CuPackedMatrix< Real > Class Template Reference

Matrix for CUDA computing. More...

#include <matrix-common.h>

Inheritance diagram for CuPackedMatrix< Real >:

Collaboration diagram for CuPackedMatrix< Real >:

Public Member Functions
	CuPackedMatrix ()
	CuPackedMatrix (MatrixIndexT r, MatrixResizeType resize_type=kSetZero)
	CuPackedMatrix (const PackedMatrix< Real > &orig)
	CuPackedMatrix (const CuPackedMatrix< Real > &orig)
void	SetZero ()
void	SetUnit ()
	< Set to zero More...
void	SetRandn ()
	< Set to unit matrix. More...
void	SetDiag (Real alpha)
	< Set to random values of a normal distribution More...
void	AddToDiag (Real r)
	< Set the diagonal value to alpha More...
void	Scale (Real alpha)
void	ScaleDiag (Real alpha)
Real	Trace () const
	~CuPackedMatrix ()
void	Resize (MatrixIndexT nRows, MatrixResizeType resize_type=kSetZero)
	Set packed matrix to a specified size (can be zero). More...
void	CopyFromPacked (const CuPackedMatrix< Real > &src)
void	CopyFromPacked (const PackedMatrix< Real > &src)
void	CopyToPacked (PackedMatrix< Real > *dst) const
void	Read (std::istream &in, bool binary)
void	Write (std::ostream &out, bool binary) const
void	Destroy ()
void	Swap (CuPackedMatrix< Real > *other)
	Swaps the contents of *this and *other. Shallow swap. More...
void	Swap (PackedMatrix< Real > *other)
	Swaps the contents of *this and *other. More...
Real *	Data ()
const Real *	Data () const
Real	operator() (MatrixIndexT r, MatrixIndexT c) const
MatrixIndexT	NumRows () const
MatrixIndexT	NumCols () const
size_t	SizeInBytes () const
	Returns size in bytes of the data held by the matrix. More...

Protected Member Functions
const PackedMatrix< Real > &	Mat () const
PackedMatrix< Real > &	Mat ()
void	AddPacked (const Real alpha, const CuPackedMatrix< Real > &M)

Protected Attributes
Real *	data_
MatrixIndexT	num_rows_

Private Member Functions
PackedMatrix< Real > &	operator= (const PackedMatrix< Real > &other)

Friends
class	CuMatrixBase< Real >
class	CuVectorBase< Real >
class	CuSubMatrix< Real >
class	CuRand< Real >

Detailed Description

template<typename Real>
class kaldi::CuPackedMatrix< Real >

Matrix for CUDA computing.

This is a base class for packed triangular and symmetric matrices. Does the computation on the CUDA card when CUDA is compiled in and we have a suitable GPU (CuDevice::Instantiate().Enabled() == true); otherwise, does it on the CPU.Packed CUDA matrix: base class for triangular and symmetric matrices on a GPU card.

Definition at line 75 of file matrix-common.h.

Constructor & Destructor Documentation

CuPackedMatrix() [1/4]

CuPackedMatrix ( )

inline

Definition at line 60 of file cu-packed-matrix.h.

: data_(NULL), num_rows_(0) {}

CuPackedMatrix() [2/4]

CuPackedMatrix ( MatrixIndexT  r, MatrixResizeType  resize_type = kSetZero  )

inlineexplicit

Definition at line 62 of file cu-packed-matrix.h.

References CuPackedMatrix< Real >::Resize().

                                                                  :
      data_(NULL), num_rows_(0) {  Resize(r, resize_type);  }

CuPackedMatrix() [3/4]

CuPackedMatrix ( const PackedMatrix< Real > &  orig )

inlineexplicit

Definition at line 66 of file cu-packed-matrix.h.

References CuPackedMatrix< Real >::CopyFromPacked(), kaldi::kUndefined, PackedMatrix< Real >::num_rows_, and CuPackedMatrix< Real >::Resize().

