tensor_lib

A smart pointer based wrapper for CUDA memory allocations in c++ 17.

Dependencies:

1. CMake v3.8+ [for CUDA support within CMake]
2. CUDA v9.0+

Build

mkdir build
cd build
cmake ..
make install

Example usage:

1. To integrate the library in your code check out the examples/CMakeLists.txt

#include "host_tensor.hpp"	// managed ptr to data stored on the CPU
#include "device_tensor.hpp"	// managed ptr to data stored on the GPU

....
    
int main(){
    std::array<int, 1> size{1024};
    
    /* create a 1D tensor holding 1024 floats.*/
    host_tensor</*DIMS=*/ 1> a(/*size=*/ size, /*rand=*/ true);
    
    /* copy data to device. */
    device_tensor<1> b(a, /*copy=*/ true);
    
    /* copy data from the device to host. */
    host_tensor<1> c(b, /*copy=*/ true);
    
   .... 
         
   /* access data at i-th location. */
   std::cout << device_tensor.at(i);	/* done in device code */
   std::cout << host_tensor.at(i);	/* done in host code */
}

Road Map

• Operations
  • Broadcast binary ops
  • Pointwise binary ops
  • Fill
  • Reduction
• 1D support
• 2D support
• 3D support