From ecae1389dff8bcb6643685aedb1fbaeb9b26d79d Mon Sep 17 00:00:00 2001 From: Chris Sidebottom Date: Sat, 23 Dec 2023 12:21:48 +0000 Subject: [PATCH] Reduce duplication in kernel definitions These files are exactly the same, so I believe we can reduce these files down. Other files require a slightly more complex unpicking. --- kernel/arm64/KERNEL.A64FX | 207 +----------------------------- kernel/arm64/KERNEL.CORTEXA55 | 197 +--------------------------- kernel/arm64/KERNEL.THUNDERX3T110 | 185 +------------------------- 3 files changed, 3 insertions(+), 586 deletions(-) diff --git a/kernel/arm64/KERNEL.A64FX b/kernel/arm64/KERNEL.A64FX index ccbce27e1b..bc59990979 100644 --- a/kernel/arm64/KERNEL.A64FX +++ b/kernel/arm64/KERNEL.A64FX @@ -1,206 +1 @@ -SAMINKERNEL = ../arm/amin.c -DAMINKERNEL = ../arm/amin.c -CAMINKERNEL = ../arm/zamin.c -ZAMINKERNEL = ../arm/zamin.c - -SMAXKERNEL = ../arm/max.c -DMAXKERNEL = ../arm/max.c - -SMINKERNEL = ../arm/min.c -DMINKERNEL = ../arm/min.c - -ISAMINKERNEL = ../arm/iamin.c -IDAMINKERNEL = ../arm/iamin.c -ICAMINKERNEL = ../arm/izamin.c -IZAMINKERNEL = ../arm/izamin.c - -ISMAXKERNEL = ../arm/imax.c -IDMAXKERNEL = ../arm/imax.c - -ISMINKERNEL = ../arm/imin.c -IDMINKERNEL = ../arm/imin.c - -STRSMKERNEL_LN = trsm_kernel_LN_sve.c -STRSMKERNEL_LT = trsm_kernel_LT_sve.c -STRSMKERNEL_RN = trsm_kernel_RN_sve.c -STRSMKERNEL_RT = trsm_kernel_RT_sve.c - -DTRSMKERNEL_LN = trsm_kernel_LN_sve.c -DTRSMKERNEL_LT = trsm_kernel_LT_sve.c -DTRSMKERNEL_RN = trsm_kernel_RN_sve.c -DTRSMKERNEL_RT = trsm_kernel_RT_sve.c - -TRSMCOPYLN_M = trsm_lncopy_sve.c -TRSMCOPYLT_M = trsm_ltcopy_sve.c -TRSMCOPYUN_M = trsm_uncopy_sve.c -TRSMCOPYUT_M = trsm_utcopy_sve.c - -CTRSMKERNEL_LN = trsm_kernel_LN_sve.c -CTRSMKERNEL_LT = trsm_kernel_LT_sve.c -CTRSMKERNEL_RN = trsm_kernel_RN_sve.c -CTRSMKERNEL_RT = trsm_kernel_RT_sve.c - -ZTRSMKERNEL_LN = trsm_kernel_LN_sve.c -ZTRSMKERNEL_LT = trsm_kernel_LT_sve.c -ZTRSMKERNEL_RN = trsm_kernel_RN_sve.c -ZTRSMKERNEL_RT = trsm_kernel_RT_sve.c - -ZTRSMCOPYLN_M = ztrsm_lncopy_sve.c -ZTRSMCOPYLT_M = ztrsm_ltcopy_sve.c -ZTRSMCOPYUN_M = ztrsm_uncopy_sve.c -ZTRSMCOPYUT_M = ztrsm_utcopy_sve.c - - -SAMAXKERNEL = amax.S -DAMAXKERNEL = amax.S -CAMAXKERNEL = zamax.S -ZAMAXKERNEL = zamax.S - -SAXPYKERNEL = axpy.S -DAXPYKERNEL = daxpy_thunderx2t99.S -CAXPYKERNEL = zaxpy.S -ZAXPYKERNEL = zaxpy.S - -SROTKERNEL = rot.S -DROTKERNEL = rot.S -CROTKERNEL = zrot.S -ZROTKERNEL = zrot.S - -SSCALKERNEL = scal.S -DSCALKERNEL = scal.S -CSCALKERNEL = zscal.S -ZSCALKERNEL = zscal.S - -SGEMVNKERNEL = gemv_n.S -DGEMVNKERNEL = gemv_n.S -CGEMVNKERNEL = zgemv_n.S -ZGEMVNKERNEL = zgemv_n.S - -SGEMVTKERNEL = gemv_t.S -DGEMVTKERNEL = gemv_t.S -CGEMVTKERNEL = zgemv_t.S -ZGEMVTKERNEL = zgemv_t.S - -SASUMKERNEL = sasum_thunderx2t99.c -DASUMKERNEL = dasum_thunderx2t99.c -CASUMKERNEL = casum_thunderx2t99.c -ZASUMKERNEL = zasum_thunderx2t99.c - -SCOPYKERNEL = copy_thunderx2t99.c -DCOPYKERNEL = copy_thunderx2t99.c -CCOPYKERNEL = copy_thunderx2t99.c -ZCOPYKERNEL = copy_thunderx2t99.c - -SSWAPKERNEL = swap_thunderx2t99.S -DSWAPKERNEL = swap_thunderx2t99.S -CSWAPKERNEL = swap_thunderx2t99.S -ZSWAPKERNEL = swap_thunderx2t99.S - -ISAMAXKERNEL = iamax_thunderx2t99.c -IDAMAXKERNEL = iamax_thunderx2t99.c -ICAMAXKERNEL = izamax_thunderx2t99.c -IZAMAXKERNEL = izamax_thunderx2t99.c - -SNRM2KERNEL = scnrm2_thunderx2t99.c -DNRM2KERNEL = dznrm2_thunderx2t99.c -CNRM2KERNEL = scnrm2_thunderx2t99.c -ZNRM2KERNEL = dznrm2_thunderx2t99.c - -DDOTKERNEL = dot.c -SDOTKERNEL = dot.c -CDOTKERNEL = zdot_thunderx2t99.c -ZDOTKERNEL = zdot_thunderx2t99.c -DSDOTKERNEL = dot.S - -DGEMM_BETA = dgemm_beta.S -SGEMM_BETA = sgemm_beta.S - -SGEMMKERNEL = sgemm_kernel_sve_v2x$(SGEMM_UNROLL_N).S -STRMMKERNEL = strmm_kernel_sve_v1x$(SGEMM_UNROLL_N).S - -SGEMMINCOPY = gemm_ncopy_sve_v1x$(SGEMM_UNROLL_N).c -SGEMMITCOPY = gemm_tcopy_sve_v1x$(SGEMM_UNROLL_N).c -SGEMMONCOPY = sgemm_ncopy_$(SGEMM_UNROLL_N).S -SGEMMOTCOPY = sgemm_tcopy_$(SGEMM_UNROLL_N).S - -SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX) -SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX) -SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX) -SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX) - -STRMMUNCOPY_M = trmm_uncopy_sve_v1.c -STRMMLNCOPY_M = trmm_lncopy_sve_v1.c -STRMMUTCOPY_M = trmm_utcopy_sve_v1.c -STRMMLTCOPY_M = trmm_ltcopy_sve_v1.c - -SSYMMUCOPY_M = symm_ucopy_sve.c -SSYMMLCOPY_M = symm_lcopy_sve.c - -DGEMMKERNEL = dgemm_kernel_sve_v2x$(DGEMM_UNROLL_N).S -DTRMMKERNEL = dtrmm_kernel_sve_v1x$(DGEMM_UNROLL_N).S - -DGEMMINCOPY = gemm_ncopy_sve_v1x$(DGEMM_UNROLL_N).c -DGEMMITCOPY = gemm_tcopy_sve_v1x$(DGEMM_UNROLL_N).c -DGEMMONCOPY = dgemm_ncopy_$(DGEMM_UNROLL_N).S -DGEMMOTCOPY = dgemm_tcopy_$(DGEMM_UNROLL_N).S - -DGEMMINCOPYOBJ = dgemm_incopy$(TSUFFIX).$(SUFFIX) -DGEMMITCOPYOBJ = dgemm_itcopy$(TSUFFIX).$(SUFFIX) -DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX) -DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX) - -DTRMMUNCOPY_M = trmm_uncopy_sve_v1.c -DTRMMLNCOPY_M = trmm_lncopy_sve_v1.c -DTRMMUTCOPY_M = trmm_utcopy_sve_v1.c -DTRMMLTCOPY_M = trmm_ltcopy_sve_v1.c - -DSYMMUCOPY_M = symm_ucopy_sve.c -DSYMMLCOPY_M = symm_lcopy_sve.c - -CGEMMKERNEL = cgemm_kernel_sve_v1x$(ZGEMM_UNROLL_N).S -CTRMMKERNEL = ctrmm_kernel_sve_v1x$(ZGEMM_UNROLL_N).S - -CGEMMINCOPY = gemm_ncopy_complex_sve_v1x$(ZGEMM_UNROLL_N).c -CGEMMITCOPY = gemm_tcopy_complex_sve_v1x$(ZGEMM_UNROLL_N).c -CGEMMONCOPY = ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_N).c -CGEMMOTCOPY = ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_N).c - -CGEMMINCOPYOBJ = cgemm_incopy$(TSUFFIX).$(SUFFIX) -CGEMMITCOPYOBJ = cgemm_itcopy$(TSUFFIX).