From 3c451b99639a984126c6a618d4394ee582e872a2 Mon Sep 17 00:00:00 2001 From: Alexis Montoison <35051714+amontoison@users.noreply.github.com> Date: Wed, 1 May 2024 17:21:00 -0400 Subject: [PATCH] Upgrade gensymbol (#125) * Upgrade gensymbol * Update CI scripts --- .buildkite/pipeline.yml | 2 +- .cirrus.yml | 2 +- ext/gensymbol/gensymbol | 204 ++++++++++++++++++++++++++++++++++++---- src/exported_funcs.inc | 192 +++++++++++++++++++++++++++++++------ 4 files changed, 351 insertions(+), 49 deletions(-) mode change 100755 => 100644 ext/gensymbol/gensymbol diff --git a/.buildkite/pipeline.yml b/.buildkite/pipeline.yml index e94f3bb..f84eb80 100644 --- a/.buildkite/pipeline.yml +++ b/.buildkite/pipeline.yml @@ -9,7 +9,7 @@ steps: } # We'll test on the following julia versions - for JULIA_VERSION in 1.8 nightly; do + for JULIA_VERSION in 1.10 nightly; do # First, linux OS="linux" diff --git a/.cirrus.yml b/.cirrus.yml index 0d7b745..7559356 100644 --- a/.cirrus.yml +++ b/.cirrus.yml @@ -2,7 +2,7 @@ task: matrix: - name: FreeBSD freebsd_instance: - image: freebsd-13-1-release-amd64 + image: freebsd-13-2-release-amd64 env: matrix: - JULIA_VERSION: 1 diff --git a/ext/gensymbol/gensymbol b/ext/gensymbol/gensymbol old mode 100755 new mode 100644 index 551382f..1bbf2ab --- a/ext/gensymbol/gensymbol +++ b/ext/gensymbol/gensymbol @@ -17,38 +17,38 @@ # removed blas_thread_shutdown_ # @blasobjsc = ( - caxpy,caxpby,ccopy,cdotc,cdotu,cgbmv,cgemm,cgemmt,cgemv,cgerc,cgeru, + caxpy,caxpby,ccopy,cdotc,cdotu,cgbmv,cgemm,cgemv,cgerc,cgeru, chbmv,chemm,chemv,cher2,cher2k,cher,cherk,scabs1,scamax, chpmv,chpr2,chpr,crotg,cscal,csrot,csscal,cswap,scamin,scasum,scnrm2, csymm,csyr2k,csyrk,ctbmv,ctbsv,ctpmv,ctpsv,ctrmm,ctrmv,ctrsm, - ctrsv,icamax,icamin,cimatcopy,comatcopy,cgeadd,scsum); + ctrsv,icamax,icamin,cimatcopy,comatcopy,cgeadd,scsum,cgemmt); @blasobjsd = ( - damax,damin,dasum,daxpy,daxpby,dcabs1,dcopy,ddot,dgbmv,dgemm,dgemmt, + damax,damin,dasum,daxpy,daxpby,dcabs1,dcopy,ddot,dgbmv,dgemm, dgemv,dger,dmax,dmin,dnrm2,drot,drotg,drotm,drotmg,dsbmv, dscal,dsdot,dspmv,dspr2,dimatcopy,domatcopy, dspr,dswap,dsymm,dsymv,dsyr2,dsyr2k,dsyr,dsyrk,dtbmv,dtbsv, dtpmv,dtpsv,dtrmm,dtrmv,dtrsm,dtrsv, - idamax,idamin,idmax,idmin,dgeadd,dsum); + idamax,idamin,idmax,idmin,dgeadd,dsum,dgemmt); @blasobjss = ( isamax,isamin,ismax,ismin, samax,samin,sasum,saxpy, saxpby, - scopy,sdot,sdsdot,sgbmv,sgemm,sgemmt,sgemv,sger, + scopy,sdot,sdsdot,sgbmv,sgemm,sgemv,sger, smax,smin,snrm2,simatcopy,somatcopy, srot,srotg,srotm,srotmg,ssbmv,sscal,sspmv,sspr2,sspr,sswap, ssymm,ssymv,ssyr2,ssyr2k,ssyr,ssyrk,stbmv,stbsv,stpmv,stpsv, - strmm,strmv,strsm,strsv, sgeadd,ssum); + strmm,strmv,strsm,strsv, sgeadd,ssum,sgemmt); @blasobjsz = ( izamax,izamin,, zaxpy,zaxpby,zcopy,zdotc,zdotu,zdrot, - zdscal,zgbmv,zgemm,zgemmt,zgemv,zgerc,zgeru, + zdscal,zgbmv,zgemm,zgemv,zgerc,zgeru, zhbmv,zhemm,zhemv,zher2,zher2k,zher,zherk,zhpmv,zhpr2, zhpr,zrotg,zscal,zswap,zsymm,zsyr2k,zsyrk,ztbmv, ztbsv,ztpmv,ztpsv,ztrmm,ztrmv,ztrsm,ztrsv, - zomatcopy, zimatcopy,dzamax,dzamin,dzasum,dznrm2, - zgeadd, dzsum); + zomatcopy,zimatcopy,dzamax,dzamin,dzasum,dznrm2, + zgeadd,dzsum,zgemmt); @blasobjs = (lsame, xerbla); @bfblasobjs = (sbgemm, sbgemv, sbdot, sbstobf16, sbdtobf16, sbf16tos, dbf16tod); @@ -59,8 +59,8 @@ cblas_csscal, cblas_cswap, cblas_csymm, cblas_csyr2k, cblas_csyrk, cblas_ctbmv, cblas_cgeadd, cblas_ctbsv, cblas_ctpmv, cblas_ctpsv, cblas_ctrmm, cblas_ctrmv, cblas_ctrsm, cblas_ctrsv, cblas_scnrm2, cblas_scasum, - cblas_icamax, cblas_icamin, cblas_icmin, cblas_icmax, cblas_scsum,cblas_cimatcopy,cblas_comatcopy - ); + cblas_icamax, cblas_icamin, cblas_icmin, cblas_icmax, cblas_scsum,cblas_cimatcopy,cblas_comatcopy, + cblas_cgemmt, cblas_caxpyc, cblas_crotg, cblas_csrot, cblas_scamax, cblas_scamin); @cblasobjsd = ( cblas_dasum, cblas_daxpy, cblas_dcopy, cblas_ddot, cblas_dgbmv, cblas_dgemm, cblas_dgemv, cblas_dger, cblas_dnrm2, @@ -68,8 +68,8 @@ cblas_dspmv, cblas_dspr2, cblas_dspr, cblas_dswap, cblas_dsymm, cblas_dsymv, cblas_dsyr2, cblas_dsyr2k, cblas_dsyr, cblas_dsyrk, cblas_dtbmv, cblas_dtbsv, cblas_dtpmv, cblas_dtpsv, cblas_dtrmm, cblas_dtrmv, cblas_dtrsm, cblas_dtrsv, cblas_daxpby, cblas_dgeadd, - cblas_idamax, cblas_idamin, cblas_idmin, cblas_idmax, cblas_dsum,cblas_dimatcopy,cblas_domatcopy - ); + cblas_idamax, cblas_idamin, cblas_idmin, cblas_idmax, cblas_dsum,cblas_dimatcopy,cblas_domatcopy, + cblas_dgemmt, cblas_damax, cblas_damin); @cblasobjss = ( cblas_sasum, cblas_saxpy, cblas_saxpby, @@ -79,8 +79,8 @@ cblas_sswap, cblas_ssymm, cblas_ssymv, cblas_ssyr2, cblas_ssyr2k, cblas_ssyr, cblas_ssyrk, cblas_stbmv, cblas_stbsv, cblas_stpmv, cblas_stpsv, cblas_strmm, cblas_strmv, cblas_strsm, cblas_strsv, cblas_sgeadd, - cblas_isamax, cblas_isamin, cblas_ismin, cblas_ismax, cblas_ssum,cblas_simatcopy,cblas_somatcopy - ); + cblas_isamax, cblas_isamin, cblas_ismin, cblas_ismax, cblas_ssum,cblas_simatcopy,cblas_somatcopy, + cblas_sgemmt, cblas_samax, cblas_samin); @cblasobjsz = ( cblas_dzasum, cblas_dznrm2, cblas_zaxpy, cblas_zcopy, cblas_zdotc, cblas_zdotu, cblas_zdscal, cblas_zgbmv, cblas_zgemm, cblas_zgemv, cblas_zgerc, cblas_zgeru, cblas_zhbmv, cblas_zhemm, @@ -89,8 +89,8 @@ cblas_ztbmv, cblas_ztbsv, cblas_ztpmv, cblas_ztpsv, cblas_ztrmm, cblas_ztrmv, cblas_ztrsm, cblas_ztrsv, cblas_cdotc_sub, cblas_cdotu_sub, cblas_zdotc_sub, cblas_zdotu_sub, cblas_zaxpby, cblas_zgeadd, - cblas_izamax, cblas_izamin, cblas_izmin, cblas_izmax, cblas_dzsum,cblas_zimatcopy,cblas_zomatcopy -); + cblas_izamax, cblas_izamin, cblas_izmin, cblas_izmax, cblas_dzsum,cblas_zimatcopy,cblas_zomatcopy, + cblas_zgemmt, cblas_dzamax, cblas_dzamin); @cblasobjs = ( cblas_xerbla ); @@ -709,6 +709,7 @@ zpotri, # functions added for lapack-3.7.0 @lapackobjs2s = (@lapackobjs2s, slarfy, + ssyconvf, strevc3, sgelqt, sgelqt3, @@ -813,7 +814,86 @@ zpotri, @lapackobjs2z = (@lapackobjs2z, zgesvdq, zungtsqr -); + ); + #functions added for lapack-3.10 +@lapackobjs2c = (@lapackobjs2c, + cgetsqrhrt, + cungtsqr_row + ); +@lapackobjs2d = (@lapackobjs2d, + dgetsqrhrt, + dorgtsqr_row + ); +@lapackobjs2s = (@lapackobjs2s, + sgetsqrhrt, + sorgtsqr_row + ); +@lapackobjs2z = (@lapackobjs2z, + zgetsqrhrt, + zungtsqr_row + ); + #functions added for lapack-3.11 +@lapackobjs2c = (@lapackobjs2c, + cgedmd, + cgedmdq + ); +@lapackobjs2d = (@lapackobjs2d, + dgedmd, + dgedmdq + ); +@lapackobjs2s = (@lapackobjs2s, + sgedmd, + sgedmdq + ); +@lapackobjs2z = (@lapackobjs2z, + zgedmd, + zgedmdq + ); + #functions added for lapack-3.12 +@lapackobjs2c = (@lapackobjs2c, + cgelqs, + cgelst, + cgeqp3rk, + cgeqrs, + claqp2rk, + claqp3rk, + clatrs3, + ctrsyl3, + crscl, + ); +@lapackobjs2d = (@lapackobjs2d, + dgelqs, + dgelst, + dgeqp3rk, + dgeqrs, + dlaqp2rk, + dlaqp3rk, + dlatrs3, + dtrsyl3, + dlarmm, + ); +@lapackobjs2s = (@lapackobjs2s, + sgelqs, + sgelst, + sgeqp3rk, + sgeqrs, + slaqp2rk, + slaqp3rk, + slatrs3, + strsyl3, + slarmm, + ); +@lapackobjs2z = (@lapackobjs2z, + zgelqs, + zgelst, + zgeqp3rk, + zgeqrs, + zlaqp2rk, + zlaqp3rk, + zlatrs3, + ztrsyl3, + zrscl, + ); @lapack_extendedprecision_objs = ( zposvxx, clagge, clatms, chesvxx, cposvxx, cgesvxx, ssyrfssx, csyrfsx, @@ -841,7 +921,8 @@ zpotri, dggsvp, dlahrd, dlatzm, - dtzrqf); + dtzrqf, + ); @lapack_deprecated_objss = ( sgelsx, @@ -989,6 +1070,10 @@ zpotri, LAPACKE_cgebrd_work, LAPACKE_cgecon, LAPACKE_cgecon_work, + LAPACKE_cgedmd, + LAPACKE_cgedmd_work, + LAPACKE_cgedmdq, + LAPACKE_cgedmdq_work, LAPACKE_cgeequ, LAPACKE_cgeequ_work, LAPACKE_cgeequb, @@ -1571,6 +1656,20 @@ zpotri, LAPACKE_csytrf_aa_2stage_work, LAPACKE_csytrs_aa_2stage, LAPACKE_csytrs_aa_2stage_work, + LAPACKE_cgesvdq, + LAPACKE_cgesvdq_work, + LAPACKE_cgetsqrhrt, + LAPACKE_cgetsqrhrt_work, + LAPACKE_cungtsqr_row, + LAPACKE_cungtsqr_row_work, + LAPACKE_clangb, + LAPACKE_clangb_work, + LAPACKE_ctrsyl3, + LAPACKE_ctrsyl3_work, + LAPACKE_ctz_nancheck, + LAPACKE_ctz_trans, + LAPACKE_cunhr_col, + LAPACKE_cunhr_col_work, ); @lapackeobjsd = ( LAPACKE_dgb_nancheck, @@ -1641,6 +1740,10 @@ zpotri, LAPACKE_dgebrd_work, LAPACKE_dgecon, LAPACKE_dgecon_work, + LAPACKE_dgedmd, + LAPACKE_dgedmd_work, + LAPACKE_dgedmdq, + LAPACKE_dgedmdq_work, LAPACKE_dgeequ, LAPACKE_dgeequ_work, LAPACKE_dgeequb, @@ -2179,6 +2282,18 @@ zpotri, LAPACKE_dgesvdq_work, LAPACKE_slag2d, LAPACKE_slag2d_work, + LAPACKE_dgetsqrhrt, + LAPACKE_dgetsqrhrt_work, + LAPACKE_dorgtsqr_row, + LAPACKE_dorgtsqr_row_work, + LAPACKE_dlangb, + LAPACKE_dlangb_work, + LAPACKE_dorhr_col, + LAPACKE_dorhr_col_work, + LAPACKE_dtrsyl3, + LAPACKE_dtrsyl3_work, + LAPACKE_dtz_nancheck, + LAPACKE_dtz_trans, ); @lapackeobjss = ( LAPACKE_sgb_nancheck, @@ -2249,6 +2364,10 @@ zpotri, LAPACKE_sgebrd_work, LAPACKE_sgecon, LAPACKE_sgecon_work, + LAPACKE_sgedmd, + LAPACKE_sgedmd_work, + LAPACKE_sgedmdq, + LAPACKE_sgedmdq_work, LAPACKE_sgeequ, LAPACKE_sgeequ_work, LAPACKE_sgeequb, @@ -2779,6 +2898,18 @@ zpotri, LAPACKE_ssytrs_aa_2stage_work, LAPACKE_sgesvdq, LAPACKE_sgesvdq_work, + LAPACKE_sgetsqrhrt, + LAPACKE_sgetsqrhrt_work, + LAPACKE_sorgtsqr_row, + LAPACKE_sorgtsqr_row_work, + LAPACKE_slangb, + LAPACKE_slangb_work, + LAPACKE_sorhr_col, + LAPACKE_sorhr_col_work, + LAPACKE_strsyl3, + LAPACKE_strsyl3_work, + LAPACKE_stz_nancheck, + LAPACKE_stz_trans, ); @lapackeobjsz = ( LAPACKE_zgb_nancheck, @@ -2853,6 +2984,10 @@ zpotri, LAPACKE_zgebrd_work, LAPACKE_zgecon, LAPACKE_zgecon_work, + LAPACKE_zgedmd, + LAPACKE_zgedmd_work, + LAPACKE_zgedmdq, + LAPACKE_zgedmdq_work, LAPACKE_zgeequ, LAPACKE_zgeequ_work, LAPACKE_zgeequb, @@ -3317,6 +3452,19 @@ zpotri, LAPACKE_zupmtr_work, LAPACKE_zsyr, LAPACKE_zsyr_work, + LAPACKE_zgetsqrhrt, + LAPACKE_zgetsqrhrt_work, + LAPACKE_zungtsqr_row, + LAPACKE_zungtsqr_row_work, + LAPACKE_zlangb, + LAPACKE_zlangb_work, + LAPACKE_ztrsyl3, + LAPACKE_ztrsyl3_work, + LAPACKE_ztz_nancheck, + LAPACKE_ztz_trans, + LAPACKE_zunhr_col, + LAPACKE_zunhr_col_work, + ## @(SRCX_OBJ) from `lapack-3.4.1/lapacke/src/Makefile` ## Not exported: requires LAPACKE_EXTENDED to be set and depends on the ## corresponding LAPACK extended precision routines. @@ -3544,6 +3692,7 @@ zpotri, ssysv_aa_2stage, ssytrf_aa_2stage, ssytrs_aa_2stage, slaorhr_col_getrfnp, slaorhr_col_getrfnp2, sorhr_col, + slarfb_gett, ); @lapack_embeded_underscore_objs_c=( chetf2_rook, chetrf_rook, chetri_rook, @@ -3569,6 +3718,7 @@ zpotri, csysv_aa_2stage, csytrf_aa_2stage, csytrs_aa_2stage, claunhr_col_getrfnp, claunhr_col_getrfnp2, cunhr_col, + clarfb_gett, ); @lapack_embeded_underscore_objs_d=( dlasyf_rook, @@ -3586,6 +3736,7 @@ zpotri, dsysv_aa_2stage, dsytrf_aa_2stage, dsytrs_aa_2stage, dlaorhr_col_getrfnp, dlaorhr_col_getrfnp2, dorhr_col, + dlarfb_gett, ); @lapack_embeded_underscore_objs_z=( zhetf2_rook, zhetrf_rook, zhetri_rook, @@ -3609,7 +3760,8 @@ zpotri, zhesv_aa_2stage, zhetrf_aa_2stage, zhetrs_aa_2stage, zsysv_aa_2stage, zsytrf_aa_2stage, zsytrs_aa_2stage, - zlaunhr_col_getrfnp, zlaunhr_col_getrfnp2, zunhr_col + zlaunhr_col_getrfnp, zlaunhr_col_getrfnp2, zunhr_col, + zlarfb_gett, ); @@ -3930,6 +4082,18 @@ if ($ARGV[0] eq "linktest") { @no_underscore_objs = (@no_underscore_objs, @misc_common_objs); print "int main(void){\n"; + foreach $objs (@underscore_objs) { + print "extern void ", $symbolprefix, $objs, $bu, $symbolsuffix, "();\n" if $objs ne "xerbla"; + } + + foreach $objs (@need_2underscore_objs) { + print "extern void ", $symbolprefix, $objs, $bu, $bu, $symbolsuffix, "();\n"; + } + + foreach $objs (@no_underscore_objs) { + print "extern void ", $symbolprefix, $objs, $symbolsuffix, "();\n"; + } + foreach $objs (@underscore_objs) { print $symbolprefix, $objs, $bu, $symbolsuffix, "();\n" if $objs ne "xerbla"; } diff --git a/src/exported_funcs.inc b/src/exported_funcs.inc index 35b499e..453efc4 100644 --- a/src/exported_funcs.inc +++ b/src/exported_funcs.inc @@ -37,6 +37,10 @@ XX(LAPACKE_cgebrd_work) \ XX(LAPACKE_cgecon) \ XX(LAPACKE_cgecon_work) \ + XX(LAPACKE_cgedmd) \ + XX(LAPACKE_cgedmd_work) \ + XX(LAPACKE_cgedmdq) \ + XX(LAPACKE_cgedmdq_work) \ XX(LAPACKE_cgeequ) \ XX(LAPACKE_cgeequ_work) \ XX(LAPACKE_cgeequb) \ @@ -103,6 +107,8 @@ XX(LAPACKE_cgesv_work) \ XX(LAPACKE_cgesvd) \ XX(LAPACKE_cgesvd_work) \ + XX(LAPACKE_cgesvdq) \ + XX(LAPACKE_cgesvdq_work) \ XX(LAPACKE_cgesvdx) \ XX(LAPACKE_cgesvdx_work) \ XX(LAPACKE_cgesvj) \ @@ -121,6 +127,8 @@ XX(LAPACKE_cgetrs_work) \ XX(LAPACKE_cgetsls) \ XX(LAPACKE_cgetsls_work) \ + XX(LAPACKE_cgetsqrhrt) \ + XX(LAPACKE_cgetsqrhrt_work) \ XX(LAPACKE_cgg_nancheck) \ XX(LAPACKE_cgg_trans) \ XX(LAPACKE_cggbak) \ @@ -336,6 +344,8 @@ XX(LAPACKE_claghe_work) \ XX(LAPACKE_clagsy) \ XX(LAPACKE_clagsy_work) \ + XX(LAPACKE_clangb) \ + XX(LAPACKE_clangb_work) \ XX(LAPACKE_clange) \ XX(LAPACKE_clange_work) \ XX(LAPACKE_clanhe) \ @@ -607,6 +617,8 @@ XX(LAPACKE_ctrsna) \ XX(LAPACKE_ctrsna_work) \ XX(LAPACKE_ctrsyl) \ + XX(LAPACKE_ctrsyl3) \ + XX(LAPACKE_ctrsyl3_work) \ XX(LAPACKE_ctrsyl_work) \ XX(LAPACKE_ctrtri) \ XX(LAPACKE_ctrtri_work) \ @@ -616,6 +628,8 @@ XX(LAPACKE_ctrttf_work) \ XX(LAPACKE_ctrttp) \ XX(LAPACKE_ctrttp_work) \ + XX(LAPACKE_ctz_nancheck) \ + XX(LAPACKE_ctz_trans) \ XX(LAPACKE_ctzrzf) \ XX(LAPACKE_ctzrzf_work) \ XX(LAPACKE_cunbdb) \ @@ -638,6 +652,10 @@ XX(LAPACKE_cungrq_work) \ XX(LAPACKE_cungtr) \ XX(LAPACKE_cungtr_work) \ + XX(LAPACKE_cungtsqr_row) \ + XX(LAPACKE_cungtsqr_row_work) \ + XX(LAPACKE_cunhr_col) \ + XX(LAPACKE_cunhr_col_work) \ XX(LAPACKE_cunmbr) \ XX(LAPACKE_cunmbr_work) \ XX(LAPACKE_cunmhr) \ @@ -699,6 +717,10 @@ XX(LAPACKE_dgebrd_work) \ XX(LAPACKE_dgecon) \ XX(LAPACKE_dgecon_work) \ + XX(LAPACKE_dgedmd) \ + XX(LAPACKE_dgedmd_work) \ + XX(LAPACKE_dgedmdq) \ + XX(LAPACKE_dgedmdq_work) \ XX(LAPACKE_dgeequ) \ XX(LAPACKE_dgeequ_work) \ XX(LAPACKE_dgeequb) \ @@ -785,6 +807,8 @@ XX(LAPACKE_dgetrs_work) \ XX(LAPACKE_dgetsls) \ XX(LAPACKE_dgetsls_work) \ + XX(LAPACKE_dgetsqrhrt) \ + XX(LAPACKE_dgetsqrhrt_work) \ XX(LAPACKE_dgg_nancheck) \ XX(LAPACKE_dgg_trans) \ XX(LAPACKE_dggbak) \ @@ -856,6 +880,8 @@ XX(LAPACKE_dlagsy_work) \ XX(LAPACKE_dlamch) \ XX(LAPACKE_dlamch_work) \ + XX(LAPACKE_dlangb) \ + XX(LAPACKE_dlangb_work) \ XX(LAPACKE_dlange) \ XX(LAPACKE_dlange_work) \ XX(LAPACKE_dlansy) \ @@ -922,6 +948,10 @@ XX(LAPACKE_dorgrq_work) \ XX(LAPACKE_dorgtr) \ XX(LAPACKE_dorgtr_work) \ + XX(LAPACKE_dorgtsqr_row) \ + XX(LAPACKE_dorgtsqr_row_work) \ + XX(LAPACKE_dorhr_col) \ + XX(LAPACKE_dorhr_col_work) \ XX(LAPACKE_dormbr) \ XX(LAPACKE_dormbr_work) \ XX(LAPACKE_dormhr) \ @@ -1257,6 +1287,8 @@ XX(LAPACKE_dtrsna) \ XX(LAPACKE_dtrsna_work) \ XX(LAPACKE_dtrsyl) \ + XX(LAPACKE_dtrsyl3) \ + XX(LAPACKE_dtrsyl3_work) \ XX(LAPACKE_dtrsyl_work) \ XX(LAPACKE_dtrtri) \ XX(LAPACKE_dtrtri_work) \ @@ -1266,6 +1298,8 @@ XX(LAPACKE_dtrttf_work) \ XX(LAPACKE_dtrttp) \ XX(LAPACKE_dtrttp_work) \ + XX(LAPACKE_dtz_nancheck) \ + XX(LAPACKE_dtz_trans) \ XX(LAPACKE_dtzrzf) \ XX(LAPACKE_dtzrzf_work) \ XX(LAPACKE_get_nancheck) \ @@ -1313,6 +1347,10 @@ XX(LAPACKE_sgebrd_work) \ XX(LAPACKE_sgecon) \ XX(LAPACKE_sgecon_work) \ + XX(LAPACKE_sgedmd) \ + XX(LAPACKE_sgedmd_work) \ + XX(LAPACKE_sgedmdq) \ + XX(LAPACKE_sgedmdq_work) \ XX(LAPACKE_sgeequ) \ XX(LAPACKE_sgeequ_work) \ XX(LAPACKE_sgeequb) \ @@ -1399,6 +1437,8 @@ XX(LAPACKE_sgetrs_work) \ XX(LAPACKE_sgetsls) \ XX(LAPACKE_sgetsls_work) \ + XX(LAPACKE_sgetsqrhrt) \ + XX(LAPACKE_sgetsqrhrt_work) \ XX(LAPACKE_sgg_nancheck) \ XX(LAPACKE_sgg_trans) \ XX(LAPACKE_sggbak) \ @@ -1470,6 +1510,8 @@ XX(LAPACKE_slagsy_work) \ XX(LAPACKE_slamch) \ XX(LAPACKE_slamch_work) \ + XX(LAPACKE_slangb) \ + XX(LAPACKE_slangb_work) \ XX(LAPACKE_slange) \ XX(LAPACKE_slange_work) \ XX(LAPACKE_slansy) \ @@ -1536,6 +1578,10 @@ XX(LAPACKE_sorgrq_work) \ XX(LAPACKE_sorgtr) \ XX(LAPACKE_sorgtr_work) \ + XX(LAPACKE_sorgtsqr_row) \ + XX(LAPACKE_sorgtsqr_row_work) \ + XX(LAPACKE_sorhr_col) \ + XX(LAPACKE_sorhr_col_work) \ XX(LAPACKE_sormbr) \ XX(LAPACKE_sormbr_work) \ XX(LAPACKE_sormhr) \ @@ -1867,6 +1913,8 @@ XX(LAPACKE_strsna) \ XX(LAPACKE_strsna_work) \ XX(LAPACKE_strsyl) \ + XX(LAPACKE_strsyl3) \ + XX(LAPACKE_strsyl3_work) \ XX(LAPACKE_strsyl_work) \ XX(LAPACKE_strtri) \ XX(LAPACKE_strtri_work) \ @@ -1876,6 +1924,8 @@ XX(LAPACKE_strttf_work) \ XX(LAPACKE_strttp) \ XX(LAPACKE_strttp_work) \ + XX(LAPACKE_stz_nancheck) \ + XX(LAPACKE_stz_trans) \ XX(LAPACKE_stzrzf) \ XX(LAPACKE_stzrzf_work) \ XX(LAPACKE_xerbla) \ @@ -1918,6 +1968,10 @@ XX(LAPACKE_zgebrd_work) \ XX(LAPACKE_zgecon) \ XX(LAPACKE_zgecon_work) \ + XX(LAPACKE_zgedmd) \ + XX(LAPACKE_zgedmd_work) \ + XX(LAPACKE_zgedmdq) \ + XX(LAPACKE_zgedmdq_work) \ XX(LAPACKE_zgeequ) \ XX(LAPACKE_zgeequ_work) \ XX(LAPACKE_zgeequb) \ @@ -2004,6 +2058,8 @@ XX(LAPACKE_zgetrs_work) \ XX(LAPACKE_zgetsls) \ XX(LAPACKE_zgetsls_work) \ + XX(LAPACKE_zgetsqrhrt) \ + XX(LAPACKE_zgetsqrhrt_work) \ XX(LAPACKE_zgg_nancheck) \ XX(LAPACKE_zgg_trans) \ XX(LAPACKE_zggbak) \ @@ -2219,6 +2275,8 @@ XX(LAPACKE_zlaghe_work) \ XX(LAPACKE_zlagsy) \ XX(LAPACKE_zlagsy_work) \ + XX(LAPACKE_zlangb) \ + XX(LAPACKE_zlangb_work) \ XX(LAPACKE_zlange) \ XX(LAPACKE_zlange_work) \ XX(LAPACKE_zlanhe) \ @@ -2490,6 +2548,8 @@ XX(LAPACKE_ztrsna) \ XX(LAPACKE_ztrsna_work) \ XX(LAPACKE_ztrsyl) \ + XX(LAPACKE_ztrsyl3) \ + XX(LAPACKE_ztrsyl3_work) \ XX(LAPACKE_ztrsyl_work) \ XX(LAPACKE_ztrtri) \ XX(LAPACKE_ztrtri_work) \ @@ -2499,6 +2559,8 @@ XX(LAPACKE_ztrttf_work) \ XX(LAPACKE_ztrttp) \ XX(LAPACKE_ztrttp_work) \ + XX(LAPACKE_ztz_nancheck) \ + XX(LAPACKE_ztz_trans) \ XX(LAPACKE_ztzrzf) \ XX(LAPACKE_ztzrzf_work) \ XX(LAPACKE_zunbdb) \ @@ -2521,6 +2583,10 @@ XX(LAPACKE_zungrq_work) \ XX(LAPACKE_zungtr) \ XX(LAPACKE_zungtr_work) \ + XX(LAPACKE_zungtsqr_row) \ + XX(LAPACKE_zungtsqr_row_work) \ + XX(LAPACKE_zunhr_col) \ + XX(LAPACKE_zunhr_col_work) \ XX(LAPACKE_zunmbr) \ XX(LAPACKE_zunmbr_work) \ XX(LAPACKE_zunmhr) \ @@ -2547,6 +2613,7 @@ XX(cbdsqr_) \ XX(cblas_caxpby) \ XX(cblas_caxpy) \ + XX(cblas_caxpyc) \ XX(cblas_ccopy) \ XX(cblas_cdotc) \ XX(cblas_cdotc_sub) \ @@ -2556,6 +2623,7 @@ XX(cblas_cgeadd) \ XX(cblas_cgemm) \ XX(cblas_cgemm3m) \ + XX(cblas_cgemmt) \ XX(cblas_cgemv) \ XX(cblas_cgerc) \ XX(cblas_cgeru) \ @@ -2571,7 +2639,9 @@ XX(cblas_chpr2) \ XX(cblas_cimatcopy) \ XX(cblas_comatcopy) \ + XX(cblas_crotg) \ XX(cblas_cscal) \ + XX(cblas_csrot) \ XX(cblas_csscal) \ XX(cblas_cswap) \ XX(cblas_csymm) \ @@ -2585,6 +2655,8 @@ XX(cblas_ctrmv) \ XX(cblas_ctrsm) \ XX(cblas_ctrsv) \ + XX(cblas_damax) \ + XX(cblas_damin) \ XX(cblas_dasum) \ XX(cblas_daxpby) \ XX(cblas_daxpy) \ @@ -2594,6 +2666,7 @@ XX(cblas_dgbmv) \ XX(cblas_dgeadd) \ XX(cblas_dgemm) \ + XX(cblas_dgemmt) \ XX(cblas_dgemv) \ XX(cblas_dger) \ XX(cblas_dimatcopy) \ @@ -2625,6 +2698,8 @@ XX(cblas_dtrmv) \ XX(cblas_dtrsm) \ XX(cblas_dtrsv) \ + XX(cblas_dzamax) \ + XX(cblas_dzamin) \ XX(cblas_dzasum) \ XX(cblas_dznrm2) \ XX(cblas_dzsum) \ @@ -2644,6 +2719,8 @@ XX(cblas_izamin) \ XX(cblas_izmax) \ XX(cblas_izmin) \ + XX(cblas_samax) \ + XX(cblas_samin) \ XX(cblas_sasum) \ XX(cblas_saxpby) \ XX(cblas_saxpy) \ @@ -2653,6 +2730,8 @@ XX(cblas_sbgemm) \ XX(cblas_sbgemv) \ XX(cblas_sbstobf16) \ + XX(cblas_scamax) \ + XX(cblas_scamin) \ XX(cblas_scasum) \ XX(cblas_scnrm2) \ XX(cblas_scopy) \ @@ -2662,6 +2741,7 @@ XX(cblas_sgbmv) \ XX(cblas_sgeadd) \ XX(cblas_sgemm) \ + XX(cblas_sgemmt) \ XX(cblas_sgemv) \ XX(cblas_sger) \ XX(cblas_simatcopy) \ @@ -2705,6 +2785,7 @@ XX(cblas_zgeadd) \ XX(cblas_zgemm) \ XX(cblas_zgemm3m) \ + XX(cblas_zgemmt) \ XX(cblas_zgemv) \ XX(cblas_zgerc) \ XX(cblas_zgeru) \ @@ -2753,6 +2834,8 @@ XX(cgebd2_) \ XX(cgebrd_) \ XX(cgecon_) \ + XX(cgedmd_) \ + XX(cgedmdq_) \ XX(cgeequ_) \ XX(cgeequb_) \ XX(cgees_) \ @@ -2767,11 +2850,13 @@ XX(cgelq2_) \ XX(cgelq_) \ XX(cgelqf_) \ + XX(cgelqs_) \ XX(cgelqt3_) \ XX(cgelqt_) \ XX(cgels_) \ XX(cgelsd_) \ XX(cgelss_) \ + XX(cgelst_) \ XX(cgelsx_) \ XX(cgelsy_) \ XX(cgemlq_) \ @@ -2785,12 +2870,14 @@ XX(cgeql2_) \ XX(cgeqlf_) \ XX(cgeqp3_) \ + XX(cgeqp3rk_) \ XX(cgeqpf_) \ XX(cgeqr2_) \ XX(cgeqr2p_) \ XX(cgeqr_) \ XX(cgeqrf_) \ XX(cgeqrfp_) \ + XX(cgeqrs_) \ XX(cgeqrt2_) \ XX(cgeqrt3_) \ XX(cgeqrt_) \ @@ -2814,6 +2901,7 @@ XX(cgetri_) \ XX(cgetrs_) \ XX(cgetsls_) \ + XX(cgetsqrhrt_) \ XX(cggbak_) \ XX(cggbal_) \ XX(cgges3_) \ @@ -2995,6 +3083,8 @@ XX(claqhe_) \ XX(claqhp_) \ XX(claqp2_) \ + XX(claqp2rk_) \ + XX(claqp3rk_) \ XX(claqps_) \ XX(claqr0_) \ XX(claqr1_) \ @@ -3010,6 +3100,7 @@ XX(clarcm_) \ XX(clarf_) \ XX(clarfb_) \ + XX(clarfb_gett_) \ XX(clarfg_) \ XX(clarfgp_) \ XX(clarft_) \ @@ -3050,6 +3141,7 @@ XX(clatmt_) \ XX(clatps_) \ XX(clatrd_) \ + XX(clatrs3_) \ XX(clatrs_) \ XX(clatrz_) \ XX(clatsqr_) \ @@ -3102,6 +3194,7 @@ XX(cptts2_) \ XX(crot_) \ XX(crotg_) \ + XX(crscl_) \ XX(csbmv_) \ XX(cscal_) \ XX(cspcon_) \ @@ -3207,6 +3300,7 @@ XX(ctrsm_) \ XX(ctrsna_) \ XX(ctrsv_) \ + XX(ctrsyl3_) \ XX(ctrsyl_) \ XX(ctrti2_) \ XX(ctrtri_) \ @@ -3236,6 +3330,7 @@ XX(cungrq_) \ XX(cungtr_) \ XX(cungtsqr_) \ + XX(cungtsqr_row_) \ XX(cunhr_col_) \ XX(cunm22_) \ XX(cunm2l_) \ @@ -3285,6 +3380,8 @@ XX(dgebd2_) \ XX(dgebrd_) \ XX(dgecon_) \ + XX(dgedmd_) \ + XX(dgedmdq_) \ XX(dgeequ_) \ XX(dgeequb_) \ XX(dgees_) \ @@ -3299,11 +3396,13 @@ XX(dgelq2_) \ XX(dgelq_) \ XX(dgelqf_) \ + XX(dgelqs_) \ XX(dgelqt3_) \ XX(dgelqt_) \ XX(dgels_) \ XX(dgelsd_) \ XX(dgelss_) \ + XX(dgelst_) \ XX(dgelsx_) \ XX(dgelsy_) \ XX(dgemlq_) \ @@ -3316,12 +3415,14 @@ XX(dgeql2_) \ XX(dgeqlf_) \ XX(dgeqp3_) \ + XX(dgeqp3rk_) \ XX(dgeqpf_) \ XX(dgeqr2_) \ XX(dgeqr2p_) \ XX(dgeqr_) \ XX(dgeqrf_) \ XX(dgeqrfp_) \ + XX(dgeqrs_) \ XX(dgeqrt2_) \ XX(dgeqrt3_) \ XX(dgeqrt_) \ @@ -3344,6 +3445,7 @@ XX(dgetri_) \ XX(dgetrs_) \ XX(dgetsls_) \ + XX(dgetsqrhrt_) \ XX(dggbak_) \ XX(dggbal_) \ XX(dgges3_) \ @@ -3449,6 +3551,8 @@ XX(dlaqgb_) \ XX(dlaqge_) \ XX(dlaqp2_) \ + XX(dlaqp2rk_) \ + XX(dlaqp3rk_) \ XX(dlaqps_) \ XX(dlaqr0_) \ XX(dlaqr1_) \ @@ -3465,6 +3569,7 @@ XX(dlaran_) \ XX(dlarf_) \ XX(dlarfb_) \ + XX(dlarfb_gett_) \ XX(dlarfg_) \ XX(dlarfgp_) \ XX(dlarft_) \ @@ -3472,6 +3577,7 @@ XX(dlarfy_) \ XX(dlarge_) \ XX(dlargv_) \ + XX(dlarmm_) \ XX(dlarnd_) \ XX(dlarnv_) \ XX(dlaror_) \ @@ -3541,6 +3647,7 @@ XX(dlatmt_) \ XX(dlatps_) \ XX(dlatrd_) \ + XX(dlatrs3_) \ XX(dlatrs_) \ XX(dlatrz_) \ XX(dlatsqr_) \ @@ -3574,6 +3681,7 @@ XX(dorgrq_) \ XX(dorgtr_) \ XX(dorgtsqr_) \ + XX(dorgtsqr_row_) \ XX(dorhr_col_) \ XX(dorm22_) \ XX(dorm2l_) \ @@ -3788,6 +3896,7 @@ XX(dtrsm_) \ XX(dtrsna_) \ XX(dtrsv_) \ + XX(dtrsyl3_) \ XX(dtrsyl_) \ XX(dtrti2_) \ XX(dtrtri_) \ @@ -3921,6 +4030,8 @@ XX(sgebd2_) \ XX(sgebrd_) \ XX(sgecon_) \ + XX(sgedmd_) \ + XX(sgedmdq_) \ XX(sgeequ_) \ XX(sgeequb_) \ XX(sgees_) \ @@ -3935,11 +4046,13 @@ XX(sgelq2_) \ XX(sgelq_) \ XX(sgelqf_) \ + XX(sgelqs_) \ XX(sgelqt3_) \ XX(sgelqt_) \ XX(sgels_) \ XX(sgelsd_) \ XX(sgelss_) \ + XX(sgelst_) \ XX(sgelsx_) \ XX(sgelsy_) \ XX(sgemlq_) \ @@ -3952,12 +4065,14 @@ XX(sgeql2_) \ XX(sgeqlf_) \ XX(sgeqp3_) \ + XX(sgeqp3rk_) \ XX(sgeqpf_) \ XX(sgeqr2_) \ XX(sgeqr2p_) \ XX(sgeqr_) \ XX(sgeqrf_) \ XX(sgeqrfp_) \ + XX(sgeqrs_) \ XX(sgeqrt2_) \ XX(sgeqrt3_) \ XX(sgeqrt_) \ @@ -3980,6 +4095,7 @@ XX(sgetri_) \ XX(sgetrs_) \ XX(sgetsls_) \ + XX(sgetsqrhrt_) \ XX(sggbak_) \ XX(sggbal_) \ XX(sgges3_) \ @@ -4085,6 +4201,8 @@ XX(slaqgb_) \ XX(slaqge_) \ XX(slaqp2_) \ + XX(slaqp2rk_) \ + XX(slaqp3rk_) \ XX(slaqps_) \ XX(slaqr0_) \ XX(slaqr1_) \ @@ -4101,6 +4219,7 @@ XX(slaran_) \ XX(slarf_) \ XX(slarfb_) \ + XX(slarfb_gett_) \ XX(slarfg_) \ XX(slarfgp_) \ XX(slarft_) \ @@ -4108,6 +4227,7 @@ XX(slarfy_) \ XX(slarge_) \ XX(slargv_) \ + XX(slarmm_) \ XX(slarnd_) \ XX(slarnv_) \ XX(slaror_) \ @@ -4176,6 +4296,7 @@ XX(slatmt_) \ XX(slatps_) \ XX(slatrd_) \ + XX(slatrs3_) \ XX(slatrs_) \ XX(slatrz_) \ XX(slatsqr_) \ @@ -4209,6 +4330,7 @@ XX(sorgrq_) \ XX(sorgtr_) \ XX(sorgtsqr_) \ + XX(sorgtsqr_row_) \ XX(sorhr_col_) \ XX(sorm22_) \ XX(sorm2l_) \ @@ -4320,6 +4442,7 @@ XX(ssycon_3_) \ XX(ssycon_rook_) \ XX(ssyconv_) \ + XX(ssyconvf_) \ XX(ssyconvf_rook_) \ XX(ssyequb_) \ XX(ssyev_) \ @@ -4418,6 +4541,7 @@ XX(strsm_) \ XX(strsna_) \ XX(strsv_) \ + XX(strsyl3_) \ XX(strsyl_) \ XX(strti2_) \ XX(strtri_) \ @@ -4490,6 +4614,8 @@ XX(zgebd2_) \ XX(zgebrd_) \ XX(zgecon_) \ + XX(zgedmd_) \ + XX(zgedmdq_) \ XX(zgeequ_) \ XX(zgeequb_) \ XX(zgees_) \ @@ -4504,11 +4630,13 @@ XX(zgelq2_) \ XX(zgelq_) \ XX(zgelqf_) \ + XX(zgelqs_) \ XX(zgelqt3_) \ XX(zgelqt_) \ XX(zgels_) \ XX(zgelsd_) \ XX(zgelss_) \ + XX(zgelst_) \ XX(zgelsx_) \ XX(zgelsy_) \ XX(zgemlq_) \ @@ -4522,12 +4650,14 @@ XX(zgeql2_) \ XX(zgeqlf_) \ XX(zgeqp3_) \ + XX(zgeqp3rk_) \ XX(zgeqpf_) \ XX(zgeqr2_) \ XX(zgeqr2p_) \ XX(zgeqr_) \ XX(zgeqrf_) \ XX(zgeqrfp_) \ + XX(zgeqrs_) \ XX(zgeqrt2_) \ XX(zgeqrt3_) \ XX(zgeqrt_) \ @@ -4551,6 +4681,7 @@ XX(zgetri_) \ XX(zgetrs_) \ XX(zgetsls_) \ + XX(zgetsqrhrt_) \ XX(zggbak_) \ XX(zggbal_) \ XX(zgges3_) \ @@ -4731,6 +4862,8 @@ XX(zlaqhe_) \ XX(zlaqhp_) \ XX(zlaqp2_) \ + XX(zlaqp2rk_) \ + XX(zlaqp3rk_) \ XX(zlaqps_) \ XX(zlaqr0_) \ XX(zlaqr1_) \ @@ -4746,6 +4879,7 @@ XX(zlarcm_) \ XX(zlarf_) \ XX(zlarfb_) \ + XX(zlarfb_gett_) \ XX(zlarfg_) \ XX(zlarfgp_) \ XX(zlarft_) \ @@ -4787,6 +4921,7 @@ XX(zlatmt_) \ XX(zlatps_) \ XX(zlatrd_) \ + XX(zlatrs3_) \ XX(zlatrs_) \ XX(zlatrz_) \ XX(zlatsqr_) \ @@ -4839,6 +4974,7 @@ XX(zptts2_) \ XX(zrot_) \ XX(zrotg_) \ + XX(zrscl_) \ XX(zsbmv_) \ XX(zscal_) \ XX(zspcon_) \ @@ -4941,6 +5077,7 @@ XX(ztrsm_) \ XX(ztrsna_) \ XX(ztrsv_) \ + XX(ztrsyl3_) \ XX(ztrsyl_) \ XX(ztrti2_) \ XX(ztrtri_) \ @@ -4970,6 +5107,7 @@ XX(zungrq_) \ XX(zungtr_) \ XX(zungtsqr_) \ + XX(zungtsqr_row_) \ XX(zunhr_col_) \ XX(zunm22_) \ XX(zunm2l_) \ @@ -4988,51 +5126,51 @@ XX(zupgtr_) \ XX(zupmtr_) \ -#define NUM_EXPORTED_FUNCS 4985 +#define NUM_EXPORTED_FUNCS 5123 #endif #ifndef FLOAT32_FUNCS #define FLOAT32_FUNCS(XX) \ - XX(sdot_, 3899) \ - XX(sdsdot_, 3900) \ - XX(sasum_, 3876) \ - XX(scasum_, 3892) \ - XX(ssum_, 4312) \ - XX(scsum_, 3897) \ - XX(samax_, 3874) \ - XX(scamax_, 3890) \ - XX(samin_, 3875) \ - XX(scamin_, 3891) \ - XX(smax_, 4180) \ - XX(smin_, 4181) \ - XX(snrm2_, 4182) \ - XX(scnrm2_, 3893) \ - XX(slamch_, 4055) \ - XX(slamc3_, 4054) \ + XX(sdot_, 4008) \ + XX(sdsdot_, 4009) \ + XX(sasum_, 3985) \ + XX(scasum_, 4001) \ + XX(ssum_, 4434) \ + XX(scsum_, 4006) \ + XX(samax_, 3983) \ + XX(scamax_, 3999) \ + XX(samin_, 3984) \ + XX(scamin_, 4000) \ + XX(smax_, 4301) \ + XX(smin_, 4302) \ + XX(snrm2_, 4303) \ + XX(scnrm2_, 4002) \ + XX(slamch_, 4171) \ + XX(slamc3_, 4170) \ #endif #ifndef COMPLEX64_FUNCS #define COMPLEX64_FUNCS(XX) \ - XX(cdotu_, 2733) \ - XX(cdotc_, 2732) \ + XX(cdotu_, 2814) \ + XX(cdotc_, 2813) \ #endif #ifndef COMPLEX128_FUNCS #define COMPLEX128_FUNCS(XX) \ - XX(zdotu_, 4467) \ - XX(zdotc_, 4466) \ + XX(zdotu_, 4591) \ + XX(zdotc_, 4590) \ #endif #ifndef CBLAS_WORKAROUND_FUNCS #define CBLAS_WORKAROUND_FUNCS(XX) \ - XX(cblas_cdotc_sub, 2547) \ - XX(cblas_cdotu_sub, 2549) \ - XX(cblas_zdotc_sub, 2695) \ - XX(cblas_zdotu_sub, 2697) \ - XX(cblas_ddot, 2588) \ - XX(cblas_sdot, 2655) \ + XX(cblas_cdotc_sub, 2614) \ + XX(cblas_cdotu_sub, 2616) \ + XX(cblas_zdotc_sub, 2775) \ + XX(cblas_zdotu_sub, 2777) \ + XX(cblas_ddot, 2660) \ + XX(cblas_sdot, 2734) \ #endif