Re: [eigen] Mac CUDA build failure question

On Thu, Jun 20, 2019 at 3:11 PM Artem Belevich <tra@xxxxxxxxxxx> wrote:

The changes look reasonable to me. Thank you for helping to sort this out.

On Thu, Jun 20, 2019 at 3:05 PM Eric Klein <elklein@xxxxxxxxx> wrote:
Ok. Looks like the warnings are there with and without my hack(s), and the minimal set of edits needed to get Eigen to build on Mac with nvcc consists of:

Half.h. Change this:
#if !defined(EIGEN_HAS_NATIVE_FP16) || EIGEN_COMP_CLANG // Emulate support for half floats
to this:
#if !defined(EIGEN_HAS_NATIVE_FP16) || (EIGEN_COMP_CLANG && !EIGEN_COMP_NVCC) // Emulate support for half floats

And in PacketMath.h, change this:
#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE) || (defined(EIGEN_CUDACC) && EIGEN_COMP_CLANG)
to this:
#if defined(EIGEN_CUDA_ARCH) || defined(EIGEN_HIP_DEVICE_COMPILE) || (defined(EIGEN_CUDACC) && EIGEN_COMP_CLANG && !EIGEN_COMP_NVCC)

Obviously that's excluding any enlightening comments about why that's being done.

Would you like me to prepare a patch file, or is this something that would be better handled by one of the regular contributors?

Thank you!
---
Eric Klein
elklein@xxxxxxxxx

On Thu, Jun 20, 2019 at 12:01 AM Eric Klein <elklein@xxxxxxxxx> wrote:
That appears to work, although there are 2-3 other places that need similar modifications in order to work. I'll try to get you a more complete list tomorrow.

I'm paying more attention tonight to warnings coming from Eigen than I had been previously ignoring, and both with my old Macros.h based hack and the newer modifications, I'm seeing some of these: "warning: calling a __host__ function from a __host__ __device__ function is not allowed". A representative one is:

