35
35
#define UNALIASED_CUSTOM_BUILTIN (ID, TYPES, ACCUMULATE, FEATURE ) \
36
36
CUSTOM_BUILTIN (ID, ID, TYPES, ACCUMULATE, FEATURE)
37
37
38
+ #define UNALIASED_CUSTOM_ACCUMULATE_BUILTIN (ID, TYPES, FEATURE ) \
39
+ UNALIASED_CUSTOM_BUILTIN (ID, TYPES, false , FEATURE) \
40
+ UNALIASED_CUSTOM_BUILTIN(ID##nn, TYPES, true , FEATURE) \
41
+ UNALIASED_CUSTOM_BUILTIN(ID##np, TYPES, true , FEATURE) \
42
+ UNALIASED_CUSTOM_BUILTIN(ID##pn, TYPES, true , FEATURE) \
43
+ UNALIASED_CUSTOM_BUILTIN(ID##pp, TYPES, true , FEATURE)
44
+
38
45
// GCC predefined macros to rename builtins, undef them to keep original names.
39
46
#if defined(__GNUC__) && !defined(__clang__)
40
47
#undef __builtin_vsx_xvnmaddadp
@@ -1032,12 +1039,6 @@ UNALIASED_CUSTOM_BUILTIN(mma_xvi16ger2, "vW512*VV", false,
1032
1039
" mma,paired-vector-memops" )
1033
1040
UNALIASED_CUSTOM_BUILTIN(mma_xvi16ger2s, " vW512*VV" , false ,
1034
1041
" mma,paired-vector-memops" )
1035
- UNALIASED_CUSTOM_BUILTIN(mma_xvf16ger2, " vW512*VV" , false ,
1036
- " mma,paired-vector-memops" )
1037
- UNALIASED_CUSTOM_BUILTIN(mma_xvf32ger, " vW512*VV" , false ,
1038
- " mma,paired-vector-memops" )
1039
- UNALIASED_CUSTOM_BUILTIN(mma_xvf64ger, " vW512*W256V" , false ,
1040
- " mma,paired-vector-memops" )
1041
1042
UNALIASED_CUSTOM_BUILTIN(mma_pmxvi4ger8, " vW512*VVi15i15i255" , false ,
1042
1043
" mma,paired-vector-memops" )
1043
1044
UNALIASED_CUSTOM_BUILTIN(mma_pmxvi8ger4, " vW512*VVi15i15i15" , false ,
@@ -1046,12 +1047,6 @@ UNALIASED_CUSTOM_BUILTIN(mma_pmxvi16ger2, "vW512*VVi15i15i3", false,
1046
1047
" mma,paired-vector-memops" )
1047
1048
UNALIASED_CUSTOM_BUILTIN(mma_pmxvi16ger2s, " vW512*VVi15i15i3" , false ,
1048
1049
" mma,paired-vector-memops" )
1049
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf16ger2, " vW512*VVi15i15i3" , false ,
1050
- " mma,paired-vector-memops" )
1051
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf32ger, " vW512*VVi15i15" , false ,
1052
- " mma,paired-vector-memops" )
1053
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf64ger, " vW512*W256Vi15i3" , false ,
1054
- " mma,paired-vector-memops" )
1055
1050
UNALIASED_CUSTOM_BUILTIN(mma_xvi4ger8pp, " vW512*VV" , true ,
1056
1051
" mma,paired-vector-memops" )
1057
1052
UNALIASED_CUSTOM_BUILTIN(mma_xvi8ger4pp, " vW512*VV" , true ,
@@ -1072,85 +1067,33 @@ UNALIASED_CUSTOM_BUILTIN(mma_pmxvi16ger2pp, "vW512*VVi15i15i3", true,
1072
1067
" mma,paired-vector-memops" )
1073
1068
UNALIASED_CUSTOM_BUILTIN(mma_pmxvi16ger2spp, " vW512*VVi15i15i3" , true ,
1074
1069
" mma,paired-vector-memops" )
1075
- UNALIASED_CUSTOM_BUILTIN(mma_xvf16ger2pp, " vW512*VV" , true ,
1076
- " mma,paired-vector-memops" )
1077
- UNALIASED_CUSTOM_BUILTIN(mma_xvf16ger2pn, " vW512*VV" , true ,
1078
- " mma,paired-vector-memops" )
1079
- UNALIASED_CUSTOM_BUILTIN(mma_xvf16ger2np, " vW512*VV" , true ,
1080
- " mma,paired-vector-memops" )
