diff --git a/cmake/kernel.cmake b/cmake/kernel.cmake index 211da229d..3a4d13837 100644 --- a/cmake/kernel.cmake +++ b/cmake/kernel.cmake @@ -1,110 +1,158 @@ -# helper functions for the kernel CMakeLists.txt - - -# Set the default filenames for L1 objects. Most of these will be overriden by the appropriate KERNEL file. -macro(SetDefaultL1) - set(SAMAXKERNEL amax.S) - set(DAMAXKERNEL amax.S) - set(QAMAXKERNEL amax.S) - set(CAMAXKERNEL zamax.S) - set(ZAMAXKERNEL zamax.S) - set(XAMAXKERNEL zamax.S) - set(SAMINKERNEL amin.S) - set(DAMINKERNEL amin.S) - set(QAMINKERNEL amin.S) - set(CAMINKERNEL zamin.S) - set(ZAMINKERNEL zamin.S) - set(XAMINKERNEL zamin.S) - set(SMAXKERNEL max.S) - set(DMAXKERNEL max.S) - set(QMAXKERNEL max.S) - set(SMINKERNEL min.S) - set(DMINKERNEL min.S) - set(QMINKERNEL min.S) - set(ISAMAXKERNEL iamax.S) - set(IDAMAXKERNEL iamax.S) - set(IQAMAXKERNEL iamax.S) - set(ICAMAXKERNEL izamax.S) - set(IZAMAXKERNEL izamax.S) - set(IXAMAXKERNEL izamax.S) - set(ISAMINKERNEL iamin.S) - set(IDAMINKERNEL iamin.S) - set(IQAMINKERNEL iamin.S) - set(ICAMINKERNEL izamin.S) - set(IZAMINKERNEL izamin.S) - set(IXAMINKERNEL izamin.S) - set(ISMAXKERNEL iamax.S) - set(IDMAXKERNEL iamax.S) - set(IQMAXKERNEL iamax.S) - set(ISMINKERNEL iamin.S) - set(IDMINKERNEL iamin.S) - set(IQMINKERNEL iamin.S) - set(SASUMKERNEL asum.S) - set(DASUMKERNEL asum.S) - set(CASUMKERNEL zasum.S) - set(ZASUMKERNEL zasum.S) - set(QASUMKERNEL asum.S) - set(XASUMKERNEL zasum.S) - set(SAXPYKERNEL axpy.S) - set(DAXPYKERNEL axpy.S) - set(CAXPYKERNEL zaxpy.S) - set(ZAXPYKERNEL zaxpy.S) - set(QAXPYKERNEL axpy.S) - set(XAXPYKERNEL zaxpy.S) - set(SCOPYKERNEL copy.S) - set(DCOPYKERNEL copy.S) - set(CCOPYKERNEL zcopy.S) - set(ZCOPYKERNEL zcopy.S) - set(QCOPYKERNEL copy.S) - set(XCOPYKERNEL zcopy.S) - set(SDOTKERNEL dot.S) - set(DDOTKERNEL dot.S) - set(CDOTKERNEL zdot.S) - set(ZDOTKERNEL zdot.S) - set(QDOTKERNEL dot.S) - set(XDOTKERNEL zdot.S) - set(SNRM2KERNEL nrm2.S) - set(DNRM2KERNEL nrm2.S) - set(QNRM2KERNEL nrm2.S) - set(CNRM2KERNEL znrm2.S) - set(ZNRM2KERNEL znrm2.S) - set(XNRM2KERNEL znrm2.S) - set(SROTKERNEL rot.S) - set(DROTKERNEL rot.S) - set(QROTKERNEL rot.S) - set(CROTKERNEL zrot.S) - set(ZROTKERNEL zrot.S) - set(XROTKERNEL zrot.S) - set(SSCALKERNEL scal.S) - set(DSCALKERNEL scal.S) - set(CSCALKERNEL zscal.S) - set(ZSCALKERNEL zscal.S) - set(QSCALKERNEL scal.S) - set(XSCALKERNEL zscal.S) - set(SSWAPKERNEL swap.S) - set(DSWAPKERNEL swap.S) - set(CSWAPKERNEL zswap.S) - set(ZSWAPKERNEL zswap.S) - set(QSWAPKERNEL swap.S) - set(XSWAPKERNEL zswap.S) - set(SGEMVNKERNEL gemv_n.S) - set(SGEMVTKERNEL