external/eigen_archive/unsupported/Eigen/CXX11/src/Tensor/TensorReduction.h(735): warning: calling a __host__ function from a __host__ __device__ function is not allowed
detected during:
instantiation of "__nv_bool Eigen::TensorEvaluator<const Eigen::TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device>::evalSubExprsIfNeeded(MakePointer_<Eigen::TensorEvaluator<const Eigen::TensorReductionOp<Op, Dims, ArgType, MakePointer_>, Device>::CoeffReturnType>::Type) [with Op=Eigen::internal::AvgPoolMeanReducer<double>, Dims=const Eigen::IndexList<Eigen::type2index<1L>, Eigen::type2index<2L>>, ArgType=const Eigen::TensorImagePatchOp<-1L, -1L, const Eigen::TensorLayoutSwapOp<const Eigen::TensorMap<Eigen::Tensor<const double, 4, 1, Eigen::DenseIndex>, 16, Eigen::MakePointer>>>, MakePointer_=Eigen::MakePointer, Device=Eigen::GpuDevice]"
external/eigen_archive/unsupported/Eigen/CXX11/src/Tensor/TensorMorphing.h(172): here
instantiation of "__nv_bool Eigen::TensorEvaluator<const Eigen::TensorReshapingOp<NewDimensions, ArgType>, Device>::evalSubExprsIfNeeded(Eigen::TensorEvaluator<const Eigen::TensorReshapingOp<NewDimensions, ArgType>, Device>::CoeffReturnType *) [with NewDimensions=const Eigen::DSizes<Eigen::DenseIndex, 4>, ArgType=const Eigen::TensorReductionOp<Eigen::internal::AvgPoolMeanReducer<double>, const Eigen::IndexList<Eigen::type2index<1L>, Eigen::type2index<2L>>, const Eigen::TensorImagePatchOp<-1L, -1L, const Eigen::TensorLayoutSwapOp<const Eigen::TensorMap<Eigen::Tensor<const double, 4, 1, Eigen::DenseIndex>, 16, Eigen::MakePointer>>>, Eigen::MakePointer>, Device=Eigen::GpuDevice]"
external/eigen_archive/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h(146): here
instantiation of "__nv_bool Eigen::TensorEvaluator<const Eigen::TensorAssignOp<LeftArgType, RightArgType>, Device>::evalSubExprsIfNeeded(Eigen::TensorEvaluator<const Eigen::TensorAssignOp<LeftArgType, RightArgType>, Device>::Scalar *) [with LeftArgType=Eigen::TensorLayoutSwapOp<Eigen::TensorMap<Eigen::Tensor<double, 4, 1, Eigen::DenseIndex>, 16, Eigen::MakePointer>>, RightArgType=const Eigen::TensorReshapingOp<const Eigen::DSizes<Eigen::DenseIndex, 4>, const Eigen::TensorReductionOp<Eigen::internal::AvgPoolMeanReducer<double>, const Eigen::IndexList<Eigen::type2index<1L>, Eigen::type2index<2L>>, const Eigen::TensorImagePatchOp<-1L, -1L, const Eigen::TensorLayoutSwapOp<const Eigen::TensorMap<Eigen::Tensor<const double, 4, 1, Eigen::DenseIndex>, 16, Eigen::MakePointer>>>, Eigen::MakePointer>>, Device=Eigen::GpuDevice]"
external/eigen_archive/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor..h(422): here
instantiation of "void Eigen::internal::TensorExecutor<_expression_, Eigen::GpuDevice, Vectorizable, Tileable>::run(const _expression_ &, const Eigen::GpuDevice &) [with _expression_=const Eigen::TensorAssignOp<Eigen::TensorLayoutSwapOp<Eigen::TensorMap<Eigen::Tensor<double, 4, 1, Eigen::DenseIndex>, 16, Eigen::MakePointer>>, const Eigen::TensorReshapingOp<const Eigen::DSizes<Eigen::DenseIndex, 4>, const Eigen::TensorReductionOp<Eigen::internal::AvgPoolMeanReducer<double>, const Eigen::IndexList<Eigen::type2index<1L>, Eigen::type2index<2L>>, const Eigen::TensorImagePatchOp<-1L, -1L, const Eigen::TensorLayoutSwapOp<const Eigen::TensorMap<Eigen::Tensor<const double, 4, 1, Eigen::DenseIndex>, 16, Eigen::MakePointer>>>, Eigen::MakePointer>>>, Vectorizable=false, Tileable=false]"
external/eigen_archive/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h(35): here
instantiation of "Eigen::TensorDevice<ExpressionType, DeviceType> &Eigen::TensorDevice<ExpressionType, DeviceType>::operator=(const OtherDerived &) [with ExpressionType=Eigen::TensorLayoutSwapOp<Eigen::TensorMap<Eigen::Tensor<double, 4, 1, Eigen::DenseIndex>, 16, Eigen::MakePointer>>, DeviceType=tensorflow::GPUDevice, OtherDerived=Eigen::TensorReshapingOp<const Eigen::DSizes<Eigen::DenseIndex, 4>, const Eigen::TensorReductionOp<Eigen::internal::AvgPoolMeanReducer<double>, const Eigen::IndexList<Eigen::type2index<1L>, Eigen::type2index<2L>>, const Eigen::TensorImagePatchOp<-1L, -1L, const Eigen::TensorLayoutSwapOp<const Eigen::TensorMap<Eigen::Tensor<const double, 4, 1, Eigen::DenseIndex>, 16, Eigen::MakePointer>>>, Eigen::MakePointer>>]"
./tensorflow/core/kernels/avgpooling_op.h(42): here
instantiation of "void tensorflow::functor::SpatialAvgPooling<Device, T>::operator()(const Device &, tensorflow::TTypes<T, 4, Eigen::DenseIndex>::Tensor, tensorflow::TTypes<T, 4, Eigen::DenseIndex>::ConstTensor, int, int, int, int, const Eigen::PaddingType &) [with Device=tensorflow::GPUDevice, T=double]"
tensorflow/core/kernels/avgpooling_op_gpu.cu.cc(38): here

I'm not sure how concerned I should be about these. The build will succeed, but.... I wouldn't be at all surprised to get weird results eventually.

In this particular case, it looks like it's complaining because Eigen::GpuDevice::allocate_temp appears to be __host__ rather than __host__ __device__ (i.e. missing EIGEN_DEVICE_FUNC). I fully admit that I could be misinterpreting that or otherwise misunderstanding something basic.

