File: README

package info (click to toggle)
libxsmm 1.17-4
  • links: PTS, VCS
  • area: main
  • in suites: sid, trixie
  • size: 14,976 kB
  • sloc: ansic: 119,587; cpp: 27,680; fortran: 9,179; sh: 5,765; makefile: 5,040; pascal: 2,312; python: 1,812; f90: 1,773
file content (15 lines) | stat: -rw-r--r-- 1,122 bytes parent folder | download | duplicates (2) 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
 

*Please install TVM by fetching the following commit id from the master branch                                                                                             * Commit ID:  5a27632e274fff57087ed0b6eb2856b6e5946cfb

* Please follow the instuctions in ./libxsmm_wrapper/README to install the libxsmm implementation and wrapper
* Run the script 'mb1_tuned_latest.py' as follows
  * $LD_PRELOAD=./libxsmm_wrapper/libxsmm_wrapper.so python -u mb1_tuned_latest.py -d <layer_name> 
  * layer_name can be any layer from resnet2,resnet3, ..., resnet20
  * These layers are from resnet-50 with minibatch size =1
  * Eg.
    $LD_PRELOAD=./libxsmm_wrapper/libxsmm_wrapper.so python -u mb1_tuned_latest.py -d resnet3 
* A sample slurm job script 'resnet3.slurm' is given to run on a cluster. Please run using:
    $sbatch resnet3.slurm
* Tuning can be paused, by commenting out the tuning function in the script and just measuring the best parameters
* contained in log file. The place to comment out is line no. 406 in mb1_tuned_latest.py
* Result containing the best performance is recorded in a generated excel sheet.