ADLXPybind

This document describes how to build the Python bindings for ADLX, as well as the environment in which they can be executed.

Documentation to C/Cpp ADLX can be found in https://gpuopen.com/manuals/adlx/adlx-page_interfaces/

1. Building

Supported Operating Systems

• Windows® 11 (32- and 64-bit versions) and up.

Build Tool

• Visual Studio 2022 with C and C++ components for desktop.

Build Prerequisites

• ADLX SDK is installed.
• Python 3.12 is installed.
• pybind11 is installed.

The path to the Python 3.12 installation is set to the PYHOME system environment variable.

Note: The ADLX library is installed along with the AMD display driver.

Build Instructions

• Run Visual Studio 2022.
• Open ~ADLXPybind\ADLXPybind.sln.
• Rebuild the project.

2. Packing into wheel

Note: before packing bindings into wheel please copy compiled ADLX.pyd file to unit test folder: ~ADLXPybind\adlx_tests and run the unit tests, be sure that all tests passed, infomation how to run unit tests can be found here: ~ADLXPybind\adlx_tests\_readme.txt

• copy ADLX.pyd file into: ~ADLXPybind\build_wheel\adlx
• update bidnings verion in file: ~ADLXPybind\build_wheel\libversion.txt
• run bat file: ~ADLXPybind\build_wheel\build_wheel.bat
• wheel file will be generated in folder: ~ADLXPybind\build_wheel\build_wheel\dist

3. Usage example

• install wheel in your Python environment, example: $pip install <path_to_wheel>
• now you can import and use the bindings:

from typing import List
from adlx import ADLX


def check_tunning_services(system:ADLX.IADLXSystem):
    """Prints if gpu support autotingin

    :param system: ADLX System Interface
    :type system: ADLX.IADLXSystem
    """
    tuning_services: ADLX.IADLXGPUTuningServices = system.GetGPUTuningServices()
    gpus: List[ADLX.IADLXGPU] = system.GetGPUs()
    
    for gpu in gpus:
        is_supported_auto_tuning = tuning_services.IsSupportedAutoTuning(gpu)
        asic_name = gpu.Name()
        if is_supported_auto_tuning:
            print(f"GPU Auto Tuning is supported on asic: {asic_name}")
        else:
            print(f"GPU Auto Tuning is NOT supported on asic: {asic_name}")


def main():
    adlx: ADLX = ADLX.ADLXHelper()
    ret: ADLX.ADLX_RESULT = adlx.Initialize()
    assert ret == ADLX.ADLX_RESULT.ADLX_OK

    system: ADLX.IADLXSystem = adlx.GetSystemServices()
    check_tunning_services(system=system)
    ret: ADLX.ADLX_RESULT = adlx.Terminate()
    assert ret == ADLX.ADLX_RESULT.ADLX_OK

if __name__ == "__main__":
    main()

Contribution process

• Expose the desired API (C++ bind.cpp)
• Write pytest tests for the added API ~ADLXPybind\adlxtest
• Build the .pyd binary and store it in ~ADLXPybind\adlxtest
• Run the new API test and if possible, run regression tests
• Create wheel package, don't forgate to increase version number
• Create a PR in this repo, showing the test case passed
• after PR is approved, deploy the .whl package in the ADLX