* examples/Makefile.am:

* examples/memcpy-speed.c: (main):
New memory speed test program
* examples/oil-test.c: (main):
Use n=100 for test.
* liboil/copy/Makefile.am:
* liboil/copy/copy_i386.c: (copy_u8_mmx3):
movntq based copy
* liboil/copy/splat_i386.c: (splat_u32_ns_mmx):
movntq based splat

3 files changed, 78 insertions, 2 deletions

diff --git a/examples/Makefile.am b/examples/Makefile.am
index d648f68..f38c12a 100644
--- a/examples/Makefile.am
+++ b/examples/Makefile.am
@@ -1,7 +1,7 @@
 
 SUBDIRS = jpeg md5 uberopt work huffman
 
-noinst_PROGRAMS = example1 oil-inspect oil-test report
+noinst_PROGRAMS = example1 oil-inspect oil-test report memcpy-speed
 
 
 example1_SOURCES = example1.c
@@ -20,3 +20,7 @@ report_SOURCES = report.c
 report_CFLAGS = $(LIBOIL_CFLAGS)
 report_LDADD = $(LIBOIL_LIBS)
 
+memcpy_speed_SOURCES = memcpy-speed.c
+memcpy_speed_CFLAGS = $(LIBOIL_CFLAGS)
+memcpy_speed_LDADD = $(LIBOIL_LIBS)
+
diff --git a/examples/memcpy-speed.c b/examples/memcpy-speed.c
new file mode 100644
index 0000000..394625c
--- /dev/null
+++ b/examples/memcpy-speed.c
@@ -0,0 +1,70 @@
+
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+
+#include <liboil/liboil.h>
+#include <liboil/liboilprofile.h>
+#include <liboil/liboilfunction.h>
+
+#define ALIGN(ptr,n) ((void *)((unsigned long)(ptr) & (~(unsigned long)(n-1))))
+
+int
+main(int argc, char *argv[])
+{
+  char *s, *d;
+  uint32_t *src, *dest;
+  OilProfile prof;
+  double ave, std;
+  int i,j;
+  double cpufreq;
+  OilFunctionClass *klass;
+  OilFunctionImpl *impl;
+  int use_memset = 0;
+
+  //use_memset = 1;
+
+  oil_init ();
+
+  cpufreq = 1788.699e6;
+
+  s = malloc(1024*1024*64+1024);
+  d = malloc(1024*1024*64+1024);
+  src = ALIGN(s,128);
+  dest = ALIGN(d,128);
+
+  if (use_memset) {
+    klass = oil_class_get ("splat_u32_ns");
+  } else {
+    klass = oil_class_get ("copy_u8");
+  }
+
+  for(impl=klass->first_impl;impl;impl=impl->next) {
+    printf("impl %s\n", impl->name);
+
+    oil_class_choose_by_name (klass, impl->name);
+
+    for(i=10;i<26;i++){
+      oil_profile_init (&prof);
+      for(j=0;j<10;j++){
+        if (use_memset) {
+          oil_profile_start(&prof);
+          oil_splat_u32_ns (dest, src, 1<<(i-2));
+          oil_profile_stop(&prof);
+        } else {
+          oil_profile_start(&prof);
+          oil_memcpy (dest, src, 1<<i);
+          oil_profile_stop(&prof);
+        }
+      }
+
+      oil_profile_get_ave_std (&prof, &ave, &std);
+
+      printf("%d: %10.4g %10.4g %10.4g %10.4g\n", i, ave, std,
+          ave/(1<<i), cpufreq/(ave/(1<<i)));
+    }
+  }
+
+  return 0;
+}
+
diff --git a/examples/oil-test.c b/examples/oil-test.c
index c7f369c..9072da4 100644
--- a/examples/oil-test.c
+++ b/examples/oil-test.c
@@ -173,7 +173,7 @@ int main (int argc, char *argv[])
 
   test = oil_test_new(klass);
   oil_test_set_iterations(test, 1);
-  test->n = 10;
+  test->n = 100;
   test->m = 10;
 
   impl = klass->reference_impl;
@@ -184,6 +184,7 @@ int main (int argc, char *argv[])
   dump_source(test);
   printf ("reference impl %s\n", impl->name);
   printf("  ave=%g std=%g\n", ave, std);
+  printf("  (this test) ave=%g std=%g\n", impl->profile_ave, impl->profile_std);
   dump_test(test);
 
   for (impl = klass->first_impl; impl; impl = impl->next) {
@@ -192,6 +193,7 @@ int main (int argc, char *argv[])
     if (oil_impl_is_runnable (impl)) {
       printf("  ave=%g std=%g\n", impl->profile_ave, impl->profile_std);
       oil_test_check_impl (test, impl);
+      printf("  (this test) ave=%g std=%g\n", impl->profile_ave, impl->profile_std);
       dump_test(test);
     }
   }