Join the eight fp32 head-major context matrices [head][position][64] into
one fp16 token-major matrix [position][512]. A 256-thread block owns one
position; each thread converts one adjacent component pair.
787def attentionHeadLayoutSM86HeadToTokenProgram : (family SM86Program) =
788 (attentionHeadLayoutSM86Prologue
789 (byte 5)
790 attentionHeadLayoutSM86U31
791 (attentionHeadLayoutSM86Next
792 (attentionHeadLayoutSM86IMAD (byte 6) (byte 3) attentionHeadLayoutSM86U2 (byte 7))
793 (attentionHeadLayoutSM86Next
794 (attentionHeadLayoutSM86IMAD (byte 8) (byte 2) attentionHeadLayoutSM86U65536 (byte 7))
795 (attentionHeadLayoutSM86Next
796 (attentionHeadLayoutSM86IMAD (byte 8) (byte 1) attentionHeadLayoutSM86U64 (byte 8))
797 (attentionHeadLayoutSM86Next
798 (attentionHeadLayoutSM86IMAD (byte 8) (byte 6) attentionHeadLayoutSM86U1 (byte 8))
799 (attentionHeadLayoutSM86Next
800 (attentionHeadLayoutSM86Wide
801 (byte 10)
802 (byte 8)
803 (byte 4)
804 attentionHeadLayoutSM86InputPointer)
805 (attentionHeadLayoutSM86Next
806 (attentionHeadLayoutSM86Load (byte 12) (byte 10) attentionHeadLayoutSM86U0)
807 (attentionHeadLayoutSM86Next
808 (attentionHeadLayoutSM86Load (byte 13) (byte 10) attentionHeadLayoutSM86U4)
809 (attentionHeadLayoutSM86Next
810 (attentionHeadLayoutSM86IMAD
811 (byte 16)
812 (byte 1)
813 attentionHeadLayoutSM86U512
814 (byte 7))
815 (attentionHeadLayoutSM86Next
816 (attentionHeadLayoutSM86IMAD
817 (byte 16)
818 (byte 2)
819 attentionHeadLayoutSM86U64
820 (byte 16))
821 (attentionHeadLayoutSM86Next
822 (attentionHeadLayoutSM86IMAD
823 (byte 16)
824 (byte 6)
825 attentionHeadLayoutSM86U1
826 (byte 16))
827 (attentionHeadLayoutSM86Next
828 (attentionHeadLayoutSM86Wide
829 (byte 18)
830 (byte 16)
831 (byte 9)
832 attentionHeadLayoutSM86FirstOutputPointer)
833 (attentionHeadLayoutSM86Next
834 (attentionHeadLayoutSM86Pack
835 (byte 14)
836 (byte 13)
837 (byte 12)
838 attentionHeadLayoutSM86WaitLoads)
839 (attentionHeadLayoutSM86Next
840 (attentionHeadLayoutSM86Store
841 (byte 18)
842 (byte 14)
843 attentionHeadLayoutSM86U0)
844 (attentionHeadLayoutSM86Next
845 (constructor SM86InstructionBody SM86Exit sm86SafeControl)
846 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.