OpenBLAS: aarch64: Add neoverse-v1/n2 architecture specifics

Makefile.arm64

@@ -78,6 +78,66 @@ endif
 endif
 endif
 
+# Use a72 tunings because Neoverse-V1 is only available
+# in GCC>=9.4
+ifeq ($(CORE), NEOVERSEV1)
+ifeq (1, $(filter 1,$(GCCVERSIONGTEQ7) $(ISCLANG)))
+ifeq ($(GCCVERSIONGTEQ9), 1)
+ifeq (1, $(filter 1,$(GCCMINORVERSIONGTEQ4) $(GCCVERSIONGTEQ10)))
+CCOMMON_OPT += -march=armv8.4-a -mtune=neoverse-v1
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=armv8.4-a -mtune=neoverse-v1
+endif
+else
+CCOMMON_OPT += -march=armv8.4-a -mtune=native
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=armv8.4-a -mtune=native
+endif
+endif
+else
+CCOMMON_OPT += -march=armv8.2-a -mtune=cortex-a72
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=armv8.2-a -mtune=cortex-a72
+endif
+endif
+else
+CCOMMON_OPT += -march=armv8-a -mtune=cortex-a72
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=armv8-a -mtune=cortex-a72
+endif
+endif
+endif
+
+# Use a72 tunings because Neoverse-N2 is only available
+# in GCC>=9.4
+ifeq ($(CORE), NEOVERSEN2)
+ifeq (1, $(filter 1,$(GCCVERSIONGTEQ7) $(ISCLANG)))
+ifeq ($(GCCVERSIONGTEQ9), 1)
+ifeq (1, $(filter 1,$(GCCMINORVERSIONGTEQ4) $(GCCVERSIONGTEQ10)))
+CCOMMON_OPT += -march=armv8.5-a -mtune=neoverse-n2
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=armv8.5-a -mtune=neoverse-n2
+endif
+else
+CCOMMON_OPT += -march=armv8.5-a -mtune=native
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=armv8.5-a -mtune=native
+endif
+endif
+else
+CCOMMON_OPT += -march=armv8.2-a -mtune=cortex-a72
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=armv8.2-a -mtune=cortex-a72
+endif
+endif
+else
+CCOMMON_OPT += -march=armv8-a -mtune=cortex-a72
+ifneq ($(F_COMPILER), NAG)
+FCOMMON_OPT += -march=armv8-a -mtune=cortex-a72
+endif
+endif
+endif
+
 # Use a53 tunings because a55 is only available in GCC>=8.1
 ifeq ($(CORE), CORTEXA55)
 ifeq (1, $(filter 1,$(GCCVERSIONGTEQ7) $(ISCLANG)))

Makefile.system

@@ -374,6 +374,7 @@ else
 endif
 GCCMINORVERSIONGTEQ1 := $(shell expr `$(CC) $(GCCDUMPVERSION_PARAM) | cut -f2 -d.` \>= 1)
 GCCMINORVERSIONGTEQ2 := $(shell expr `$(CC) $(GCCDUMPVERSION_PARAM) | cut -f2 -d.` \>= 2)
+GCCMINORVERSIONGTEQ4 := $(shell expr `$(CC) $(GCCDUMPVERSION_PARAM) | cut -f2 -d.` \>= 4)
 GCCMINORVERSIONGTEQ7 := $(shell expr `$(CC) $(GCCDUMPVERSION_PARAM) | cut -f2 -d.` \>= 7)
 endif
 
@@ -654,6 +655,8 @@ DYNAMIC_CORE += CORTEXA57
 DYNAMIC_CORE += CORTEXA72
 DYNAMIC_CORE += CORTEXA73
 DYNAMIC_CORE += NEOVERSEN1
+DYNAMIC_CORE += NEOVERSEV1
+DYNAMIC_CORE += NEOVERSEN2
 DYNAMIC_CORE += CORTEXA55
 DYNAMIC_CORE += FALKOR
 DYNAMIC_CORE += THUNDERX

