From 2fd0a129c3b9eaf8dea25b32ee4fe9deef629963 Mon Sep 17 00:00:00 2001
From: Martin Kroeker <martin@ruby.chemie.uni-freiburg.de>
Date: Thu, 26 Dec 2024 10:45:57 +0100
Subject: [PATCH] count and sort cpu models on big.little systems

---
 cpuid_arm64.c | 77 +++++++++++++++++++--------------------------------
 1 file changed, 28 insertions(+), 49 deletions(-)

diff --git a/cpuid_arm64.c b/cpuid_arm64.c
index fbb78e794..0d9254888 100644
--- a/cpuid_arm64.c
+++ b/cpuid_arm64.c
@@ -25,6 +25,7 @@
   USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
   *****************************************************************************/
 
+#include <stdlib.h>
 #include <string.h>
 #ifdef __APPLE__
 #include <sys/sysctl.h>
@@ -42,11 +43,9 @@ size_t length64=sizeof(value64);
 #define CPU_CORTEXA57     3
 #define CPU_CORTEXA72     4
 #define CPU_CORTEXA73     5
-#define CPU_CORTEXA76	  23
 #define CPU_NEOVERSEN1    11
 #define CPU_NEOVERSEV1    16
 #define CPU_NEOVERSEN2    17
-#define CPU_NEOVERSEV2   24
 #define CPU_CORTEXX1      18
 #define CPU_CORTEXX2	  19
 #define CPU_CORTEXA510	  20
@@ -91,9 +90,7 @@ static char *cpuname[] = {
   "CORTEXX2",
   "CORTEXA510",
   "CORTEXA710",
-  "FT2000",
-  "CORTEXA76",
-	"NEOVERSEV2"
+  "FT2000"
 };
 
 static char *cpuname_lower[] = {
@@ -119,15 +116,13 @@ static char *cpuname_lower[] = {
   "cortexx2",
   "cortexa510",
   "cortexa710",
-  "ft2000",
-  "cortexa76",
-	"neoversev2"
+  "ft2000"
 };
 
 int get_feature(char *search)
 {
 
-#if defined( __linux ) || defined( __NetBSD__ )
+#ifdef __linux
 	FILE *infile;
   	char buffer[2048], *p,*t;
   	p = (char *) NULL ;
@@ -158,19 +153,29 @@ int get_feature(char *search)
 #endif
 	return(0);
 }
-
+static int cpusort(const void *model1, const void *model2)
+{
+	return (*(int*)model2-*(int*)model1);
+}
 
 int detect(void)
 {
 
-#if defined( __linux ) || defined( __NetBSD__ )
-
+#ifdef __linux
+	int n,i;
+	int cpucores[1024];
 	FILE *infile;
-	char buffer[512], *p, *cpu_part = NULL, *cpu_implementer = NULL;
+	char cpupart[6];
+	char buffer[2048], *p, *cpu_part = NULL, *cpu_implementer = NULL;
 	p = (char *) NULL ;
-
+	infile = fopen("/proc/cpuinfo", "r");                                                                                   while (fgets(buffer, sizeof(buffer), infile))               {                                                                                                                               if (!strncmp("processor", buffer, 9))                       n++;                                                }
+                                                                    fclose(infile);
+								    for (i=0;i<1024;i++)cpucores[i]=0;
+	cpu_implementer=NULL;
 	infile = fopen("/proc/cpuinfo", "r");
+	for (i=0;i<n-1;i++){
 	while (fgets(buffer, sizeof(buffer), infile)) {
+		cpu_part=NULL;//cpu_implementer=NULL;
 		if ((cpu_part != NULL) && (cpu_implementer != NULL)) {
 			break;
 		}
@@ -178,13 +183,19 @@ int detect(void)
 		if ((cpu_part == NULL) && !strncmp("CPU part", buffer, 8)) {
 			cpu_part = strchr(buffer, ':') + 2;
 			cpu_part = strdup(cpu_part);
+			cpucores[i]=strtol(cpu_part,NULL,0);
+
 		} else if ((cpu_implementer == NULL) && !strncmp("CPU implementer", buffer, 15)) {
 			cpu_implementer = strchr(buffer, ':') + 2;
 			cpu_implementer = strdup(cpu_implementer);
 		}
-	}
 
