ctuning / ck-autotuning Goto Github PK

Add OpenCL v1.1. stab for compilation opencl caffe conpatible with old phones

Running a compile program on Windows e.g.

$ ck compile program:tool-print-cuda-devices

fails with "md5sum is not recognized as an internal or external command..."

I've nearly solved it by downloading "File Checksum Integrity Verifier" (fciv.exe) from Microsoft into ~CK/ck/bin and placing a file md5sum.bat there:

fciv.exe -md5 %1

But the Microsoft tool prints a comment before the actual checksum:

//
// File Checksum Integrity Verifier version 2.05.
//
2a084556e80926ce59f81cd51520ae08 a.exe

If this is an acceptable solution, CK should ignore the lines starting with //.

We currently read CPU frequency to check if benchmarking experiments changed, but we do not do the same for GPU frequency (we can only set it) - we need to read and report it too ...

Moved from CK: mlcommons/ck#28

Add functionality to compile and run program with each available data set and data set file for testing ...

At some points extend Collective Mind Node (https://play.google.com/store/apps/details?id=com.collective_mind.node) to support OpenCL benchmarks such as SLAMbench and HOG. CMN is used to crowdsource benchmarking, autotuning and predictive analytics.

Moved from CK: mlcommons/ck#43

3.0.0 -> 3.1.0

See related issue from ck-env repo: ctuning/ck-env#7

My issue: I have added a new android machine to my local ck repo. I want to use the update_target_os_dict key inside the machines meta file to update the adb commands to connect via an adb server. This seems to work when I use 'ck run program:...' but when I use 'ck pipeline program:...' I get some issues. The problems relate to the 'target' not being passed to some actions in the program module such as set_freq. These actions seem to handle being passed a 'target' and extracting the device_cfg from it.
I have attached a patch which seems to fix the current issues I am seeing. Hopefully you're happy with the changes?

Thanks,
Stuart

program_module.patch.txt

For now, I explore optimization spaces randomly mainly (the best when you do not have any knowledge about the space). I started adding ML-based exploration via CK plugins as I did in my previous project including CM, but need help (do not have time at the moment).

Moved from CK: mlcommons/ck#16

I installed this android app: https://play.google.com/store/apps/details?id=openscience.crowdsource.experiments, pressed "start" and got this error:

Hi all,

I am using the machine update_target_os_dict parameter to override some target_os parameters. Unfortunately program/module.py was not passing the target when calling platform.gpgpu detect which meant that the update_target_os_dict overrides were not available. This is only an issue for programs which use the need_compute_device flag. The detect action does accept a target parameter so I've provided a patch which provides it to the action if not empty.

Note: I have only tested this when passing a target to the pipeline.

Let me know if you require more information for testing etc.
platform_gpgpu_detect_with_target.patch.txt

Regards,
Stuart

Hi,

Looks like there is a problem with compiling or linking shared libs while doing ck run program on OSX. I do compile and get this:

$ ck compile program:cbench-automotive-susan

***************************************************************************************
Current directory: /Users/dsavenko/CK/ctuning-programs/program/cbench-automotive-susan/tmp
***************************************************************************************

Resolving software dependencies ...

*** Dependency 1 = compiler (C compiler):

    Resolved. CK environment UID = 2a0247258e6fae60 (detected version 3.8.0)

*** Dependency 2 = xopenme (xOpenME library):

    Resolved. CK environment UID = 0dc79501048b7bb2

***************************************************************************************
Detected compiler version: 3.8.0

***************************************************************************************
Compiler vars:
  XOPENME=
***************************************************************************************

Executing prepared batch file tmp-5qVXpo.sh ...

clang -c   -I../ -DXOPENME -I/Users/dsavenko/si/checker/include -I/Users/dsavenko/CK-TOOLS/lib-rtl-xopenme-0.3-llvm-3.8.0-macos-64/include  ../susan.c  -o susan.o
../susan.c:688:9: warning: implicit conversion from 'double' to 'TOTAL_TYPE' (aka 'int') changes value from 0.1 to 0 [-Wliteral-conversion]
  total=0.1; /* test for total's type */
       ~^~~
1 warning generated.
clang    susan.o  -o a.out    -L/Users/dsavenko/CK-TOOLS/lib-rtl-xopenme-0.3-llvm-3.8.0-macos-64/lib -lrtlxopenme -lm
***************************************************************************************
Compilation time: 0.201 sec.; Object size: 47244; MD5: 

