50 lines
1.4 KiB
Markdown
50 lines
1.4 KiB
Markdown
|
---
|
|||
|
|
tags: []
|
|||
|
|
date created: 星期四, 十一月 20日 2025, 8:52:43 晚上
|
|||
|
|
date modified: 星期日, 十二月 7日 2025, 9:26:50 晚上
|
|||
|
|
---
|
|||
|
|
流的设计模式：B. 多流乒乓/多缓冲 (Multi-Stream Ping-Pong) (推荐)
|
|||
|
|
但是 先设计 A. 单流串行 (Serial Stream) 作为代码调试阶段的轻量级。
|
|||
|
|
|
|||
|
|
1. 缓冲区管理状态机
|
|||
|
|
|
|||
|
|
```mermaid
|
|||
|
|
stateDiagram-v2
|
|||
|
|
%% 状态定义
|
|||
|
|
state "HOST_OWNED<br/>(主机所有)" as HOST
|
|||
|
|
state "DEVICE_OWNED_H2D<br/>(传输中: H->D)" as H2D
|
|||
|
|
state "DEVICE_OWNED_COMPUTE<br/>(计算中: Kernel)" as COMPUTE
|
|||
|
|
state "DEVICE_OWNED_D2H<br/>(传输中: D->H)" as D2H
|
|||
|
|
state "RELEASED<br/>(待归还)" as RELEASED
|
|||
|
|
|
|||
|
|
%% 流程流转
|
|||
|
|
[*] --> HOST : 从 MemoryPool 申请
|
|||
|
|
|
|||
|
|
HOST --> H2D : I/O线程填充数据\n并调用 cudaMemcpyAsync
|
|||
|
|
note right of HOST
|
|||
|
|
此时数据位于页锁定内存
|
|||
|
|
CPU 写入完成
|
|||
|
|
end note
|
|||
|
|
|
|||
|
|
H2D --> COMPUTE : 记录 H2D_Event\nStreamWaitEvent
|
|||
|
|
note right of H2D
|
|||
|
|
DMA 引擎正在搬运
|
|||
|
|
CPU 不阻塞
|
|||
|
|
end note
|
|||
|
|
|
|||
|
|
COMPUTE --> D2H : Kernel 执行完毕\n自动触发 D2H
|
|||
|
|
note right of COMPUTE
|
|||
|
|
GPU 核心正在计算
|
|||
|
|
数据驻留显存
|
|||
|
|
end note
|
|||
|
|
|
|||
|
|
D2H --> RELEASED : D2H 完成回调\n或 Event 同步
|
|||
|
|
note right of D2H
|
|||
|
|
结果已写回 Host
|
|||
|
|
end note
|
|||
|
|
|
|||
|
|
RELEASED --> HOST : DataPacket 析构\n自动归还 Pool
|
|||
|
|
|
|||
|
|
RELEASED --> [*]
|
|||
|
|
```
|