Building and Running a cuDNN Dependent Program#

Building a cuDNN Dependent Program#

Because cuDNN uses symbols defined in external libraries, you need to ensure that the linker can locate these libraries while building a cuDNN dependent program. One way to achieve this is by explicitly specifying them on the linker command.

Linker Dependencies for the Dynamic cuDNN Libs

Linux

Add -lz to the linker command.

Windows

No additional linker flags are required for Windows.

Linker Dependencies for the Static cuDNN Libs

Linux

Add -lcublasLt_static -lz -lnvrtc_static -lnvrtc-builtins_static -lnvptxcompiler_static -lcudart_static to the linker command.

Windows

Not applicable. Static cuDNN libs for Windows are not supported.

Running a cuDNN Dependent Program#

While running a cuDNN dependent program, in addition to the paths to the directory of cuDNN, make sure to add the path to the directory of the following libraries to the environment variable $LD_LIBRARY_PATH on Linux, or to the environment variable %PATH% on Windows.

Linux

  • libz.so (required by ld.so at load-time)

  • libcublasLt.so (required by dlopen() at runtime when using certain matmul engines, convolution engines, legacy convolution APIs, RNN APIs, and legacy multihead attention APIs)

  • libnvrtc.so (required by dlopen() at runtime when using runtime-compiled engines)

Windows

  • cublasLt*.dll (required by LoadLibraryA() at runtime when using certain matmul engines, convolution engines, legacy convolution APIs, RNN APIs, and legacy multihead attention APIs)

  • nvrtc*.dll (required by LoadLibraryA() at runtime when using runtime-compiled engines)

Note

When linking cuDNN statically, wrap the needed static sub-libraries with the --whole-archive and --no-whole-archive linker flags within the linker command.

Inter-Library Dependencies#

Since cuDNN is split into several libraries, dependencies between them need to be taken into account. For more information, refer to the API Library Overview.

On this page