floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Fix CPU identification to work on AIX.

This commit is contained in:
Chip-Kerchner 2023-10-02 12:00:48 -05:00
parent a69367c43b
commit 3b1150fcee
1 changed files with 58 additions and 158 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

216
driver/others/dynamic_power.c
View File

@ -20,12 +20,10 @@ static char *corename[] = {
"POWER10" "POWER10"
}; };
#define NUM_CORETYPES 4 #define NUM_CORETYPES 5
char *gotoblas_corename(void) { char *gotoblas_corename(void) {
#ifndef C_PGI
if (gotoblas == &gotoblas_POWER6) return corename[1]; if (gotoblas == &gotoblas_POWER6) return corename[1];
#endif
if (gotoblas == &gotoblas_POWER8) return corename[2]; if (gotoblas == &gotoblas_POWER8) return corename[2];
#if (!defined __GNUC__) || ( __GNUC__ >= 6) #if (!defined __GNUC__) || ( __GNUC__ >= 6)
if (gotoblas == &gotoblas_POWER9) return corename[3]; if (gotoblas == &gotoblas_POWER9) return corename[3];
@ -36,177 +34,81 @@ char *gotoblas_corename(void) {
return corename[0]; return corename[0];
} }
#if defined(__clang__) #ifdef _AIX
static int __builtin_cpu_supports(char* arg) #include <sys/systemcfg.h>
{
return 0;
}
#endif
#if defined(C_PGI) || defined(__clang__) #define CPU_UNKNOWN 0
/* #define CPU_POWER6 6
* NV HPC compilers do not yet implement __builtin_cpu_is(). #define CPU_POWER7 7
* Fake a version here for use in the CPU detection code below. #define CPU_POWER8 8
* #define CPU_POWER9 9
* Strategy here is to first check the CPU to see what it actually is,
* and then test the input to see if what the CPU actually is matches
* what was requested.
*/
#include <string.h>
/*
* Define POWER processor version table.
*
* NOTE NV HPC SDK compilers only support POWER8 and POWER9 at this time
*/
#define CPU_UNKNOWN 0
#define CPU_POWER5 5
#define CPU_POWER6 6
#define CPU_POWER8 8
#define CPU_POWER9 9
#define CPU_POWER10 10 #define CPU_POWER10 10
static struct { int cpuid()
uint32_t pvr_mask; {
uint32_t pvr_value; int arch = _system_configuration.implementation;
const char* cpu_name; #ifdef POWER_6
uint32_t cpu_type; if (arch == POWER_6) return CPU_POWER6;
} pvrPOWER [] = {
{ /* POWER6 in P5+ mode; 2.04-compliant processor */
.pvr_mask = 0xffffffff,
.pvr_value = 0x0f000001,
.cpu_name = "POWER5+",
.cpu_type = CPU_POWER5,
},
{ /* Power6 aka POWER6X*/
.pvr_mask = 0xffff0000,
.pvr_value = 0x003e0000,
.cpu_name = "POWER6 (raw)",
.cpu_type = CPU_POWER6,
},
{ /* Power7 */
.pvr_mask = 0xffff0000,
.pvr_value = 0x003f0000,
.cpu_name = "POWER7 (raw)",
.cpu_type = CPU_POWER6,
},
{ /* Power7+ */
.pvr_mask = 0xffff0000,
.pvr_value = 0x004A0000,
.cpu_name = "POWER7+ (raw)",
.cpu_type = CPU_POWER6,
},
{ /* Power8E */
.pvr_mask = 0xffff0000,
.pvr_value = 0x004b0000,
.cpu_name = "POWER8E (raw)",
.cpu_type = CPU_POWER8,
},
{ /* Power8NVL */
.pvr_mask = 0xffff0000,
.pvr_value = 0x004c0000,
.cpu_name = "POWER8NVL (raw)",
.cpu_type = CPU_POWER8,
},
{ /* Power8 */
.pvr_mask = 0xffff0000,
.pvr_value = 0x004d0000,
.cpu_name = "POWER8 (raw)",
.cpu_type = CPU_POWER8,
},
{ /* Power9 DD2.0 */
.pvr_mask = 0xffffefff,
.pvr_value = 0x004e0200,
.cpu_name = "POWER9 (raw)",
.cpu_type = CPU_POWER9,
},
{ /* Power9 DD 2.1 */
.pvr_mask = 0xffffefff,
.pvr_value = 0x004e0201,
.cpu_name = "POWER9 (raw)",
.cpu_type = CPU_POWER9,
},
{ /* Power9 DD2.2 or later */
.pvr_mask = 0xffff0000,
.pvr_value = 0x004e0000,
.cpu_name = "POWER9 (raw)",
.cpu_type = CPU_POWER9,
},