$(SUFFIX) -CGEMMONCOPYOBJ = cgemm_oncopy$(TSUFFIX).$(SUFFIX) -CGEMMOTCOPYOBJ = cgemm_otcopy$(TSUFFIX).$(SUFFIX) - -CTRMMUNCOPY_M = ztrmm_uncopy_sve_v1.c -CTRMMLNCOPY_M = ztrmm_lncopy_sve_v1.c -CTRMMUTCOPY_M = ztrmm_utcopy_sve_v1.c -CTRMMLTCOPY_M = ztrmm_ltcopy_sve_v1.c - -CHEMMLTCOPY_M = zhemm_ltcopy_sve.c -CHEMMUTCOPY_M = zhemm_utcopy_sve.c - -CSYMMUCOPY_M = zsymm_ucopy_sve.c -CSYMMLCOPY_M = zsymm_lcopy_sve.c - -ZGEMMKERNEL = zgemm_kernel_sve_v1x$(ZGEMM_UNROLL_N).S -ZTRMMKERNEL = ztrmm_kernel_sve_v1x$(ZGEMM_UNROLL_N).S - -ZGEMMINCOPY = gemm_ncopy_complex_sve_v1x$(ZGEMM_UNROLL_N).c -ZGEMMITCOPY = gemm_tcopy_complex_sve_v1x$(ZGEMM_UNROLL_N).c -ZGEMMONCOPY = ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_N).c -ZGEMMOTCOPY = ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_N).c - -ZGEMMINCOPYOBJ = zgemm_incopy$(TSUFFIX).$(SUFFIX) -ZGEMMITCOPYOBJ = zgemm_itcopy$(TSUFFIX).$(SUFFIX) -ZGEMMONCOPYOBJ = zgemm_oncopy$(TSUFFIX).$(SUFFIX) -ZGEMMOTCOPYOBJ = zgemm_otcopy$(TSUFFIX).$(SUFFIX) - -ZTRMMUNCOPY_M = ztrmm_uncopy_sve_v1.c -ZTRMMLNCOPY_M = ztrmm_lncopy_sve_v1.c -ZTRMMUTCOPY_M = ztrmm_utcopy_sve_v1.c -ZTRMMLTCOPY_M = ztrmm_ltcopy_sve_v1.c - -ZHEMMLTCOPY_M = zhemm_ltcopy_sve.c -ZHEMMUTCOPY_M = zhemm_utcopy_sve.c - -ZSYMMUCOPY_M = zsymm_ucopy_sve.c -ZSYMMLCOPY_M = zsymm_lcopy_sve.c +include $(KERNELDIR)/KERNEL.ARMV8SVE diff --git a/kernel/arm64/KERNEL.CORTEXA55 b/kernel/arm64/KERNEL.CORTEXA55 index e2e0067707..574e98b8cd 100644 --- a/kernel/arm64/KERNEL.CORTEXA55 +++ b/kernel/arm64/KERNEL.CORTEXA55 @@ -1,196 +1 @@ -SAMINKERNEL = ../arm/amin.c -DAMINKERNEL = ../arm/amin.c -CAMINKERNEL = ../arm/zamin.c -ZAMINKERNEL = ../arm/zamin.c - -SMAXKERNEL = ../arm/max.c -DMAXKERNEL = ../arm/max.c - -SMINKERNEL = ../arm/min.c -DMINKERNEL = ../arm/min.c - -ISAMINKERNEL = ../arm/iamin.c -IDAMINKERNEL = ../arm/iamin.c -ICAMINKERNEL = ../arm/izamin.c -IZAMINKERNEL = ../arm/izamin.c - -ISMAXKERNEL = ../arm/imax.c -IDMAXKERNEL = ../arm/imax.c - -ISMINKERNEL = ../arm/imin.c -IDMINKERNEL = ../arm/imin.c - -STRSMKERNEL_LN = ../generic/trsm_kernel_LN.c -STRSMKERNEL_LT = ../generic/trsm_kernel_LT.c -STRSMKERNEL_RN = ../generic/trsm_kernel_RN.c -STRSMKERNEL_RT = ../generic/trsm_kernel_RT.c - -DTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c -DTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c -DTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c -DTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c - -CTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c -CTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c -CTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c -CTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c - -ZTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c -ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c -ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c -ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c - -SAMAXKERNEL = amax.S -DAMAXKERNEL = amax.S -CAMAXKERNEL = zamax.S -ZAMAXKERNEL = zamax.S - -SAXPYKERNEL = axpy.S -DAXPYKERNEL = axpy.S -CAXPYKERNEL = zaxpy.S -ZAXPYKERNEL = zaxpy.S - -SROTKERNEL = rot.S -DROTKERNEL = rot.S -CROTKERNEL = zrot.S -ZROTKERNEL = zrot.S - -SSCALKERNEL = scal.S -DSCALKERNEL = scal.S -CSCALKERNEL = zscal.S -ZSCALKERNEL = zscal.S - -SGEMVNKERNEL = gemv_n.S -DGEMVNKERNEL = gemv_n.S -CGEMVNKERNEL = zgemv_n.S -ZGEMVNKERNEL = zgemv_n.S - -SGEMVTKERNEL = gemv_t.S -DGEMVTKERNEL = gemv_t.S -CGEMVTKERNEL = zgemv_t.S -ZGEMVTKERNEL = zgemv_t.S - - -SASUMKERNEL = asum.S -DASUMKERNEL = asum.S -CASUMKERNEL = casum.S -ZASUMKERNEL = zasum.S - -SCOPYKERNEL = copy.S -DCOPYKERNEL = copy.S -CCOPYKERNEL = copy.S -ZCOPYKERNEL = copy.S - -SSWAPKERNEL = swap.S -DSWAPKERNEL = swap.S -CSWAPKERNEL = swap.S -ZSWAPKERNEL = swap.S - -ISAMAXKERNEL = iamax.S -IDAMAXKERNEL = iamax.S -ICAMAXKERNEL = izamax.S -IZAMAXKERNEL = izamax.S - -SNRM2KERNEL = nrm2.S -DNRM2KERNEL = nrm2.S -CNRM2KERNEL = znrm2.S -ZNRM2KERNEL = znrm2.S - -ifneq ($(C_COMPILER), PGI) -SDOTKERNEL = ../generic/dot.c -else -SDOTKERNEL = dot.S -endif -DDOTKERNEL = dot.S -ifneq ($(C_COMPILER), PGI) -CDOTKERNEL = zdot.S -ZDOTKERNEL = zdot.S -else -CDOTKERNEL = ../arm/zdot.c -ZDOTKERNEL = ../arm/zdot.c -endif -DSDOTKERNEL = dot.S - -DGEMM_BETA = dgemm_beta.S -SGEMM_BETA = sgemm_beta.S - -ifeq ($(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N), 8x8) -SGEMMKERNEL = sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N)_cortexa53.S -STRMMKERNEL = strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N)_cortexa53.S -else -SGEMMKERNEL = sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S -STRMMKERNEL = strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S -endif -ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N)) -ifeq ($(SGEMM_UNROLL_M), 16) -SGEMMITCOPY = sgemm_tcopy_$(SGEMM_UNROLL_M).S -else -SGEMMITCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c -endif -ifeq ($(SGEMM_UNROLL_M), 4) -SGEMMINCOPY = sgemm_ncopy_$(SGEMM_UNROLL_M).S -else -SGEMMINCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c -endif -SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX) -SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX) -endif - -SGEMMOTCOPY = sgemm_tcopy_$(SGEMM_UNROLL_N).S -SGEMMONCOPY = sgemm_ncopy_$(SGEMM_UNROLL_N).S -SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX) -SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX) - -DGEMMKERNEL = dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N)_cortexa53.c -DTRMMKERNEL = dtrmm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S - -ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N)) - -ifeq ($(DGEMM_UNROLL_M), 8) -DGEMMINCOPY = dgemm_ncopy_$(DGEMM_UNROLL_M).S -DGEMMITCOPY = dgemm_tcopy_$(DGEMM_UNROLL_M).S -else -DGEMMINCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_M).c -DGEMMITCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_M).c -endif - -DGEMMINCOPYOBJ = dgemm_incopy$(TSUFFIX).