Merge pull request #4811 from yamazakimitsufumi/add_a64fx_to_dynamic_arch

Add A64FX to the list of CPUs supported by DYNAMIC_ARCH

Author: martin-frbg

Makefile.system

@@ -689,6 +689,7 @@ ifneq ($(NO_SVE), 1)
 DYNAMIC_CORE += NEOVERSEV1
 DYNAMIC_CORE += NEOVERSEN2
 DYNAMIC_CORE += ARMV8SVE
+DYNAMIC_CORE += A64FX
 endif
 DYNAMIC_CORE += THUNDERX
 DYNAMIC_CORE += THUNDERX2T99

cmake/arch.cmake

@@ -46,7 +46,7 @@ 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)
+	    set(DYNAMIC_CORE ${DYNAMIC_CORE} NEOVERSEV1 NEOVERSEN2 ARMV8SVE A64FX)
     endif ()
     if (DYNAMIC_LIST)
 	  set(DYNAMIC_CORE ARMV8 ${DYNAMIC_LIST})

cmake/prebuild.cmake

@@ -1218,6 +1218,37 @@ endif ()
     set(ZGEMM_UNROLL_M 4)
     set(ZGEMM_UNROLL_N 4)
     set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "A64FX")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_CODE_SIZE\t65536\n"
+      "#define L1_CODE_LINESIZE\t256\n"
+      "#define L1_CODE_ASSOCIATIVE\t8\n"
+      "#define L1_DATA_SIZE\t32768\n"
+      "#define L1_DATA_LINESIZE\t256\n"
+      "#define L1_DATA_ASSOCIATIVE\t8\n"
+      "#define L2_SIZE\t8388608\n\n"
+      "#define L2_LINESIZE\t256\n"
+      "#define L2_ASSOCIATIVE\t8\n"
+      "#define L3_SIZE\t0\n\n"
+      "#define L3_LINESIZE\t0\n\n"
+      "#define L3_ASSOCIATIVE\t0\n\n"
+      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define HAVE_VFPV4\n"
+      "#define HAVE_VFPV3\n"
+      "#define HAVE_VFP\n"
+      "#define HAVE_NEON\n"
+      "#define HAVE_SVE\n"
+      "#define ARMV8\n")
+    set(SGEMM_UNROLL_M 4)
+    set(SGEMM_UNROLL_N 8)
+    set(DGEMM_UNROLL_M 2)
+    set(DGEMM_UNROLL_N 8)
+    set(CGEMM_UNROLL_M 2)
+    set(CGEMM_UNROLL_N 4)
+    set(ZGEMM_UNROLL_M 2)
+    set(ZGEMM_UNROLL_N 4)
+    set(SYMV_P 16)
   elseif ("${TCORE}" STREQUAL "P5600")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define L2_SIZE 1048576\n"

cmake/system.cmake

@@ -310,6 +310,18 @@ if (${TARGET} STREQUAL NEOVERSEV1)
       set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv8.2-a+sve")
     endif()
   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")
+    else ()
+      execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
+      if (${GCC_VERSION} VERSION_GREATER 10.4 OR ${GCC_VERSION} VERSION_EQUAL 10.4)
+        set (KERNEL_DEFINITIONS  "${KERNEL_DEFINITIONS} -march=armv8.2-a+sve -mtune=a64fx")
+      else ()
+        message(FATAL_ERROR "Compiler $${CMAKE_C_COMPILER} {GCC_VERSION} does not support A64FX.")
+      endif()
+    endif()
+  endif()
 
 endif()
 

