Compile JIT from source to support CUDA

C++
1

Could you tell the bug in the following codes?

pytorch/torch/csrc/jit/codegen/cuda/manager.cpp

namespace {
std::unique_ptr makePWKernelSupport(
const at::ArrayRef& inputs) {
auto req_ptr = std::make_unique();
for (const auto& input : inputs) {
req_ptr->dims_.push_back(input.isTensor() ? input.toTensor().dim() : -1);
}
return req_ptr;
}

C++ compiler told me there is an error resulted from copying unique_ptr?
I am not a C++ expert, so I cannot tell where the wrong code is and how to fix it.

2

What kind of error did you see?
Could you post the complete error message, please?

3

Thanks for your reply.
To compile cuda-support pytorch from source, I encountered some problems.

First, by following the latest guide in github README, the error occurred.

…/torch/csrc/jit/codegen/cuda/manager.cpp:75:58: required from here
/usr/include/c++/5/bits/stl_construct.h:75:7: error: use of deleted function ‘constexpr std::pair<_T1, _T2>::pair(const std::pair<_T1, _T2>&) [with _T1 = std::unique_ptrtorch::jit::fuser::cuda::KernelArgsReq; _T2 = torch::jit::fuser::cuda::CudaKernel]’
{ ::new(static_cast<void*>(__p)) _T1(std::forward<_Args>(__args)…); }

I modified /torch/csrc/jit/codegen/cuda/manager.cpp:75
from

kernel_cache_.insert({kernel_id, CudaKernelCache()});

to

kernel_cache_.insert(std::make_pair(std::move(kernel_id), CudaKernelCache()));

After modification, the compilation can go on, but a new error comes out.

…/test/cpp/jit/test_gpu.cpp:1633:47: error: converting to ‘std::tuple<at::Tensor ()(const at::Tensor&), torch::jit::fuser::UnaryOpType, std::__cxx11::basic_string<char, std::char_traits, std::allocator > >’ from initializer list would use explicit constructor ‘constexpr std::tuple< >::tuple(_UElements&& …) [with _UElements = {at::Tensor (&)(const at::Tensor&), torch::jit::fuser::UnaryOpType, const char (&)[4]}; = void; _Elements = {at::Tensor ()(const at::Tensor&), torch::jit::fuser::UnaryOpType, std::__cxx11::basic_string<char, std::char_traits, std::allocator >}]’
{at::trunc, UnaryOpType::Trunc, “trunc”}};

I noticed std::__cxx11::basic_string, so I guess cuda-supported pytorch should be compiled with C++11 instead of C++14. I am about to verify my guess.

Do you have any suggestions?

4

You would need a C++14 compiler as given in the build instructions:

If you are installing from source, you will need a C++14 compiler.

5

Now I am using C++14, but errors comes out.
So what do you think about the second error?

…/test/cpp/jit/test_gpu.cpp:1633:47: error: converting to ‘std::tuple<at::Tensor ( )(const at::Tensor&), torch::jit::fuser::UnaryOpType, std::__cxx11::basic_string<char, std::char_traits, std::allocator > >’ from initializer list would use explicit constructor ‘constexpr std::tuple< >::tuple(_UElements&& …) [with _UElements = {at::Tensor (&)(const at::Tensor&), torch::jit::fuser::UnaryOpType, const char (&)[4]}; = void; _Elements = {at::Tensor ( )(const at::Tensor&), torch::jit::fuser::UnaryOpType, std::__cxx11::basic_string<char, std::char_traits, std::allocator >}]’
{at::trunc, UnaryOpType::Trunc, “trunc”}};

Is there a solution?
Thank you!

6

I have passed the compilation and the CUDA can be detected correctly.
For the second error, I commented all the related tuple initialization in /home/lim/Desktop/GithubRepo/pytorch/test/cpp/jit/test_gpu.cpp.