Use argparse in taichi/main.py to implement #600

[archibate]

Related issue id = #600

[archibate]

Note that no more test = test_python + test_cpp, we use test = test_python now for simplicity and since we usually want test_python. So I use ti test && ti test_cpp in CI ymls.
Will be fixed with further --cpp arguments.

[archibate]

plan A, lower into TI_ARCH:
ti test --arch=opengl numpy
ti debug --arch=opengl a.py

plan B, hack in all_archs:
ti test --arch=opengl numpy
ti test --arch=opengl,metal numpy

AFK take eye rest.

[archibate]

Guess what? Appveyor:

================================== FAILURES ===================================
_______________________ tests/python/test_ad_atomic.py ________________________
[gw1] win32 -- Python 3.6.8 C:\Python36-x64\python.exe
worker 'gw1' crashed while running 'tests/python/test_ad_atomic.py::test_ad_reduce'
=========================== short test summary info ===========================
FAILED tests/python/test_ad_atomic.py::test_ad_reduce
================== 1 failed, 279 passed in 985.21s (0:16:25) ==================

Confusing crash again.. We may want to have an issue for this. Also notable test time 16m25s.. We used to have this time to build&test all done..

[yuanming-hu]

Note that no more test = test_python + test_cpp, we use test = test_python now for simplicity and since we usually want test_python. So I use ti test && ti test_cpp in CI ymls.
Will be fixed with further --cpp arguments.

test_cpp takes < 0.1 second, so I think we can actually enable it by default.

[yuanming-hu]

Guess what? Appveyor:

================================== FAILURES ===================================
_______________________ tests/python/test_ad_atomic.py ________________________
[gw1] win32 -- Python 3.6.8 C:\Python36-x64\python.exe
worker 'gw1' crashed while running 'tests/python/test_ad_atomic.py::test_ad_reduce'
=========================== short test summary info ===========================
FAILED tests/python/test_ad_atomic.py::test_ad_reduce
================== 1 failed, 279 passed in 985.21s (0:16:25) ==================

Confusing crash again.. We may want to have an issue for this. Also notable test time 16m25s.. We used to have this time to build&test all done..

Yeah for some reason on Windows sometimes it fails randomly. Haven't figured out why.

Also it's true that test time needs improvement.

[yuanming-hu]

plan A, lower into TI_ARCH:
ti test --arch=opengl numpy
ti debug --arch=opengl a.py

plan B, hack in all_archs:
ti test --arch=opengl numpy
ti test --arch=opengl,metal numpy

AFK take eye rest.

Plan B sounds better :-)

[archibate]

Already used plan B now :)

[yuanming-hu]

Is this one ready for review?

[archibate]

Is this one ready for review?

Yes!

[archibate]

I've thought about a good idea eval(f'do_taichi_{args.action}()') to make something same as dict, but replaces if-else-if.

[yuanming-hu]

Looks great basically. The only thing to change: make test cover both the old test_python and test_cpp?

[yuanming-hu]

I suggest that we merge test_cpp into test: #601 (comment)

[yuanming-hu]

I've thought about a good idea eval(f'do_taichi_{args.action}()') to make something same as dict, but replaces if-else-if.

I think this is a great idea! We should do it in the future - please feel free to open up an issue to track it. I'm working on the code formatting issues right now.

[archibate]

Btw, isn't it better use return -1 instead of exit(-1)? Might call taichi.main and return codes.

[yuanming-hu]

Btw, isn't it better use return -1 instead of exit(-1)? Might call taichi.main and return codes.

Good point!

[archibate]

Here comes the problem: How can we distinct cpp test_files and python test_files? Still need test_cpp cmd?

[yuanming-hu]

Hmm... Maybe use test -py [files] and test -cpp [files...] to specify the files? It will be much cleaner if a single root command test can be used.

[archibate]

I've thought about a good idea eval(f'do_taichi_{args.action}()') to make something same as dict, but replaces if-else-if.

I think this is a great idea! We should do it in the future - please feel free to open up an issue to track it. I'm working on the code formatting issues right now.

Something even better is to use class TaichiMain, contains do_{args.action} (access using getattr acts like visit_Return), and self.parser (currently so many dirty global..), enable us to have multiple taichi instances, no more re-ti.init for all_archs serialized test.

[yuanming-hu]

I've thought about a good idea eval(f'do_taichi_{args.action}()') to make something same as dict, but replaces if-else-if.

I think this is a great idea! We should do it in the future - please feel free to open up an issue to track it. I'm working on the code formatting issues right now.

Something even better is to use class TaichiMain, contains do_{args.action} (access using getattr acts like visit_Return), and self.parser (currently so many dirty global..), enable us to have multiple taichi instances, no more re-ti.init for all_archs serialized test.

Sounds like a good plan for a future PR. I would merge this in as soon as possible to prevent the changesets of this PR becoming too big...

[yuanming-hu]

Looks great now!! Thank you so much.