[openmp] porting affinity feature to netbsd. (#84618)

netbsd supports the portable hwloc's layer as well. for a hardware with
4 cpus, a cpu set is 4 and maxcpus is 256.

Author: devnexen

openmp/runtime/src/kmp.h

@@ -3912,7 +3912,7 @@ extern void __kmp_balanced_affinity(kmp_info_t *th, int team_size);
 #if KMP_WEIGHTED_ITERATIONS_SUPPORTED
 extern int __kmp_get_first_osid_with_ecore(void);
 #endif
-#if KMP_OS_LINUX || KMP_OS_FREEBSD
+#if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD
 extern int kmp_set_thread_affinity_mask_initial(void);
 #endif
 static inline void __kmp_assign_root_init_mask() {

openmp/runtime/src/kmp_affinity.cpp

@@ -2828,7 +2828,8 @@ static void __kmp_dispatch_set_hierarchy_values() {
   __kmp_hier_max_units[kmp_hier_layer_e::LAYER_THREAD + 1] =
       nPackages * nCoresPerPkg * __kmp_nThreadsPerCore;
   __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L1 + 1] = __kmp_ncores;
-#if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS) &&   \
+#if KMP_ARCH_X86_64 &&                                                         \
+    (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_WINDOWS) &&     \
     KMP_MIC_SUPPORTED
   if (__kmp_mic_type >= mic3)
     __kmp_hier_max_units[kmp_hier_layer_e::LAYER_L2 + 1] = __kmp_ncores / 2;
@@ -2843,7 +2844,8 @@ static void __kmp_dispatch_set_hierarchy_values() {
   __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_THREAD + 1] = 1;
   __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L1 + 1] =
       __kmp_nThreadsPerCore;
-#if KMP_ARCH_X86_64 && (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_WINDOWS) &&   \
+#if KMP_ARCH_X86_64 &&                                                         \
+    (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD || KMP_OS_WINDOWS) &&     \
     KMP_MIC_SUPPORTED
   if (__kmp_mic_type >= mic3)
     __kmp_hier_threads_per[kmp_hier_layer_e::LAYER_L2 + 1] =
@@ -5557,7 +5559,7 @@ void __kmp_balanced_affinity(kmp_info_t *th, int nthreads) {
   }
 }
 
-#if KMP_OS_LINUX || KMP_OS_FREEBSD
+#if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD
 // We don't need this entry for Windows because
 // there is GetProcessAffinityMask() api
 //

openmp/runtime/src/kmp_affinity.h

@@ -191,7 +191,7 @@ class KMPHwlocAffinity : public KMPAffinity {
 };
 #endif /* KMP_USE_HWLOC */
 
-#if KMP_OS_LINUX || KMP_OS_FREEBSD
+#if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD
 #if KMP_OS_LINUX
 /* On some of the older OS's that we build on, these constants aren't present
    in <asm/unistd.h> #included from <sys.syscall.h>. They must be the same on
@@ -314,6 +314,9 @@ class KMPHwlocAffinity : public KMPAffinity {
 #elif KMP_OS_FREEBSD
 #include <pthread.h>
 #include <pthread_np.h>
+#elif KMP_OS_NETBSD
+#include <pthread.h>
+#include <sched.h>
 #endif
 class KMPNativeAffinity : public KMPAffinity {
   class Mask : public KMPAffinity::Mask {
@@ -407,7 +410,7 @@ class KMPNativeAffinity : public KMPAffinity {
 #if KMP_OS_LINUX
       long retval =
           syscall(__NR_sched_getaffinity, 0, __kmp_affin_mask_size, mask);
-#elif KMP_OS_FREEBSD
+#elif KMP_OS_FREEBSD || KMP_OS_NETBSD
       int r = pthread_getaffinity_np(pthread_self(), __kmp_affin_mask_size,
                                      reinterpret_cast<cpuset_t *>(mask));
       int retval = (r == 0 ? 0 : -1);
@@ -428,7 +431,7 @@ class KMPNativeAffinity : public KMPAffinity {
 #if KMP_OS_LINUX
       long retval =
           syscall(__NR_sched_setaffinity, 0, __kmp_affin_mask_size, mask);
-#elif KMP_OS_FREEBSD
+#elif KMP_OS_FREEBSD || KMP_OS_NETBSD
       int r = pthread_setaffinity_np(pthread_self(), __kmp_affin_mask_size,
                                      reinterpret_cast<cpuset_t *>(mask));
       int retval = (r == 0 ? 0 : -1);
@@ -471,7 +474,7 @@ class KMPNativeAffinity : public KMPAffinity {
   }
   api_type get_api_type() const override { return NATIVE_OS; }
 };
-#endif /* KMP_OS_LINUX || KMP_OS_FREEBSD */
+#endif /* KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD */
 
