已合并
[AscendNPU IR][CV][CVPipelining] Clarify CV lazy load tensor hint #1324
huangyujun创建于 7月3日
[AscendNPU IR][CV][CVPipelining] Clarify CV lazy load tensor hint #1324
已合并
共 2 个文件变更+14-0
| @@ -58,6 +58,13 @@ scf.for 0 to N step 3*S { | |||
| 58 | | Option | Default | Description | | 58 | | Option | Default | Description | |
| 59 | |--------|--------|-------------| | 59 | |--------|--------|-------------| |
| 60 | | `set-workspace-multibuffer` | 2 | Number of software pipeline stages and multi-buffering count | | 60 | | `set-workspace-multibuffer` | 2 | Number of software pipeline stages and multi-buffering count | |
| 61 | +| `--enable-lazy-loading` | false | Enable lazy load in CV pipelining. Load ops can be cloned into multiple work items to reduce intermediate buffer expansion. | | ||
| 62 | + | ||
| 63 | +Lazy load can also be enabled for a specific tensor by adding the `cv_pipeline_lazy_load` compile hint: | ||
| 64 | + | ||
| 65 | +```python | ||
| 66 | +extension.compile_hint(t, "cv_pipeline_lazy_load", True) | ||
| 67 | +``` | ||
| 61 | 68 | ||
| 62 | ## Constraints | 69 | ## Constraints |
| 63 | 70 | ||
| @@ -59,6 +59,13 @@ scf.for 0 to N step 3*S { | |||
| 59 | | 选项 | 默认值 | 含义 | | 59 | | 选项 | 默认值 | 含义 | |
| 60 | |------|--------|------| | 60 | |------|--------|------| |
| 61 | | `set-workspace-multibuffer` | 2 | 软件流水的阶段数，同时也是Multi-Buffering的数量 | | 61 | | `set-workspace-multibuffer` | 2 | 软件流水的阶段数，同时也是Multi-Buffering的数量 | |
| 62 | +| `--enable-lazy-loading` | false | 开启CV Pipelining中的Lazy Load功能，允许将Load op克隆到多个`Work Item`中，以减少中间buffer扩展 | | ||
| 63 | + | ||
| 64 | +也可以在算子侧通过`cv_pipeline_lazy_load`编译提示为指定tensor开启Lazy Load功能： | ||
| 65 | + | ||
| 66 | +```python | ||
| 67 | +extension.compile_hint(t, "cv_pipeline_lazy_load", True) | ||
| 68 | +``` | ||
| 62 | 69 | ||
| 63 | ## 约束能力 | 70 | ## 约束能力 |
| 64 | 71 | ||