color_spinor_field.h

Go to the documentation of this file.

#ifndef _COLOR_SPINOR_FIELD_H
#define _COLOR_SPINOR_FIELD_H

#include <quda_internal.h>
#include <quda.h>

#include <iostream>

#include <lattice_field.h>
#include <random_quda.h>
#include <fast_intdiv.h>

namespace quda {

  enum MemoryLocation { Device = 1, Host = 2, Remote = 4 };

  struct FullClover;

  typedef std::vector<ColorSpinorField*> CompositeColorSpinorField;

  struct CompositeColorSpinorFieldDescriptor {

     bool is_composite; //set to 'false' for a regular spinor field
     bool is_component; //set to 'true' if we want to work with an individual component (otherwise will work with the whole set)

     int  dim;//individual component has dim = 0
     int  id;

     int volume;       // volume of a single eigenvector
     int volumeCB;     // CB volume of a single eigenvector
     int stride;       // stride of a single eigenvector
     size_t real_length;  // physical length of a single eigenvector
     size_t length;       // length including pads (but not ghost zones)

     size_t bytes;      // size in bytes of spinor field
     size_t norm_bytes; // makes no sense but let's keep it...

     CompositeColorSpinorFieldDescriptor()
     : is_composite(false), is_component(false), dim(0), id(0), volume(0), volumeCB(0),
       stride(0), real_length(0), length(0), bytes(0), norm_bytes(0)  {};

     CompositeColorSpinorFieldDescriptor(bool is_composite, int dim, bool is_component = false, int id = 0)
     : is_composite(is_composite), is_component(is_component), dim(dim), id(id), volume(0), volumeCB(0),
       stride(0), real_length(0), length(0), bytes(0), norm_bytes(0)
     {
        if(is_composite && is_component) errorQuda("\nComposite type is not implemented.\n");
        else if(is_composite && dim == 0) is_composite = false;
     }

     CompositeColorSpinorFieldDescriptor(const CompositeColorSpinorFieldDescriptor &descr)
     {
       is_composite = descr.is_composite;
       is_component = descr.is_component;

       if(is_composite && is_component) errorQuda("\nComposite type is not implemented.\n");

       dim = descr.dim;
       id  = descr.id;

       volume   = descr.volume;
       volumeCB = descr.volumeCB;
       stride   = descr.stride;       // stride of a single eigenvector
       real_length = descr.real_length;  // physical length of a single eigenvector
       length      = descr.length;       // length including pads (but not ghost zones)

       bytes = descr.bytes;      // size in bytes of spinor field
       norm_bytes = descr.norm_bytes; // makes no sense but let's keep it...
     }

  };

  class ColorSpinorParam : public LatticeFieldParam {

  public:
    QudaFieldLocation location; // where are we storing the field (CUDA or CPU)?

    int nColor; // Number of colors of the field
    int nSpin; // =1 for staggered, =2 for coarse Dslash, =4 for 4d spinor
    int nVec;  // number of packed vectors (for multigrid transfer operator)

    QudaTwistFlavorType twistFlavor; // used by twisted mass

    QudaSiteOrder siteOrder; // defined for full fields

    QudaFieldOrder fieldOrder; // Float, Float2, Float4 etc.
    QudaGammaBasis gammaBasis;
    QudaFieldCreate create; //

    QudaPCType pc_type; // used to select preconditioning method in DWF

    void *v; // pointer to field
    void *norm;

    bool is_composite;
    int composite_dim;    //e.g., number of eigenvectors in the set
    bool is_component;
    int component_id;          //eigenvector index

    ColorSpinorParam(const ColorSpinorField &a);

    ColorSpinorParam() :
        LatticeFieldParam(),
        location(QUDA_INVALID_FIELD_LOCATION),
        nColor(0),
        nSpin(0),
        nVec(1),
        twistFlavor(QUDA_TWIST_INVALID),
        siteOrder(QUDA_INVALID_SITE_ORDER),
        fieldOrder(QUDA_INVALID_FIELD_ORDER),
        gammaBasis(QUDA_INVALID_GAMMA_BASIS),
        create(QUDA_INVALID_FIELD_CREATE),
        pc_type(QUDA_PC_INVALID),
        is_composite(false),
        composite_dim(0),
        is_component(false),
        component_id(0)
    {
      ;
    }

