In the Cayman documentation:
This sentence is malformed and confusing: page 34: "The maximum number shared register is 128 less two times the number of clause temp registers used."
page 67: "A constant-register base address is always be in-bounds."
Page 79: "When any of the reduction instructions (DOT4, DOT4_IEEE, CUBE, and MAX4) is (-> are) used, it (-> they) must be executed on all four elements of a single vector."
Also from the APP SDK FAQ:
69. Is LDS memory persistent?
Within a single kernel execution, LDS memory is persistent; however, between kernel executions, memory peristency is not gauranteed.
Guarantee is spelled worng