gemv_t.S) - set(DGEMVNKERNEL gemv_n.S) - set(DGEMVTKERNEL gemv_t.S) - set(CGEMVNKERNEL zgemv_n.S) - set(CGEMVTKERNEL zgemv_t.S) - set(ZGEMVNKERNEL zgemv_n.S) - set(ZGEMVTKERNEL zgemv_t.S) - set(QGEMVNKERNEL gemv_n.S) - set(QGEMVTKERNEL gemv_t.S) - set(XGEMVNKERNEL zgemv_n.S) - set(XGEMVTKERNEL zgemv_t.S) - set(SCABS_KERNEL cabs.S) - set(DCABS_KERNEL cabs.S) - set(QCABS_KERNEL cabs.S) - set(LSAME_KERNEL lsame.S) - set(SAXPBYKERNEL ../arm/axpby.c) - set(DAXPBYKERNEL ../arm/axpby.c) - set(CAXPBYKERNEL ../arm/zaxpby.c) - set(ZAXPBYKERNEL ../arm/zaxpby.c) -endmacro () +# helper functions for the kernel CMakeLists.txt + + +# Set the default filenames for L1 objects. Most of these will be overriden by the appropriate KERNEL file. +macro(SetDefaultL1) + set(SAMAXKERNEL amax.S) + set(DAMAXKERNEL amax.S) + set(QAMAXKERNEL amax.S) + set(CAMAXKERNEL zamax.S) + set(ZAMAXKERNEL zamax.S) + set(XAMAXKERNEL zamax.S) + set(SAMINKERNEL amin.S) + set(DAMINKERNEL amin.S) + set(QAMINKERNEL amin.S) + set(CAMINKERNEL zamin.S) + set(ZAMINKERNEL zamin.S) + set(XAMINKERNEL zamin.S) + set(SMAXKERNEL max.S) + set(DMAXKERNEL max.S) + set(QMAXKERNEL max.S) + set(SMINKERNEL min.S) + set(DMINKERNEL min.S) + set(QMINKERNEL min.S) + set(ISAMAXKERNEL iamax.S) + set(IDAMAXKERNEL iamax.S) + set(IQAMAXKERNEL iamax.S) + set(ICAMAXKERNEL izamax.S) + set(IZAMAXKERNEL izamax.S) + set(IXAMAXKERNEL izamax.S) + set(ISAMINKERNEL iamin.S) + set(IDAMINKERNEL iamin.S) + set(IQAMINKERNEL iamin.S) + set(ICAMINKERNEL izamin.S) + set(IZAMINKERNEL izamin.S) + set(IXAMINKERNEL izamin.S) + set(ISMAXKERNEL iamax.S) + set(IDMAXKERNEL iamax.S) + set(IQMAXKERNEL iamax.S) + set(ISMINKERNEL iamin.S) + set(IDMINKERNEL iamin.S) + set(IQMINKERNEL iamin.S) + set(SASUMKERNEL asum.S) + set(DASUMKERNEL asum.S) + set(CASUMKERNEL zasum.S) + set(ZASUMKERNEL zasum.S) + set(QASUMKERNEL asum.S) + set(XASUMKERNEL zasum.S) + set(SAXPYKERNEL axpy.S) + set(DAXPYKERNEL axpy.S) + set(CAXPYKERNEL zaxpy.S) + set(ZAXPYKERNEL zaxpy.S) + set(QAXPYKERNEL axpy.S) + set(XAXPYKERNEL zaxpy.S) + set(SCOPYKERNEL copy.S) + set(DCOPYKERNEL copy.S) + set(CCOPYKERNEL zcopy.S) + set(ZCOPYKERNEL zcopy.S) + set(QCOPYKERNEL copy.S) + set(XCOPYKERNEL zcopy.S) + set(SDOTKERNEL dot.S) + set(DDOTKERNEL dot.S) + set(CDOTKERNEL zdot.S) + set(ZDOTKERNEL zdot.S) + set(QDOTKERNEL dot.S) + set(XDOTKERNEL zdot.S) + set(SNRM2KERNEL nrm2.S) + set(DNRM2KERNEL nrm2.S) + set(QNRM2KERNEL nrm2.S) + set(CNRM2KERNEL znrm2.S) + set(ZNRM2KERNEL znrm2.S) + set(XNRM2KERNEL znrm2.S) + set(SROTKERNEL rot.S) + set(DROTKERNEL rot.S) + set(QROTKERNEL rot.S) + set(CROTKERNEL zrot.S) + set(ZROTKERNEL