เวกเตอร์และการแปลงของ Thrust

Thrust is a high-level library that feels like the C++ STL but runs on the GPU. You write almost no kernel code yourself. 🎉

Thrust gives you host_vector for CPU memory and device_vector for GPU memory. They look identical but live in different worlds.

Assigning a host_vector to a device_vector triggers a memory transfer automatically. No cudaMemcpy boilerplate to write by hand.

A device_vector allocates and frees GPU memory for you using RAII. When it goes out of scope, the memory is released cleanly.

Helpers like thrust::fill and thrust::sequence initialize a whole device_vector in one parallel call instead of a manual loop.

thrust::transform applies a function to every element in parallel and writes the result to an output range. This is the classic map pattern.

The operation you pass is a functor: a struct with an operator() marked __host__ __device__ so it can run on the GPU.

For common math, Thrust ships functors like thrust::plus and thrust::multiplies, so you skip writing a functor for simple jobs.

The binary form of transform takes two input ranges and combines them pairwise, perfect for element-wise vector addition.

Need to drop into a hand-written kernel? thrust::raw_pointer_cast hands you the underlying device pointer to pass along.

Thrust hides allocation, copies, and launch configs. You trade a little control for readable, safe GPU code that just works.

Recall how Thrust applies an operation to a whole vector.

You used device_vector for automatic memory, transferred data by assignment, and ran transform with functors. STL-style GPU power. 💪