CUDA details GPU architecture

         Each thread has its own private local memory （Local Memory）, Each thread block contains shared memory （Shared Memory）, Can be shared by all threads in the thread block , Its life cycle is consistent with the thread block . Besides , All threads have access to global memory （Global Memory）. You can also access some read-only memory blocks ： Constant memory （Constant Memory） And texture memory （Texture Memory）. Memory structure involves program optimization , They are not discussed in depth here .

        GPU One of the core components of hardware is SM, I've said that before ,SM The English name is Streaming Multiprocessor, Stream multiprocessor .SM Its core components include CUDA The core , Shared memory , Register, etc ,SM Hundreds of threads can be executed concurrently , Concurrency depends on SM Number of resources owned . When one kernel When executed , its gird Thread blocks in are allocated to SM On , A thread block can only be in one SM Was dispatched to .SM In general, multiple thread blocks can be scheduled , It depends SM Own ability . So it's possible that one kernel Each thread block of is allocated multiple SM, therefore grid It's just logic , and SM It's the physical layer of execution .SM It's using SIMT (Single-Instruction, Multiple-Thread, Single instruction multithreading ) framework , The basic execution unit is