|
|
@@ -0,0 +1,106 @@ |
|
|
|
include_directories(${CMAKE_SOURCE_DIR}) |
|
|
|
|
|
|
|
set(USE_GEMM3M 0) |
|
|
|
|
|
|
|
if (DEFINED ARCH) |
|
|
|
if (${ARCH} STREQUAL "x86") |
|
|
|
set(USE_GEMM3M 1) |
|
|
|
endif () |
|
|
|
|
|
|
|
if (${ARCH} STREQUAL "x86_64") |
|
|
|
set(USE_GEMM3M 1) |
|
|
|
endif () |
|
|
|
|
|
|
|
if (${ARCH} STREQUAL "ia64") |
|
|
|
set(USE_GEMM3M 1) |
|
|
|
endif () |
|
|
|
|
|
|
|
if (${ARCH} STREQUAL "MIPS") |
|
|
|
set(USE_GEMM3M 1) |
|
|
|
endif () |
|
|
|
endif () |
|
|
|
|
|
|
|
# N.B. In the original makefile there was a BLOCKS define used in the compilation of these files but I don't see any evidence of it being set anywhere. -hpa |
|
|
|
|
|
|
|
# loop through gemm.c defines |
|
|
|
set(GEMM_DEFINES NN NT TN TT) |
|
|
|
foreach (GEMM_DEFINE ${GEMM_DEFINES}) |
|
|
|
add_library(GEMM_${GEMM_DEFINE}_OBJS OBJECT gemm.c) |
|
|
|
set_target_properties(GEMM_${GEMM_DEFINE}_OBJS PROPERTIES COMPILE_DEFINITIONS ${GEMM_DEFINE}) |
|
|
|
endforeach () |
|
|
|
|
|
|
|
# Returns all combinations of the input list, as a list with colon-separated combinations |
|
|
|
# E.g. input of A B C returns A B C A:B A:C B:C |
|
|
|
# N.B. The input is meant to be a list, and to past a list to a function in CMake you must quote it (e.g. AllCombinations("${LIST_VAR}")). |
|
|
|
function(AllCombinations list_in) |
|
|
|
list(LENGTH list_in list_count) |
|
|
|
set(num_combos 1) |
|
|
|
math(EXPR num_combos "${num_combos} << ${list_count}") |
|
|
|
set(LIST_OUT "") |
|
|
|
foreach (c RANGE ${num_combos}) |
|
|
|
set(current_combo "") |
|
|
|
# this is a little ridiculous just to iterate through a list w/ indices |
|
|
|
math(EXPR last_list_index "${list_count} - 1") |
|
|
|
foreach (list_index RANGE 0 ${last_list_index}) |
|
|
|
math(EXPR bit "1 << ${list_index}") |
|
|
|
math(EXPR combo_has_bit "${c} & ${bit}") |
|
|
|
list(GET list_in ${list_index} list_elem) |
|
|
|
if (combo_has_bit) |
|
|
|
if (current_combo) |
|
|
|
set(current_combo "${current_combo}:${list_elem}") |
|
|
|
else () |
|
|
|
set(current_combo ${list_elem}) |
|
|
|
endif () |
|
|
|
endif () |
|
|
|
endforeach () |
|
|
|
list(APPEND LIST_OUT ${current_combo}) |
|
|
|
endforeach () |
|
|
|
set(LIST_OUT ${LIST_OUT} PARENT_SCOPE) |
|
|
|
endfunction () |
|
|
|
|
|
|
|
# these sources are compiled with combinations of TRANS, UPPER, and UNIT, for 32 combinations total |
|
|
|
set(TRM_SOURCES trmm_L.c trmm_R.c trsm_L.c trsm_R.c) |
|
|
|
AllCombinations("TRANS UPPER UNIT") |
|
|
|
set(TRM_DEFINE_COMBOS LIST_OUT) |
|
|
|
message(STATUS "alcombos result: ${LIST_OUT}") |
|
|
|
foreach (TRM_SOURCE ${TRM_SOURCES}) |
|
|
|
foreach (TRM_DEFINES ${TRM_DEFINE_COMBOS}) |
|
|
|
string(REGEX MATCH "[a-z]+_[LR]" TRM_NAME ${TRM_SOURCE}) |
|
|
|
string(TOUPPER ${TRM_NAME} TRM_NAME) |
|
|
|
# TODO: TRM_DEFINES is a colon-separated list of defines to set for this object - need to parse it and set them using set_target_properties, and also come up with a unique id for the lib name (e.g. first letter of each define, so TRANS UPPER UNIT is TUU) |
|
|
|
#add_library(${TRM_NAME}_${TRM_DEFINE}_OBJS OBJECT ${TRM_SOURCE}) |
|
|
|
#set_target_properties(${TRM_NAME}_${TRM_DEFINE}_OBJS PROPERTIES COMPILE_DEFINITIONS ${TRM_DEFINE}) |
|
|
|
endforeach () |
|
|
|
endforeach () |
|
|
|
|
|
|
|
# dsymm_LU.c dsymm_LL.c dsymm_RU.c dsymm_RL.c |
|
|
|
# dsyrk_UN.c dsyrk_UT.c dsyrk_LN.c dsyrk_LT.c |
|
|
|
# dsyr2k_UN.c dsyr2k_UT.c dsyr2k_LN.c dsyr2k_LT.c |
|
|
|
# dsyrk_kernel_U.c dsyrk_kernel_L.c |
|
|
|
# dsyr2k_kernel_U.c dsyr2k_kernel_L.c |
|
|
|
|
|
|
|
#if (SMP) |
|
|
|
# |
|
|
|
# COMMONOBJS += gemm_thread_m.c gemm_thread_n.c gemm_thread_mn.c gemm_thread_variable.c |
|
|
|
# COMMONOBJS += syrk_thread.c |
|
|
|
# |
|
|
|
# if (USE_SIMPLE_THREADED_LEVEL3) |
|
|
|
# DBLASOBJS += dgemm_thread_nn.c dgemm_thread_nt.c dgemm_thread_tn.c dgemm_thread_tt.c |
|
|
|
# DBLASOBJS += dsymm_thread_LU.c dsymm_thread_LL.c dsymm_thread_RU.c dsymm_thread_RL.c |
|
|
|
# DBLASOBJS += dsyrk_thread_UN.c dsyrk_thread_UT.c dsyrk_thread_LN.c dsyrk_thread_LT.c |
|
|
|
# |
|
|
|
# endif () |
|
|
|
#endif () |
|
|
|
# |
|
|
|
#HPLOBJS = |
|
|
|
# dgemm_nn.c dgemm_nt.c dgemm_tn.c dgemm_tt.c |
|
|
|
# dtrsm_LNUU.c dtrsm_LNUN.c dtrsm_LNLU.c dtrsm_LNLN.c |
|
|
|
# dtrsm_LTUU.c dtrsm_LTUN.c dtrsm_LTLU.c dtrsm_LTLN.c |
|
|
|
# dtrsm_RNUU.c dtrsm_RNUN.c dtrsm_RNLU.c dtrsm_RNLN.c |
|
|
|
# dtrsm_RTUU.c dtrsm_RTUN.c dtrsm_RTLU.c dtrsm_RTLN.c |
|
|
|
# |
|
|
|
#if (USE_SIMPLE_THREADED_LEVEL3) |
|
|
|
# HPLOBJS += dgemm_thread_nn.c dgemm_thread_nt.c |
|
|
|
# dgemm_thread_tn.c dgemm_thread_tt.c |
|
|
|
#endif |
|
|
|
# |