Browse Source

Add initial support for M1 on Linux, Phytium FT2xxx series, ARM Cortex 510/710/X1/X2

tags/v0.3.21
Martin Kroeker GitHub 3 years ago
parent
commit
09b8545fc5
No known key found for this signature in database GPG Key ID: 4AEE18F83AFDEB23
6 changed files with 193 additions and 25 deletions
  1. +44
    -0
      Makefile.arm64
  2. +7
    -0
      TargetList.txt
  3. +1
    -0
      c_check
  4. +67
    -9
      cpuid_arm64.c
  5. +72
    -14
      getarch.c
  6. +2
    -2
      param.h

+ 44
- 0
Makefile.arm64 View File

@@ -55,6 +55,13 @@ FCOMMON_OPT += -march=armv8-a -mtune=cortex-a73
endif
endif

ifeq ($(CORE), FT2000)
CCOMMON_OPT += -march=armv8-a -mtune=cortex-a72
ifneq ($(F_COMPILER), NAG)
FCOMMON_OPT += -march=armv8-a -mtune=cortex-a72
endif
endif

# Use a72 tunings because Neoverse-N1 is only available
# in GCC>=9
ifeq ($(CORE), NEOVERSEN1)
@@ -229,6 +236,43 @@ endif
endif
endif

ifeq (1, $(filter 1,$(GCCVERSIONGTEQ11) $(ISCLANG)))
ifeq ($(CORE), CORTEXX1)
CCOMMON_OPT += -march=armv9 -mtune=cortexx1
ifneq ($(F_COMPILER), NAG)
FCOMMON_OPT += -march=armv9 -mtune=cortexx1
endif
endif
endif

ifeq (1, $(filter 1,$(GCCVERSIONGTEQ11) $(ISCLANG)))
ifeq ($(CORE), CORTEXX2)
CCOMMON_OPT += -march=armv9 -mtune=cortexx2
ifneq ($(F_COMPILER), NAG)
FCOMMON_OPT += -march=armv9 -mtune=cortexx2
endif
endif
endif

#ifeq (1, $(filter 1,$(ISCLANG)))
ifeq (1, $(filter 1,$(GCCVERSIONGTEQ11) $(ISCLANG)))
ifeq ($(CORE), CORTEXA510)
CCOMMON_OPT += -march=armv8.4-a+sve
ifneq ($(F_COMPILER), NAG)
FCOMMON_OPT += -march=armv8.4-a+sve
endif
endif
endif

ifeq (1, $(filter 1,$(GCCVERSIONGTEQ11) $(ISCLANG)))
ifeq ($(CORE), CORTEXA710)
CCOMMON_OPT += -march=armv8.2-a+sve -mtune=cortexa710
ifneq ($(F_COMPILER), NAG)
FCOMMON_OPT += -march=armv8.2-a+sve -mtune=cortexa710
endif
endif
endif

endif

endif

+ 7
- 0
TargetList.txt View File

@@ -92,6 +92,10 @@ CORTEXA53
CORTEXA57
CORTEXA72
CORTEXA73
CORTEXA510
CORTEXA710
CORTEXX1
CORTEXX2
NEOVERSEN1
NEOVERSEV1
NEOVERSEN2
@@ -103,6 +107,9 @@ THUNDERX2T99
TSV110
THUNDERX3T110
VORTEX
A64FX
ARMV8SVE
FT2000

9.System Z:
ZARCH_GENERIC


+ 1
- 0
c_check View File

@@ -316,6 +316,7 @@ if ($architecture ne $hostarch) {
}

$cross = 1 if ($os ne $hostos);
$cross = 0 if (($os eq "Android") && ($hostos eq "Linux") && ($ENV{TERMUX_APP_PID} != ""));

$openmp = "" if $ENV{USE_OPENMP} != 1;