                                                          : data_(NULL), num_rows_(0) {
    Resize(orig.num_rows_, kUndefined);
    CopyFromPacked(orig);
  }

CuPackedMatrix() [4/4]

CuPackedMatrix ( const CuPackedMatrix< Real > &  orig )

inlineexplicit

Definition at line 71 of file cu-packed-matrix.h.

References CuPackedMatrix< Real >::AddToDiag(), CuPackedMatrix< Real >::CopyFromPacked(), kaldi::kUndefined, CuPackedMatrix< Real >::Resize(), CuPackedMatrix< Real >::Scale(), CuPackedMatrix< Real >::ScaleDiag(), CuPackedMatrix< Real >::SetDiag(), CuPackedMatrix< Real >::SetRandn(), CuPackedMatrix< Real >::SetUnit(), CuPackedMatrix< Real >::SetZero(), and CuPackedMatrix< Real >::Trace().

                                                            : data_(NULL), num_rows_(0) {
    Resize(orig.NumRows(), kUndefined);
    CopyFromPacked(orig);
  }

~CuPackedMatrix()

~CuPackedMatrix ( )

inline

Definition at line 86 of file cu-packed-matrix.h.

References CuPackedMatrix< Real >::CopyFromPacked(), CuPackedMatrix< Real >::CopyToPacked(), CuPackedMatrix< Real >::Destroy(), kaldi::kSetZero, CuPackedMatrix< Real >::Read(), CuPackedMatrix< Real >::Resize(), CuPackedMatrix< Real >::Swap(), and CuPackedMatrix< Real >::Write().

{ Destroy(); }

Member Function Documentation

AddPacked()

void AddPacked ( const Real  alpha, const CuPackedMatrix< Real > &  M  )

protected

Definition at line 333 of file cu-packed-matrix.cc.

References CuPackedMatrix< Real >::Data(), data_, KALDI_ASSERT, CuPackedMatrix< Real >::Mat(), and CuPackedMatrix< Real >::NumRows().

Referenced by CuSpMatrix< Real >::AddSp().

                                                                                    {
  KALDI_ASSERT(num_rows_ == M.NumRows());
#if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    if (num_rows_ == 0) return;
    CuTimer tim;
    size_t nr = num_rows_,
        sz = (nr * (nr + 1)) / 2;
    cublas_axpy(GetCublasHandle(), sz, alpha, M.Data(), 1, data_, 1);
    CuDevice::Instantiate().AccuProfile("CuPackedMatrix::AddPacked", tim);
  } else
#endif
  {
    Mat().AddPacked(alpha, M.Mat());
  }
}

AddToDiag()

void AddToDiag

(

Real

r

)

< Set the diagonal value to alpha

Add this quantity to the diagonal of the matrix.

Definition at line 351 of file cu-packed-matrix.cc.

References CU1DBLOCK, and data_.

Referenced by CuPackedMatrix< Real >::CuPackedMatrix(), kaldi::UnitTestCuPackedMatrixAddToDiag(), kaldi::UnitTestCuSpMatrixAddToDiag(), kaldi::UnitTestCuSpMatrixInvert(), and kaldi::UnitTestCuSpMatrixSetUnit().

                                           {
#if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    if (num_rows_ == 0) return;
    CuTimer tim;
    int dimBlock(CU1DBLOCK);
    int dimGrid(n_blocks(NumRows(),CU1DBLOCK));
    cuda_add_diag_packed(dimGrid,dimBlock,data_,r,num_rows_);
    CU_SAFE_CALL(cudaGetLastError());
    CuDevice::Instantiate().AccuProfile("CuPackedMatrix::AddToDiag", tim);
  } else
#endif
  {
    // TODO
    Mat().AddToDiag(r);
  }
}

CopyFromPacked() [1/2]

void CopyFromPacked

(

const CuPackedMatrix< Real > &

src

)

Definition at line 137 of file cu-packed-matrix.cc.

References data_, CuPackedMatrix< Real >::data_, KALDI_ASSERT, CuPackedMatrix< Real >::Mat(), and CuPackedMatrix< Real >::NumRows().

Referenced by CuSpMatrix< Real >::CopyFromSp(), CuTpMatrix< Real >::CopyFromTp(), CuPackedMatrix< Real >::CuPackedMatrix(), CuTpMatrix< Real >::operator=(), kaldi::UnitTestCuPackedMatrixCopy(), and CuPackedMatrix< Real >::~CuPackedMatrix().

