rocSPARSE User Guide#
HIP Device Management#
Before starting a HIP kernel you can call hipSetDevice()
to set the device to run the kernel on, for example device 2. Unless you explicitly specify a different device HIP kernels always run on device 0. This is a HIP (and CUDA) device management approach and is not specific to the rocSPARSE library. rocSPARSE honors this approach and assumes you have set the preferred device before a rocSPARSE routine call.
Once you set the device, you can create a handle with rocsparse_create_handle(). Subsequent rocSPARSE routines take this handle as an input parameter. rocSPARSE ONLY queries (by hipGetDevice()
) the user’s device; rocSPARSE does NOT set the device for users. If rocSPARSE does not see a valid device, it returns an error message. It is your responsibility to provide a valid device to rocSPARSE and ensure the device safety.
The handle should be destroyed at the end using rocsparse_destroy_handle() to release the resources consumed by the rocSPARSE library. You CANNOT switch devices between rocsparse_create_handle() and rocsparse_destroy_handle(). If you want to change the device, you must destroy the current handle and create another rocSPARSE handle on a new device.
Note
hipSetDevice()
and hipGetDevice()
are NOT part of the rocSPARSE API. They are part of the HIP Runtime API - Device Management.
HIP Stream Management#
HIP kernels are always launched in a queue (also known as a stream). If you do not explicitly specify a stream, the system provides and maintains a default stream. You cannot create or destroy the default stream. However, you can freely create new streams (with hipStreamCreate()
) and bind it to a rocSPARSE handle using rocsparse_set_stream(). HIP kernels are invoked in rocSPARSE routines. The rocSPARSE handle is always associated with a stream, and rocSPARSE passes its stream to the kernels inside the routine. One rocSPARSE routine only takes one stream in a single invocation. If you create a stream, you are responsible for destroying it. Refer to HIP Runtime API - Stream Management for more information.
Asynchronous Execution#
All rocSPARSE library functions are non-blocking and executed asynchronously with respect to the host, except functions having memory allocation inside preventing asynchronicity. The function may return immediately, or before the actual computation has finished. To force synchronization, use either hipDeviceSynchronize()
or hipStreamSynchronize()
. This will ensure that all previously executed rocSPARSE functions on the device, or in the particular stream, have completed.
Multiple Streams and Multiple Devices#
If a system has multiple HIP devices, you can run multiple rocSPARSE handles concurrently. However, you can NOT run a single rocSPARSE handle concurrently on multiple discrete devices. Each handle can only be associated with a single device, and a new handle should be created for each additional device.
Storage Formats#
The following describes supported matrix storage formats.
Note
The different storage formats support indexing from a base of 0 or 1 as described in Storage schemes and indexing base.
COO storage format#
The Coordinate (COO) storage format represents a \(m \times n\) matrix by
m |
number of rows (integer). |
n |
number of columns (integer). |
nnz |
number of non-zero elements (integer). |
coo_val |
array of |
coo_row_ind |
array of |
coo_col_ind |
array of |
The COO matrix is expected to be sorted by row indices and column indices per row. Furthermore, each pair of indices should appear only once. Consider the following \(3 \times 5\) matrix and the corresponding COO structures, with \(m = 3, n = 5\) and \(\text{nnz} = 8\) using zero based indexing:
where
COO (AoS) storage format#
The Coordinate (COO) Array of Structure (AoS) storage format represents a \(m \times n\) matrix by
m |
number of rows (integer). |
n |
number of columns (integer). |
nnz |
number of non-zero elements (integer). |
coo_val |
array of |
coo_ind |
array of |
The COO (AoS) matrix is expected to be sorted by row indices and column indices per row. Furthermore, each pair of indices should appear only once. Consider the following \(3 \times 5\) matrix and the corresponding COO (AoS) structures, with \(m = 3, n = 5\) and \(\text{nnz} = 8\) using zero based indexing:
where
CSR storage format#
The Compressed Sparse Row (CSR) storage format represents a \(m \times n\) matrix by
m |
number of rows (integer). |
n |
number of columns (integer). |
nnz |
number of non-zero elements (integer). |
csr_val |
array of |
csr_row_ptr |
array of |
csr_col_ind |
array of |
The CSR matrix is expected to be sorted by column indices within each row. Furthermore, each pair of indices should appear only once. Consider the following \(3 \times 5\) matrix and the corresponding CSR structures, with \(m = 3, n = 5\) and \(\text{nnz} = 8\) using one based indexing:
where
CSC storage format#
The Compressed Sparse Column (CSC) storage format represents a \(m \times n\) matrix by