TargetList.txt

@@ -93,6 +93,8 @@ CORTEXA57
 CORTEXA72
 CORTEXA73
 NEOVERSEN1
+NEOVERSEV1
+NEOVERSEN2
 CORTEXA55
 EMAG8180
 FALKOR

cmake/arch.cmake

@@ -44,7 +44,7 @@ endif ()
 
 if (DYNAMIC_ARCH)
   if (ARM64)
-	  set(DYNAMIC_CORE ARMV8 CORTEXA53 CORTEXA55 CORTEXA57 CORTEXA72 CORTEXA73 FALKOR THUNDERX THUNDERX2T99 TSV110 EMAG8180 NEOVERSEN1 THUNDERX3T110)
+	  set(DYNAMIC_CORE ARMV8 CORTEXA53 CORTEXA55 CORTEXA57 CORTEXA72 CORTEXA73 FALKOR THUNDERX THUNDERX2T99 TSV110 EMAG8180 NEOVERSEN1 NEOVERSEV1 NEOVERSEN2 THUNDERX3T110)
     if (DYNAMIC_LIST)
 	    set(DYNAMIC_CORE ARMV8 ${DYNAMIC_LIST})
     endif ()

cmake/prebuild.cmake

@@ -237,6 +237,61 @@ endif ()
     set(ZGEMM_UNROLL_N 4)
     set(SYMV_P 16)
   elseif ("${TCORE}" STREQUAL "NEOVERSEN1")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_CODE_SIZE\t65536\n"
+      "#define L1_CODE_LINESIZE\t64\n"
+      "#define L1_CODE_ASSOCIATIVE\t4\n"
+      "#define L1_DATA_SIZE\t65536\n"
+      "#define L1_DATA_LINESIZE\t64\n"
+      "#define L1_DATA_ASSOCIATIVE\t4\n"
+      "#define L2_SIZE\t1048576\n\n"
+      "#define L2_LINESIZE\t64\n"
+      "#define L2_ASSOCIATIVE\t8\n"
+      "#define DTB_DEFAULT_ENTRIES\t48\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define HAVE_VFPV4\n"
+      "#define HAVE_VFPV3\n"
+      "#define HAVE_VFP\n"
+      "#define HAVE_NEON\n"
+      "#define ARMV8\n")
+    set(SGEMM_UNROLL_M 16)
+    set(SGEMM_UNROLL_N 4)
+    set(DGEMM_UNROLL_M 8)
+    set(DGEMM_UNROLL_N 4)
+    set(CGEMM_UNROLL_M 8)
+    set(CGEMM_UNROLL_N 4)
+    set(ZGEMM_UNROLL_M 4)
+    set(ZGEMM_UNROLL_N 4)
+    set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "NEOVERSEV1")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_CODE_SIZE\t65536\n"
+      "#define L1_CODE_LINESIZE\t64\n"
+      "#define L1_CODE_ASSOCIATIVE\t4\n"
+      "#define L1_DATA_SIZE\t65536\n"
+      "#define L1_DATA_LINESIZE\t64\n"
+      "#define L1_DATA_ASSOCIATIVE\t4\n"
+      "#define L2_SIZE\t1048576\n\n"
+      "#define L2_LINESIZE\t64\n"
+      "#define L2_ASSOCIATIVE\t8\n"
+      "#define DTB_DEFAULT_ENTRIES\t48\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define HAVE_VFPV4\n"
+      "#define HAVE_VFPV3\n"
+      "#define HAVE_VFP\n"
+      "#define HAVE_NEON\n"
+      "#define HAVE_SVE\n"
+      "#define ARMV8\n")
+    set(SGEMM_UNROLL_M 16)
+    set(SGEMM_UNROLL_N 4)
+    set(DGEMM_UNROLL_M 8)
+    set(DGEMM_UNROLL_N 4)
+    set(CGEMM_UNROLL_M 8)
+    set(CGEMM_UNROLL_N 4)
+    set(ZGEMM_UNROLL_M 4)
+    set(ZGEMM_UNROLL_N 4)
+    set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "NEOVERSEN2")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_CODE_SIZE\t65536\n"
       "#define L1_CODE_LINESIZE\t64\n"
@@ -246,13 +301,14 @@ endif ()
       "#define L1_DATA_ASSOCIATIVE\t2\n"
       "#define L2_SIZE\t1048576\n\n"
       "#define L2_LINESIZE\t64\n"
-      "#define L2_ASSOCIATIVE\t16\n"
-      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define L2_ASSOCIATIVE\t8\n"
+      "#define DTB_DEFAULT_ENTRIES\t48\n"
       "#define DTB_SIZE\t4096\n"
       "#define HAVE_VFPV4\n"
       "#define HAVE_VFPV3\n"
       "#define HAVE_VFP\n"
       "#define HAVE_NEON\n"
+      "#define HAVE_SVE\n"
       "#define ARMV8\n")
     set(SGEMM_UNROLL_M 16)
     set(SGEMM_UNROLL_N 4)