                                                                         {
  KALDI_ASSERT(src.NumRows() == num_rows_);
#if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    if (num_rows_ == 0) return; // Nothing to do.
    CuTimer tim;
    size_t nr = static_cast<size_t>(num_rows_),
        num_bytes = ((nr * (nr+1)) / 2) * sizeof(Real);

    CU_SAFE_CALL(
      cudaMemcpyAsync(data_, src.data_, num_bytes, cudaMemcpyDeviceToDevice,
                      cudaStreamPerThread));
    CuDevice::Instantiate().AccuProfile("CuPackedMatrix::CopyFromPacked1",
                                        tim);
  } else
#endif
  {
    Mat().CopyFromPacked(src.Mat());
  }
}

CopyFromPacked() [2/2]

void CopyFromPacked

(

const PackedMatrix< Real > &

src

)

Definition at line 159 of file cu-packed-matrix.cc.

References data_, PackedMatrix< Real >::data_, KALDI_ASSERT, PackedMatrix< Real >::NumRows(), and PackedMatrix< Real >::SizeInBytes().

                                                                       {
  KALDI_ASSERT(src.NumRows() == num_rows_);
#if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    if (num_rows_ == 0) return; // Nothing to do.
    CuTimer tim;
    CU_SAFE_CALL(cudaMemcpyAsync(data_, src.data_, src.SizeInBytes(),
                                 cudaMemcpyHostToDevice, cudaStreamPerThread));
    CU_SAFE_CALL(cudaStreamSynchronize(cudaStreamPerThread));
    CuDevice::Instantiate().AccuProfile("CuPackedMatrix::CopyFromPacked2", tim);
  } else
#endif
  {
    Mat().CopyFromPacked(src);
    //memcpy(data_, src.Data(), SizeInBytes());
  }
}

CopyToPacked()

void CopyToPacked

(

PackedMatrix< Real > *

dst

)

const

Definition at line 178 of file cu-packed-matrix.cc.

References PackedMatrix< Real >::CopyFromPacked(), data_, PackedMatrix< Real >::data_, KALDI_ASSERT, and PackedMatrix< Real >::NumRows().

Referenced by TpMatrix< float >::CopyFromMat(), CuSpMatrix< Real >::CopyToSp(), kaldi::UnitTestCuPackedMatrixCopy(), and CuPackedMatrix< Real >::~CuPackedMatrix().

                                                                     {
  KALDI_ASSERT(dst->NumRows() == NumRows());

#if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    if (num_rows_ == 0) return; // Nothing to do.
    CuTimer tim;
    size_t nr = static_cast<size_t>(num_rows_),
      num_bytes = ((nr * (nr+1)) / 2) * sizeof(Real);

    CU_SAFE_CALL(cudaMemcpyAsync(dst->data_, data_, num_bytes,
                                 cudaMemcpyDeviceToHost, cudaStreamPerThread));
    CU_SAFE_CALL(cudaStreamSynchronize(cudaStreamPerThread));
    CuDevice::Instantiate().AccuProfile("CuPackedMatrix::CopyToPackedD2H", tim);
  } else
#endif
  {
    //memcpy(data_, dst->Data(), SizeInBytes());
    dst->CopyFromPacked(Mat());
  }
}

Data() [1/2]

Real* Data ( )

inline

Definition at line 113 of file cu-packed-matrix.h.

References CuPackedMatrix< Real >::data_.

Referenced by CuPackedMatrix< Real >::AddPacked(), CuVectorBase< float >::AddSpVec(), CuVectorBase< float >::CopyDiagFromPacked(), CuMatrixBase< float >::CopyFromSp(), CuMatrixBase< float >::CopyFromTp(), CuVectorBase< float >::MulTp(), and kaldi::TraceSpSp().

{ return data_; }  

Data() [2/2]

const Real* Data ( ) const

inline

Definition at line 114 of file cu-packed-matrix.h.

References CuPackedMatrix< Real >::data_.

