Merge pull request OpenMathLib#1061 from ashwinyes/develop_aarch64_vu…

…lcan_thunderx_patch

Add new targets for ARM64

Makefile.arm64

@@ -9,3 +9,17 @@ CCOMMON_OPT += -march=armv8-a+crc+crypto+fp+simd -mtune=cortex-a57
 FCOMMON_OPT += -march=armv8-a+crc+crypto+fp+simd -mtune=cortex-a57
 endif
 
+ifeq ($(CORE), VULCAN)
+CCOMMON_OPT += -mtune=vulcan -mcpu=vulcan
+FCOMMON_OPT += -mtune=vulcan -mcpu=vulcan
+endif
+
+ifeq ($(CORE), THUNDERX)
+CCOMMON_OPT += -mtune=thunderx -mcpu=thunderx
+FCOMMON_OPT += -mtune=thunderx -mcpu=thunderx
+endif
+
+ifeq ($(CORE), THUNDERX2T99)
+CCOMMON_OPT += -mtune=vulcan -mcpu=vulcan
+FCOMMON_OPT += -mtune=vulcan -mcpu=vulcan
+endif

TargetList.txt

@@ -80,4 +80,7 @@ ARMV5
 8.ARM 64-bit CPU:
 ARMV8
 CORTEXA57
+VULCAN
+THUNDERX
+THUNDERX2T99
 

common_macro.h

@@ -2193,7 +2193,7 @@
 #endif
 
 #ifndef ASSEMBLER
-#if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64)
+#if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64) || defined(ARCH_ARM64)
 extern BLASLONG gemm_offset_a;
 extern BLASLONG gemm_offset_b;
 extern BLASLONG sgemm_p;

cpuid_arm64.c

@@ -30,17 +30,26 @@
 #define CPU_UNKNOWN     	0
 #define CPU_ARMV8       	1
 #define CPU_CORTEXA57       	2
+#define CPU_VULCAN       	3
+#define CPU_THUNDERX    	4
+#define CPU_THUNDERX2T99   	5
 
 static char *cpuname[] = {
   "UNKNOWN",
   "ARMV8" ,
-  "CORTEXA57"
+  "CORTEXA57",
+  "VULCAN",
+  "THUNDERX",
+  "THUNDERX2T99"
 };
 
 static char *cpuname_lower[] = {
   "unknown",
   "armv8" ,
-  "cortexa57"
+  "cortexa57",
+  "vulcan",
+  "thunderx",
+  "thunderx2t99"
 };
 
 int get_feature(char *search)
@@ -85,25 +94,34 @@ int detect(void)
 #ifdef linux
 
 	FILE *infile;
-  	char buffer[512], *p;
-  	p = (char *) NULL ;
-
-  	infile = fopen("/proc/cpuinfo", "r");
-	while (fgets(buffer, sizeof(buffer), infile))
-	{
+	char buffer[512], *p, *cpu_part = NULL, *cpu_implementer = NULL;
+	p = (char *) NULL ;
 
-		if (!strncmp("CPU part", buffer, 8))
-		{
-			p = strchr(buffer, ':') + 2;
+	infile = fopen("/proc/cpuinfo", "r");
+	while (fgets(buffer, sizeof(buffer), infile)) {
+		if ((cpu_part != NULL) && (cpu_implementer != NULL)) {
 			break;
 		}
+
+		if ((cpu_part == NULL) && !strncmp("CPU part", buffer, 8)) {
+			cpu_part = strchr(buffer, ':') + 2;
+			cpu_part = strdup(cpu_part);
+		} else if ((cpu_implementer == NULL) && !strncmp("CPU implementer", buffer, 15)) {
+			cpu_implementer = strchr(buffer, ':') + 2;
+			cpu_implementer = strdup(cpu_implementer);
+		}
 	}
 
 	fclose(infile);
-	if(p != NULL) {
-	  if (strstr(p, "0xd07")) {
-	    return CPU_CORTEXA57;
-	  }
+	if(cpu_part != NULL && cpu_implementer != NULL) {
+		if (strstr(cpu_part, "0xd07") && strstr(cpu_implementer, "0x41"))
+			return CPU_CORTEXA57;
+		else if (strstr(cpu_part, "0x516") && strstr(cpu_implementer, "0x42"))
+			return CPU_VULCAN;
+		else if (strstr(cpu_part, "0x0a1") && strstr(cpu_implementer, "0x43"))
+			return CPU_THUNDERX;
+		else if (strstr(cpu_part, "0xFFF") && strstr(cpu_implementer, "0x43")) /* TODO */
+			return CPU_THUNDERX2T99;
 	}
 
 	p = (char *) NULL ;
@@ -176,6 +194,28 @@ void get_cpuconfig(void)
     			printf("#define L2_ASSOCIATIVE 4\n");
 			break;
 