cpuid_arm64.c

@@ -43,6 +43,8 @@ size_t length64=sizeof(value64);
 #define CPU_CORTEXA72     4
 #define CPU_CORTEXA73     5
 #define CPU_NEOVERSEN1    11
+#define CPU_NEOVERSEV1    16
+#define CPU_NEOVERSEN2    17
 // Qualcomm
 #define CPU_FALKOR        6
 // Cavium
@@ -71,6 +73,8 @@ static char *cpuname[] = {
   "TSV110",
   "EMAG8180",
   "NEOVERSEN1",
+  "NEOVERSEV1"
+  "NEOVERSEN2"
   "THUNDERX3T110",
   "VORTEX",
   "CORTEXA55",
@@ -90,6 +94,8 @@ static char *cpuname_lower[] = {
   "tsv110",
   "emag8180",
   "neoversen1",
+  "neoversev1",
+  "neoversen2",
   "thunderx3t110",
   "vortex",
   "cortexa55",
@@ -170,6 +176,10 @@ int detect(void)
         return CPU_CORTEXA73;
       else if (strstr(cpu_part, "0xd0c"))
         return CPU_NEOVERSEN1;
+      else if (strstr(cpu_part, "0xd40"))
+        return CPU_NEOVERSEV1;
+      else if (strstr(cpu_part, "0xd49"))
+        return CPU_NEOVERSEN2;
       else if (strstr(cpu_part, "0xd05"))
 	return CPU_CORTEXA55;
     }
@@ -338,11 +348,41 @@ void get_cpuconfig(void)
 		printf("#define L1_DATA_ASSOCIATIVE 4\n");
 		printf("#define L2_SIZE 1048576\n");
 		printf("#define L2_LINESIZE 64\n");
-		printf("#define L2_ASSOCIATIVE 16\n");
-		printf("#define DTB_DEFAULT_ENTRIES 64\n");
+		printf("#define L2_ASSOCIATIVE 8\n");
+		printf("#define DTB_DEFAULT_ENTRIES 48\n");
 		printf("#define DTB_SIZE 4096\n");
 		break;
 
+	    case CPU_NEOVERSEV1:
+                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 48\n");
+                printf("#define DTB_SIZE 4096\n");
+                break;
+
+	    case CPU_NEOVERSEN2:
+                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 48\n");
+                printf("#define DTB_SIZE 4096\n");
+                break;
+
 	    case CPU_FALKOR:
 	        printf("#define FALKOR\n");
 	        printf("#define L1_CODE_SIZE 65536\n");

driver/others/dynamic_arm64.c

@@ -147,6 +147,8 @@ static char *corename[] = {
   "tsv110",
   "emag8180",
   "neoversen1",
+  "neoversev1",
+  "neoversen2",
   "thunderx3t110",
   "cortexa55",
   "unknown"

getarch.c

@@ -1302,12 +1302,47 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
        "-DL2_SIZE=1048576 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=16 " \
        "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
        "-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DARMV8 " \
-       "-march=armv8.2-a -mtune=cortex-a72"
+       "-march=armv8.2-a -mtune=neoverse-n1"
 #define LIBNAME   "neoversen1"
 #define CORENAME  "NEOVERSEN1"
 #else
 #endif
 
