Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[RFC][AutoTVM] Tuning with subgraph granularity #64

Open
wants to merge 1 commit into
base: main
Choose a base branch
from
Open
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
135 changes: 135 additions & 0 deletions rfcs/0064-autotvm-tune-with-subgraph-granularity.md
Original file line number Diff line number Diff line change
@@ -0,0 +1,135 @@
- Feature Name: AutoTVM tuning with Subgraph Granularity
- Start Date: 2022-3-17
- RFC PR: [apache/tvm-rfcs#0064](https://github.com/apache/tvm-rfcs/pull/0064)
- GitHub Issue: N/A

# Summary
[summary]: #summary

This RFC introduces why and how we tune with subgraph granularity.

# Motivation
[motivation]: #motivation

During performance optimization for platform Xavier which has a Volta GPU in it, we found that tuning by AutoTVM with subgraph as granularity could bring performance improvement. Because the data type of the subgraph's output may be different from the data type of the subgraph's anchor operator's output, this may change the task from memory-bound to compute-bound or change in reverse.
Let's take the subgraph in the figure below as an example. If we tune with the single convolution the output data type is 'Int32' but if we tune with the subgraph the output data type is 'Int8'. The former's data size is four times the latter one. But in the actual inference, the data type is 'Int8' same as the latter one. So the best config searched by tuning with a single operator maybe not be the best for the subgraph.
![image](assets/9999/subgraph-example.png)
We also run an example to verify the theory above.

- We wrote a schedule marked as ```ScheduleA``` for the subgraph above by hardcode and the latency of subgraph inferencing is 104 microseconds. Then we tuned the subgraph with single op as granularity. In the tuning log we found ```ScheduleA``` and the latency recorded in the measurement result is 329 microseconds.
- The best schedule from the tuned log in the step above is marked as ```ScheduleB``` and the latency recorded in the measurement result is 237 microseconds. The latency of subgraph inferencing with ```ScheduleB``` is 224 microseconds.

From the example above we can tell AutoTVM would not find ```ScheduleA```, the obvious better schedule. This means the tuning result is distorted, the distortion would be more obvious if the shape of the output became bigger.
Comment on lines +19 to +22
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is a great motivating example. I believe this is also an important reason that auto-scheduler could have end-to-end speedup over AutoTVM.


# Guide-level explanation
[guide-level-explanation]: #guide-level-explanation

We propose a method to tune with subgraph as granularity by AUTOTVM. As we know, tuning with a single operator we need function ```fcompute``` which computes the output tensors of the operator and function ```fschedule``` which schedules the computation. With these two functions, we can build a measurable program running on the device. So the key problem is assigning function ```fcompute``` and function ```fschedule``` for subgraph. In this PR, we use the fucntion ```fschedule``` of the anchor operator as the subgraph's function ```fschedule```. As for function ```fcompute```, its purpose is getting output tensors of the subgraph. The function ```LowerToTECompute::Lower``` can get the output tensors of the subgraph, so we can use these tensors as output of subgraph's function ```fcompute```. And a GLOBAL_SCOPE.tune_subgraph option is introduced to control tuning with single operator or subgraph, default vaule is `False`.
The whole process can breakdown into two major phases.
1. Task extracting and tuning.
1.1 Select best implementation to compute output for anchor operator and record implementation name `best_impl_name`
1.2 Lower subgraph to `outputs` by `LowerToTECompute`
1.3 Create subgraph tuning task name `task_name` with subgraph name and `iotensors` extracted from `outputs`.
1.4 Add subgraph task with `task_name`, `iotensors` and `best_impl_name`.
1.5 Create `workload` for subgraph tuning task with `task_name` and `iotensors`.
1.6 Set function `fcompute` for subgraph by returning `outputs` .
1.7 Set function `fschedule` for subgraph by querying table with `best_impl_name`.
1.8 Tune.

1. Building.
2.1 Apply the best history.
2.2 Select best implementation to compute output for anchor operator and record implementation name `best_impl_name`
2.3 Lower subgraph to `outputs` by `LowerToTECompute`
2.4 Create subgraph tuning task name `task_name` with subgraph name and `iotensors` extracted from `outputs`.
2.5 Create `workload` for subgraph tuning task with `task_name` and `iotensors`.
2.6 Lower schedule with the best config queried by `workload`.
2.7 Codegen.

# Reference-level explanation
[reference-level-explanation]: #reference-level-explanation

In steps 1.1 and 2.2 mentioned in the previous section, returning the best implementation name is kind of tricky. We can get the `best_plevel_impl` in function `select_implementation`, but sometimes the actual implementation name is not `best_plevel_impl.name`. For example, implementation for `conv2d_nchw.x86` is added like this.

```python
@conv2d_NCHWc_strategy.register("cpu")
def conv2d_NCHWc_strategy_cpu(attrs, inputs, out_type, target):
"""conv2d_NCHWc x86 strategy"""
strategy = _op.OpStrategy()
strategy.add_implementation(
wrap_compute_conv2d(topi.x86.conv2d_nchw),
wrap_topi_schedule(topi.x86.schedule_conv2d_nchw),
name="conv2d_nchw.x86",
)
return strategy
```

But `topi.x86.conv2d_nchw` wrappers another implementation.

```python
def conv2d_nchw(data, kernel, strides, padding, dilation, out_dtype):
layout = "NCHW"
packed_out = conv2d_NCHWc(data, kernel, strides, padding, dilation, layout, layout, out_dtype)
return unpack_NCHWc_to_nchw(packed_out, out_dtype)
```

