Summary
Support user-defined input/output tensors for kernels, instead of using randomly generated input tensors.
Motivation / Use Case
Currently, PyPTO has limitations in kernel input/output tensors: the kernel's input tensors are randomly generated, and users cannot customize input tensors.
Need to support user-defined inputs and outputs to specific tensors (such as specifying virtual addresses), so that users can:
- Control the data source of kernel inputs
- Specify the output target location of the kernel
- Integrate more flexibly into existing systems
Alternatives Considered
- Manually modifying generated random data (not feasible, as data is generated at compile time)
- Using external tools to inject data (complex and not elegant)
- Passing data through file I/O (high performance overhead)
Additional Context
- Scope: kernel definition, compilation process, runtime execution
- Modules that may need modification: python/pypto/, src/, python/bindings/
Summary
Support user-defined input/output tensors for kernels, instead of using randomly generated input tensors.
Motivation / Use Case
Currently, PyPTO has limitations in kernel input/output tensors: the kernel's input tensors are randomly generated, and users cannot customize input tensors.
Need to support user-defined inputs and outputs to specific tensors (such as specifying virtual addresses), so that users can:
Alternatives Considered
Additional Context