Adapt CMake for SVE

cmake/cc.cmake

@@ -139,6 +139,24 @@ if (${CORE} STREQUAL SAPPHIRERAPIDS)
   endif ()
 endif ()
 
+if (${CORE} STREQUAL A64FX)
+  if (NOT DYNAMIC_ARCH)
+    execute_process(COMMAND ${CMAKE_C_COMPILER} -dumpversion OUTPUT_VARIABLE GCC_VERSION)
+    if (${GCC_VERSION} VERSION_GREATER 11.0 OR ${GCC_VERSION} VERSION_EQUAL 11.0)
+      set (CCOMMON_OPT  "${CCOMMON_OPT} -march=armv8.2-a+sve -mtune=a64fx")
+    else ()
+      set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8.2-a+sve")
+    endif()
+  endif ()
+endif ()
+
+if (${CORE} STREQUAL ARMV8SVE)
+  if (NOT DYNAMIC_ARCH)
+    set (CCOMMON_OPT "${CCOMMON_OPT} -march=armv8-a+sve")
+  endif ()
+endif ()
+
+
 if (NOT DYNAMIC_ARCH)
 	if (HAVE_AVX2)
         set (CCOMMON_OPT  "${CCOMMON_OPT} -mavx2")

kernel/CMakeLists.txt

@@ -418,32 +418,50 @@ function (build_core TARGET_CORE KDIR TSUFFIX KERNEL_DEFINITIONS)
         GenerateCombinationObjects("${KERNELDIR}/${TRMM_KERNEL}" "LEFT;TRANSA" "R;N" "TRMMKERNEL" 2 "trmm_kernel" false ${float_type})
 
         # symm for s and d
+if (NOT DEFINED ${float_char}SYMMUCOPY_M)
+	set(SYMMUCOPY_M "generic/symm_ucopy_${${float_char}GEMM_UNROLL_M}.c")
+	set(SYMMLCOPY_M "generic/symm_lcopy_${${float_char}GEMM_UNROLL_M}.c")
+else ()
+	set(SYMMUCOPY_M "${KERNELDIR}/${${float_char}SYMMUCOPY_M}")
+	set(SYMMLCOPY_M "${KERNELDIR}/${${float_char}SYMMLCOPY_M}")
+endif()
         GenerateNamedObjects("generic/symm_ucopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "symm_outcopy" false "" "" false ${float_type})
-        GenerateNamedObjects("generic/symm_ucopy_${${float_char}GEMM_UNROLL_M}.c" "" "symm_iutcopy" false "" "" false ${float_type})
+        GenerateNamedObjects(${SYMMUCOPY_M} "" "symm_iutcopy" false "" "" false ${float_type})
 
         GenerateNamedObjects("generic/symm_lcopy_${${float_char}GEMM_UNROLL_N}.c" "LOWER;OUTER" "symm_oltcopy" false "" "" false ${float_type})
-        GenerateNamedObjects("generic/symm_lcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "symm_iltcopy" false "" "" false ${float_type})
+        GenerateNamedObjects(${SYMMLCOPY_M} "LOWER" "symm_iltcopy" false "" "" false ${float_type})
 
       # These don't use a scheme that is easy to iterate over - the filenames have part of the DEFINE codes in them, for UPPER/TRANS but not for UNIT/OUTER. Also TRANS is not passed in as a define.
       # Could simplify it a bit by pairing up by -UUNIT/-DUNIT.
 
