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

Makefile.arm64

@@ -303,6 +303,11 @@ FCOMMON_OPT += -march=armv8.3-a
 endif
 endif
 
+ifeq ($(CORE), VORTEXM4)
+CCOMMON_OPT += -march=armv8.4-a+sme
+FCOMMON_OPT += -march=armv8.4-a+sme
+endif
+
 ifeq (1, $(filter 1,$(GCCVERSIONGTEQ9) $(ISCLANG)))
 ifeq ($(CORE), TSV110)
 CCOMMON_OPT += -march=armv8.2-a -mtune=tsv110

Makefile.system

@@ -427,7 +427,7 @@ ifndef MACOSX_DEPLOYMENT_TARGET
 ifeq ($(ARCH), arm64)
 export MACOSX_DEPLOYMENT_TARGET=11.0
 export NO_SVE = 1
-export NO_SME = 1
+# export NO_SME = 1
 else
 export MACOSX_DEPLOYMENT_TARGET=10.8
 endif
@@ -723,6 +723,7 @@ DYNAMIC_CORE += A64FX
 endif
 ifneq ($(NO_SME), 1)
 DYNAMIC_CORE += ARMV9SME
+DYNAMIC_CORE += VORTEXM4
 endif
 DYNAMIC_CORE += THUNDERX
 DYNAMIC_CORE += THUNDERX2T99

TargetList.txt

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

cmake/arch.cmake

@@ -39,14 +39,14 @@ if (DYNAMIC_ARCH)
         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)
+        set(DYNAMIC_CORE ${DYNAMIC_CORE} ARMV9SME VORTEXM4)
       endif()
     elseif (${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

@@ -315,6 +315,16 @@ if (${CORE} STREQUAL ARMV9SME)
   endif ()
 endif ()
 
+if (${CORE} STREQUAL VORTEXM4)
+  if (NOT DYNAMIC_ARCH)
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "NVC" AND NOT NO_SVE)
+	    set (CCOMMON_OPT  "${CCOMMON_OPT} -tp=host")
+    else ()
+    set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.4-a+sme")
+    endif ()
+  endif ()
+endif ()
+
 if (${CORE} STREQUAL CORTEXA510)
   if (NOT DYNAMIC_ARCH)
     set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.4-a+sve")

cmake/prebuild.cmake

@@ -1252,7 +1252,7 @@ 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"

cmake/system.cmake

@@ -361,6 +361,9 @@ if (${TARGET} STREQUAL NEOVERSEV1)
   if (${TARGET} STREQUAL ARMV9SME)
       set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv9-a+sme -O3")
   endif()
+  if (${TARGET} STREQUAL VORTEXM4)
+      set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv8.4-a+sme -O3")
+  endif()
   if (${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")

cmake/system_check.cmake

@@ -142,7 +142,7 @@ endif()
 if (ARM64)
 if (NOT NO_SME)
   file(WRITE ${PROJECT_BINARY_DIR}/sme.c ".text \n.global sme_test\n\nsme_test:\nsmstart\nsmstop\nret\n")
-  execute_process(COMMAND ${CMAKE_C_COMPILER} -march=armv9-a+sve2+sme -c -v -o ${PROJECT_BINARY_DIR}/sme.o ${PROJECT_BINARY_DIR}/sme.c OUTPUT_QUIET ERROR_QUIET RESULT_VARIABLE NO_SME)
+  execute_process(COMMAND ${CMAKE_C_COMPILER} -march=armv8.4-a+sme -c -v -o ${PROJECT_BINARY_DIR}/sme.o ${PROJECT_BINARY_DIR}/sme.c OUTPUT_QUIET ERROR_QUIET RESULT_VARIABLE NO_SME)
 if (NO_SME EQUAL 1)
 set (CCOMMON_OPT "${CCOMMON_OPT} -DNO_SME")
 endif()

common_param.h

@@ -257,6 +257,7 @@ int (*shgemm_otcopy   )(BLASLONG, BLASLONG, hfloat16 *, BLASLONG, hfloat16 *);
 #ifdef ARCH_ARM64
   void (*sgemm_direct) (BLASLONG, BLASLONG, BLASLONG, float *, BLASLONG , float *, BLASLONG , float * , BLASLONG);
   void (*sgemm_direct_alpha_beta) (BLASLONG, BLASLONG, BLASLONG, float, float *, BLASLONG, float *, BLASLONG, float, float * , BLASLONG);
+  int  (*sgemm_direct_performant) (BLASLONG M, BLASLONG N, BLASLONG K);
 #endif
 
