attachment:memspeed.c von AMD64

   1 /*
   2 gcc -O2 memspeed.c -o memspeed
   3 */
   4 
   5 #include <stdlib.h>
   6 #include <stdio.h>
   7 
   8 #define SIZEMB (32)
   9 #define SIZE (1024*1024*SIZEMB)
  10 #define ITER (20)
  11 #define ITER2 (10)
  12 
  13 double this_gettime(void)
  14 {
  15     struct timeval tv;
  16     double time;
  17     
  18     gettimeofday(&tv, 0);
  19     time = tv.tv_sec;
  20     time += ((double)tv.tv_usec)/1000000.0;
  21     
  22     return time;
  23 }
  24 
  25 
  26 inline void prefetch(void *x)
  27 {
  28         asm volatile("prefetcht0 %0\n\t"
  29 	:: "m" (*(unsigned long *)x));
  30 }
  31 
  32 
  33 int main(int argc, char *argv[])
  34 {
  35     int iter, i, j;
  36     void *aptr;
  37     
  38     float *a, *b;
  39     volatile float read;
  40     
  41     double *da, *db;
  42     double t1, t2, t3, t4;
  43     volatile double dread;
  44     
  45     double time1, time2;
  46     
  47     aptr = malloc(SIZE*sizeof(float) + 64);
  48     a = (float *) ((long int)(aptr+((long int)63))& ((long int)~0x3f));
  49     
  50     aptr = malloc(SIZE*sizeof(double) + 64);
  51     da = (double *) ((long int)(aptr+((long int)63))& ((long int)~0x3f));
  52     
  53     printf("readbuffer alignment: %d %d\n", ((long int)a)&0x3f, ((long int)da)&0x3f);
  54 
  55     for(j = 0; j < SIZE; j++)
  56     {
  57 	a[j] = 0.0;
  58 	da[j] = 0.0;
  59     }
  60     
  61     
  62     for(iter = 0; iter < ITER; iter++)
  63     {
  64 	time1 = this_gettime();
  65 	for(j = 0; j < (SIZE/16); j++)
  66 	{
  67 	    prefetch(a+128);
  68 	    read = a[0];
  69 	    read = a[1];
  70 	    read = a[2];
  71 	    read = a[3];
  72 	    read = a[4];
  73 	    read = a[5];
  74 	    read = a[6];
  75 	    read = a[7];
  76 	    read = a[8];
  77 	    read = a[9];
  78 	    read = a[10];
  79 	    read = a[11];
  80 	    read = a[12];
  81 	    read = a[13];
  82 	    read = a[14];
  83 	    read = a[15];
  84 	    a += 16;
  85 	}
  86 	time2 = this_gettime();
  87 	printf("c int: %f mb/s\n", ((double)(SIZEMB*4))/(time2-time1));
  88 	a -= SIZE;
  89     }
  90 
  91 
  92 
  93     for(iter = 0; iter < ITER; iter++)
  94     {
  95 	time1 = this_gettime();
  96 	for(i = 0; i < ITER2; i++)
  97 	{
  98 	    for(j = 0; j < (SIZE/64); j++)
  99 	    {
 100 	    
 101     	    asm volatile(
 102 		"prefetcht0  0x800(%0)		\n\t"
 103 		"movaps 0x00(%0), %%xmm0	;\n\t"
 104 		"movaps 0x10(%0), %%xmm1	;\n\t"
 105 		"movaps 0x20(%0), %%xmm2	;\n\t"
 106 		"movaps 0x30(%0), %%xmm3	;\n\t"
 107 		"prefetcht0  0x840(%0)		\n\t"
 108 		"movaps 0x40(%0), %%xmm4	;\n\t"
 109 		"movaps 0x50(%0), %%xmm5	;\n\t"
 110 		"movaps 0x60(%0), %%xmm6	;\n\t"
 111 		"movaps 0x70(%0), %%xmm7	;\n\t"
 112 		"prefetcht0  0x880(%0)		\n\t"
 113 		"movaps 0x80(%0), %%xmm8	;\n\t"
 114 		"movaps 0x90(%0), %%xmm9	;\n\t"
 115 		"movaps 0xa0(%0), %%xmm10	;\n\t"
 116 		"movaps 0xb0(%0), %%xmm11	;\n\t"
 117 		"prefetcht0  0x8c0(%0)		\n\t"
 118 		"movaps 0xc0(%0), %%xmm12	;\n\t"
 119 		"movaps 0xd0(%0), %%xmm13	;\n\t"
 120 		"movaps 0xe0(%0), %%xmm14	;\n\t"
 121 		"movaps 0xf0(%0), %%xmm15	;\n\t"
 122 		"prefetcht0  0x900(%0)		\n\t"
 123 		"movaps 0x100(%0), %%xmm0	;\n\t"
 124 		"movaps 0x110(%0), %%xmm1	;\n\t"
 125 		"movaps 0x120(%0), %%xmm2	;\n\t"
 126 		"movaps 0x130(%0), %%xmm3	;\n\t"
 127 		"prefetcht0  0x940(%0)		\n\t"
 128 		"movaps 0x140(%0), %%xmm4	;\n\t"
 129 		"movaps 0x150(%0), %%xmm5	;\n\t"
 130 		"movaps 0x160(%0), %%xmm6	;\n\t"
 131 		"movaps 0x170(%0), %%xmm7	;\n\t"
 132 		"prefetcht0  0x980(%0)		\n\t"
 133 		"movaps 0x180(%0), %%xmm8	;\n\t"
 134 		"movaps 0x190(%0), %%xmm9	;\n\t"
 135 		"movaps 0x1a0(%0), %%xmm10	;\n\t"
 136 		"movaps 0x1b0(%0), %%xmm11	;\n\t"
 137 		"prefetcht0  0x9c0(%0)		\n\t"
 138 		"movaps 0x1c0(%0), %%xmm12	;\n\t"
 139 		"movaps 0x1d0(%0), %%xmm13	;\n\t"
 140 		"movaps 0x1e0(%0), %%xmm14	;\n\t"
 141 		"movaps 0x1f0(%0), %%xmm15	;\n\t"
 142 	    : : "r" ((unsigned long *)da));
 143 	    
 144 	    da += 64;
 145 
 146 	    }
 147 	    da -= SIZE;
 148 	}
 149 	
 150 	time2 = this_gettime();
 151 	printf("asm xmms: %f mb/s\n", ((double)(SIZEMB*8*ITER2))/(time2-time1));
 152     }
 153 
 154 /*    printf("xmms: %f mb/s\n", ((double)(SIZEMB*8*ITER))/(time2-time1)); */
 155 }