[test] Test test_ad_basic.py failed on MacOS Mojave

[Rullec]

Describe the bug
I have a developement build on my MacOS, and didn't manage to pass the test of test_ad_basic.py.

Logs shown as below:

To Reproduce
pytest test_ad_basic.py

Logs

========================= FAILURES ==========================
____________________ test_trigonometric _____________________

args = (), kwargs = {}

    def wrapper(*args, **kwargs):
        if has_autograd:
>           func(*args, **kwargs)

test_ad_basics.py:17: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
test_ad_basics.py:80: in test_trigonometric
    grad_test(lambda x: ti.sin(x), lambda x: np.sin(x))
test_ad_basics.py:50: in grad_test
    impl()
../../python/taichi/lang/__init__.py:449: in wrapped
    test(*test_args, **test_kwargs)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _

    @ti.all_archs_with(default_fp=default_fp)
    def impl():
        print(f'arch={ti.cfg.arch} default_fp={ti.cfg.default_fp}')
        x = ti.var(default_fp)
        y = ti.var(default_fp)
    
        ti.root.dense(ti.i, 1).place(x, x.grad, y, y.grad)
    
        @ti.kernel
        def func():
            for i in x:
                y[i] = tifunc(x[i])
    
        v = 0.234
    
        y.grad[0] = 1
        x[0] = v
        func()
        func.grad()
    
>       assert y[0] == approx(npfunc(v))
E       assert 0.23188334703445435 == 0.2318703549116868 ± 2.3e-07
E        +  where 0.2318703549116868 ± 2.3e-07 = approx(0.2318703549116868)
E        +    where 0.2318703549116868 = <function test_trigonometric.<locals>.<lambda> at 0x123d80378>(0.234)

test_ad_basics.py:47: AssertionError
------------------- Captured stdout call --------------------
Running test on arch=Arch.x64
[Taichi] Starting on arch=x64
arch=Arch.x64 default_fp=DataType.float32
[Taichi] materializing...
Running test on arch=Arch.metal
[Taichi] Starting on arch=metal
arch=Arch.metal default_fp=DataType.float32
[Taichi] materializing...
Running test on arch=Arch.x64
[Taichi] Starting on arch=x64
arch=Arch.x64 default_fp=DataType.float32
[Taichi] materializing...
Running test on arch=Arch.metal
[Taichi] Starting on arch=metal
arch=Arch.metal default_fp=DataType.float32
[Taichi] materializing...

[archibate]

assert 0.23188334703445435 == 0.2318703549116868 ± 2.3e-07

As you can see the error here is really small, that it should not fail at all.

I think we can do a mock in ti.approx to increase the rel tolerance a little bit like OpenGL does:

taichi/python/taichi/testing.py

Lines 22 to 23 in bf403b5

	if ti.cfg.arch == ti.opengl:
	kwargs['rel'] = max(kwargs.get('rel', 1e-6), 1e-3)

@Rullec would you do the same thing to metal? A good issue for you :)

[Rullec]

``

As you can see the error here is really small, that it should not fail at all.

Yes! It's the same as I guess. It's always hard to find a nice "epsilon".

@Rullec would you do the same thing to metal? A good issue for you :)

Thanks! I will try to handle it under all of your guidence.

[Rullec]

Btw, The other 3 tests failed on my mac as well:

[error] error when testing test_bitmasked.py, return 256
[error] error when testing test_global_store_grad.py, return 1280
[error] error when testing test_ad_basics.py, return 256
[error] error when testing test_element_wise.py, return 256

I will double check them later, and if they are not my wrong, would it be proper to report them as well in this issue?