[eigen] Issue with CUDA

[Vladimir Florinski]

I hope one of the developers could help me with using Eigen in CUDA code. I
am using version 9.1 of the CUDA toolkit and Eigen 3.3.5. This simple code

#include <cuda_runtime.h>
#include <Eigen/Dense>

__device__ void EigenMVCuda(Eigen::MatrixXd& mat, Eigen::VectorXd& vec,
Eigen::VectorXd& res)
{
res = mat * vec;
};

int main() {};

Generates the following type of errors:

../include/Eigen/src/Core/MathFunctions.h(1067): warning: calling a
__host__ function from a __host__ __device__ function is not allowed

../include/Eigen/src/Core/arch/CUDA/Half.h(96): error: identifier "x" is
undefined

../include/Eigen/src/Core/arch/CUDA/PacketMathHalf.h(102): error: more than
one conversion function from "const __half" to a built-in type applies:
function "__half::operator short() const"
function "__half::operator unsigned short() const"
function "__half::operator int() const"
function "__half::operator unsigned int() const"
function "__half::operator long long() const"
function "__half::operator unsigned long long() const"
function "__half::operator __nv_bool() const"

Looks like there are three different issues here, but I am not sufficiently
familiar with the inner workings of either library to make much sense out
of them. Please advise how to resolve.

Thanks,

[Gael Guennebaud]

For CUDA 9 better use the head of the default branch, or maybe the head of
the 3.3 branch which also contains a few CUDA9 related fixes. Then, we
don't alllow dynamic allocation on device code, so either pass fixed size
matrices of Map<> referencing existing buffers:

typedef Eigen::Map<Eigen::MatrixXd> MapMatXd;
typedef Eigen::Map<Eigen::VectorXd> MapVecXd;
typedef Eigne::Ref<Eigen::MatrixXd> RefMatXd;
typedef Eigen::Ref<Eigen::VectorXd> RefVecXd;

__device__ void EigenMVCuda(RefMatXd mat, RefVecXd& vec, RefVecXd& res)
{
res.noalias() = mat * vec;
};

and in your kernel use MapMatXd/MapVecXd to wrap existing buffers.

Post by Vladimir Florinski
I hope one of the developers could help me with using Eigen in CUDA code.
I am using version 9.1 of the CUDA toolkit and Eigen 3.3.5. This simple code
#include <cuda_runtime.h>
#include <Eigen/Dense>
__device__ void EigenMVCuda(Eigen::MatrixXd& mat, Eigen::VectorXd& vec,
Eigen::VectorXd& res)
{
res = mat * vec;
};
int main() {};
../include/Eigen/src/Core/MathFunctions.h(1067): warning: calling a
__host__ function from a __host__ __device__ function is not allowed
../include/Eigen/src/Core/arch/CUDA/Half.h(96): error: identifier "x" is
undefined
../include/Eigen/src/Core/arch/CUDA/PacketMathHalf.h(102): error: more
function "__half::operator short() const"
function "__half::operator unsigned short() const"
function "__half::operator int() const"
function "__half::operator unsigned int() const"
function "__half::operator long long() const"
function "__half::operator unsigned long long() const"
function "__half::operator __nv_bool() const"
Looks like there are three different issues here, but I am not
sufficiently familiar with the inner workings of either library to make
much sense out of them. Please advise how to resolve.
Thanks,
--
Vladimir Florinski