CUDA basic knowledge

GPU The problem that computation suits to solve is simple logic , Calculations are independent and parallel , and CPU The suitable problem to solve is , Complex logic , The problem of interdependency between calculations .

GPU Our design idea is high concurrent throughput , and CPU The design idea of is low delay . So we can see , stay CPU It has multi-level cache and powerful control unit , The purpose of introducing multi-level cache is to reduce latency , and GPU in , It has a large number of computing single clouds , The control unit is relatively simple , The number of caches is also relatively small .

1. Memory model

Hardware side ：

        SP： Thread processor , Have your own registers（ register ） And local memory（ Local memory ）.registers and local memory Only by SP Visit by yourself .

        SM： Multicore processor , By multiple SP as well as shared memory（ Shared memory ） form .shared memory Can be SM All in SP visit .

        GPU： The graphics card , By multiple SM as well as global memory（ Global memory ） form .global memory Can be used by all SP visit .

Software side ：

        thread： Threads , Corresponding SP.

        block： Thread block , By multiple tread form , Corresponding SM.

        grid： By multiple block form , One GPU You can have multiple grid.

        warp： Thread bundles , from 32 individual tread form . generally ,block yes 32 Multiple .

The index of the thread ：

        For this concept , It is actually calculating the index of the current thread in the defined computing unit , Index starts from zero . The key to understanding this calculation is to introduce spatial imagination , One dimension is a line , Two dimensions are faces , Three dimensional is three-dimensional .

dim3 grid1(2, 1, 1); // x=2, y=1, z=1, It's one-dimensional ,yz by 1
dim3 grid2(4, 2, 1); // x=4, y=2, z=1, It's two-dimensional ,z by 1
dim3 grid3(2, 3, 4); // x=2, y=3, z=4, This is three-dimensional ,xyz Are not as 1