Support for SME1 based sgemm_direct kernel for cblas_sgemm level 3 API

* Added ARMV9SME target
* Added SGEMM_DIRECT kernel based on SME1

Author: vaiskv

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

@@ -21,7 +21,15 @@ endif()
 # Other files expect CORE, which is actually TARGET and will become TARGET_CORE for kernel build. Confused yet?
 # It seems we are meant to use TARGET as input and CORE internally as kernel.
 if(NOT DEFINED CORE AND DEFINED TARGET)
-  set(CORE ${TARGET})
+  if (${TARGET} STREQUAL "LOONGSON3R5")
+    set(CORE "LA464")
+  elseif (${TARGET} STREQUAL "LOONGSON2K1000")
+    set(CORE "LA264")
+  elseif (${TARGET} STREQUAL "LOONGSONGENERIC")
+    set(CORE "LA64_GENERIC)")
+  else ()
+    set(CORE ${TARGET})
+  endif()
 endif()
 
 # TARGET_CORE will override TARGET which is used in DYNAMIC_ARCH=1.
@@ -310,6 +318,9 @@ 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  "${KERNEL_DEFINITIONS} -Msve-intrinsics -march=armv8.2-a+sve -mtune=a64fx")
@@ -382,6 +393,8 @@ if (NEED_PIC)
   if (NOT NOFORTRAN)
     if (${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 +653,17 @@ 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 ("${F_COMPILER}" STREQUAL "GFORTRAN")
@@ -670,6 +683,9 @@ endif ()
 if (${CMAKE_C_COMPILER} STREQUAL "LSB" OR ${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
   set(LAPACK_CFLAGS "${LAPACK_CFLAGS} -DLAPACK_COMPLEX_STRUCTURE")
 endif ()
+if (${CMAKE_C_COMPILER_ID} MATCHES "IntelLLVM" AND ${CMAKE_SYSTEM_NAME} STREQUAL "Windows")
+	set(LAPACK_CFLAGS "${LAPACK_CFLAGS} -DNOCHANGE")
+endif ()
 
 if ("${CMAKE_BUILD_TYPE}" STREQUAL "Release")
 if ("${F_COMPILER}" STREQUAL "FLANG")

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 +221,12 @@ 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
+#ifdef HAVE_SME
+  void (*sgemm_direct) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG , float *, BLASLONG , float * , BLASLONG);
+#endif
+#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);