Skip to content

Commit be807c9

Browse files
martin-frbgmartin-frbg
authored
Identify all cores, group by performance and report the fastest TARGET
1 parent a63282a commit be807c9

File tree

1 file changed

+123
-61
lines changed

1 file changed

+123
-61
lines changed

‎cpuid_arm64.c

Lines changed: 123 additions & 61 deletions
Original file line numberDiff line numberDiff line change
@@ -25,6 +25,7 @@
2525
USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
2626
*****************************************************************************/
2727

28+
#include <stdlib.h>
2829
#include <string.h>
2930
#ifdef __APPLE__
3031
#include <sys/sysctl.h>
@@ -33,6 +34,20 @@ size_t length=sizeof(value);
3334
int64_t value64;
3435
size_t length64=sizeof(value64);
3536
#endif
37+
#if (defined OS_LINUX || defined OS_ANDROID)
38+
#include <asm/hwcap.h>
39+
#include <sys/auxv.h>
40+
#ifndef HWCAP_CPUID
41+
#define HWCAP_CPUID (1 << 11)
42+
#endif
43+
#ifndef HWCAP_SVE
44+
#define HWCAP_SVE (1 << 22)
45+
#endif
46+
47+
#define get_cpu_ftr(id, var) ({ \
48+
__asm__ __volatile__ ("mrs %0, "#id : "=r" (var)); \
49+
})
50+
#endif
3651

3752
#define CPU_UNKNOWN 0
3853
#define CPU_ARMV8 1
@@ -42,11 +57,9 @@ size_t length64=sizeof(value64);
4257
#define CPU_CORTEXA57 3
4358
#define CPU_CORTEXA72 4
4459
#define CPU_CORTEXA73 5
45-
#define CPU_CORTEXA76 23
4660
#define CPU_NEOVERSEN1 11
4761
#define CPU_NEOVERSEV1 16
4862
#define CPU_NEOVERSEN2 17
49-
#define CPU_NEOVERSEV2 24
5063
#define CPU_CORTEXX1 18
5164
#define CPU_CORTEXX2 19
5265
#define CPU_CORTEXA510 20
@@ -91,9 +104,7 @@ static char *cpuname[] = {
91104
"CORTEXX2",
92105
"CORTEXA510",
93106
"CORTEXA710",
94-
"FT2000",
95-
"CORTEXA76",
96-
"NEOVERSEV2"
107+
"FT2000"
97108
};
98109

99110
static char *cpuname_lower[] = {
@@ -119,15 +130,17 @@ static char *cpuname_lower[] = {
119130
"cortexx2",
120131
"cortexa510",
121132
"cortexa710",
122-
"ft2000",
123-
"cortexa76",
124-
"neoversev2"
133+
"ft2000"
125134
};
126135

136+
static int cpulowperf=0;
137+
static int cpumidperf=0;
138+
static int cpuhiperf=0;
139+
127140
int get_feature(char *search)
128141
{
129142

130-
#if defined( __linux ) || defined( __NetBSD__ )
143+
#ifdef __linux
131144
FILE *infile;
132145
char buffer[2048], *p,*t;
133146
p = (char *) NULL ;
@@ -158,33 +171,108 @@ int get_feature(char *search)
158171
#endif
159172
return(0);
160173
}
161-
174+
static int cpusort(const void *model1, const void *model2)
175+
{
176+
return (*(int*)model2-*(int*)model1);
177+
}
162178