{ return data_; }

Destroy()

void Destroy

(

)

Definition at line 85 of file cu-packed-matrix.cc.

References data_, and KALDI_MEMALIGN_FREE.

Referenced by CuPackedMatrix< Real >::~CuPackedMatrix().

                                   {
#if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    if (this->data_ != NULL) {
      CuDevice::Instantiate().Free(this->data_);
    }
  } else
#endif
  {
    if (this->data_ != NULL) KALDI_MEMALIGN_FREE(this->data_);
  }
  this->data_ = NULL;
  this->num_rows_ = 0;
}

Mat() [1/2]

const PackedMatrix<Real>& Mat ( ) const

inlineprotected

Definition at line 150 of file cu-packed-matrix.h.

Referenced by CuPackedMatrix< Real >::AddPacked(), CuVectorBase< float >::CopyDiagFromPacked(), and CuPackedMatrix< Real >::CopyFromPacked().

                                               {
    return *(reinterpret_cast<const PackedMatrix<Real>* >(this));
  }

Mat() [2/2]

PackedMatrix<Real>& Mat ( )

inlineprotected

Definition at line 153 of file cu-packed-matrix.h.

                                   {
    return *(reinterpret_cast<PackedMatrix<Real>* >(this));
  }

NumCols()

MatrixIndexT NumCols ( ) const

inline

Definition at line 136 of file cu-packed-matrix.h.

References CuPackedMatrix< Real >::num_rows_.

Referenced by CuVectorBase< float >::AddSpVec(), and CuMatrix< float >::CuMatrix().

{ return num_rows_; }

NumRows()

MatrixIndexT NumRows ( ) const

inline

Definition at line 135 of file cu-packed-matrix.h.

References CuPackedMatrix< Real >::num_rows_.

Referenced by CuPackedMatrix< Real >::AddPacked(), CuVectorBase< float >::AddSpVec(), CuVectorBase< float >::AddTpVec(), CuSpMatrix< Real >::ApproxEqual(), kaldi::AssertEqual(), CuVectorBase< float >::CopyDiagFromPacked(), CuPackedMatrix< Real >::CopyFromPacked(), CuMatrixBase< float >::CopyFromSp(), CuMatrixBase< float >::CopyFromTp(), CuMatrix< float >::CuMatrix(), CuTpMatrix< Real >::Invert(), CuVectorBase< float >::MulTp(), CuSpMatrix< Real >::operator=(), CuTpMatrix< Real >::operator=(), TpMatrix< float >::TpMatrix(), kaldi::TraceSpSp(), and kaldi::UnitTestCuPackedMatrixSetUnit().

{ return num_rows_; }

operator()()

Real operator() ( MatrixIndexT  r, MatrixIndexT  c  ) const

inline

Definition at line 116 of file cu-packed-matrix.h.

References CuPackedMatrix< Real >::data_, KALDI_ASSERT, CuPackedMatrix< Real >::num_rows_, and kaldi::swap().

                                                                {
    if (static_cast<UnsignedMatrixIndexT>(c) >
        static_cast<UnsignedMatrixIndexT>(r))
      std::swap(c, r);
    KALDI_ASSERT(static_cast<UnsignedMatrixIndexT>(r) <
                 static_cast<UnsignedMatrixIndexT>(this->num_rows_));
#if HAVE_CUDA == 1
    if (CuDevice::Instantiate().Enabled()) {    
      Real value;
      CU_SAFE_CALL(cudaMemcpyAsync(&value, this->data_ + (r * (r+1)) / 2 + c,
                                   sizeof(Real), cudaMemcpyDeviceToHost,
                                   cudaStreamPerThread));
      CU_SAFE_CALL(cudaStreamSynchronize(cudaStreamPerThread));
      return value;
    } else
#endif
    return this->data_[(r * (r+1)) / 2 + c];
  }

operator=()

PackedMatrix<Real>& operator= ( const PackedMatrix< Real > &  other )

private

Read()

void Read	(	std::istream &	in,
		bool	binary
	)

Definition at line 231 of file cu-packed-matrix.cc.

References PackedMatrix< Real >::Read().

