Split/merge of size-0 slices

[jacobhinkle]

In a very similar test to that in #365, we hit a different error in FusionKernelRuntime::getKernelConfig.

TEST_F(NVFuserTest, SizeZeroSlice) {
  auto fusion = std::make_unique<Fusion>();
  FusionGuard fg(fusion.get());

  std::vector<int64_t> shape({8});
  auto tv0 = makeContigConcreteTensor(shape);
  fusion->addInput(tv0);

  auto tv1 = slice(
      tv0,
      {{IrBuilder::create<Int>(0),
        IrBuilder::create<Int>(2),
        IrBuilder::create<Int>(1)}});
  auto tv2 = slice(
      tv0,
      {{IrBuilder::create<Int>(2),
        IrBuilder::create<Int>(4),
        IrBuilder::create<Int>(1)}});
  auto tv3 = slice(
      tv0,
      {{IrBuilder::create<Int>(4),
        IrBuilder::create<Int>(6),
        IrBuilder::create<Int>(1)}});
  auto tv4 = slice(
      tv0,
      {{IrBuilder::create<Int>(6),
        IrBuilder::create<Int>(6),
        IrBuilder::create<Int>(1)}});
  auto tv5 = slice(
      tv0,
      {{IrBuilder::create<Int>(6),
        IrBuilder::create<Int>(6),
        IrBuilder::create<Int>(1)}});
  auto tv6 = slice(
      tv0,
      {{IrBuilder::create<Int>(6),
        IrBuilder::create<Int>(8),
        IrBuilder::create<Int>(1)}});
  fusion->addOutput(tv1);
  fusion->addOutput(tv2);
  fusion->addOutput(tv3);
  fusion->addOutput(tv4);
  fusion->addOutput(tv5);
  fusion->addOutput(tv6);

  auto options = at::TensorOptions().dtype(at::kFloat).device(at::kCUDA, 0);
  at::manual_seed(0);

  auto t0 = at::randn(shape, options);
  std::vector<c10::IValue> aten_inputs({t0});

  FusionExecutorCache executor_cache(std::move(fusion));
  executor_cache.runFusionWithInputs(aten_inputs);
}

The error is

[E thread_pool.cpp:109] Exception in thread pool task: Splitting IterDomains with ending values that are 0 is not supported at this time.
Exception raised from split at /opt/pytorch/nvfuser/csrc/ir_nodes.cpp:2333

This happens when we choose a reference tensor that's not empty, but attempt to replay transforms on a size-0 tensor. We probably just need to carefully add size-0 support in IterDomain::split. Note that other incantations also expose this limitation in Merge so they should both be tested.

[jacobhinkle]

Test passes and outputs are correct if I just comment out the check.