Migrate NVIDIA CUDA cuDNN projects to hipDNN

2026-03-31

3 min read time

Applies to Linux and Windows

This topic demonstrates how to migrate a NVIDIA CUDA cuDNN project to hipDNN.

Before you begin, ensure hipDNN (ROCm) is installed. See hipDNN installation for more information.

Here’s a minimal example of a hipDNN project in CMakeLists.txt:

project(my_app LANGUAGES CXX)
set(CMAKE_CXX_STANDARD 17)

find_package(hipdnn_frontend CONFIG REQUIRED)

add_executable(my_app main.cpp)
target_link_libraries(my_app PRIVATE hipdnn_frontend)

Tip

See Working examples in the Porting Guide for ported code samples.

Key differences between cuDNN and hipDNN

This table provides a high-level overview of the differences between cuDNN and hipDNN:

Aspect	cuDNN frontend	hipDNN frontend
Namespace	cudnn_frontend	hipdnn_frontend
Handle creation	cudnnCreate(&handle)	hipdnnCreate(&handle
Handle destruction	cudnnDestroy(handle)	hipdnnDestroy(handle)
Heuristics modes	All cuDNN heuristic modes	Currently only HeurMode_t::FALLBACK
Operation support	All cuDNN operations	See hipDNN operation support for more information.
Device memory utility	Surface<type>	MigratableMemory<type>
Device memory access	Surface<type>::devPtr	MigratableMemory<type>::deviceData()

See the Tensor dimensions and layouts section of the Operation Support doccument for operation-specifc hipDNN tensor layout details.

Troubleshooting

Error: Missing Heuristic modes A and B

The heuristic implementation in hipDNN is not implemented.

To fix the problem, use a combination of graph::get_ranked_engine_ids() and graph::set_preferred_engine_id_ext() if you need more detailed control over engine selection.

Error: Different memory utilities for allocating device memory

The memory utilities are typically consumer dependent and written on an as-needed basis. cuDNN provides a surface utility for their samples, for example.

To fix the issue, use MigratableMemory<type>, a utility that can automatically migrate data between the host and device (it also works as a stand-in). If you want to manage dims/strides more carefully, use the Tensor utility class. Both of these classes can be found in the hipdnn_data_sdk::utilities namespace.