Mac affinity

commonb.c

@@ -599,6 +599,7 @@ void SetPriority (
 /* Set affinity for this thread to a specific CPU core */
 
 	if (bind_type == 0) {
+#ifndef __APPLE__
 		int	num_cores;
 		hwloc_obj_t obj;
 		num_cores = hwloc_get_nbobjs_by_type (hwloc_topology, HWLOC_OBJ_CORE);
@@ -630,6 +631,17 @@ void SetPriority (
 			sprintf (buf, "Error getting hwloc object for core #%d.  Affinity not set.\n", core+1);
 			OutputStr (info->worker_num, buf);
 		}
+#else
+		int merror = mach_set_thread_cpubind(pthread_self(), core);
+		if (merror) {
+			sprintf (buf, "Error setting affinity to affinity-set #%d: %s\n", core+1, strerror (merror));
+			OutputStr (info->worker_num, buf);
+		}
+		else if (info->verbosity >= 2) {
+			sprintf (buf, "Affinity set to affinity-set #%d\n", core+1);
+			OutputStr (info->worker_num, buf);
+		}
+#endif
 	}
 
 /* Set affinity for this thread to a specific logical CPU */

commonc.c

@@ -1323,8 +1323,10 @@ void nameAndReadIniFiles (
 	{
 		const struct hwloc_topology_support *support;
 		OS_CAN_SET_AFFINITY = 1;
+#ifndef __APPLE__
 		support = hwloc_topology_get_support (hwloc_topology);
 		if (support == NULL || ! support->cpubind->set_thread_cpubind) OS_CAN_SET_AFFINITY = 0;
+#endif
 	}
 
 /* Initialize mutexes */

macosx64/aff.c

@@ -0,0 +1,59 @@
+#ifdef __APPLE__
+#include <sys/types.h>
+#include <sys/sysctl.h>
+#include <pthread.h>
+#include <stdint.h>
+#include <mach/thread_policy.h>
+#include <mach/thread_act.h>
+#include <errno.h>
+
+/* Alternatively, this can be simplified to set-only without checking core count */
+int32_t pu_per_core = 0;
+int32_t ncore = 0;
+
+__attribute__((constructor))
+static inline void initialize_pu_per_core() {
+    int64_t cacheconfig[32];
+    int32_t *cacheconfig32 = cacheconfig;
+    size_t size = sizeof(cacheconfig);
+    sysctlbyname("hw.cacheconfig", &cacheconfig, &size, NULL, 0);
+    /* per hwloc, there are 32 and 64-bit configs... */
+    if (cacheconfig[0] <= 0xFFFFFFFFUL) {
+        pu_per_core = cacheconfig[2];
+    } else {
+        pu_per_core = cacheconfig32[2];
+    }
+}
+
+__attribute__((constructor))
+static inline void initialize_ncore() {
+    size_t size = sizeof(ncore);
+    sysctlbyname("hw.ncpu", &ncore, &size, NULL, 0);
+    // assert: 0 == ncore % pu_per_core
+    ncore /= pu_per_core;
+}
+
+/* Index is 1-based! 0 means no binding. Return value is positive-errno. Most commonly you get EPFNOSUPPORT (46) on ARM */
+int mach_set_thread_cpubind(pthread_t thread, int cpu) {
+    if (cpu < 0 || cpu > ncore) {
+        return EINVAL;
+    }
+
+    thread_act_t handle = pthread_mach_thread_np(thread);
+    thread_affinity_policy_data_t policy = {.affinity_tag = cpu};
+    return thread_policy_set(handle, THREAD_AFFINITY_POLICY, (thread_policy_t)&policy, THREAD_AFFINITY_POLICY_COUNT);
+}
+
+/* Return value is tag or negative-errno */
+int mach_get_thread_cpubind(pthread_t thread) {
+    thread_act_t handle = pthread_mach_thread_np(thread);
+    thread_affinity_policy_data_t policy;
+    mach_msg_type_number_t count = THREAD_AFFINITY_POLICY_COUNT;
+    kern_return_t kr = thread_policy_get(handle, THREAD_AFFINITY_POLICY, (thread_policy_t)&policy, &count, 0);
+    if (kr != KERN_SUCCESS) {
+        return -kr;
+    }
+
+    return policy.affinity_tag;
+}
+#endif

macosx64/makefile

@@ -23,7 +23,7 @@ LFLAGS = -m64 -Wl,-no_pie -L/usr/local/lib
 LIBS   = ../gwnum/gwnum.a ../gwnum/polymult.a -lm -lpthread /usr/local/lib/libhwloc.a /usr/local/lib/libgmp.a -lcurl -framework IOKit -framework CoreFoundation -lc++
 
 FACTOROBJ = ../prime95/macosx64/factor64.o
-OBJS = prime.o menu.o cJSON.o ecm.o exponentiate.o pair.o pm1prob.o
+OBJS = aff.o prime.o menu.o cJSON.o ecm.o exponentiate.o pair.o pm1prob.o
 EXE = mprime
 
 #########################################################################

macosx64/prime.h

@@ -85,3 +85,10 @@ void test_status(void);
 
 void rangeStatus (void);
 void options_cpu (void);
+
+#ifdef __APPLE__
+#include <pthread.h>
+int mach_set_thread_cpubind(pthread_t thread, int cpu);
+int mach_get_thread_cpubind(pthread_t thread);
+#endif
+