Understanding of several important concepts of Cuda learning notes

Today we will talk about several cuda-related concepts in the GPU hardware structure: thread block grid warp SP Sm

SP: the most basic processing unit. The specific commands and tasks of streaming processor are processed on the SP. GPU for parallel computing, that is, multiple SPs simultaneously Process

SM: multiple SPs and other resources form an SM, streaming multiprocessor. Other resources are storage resources, shared memory, and storage devices.

Warp: The scheduling unit when the GPU executes the program. The current Cuda Warp Size is 32, and the same command is executed in different data resources in the same warp thread.

Thread --> block --> grid: When Cuda is used for programming, a grid is divided into multiple blocks, and a block is divided into multiple threads. Whether the division of tasks affects the final execution result. The division is based on task features and GPU hardware features.

The following hardware structure diagram is easy to understand (picture comes from the Internet)

The above two figures can clearly show the relationship between SM and sp.

This figure reflects the role of warp as the scheduling unit. Each time the GPU schedules 32 threads in a warp to execute the same command

Different threads correspond to different data resources.

This figure is an example of a warp schedule.

One SM will only execute the warp in one block, and the warp in other blocks will be executed after the warp in the block is executed.

When partitioning, it is best to ensure that the warp in each block is more reasonable, so that an SM can execute the warp in turn to improve

Efficiency. In addition, when allocating blocks, we need to allocate a reasonable number of blocks based on the number of GPUs SM, so that the GPU Sm can be used to improve the efficiency.

Utilization. When allocating resources, you must also consider the resources of the same thread block. Do not make sure the corresponding resources are insufficient.

Reprinted on http://blog.csdn.net/pizi0475/article/details/7630248