The attention adjoint produces dQ and dK as fp32
[head][position][component], and dV transposed as fp32
[head][component][position]. Apply the inverse rotary transform to Q/K,
merge all three planes into the projection's token-major QKV order, and
pack the result to fp16 for the QKV input/weight GEMMs. Pointer ABI:
0x160 output QKV, 0x168 dQ, 0x170 dK, 0x178 cos, 0x180 sin, 0x188 dV^T.
917def attentionHeadLayoutSM86InverseRoPEQKVMergeProgram : (family SM86Program) =
918 (attentionHeadLayoutSM86Prologue
919 (byte 4)
920 attentionHeadLayoutSM86U15
921 (attentionHeadLayoutSM86Next
922 (attentionHeadLayoutSM86IMAD (byte 6) (byte 3) attentionHeadLayoutSM86U2 (byte 7))
923 (attentionHeadLayoutSM86Next
924 (attentionHeadLayoutSM86IMAD (byte 8) (byte 2) attentionHeadLayoutSM86U65536 (byte 7))
925 (attentionHeadLayoutSM86Next
926 (attentionHeadLayoutSM86IMAD (byte 8) (byte 1) attentionHeadLayoutSM86U64 (byte 8))
927 (attentionHeadLayoutSM86Next
928 (attentionHeadLayoutSM86IMAD (byte 8) (byte 6) attentionHeadLayoutSM86U1 (byte 8))
929 (attentionHeadLayoutSM86Next
930 (attentionHeadLayoutSM86Wide
931 (byte 10)
932 (byte 8)
933 (byte 4)
934 attentionHeadLayoutSM86SecondOutputPointer)
935 (attentionHeadLayoutSM86Next
936 (attentionHeadLayoutSM86Load (byte 12) (byte 10) attentionHeadLayoutSM86U0)
937 (attentionHeadLayoutSM86Next
938 (attentionHeadLayoutSM86Load (byte 13) (byte 10) attentionHeadLayoutSM86U4)
939 (attentionHeadLayoutSM86Next
940 (attentionHeadLayoutSM86Load (byte 14) (byte 10) attentionHeadLayoutSM86U128)
941 (attentionHeadLayoutSM86Next
942 (attentionHeadLayoutSM86Load (byte 15) (byte 10) attentionHeadLayoutSM86U132)
943 (attentionHeadLayoutSM86Next
944 (attentionHeadLayoutSM86Wide
945 (byte 16)
946 (byte 8)
947 (byte 4)
948 attentionHeadLayoutSM86InputPointer)
949 (attentionHeadLayoutSM86Next
950 (attentionHeadLayoutSM86Load
951 (byte 18)
952 (byte 16)
953 attentionHeadLayoutSM86U0)
954 (attentionHeadLayoutSM86Next
955 (attentionHeadLayoutSM86Load
956 (byte 19)
957 (byte 16)
958 attentionHeadLayoutSM86U4)
959 (attentionHeadLayoutSM86Next
960 (attentionHeadLayoutSM86Load
961 (byte 20)
962 (byte 16)
963 attentionHeadLayoutSM86U128)
964 (attentionHeadLayoutSM86Next
965 (attentionHeadLayoutSM86Load
966 (byte 21)
967 (byte 16)
968 attentionHeadLayoutSM86U132)
969 (attentionHeadLayoutSM86Next
970 (attentionHeadLayoutSM86IMAD
971 (byte 22)
972 (byte 1)
973 attentionHeadLayoutSM86U32
974 (byte 6))
975 (attentionHeadLayoutSM86Next
976 (attentionHeadLayoutSM86Wide
977 (byte 24)
978 (byte 22)
979 (byte 4)
980 attentionHeadLayoutSM86CosinePointer)
981 (attentionHeadLayoutSM86Next
982 (attentionHeadLayoutSM86Load
983 (byte 26)
984 (byte 24)
985 attentionHeadLayoutSM86U0)
986 (attentionHeadLayoutSM86Next
987 (attentionHeadLayoutSM86Load
988 (byte 27)
989 (byte 24)
990 attentionHeadLayoutSM86U4)
991 (attentionHeadLayoutSM86Next
992 (attentionHeadLayoutSM86Wide
993 (byte 28)
994 (byte 22)
995 (byte 4)
996 attentionHeadLayoutSM86SinePointer)
997 (attentionHeadLayoutSM86Next
998 (attentionHeadLayoutSM86Load
999 (byte 30)
1000 (byte 28)
1001 attentionHeadLayoutSM86U0)
1002 (attentionHeadLayoutSM86Next
1003 (attentionHeadLayoutSM86Load
1004 (byte 31)
1005 (byte 28)
1006 attentionHeadLayoutSM86U4)
1007 (attentionHeadLayoutSM86Next
1008 (attentionHeadLayoutSM86NegateWithControl
1009 (byte 30)
1010 attentionHeadLayoutSM86WaitLoads)
1011 (attentionHeadLayoutSM86Next
1012 (attentionHeadLayoutSM86Negate (byte 31))
1013 (attentionHeadLayoutSM86Rotate
1014 (byte 32)
1015 (byte 33)
1016 (byte 12)
1017 (byte 14)
1018 (byte 26)
1019 (byte 30)
1020 attentionHeadLayoutSM86WaitLoads
1021 (attentionHeadLayoutSM86Rotate
1022 (byte 34)
1023 (byte 35)
1024 (byte 13)
1025 (byte 15)
1026 (byte 27)
1027 (byte 31)
1028 sm86SafeControl
1029 (attentionHeadLayoutSM86Next
1030 (attentionHeadLayoutSM86Pack
1031 (byte 36)
1032 (byte 34)
1033 (byte 32)
