Dslash
This module implements the Dslash operator for staggered and highly improved staggered quark (HISQ) fermions. It is a sparse matrix-vector product that is implemented as a 4 dimensional stencil kernel with nearest and third-nearest neighbor terms. For HISQ fermions, the matrix-vector product takes the form \( \begin{align} D[U]\psi_{x}&=\sum_{\mu=0}^{4}\left[c_{1}\left(V_{x,\mu}\psi_{x+\hat{\mu}}-V^{\dagger}_{x-\hat{\mu},\mu}\psi_{x-\hat{\mu}}\right)+c_{3}\left(W_{x,\mu}\psi_{x+3\hat{\mu}}-W^{\dagger}_{x-3\hat{\mu},\mu}\psi_{x-3\hat{\mu}}\right)\right], \end{align}\) where \(V_{x,\mu}\) and \(W_{x,\mu}\) are the HISQ smeared fields described in Hisq Smearing. For staggered fermions, only the nearest neighbor term is present and the basic gauge field \(U_{x,\mu}\) is used in place of \(V_{x,\mu}\).
In the code, Dslash operators are derived from an abstract base class that defines the interface for Dslash operators.
template<typename SpinorLHS_t, typename SpinorRHS_t>
class DSlash : public LinearOperator<SpinorRHS_t> {
public:
virtual void Dslash(SpinorLHS_t &lhs, SpinorRHS_t &rhs, bool update = true);
virtual void applyMdaggM(SpinorRHS_t &, SpinorRHS_t &, bool update = true) = 0;
};
The method void Dslash(SpinorLHS_t &lhs, SpinorRHS_t &rhs, bool update = true)
applies the stencil operator described above to an input vector rhs
and write the result into the vector lhs
.
\(\begin{align} \chi=D\psi \end{align}\)
The method void applyMdaggM(SpinorRHS_t &lhs, SpinorRHS_t &rhs, bool update = true)
computes
\(\begin{align} \chi=M^{\dagger}M\psi \;\;\;\mathrm{where} \;\;M[U]=D[U]+m_{f}, \end{align}\)
and \(m_{f}\) is the quark mass. In both methods, bool update
toggles wether or not a halo update should be performed on the output spinor after the kernel is applied.
The derived class for HISQ fermions is:
template<typename floatT, bool onDevice, Layout LatLayoutRHS, size_t HaloDepthGauge, size_t HaloDepthSpin, size_t NStacks = 1>
class HisqDSlash : public DSlash<Spinorfield<floatT, onDevice, LayoutSwitcher<LatLayoutRHS>(), HaloDepthSpin, NStacks>,
Spinorfield<floatT, onDevice, LatLayoutRHS, HaloDepthSpin, NStacks> >
floatT
specifies which floating point type to use and onDevice
toggles whether or not gauge and spinor fields are residing on the host or device. The template parameter Layout LatLayoutRHS
specifies the lattice layout of the input vector and can take the values Even
Odd
and All
. HaloDepthGauge
and HaloDepthSpin
specify the depth of the halo buffers needed for multi-gpu calculations. Usually, HaloDepthGauge=2
and HaloDepthSpin=4
should be chosen.
The last template parameter size_t NStacks
specifies how many rhs vectors are used simultaneously. Loading a gauge link once and multiplying it to multiple vectors within the same kernel call significantly increases performance as the DSlash kernels performance is mostly bandwidth bound.
A HisqDSlash object is constructed with:
HisqDSlash(Gauge_t<R18> &gaugefield_smeared, Gauge_t<U3R14> &gaugefield_Naik, const double mass, floatT naik_epsilon = 0.0,
std::string spinorName = "SHARED_HisqDSlashSpinor")
Here gaugefield_smeared
refers to \(V_{x,\mu}\) and gaugefield_Naik
to \(W_{x,\mu}\). Note the different compression types for the gauge fields: the field entering the third-nearest neighbor hopping term uses CompressionType comp=U3R14
in order to save memory bandwidth. const double mass
specifies the quark mass that enters void MdaggM
and floatT naik_epsilon
specifies the coefficient \(\epsilon\) that can be included in the Naik term. The last parameter that can be given to the constructor is a string identifiying the memory allocated for a temporary spinor that is used internally. By default multiple instances of HisqDslash will share the same memory for this temporary spinor. See the documentation on MemoryManagement
for further information.