Merge pull request #5423 from martin-frbg/issue5414

Split VORTEXM4 from VORTEX target and fix SGEMM_DIRECT support for SME-capable targets

Author: martin-frbg

Makefile.arm64

@@ -61,6 +61,11 @@ endif
 ifeq ($(CORE), ARMV9SME)
 CCOMMON_OPT += -march=armv9-a+sve2+sme
 FCOMMON_OPT += -march=armv9-a+sve2
+ifdef OS_WINDOWS
+ifeq ($(C_COMPILER), CLANG)
+CCOMMON_OPT += --aarch64-stack-hazard-size=0
+endif
+endif
 endif
 
 ifeq ($(CORE), CORTEXA53)
@@ -303,6 +308,20 @@ FCOMMON_OPT += -march=armv8.3-a
 endif
 endif
 
+ifeq ($(CORE), VORTEXM4)
+ifneq ($(C_COMPILER), GCC)
+ifeq ($(APPLECLANG),1)
+CCOMMON_OPT += -march=armv8.4-a+sme
+else
+CCOMMON_OPT += -march=armv8.4-a+sme
+override LDFLAGS += -lclang_rt_builtins-aarch64
+endif
+else
+CCOMMON_OPT += -march=armv8.4-a
+endif
+FCOMMON_OPT += -march=armv8.4-a
+endif
+
 ifeq (1, $(filter 1,$(GCCVERSIONGTEQ9) $(ISCLANG)))
 ifeq ($(CORE), TSV110)
 CCOMMON_OPT += -march=armv8.2-a -mtune=tsv110

Makefile.system

@@ -331,6 +331,7 @@ HAVE_SSE5=
 HAVE_AVX=
 HAVE_AVX2=
 HAVE_FMA3=
+HAVE_SME=
 include $(TOPDIR)/Makefile_kernel.conf
 endif
 
@@ -427,7 +428,7 @@ ifndef MACOSX_DEPLOYMENT_TARGET
 ifeq ($(ARCH), arm64)
 export MACOSX_DEPLOYMENT_TARGET=11.0
 export NO_SVE = 1
-export NO_SME = 1
+# export NO_SME = 1
 else
 export MACOSX_DEPLOYMENT_TARGET=10.8
 endif
@@ -721,6 +722,11 @@ DYNAMIC_CORE += A64FX
 endif
 ifneq ($(NO_SME), 1)
 DYNAMIC_CORE += ARMV9SME
+ifeq ($(OSNAME), Darwin)
+ifneq ($(C_COMPILER), GCC)
+DYNAMIC_CORE += VORTEXM4
+endif
+endif
 endif
 DYNAMIC_CORE += THUNDERX
 DYNAMIC_CORE += THUNDERX2T99
@@ -1896,6 +1902,7 @@ ifndef NO_MSA
   export HAVE_MSA
   export MSA_FLAGS
 endif
+export HAVE_SME
 export KERNELDIR
 export FUNCTION_PROFILE
 export TARGET_CORE

TargetList.txt

@@ -111,6 +111,7 @@ THUNDERX2T99
 TSV110
 THUNDERX3T110
 VORTEX
+VORTEXM4
 A64FX
 ARMV8SVE
 ARMV9SME

c_check

@@ -335,7 +335,14 @@ if [ "$architecture" = "arm64" ]; then
 fi
 
 no_sme=0
+is_appleclang=0
 if [ "$architecture" = "arm64" ]; then
+    if [ "$compiler" = "CLANG" ]; then
+        data=`$compiler_name --version`
+	case "$data" in Apple*)
+	is_appleclang=1
+	esac
+    fi
     tmpd=$(mktemp -d 2>/dev/null || mktemp -d -t 'OBC')
     tmpf="$tmpd/a.S"
     printf ".text \n.global sme_test\n\nsme_test:\nsmstart\nsmstop\nret\n">> "$tmpf"
@@ -469,6 +476,7 @@ done
     [ "$no_avx512bf" -eq 1 ] && printf "NO_AVX512BF16=1\n"
     [ "$no_avx2" -eq 1 ] && printf "NO_AVX2=1\n"
     [ "$oldgcc" -eq 1 ] && printf "OLDGCC=1\n"
+    [ "$is_appleclang" -eq 1 ] && printf "APPLECLANG=1\n"
     exit 0
 }
 
@@ -499,6 +507,7 @@ done
     [ "$no_avx512bf" -eq 1 ] && printf "NO_AVX512BF16=1\n"
     [ "$no_avx2" -eq 1 ] && printf "NO_AVX2=1\n"
     [ "$oldgcc" -eq 1 ] && printf "OLDGCC=1\n"
+    [ "$is_appleclang" -eq 1 ] && printf "APPLECLANG=1\n"
     [ "$no_lsx" -eq 1 ] && printf "NO_LSX=1\n"
     [ "$no_lasx" -eq 1 ] && printf "NO_LASX=1\n"
 } >> "$makefile"

