annotate TOOLS/fastmemcpybench.c @ 27178:023835d5dfb9

Surround stream cache specific code by an appropriate #ifdef; fixes linking when stream cache is disabled. noticed by Andrea Palmat, andrea amigasoft net
author diego
date Sat, 05 Jul 2008 14:22:57 +0000
parents 8eff880f638c
children 13c920fdf7f5
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
1 /*
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
2 fastmemcpybench.c used to benchmark fastmemcpy.h code from libvo.
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
3
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
4 Note: this code can not be used on PentMMX-PII because they contain
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
5 a bug in rdtsc. For Intel processors since P6(PII) rdpmc should be used
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
6 instead. For PIII it's disputable and seems bug was fixed but I don't
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
7 tested it.
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
8 */
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
9
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
10 #include <stdio.h>
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
11 #include <stdlib.h>
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
12 #include <string.h>
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
13 #include <sys/ioctl.h>
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
14 #include <unistd.h>
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
15 #include <fcntl.h>
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
16 #include <sys/mman.h>
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
17 #include <sys/time.h>
13839
40391656ae23 small compilation fix
rathann
parents: 3082
diff changeset
18 #include <inttypes.h>
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
19
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
20 //#define ARR_SIZE 100000
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
21 #define ARR_SIZE (1024*768*2)
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
22
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
23 #ifdef HAVE_MGA
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
24
17017
dd053aeb5442 Unify include path handling by using -I.
diego
parents: 13839
diff changeset
25 #include "drivers/mga_vid.h"
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
26
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
27 static mga_vid_config_t mga_vid_config;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
28 static unsigned char* frame=NULL;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
29 static int f;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
30
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
31 static int mga_init(){
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
32
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
33 f = open("/dev/mga_vid",O_RDWR);
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
34 if(f == -1)
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
35 {
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
36 fprintf(stderr,"Couldn't open /dev/mga_vid\n");
26759
8eff880f638c cosmetics: Remove useless parentheses from return statements.
diego
parents: 26575
diff changeset
37 return -1;
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
38 }
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
39
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
40 mga_vid_config.num_frames=1;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
41 mga_vid_config.frame_size=ARR_SIZE;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
42 mga_vid_config.format=MGA_VID_FORMAT_YUY2;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
43
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
44 mga_vid_config.colkey_on=0;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
45 mga_vid_config.src_width = 640;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
46 mga_vid_config.src_height= 480;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
47 mga_vid_config.dest_width = 320;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
48 mga_vid_config.dest_height= 200;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
49 mga_vid_config.x_org= 0;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
50 mga_vid_config.y_org= 0;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
51
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
52 mga_vid_config.version=MGA_VID_VERSION;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
53 if (ioctl(f,MGA_VID_CONFIG,&mga_vid_config))
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
54 {
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
55 perror("Error in mga_vid_config ioctl()");
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
56 printf("Your mga_vid driver version is incompatible with this MPlayer version!\n");
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
57 exit(1);
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
58 }
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
59 ioctl(f,MGA_VID_ON,0);
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
60
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
61 frame = (char*)mmap(0,mga_vid_config.frame_size*mga_vid_config.num_frames,PROT_WRITE,MAP_SHARED,f,0);
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
62 if(!frame){
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
63 printf("Can't mmap mga frame\n");
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
64 exit(1);
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
65 }
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
66
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
67 //clear the buffer
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
68 //memset(frames[0],0x80,mga_vid_config.frame_size*mga_vid_config.num_frames);
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
69
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
70 return 0;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
71
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
72 }
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
73
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
74 #endif
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
75
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
76 // Returns current time in microseconds
26575
1ca484e74f18 Mark all functions that are only used within the file as static.
diego
parents: 24211
diff changeset
77 static unsigned int GetTimer(){
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
78 struct timeval tv;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
79 struct timezone tz;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
80 // float s;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
81 gettimeofday(&tv,&tz);
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
82 // s=tv.tv_usec;s*=0.000001;s+=tv.tv_sec;
26759
8eff880f638c cosmetics: Remove useless parentheses from return statements.
diego
parents: 26575
diff changeset
83 return tv.tv_sec * 1000000 + tv.tv_usec;
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
84 }
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
85
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
86 static inline unsigned long long int read_tsc( void )
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
87 {
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
88 unsigned long long int retval;
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
89 __asm __volatile ("rdtsc":"=A"(retval)::"memory");
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
90 return retval;
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
91 }
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
92
3076
bfc881c0e591 newly allocated memory seems to point to only 1 zero filled 4k page or something so there is a near 100% cache hit rate ... not very realistic, writeing something in the source array fixes that so the benchmark scores are meaningfull now
michael
parents: 687
diff changeset
93 unsigned char __attribute__((aligned(4096)))arr1[ARR_SIZE],arr2[ARR_SIZE];
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
94
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
95 int main( void )
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
96 {
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
97 unsigned long long int v1,v2;
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
98 unsigned char * marr1,*marr2;
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
99 int i;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
100 unsigned int t;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
101 #ifdef HAVE_MGA
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
102 mga_init();
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
103 marr1 = &frame[3];
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
104 #else
3081
aa361482f2f1 missaligned arrays, as nick requested
michael
parents: 3076
diff changeset
105 marr1 = &arr1[3];
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
106 #endif
3081
aa361482f2f1 missaligned arrays, as nick requested
michael
parents: 3076
diff changeset
107 marr2 = &arr2[9];
3076
bfc881c0e591 newly allocated memory seems to point to only 1 zero filled 4k page or something so there is a near 100% cache hit rate ... not very realistic, writeing something in the source array fixes that so the benchmark scores are meaningfull now
michael
parents: 687
diff changeset
108
3082
michael
parents: 3081
diff changeset
109 for(i=0; i<ARR_SIZE-16; i++) marr1[i] = marr2[i] = i;
3076
bfc881c0e591 newly allocated memory seems to point to only 1 zero filled 4k page or something so there is a near 100% cache hit rate ... not very realistic, writeing something in the source array fixes that so the benchmark scores are meaningfull now
michael
parents: 687
diff changeset
110
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
111 t=GetTimer();
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
112 v1 = read_tsc();
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
113 for(i=0;i<100;i++) memcpy(marr1,marr2,ARR_SIZE-16);
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
114 v2 = read_tsc();
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
115 t=GetTimer()-t;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
116 // ARR_SIZE*100/(1024*1024)/(t/1000000) = ARR_SIZE*95.36743/t
687
bcb46c9eaae1 slight proofreading
nickols_k
parents: 572
diff changeset
117 printf(NAME": cpu clocks=%llu = %dus (%5.3ffps) %5.1fMB/s\n",v2-v1,t,100000000.0f/(float)t,(float)ARR_SIZE*95.36743f/(float)t);
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
118 return 0;
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
119 }