 

common_s.h

@@ -217,7 +217,7 @@
 #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
 #endif

cpuid_arm64.c

@@ -82,6 +82,7 @@ size_t length64=sizeof(value64);
 #define CPU_AMPERE1      25
 // Apple
 #define CPU_VORTEX       13
+#define CPU_VORTEXM4     26 
 // Fujitsu
 #define CPU_A64FX	 15
 // Phytium
@@ -113,7 +114,8 @@ static char *cpuname[] = {
   "FT2000",
   "CORTEXA76",
   "NEOVERSEV2",
-  "AMPERE1"
+  "AMPERE1",
+  "VORTEXM4",
 };
 
 static char *cpuname_lower[] = {
@@ -143,7 +145,7 @@ static char *cpuname_lower[] = {
   "cortexa76",
   "neoversev2",
   "ampere1",
-  "ampere1a"
+  "vortexm4"
 };
 
 static int cpulowperf=0;
@@ -400,7 +402,7 @@ int detect(void)
 	if (value64 ==131287967|| value64 == 458787763 ) return CPU_VORTEX; //A12/M1
 	if (value64 == 3660830781) return CPU_VORTEX; //A15/M2
         if (value64 == 2271604202) return CPU_VORTEX; //A16/M3
-        if (value64 == 1867590060) return CPU_VORTEX; //M4
+        if (value64 == 1867590060) return CPU_VORTEXM4; //M4
 #else
 #ifdef OS_WINDOWS
 	HKEY reghandle;
@@ -740,6 +742,27 @@ void get_cpuconfig(void)
 		length64 = sizeof(value64);
 		sysctlbyname("hw.l2cachesize",&value64,&length64,NULL,0);
 		printf("#define L2_SIZE	     %lld       \n",value64);
+#endif	
+		printf("#define DTB_DEFAULT_ENTRIES  64       \n");
+		printf("#define DTB_SIZE             4096     \n");
+		break;
+	    case CPU_VORTEXM4:
+		printf("#define VORTEXM4		      \n");
+		printf("#define HAVE_SME 1		      \n");
+#ifdef __APPLE__
+		length64 = sizeof(value64);
+		sysctlbyname("hw.l1icachesize",&value64,&length64,NULL,0);
+		printf("#define L1_CODE_SIZE	     %lld       \n",value64);
+		length64 = sizeof(value64);
+		sysctlbyname("hw.cachelinesize",&value64,&length64,NULL,0);
+		printf("#define L1_CODE_LINESIZE     %lld       \n",value64);
+		printf("#define L1_DATA_LINESIZE     %lld       \n",value64);
+		length64 = sizeof(value64);
+		sysctlbyname("hw.l1dcachesize",&value64,&length64,NULL,0);
+		printf("#define L1_DATA_SIZE	     %lld       \n",value64);
+		length64 = sizeof(value64);
+		sysctlbyname("hw.l2cachesize",&value64,&length64,NULL,0);
+		printf("#define L2_SIZE	     %lld       \n",value64);
 #endif	
 		printf("#define DTB_DEFAULT_ENTRIES  64       \n");
 		printf("#define DTB_SIZE             4096     \n");

