torch.compile (inductor) bug random signed number generation about pytorch HOT 6 CLOSED

Hi @hgreving2304, thanks for reporting the issue. Will take a look.

from pytorch.

Test with the latest PyTorch: 705346b, the result is correct as cpu is tensor([[[-1]]]). @hgreving2304 could you take a try of it?

from pytorch.

I don't have access right now to the latest version, I can try it out later at home. In the meantime, you could also check whether the emitted C-code correctly shifts a signed vs. an unsigned number in the latest version and whether there was an explicit fix.

from pytorch.

Here are the generated code, randint64_cpu returns type of int64 and we use it to do the bit shift.

cpp_fused_bitwise_right_shift_0 = async_compile.cpp_pybinding(['int64_t*'], '''
#include "/tmp/torchinductor_leslie/sk/cskh5dx62fglpphcrl6723dnmowdabouerrzy3dmqcngbxwfa7bv.h"
extern "C" void kernel(int64_t* in_out_ptr0)
{
    {
        auto tmp0 = in_out_ptr0[static_cast<long>(0L)];
        auto tmp1 = static_cast<int32_t>(0);
        auto tmp2 = static_cast<int64_t>(-1);
        auto tmp3 = static_cast<int64_t>(0);
        auto tmp4 = randint64_cpu(tmp0, tmp1, tmp2, tmp3);
        auto tmp5 = static_cast<int64_t>(31);
        auto tmp6 = decltype(tmp4)(tmp4 >> tmp5);
        in_out_ptr0[static_cast<long>(0L)] = tmp6;
    }
}
''')

from pytorch.

Just tried, with the latest version from Ubuntu I am getting a signed shift, fwiw that's fine. I recommend verifying that there was an explicit fix, thanks..

from pytorch.

Hi @hgreving2304, I am going to close this ticket since the UT has added in PyTorch Main. Feel free to reopen if any further discussion.

from pytorch.