      // used to create cpu params
    ColorSpinorParam(void *V, QudaInvertParam &inv_param, const int *X, const bool pc_solution,
        QudaFieldLocation location = QUDA_CPU_FIELD_LOCATION) :
        LatticeFieldParam(4, X, 0, inv_param.cpu_prec),
        location(location),
        nColor(3),
        nSpin((inv_param.dslash_type == QUDA_ASQTAD_DSLASH || inv_param.dslash_type == QUDA_STAGGERED_DSLASH
                  || inv_param.dslash_type == QUDA_LAPLACE_DSLASH) ?
                1 :
                4),
        nVec(1),
        twistFlavor(inv_param.twist_flavor),
        siteOrder(QUDA_INVALID_SITE_ORDER),
        fieldOrder(QUDA_INVALID_FIELD_ORDER),
        gammaBasis(inv_param.gamma_basis),
        create(QUDA_REFERENCE_FIELD_CREATE),
        pc_type(inv_param.dslash_type == QUDA_DOMAIN_WALL_DSLASH ? QUDA_5D_PC : QUDA_4D_PC),
        v(V),
        is_composite(false),
        composite_dim(0),
        is_component(false),
        component_id(0)
    {

      if (nDim > QUDA_MAX_DIM) errorQuda("Number of dimensions too great");
      for (int d = 0; d < nDim; d++) x[d] = X[d];

      if (!pc_solution) {
        siteSubset = QUDA_FULL_SITE_SUBSET;
      } else {
        x[0] /= 2; // X defined the full lattice dimensions
        siteSubset = QUDA_PARITY_SITE_SUBSET;
      }

      if (inv_param.dslash_type == QUDA_DOMAIN_WALL_DSLASH || inv_param.dslash_type == QUDA_DOMAIN_WALL_4D_DSLASH
          || inv_param.dslash_type == QUDA_MOBIUS_DWF_DSLASH) {
        nDim++;
        x[4] = inv_param.Ls;
      } else if (inv_param.dslash_type == QUDA_TWISTED_MASS_DSLASH && (twistFlavor == QUDA_TWIST_NONDEG_DOUBLET)) {
        nDim++;
        x[4] = 2; // for two flavors
      } else if (inv_param.dslash_type == QUDA_STAGGERED_DSLASH || inv_param.dslash_type == QUDA_ASQTAD_DSLASH) {
        nDim++;
        x[4] = inv_param.Ls;
      }

      if (inv_param.dirac_order == QUDA_INTERNAL_DIRAC_ORDER) {
        fieldOrder
          = (precision == QUDA_DOUBLE_PRECISION || nSpin == 1 || nSpin == 2) ? QUDA_FLOAT2_FIELD_ORDER : QUDA_FLOAT4_FIELD_ORDER;
        siteOrder = QUDA_EVEN_ODD_SITE_ORDER;
      } else if (inv_param.dirac_order == QUDA_CPS_WILSON_DIRAC_ORDER) {
        fieldOrder = QUDA_SPACE_SPIN_COLOR_FIELD_ORDER;
        siteOrder = QUDA_ODD_EVEN_SITE_ORDER;
      } else if (inv_param.dirac_order == QUDA_QDP_DIRAC_ORDER) {
        fieldOrder = QUDA_SPACE_COLOR_SPIN_FIELD_ORDER;
        siteOrder = QUDA_EVEN_ODD_SITE_ORDER;
      } else if (inv_param.dirac_order == QUDA_DIRAC_ORDER) {
        fieldOrder = QUDA_SPACE_SPIN_COLOR_FIELD_ORDER;
        siteOrder = QUDA_EVEN_ODD_SITE_ORDER;
      } else if (inv_param.dirac_order == QUDA_QDPJIT_DIRAC_ORDER) {
        fieldOrder = QUDA_QDPJIT_FIELD_ORDER;
        siteOrder = QUDA_EVEN_ODD_SITE_ORDER;
      } else if (inv_param.dirac_order == QUDA_TIFR_PADDED_DIRAC_ORDER) {
        fieldOrder = QUDA_PADDED_SPACE_SPIN_COLOR_FIELD_ORDER;
        siteOrder = QUDA_EVEN_ODD_SITE_ORDER;
      } else {
        errorQuda("Dirac order %d not supported", inv_param.dirac_order);
      }
    }