driver/others/dynamic_arm64.c

@@ -120,6 +120,11 @@ extern gotoblas_t  gotoblas_CORTEXA55;
 #else
 #define gotoblas_CORTEXA55 gotoblas_ARMV8
 #endif
+#ifdef DYN_A64FX
+extern gotoblas_t gotoblas_A64FX;
+#else
+#define gotoblas_A64FX gotoblas_ARMV8
+#endif
 #else
 extern gotoblas_t  gotoblas_CORTEXA53;
 #define gotoblas_CORTEXA55 gotoblas_CORTEXA53
@@ -136,10 +141,12 @@ extern gotoblas_t  gotoblas_NEOVERSEN1;
 extern gotoblas_t  gotoblas_NEOVERSEV1;
 extern gotoblas_t  gotoblas_NEOVERSEN2;
 extern gotoblas_t  gotoblas_ARMV8SVE;
+extern gotoblas_t  gotoblas_A64FX;
 #else
 #define gotoblas_NEOVERSEV1 gotoblas_ARMV8
 #define gotoblas_NEOVERSEN2 gotoblas_ARMV8
 #define gotoblas_ARMV8SVE   gotoblas_ARMV8
+#define gotoblas_A64FX      gotoblas_ARMV8
 #endif
 extern gotoblas_t  gotoblas_THUNDERX3T110;
 #endif
@@ -149,7 +156,7 @@ 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   17
+#define NUM_CORETYPES   18
 
 /*
  * In case asm/hwcap.h is outdated on the build system, make sure
@@ -184,6 +191,7 @@ static char *corename[] = {
   "thunderx3t110",
   "cortexa55",
   "armv8sve",
+  "a64fx",
   "unknown"
 };
 
@@ -205,6 +213,7 @@ char *gotoblas_corename(void) {
   if (gotoblas == &gotoblas_THUNDERX3T110) return corename[14];
   if (gotoblas == &gotoblas_CORTEXA55)    return corename[15];
   if (gotoblas == &gotoblas_ARMV8SVE)     return corename[16];
+  if (gotoblas == &gotoblas_A64FX)        return corename[17];
   return corename[NUM_CORETYPES];
 }
 
@@ -241,6 +250,7 @@ static gotoblas_t *force_coretype(char *coretype) {
     case 14: return (&gotoblas_THUNDERX3T110);
     case 15: return (&gotoblas_CORTEXA55);
     case 16: return (&gotoblas_ARMV8SVE);
+    case 17: return (&gotoblas_A64FX);
   }
   snprintf(message, 128, "Core not found: %s\n", coretype);
   openblas_warning(1, message);
@@ -346,6 +356,15 @@ static gotoblas_t *get_coretype(void) {
           return &gotoblas_THUNDERX3T110;
       }
       break;
+    case 0x46: // Fujitsu
+      switch (part)
+      {
+#ifndef NO_SVE
+        case 0x001: // A64FX
+          return &gotoblas_A64FX;
+#endif
+      }
+      break;
     case 0x48: // HiSilicon
       switch (part)
       {

kernel/arm64/gemv_n_sve.c

@@ -37,13 +37,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SV_COUNT svcntd
 #define SV_TYPE svfloat64_t
 #define SV_TRUE svptrue_b64
-#define SV_WHILE svwhilelt_b64
+#define SV_WHILE svwhilelt_b64_s64
 #define SV_DUP svdup_f64
 #else
 #define SV_COUNT svcntw
 #define SV_TYPE svfloat32_t
 #define SV_TRUE svptrue_b32
-#define SV_WHILE svwhilelt_b32
+#define SV_WHILE svwhilelt_b32_s64
 #define SV_DUP svdup_f32
 #endif
 

kernel/arm64/gemv_t_sve.c

@@ -37,13 +37,13 @@ USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 #define SV_COUNT svcntd
 #define SV_TYPE svfloat64_t
 #define SV_TRUE svptrue_b64
-#define SV_WHILE svwhilelt_b64
+#define SV_WHILE svwhilelt_b64_s64
 #define SV_DUP svdup_f64
 #else
 #define SV_COUNT svcntw
 #define SV_TYPE svfloat32_t
 #define SV_TRUE svptrue_b32
-#define SV_WHILE svwhilelt_b32
+#define SV_WHILE svwhilelt_b32_s64
 #define SV_DUP svdup_f32
 #endif