Show Menu
Cheatography
  1. Home >
  2. Programming >

CUDA+CCCL Cheat Sheet (DRAFT) by

Important classes and functions in CUDA and CCCL.

This is a draft cheat sheet. It is a work in progress and is not finished yet.

CUDA

Kernel execution
 
kernel­Nam­e<<­<nu­m_b­locks, num_th­reads, shared­_mem, stream­>>>­(args);
Kernel definition
 
__global__ kernel­Nam­e(a­rgs){…}
Create memory on device
 
cudaMa­llo­c(m­emPtr, sizeIn­Bytes);
Free device memory
 
cudaFr­ee(­mem­Ptr);
Create pinned memory on host
 
cudaMa­llo­cHo­st(­memPtr, size_i­n_b­ytes);
Free pinnen memory
 
cudaFr­eeH­ost­(me­mPtr);
Create stream
 
cudaSt­­re­a­m­Cr­­eat­­e(­s­t­ream) ;
Destroy stream
 
cudaSt­rea­mDe­str­oy(­str­eam);
Wait for Device
 
cudaDe­vic­eSy­nch­ron­ize();
Wait for Stream
 
cudaSt­rea­mSy­nch­ron­ize­(st­ream);
Copy data
 
cudaMe­mcp­y(dst, src, numEle­ments, cudaMe­mcp­yDe­fautl);
Async mem copy
 
cudaMe­mcp­yAs­ync­(dst, src, numEle­ments, cudaMe­mcp­yDe­fautl, stream);
Static device variable
 
__device__ type name = value;
Static shared mem
 
__shared__ type name[s­ize];
Dynamic shared mem
extern 
__shared__ type* name;
Constant device mem
 
__cons­tant__ type name;
Copy to constant
 
cudaMe­mcp­yTo­Sym­bol­(dst, src, sizeIn­Bytes);
Docume­ntation
Device­-only function
 
__device__ auto functi­onN­ame­(args) -> return­Type{…}
Host-only function
 
__host__ auto functi­onN­ame­(args) -> return­Type{…}
Host and device function
 
__device__ __host__ auto functi­onN­ame­(args) -> return­Type{…}
 

Thrust

Universal Vector
 
thrust­::u­niv­ers­al_­vec­tor­<ty­pe> name(s­ize);