computeAt with multiple use sites

[naoyam]

❓ Questions and Help

What is expected behavior when a single Val is computed at multiple sites? Here's a simple example:

TensorView* t0 = makeDummyTensor(2);
fusion.addInput(t0);
auto t1 = unaryOp(UnaryOpType::Exp, t0);
auto t2 = reductionOp(BinaryOpType::Max, {-1}, new Float(0), t1);
auto t3 = reductionOp(BinaryOpType::Min, {-1}, new Float(0), t1);
auto t4 = add(t2, t3);
fusion.addOutput(t4);
t1->computeAt(t2, -1);

At the computeAt, this is what I got:

C++ exception with description "Tried to access position 1 in domain: [ iS{i1} ]
Exception raised from axis at ../torch/csrc/jit/codegen/cuda/tensor_view.cpp:73 (most recent call first):
frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >) + 0x68 (0x7f3bb891ed78 in /home/nmaruyama/pytorch/src/csarofeen.2/build/lib/libc10.so)

Is this a user error? If not, what is the expected result?

[csarofeen]

Looks like a bug.
If I'm reading this right, it should effectively run

t1->computeAt(t2, -1)
t1->computeAt(t3, -1)
t3->computeAt(t4, -1)
```

[naoyam]

OK. I'll create a reproducer and file it as a bug.

[naoyam]

To reproduce the bug, run testGPU_FusionComputeAtMultiSites, which can be found at:

https://github.com/naoyam/pytorch/tree/computeAt-multi-use-sites

Here's what I get:

[ RUN      ] JitTest.GPU_FusionComputeAtMultiSites_CUDA
unknown file: Failure
C++ exception with description "Tried to access position 1 in domain: [ iS{i1} ]
Exception raised from axis at ../torch/csrc/jit/codegen/cuda/tensor_view.cpp:73 (most recent call first):
frame #0: c10::Error::Error(c10::SourceLocation, std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> >) + 0x68 (0x7fc31dc32d78 in /home/nmaruyama/pytorch/src/csarofeen.3/build/lib/libc10.so)
frame #1: torch::jit::fuser::TensorView::axis(int) const + 0x400 (0x7fc3216274e0 in /home/nmaruyama/pytorch/src/csarofeen.3/build/lib/libtorch_cuda.so)
frame #2: torch::jit::fuser::TensorView::setThisComputeAtAxis() + 0x5b (0x7fc32162897b in /home/nmaruyama/pytorch/src/csarofeen.3/build/lib/libtorch_cuda.so)
frame #3: torch::jit::fuser::TensorView::setComputeAt(torch::jit::fuser::TensorView*, int) + 0x32 (0x7fc321628dc2 in /home/nmaruyama/pytorch/src/csarofeen.3/build/lib/libtorch_cuda.so)
frame #4: torch::jit::fuser::TensorView::computeAt(torch::jit::fuser::TensorView*, int) + 0x1c3a (0x7fc32162b90a in /home/nmaruyama/pytorch/src/csarofeen.3/build/lib/libtorch_cuda.so)
frame #5: torch::jit::testGPU_FusionComputeAtMultiSites() + 0x1cc (0x55a5706566cc in ./build/bin/test_jit)
frame #6: void testing::internal::HandleExceptionsInMethodIfSupported<testing::Test, void>(testing::Test*, void (testing::Test::*)(), char const*) + 0x4a (0x55a5706955fa in ./build/bin/test_jit)
frame #7: <unknown function> + 0x21735e (0x55a57068b35e in ./build/bin/test_jit)
frame #8: <unknown function> + 0x21781d (0x55a57068b81d in ./build/bin/test_jit)
frame #9: <unknown function> + 0x217a3d (0x55a57068ba3d in ./build/bin/test_jit)
frame #10: testing::internal::UnitTestImpl::RunAllTests() + 0xc59 (0x55a57068c919 in ./build/bin/test_jit)
frame #11: testing::UnitTest::Run() + 0x98 (0x55a57068cc18 in ./build/bin/test_jit)
frame #12: main + 0xc8 (0x55a570525558 in ./build/bin/test_jit)
frame #13: __libc_start_main + 0xe7 (0x7fc31d479b97 in /lib/x86_64-linux-gnu/libc.so.6)
frame #14: _start + 0x2a (0x55a57052e64a in ./build/bin/test_jit)
" thrown in the test body.
[  FAILED  ] JitTest.GPU_FusionComputeAtMultiSites_CUDA (2 ms)

[csarofeen]

@naoyam can we close this or update it?

[naoyam]

This works fine now.