GPU CUDA API Guide
CUDA requires that all functions be declared according to where they are going to be called and executed.
Given that the custom API scripts are called and run on the GPU, they must be declared as
__device__ functions as follows:
__device__
void calculateForce(NCudaApiTypesV1_0_0::CApiElement& element1,
NCudaApiTypesV1_0_0::CApiElement& element2,
…Other function declarations such as __Global__ and __Host__ are
available. However, they are not required.
CUDA Programming Style
The CUDA API is closer to the C++ programming style and the basic operations usually
take a much simpler form. Some examples of functions are as follows:
| Function Type | CUDA Function |
|---|---|
| Vector arithmetic | a-b |
| a+b | |
| a*b | |
| a/b | |
| Vector operations | a.dot(b) |
| a.cross(b) | |
| a.normalize() |
Note: For more information about the available
functions, see the API Help documentation at the following location:
<EDEM install
directory>\202X.X\EDEM\src\Api\Help.CAUTION: API users should be cautious when implementing contact models for
the CUDA solver to not rely on the contact normal being in the same direction as
the vector from the API element 'position' member to the contact point. While
this does work for Multi-Sphere shapes, it will not work for Sphero-Cylinder or
Polyhedral shapes. You must instead use the
contactNormal
member which is passed in the overlapData API structure. Apart
from this, all contact properties for Sphero-Cylinders have a similar definition
to those for Multi-Sphere particles.