Referenced by kaldi::UnitTestCuSpMatrixIO(), kaldi::UnitTestCuTpMatrixIO(), and CuPackedMatrix< Real >::~CuPackedMatrix().

                                                           {
  PackedMatrix<Real> temp;
  temp.Read(is, binary);
  Destroy();
  Swap(&temp);
}

Resize()

void Resize	(	MatrixIndexT	nRows,
		MatrixResizeType	resize_type = kSetZero
	)

Set packed matrix to a specified size (can be zero).

The value of the new data depends on resize_type: -if kSetZero, the new data will be zero -if kUndefined, the new data will be undefined -if kCopyData, the new data will be the same as the old data in any shared positions, and zero elsewhere. This function takes time proportional to the number of data elements.

Definition at line 40 of file cu-packed-matrix.cc.

References data_, KALDI_ASSERT, kaldi::kSetZero, and kaldi::kUndefined.

Referenced by CuPackedMatrix< Real >::CuPackedMatrix(), CuTpMatrix< Real >::operator=(), CuSpMatrix< Real >::Resize(), TpMatrix< float >::TpMatrix(), and CuPackedMatrix< Real >::~CuPackedMatrix().

                                                                {
  // This code does not currently support the other resize_type options.
  KALDI_ASSERT(resize_type == kSetZero || resize_type == kUndefined);

  if (this->num_rows_ == rows) {
    if (resize_type == kSetZero) this->SetZero();
    return;
  }

  if (this->num_rows_ != 0)
    this->Destroy();
  if (rows == 0) return;
#if HAVE_CUDA == 1
  CuDevice &device = CuDevice::Instantiate();
  if (device.Enabled()) {
    CuTimer tim;
    this->num_rows_ = rows;
    size_t nr = static_cast<size_t>(num_rows_),
        num_bytes = ((nr * (nr+1)) / 2) * sizeof(Real);
    this->data_ = static_cast<Real*>(device.Malloc(num_bytes));

    if (resize_type == kSetZero) this->SetZero();
    device.AccuProfile("CuPackedMatrix::Resize", tim);
  } else
#endif
  { // Let the initializer of SpMatrix<Real> handle the allocation,
    // and then just do Swap which will switch the pointers.
    // This wastes a few instructions but is simple to code.
    SpMatrix<Real> mat(rows, resize_type);
    this->Swap(&mat);
  }
}

Scale()

void Scale

(

Real

alpha

)

Definition at line 299 of file cu-packed-matrix.cc.

References data_.

Referenced by CuPackedMatrix< Real >::CuPackedMatrix(), and kaldi::UnitTestCuPackedMatrixScale().

                                           {
#if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    CuTimer tim;
    size_t nr = static_cast<size_t>(num_rows_),
        num_elements = ((nr * (nr+1)) / 2);
    CUBLAS_SAFE_CALL(cublas_scal(GetCublasHandle(), num_elements, alpha, data_, 1));

    CuDevice::Instantiate().AccuProfile("CuPackedMatrix::Scale", tim);
  } else
#endif
  {
    Mat().Scale(alpha);
  }
}

ScaleDiag()

void ScaleDiag

(

Real

alpha

)

Definition at line 316 of file cu-packed-matrix.cc.

References CU1DBLOCK, and data_.

Referenced by CuPackedMatrix< Real >::CuPackedMatrix(), kaldi::UnitTestCuPackedMatrixScaleDiag(), and kaldi::nnet2::UnitTestPreconditionDirections().

                                               {
#if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    CuTimer tim;
    int dimBlock(CU1DBLOCK);
    int dimGrid(n_blocks(NumRows(),CU1DBLOCK));
    cuda_scale_diag_packed(dimGrid,dimBlock,data_,alpha,num_rows_);
    CU_SAFE_CALL(cudaGetLastError());
    CuDevice::Instantiate().AccuProfile("CuPackedMatrix::ScaleDiag", tim);
  } else
#endif
  {
    Mat().ScaleDiag(alpha);
  }
}

SetDiag()

void SetDiag

(

Real

alpha

)

< Set to random values of a normal distribution

Definition at line 281 of file cu-packed-matrix.cc.

