Skip to content

Commit 783ba80

Browse files
authored
HiSilicon tsv110 CPUs optimization branch
add HiSilicon tsv110 CPUs optimization branch
1 parent edb8143 commit 783ba80

File tree

1 file changed

+175
-0
lines changed

1 file changed

+175
-0
lines changed

kernel/arm64/KERNEL.TSV110

Lines changed: 175 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,175 @@
1+
SAMINKERNEL = ../arm/amin.c
2+
DAMINKERNEL = ../arm/amin.c
3+
CAMINKERNEL = ../arm/zamin.c
4+
ZAMINKERNEL = ../arm/zamin.c
5+
6+
SMAXKERNEL = ../arm/max.c
7+
DMAXKERNEL = ../arm/max.c
8+
9+
SMINKERNEL = ../arm/min.c
10+
DMINKERNEL = ../arm/min.c
11+
12+
ISAMINKERNEL = ../arm/iamin.c
13+
IDAMINKERNEL = ../arm/iamin.c
14+
ICAMINKERNEL = ../arm/izamin.c
15+
IZAMINKERNEL = ../arm/izamin.c
16+
17+
ISMAXKERNEL = ../arm/imax.c
18+
IDMAXKERNEL = ../arm/imax.c
19+
20+
ISMINKERNEL = ../arm/imin.c
21+
IDMINKERNEL = ../arm/imin.c
22+
23+
STRMMKERNEL = ../generic/trmmkernel_4x4.c
24+
DTRMMKERNEL = ../generic/trmmkernel_2x2.c
25+
CTRMMKERNEL = ../generic/ztrmmkernel_2x2.c
26+
ZTRMMKERNEL = ../generic/ztrmmkernel_2x2.c
27+
28+
STRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
29+
STRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
30+
STRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
31+
STRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
32+
33+
DTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
34+
DTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
35+
DTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
36+
DTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
37+
38+
CTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
39+
CTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
40+
CTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
41+
CTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
42+
43+
ZTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
44+
ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
45+
ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
46+
ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
47+
48+
SAMAXKERNEL = amax.S
49+
DAMAXKERNEL = amax.S
50+
CAMAXKERNEL = zamax.S
51+
ZAMAXKERNEL = zamax.S
52+
53+
ISAMAXKERNEL = iamax.S
54+
IDAMAXKERNEL = iamax.S
55+
ICAMAXKERNEL = izamax.S
56+
IZAMAXKERNEL = izamax.S
57+
58+
SASUMKERNEL = asum.S
59+
DASUMKERNEL = asum.S
60+
CASUMKERNEL = casum.S
61+
ZASUMKERNEL = zasum.S
62+
63+
SAXPYKERNEL = axpy.S
64+
DAXPYKERNEL = axpy.S
65+
CAXPYKERNEL = zaxpy.S
66+
ZAXPYKERNEL = zaxpy.S
67+
68+
SCOPYKERNEL = copy.S
69+
DCOPYKERNEL = copy.S
70+
CCOPYKERNEL = copy.S
71+
ZCOPYKERNEL = copy.S
72+
73+
SDOTKERNEL = dot.S
74+
DDOTKERNEL = dot.S
75+
CDOTKERNEL = zdot.S
76+
ZDOTKERNEL = zdot.S
77+
DSDOTKERNEL = dot.S
78+
79+
SNRM2KERNEL = nrm2.S
80+
DNRM2KERNEL = nrm2.S
81+
CNRM2KERNEL = znrm2.S
82+
ZNRM2KERNEL = znrm2.S
83+
84+
SROTKERNEL = rot.S
85+
DROTKERNEL = rot.S
86+
CROTKERNEL = zrot.S
87+
ZROTKERNEL = zrot.S
88+
89+
SSCALKERNEL = scal.S
90+
DSCALKERNEL = scal.S
91+
CSCALKERNEL = zscal.S
92+
ZSCALKERNEL = zscal.S
93+
94+
SSWAPKERNEL = swap.S
95+
DSWAPKERNEL = swap.S
96+
CSWAPKERNEL = swap.S
97+
ZSWAPKERNEL = swap.S
98+
99+
SGEMVNKERNEL = gemv_n.S
100+
DGEMVNKERNEL = gemv_n.S
101+
CGEMVNKERNEL = zgemv_n.S
102+
ZGEMVNKERNEL = zgemv_n.S
103+
104+
SGEMVTKERNEL = gemv_t.S
