Added yet another naming scheme for lapack functions.

cmake/utils.cmake

@@ -136,6 +136,7 @@ endfunction ()
 #                  If 0, it will simply append the code, e.g. symm_L.c with TRANS and UNIT will be symm_LTU.
 #                  If 2, it will append the code with an underscore, e.g. symm.c with TRANS and UNIT will be symm_TU.
 #                  If 3, it will insert the code *around* the last character with an underscore, e.g. symm_L.c with TRANS and UNIT will be symm_TLU (required by BLAS level2 objects).
+#                  If 4, it will insert the code before the last underscore. E.g. trtri_U_parallel with TRANS will be trtri_UT_parallel
 # @param alternate_name replaces the source name as the object name (define codes are still appended)
 function(GenerateCombinationObjects sources_in defines_in absent_codes_in float_type_in all_defines_in replace_scheme)
 
@@ -184,6 +185,18 @@ function(GenerateCombinationObjects sources_in defines_in absent_codes_in float_
           string(SUBSTRING ${define_code} 0 1 define_code_first)
           string(SUBSTRING ${define_code} 1 -1 define_code_second)
           set(replace_code "${define_code_first}${last_letter}${define_code_second}")
+        elseif (replace_scheme EQUAL 4)
+          # insert code before the last underscore and pass that in as the alternate_name
+          get_filename_component(alternate_name ${source_file} NAME_WE)
+          set(extra_underscore "")
+          # check if filename has two underscores, insert another if not (e.g. getrs_parallel needs to become getrs_U_parallel not getrsU_parallel)
+          string(REGEX MATCH "_[a-zA-Z]+_" underscores ${alternate_name})
+          string(LENGTH "${underscores}" underscores)
+          if (underscores EQUAL 0)
+            set(extra_underscore "_")
+          endif ()
+          string(REGEX REPLACE "(.+)(_[^_]+)$" "\\1${extra_underscore}${define_code}\\2" alternate_name ${alternate_name})
+          message(STATUS ${alternate_name})
         else()
           set(append_code ${define_code}) # replace_scheme should be 0
         endif ()

lapack/CMakeLists.txt

@@ -4,12 +4,10 @@ include_directories(${CMAKE_SOURCE_DIR})
 # TODO: laswp needs arch specific code
 # TODO: getrs needs to be compiled with and without TRANS (and up to TRANS=4 in the complex case)
 # TODO: trti2 needs to be compiled with and without UNIT
-# TODO: trtri needs to be compiled with and without UNIT
 
 set(LAPACK_SOURCES
   getf2/getf2_k.c
   getrf/getrf_single.c
-  getrs/getrs_single.c
   potrf/potrf_U_single.c
   potrf/potrf_L_single.c
   potf2/potf2_U.c
@@ -20,6 +18,15 @@ set(LAPACK_SOURCES
   lauum/lauum_L_single.c
   trti2/trti2_U.c
   trti2/trti2_L.c
+)
+
+# sources that need TRANS set
+set(TRANS_SOURCES
+  getrs/getrs_single.c
+)
+
+# sources that need UNIT set
+set(UNIT_SOURCES
   trtri/trtri_U_single.c
   trtri/trtri_L_single.c
 )
@@ -55,11 +62,17 @@ if (SMP)
 
   set(PARALLEL_SOURCES
     ${GETRF_SRC}
-    getrs/getrs_parallel.c
     potrf/potrf_U_parallel.c
     potrf/potrf_L_parallel.c
     lauum/lauum_U_parallel.c
     lauum/lauum_L_parallel.c
+  )
+
+  list(APPEND TRANS_SOURCES
+    getrs/getrs_parallel.c
+  )
+
+  list(APPEND UNIT_SOURCES
     trtri/trtri_U_parallel.c
     trtri/trtri_L_parallel.c
   )
@@ -79,5 +92,10 @@ if (SMP)
   list(APPEND DBLAS_OBJS ${OBJ_LIST_OUT})
 endif ()
 
+GenerateCombinationObjects("${TRANS_SOURCES}" "TRANS" "N" "DOUBLE" "" 4)
+list(APPEND DBLAS_OBJS ${COMBO_OBJ_LIST_OUT})
+GenerateCombinationObjects("${UNIT_SOURCES}" "UNIT" "N" "DOUBLE" "" 4)
+list(APPEND DBLAS_OBJS ${COMBO_OBJ_LIST_OUT})
+
 set(DBLAS_OBJS ${DBLAS_OBJS} PARENT_SCOPE) # list append removes the scope from DBLAS_OBJS