cmake/arch.cmake

@@ -40,13 +40,16 @@ if (DYNAMIC_ARCH)
       endif ()
       if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER_EQUAL 14) # SME ACLE supported in GCC >= 14
         set(DYNAMIC_CORE ${DYNAMIC_CORE} ARMV9SME)
+      endif()
+	  if (${CMAKE_C_COMPILER_ID} MATCHES "Clang" AND ${CMAKE_SYSTEM_NAME} STREQUAL "Darwin")
+	    set(DYNAMIC_CORE ${DYNAMIC_CORE} VORTEXM4)
       endif()
     elseif (${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 ${DYNAMIC_CORE} ARMV9SME)
+      if (${CMAKE_C_COMPILER_VERSION} VERSION_GREATER_EQUAL 19 OR (${CMAKE_C_COMPILER_ID} MATCHES AppleClang AND ${CMAKE_C_COMPILER_VERSION} VERSION_GREATER_EQUAL 17) ) # SME ACLE supported in LLVM >= 19 and AppleClang >= 17
+        set(DYNAMIC_CORE ${DYNAMIC_CORE} ARMV9SME VORTEXM4)
       endif()
     endif ()
     if (DYNAMIC_LIST)

cmake/cc.cmake

@@ -315,7 +315,24 @@ if (${CORE} STREQUAL ARMV9SME)
 	    set (CCOMMON_OPT  "${CCOMMON_OPT} -tp=host")
     else ()
     set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv9-a+sme")
+	if (${OSNAME} STREQUAL Windows AND ${CMAKE_C_COMPILER_ID} MATCHES "Clang" )
+	set (CCOMMON_OPT "${CCOMMON_OPT} --aarch64-stack-hazard-size=0")
     endif ()
+	endif ()
+  endif ()
+endif ()
+
+if (${CORE} STREQUAL VORTEXM4)
+  if (NOT DYNAMIC_ARCH)
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "NVC" AND NOT NO_SVE)
+	    set (CCOMMON_OPT  "${CCOMMON_OPT} -tp=host")
+    else ()
+	  if (${CMAKE_C_COMPILER_ID} STREQUAL "AppleClang")
+        set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.4-a+sme -mcpu=apple-m4")
+	  else ()
+	    set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.4-a -mcpu=apple-m4")
+      endif ()
+	endif ()
   endif ()
 endif ()
 

cmake/prebuild.cmake

@@ -1255,7 +1255,7 @@ endif ()
     set(ZGEMM_UNROLL_M 4)
     set(ZGEMM_UNROLL_N 4)
     set(SYMV_P 16)
-  elseif ("${TCORE}" STREQUAL "VORTEX")
+  elseif ("${TCORE}" STREQUAL "VORTEX" OR "${TCORE}" STREQUAL "VORTEXM4")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define ARMV8\n"
       "#define L1_CODE_SIZE\t32768\n"
@@ -1639,6 +1639,8 @@ else(NOT CMAKE_CROSSCOMPILING)
   unset (HAVE_VFP)
   unset (HAVE_VFPV3)
   unset (HAVE_VFPV4)
+  unset (HAVE_SVE)
+  unset (HAVE_SME)
   message(STATUS "Running getarch")
 
   # use the cmake binary w/ the -E param to run a shell command in a cross-platform way

cmake/system.cmake

@@ -367,6 +367,15 @@ if (${TARGET} STREQUAL NEOVERSEV1)
   endif()
   if (${TARGET} STREQUAL ARMV9SME)
       set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv9-a+sme -O3")
+	  if (${CMAKE_SYSTEM_NAME} STREQUAL Windows AND ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
+	  set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} --aarch64-stack-hazard-size=0")
+	  endif()
+  endif()
+  if (${TARGET} STREQUAL VORTEXM4)
+      set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv8.4-a+sme -O3")
+	  if (${CMAKE_SYSTEM_NAME} STREQUAL Windows AND ${CMAKE_C_COMPILER_ID} MATCHES "Clang")
+	  set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} --aarch64-stack-hazard-size=0")
+	  endif()
   endif()
   if (${TARGET} STREQUAL A64FX)
     if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)

common_param.h

@@ -257,6 +257,7 @@ int (*shgemv_t) (BLASLONG, BLASLONG, float, hfloat16 *, BLASLONG, hfloat16 *, BL
 #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);
   void (*sgemm_direct_alpha_beta) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float, float * , BLASLONG);
   void (*ssymm_direct_alpha_betaLU) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float, float * , BLASLONG);
   void (*ssymm_direct_alpha_betaLL) (BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float, float * , BLASLONG);

common_s.h

@@ -231,7 +231,7 @@
 #define SGEMM_DIRECT_PERFORMANT gotoblas -> sgemm_direct_performant
 #define  SGEMM_DIRECT		gotoblas -> sgemm_direct
 #elif ARCH_ARM64
-#define SGEMM_DIRECT_PERFORMANT    sgemm_direct_performant
+#define SGEMM_DIRECT_PERFORMANT gotoblas -> sgemm_direct_performant
 #define  SGEMM_DIRECT		gotoblas -> sgemm_direct
 #define  SGEMM_DIRECT_ALPHA_BETA		gotoblas -> sgemm_direct_alpha_beta
 #define  SSYMM_DIRECT_ALPHA_BETA_LU		gotoblas -> ssymm_direct_alpha_betaLU