Home | History | Annotate | Download | only in bench
      1 
      2 #define EIGEN_INTERNAL_DEBUG_CACHE_QUERY
      3 #include <iostream>
      4 #include "../Eigen/Core"
      5 
      6 using namespace Eigen;
      7 using namespace std;
      8 
      9 #define DUMP_CPUID(CODE) {\
     10   int abcd[4]; \
     11   abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;\
     12   EIGEN_CPUID(abcd, CODE, 0); \
     13   std::cout << "The code " << CODE << " gives " \
     14               << (int*)(abcd[0]) << " " << (int*)(abcd[1]) << " " \
     15               << (int*)(abcd[2]) << " " << (int*)(abcd[3]) << " " << std::endl; \
     16   }
     17 
     18 int main()
     19 {
     20   cout << "Eigen's L1    = " << internal::queryL1CacheSize() << endl;
     21   cout << "Eigen's L2/L3 = " << internal::queryTopLevelCacheSize() << endl;
     22   int l1, l2, l3;
     23   internal::queryCacheSizes(l1, l2, l3);
     24   cout << "Eigen's L1, L2, L3       = " << l1 << " " << l2 << " " << l3 << endl;
     25 
     26   #ifdef EIGEN_CPUID
     27 
     28   int abcd[4];
     29   int string[8];
     30   char* string_char = (char*)(string);
     31 
     32   // vendor ID
     33   EIGEN_CPUID(abcd,0x0,0);
     34   string[0] = abcd[1];
     35   string[1] = abcd[3];
     36   string[2] = abcd[2];
     37   string[3] = 0;
     38   cout << endl;
     39   cout << "vendor id = " << string_char << endl;
     40   cout << endl;
     41   int max_funcs = abcd[0];
     42 
     43   internal::queryCacheSizes_intel_codes(l1, l2, l3);
     44   cout << "Eigen's intel codes L1, L2, L3 = " << l1 << " " << l2 << " " << l3 << endl;
     45   if(max_funcs>=4)
     46   {
     47     internal::queryCacheSizes_intel_direct(l1, l2, l3);
     48     cout << "Eigen's intel direct L1, L2, L3 = " << l1 << " " << l2 << " " << l3 << endl;
     49   }
     50   internal::queryCacheSizes_amd(l1, l2, l3);
     51   cout << "Eigen's amd L1, L2, L3         = " << l1 << " " << l2 << " " << l3 << endl;
     52   cout << endl;
     53 
     54   // dump Intel direct method
     55   if(max_funcs>=4)
     56   {
     57     l1 = l2 = l3 = 0;
     58     int cache_id = 0;
     59     int cache_type = 0;
     60     do {
     61       abcd[0] = abcd[1] = abcd[2] = abcd[3] = 0;
     62       EIGEN_CPUID(abcd,0x4,cache_id);
     63       cache_type  = (abcd[0] & 0x0F) >> 0;
     64       int cache_level = (abcd[0] & 0xE0) >> 5;  // A[7:5]
     65       int ways        = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
     66       int partitions  = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
     67       int line_size   = (abcd[1] & 0x00000FFF) >>  0; // B[11:0]
     68       int sets        = (abcd[2]);                    // C[31:0]
     69       int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
     70 
     71       cout << "cache[" << cache_id << "].type       = " << cache_type << "\n";
     72       cout << "cache[" << cache_id << "].level      = " << cache_level << "\n";
     73       cout << "cache[" << cache_id << "].ways       = " << ways << "\n";
     74       cout << "cache[" << cache_id << "].partitions = " << partitions << "\n";
     75       cout << "cache[" << cache_id << "].line_size  = " << line_size << "\n";
     76       cout << "cache[" << cache_id << "].sets       = " << sets << "\n";
     77       cout << "cache[" << cache_id << "].size       = " << cache_size << "\n";
     78 
     79       cache_id++;
     80     } while(cache_type>0 && cache_id<16);
     81   }
     82 
     83   // dump everything
     84   std::cout << endl <<"Raw dump:" << endl;
     85   for(int i=0; i<max_funcs; ++i)
     86     DUMP_CPUID(i);
     87 
     88   DUMP_CPUID(0x80000000);
     89   DUMP_CPUID(0x80000001);
     90   DUMP_CPUID(0x80000002);
     91   DUMP_CPUID(0x80000003);
     92   DUMP_CPUID(0x80000004);
     93   DUMP_CPUID(0x80000005);
     94   DUMP_CPUID(0x80000006);
     95   DUMP_CPUID(0x80000007);
     96   DUMP_CPUID(0x80000008);
     97   #else
     98   cout << "EIGEN_CPUID is not defined" << endl;
     99   #endif
    100   return 0;
    101 }
    102