-      GenerateNamedObjects("generic/trmm_uncopy_${${float_char}GEMM_UNROLL_M}.c" "UNIT" "trmm_iunucopy" false "" "" false ${float_type})
-      GenerateNamedObjects("generic/trmm_uncopy_${${float_char}GEMM_UNROLL_M}.c" "" "trmm_iunncopy" false "" "" false ${float_type})
+if (NOT DEFINED ${float_char}TRMMUNCOPY_M)
+	set(TRMMUNCOPY_M "generic/trmm_uncopy_${${float_char}GEMM_UNROLL_M}.c")
+	set(TRMMLNCOPY_M "generic/trmm_lncopy_${${float_char}GEMM_UNROLL_M}.c")
+	set(TRMMUTCOPY_M "generic/trmm_utcopy_${${float_char}GEMM_UNROLL_M}.c")
+	set(TRMMLTCOPY_M "generic/trmm_ltcopy_${${float_char}GEMM_UNROLL_M}.c")
+else ()
+	set(TRMMUNCOPY_M "${KERNELDIR}/${${float_char}TRMMUNCOPY_M}")
+	set(TRMMLNCOPY_M "${KERNELDIR}/${${float_char}TRMMLNCOPY_M}")
+	set(TRMMUTCOPY_M "${KERNELDIR}/${${float_char}TRMMUTCOPY_M}")
+	set(TRMMLTCOPY_M "${KERNELDIR}/${${float_char}TRMMLTCOPY_M}")
+endif ()
+      GenerateNamedObjects(${TRMMUNCOPY_M} "UNIT" "trmm_iunucopy" false "" "" false ${float_type})
+      GenerateNamedObjects(${TRMMUNCOPY_M} "" "trmm_iunncopy" false "" "" false ${float_type})
       GenerateNamedObjects("generic/trmm_uncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;UNIT" "trmm_ounucopy" false "" "" false ${float_type})
       GenerateNamedObjects("generic/trmm_uncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "trmm_ounncopy" false "" "" false ${float_type})
 
-      GenerateNamedObjects("generic/trmm_lncopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER;UNIT" "trmm_ilnucopy" false "" "" false ${float_type})
-      GenerateNamedObjects("generic/trmm_lncopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "trmm_ilnncopy" false "" "" false ${float_type})
+      GenerateNamedObjects(${TRMMLNCOPY_M} "LOWER;UNIT" "trmm_ilnucopy" false "" "" false ${float_type})
+      GenerateNamedObjects(${TRMMLNCOPY_M} "LOWER" "trmm_ilnncopy" false "" "" false ${float_type})
       GenerateNamedObjects("generic/trmm_lncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER;UNIT" "trmm_olnucopy" false "" "" false ${float_type})
       GenerateNamedObjects("generic/trmm_lncopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER" "trmm_olnncopy" false "" "" false ${float_type})
 
-      GenerateNamedObjects("generic/trmm_utcopy_${${float_char}GEMM_UNROLL_M}.c" "UNIT" "trmm_iutucopy" false "" "" false ${float_type})
-      GenerateNamedObjects("generic/trmm_utcopy_${${float_char}GEMM_UNROLL_M}.c" "" "trmm_iutncopy" false "" "" false ${float_type})
+      GenerateNamedObjects(${TRMMUTCOPY_M} "UNIT" "trmm_iutucopy" false "" "" false ${float_type})
+      GenerateNamedObjects(${TRMMUTCOPY_M} "" "trmm_iutncopy" false "" "" false ${float_type})
       GenerateNamedObjects("generic/trmm_utcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;UNIT" "trmm_outucopy" false "" "" false ${float_type})
       GenerateNamedObjects("generic/trmm_utcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER" "trmm_outncopy" false "" "" false ${float_type})
 
-      GenerateNamedObjects("generic/trmm_ltcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER;UNIT" "trmm_iltucopy" false "" "" false ${float_type})
-      GenerateNamedObjects("generic/trmm_ltcopy_${${float_char}GEMM_UNROLL_M}.c" "LOWER" "trmm_iltncopy" false "" "" false ${float_type})
+      GenerateNamedObjects(${TRMMLTCOPY_M} "LOWER;UNIT" "trmm_iltucopy" false "" "" false ${float_type})
+      GenerateNamedObjects(${TRMMLTCOPY_M} "LOWER" "trmm_iltncopy" false "" "" false ${float_type})
       GenerateNamedObjects("generic/trmm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER;UNIT" "trmm_oltucopy" false "" "" false ${float_type})
       GenerateNamedObjects("generic/trmm_ltcopy_${${float_char}GEMM_UNROLL_N}.c" "OUTER;LOWER" "trmm_oltncopy" false "" "" false ${float_type})