 #if KMP_OS_WINDOWS
 class KMPNativeAffinity : public KMPAffinity {

openmp/runtime/src/kmp_os.h

@@ -75,7 +75,8 @@
 #error Unknown compiler
 #endif
 
-#if (KMP_OS_LINUX || KMP_OS_WINDOWS || KMP_OS_FREEBSD) && !KMP_OS_WASI
+#if (KMP_OS_LINUX || KMP_OS_WINDOWS || KMP_OS_FREEBSD || KMP_OS_NETBSD) &&     \
+    !KMP_OS_WASI
 #define KMP_AFFINITY_SUPPORTED 1
 #if KMP_OS_WINDOWS && KMP_ARCH_X86_64
 #define KMP_GROUP_AFFINITY 1

openmp/runtime/src/kmp_runtime.cpp

@@ -5376,7 +5376,7 @@ __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
           __kmp_reinitialize_team(team, new_icvs, NULL);
         }
 
-#if (KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED
+#if (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD) && KMP_AFFINITY_SUPPORTED
         /* Temporarily set full mask for primary thread before creation of
            workers. The reason is that workers inherit the affinity from the
            primary thread, so if a lot of workers are created on the single
@@ -5412,7 +5412,7 @@ __kmp_allocate_team(kmp_root_t *root, int new_nproc, int max_nproc,
           }
         }
 
-#if (KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED
+#if (KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD) && KMP_AFFINITY_SUPPORTED
         /* Restore initial primary thread's affinity mask */
         new_temp_affinity.restore();
 #endif

openmp/runtime/src/z_Linux_util.cpp

@@ -65,6 +65,9 @@
 #elif KMP_OS_NETBSD || KMP_OS_OPENBSD
 #include <sys/types.h>
 #include <sys/sysctl.h>
+#if KMP_OS_NETBSD
+#include <sched.h>
+#endif
 #elif KMP_OS_SOLARIS
 #include <libproc.h>
 #include <procfs.h>
@@ -122,7 +125,8 @@ static void __kmp_print_cond(char *buffer, kmp_cond_align_t *cond) {
 }
 #endif
 
-#if ((KMP_OS_LINUX || KMP_OS_FREEBSD) && KMP_AFFINITY_SUPPORTED)
+#if ((KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD) &&                      \
+     KMP_AFFINITY_SUPPORTED)
 
 /* Affinity support */
 
@@ -149,6 +153,8 @@ void __kmp_affinity_determine_capable(const char *env_var) {
 #define KMP_CPU_SET_TRY_SIZE CACHE_LINE
 #elif KMP_OS_FREEBSD
 #define KMP_CPU_SET_SIZE_LIMIT (sizeof(cpuset_t))
+#elif KMP_OS_NETBSD
+#define KMP_CPU_SET_SIZE_LIMIT (256)
 #endif
 
   int verbose = __kmp_affinity.flags.verbose;
@@ -236,7 +242,7 @@ void __kmp_affinity_determine_capable(const char *env_var) {
     KMP_INTERNAL_FREE(buf);
     return;
   }
-#elif KMP_OS_FREEBSD
+#elif KMP_OS_FREEBSD || KMP_OS_NETBSD
   long gCode;
   unsigned char *buf;
   buf = (unsigned char *)KMP_INTERNAL_MALLOC(KMP_CPU_SET_SIZE_LIMIT);
@@ -1262,7 +1268,7 @@ static void __kmp_atfork_child(void) {
   ++__kmp_fork_count;
 
 #if KMP_AFFINITY_SUPPORTED
-#if KMP_OS_LINUX || KMP_OS_FREEBSD
+#if KMP_OS_LINUX || KMP_OS_FREEBSD || KMP_OS_NETBSD
   // reset the affinity in the child to the initial thread
   // affinity in the parent
   kmp_set_thread_affinity_mask_initial();