[CODEGEN] LLVM PReLU Error

[nttstar]

Related discussion: https://discuss.tvm.ai/t/debug-failed-to-export-library-with-llvm-on-arm-cpu-when-using-prelu/1006.

It is ok on arm mali gpu and x86 cpu. Only failed on arm cpu. It could be critical as lots of CNN models using prelu.

[tqchen]

Thanks for reporting this issue. It would be helpful to narrow the case down, for example, try to compose a single layer Prelu and see if the error still occurs, this will help to give evidence to other developers.

Likely this problem is due to LLVM or low-level lowering type check issues. cc @masahi @grwlf @ZihengJiang @denis0x0Dwho might be familiar with this part of the code

[denis0x0D]

@nttstar can you please add simple test case, which fails. In this case it would be easy to reduce problem in llvm codegen part.
Thanks.

[FrozenGene]

I can reproduce this issue after 100 trials auto tvm tunning on ARM CPU, but can not reproduce without auto tvm tunning.

input_sym = nnvm.sym.Variable('data')
alpha = nnvm.sym.Variable('alpha')
net = nnvm.sym.conv2d(input_sym, channels=3, kernel_size=(1,1), padding=(0,0), strides=(1, 1), use_bias=False)
net = nnvm.sym.prelu(data=net, alpha=alpha, axis=1)
params = {'alpha': tvm.nd.array(np.ones((3,), dtype=dtype))}

The issue is that when we apply schedule of conv2d combined prelu we emit wrong LLVM IR. For example, we would emit this kind of LLVM IR:

%72 = fcmp ogt <16 x float> %68, zeroinitializer
br <16 x i1> %72, label %if_end, label %if_else

However, br instruction can not accept <N x i1> type, which only accept i1 type. Please see this link: https://llvm.org/docs/LangRef.html#br-instruction

So, we can add condition type check of UnsafeSelectRewriter to avoid this issue. i.e. to make sure condition type is just i1, not other things else.

I can help make a PR to resolve this issue soon.

[denis0x0D]

@FrozenGene, thanks for the answer, good to know you've resolved the issue.