a generic add)
             ],
 dnl the following settings seems to be better on ppc processors
 dnl testet on: ppc_Linux, 740/750 PowerMac G3, 512k L2 cache
-  [powerpc*|ppc*], [AC_DEFINE(HAVE_MULT_MOD,1,multiplication is fast on the cpu: a*b is with mod otherwise using tables of logartihms)],
+  [powerpc*|yyppc*], [AC_DEFINE(HAVE_GENERIC_MULT,1,multiplication is fast on the cpu: a*b is with mod otherwise using tables of logartihms)],
   []
 )