coppelius S5 rung 4d: the softmax BACKWARD over 1024-wide rows, the adjoint
of SoftmaxSM86CausalForward1024 (256 threads x four elements, one block per
row): dS = P * (dP - sum(P * dP)). Probabilities P at the primary input
(0x168), adjoints dP at the secondary input (0x170), score adjoints to the
output (0x160). Masked entries carry P = 0 and so dS = 0 with no mask logic.
1234def softmaxSM86Backward1024Load =
1235 (lambda unrestricted element : (family SM86Register) .
1236 (lambda unrestricted address : (family SM86Register) .
1237 (lambda unrestricted offset : (family SM86Unsigned32) .
1238 (lambda unrestricted tail : (family SM86Program) .
1239 (softmaxSM86Next
1240 (constructor SM86InstructionBody SM86LoadGlobal element address offset softmaxSM86Set1)
1241 tail)))))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.