+		case CPU_VULCAN:
+			printf("#define VULCAN                        \n");
+			printf("#define HAVE_VFP                      \n");
+			printf("#define HAVE_VFPV3                    \n");
+			printf("#define HAVE_NEON                     \n");
+			printf("#define HAVE_VFPV4                    \n");
+			printf("#define L1_CODE_SIZE         32768    \n");
+			printf("#define L1_CODE_LINESIZE     64       \n");
+			printf("#define L1_CODE_ASSOCIATIVE  8        \n");
+			printf("#define L1_DATA_SIZE         32768    \n");
+			printf("#define L1_DATA_LINESIZE     64       \n");
+			printf("#define L1_DATA_ASSOCIATIVE  8        \n");
+			printf("#define L2_SIZE              262144   \n");
+			printf("#define L2_LINESIZE          64       \n");
+			printf("#define L2_ASSOCIATIVE       8        \n");
+			printf("#define L3_SIZE              33554432 \n");
+			printf("#define L3_LINESIZE          64       \n");
+			printf("#define L3_ASSOCIATIVE       32       \n");
+			printf("#define DTB_DEFAULT_ENTRIES  64       \n");
+			printf("#define DTB_SIZE             4096     \n");
+			break;
+
 		case CPU_CORTEXA57:
 			printf("#define CORTEXA57\n");
 			printf("#define HAVE_VFP\n");
@@ -191,8 +231,42 @@ void get_cpuconfig(void)
 			printf("#define L2_SIZE 2097152\n");
 			printf("#define L2_LINESIZE 64\n");
 			printf("#define L2_ASSOCIATIVE 16\n");
-    			printf("#define DTB_DEFAULT_ENTRIES 64\n");
-    			printf("#define DTB_SIZE 4096\n");
+			printf("#define DTB_DEFAULT_ENTRIES 64\n");
+			printf("#define DTB_SIZE 4096\n");
+			break;
+
+		case CPU_THUNDERX:
+			printf("#define ARMV8\n");
+			printf("#define THUNDERX\n");
+			printf("#define L1_DATA_SIZE 32768\n");
+			printf("#define L1_DATA_LINESIZE 128\n");
+			printf("#define L2_SIZE 16777216\n");
+			printf("#define L2_LINESIZE 128\n");
+			printf("#define DTB_DEFAULT_ENTRIES 64\n");
+			printf("#define DTB_SIZE 4096\n");
+			printf("#define L2_ASSOCIATIVE 16\n");
+			break;
+
+		case CPU_THUNDERX2T99:
+			printf("#define VULCAN                        \n");
+			printf("#define HAVE_VFP                      \n");
+			printf("#define HAVE_VFPV3                    \n");
+			printf("#define HAVE_NEON                     \n");
+			printf("#define HAVE_VFPV4                    \n");
+			printf("#define L1_CODE_SIZE         32768    \n");
+			printf("#define L1_CODE_LINESIZE     64       \n");
+			printf("#define L1_CODE_ASSOCIATIVE  8        \n");
+			printf("#define L1_DATA_SIZE         32768    \n");
+			printf("#define L1_DATA_LINESIZE     64       \n");
+			printf("#define L1_DATA_ASSOCIATIVE  8        \n");
+			printf("#define L2_SIZE              262144   \n");
+			printf("#define L2_LINESIZE          64       \n");
+			printf("#define L2_ASSOCIATIVE       8        \n");
+			printf("#define L3_SIZE              33554432 \n");
+			printf("#define L3_LINESIZE          64       \n");
+			printf("#define L3_ASSOCIATIVE       32       \n");
+			printf("#define DTB_DEFAULT_ENTRIES  64       \n");
+			printf("#define DTB_SIZE             4096     \n");
 			break;
 	}
 }

driver/others/memory.c

@@ -995,7 +995,7 @@ void *blas_memory_alloc(int procpos){
     if (!blas_num_threads) blas_cpu_number = blas_get_cpu_number();
 #endif
 
-#if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64)
+#if defined(ARCH_X86) || defined(ARCH_X86_64) || defined(ARCH_IA64) || defined(ARCH_MIPS64) || defined(ARCH_ARM64)
 #ifndef DYNAMIC_ARCH
     blas_set_parameter();
 #endif

driver/others/parameter.c

@@ -727,3 +727,26 @@ void blas_set_parameter(void){
 
 }
 #endif
+
+#if defined(ARCH_ARM64)
+
+#if defined(VULCAN) || defined(THUNDERX2T99)
+unsigned long dgemm_prefetch_size_a;
+unsigned long dgemm_prefetch_size_b;
+unsigned long dgemm_prefetch_size_c;
+#endif
+
+void blas_set_parameter(void)
+{
+#if defined(VULCAN) || defined(THUNDERX2T99)
+  dgemm_p = 160;
+  dgemm_q = 128;
+  dgemm_r = 4096;
+
+  dgemm_prefetch_size_a = 3584;
+  dgemm_prefetch_size_b = 512;
+  dgemm_prefetch_size_c = 128;
+#endif
+}
+
+#endif

getarch.c

@@ -884,7 +884,7 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #ifdef FORCE_CORTEXA57
 #define FORCE
 #define ARCHITECTURE    "ARM64"
-#define SUBARCHITECTURE "ARMV8"
+#define SUBARCHITECTURE "CORTEXA57"
 #define SUBDIRNAME      "arm64"
 #define ARCHCONFIG   "-DCORTEXA57 " \
        "-DL1_CODE_SIZE=49152 -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=3 " \
@@ -897,6 +897,54 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #else
 #endif
 