1081
- UNALIASED_CUSTOM_BUILTIN(mma_xvf16ger2nn, " vW512*VV" , true ,
1082
- " mma,paired-vector-memops" )
1083
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf16ger2pp, " vW512*VVi15i15i3" , true ,
1084
- " mma,paired-vector-memops" )
1085
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf16ger2pn, " vW512*VVi15i15i3" , true ,
1086
- " mma,paired-vector-memops" )
1087
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf16ger2np, " vW512*VVi15i15i3" , true ,
1088
- " mma,paired-vector-memops" )
1089
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf16ger2nn, " vW512*VVi15i15i3" , true ,
1090
- " mma,paired-vector-memops" )
1091
- UNALIASED_CUSTOM_BUILTIN(mma_xvf32gerpp, " vW512*VV" , true ,
1092
- " mma,paired-vector-memops" )
1093
- UNALIASED_CUSTOM_BUILTIN(mma_xvf32gerpn, " vW512*VV" , true ,
1094
- " mma,paired-vector-memops" )
1095
- UNALIASED_CUSTOM_BUILTIN(mma_xvf32gernp, " vW512*VV" , true ,
1096
- " mma,paired-vector-memops" )
1097
- UNALIASED_CUSTOM_BUILTIN(mma_xvf32gernn, " vW512*VV" , true ,
1098
- " mma,paired-vector-memops" )
1099
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf32gerpp, " vW512*VVi15i15" , true ,
1100
- " mma,paired-vector-memops" )
1101
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf32gerpn, " vW512*VVi15i15" , true ,
1102
- " mma,paired-vector-memops" )
1103
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf32gernp, " vW512*VVi15i15" , true ,
1104
- " mma,paired-vector-memops" )
1105
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf32gernn, " vW512*VVi15i15" , true ,
1106
- " mma,paired-vector-memops" )
1107
- UNALIASED_CUSTOM_BUILTIN(mma_xvf64gerpp, " vW512*W256V" , true ,
1108
- " mma,paired-vector-memops" )
1109
- UNALIASED_CUSTOM_BUILTIN(mma_xvf64gerpn, " vW512*W256V" , true ,
1110
- " mma,paired-vector-memops" )
1111
- UNALIASED_CUSTOM_BUILTIN(mma_xvf64gernp, " vW512*W256V" , true ,
1112
- " mma,paired-vector-memops" )
1113
- UNALIASED_CUSTOM_BUILTIN(mma_xvf64gernn, " vW512*W256V" , true ,
1114
- " mma,paired-vector-memops" )
1115
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf64gerpp, " vW512*W256Vi15i3" , true ,
1116
- " mma,paired-vector-memops" )
1117
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf64gerpn, " vW512*W256Vi15i3" , true ,
1118
- " mma,paired-vector-memops" )
1119
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf64gernp, " vW512*W256Vi15i3" , true ,
1120
- " mma,paired-vector-memops" )
1121
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvf64gernn, " vW512*W256Vi15i3" , true ,
1122
- " mma,paired-vector-memops" )
1123
- UNALIASED_CUSTOM_BUILTIN(mma_xvbf16ger2, " vW512*VV" , false ,
1124
- " mma,paired-vector-memops" )
1125
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvbf16ger2, " vW512*VVi15i15i3" , false ,
1126
- " mma,paired-vector-memops" )
1127
- UNALIASED_CUSTOM_BUILTIN(mma_xvbf16ger2pp, " vW512*VV" , true ,
