NVIDIALibraries.jl was designed to allow full low level access by providing wrapper functions to the CUDA Toolkit libraries.

This package is a Julia library written for Julia 1.0 and onwards.

NVIDIALibraries.jl defines C definitions and functions as they would appear in a C project.

• CUDA library (NVIDIALibraries.CUDA)
• CUDA Runtime library (NVIDIALibraries.CUDARuntime)
• CUBLAS library (NVIDIALibraries.CUBLAS)

• 8.0
• 9.0
• 9.1
• 9.2
• 10.0
• 10.1

Enter the Pkg REPL-mode from the Julia REPL by using the key ].

Now add the Julia package by using the add command:

(v1.0) pkg> add https://github.com/Blackbody-Research/NVIDIALibraries.jl

The CUDAArray datatype uses the CUDA Runtime library to allocate and free GPU device memory.

CUDAArrays can be constructed with initial values by passing a Julia array to cuda_allocate(::Array).

In addition, CUDAArrays can be explicitly freed by calling deallocate!(::CUDAArray).

Use cuModuleLoad() to load kernel by filename.

example_module = cuModuleLoad("example.ptx")

Alternatively, you can use the CUDA JIT linker to compile a PTX file into a cubin format.

jit_options = [CU_JIT_TARGET, CU_JIT_OPTIMIZATION_LEVEL]
jit_option_values = [CU_TARGET_COMPUTE_70, Cuint(4)]
jit_link_state = cuLinkCreate(2, jit_options, jit_option_values)
cuLinkAddFile(jit_link_state, CU_JIT_INPUT_PTX, "example.ptx")
cubin_array = cuLinkComplete(jit_link_state)

# cleanup CUDA JIT linker
result = cuLinkDestroy(jit_link_state)
@assert (result == CUDA_SUCCESS) ("cuLinkDestroy() error: " * cuGetErrorString(result))

Then, load the kernel from the new cubin image as a CUDA module.

example_module = cuModuleLoadData(Base.unsafe_convert(Ptr{Nothing}, cubin_array))

The update command in Pkg REPL-mode can be used to add recent commits made to NVIDIALibraries.jl:

(v1.0) pkg> update NVIDIALibraries

The license can be found in LICENSE.