# MLCommon Namespace

#### WARNING
Primarily internal API: it may change or disappear without notice and has no
stability, deprecation, backward-compatibility, or input-validation
guarantees. Callers must validate inputs and satisfy all memory, stream, and
lifetime preconditions. Prefer the supported [Python API](../../../../api/index.html.md).

### namespace MLCommon

### class TimerCPU

 *#include <Timer.h>*

### namespace Datasets

### namespace Boston

### Variables

### const std::vector<float> boston

### static const int n_samples = 506

### static const int n_features = 13

### namespace BreastCancer

### Variables

### const std::vector<float> breast_cancer

### static const int n_samples = 569

### static const int n_features = 30

### namespace Diabetes

### Variables

### const std::vector<float> diabetes

### static const int n_samples = 442

### static const int n_features = 10

### namespace Digits

### Variables

### const std::vector<float> digits

### static const std::vector<float> distances

### static const std::vector<long> indices

### static const std::vector<float> Y

### static const std::vector<float> VAL_vector

### static const std::vector<int> ROW_vector

### static const std::vector<int> COL_vector

### static const int n_samples = 1797

### static const int n_features = 64

### namespace LinAlg

### namespace opg

### Functions

### void eigDC(const raft::handle_t &h, float \*eigenValues, float \*eigenVectors, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &inParts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream)

Multi-GPU version of Eigen decomposition. This function works for symmetric matrices only. Whole input matrix is gathered at rank 0 and Eigen decomposition is carried out sequentially. 

* **Parameters:**
  * **h** – cuML handle object. 
  * **eigenValues** – **[out]** Output N Eigen values. 
  * **eigenVectors** – **[out]** Output N Eigen vectors of size N x 1. 
  * **inParts** – **[in]** Input symmetric matrix of size N x N. 
  * **desc** – **[in]** Descriptor of input matrix. 
  * **myRank** – **[in]** MPI rank of the process. 
  * **stream** – **[in]** CUDA stream. 

### void eigDC(const raft::handle_t &h, double \*eigenValues, double \*eigenVectors, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &inParts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream)

### void eigJacobi(const raft::handle_t &h, float \*eigenValues, float \*eigenVectors, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &inParts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream)

### void eigJacobi(const raft::handle_t &h, double \*eigenValues, double \*eigenVectors, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &inParts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream)

### void gemm(const raft::handle_t &h, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &outZParts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &outZDesc, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &inXParts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inXDesc, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &inYParts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inYDesc, int myRank, cudaStream_t stream)

A multi gpu generalized matrix multiplication function. This function performs Z = X \* Y The X and Y matrix are distributed in blocks on different ranks. First Y matrix is duplicated at each rank. It is multiplied with blocks of X local to the rank. 

* **Parameters:**
  * **h** – cuML handle object. 
  * **outZParts** – **[out]** Result of the multiplication with size M x N. Distributed across ranks according to inXDesc/outZDesc descriptor. 
  * **outZDesc** – **[out]** Descriptor for outZParts matrix. It has to be same as inXDesc. 
  * **inXParts** – **[in]** Input matrix X with dimensions M x K. Distributed across ranks according to inXDesc descriptor. 
  * **inXDesc** – **[in]** Descriptor for X matrix. 
  * **inYParts** – **[in]** Input matrix Y with dimensions K x N. Distributed across ranks according to inYDesc descriptor. 
  * **inYDesc** – **[in]** Descriptor for Y matrix. 
  * **myRank** – **[in]** Rank of calling process. 
  * **stream** – **[in]** cuda stream on which work is launched. 

### void gemm(const raft::handle_t &h, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &outZParts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &outZDesc, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &inXParts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inXDesc, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &inYParts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inYDesc, int myRank, cudaStream_t stream)

### void lstsqEig(const raft::handle_t &handle, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &A, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &ADesc, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &b, float \*w, cudaStream_t \*streams, int n_streams)

performs MNMG Least squares calculation. 