+ 67
- 9
cpuid_arm64.c View File

@@ -45,6 +45,10 @@ size_t length64=sizeof(value64);
#define CPU_NEOVERSEN1 11
#define CPU_NEOVERSEV1 16
#define CPU_NEOVERSEN2 17
#define CPU_CORTEXX1 18
#define CPU_CORTEXX2 19
#define CPU_CORTEXA510 20
#define CPU_CORTEXA710 21
// Qualcomm
#define CPU_FALKOR 6
// Cavium
@@ -59,6 +63,8 @@ size_t length64=sizeof(value64);
#define CPU_VORTEX 13
// Fujitsu
#define CPU_A64FX 15
// Phytium
#define CPU_FT2000 22

static char *cpuname[] = {
"UNKNOWN",
@@ -73,12 +79,17 @@ static char *cpuname[] = {
"TSV110",
"EMAG8180",
"NEOVERSEN1",
"NEOVERSEV1"
"NEOVERSEN2"
"THUNDERX3T110",
"VORTEX",
"CORTEXA55",
"A64FX"
"A64FX",
"NEOVERSEV1",
"NEOVERSEN2",
"CORTEXX1",
"CORTEXX2",
"CORTEXA510",
"CORTEXA710",
"FT2000"
};

static char *cpuname_lower[] = {
@@ -94,12 +105,17 @@ static char *cpuname_lower[] = {
"tsv110",
"emag8180",
"neoversen1",
"neoversev1",
"neoversen2",
"thunderx3t110",
"vortex",
"cortexa55",
"a64fx"
"a64fx",
"neoversev1",
"neoversen2",
"cortexx1",
"cortexx2",
"cortexa510",
"cortexa710",
"ft2000"
};

int get_feature(char *search)
@@ -182,6 +198,14 @@ int detect(void)
return CPU_NEOVERSEN2;
else if (strstr(cpu_part, "0xd05"))
return CPU_CORTEXA55;
else if (strstr(cpu_part, "0xd46"))
return CPU_CORTEXA510;
else if (strstr(cpu_part, "0xd47"))
return CPU_CORTEXA710;
else if (strstr(cpu_part, "0xd44"))
return CPU_CORTEXX1;
else if (strstr(cpu_part, "0xd4c"))
return CPU_CORTEXX2;
}
// Qualcomm
else if (strstr(cpu_implementer, "0x51") && strstr(cpu_part, "0xc00"))
@@ -202,6 +226,13 @@ int detect(void)
// Fujitsu
else if (strstr(cpu_implementer, "0x46") && strstr(cpu_part, "0x001"))
return CPU_A64FX;
// Apple
else if (strstr(cpu_implementer, "0x61") && strstr(cpu_part, "0x022"))
return CPU_VORTEX;
// Phytium
else if (strstr(cpu_implementer, "0x70") && (strstr(cpu_part, "0x660") || strstr(cpu_part, "0x661")
|| strstr(cpu_part, "0x662") || strstr(cpu_part, "0x663")))
return CPU_FT2000;
}

p = (char *) NULL ;
@@ -382,7 +413,24 @@ void get_cpuconfig(void)
printf("#define DTB_DEFAULT_ENTRIES 48\n");
printf("#define DTB_SIZE 4096\n");
break;

