源码编译ARM机器时，make TARGET=ARMV8 -j$(nproc)卡住

[kane-hu]

问题描述 Issue Description

在鲲鹏920的机器上源码编译Paddle时，执行make TARGET=ARMV8 -j$(nproc)后出现

In file included from /opt/aPaddle/third_party/eigen3/unsupported/Eigen/CXX11/../../../Eigen/Core:253,
from /opt/aPaddle/third_party/eigen3/unsupported/Eigen/CXX11/Tensor:14,
from /opt/aPaddle/paddle/phi/kernels/funcs/eigen/extensions.h:23,
from /opt/aPaddle/paddle/phi/kernels/funcs/eigen/common.h:20,
from /opt/aPaddle/paddle/phi/kernels/funcs/common_shape.h:18,
from /opt/aPaddle/paddle/phi/kernels/funcs/elementwise_base.h:21,
from /opt/aPaddle/paddle/phi/kernels/funcs/broadcast_function.h:18,
from /opt/aPaddle/paddle/phi/kernels/cpu/elementwise.h:20,
from /opt/aPaddle/paddle/phi/kernels/legacy/cpu/elementwise_multiply_kernel.cc:20:
/opt/aPaddle/third_party/eigen3/unsupported/Eigen/CXX11/../../../Eigen/src/Core/functors/BinaryFunctors.h: In instantiation of ‘Eigen::internal::scalar_product_op<LhsScalar, RhsScalar>::result_type Eigen::internal::scalar_product_op<LhsScalar, RhsScalar>::operator()(const LhsScalar&, const RhsScalar&) const [with LhsScalar = const bool; RhsScalar = const bool; Eigen::internal::scalar_product_op<LhsScalar, RhsScalar>::result_type = const bool]’:
/opt/aPaddle/third_party/eigen3/unsupported/Eigen/CXX11/src/Tensor/TensorEvaluator.h:641:71: required from ‘Eigen::TensorEvaluator<const Eigen::TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArgType>, Device>::CoeffReturnType Eigen::TensorEvaluator<const Eigen::TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArgType>, Device>::coeff(Eigen::TensorEvaluator<const Eigen::TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArgType>, Device>::Index) const [with BinaryOp = Eigen::internal::scalar_product_op<const bool, const bool>; LeftArgType = const Eigen::TensorMap<Eigen::Tensor<const bool, 1, 1, long int>, 0, Eigen::MakePointer>; RightArgType = const Eigen::TensorMap<Eigen::Tensor<const bool, 1, 1, long int>, 0, Eigen::MakePointer>; Device = Eigen::DefaultDevice; Eigen::TensorEvaluator<const Eigen::TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArgType>, Device>::CoeffReturnType = bool; Eigen::TensorEvaluator<const Eigen::TensorCwiseBinaryOp<BinaryOp, LeftArgType, RightArgType>, Device>::Index = long int]’
/opt/aPaddle/third_party/eigen3/unsupported/Eigen/CXX11/src/Tensor/TensorAssign.h:172:28: required from ‘void Eigen::TensorEvaluator<const Eigen::TensorAssignOp<LhsXprType, RhsXprType>, Device>::evalScalar(Eigen::TensorEvaluator<const Eigen::TensorAssignOp<LhsXprType, RhsXprType>, Device>::Index) [with LeftArgType = Eigen::TensorMap<Eigen::Tensor<bool, 1, 1, long int>, 0, Eigen::MakePointer>; RightArgType = const Eigen::TensorCwiseBinaryOp<Eigen::internal::scalar_product_op<const bool, const bool>, const Eigen::TensorMap<Eigen::Tensor<const bool, 1, 1, long int>, 0, Eigen::MakePointer>, const Eigen::TensorMap<Eigen::Tensor<const bool, 1, 1, long int>, 0, Eigen::MakePointer> >; Device = Eigen::DefaultDevice; Eigen::TensorEvaluator<const Eigen::TensorAssignOp<LhsXprType, RhsXprType>, Device>::Index = long int]’
