What is this?

This is a clone of the deprecated LLVM-based static analyzer from the PyTorch repo, which can be used to produce the PyTorch operator dependency graph.

Where to download the op dependency graph?

GitHub Actions CI job will automatically generate and upload the dependency graph to the Releases page.

How to read the graph?

• In the output yaml file, it lists all the ops that an op might call. For example, the following section shows that quantized::add might call aten::_empty_affine_quantized and etc.

- name: quantized::add
  depends:
  - name: aten::_empty_affine_quantized
  - name: aten::as_strided_
  - name: aten::contiguous
  - name: aten::copy_
...

• If you turn on the debug mode, it will also show why op A calls op B. For example, the following section shows a call path between quantized::add and aten::_empty_affine_quantized.

- name: quantized::add
  depends:
  - name: aten::_empty_affine_quantized
    path:
    - quantized::add
    - at::Tensor at::native::(anonymous namespace)::qadd<false>(at::Tensor, at::Tensor, double, long)
    - at::_empty_affine_quantized(c10::ArrayRef<long>, c10::TensorOptions, double, long, c10::optional<c10::MemoryFormat>)
    - at::_ops::_empty_affine_quantized::call(c10::ArrayRef<long>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>, double, long, c10::optional<c10::MemoryFormat>)
    - at::_ops::create__empty_affine_quantized_typed_handle()
    - aten::_empty_affine_quantized

• It also includes a __ROOT__ section at the beginning, which shows operators that might be called by the runtime directly. For example, the following snippet shows why torch::jit::Unpickler::readInstruction might call aten::_empty_affine_quantized.

- name: __ROOT__
  depends:
  - name: aten::_empty_affine_quantized
    path:
    - __ROOT__
    - torch::jit::Unpickler::readInstruction()
    - at::_empty_affine_quantized(c10::ArrayRef<long>, c10::TensorOptions, double, long, c10::optional<c10::MemoryFormat>)
    - at::_ops::_empty_affine_quantized::call(c10::ArrayRef<long>, c10::optional<c10::ScalarType>, c10::optional<c10::Layout>, c10::optional<c10::Device>, c10::optional<bool>, double, long, c10::optional<c10::MemoryFormat>)
    - at::_ops::create__empty_affine_quantized_typed_handle()
    - aten::_empty_affine_quantized

How to run it locally?

1. Analyze torch and generate yaml file of op dependency transitive closure:

LLVM_DIR=/usr/lib/llvm-8 \
ANALYZE_TORCH=1 ./build.sh

2. Analyze test project and compare with expected result:

LLVM_DIR=/usr/lib/llvm-8 \
ANALYZE_TEST=1 ./build.sh

3. Analyze torch and generate yaml file of op dependency with debug path:

LLVM_DIR=/usr/lib/llvm-8 \
ANALYZE_TORCH=1 ./build.sh -debug_path=true

How is it implemented?

See original PRs:

• LLVM pass: pytorch/pytorch#29550
• Test project: pytorch/pytorch#29716
• Bash driver: pytorch/pytorch#29718