Package org.apache.sysds.runtime.instructions.gpu.context

Class CSRPointer

java.lang.Object

org.apache.sysds.runtime.instructions.gpu.context.CSRPointer

public class CSRPointer
extends Object

Compressed Sparse Row (CSR) format for CUDA Generalized matrix multiply is implemented for CSR format in the cuSparse library among other operations Since we assume that the matrix is stored with zero-based indexing (i.e. CUSPARSE_INDEX_BASE_ZERO), the matrix 1.0 4.0 0.0 0.0 0.0 0.0 2.0 3.0 0.0 0.0 5.0 0.0 0.0 7.0 8.0 0.0 0.0 9.0 0.0 6.0 is stored as val = 1.0 4.0 2.0 3.0 5.0 7.0 8.0 9.0 6.0 rowPtr = 0.0 2.0 4.0 7.0 9.0 colInd = 0.0 1.0 1.0 2.0 0.0 3.0 4.0 2.0 4.0

Field Summary

Fields

Modifier and Type	Field	Description
jcuda.Pointer	colInd	integer array of nnz values' column indices
jcuda.jcusparse.cusparseMatDescr	descr	descriptor of matrix, only CUSPARSE_MATRIX_TYPE_GENERAL supported
static jcuda.jcusparse.cusparseMatDescr	matrixDescriptor
long	nnz	Number of non zeroes
jcuda.Pointer	rowPtr	integer array of start of all rows and end of last row + 1
jcuda.Pointer	val	double array of non zero values

Method Summary

Modifier and Type	Method	Description
static CSRPointer	allocateEmpty(GPUContext gCtx, long nnz2, long rows)
static CSRPointer	allocateEmpty(GPUContext gCtx, long nnz2, long rows, boolean initialize)	Factory method to allocate an empty CSR Sparse matrix on the GPU
static CSRPointer	allocateForDgeam(GPUContext gCtx, jcuda.jcusparse.cusparseHandle handle, CSRPointer A, CSRPointer B, int m, int n)	Estimates the number of non zero elements from the results of a sparse cusparseDgeam operation C = a op(A) + b op(B)
static CSRPointer	allocateForMatrixMultiply(GPUContext gCtx, jcuda.jcusparse.cusparseHandle handle, CSRPointer A, int transA, CSRPointer B, int transB, int m, int n, int k)	Estimates the number of non-zero elements from the result of a sparse matrix multiplication C = A * B and returns the CSRPointer to C with the appropriate GPU memory.
CSRPointer	clone(int rows)
static void	copyPtrToHost(CSRPointer src, int rows, long nnz, int[] rowPtr, int[] colInd)	Static method to copy a CSR sparse matrix from Device to host
static void	copyToDevice(GPUContext gCtx, CSRPointer dest, int rows, long nnz, int[] rowPtr, int[] colInd, double[] values)	Static method to copy a CSR sparse matrix from Host to Device
void	deallocate()	Calls cudaFree lazily on the allocated Pointer instances
void	deallocate(boolean eager)	Calls cudaFree lazily or eagerly on the allocated Pointer instances
static long	estimateSize(long nnz2, long rows)	Estimate the size of a CSR matrix in GPU memory Size of pointers is not needed and is not added in
static jcuda.jcusparse.cusparseMatDescr	getDefaultCuSparseMatrixDescriptor()
boolean	isUltraSparse(int rows, int cols)	Check for ultra sparsity
jcuda.Pointer	toColumnMajorDenseMatrix(jcuda.jcusparse.cusparseHandle cusparseHandle, jcuda.jcublas.cublasHandle cublasHandle, int rows, int cols, String instName)	Copies this CSR matrix on the GPU to a dense column-major matrix on the GPU.
static int	toIntExact(long l)
String	toString()

Methods inherited from class java.lang.Object

equals, getClass, hashCode, notify, notifyAll, wait, wait, wait

Field Detail

matrixDescriptor

public static jcuda.jcusparse.cusparseMatDescr matrixDescriptor

nnz

public long nnz

Number of non zeroes

val

public jcuda.Pointer val

double array of non zero values

rowPtr

public jcuda.Pointer rowPtr

integer array of start of all rows and end of last row + 1

colInd

public jcuda.Pointer colInd

integer array of nnz values' column indices

descr

public jcuda.jcusparse.cusparseMatDescr descr

descriptor of matrix, only CUSPARSE_MATRIX_TYPE_GENERAL supported

Method Detail

toIntExact

public static int toIntExact(long l)

getDefaultCuSparseMatrixDescriptor

public static jcuda.jcusparse.cusparseMatDescr getDefaultCuSparseMatrixDescriptor()

Returns:

Singleton default matrix descriptor object (set with CUSPARSE_MATRIX_TYPE_GENERAL, CUSPARSE_INDEX_BASE_ZERO)

estimateSize

public static long estimateSize(long nnz2,
                                long rows)

