DenseVector/Kernels/Morpheus_Copy_Impl.hpp

 
#ifndef MORPHEUS_DENSEVECTOR_KERNELS_COPY_IMPL_HPP
#define MORPHEUS_DENSEVECTOR_KERNELS_COPY_IMPL_HPP
 
#if defined(MORPHEUS_ENABLE_HIP)
#include <impl/Morpheus_HIPUtils.hpp>
#elif defined(MORPHEUS_ENABLE_CUDA)
#include <impl/Morpheus_CudaUtils.hpp>
#endif
 
namespace Morpheus {
namespace Impl {
namespace Kernels {
 
template <typename ValueType, typename IndexType, typename SizeType>
__global__ void copy_by_key_kernel(SizeType n, const IndexType* keys,
                                   const ValueType* src, ValueType* dst) {
  const SizeType tid = blockDim.x * blockIdx.x + threadIdx.x;
  if (tid > n) return;
 
  dst[tid] = src[keys[tid]];
}
 
}  // namespace Kernels
}  // namespace Impl
}  // namespace Morpheus
 
#endif  // MORPHEUS_DENSEVECTOR_KERNELS_COPY_IMPL_HPP