+#ifdef FORCE_NEOVERSEV1
+#define FORCE
+#define ARCHITECTURE    "ARM64"
+#define SUBARCHITECTURE "NEOVERSEV1"
+#define SUBDIRNAME      "arm64"
+#define ARCHCONFIG   "-DNEOVERSEV1 " \
+       "-DL1_CODE_SIZE=65536 -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=4 " \
+       "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 -DL1_DATA_ASSOCIATIVE=4 " \
+       "-DL2_SIZE=1048576 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=16 " \
+       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
+       "-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DHAVE_SVE -DARMV8 " \
+       "-march=armv8.4-a -mtune=neoverse-v1"
+#define LIBNAME   "neoversev1"
+#define CORENAME  "NEOVERSEV1"
+#else
+#endif
+
+
+#ifdef FORCE_NEOVERSEN2
+#define FORCE
+#define ARCHITECTURE    "ARM64"
+#define SUBARCHITECTURE "NEOVERSEN2"
+#define SUBDIRNAME      "arm64"
+#define ARCHCONFIG   "-DNEOVERSEN2 " \
+       "-DL1_CODE_SIZE=65536 -DL1_CODE_LINESIZE=64 -DL1_CODE_ASSOCIATIVE=4 " \
+       "-DL1_DATA_SIZE=65536 -DL1_DATA_LINESIZE=64 -DL1_DATA_ASSOCIATIVE=4 " \
+       "-DL2_SIZE=1048576 -DL2_LINESIZE=64 -DL2_ASSOCIATIVE=16 " \
+       "-DDTB_DEFAULT_ENTRIES=64 -DDTB_SIZE=4096 " \
+       "-DHAVE_VFPV4 -DHAVE_VFPV3 -DHAVE_VFP -DHAVE_NEON -DHAVE_SVE -DARMV8 " \
+       "-march=armv8.5-a -mtune=neoverse-n2"
+#define LIBNAME   "neoversen2"
+#define CORENAME  "NEOVERSEN2"
+#else
+#endif
+
 #ifdef FORCE_CORTEXA55
 #define FORCE
 #define ARCHITECTURE    "ARM64"

