1.4 KiB
1.4 KiB
tags, date created, date modified
| tags | date created | date modified |
|---|---|---|
| 星期四, 十一月 20日 2025, 8:52:43 晚上 | 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。
- 缓冲区管理状态机
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 --> [*]