Memory allocation in distributed memories for multiprocessing

摘要

In some aspects, finer grained parallelism is achieved by segmenting programmatic workloads into smaller discretized portions, where a first element can be indicative both of a configuration or program to be executed, and a first data set to be used in such execution, while a second element can be indicative of a second data element or group. The discretized portions can cause program execute on distributed processors. Approaches to selecting processors, and allocating local memory associated with those processors are disclosed. In one example, discretized portions that share a program have an anti-affinity to cause dispersion, for initial execution assignment. Flags, such as programmer and compiler generated flags can be used in determining such allocations. Workloads can be grouped according to compatibility of memory usage requirements.

主权项

1. A processor for graphics rendering, comprising:
a plurality of processing units, each processing unit capable of executing at least one program instance independently of the other processing units of the plurality; a plurality of local memories used by the plurality of processing units; an allocator shared among the plurality of processing units, the allocator operable to allocate space in the local memories for data associated with program instances and create a mapping between the program instances and the allocated space in the local memories; a collector comprising packet storage configured to store a plurality of packets of program instances, wherein the collector is configured to collect program instances into packets in the packet storage by, for each of a plurality of program instances, selecting a packet and adding the program instance to the selected packet, the selected packet being stored in the packet storage; and a distributor configured to distribute program instances, obtained from the packets, to be executed on the plurality of processing units according to the mapping.