1034 sm86SafeControl)
1035 (attentionHeadLayoutSM86Next
1036 (attentionHeadLayoutSM86Pack
1037 (byte 37)
1038 (byte 35)
1039 (byte 33)
1040 sm86SafeControl)
1041 (attentionHeadLayoutSM86Rotate
1042 (byte 38)
1043 (byte 39)
1044 (byte 18)
1045 (byte 20)
1046 (byte 26)
1047 (byte 30)
1048 sm86SafeControl
1049 (attentionHeadLayoutSM86Rotate
1050 (byte 40)
1051 (byte 41)
1052 (byte 19)
1053 (byte 21)
1054 (byte 27)
1055 (byte 31)
1056 sm86SafeControl
1057 (attentionHeadLayoutSM86Next
1058 (attentionHeadLayoutSM86Pack
1059 (byte 42)
1060 (byte 40)
1061 (byte 38)
1062 sm86SafeControl)
1063 (attentionHeadLayoutSM86Next
1064 (attentionHeadLayoutSM86Pack
1065 (byte 43)
1066 (byte 41)
1067 (byte 39)
1068 sm86SafeControl)
1069 (attentionHeadLayoutSM86Next
1070 (attentionHeadLayoutSM86IMAD
1071 (byte 44)
1072 (byte 1)
1073 attentionHeadLayoutSM86U1536
1074 (byte 7))
1075 (attentionHeadLayoutSM86Next
1076 (attentionHeadLayoutSM86IMAD
1077 (byte 44)
1078 (byte 2)
1079 attentionHeadLayoutSM86U64
1080 (byte 44))
1081 (attentionHeadLayoutSM86Next
1082 (attentionHeadLayoutSM86IMAD
1083 (byte 44)
1084 (byte 6)
1085 attentionHeadLayoutSM86U1
1086 (byte 44))
1087 (attentionHeadLayoutSM86Next
1088 (attentionHeadLayoutSM86Wide
1089 (byte 46)
1090 (byte 44)
1091 (byte 9)
1092 attentionHeadLayoutSM86FirstOutputPointer)
1093 (attentionHeadLayoutSM86Next
1094 (attentionHeadLayoutSM86Store
1095 (byte 46)
1096 (byte 36)
1097 attentionHeadLayoutSM86U0)
1098 (attentionHeadLayoutSM86Next
1099 (attentionHeadLayoutSM86Store
1100 (byte 46)
1101 (byte 37)
1102 attentionHeadLayoutSM86U64)
1103 (attentionHeadLayoutSM86Next
1104 (attentionHeadLayoutSM86Store
1105 (byte 46)
1106 (byte 42)
1107 attentionHeadLayoutSM86U1024)
1108 (attentionHeadLayoutSM86Next
1109 (attentionHeadLayoutSM86Store
1110 (byte 46)
1111 (byte 43)
1112 attentionHeadLayoutSM86U1088)
1113 (attentionHeadLayoutSM86Next
1114 (attentionHeadLayoutSM86IMAD
1115 (byte 48)
1116 (byte 2)
1117 attentionHeadLayoutSM86U65536
1118 (byte 7))
1119 (attentionHeadLayoutSM86Next
1120 (attentionHeadLayoutSM86IMAD
1121 (byte 48)
1122 (byte 3)
1123 attentionHeadLayoutSM86U2048
1124 (byte 48))
1125 (attentionHeadLayoutSM86Next
1126 (attentionHeadLayoutSM86IMAD
1127 (byte 48)
1128 (byte 1)
1129 attentionHeadLayoutSM86U1
1130 (byte 48))
1131 (attentionHeadLayoutSM86Next
1132 (attentionHeadLayoutSM86Wide
1133 (byte 50)
1134 (byte 48)
1135 (byte 4)
1136 attentionHeadLayoutSM86ValueGradientPointer)
1137 (attentionHeadLayoutSM86Next
1138 (attentionHeadLayoutSM86Load
1139 (byte 52)
1140 (byte 50)
1141 attentionHeadLayoutSM86U0)
1142 (attentionHeadLayoutSM86Next
1143 (attentionHeadLayoutSM86Load
1144 (byte 53)
1145 (byte 50)
1146 attentionHeadLayoutSM86U4096)
1147 (attentionHeadLayoutSM86Next
1148 (attentionHeadLayoutSM86Load
1149 (byte 55)
1150 (byte 50)
1151 attentionHeadLayoutSM86U131072)
1152 (attentionHeadLayoutSM86Next
1153 (attentionHeadLayoutSM86Load
1154 (byte 56)
1155 (byte 50)
1156 attentionHeadLayoutSM86U135168)
1157 (attentionHeadLayoutSM86Next
1158 (attentionHeadLayoutSM86Pack
1159 (byte 54)
1160 (byte 53)
1161 (byte 52)
1162 attentionHeadLayoutSM86WaitLoads)
1163 (attentionHeadLayoutSM86Next
1164 (attentionHeadLayoutSM86Pack
1165 (byte 58)
1166 (byte 56)
1167 (byte 55)
1168 sm86SafeControl)
1169 (attentionHeadLayoutSM86Next
1170 (attentionHeadLayoutSM86Store
1171 (byte 46)
1172 (byte 54)
1173 attentionHeadLayoutSM86U2048)
1174 (attentionHeadLayoutSM86Next
1175 (attentionHeadLayoutSM86Store
1176 (byte 46)
1177 (byte 58)
1178 attentionHeadLayoutSM86U2112)
1179 (attentionHeadLayoutSM86Next
1180 (constructor SM86InstructionBody SM86Exit sm86SafeControl)
1181 attentionHeadLayoutSM86End)))))))))))))))))))))))))))))))))))))))))))))))))))))The compiler supplied declaration spans and resolved links from this source snapshot. This page does not assert that this file belongs to a checked closure.