Estimate the size of a CSR matrix in GPU memory Size of pointers is not needed and is not added in

Parameters:

nnz2 - number of non zeroes

rows - number of rows

Returns:

size estimate

copyToDevice

public static void copyToDevice(GPUContext gCtx,
                                CSRPointer dest,
                                int rows,
                                long nnz,
                                int[] rowPtr,
                                int[] colInd,
                                double[] values)

Static method to copy a CSR sparse matrix from Host to Device

Parameters:

gCtx - GPUContext

dest - [input] destination location (on GPU)

rows - number of rows

nnz - number of non-zeroes

rowPtr - integer array of row pointers

colInd - integer array of column indices

values - double array of non zero values

copyPtrToHost

public static void copyPtrToHost(CSRPointer src,
                                 int rows,
                                 long nnz,
                                 int[] rowPtr,
                                 int[] colInd)

Static method to copy a CSR sparse matrix from Device to host

Parameters:

src - [input] source location (on GPU)

rows - [input] number of rows

nnz - [input] number of non-zeroes

rowPtr - [output] pre-allocated integer array of row pointers of size (rows+1)

colInd - [output] pre-allocated integer array of column indices of size nnz

allocateForDgeam

public static CSRPointer allocateForDgeam(GPUContext gCtx,
                                          jcuda.jcusparse.cusparseHandle handle,
                                          CSRPointer A,
                                          CSRPointer B,
                                          int m,
                                          int n)

Estimates the number of non zero elements from the results of a sparse cusparseDgeam operation C = a op(A) + b op(B)

Parameters:

gCtx - a valid GPUContext

handle - a valid cusparseHandle

A - Sparse Matrix A on GPU

B - Sparse Matrix B on GPU

m - Rows in A

n - Columns in Bs

Returns:

CSR (compressed sparse row) pointer

allocateForMatrixMultiply

public static CSRPointer allocateForMatrixMultiply(GPUContext gCtx,
                                                   jcuda.jcusparse.cusparseHandle handle,
                                                   CSRPointer A,
                                                   int transA,
                                                   CSRPointer B,
                                                   int transB,
                                                   int m,
                                                   int n,
                                                   int k)

Estimates the number of non-zero elements from the result of a sparse matrix multiplication C = A * B and returns the CSRPointer to C with the appropriate GPU memory.

Parameters:

gCtx - a valid GPUContext

handle - a valid cusparseHandle

A - Sparse Matrix A on GPU

transA - 'T' if A is to be transposed, 'N' otherwise

B - Sparse Matrix B on GPU

transB - 'T' if B is to be transposed, 'N' otherwise

m - Rows in A

n - Columns in B

k - Columns in A / Rows in B

Returns:

a CSRPointer instance that encapsulates the CSR matrix on GPU

allocateEmpty

public static CSRPointer allocateEmpty(GPUContext gCtx,
                                       long nnz2,
                                       long rows,
                                       boolean initialize)

Factory method to allocate an empty CSR Sparse matrix on the GPU

Parameters:

gCtx - a valid GPUContext

nnz2 - number of non-zeroes

rows - number of rows

initialize - memset to zero?

Returns:

a CSRPointer instance that encapsulates the CSR matrix on GPU

allocateEmpty

public static CSRPointer allocateEmpty(GPUContext gCtx,
                                       long nnz2,
                                       long rows)

clone

public CSRPointer clone(int rows)

isUltraSparse

public boolean isUltraSparse(int rows,
                             int cols)

Check for ultra sparsity

Parameters:

rows - number of rows

cols - number of columns

Returns:

true if ultra sparse

toColumnMajorDenseMatrix

public jcuda.Pointer toColumnMajorDenseMatrix(jcuda.jcusparse.cusparseHandle cusparseHandle,
                                              jcuda.jcublas.cublasHandle cublasHandle,
                                              int rows,
                                              int cols,
                                              String instName)

Copies this CSR matrix on the GPU to a dense column-major matrix on the GPU. This is a temporary matrix for operations such as cusparseDcsrmv. Since the allocated matrix is temporary, bookkeeping is not updated. The caller is responsible for calling "free" on the returned Pointer object

Parameters:

cusparseHandle - a valid cusparseHandle

cublasHandle - a valid cublasHandle

rows - number of rows in this CSR matrix

cols - number of columns in this CSR matrix

instName - name of the invoking instruction to recordStatistics.

Returns:

A Pointer to the allocated dense matrix (in column-major format)

deallocate

public void deallocate()

Calls cudaFree lazily on the allocated Pointer instances

deallocate

public void deallocate(boolean eager)

Calls cudaFree lazily or eagerly on the allocated Pointer instances

Parameters:

eager - whether to do eager or lazy cudaFrees

toString

public String toString()

Overrides:

toString in class Object