From 3b1150fcee164922ed932c7d46b28a8ffec744a8 Mon Sep 17 00:00:00 2001 From: Chip-Kerchner Date: Mon, 2 Oct 2023 12:00:48 -0500 Subject: [PATCH 1/7] Fix CPU identification to work on AIX. --- driver/others/dynamic_power.c | 216 +++++++++------------------------- 1 file changed, 58 insertions(+), 158 deletions(-) diff --git a/driver/others/dynamic_power.c b/driver/others/dynamic_power.c index 2847ea9ae..7f8bfd5b9 100644 --- a/driver/others/dynamic_power.c +++ b/driver/others/dynamic_power.c @@ -20,12 +20,10 @@ static char *corename[] = { "POWER10" }; -#define NUM_CORETYPES 4 +#define NUM_CORETYPES 5 char *gotoblas_corename(void) { -#ifndef C_PGI if (gotoblas == &gotoblas_POWER6) return corename[1]; -#endif if (gotoblas == &gotoblas_POWER8) return corename[2]; #if (!defined __GNUC__) || ( __GNUC__ >= 6) if (gotoblas == &gotoblas_POWER9) return corename[3]; @@ -36,177 +34,81 @@ char *gotoblas_corename(void) { return corename[0]; } -#if defined(__clang__) -static int __builtin_cpu_supports(char* arg) -{ - return 0; -} -#endif +#ifdef _AIX +#include -#if defined(C_PGI) || defined(__clang__) -/* - * NV HPC compilers do not yet implement __builtin_cpu_is(). - * Fake a version here for use in the CPU detection code below. - * - * 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 - -/* - * 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_UNKNOWN 0 +#define CPU_POWER6 6 +#define CPU_POWER7 7 +#define CPU_POWER8 8 +#define CPU_POWER9 9 #define CPU_POWER10 10 -static struct { - uint32_t pvr_mask; - uint32_t pvr_value; - const char* cpu_name; - uint32_t cpu_type; -} 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); +int cpuid() +{ + int arch = _system_configuration.implementation; +#ifdef POWER_6 + if (arch == POWER_6) return CPU_POWER6; #endif - cpu_type = pvrPOWER[i].cpu_type; - - if (!strcmp(cpu, "power8")) - return cpu_type == CPU_POWER8; - if (!strcmp(cpu, "power9")) - return cpu_type == CPU_POWER9; - return 0; +#ifdef POWER_7 + else if (arch == POWER_7) return CPU_POWER7; +#endif +#ifdef POWER_8 + else if (arch == POWER_8) return CPU_POWER8; +#endif +#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) { -#ifndef C_PGI - if (__builtin_cpu_is("power6") || __builtin_cpu_is("power6x")) + if (__builtin_cpu_supports("power6")) return &gotoblas_POWER6; -#endif - if (__builtin_cpu_is("power8")) + if (__builtin_cpu_supports("power8")) return &gotoblas_POWER8; #if (!defined __GNUC__) || ( __GNUC__ >= 6) - if (__builtin_cpu_is("power9")) + if (__builtin_cpu_supports("power9")) return &gotoblas_POWER9; #endif #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; #endif /* 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 (__builtin_cpu_is("power10")) +#if (!defined __GNUC__) || ( __GNUC__ < 11) || (__GNUC__ == 10 && __GNUC_MINOR__ < 2) + if (__builtin_cpu_supports("power10")) return &gotoblas_POWER9; #endif return NULL; @@ -229,9 +131,7 @@ static gotoblas_t *force_coretype(char * coretype) { switch (found) { -#ifndef C_PGI case 1: return (&gotoblas_POWER6); -#endif case 2: return (&gotoblas_POWER8); #if (!defined __GNUC__) || ( __GNUC__ >= 6) case 3: return (&gotoblas_POWER9); From eb738d99293dc658bd6941cc6c2b76cd6ece0c11 Mon Sep 17 00:00:00 2001 From: Chip-Kerchner Date: Mon, 2 Oct 2023 12:14:46 -0500 Subject: [PATCH 2/7] Minor changes. --- Makefile.system | 2 -- driver/others/dynamic_power.c | 2 +- 2 files changed, 1 insertion(+), 3 deletions(-) diff --git a/Makefile.system b/Makefile.system index 1fd47e68e..b1a357fdf 100644 --- a/Makefile.system +++ b/Makefile.system @@ -1170,8 +1170,6 @@ CCOMMON_OPT += -DF_INTERFACE_IBM FEXTRALIB += -lxlf90 ifeq ($(C_COMPILER), $(filter $(C_COMPILER),GCC CLANG)) FCOMMON_OPT += -qextname -else ifeq ($(C_COMPILER), CLANG) -FCOMMON_OPT += -qextname endif # FCOMMON_OPT += -qarch=440 ifdef BINARY64 diff --git a/driver/others/dynamic_power.c b/driver/others/dynamic_power.c index 7f8bfd5b9..1d3f36875 100644 --- a/driver/others/dynamic_power.c +++ b/driver/others/dynamic_power.c @@ -44,7 +44,7 @@ char *gotoblas_corename(void) { #define CPU_POWER9 9 #define CPU_POWER10 10 -int cpuid() +static int cpuid(void) { int arch = _system_configuration.implementation; #ifdef POWER_6 From 12130ee9613936f2fa49fd58a7f6bf8210a65552 Mon Sep 17 00:00:00 2001 From: Chip-Kerchner Date: Mon, 2 Oct 2023 12:19:22 -0500 Subject: [PATCH 3/7] Remove tab. --- driver/others/dynamic_power.c | 2 +- 1 file changed, 1 insertion(+), 1 deletion(-) diff --git a/driver/others/dynamic_power.c b/driver/others/dynamic_power.c index 1d3f36875..3c5f1f3c1 100644 --- a/driver/others/dynamic_power.c +++ b/driver/others/dynamic_power.c @@ -102,7 +102,7 @@ static gotoblas_t *get_coretype(void) { #ifdef _AIX if (__builtin_cpu_supports("power10")) #else - if (__builtin_cpu_supports("arch_3_1") && __builtin_cpu_supports("mma")) + if (__builtin_cpu_supports("arch_3_1") && __builtin_cpu_supports("mma")) #endif return &gotoblas_POWER10; #endif From a922a07e610e0508e2f2f84ae158c46e2e3d7a0e Mon Sep 17 00:00:00 2001 From: Chip-Kerchner Date: Mon, 2 Oct 2023 12:24:30 -0500 Subject: [PATCH 4/7] Cleanup white spaces. --- driver/others/dynamic_power.c | 144 +++++++++++++++++----------------- 1 file changed, 72 insertions(+), 72 deletions(-) diff --git a/driver/others/dynamic_power.c b/driver/others/dynamic_power.c index 3c5f1f3c1..40f00a634 100644 --- a/driver/others/dynamic_power.c +++ b/driver/others/dynamic_power.c @@ -13,25 +13,25 @@ extern gotoblas_t gotoblas_POWER10; extern void openblas_warning(int verbose, const char *msg); static char *corename[] = { - "unknown", - "POWER6", - "POWER8", - "POWER9", - "POWER10" + "unknown", + "POWER6", + "POWER8", + "POWER9", + "POWER10" }; #define NUM_CORETYPES 5 char *gotoblas_corename(void) { - if (gotoblas == &gotoblas_POWER6) return corename[1]; - if (gotoblas == &gotoblas_POWER8) return corename[2]; + if (gotoblas == &gotoblas_POWER6) return corename[1]; + if (gotoblas == &gotoblas_POWER8) return corename[2]; #if (!defined __GNUC__) || ( __GNUC__ >= 6) - if (gotoblas == &gotoblas_POWER9) return corename[3]; + if (gotoblas == &gotoblas_POWER9) return corename[3]; #endif #ifdef HAVE_P10_SUPPORT - if (gotoblas == &gotoblas_POWER10) return corename[4]; + if (gotoblas == &gotoblas_POWER10) return corename[4]; #endif - return corename[0]; + return corename[0]; } #ifdef _AIX @@ -90,13 +90,13 @@ static int __builtin_cpu_supports(char* arg) static gotoblas_t *get_coretype(void) { - if (__builtin_cpu_supports("power6")) - return &gotoblas_POWER6; - if (__builtin_cpu_supports("power8")) - return &gotoblas_POWER8; + if (__builtin_cpu_supports("power6")) + return &gotoblas_POWER6; + if (__builtin_cpu_supports("power8")) + return &gotoblas_POWER8; #if (!defined __GNUC__) || ( __GNUC__ >= 6) - if (__builtin_cpu_supports("power9")) - return &gotoblas_POWER9; + if (__builtin_cpu_supports("power9")) + return &gotoblas_POWER9; #endif #ifdef HAVE_P10_SUPPORT #ifdef _AIX @@ -104,84 +104,84 @@ static gotoblas_t *get_coretype(void) { #else if (__builtin_cpu_supports("arch_3_1") && __builtin_cpu_supports("mma")) #endif - return &gotoblas_POWER10; + return &gotoblas_POWER10; #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 (__builtin_cpu_supports("power10")) - return &gotoblas_POWER9; -#endif - return NULL; + if (__builtin_cpu_supports("power10")) + return &gotoblas_POWER9; +#endif + return NULL; } static gotoblas_t *force_coretype(char * coretype) { - int i ; - int found = -1; - char message[128]; + int i ; + int found = -1; + char message[128]; - for ( i = 0 ; i < NUM_CORETYPES; i++) - { - if (!strncasecmp(coretype, corename[i], 20)) - { - found = i; - break; - } - } + for ( i = 0 ; i < NUM_CORETYPES; i++) + { + if (!strncasecmp(coretype, corename[i], 20)) + { + found = i; + break; + } + } - switch (found) - { - case 1: return (&gotoblas_POWER6); - case 2: return (&gotoblas_POWER8); + switch (found) + { + case 1: return (&gotoblas_POWER6); + case 2: return (&gotoblas_POWER8); #if (!defined __GNUC__) || ( __GNUC__ >= 6) - case 3: return (&gotoblas_POWER9); + case 3: return (&gotoblas_POWER9); #endif #ifdef HAVE_P10_SUPPORT - case 4: return (&gotoblas_POWER10); + case 4: return (&gotoblas_POWER10); #endif - default: return NULL; - } - snprintf(message, 128, "Core not found: %s\n", coretype); - openblas_warning(1, message); + default: return NULL; + } + snprintf(message, 128, "Core not found: %s\n", coretype); + openblas_warning(1, message); } void gotoblas_dynamic_init(void) { - char coremsg[128]; - char coren[22]; - char *p; + char coremsg[128]; + char coren[22]; + char *p; - if (gotoblas) return; + if (gotoblas) return; - p = getenv("OPENBLAS_CORETYPE"); - if ( p ) - { - gotoblas = force_coretype(p); - } - else - { - gotoblas = get_coretype(); - } + p = getenv("OPENBLAS_CORETYPE"); + if ( p ) + { + gotoblas = force_coretype(p); + } + else + { + gotoblas = get_coretype(); + } - if (gotoblas == NULL) - { - snprintf(coremsg, 128, "Falling back to POWER8 core\n"); - openblas_warning(1, coremsg); - gotoblas = &gotoblas_POWER8; - } + if (gotoblas == NULL) + { + snprintf(coremsg, 128, "Falling back to POWER8 core\n"); + openblas_warning(1, coremsg); + gotoblas = &gotoblas_POWER8; + } - if (gotoblas && gotoblas -> init) { - strncpy(coren,gotoblas_corename(),20); - sprintf(coremsg, "Core: %s\n",coren); - openblas_warning(2, coremsg); - gotoblas -> init(); - } else { - openblas_warning(0, "OpenBLAS : Architecture Initialization failed. No initialization function found.\n"); - exit(1); - } + if (gotoblas && gotoblas -> init) { + strncpy(coren,gotoblas_corename(),20); + sprintf(coremsg, "Core: %s\n",coren); + openblas_warning(2, coremsg); + gotoblas -> init(); + } else { + openblas_warning(0, "OpenBLAS : Architecture Initialization failed. No initialization function found.\n"); + exit(1); + } } void gotoblas_dynamic_quit(void) { - gotoblas = NULL; + gotoblas = NULL; } From 10210748de17a217fd67f6cb8501272b8bfa88c2 Mon Sep 17 00:00:00 2001 From: Chip-Kerchner Date: Mon, 2 Oct 2023 12:44:07 -0500 Subject: [PATCH 5/7] Revert PGI changes. --- driver/others/dynamic_power.c | 308 +++++++++++++++++++++++++--------- 1 file changed, 233 insertions(+), 75 deletions(-) diff --git a/driver/others/dynamic_power.c b/driver/others/dynamic_power.c index 40f00a634..0f5b06be5 100644 --- a/driver/others/dynamic_power.c +++ b/driver/others/dynamic_power.c @@ -13,27 +13,181 @@ extern gotoblas_t gotoblas_POWER10; extern void openblas_warning(int verbose, const char *msg); static char *corename[] = { - "unknown", - "POWER6", - "POWER8", - "POWER9", - "POWER10" + "unknown", + "POWER6", + "POWER8", + "POWER9", + "POWER10" }; #define