Support for SME1 based sgemm_direct kernel for cblas_sgemm level 3 API

* Added ARMV9SME target
* Added SGEMM_DIRECT kernel based on SME1

Makefile.arm64

@@ -30,6 +30,11 @@ FCOMMON_OPT += -march=armv8-a+sve
 endif
 endif
 
+ifeq ($(CORE), ARMV9SME)
+CCOMMON_OPT += -march=armv9-a+sve2+sme
+FCOMMON_OPT += -march=armv9-a+sve2
+endif
+
 ifeq ($(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 +710,9 @@ DYNAMIC_CORE += NEOVERSEN2
 DYNAMIC_CORE += ARMV8SVE
 DYNAMIC_CORE += A64FX
 endif
+ifneq ($(NO_SME), 1)
+DYNAMIC_CORE += ARMV9SME
+endif
 DYNAMIC_CORE += THUNDERX
 DYNAMIC_CORE += THUNDERX2T99
 DYNAMIC_CORE += TSV110
@@ -1474,6 +1478,10 @@ ifeq ($(NO_SVE), 1)
 CCOMMON_OPT     += -DNO_SVE
 endif
 
+ifeq ($(NO_SME), 1)
+CCOMMON_OPT     += -DNO_SME
+endif
+
 ifdef SMP
 CCOMMON_OPT	+= -DSMP_SERVER
 

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+sme")
+  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+sme -O3")
+  endif()
   if (${TARGET} STREQUAL A64FX)
     if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)
       set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -Msve-intrinsics -march=armv8.2-a+sve -mtune=a64fx")
@@ -382,6 +385,8 @@ if (NEED_PIC)
   if (NOT NOFORTRAN)
     if (${F_COMPILER} STREQUAL "SUN")
       set(FCOMMON_OPT "${FCOMMON_OPT} -pic")
+    elseif (${F_COMPILER} STREQUAL "NAGFOR")
+      set(FCOMMON_OPT "${FCOMMON_OPT} -PIC")
     else ()
       set(FCOMMON_OPT "${FCOMMON_OPT} -fPIC")
     endif ()
@@ -640,17 +645,17 @@ if (${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
 endif ()
 
 if (CMAKE_Fortran_COMPILER)
-if ("${F_COMPILER}" STREQUAL "NAG" OR "${F_COMPILER}" STREQUAL "CRAY" OR CMAKE_Fortran_COMPILER_ID MATCHES "LLVMFlang.*")
-  set(FILTER_FLAGS "-msse3;-mssse3;-msse4.1;-mavx;-mavx2,-mskylake-avx512")
-  if (CMAKE_Fortran_COMPILER_ID MATCHES "LLVMFlang.*")
-message(STATUS "removing fortran flags")
-    set(FILTER_FLAGS "${FILTER_FLAGS};-m32;-m64")
+  if ("${F_COMPILER}" STREQUAL "NAGFOR" OR "${F_COMPILER}" STREQUAL "CRAY" OR CMAKE_Fortran_COMPILER_ID MATCHES "LLVMFlang.*")
+    set(FILTER_FLAGS "-msse3;-mssse3;-msse4.1;-mavx;-mavx2,-mskylake-avx512")
+    if (CMAKE_Fortran_COMPILER_ID MATCHES "LLVMFlang.*")
+      message(STATUS "removing fortran flags")
+      set(FILTER_FLAGS "${FILTER_FLAGS};-m32;-m64")
+    endif ()
+    foreach (FILTER_FLAG ${FILTER_FLAGS})
+      string(REPLACE ${FILTER_FLAG} "" LAPACK_FFLAGS ${LAPACK_FFLAGS})
+      string(REPLACE ${FILTER_FLAG} "" LAPACK_FPFLAGS ${LAPACK_FPFLAGS})
+    endforeach ()
   endif ()
-  foreach (FILTER_FLAG ${FILTER_FLAGS})
-    string(REPLACE ${FILTER_FLAG} "" LAPACK_FFLAGS ${LAPACK_FFLAGS})
-    string(REPLACE ${FILTER_FLAG} "" LAPACK_FPFLAGS ${LAPACK_FPFLAGS})
-  endforeach ()
-endif ()
 endif ()
 
 if ("${F_COMPILER}" STREQUAL "GFORTRAN")

common.h

@@ -696,6 +696,7 @@ void gotoblas_profile_init(void);
 void gotoblas_profile_quit(void);
 
 int support_avx512(void);	
+int support_sme1(void);	
 