$(SUFFIX) -DGEMMITCOPYOBJ = dgemm_itcopy$(TSUFFIX).$(SUFFIX) -endif - -ifeq ($(DGEMM_UNROLL_N), 4) -DGEMMONCOPY = dgemm_ncopy_$(DGEMM_UNROLL_N).S -DGEMMOTCOPY = dgemm_tcopy_$(DGEMM_UNROLL_N).S -else -DGEMMONCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c -DGEMMOTCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_N).c -endif - -DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX) -DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX) - -CGEMMKERNEL = cgemm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N)_cortexa53.c -CTRMMKERNEL = ctrmm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S -ifneq ($(CGEMM_UNROLL_M), $(CGEMM_UNROLL_N)) -CGEMMINCOPY = ../generic/zgemm_ncopy_$(CGEMM_UNROLL_M).c -CGEMMITCOPY = ../generic/zgemm_tcopy_$(CGEMM_UNROLL_M).c -CGEMMINCOPYOBJ = cgemm_incopy$(TSUFFIX).$(SUFFIX) -CGEMMITCOPYOBJ = cgemm_itcopy$(TSUFFIX).$(SUFFIX) -endif -CGEMMONCOPY = ../generic/zgemm_ncopy_$(CGEMM_UNROLL_N).c -CGEMMOTCOPY = ../generic/zgemm_tcopy_$(CGEMM_UNROLL_N).c -CGEMMONCOPYOBJ = cgemm_oncopy$(TSUFFIX).$(SUFFIX) -CGEMMOTCOPYOBJ = cgemm_otcopy$(TSUFFIX).$(SUFFIX) - -ZGEMMKERNEL = zgemm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N)_cortexa53.c -ZTRMMKERNEL = ztrmm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S -ifneq ($(ZGEMM_UNROLL_M), $(ZGEMM_UNROLL_N)) -ZGEMMINCOPY = ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_M).c -ZGEMMITCOPY = ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_M).c -ZGEMMINCOPYOBJ = zgemm_incopy$(TSUFFIX).$(SUFFIX) -ZGEMMITCOPYOBJ = zgemm_itcopy$(TSUFFIX).$(SUFFIX) -endif -ZGEMMONCOPY = ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_N).c -ZGEMMOTCOPY = ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_N).c -ZGEMMONCOPYOBJ = zgemm_oncopy$(TSUFFIX).$(SUFFIX) -ZGEMMOTCOPYOBJ = zgemm_otcopy$(TSUFFIX).$(SUFFIX) +include $(KERNELDIR)/KERNEL.CORTEXA53 diff --git a/kernel/arm64/KERNEL.THUNDERX3T110 b/kernel/arm64/KERNEL.THUNDERX3T110 index 41cedc8519..5d3bd69f79 100644 --- a/kernel/arm64/KERNEL.THUNDERX3T110 +++ b/kernel/arm64/KERNEL.THUNDERX3T110 @@ -1,184 +1 @@ -SAMINKERNEL = ../arm/amin.c -DAMINKERNEL = ../arm/amin.c -CAMINKERNEL = ../arm/zamin.c -ZAMINKERNEL = ../arm/zamin.c - -SMAXKERNEL = ../arm/max.c -DMAXKERNEL = ../arm/max.c - -SMINKERNEL = ../arm/min.c -DMINKERNEL = ../arm/min.c - -ISAMINKERNEL = ../arm/iamin.c -IDAMINKERNEL = ../arm/iamin.c -ICAMINKERNEL = ../arm/izamin.c -IZAMINKERNEL = ../arm/izamin.c - -ISMAXKERNEL = ../arm/imax.c -IDMAXKERNEL = ../arm/imax.c - -ISMINKERNEL = ../arm/imin.c -IDMINKERNEL = ../arm/imin.c - -STRSMKERNEL_LN = ../generic/trsm_kernel_LN.c -STRSMKERNEL_LT = ../generic/trsm_kernel_LT.c -STRSMKERNEL_RN = ../generic/trsm_kernel_RN.c -STRSMKERNEL_RT = ../generic/trsm_kernel_RT.c - -DTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c -DTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c -DTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c -DTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c - -CTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c -CTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c -CTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c -CTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c - -ZTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c -ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c -ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c -ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c - -SAMAXKERNEL = amax.S -DAMAXKERNEL = amax.S -CAMAXKERNEL = zamax.S -ZAMAXKERNEL = zamax.S - -SAXPYKERNEL = axpy.S -DAXPYKERNEL = daxpy_thunderx2t99.S -CAXPYKERNEL = zaxpy.S -ZAXPYKERNEL = zaxpy.S - -SROTKERNEL = rot.S -DROTKERNEL = rot.S -CROTKERNEL = zrot.S -ZROTKERNEL = zrot.S - -SSCALKERNEL = scal.S -DSCALKERNEL = scal.S -CSCALKERNEL = zscal.S -ZSCALKERNEL = zscal.S - -SGEMVNKERNEL = gemv_n.S -DGEMVNKERNEL = gemv_n.S -CGEMVNKERNEL = zgemv_n.S -ZGEMVNKERNEL = zgemv_n.S - -SGEMVTKERNEL = gemv_t.S -DGEMVTKERNEL = gemv_t.S -CGEMVTKERNEL = zgemv_t.S -ZGEMVTKERNEL = zgemv_t.S - -STRMMKERNEL = strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S -ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N)) -SGEMMINCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c -SGEMMITCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c -SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX) -SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX) -endif -SGEMMONCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c -SGEMMOTCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c -SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX) -SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX) - -DTRMMKERNEL = dtrmm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S - -ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N)) - -ifeq ($(DGEMM_UNROLL_M), 8) -DGEMMINCOPY = dgemm_ncopy_$(DGEMM_UNROLL_M).S -DGEMMITCOPY = dgemm_tcopy_$(DGEMM_UNROLL_M).S -else -DGEMMINCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_M).c -DGEMMITCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_M).c -endif - -DGEMMINCOPYOBJ = dgemm_incopy$(TSUFFIX).$(SUFFIX) -DGEMMITCOPYOBJ = dgemm_itcopy$(TSUFFIX).