    // normally used to create cuda param from a cpu param
    ColorSpinorParam(
        ColorSpinorParam &cpuParam, QudaInvertParam &inv_param, QudaFieldLocation location = QUDA_CUDA_FIELD_LOCATION) :
        LatticeFieldParam(cpuParam.nDim, cpuParam.x, inv_param.sp_pad, inv_param.cuda_prec),
        location(location),
        nColor(cpuParam.nColor),
        nSpin(cpuParam.nSpin),
        nVec(cpuParam.nVec),
        twistFlavor(cpuParam.twistFlavor),
        siteOrder(QUDA_EVEN_ODD_SITE_ORDER),
        fieldOrder(QUDA_INVALID_FIELD_ORDER),
        gammaBasis(nSpin == 4 ? QUDA_UKQCD_GAMMA_BASIS : QUDA_DEGRAND_ROSSI_GAMMA_BASIS),
        create(QUDA_COPY_FIELD_CREATE),
        pc_type(cpuParam.pc_type),
        v(0),
        is_composite(cpuParam.is_composite),
        composite_dim(cpuParam.composite_dim),
        is_component(false),
        component_id(0)
    {
      siteSubset = cpuParam.siteSubset;
      fieldOrder = (precision == QUDA_DOUBLE_PRECISION || nSpin == 1 || nSpin == 2) ? QUDA_FLOAT2_FIELD_ORDER : QUDA_FLOAT4_FIELD_ORDER;
    }

    void setPrecision(QudaPrecision precision, QudaPrecision ghost_precision=QUDA_INVALID_PRECISION, bool force_native=false) {
      // is the current status in native field order?
      bool native = force_native ? true : false;
      if ( ((this->precision == QUDA_DOUBLE_PRECISION || nSpin==1 || nSpin==2) &&
      (fieldOrder == QUDA_FLOAT2_FIELD_ORDER)) ||
     ((this->precision == QUDA_SINGLE_PRECISION || this->precision == QUDA_HALF_PRECISION || this->precision == QUDA_QUARTER_PRECISION) &&
      (nSpin==4) && fieldOrder == QUDA_FLOAT4_FIELD_ORDER) ) { native = true; }

      this->precision = precision;
      this->ghost_precision = (ghost_precision == QUDA_INVALID_PRECISION) ? precision : ghost_precision;

      // if this is a native field order, let's preserve that status, else keep the same field order
      if (native) fieldOrder = (precision == QUDA_DOUBLE_PRECISION || nSpin == 1 || nSpin == 2) ?
  QUDA_FLOAT2_FIELD_ORDER : QUDA_FLOAT4_FIELD_ORDER;
    }

    void print() {
      printfQuda("nColor = %d\n", nColor);
      printfQuda("nSpin = %d\n", nSpin);
      printfQuda("twistFlavor = %d\n", twistFlavor);
      printfQuda("nDim = %d\n", nDim);
      for (int d=0; d<nDim; d++) printfQuda("x[%d] = %d\n", d, x[d]);
      printfQuda("precision = %d\n", precision);
      printfQuda("ghost_precision = %d\n", ghost_precision);
      printfQuda("pad = %d\n", pad);
      printfQuda("siteSubset = %d\n", siteSubset);
      printfQuda("siteOrder = %d\n", siteOrder);
      printfQuda("fieldOrder = %d\n", fieldOrder);
      printfQuda("gammaBasis = %d\n", gammaBasis);
      printfQuda("create = %d\n", create);
      printfQuda("v = %lx\n", (unsigned long)v);
      printfQuda("norm = %lx\n", (unsigned long)norm);
      if(is_composite) printfQuda("Number of elements = %d\n", composite_dim);
    }

