+ "message": "Make trace generation more GPU ready (#3356)\n\nCurrently, our `generate_trace` method which is supposed to be generic\nuses an internal concrete type `Trace` to represent the traces of the\noriginal instructions. This type is based on p3 which represents traces\nas a contiguous piece of host memory.\n\nTo support GPU tracegen, this is not generic enough: once we call trace\ngen for the original instructions, we receive some pointers to device\n(GPU) memory. The plan is to then do something similar as on CPU: split\nthese original traces, grouping rows which correspond to the same apc\ncall, and map them to a single apc row on GPU.\n\nThis PR makes `generate_trace` generic over the type of the original\ntraces it is passed, with a flexible trait bound `TraceTrait` which only\nexposes the width of the trace. For now, we also require `Send + Sync`,\nbecause this works for the CPU traces. This might break with GPU, but\nwe'll see when we get there.",
0 commit comments