(Note there's no MD5 hash which is unusual)

Then, I do run:

$ ck run program:cbench-automotive-susan


OS CK UOA:            macos-64 (fb7d7f63b44b9ea3)

OS name:              Mac OS X 10.11.6
Short OS name:        Darwin 15.6.0
Long OS name:         Darwin-15.6.0-x86_64-i386-64bit
OS bits:              64

Platform init UOA:    42818da3a0789331
***************************************************************************************
Current directory: /Users/dsavenko/CK/ctuning-programs/program/cbench-automotive-susan/tmp

Reloading depedencies from compilation tmp-deps.json ...
***************************************************************************************

Resolving software dependencies ...

*** Dependency 1 = compiler (C compiler):

    Resolved. CK environment UID = 2a0247258e6fae60 (detected version 3.8.0)

*** Dependency 2 = xopenme (xOpenME library):

    Resolved. CK environment UID = 0dc79501048b7bb2


More than one commmand line is found to run this program:

0) corners ($#BIN_FILE#$ $#dataset_path#$$#dataset_filename#$ tmp-output.tmp -c)
1) edges ($#BIN_FILE#$ $#dataset_path#$$#dataset_filename#$ tmp-output.tmp -e)
2) smoothing ($#BIN_FILE#$ $#dataset_path#$$#dataset_filename#$ tmp-output.tmp -s)

Select command line (or Enter to select 0): 

More than one dataset entry is found for this program:

0) image-pgm-0001 (b2130844c38e4a56)
1) image-pgm-clean-gray-square-600-450-8 (c5a378cde441bbb7)
2) image-pgm-from-pamela-100 (d11ad4b49f2b9f45)

Select dataset UOA (or Enter to select 0): 
  Cleaning output files and directories:
    tmp-output1.tmp
    tmp-output2.tmp
    tmp-output.tmp
    tmp-ck-timer.json

***************************************************************************************
### Calibration 1 out of 10 ; Kernel repeat number = 1
***************************************************************************************
***************************************************************************************
Prepared script:

#! /bin/bash


export PATH=/Users/dsavenko/CK/ck-env/platform.init/generic-linux:$PATH


. /Users/dsavenko/CK/local/env/2a0247258e6fae60/env.sh
. /Users/dsavenko/CK/local/env/0dc79501048b7bb2/env.sh

. /Users/dsavenko/CK/local/env/2a0247258e6fae60/env.sh 1

export CT_REPEAT_MAIN=1


export CK_DATASET_PATH=/Users/dsavenko/CK/ctuning-datasets-min/dataset/image-pgm-0001/

export CK_DATASET_FILENAME=data.pgm

echo    executing code ...
 ./a.out /Users/dsavenko/CK/ctuning-datasets-min/dataset/image-pgm-0001/data.pgm tmp-output.tmp -c > tmp-output1.tmp 2> tmp-output2.tmp

***************************************************************************************
  (bash -c "chmod 755 ./tmp-i7Rhe2.sh; . ./tmp-i7Rhe2.sh")

  (sleep 0.5 sec ...)

  (run ...)
executing code ...
./tmp-i7Rhe2.sh: line 20: 20215 Trace/BPT trap: 5       ./a.out /Users/dsavenko/CK/ctuning-datasets-min/dataset/image-pgm-0001/data.pgm tmp-output.tmp -c > tmp-output1.tmp 2> tmp-output2.tmp

  (reading fine grain timers from tmp-ck-timer.json ...)


Program execution likely failed (can't find fine grain timers)!

Stderr (the tmp-output2.tmp file) contains this:

dyld: Library not loaded: librtlxopenme.so
  Referenced from: /Users/dsavenko/CK/ctuning-programs/program/cbench-automotive-susan/tmp/./a.out
  Reason: image not found

The problem reproduces for each program with shared dependencies.

I suspect the library either compiled improperly (e.g. this is no .dylib file), or there may be a problem with linking.

See Interactive Compilation Interface (or plugin framework) for GCC.

Moved from CK: mlcommons/ck#17

Connect CK with GCC and LLVM (LNT) benchmarking infrastructure to crowdsource benchmarking and connect to CK autotuning and predictive analytics ...