{ /* Power10 */
.pvr_mask = 0xffff0000,
.pvr_value = 0x00800000,
.cpu_name = "POWER10 (raw)",
.cpu_type = CPU_POWER10,
},
{ /* End of table, pvr_mask and pvr_value must be zero */
.pvr_mask = 0x0,
.pvr_value = 0x0,
.cpu_name = "Unknown",
.cpu_type = CPU_UNKNOWN,
},
};
static int __builtin_cpu_is(const char *cpu) {
int i;
uint32_t pvr;
uint32_t cpu_type;
asm("mfpvr %0" : "=r"(pvr));
for (i = 0 ; i < sizeof pvrPOWER / sizeof *pvrPOWER ; ++i) {
if ((pvr & pvrPOWER[i].pvr_mask) == pvrPOWER[i].pvr_value) {
break;
}
}
#if defined(DEBUG)
printf("%s: returning CPU=%s, cpu_type=%p\n", __func__,
pvrPOWER[i].cpu_name, pvrPOWER[i].cpu_type);
#endif #endif
cpu_type = pvrPOWER[i].cpu_type; #ifdef POWER_7
else if (arch == POWER_7) return CPU_POWER7;
if (!strcmp(cpu, "power8")) #endif
return cpu_type == CPU_POWER8; #ifdef POWER_8
if (!strcmp(cpu, "power9")) else if (arch == POWER_8) return CPU_POWER8;
return cpu_type == CPU_POWER9; #endif
return 0; #ifdef POWER_9
else if (arch == POWER_9) return CPU_POWER9;
#endif
#ifdef POWER_10
else if (arch == POWER_10) return CPU_POWER10;
#endif
return CPU_UNKNOWN;
} }
#endif /* C_PGI */ #ifndef __BUILTIN_CPU_SUPPORTS__
static int __builtin_cpu_supports(char* arg)
{
static int ipinfo = -1;
if (ipinfo < 0) {
ipinfo = cpuid();
}
if (ipinfo >= CPU_POWER10) {
if (!strcmp(arg, "power10")) return 1;
}
if (ipinfo >= CPU_POWER9) {
if (!strcmp(arg, "power9")) return 1;
}
if (ipinfo >= CPU_POWER8) {
if (!strcmp(arg, "power8")) return 1;
}
if (ipinfo >= CPU_POWER6) {
if (!strcmp(arg, "power6")) return 1;
}
return 0;
}
#endif
static gotoblas_t *get_coretype(void) { static gotoblas_t *get_coretype(void) {
#ifndef C_PGI if (__builtin_cpu_supports("power6"))
if (__builtin_cpu_is("power6") || __builtin_cpu_is("power6x"))
return &gotoblas_POWER6; return &gotoblas_POWER6;
#endif if (__builtin_cpu_supports("power8"))
if (__builtin_cpu_is("power8"))
return &gotoblas_POWER8; return &gotoblas_POWER8;
#if (!defined __GNUC__) || ( __GNUC__ >= 6) #if (!defined __GNUC__) || ( __GNUC__ >= 6)
if (__builtin_cpu_is("power9")) if (__builtin_cpu_supports("power9"))
return &gotoblas_POWER9; return &gotoblas_POWER9;
#endif #endif
#ifdef HAVE_P10_SUPPORT #ifdef HAVE_P10_SUPPORT
if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")) #ifdef _AIX
if (__builtin_cpu_supports("power10"))
#else
if (__builtin_cpu_supports("arch_3_1") && __builtin_cpu_supports("mma"))
#endif
return &gotoblas_POWER10; return &gotoblas_POWER10;
#endif #endif
/* Fall back to the POWER9 implementation if the toolchain is too old or the MMA feature is not set */ /* Fall back to the POWER9 implementation if the toolchain is too old or the MMA feature is not set */
#if (!defined __GNUC__) || ( __GNUC__ >= 11) || (__GNUC__ == 10 && __GNUC_MINOR__ >= 2) #if (!defined __GNUC__) || ( __GNUC__ < 11) || (__GNUC__ == 10 && __GNUC_MINOR__ < 2)
if (__builtin_cpu_is("power10")) if (__builtin_cpu_supports("power10"))
return &gotoblas_POWER9; return &gotoblas_POWER9;
#endif #endif
return NULL; return NULL;
@ -229,9 +131,7 @@ static gotoblas_t *force_coretype(char * coretype) {
switch (found) switch (found)
{ {
#ifndef C_PGI
case 1: return (&gotoblas_POWER6); case 1: return (&gotoblas_POWER6);
#endif
case 2: return (&gotoblas_POWER8); case 2: return (&gotoblas_POWER8);
#if (!defined __GNUC__) || ( __GNUC__ >= 6) #if (!defined __GNUC__) || ( __GNUC__ >= 6)
case 3: return (&gotoblas_POWER9); case 3: return (&gotoblas_POWER9);