Revert "Honor cgroup/cpuset limits when enumerating cpus"

Use cpuid 4 with subleafs to query L1 cache size on Intel processors

CMakeLists.txt

@@ -5,8 +5,8 @@
 cmake_minimum_required(VERSION 2.8.5)
 project(OpenBLAS)
 set(OpenBLAS_MAJOR_VERSION 0)
-set(OpenBLAS_MINOR_VERSION 2)
-set(OpenBLAS_PATCH_VERSION 20)
+set(OpenBLAS_MINOR_VERSION 3)
+set(OpenBLAS_PATCH_VERSION 0.dev)
 set(OpenBLAS_VERSION "${OpenBLAS_MAJOR_VERSION}.${OpenBLAS_MINOR_VERSION}.${OpenBLAS_PATCH_VERSION}")
 
 enable_language(ASM)

Makefile.rule

@@ -3,7 +3,7 @@
 #
 
 # This library's version
-VERSION = 0.2.20
+VERSION = 0.3.0.dev
 
 # If you set the suffix, the library name will be libopenblas_$(LIBNAMESUFFIX).a
 # and libopenblas_$(LIBNAMESUFFIX).so. Meanwhile, the soname in shared library

cpuid_x86.c

@@ -71,12 +71,23 @@ void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx)
   *edx = cpuInfo[3];
 }
 
+void cpuid_count(int op, int count, int *eax, int *ebx, int *ecx, int *edx)
+{
+  int cpuInfo[4] = {-1};
+  __cpuidex(cpuInfo, op, count);
+  *eax = cpuInfo[0];
+  *ebx = cpuInfo[1];
+  *ecx = cpuInfo[2];
+  *edx = cpuInfo[3];
+}
+
 #else
 
 #ifndef CPUIDEMU
 
 #if defined(__APPLE__) && defined(__i386__)
 void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx);
+void cpuid_count(int op, int count, int *eax, int *ebx, int *ecx, int *edx);
 #else
 static C_INLINE void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx){
 #if defined(__i386__) && defined(__PIC__)
@@ -90,6 +101,19 @@ static C_INLINE void cpuid(int op, int *eax, int *ebx, int *ecx, int *edx){
     ("cpuid": "=a" (*eax), "=b" (*ebx), "=c" (*ecx), "=d" (*edx) : "a" (op) : "cc");
 #endif
 }
+
+static C_INLINE void cpuid_count(int op, int count ,int *eax, int *ebx, int *ecx, int *edx){
+#if defined(__i386__) && defined(__PIC__)
+  __asm__ __volatile__
+    ("mov %%ebx, %%edi;"
+     "cpuid;"
+     "xchgl %%ebx, %%edi;"
+     : "=a" (*eax), "=D" (*ebx), "=c" (*ecx), "=d" (*edx) : "0" (op), "2" (count) : "cc");
+#else
+  __asm__ __volatile__
+    ("cpuid": "=a" (*eax), "=b" (*ebx), "=c" (*ecx), "=d" (*edx) : "0" (op), "2" (count) : "cc");
+#endif
+}
 #endif
 
 #else
@@ -133,6 +157,10 @@ void cpuid(unsigned int op, unsigned int *eax, unsigned int *ebx, unsigned int *
   *edx = idlist[current].d;
 }
 
+void cpuid_count (unsigned int op, unsigned int count, unsigned int *eax, unsigned int *ebx, unsigned int *ecx, unsigned int *edx) {
+  return cpuid (op, eax, ebx, ecx, edx);
+}
+
 #endif
 
 #endif // _MSC_VER
@@ -312,9 +340,9 @@ int get_cacheinfo(int type, cache_info_t *cacheinfo){
   cpuid(0, &cpuid_level, &ebx, &ecx, &edx);
 
   if (cpuid_level > 1) {
-
+    int numcalls =0 ;
     cpuid(2, &eax, &ebx, &ecx, &edx);
-
+    numcalls = BITMASK(eax, 0, 0xff); //FIXME some systems may require repeated calls to read all entries
     info[ 0] = BITMASK(eax,  8, 0xff);
     info[ 1] = BITMASK(eax, 16, 0xff);
     info[ 2] = BITMASK(eax, 24, 0xff);
@@ -335,7 +363,6 @@ int get_cacheinfo(int type, cache_info_t *cacheinfo){
     info[14] = BITMASK(edx, 24, 0xff);
 