$(SUFFIX) -endif - -ifeq ($(DGEMM_UNROLL_N), 4) -DGEMMONCOPY = dgemm_ncopy_$(DGEMM_UNROLL_N).S -DGEMMOTCOPY = dgemm_tcopy_$(DGEMM_UNROLL_N).S -else -DGEMMONCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c -DGEMMOTCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_N).c -endif - -DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX) -DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX) - -CTRMMKERNEL = ctrmm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S -ifneq ($(CGEMM_UNROLL_M), $(CGEMM_UNROLL_N)) -CGEMMINCOPY = ../generic/zgemm_ncopy_$(CGEMM_UNROLL_M).c -CGEMMITCOPY = ../generic/zgemm_tcopy_$(CGEMM_UNROLL_M).c -CGEMMINCOPYOBJ = cgemm_incopy$(TSUFFIX).$(SUFFIX) -CGEMMITCOPYOBJ = cgemm_itcopy$(TSUFFIX).$(SUFFIX) -endif -CGEMMONCOPY = ../generic/zgemm_ncopy_$(CGEMM_UNROLL_N).c -CGEMMOTCOPY = ../generic/zgemm_tcopy_$(CGEMM_UNROLL_N).c -CGEMMONCOPYOBJ = cgemm_oncopy$(TSUFFIX).$(SUFFIX) -CGEMMOTCOPYOBJ = cgemm_otcopy$(TSUFFIX).$(SUFFIX) - -ZTRMMKERNEL = ztrmm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S -ifneq ($(ZGEMM_UNROLL_M), $(ZGEMM_UNROLL_N)) -ZGEMMINCOPY = ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_M).c -ZGEMMITCOPY = ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_M).c -ZGEMMINCOPYOBJ = zgemm_incopy$(TSUFFIX).$(SUFFIX) -ZGEMMITCOPYOBJ = zgemm_itcopy$(TSUFFIX).$(SUFFIX) -endif -ZGEMMONCOPY = ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_N).c -ZGEMMOTCOPY = ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_N).c -ZGEMMONCOPYOBJ = zgemm_oncopy$(TSUFFIX).$(SUFFIX) -ZGEMMOTCOPYOBJ = zgemm_otcopy$(TSUFFIX).$(SUFFIX) - -SASUMKERNEL = sasum_thunderx2t99.c -DASUMKERNEL = dasum_thunderx2t99.c -CASUMKERNEL = casum_thunderx2t99.c -ZASUMKERNEL = zasum_thunderx2t99.c - -SCOPYKERNEL = copy_thunderx2t99.c -DCOPYKERNEL = copy_thunderx2t99.c -CCOPYKERNEL = copy_thunderx2t99.c -ZCOPYKERNEL = copy_thunderx2t99.c - -SSWAPKERNEL = swap_thunderx2t99.S -DSWAPKERNEL = swap_thunderx2t99.S -CSWAPKERNEL = swap_thunderx2t99.S -ZSWAPKERNEL = swap_thunderx2t99.S - -ISAMAXKERNEL = iamax_thunderx2t99.c -IDAMAXKERNEL = iamax_thunderx2t99.c -ICAMAXKERNEL = izamax_thunderx2t99.c -IZAMAXKERNEL = izamax_thunderx2t99.c - -SNRM2KERNEL = scnrm2_thunderx2t99.c -CNRM2KERNEL = scnrm2_thunderx2t99.c -#DNRM2KERNEL = dznrm2_thunderx2t99_fast.c -#ZNRM2KERNEL = dznrm2_thunderx2t99_fast.c -DNRM2KERNEL = dznrm2_thunderx2t99.c -ZNRM2KERNEL = dznrm2_thunderx2t99.c - - -DDOTKERNEL = dot.c -SDOTKERNEL = dot.c -CDOTKERNEL = zdot_thunderx2t99.c -ZDOTKERNEL = zdot_thunderx2t99.c -DSDOTKERNEL = dot.S - -ifeq ($(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N), 8x4) -DGEMMKERNEL = dgemm_kernel_8x4_thunderx2t99.S -endif - -ifeq ($(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N), 16x4) -SGEMMKERNEL = sgemm_kernel_16x4_thunderx2t99.S -endif - -ifeq ($(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N), 8x4) -CGEMMKERNEL = cgemm_kernel_8x4_thunderx2t99.S -endif - -ifeq ($(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N), 4x4) -ZGEMMKERNEL = zgemm_kernel_4x4_thunderx2t99.S -endif +include $(KERNELDIR)/KERNEL.THUNDERX2T99