### void lstsqEig(const raft::handle_t &handle, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &A, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &ADesc, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &b, double \*w, cudaStream_t \*streams, int n_streams)

### void meanSquaredError(double \*out, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &in1, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &in1Desc, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &in2, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &in2Desc, const raft::comms::comms_t &comm, cudaStream_t stream, int root = 0, bool broadcastResult = true)

multi-gpu mean squared error 

* **Parameters:**
  * **out** – **[out]** the output data (device pointer) 
  * **in1** – **[in]** the first input data 
  * **in1Desc** – **[in]** descriptor for the first input data 
  * **in2** – **[in]** the second input data 
  * **in2Desc** – **[in]** descriptor for the second input data 
  * **comm** – **[in]** the communicator object 
  * **stream** – **[in]** cuda stream where to launch work 
  * **root** – **[in]** worker ID which is supposed to be considered as root 
  * **broadcastResult** – **[in]** if false, only root process will have the result, else all ranks 

### void meanSquaredError(float \*out, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &in1, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &in1Desc, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &in2, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &in2Desc, const raft::comms::comms_t &comm, cudaStream_t stream, int root = 0, bool broadcastResult = true)

### void mm_aTa(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &out, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &in1, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &in1Desc, cudaStream_t \*streams, int n_streams)

performs MNMG A^T x A calculation. 

* **Parameters:**
  * **handle** – **[in]** cuML handle object 
  * **out** – **[out]** resulting A^T x A matrix 
  * **in1** – **[in]** input data 
  * **in1Desc** – **[in]** MNMG description of the input 
  * **streams** – **[in]** cuda streams 
  * **n_streams** – **[in]** number of streams 

### void mm_aTa(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &out, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &in1, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &in1Desc, cudaStream_t \*streams, int n_streams)

### void mv_aTb(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &out, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &A, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &ADesc, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &b, cudaStream_t \*streams, int n_streams)

performs MNMG A^T x b calculation. 

* **Parameters:**
  * **handle** – **[in]** cuML handle object 
  * **out** – **[out]** resulting A^T x b matrix 
  * **A** – **[in]** input data matrix 
  * **ADesc** – **[in]** MNMG description of A 
  * **b** – **[in]** input vector 
  * **streams** – **[in]** cuda streams 
  * **n_streams** – **[in]** number of streams 

### void mv_aTb(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &out, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &A, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &ADesc, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &b, cudaStream_t \*streams, int n_streams)

### void colNorm2(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &out, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &in, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inDesc, cudaStream_t \*streams, int n_streams)

performs MNMG column-wise L2 norm calculation. 

### void colNorm2(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &out, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &in, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inDesc, cudaStream_t \*streams, int n_streams)

### void colNorm2NoSeq(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &out, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &in, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inDesc, cudaStream_t \*streams, int n_streams)

### void colNorm2NoSeq(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &out, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &in, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inDesc, cudaStream_t \*streams, int n_streams)

### void svdEig(const raft::handle_t &handle, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &A, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &ADesc, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &U, float \*S, float \*V, cudaStream_t \*streams, int n_streams)

performs MNMG Least squares calculation. 

### void svdEig(const raft::handle_t &handle, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &A, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &ADesc, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &U, double \*S, double \*V, cudaStream_t \*streams, int n_streams)

### namespace Matrix

### Typedefs

### typedef [Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> floatData_t

### typedef [Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> doubleData_t

### Enums

### enum Layout

Describes the data layout 

*Values:*

### enumerator LayoutRowMajor

row major layout 

### enumerator LayoutColMajor

column major layout 

### Functions

### std::ostream &operator<<(std::ostream &os, const [PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc)

Print matrix descriptor in human readable form 

### bool operator==(const [PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &a, const [PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &b)

compare 2 descriptor objects 

### template<typename Type>struct Data

 *#include <data.hpp>*

This is a *helper* wrapper around the multi-gpu data blocks owned by a worker. It’s design is NOT final. Its so written this way to get something concrete in a short span of time. 

### *Todo:*

add support for custom allocators 

### Public Functions

### inline size_t numElements() const

Return the number of elements of Type in ptr. 

### Public Members

### [Type](#_CPPv4I0EN8MLCommon6Matrix4DataE) \*ptr = nullptr

actual data block. This is just a linearly laid out buffer of all blocks owned by this worker 

### size_t totalSize = (size_t)0

total size (in bytes) of this buffer. In future, this will be passed to the dealloc function underneath 

### struct PartDescriptor

 *#include <part_descriptor.hpp>*

### Public Functions

### PartDescriptor(size_t \_M, size_t \_N, const std::vector<[RankSizePair](#_CPPv4N8MLCommon6Matrix12RankSizePairE)\*> &\_partsToRanks, int rank, [Layout](#_CPPv4N8MLCommon6Matrix6LayoutE) \_layout = LayoutColMajor)

For a given matrix and block-sizes construct the corresponding descriptor for it. This is useful when we are dealing with standard row/column-wise block-cyclic data distribution, as seen in other popular multi-node packages like magma etc. 

