Wrong vector shift results on PowerPC

[Flamefire]

🐛 Describe the bug

This is basically #70904 which is still present for PPC after #98511 added vector operations.

Basically: Shifting a vector with out-of-bounds values results in undefined behavior and sporadic wrong results.
E.g. test_non_contig_bitwise_right_shift_cpu_int32 fails because of right shifts with too-large values and test_shift_limits_cpu_int16 fails due to left shifts with negative values. The latter is very much understandable given that #98511 uses a reinterpret_cast to the unsigned vector type

Reproduced with python run_test.py -i test_binary_ufuncs which fails

• test_contig_vs_every_other_bitwise_right_shift_cpu_int16
• test_contig_vs_every_other_bitwise_right_shift_cpu_int32
• test_contig_vs_every_other_bitwise_right_shift_cpu_int64
• test_non_contig_bitwise_right_shift_cpu_int16
• test_non_contig_bitwise_right_shift_cpu_int32
• test_non_contig_bitwise_right_shift_cpu_int64
• test_shift_limits_cpu_int16
• test_shift_limits_cpu_int32
• test_shift_limits_cpu_int64

Versions

Using PyTorch 2.0.1 with the above PRs applied as patches.

[quickwritereader]

I see. could you remove the reinterpret_cast s and see if it complies to the desired.

[Flamefire]

I see. could you remove the reinterpret_cast s and see if it complies to the desired.

No that does not work because vec_sl is only supported for unsigned rhs which matches C++ specs where negative shifts and shifts greater than the bitsize are UB.

My current workaround is the first approach of #98511 with an #ifdef for the shift kernels using cpu_kernel with the explicit checks for wrong shift values introduced in #96659

An alternative I can see is to implement the vectorized version of the checks present in e.g. the AVX code for VSX. And likely use vec_sla/vec_sra instead of vec_sl/vec_sr as seemingly arithmetic shifts are intended