THCTensorMath.cuh

#ifndef THC_TENSORMATH_CUH
#define THC_TENSORMATH_CUH

// Copy the kth diagonal of a matrix B to a vector A.
template <typename T>
__global__ void THCTensor_copyFromDiagonal(T* a, T* b, ptrdiff_t start, ptrdiff_t size, ptrdiff_t strideSum, ptrdiff_t strideA) {
  for (ptrdiff_t linearIndex = blockIdx.x * blockDim.x + threadIdx.x;
       linearIndex < size;
       linearIndex += gridDim.x * blockDim.x) {
    const ptrdiff_t bOffset = start + strideSum * linearIndex;
    a[strideA * linearIndex] = b[bOffset];
  }
}

// Copy vector B to the kth diagonal of a matrix A
template <typename T>
__global__ void THCTensor_copyToDiagonal(T* a, T* b, ptrdiff_t start, ptrdiff_t size, ptrdiff_t strideSum, ptrdiff_t strideB) {
  for (ptrdiff_t linearIndex = blockIdx.x * blockDim.x + threadIdx.x;
       linearIndex < size;
       linearIndex += gridDim.x * blockDim.x) {
    const ptrdiff_t aOffset = start + strideSum * linearIndex;
    a[aOffset] = b[strideB * linearIndex];
  }
}

#endif