* **Parameters:**
  * **\_M** – total number of rows of this matrix 
  * **\_N** – total number of columns 
  * **\_partsToRanks** – mapping of ranks to parts and sizes 
  * **rank** – MPI rank of the process 
  * **\_layout** – data layout (row or column major) 

### inline int totalBlocks() const

total number of blocks across all workers 

### int totalBlocksOwnedBy(int rank) const

Count the total number of blocks owned by a given rank 

### std::vector<[RankSizePair](#_CPPv4N8MLCommon6Matrix12RankSizePairE)\*> blocksOwnedBy(int rank) const

Returns the vector of blocks (each identified by linearBLockIndex) owned by the given rank. 

### size_t totalElementsOwnedBy(int rank) const

Count the total number of matrix elements owned by a given rank 

### Public Members

### size_t M

total number of rows 

### size_t N

total number of columns 

### std::vector<[RankSizePair](#_CPPv4N8MLCommon6Matrix12RankSizePairE)\*> partsToRanks

mapping of each block (in col-major order) to the device that owns it 

### Friends

### friend std::ostream &operator<<(std::ostream &os, const [PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc)

Print matrix descriptor in human readable form 

### friend bool operator==(const [PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &a, const [PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &b)

compare 2 descriptor objects 

### struct RankSizePair

 *#include <part_descriptor.hpp>*

### Public Members

### size_t size

Total number of rows 

### namespace opg

### Functions

### template<bool rowMajor, bool bcastAlongRows>void matrixVectorBinaryDivSkipZero(std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &data, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inDesc, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &vec, bool return_zero, const raft::comms::comms_t &comm, cudaStream_t \*streams, int n_streams)

### template<bool rowMajor, bool bcastAlongRows>void matrixVectorBinaryDivSkipZero(std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &data, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inDesc, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &vec, bool return_zero, const raft::comms::comms_t &comm, cudaStream_t \*streams, int n_streams)

### template<bool rowMajor, bool bcastAlongRows>void matrixVectorBinaryMult(std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &data, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inDesc, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &vec, const raft::comms::comms_t &comm, cudaStream_t \*streams, int n_streams)

### template<bool rowMajor, bool bcastAlongRows>void matrixVectorBinaryMult(std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &data, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inDesc, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &vec, const raft::comms::comms_t &comm, cudaStream_t \*streams, int n_streams)

### void gatherPart(const raft::handle_t &h, float \*gatheredPart, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int partIndex, int rootRank, int myRank, cudaStream_t stream)

### void allGatherPart(const raft::handle_t &h, float \*gatheredPart, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int partIndex, int myRank, cudaStream_t stream)

### void gather(const raft::handle_t &h, float \*gatheredMatrix, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int rootRank, int myRank, cudaStream_t stream)

### void allGather(const raft::handle_t &h, float \*gatheredMatrix, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream)

### void allocate(const raft::handle_t &h, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream)

### void deallocate(const raft::handle_t &h, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream)

### void randomize(const raft::handle_t &h, raft::random::Rng &r, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream, float low = -1.0f, float high = 1.0f)

### void reset(const raft::handle_t &h, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream)

### void printRaw2D(float \*buffer, int rows, int cols, bool isColMajor, cudaStream_t stream)

### void print(const raft::handle_t &h, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, const char \*matrixName, int myRank, cudaStream_t stream)

### void gatherPart(const raft::handle_t &h, double \*gatheredPart, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int partIndex, int rootRank, int myRank, cudaStream_t stream)

### void allGatherPart(const raft::handle_t &h, double \*gatheredPart, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int partIndex, int myRank, cudaStream_t stream)

### void gather(const raft::handle_t &h, double \*gatheredMatrix, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int rootRank, int myRank, cudaStream_t stream)

### void allGather(const raft::handle_t &h, double \*gatheredMatrix, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream)

### void allocate(const raft::handle_t &h, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream)

### void deallocate(const raft::handle_t &h, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream)

### void randomize(const raft::handle_t &h, raft::random::Rng &r, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream, double low = -1.0, double high = 1.0)

### void reset(const raft::handle_t &h, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, int myRank, cudaStream_t stream)

### void printRaw2D(double \*buffer, int rows, int cols, bool isColMajor, cudaStream_t stream)

### void print(const raft::handle_t &h, std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &parts, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &desc, const char \*matrixName, int myRank, cudaStream_t stream)

