Add Arm®v9-A architecture SME target

Add a new target, ARMV9SME, for Arm®v9-A architecture systems that
support the Scalable Matrix Extension (SME) [1].

Initially inherits ARMV8SVE settings with updated compiler flags. This
target can only be built with an SME-capable toolchain such as GCC 14 or
LLVM 19.

Includes some initial FEAT_SME2 feature detection on Linux targets via
hwcaps. Target is disabled in DYNAMIC_ARCH builds by default.

This is intended as a base target for SME2 kernels.

[1] https://developer.arm.com/documentation/109246/0100/SME-Overview/SME-and-SME2

Makefile.arm64

@@ -30,6 +30,11 @@ FCOMMON_OPT += -march=armv8-a+sve
 endif
 endif
 
+ifeq ($(CORE), ARMV9SME)
+CCOMMON_OPT += -march=armv9-a+sme2 -O3
+FCOMMON_OPT += -march=armv9-a+sve2 -O3
+endif
+
 ifeq ($(CORE), CORTEXA53)
 CCOMMON_OPT += -march=armv8-a -mtune=cortex-a53
 ifneq ($(F_COMPILER), NAG)

Makefile.system

@@ -420,6 +420,7 @@ ifeq ($(ARCH), arm64)
 export MACOSX_DEPLOYMENT_TARGET=11.0
 ifeq ($(C_COMPILER), GCC)
 export NO_SVE = 1
+export NO_SME = 1
 endif
 else
 export MACOSX_DEPLOYMENT_TARGET=10.8
@@ -709,6 +710,11 @@ DYNAMIC_CORE += NEOVERSEN2
 DYNAMIC_CORE += ARMV8SVE
 DYNAMIC_CORE += A64FX
 endif
+# Disabled by default while ARMV9SME is WIP
+NO_SME ?= 1
+ifneq ($(NO_SME), 1)
+DYNAMIC_CORE += ARMV9SME
+endif
 DYNAMIC_CORE += THUNDERX
 DYNAMIC_CORE += THUNDERX2T99
 DYNAMIC_CORE += TSV110
@@ -1474,6 +1480,10 @@ ifeq ($(NO_SVE), 1)
 CCOMMON_OPT     += -DNO_SVE
 endif
 
+ifeq ($(NO_SME), 1)
+CCOMMON_OPT     += -DNO_SME
+endif
+
 ifdef SMP
 CCOMMON_OPT	+= -DSMP_SERVER
 

README.md

@@ -188,6 +188,7 @@ Please read `GotoBLAS_01Readme.txt` for older CPU models already supported by th
 - **Apple Vortex**: preliminary support based on ThunderX2/3
 - **A64FX**:  preliminary support, optimized Level-3 BLAS
 - **ARMV8SVE**: any ARMV8 cpu with SVE extensions 
+- **ARMV9SME**: WIP target, any Arm®v9-A core with SME2 support. Only functional for GEMM.
 
 #### PPC/PPC64
 

TargetList.txt

@@ -111,6 +111,7 @@ THUNDERX3T110
 VORTEX
 A64FX
 ARMV8SVE
+ARMV9SME
 FT2000
 
 9.System Z:

cmake/arch.cmake

@@ -44,9 +44,21 @@ endif ()
 
 if (DYNAMIC_ARCH)
   if (ARM64)
-	  set(DYNAMIC_CORE ARMV8 CORTEXA53 CORTEXA57 THUNDERX THUNDERX2T99 TSV110 EMAG8180 NEOVERSEN1 THUNDERX3T110)
-    if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER 9.99)
-	    set(DYNAMIC_CORE ${DYNAMIC_CORE} NEOVERSEV1 NEOVERSEN2 ARMV8SVE A64FX)
+    set(DYNAMIC_CORE ARMV8 CORTEXA53 CORTEXA57 THUNDERX THUNDERX2T99 TSV110 EMAG8180 NEOVERSEN1 THUNDERX3T110)
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "GNU")
+      if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER_EQUAL 10) # SVE ACLE supported in GCC >= 10
+        set(DYNAMIC_CORE ${DYNAMIC_CORE} NEOVERSEV1 NEOVERSEN2 ARMV8SVE A64FX)
+      endif ()
+      if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER_EQUAL 14) # SME ACLE supported in GCC >= 14
+        set(DYNAMIC_CORE ${DYNAMIC_CORE} ARMV9SME)
+      endif()
+    elseif (${CMAKE_C_COMPILER_ID} MATCHES "Clang")
+      if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER_EQUAL 11) # SVE ACLE supported in LLVM >= 11
+        set(DYNAMIC_CORE ${DYNAMIC_CORE} NEOVERSEV1 NEOVERSEN2 ARMV8SVE A64FX)
+      endif ()
+      if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER_EQUAL 19) # SME ACLE supported in LLVM >= 19
+        set(DYNAMIC_CORE ${DYNAMIC_CORE} ARMV9SME)
+      endif()
     endif ()
     if (DYNAMIC_LIST)
 	  set(DYNAMIC_CORE ARMV8 ${DYNAMIC_LIST})

cmake/cc.cmake

@@ -238,6 +238,12 @@ if (${CORE} STREQUAL ARMV8SVE)
   endif ()
 endif ()
 
