@@ -3,57 +3,83 @@ ifndef NO_LASX
33SDOTKERNEL = dot_lasx.S
44DSDOTKERNEL = dot_lasx.S
55DDOTKERNEL = dot_lasx.S
6+ CDOTKERNEL = cdot_lasx.S
7+ ZDOTKERNEL = cdot_lasx.S
68
7- SSCALKERNEL = sscal_lasx.S
8- DSCALKERNEL = dscal_lasx.S
9+ SSCALKERNEL = scal_lasx.S
10+ DSCALKERNEL = scal_lasx.S
11+ CSCALKERNEL = cscal_lasx.S
12+ ZSCALKERNEL = cscal_lasx.S
913
10- SAMAXKERNEL = samax_lasx.S
11- DAMAXKERNEL = damax_lasx.S
14+ SAMAXKERNEL = amax_lasx.S
15+ DAMAXKERNEL = amax_lasx.S
16+ CAMAXKERNEL = camax_lasx.S
1217
13- SAMINKERNEL = samin_lasx.S
14- DAMINKERNEL = damin_lasx.S
18+ SAMINKERNEL = amin_lasx.S
19+ DAMINKERNEL = amin_lasx.S
20+ CAMINKERNEL = camin_lasx.S
1521
16- SMAXKERNEL = smax_lasx .S
17- DMAXKERNEL = dmax_lasx .S
22+ SMAXKERNEL = max_lsx .S
23+ DMAXKERNEL = max_lsx .S
1824
19- SMINKERNEL = smin_lasx .S
20- DMINKERNEL = dmin_lasx .S
25+ SMINKERNEL = min_lsx .S
26+ DMINKERNEL = min_lsx .S
2127
22- ISMAXKERNEL = ismax_lasx .S
23- IDMAXKERNEL = idmax_lasx .S
28+ ISMAXKERNEL = imax_lasx .S
29+ IDMAXKERNEL = imax_lasx .S
2430
25- ISMINKERNEL = ismin_lasx .S
26- IDMINKERNEL = idmin_lasx .S
31+ ISMINKERNEL = imin_lasx .S
32+ IDMINKERNEL = imin_lasx .S
2733
28- ISAMAXKERNEL = isamax_lasx.S
29- IDAMAXKERNEL = idamax_lasx.S
34+ ISAMAXKERNEL = iamax_lasx.S
35+ IDAMAXKERNEL = iamax_lasx.S
36+ ICAMAXKERNEL = icamax_lasx.S
37+ IZAMAXKERNEL = icamax_lasx.S
3038
31- ISAMINKERNEL = isamin_lasx.S
32- IDAMINKERNEL = idamin_lasx.S
39+ ISAMINKERNEL = iamin_lasx.S
40+ IDAMINKERNEL = iamin_lasx.S
41+ ICAMINKERNEL = icamin_lasx.S
42+ IZAMINKERNEL = icamin_lasx.S
3343
34- SCOPYKERNEL = scopy_lasx.S
35- DCOPYKERNEL = dcopy_lasx.S
44+ SCOPYKERNEL = copy_lasx.S
45+ DCOPYKERNEL = copy_lasx.S
46+ CCOPYKERNEL = ccopy_lasx.S
47+ ZCOPYKERNEL = ccopy_lasx.S
3648
37- SSWAPKERNEL = sswap_lasx .S
38- DSWAPKERNEL = dswap_lasx .S
49+ SSWAPKERNEL = swap_lasx .S
50+ DSWAPKERNEL = swap_lasx .S
3951
40- SAXPYKERNEL = saxpy_lasx.S
41- DAXPYKERNEL = daxpy_lasx.S
52+ SAXPYKERNEL = axpy_lasx.S
53+ DAXPYKERNEL = axpy_lasx.S
54+ CAXPYKERNEL = caxpy_lasx.S
55+ ZAXPYKERNEL = caxpy_lasx.S
4256
43- SAXPBYKERNEL = saxpby_lasx .S
44- DAXPBYKERNEL = daxpby_lasx .S
57+ SAXPBYKERNEL = axpby_lasx .S
58+ DAXPBYKERNEL = axpby_lasx .S
4559
46- SSUMKERNEL = ssum_lasx .S
47- DSUMKERNEL = dsum_lasx .S
60+ SSUMKERNEL = sum_lasx .S