     for (i = 0; i < 15; i++){
-
       switch (info[i]){
 
 	/* This table is from http://www.sandpile.org/ia32/cpuid.htm */
@@ -637,12 +664,13 @@ int get_cacheinfo(int type, cache_info_t *cacheinfo){
 	LD1.linesize    = 64;
 	break;
       case 0x63 :
-  DTB.size        = 2048;
-  DTB.associative = 4;
-  DTB.linesize    = 32;
-  LDTB.size       = 4096;
-  LDTB.associative= 4;
-  LDTB.linesize   = 32;
+  	DTB.size        = 2048;
+  	DTB.associative = 4;
+  	DTB.linesize    = 32;
+  	LDTB.size       = 4096;
+  	LDTB.associative= 4;
+  	LDTB.linesize   = 32;
+	break;
       case 0x66 :
 	LD1.size        = 8;
 	LD1.associative = 4;
@@ -675,12 +703,13 @@ int get_cacheinfo(int type, cache_info_t *cacheinfo){
 	LC1.associative = 8;
 	break;
       case 0x76 :
-  ITB.size        = 2048;
-  ITB.associative = 0;
-  ITB.linesize    = 8;
-  LITB.size       = 4096;
-  LITB.associative= 0;
-  LITB.linesize   = 8;
+  	ITB.size        = 2048;
+  	ITB.associative = 0;
+  	ITB.linesize    = 8;
+  	LITB.size       = 4096;
+  	LITB.associative= 0;
+  	LITB.linesize   = 8;
+	break;
       case 0x77 :
 	LC1.size        = 16;
 	LC1.associative = 4;
@@ -891,6 +920,67 @@ int get_cacheinfo(int type, cache_info_t *cacheinfo){
   }
 
   if (get_vendor() == VENDOR_INTEL) {
+      if(LD1.size<=0 || LC1.size<=0){
+	//If we didn't detect L1 correctly before,
+	int count;
+	for (count=0;count <4;count++) {
+	cpuid_count(4, count, &eax, &ebx, &ecx, &edx);
+        switch (eax &0x1f) {
+        case 0:
+          continue;
+          case 1:
+          case 3:
+          {
+            switch ((eax >>5) &0x07)
+            {
+            case 1:
+            {
+//            fprintf(stderr,"L1 data cache...\n");
+            int sets = ecx+1;
+            int lines = (ebx & 0x0fff) +1;
+            ebx>>=12;
+            int part = (ebx&0x03ff)+1;
+            ebx >>=10;
+            int assoc = (ebx&0x03ff)+1;
+            LD1.size = (assoc*part*lines*sets)/1024;
+            LD1.associative = assoc;
+            LD1.linesize= lines;
+            break;
+            }
+            default: 
+              break;
+           }
+          break;
+          }
+         case 2:
+          {
+            switch ((eax >>5) &0x07)
+            {
+            case 1:
+            {
+//            fprintf(stderr,"L1 instruction cache...\n");
+            int sets = ecx+1;
+            int lines = (ebx & 0x0fff) +1;
+            ebx>>=12;
+            int part = (ebx&0x03ff)+1;
+            ebx >>=10;
+            int assoc = (ebx&0x03ff)+1;
+            LC1.size = (assoc*part*lines*sets)/1024;
+            LC1.associative = assoc;
+            LC1.linesize= lines;
+            break;
+            }
+            default: 
+              break;
+           }
+          break;
+          
+          }
+          default:
+          break;
+        }
+      }
+    }
     cpuid(0x80000000, &cpuid_level, &ebx, &ecx, &edx);
     if (cpuid_level >= 0x80000006) {
       if(L2.size<=0){

driver/others/init.c

@@ -778,11 +778,11 @@ static int initialized = 0;
 void gotoblas_affinity_init(void) {
 
   int cpu, num_avail;
-#ifndef USE_OPENMP	
+#ifndef USE_OPENMP
   cpu_set_t cpu_mask;
 #endif
   int i;
-	
+
   if (initialized) return;
 
   initialized = 1;
@@ -826,54 +826,15 @@ void gotoblas_affinity_init(void) {
   common -> shmid = pshmid;
 
   if (common -> magic != SH_MAGIC) {
-    cpu_set_t *cpusetp;
-    int nums;
-    int ret;
-
 #ifdef DEBUG
     fprintf(stderr, "Shared Memory Initialization.\n");
 #endif
 