Should I be concerned about these?

Thanks!
---
Eric Klein
elklein@xxxxxxxxx

On Wed, Jun 19, 2019 at 5:21 PM Rasmus Munk Larsen <rmlarsen@xxxxxxxxxx> wrote:
Erik, does Artem's suggestion work for you?

On Wed, Jun 19, 2019 at 2:52 PM Artem Belevich <tra@xxxxxxxxxx> wrote:

On Wed, Jun 19, 2019 at 1:47 PM Rasmus Munk Larsen <rmlarsen@xxxxxxxxxx> wrote:
It looks like we broke the Eigen Cuda build on Mac. What do you think about his workaround?

---------- Forwarded message ---------
From: Eric Klein <elklein@xxxxxxxxx>
Date: Wed, Jun 19, 2019 at 1:39 PM
Subject: [eigen] Mac CUDA build failure question
To: <eigen@xxxxxxxxxxxxxxxxxxx>

Hello all,

I posted a question on the forums several days back, but suspect that might not be the right place to be asking what I'm asking, so I'm trying the mailing list as well.

I'll just repost here what I put in the forums, but the link to that is here: https://forum.kde.org/viewtopic.php?f=74&t=161199

I'm trying to build Eigen on Mac for CUDA (using the nvcc compiler), and getting build errors. I understand the errors, and I have a change that lets me dodge the build failures, but I suspect it's not the right change for checkin, and so I'm looking for feedback.

So the issue I have is in Half.h. I wind up getting errors about a bunch of operators being already defined. The core issue is that on Mac, nvcc (the CUDA compliler) is using gcc as the host compiler, but gcc on Mac is built on top of clang. Eigen seems to be implicitly assuming that the presence of clang implies that absence of CUDA (or at least the absence of nvcc CUDA support).

In my build I'm hitting this block:

#if (defined(EIGEN_HAS_CUDA_FP16) && defined(EIGEN_CUDA_ARCH) && \
EIGEN_CUDA_ARCH >= 530) || \
(defined(EIGEN_HAS_HIP_FP16) && defined(HIP_DEVICE_COMPILE))
#define EIGEN_HAS_NATIVE_FP16
#endif

which results in EIGEN_HAS_NATIVE_FP16 being set, and so we wind up compiling in all the operators from Half.h:253-313. That's fine so far.

This assumes device-side compilation.

What happens next is we hit this line:

#if !defined(EIGEN_HAS_NATIVE_FP16) || EIGEN_COMP_CLANG // Emulate support for half floats

which is followed shortly after by (roughly) the same operator functions (but... emulated), and I get errors because those operator functions were defined above.

If Clang were CUDA compiler, that would not be a problem. This implies that it's a CUDA compilation with NVCC. What puzzles me is how did we end up with EIGEN_COMP_CLANG defined for the *device* side of the compilation. I suspect it's the side effect of NVCC doing device-side preprocessing with clang, but actually compiling with cicc, which is obviously not clang.

I guess what we need to do here is something like this:
#if !defined(EIGEN_HAS_NATIVE_FP16) || (EIGEN_COMP_CLANG && !EIGEN_COMP_NVCC)

That, and a comment explaining what's going on.

If that does not help, it would be great to compile with '-keep -verbose' and check which compilation phase is failing and what exactly is it trying to compile.

--Artem

So. My hack to work around this is to ensure that EIGEN_COMP_CLANG gets set to 0 in Macros.h if __NVCC__ is defined. That works fine for me locally, and gets Eigen building fine (and thus unblocks me on getting TensorFlow building for Mac, or at least unblocks this issue).

I'm willing to bet however that this is the wrong thing to do in general. I don't understand enough of what this second code block is doing to really understand why clang is being treated differently than nvcc here (and specifically why half support needs to be emulated in the presence of clang). I believe there is a version of clang that supports CUDA (at least on some platforms?). Presumably this is for that, but I don't know enough about how that differs from nvcc to fully grok this.

Can anyone help enlighten me about the best way to fix this?

Thanks!
---
Eric Klein
elklein@xxxxxxxxx

--
--Artem Belevich

--
--Artem Belevich