Re: [PATCH] Optimize MIPS memcpy

[Steve Ellcey <sellcey at mips dot com>]

> Exactly what benchmarks did you run to verify the performance gains?
> 
> The one thing I'd like to continue seeing is strong rationalization for
> performance patches such that we have reproducible data in the event that
> someone else comes along and wants to make a change.
> 
> For example see:
> http://sourceware.org/glibc/wiki/benchmarking/results_2_17
> 
> and:
> http://sourceware.org/glibc/wiki/benchmarking/benchmarks
> 
> Cheers,
> Carlos.

We had a few tests around here that I used and I wrote one of my own
too.  I have attached my test, when using it with -UVERIFY and testing
on a 74K system I got the following timings (32 bits, little-endian):

The FSF memcpy: 3m9.34s
Maxim's memcpy: 2m0.41s
My memcpy:      1m22.20s

If there are any official or recommended memcpy benchmarks I'd be happy
to try them as well.

Steve Ellcey
sellcey@mips.com

#include <string.h>
#include <stdio.h>

#define SIZE 1024*100
#define MAXCOPYSIZE 1024*50
#define MAXSRCOFFSET 13
#define MAXDSTOFFSET 18
#define SRCVAL(N) ((N+10000) % 13)
#define DSTVAL(N) ((N+20001) % 17)
char src[SIZE], dst[SIZE];

#ifndef MEMCPY_NAME
#define MEMCPY_NAME memcpy
#endif

extern void *MEMCPY_NAME(void *, const void *, size_t);

test(int src_offset, int dst_offset, int size)
{
  int i;
  char *x, *y;
  for (i = 0; i < SIZE; i++) {
    src[i] = SRCVAL(i);
    dst[i] = DSTVAL(i);
  }
  x = src;
  y = dst;
  x = x + src_offset;
  y = y + dst_offset;
  MEMCPY_NAME(&dst[dst_offset], &src[src_offset], size);
  for (i = 0; i < SIZE; i++) {
    if (src[i] != SRCVAL(i)) printf("FAIL, src got changed\n");
    if (i < dst_offset) {
      if (dst[i] != DSTVAL(i))
	printf("FAIL, dst got changed before it should be\n");
    } else if (i >= (dst_offset+size)) {
      if (dst[i] != DSTVAL(i))
	printf("FAIL, dst got changed after it should be (%d %d %d %d)\n", src_offset, dst_offset, size, i);
    } else {
      if (dst[i] != SRCVAL(i-dst_offset+src_offset)) {
	printf("FAIL, dst was not changed when it should be (%d %d %d %d)\n", src_offset, dst_offset, size, i);
      }
    }
  }
}

main()
{
  int i, j, k;
  for (i = 8; i < MAXDSTOFFSET; i++)
    for (j = 8; j < MAXSRCOFFSET; j++)
      for (k = MAXCOPYSIZE-20; k < MAXCOPYSIZE; k++)
      test(i, j, k);
}