    virtual ~ColorSpinorParam() {
    }

  };

  class cpuColorSpinorField;
  class cudaColorSpinorField;

  struct DslashConstant {
    int Vh;
    int_fastdiv X[QUDA_MAX_DIM];
    int_fastdiv Xh[QUDA_MAX_DIM];
    int Ls;

    int volume_4d;
    int_fastdiv volume_4d_cb;

    int_fastdiv face_X[4];
    int_fastdiv face_Y[4];
    int_fastdiv face_Z[4];
    int_fastdiv face_T[4];
    int_fastdiv face_XY[4];
    int_fastdiv face_XYZ[4];
    int_fastdiv face_XYZT[4];

    int ghostFace[QUDA_MAX_DIM+1];
    int ghostFaceCB[QUDA_MAX_DIM + 1];

    int X2X1;
    int X3X2X1;
    int X4X3X2X1;

    int X2X1mX1;
    int X3X2X1mX2X1;
    int X4X3X2X1mX3X2X1;
    int X5X4X3X2X1mX4X3X2X1;
    int X4X3X2X1hmX3X2X1h;

    int_fastdiv dims[4][3];
  };

  class ColorSpinorField : public LatticeField {

  private:
      void create(int nDim, const int *x, int Nc, int Ns, int Nvec, QudaTwistFlavorType Twistflavor,
          QudaPrecision precision, int pad, QudaSiteSubset subset, QudaSiteOrder siteOrder, QudaFieldOrder fieldOrder,
          QudaGammaBasis gammaBasis, QudaPCType pc_type);
      void destroy();

  protected:
    bool init;

    mutable QudaPrecision ghost_precision_allocated;

    int nColor;
    int nSpin;
    int nVec;

    int nDim;
    int x[QUDA_MAX_DIM];

    int volume;
    int volumeCB;
    int pad;
    int stride;

    QudaTwistFlavorType twistFlavor;

    QudaPCType pc_type; // used to select preconditioning method in DWF

    size_t real_length; // physical length only
    size_t length; // length including pads, but not ghost zone - used for BLAS

    void *v; // the field elements
    void *norm; // the normalization field

    void *v_h; // the field elements
    void *norm_h; // the normalization field

    // multi-GPU parameters

    void* ghost[2][QUDA_MAX_DIM]; // pointers to the ghost regions - NULL by default
    void* ghostNorm[2][QUDA_MAX_DIM]; // pointers to ghost norms - NULL by default

    mutable int ghostFace[QUDA_MAX_DIM];   // the size of each face
    mutable int ghostFaceCB[QUDA_MAX_DIM]; // the size of each checkboarded face

    mutable void *ghost_buf[2*QUDA_MAX_DIM]; // wrapper that points to current ghost zone

    mutable DslashConstant dslash_constant; // constants used by dslash and packing kernels

    size_t bytes; // size in bytes of spinor field
    size_t norm_bytes; // size in bytes of norm field

    QudaSiteSubset siteSubset;
    QudaSiteOrder siteOrder;
    QudaFieldOrder fieldOrder;
    QudaGammaBasis gammaBasis;

    // in the case of full fields, these are references to the even / odd sublattices
    ColorSpinorField *even;
    ColorSpinorField *odd;

    CompositeColorSpinorFieldDescriptor composite_descr;//containes info about the set
    //
    CompositeColorSpinorField components;

    void createGhostZone(int nFace, bool spin_project=true) const;

