Generic non-SIMD, platform-agnostic version?

[barracuda156]

The website mentions that "EigenRand currently supports only x86-64 architecture (SSE, AVX, AVX2) and ARM64 NEON". Can non-SIMD version be added? Or it won't make sense?
Otherwise other platforms (PowerPC, RISC-V) cannot be supported.

[bab2min]

Hi @barracuda156 ,
In short, EigenRand already has non-SIMD version, and it may run on other platforms, but without acceleration. The reason I marked this as unsupported on other platforms is because I haven't (and don't plan on) testing that it works on all platforms, it doesn't mean that it's actually impossible to compile on PowerPC or RISC-V.

EigenRand actually has non-SIMD implementations like Eigen does. Every RNG of EigenRand has both scalarOp implementation (non-SIMD) & packetOp implementation (SIMD) and proper operation would be selected by Eigen's logic.

For example, StdNormalGen has both scalar version:

EigenRand/EigenRand/Dists/NormalExp.h

Lines 35 to 59 in f3190cd

	template<typename Rng>
	EIGEN_STRONG_INLINE const _Scalar operator() (Rng&& rng)
	{
	using namespace Eigen::internal;
	if (valid)
	{
	valid = false;
	return OptCacheStore::get<_Scalar>();
	}
	valid = true;
	_Scalar v1, v2, sx;
	for (int _i = 0; ; ++_i)
	{
	EIGENRAND_CHECK_INFINITY_LOOP();
	v1 = 2 * ur(rng) - 1;
	v2 = 2 * ur(rng) - 1;
	sx = v1 * v1 + v2 * v2;
	if (sx && sx < 1) break;
	}
	_Scalar fx = std::sqrt((_Scalar)-2.0 * std::log(sx) / sx);
	OptCacheStore::get<_Scalar>() = fx * v2;
	return fx * v1;
	}

and packet version:

EigenRand/EigenRand/Dists/NormalExp.h

Lines 60 to 83 in f3190cd

	template<typename Packet, typename Rng>
	EIGEN_STRONG_INLINE const Packet packetOp(Rng&& rng)
	{
	using namespace Eigen::internal;
	if (valid)
	{
	valid = false;
	return OptCacheStore::template get<Packet>();
	}
	valid = true;
	Packet u1 = ur.template packetOp<Packet>(rng),
	u2 = ur.template packetOp<Packet>(rng);
	u1 = psub(pset1<Packet>(1), u1);
	auto radius = psqrt(pmul(pset1<Packet>(-2), plog(u1)));
	auto theta = pmul(pset1<Packet>(2 * constant::pi), u2);
	Packet sintheta, costheta;
	psincos(theta, sintheta, costheta);
	OptCacheStore::template get<Packet>() = pmul(radius, costheta);
	return pmul(radius, sintheta);
	}
	};

And only when packet operation is available, the latter one is selected by PacketAccess attribute.

EigenRand/EigenRand/RandUtils.h

Line 137 in f3190cd

enum { Cost = HugeCost, PacketAccess = std::is_floating_point<_Scalar>::value ? packet_traits<_Scalar>::HasExp : packet_traits<_Scalar>::Vectorizable, IsRepeatable = false };

So, if you compile the code using EigenRand on other platforms, PacketAccess would be 0 and the scalar one would be selected. If any compile errors appears, please report because it must be a bug of EigenRand.

[barracuda156]

Thank you very much for a detailed reply! I will try it.