163179
int detect(void)
164180
{
165181

166-
#if defined( __linux ) || defined( __NetBSD__ )
167-
182+
#ifdef __linux
183+
int n,i,ii;
184+
int midr_el1;
185+
int implementer;
186+
int cpucap[1024];
187+
int cpucores[1024];
168188
FILE *infile;
169-
char buffer[512], *p, *cpu_part = NULL, *cpu_implementer = NULL;
189+
char cpupart[6],cpuimpl[6];
190+
char *cpu_impl=NULL,*cpu_pt=NULL;
191+
char buffer[2048], *p, *cpu_part = NULL, *cpu_implementer = NULL;
170192
p = (char *) NULL ;
171-
172-
infile = fopen("/proc/cpuinfo", "r");
173-
while (fgets(buffer, sizeof(buffer), infile)) {
174-
if ((cpu_part != NULL) && (cpu_implementer != NULL)) {
175-
break;
193+
cpulowperf=cpumidperf=cpuhiperf=0;
194+
for (i=0;i<1024;i++)cpucores[i]=0;
195+
n=0;
196+
infile = fopen("/sys/devices/system/cpu/possible", "r");
197+
if (!infile) {
198+
infile = fopen("/proc/cpuinfo", "r");
199+
while (fgets(buffer, sizeof(buffer), infile)) {
200+
if (!strncmp("processor", buffer, 9))
201+
n++;
176202
}
177-
178-
if ((cpu_part == NULL) && !strncmp("CPU part", buffer, 8)) {
179-
cpu_part = strchr(buffer, ':') + 2;
180-
cpu_part = strdup(cpu_part);
181-
} else if ((cpu_implementer == NULL) && !strncmp("CPU implementer", buffer, 15)) {
182-
cpu_implementer = strchr(buffer, ':') + 2;
183-
cpu_implementer = strdup(cpu_implementer);
203+
} else {
204+
fgets(buffer, sizeof(buffer), infile);
205+
sscanf(buffer,"0-%d",&n);
206+
n++;
207+
}
208+
fclose(infile);
209+
210+
cpu_implementer=NULL;
211+
for (i=0;i<n;i++){
212+
sprintf(buffer,"/sys/devices/system/cpu/cpu%d/regs/identification/midr_el1",i);
213+
infile= fopen(buffer,"r");
214+
if (!infile) {
215+
infile = fopen("/proc/cpuinfo", "r");
216+
for (ii=0;ii<n;ii++){
217+
cpu_part=NULL;cpu_implementer=NULL;
218+
while (fgets(buffer, sizeof(buffer), infile)) {
219+
if ((cpu_part != NULL) && (cpu_implementer != NULL)) {
220+
break;
221+
}
222+
223+
if ((cpu_part == NULL) && !strncmp("CPU part", buffer, 8)) {
224+
cpu_pt = strchr(buffer, ':') + 2;
225+
cpu_part = strdup(cpu_pt);
226+
cpucores[i]=strtol(cpu_part,NULL,0);
227+
228+
} else if ((cpu_implementer == NULL) && !strncmp("CPU implementer", buffer, 15)) {
229+
cpu_impl = strchr(buffer, ':') + 2;
230+
cpu_implementer = strdup(cpu_impl);
231+
}
232+
233+
}
234+
if (strstr(cpu_implementer, "0x41")) {
235+
if (cpucores[ii] >= 0xd4b) cpuhiperf++;
236+
else
237+
if (cpucores[ii] >= 0xd07) cpumidperf++;
238+
else cpulowperf++;
239+
}
240+
else cpulowperf++;
241+
}
242+
fclose(infile);
243+
break;
244+
} else {
245+
(void)fgets(buffer, sizeof(buffer), infile);
246+
midr_el1=strtoul(buffer,NULL,16);
247+
fclose(infile);
248+
implementer = (midr_el1 >> 24) & 0xFF;
249+
cpucores[i] = (midr_el1 >> 4) & 0xFFF;
250+
sprintf(buffer,"/sys/devices/system/cpu/cpu%d/cpu_capacity",i);
251+
infile= fopen(buffer,"r");
252+
if (!infile) {
253+
if (implementer== 65) {
254+
if (cpucores[i] >= 0xd4b) cpuhiperf++;
255+
else
256+
if (cpucores[i] >= 0xd07) cpumidperf++;
257+
else cpulowperf++;
258+
}
259+
else cpulowperf++;
260+
} else {
261+
(void)fgets(buffer, sizeof(buffer), infile);
262+
sscanf(buffer,"%d",&cpucap[i]);
263+
if (cpucap[i] >= 1000) cpuhiperf++;
264+
else
265+
if (cpucap[i] >= 500) cpumidperf++;
266+
else cpulowperf++;
267+
fclose(infile);
268+
}
184269
}
270+
sprintf(cpuimpl,"0x%2x",implementer);
271+
cpu_implementer=strdup(cpuimpl);
185272
}
186-
187-
fclose(infile);
273+
qsort(cpucores,1024,sizeof(int),cpusort);
274+
sprintf(cpupart,"0x%3x",cpucores[0]);
275+
cpu_part=strdup(cpupart);
188276
if(cpu_part != NULL && cpu_implementer != NULL) {
189277
// Arm
190278
if (strstr(cpu_implementer, "0x41")) {
@@ -216,10 +304,6 @@ int detect(void)
216304
return CPU_CORTEXX2;
217305
else if (strstr(cpu_part, "0xd4e")) //X3
218306
return CPU_CORTEXX2;
219-
else if (strstr(cpu_part, "0xd4f")) //NVIDIA Grace et al.
220-
return CPU_NEOVERSEV2;
221-
else if (strstr(cpu_part, "0xd0b"))
222-
return CPU_CORTEXA76;
223307
}
224308
// Qualcomm
225309
else if (strstr(cpu_implementer, "0x51") && strstr(cpu_part, "0xc00"))
@@ -280,8 +364,6 @@ int detect(void)
280364
sysctlbyname("hw.cpufamily",&value64,&length64,NULL,0);
281365
if (value64 ==131287967|| value64 == 458787763 ) return CPU_VORTEX; //A12/M1
282366
if (value64 == 3660830781) return CPU_VORTEX; //A15/M2
283-
if (value64 == 2271604202) return CPU_VORTEX; //A16/M3
284-
if (value64 == 1867590060) return CPU_VORTEX; //M4
285367
#endif
286368
return CPU_ARMV8;
287369
#endif
@@ -314,7 +396,7 @@ void get_cpucount(void)
314396
{
315397
int n=0;
316398