1070
+ UNALIASED_CUSTOM_BUILTIN(mma_dmxvi8gerx4, " vW1024*W256V" , false ,
1128
1071
" mma,paired-vector-memops" )
1129
- UNALIASED_CUSTOM_BUILTIN(mma_xvbf16ger2pn , " vW512*VV " , true ,
1072
+ UNALIASED_CUSTOM_BUILTIN(mma_pmdmxvi8gerx4 , " vW1024*W256Vi255i15i15 " , false ,
1130
1073
" mma,paired-vector-memops" )
1131
- UNALIASED_CUSTOM_BUILTIN(mma_xvbf16ger2np , " vW512*VV " , true ,
1074
+ UNALIASED_CUSTOM_BUILTIN(mma_dmxvi8gerx4pp , " vW1024*W256V " , true ,
1132
1075
" mma,paired-vector-memops" )
1133
- UNALIASED_CUSTOM_BUILTIN(mma_xvbf16ger2nn , " vW512*VV " , true ,
1076
+ UNALIASED_CUSTOM_BUILTIN(mma_pmdmxvi8gerx4pp , " vW1024*W256Vi255i15i15 " , true ,
1134
1077
" mma,paired-vector-memops" )
1135
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvbf16ger2pp, " vW512*VVi15i15i3 " , true ,
1078
+ UNALIASED_CUSTOM_BUILTIN(mma_dmxvi8gerx4spp, " vW1024*W256V " , true ,
1136
1079
" mma,paired-vector-memops" )
1137
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvbf16ger2pn , " vW512*VVi15i15i3 " , true ,
1080
+ UNALIASED_CUSTOM_BUILTIN(mma_pmdmxvi8gerx4spp , " vW1024*W256Vi255i15i15 " , true ,
1138
1081
" mma,paired-vector-memops" )
1139
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvbf16ger2np , " vW512*VVi15i15i3 " , true ,
1082
+ UNALIASED_CUSTOM_ACCUMULATE_BUILTIN(mma_xvf16ger2 , " vW512*VV " ,
1140
1083
" mma,paired-vector-memops" )
1141
- UNALIASED_CUSTOM_BUILTIN(mma_pmxvbf16ger2nn , " vW512*VVi15i15i3 " , true ,
1084
+ UNALIASED_CUSTOM_ACCUMULATE_BUILTIN(mma_xvf32ger , " vW512*VV " ,
1142
1085
" mma,paired-vector-memops" )
1143
- UNALIASED_CUSTOM_BUILTIN(mma_dmxvi8gerx4 , " vW1024 *W256V" , false ,
1086
+ UNALIASED_CUSTOM_ACCUMULATE_BUILTIN(mma_xvf64ger , " vW512 *W256V" ,
1144
1087
" mma,paired-vector-memops" )
1145
- UNALIASED_CUSTOM_BUILTIN(mma_pmdmxvi8gerx4 , " vW1024*W256Vi255i15i15 " , false ,
1088
+ UNALIASED_CUSTOM_ACCUMULATE_BUILTIN(mma_pmxvf16ger2 , " vW512*VVi15i15i3 " ,
1146
1089
" mma,paired-vector-memops" )
1147
- UNALIASED_CUSTOM_BUILTIN(mma_dmxvi8gerx4pp , " vW1024*W256V " , true ,
1090
+ UNALIASED_CUSTOM_ACCUMULATE_BUILTIN(mma_pmxvf32ger , " vW512*VVi15i15 " ,
1148
1091
" mma,paired-vector-memops" )
1149
- UNALIASED_CUSTOM_BUILTIN(mma_pmdmxvi8gerx4pp , " vW1024*W256Vi255i15i15 " , true ,
1092
+ UNALIASED_CUSTOM_ACCUMULATE_BUILTIN(mma_pmxvf64ger , " vW512*W256Vi15i3 " ,
1150
1093
" mma,paired-vector-memops" )
1151
- UNALIASED_CUSTOM_BUILTIN(mma_dmxvi8gerx4spp, " vW1024*W256V " , true ,
1094
+ UNALIASED_CUSTOM_ACCUMULATE_BUILTIN(mma_xvbf16ger2, " vW512*VV " ,
1152
1095
" mma,paired-vector-memops" )
1153
- UNALIASED_CUSTOM_BUILTIN(mma_pmdmxvi8gerx4spp , " vW1024*W256Vi255i15i15 " , true ,
1096
+ UNALIASED_CUSTOM_ACCUMULATE_BUILTIN(mma_pmxvbf16ger2 , " vW512*VVi15i15i3 " ,
1154
1097
" mma,paired-vector-memops" )
1155
1098
1156
1099
// FIXME: Obviously incomplete.
0 commit comments