lib/int_sqrt.c: optimize square root algorithm
authorDavidlohr Bueso <davidlohr.bueso@hp.com>
Mon, 29 Apr 2013 23:18:09 +0000 (16:18 -0700)
committerLinus Torvalds <torvalds@linux-foundation.org>
Tue, 30 Apr 2013 01:28:19 +0000 (18:28 -0700)
Optimize the current version of the shift-and-subtract (hardware)
algorithm, described by John von Newmann[1] and Guy L Steele.

Iterating 1,000,000 times, perf shows for the current version:

 Performance counter stats for './sqrt-curr' (10 runs):

         27.170996 task-clock                #    0.979 CPUs utilized            ( +-  3.19% )
                 3 context-switches          #    0.103 K/sec                    ( +-  4.76% )
                 0 cpu-migrations            #    0.004 K/sec                    ( +-100.00% )
               104 page-faults               #    0.004 M/sec                    ( +-  0.16% )
        64,921,199 cycles                    #    2.389 GHz                      ( +-  0.03% )
        28,967,789 stalled-cycles-frontend   #   44.62% frontend cycles idle     ( +-  0.18% )
   <not supported> stalled-cycles-backend
       104,502,623 instructions              #    1.61  insns per cycle
                                             #    0.28  stalled cycles per insn  ( +-  0.00% )
        34,088,368 branches                  # 1254.587 M/sec                    ( +-  0.00% )
             4,901 branch-misses             #    0.01% of all branches          ( +-  1.32% )

       0.027763015 seconds time elapsed                                          ( +-  3.22% )

And for the new version:

Performance counter stats for './sqrt-new' (10 runs):

          0.496869 task-clock                #    0.519 CPUs utilized            ( +-  2.38% )
                 0 context-switches          #    0.000 K/sec
                 0 cpu-migrations            #    0.403 K/sec                    ( +-100.00% )
               104 page-faults               #    0.209 M/sec                    ( +-  0.15% )
           590,760 cycles                    #    1.189 GHz                      ( +-  2.35% )
           395,053 stalled-cycles-frontend   #   66.87% frontend cycles idle     ( +-  3.67% )
   <not supported> stalled-cycles-backend
           398,963 instructions              #    0.68  insns per cycle
                                             #    0.99  stalled cycles per insn  ( +-  0.39% )
            70,228 branches                  #  141.341 M/sec                    ( +-  0.36% )
             3,364 branch-misses             #    4.79% of all branches          ( +-  5.45% )

       0.000957440 seconds time elapsed                                          ( +-  2.42% )

Furthermore, this saves space in instruction text:

   text    data     bss     dec     hex filename
    111       0       0     111      6f lib/int_sqrt-baseline.o
     89       0       0      89      59 lib/int_sqrt.o

[1] http://en.wikipedia.org/wiki/First_Draft_of_a_Report_on_the_EDVAC

Signed-off-by: Davidlohr Bueso <davidlohr.bueso@hp.com>
Reviewed-by: Jonathan Gonzalez <jgonzlez@linets.cl>
Tested-by: Jonathan Gonzalez <jgonzlez@linets.cl>
Signed-off-by: Andrew Morton <akpm@linux-foundation.org>
Signed-off-by: Linus Torvalds <torvalds@linux-foundation.org>
lib/int_sqrt.c

index fc2eeb7cb2eaf654b8d94b0369906d42ca023e21..1ef4cc344977cda2886ec1033a0868cec4a8ec41 100644 (file)
@@ -1,3 +1,9 @@
+/*
+ * Copyright (C) 2013 Davidlohr Bueso <davidlohr.bueso@hp.com>
+ *
+ *  Based on the shift-and-subtract algorithm for computing integer
+ *  square root from Guy L. Steele.
+ */
 
 #include <linux/kernel.h>
 #include <linux/export.h>
  */
 unsigned long int_sqrt(unsigned long x)
 {
-       unsigned long op, res, one;
+       unsigned long b, m, y = 0;
 
-       op = x;
-       res = 0;
+       if (x <= 1)
+               return x;
 
-       one = 1UL << (BITS_PER_LONG - 2);
-       while (one > op)
-               one >>= 2;
+       m = 1UL << (BITS_PER_LONG - 2);
+       while (m != 0) {
+               b = y + m;
+               y >>= 1;
 
-       while (one != 0) {
-               if (op >= res + one) {
-                       op = op - (res + one);
-                       res = res +  2 * one;
+               if (x >= b) {
+                       x -= b;
+                       y += m;
                }
-               res /= 2;
-               one /= 4;
+               m >>= 2;
        }
-       return res;
+
+       return y;
 }
 EXPORT_SYMBOL(int_sqrt);