ops.masked with bool inputs leads to compile failure on CPU about pytorch

[lezcano]

cc @jgong5

[isuruf]

The failure seems to come from special casing of bool at

pytorch/torch/_inductor/codegen/cpp.py

Lines 2077 to 2085 in 91bf952

	if dtype == torch.bool:
	# TODO: should we consider load mask here?
	line = f"{self._get_mask_type()}::from({loadbuf})"
	else:
	line = (
	f"{load_mask_str}.template loadu<{cpp_type},{num_vectors}>({loadbuf})"
	if load_mask_str
	else f"{self._get_vec_type(dtype)}::loadu({loadbuf}, {self.tiling_factor})"
	)

Tried fixing this in #126406, but that's not the correct solution.

[isuruf]

Is there a CI job with PYTORCH_ALL_SAMPLES=1 that runs periodically or is that too costly?

Running

PYTORCH_ALL_SAMPLES=1 python test/inductor/test_torchinductor_opinfo.py -k test_comprehensive_constant_pad_nd_cpu_bool -vv

gives us a test case that fails.

[peterbell10]

or is that too costly?

I think it's too broken. Trying to enable all samples testing on as many operators as possible could be a worthwhile project though.

[leslie-fang-intel]

Will take a look for this failure.

[isuruf]

A similar failure can be seen with

PYTORCH_ALL_SAMPLES=1 python test/inductor/test_torchinductor_opinfo.py -k test_comprehensive__chunk_cat_cpu_bool

It fails with

Output:
/tmp/tmp7vma6sgm/xo/cxojdq4ofjj257z6cmt43wndg43l3kt7mx2kvr6mwuq6haam22qj.cpp: In function 'void kernel(const bool*, const bool*, const bool*, bool*, bool*, bool*)':
/tmp/tmp7vma6sgm/xo/cxojdq4ofjj257z6cmt43wndg43l3kt7mx2kvr6mwuq6haam22qj.cpp:86:34: error: operands to '?:' have different types 'at::vec::CPU_CAPABILITY::VecMask<float, 1>' and 'at::vec::CPU_CAPABILITY::Vectorized<bool>'
   86 |                 auto tmp6 = tmp3 ? tmp4() : at::vec::Vectorized<bool>(static_cast<bool>(false));
      |                             ~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

[leslie-fang-intel]

A similar failure can be seen with

PYTORCH_ALL_SAMPLES=1 python test/inductor/test_torchinductor_opinfo.py -k test_comprehensive__chunk_cat_cpu_bool

It fails with

Output:
/tmp/tmp7vma6sgm/xo/cxojdq4ofjj257z6cmt43wndg43l3kt7mx2kvr6mwuq6haam22qj.cpp: In function 'void kernel(const bool*, const bool*, const bool*, bool*, bool*, bool*)':
/tmp/tmp7vma6sgm/xo/cxojdq4ofjj257z6cmt43wndg43l3kt7mx2kvr6mwuq6haam22qj.cpp:86:34: error: operands to '?:' have different types 'at::vec::CPU_CAPABILITY::VecMask<float, 1>' and 'at::vec::CPU_CAPABILITY::Vectorized<bool>'
   86 |                 auto tmp6 = tmp3 ? tmp4() : at::vec::Vectorized<bool>(static_cast<bool>(false));
      |                             ~~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Thanks for reporting it. Will also fix it in the same PR.

[isuruf]

Thanks @leslie-fang-intel for the PR. Opened another issue at #126824