Merge pull request #5342 from martin-frbg/cmake_ampere

Add CMake build settings for the Ampere One cpu

cmake/cc.cmake

@@ -236,6 +236,18 @@ if (${CORE} STREQUAL NEOVERSEN1)
   endif ()
 endif ()
 
+if (${CORE} STREQUAL AMPEREONE)
+  if (NOT DYNAMIC_ARCH)
+    if (${CMAKE_C_COMPILER_ID} STREQUAL "NVC")
+	set (CCOMMON_OPT  "${CCOMMON_OPT} -tp=neoverse-n1")
+    elseif (${GCC_VERSION} VERSION_GREATER 12.1)
+      set (CCOMMON_OPT  "${CCOMMON_OPT} -march=armv8.6-a+crypto+crc+fp16+sha3+rng -mtune=ampereone")
+    else ()
+      set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.6-a+fp16")
+    endif()
+  endif ()
+endif ()
+
 if (${CORE} STREQUAL ARMV8SVE)
   if (NOT DYNAMIC_ARCH)
     if (${CMAKE_C_COMPILER_ID} STREQUAL "PGI" AND NOT NO_SVE)

cmake/prebuild.cmake

@@ -1195,6 +1195,33 @@ endif ()
     set(ZGEMM_UNROLL_M 4)
     set(ZGEMM_UNROLL_N 4)
     set(SYMV_P 16)
+  elseif ("${TCORE}" STREQUAL "AMPEREONE")
+    file(APPEND ${TARGET_CONF_TEMP}
+      "#define L1_CODE_SIZE\t16384\n"
+      "#define L1_CODE_LINESIZE\t64\n"
+      "#define L1_CODE_ASSOCIATIVE\t4\n"
+      "#define L1_DATA_SIZE\t65536\n"
+      "#define L1_DATA_LINESIZE\t64\n"
+      "#define L1_DATA_ASSOCIATIVE\t4\n"
+      "#define L2_SIZE\t2097152\n\n"
+      "#define L2_LINESIZE\t64\n"
+      "#define L2_ASSOCIATIVE\t8\n"
+      "#define DTB_DEFAULT_ENTRIES\t64\n"
+      "#define DTB_SIZE\t4096\n"
+      "#define HAVE_VFPV4\n"
+      "#define HAVE_VFPV3\n"
+      "#define HAVE_VFP\n"
+      "#define HAVE_NEON\n"
+      "#define ARMV8\n")
+    set(SGEMM_UNROLL_M 16)
+    set(SGEMM_UNROLL_N 4)
+    set(DGEMM_UNROLL_M 8)
+    set(DGEMM_UNROLL_N 4)
+    set(CGEMM_UNROLL_M 8)
+    set(CGEMM_UNROLL_N 4)
+    set(ZGEMM_UNROLL_M 4)
+    set(ZGEMM_UNROLL_N 4)
+    set(SYMV_P 16)
   elseif ("${TCORE}" STREQUAL "VORTEX")
     file(APPEND ${TARGET_CONF_TEMP}
       "#define ARMV8\n"