317-
#if defined( __linux ) || defined( __NetBSD__ )
399+
#ifdef __linux
318400
FILE *infile;
319401
char buffer[2048], *p,*t;
320402
p = (char *) NULL ;
@@ -331,6 +413,12 @@ int n=0;
331413
fclose(infile);
332414

333415
printf("#define NUM_CORES %d\n",n);
416+
if (cpulowperf >0)
417+
printf("#define NUM_CORES_LP %d\n",cpulowperf);
418+
if (cpumidperf >0)
419+
printf("#define NUM_CORES_MP %d\n",cpumidperf);
420+
if (cpuhiperf >0)
421+
printf("#define NUM_CORES_HP %d\n",cpuhiperf);
334422
#endif
335423
#ifdef __APPLE__
336424
sysctlbyname("hw.physicalcpu_max",&value,&length,NULL,0);
@@ -347,7 +435,6 @@ void get_cpuconfig(void)
347435
printf("#define ARMV8\n");
348436
printf("#define HAVE_NEON\n"); // This shouldn't be necessary
349437
printf("#define HAVE_VFPV4\n"); // This shouldn't be necessary
350-
351438
int d = detect();
352439
switch (d)
353440
{
@@ -402,8 +489,6 @@ void get_cpuconfig(void)
402489
break;
403490

404491
case CPU_NEOVERSEV1:
405-
printf("#define HAVE_SVE 1\n");
406-
case CPU_CORTEXA76:
407492
printf("#define %s\n", cpuname[d]);
408493
printf("#define L1_CODE_SIZE 65536\n");
409494
printf("#define L1_CODE_LINESIZE 64\n");
@@ -431,32 +516,12 @@ void get_cpuconfig(void)
431516
printf("#define L2_ASSOCIATIVE 8\n");
432517
printf("#define DTB_DEFAULT_ENTRIES 48\n");
433518
printf("#define DTB_SIZE 4096\n");
434-
printf("#define HAVE_SVE 1\n");
435-
break;
436-
case CPU_NEOVERSEV2:
437-
printf("#define ARMV9\n");
438-
printf("#define HAVE_SVE 1\n");
439-
printf("#define %s\n", cpuname[d]);
440-
printf("#define L1_CODE_SIZE 65536\n");
441-
printf("#define L1_CODE_LINESIZE 64\n");
442-
printf("#define L1_CODE_ASSOCIATIVE 4\n");
443-
printf("#define L1_DATA_SIZE 65536\n");
444-
printf("#define L1_DATA_LINESIZE 64\n");
445-
printf("#define L1_DATA_ASSOCIATIVE 4\n");
446-
printf("#define L2_SIZE 1048576\n");
447-
printf("#define L2_LINESIZE 64\n");
448-
printf("#define L2_ASSOCIATIVE 8\n");
449-
// L1 Data TLB = 48 entries
450-
// L2 Data TLB = 2048 entries
451-
printf("#define DTB_DEFAULT_ENTRIES 48\n");
452-
printf("#define DTB_SIZE 4096\n"); // Set to 4096 for symmetry with other configs.
453519
break;
454520
case CPU_CORTEXA510:
455521
case CPU_CORTEXA710:
456522
case CPU_CORTEXX1:
457523
case CPU_CORTEXX2:
458524
printf("#define ARMV9\n");
459-
printf("#define HAVE_SVE 1\n");
460525
printf("#define %s\n", cpuname[d]);
461526
printf("#define L1_CODE_SIZE 65536\n");
462527
printf("#define L1_CODE_LINESIZE 64\n");
@@ -559,8 +624,6 @@ void get_cpuconfig(void)
559624
case CPU_VORTEX:
560625
printf("#define VORTEX \n");
561626
#ifdef __APPLE__
562-
sysctlbyname("hw.cpufamily",&value64,&length64,NULL,0);
563-
if (value64 == 1867590060) printf("#define HAVE_SME 1\n");; //M4
564627
sysctlbyname("hw.l1icachesize",&value64,&length64,NULL,0);
565628
printf("#define L1_CODE_SIZE %lld \n",value64);
566629
sysctlbyname("hw.cachelinesize",&value64,&length64,NULL,0);
@@ -575,7 +638,6 @@ void get_cpuconfig(void)
575638
break;
576639
case CPU_A64FX:
577640
printf("#define A64FX\n");
578-
printf("#define HAVE_SVE 1\n");
579641
printf("#define L1_CODE_SIZE 65535\n");
580642
printf("#define L1_DATA_SIZE 65535\n");
581643
printf("#define L1_DATA_LINESIZE 256\n");
@@ -608,7 +670,7 @@ void get_libname(void)
608670
void get_features(void)
609671
{
610672

611-
#if defined( __linux ) || defined( __NetBSD__ )
673+
#ifdef __linux
612674
FILE *infile;
613675
char buffer[2048], *p,*t;
614676
p = (char *) NULL ;

0 commit comments

Comments
 (0)