driver/others/dynamic_arm64.c

@@ -128,6 +128,12 @@ extern gotoblas_t gotoblas_ARMV9SME;
 #else
 #define gotoblas_ARMV9SME gotoblas_ARMV8
 #endif
+#ifdef DYN_VORTEXM4
+extern gotoblas_t gotoblas_VORTEXM4;
+#else
+#error "dont have vortexm4"
+#define gotoblas_VORTEXM4 gotoblas_ARMV8
+#endif
 #ifdef DYN_CORTEXA55
 extern gotoblas_t  gotoblas_CORTEXA55;
 #else
@@ -155,17 +161,22 @@ extern gotoblas_t  gotoblas_NEOVERSEV1;
 extern gotoblas_t  gotoblas_NEOVERSEN2;
 extern gotoblas_t  gotoblas_ARMV8SVE;
 extern gotoblas_t  gotoblas_A64FX;
-#ifndef NO_SME
-extern gotoblas_t  gotoblas_ARMV9SME;
-#else
-#define gotoblas_ARMV9SME gotoblas_ARMV8SVE
-#endif
 #else
 #define gotoblas_NEOVERSEV1 gotoblas_ARMV8
 #define gotoblas_NEOVERSEN2 gotoblas_ARMV8
 #define gotoblas_ARMV8SVE   gotoblas_ARMV8
 #define gotoblas_A64FX      gotoblas_ARMV8
-#define gotoblas_ARMV9SME   gotoblas_ARMV8
+#endif
+#ifndef NO_SME
+extern gotoblas_t  gotoblas_ARMV9SME;
+extern gotoblas_t  gotoblas_VORTEXM4;
+#else
+#ifndef NO_SVE
+#define gotoblas_ARMV9SME gotoblas_ARMV8SVE
+#else
+#define gotoblas_ARMV9SME gotoblas_NEOVERSEN1
+#endif
+#define gotoblas_VORTEXM4 gotoblas_NEOVERSEN1
 #endif
 
 extern gotoblas_t  gotoblas_THUNDERX3T110;
@@ -176,7 +187,7 @@ extern void openblas_warning(int verbose, const char * msg);
 #define FALLBACK_VERBOSE 1
 #define NEOVERSEN1_FALLBACK "OpenBLAS : Your OS does not support SVE instructions. OpenBLAS is using Neoverse N1 kernels as a fallback, which may give poorer performance.\n"
 
-#define NUM_CORETYPES   19
+#define NUM_CORETYPES   20
 
 /*
  * In case asm/hwcap.h is outdated on the build system, make sure
@@ -216,6 +227,7 @@ static char *corename[] = {
   "armv8sve",
   "a64fx",
   "armv9sme",
+  "vortexm4",
   "unknown"
 };
 
@@ -239,6 +251,7 @@ char *gotoblas_corename(void) {
   if (gotoblas == &gotoblas_ARMV8SVE)     return corename[16];
   if (gotoblas == &gotoblas_A64FX)        return corename[17];
   if (gotoblas == &gotoblas_ARMV9SME)     return corename[18];
+  if (gotoblas == &gotoblas_VORTEXM4)     return corename[19];
   return corename[NUM_CORETYPES];
 }
 
@@ -277,6 +290,7 @@ static gotoblas_t *force_coretype(char *coretype) {
     case 16: return (&gotoblas_ARMV8SVE);
     case 17: return (&gotoblas_A64FX);
     case 18: return (&gotoblas_ARMV9SME);
+    case 19: return (&gotoblas_VORTEXM4);
   }
   snprintf(message, 128, "Core not found: %s\n", coretype);
   openblas_warning(1, message);
@@ -288,11 +302,11 @@ static gotoblas_t *get_coretype(void) {
   char coremsg[128];
 
 #if defined (OS_DARWIN)
-//future	#if !defined(NO_SME)
-//  		if (support_sme1()) {
-//    			return &gotoblas_ARMV9SME;
-//		  }
-//		#endif
+#if !defined(NO_SME)
+  if (support_sme1()) {
+   	return &gotoblas_VORTEXM4;
+  }
+#endif
   return &gotoblas_NEOVERSEN1;
 #endif
 
@@ -463,7 +477,7 @@ static gotoblas_t *get_coretype(void) {
       }
       break;
     case 0x61: // Apple
-//future	if (support_sme1()) return &gotoblas_ARMV9SME;
+	if (support_sme1()) return &gotoblas_VORTEXM4;
 	return &gotoblas_NEOVERSEN1;
       break;
     default:

getarch.c

@@ -1654,6 +1654,20 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define CORENAME  "VORTEX"
 #endif
 
