Virtualization of multiple coprocessors

A non-transitory computer readable medium comprising instructions that are executable in a computer system to cause a hardware platform of the computer system that includes at least one processor and a plurality of coprocessors to carry out a data processing method while an application is running thereon via system software, the data processing method comprising:

• intercepting, by an intermediate software layer, which is installed in a non-privileged, user space to run logically between the application and the system software, without modification of the application or of the system software, first, second, and third kernels dispatched within a data and command stream issued by the application, each said kernel corresponding to instructions to the plurality of coprocessors;
• automatically, and transparently to the application, determining data dependencies among the kernels including determining that an input to the first kernel is an output from the second kernel;
• automatically, and transparently to the application, determining that the third kernel has no data dependencies on the first and second kernels;
• selecting at least one coprocessor to which the first and second kernels are to be dispatched based at least in part on the determined data dependencies;
• selecting another coprocessor to which the third kernel is to be dispatched based at least in part on the third kernel having no data dependencies on the first and second kernels;
• dispatching the first and second kernels to the selected at least one coprocessor and the third kernel to the selected another coprocessor;
• and
• returning kernel results to the application.

A data processing system comprising:

• a hardware platform that includes at least one processor and a plurality of coprocessors;
• an application running on the hardware platform via system software;
• and
• a hardware abstraction layer installed in a non-privileged, user space and running logically between the application and the system software, wherein the hardware abstraction layer is configured to:
• intercept, without modification of the application or of the system software, first, second, and third kernels dispatched within a data and command stream issued by the application, each said kernel corresponding to instructions to the plurality of coprocessors,
• automatically, and transparently to the application, determine data dependencies among the kernels including determining that an input to the first kernel is an output from the second kernel,
• automatically, and transparently to the application, determine that the third kernel has no data dependencies on the first and second kernels,
• select at least one coprocessor to which the first and second kernels are to be dispatched based at least in part on the determined data dependencies,
• select another coprocessor to which the third kernel is to be dispatched based at least in part on the third kernel having no data dependencies on the first and second kernels,
• dispatch the first and second kernels to the selected at least one coprocessor and the third kernel to the selected another coprocessor, and
• return kernel results to the application.