References CU1DBLOCK, and data_.

Referenced by CuPackedMatrix< Real >::CuPackedMatrix().

                                             {
#if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    if (num_rows_ == 0) return;
    CuTimer tim;
    int dimBlock(CU1DBLOCK);
    int dimGrid(n_blocks(NumRows(),CU1DBLOCK));
    cuda_set_diag_packed(dimGrid,dimBlock,data_,alpha,num_rows_);
    CU_SAFE_CALL(cudaGetLastError());
    CuDevice::Instantiate().AccuProfile("CuPackedMatrix::SetDiag", tim);
  } else
#endif
  {
    Mat().SetDiag(alpha);
  }
}

SetRandn()

void SetRandn

(

)

< Set to unit matrix.

Definition at line 75 of file cu-packed-matrix.cc.

References data_, and CuRand< Real >::RandGaussian().

Referenced by CuPackedMatrix< Real >::CuPackedMatrix(), kaldi::CuVectorUnitTestAddSpVec(), kaldi::CuVectorUnitTestCopyDiagFromPacked(), kaldi::TestCuMatrixCopyFromSp(), kaldi::TestCuMatrixCopyFromTp(), kaldi::UnitTestCuSpMatrixInvert(), kaldi::UnitTestCuSpMatrixIO(), kaldi::UnitTestCuSpMatrixSetUnit(), and kaldi::UnitTestCuTpMatrixIO().

                                    {
  if (num_rows_ != 0) {
    MatrixIndexT size = num_rows_ * (num_rows_ + 1) / 2;
    CuSubVector<Real> tmp(data_, size);
    CuRand<Real> rand;
    rand.RandGaussian(&tmp);
  }
}

SetUnit()

void SetUnit

(

)

< Set to zero

Definition at line 370 of file cu-packed-matrix.cc.

Referenced by CuPackedMatrix< Real >::CuPackedMatrix(), kaldi::UnitTestCuPackedMatrixSetUnit(), and kaldi::nnet2::UnitTestPreconditionDirections().

                                   {
#if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    this->SetZero();
    this->SetDiag(1.0);
  } else
#endif
  {
    Mat().SetUnit();
  }
}

SetZero()

void SetZero

(

)

Definition at line 246 of file cu-packed-matrix.cc.

References data_.

Referenced by CuPackedMatrix< Real >::CuPackedMatrix(), and kaldi::UnitTestCuSpMatrixSetUnit().

                                   {
  #if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    CuTimer tim;
    size_t nr = static_cast<size_t>(num_rows_),
      num_bytes = ((nr * (nr+1)) / 2) * sizeof(Real);

    CU_SAFE_CALL(cudaMemsetAsync(reinterpret_cast<void*>(this->data_), 0, 
          num_bytes, cudaStreamPerThread));
    CuDevice::Instantiate().AccuProfile("CuPackedMatrix::SetZero", tim);
  } else
  #endif
  {
    Mat().SetZero();
  }
}

SizeInBytes()

size_t SizeInBytes ( ) const

inline

Returns size in bytes of the data held by the matrix.

Definition at line 139 of file cu-packed-matrix.h.

References CuPackedMatrix< Real >::num_rows_.

                              {
    size_t nr = static_cast<size_t>(num_rows_),
      num_bytes = ((nr * (nr+1)) / 2) * sizeof(Real);
    return num_bytes;
  }

Swap() [1/2]

void Swap

(

CuPackedMatrix< Real > *

other

)

Swaps the contents of *this and *other. Shallow swap.

Referenced by CuPackedMatrix< Real >::~CuPackedMatrix().

Swap() [2/2]

void Swap

(

PackedMatrix< Real > *

other

)

Swaps the contents of *this and *other.

Definition at line 101 of file cu-packed-matrix.cc.

References PackedMatrix< Real >::data_, kaldi::kUndefined, PackedMatrix< Real >::num_rows_, PackedMatrix< Real >::Resize(), PackedMatrix< Real >::Swap(), and kaldi::swap().