+if (${CORE} STREQUAL ARMV9SME)
+  if (NOT DYNAMIC_ARCH)
+    set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv9-a+sme2")
+  endif ()
+endif ()
+
 if (${CORE} STREQUAL CORTEXA510)
   if (NOT DYNAMIC_ARCH)
     set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8-a+sve")

cmake/prebuild.cmake

@@ -1014,7 +1014,7 @@ endif ()
     set(ZGEMM_UNROLL_M 4)
     set(ZGEMM_UNROLL_N 4)
     set(SYMV_P 16)
-  elseif ("${TCORE}" STREQUAL "NEOVERSEN2")
+  elseif ("${TCORE}" STREQUAL "NEOVERSEN2" or "${TCORE}" STREQUAL "ARMV9SME")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L1_CODE_SIZE\t65536\n"
       "#define L1_CODE_LINESIZE\t64\n"

cmake/system.cmake

@@ -310,6 +310,9 @@ if (${TARGET} STREQUAL NEOVERSEV1)
       set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv8.2-a+sve")
     endif()
   endif()
+  if (${TARGET} STREQUAL ARMV9SME)
+      set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv9-a+sme2 -O3")
+  endif()
   if (${TARGET} STREQUAL A64FX)
     if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
       set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -Msve-intrinsics -march=armv8.2-a+sve -mtune=a64fx")

common_arm64.h

@@ -175,7 +175,7 @@ static inline int blas_quickdivide(blasint x, blasint y){
 #define HUGE_PAGESIZE   ( 4 << 20)
 
 #ifndef BUFFERSIZE
-#if defined(NEOVERSEN1) || defined(NEOVERSEN2) || defined(NEOVERSEV1) || defined(A64FX) || defined(ARMV8SVE)
+#if defined(NEOVERSEN1) || defined(NEOVERSEN2) || defined(NEOVERSEV1) || defined(A64FX) || defined(ARMV8SVE) || defined(ARMV9SME)
 #define BUFFER_SIZE     (32 << 22)
 #else
 #define BUFFER_SIZE     (32 << 20)

driver/others/dynamic_arm64.c

@@ -115,6 +115,11 @@ extern gotoblas_t gotoblas_ARMV8SVE;
 #else
 #define gotoblas_ARMV8SVE gotoblas_ARMV8
 #endif
+#ifdef DYN_ARMV9SME
+extern gotoblas_t gotoblas_ARMV9SME;
+#else
+#define gotoblas_ARMV9SME gotoblas_ARMV8
+#endif
 #ifdef DYN_CORTEX_A55
 extern gotoblas_t  gotoblas_CORTEXA55;
 #else
@@ -148,6 +153,13 @@ extern gotoblas_t  gotoblas_A64FX;
 #define gotoblas_ARMV8SVE   gotoblas_ARMV8
 #define gotoblas_A64FX      gotoblas_ARMV8
 #endif
+
+#ifndef NO_SME
+extern gotoblas_t  gotoblas_ARMV9SME;
+#else
+#define gotoblas_ARMV9SME gotoblas_ARMV8SVE
+#endif
+
 extern gotoblas_t  gotoblas_THUNDERX3T110;
 #endif
 #define gotoblas_NEOVERSEV2 gotoblas_NEOVERSEV1
@@ -393,6 +405,13 @@ static gotoblas_t *get_coretype(void) {
       snprintf(coremsg, 128, "Unknown CPU model - implementer %x part %x\n",implementer,part);
       openblas_warning(1, coremsg);
   }
+
+#if !defined(NO_SME) && defined(HWCAP2_SME2)
+  if ((getauxval(AT_HWCAP2) & HWCAP2_SME2)) {
+    return &gotoblas_ARMV9SME;
+  }
+#endif
+
 #ifndef NO_SVE
   if ((getauxval(AT_HWCAP) & HWCAP_SVE)) {
     return &gotoblas_ARMV8SVE;

getarch.c

@@ -1289,6 +1289,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CORENAME  "ARMV8SVE"
 #endif
 
+#ifdef FORCE_ARMV9SME
+#define FORCE
+#define ARCHITECTURE    "ARM64"
+#define SUBARCHITECTURE "ARMV9SME"
+#define SUBDIRNAME      "arm64"
+#define ARCHCONFIG   "-DARMV9SME " \
+       "-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 -DHAVE_SME -DARMV8 -DARMV9"
+#define LIBNAME   "armv9sme"
+#define CORENAME  "ARMV9SME"
+#endif
+
 
 #ifdef FORCE_ARMV8
 #define FORCE

kernel/arm64/KERNEL.ARMV9SME

@@ -0,0 +1 @@
+include $(KERNELDIR)/KERNEL.ARMV8SVE

param.h

@@ -3667,7 +3667,7 @@ Until then, just keep it different than DGEMM_DEFAULT_UNROLL_N to keep copy rout
 #define CGEMM_DEFAULT_R 4096
 #define ZGEMM_DEFAULT_R 4096
 
-#elif defined(ARMV8SVE) || defined(ARMV9) || defined(CORTEXA510)|| defined(CORTEXA710) || defined(CORTEXX2) // 128-bit SVE
+#elif defined(ARMV8SVE) || defined(ARMV9SME) || defined(ARMV9) || defined(CORTEXA510)|| defined(CORTEXA710) || defined(CORTEXX2) // 128-bit SVE
 
 #if defined(XDOUBLE) || defined(DOUBLE)
 #define SWITCH_RATIO            8