NUM_CORETYPES 5 char *gotoblas_corename(void) { - if (gotoblas == &gotoblas_POWER6) return corename[1]; - if (gotoblas == &gotoblas_POWER8) return corename[2]; +#ifndef C_PGI + if (gotoblas == &gotoblas_POWER6) return corename[1]; +#endif + if (gotoblas == &gotoblas_POWER8) return corename[2]; #if (!defined __GNUC__) || ( __GNUC__ >= 6) - if (gotoblas == &gotoblas_POWER9) return corename[3]; + if (gotoblas == &gotoblas_POWER9) return corename[3]; #endif #ifdef HAVE_P10_SUPPORT - if (gotoblas == &gotoblas_POWER10) return corename[4]; + if (gotoblas == &gotoblas_POWER10) return corename[4]; #endif - return corename[0]; + return corename[0]; } +#if defined(__clang__) +static int __builtin_cpu_supports(char* arg) +{ + return 0; +} +#endif + +#if defined(C_PGI) || defined(__clang__) +/* + * NV HPC compilers do not yet implement __builtin_cpu_is(). + * Fake a version here for use in the CPU detection code below. + * + * 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 + +/* + * 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 + +static struct { + uint32_t pvr_mask; + uint32_t pvr_value; + const char* cpu_name; + uint32_t cpu_type; +} 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 + cpu_type = pvrPOWER[i].cpu_type; + + if (!strcmp(cpu, "power8")) + return cpu_type == CPU_POWER8; + if (!strcmp(cpu, "power9")) + return cpu_type == CPU_POWER9; + return 0; +} + +#endif /* C_PGI */ + #ifdef _AIX #include @@ -90,98 +244,102 @@ static int __builtin_cpu_supports(char* arg) static gotoblas_t *get_coretype(void) { - if (__builtin_cpu_supports("power6")) - return &gotoblas_POWER6; - if (__builtin_cpu_supports("power8")) - return &gotoblas_POWER8; +#ifndef C_PGI + if (__builtin_cpu_is("power6") || __builtin_cpu_is("power6x")) + return &gotoblas_POWER6; +#endif + if (__builtin_cpu_is("power8")) + return &gotoblas_POWER8; #if (!defined __GNUC__) || ( __GNUC__ >= 6) - if (__builtin_cpu_supports("power9")) - return &gotoblas_POWER9; + if (__builtin_cpu_is("power9")) + return &gotoblas_POWER9; #endif #ifdef HAVE_P10_SUPPORT #ifdef _AIX - if (__builtin_cpu_supports("power10")) + if (__builtin_cpu_supports("power10")) #else - if (__builtin_cpu_supports("arch_3_1") && __builtin_cpu_supports("mma")) + if (__builtin_cpu_supports ("arch_3_1") && __builtin_cpu_supports ("mma")) #endif - return &gotoblas_POWER10; + return &gotoblas_POWER10; #endif - /* 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 (__builtin_cpu_supports("power10")) - return &gotoblas_POWER9; -#endif - return NULL; + /* 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 (__builtin_cpu_is("power10")) + return &gotoblas_POWER9; +#endif + return NULL; } static gotoblas_t *force_coretype(char * coretype) { - int i ; - int found = -1; - char message[128]; + int i ; + int found = -1; + char message[128]; - for ( i = 0 ; i < NUM_CORETYPES; i++) - { - if (!strncasecmp(coretype, corename[i], 20)) - { - found = i; - break; - } - } + for ( i = 0 ; i < NUM_CORETYPES; i++) + { + if (!strncasecmp(coretype, corename[i], 20)) + { + found = i; + break; + } + } - switch (found) - { - case 1: return (&gotoblas_POWER6); - case 2: return (&gotoblas_POWER8); + switch (found) + { +#ifndef C_PGI + case 1: return (&gotoblas_POWER6); +#endif + case 2: return (&gotoblas_POWER8); #if (!defined __GNUC__) || ( __GNUC__ >= 6) - case 3: return (&gotoblas_POWER9); + case 3: return (&gotoblas_POWER9); #endif #ifdef HAVE_P10_SUPPORT - case 4: return (&gotoblas_POWER10); + case 4: return (&gotoblas_POWER10); #endif - default: return NULL; - } - snprintf(message, 128, "Core not found: %s\n", coretype); - openblas_warning(1, message); + default: return NULL; + } + snprintf(message, 128, "Core not found: %s\n", coretype); + openblas_warning(1, message); } void gotoblas_dynamic_init(void) { - char coremsg[128]; - char coren[22]; - char *p; + char coremsg[128]; + char coren[22]; + char *p; - if (gotoblas) return; + if (gotoblas) return; - p = getenv("OPENBLAS_CORETYPE"); - if ( p ) - { - gotoblas = force_coretype(p); - } - else - { - gotoblas = get_coretype(); - } + p = getenv("OPENBLAS_CORETYPE"); + if ( p ) + { + gotoblas = force_coretype(p); + } + else + { + gotoblas = get_coretype(); + } - if (gotoblas == NULL) - { - snprintf(coremsg, 128, "Falling back to POWER8 core\n"); - openblas_warning(1, coremsg); - gotoblas = &gotoblas_POWER8; - } + if (gotoblas == NULL) + { + snprintf(coremsg, 128, "Falling back to POWER8 core\n"); + openblas_warning(1, coremsg); + gotoblas = &gotoblas_POWER8; + } - if (gotoblas && gotoblas -> init) { - strncpy(coren,gotoblas_corename(),20); - sprintf(coremsg, "Core: %s\n",coren); - openblas_warning(2, coremsg); - gotoblas -> init(); - } else { - openblas_warning(0, "OpenBLAS : Architecture Initialization failed. No initialization function found.\n"); - exit(1); - } + if (gotoblas && gotoblas -> init) { + strncpy(coren,gotoblas_corename(),20); + sprintf(coremsg, "Core: %s\n",coren); + openblas_warning(2, coremsg); + gotoblas -> init(); + } else { + openblas_warning(0, "OpenBLAS : Architecture Initialization failed. No initialization function found.\n"); + exit(1); + } } void gotoblas_dynamic_quit(void) { - gotoblas = NULL; + gotoblas = NULL; } From e5dc376912dab278afdf677cb112008d36ead0fe Mon Sep 17 00:00:00 2001 From: Chip-Kerchner Date: Mon, 2 Oct 2023 12:48:47 -0500 Subject: [PATCH 6/7] Remove duplicate defines. --- driver/others/dynamic_power.c | 21 +++++++-------------- 1 file changed, 7 insertions(+), 14 deletions(-) diff --git a/driver/others/dynamic_power.c b/driver/others/dynamic_power.c index 0f5b06be5..7b0b4ea01 100644 --- a/driver/others/dynamic_power.c +++ b/driver/others/dynamic_power.c @@ -43,6 +43,13 @@ static int __builtin_cpu_supports(char* arg) } #endif +#define CPU_UNKNOWN 0 +#define CPU_POWER6 6 +#define CPU_POWER7 7 +#define CPU_POWER8 8 +#define CPU_POWER9 9 +#define CPU_POWER10 10 + #if defined(C_PGI) || defined(__clang__) /* * NV HPC compilers do not yet implement __builtin_cpu_is(). @@ -61,13 +68,6 @@ static int __builtin_cpu_supports(char* arg) * 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 - static struct { uint32_t pvr_mask; uint32_t pvr_value; @@ -191,13 +191,6 @@ static int __builtin_cpu_is(const char *cpu) { #ifdef _AIX #include -#define CPU_UNKNOWN 0 -#define CPU_POWER6 6 -#define CPU_POWER7 7 -#define CPU_POWER8 8 -#define CPU_POWER9 9 -#define CPU_POWER10 10 - static int cpuid(void) { int arch = _system_configuration.implementation; From b677d0d5fd175768e63d02253b12d1b0ccb2d242 Mon Sep 17 00:00:00 2001 From: Chip-Kerchner Date: Mon, 2 Oct 2023 13:09:12 -0500 Subject: [PATCH 7/7] Adding missing endif --- driver/others/dynamic_power.c | 1 + 1 file changed, 1 insertion(+) diff --git a/driver/others/dynamic_power.c b/driver/others/dynamic_power.c index 7b0b4ea01..6ed26ad1e 100644 --- a/driver/others/dynamic_power.c +++ b/driver/others/dynamic_power.c @@ -234,6 +234,7 @@ static int __builtin_cpu_supports(char* arg) return 0; } #endif +#endif static gotoblas_t *get_coretype(void) {