floraachy
/
OpenBLAS
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

Merge 682d9be199 into a6515bb858

This commit is contained in:
akira-miasato 2017-07-03 05:48:46 +00:00 committed by GitHub
parent a6515bb858 682d9be199
commit 3153aa734d
1 changed files with 177 additions and 71 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

220
kernel/arm/KERNEL.ARMV7
View File

@ -1,91 +1,42 @@
################################################################################# #################################################################################
############################################################
# Files already implemented in ARM Assembly for softfp ABI #
############################################################
SAMAXKERNEL = iamax_vfp.S SAMAXKERNEL = iamax_vfp.S
DAMAXKERNEL = iamax_vfp.S
CAMAXKERNEL = iamax_vfp.S
ZAMAXKERNEL = iamax_vfp.S
SAMINKERNEL = iamax_vfp.S SAMINKERNEL = iamax_vfp.S
DAMINKERNEL = iamax_vfp.S
CAMINKERNEL = iamax_vfp.S
ZAMINKERNEL = iamax_vfp.S
SMAXKERNEL = iamax_vfp.S SMAXKERNEL = iamax_vfp.S
DMAXKERNEL = iamax_vfp.S
SMINKERNEL = iamax_vfp.S SMINKERNEL = iamax_vfp.S
DMINKERNEL = iamax_vfp.S
ISAMAXKERNEL = iamax_vfp.S ISAMAXKERNEL = iamax_vfp.S
IDAMAXKERNEL = iamax_vfp.S
ICAMAXKERNEL = iamax_vfp.S
IZAMAXKERNEL = iamax_vfp.S
ISAMINKERNEL = iamax_vfp.S ISAMINKERNEL = iamax_vfp.S
IDAMINKERNEL = iamax_vfp.S
ICAMINKERNEL = iamax_vfp.S
IZAMINKERNEL = iamax_vfp.S
ISMAXKERNEL = iamax_vfp.S ISMAXKERNEL = iamax_vfp.S
IDMAXKERNEL = iamax_vfp.S
ISMINKERNEL = iamax_vfp.S ISMINKERNEL = iamax_vfp.S
IDMINKERNEL = iamax_vfp.S
SSWAPKERNEL = swap_vfp.S SSWAPKERNEL = swap_vfp.S
DSWAPKERNEL = swap_vfp.S
CSWAPKERNEL = swap_vfp.S
ZSWAPKERNEL = swap_vfp.S
SASUMKERNEL = asum_vfp.S SASUMKERNEL = asum_vfp.S
DASUMKERNEL = asum_vfp.S
CASUMKERNEL = asum_vfp.S
ZASUMKERNEL = asum_vfp.S
SAXPYKERNEL = axpy_vfp.S SAXPYKERNEL = axpy_vfp.S
DAXPYKERNEL = axpy_vfp.S
CAXPYKERNEL = axpy_vfp.S
ZAXPYKERNEL = axpy_vfp.S
SCOPYKERNEL = copy.c
DCOPYKERNEL = copy.c
CCOPYKERNEL = zcopy.c
ZCOPYKERNEL = zcopy.c
SDOTKERNEL = sdot_vfp.S SDOTKERNEL = sdot_vfp.S
DDOTKERNEL = ddot_vfp.S
CDOTKERNEL = cdot_vfp.S
ZDOTKERNEL = zdot_vfp.S
SNRM2KERNEL = nrm2_vfpv3.S SNRM2KERNEL = nrm2_vfpv3.S
DNRM2KERNEL = nrm2_vfpv3.S
CNRM2KERNEL = nrm2_vfpv3.S
ZNRM2KERNEL = nrm2_vfpv3.S
SROTKERNEL = rot_vfp.S SROTKERNEL = rot_vfp.S
DROTKERNEL = rot_vfp.S
CROTKERNEL = rot_vfp.S
ZROTKERNEL = rot_vfp.S
SSCALKERNEL = scal.c
DSCALKERNEL = scal.c
CSCALKERNEL = zscal.c
ZSCALKERNEL = zscal.c
