Command Line Arguments

The WFA uses the argparse library to automatically add compiler options to any script that calls the WFA.WSE_Interface.WSE_Interface.make_WSE method. This feature enables rapid development, supports debugging, and compiling larges scale executables for hardware runs with just a few keystrokes.

Just add -h behind any python script with the make_WSE method called in it to get a help menu. It is possible to add domain specific command line arguments to a program with the get_cmd_line_parser and get_available_cmd_args methods as seen in the above examples. Use the WSE_Interface.str_to_bool method in as the input to the type keyword argument in the parser.add_argument method for boolean inputs.

The -h option will produce an output similar to:

usage: test_center_add.py [-h] [-dl DEBUG_LEVEL] [-lp LAUNCH_PORTRAIT]
              [-mg MAKE_GOAL] [-vr VERIFY_RESULT]
              [-ptv PRINT_TILE_VERIFICATION]
              [-hin HARDWARE_IMAGE_NAME]
              [-ast ANALYZE_SIMULATOR_TIMESTAMPS]
              [-hon HARDWARE_OUTPUT_NAME] [-to TIMEOUT]
              [-att ADD_TASK_TAGS] [-ts32 TIME_STAMPS_32BIT]
              [-ts16 TIME_STAMPS_16BIT] [-tsm TIME_STAMP_MODULO]
              [-ppc POWER_PERFORMANCE_COUNTER]
              [-dml DISPLAY_MAKE_LOG] [-x X] [-y Y] [-z Z]

Command Line Parser for WSE Interface

optional arguments:
  -h, --help            show this help message and exit
  -dl DEBUG_LEVEL, --debug_level DEBUG_LEVEL
                        Sets the level of debug. 0:off 1:create portrait files
                        2:JOB=-
  -lp LAUNCH_PORTRAIT, --launch_portrait LAUNCH_PORTRAIT
                        launches portrait
  -mg MAKE_GOAL, --make_goal MAKE_GOAL
                        specify make goal
  -vr VERIFY_RESULT, --verify_result VERIFY_RESULT
                        toggle verification of results through NumPy
  -ptv PRINT_TILE_VERIFICATION, --print_tile_verification PRINT_TILE_VERIFICATION
                        enables printing the memory space from the host during
                        validation. Specify Tx.y where x and y are the tile
                        coordinates from passing
  -hin HARDWARE_IMAGE_NAME, --hardware_image_name HARDWARE_IMAGE_NAME
                        specifying a name creates files for a hardware run
  -ast ANALYZE_SIMULATOR_TIMESTAMPS, --analyze_simulator_timestamps ANALYZE_SIMULATOR_TIMESTAMPS
                        turns on analysis of time stamps from Simfabric
  -hon HARDWARE_OUTPUT_NAME, --hardware_output_name HARDWARE_OUTPUT_NAME
                        specifying on output image name turns on hardware
                        analysis
  -to TIMEOUT, --timeout TIMEOUT
                        specifies the allowable time for a Simfabric run
  -att ADD_TASK_TAGS, --add_task_tags ADD_TASK_TAGS
                        compiles task tags into ag files for debugging
  -ts32 TIME_STAMPS_32BIT, --time_stamps_32bit TIME_STAMPS_32BIT
                        set time stamps to 32 bit recording
  -ts16 TIME_STAMPS_16BIT, --time_stamps_16bit TIME_STAMPS_16BIT
                        set time stamps to 16 bit recording
  -tsm TIME_STAMP_MODULO, --time_stamp_modulo TIME_STAMP_MODULO
                        specify the frequency of recording time stamps
  -ppc POWER_PERFORMANCE_COUNTER, --power_performance_counter POWER_PERFORMANCE_COUNTER
                        turns on power performance counter recording
  -dml DISPLAY_MAKE_LOG, --display_make_log DISPLAY_MAKE_LOG
                        displays make log from wse make
  -x X, --x X           specify x dimension
  -y Y, --y Y           specify y dimension
  -z Z, --z Z           specify z dimension

Running Scripts for Development

Running the script without any command line options will cause the API to create the .ag, .paint, and Makefile, compile the code into a binary, run the hardware simulator (Simfabric) and compare the output to a NumPy based validation code within the API. The correct result is specified in make_wse through the answer keyword argument.

This configuration is a great way to debug the code during development. It is not recommended that the problem size be very large if run in this mode. Simfabric is very useful, but it is not very fast and can consume considerable memory if the problem size is set to a large size.

If the code ran successfully and produced the same answer as the validation code it will produce an output similar to:

********************* Starting Make ***********************
********************* Writing binary input files ***********************
    Total vectors to write:  6
    Total Data to Write (MB):  0.024
    Total T_bin to Write (MB):  0.01536
**************** Done writing binary input files ***********************
********************* Started Host Validation ***********************
********************* Done Host Validation ***********************
********************* Running Cerebras Tools ***********************
examples                :  0:03.71 PASS smoke [ H=4 W=4 D=100 NC=8 NArgs=6 COMPILER=angler ]
********************* Finished Running Cerebras Tools ***********************
********************* Done Make ***********************

The PASS indicates that Simfabric ran and the validation/Simfabric output provided in answer were identical.

Debugging

The following command line set is quite useful. The -ptv T4.5 option causes the validation code to print out the source and destination memory space for the tile at coordinates of (4,5) in validation.txt for every operation. This can then be compared to the Cerebras debug tool, Portrait, for each operation to debug the code. The -dl 1 option launches the portrait sever which can be connected to in a browser at http://localhost:8080/paint.html

$ python test_add.py -ptv T4.5 -dl 1

Setting debug_level to 2 is equivalent to setting GOAL=debug JOB=- flags in the Cerebras make process. This will cause Simfabric to dump out detailed log files and launch the Portrait server. It will also ignore any tests and force a pass to allow portrait to launch even if a fatal condition was detected in the Simfabric run. This is useful when debugging failing microcode.

Setting debug_level to 1 or 2 will also turn on more printing to screen for the API

Compiling for Hardware

It is not recommended to run the validation or Simfabric for large scale runs. Running either the NumPy validation and/or Simfabric at the maximum scales that the WSE can run can take many days to complete on a single node. It is thus recommended that development be done on the smallest size problem that is still meaningful and that it then be scaled up for a hardware run and all validation be turned off.

This can be easily done through the command line as follows

$ python rectangular_diffusion.py -x 800 -y 1000 -z 2000 -vr 0 -mg bench.img -hin full_scale_implicit_diffusion

If the code contains any timers, they can be turned on with the -ts16 or -ts32 arguments