The actual implementation name is `conv2d_NCHWc.x86`. So in function `select_implementation` we fixed this problem by getting the actual implementation name from the workload which is created in function `register_topi_compute`.

```python
def select_implementation():
# ignore some codes

if GLOBAL_SCOPE.tune_subgraph:
# In some cases, one strategy's compute may call another compute.
# So the impl name need to match with actual compute.
if workloads[best_plevel_impl]:
workload = workloads[best_plevel_impl]
if best_plevel_impl.name != "injective.cpu" and best_plevel_impl.name != workload[0]:
best_plevel_impl.name = workload[0]
# value changed in python side will not effect C++ side,
# so here need to pass new name to C++
return best_plevel_impl, outputs[best_plevel_impl], best_plevel_impl.name
```

Because the value changed on the python side will not affect the C++ side, we need to pass the new name to C++. This results in the function `select_implementation` number of returned values changing to three. This should be noticed.


# Drawbacks
[drawbacks]: #drawbacks

There are three cases in which tuning with subgraph may not get better performance.

Case 1: Anchor operator has more than one implementation.
We register the subgraph tuning task by `outputs` in step 1.4. No matter how many implementations the anchor operator has, step 1.1 will only pick the implementation with the highest level and return outputs computed by it. So the subgraph tuning tasks may not contain the potential best implementation.
Comment on lines +101 to +102
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

This is my major concern too, as one important feature in AutoTVM is transparent the implementation selection to the tuning results. This is also a challenge issue that auto-scheduler hasn't figured out yet. As you pointed out, selecting an implementation based on the plevel before tuning could result in sub-optimal, is there anyway to still generate N subgraph tasks (N=#impl) for the anchor op?

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Generate N subgraph tasks means we need N subgraph outputs and that means we need to run LowerToTECompute::Lower N times. It will be a big change to the current framework but I think it will work. Even we can resolve this case, case 2 is still a problem preventing anchor op to find the best config from N tuning results.


Case 2: Anchor operator's function `fcompute` needs value from config such as code block below. In step 2.2, computing output will call function `_pack_data` and the `cfg` suppose to be the best config of the subgraph. But in step 2.2 we don't know which subgraph the anchor operator belongs to yet, so we cannot get the right config from the best history and fallback to the default one. This may bring great performance regression.
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Do you see a potential way around this. All the cases I've wanted subgraph tuning in autotvm are around layout rewriting.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I feel it's the same issue as my previous comment. If we could generate N tasks based on the implementations, we should be able to workaround this.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Do you see a potential way around this. All the cases I've wanted subgraph tuning in autotvm are around layout rewriting.

Do not yet.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I feel it's the same issue as my previous comment. If we could generate N tasks based on the implementations, we should be able to workaround this.

I think they are not the same. The key problem, in this case, is the tunning results are recorded with the subgraph task name as workload key but in step 2.2 the subgraph task name is not created yet because the ops after the anchor op are not visited yet. So we don't know which tuning log we should employ.


```python
def _pack_data(cfg, data, kernel):
n, _, ih, iw = get_const_tuple(data.shape)
oc, ic, kh, kw = get_const_tuple(kernel.shape)
ic_bn, oc_bn = cfg["tile_ic"].size[-1], cfg["tile_oc"].size[-1]
......
```

Case 3: During task extraction subgraph is lowered by VMCompile, and pass `AlterOpLayout` is disabled, see [code](https://github.com/apache/tvm/blob/main/python/tvm/autotvm/task/relay_integration.py#:~:text=with%20tvm.transform.PassContext(opt_level%3Dopt_level%2C%20disabled_pass%3D%7B%22AlterOpLayout%22%7D)%3A). So during building phash we need to disable pass `AlterOpLayout` too, otherwise the subgraphs generated in the task extracting phase may be different from those generated in the building phase.


# Rationale and alternatives
[rationale-and-alternatives]: #rationale-and-alternatives

The core problem of tune subgraphs is getting output tensors of subgraphs. Employing function `LowerToTECompute::Lower` can achieve minimal change to the current framework.

# Prior art
[prior-art]: #prior-art

Our implementation is inspired by auto-schedule.

# Unresolved questions
[unresolved-questions]: #unresolved-questions
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

How are you handling serialization of the tuning results? Autotvm logs just contain the implementation name, but that is not specific enough for subgraphs. Will you use the structural hash like auto scheduler?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Same question here. This also determines how general the tuned schedule could be. For example, considering the following hashing approaches:

H1: Similar to auto-scheduler, such as ["impl-name", "args", "compute DAG MD5"]. This guarantees that the tuned schedule will be applied to ONLY exactly the same subgraph.

H2 Canonicalize compute DAG. For example, canonicalize a compute DAG cast - dense - cast to be elementwise - dense - elementwise. Since different elementwise ops in this DAG should result in almost the same performance, canonicalization makes the tuned result more general applicable. The drawback is we need to be careful to make sure the canonicalized compute DAG won't cover too many undesired cases.

Copy link
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

H1: Similar to auto-scheduler, such as ["impl-name", "args", "compute DAG MD5"]. This guarantees that the tuned schedule will be applied to ONLY exactly the same subgraph.

We use the first approach. We serialize the iotensors of the subgraph to a hashable tuple and compute the tuple MD5 for getting a hash key. Then use subgraph name + hash key as the subgraph task name.


See Section Drawbacks.

# Future possibilities
[future-possibilities]: #future-possibilities

Resolve the 3 drawbacks listed above.
Binary file added rfcs/assets/0064/subgraph-example.png
Loading
Sorry, something went wrong. Reload?
Sorry, we cannot display this file.
Sorry, this file is invalid so it cannot be displayed.