APMIDG: Argo Power Management Glue Layer for Intel Discrete GPUs

Requirements:

• Intel Discrete GPUs
• oneAPI environment (header files and shared libraries)
• cmake 3.9 or later

To build shared libraries and testcodes

$ mkdir build && cd build
$ cmake -DCMAKE_INSTALL_PREFIX:PATH=__LIBAPMIDG_INSTALL_PATH__ ..
$ make install

NOTE: please replace __LIBAPMIDG_INSTALL_PATH__ with your preferred path.

To run test codes

$ export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:__LIBAPMIDG_INSTALL_PATH__/lib64
$ export PATH=$PATH:__LIBAPMIDG_INSTALL_PATH__/bin

C-based testcode
$ apmidgstats

Python binding demo
$ python3
>>> import pyapmidg
>>> pm = pyapmidg.clr_apmidg()
>>> pm.readpoweravg()  # read the average power of domain0 on device0
xx.x
>>> pm.getndevs() # return the number of GPUs
x
>>> pm.getnpwrdoms(0) # return the number of the power domains on device0. Note the number starts at 0
x
>>> pm.getnfreqdoms(1) # return the number of the freq domains on device1
x
>>> pm.getntempsensors(0) # return the number of the temp sensors on device0
x
>>> pm.readpoweravg(1,0)  # read the average power of domain0 on device1
>>> flims = pm.getfreqlims(1,0) # read the current freq limit of domain0 on device1
>>> flims.max_MHz
xxxx.x
>>> flims.max_MHz = xxxx
>>>> pm.setfreqlims(1, 0, flims.min_MHz, flims.max_MHz)
>>> pm.reset2default() # reset back to the default setting

NOTE:

• See src/pyapmidg/demo_monitor_articus for Python API usages
• C examples are available in src/c_examples

-- Contact: Kazutomo Yoshii [email protected]