covDev.cu

Go to the documentation of this file.

#include <dslash.h>
#include <worker.h>
#include <dslash_helper.cuh>
#include <color_spinor_field_order.h>
#include <gauge_field_order.h>
#include <color_spinor.h>
#include <dslash_helper.cuh>
#include <index_helper.cuh>
#include <gauge_field.h>
#include <uint_to_char.h>

#include <dslash_policy.cuh>
#include <kernels/covDev.cuh>

namespace quda
{

#ifdef GPU_COVDEV

  template <typename Float, int nDim, int nColor, int nParity, bool dagger, bool xpay, KernelType kernel_type, typename Arg>
  struct CovDevLaunch {

    // kernel name for jit compilation
    static constexpr const char *kernel = "quda::covDevGPU";

    template <typename Dslash>
    inline static void launch(Dslash &dslash, TuneParam &tp, Arg &arg, const cudaStream_t &stream)
    {
      static_assert(xpay == false, "Covariant derivative operator only defined without xpay");
      static_assert(nParity == 2, "Covariant derivative operator only defined for full field");
      dslash.launch(covDevGPU<Float, nDim, nColor, nParity, dagger, xpay, kernel_type, Arg>, tp, arg, stream);
    }
  };

  template <typename Float, int nDim, int nColor, typename Arg> class CovDev : public Dslash<Float>
  {

protected:
    Arg &arg;
    const ColorSpinorField &in;

public:
    CovDev(Arg &arg, const ColorSpinorField &out, const ColorSpinorField &in) :
      Dslash<Float>(arg, out, in, "kernels/covDev.cuh"),
      arg(arg),
      in(in)
    {
    }

    virtual ~CovDev() {}

    void apply(const cudaStream_t &stream)
    {
      TuneParam tp = tuneLaunch(*this, getTuning(), getVerbosity());
      Dslash<Float>::setParam(arg);
      if (arg.xpay) errorQuda("Covariant derivative operator only defined without xpay");
      if (arg.nParity != 2) errorQuda("Covariant derivative operator only defined for full field");

      constexpr bool xpay = false;
      constexpr int nParity = 2;
      Dslash<Float>::template instantiate<CovDevLaunch, nDim, nColor, nParity, xpay>(tp, arg, stream);
    }

    long long flops() const
    {
      int mv_flops = (8 * in.Ncolor() - 2) * in.Ncolor(); // SU(3) matrix-vector flops
      int num_mv_multiply = in.Nspin();
      int ghost_flops = num_mv_multiply * mv_flops;
      int dim = arg.mu % 4;
      long long flops_ = 0;

      switch (arg.kernel_type) {
      case EXTERIOR_KERNEL_X:
      case EXTERIOR_KERNEL_Y:
      case EXTERIOR_KERNEL_Z:
      case EXTERIOR_KERNEL_T:
        if (arg.kernel_type != dim) break;
        flops_ = (ghost_flops)*in.GhostFace()[dim];
        break;
      case EXTERIOR_KERNEL_ALL: {
        long long ghost_sites = in.GhostFace()[dim];
        flops_ = ghost_flops * ghost_sites;
        break;
      }
      case INTERIOR_KERNEL:
      case KERNEL_POLICY: {
        long long sites = in.Volume();
        flops_ = num_mv_multiply * mv_flops * sites; // SU(3) matrix-vector multiplies

        if (arg.kernel_type == KERNEL_POLICY) break;
        // now correct for flops done by exterior kernel
        long long ghost_sites = arg.commDim[dim] ? in.GhostFace()[dim] : 0;
        flops_ -= ghost_flops * ghost_sites;

        break;
      }
      }

      return flops_;
    }

    long long bytes() const
    {
      int gauge_bytes = arg.reconstruct * in.Precision();
      bool isFixed = (in.Precision() == sizeof(short) || in.Precision() == sizeof(char)) ? true : false;
      int spinor_bytes = 2 * in.Ncolor() * in.Nspin() * in.Precision() + (isFixed ? sizeof(float) : 0);
      int ghost_bytes = gauge_bytes + 3 * spinor_bytes; // 3 since we have to load the partial
      int dim = arg.mu % 4;
      long long bytes_ = 0;

      switch (arg.kernel_type) {
      case EXTERIOR_KERNEL_X:
      case EXTERIOR_KERNEL_Y:
      case EXTERIOR_KERNEL_Z:
      case EXTERIOR_KERNEL_T:
        if (arg.kernel_type != dim) break;
        bytes_ = ghost_bytes * in.GhostFace()[dim];
        break;
      case EXTERIOR_KERNEL_ALL: {
        long long ghost_sites = in.GhostFace()[dim];
        bytes_ = ghost_bytes * ghost_sites;
        break;
      }
      case INTERIOR_KERNEL:
      case KERNEL_POLICY: {
        long long sites = in.Volume();
        bytes_ = (gauge_bytes + 2 * spinor_bytes) * sites;

        if (arg.kernel_type == KERNEL_POLICY) break;
        // now correct for bytes done by exterior kernel
        long long ghost_sites = arg.commDim[dim] ? in.GhostFace()[dim] : 0;
        bytes_ -= ghost_bytes * ghost_sites;

        break;
      }
      }
      return bytes_;
    }

    TuneKey tuneKey() const
    {
      // add mu to the key
      char aux[TuneKey::aux_n];
      strcpy(aux, Dslash<Float>::aux[arg.kernel_type]);
      strcat(aux, ",mu=");
      char mu[8];
      u32toa(mu, arg.mu);
      strcat(aux, mu);
      return TuneKey(in.VolString(), typeid(*this).name(), aux);
    }
  };

  template <typename Float, int nColor, QudaReconstructType recon> struct CovDevApply {

    inline CovDevApply(ColorSpinorField &out, const ColorSpinorField &in, const GaugeField &U, int mu, int parity,
                       bool dagger, const int *comm_override, TimeProfile &profile)

    {
      constexpr int nDim = 4;
      CovDevArg<Float, nColor, recon> arg(out, in, U, mu, parity, dagger, comm_override);
      CovDev<Float, nDim, nColor, CovDevArg<Float, nColor, recon>> covDev(arg, out, in);

      dslash::DslashPolicyTune<decltype(covDev)> policy(
        covDev, const_cast<cudaColorSpinorField *>(static_cast<const cudaColorSpinorField *>(&in)), in.VolumeCB(),
        in.GhostFaceCB(), profile);
      policy.apply(0);

      checkCudaError();
    }
  };

#endif

  // Apply the covariant derivative operator
  // out(x) = U_{\mu}(x)in(x+mu) for mu = 0...3
  // out(x) = U^\dagger_mu'(x-mu')in(x-mu') for mu = 4...7 and we set mu' = mu-4
  void ApplyCovDev(ColorSpinorField &out, const ColorSpinorField &in, const GaugeField &U, int mu, int parity,
                   bool dagger, const int *comm_override, TimeProfile &profile)
  {
#ifdef GPU_COVDEV
    if (in.V() == out.V()) errorQuda("Aliasing pointers");
    if (in.FieldOrder() != out.FieldOrder())
      errorQuda("Field order mismatch in = %d, out = %d", in.FieldOrder(), out.FieldOrder());

    // check all precisions match
    checkPrecision(out, in, U);

    // check all locations match
    checkLocation(out, in, U);

    pushKernelPackT(true); // non-spin projection requires kernel packing

    instantiate<CovDevApply>(out, in, U, mu, parity, dagger, comm_override, profile);

    popKernelPackT();
#else
    errorQuda("Covariant derivative kernels have not been built");
#endif
  }
} // namespace quda