@@ -60,32 +60,55 @@ DGEMVTKERNEL = gemv_t.S
6060CGEMVTKERNEL = zgemv_t.S
6161ZGEMVTKERNEL = zgemv_t.S
6262
63- STRMMKERNEL = strmm_kernel_4x4.S
64- DTRMMKERNEL = dtrmm_kernel_4x4.S
65- CTRMMKERNEL = ctrmm_kernel_4x4.S
66- ZTRMMKERNEL = ztrmm_kernel_4x4.S
67-
68- SGEMMKERNEL = sgemm_kernel_4x4.S
69- SGEMMONCOPY = ../generic/gemm_ncopy_4.c
70- SGEMMOTCOPY = ../generic/gemm_tcopy_4.c
63+ SGEMMKERNEL = sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
64+ STRMMKERNEL = strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
65+ ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
66+ SGEMMINCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
67+ SGEMMITCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
68+ SGEMMINCOPYOBJ = sgemm_incopy.o
69+ SGEMMITCOPYOBJ = sgemm_itcopy.o
70+ endif
71+ SGEMMONCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c
72+ SGEMMOTCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c
7173SGEMMONCOPYOBJ = sgemm_oncopy.o
7274SGEMMOTCOPYOBJ = sgemm_otcopy.o
7375
74- DGEMMKERNEL = dgemm_kernel_4x4.S
75- DGEMMONCOPY = ../generic/gemm_ncopy_4.c
76- DGEMMOTCOPY = ../generic/gemm_tcopy_4.c
76+ DGEMMKERNEL = dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S
77+ DTRMMKERNEL = dtrmm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S
78+ ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N))
79+ DGEMMINCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_M).c
80+ DGEMMITCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_M).c
81+ DGEMMINCOPYOBJ = dgemm_incopy.o
82+ DGEMMITCOPYOBJ = dgemm_itcopy.o
83+ endif
84+ DGEMMONCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c
85+ DGEMMOTCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_N).c
7786DGEMMONCOPYOBJ = dgemm_oncopy.o
7887DGEMMOTCOPYOBJ = dgemm_otcopy.o
7988
80- CGEMMKERNEL = cgemm_kernel_4x4.S
81- CGEMMONCOPY = ../generic/zgemm_ncopy_4.c
82- CGEMMOTCOPY = ../generic/zgemm_tcopy_4.c
89+ CGEMMKERNEL = cgemm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
90+ CTRMMKERNEL = ctrmm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
91+ ifneq ($(CGEMM_UNROLL_M), $(CGEMM_UNROLL_N))
92+ CGEMMINCOPY = ../generic/zgemm_ncopy_$(CGEMM_UNROLL_M).c
93+ CGEMMITCOPY = ../generic/zgemm_tcopy_$(CGEMM_UNROLL_M).c
94+ CGEMMINCOPYOBJ = cgemm_incopy.o
95+ CGEMMITCOPYOBJ = cgemm_itcopy.o
96+ endif
97+ CGEMMONCOPY = ../generic/zgemm_ncopy_$(CGEMM_UNROLL_N).c
98+ CGEMMOTCOPY = ../generic/zgemm_tcopy_$(CGEMM_UNROLL_N).c
8399CGEMMONCOPYOBJ = cgemm_oncopy.o
84100CGEMMOTCOPYOBJ = cgemm_otcopy.o
85101
86- ZGEMMKERNEL = zgemm_kernel_4x4.S
87- ZGEMMONCOPY = ../generic/zgemm_ncopy_4.c
88- ZGEMMOTCOPY = ../generic/zgemm_tcopy_4.c
102+ ZGEMMKERNEL = zgemm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S
103+ ZTRMMKERNEL = ztrmm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S
104+ ifneq ($(ZGEMM_UNROLL_M), $(ZGEMM_UNROLL_N))
105+ ZGEMMINCOPY = ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_M).c
106+ ZGEMMITCOPY = ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_M).c
107+ ZGEMMINCOPYOBJ = zgemm_incopy.o
108+ ZGEMMITCOPYOBJ = zgemm_itcopy.o
109+ endif
110+ ZGEMMONCOPY = ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_N).c
111+ ZGEMMOTCOPY = ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_N).c
89112ZGEMMONCOPYOBJ = zgemm_oncopy.o
90113ZGEMMOTCOPYOBJ = zgemm_otcopy.o
91114
0 commit comments