TOOLS/fastmemcpybench.c

   1 /*
   2  * benchmark tool for fast_memcpy code from libvo
   3  *
   4  * NOTE: This code can not be used on Pentium MMX / II because they contain
   5  * a bug in rdtsc. For Intel processors since P6(PII) rdpmc should be used
   6  * instead. For PIII it's disputable and it seems the bug was fixed but this
   7  * was not confirmed through testing.
   8 */
   9
  10 #include <stdio.h>
  11 #include <stdlib.h>
  12 #include <string.h>
  13 #include <sys/ioctl.h>
  14 #include <unistd.h>
  15 #include <fcntl.h>
  16 #include <sys/mman.h>
  17 #include <sys/time.h>
  18 #include <inttypes.h>
  19
  20 #include "config.h"
  21 #include "cpudetect.h"
  22
  23 #define BLOCK_SIZE 4096
  24 #define CONFUSION_FACTOR 0
  25
  26 #if HAVE_MMX
  27 #define COMPILE_MMX
  28 #endif
  29
  30 #if HAVE_MMX2
  31 #define COMPILE_MMX2
  32 #endif
  33
  34 #if HAVE_AMD3DNOW
  35 #define COMPILE_AMD3DNOW
  36 #endif
  37
  38 #if HAVE_SSE
  39 #define COMPILE_SSE
  40 #endif
  41
  42 #ifdef COMPILE_MMX
  43 #undef RENAME
  44 #undef HAVE_MMX
  45 #undef HAVE_MMX2
  46 #undef HAVE_AMD3DNOW
  47 #undef HAVE_SSE
  48 #undef HAVE_SSE2
  49 #define HAVE_MMX 1
  50 #define HAVE_MMX2 0
  51 #define HAVE_AMD3DNOW 0
  52 #define HAVE_SSE 0
  53 #define HAVE_SSE2 0
  54 #define RENAME(a) a ## _MMX
  55 #include "libvo/aclib_template.c"
  56 #endif
  57
  58 #ifdef COMPILE_MMX2
  59 #undef RENAME
  60 #undef HAVE_MMX
  61 #undef HAVE_MMX2
  62 #undef HAVE_AMD3DNOW
  63 #undef HAVE_SSE
  64 #undef HAVE_SSE2
  65 #define HAVE_MMX 1
  66 #define HAVE_MMX2 1
  67 #define HAVE_AMD3DNOW 0
  68 #define HAVE_SSE 0
  69 #define HAVE_SSE2 0
  70 #define RENAME(a) a ## _MMX2
  71 #include "libvo/aclib_template.c"
  72 #endif
  73
  74 #ifdef COMPILE_AMD3DNOW
  75 #undef RENAME
  76 #undef HAVE_MMX
  77 #undef HAVE_MMX2
  78 #undef HAVE_AMD3DNOW
  79 #undef HAVE_SSE
  80 #undef HAVE_SSE2
  81 #define HAVE_MMX 1
  82 #define HAVE_MMX2 0
  83 #define HAVE_AMD3DNOW 1
  84 #define HAVE_SSE 0
  85 #define HAVE_SSE2 0
  86 #define RENAME(a) a ## _3DNow
  87 #include "libvo/aclib_template.c"
  88 #endif
  89
  90 #ifdef COMPILE_SSE
  91 #undef RENAME
  92 #undef HAVE_MMX
  93 #undef HAVE_MMX2
  94 #undef HAVE_AMD3DNOW
  95 #undef HAVE_SSE
  96 #undef HAVE_SSE2
  97 #define HAVE_MMX 1
  98 #define HAVE_MMX2 1
  99 #define HAVE_AMD3DNOW 0
 100 #define HAVE_SSE 1
 101 #define HAVE_SSE2 1
 102 #define RENAME(a) a ## _SSE
 103 #include "libvo/aclib_template.c"
 104 #endif
 105
 106 //#define ARR_SIZE 100000
 107 #define ARR_SIZE (1024*768*2)
 108
 109 #ifdef CONFIG_MGA
 110
 111 #include "drivers/mga_vid.h"
 112
 113 static mga_vid_config_t mga_vid_config;
 114 static unsigned char* frame = NULL;
 115 static int f;
 116
 117 static int mga_init(void)
 118 {
 119     f = open("/dev/mga_vid", O_RDWR);
 120     if (f == -1) {
 121         fprintf(stderr, "Couldn't open /dev/mga_vid.\n");
 122         return -1;
 123     }
 124
 125     mga_vid_config.num_frames = 1;
 126     mga_vid_config.frame_size = ARR_SIZE;
 127     mga_vid_config.format     = MGA_VID_FORMAT_YUY2;
 128
 129     mga_vid_config.colkey_on   =   0;
 130     mga_vid_config.src_width   = 640;
