[Relay] support i64 indices

[hzfan]

Remove i32 casting at https://github.com/apache/incubator-tvm/blob/master/src/relay/backend/compile_engine.cc#L83

[tqchen]

@hzfan is this PR ready for review? cc @yzhliu please followup

[hzfan]

@yzhliu @tqchen It's ready for review.

I'm not sure if this pr introduces perf. regression (because some models may contain i64 indexing, and not all i64 can be narrowed).

Do we need a compilation flag like TVM_USE_I64 (defaults to ON) to roll back if regression is detected by users? We can get it removed if no regression is reported for quite some time.

const int64_t* pval = tir::as_const_int(val);
if (pval != nullptr) {
#ifdef TVM_USE_I64  // new behavior introduced by this pr
  res.push_back(val);
#else  // original behavior
  CHECK_LE(pval[0], std::numeric_limits<int32_t>::max());
  CHECK_GE(pval[0], std::numeric_limits<int32_t>::min());
  res.push_back(IntImm(DataType::Int(32), *pval));
#endif
}

[tqchen]

Let us confirm the major networks and also notify the parties. We could have an option TVM_INDEX_DEFAULT_I64 and keep that option for a bit.

cc @jwfromm @kevinthesun @mbaret @merrymercy @FrozenGene

[jwfromm]

LGTM, excited for i64 support to finally make it in!

[tqchen]

Thanks @hzfan @yzhliu @jwfromm !

[trevor-m]

@hzfan FYI, I found that Keras MobileNetV2 model experiences a heavy perf regression with i64 indices enabled.
With ON: 66.55801918395055 FPS
With OFF: 435.48951121558594 FPS

This is on an AWS m5.12xlarge instance.