### namespace opg

### Functions

### template<typename T>void reduce_single_sum([T](#_CPPv4I0EN8MLCommon3opg17reduce_single_sumEvP1TPK1TRKN4raft5comms7comms_tE12cudaStream_ti) \*out, const [T](#_CPPv4I0EN8MLCommon3opg17reduce_single_sumEvP1TPK1TRKN4raft5comms7comms_tE12cudaStream_ti) \*in, const raft::comms::comms_t &comm, cudaStream_t stream, int root = 0)

sum-reduce single float numbers across workers 

### template<typename T>void allreduce_single_sum([T](#_CPPv4I0EN8MLCommon3opg20allreduce_single_sumEvP1TPK1TRKN4raft5comms7comms_tE12cudaStream_t) \*out, const [T](#_CPPv4I0EN8MLCommon3opg20allreduce_single_sumEvP1TPK1TRKN4raft5comms7comms_tE12cudaStream_t) \*in, const raft::comms::comms_t &comm, cudaStream_t stream)

all-reduce single float numbers across workers in Comm 

### namespace Stats

### namespace opg

### Functions

### void cov(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &covar, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &data, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &dataDesc, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &mu, bool sample, cudaStream_t \*streams, int n_streams)

performs MNMG covariance calculation. 

* **Parameters:**
  * **handle** – **[in]** cuML handle object 
  * **covar** – **[out]** resulting covariance matrix 
  * **data** – **[in]** the data that cov matrix is calculated for 
  * **dataDesc** – **[in]** MNMG description of the input data 
  * **mu** – **[in]** mean of every column in data 
  * **sample** – **[in]** whether to compute sample covariance 
  * **streams** – **[in]** cuda streams 
  * **n_streams** – **[in]** number of streams 

### void cov(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &covar, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &data, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &dataDesc, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &mu, bool sample, cudaStream_t \*streams, int n_streams)

### void mean(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &out, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &in, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inDesc, cudaStream_t \*streams, int n_streams)

performs MNMG mean calculation. 

* **Parameters:**
  * **handle** – **[in]** cuML handle object 
  * **out** – **[out]** mean of every column of the input 
  * **in** – **[in]** all the data partitions 
  * **inDesc** – **[in]** MNMG description of the input data 
  * **streams** – **[in]** cuda streams 
  * **n_streams** – **[in]** number of streams 

### void mean(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &out, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &in, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inDesc, cudaStream_t \*streams, int n_streams)

### void mean_center(const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &data, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &dataDesc, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &mu, const raft::comms::comms_t &comm, cudaStream_t \*streams, int n_streams)

performs MNMG mean subtraction calculation. 

* **Parameters:**
  * **data** – **[inout]** the data that mean of every column is added 
  * **dataDesc** – **[in]** MNMG description of the input data 
  * **mu** – **[in]** mean of every column in data 
  * **comm** – **[in]** communicator 
  * **streams** – **[in]** cuda streams 
  * **n_streams** – **[in]** number of streams 

### void mean_center(const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &data, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &dataDesc, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &mu, const raft::comms::comms_t &comm, cudaStream_t \*streams, int n_streams)

### void mean_add(const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &data, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &dataDesc, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &mu, const raft::comms::comms_t &comm, cudaStream_t \*streams, int n_streams)

performs MNMG mean add calculation. 

* **Parameters:**
  * **data** – **[inout]** the data that mean of every column is added 
  * **dataDesc** – **[in]** MNMG description of the input data 
  * **mu** – **[in]** mean of every column in data 
  * **comm** – **[in]** communicator 
  * **streams** – **[in]** cuda streams 
  * **n_streams** – **[in]** number of streams 

### void mean_add(const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &data, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &dataDesc, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &mu, const raft::comms::comms_t &comm, cudaStream_t \*streams, int n_streams)

### void var(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float> &out, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<float>\*> &in, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inDesc, const float \*mu, cudaStream_t \*streams, int n_streams)

performs MNMG var calculation. 

* **Parameters:**
  * **handle** – **[in]** cuML handle object 
  * **out** – **[out]** var of every column of the input 
  * **in** – **[in]** all the data partitions 
  * **inDesc** – **[in]** MNMG description of the input data 
  * **mu** – **[in]** mean of every column in data 
  * **streams** – **[in]** cuda streams 
  * **n_streams** – **[in]** number of streams 

### void var(const raft::handle_t &handle, [Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double> &out, const std::vector<[Matrix](#_CPPv4N8MLCommon6MatrixE)::[Data](#_CPPv4I0EN8MLCommon6Matrix4DataE)<double>\*> &in, const [Matrix](#_CPPv4N8MLCommon6MatrixE)::[PartDescriptor](#_CPPv4N8MLCommon6Matrix14PartDescriptorE) &inDesc, const double \*mu, cudaStream_t \*streams, int n_streams)