fix auto-tiling output mem-flow bug · Pull Request !504 · MindSpore/akg

<!--  Thanks for sending a pull request!  Here are some tips for you:

If this is your first time, please read our contributor guidelines: https://gitee.com/mindspore/akg/blob/master/CONTRIBUTING.md
-->

**What type of PR is this?**
> Uncomment only one ` /kind <>` line, hit enter to put that in a new line, and remove leading whitespaces from that line:

kind bug

> /kind task
> /kind feature

**What does this PR do / why do we need it**:
This PR fixes three issues found in AutoPoly:

1. `BufferUsageTimetable` erase -> alloc
The "Random tiling" described in issue comes from the use of `BufferUsageTimetable`:

In one version, `T_batchmatmul_input_0_input_1_local_UB` was erased before alloctaing `T_cast_local_UB_0`, so the memory infer size is `786432`; 
![输入图片说明](https://images.gitee.com/uploads/images/2022/0104/103400_dd3e10e9_6572207.png "屏幕截图.png")

while in other version, the erase happened atfer allocation, leading to `1179648` memory infer size. 
![输入图片说明](https://images.gitee.com/uploads/images/2022/0104/103426_aa120ab7_6572207.png "屏幕截图.png")

And the difference in memory size results in different tiling factors.
![输入图片说明](https://images.gitee.com/uploads/images/2022/0104/103440_2494901c_6572207.png "屏幕截图.png")

Therefore, we construct the erase and alloc buffer map ahead of memory infer and do the erase before allocation to ensure consistent memory infer size.

2. `BufferUsageTimetable` output tensor's memory flow
During the construction of `BufferUsageTimetable`, we follow the analysis result of `tensor_mem_flow` to estimate the statements after memory promotion. However, we found that output tensor has a reversed memory flow: 
        `T_batchmatmul_input_0_input_1`(DDR) -> 
        `T_batchmatmul_input_0_input_1_local_UB`(UB) -> 
        `T_batchmatmul_input_0_input_1_local_UB_local_L0C`(L0C)

Therefore, we identify those output tensors during tiling analysis and build the statements for output tensors in a reversed order.

3. GEMM's constraint
We modify the GEMM's auto-tiling strategy to match the actual target constraint.

**Which issue(s) this PR fixes**:

Fixes https://gitee.com/mindspore/akg/issues/I4P75Y

**Special notes for your reviewers**:

MindSpore/akg关闭 .gitee-modal { width: 500px !important; }

搜索帮助

MindSpore/akg
关闭