+#ifdef FORCE_VULCAN
+#define FORCE
+#define ARCHITECTURE    "ARM64"
+#define SUBARCHITECTURE "VULCAN"
+#define SUBDIRNAME      "arm64"
+#define ARCHCONFIG   "-DVULCAN " \
+       "-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=262144 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=8 " \
+       "-DL3_SIZE=33554432 -DL3_LINESIZE=64 -DL3_ASSOCIATIVE=32 " \
+       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
+       "-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON"
+#define LIBNAME   "vulcan"
+#define CORENAME  "VULCAN"
+#else
+#endif
+
+#ifdef FORCE_THUNDERX
+#define FORCE
+#define ARCHITECTURE    "ARM64"
+#define SUBARCHITECTURE "THUNDERX"
+#define SUBDIRNAME      "arm64"
+#define ARCHCONFIG   "-DTHUNDERX " \
+       "-DL1_DATA_SIZE=32768 -DL1_DATA_LINESIZE=128 " \
+       "-DL2_SIZE=16777216 -DL2_LINESIZE=128 -DL2_ASSOCIATIVE=16 " \
+       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 "
+#define LIBNAME   "thunderx"
+#define CORENAME  "THUNDERX"
+#else
+#endif
+
+#ifdef FORCE_THUNDERX2T99
+#define FORCE
+#define ARCHITECTURE    "ARM64"
+#define SUBARCHITECTURE "THUNDERX2T99"
+#define SUBDIRNAME      "arm64"
+#define ARCHCONFIG   "-DTHUNDERX2T99 " \
+       "-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=262144 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=8 " \
+       "-DL3_SIZE=33554432 -DL3_LINESIZE=64 -DL3_ASSOCIATIVE=32 " \
+       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
+       "-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON"
+#define LIBNAME   "thunderx2t99"
+#define CORENAME  "THUNDERX2T99"
+#else
+#endif
+
 #ifndef FORCE
 
 #if defined(__powerpc__) || defined(__powerpc) || defined(powerpc) || \

kernel/arm64/KERNEL.CORTEXA57

@@ -75,14 +75,29 @@ SGEMMOTCOPYOBJ =  sgemm_otcopy.o
 
 DGEMMKERNEL    =  dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S
 DTRMMKERNEL    =  dtrmm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S
+
 ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N))
+
+ifeq ($(DGEMM_UNROLL_M), 8)
+DGEMMINCOPY    =  dgemm_ncopy_$(DGEMM_UNROLL_M).S
+DGEMMITCOPY    =  dgemm_tcopy_$(DGEMM_UNROLL_M).S
+else
 DGEMMINCOPY    =  ../generic/gemm_ncopy_$(DGEMM_UNROLL_M).c
 DGEMMITCOPY    =  ../generic/gemm_tcopy_$(DGEMM_UNROLL_M).c
+endif
+
 DGEMMINCOPYOBJ =  dgemm_incopy.o
 DGEMMITCOPYOBJ =  dgemm_itcopy.o
 endif
+
+ifeq ($(DGEMM_UNROLL_N), 4)
+DGEMMONCOPY    =  dgemm_ncopy_$(DGEMM_UNROLL_N).S
+DGEMMOTCOPY    =  dgemm_tcopy_$(DGEMM_UNROLL_N).S
+else
 DGEMMONCOPY    =  ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c
 DGEMMOTCOPY    =  ../generic/gemm_tcopy_$(DGEMM_UNROLL_N).c
+endif
+
 DGEMMONCOPYOBJ =  dgemm_oncopy.o
 DGEMMOTCOPYOBJ =  dgemm_otcopy.o
 

kernel/arm64/KERNEL.THUNDERX

@@ -0,0 +1,6 @@
+include $(KERNELDIR)/KERNEL.ARMV8
+
+SDOTKERNEL=dot-thunderx.c
+DDOTKERNEL=ddot-thunderx.c
+DAXPYKERNEL=daxpy-thunderx.c
+

kernel/arm64/KERNEL.THUNDERX2T99

@@ -0,0 +1,2 @@
+include $(KERNELDIR)/KERNEL.VULCAN
+

kernel/arm64/KERNEL.VULCAN

@@ -0,0 +1,4 @@
+include $(KERNELDIR)/KERNEL.CORTEXA57
+
+DGEMMKERNEL    =  dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N)_vulcan.S
+