# NVIDIA


The set of samples demonstrates how to interact with an external graphics API via Unigine API (*[ResourceExternalMemory](../../../../api/library/rendering/class.resourceexternalmemory_cpp.md)* and *[ResourceFence](../../../../api/library/rendering/class.resourcefence_cpp.md)*). Although the samples show interaction with *NVIDIA CUDA Toolkit*, you may use them as the basis for integrating the Engine GAPI (*Vulkan, DX11, DX12*) with other GAPI (OpenGL, DX, CUDA, etc.) that can accept the handle and import a synchronization primitive (such as external semaphore in CUDA).


![](cuda_interop.jpg)


This approach allows interaction of various APIs using different GAPI, such as interaction of *MediaFoundation* using DX11 with the Engine using DX12 or *Vulkan*.


The samples also illustrate how to take the data directly from the video memory without CPU roundtrip and processing it using non-engine graphics API.


![](cuda_data_exchange.jpg)


The difference in *GPU -> CPU* texture transfer performance made via CUDA 12.3 and by the Engine are shown in the table below (tested on our working configurations). Average performance gain with CUDA comprised **x10** to **x24** depending on the viewport resolution.


![](cuda_engine_comparison.png)


You can check the results on your PC just by launching the [CUDATextureTransfer](#texture_transfer) sample and enabling the *[Profiler](../../../../tools/profiling/profiler/index.md)*.


## CUDAMeshDynamic


![](../../../../samples/img/cuda_meshdynamic.jpg)

Vertices of a shared dynamic mesh are processed on GPU using CUDA.


To build the sample, use CUDA [Toolkit v12.4](https://developer.nvidia.com/cuda-12-4-0-download-archive).


**On Linux:** after installing the CUDA Toolkit, make sure that the **PATH** environment variable includes `/usr/local/cuda-12.4/bin`.


**SDK Path:***<SAMPLES_PROJECT_PATH>/source\samples\3rdparty\CUDAMeshDynamic*
## CUDAStructuredBufferWrite


![](../../../../samples/img/cuda_structured_buffer_write.jpg)

Particles in a shared structured buffer are updated on GPU using CUDA, and rasterized manually using Unigine Material with Compute Shader.


To build the sample, use CUDA [Toolkit v12.4](https://developer.nvidia.com/cuda-12-4-0-download-archive).


**On Linux:** after installing the CUDA Toolkit, make sure that the **PATH** environment variable includes `/usr/local/cuda-12.4/bin`.


**SDK Path:***<SAMPLES_PROJECT_PATH>/source\samples\3rdparty\StructuredBufferWrite*
## CUDATextureTransfer


![](../../../../samples/img/cuda_texture_transfer.jpg)

A shared texture is processed with CUDA and copied from video memory to RAM into Unigine *Image*.


To build the sample, use CUDA [Toolkit v12.4](https://developer.nvidia.com/cuda-12-4-0-download-archive).


**On Linux:** after installing the CUDA Toolkit, make sure that the **PATH** environment variable includes `/usr/local/cuda-12.4/bin`.


**SDK Path:***<SAMPLES_PROJECT_PATH>/source\samples\3rdparty\CUDATextureTransfer*
## CUDATextureWrite


![](../../../../samples/img/cuda_texture_write.jpg)

A shared texture is processed with CUDA and displayed on the Unigine *Object* as a texture on a cube.


To build the sample, use CUDA [Toolkit v12.4](https://developer.nvidia.com/cuda-12-4-0-download-archive).


**On Linux:** after installing the CUDA Toolkit, make sure that the **PATH** environment variable includes `/usr/local/cuda-12.4/bin`.


**SDK Path:***<SAMPLES_PROJECT_PATH>/source\samples\3rdparty\CUDATextureWrite*