Skip to content

FrzMtrsprt/PyTorchTraceAnalyzer

 
 

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

37 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

CallingTreeBuilder

API reference

Query class

convert.Query(
    torch_trace="",
    gpu_trace="",
    gpu_trace_count=1
)

The main component of this library.
You can initialize an empty instance by not providing any argument, but it has to be initialized later by calling Query.init().

Parameters

  • torch_trace: str, path to the PyTorch trace file.
  • gpu_trace: str, path to the GPU trace file.
  • gpu_trace_count: int, number of iterations included in the GPU trace file.

Return value

An empty instance of the Query class.

init method

convert.Query.init(
    torch_trace,
    gpu_trace,
    gpu_trace_count=1
)

Loads and analyzes the traces.

Parameters

  • torch_trace: str, path to the PyTorch trace file.
  • gpu_trace: str, path to the GPU trace file.
  • gpu_trace_count: int, number of iterations included in the GPU trace file.

Return value

An loaded instance of the Query class.

loaded method

convert.Query.loaded()

Return value

If the Query class is loaded or not.

query method

convert.Query.query(
    func_name,
    usage_query_type="KernelUsage",
    time_query_type="RangeTime",
    name_query_type="PreciseName"
)

Queries the specifies function, with selected usage calculation method and name matching method.

Parameters

  • func_name: str, the function name to query.
  • usage_query_type: str, method used to calculate the GPU usage sum. Can be "KernelUsage" or "RangeUsage".
  • time_query_type: str, method used to calculate the time sum. Can be "KernelTime" or "RangeTime".
  • name_query_type: str, name matching mathod. Can be "FuzzyName" or "PreciseName".

Return value

The average GPU usage of matching functions, in float.

query_forward method

convert.Query.query_forward(
    usage_query_type="KernelUsage",
    time_query_type="RangeTime"
)

Queries the forward layer, with selected usage calculation method.

Parameters

  • usage_query_type: str, method used to calculate the GPU usage sum. Can be "KernelUsage" or "RangeUsage".
  • time_query_type: str, method used to calculate the time sum. Can be "KernelTime" or "RangeTime".

Return value

The average GPU usage of forward layer, in float.

query_backward method

convert.Query.query_backward(
    usage_query_type="KernelUsage",
    time_query_type="RangeTime"
)

Queries the backward layer, with selected usage calculation method.

Parameters

  • usage_query_type: str, method used to calculate the GPU usage sum. Can be "KernelUsage" or "RangeUsage".
  • time_query_type: str, method used to calculate the time sum. Can be "KernelTime" or "RangeTime".

Return value

The average GPU usage of backward layer, in float.

query_optimizer method

convert.Query.query_optimizer(
    usage_query_type="KernelUsage",
    time_query_type="RangeTime"
)

Queries the optimizer, with selected usage calculation method.

Parameters

  • usage_query_type: str, method used to calculate the GPU usage sum. Can be "KernelUsage" or "RangeUsage".
  • time_query_type: str, method used to calculate the time sum. Can be "KernelTime" or "RangeTime".

Return value

The average GPU usage of optimizer, in float.

query_module method

convert.Query.query_modules(
    module_name,
    usage_query_type="KernelUsage",
    time_query_type="RangeTime"
)

Queries the specified module, with selected usage calculation method.

Parameters

  • modules_name: str, the module name to query.
  • usage_query_type: str, method used to calculate the GPU usage sum. Can be "KernelUsage" or "RangeUsage".
  • time_query_type: str, method used to calculate the time sum. Can be "KernelTime" or "RangeTime".

Return value

The average GPU usage of the specified module, in float.

Build

  1. Compile CMakeLists.txt
mkdir build
cd build
cmake ..
make
cd ..
  1. Build Python package
python3 -m build --wheel

About

No description, website, or topics provided.

Resources

Stars

Watchers

Forks

Releases

No releases published

Packages

No packages published

Languages

  • C++ 78.0%
  • Python 14.0%
  • C 4.1%
  • CMake 3.3%
  • Shell 0.6%