 #ifdef USE_OPENMP
 

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)

common_param.h

@@ -221,6 +221,11 @@ BLASLONG (*ismin_k) (BLASLONG, float *, BLASLONG);
   void (*sgemm_direct) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG , float *, BLASLONG , float * , BLASLONG);
   int  (*sgemm_direct_performant) (BLASLONG M, BLASLONG N, BLASLONG K);
 #endif
+#ifdef ARCH_ARM64
+  void (*sgemm_direct) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG , float *, BLASLONG , float * , BLASLONG);
+  int  (*sgemm_direct_performant) (BLASLONG M, BLASLONG N, BLASLONG K);
+#endif
+
 
   int    (*sgemm_kernel   )(BLASLONG, BLASLONG, BLASLONG, float, float *, float *, float *, BLASLONG);
   int    (*sgemm_beta     )(BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float  *, BLASLONG);

common_s.h

@@ -213,9 +213,9 @@
 #ifdef ARCH_X86_64
 #define SGEMM_DIRECT_PERFORMANT gotoblas -> sgemm_direct_performant
 #define  SGEMM_DIRECT		gotoblas -> sgemm_direct
-#else
+#elif ARCH_ARM64
 #define SGEMM_DIRECT_PERFORMANT    sgemm_direct_performant
-#define  SGEMM_DIRECT		sgemm_direct
+#define  SGEMM_DIRECT		gotoblas -> sgemm_direct
 #endif
 
 #define	SGEMM_ONCOPY		gotoblas -> sgemm_oncopy

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_SME)
+  if ((getauxval(AT_HWCAP2) & HWCAP2_SME)) {
+    return &gotoblas_ARMV9SME;
+  }
+#endif
+
 #ifndef NO_SVE
   if ((getauxval(AT_HWCAP) & HWCAP_SVE)) {
     return &gotoblas_ARMV8SVE;
@@ -443,3 +462,15 @@ void gotoblas_dynamic_init(void) {
 void gotoblas_dynamic_quit(void) {
   gotoblas = NULL;
 }
+
+int support_sme1(void) {
+        int ret = 0;
+
+#if (defined OS_LINUX || defined OS_ANDROID)
+        ret = getauxval(AT_HWCAP2) & HWCAP2_SME;
+        if(getauxval(AT_HWCAP2) & HWCAP2_SME){
+                ret = 1;
+        }
+#endif
+       return ret;
+}

getarch.c

@@ -1289,6 +1289,19 @@ 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

interface/gemm.c

@@ -45,6 +45,7 @@
 #include "functable.h"
 #endif
 
+
 #ifndef COMPLEX
 #define SMP_THRESHOLD_MIN 65536.0
 #ifdef XDOUBLE
@@ -85,6 +86,7 @@
 #define GEMM_MULTITHREAD_THRESHOLD 4
 #endif
 
+
 static int (*gemm[])(blas_arg_t *, BLASLONG *, BLASLONG *, IFLOAT *, IFLOAT *, BLASLONG) = {
 #ifndef GEMM3M
   GEMM_NN, GEMM_TN, GEMM_RN, GEMM_CN,
@@ -347,17 +349,26 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
   int nodes;
 #endif
 
+
   PRINT_DEBUG_CNAME;
 
 #if !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16) && defined(USE_SGEMM_KERNEL_DIRECT)
-#ifdef DYNAMIC_ARCH
+#if defined(DYNAMIC_ARCH) && defined(ARCH_x86)
  if (support_avx512() )
-#endif  
+
   if (beta == 0 && alpha == 1.0 && order == CblasRowMajor && TransA == CblasNoTrans && TransB == CblasNoTrans && SGEMM_DIRECT_PERFORMANT(m,n,k)) {
 	SGEMM_DIRECT(m, n, k, a, lda, b, ldb, c, ldc);
 	return;
   }
-
+#endif
+#if defined(DYNAMIC_ARCH) && defined(ARCH_ARM64)
+ if (support_sme1()){
+  if (beta == 0 && alpha == 1.0 && order == CblasRowMajor && TransA == CblasNoTrans && TransB == CblasNoTrans) {
+	SGEMM_DIRECT(m, n, k, a, lda, b, ldb, c, ldc);
+	return;
+  }
+ }
+#endif
 #endif
 
