Skip to content

Commit 9c2a834

Browse files
committed
compile kernel_align also for reference and wrapper backends
1 parent 875bb12 commit 9c2a834

File tree

5 files changed

+103
-69
lines changed

5 files changed

+103
-69
lines changed

CMakeLists.txt

Lines changed: 46 additions & 22 deletions
Original file line numberDiff line numberDiff line change
@@ -691,12 +691,14 @@ file(GLOB KERNEL_SRC
691691
${PROJECT_SOURCE_DIR}/kernel/avx512/kernel_dgeqrf_8_lib8.c
692692
${PROJECT_SOURCE_DIR}/kernel/avx512/kernel_dgelqf_lib8.S
693693

694-
${PROJECT_SOURCE_DIR}/kernel/sse3/kernel_align_x64.S
695-
696694
${PROJECT_SOURCE_DIR}/kernel/avx2/kernel_dgemm_4x4_lib4.S
697695
${PROJECT_SOURCE_DIR}/kernel/avx/kernel_dpack_lib4.S
698696
)
699697

698+
file(GLOB KERNEL_ALIGN_SRC
699+
${PROJECT_SOURCE_DIR}/kernel/sse3/kernel_align_x64.S
700+
)
701+
700702
endif(${TARGET} MATCHES X64_INTEL_SKYLAKE_X)
701703

702704
if(${TARGET} MATCHES X64_INTEL_HASWELL)
@@ -743,11 +745,13 @@ file(GLOB KERNEL_SRC
743745
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_sdot_lib.c
744746
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_saxpy_lib.c
745747

746-
${PROJECT_SOURCE_DIR}/kernel/sse3/kernel_align_x64.S
747-
748748
${PROJECT_SOURCE_DIR}/kernel/avx/kernel_d_aux_lib.S
749749
)
750750

751+
file(GLOB KERNEL_ALIGN_SRC
752+
${PROJECT_SOURCE_DIR}/kernel/sse3/kernel_align_x64.S
753+
)
754+
751755
endif(${TARGET} MATCHES X64_INTEL_HASWELL)
752756
if(${TARGET} MATCHES X64_INTEL_SANDY_BRIDGE)
753757

@@ -791,11 +795,13 @@ file(GLOB KERNEL_SRC
791795
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_sdot_lib.c
792796
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_saxpy_lib.c
793797

794-
${PROJECT_SOURCE_DIR}/kernel/sse3/kernel_align_x64.S
795-
796798
${PROJECT_SOURCE_DIR}/kernel/avx/kernel_d_aux_lib.S
797799
)
798800

801+
file(GLOB KERNEL_ALIGN_SRC
802+
${PROJECT_SOURCE_DIR}/kernel/sse3/kernel_align_x64.S
803+
)
804+
799805
endif(${TARGET} MATCHES X64_INTEL_SANDY_BRIDGE)
800806
if(${TARGET} MATCHES X64_INTEL_CORE)
801807

@@ -829,11 +835,13 @@ file(GLOB KERNEL_SRC
829835
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_sdot_lib.c
830836
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_saxpy_lib.c
831837

832-
${PROJECT_SOURCE_DIR}/kernel/sse3/kernel_align_x64.S
833-
834838
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_d_aux_lib.c
835839
)
836840

841+
file(GLOB KERNEL_ALIGN_SRC
842+
${PROJECT_SOURCE_DIR}/kernel/sse3/kernel_align_x64.S
843+
)
844+
837845
endif(${TARGET} MATCHES X64_INTEL_CORE)
838846
if(${TARGET} MATCHES X64_AMD_BULLDOZER)
839847

@@ -865,11 +873,13 @@ file(GLOB KERNEL_SRC
865873
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_sdot_lib.c
866874
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_saxpy_lib.c
867875

868-
${PROJECT_SOURCE_DIR}/kernel/sse3/kernel_align_x64.S
869-
870876
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_d_aux_lib.c
871877
)
872878

879+
file(GLOB KERNEL_ALIGN_SRC
880+
${PROJECT_SOURCE_DIR}/kernel/sse3/kernel_align_x64.S
881+
)
882+
873883
endif(${TARGET} MATCHES X64_AMD_BULLDOZER)
874884
if(${TARGET} MATCHES ARMV8A_ARM_CORTEX_A57 OR ${TARGET2} MATCHES ARMV8A_ARM_CORTEX_A57)
875885

@@ -914,11 +924,13 @@ file(GLOB KERNEL_SRC
914924
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_sdot_lib.c
915925
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_saxpy_lib.c
916926

917-
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_align_generic.c
918-
919927
${PROJECT_SOURCE_DIR}/kernel/armv8a/kernel_d_aux_lib.S
920928
)
921929