                                                       {
#if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    if (this->num_rows_ == 0) {
      if (mat->num_rows_ != 0) {
        // *this is empty, but mat is nonempty.
        Resize(mat->num_rows_, kUndefined);
        CopyFromPacked(*mat);
        mat->Resize(0);
      }
      // else both are empty.
    } else { // *this is nonempty.
      if (mat->num_rows_ != 0) {
        // Both *this and *mat are nonempty.  Recurse to simpler cases.
        // this could be done more efficiently in the case where
        // the size does not change.
        PackedMatrix<Real> temp;
        this->Swap(&temp); // now temp is full, *this is empty.
        mat->Swap(&temp); // now mat has data from *this, temp has
        // data from mat.
        this->Swap(&temp); // copy data in mat to *this, which is now empty.
      } else { // *this is full but *mat is empty.
        mat->Resize(this->num_rows_, kUndefined);
        this->CopyToPacked(mat);
        this->Destroy();
      }
    }
  } else
#endif
  {
    std::swap(mat->data_, this->data_);
    std::swap(mat->num_rows_, this->num_rows_);
  }
}

Trace()

Real Trace

(

)

const

Definition at line 264 of file cu-packed-matrix.cc.

References CuVectorBase< Real >::CopyDiagFromPacked(), kaldi::kUndefined, and CuVectorBase< Real >::Sum().

Referenced by CuPackedMatrix< Real >::CuPackedMatrix(), kaldi::UnitTestCuPackedMatrixTrace(), kaldi::UnitTestCuSpMatrixInvert(), and kaldi::UnitTestTrace().

                                       {
  Real result = 0.0;
#if HAVE_CUDA == 1
  if (CuDevice::Instantiate().Enabled()) {
    if (num_rows_ == 0) return 0.0;
    CuVector<Real> tmp(num_rows_, kUndefined);
    tmp.CopyDiagFromPacked(*this);
    return tmp.Sum();
  } else
#endif
  {
    result = Mat().Trace();
  }
  return result;
}

Write()

void Write	(	std::ostream &	out,
		bool	binary
	)		const

Definition at line 239 of file cu-packed-matrix.cc.

References kaldi::kUndefined, and PackedMatrix< Real >::Write().

Referenced by kaldi::UnitTestCuSpMatrixIO(), kaldi::UnitTestCuTpMatrixIO(), and CuPackedMatrix< Real >::~CuPackedMatrix().

                                                                  {
  PackedMatrix<Real> temp(this->num_rows_, kUndefined);
  this->CopyToPacked(&temp);
  temp.Write(os, binary);
}

Friends And Related Function Documentation

CuMatrixBase< Real >

friend class CuMatrixBase< Real >

friend

Definition at line 55 of file cu-packed-matrix.h.

CuRand< Real >

friend class CuRand< Real >

friend

Definition at line 58 of file cu-packed-matrix.h.

CuSubMatrix< Real >

friend class CuSubMatrix< Real >

friend

Definition at line 57 of file cu-packed-matrix.h.

CuVectorBase< Real >

friend class CuVectorBase< Real >

friend

Definition at line 56 of file cu-packed-matrix.h.

Member Data Documentation

data_

Real* data_

protected

Definition at line 160 of file cu-packed-matrix.h.

Referenced by CuTpMatrix< Real >::CopyFromMat(), CuPackedMatrix< Real >::CopyFromPacked(), CuPackedMatrix< Real >::Data(), CuSpMatrix< Real >::operator()(), and CuPackedMatrix< Real >::operator()().

num_rows_

MatrixIndexT num_rows_

protected

Definition at line 161 of file cu-packed-matrix.h.

Referenced by CuTpMatrix< Real >::CopyFromMat(), CuTpMatrix< Real >::Invert(), CuPackedMatrix< Real >::NumCols(), CuPackedMatrix< Real >::NumRows(), CuSpMatrix< Real >::operator()(), CuPackedMatrix< Real >::operator()(), and CuPackedMatrix< Real >::SizeInBytes().

---

The documentation for this class was generated from the following files:

• matrix/matrix-common.h
• cudamatrix/cu-packed-matrix.h
• cudamatrix/cu-packed-matrix.cc