SGEMVNKERNEL = gemv_n_vfpv3.S SGEMVNKERNEL = gemv_n_vfpv3.S
DGEMVNKERNEL = gemv_n_vfpv3.S
CGEMVNKERNEL = cgemv_n_vfp.S
ZGEMVNKERNEL = zgemv_n_vfp.S
SGEMVTKERNEL = gemv_t_vfp.S SGEMVTKERNEL = gemv_t_vfp.S
DGEMVTKERNEL = gemv_t_vfp.S
CGEMVTKERNEL = cgemv_t_vfp.S
ZGEMVTKERNEL = zgemv_t_vfp.S
STRMMKERNEL = strmm_kernel_4x4_vfpv3.S STRMMKERNEL = strmm_kernel_4x4_vfpv3.S
DTRMMKERNEL = dtrmm_kernel_4x4_vfpv3.S
CTRMMKERNEL = ctrmm_kernel_2x2_vfpv3.S
ZTRMMKERNEL = ztrmm_kernel_2x2_vfpv3.S
SGEMMKERNEL = sgemm_kernel_4x4_vfpv3.S SGEMMKERNEL = sgemm_kernel_4x4_vfpv3.S
SGEMMONCOPY = sgemm_ncopy_4_vfp.S SGEMMONCOPY = sgemm_ncopy_4_vfp.S
@ -93,11 +44,153 @@ SGEMMOTCOPY = sgemm_tcopy_4_vfp.S
SGEMMONCOPYOBJ = sgemm_oncopy.o SGEMMONCOPYOBJ = sgemm_oncopy.o
SGEMMOTCOPYOBJ = sgemm_otcopy.o SGEMMOTCOPYOBJ = sgemm_otcopy.o
###############################################################
# Files yet to be implemented in ARM Assembly for softfp ABI #
###############################################################
ifeq ($(ARM_SOFTFP_ABI), 1)
DAMAXKERNEL = amax.c
CAMAXKERNEL = zamax.c
ZAMAXKERNEL = zamax.c
DAMINKERNEL = amin.c
CAMINKERNEL = zamin.c
ZAMINKERNEL = zamin.c
DMAXKERNEL = max.c
DMINKERNEL = min.c
IDAMAXKERNEL = iamax.c
ICAMAXKERNEL = izamax.c
IZAMAXKERNEL = izamax.c
IDAMINKERNEL = iamin.c
ICAMINKERNEL = izamin.c
IZAMINKERNEL = izamin.c
IDMAXKERNEL = imax.c
IDMINKERNEL = imin.c
DSWAPKERNEL = swap.c
CSWAPKERNEL = zswap.c
ZSWAPKERNEL = zswap.c
DASUMKERNEL = asum.c
CASUMKERNEL = zasum.c
ZASUMKERNEL = zasum.c
DAXPYKERNEL = axpy.c
CAXPYKERNEL = zaxpy.c
ZAXPYKERNEL = zaxpy.c
DDOTKERNEL = dot.c
CDOTKERNEL = zdot.c
ZDOTKERNEL = zdot.c
DNRM2KERNEL = nrm2.c
CNRM2KERNEL = znrm2.c
ZNRM2KERNEL = znrm2.c
DROTKERNEL = rot.c
CROTKERNEL = zrot.c
ZROTKERNEL = zrot.c
DGEMVNKERNEL = gemv_n.c
CGEMVNKERNEL = zgemv_n.c
ZGEMVNKERNEL = zgemv_n.c
DGEMVTKERNEL = gemv_t.c
CGEMVTKERNEL = zgemv_t.c
ZGEMVTKERNEL = zgemv_t.c
DTRMMKERNEL = ../generic/trmmkernel_2x2.c
CTRMMKERNEL = ../generic/ztrmmkernel_2x2.c
ZTRMMKERNEL = ../generic/ztrmmkernel_2x2.c
DGEMMKERNEL = ../generic/gemmkernel_2x2.c
DGEMMONCOPY = ../generic/gemm_ncopy_2.c
DGEMMOTCOPY = ../generic/gemm_tcopy_2.c
DGEMMONCOPYOBJ = dgemm_oncopy.o
DGEMMOTCOPYOBJ = dgemm_otcopy.o
CGEMMKERNEL = ../generic/zgemmkernel_2x2.c
CGEMMONCOPY = ../generic/zgemm_ncopy_2.c
CGEMMOTCOPY = ../generic/zgemm_tcopy_2.c
CGEMMONCOPYOBJ = cgemm_oncopy.o
CGEMMOTCOPYOBJ = cgemm_otcopy.o
ZGEMMKERNEL = ../generic/zgemmkernel_2x2.c
ZGEMMONCOPY = ../generic/zgemm_ncopy_2.c
ZGEMMOTCOPY = ../generic/zgemm_tcopy_2.c
