@@ -96,27 +96,47 @@ CGEMVTKERNEL = zgemv_t_rvv.c
9696ZGEMVTKERNEL = zgemv_t_rvv.c
9797
9898SGEMMKERNEL = sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N)_zvl128b.c
99+ ifneq ($(filter $(SGEMM_UNROLL_N),4 8 16),)
100+ SGEMMONCOPY = gemm_ncopy_$(SGEMM_UNROLL_N)_rvv.c
101+ SGEMMOTCOPY = gemm_tcopy_$(SGEMM_UNROLL_N)_rvv.c
102+ else
99103SGEMMONCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c
100104SGEMMOTCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c
105+ endif
101106SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX)
102107SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)
103108
104109ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
110+ ifneq ($(filter $(SGEMM_UNROLL_M),4 8 16),)
111+ SGEMMINCOPY = gemm_ncopy_$(SGEMM_UNROLL_M)_rvv.c
112+ SGEMMITCOPY = gemm_tcopy_$(SGEMM_UNROLL_M)_rvv.c
113+ else
105114SGEMMINCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
106115SGEMMITCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
116+ endif
107117SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX)
108118SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX)
109119endif
110120
111121DGEMMKERNEL = dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N)_zvl128b.c
122+ ifneq ($(filter $(DGEMM_UNROLL_N),4 8 16),)
123+ DGEMMONCOPY = gemm_ncopy_$(DGEMM_UNROLL_N)_rvv.c
124+ DGEMMOTCOPY = gemm_tcopy_$(DGEMM_UNROLL_N)_rvv.c
125+ else
112126DGEMMONCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c
113127DGEMMOTCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_N).c
128+ endif
114129DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
115130DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
116131
117132ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N))
133+ ifneq ($(filter $(DGEMM_UNROLL_M),4 8 16),)
134+ DGEMMINCOPY = gemm_ncopy_$(DGEMM_UNROLL_M)_rvv.c
135+ DGEMMITCOPY = gemm_tcopy_$(DGEMM_UNROLL_M)_rvv.c
136+ else
118137DGEMMINCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_M).c
119138DGEMMITCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_M).c
139+ endif
120140DGEMMINCOPYOBJ = dgemm_incopy$(TSUFFIX).$(SUFFIX)
121141DGEMMITCOPYOBJ = dgemm_itcopy$(TSUFFIX).$(SUFFIX)
122142endif
@@ -255,4 +275,4 @@ SHGEMMOTCOPYOBJ = shgemm_otcopy$(TSUFFIX).$(SUFFIX)
255275ifndef SHGEMM_BETA
256276SHGEMM_BETA = gemm_beta_rvv.c
257277endif
258- endif
278+ endif
0 commit comments