add startup numa sense

StgSharp.Native/sn-src/simdid.x86.c

@@ -340,16 +340,6 @@ SN_API SIMDID SN_DECL sn_get_simd_level_local()
         return simdid;
 }
 
-SN_API SIMDID SN_DECL sn_get_simd_level_global()
-{
-        if (global_simdid.mask) {
-                return global_simdid;
-        }
-
-        global_simdid = sn_get_simd_level_local();
-        return global_simdid;
-}
-
 #else
 
 SN_API SIMDID SN_DECL sn_get_simd_level_local()

StgSharp.Native/sn-src/thread.linux.c

@@ -0,0 +1,39 @@
+#include "sn_target.h"
+#include "StgSharpNative.h"
+
+#if SN_IS_OS(SN_OS_LINUX)
+#define _GNU_SOURCE
+#include <pthread.h>
+#include <sched.h>
+#include <unistd.h>
+
+/* Pin the calling thread to a single logical processor.
+ * core_id >= 0 : bind to that logical processor (global index, across groups).
+ * core_id <  0 : clear affinity constraint within the current group. */
+SN_API void SN_DECL sn_set_thread_affinity(int core_id)
+{
+        long ncpus = sysconf(_SC_NPROCESSORS_ONLN);
+
+        cpu_set_t *set = CPU_ALLOC(ncpus);
+        size_t size = CPU_ALLOC_SIZE(ncpus);
+        CPU_ZERO_S(size, set);
+
+        if (core_id < 0) {
+                for (long i = 0; i < ncpus; i++)
+                        CPU_SET_S((int)i, size, set);
+        } else {
+                CPU_SET_S(core_id, size, set);
+        }
+
+        pthread_setaffinity_np(pthread_self(), size, set);
+        CPU_FREE(set);
+}
+
+SN_API void SN_DECL sn_get_numa_overall(int *out_numa_count, int *count_of_each_numa)
+{
+}
+
+SN_API void SN_DECL sn_get_numa_each(int numa_id, int *in_numa_cpuid, int *global_cpuid)
+{
+}
+#endif

StgSharp.Native/sn-src/thread.windows.c

@@ -1,19 +1,13 @@
 #include "sn_target.h"
 #include "StgSharpNative.h"
 
-#if SN_IS_OS(SN_OS_LINUX)
-#define _GNU_SOURCE
-#include <pthread.h>
-#include <sched.h>
-#include <unistd.h>
-#endif
+#if SN_IS_OS(SN_OS_WINDOWS)
 
 /* Pin the calling thread to a single logical processor.
  * core_id >= 0 : bind to that logical processor (global index, across groups).
  * core_id <  0 : clear affinity constraint within the current group. */
 SN_API void SN_DECL sn_set_thread_affinity(int core_id)
 {
-#if SN_IS_OS(SN_OS_WINDOWS)
         if (core_id < 0) {
                 /* Query which group the thread currently belongs to */
                 GROUP_AFFINITY current;
@@ -23,33 +17,53 @@ SN_API void SN_DECL sn_set_thread_affinity(int core_id)
                 GROUP_AFFINITY ga;
                 memset(&ga, 0, sizeof(ga));
                 ga.Group = current.Group;
-                ga.Mask  = (count >= 64) ? ~(KAFFINITY)0
-                                         : ((KAFFINITY)1 << count) - 1;
+                ga.Mask = (count >= 64) ? ~(KAFFINITY)0 : ((KAFFINITY)1 << count) - 1;
                 SetThreadGroupAffinity(GetCurrentThread(), &ga, NULL);
                 return;
         }
         WORD group = (WORD)(core_id / 64);
-        int  index = core_id % 64;
+        int index = core_id % 64;
         DWORD count = GetActiveProcessorCount(group);
         if ((DWORD)index >= count)
                 return;
         GROUP_AFFINITY ga;
         memset(&ga, 0, sizeof(ga));
         ga.Group = group;
-        ga.Mask  = (KAFFINITY)1 << index;
+        ga.Mask = (KAFFINITY)1 << index;
         SetThreadGroupAffinity(GetCurrentThread(), &ga, NULL);
-
-#elif SN_IS_OS(SN_OS_LINUX)
-        cpu_set_t set;
-        if (core_id < 0) {
-                CPU_ZERO(&set);
-                long ncpus = sysconf(_SC_NPROCESSORS_ONLN);
-                for (long i = 0; i < ncpus; i++)
-                        CPU_SET((int)i, &set);
-        } else {
-                CPU_ZERO(&set);
-                CPU_SET(core_id, &set);
-        }
-        pthread_setaffinity_np(pthread_self(), sizeof(set), &set);
-#endif
 }
+
+SN_API int SN_DECL sn_get_numa_overall(int *out_numa_count, int *count_of_each_numa)
+{
+        ULONG highest;
+        if (!GetNumaHighestNodeNumber(&highest))
+                return 0;
+        *out_numa_count = (int)highest + 1;
+
+        for (UCHAR n = 0; n <= (UCHAR)highest; n++) {
+                ULONGLONG mask = 0;
+                GetNumaNodeProcessorMask(n, &mask);
+                count_of_each_numa[n] = (int)__popcnt64(mask);
+        }
+        return 1;
+}
+
+SN_API int SN_DECL sn_get_numa_each(int numa_id, uint64_t *internal, uint64_t *global)
+{
+        ULONGLONG mask = 0;
+        if (!GetNumaNodeProcessorMask((UCHAR)numa_id, &mask))
+                return 0;
+
+        int intra = 0;
+        while (mask) {
+                unsigned long idx;
+                _BitScanForward64(&idx, mask);
+                internal[intra] = intra;
+                global[intra] = (int)idx;
+                intra++;
+                mask &= mask - 1;
+        }
+        return 1;
+}
+
+#endif