105+
DGEMVTKERNEL = gemv_t.S
106+
CGEMVTKERNEL = zgemv_t.S
107+
ZGEMVTKERNEL = zgemv_t.S
108+
109+
SGEMMKERNEL = sgemm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
110+
STRMMKERNEL = strmm_kernel_$(SGEMM_UNROLL_M)x$(SGEMM_UNROLL_N).S
111+
ifneq ($(SGEMM_UNROLL_M), $(SGEMM_UNROLL_N))
112+
SGEMMINCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_M).c
113+
SGEMMITCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_M).c
114+
SGEMMINCOPYOBJ = sgemm_incopy$(TSUFFIX).$(SUFFIX)
115+
SGEMMITCOPYOBJ = sgemm_itcopy$(TSUFFIX).$(SUFFIX)
116+
endif
117+
SGEMMONCOPY = ../generic/gemm_ncopy_$(SGEMM_UNROLL_N).c
118+
SGEMMOTCOPY = ../generic/gemm_tcopy_$(SGEMM_UNROLL_N).c
119+
SGEMMONCOPYOBJ = sgemm_oncopy$(TSUFFIX).$(SUFFIX)
120+
SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)
121+
122+
DGEMMKERNEL = dgemm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S
123+
DTRMMKERNEL = dtrmm_kernel_$(DGEMM_UNROLL_M)x$(DGEMM_UNROLL_N).S
124+
125+
ifneq ($(DGEMM_UNROLL_M), $(DGEMM_UNROLL_N))
126+
127+
ifeq ($(DGEMM_UNROLL_M), 8)
128+
DGEMMINCOPY = dgemm_ncopy_$(DGEMM_UNROLL_M).S
129+
DGEMMITCOPY = dgemm_tcopy_$(DGEMM_UNROLL_M).S
130+
else
131+
DGEMMINCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_M).c
132+
DGEMMITCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_M).c
133+
endif
134+
135+
DGEMMINCOPYOBJ = dgemm_incopy$(TSUFFIX).$(SUFFIX)
136+
DGEMMITCOPYOBJ = dgemm_itcopy$(TSUFFIX).$(SUFFIX)
137+
endif
138+
139+
ifeq ($(DGEMM_UNROLL_N), 4)
140+
DGEMMONCOPY = dgemm_ncopy_$(DGEMM_UNROLL_N).S
141+
DGEMMOTCOPY = dgemm_tcopy_$(DGEMM_UNROLL_N).S
142+
else
143+
DGEMMONCOPY = ../generic/gemm_ncopy_$(DGEMM_UNROLL_N).c
144+
DGEMMOTCOPY = ../generic/gemm_tcopy_$(DGEMM_UNROLL_N).c
145+
endif
146+
147+
DGEMMONCOPYOBJ = dgemm_oncopy$(TSUFFIX).$(SUFFIX)
148+
DGEMMOTCOPYOBJ = dgemm_otcopy$(TSUFFIX).$(SUFFIX)
149+
150+
CGEMMKERNEL = cgemm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
151+
CTRMMKERNEL = ctrmm_kernel_$(CGEMM_UNROLL_M)x$(CGEMM_UNROLL_N).S
152+
ifneq ($(CGEMM_UNROLL_M), $(CGEMM_UNROLL_N))
153+
CGEMMINCOPY = ../generic/zgemm_ncopy_$(CGEMM_UNROLL_M).c
154+
CGEMMITCOPY = ../generic/zgemm_tcopy_$(CGEMM_UNROLL_M).c
155+
CGEMMINCOPYOBJ = cgemm_incopy$(TSUFFIX).$(SUFFIX)
156+
CGEMMITCOPYOBJ = cgemm_itcopy$(TSUFFIX).$(SUFFIX)
157+
endif
158+
CGEMMONCOPY = ../generic/zgemm_ncopy_$(CGEMM_UNROLL_N).c
159+
CGEMMOTCOPY = ../generic/zgemm_tcopy_$(CGEMM_UNROLL_N).c
160+
CGEMMONCOPYOBJ = cgemm_oncopy$(TSUFFIX).$(SUFFIX)
161+
CGEMMOTCOPYOBJ = cgemm_otcopy$(TSUFFIX).$(SUFFIX)
162+
163+
ZGEMMKERNEL = zgemm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S
164+
ZTRMMKERNEL = ztrmm_kernel_$(ZGEMM_UNROLL_M)x$(ZGEMM_UNROLL_N).S
165+
ifneq ($(ZGEMM_UNROLL_M), $(ZGEMM_UNROLL_N))
166+
ZGEMMINCOPY = ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_M).c
167+
ZGEMMITCOPY = ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_M).c
168+
ZGEMMINCOPYOBJ = zgemm_incopy$(TSUFFIX).$(SUFFIX)
169+
ZGEMMITCOPYOBJ = zgemm_itcopy$(TSUFFIX).$(SUFFIX)
170+
endif
171+
ZGEMMONCOPY = ../generic/zgemm_ncopy_$(ZGEMM_UNROLL_N).c
172+
ZGEMMOTCOPY = ../generic/zgemm_tcopy_$(ZGEMM_UNROLL_N).c
173+
ZGEMMONCOPYOBJ = zgemm_oncopy$(TSUFFIX).$(SUFFIX)
174+
ZGEMMOTCOPYOBJ = zgemm_otcopy$(TSUFFIX).$(SUFFIX)
175+

0 commit comments

Comments
 (0)