annotate TOOLS/fastmemcpybench.c @ 28869:04a232eb5f0d

comment/output cosmetics
author diego
date Mon, 09 Mar 2009 11:32:43 +0000
parents a10ddae3de31
children 5643b5e6bee0
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 /*
28869
04a232eb5f0d comment/output cosmetics
diego
parents: 28868
diff changeset
2 * benchmark tool for fast_memcpy code from libvo
04a232eb5f0d comment/output cosmetics
diego
parents: 28868
diff changeset
3 *
04a232eb5f0d comment/output cosmetics
diego
parents: 28868
diff changeset
4 * NOTE: This code can not be used on Pentium MMX / II because they contain
04a232eb5f0d comment/output cosmetics
diego
parents: 28868
diff changeset
5 * a bug in rdtsc. For Intel processors since P6(PII) rdpmc should be used
04a232eb5f0d comment/output cosmetics
diego
parents: 28868
diff changeset
6 * instead. For PIII it's disputable and it seems the bug was fixed but this
04a232eb5f0d comment/output cosmetics
diego
parents: 28868
diff changeset
7 * was not confirmed through testing.
562
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
28232
8df85ad26746 Add missing 'void' keyword to parameterless function declarations.
diego
parents: 27757
diff changeset
10 /* According to Uoti this code is broken. */
8df85ad26746 Add missing 'void' keyword to parameterless function declarations.
diego
parents: 27757
diff changeset
11
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
12 #include <stdio.h>
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
13 #include <stdlib.h>
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
14 #include <string.h>
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
15 #include <sys/ioctl.h>
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
16 #include <unistd.h>
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
17 #include <fcntl.h>
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
18 #include <sys/mman.h>
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
19 #include <sys/time.h>
13839
40391656ae23 small compilation fix
rathann
parents: 3082
diff changeset
20 #include <inttypes.h>
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
21
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
22 //#define ARR_SIZE 100000
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
23 #define ARR_SIZE (1024*768*2)
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
24
27365
13c920fdf7f5 Change a bunch of video-output-specific preprocessor directives from a HAVE_
diego
parents: 26759
diff changeset
25 #ifdef CONFIG_MGA
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
26
17017
dd053aeb5442 Unify include path handling by using -I.
diego
parents: 13839
diff changeset
27 #include "drivers/mga_vid.h"
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
28
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
29 static mga_vid_config_t mga_vid_config;
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
30 static unsigned char* frame = NULL;
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
31 static int f;
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
32
28232
8df85ad26746 Add missing 'void' keyword to parameterless function declarations.
diego
parents: 27757
diff changeset
33 static int mga_init(void)
8df85ad26746 Add missing 'void' keyword to parameterless function declarations.
diego
parents: 27757
diff changeset
34 {
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
35 f = open("/dev/mga_vid", O_RDWR);
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
36 if (f == -1) {
28869
04a232eb5f0d comment/output cosmetics
diego
parents: 28868
diff changeset
37 fprintf(stderr, "Couldn't open /dev/mga_vid.\n");
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
38 return -1;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
39 }
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
40
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
41 mga_vid_config.num_frames = 1;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
42 mga_vid_config.frame_size = ARR_SIZE;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
43 mga_vid_config.format = MGA_VID_FORMAT_YUY2;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
44
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
45 mga_vid_config.colkey_on = 0;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
46 mga_vid_config.src_width = 640;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
47 mga_vid_config.src_height = 480;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
48 mga_vid_config.dest_width = 320;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
49 mga_vid_config.dest_height = 200;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
50 mga_vid_config.x_org = 0;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
51 mga_vid_config.y_org = 0;
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
52
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
53 mga_vid_config.version = MGA_VID_VERSION;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
54 if (ioctl(f, MGA_VID_CONFIG, &mga_vid_config)) {
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
55 perror("Error in mga_vid_config ioctl()");
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
56 printf("Your mga_vid driver version is incompatible with this MPlayer version!\n");
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
57 exit(1);
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
58 }
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
59 ioctl(f, MGA_VID_ON, 0);
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
60
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
61 frame = (char*)mmap(0, mga_vid_config.frame_size*mga_vid_config.num_frames,
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
62 PROT_WRITE,MAP_SHARED, f, 0);
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
63 if (!frame) {
28869
04a232eb5f0d comment/output cosmetics
diego
parents: 28868
diff changeset
64 printf("Can't mmap MGA frame.\n");
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
65 exit(1);
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
66 }
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
67
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
68 //clear the buffer
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
69 //memset(frames[0], 0x80, mga_vid_config.frame_size*mga_vid_config.num_frames);
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
70
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
71 return 0;
572
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
28232
8df85ad26746 Add missing 'void' keyword to parameterless function declarations.
diego
parents: 27757
diff changeset
77 static unsigned int GetTimer(void)
8df85ad26746 Add missing 'void' keyword to parameterless function declarations.
diego
parents: 27757
diff changeset
78 {
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
79 struct timeval tv;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
80 struct timezone tz;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
81 //float s;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
82 gettimeofday(&tv, &tz);
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
83 //s = tv.tv_usec; s *= 0.000001; s += tv.tv_sec;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
84 return tv.tv_sec * 1000000 + tv.tv_usec;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
85 }
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
86
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
87 static inline unsigned long long int read_tsc(void)
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
88 {
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
89 unsigned long long int retval;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
90 __asm__ volatile ("rdtsc":"=A" (retval)::"memory");
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
91 return retval;
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
92 }
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
93
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
94 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
95
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
96 int main(void)
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
97 {
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
98 unsigned long long int v1, v2;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
99 unsigned char *marr1, *marr2;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
100 int i;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
101 unsigned int t;
27365
13c920fdf7f5 Change a bunch of video-output-specific preprocessor directives from a HAVE_
diego
parents: 26759
diff changeset
102 #ifdef CONFIG_MGA
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
103 mga_init();
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
104 marr1 = &frame[3];
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
105 #else
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
106 marr1 = &arr1[0];
572
9935c60a15a0 added mga_vid support (systemram->videoram tests)
arpi_esp
parents: 562
diff changeset
107 #endif
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
108 marr2 = &arr2[0];
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
109
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
110 for (i = 0; i < ARR_SIZE - 16; i++)
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
111 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
112
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
113 t = GetTimer();
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
114 v1 = read_tsc();
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
115 for (i = 0; i < 100; i++)
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
116 memcpy(marr1, marr2, ARR_SIZE - 16);
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
117 v2 = read_tsc();
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
118 t = GetTimer() - t;
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
119 // ARR_SIZE*100 / (1024*1024) / (t/1000000) = ARR_SIZE*95.36743 / t
28869
04a232eb5f0d comment/output cosmetics
diego
parents: 28868
diff changeset
120 printf(NAME ": CPU clocks=%llu = %dus (%5.3ffps) %5.1fMB/s\n", v2-v1, t,
28868
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
121 100000000.0f/(float)t, (float)ARR_SIZE*95.36743f/(float)t);
a10ddae3de31 whitespace cosmetics:
diego
parents: 28855
diff changeset
122 return 0;
562
312fee2a6816 Initial release, used to benchmark fastmemcpy.h code from libvo.
atmosfear
parents:
diff changeset
123 }