     //returns the number of processors which are currently online
-
-    nums = sysconf(_SC_NPROCESSORS_CONF);
-     
-#if !defined(__GLIBC_PREREQ) || !__GLIBC_PREREQ(2, 3)
-    common->num_procs = nums;
-#elif __GLIBC_PREREQ(2, 7)
-    cpusetp = CPU_ALLOC(nums);
-    if (cpusetp == NULL) {
-        common->num_procs = nums;
-    } else {
-        size_t size;
-        size = CPU_ALLOC_SIZE(nums);
-        ret = sched_getaffinity(0,size,cpusetp);
-        if (ret!=0) 
-            common->num_procs = nums;
-        else
-            common->num_procs = CPU_COUNT_S(size,cpusetp);
-    }
-    CPU_FREE(cpusetp);
-#else
-    ret = sched_getaffinity(0,sizeof(cpu_set_t), cpusetp);
-    if (ret!=0) {
-        common->num_procs = nums;
-    } else {
-#if !__GLIBC_PREREQ(2, 6)  
-    int i;
-    int n = 0;
-    for (i=0;i<nums;i++)
-        if (CPU_ISSET(i,cpusetp)) n++;
-    common->num_procs = n;
-    }
-#else
-    common->num_procs = CPU_COUNT(sizeof(cpu_set_t),cpusetp);
-#endif
-
-#endif 
+    common -> num_procs = sysconf(_SC_NPROCESSORS_CONF);;
 
     if(common -> num_procs > MAX_CPUS) {
-      fprintf(stderr, "\nOpenBLAS Warning : The number of CPU/Cores(%d) is beyond the limit(%d). Terminated.\n", common->num_procs, MAX_CPUS);
+      fprintf(stderr, "\nOpenBLAS Warining : The number of CPU/Cores(%d) is beyond the limit(%d). Terminated.\n", common->num_procs, MAX_CPUS);
       exit(1);
     }
 
@@ -886,7 +847,7 @@ void gotoblas_affinity_init(void) {
     if (common -> num_nodes > 1) numa_mapping();
 
     common -> final_num_procs = 0;
-    for(i = 0; i < common -> avail_count; i++) common -> final_num_procs += rcount(common -> avail[i]) + 1;   //Make the max cpu number.
+    for(i = 0; i < common -> avail_count; i++) common -> final_num_procs += rcount(common -> avail[i]) + 1;   //Make the max cpu number. 
 
     for (cpu = 0; cpu < common -> final_num_procs; cpu ++) common -> cpu_use[cpu] =  0;
 

driver/others/memory.c

@@ -175,44 +175,7 @@ int get_num_procs(void);
 #else
 int get_num_procs(void) {
   static int nums = 0;
-cpu_set_t *cpusetp;
-size_t size;
-int ret;
-int i,n;
-
   if (!nums) nums = sysconf(_SC_NPROCESSORS_CONF);
-#if !defined(OS_LINUX)
-     return nums;
-#endif
-     
-#if !defined(__GLIBC_PREREQ)
-   return nums;
-#endif   
-#if !__GLIBC_PREREQ(2, 3)
-   return nums;
-#endif   
-
-#if !__GLIBC_PREREQ(2, 7)
-  ret = sched_getaffinity(0,sizeof(cpu_set_t), cpusetp);
-  if (ret!=0) return nums;
-  n=0;
-#if !__GLIBC_PREREQ(2, 6)  
-  for (i=0;i<nums;i++)
-     if (CPU_ISSET(i,cpusetp)) n++;
-  nums=n;   
-#else
-  nums = CPU_COUNT(sizeof(cpu_set_t),cpusetp);
-#endif
-  return nums;
-#endif
-
-  cpusetp = CPU_ALLOC(nums);
-  if (cpusetp == NULL) return nums;
-  size = CPU_ALLOC_SIZE(nums);
-  ret = sched_getaffinity(0,size,cpusetp);
-  if (ret!=0) return nums;
-  nums = CPU_COUNT_S(size,cpusetp);
-  CPU_FREE(cpusetp);
   return nums;
 }
 #endif