 #ifndef COMPLEX

kernel/Makefile

@@ -24,7 +24,11 @@ ifdef NO_AVX2
  AVX2OPT=
 endif
 
+
 ifdef TARGET_CORE
+ifeq ($(TARGET_CORE), ARMV9SME)
+ override CFLAGS += -march=armv9-a+sve2+sme
+endif
 ifeq ($(TARGET_CORE), SAPPHIRERAPIDS)
  override CFLAGS += -DBUILD_KERNEL -DTABLE_NAME=gotoblas_$(TARGET_CORE)
  ifeq (1, $(filter 1,$(GCCVERSIONGTEQ11) $(CLANGVERSIONGTEQ12)))

kernel/Makefile.L3

@@ -24,6 +24,7 @@ endif
 
 ifeq ($(ARCH), arm64)
 USE_TRMM = 1
+USE_DIRECT_SGEMM = 1
 endif
 
 ifeq ($(ARCH), riscv64)
@@ -95,9 +96,16 @@ endif
 
 ifdef USE_DIRECT_SGEMM
 ifndef SGEMMDIRECTKERNEL
+ifeq ($(ARCH), x86_64)
 SGEMMDIRECTKERNEL = sgemm_direct_skylakex.c
 SGEMMDIRECTPERFORMANT = sgemm_direct_performant.c
 endif
+ifeq ($(ARCH), arm64)
+ifdef HAVE_SME
+SGEMMDIRECTKERNEL = sgemm_direct_arm64_sme1.c
+endif
+endif
+endif
 endif
 
 ifeq ($(BUILD_BFLOAT16), 1)
@@ -128,9 +136,19 @@ SKERNELOBJS	+= \
 	$(SGEMMONCOPYOBJ) $(SGEMMOTCOPYOBJ)
 
 ifdef USE_DIRECT_SGEMM
+ifeq ($(ARCH), x86_64)
 SKERNELOBJS += \
 	sgemm_direct$(TSUFFIX).$(SUFFIX) \
-	sgemm_direct_performant$(TSUFFIX).$(SUFFIX) 
+	sgemm_direct_performant$(TSUFFIX).$(SUFFIX)
+endif
+ifeq ($(ARCH), arm64)
+ifdef HAVE_SME
+SKERNELOBJS += \
+	sgemm_direct.$(SUFFIX) \
+	sgemm_direct_sme1.$(SUFFIX) \
+	sgemm_direct_sme1_preprocess.$(SUFFIX)
+endif
+endif
 endif
 endif
 
@@ -809,11 +827,23 @@ else
 endif
 
 ifdef USE_DIRECT_SGEMM
+ifeq ($(ARCH), x86_64)
 $(KDIR)sgemm_direct_performant$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMMDIRECTPERFORMANT)
 	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 $(KDIR)sgemm_direct$(TSUFFIX).$(SUFFIX) : $(KERNELDIR)/$(SGEMMDIRECTKERNEL)
 	$(CC) $(CFLAGS) -c -UDOUBLE -UCOMPLEX $< -o $@
 endif
+ifeq ($(ARCH), arm64)
+ifdef HAVE_SME
+$(KDIR)sgemm_direct_sme1.$(SUFFIX) : 
+	$(CC) $(CFLAGS) -c $(KERNELDIR)/sgemm_direct_sme1.S -UDOUBLE -UCOMPLEX  -o $@
+$(KDIR)sgemm_direct_sme1_preprocess.$(SUFFIX) : 
+	$(CC) $(CFLAGS) -c $(KERNELDIR)/sgemm_direct_sme1_preprocess.S -UDOUBLE -UCOMPLEX  -o $@
+$(KDIR)sgemm_direct.$(SUFFIX) : 
+	$(CC) $(CFLAGS) -c $(KERNELDIR)/sgemm_direct_arm64_sme1.c -UDOUBLE -UCOMPLEX  -o $@
+endif
+endif
+endif
 
 ifeq ($(BUILD_BFLOAT16), 1)
 

kernel/arm64/KERNEL.ARMV9SME

@@ -0,0 +1,3 @@
+include $(KERNELDIR)/KERNEL.ARMV8SVE
+
+