Systems and methods for dynamically choosing a processing element for a compute kernel
A runtime system implemented in accordance with the present invention provides an application platform for parallel-processing computer systems. Such a runtime system enables users to leverage the computational power of parallel-processing computer systems to accelerate/optimize numeric and array-intensive computations in their application programs. This enables greatly increased performance of high-performance computing (HPC) applications.
1. A computer-implemented method configured to be performed by a runtime system at a parallel-processing computer system that includes multiple types of processing elements, comprising:
at runtime:
receiving one or more operation requests issued by an application;
dynamically choosing a respective type of processing element identifying at least one of the one or more types of processing elements for at least one of the one or more operation requests, the at least one operation request corresponding to an intrinsic operation, wherein the intrinsic operation is an arithmetic operation, and the respective type of processing element is one of the multiple types of processing elements, which include sing-core/multi-core central processing units, graphics processing units and single-core/multi-core co-processors;
dynamically identifying, from a library, a precompiled, processor-specific compute kernel corresponding to the intrinsic operation of the at least one operation request, the library comprising a plurality of processor-specific compute kernels corresponding to a plurality of intrinsic operations; and
dynamically preparing one or more compute kernels for the at least one operation request, wherein the one or more compute kernels include the at least one precompiled, processor-specific compute kernel, and are configured to execute on the respective type of processing element.
2. The computer-implemented method of claim 1 , wherein the precompiled, processor-specific compute kernels in the library are compressed and/or encrypted.
3. The computer-implemented method of claim 1 , wherein the precompiled, processor-specific compute kernels are hand-coded for a specific type of processor.
4. The computer-implemented method of claim 1 , wherein the intrinsic operation corresponding to the at least one operation request is selected from a group consisting of:
matrix multiplication, matrix solvers, fast fourier transforms, convolutions, and LU decomposition.
5. The computer-implemented method of claim 1 , wherein the respective type of processing element is dynamically chosen based, at least in part, on one of a predefined computer resource requirement metric of the one or more operation requests and a predefined workload requirement metric of the parallel-processing computer system running the runtime system.
6. The computer-implemented method of claim 1 , wherein the one or more operation requests correspond to application program interface function calls.
7. The computer-implemented method of claim 1 , wherein the identified precompiled, processor-specific compute kernel corresponding to the intrinsic operation is in a GPU binary instruction set architecture format.
8. A parallel-processing computer system, comprising:
memory;
multiple types of processing elements; and
at least one program stored in the memory and executed by the multiple types of processing elements, the at least one program including a runtime system comprising instructions for:
at runtime:
receiving one or more operation requests issued by an application;
dynamically choosing a respective type of processing element identifying at least one of the one or more types of processing elements for at least one of the one or more operation requests, the at least one operation request corresponding to an intrinsic operation, wherein the intrinsic operation is an arithmetic operation, and the respective type of processing element is one of the multiple types of processing elements, which include sing-core/multi-core central processing units, graphics processing units and single-core/multi-core co-processors;
dynamically identifying, from a library, a precompiled, processor-specific compute kernel corresponding to the intrinsic operation of the at least one operation request, the library comprising a plurality of processor-specific compute kernels corresponding to a plurality of intrinsic operations; and
dynamically preparing one or more compute kernels for the at least one operation request, wherein the one or more compute kernels include the at least one precompiled, processor-specific compute kernel, and are configured to execute on the respective type of processing element.
9. The parallel-processing computer system of claim 8 , wherein the precompiled, processor-specific compute kernels in the library are compressed and/or encrypted.
10. The parallel-processing computer system of claim 8 , wherein the precompiled, processor-specific compute kernels are hand-coded for a specific type of processor.
11. The parallel-processing computer system of claim 8 , wherein the intrinsic operation corresponding to the at least one operation request is selected from a group consisting of: matrix multiplication, matrix solvers, fast fourier transforms, convolutions, and LU decomposition.
12. The parallel-processing computer system of claim 8 , wherein the respective type of processing element is dynamically chosen based, at least in part, on one of a predefined computer resource requirement metric of the one or more operation requests and a predefined workload requirement metric of the parallel-processing computer system running the runtime system.
13. The parallel-processing computer system of claim 8 , wherein the one or more operation requests correspond to application program interface function calls.
14. The parallel-processing computer system of claim 8 , wherein the identified precompiled, processor-specific compute kernel corresponding to the intrinsic operation is in a GPU binary instruction set architecture format.
15. A non-transitory computer readable storage medium storing one or more programs configured to be executed by computer program product for use in conjunction with a parallel-processing computer system that includes multiple types of processing elements, the one or more programs comprising instructions for:
at runtime:
receiving one or more operation requests issued by an application;
dynamically choosing a respective type of processing element identifying at least one of the one or more types of processing elements for at least one of the one or more operation requests, the at least one operation request corresponding to an intrinsic operation, wherein the intrinsic operation is an arithmetic operation, and the respective type of processing element is one of the multiple types of processing elements, which include sing-core/multi-core central processing units, graphics processing units and single-core/multi-core co-processors;
dynamically identifying, from a library, a precompiled, processor-specific compute kernel corresponding to the intrinsic operation of the at least one operation request, the library comprising a plurality of processor-specific compute kernels corresponding to a plurality of intrinsic operations; and
dynamically preparing one or more compute kernels for the at least one operation request, wherein the one or more compute kernels include the at least one precompiled, processor-specific compute kernel, and are configured to execute on the respective type of processing element.
16. The non-transitory computer readable storage medium of claim 15 , wherein the precompiled, processor-specific compute kernels in the library are compressed and/or encrypted.
17. The non-transitory computer readable storage medium of claim 15 , wherein the precompiled, processor-specific compute kernels are hand-coded for a specific type of processor.
18. The non-transitory computer readable storage medium of claim 15 , wherein the intrinsic operation corresponding to the at least one operation request is selected from a group consisting of: matrix multiplication, matrix solvers, fast fourier transforms, convolutions, and LU decomposition.
19. The non-transitory computer readable storage medium of claim 15 , wherein the respective type of processing element is dynamically chosen based, at least in part, on one of a predefined computer resource requirement metric of the one or more operation requests and a predefined workload requirement metric of the parallel-processing computer system running the runtime system.
20. The non-transitory computer readable storage medium of claim 15 , wherein the one or more operation requests correspond to application program interface function calls.
21. The non-transitory computer readable storage medium of claim 15 , wherein the identified precompiled, processor-specific compute kernel corresponding to the intrinsic operation is in a GPU binary instruction set architecture format.