add startup numa sense

2026-04-02 18:13:32 +08:00
parent ba30c02cb3
commit 2f74e5e8a9
10 changed files with 273 additions and 45 deletions
@@ -0,0 +1,69 @@
+#include "sn_target.h"
+#include "StgSharpNative.h"
+
+#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 (core_id < 0) {
+                /* Query which group the thread currently belongs to */
+                GROUP_AFFINITY current;
+                GetThreadGroupAffinity(GetCurrentThread(), &current);
+                /* Build a full mask for that group's actual core count */
+                DWORD count = GetActiveProcessorCount(current.Group);
+                GROUP_AFFINITY ga;
+                memset(&ga, 0, sizeof(ga));
+                ga.Group = current.Group;
+                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;
+        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;
+        SetThreadGroupAffinity(GetCurrentThread(), &ga, NULL);
+}
+
+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