+		}
+	}
 	fclose(infile);
+	qsort(cpucores,1024,sizeof(int),cpusort);
+	sprintf(cpupart,"0x%3x",cpucores[0]);
+	cpu_part=strdup(cpupart);
 	if(cpu_part != NULL && cpu_implementer != NULL) {
     // Arm
     if (strstr(cpu_implementer, "0x41")) {
@@ -216,10 +227,6 @@ int detect(void)
 	return CPU_CORTEXX2;
       else if (strstr(cpu_part, "0xd4e")) //X3
 	return CPU_CORTEXX2;
-      else if (strstr(cpu_part, "0xd4f")) //NVIDIA Grace et al.
-        return CPU_NEOVERSEV2;
-      else if (strstr(cpu_part, "0xd0b")) 
-	return CPU_CORTEXA76;
     }
     // Qualcomm
     else if (strstr(cpu_implementer, "0x51") && strstr(cpu_part, "0xc00"))
@@ -280,8 +287,6 @@ int detect(void)
 	sysctlbyname("hw.cpufamily",&value64,&length64,NULL,0);
 	if (value64 ==131287967|| value64 == 458787763 ) return CPU_VORTEX; //A12/M1
 	if (value64 == 3660830781) return CPU_VORTEX; //A15/M2
-	if (value64 == 2271604202) return CPU_VORTEX; //A16/M3
-	if (value64 == 1867590060) return CPU_VORTEX; //M4
 #endif
 	return CPU_ARMV8;	
 #endif
@@ -314,7 +319,7 @@ void get_cpucount(void)
 {
 int n=0;
 
-#if defined( __linux ) || defined( __NetBSD__ )
+#ifdef __linux
 	FILE *infile;
   	char buffer[2048], *p,*t;
   	p = (char *) NULL ;
@@ -347,7 +352,6 @@ void get_cpuconfig(void)
   printf("#define ARMV8\n");
   printf("#define HAVE_NEON\n"); // This shouldn't be necessary
   printf("#define HAVE_VFPV4\n"); // This shouldn't be necessary
-
 	int d = detect();
 	switch (d)
 	{
@@ -402,8 +406,6 @@ void get_cpuconfig(void)
 		break;
 
 	    case CPU_NEOVERSEV1:
-		printf("#define HAVE_SVE 1\n");
-	    case CPU_CORTEXA76:
                 printf("#define %s\n", cpuname[d]);
                 printf("#define L1_CODE_SIZE 65536\n");
                 printf("#define L1_CODE_LINESIZE 64\n");
@@ -431,32 +433,12 @@ void get_cpuconfig(void)
                 printf("#define L2_ASSOCIATIVE 8\n");
                 printf("#define DTB_DEFAULT_ENTRIES 48\n");
                 printf("#define DTB_SIZE 4096\n");
-		printf("#define HAVE_SVE 1\n");
-                break;
-      case CPU_NEOVERSEV2:
-                printf("#define ARMV9\n");
-	        printf("#define HAVE_SVE 1\n");
-                printf("#define %s\n", cpuname[d]);
-                printf("#define L1_CODE_SIZE 65536\n");
-                printf("#define L1_CODE_LINESIZE 64\n");
-                printf("#define L1_CODE_ASSOCIATIVE 4\n");
-                printf("#define L1_DATA_SIZE 65536\n");
-                printf("#define L1_DATA_LINESIZE 64\n");
-                printf("#define L1_DATA_ASSOCIATIVE 4\n");
-                printf("#define L2_SIZE 1048576\n");
-                printf("#define L2_LINESIZE 64\n");
-                printf("#define L2_ASSOCIATIVE 8\n");
-								// L1 Data TLB = 48 entries
-								// L2 Data TLB = 2048 entries
-                printf("#define DTB_DEFAULT_ENTRIES 48\n");
-                printf("#define DTB_SIZE 4096\n");  // Set to 4096 for symmetry with other configs.
                 break;
 	    case CPU_CORTEXA510:
 	    case CPU_CORTEXA710:
 	    case CPU_CORTEXX1:
 	    case CPU_CORTEXX2:
 		printf("#define ARMV9\n");
-		printf("#define HAVE_SVE 1\n");
                 printf("#define %s\n", cpuname[d]);
                 printf("#define L1_CODE_SIZE 65536\n");
                 printf("#define L1_CODE_LINESIZE 64\n");
@@ -559,8 +541,6 @@ void get_cpuconfig(void)
 	    case CPU_VORTEX:
 		printf("#define VORTEX			      \n");
 #ifdef __APPLE__
-		sysctlbyname("hw.cpufamily",&value64,&length64,NULL,0);
-		if (value64 == 1867590060) printf("#define HAVE_SME 1\n");; //M4
 		sysctlbyname("hw.l1icachesize",&value64,&length64,NULL,0);
 		printf("#define L1_CODE_SIZE	     %lld       \n",value64);
 		sysctlbyname("hw.cachelinesize",&value64,&length64,NULL,0);
@@ -575,7 +555,6 @@ void get_cpuconfig(void)
 		break;
 	    case CPU_A64FX:
 		printf("#define A64FX\n");
-		printf("#define HAVE_SVE 1\n");
     		printf("#define L1_CODE_SIZE 65535\n");
     		printf("#define L1_DATA_SIZE 65535\n");
     		printf("#define L1_DATA_LINESIZE 256\n");
@@ -608,7 +587,7 @@ void get_libname(void)
 void get_features(void)
 {
 
-#if defined( __linux ) || defined( __NetBSD__ )
+#ifdef __linux
 	FILE *infile;
   	char buffer[2048], *p,*t;
   	p = (char *) NULL ;