| 主权项 |
1. A computerized method comprising:
receiving, from a host application executing on a host processor, a request to execute a compute program on a selected compute device, the compute program corresponding to source code for a task in the host application and comprising a compute kernel containing executable code for the task,
wherein the selected compute device corresponds to a compute identifier selected by the host application selected by the host application from one or more compute identifiers received by the host application from the host processor in response to a previous request from the host application to the host processor, the previous request including a processing requirement specified by the host application for the task to cause the host processor to identify compute devices that match the processing requirement for the task and to return compute identifiers for matching compute devices to the host application, the compute devices comprising at least one central processing unit (CPU) and at least one graphics processing unit (GPU) physically coupled to the host processor; if a compute library does not exist, or the compute program does not exist in an existing compute library, or the compute program in the existing compute library is not optimized for the selected compute device, creating, by the host processor, the compute program for execution on the selected compute device, including compiling the compute kernel from the source; and scheduling, by the host processor, the compute kernel for execution. |