    // resets the above attributes based on contents of param
    void reset(const ColorSpinorParam &);
    void fill(ColorSpinorParam &) const;
    static void checkField(const ColorSpinorField &, const ColorSpinorField &);

    void setTuningString();

  public:
    //ColorSpinorField();
    ColorSpinorField(const ColorSpinorField &);
    ColorSpinorField(const ColorSpinorParam &);

    virtual ~ColorSpinorField();

    virtual ColorSpinorField& operator=(const ColorSpinorField &);

    int Ncolor() const { return nColor; }
    int Nspin() const { return nSpin; }
    int Nvec() const { return nVec; }
    QudaTwistFlavorType TwistFlavor() const { return twistFlavor; }
    int Ndim() const { return nDim; }
    const int* X() const { return x; }
    int X(int d) const { return x[d]; }
    size_t RealLength() const { return real_length; }
    size_t Length() const { return length; }
    int Stride() const { return stride; }
    int Volume() const { return volume; }
    int VolumeCB() const { return siteSubset == QUDA_PARITY_SITE_SUBSET ? volume : volume / 2; }
    int Pad() const { return pad; }
    size_t Bytes() const { return bytes; }
    size_t NormBytes() const { return norm_bytes; }
    size_t GhostBytes() const { return ghost_bytes; }
    size_t GhostNormBytes() const { return ghost_bytes; }
    void PrintDims() const { printfQuda("dimensions=%d %d %d %d\n", x[0], x[1], x[2], x[3]); }

    void* V() {return v;}
    const void* V() const {return v;}
    void* Norm(){return norm;}
    const void* Norm() const {return norm;}
    virtual const void* Ghost2() const { return nullptr; }

    void exchange(void **ghost, void **sendbuf, int nFace=1) const;

    virtual void exchangeGhost(QudaParity parity, int nFace, int dagger, const MemoryLocation *pack_destination=nullptr,
             const MemoryLocation *halo_location=nullptr, bool gdr_send=false, bool gdr_recv=false,
             QudaPrecision ghost_precision=QUDA_INVALID_PRECISION) const = 0;

    bool isNative() const;

    bool IsComposite() const { return composite_descr.is_composite; }
    bool IsComponent() const { return composite_descr.is_component; }

    int CompositeDim() const { return composite_descr.dim; }
    int ComponentId() const { return composite_descr.id; }
    int ComponentVolume() const { return composite_descr.volume; }
    int ComponentVolumeCB() const { return composite_descr.volumeCB; }
    int ComponentStride() const { return composite_descr.stride; }
    size_t ComponentLength() const { return composite_descr.length; }
    size_t ComponentRealLength() const { return composite_descr.real_length; }

    size_t ComponentBytes() const { return composite_descr.bytes; }
    size_t ComponentNormBytes() const { return composite_descr.norm_bytes; }

    QudaPCType PCType() const { return pc_type; }

    QudaSiteSubset SiteSubset() const { return siteSubset; }
    QudaSiteOrder SiteOrder() const { return siteOrder; }
    QudaFieldOrder FieldOrder() const { return fieldOrder; }
    QudaGammaBasis GammaBasis() const { return gammaBasis; }

    const int *GhostFace() const { return ghostFace; }
    const int *GhostFaceCB() const { return ghostFaceCB; }
    int GhostOffset(const int i) const { return ghostOffset[i][0]; }
    int GhostOffset(const int i, const int j) const { return ghostOffset[i][j]; }
    int GhostNormOffset(const int i ) const { return ghostNormOffset[i][0]; }
    int GhostNormOffset(const int i, const int j) const { return ghostNormOffset[i][j]; }

    void* Ghost(const int i);
    const void* Ghost(const int i) const;
    void* GhostNorm(const int i);
    const void* GhostNorm(const int i) const;

    void* const* Ghost() const;

    const DslashConstant& getDslashConstant() const { return dslash_constant; }

    const ColorSpinorField& Even() const;
    const ColorSpinorField& Odd() const;

    ColorSpinorField& Even();
    ColorSpinorField& Odd();

    ColorSpinorField& Component(const int idx) const;
    ColorSpinorField& Component(const int idx);