case CPU_CORTEXA510:
case CPU_CORTEXA710:
case CPU_CORTEXX1:
case CPU_CORTEXX2:
printf("#define ARMV9\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");
printf("#define DTB_DEFAULT_ENTRIES 64\n");
printf("#define DTB_SIZE 4096\n");
break;
case CPU_FALKOR:
printf("#define FALKOR\n");
printf("#define L1_CODE_SIZE 65536\n");
@@ -469,9 +517,9 @@ void get_cpuconfig(void)
printf("#define DTB_DEFAULT_ENTRIES 64 \n");
printf("#define DTB_SIZE 4096 \n");
break;
#ifdef __APPLE__
case CPU_VORTEX:
printf("#define VORTEX \n");
#ifdef __APPLE__
sysctlbyname("hw.l1icachesize",&value64,&length64,NULL,0);
printf("#define L1_CODE_SIZE %lld \n",value64);
sysctlbyname("hw.cachelinesize",&value64,&length64,NULL,0);
@@ -480,10 +528,10 @@ void get_cpuconfig(void)
printf("#define L1_DATA_SIZE %lld \n",value64);
sysctlbyname("hw.l2cachesize",&value64,&length64,NULL,0);
printf("#define L2_SIZE %lld \n",value64);
#endif
printf("#define DTB_DEFAULT_ENTRIES 64 \n");
printf("#define DTB_SIZE 4096 \n");
break;
#endif
case CPU_A64FX:
printf("#define A64FX\n");
printf("#define L1_CODE_SIZE 65535\n");
@@ -494,6 +542,16 @@ void get_cpuconfig(void)
printf("#define DTB_DEFAULT_ENTRIES 64\n");
printf("#define DTB_SIZE 4096\n");
break;
case CPU_FT2000:
printf("#define FT2000\n");
printf("#define L1_CODE_SIZE 32768\n");
printf("#define L1_DATA_SIZE 32768\n");
printf("#define L1_DATA_LINESIZE 64\n");
printf("#define L2_SIZE 33554432\n");
printf("#define L2_LINESIZE 64\n");
printf("#define DTB_DEFAULT_ENTRIES 64\n");
printf("#define DTB_SIZE 4096\n");
break;
}
get_cpucount();
}


+ 72
- 14
getarch.c View File

@@ -1232,7 +1232,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8"
#define LIBNAME "cortexa53"
#define CORENAME "CORTEXA53"
#else
#endif

#ifdef FORCE_CORTEXA57
@@ -1248,7 +1247,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8"
#define LIBNAME "cortexa57"
#define CORENAME "CORTEXA57"
#else
#endif

#ifdef FORCE_CORTEXA72
@@ -1264,7 +1262,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8"
#define LIBNAME "cortexa72"
#define CORENAME "CORTEXA72"
#else
#endif

#ifdef FORCE_CORTEXA73
@@ -1280,7 +1277,62 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8"
#define LIBNAME "cortexa73"
#define CORENAME "CORTEXA73"
#else
#endif

#ifdef FORCE_CORTEXX1
#define FORCE
#define ARCHITECTURE "ARM64"
#define SUBARCHITECTURE "CORTEXX1"
#define SUBDIRNAME "arm64"
#define ARCHCONFIG "-DCORTEXX1 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=262144 -DL2_LINESIZE=64 " \
"-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=32 " \
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DHAVE_SVE -DARMV8 -DARMV9"
#define LIBNAME "cortexx1"
#define CORENAME "CORTEXX1"
#endif

#ifdef FORCE_CORTEXX2
#define FORCE
#define ARCHITECTURE "ARM64"
#define SUBARCHITECTURE "CORTEXX2"
#define SUBDIRNAME "arm64"
#define ARCHCONFIG "-DCORTEXX2 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=262144 -DL2_LINESIZE=64 " \
"-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=32 " \
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DHAVE_SVE -DARMV8 -DARMV9"
#define LIBNAME "cortexx2"
#define CORENAME "CORTEXX2"
#endif

#ifdef FORCE_CORTEXA510
#define FORCE
#define ARCHITECTURE "ARM64"
#define SUBARCHITECTURE "CORTEXA510"
#define SUBDIRNAME "arm64"
#define ARCHCONFIG "-DCORTEXA510 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=262144 -DL2_LINESIZE=64 " \
"-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=32 " \
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DHAVE_SVE -DARMV8 -DARMV9"
#define LIBNAME "cortexa510"
#define CORENAME "CORTEXA510"
#endif

#ifdef FORCE_CORTEXA710
#define FORCE
#define ARCHITECTURE "ARM64"
#define SUBARCHITECTURE "CORTEXA710"
#define SUBDIRNAME "arm64"
#define ARCHCONFIG "-DCORTEXA710 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 " \
"-DL2_SIZE=262144 -DL2_LINESIZE=64 " \
"-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 -DL2_ASSOCIATIVE=32 " \
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DHAVE_SVE -DARMV8 -DARMV9"
#define LIBNAME "cortexa710"
#define CORENAME "CORTEXA710"
#endif