61+ DSUMKERNEL = sum_lasx .S
4862
49- SASUMKERNEL = sasum_lasx.S
50- DASUMKERNEL = dasum_lasx.S
63+ SASUMKERNEL = asum_lasx.S
64+ DASUMKERNEL = asum_lasx.S
65+ CASUMKERNEL = casum_lasx.S
66+ ZASUMKERNEL = casum_lasx.S
5167
52- SROTKERNEL = srot_lasx.S
53- DROTKERNEL = drot_lasx.S
68+ SROTKERNEL = rot_lasx.S
69+ DROTKERNEL = rot_lasx.S
70+ CROTKERNEL = crot_lasx.S
71+ ZROTKERNEL = crot_lasx.S
5472
5573SNRM2KERNEL = snrm2_lasx.S
5674DNRM2KERNEL = dnrm2_lasx.S
75+ CNRM2KERNEL = cnrm2_lasx.S
76+ ZNRM2KERNEL = znrm2_lasx.S
77+
78+ CSWAPKERNEL = cswap_lasx.S
79+ ZSWAPKERNEL = cswap_lasx.S
80+
81+ CSUMKERNEL = csum_lasx.S
82+ ZSUMKERNEL = csum_lasx.S
5783
5884DGEMMKERNEL = dgemm_kernel_16x4.S
5985DGEMMINCOPY = dgemm_ncopy_16.S
@@ -81,13 +107,35 @@ SGEMMOTCOPYOBJ = sgemm_otcopy$(TSUFFIX).$(SUFFIX)
81107SGEMVNKERNEL = sgemv_n_8_lasx.S
82108SGEMVTKERNEL = sgemv_t_8_lasx.S
83109
110+ CGEMMKERNEL = cgemm_kernel_2x2_lsx.S
111+ CGEMMONCOPY = cgemm_ncopy_2_lsx.S
112+ CGEMMOTCOPY = cgemm_tcopy_2_lsx.S
113+ CGEMMONCOPYOBJ = cgemm_oncopy$(TSUFFIX).$(SUFFIX)
114+ CGEMMOTCOPYOBJ = cgemm_otcopy$(TSUFFIX).$(SUFFIX)
115+
116+ CTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
117+ CTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
118+ CTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
119+ CTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
120+
121+ ZGEMMKERNEL = zgemm_kernel_2x2_lasx.S
122+ ZGEMMONCOPY = zgemm_ncopy_2_lasx.S
123+ ZGEMMOTCOPY = zgemm_tcopy_2_lasx.S
124+ ZGEMMONCOPYOBJ = zgemm_oncopy$(TSUFFIX).$(SUFFIX)
125+ ZGEMMOTCOPYOBJ = zgemm_otcopy$(TSUFFIX).$(SUFFIX)
126+
127+ ZTRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
128+ ZTRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
129+ ZTRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
130+ ZTRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
131+
84132DTRSMKERNEL_LN = dtrsm_kernel_LN_16x4_lasx.S
85133DTRSMKERNEL_LT = dtrsm_kernel_LT_16x4_lasx.S
86134DTRSMKERNEL_RN = dtrsm_kernel_RN_16x4_lasx.S
87135DTRSMKERNEL_RT = dtrsm_kernel_RT_16x4_lasx.S
88- endif
89136
90137STRSMKERNEL_LN = ../generic/trsm_kernel_LN.c
91138STRSMKERNEL_LT = ../generic/trsm_kernel_LT.c
92139STRSMKERNEL_RN = ../generic/trsm_kernel_RN.c
93140STRSMKERNEL_RT = ../generic/trsm_kernel_RT.c
141+ endif
0 commit comments