    CompositeColorSpinorField& Components(){
      return components;
    };

    virtual void Source(const QudaSourceType sourceType, const int st=0, const int s=0, const int c=0) = 0;

    virtual void PrintVector(unsigned int x) const = 0;

    void LatticeIndex(int *y, int i) const;

    void OffsetIndex(int &i, int *y) const;

    static ColorSpinorField* Create(const ColorSpinorParam &param);
    static ColorSpinorField* Create(const ColorSpinorField &src, const ColorSpinorParam &param);

    ColorSpinorField* CreateCoarse(const int *geoBlockSize, int spinBlockSize, int Nvec,
                                   QudaPrecision precision=QUDA_INVALID_PRECISION,
           QudaFieldLocation location=QUDA_INVALID_FIELD_LOCATION,
                                   QudaMemoryType mem_Type=QUDA_MEMORY_INVALID);

    ColorSpinorField* CreateFine(const int *geoblockSize, int spinBlockSize, int Nvec,
                                 QudaPrecision precision=QUDA_INVALID_PRECISION,
         QudaFieldLocation location=QUDA_INVALID_FIELD_LOCATION,
                                 QudaMemoryType mem_type=QUDA_MEMORY_INVALID);

    friend std::ostream& operator<<(std::ostream &out, const ColorSpinorField &);
    friend class ColorSpinorParam;
  };

  // CUDA implementation
  class cudaColorSpinorField : public ColorSpinorField {

    friend class cpuColorSpinorField;

  private:
    bool alloc; // whether we allocated memory
    bool init;

    bool texInit; // whether a texture object has been created or not
    mutable bool ghostTexInit; // whether the ghost texture object has been created
    mutable QudaPrecision ghost_precision_tex; 
#ifdef USE_TEXTURE_OBJECTS
    cudaTextureObject_t tex;
    cudaTextureObject_t texNorm;
    void createTexObject();
    void destroyTexObject();
    mutable cudaTextureObject_t ghostTex[4]; // these are double buffered and variants to host-mapped buffers
    mutable cudaTextureObject_t ghostTexNorm[4];
    void createGhostTexObject() const;
    void destroyGhostTexObject() const;
#endif

    bool reference; // whether the field is a reference or not

    mutable void *ghost_field_tex[4]; // instance pointer to GPU halo buffer (used to check if static allocation has changed)

    void create(const QudaFieldCreate);
    void destroy();
    void copy(const cudaColorSpinorField &);

    void zeroPad();

    void copySpinorField(const ColorSpinorField &src);

    void loadSpinorField(const ColorSpinorField &src);
    void saveSpinorField (ColorSpinorField &src) const;

  public:

    //cudaColorSpinorField();
    cudaColorSpinorField(const cudaColorSpinorField&);
    cudaColorSpinorField(const ColorSpinorField&, const ColorSpinorParam&);
    cudaColorSpinorField(const ColorSpinorField&);
    cudaColorSpinorField(const ColorSpinorParam&);
    virtual ~cudaColorSpinorField();

    ColorSpinorField& operator=(const ColorSpinorField &);
    cudaColorSpinorField& operator=(const cudaColorSpinorField&);
    cudaColorSpinorField& operator=(const cpuColorSpinorField&);

    void switchBufferPinned();

    void createComms(int nFace, bool spin_project=true);

    void allocateGhostBuffer(int nFace, bool spin_project=true) const;

    void packGhost(const int nFace, const QudaParity parity, const int dim, const QudaDirection dir, const int dagger,
                   cudaStream_t *stream, MemoryLocation location[2 * QUDA_MAX_DIM], MemoryLocation location_label,
                   bool spin_project, double a = 0, double b = 0, double c = 0);

    void packGhostExtended(const int nFace, const int R[], const QudaParity parity, const int dim, const QudaDirection dir,
         const int dagger,cudaStream_t* stream, bool zero_copy=false);

    void sendGhost(void *ghost_spinor, const int nFace, const int dim, const QudaDirection dir,
        const int dagger, cudaStream_t *stream);

    void unpackGhost(const void* ghost_spinor, const int nFace, const int dim,
        const QudaDirection dir, const int dagger, cudaStream_t* stream);

