update armv8sve + contributors

diff --git a/CONTRIBUTORS.md b/CONTRIBUTORS.md
index 879aaeb..5378c79 100644 (file)
--- a/CONTRIBUTORS.md
+++ b/CONTRIBUTORS.md
@@ -203,3 +203,4 @@ In chronological order:
   * [2021-11-20] Vector-length-agnostic Arm SVE copy routines for DGEMM, DTRMM, DSYMM
   * [2021-11-12] SVE kernels for SGEMM, STRMM and corresponding SVE copy functions
   * [2022-01-06] SVE kernels for CGEMM, ZGEMM, CTRMM, ZTRMM and corresponding SVE copy functions
+  * [2022-01-18] SVE kernels and copy functions for TRSM

diff --git a/kernel/arm64/KERNEL.ARMV8SVE b/kernel/arm64/KERNEL.ARMV8SVE
index 66de642..bd25f7c 100644 (file)
--- a/kernel/arm64/KERNEL.ARMV8SVE
+++ b/kernel/arm64/KERNEL.ARMV8SVE
@@ -20,25 +20,36 @@ 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
+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
@@ -140,8 +151,8 @@ DTRMMKERNEL    =  dtrmm_kernel_sve_v1x$(DGEMM_UNROLL_N).S
 
 DGEMMINCOPY    =  dgemm_ncopy_sve_v1.c
 DGEMMITCOPY    =  dgemm_tcopy_sve_v1.c
-DGEMMONCOPY    =  ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c
-DGEMMOTCOPY    =  ../generic/gemm_tcopy_$(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)