zrot.S) + set(XROTKERNEL zrot.S) + set(SSCALKERNEL scal.S) + set(DSCALKERNEL scal.S) + set(CSCALKERNEL zscal.S) + set(ZSCALKERNEL zscal.S) + set(QSCALKERNEL scal.S) + set(XSCALKERNEL zscal.S) + set(SSWAPKERNEL swap.S) + set(DSWAPKERNEL swap.S) + set(CSWAPKERNEL zswap.S) + set(ZSWAPKERNEL zswap.S) + set(QSWAPKERNEL swap.S) + set(XSWAPKERNEL zswap.S) + set(SGEMVNKERNEL gemv_n.S) + set(SGEMVTKERNEL gemv_t.S) + set(DGEMVNKERNEL gemv_n.S) + set(DGEMVTKERNEL gemv_t.S) + set(CGEMVNKERNEL zgemv_n.S) + set(CGEMVTKERNEL zgemv_t.S) + set(ZGEMVNKERNEL zgemv_n.S) + set(ZGEMVTKERNEL zgemv_t.S) + set(QGEMVNKERNEL gemv_n.S) + set(QGEMVTKERNEL gemv_t.S) + set(XGEMVNKERNEL zgemv_n.S) + set(XGEMVTKERNEL zgemv_t.S) + set(SCABS_KERNEL cabs.S) + set(DCABS_KERNEL cabs.S) + set(QCABS_KERNEL cabs.S) + set(LSAME_KERNEL lsame.S) + set(SAXPBYKERNEL ../arm/axpby.c) + set(DAXPBYKERNEL ../arm/axpby.c) + set(CAXPBYKERNEL ../arm/zaxpby.c) + set(ZAXPBYKERNEL ../arm/zaxpby.c) +endmacro () + +macro(SetDefaultL2) + set(SGEMVNKERNEL gemv_n.S) + set(SGEMVTKERNEL gemv_t.S) + set(DGEMVNKERNEL gemv_n.S) + set(DGEMVTKERNEL gemv_t.S) + set(CGEMVNKERNEL zgemv_n.S) + set(CGEMVTKERNEL zgemv_t.S) + set(ZGEMVNKERNEL zgemv_n.S) + set(ZGEMVTKERNEL zgemv_t.S) + set(QGEMVNKERNEL gemv_n.S) + set(QGEMVTKERNEL gemv_t.S) + set(XGEMVNKERNEL zgemv_n.S) + set(XGEMVTKERNEL zgemv_t.S) + set(SGERKERNEL ../generic/ger.c) + set(DGERKERNEL ../generic/ger.c) + set(QGERKERNEL ../generic/ger.c) + set(CGERUKERNEL ../generic/zger.c) + set(CGERCKERNEL ../generic/zger.c) + set(ZGERUKERNEL ../generic/zger.c) + set(ZGERCKERNEL ../generic/zger.c) + set(XGERUKERNEL ../generic/zger.c) + set(XGERCKERNEL ../generic/zger.c) + set(SSYMV_U_KERNEL ../generic/symv_k.c) + set(SSYMV_L_KERNEL ../generic/symv_k.c) + set(DSYMV_U_KERNEL ../generic/symv_k.c) + set(DSYMV_L_KERNEL ../generic/symv_k.c) + set(QSYMV_U_KERNEL ../generic/symv_k.c) + set(QSYMV_L_KERNEL ../generic/symv_k.c) + set(CSYMV_U_KERNEL ../generic/zsymv_k.c) + set(CSYMV_L_KERNEL ../generic/zsymv_k.c) + set(ZSYMV_U_KERNEL ../generic/zsymv_k.c) + set(ZSYMV_L_KERNEL ../generic/zsymv_k.c) + set(XSYMV_U_KERNEL ../generic/zsymv_k.c) + set(XSYMV_L_KERNEL ../generic/zsymv_k.c) + set(CHEMV_U_KERNEL ../generic/zhemv_k.c) + set(CHEMV_L_KERNEL ../generic/zhemv_k.c) + set(CHEMV_V_KERNEL ../generic/zhemv_k.c) + set(CHEMV_M_KERNEL ../generic/zhemv_k.c) + set(ZHEMV_U_KERNEL ../generic/zhemv_k.c) + set(ZHEMV_L_KERNEL ../generic/zhemv_k.c) + set(ZHEMV_V_KERNEL ../generic/zhemv_k.c) + set(ZHEMV_M_KERNEL ../generic/zhemv_k.c) + set(XHEMV_U_KERNEL ../generic/zhemv_k.c) + set(XHEMV_L_KERNEL ../generic/zhemv_k.c) + set(XHEMV_V_KERNEL ../generic/zhemv_k.c) + set(XHEMV_M_KERNEL ../generic/zhemv_k.c) +endmacro () diff --git a/driver/level2/CMakeLists.txt b/driver/level2/CMakeLists.txt index 8b37917a6..54e0eb42f 100644 --- a/driver/level2/CMakeLists.txt +++ b/driver/level2/CMakeLists.txt @@ -54,14 +54,6 @@ endforeach () if (SMP) - # gbmv uses a lowercase n and t. N.B. this uses TRANSA where gbmv.c uses TRANS. Intentional? - GenerateNamedObjects("gbmv_thread.c" "" "gbmv_thread_n") - GenerateNamedObjects("gbmv_thread.c" "TRANSA" "gbmv_thread_t") - - GenerateNamedObjects("gemv_thread.c" "" "gemv_thread_n") - GenerateNamedObjects("gemv_thread.c" "TRANSA" "gemv_thread_t") - - GenerateNamedObjects("ger_thread.c") set(UL_SMP_SOURCES symv_thread.c @@ -84,10 +76,41 @@ if (SMP) GenerateCombinationObjects("${NU_SMP_SOURCES}" "TRANSA;LOWER;UNIT" "N;U;N" "" 2) foreach (float_type ${FLOAT_TYPES}) + + GenerateNamedObjects("gemv_thread.c" "" "gemv_thread_n" false "" "" false ${float_type}) + GenerateNamedObjects("gemv_thread.c" "TRANSA" "gemv_thread_t" false "" "" false ${float_type}) + + GenerateNamedObjects("gbmv_thread.c" "" "gbmv_thread_n" false "" "" false ${float_type}) + GenerateNamedObjects("gbmv_thread.c" "TRANSA" "gbmv_thread_t" false "" "" false ${float_type}) + + if (${float_type} STREQUAL "COMPLEX" OR ${float_type} STREQUAL "ZCOMPLEX") GenerateCombinationObjects("trmv_thread.c" "LOWER;UNIT" "U;N" "TRANSA=3" 0 "trmv_thread_R" false ${float_type}) GenerateCombinationObjects("trmv_thread.c" "LOWER;UNIT" "U;N" "TRANSA=4" 0 "trmv_thread_C" false ${float_type}) + + GenerateNamedObjects("gemv_thread.c" "CONJ" "gemv_thread_r" false "" "" false ${float_type}) + GenerateNamedObjects("gemv_thread.c" "CONJ;TRANSA" "gemv_thread_c" false "" "" false ${float_type}) + GenerateNamedObjects("gemv_thread.c" "XCONJ" "gemv_thread_o" false "" "" false ${float_type}) + GenerateNamedObjects("gemv_thread.c" "XCONJ;TRANSA" "gemv_thread_u" false "" "" false ${float_type}) + GenerateNamedObjects("gemv_thread.c" "XCONJ;CONJ" "gemv_thread_s" false "" "" false ${float_type}) + GenerateNamedObjects("gemv_thread.c" "XCONJ;CONJ;TRANSA" "gemv_thread_d" false "" "" false ${float_type}) + + GenerateNamedObjects("gbmv_thread.c" "CONJ" "gbmv_thread_r" false "" "" false ${float_type}) + GenerateNamedObjects("gbmv_thread.c" "CONJ;TRANSA" "gbmv_thread_c" false "" "" false ${float_type}) + GenerateNamedObjects("gbmv_thread.c" "XCONJ" "gbmv_thread_o" false "" "" false ${float_type}) + GenerateNamedObjects("gbmv_thread.c" "XCONJ;TRANSA" "gbmv_thread_u" false "" "" false ${float_type}) + GenerateNamedObjects("gbmv_thread.c" "XCONJ;CONJ" "gbmv_thread_s" false "" "" false ${float_type}) + GenerateNamedObjects("gbmv_thread.c" "XCONJ;CONJ;TRANSA" "gbmv_thread_d" false "" "" false ${float_type}) + + GenerateNamedObjects("ger_thread.c" "" "ger_thread_U" false "" "" false ${float_type}) + GenerateNamedObjects("ger_thread.c" "CONJ" "ger_thread_C" false "" "" false ${float_type}) + GenerateNamedObjects("ger_thread.c" "XCONJ" "ger_thread_V" false "" "" false ${float_type}) + GenerateNamedObjects("ger_thread.c" "XCONJ;CONJ" "ger_thread_D" false "" "" false ${float_type}) + + else () + GenerateNamedObjects("ger_thread.c" "" "" false "" "" false ${float_type}) endif () + endforeach () endif () diff --git a/kernel/CMakeLists.txt b/kernel/CMakeLists.txt index 656090cf8..cacd0f38f 100644 --- a/kernel/CMakeLists.txt +++ b/kernel/CMakeLists.txt @@ -67,8 +67,7 @@ foreach (float_type ${FLOAT_TYPES}) endforeach () # Makefile.L2 -GenerateNamedObjects("${KERNELDIR}/gemv_n.S" "DOUBLE") -GenerateNamedObjects("${KERNELDIR}/gemv_t.S" "TRANS") +SetDefaultL2() GenerateCombinationObjects("generic/symv_k.c" "LOWER" "U" "" 1 "" "" 3) GenerateNamedObjects("generic/ger.c" "" "ger_k" false "" "" "" 3) foreach (float_type ${FLOAT_TYPES}) @@ -78,6 +77,18 @@ foreach (float_type ${FLOAT_TYPES}) GenerateNamedObjects("${KERNELDIR}/${${float_char}GERCKERNEL}" "CONJ" "gerc_k" false "" "" false ${float_type}) GenerateNamedObjects("${KERNELDIR}/${${float_char}GERUKERNEL}" "XCONJ" "gerv_k" false "" "" false ${float_type}) GenerateNamedObjects("${KERNELDIR}/${${float_char}GERCKERNEL}" "CONJ;XCONJ" "gerd_k" false "" "" false ${float_type}) + + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVNKERNEL}" "" "gemv_n" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVTKERNEL}" "TRANSA" "gemv_t" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVNKERNEL}" "CONJ" "gemv_r" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVTKERNEL}" "CONJ;TRANSA" "${KERNELDIR}/${${float_char}GEMVUKERNEL}" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVNKERNEL}" "XCONJ" "gemv_o" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVTKERNEL}" "XCONJ;TRANSA" "gemv_u" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVNKERNEL}" "XCONJ;CONJ" "gemv_s" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVTKERNEL}" "XCONJ;CONJ;TRANSA" "gemv_d" false "" "" false ${float_type}) + else () + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVNKERNEL}" "" "gemv_n" false "" "" false ${float_type}) + GenerateNamedObjects("${KERNELDIR}/${${float_char}GEMVTKERNEL}" "TRANS" "gemv_t" false "" "" false ${float_type}) endif () endforeach ()