    void unpackGhostExtended(const void* ghost_spinor, const int nFace, const QudaParity parity,
           const int dim, const QudaDirection dir, const int dagger, cudaStream_t* stream, bool zero_copy);


    void streamInit(cudaStream_t *stream_p);

    void pack(int nFace, int parity, int dagger, int stream_idx, MemoryLocation location[],
              MemoryLocation location_label, bool spin_project = true, double a = 0, double b = 0, double c = 0);

    void packExtended(const int nFace, const int R[], const int parity, const int dagger,
        const int dim,  cudaStream_t *stream_p, const bool zeroCopyPack=false);

    void gather(int nFace, int dagger, int dir, cudaStream_t *stream_p=NULL);

    void recvStart(int nFace, int dir, int dagger=0, cudaStream_t *stream_p=nullptr, bool gdr=false);

    void sendStart(int nFace, int d, int dagger=0, cudaStream_t *stream_p=nullptr, bool gdr=false, bool remote_write=false);

    void commsStart(int nFace, int d, int dagger=0, cudaStream_t *stream_p=nullptr, bool gdr_send=false, bool gdr_recv=false);

    int commsQuery(int nFace, int d, int dagger=0, cudaStream_t *stream_p=nullptr, bool gdr_send=false, bool gdr_recv=false);

    void commsWait(int nFace, int d, int dagger=0, cudaStream_t *stream_p=nullptr, bool gdr_send=false, bool gdr_recv=false);

    void scatter(int nFace, int dagger, int dir, cudaStream_t *stream_p);
    void scatter(int nFace, int dagger, int dir);

    void scatterExtended(int nFace, int parity, int dagger, int dir);

    inline const void* Ghost2() const {
      if (bufferIndex < 2) {
        return ghost_recv_buffer_d[bufferIndex];
      } else {
        return ghost_pinned_recv_buffer_hd[bufferIndex%2];
      }
    }

    void exchangeGhost(QudaParity parity, int nFace, int dagger, const MemoryLocation *pack_destination=nullptr,
           const MemoryLocation *halo_location=nullptr, bool gdr_send=false, bool gdr_recv=false,
           QudaPrecision ghost_precision=QUDA_INVALID_PRECISION) const;

#ifdef USE_TEXTURE_OBJECTS
    inline const cudaTextureObject_t& Tex() const { return tex; }
    inline const cudaTextureObject_t& TexNorm() const { return texNorm; }
    inline const cudaTextureObject_t& GhostTex() const { return ghostTex[bufferIndex]; }
    inline const cudaTextureObject_t& GhostTexNorm() const { return ghostTexNorm[bufferIndex]; }
#endif

    cudaColorSpinorField& Component(const int idx) const;
    CompositeColorSpinorField& Components() const;
    void CopySubset(cudaColorSpinorField& dst, const int range, const int first_element=0) const;

    void zero();

    friend std::ostream& operator<<(std::ostream &out, const cudaColorSpinorField &);

    void getTexObjectInfo() const;

    void Source(const QudaSourceType sourceType, const int st=0, const int s=0, const int c=0);

    void PrintVector(unsigned int x) const;

    void backup() const;

    void restore() const;
  };

  // CPU implementation
  class cpuColorSpinorField : public ColorSpinorField {

    friend class cudaColorSpinorField;

  public:
    static void* fwdGhostFaceBuffer[QUDA_MAX_DIM]; //cpu memory
    static void* backGhostFaceBuffer[QUDA_MAX_DIM]; //cpu memory
    static void* fwdGhostFaceSendBuffer[QUDA_MAX_DIM]; //cpu memory
    static void* backGhostFaceSendBuffer[QUDA_MAX_DIM]; //cpu memory
    static int initGhostFaceBuffer;
    static size_t ghostFaceBytes[QUDA_MAX_DIM];

    private:
    //void *v; // the field elements
    //void *norm; // the normalization field
    bool init;
    bool reference; // whether the field is a reference or not

    void create(const QudaFieldCreate);
    void destroy();

    public:
    //cpuColorSpinorField();
    cpuColorSpinorField(const cpuColorSpinorField&);
    cpuColorSpinorField(const ColorSpinorField&);
    cpuColorSpinorField(const ColorSpinorField&, const ColorSpinorParam&);
    cpuColorSpinorField(const ColorSpinorParam&);
    virtual ~cpuColorSpinorField();