This issue was moved from CK: mlcommons/ck#44

Hi @ens-lg4. I just noticed that tests fail on this repo - is it possible to check what is happening, please? Thanks a lot!

I would like to unify how we perform checks of results (output) during auto/crowd-tuning for various types of workloads (exact match or floating point threshold). I started preparing module 'program.output' to keep sample outputs for various data sets for this purpose, but didn't finish it. It's not important to develop it!

On one of my platforms, OpenCL programs only run under super user e.g.

# ck run program:arm-mali-sgemm

It used to be the case that OpenCL programs could be launched under normal user with the --sudo flag:

$ ck run program:arm-mali-sgemm --sudo

Not anymore. It appears that CK doesn't propagate the --sudo flag to the new widget for detecting the compute platform and devices (itself an OpenCL program).

Enhance LLVM (opt, pass manager) to be able to select and reorder arbitrary passes in LLVM via external plugins and connect it to CK (similar to what I did for GCC and plugin interface).

Moved from CK: mlcommons/ck#18

https://github.com/ctuning/ck/wiki/Extensions_openme

Moved from CK: mlcommons/ck#22

Currently in order to make a new program we have to make a copy of some existed one to get non-empty meta populated with correct values: compiler dep, run-time entry and so on.

ck cp some-repo:program:some-cpp-program some-repo:program:another-cpp-program

It would be nice to have some templates, so one could call:

ck add some-repo:program:new-cpp-program --template=cpp

and get correct meta, initial source file, and may be even readme.

Just after previous command, one could call

ck compile new-cpp-program
ck run new-cpp-program

and this should be working and producing some dummy output.

Reported by @w-simon . The last fix caused problems in other pipelines, so I reverted the patch and we need to check where the problem happens. To reproduce:

We test according to the method described in the documentation(https://github.com/ctuning/ck/wiki/Autotuning).

$ rm _setup_program_pipeline_tmp.json
$ ck pipeline program:cbench-automotive-susan --cmd_key=edges --compiler_description_uoa=gcc-44x --prepare --save_to_file=_setup_program_pipeline_tmp.json
$ ck autotune pipeline:program pipeline_from_file=_setup_program_pipeline_tmp.json record=yes
The program runs for a while, but at the end, this error occurs:

CK error: [pipeline] meta is not defined - can't aggregate!

We were asked to check if it's possible to integrate LongBow with CK program workflow (and 'machine' module from 'ck-env' repository) for remote execution and tuning of CK-described workloads in supercomputers. See http://www.hecbiosim.ac.uk/longbow

We currently have support to compile code for Android devices and run them there via

$ ck pull repo:ctuning-programs
$ ck compile program:cbench-automotive-susan --target_os=android19-arm --speed
$ ck run program:cbench-automotive-susan --target_os=android19-arm

It requires Android NDK and SDK installed, as well as device connected via ADB to a host with CK.
CK then takes meta information of ''os:android19-arm'' to understand how to copy data sets, binaries and other files to the device, run them, collect stats and other files and send them back to the host.

We would like to do something similar for Windows-based devices (mobile phone or laptop/deskop) via simple CK node (possibly written in C) that just opens telnet, takes JSON input in CK format, collect files, run binary and send files back - for now that's all. In fact such approach may help us unify running workloads on any remote device (even very small ones if there is no python or java available)...

So the "CK node" should be very simple and portable ...

This ticket is related to the new 'device' module from the 'ck-autotuning' repo:

• ctuning/ck-env#12

Further discussion is required.

Originally, we connected our MILEPOST GCC with the cTuning optimization prediction service:

• http://ctuning.org/wiki/index.php?title=Special:CPredict

Later, we implemented as a Collective Mind service (cTuning3):

• http://cknowledge.ddns.net/cm/repo/?cm_menu=scenarios&cm_submenu=ctuning_pipeline_build_and_run_program_predict_opt

We would like to implement it in CK and connect it with live optimization crowd-tuning results from http://cknowledge.org/repo

It requires mostly then we getting "Segmentation fault" error for some programs.
Work around if

$ ck run program:caffe-time  --keep 
$ ./tmp-PrUGAu.sh #or some other generated file name 
$  valgrind fully_connected

it will be much better to have option and run it it one ck command

$ ck run program:caffe-time  --valgrind

For this need to add valgrind as package and soft
and pass ENV CK_ENV_VALGRIND_TOOL at env.sh

Then also valgrind could be compiled and run under android.