Merge pull request #4258 from martin-frbg/issue4257

Fix build on Fujitsu A64FX

Author: martin-frbg

kernel/arm64/KERNEL.A64FX

@@ -57,7 +57,7 @@ CAMAXKERNEL  = zamax.S
 ZAMAXKERNEL  = zamax.S
 
 SAXPYKERNEL  = axpy.S
-DAXPYKERNEL  = axpy.S
+DAXPYKERNEL  = daxpy_thunderx2t99.S
 CAXPYKERNEL  = zaxpy.S
 ZAXPYKERNEL  = zaxpy.S
 
@@ -81,45 +81,35 @@ 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
-
-DDOTKERNEL     = dot.S
-ifneq ($(C_COMPILER), PGI)
-SDOTKERNEL     = ../generic/dot.c
-else
-SDOTKERNEL = dot.S
-endif
-ifneq ($(C_COMPILER), PGI)
-CDOTKERNEL     = zdot.S
-ZDOTKERNEL     = zdot.S
-else
-CDOTKERNEL = ../arm/zdot.c
-ZDOTKERNEL = ../arm/zdot.c
-endif
+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
@@ -128,10 +118,10 @@ SGEMM_BETA     = sgemm_beta.S
 SGEMMKERNEL    =  sgemm_kernel_sve_v2x$(SGEMM_UNROLL_N).S
 STRMMKERNEL    =  strmm_kernel_sve_v1x$(SGEMM_UNROLL_N).S
 
-SGEMMINCOPY    =  sgemm_ncopy_sve_v1.c
-SGEMMITCOPY    =  sgemm_tcopy_sve_v1.c
-SGEMMONCOPY    =  sgemm_ncopy_$(DGEMM_UNROLL_N).S
-SGEMMOTCOPY    =  sgemm_tcopy_$(DGEMM_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)
@@ -149,8 +139,8 @@ 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    =  dgemm_ncopy_sve_v1.c
-DGEMMITCOPY    =  dgemm_tcopy_sve_v1.c
+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
 
@@ -170,8 +160,8 @@ 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    =  cgemm_ncopy_sve_v1.c
-CGEMMITCOPY    =  cgemm_tcopy_sve_v1.c
+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
 
@@ -194,8 +184,8 @@ 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    =  zgemm_ncopy_sve_v1.c
-ZGEMMITCOPY    =  zgemm_tcopy_sve_v1.c
+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