cupy.RawModule

class cupy.RawModule(code_or_path, options=(), backend=u'nvrtc', translate_cucomplex=False, *)

User-defined custom module.

This class can be used to either compile raw CUDA sources or load CUDA modules (*.cubin). This class is useful when a number of CUDA kernels in the same source need to be retrieved.

For the former case, the CUDA source code is compiled when initializing a new instance of this class, and the kernels can be retrieved by calling get_function(), which will return an instance of RawKernel. (Same as in RawKernel, the generated binary is also cached.)

For the latter case, an existing CUDA binary (*.cubin) can be loaded by providing its path, and kernels therein can be retrieved similarly.

Parameters:
  • code_or_path (str) – CUDA source code or path to cubin.
  • options (str) – Compiler options passed to NVRTC if compilation is needed. For details, see https://docs.nvidia.com/cuda/nvrtc/index.html#group__options.
  • backend (str) – Either nvrtc or nvcc. Defaults to nvrtc
  • translate_cucomplex (bool) – Whether the CUDA source includes the header cuComplex.h or not. If set to True, any code that uses the functions from cuComplex.h will be translated to its Thrust counterpart. Defaults to False.

Note

Each kernel in RawModule possesses independent function attributes.

Methods

get_function(self, name)

Retrieve a CUDA kernel by its name from the module.

Parameters:name (str) – Name of the kernel function.
Returns:An RawKernel instance.
Return type:RawKernel
get_texref(self, name)

Retrieve a texture reference by its name from the module.

Parameters:name (str) – Name of the texture reference.
Returns:A CUtexref handle, to be passed to TextureReference.
Return type:intptr_t

Attributes

backend
code
cubin_path
options