/opt/aPaddle/third_party/eigen3/unsupported/Eigen/CXX11/src/Tensor/TensorExecutor.h:103:9: required from ‘static void Eigen::internal::TensorExecutor<Expression, Device, Vectorizable, Tiling>::run(const Expression&, const Device&) [with Expression = const Eigen::TensorAssignOp<Eigen::TensorMap<Eigen::Tensor<bool, 1, 1, long int>, 0, Eigen::MakePointer>, const Eigen::TensorCwiseBinaryOp<Eigen::internal::scalar_product_op<const bool, const bool>, const Eigen::TensorMap<Eigen::Tensor<const bool, 1, 1, long int>, 0, Eigen::MakePointer>, const Eigen::TensorMap<Eigen::Tensor<const bool, 1, 1, long int>, 0, Eigen::MakePointer> > >; Device = Eigen::DefaultDevice; bool Vectorizable = false; Eigen::internal::TiledEvaluation Tiling = (Eigen::internal::TiledEvaluation)0]’
/opt/aPaddle/third_party/eigen3/unsupported/Eigen/CXX11/src/Tensor/TensorDevice.h:37:62: required from ‘Eigen::TensorDevice<ExpressionType, DeviceType>& Eigen::TensorDevice<ExpressionType, DeviceType>::operator=(const OtherDerived&) [with OtherDerived = Eigen::TensorCwiseBinaryOp<Eigen::internal::scalar_product_op<const bool, const bool>, const Eigen::TensorMap<Eigen::Tensor<const bool, 1, 1, long int>, 0, Eigen::MakePointer>, const Eigen::TensorMap<Eigen::Tensor<const bool, 1, 1, long int>, 0, Eigen::MakePointer> >; ExpressionType = Eigen::TensorMap<Eigen::Tensor<bool, 1, 1, long int>, 0, Eigen::MakePointer>; DeviceType = Eigen::DefaultDevice]’
/opt/aPaddle/paddle/phi/kernels/cpu/elementwise.h:187:27: required from ‘void phi::SameDimsMultiplyFunctor<DevCtx, T, typename std::enable_if<(! std::is_floating_point<_DInputType>::value)>::type>::operator()(const DevCtx&, const phi::DenseTensor&, const phi::DenseTensor&, phi::DenseTensor*) [with DevCtx = phi::CPUContext; T = bool; typename std::enable_if<(! std::is_floating_point<_DInputType>::value)>::type = void]’
/opt/aPaddle/paddle/phi/kernels/cpu/elementwise.h:197:14: required from ‘void phi::SameDimsElementwiseCompute::operator()(const phi::CPUContext&, const phi::DenseTensor&, const phi::DenseTensor&, phi::DenseTensor*) [with Functor = phi::SameDimsMultiplyFunctor<phi::CPUContext, bool, void>]’
/opt/aPaddle/paddle/phi/kernels/legacy/cpu/elementwise_multiply_kernel.cc:26:1: required from ‘void phi::MultiplyRawKernel(const Context&, const phi::DenseTensor&, const phi::DenseTensor&, int, phi::DenseTensor*) [with T = bool; Context = phi::CPUContext]’
/opt/aPaddle/paddle/phi/kernels/legacy/cpu/elementwise_multiply_kernel.cc:36:1: required from here
/opt/aPaddle/third_party/eigen3/unsupported/Eigen/CXX11/../../../Eigen/src/Core/functors/BinaryFunctors.h:80:122: 警告：‘*’ in boolean context, suggest ‘&&’ instead [-Wint-in-bool-context]
EIGEN_DEVICE_FUNC EIGEN_STRONG_INLINE result_type operator() (const LhsScalar& a, const RhsScalar& b) const { return a * b; }

随后make步骤卡住，需要大概1小时后继续

到编译95%时，出现报错

版本&环境信息 Version & Environment Information

鲲鹏920 Linux localhost.localdomain 4.19.90-52.22.v2207.ky10.aarch64 #1 SMP aarch64 aarch64 aarch64 GNU/Linux
gcc 版本 8.2.0 (GCC)
cmake version 3.27.0-rc3
GPU：nvidia T4
Cuda compilation tools, release 11.2, V11.2.67
Build cuda_11.2.r11.2/compiler.29373293_0
cmake语句：
cmake .. -DPY_VERSION=3.7 -DPYTHON_EXECUTABLE=/opt/gcc/py37_paddle/bin/python -DWITH_ARM=ON -DWITH_TESTING=OFF -DCMAKE_BUILD_TYPE=Release -DON_INFER=ON -DWITH_XBYAK=OFF -DWITH_GPU=ON -DCUDA_ARCH_NAME=All -DWITH_NCCL=OFF -DWITH_RCCL=OFF

[liuzhenhai93]

@kane-hu
全架构编译会导致编译包过大， 链接寻址溢出；
如果是 T4 卡，设置编译选项 -DCUDA_ARCH_NAME=Turing， 只变异turing 架构；