cuda.core.experimental.ObjectCode

class cuda.core.experimental.ObjectCode

Represent a compiled program to be loaded onto the device.

This object provides a unified interface for different types of compiled programs that will be loaded onto the device.

Note

This class has no default constructor. If you already have a cubin that you would like to load, use the from_cubin() alternative constructor. For all other possible code types (ex: “ptx”), only Program accepts them and returns an ObjectCode instance with its compile() method.

Note

Usage under CUDA 11.x will only load to the current device context.

Methods

__init__()
static from_cubin(module: bytes | str, *, symbol_mapping: dict | None = None) → ObjectCode

Create an ObjectCode instance from an existing cubin.

Parameters:

  • module (Union[bytes, str]) – Either a bytes object containing the in-memory cubin to load, or a file path string pointing to the on-disk cubin to load.

  • symbol_mapping (Optional[dict]) – A dictionary specifying how the unmangled symbol names (as keys) should be mapped to the mangled names before trying to retrieve them (default to no mappings).

get_kernel(name)

Return the Kernel of a specified name from this object code.

Parameters:

name (Any) – Name of the kernel to retrieve.

Returns:

Newly created kernel object.

Return type:

Kernel