kernel/arm64/KERNEL.NEOVERSEN2

@@ -0,0 +1,189 @@
+SAMINKERNEL  = ../arm/amin.c
+DAMINKERNEL  = ../arm/amin.c
+CAMINKERNEL  = ../arm/zamin.c
+ZAMINKERNEL  = ../arm/zamin.c
+
+SMAXKERNEL   = ../arm/max.c
+DMAXKERNEL   = ../arm/max.c
+
+SMINKERNEL   = ../arm/min.c
+DMINKERNEL   = ../arm/min.c
+
+ISAMINKERNEL = ../arm/iamin.c
+IDAMINKERNEL = ../arm/iamin.c
+ICAMINKERNEL = ../arm/izamin.c
+IZAMINKERNEL = ../arm/izamin.c
+
+ISMAXKERNEL  = ../arm/imax.c
+IDMAXKERNEL  = ../arm/imax.c
+
+ISMINKERNEL  = ../arm/imin.c
+IDMINKERNEL  = ../arm/imin.c
+
+STRSMKERNEL_LN	=  ../generic/trsm_kernel_LN.c
+STRSMKERNEL_LT	=  ../generic/trsm_kernel_LT.c
+STRSMKERNEL_RN	=  ../generic/trsm_kernel_RN.c
+STRSMKERNEL_RT	=  ../generic/trsm_kernel_RT.c
+
+DTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+DTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+DTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+CTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+CTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+CTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+CTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+ZTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+ZTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+ZTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+SAMAXKERNEL  = amax.S
+DAMAXKERNEL  = amax.S
+CAMAXKERNEL  = zamax.S
+ZAMAXKERNEL  = zamax.S
+
+SAXPYKERNEL  = axpy.S
+DAXPYKERNEL  = daxpy_thunderx2t99.S
+CAXPYKERNEL  = zaxpy.S
+ZAXPYKERNEL  = zaxpy.S
+
+SROTKERNEL   = rot.S
+DROTKERNEL   = rot.S
+CROTKERNEL   = zrot.S
+ZROTKERNEL   = zrot.S
+
+SSCALKERNEL  = scal.S
+DSCALKERNEL  = scal.S
+CSCALKERNEL  = zscal.S
+ZSCALKERNEL  = zscal.S
+
+SGEMVNKERNEL = gemv_n.S
+DGEMVNKERNEL = gemv_n.S
+CGEMVNKERNEL = zgemv_n.S
+ZGEMVNKERNEL = zgemv_n.S
+
+SGEMVTKERNEL = gemv_t.S
+DGEMVTKERNEL = gemv_t.S
+CGEMVTKERNEL = zgemv_t.S
+ZGEMVTKERNEL = zgemv_t.S
+
+
+SASUMKERNEL    = sasum_thunderx2t99.c
+DASUMKERNEL    = dasum_thunderx2t99.c
+CASUMKERNEL    = casum_thunderx2t99.c
+ZASUMKERNEL    = zasum_thunderx2t99.c
+
+SCOPYKERNEL    = copy_thunderx2t99.c
+DCOPYKERNEL    = copy_thunderx2t99.c
+CCOPYKERNEL    = copy_thunderx2t99.c
+ZCOPYKERNEL    = copy_thunderx2t99.c
+
+SSWAPKERNEL    = swap_thunderx2t99.S
+DSWAPKERNEL    = swap_thunderx2t99.S
+CSWAPKERNEL    = swap_thunderx2t99.S
+ZSWAPKERNEL    = swap_thunderx2t99.S
+
+ISAMAXKERNEL   = iamax_thunderx2t99.c
+IDAMAXKERNEL   = iamax_thunderx2t99.c
+ICAMAXKERNEL   = izamax_thunderx2t99.c
+IZAMAXKERNEL   = izamax_thunderx2t99.c
+
+SNRM2KERNEL    = scnrm2_thunderx2t99.c
+DNRM2KERNEL    = dznrm2_thunderx2t99.c
+CNRM2KERNEL    = scnrm2_thunderx2t99.c
+ZNRM2KERNEL    = dznrm2_thunderx2t99.c
+
+DDOTKERNEL     = dot_thunderx2t99.c
+SDOTKERNEL     = dot_thunderx2t99.c
+CDOTKERNEL     = zdot_thunderx2t99.c
+ZDOTKERNEL     = zdot_thunderx2t99.c
+DSDOTKERNEL    = dot.S
+
+DGEMM_BETA     = dgemm_beta.S
+SGEMM_BETA     = sgemm_beta.S
+
+SGEMMKERNEL    =  sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
+STRMMKERNEL    =  strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
+ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
+ifeq ($(SGEMM_UNROLL_M), 16)
+SGEMMITCOPY    =  sgemm_tcopy_$(SGEMM_UNROLL_M).S
+else
+SGEMMITCOPY    =  ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
+endif
+ifeq ($(SGEMM_UNROLL_M), 4)
+SGEMMINCOPY    =  sgemm_ncopy_$(SGEMM_UNROLL_M).S
+else
+SGEMMINCOPY    =  ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
+endif
+SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
+SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+ifeq ($(SGEMM_UNROLL_N), 16)
+SGEMMOTCOPY    =  sgemm_tcopy_$(SGEMM_UNROLL_N).S
+else
+SGEMMOTCOPY    =  ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c
+endif
+ifeq ($(SGEMM_UNROLL_N), 4)
+SGEMMONCOPY    =  sgemm_ncopy_$(SGEMM_UNROLL_N).S
+else
+SGEMMONCOPY    =  ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c
+endif
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+DGEMMKERNEL    =  dgemm_kernel_$(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$(TSUFFIX).$(SUFFIX)
+DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
+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_$(DGEMM_UNROLL_N).c
+endif
+
+DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+CGEMMKERNEL    =  cgemm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
+CTRMMKERNEL    =  ctrmm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
+ifneq ($(CGEMM_UNROLL_M), $(CGEMM_UNROLL_N))
+CGEMMINCOPY    =  ../generic/zgemm_ncopy_$(CGEMM_UNROLL_M).c
+CGEMMITCOPY    =  ../generic/zgemm_tcopy_$(CGEMM_UNROLL_M).c
+CGEMMINCOPYOBJ =  cgemm_incopy$(TSUFFIX).$(SUFFIX)
+CGEMMITCOPYOBJ =  cgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+CGEMMONCOPY    =  ../generic/zgemm_ncopy_$(CGEMM_UNROLL_N).c
+CGEMMOTCOPY    =  ../generic/zgemm_tcopy_$(CGEMM_UNROLL_N).c
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+ZGEMMKERNEL    =  zgemm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S
+ZTRMMKERNEL    =  ztrmm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S
+ifneq ($(ZGEMM_UNROLL_M), $(ZGEMM_UNROLL_N))
+ZGEMMINCOPY    =  ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_M).c
+ZGEMMITCOPY    =  ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_M).c
+ZGEMMINCOPYOBJ =  zgemm_incopy$(TSUFFIX).$(SUFFIX)
+ZGEMMITCOPYOBJ =  zgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+ZGEMMONCOPY    =  ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_N).c
+ZGEMMOTCOPY    =  ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_N).c
+ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
+ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)

