已关闭
[docs] 修正 AICPU AllGatherAdd README 中与 main.asc 不一致的函数名 #239
NUAA_wqy创建于 28 天前关闭于 6 天前
[docs] 修正 AICPU AllGatherAdd README 中与 main.asc 不一致的函数名 #239
已关闭
共 2 个文件变更+6-6
| @@ -40,7 +40,7 @@ Host: | |||
| 40 | HcclComm | 40 | HcclComm |
| 41 | -> Mc2GetCcArgs/Mc2SetCc* | 41 | -> Mc2GetCcArgs/Mc2SetCc* |
| 42 | -> Mc2AcquireCcResCtx | 42 | -> Mc2AcquireCcResCtx |
| 43 | - -> HcclGetUnfoldThread/HcclStreamAcquireWithThread (获取通信域unfold线程及其展开流) | 43 | + -> GetUnfoldThread/AcquireUnfoldStream (获取通信域unfold线程及其展开流) |
| 44 | -> Mc2CcKernelLaunch (start AICPU KFC Server on unfoldStream) | 44 | -> Mc2CcKernelLaunch (start AICPU KFC Server on unfoldStream) |
| 45 | 45 | ||
| 46 | AICore: | 46 | AICore: |
| @@ -74,8 +74,8 @@ ccResCtx | |||
| 74 | AICPU KFC Server使用通信域unfold线程绑定的展开流，无需自建。 | 74 | AICPU KFC Server使用通信域unfold线程绑定的展开流，无需自建。 |
| 75 | 3. 通过`Mc2GetCcArgs`创建MC2参数对象，设置`AICPU_TS`通信引擎和FP32源/目标数据类型。 | 75 | 3. 通过`Mc2GetCcArgs`创建MC2参数对象，设置`AICPU_TS`通信引擎和FP32源/目标数据类型。 |
| 76 | 4. 通过`Mc2AcquireCcResCtx`基于HCCL通信域申请通信资源上下文`ccResCtx`及其大小。 | 76 | 4. 通过`Mc2AcquireCcResCtx`基于HCCL通信域申请通信资源上下文`ccResCtx`及其大小。 |
| 77 | -5. 释放MC2参数对象，通过`HcclGetUnfoldThread`获取（或复用）以`"%s_unfold"`为tag持久化在 | 77 | +5. 释放MC2参数对象，通过`GetUnfoldThread`获取（或复用）以`"%s_unfold"`为tag持久化在 |
| 78 | - `COMM_ENGINE_CPU_TS`引擎上下文中的unfold线程，再通过`HcclStreamAcquireWithThread`取得其 | 78 | + `COMM_ENGINE_CPU_TS`引擎上下文中的unfold线程，再通过`AcquireUnfoldStream`取得其 |
| 79 | 绑定的展开流，最后通过`Mc2CcKernelLaunch(unfoldStream, ccResCtx, ccResCtxSize)`在展开流上 | 79 | 绑定的展开流，最后通过`Mc2CcKernelLaunch(unfoldStream, ccResCtx, ccResCtxSize)`在展开流上 |
| 80 | 启动KFC Server。 | 80 | 启动KFC Server。 |
| 81 | 6. 通过`all_gather_add_kernel<<<1, nullptr, streamAiv>>>`启动AICore kernel。kernel内部调用 | 81 | 6. 通过`all_gather_add_kernel<<<1, nullptr, streamAiv>>>`启动AICore kernel。kernel内部调用 |
| @@ -46,7 +46,7 @@ Host: | |||
| 46 | HcclComm | 46 | HcclComm |
| 47 | -> Mc2GetCcArgs/Mc2SetCc* | 47 | -> Mc2GetCcArgs/Mc2SetCc* |
| 48 | -> Mc2AcquireCcResCtx | 48 | -> Mc2AcquireCcResCtx |
| 49 | - -> HcclGetUnfoldThread/HcclStreamAcquireWithThread (acquire the comm's unfold thread and its stream) | 49 | + -> GetUnfoldThread/AcquireUnfoldStream (acquire the comm's unfold thread and its stream) |
| 50 | -> Mc2CcKernelLaunch (start AICPU KFC Server on unfoldStream) | 50 | -> Mc2CcKernelLaunch (start AICPU KFC Server on unfoldStream) |
| 51 | 51 | ||
| 52 | AICore: | 52 | AICore: |
| @@ -83,9 +83,9 @@ ccResCtx | |||
| 83 | and the FP32 source and destination data types. | 83 | and the FP32 source and destination data types. |
| 84 | 4. Call `Mc2AcquireCcResCtx` to obtain the communication resource context `ccResCtx` and its size from | 84 | 4. Call `Mc2AcquireCcResCtx` to obtain the communication resource context `ccResCtx` and its size from |
| 85 | the HCCL communication domain. | 85 | the HCCL communication domain. |
| 86 | -5. Release the MC2 argument object. Call `HcclGetUnfoldThread` to obtain (or reuse) the unfold thread | 86 | +5. Release the MC2 argument object. Call `GetUnfoldThread` to obtain (or reuse) the unfold thread |
| 87 | persisted in the `COMM_ENGINE_CPU_TS` engine context with the `"%s_unfold"` tag, and call | 87 | persisted in the `COMM_ENGINE_CPU_TS` engine context with the `"%s_unfold"` tag, and call |
| 88 | - `HcclStreamAcquireWithThread` to get its bound unfold stream. Finally, call | 88 | + `AcquireUnfoldStream` to get its bound unfold stream. Finally, call |
| 89 | `Mc2CcKernelLaunch(unfoldStream, ccResCtx, ccResCtxSize)` to start the KFC Server on the unfold | 89 | `Mc2CcKernelLaunch(unfoldStream, ccResCtx, ccResCtxSize)` to start the KFC Server on the unfold |
| 90 | stream. | 90 | stream. |
| 91 | 6. Launch the AICore kernel through `all_gather_add_kernel<<<1, nullptr, streamAiv>>>`. Inside the | 91 | 6. Launch the AICore kernel through `all_gather_add_kernel<<<1, nullptr, streamAiv>>>`. Inside the |