+#ifdef FORCE_VORTEXM4
+#define FORCE
+#define ARCHITECTURE    "ARM64"
+#define SUBARCHITECTURE "VORTEXM4"
+#define SUBDIRNAME      "arm64"
+#define ARCHCONFIG   "-DVORTEXM4 " \
+       "-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_SME -DARMV8"
+#define LIBNAME   "vortexm4"
+#define CORENAME  "VORTEXM4"
+#endif
+
 #ifdef FORCE_A64FX
 #define ARMV8
 #define FORCE

interface/gemm.c

@@ -266,6 +266,7 @@ void NAME(char *TRANSA, char *TRANSB,
 
   int transa, transb, nrowa, nrowb;
   blasint info;
+  int order = -1;
 
   char transA, transB;
   IFLOAT *buffer;
@@ -424,30 +425,6 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
 
   PRINT_DEBUG_CNAME;
 
-#if !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16)  && !defined(HFLOAT16)
-#if defined(ARCH_x86) && (defined(USE_SGEMM_KERNEL_DIRECT)||defined(DYNAMIC_ARCH))
-#if defined(DYNAMIC_ARCH)
-  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(ARCH_ARM64) && (defined(USE_SGEMM_KERNEL_DIRECT)||defined(DYNAMIC_ARCH))
-#if defined(DYNAMIC_ARCH)
- if (support_sme1())
-#endif
-  if (beta == 0 && alpha == 1.0 && order == CblasRowMajor && TransA == CblasNoTrans && TransB == CblasNoTrans) {
-	SGEMM_DIRECT(m, n, k, a, lda, b, ldb, c, ldc);
-	return;
-  }else if (order == CblasRowMajor && TransA == CblasNoTrans && TransB == CblasNoTrans) {
-	SGEMM_DIRECT_ALPHA_BETA(m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
-	return;
-  }
-#endif
-#endif
-
 #ifndef COMPLEX
   args.alpha = (void *)α
   args.beta  = (void *)β
@@ -564,6 +541,36 @@ void CNAME(enum CBLAS_ORDER order, enum CBLAS_TRANSPOSE TransA, enum CBLAS_TRANS
     return;
   }
 
+
+  if ((args.m == 0) || (args.n == 0)) return;
+#if !defined(COMPLEX) && !defined(DOUBLE) && !defined(BFLOAT16)  && !defined(HFLOAT16)
+#if defined(ARCH_x86) && (defined(USE_SGEMM_KERNEL_DIRECT)||defined(DYNAMIC_ARCH))
+#if defined(DYNAMIC_ARCH)
+  if (support_avx512() )
+#endif
+  if (order == CblasRowMajor && beta == 0 && alpha == 1.0 && 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(ARCH_ARM64) && (defined(USE_SGEMM_KERNEL_DIRECT)||defined(DYNAMIC_ARCH))
+#if defined(DYNAMIC_ARCH)
+if (strcmp(gotoblas_corename(), "armv9sme") == 0 || strcmp(gotoblas_corename(), "vortexm4") == 0)
+// if (support_sme1())
+#endif
+  if (order == CblasRowMajor && m==lda && n ==ldb && k==ldc && beta == 0 && alpha == 1.0 && TransA == CblasNoTrans && TransB == CblasNoTrans&& SGEMM_DIRECT_PERFORMANT(m,n,k)) {
+        SGEMM_DIRECT(m, n, k, a, lda, b, ldb, c, ldc);
+        return;
+  }
+else
+ if (order == CblasRowMajor && m==lda && n==ldb && k==ldc && TransA == CblasNoTrans && TransB == CblasNoTrans&& SGEMM_DIRECT_PERFORMANT(m,n,k)) {
+        SGEMM_DIRECT_ALPHA_BETA(m, n, k, alpha, a, lda, b, ldb, beta, c, ldc);
+        return;
+  }
+
+#endif
+#endif
+
 #endif
 
 #if defined(__linux__) && defined(__x86_64__) && defined(BFLOAT16)