kernel/arm64/KERNEL.NEOVERSEV1

@@ -0,0 +1,189 @@
+SAMINKERNEL  = ../arm/amin.c
+DAMINKERNEL  = ../arm/amin.c
+CAMINKERNEL  = ../arm/zamin.c
+ZAMINKERNEL  = ../arm/zamin.c
+
+SMAXKERNEL   = ../arm/max.c
+DMAXKERNEL   = ../arm/max.c
+
+SMINKERNEL   = ../arm/min.c
+DMINKERNEL   = ../arm/min.c
+
+ISAMINKERNEL = ../arm/iamin.c
+IDAMINKERNEL = ../arm/iamin.c
+ICAMINKERNEL = ../arm/izamin.c
+IZAMINKERNEL = ../arm/izamin.c
+
+ISMAXKERNEL  = ../arm/imax.c
+IDMAXKERNEL  = ../arm/imax.c
+
+ISMINKERNEL  = ../arm/imin.c
+IDMINKERNEL  = ../arm/imin.c
+
+STRSMKERNEL_LN	=  ../generic/trsm_kernel_LN.c
+STRSMKERNEL_LT	=  ../generic/trsm_kernel_LT.c
+STRSMKERNEL_RN	=  ../generic/trsm_kernel_RN.c
+STRSMKERNEL_RT	=  ../generic/trsm_kernel_RT.c
+
+DTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+DTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+DTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+DTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+CTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+CTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+CTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+CTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+ZTRSMKERNEL_LN	= ../generic/trsm_kernel_LN.c
+ZTRSMKERNEL_LT	= ../generic/trsm_kernel_LT.c
+ZTRSMKERNEL_RN	= ../generic/trsm_kernel_RN.c
+ZTRSMKERNEL_RT	= ../generic/trsm_kernel_RT.c
+
+SAMAXKERNEL  = amax.S
+DAMAXKERNEL  = amax.S
+CAMAXKERNEL  = zamax.S
+ZAMAXKERNEL  = zamax.S
+
+SAXPYKERNEL  = axpy.S
+DAXPYKERNEL  = daxpy_thunderx2t99.S
+CAXPYKERNEL  = zaxpy.S
+ZAXPYKERNEL  = zaxpy.S
+
+SROTKERNEL   = rot.S
+DROTKERNEL   = rot.S
+CROTKERNEL   = zrot.S
+ZROTKERNEL   = zrot.S
+
+SSCALKERNEL  = scal.S
+DSCALKERNEL  = scal.S
+CSCALKERNEL  = zscal.S
+ZSCALKERNEL  = zscal.S
+
+SGEMVNKERNEL = gemv_n.S
+DGEMVNKERNEL = gemv_n.S
+CGEMVNKERNEL = zgemv_n.S
+ZGEMVNKERNEL = zgemv_n.S
+
+SGEMVTKERNEL = gemv_t.S
+DGEMVTKERNEL = gemv_t.S
+CGEMVTKERNEL = zgemv_t.S
+ZGEMVTKERNEL = zgemv_t.S
+
+
+SASUMKERNEL    = sasum_thunderx2t99.c
+DASUMKERNEL    = dasum_thunderx2t99.c
+CASUMKERNEL    = casum_thunderx2t99.c
+ZASUMKERNEL    = zasum_thunderx2t99.c
+
+SCOPYKERNEL    = copy_thunderx2t99.c
+DCOPYKERNEL    = copy_thunderx2t99.c
+CCOPYKERNEL    = copy_thunderx2t99.c
+ZCOPYKERNEL    = copy_thunderx2t99.c
+
+SSWAPKERNEL    = swap_thunderx2t99.S
+DSWAPKERNEL    = swap_thunderx2t99.S
+CSWAPKERNEL    = swap_thunderx2t99.S
+ZSWAPKERNEL    = swap_thunderx2t99.S
+
+ISAMAXKERNEL   = iamax_thunderx2t99.c
+IDAMAXKERNEL   = iamax_thunderx2t99.c
+ICAMAXKERNEL   = izamax_thunderx2t99.c
+IZAMAXKERNEL   = izamax_thunderx2t99.c
+
+SNRM2KERNEL    = scnrm2_thunderx2t99.c
+DNRM2KERNEL    = dznrm2_thunderx2t99.c
+CNRM2KERNEL    = scnrm2_thunderx2t99.c
+ZNRM2KERNEL    = dznrm2_thunderx2t99.c
+
+DDOTKERNEL     = dot_thunderx2t99.c
+SDOTKERNEL     = dot_thunderx2t99.c
+CDOTKERNEL     = zdot_thunderx2t99.c
+ZDOTKERNEL     = zdot_thunderx2t99.c
+DSDOTKERNEL    = dot.S
+
+DGEMM_BETA     = dgemm_beta.S