930+
file(GLOB KERNEL_ALIGN_SRC
931+
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_align_generic.c
932+
)
933+
922934
endif(${TARGET} MATCHES ARMV8A_ARM_CORTEX_A57 OR ${TARGET2} MATCHES ARMV8A_ARM_CORTEX_A57)
923935
if(${TARGET} MATCHES ARMV8A_ARM_CORTEX_A53 OR ${TARGET2} MATCHES ARMV8A_ARM_CORTEX_A53)
924936

@@ -965,11 +977,13 @@ file(GLOB KERNEL_SRC
965977
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_sdot_lib.c
966978
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_saxpy_lib.c
967979

968-
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_align_generic.c
969-
970980
${PROJECT_SOURCE_DIR}/kernel/armv8a/kernel_d_aux_lib.S
971981
)
972982

983+
file(GLOB KERNEL_ALIGN_SRC
984+
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_align_generic.c
985+
)
986+
973987
endif(${TARGET} MATCHES ARMV8A_ARM_CORTEX_A53 OR ${TARGET2} MATCHES ARMV8A_ARM_CORTEX_A53)
974988
if(${TARGET} MATCHES ARMV7A_ARM_CORTEX_A15)
975989

@@ -1004,11 +1018,13 @@ file(GLOB KERNEL_SRC
10041018
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_sdot_lib.c
10051019
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_saxpy_lib.c
10061020

1007-
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_align_generic.c
1008-
10091021
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_d_aux_lib.c
10101022
)
10111023

1024+
file(GLOB KERNEL_ALIGN_SRC
1025+
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_align_generic.c
1026+
)
1027+
10121028
endif(${TARGET} MATCHES ARMV7A_ARM_CORTEX_A15)
10131029
if(${TARGET} MATCHES ARMV7A_ARM_CORTEX_A7)
10141030

@@ -1042,11 +1058,13 @@ file(GLOB KERNEL_SRC
10421058
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_sdot_lib.c
10431059
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_saxpy_lib.c
10441060

1045-
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_align_generic.c
1046-
10471061
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_d_aux_lib.c
10481062
)
10491063

1064+
file(GLOB KERNEL_ALIGN_SRC
1065+
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_align_generic.c
1066+
)
1067+
10501068
endif(${TARGET} MATCHES ARMV7A_ARM_CORTEX_A7)
10511069
if(${TARGET} MATCHES ARMV7A_ARM_CORTEX_A9)
10521070

@@ -1080,11 +1098,13 @@ file(GLOB KERNEL_SRC
10801098
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_sdot_lib.c
10811099
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_saxpy_lib.c
10821100

1083-
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_align_generic.c
1084-
10851101
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_d_aux_lib.c
10861102
)
10871103

1104+
file(GLOB KERNEL_ALIGN_SRC
1105+
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_align_generic.c
1106+
)
1107+
10881108
endif(${TARGET} MATCHES ARMV7A_ARM_CORTEX_A9)
10891109
if(${TARGET} MATCHES GENERIC)
10901110

@@ -1115,11 +1135,13 @@ file(GLOB KERNEL_SRC
11151135
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_sdot_lib.c
11161136
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_saxpy_lib.c
11171137

1118-
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_align_generic.c
1119-
11201138
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_d_aux_lib.c
11211139
)
11221140

1141+
file(GLOB KERNEL_ALIGN_SRC
1142+
${PROJECT_SOURCE_DIR}/kernel/generic/kernel_align_generic.c
1143+
)
1144+
11231145
endif(${TARGET} MATCHES GENERIC)
11241146

11251147

@@ -1129,6 +1151,7 @@ set(BLASFEO_SRC ${AUX_COMMON_SRC})
11291151
if(${LA} MATCHES HIGH_PERFORMANCE)
11301152

11311153
list(APPEND BLASFEO_SRC ${KERNEL_SRC})
1154+
list(APPEND BLASFEO_SRC ${KERNEL_ALIGN_SRC})
11321155

11331156
if(${MF} MATCHES PANELMAJ)
11341157

@@ -1168,6 +1191,7 @@ if(${LA} MATCHES HIGH_PERFORMANCE)
11681191
endif()
11691192
if(${LA} MATCHES REFERENCE)
11701193

1194+
list(APPEND BLASFEO_SRC ${KERNEL_ALIGN_SRC})
11711195
list(APPEND BLASFEO_SRC ${AUX_REF_SRC})
11721196
list(APPEND BLASFEO_SRC ${BLASFEO_REF_SRC})
11731197

0 commit comments

Comments
 (0)