    ColorSpinorField& operator=(const ColorSpinorField &);
    cpuColorSpinorField& operator=(const cpuColorSpinorField&);
    cpuColorSpinorField& operator=(const cudaColorSpinorField&);

    void Source(const QudaSourceType sourceType, const int st=0, const int s=0, const int c=0);

    static int Compare(const cpuColorSpinorField &a, const cpuColorSpinorField &b, const int resolution=1);

    void PrintVector(unsigned int x) const;

    void allocateGhostBuffer(int nFace) const;
    static void freeGhostBuffer(void);

    void packGhost(void **ghost, const QudaParity parity, const int nFace, const int dagger) const;
    void unpackGhost(void* ghost_spinor, const int dim,
         const QudaDirection dir, const int dagger);

    void copy(const cpuColorSpinorField&);
    void zero();

    void exchangeGhost(QudaParity parity, int nFace, int dagger, const MemoryLocation *pack_destination=nullptr,
           const MemoryLocation *halo_location=nullptr, bool gdr_send=false, bool gdr_recv=false,
           QudaPrecision ghost_precision=QUDA_INVALID_PRECISION) const;

    void backup() const;

    void restore() const;
  };

  void copyGenericColorSpinor(ColorSpinorField &dst, const ColorSpinorField &src,
      QudaFieldLocation location, void *Dst=0, void *Src=0,
      void *dstNorm=0, void*srcNorm=0);
  void genericSource(cpuColorSpinorField &a, QudaSourceType sourceType, int x, int s, int c);
  int genericCompare(const cpuColorSpinorField &a, const cpuColorSpinorField &b, int tol);

  void genericPrintVector(const cpuColorSpinorField &a, unsigned int x);
  void genericCudaPrintVector(const cudaColorSpinorField &a, unsigned x);

  void wuppertalStep(ColorSpinorField &out, const ColorSpinorField &in, int parity, const GaugeField& U, double A, double B);
  void wuppertalStep(ColorSpinorField &out, const ColorSpinorField &in, int parity, const GaugeField& U, double alpha);

  void exchangeExtendedGhost(cudaColorSpinorField* spinor, int R[], int parity, cudaStream_t *stream_p);

  void copyExtendedColorSpinor(ColorSpinorField &dst, const ColorSpinorField &src,
      QudaFieldLocation location, const int parity, void *Dst, void *Src, void *dstNorm, void *srcNorm);

  void genericPackGhost(void **ghost, const ColorSpinorField &a, QudaParity parity,
      int nFace, int dagger, MemoryLocation *destination=nullptr);

  void spinorNoise(ColorSpinorField &src, RNG& randstates, QudaNoiseType type);

  void spinorNoise(ColorSpinorField &src, unsigned long long seed, QudaNoiseType type);

  inline QudaPCType PCType_(
      const char *func, const char *file, int line, const ColorSpinorField &a, const ColorSpinorField &b)
  {
    QudaPCType type = QUDA_PC_INVALID;
    if (a.PCType() == b.PCType())
      type = a.PCType();
    else
      errorQuda("PCTypes %d %d do not match (%s:%d in %s())\n", a.PCType(), b.PCType(), file, line, func);
    return type;
  }

  template <typename... Args>
  inline QudaPCType PCType_(const char *func, const char *file, int line, const ColorSpinorField &a,
      const ColorSpinorField &b, const Args &... args)
  {
    return static_cast<QudaPCType>(PCType_(func, file, line, a, b) & PCType_(func, file, line, a, args...));
  }

#define checkPCType(...) PCType_(__func__, __FILE__, __LINE__, __VA_ARGS__)

} // namespace quda

#endif // _COLOR_SPINOR_FIELD_H