GitHub-Proxy/eigen
1
0
Fork 0
mirror of https://gitlab.com/libeigen/eigen.git synced 2025-08-12 11:49:02 +08:00
Code Issues Packages Projects Releases Wiki Activity

default to Intel's API by default

Browse Source
This commit is contained in:
Gael Guennebaud 2010-06-23 17:14:06 +02:00
parent e1a6bad087
commit 546587c7d3
1 changed files with 69 additions and 47 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

114
Eigen/src/Core/util/Memory.h
View File

@ -599,6 +599,52 @@ public:
# define EIGEN_CPUID(abcd,func) __cpuid((int*)abcd,func) # define EIGEN_CPUID(abcd,func) __cpuid((int*)abcd,func)
#endif #endif
inline bool ei_cpuid_is_vendor(int abcd[4], const char* vendor)
{
return abcd[1]==((int*)(vendor))[0] && abcd[3]==((int*)(vendor))[1] && abcd[2]==((int*)(vendor))[2];
}
inline void ei_queryCacheSizes_intel(int& l1, int& l2, int& l3)
{
int abcd[4];
l1 = l2 = l3 = 0;
int cache_id = 0;
int cache_type = 0;
do {
EIGEN_CPUID(abcd,0x4,cache_id);
cache_type = (abcd[0] & 0x0F) >> 0;
if(cache_type==1||cache_type==3) // data or unified cache
{
int cache_level = (abcd[0] & 0xE0) >> 5; // A[7:5]
int ways = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
int partitions = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
int line_size = (abcd[1] & 0x00000FFF) >> 0; // B[11:0]
int sets = (abcd[2]); // C[31:0]
int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1);
switch(cache_level)
{
case 1: l1 = cache_size; break;
case 2: l2 = cache_size; break;
case 3: l3 = cache_size; break;
default: break;
}
}
cache_id++;
} while(cache_type>0);
}
inline void ei_queryCacheSizes_amd(int& l1, int& l2, int& l3)
{
int abcd[4];
EIGEN_CPUID(abcd,0x80000005,0);
l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
EIGEN_CPUID(abcd,0x80000006,0);
l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
}
/** \internal /** \internal
* Queries and returns the cache sizes in Bytes of the L1, L2, and L3 data caches respectively */ * Queries and returns the cache sizes in Bytes of the L1, L2, and L3 data caches respectively */
inline void ei_queryCacheSizes(int& l1, int& l2, int& l3) inline void ei_queryCacheSizes(int& l1, int& l2, int& l3)
@ -606,54 +652,30 @@ inline void ei_queryCacheSizes(int& l1, int& l2, int& l3)
#ifdef EIGEN_CPUID #ifdef EIGEN_CPUID
int abcd[4]; int abcd[4];
const char GenuineIntel_char[] = "GenuntelineI"; // identify the CPU vendor
const int* GenuineIntel = (int*)GenuineIntel_char;
const char AuthenticAMD_char[] = "AuthcAMDenti";
const int* AuthenticAMD = (int*)AuthenticAMD_char;
// Step 1: identify the CPU model
EIGEN_CPUID(abcd,0x0,0); EIGEN_CPUID(abcd,0x0,0);
if(abcd[1]==GenuineIntel[0] && abcd[2]==GenuineIntel[1] && abcd[3]==GenuineIntel[2]) //if(abcd[1]==GenuineIntel[0] && abcd[2]==GenuineIntel[1] && abcd[3]==GenuineIntel[2])
{ if(ei_cpuid_is_vendor(abcd,"GenuineIntel"))
// use Intel's cpuid API ei_queryCacheSizes_intel(l1,l2,l3);
l1 = l2 = l3 = 0; else if(ei_cpuid_is_vendor(abcd,"AuthenticAMD") || ei_cpuid_is_vendor(abcd,"AMDisbetter!"))
int cache_id = 0; ei_queryCacheSizes_amd(l1,l2,l3);
int cache_type = 0; else
do { // by default let's use Intel's API
EIGEN_CPUID(abcd,0x4,cache_id); ei_queryCacheSizes_intel(l1,l2,l3);
cache_type = (abcd[0] & 0x0F) >> 0;
if(cache_type==1||cache_type==3) // data or unified cache
{
int cache_level = (abcd[0] & 0xE0) >> 5; // A[7:5]
int ways = (abcd[1] & 0xFFC00000) >> 22; // B[31:22]
int partitions = (abcd[1] & 0x003FF000) >> 12; // B[21:12]
int line_size = (abcd[1] & 0x00000FFF) >> 0; // B[11:0]
int sets = (abcd[2]); // C[31:0]
int cache_size = (ways+1) * (partitions+1) * (line_size+1) * (sets+1); // here is the list of other vendors:
// ||ei_cpuid_is_vendor(abcd,"VIA VIA VIA ")
switch(cache_level) // ||ei_cpuid_is_vendor(abcd,"CyrixInstead")
{ // ||ei_cpuid_is_vendor(abcd,"CentaurHauls")
case 1: l1 = cache_size; break; // ||ei_cpuid_is_vendor(abcd,"GenuineTMx86")
case 2: l2 = cache_size; break; // ||ei_cpuid_is_vendor(abcd,"TransmetaCPU")
case 3: l3 = cache_size; break; // ||ei_cpuid_is_vendor(abcd,"RiseRiseRise")
default: break; // ||ei_cpuid_is_vendor(abcd,"Geode by NSC")
} // ||ei_cpuid_is_vendor(abcd,"SiS SiS SiS ")
} // ||ei_cpuid_is_vendor(abcd,"UMC UMC UMC ")
cache_id++; // ||ei_cpuid_is_vendor(abcd,"NexGenDriven")
} while(cache_type>0); // ||ei_cpuid_is_vendor(abcd,"CentaurHauls")
} // ||ei_cpuid_is_vendor(abcd,"CentaurHauls")
else if(abcd[1]==AuthenticAMD[0] && abcd[2]==AuthenticAMD[1] && abcd[3]==AuthenticAMD[2])
{
// use AMD's cpuid API
EIGEN_CPUID(abcd,0x80000005,0);
l1 = (abcd[2] >> 24) * 1024; // C[31:24] = L1 size in KB
EIGEN_CPUID(abcd,0x80000006,0);
l2 = (abcd[2] >> 16) * 1024; // C[31;16] = l2 cache size in KB
l3 = ((abcd[3] & 0xFFFC000) >> 18) * 512 * 1024; // D[31;18] = l3 cache size in 512KB
}
// TODO support other vendors
#endif #endif
} }