#ifdef FORCE_NEOVERSEN1
@@ -1297,7 +1349,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-march=armv8.2-a -mtune=neoverse-n1"
#define LIBNAME "neoversen1"
#define CORENAME "NEOVERSEN1"
#else
#endif

#ifdef FORCE_NEOVERSEV1
@@ -1314,7 +1365,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-march=armv8.4-a -mtune=neoverse-v1"
#define LIBNAME "neoversev1"
#define CORENAME "NEOVERSEV1"
#else
#endif


@@ -1332,7 +1382,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-march=armv8.5-a -mtune=neoverse-n2"
#define LIBNAME "neoversen2"
#define CORENAME "NEOVERSEN2"
#else
#endif

#ifdef FORCE_CORTEXA55
@@ -1348,7 +1397,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8"
#define LIBNAME "cortexa55"
#define CORENAME "CORTEXA55"
#else
#endif

#ifdef FORCE_FALKOR
@@ -1364,7 +1412,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8"
#define LIBNAME "falkor"
#define CORENAME "FALKOR"
#else
#endif

#ifdef FORCE_THUNDERX
@@ -1379,7 +1426,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8"
#define LIBNAME "thunderx"
#define CORENAME "THUNDERX"
#else
#endif

#ifdef FORCE_THUNDERX2T99
@@ -1397,7 +1443,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8"
#define LIBNAME "thunderx2t99"
#define CORENAME "THUNDERX2T99"
#else
#endif

#ifdef FORCE_TSV110
@@ -1413,7 +1458,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8"
#define LIBNAME "tsv110"
#define CORENAME "TSV110"
#else
#endif

#ifdef FORCE_EMAG8180
@@ -1448,7 +1492,6 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8"
#define LIBNAME "thunderx3t110"
#define CORENAME "THUNDERX3T110"
#else
#endif

#ifdef FORCE_VORTEX
@@ -1480,7 +1523,22 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DHAVE_SVE -DARMV8"
#define LIBNAME "a64fx"
#define CORENAME "A64FX"
#else
#endif

#ifdef FORCE_FT2000
#define ARMV8
#define FORCE
#define ARCHITECTURE "ARM64"
#define SUBARCHITECTURE "FT2000"
#define SUBDIRNAME "arm64"
#define ARCHCONFIG "-DFT2000 " \
"-DL1_CODE_SIZE=32768 -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=8 " \
"-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=64 -DL1_DATA_ASSOCIATIVE=8 " \
"-DL2_SIZE=33554426-DL2_LINESIZE=64 -DL2_ASSOCIATIVE=8 " \
"-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
"-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8"
#define LIBNAME "ft2000"
#define CORENAME "FT2000"
#endif

#ifdef FORCE_ZARCH_GENERIC


+ 2
- 2
param.h View File

@@ -3130,7 +3130,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.

#if defined(CORTEXA57) || \
defined(CORTEXA72) || defined(CORTEXA73) || \
defined(FALKOR) || defined(TSV110) || defined(EMAG8180) || defined(VORTEX)
defined(FALKOR) || defined(TSV110) || defined(EMAG8180) || defined(VORTEX) || defined(FT2000)

#define SGEMM_DEFAULT_UNROLL_M 16
#define SGEMM_DEFAULT_UNROLL_N 4
@@ -3377,7 +3377,7 @@ is a big desktop or server with abundant cache rather than a phone or embedded d
#define CGEMM_DEFAULT_R 4096
#define ZGEMM_DEFAULT_R 4096

#elif defined(ARMV8SVE) || defined(A64FX)
#elif defined(ARMV8SVE) || defined(A64FX) || defined(ARMV9) || defined(CORTEXA510)

/* When all BLAS3 routines are implemeted with SVE, SGEMM_DEFAULT_UNROLL_M should be "sve_vl".
Until then, just keep it different than DGEMM_DEFAULT_UNROLL_N to keep copy routines in both directions seperated. */


Loading…
Cancel
Save