+SGEMM_BETA     = sgemm_beta.S
+
+SGEMMKERNEL    =  sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
+STRMMKERNEL    =  strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
+ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
+ifeq ($(SGEMM_UNROLL_M), 16)
+SGEMMITCOPY    =  sgemm_tcopy_$(SGEMM_UNROLL_M).S
+else
+SGEMMITCOPY    =  ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
+endif
+ifeq ($(SGEMM_UNROLL_M), 4)
+SGEMMINCOPY    =  sgemm_ncopy_$(SGEMM_UNROLL_M).S
+else
+SGEMMINCOPY    =  ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
+endif
+SGEMMINCOPYOBJ =  sgemm_incopy$(TSUFFIX).$(SUFFIX)
+SGEMMITCOPYOBJ =  sgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+ifeq ($(SGEMM_UNROLL_N), 16)
+SGEMMOTCOPY    =  sgemm_tcopy_$(SGEMM_UNROLL_N).S
+else
+SGEMMOTCOPY    =  ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c
+endif
+ifeq ($(SGEMM_UNROLL_N), 4)
+SGEMMONCOPY    =  sgemm_ncopy_$(SGEMM_UNROLL_N).S
+else
+SGEMMONCOPY    =  ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c
+endif
+SGEMMONCOPYOBJ =  sgemm_oncopy$(TSUFFIX).$(SUFFIX)
+SGEMMOTCOPYOBJ =  sgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+DGEMMKERNEL    =  dgemm_kernel_$(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$(TSUFFIX).$(SUFFIX)
+DGEMMITCOPYOBJ =  dgemm_itcopy$(TSUFFIX).$(SUFFIX)
+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_$(DGEMM_UNROLL_N).c
+endif
+
+DGEMMONCOPYOBJ =  dgemm_oncopy$(TSUFFIX).$(SUFFIX)
+DGEMMOTCOPYOBJ =  dgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+CGEMMKERNEL    =  cgemm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
+CTRMMKERNEL    =  ctrmm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
+ifneq ($(CGEMM_UNROLL_M), $(CGEMM_UNROLL_N))
+CGEMMINCOPY    =  ../generic/zgemm_ncopy_$(CGEMM_UNROLL_M).c
+CGEMMITCOPY    =  ../generic/zgemm_tcopy_$(CGEMM_UNROLL_M).c
+CGEMMINCOPYOBJ =  cgemm_incopy$(TSUFFIX).$(SUFFIX)
+CGEMMITCOPYOBJ =  cgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+CGEMMONCOPY    =  ../generic/zgemm_ncopy_$(CGEMM_UNROLL_N).c
+CGEMMOTCOPY    =  ../generic/zgemm_tcopy_$(CGEMM_UNROLL_N).c
+CGEMMONCOPYOBJ =  cgemm_oncopy$(TSUFFIX).$(SUFFIX)
+CGEMMOTCOPYOBJ =  cgemm_otcopy$(TSUFFIX).$(SUFFIX)
+
+ZGEMMKERNEL    =  zgemm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S
+ZTRMMKERNEL    =  ztrmm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S
+ifneq ($(ZGEMM_UNROLL_M), $(ZGEMM_UNROLL_N))
+ZGEMMINCOPY    =  ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_M).c
+ZGEMMITCOPY    =  ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_M).c
+ZGEMMINCOPYOBJ =  zgemm_incopy$(TSUFFIX).$(SUFFIX)
+ZGEMMITCOPYOBJ =  zgemm_itcopy$(TSUFFIX).$(SUFFIX)
+endif
+ZGEMMONCOPY    =  ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_N).c
+ZGEMMOTCOPY    =  ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_N).c
+ZGEMMONCOPYOBJ =  zgemm_oncopy$(TSUFFIX).$(SUFFIX)
+ZGEMMOTCOPYOBJ =  zgemm_otcopy$(TSUFFIX).$(SUFFIX)