 131     mga_vid_config.src_height  = 480;
 132     mga_vid_config.dest_width  = 320;
 133     mga_vid_config.dest_height = 200;
 134     mga_vid_config.x_org       =   0;
 135     mga_vid_config.y_org       =   0;
 136
 137     mga_vid_config.version     = MGA_VID_VERSION;
 138     if (ioctl(f, MGA_VID_CONFIG, &mga_vid_config)) {
 139         perror("Error in mga_vid_config ioctl()");
 140         printf("Your mga_vid driver version is incompatible with this MPlayer version!\n");
 141         exit(1);
 142     }
 143     ioctl(f, MGA_VID_ON, 0);
 144
 145     frame = (char*)mmap(0, mga_vid_config.frame_size*mga_vid_config.num_frames,
 146                         PROT_WRITE,MAP_SHARED, f, 0);
 147     if (!frame) {
 148         printf("Can't mmap MGA frame.\n");
 149         exit(1);
 150     }
 151
 152     //clear the buffer
 153     //memset(frames[0], 0x80, mga_vid_config.frame_size*mga_vid_config.num_frames);
 154
 155     return 0;
 156 }
 157
 158 #endif
 159
 160 // Returns current time in microseconds
 161 static unsigned int GetTimer(void)
 162 {
 163     struct timeval tv;
 164     struct timezone tz;
 165     //float s;
 166     gettimeofday(&tv, &tz);
 167     //s = tv.tv_usec; s *= 0.000001; s += tv.tv_sec;
 168     return tv.tv_sec * 1000000 + tv.tv_usec;
 169 }
 170
 171 static inline unsigned long long int read_tsc(void)
 172 {
 173     unsigned long long int retval;
 174     __asm__ volatile ("rdtsc":"=A" (retval)::"memory");
 175     return retval;
 176 }
 177
 178 unsigned char __attribute__((aligned(4096)))arr1[ARR_SIZE], arr2[ARR_SIZE];
 179
 180 int main(void)
 181 {
 182     unsigned long long int v1, v2;
 183     unsigned char *marr1, *marr2;
 184     int i;
 185     unsigned int t;
 186 #ifdef CONFIG_MGA
 187     mga_init();
 188     marr1 = &frame[3];
 189 #else
 190     marr1 = &arr1[0];
 191 #endif
 192     marr2 = &arr2[0];
 193
 194     for (i = 0; i < ARR_SIZE - 16; i++)
 195         marr1[i] = marr2[i] = i;
 196
 197 #define testblock(func, name)                                                \
 198     t  = GetTimer();                                                         \
 199     v1 = read_tsc();                                                         \
 200     for (i = 0; i < 100; i++)                                                \
 201         func(marr1, marr2, ARR_SIZE - 16);                                   \
 202     v2 = read_tsc();                                                         \
 203     t  = GetTimer() - t;                                                     \
 204     /* ARR_SIZE*100 / (1024*1024) / (t/1000000) = ARR_SIZE*95.36743 / t */   \
 205     printf(name "CPU clocks=%llu = %dus  (%5.3ffps)  %5.1fMB/s\n", v2-v1, t, \
 206            100000000.0f / (float)t, (float)ARR_SIZE*95.36743f / (float)t);
 207
 208     testblock(memcpy, "libc:   ");
 209
 210 #if HAVE_MMX
 211     testblock(fast_memcpy_MMX, "MMX:    ");
 212 #endif
 213
 214 #if HAVE_AMD3DNOW
 215     testblock(fast_memcpy_3DNow, "3DNow!: ");
 216 #endif
 217
 218 #if HAVE_MMX2
 219     testblock(fast_memcpy_MMX2, "MMX2:   ");
 220 #endif
 221
 222 #if HAVE_SSE
 223     testblock(fast_memcpy_SSE, "SSE:    ");
 224 #endif
 225
 226     return 0;
 227 }