In the A13 GPU documentation, rename discard to last_use #53
Description
You give one of the two operand hints the name discard. The Apple patents refer to this as last_use, which I think makes more sense and makes it clear how the hint is to be used. This name also clarifies your concern about the consequences of using this hint in a predicated instruction.
You can see this described here:
https://patents.google.com/patent/US9600288B1
(look for the description of Fig 3)
There is an additional "prefetch" hint. The register address field that is used by a third source (ie FMA instruction) can be used with a two src instruction (ie most of them) to specify a register that is to be "prefetched" in this cycle.
The scheme APPEARS to be flexible enough to allow for either
- one-time prefetch (get the value here when I want, so either matching a companion register already in Operand Cache, or avoiding a Register Bank collision against a second source register expected to be in Register File not Operand Cache)
- prefetch into Operand Cache, so lower latency first use, and cached subsequent use.