ZGEMMONCOPYOBJ = zgemm_oncopy.o
ZGEMMOTCOPYOBJ = zgemm_otcopy.o
else # Use optimized hard-float implementations
DAMAXKERNEL = iamax_vfp.S
CAMAXKERNEL = iamax_vfp.S
ZAMAXKERNEL = iamax_vfp.S
DAMINKERNEL = iamax_vfp.S
CAMINKERNEL = iamax_vfp.S
ZAMINKERNEL = iamax_vfp.S
DMAXKERNEL = iamax_vfp.S
DMINKERNEL = iamax_vfp.S
IDAMAXKERNEL = iamax_vfp.S
ICAMAXKERNEL = iamax_vfp.S
IZAMAXKERNEL = iamax_vfp.S
IDAMINKERNEL = iamax_vfp.S
ICAMINKERNEL = iamax_vfp.S
IZAMINKERNEL = iamax_vfp.S
IDMAXKERNEL = iamax_vfp.S
IDMINKERNEL = iamax_vfp.S
DSWAPKERNEL = swap_vfp.S
CSWAPKERNEL = swap_vfp.S
ZSWAPKERNEL = swap_vfp.S
DASUMKERNEL = asum_vfp.S
CASUMKERNEL = asum_vfp.S
ZASUMKERNEL = asum_vfp.S
DAXPYKERNEL = axpy_vfp.S
CAXPYKERNEL = axpy_vfp.S
ZAXPYKERNEL = axpy_vfp.S
DDOTKERNEL = ddot_vfp.S
CDOTKERNEL = cdot_vfp.S
ZDOTKERNEL = zdot_vfp.S
DNRM2KERNEL = nrm2_vfpv3.S
CNRM2KERNEL = nrm2_vfpv3.S
ZNRM2KERNEL = nrm2_vfpv3.S
DROTKERNEL = rot_vfp.S
CROTKERNEL = rot_vfp.S
ZROTKERNEL = rot_vfp.S
DGEMVNKERNEL = gemv_n_vfpv3.S
CGEMVNKERNEL = cgemv_n_vfp.S
ZGEMVNKERNEL = zgemv_n_vfp.S
DGEMVTKERNEL = gemv_t_vfp.S
CGEMVTKERNEL = cgemv_t_vfp.S
ZGEMVTKERNEL = zgemv_t_vfp.S
DTRMMKERNEL = dtrmm_kernel_4x4_vfpv3.S
CTRMMKERNEL = ctrmm_kernel_2x2_vfpv3.S
ZTRMMKERNEL = ztrmm_kernel_2x2_vfpv3.S
DGEMMKERNEL = dgemm_kernel_4x4_vfpv3.S DGEMMKERNEL = dgemm_kernel_4x4_vfpv3.S
DGEMMONCOPY = dgemm_ncopy_4_vfp.S DGEMMONCOPY = dgemm_ncopy_4_vfp.S
DGEMMOTCOPY = dgemm_tcopy_4_vfp.S DGEMMOTCOPY = dgemm_tcopy_4_vfp.S
DGEMMONCOPYOBJ = dgemm_oncopy.o
DGEMMOTCOPYOBJ = dgemm_otcopy.o
CGEMMKERNEL = cgemm_kernel_2x2_vfpv3.S CGEMMKERNEL = cgemm_kernel_2x2_vfpv3.S
CGEMMONCOPY = cgemm_ncopy_2_vfp.S CGEMMONCOPY = cgemm_ncopy_2_vfp.S
@ -111,6 +204,21 @@ ZGEMMOTCOPY = zgemm_tcopy_2_vfp.S
ZGEMMONCOPYOBJ = zgemm_oncopy.o ZGEMMONCOPYOBJ = zgemm_oncopy.o
ZGEMMOTCOPYOBJ = zgemm_otcopy.o ZGEMMOTCOPYOBJ = zgemm_otcopy.o
endif
############################
# Files in pure C routines #
############################
SCOPYKERNEL = copy.c
DCOPYKERNEL = copy.c
CCOPYKERNEL = zcopy.c
ZCOPYKERNEL = zcopy.c
SSCALKERNEL = scal.c
DSCALKERNEL = scal.c
CSCALKERNEL = zscal.c
ZSCALKERNEL = zscal.c
STRSMKERNEL_LN = ../generic/trsm_kernel_LN.c STRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
STRSMKERNEL_LT = ../generic/trsm_kernel_LT.c STRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
STRSMKERNEL_RN = ../generic/trsm_kernel_RN.c STRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
@ -130,5 +238,3 @@ ZTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c