param.h

@@ -3307,6 +3307,64 @@ 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(NEOVERSEV1)
+
+#define SGEMM_DEFAULT_UNROLL_M  16
+#define SGEMM_DEFAULT_UNROLL_N  4
+
+#define DGEMM_DEFAULT_UNROLL_M  8
+#define DGEMM_DEFAULT_UNROLL_N  4
+
+#define CGEMM_DEFAULT_UNROLL_M  8
+#define CGEMM_DEFAULT_UNROLL_N  4
+
+#define ZGEMM_DEFAULT_UNROLL_M  4
+#define ZGEMM_DEFAULT_UNROLL_N  4
+
+#define SGEMM_DEFAULT_P 128
+#define DGEMM_DEFAULT_P 160
+#define CGEMM_DEFAULT_P 128
+#define ZGEMM_DEFAULT_P 128
+
+#define SGEMM_DEFAULT_Q 352
+#define DGEMM_DEFAULT_Q 128
+#define CGEMM_DEFAULT_Q 224
+#define ZGEMM_DEFAULT_Q 112
+
+#define SGEMM_DEFAULT_R 4096
+#define DGEMM_DEFAULT_R 4096
+#define CGEMM_DEFAULT_R 4096
+#define ZGEMM_DEFAULT_R 4096
+
+#elif defined(NEOVERSEN2)
+
+#define SGEMM_DEFAULT_UNROLL_M  16
+#define SGEMM_DEFAULT_UNROLL_N  4
+
+#define DGEMM_DEFAULT_UNROLL_M  8
+#define DGEMM_DEFAULT_UNROLL_N  4
+
+#define CGEMM_DEFAULT_UNROLL_M  8
+#define CGEMM_DEFAULT_UNROLL_N  4
+
+#define ZGEMM_DEFAULT_UNROLL_M  4
+#define ZGEMM_DEFAULT_UNROLL_N  4
+
+#define SGEMM_DEFAULT_P 128
+#define DGEMM_DEFAULT_P 160
+#define CGEMM_DEFAULT_P 128
+#define ZGEMM_DEFAULT_P 128
+
+#define SGEMM_DEFAULT_Q 352
+#define DGEMM_DEFAULT_Q 128
+#define CGEMM_DEFAULT_Q 224
+#define ZGEMM_DEFAULT_Q 112
+
+#define SGEMM_DEFAULT_R 4096
+#define DGEMM_DEFAULT_R 4096
+#define CGEMM_DEFAULT_R 4096
+#define ZGEMM_DEFAULT_R 4096
+
 #elif defined(ARMV8SVE) || defined(A64FX)
 
 /* When all BLAS3 routines are implemeted with SVE, SGEMM_DEFAULT_UNROLL_M should be "sve_vl".