From 1d18b7687309dc120bcc943028ad359e93ce0c65 Mon Sep 17 00:00:00 2001
From: Richard Kreckel <kreckel@ginac.de>
Date: Sat, 27 May 2000 17:18:40 +0000
Subject: [PATCH] - src/float/lfloat/algebraic/cl_LF_sqrt.cc,
 src/base/digitseq/cl_DS_sqrt.cc:   Readjusted break-even points.

---
 ChangeLog                                |  7 ++++++-
 src/base/digitseq/cl_DS_sqrt.cc          | 22 ++++++++++++++++++++--
 src/float/lfloat/algebraic/cl_LF_sqrt.cc |  2 +-
 3 files changed, 27 insertions(+), 4 deletions(-)

diff --git a/ChangeLog b/ChangeLog
index fdcb027..55dadce 100644
--- a/ChangeLog
+++ b/ChangeLog
@@ -1,9 +1,14 @@
+2000-05-27  Richard Kreckel  <kreckel@ginac.de>
+
+        * src/float/lfloat/algebraic/cl_LF_sqrt.cc,
+          src/base/digitseq/cl_DS_sqrt.cc: Readjusted break-even points.
+
 2000-05-24  Richard Kreckel  <kreckel@ginac.de>
 
         * autoconf/config.*: Updated to new version from FSF
           (the new libtool wants this).
         * src/Makefile.in: added $(LDFLAGS) to link step.
-        * src/base/digitseq/cl_2DS_div.cc, cl_2DS_recip.cc: Adjusted 
+        * src/base/digitseq/cl_2DS_div.cc, cl_2DS_recip.cc: Readjusted 
           break-even points.
 
 2000-05-23  Bruno Haible  <haible@clisp.cons.org>
diff --git a/src/base/digitseq/cl_DS_sqrt.cc b/src/base/digitseq/cl_DS_sqrt.cc
index b265907..3b3f540 100644
--- a/src/base/digitseq/cl_DS_sqrt.cc
+++ b/src/base/digitseq/cl_DS_sqrt.cc
@@ -14,6 +14,24 @@
 
 // We observe the following timings:
 // Time for square root of a_len = 2*N by b_len = N digits,
+// OS: Linux 2.2, intDsize==32,        OS: TRU64/4.0, intDsize==64,
+// Machine: P-III/450MHz               Machine: EV5/300MHz:
+//      N   standard  Newton           standard  Newton
+//      30   0.00002   0.00009          0.00011   0.00027
+//     100   0.00012   0.00052          0.00057   0.0017
+//     300   0.00087   0.0031           0.0037    0.0091
+//    1000   0.0089    0.020            0.037     0.069
+//    3000   0.087     0.11  <-(~3200)  0.30      0.28  <- (~2750)
+//   10000   1.27      0.55             3.5       1.3
+//   30000  12.7       1.35            31.1       3.4
+// Newton faster for 3200<N            Newton faster for 2750<N
+// When in doubt, prefer to choose the standard algorithm.
+#if CL_USE_GMP
+  static inline cl_boolean cl_recipsqrt_suitable (uintL n)
+  { return (cl_boolean)(n >= 3200); }
+#else
+// Use the old default values from CLN version <= 1.0.3 as a crude estimate.
+// Time for square root of a_len = 2*N by b_len = N digits,
 // on a i486 33 MHz running Linux:
 //      N   standard  Newton
 //      10    0.00022 0.00132
@@ -27,9 +45,9 @@
 //    5000   24.1    10.7
 //   10000   98      23.2
 //   -----> Newton faster for 1570 <= N <= 1790 and for N >= 2100.
-// When in doubt, prefer to choose the standard algorithm.
   static inline cl_boolean cl_recipsqrt_suitable (uintL n)
-    { return (cl_boolean)(n >= 2100); }
+  { return (cl_boolean)(n >= 2100); }
+#endif
 
 // Workaround gcc-2.7.0 bug on i386.
 #if defined(__GNUC__)
diff --git a/src/float/lfloat/algebraic/cl_LF_sqrt.cc b/src/float/lfloat/algebraic/cl_LF_sqrt.cc
index 9679a40..7ca0850 100644
--- a/src/float/lfloat/algebraic/cl_LF_sqrt.cc
+++ b/src/float/lfloat/algebraic/cl_LF_sqrt.cc
@@ -64,7 +64,7 @@ const cl_LF sqrt (const cl_LF& x)
       var uintD* y_mantMSDptr = arrayMSDptr(TheLfloat(y)->data,len);
       // Wurzel ziehen:
 #ifndef CL_LF_PEDANTIC
-      if (len > 1900) // Das ist etwa 10% bis 20% schneller (im Mittel 15%).
+      if (len > 2900) // This is about 15% faster
         { // Kehrwert der Wurzel errechnen:
           var uintD* s_MSDptr;
           var uintD* s_LSDptr;