m |
number of rows (integer). |
n |
number of columns (integer). |
nnz |
number of non-zero elements (integer). |
csc_val |
array of |
csc_col_ptr |
array of |
csc_row_ind |
array of |
The CSC matrix is expected to be sorted by row indices within each column. Furthermore, each pair of indices should appear only once. Consider the following \(3 \times 5\) matrix and the corresponding CSC structures, with \(m = 3, n = 5\) and \(\text{nnz} = 8\) using one based indexing:
where
BSR storage format#
The Block Compressed Sparse Row (BSR) storage format represents a \((mb \cdot \text{bsr_dim}) \times (nb \cdot \text{bsr_dim})\) matrix by
mb |
number of block rows (integer) |
nb |
number of block columns (integer) |
nnzb |
number of non-zero blocks (integer) |
bsr_val |
array of |
bsr_row_ptr |
array of |
bsr_col_ind |
array of |
bsr_dim |
dimension of each block (integer). |
The BSR matrix is expected to be sorted by column indices within each row. If \(m\) or \(n\) are not evenly divisible by the block dimension, then zeros are padded to the matrix, such that \(mb = (m + \text{bsr_dim} - 1) / \text{bsr_dim}\) and \(nb = (n + \text{bsr_dim} - 1) / \text{bsr_dim}\). Consider the following \(4 \times 3\) matrix and the corresponding BSR structures, with \(\text{bsr_dim} = 2, mb = 2, nb = 2\) and \(\text{nnzb} = 4\) using zero based indexing and column-major storage:
with the blocks \(A_{ij}\)
such that
with arrays representation
GEBSR storage format#
The General Block Compressed Sparse Row (GEBSR) storage format represents a \((mb \cdot \text{bsr_row_dim}) \times (nb \cdot \text{bsr_col_dim})\) matrix by
mb |
number of block rows (integer) |
nb |
number of block columns (integer) |
nnzb |
number of non-zero blocks (integer) |
bsr_val |
array of |
bsr_row_ptr |
array of |
bsr_col_ind |
array of |
bsr_row_dim |
row dimension of each block (integer). |
bsr_col_dim |
column dimension of each block (integer). |
The GEBSR matrix is expected to be sorted by column indices within each row. If \(m\) is not evenly divisible by the row block dimension or \(n\) is not evenly divisible by the column block dimension, then zeros are padded to the matrix, such that \(mb = (m + \text{bsr_row_dim} - 1) / \text{bsr_row_dim}\) and \(nb = (n + \text{bsr_col_dim} - 1) / \text{bsr_col_dim}\). Consider the following \(4 \times 5\) matrix and the corresponding GEBSR structures, with \(\text{bsr_row_dim} = 2\), \(\text{bsr_col_dim} = 3\), mb = 2, nb = 2` and \(\text{nnzb} = 4\) using zero based indexing and column-major storage:
with the blocks \(A_{ij}\)
such that
with arrays representation
ELL storage format#
The Ellpack-Itpack (ELL) storage format represents a \(m \times n\) matrix by
m |
number of rows (integer). |
n |
number of columns (integer). |
ell_width |
maximum number of non-zero elements per row (integer) |
ell_val |
array of |
ell_col_ind |
array of |
The ELL matrix is assumed to be stored in column-major format. Rows with less than ell_width
non-zero elements are padded with zeros (ell_val
) and \(-1\) (ell_col_ind
).
Consider the following \(3 \times 5\) matrix and the corresponding ELL structures, with \(m = 3, n = 5\) and \(\text{ell_width} = 3\) using zero based indexing:
where
HYB storage format#
The Hybrid (HYB) storage format represents a \(m \times n\) matrix by
m |
number of rows (integer). |
n |
number of columns (integer). |
nnz |
number of non-zero elements of the COO part (integer) |
ell_width |
maximum number of non-zero elements per row of the ELL part (integer) |
ell_val |
array of |
ell_col_ind |
array of |
coo_val |
array of |
coo_row_ind |
array of |
coo_col_ind |
array of |
The HYB format is a combination of the ELL and COO sparse matrix formats. Typically, the regular part of the matrix is stored in ELL storage format, and the irregular part of the matrix is stored in COO storage format. Three different partitioning schemes can be applied when converting a CSR matrix to a matrix in HYB storage format. For further details on the partitioning schemes, see rocsparse_hyb_partition.
Storage schemes and indexing base#
rocSPARSE supports 0 and 1 based indexing.
The index base is selected by the rocsparse_index_base
type which is either passed as standalone parameter or as part of the rocsparse_mat_descr
type.
Furthermore, dense vectors are represented with a 1D array, stored linearly in memory. Sparse vectors are represented by a 1D data array stored linearly in memory that hold all non-zero elements and a 1D indexing array stored linearly in memory that hold the positions of the corresponding non-zero elements.
Pointer mode#
The auxiliary functions rocsparse_set_pointer_mode()
and rocsparse_get_pointer_mode()
are used to set and get the value of the state variable rocsparse_pointer_mode
.
If rocsparse_pointer_mode
is equal to rocsparse_pointer_mode_host
, then scalar parameters must be allocated on the host.
If rocsparse_pointer_mode
is equal to rocsparse_pointer_mode_device
, then scalar parameters must be allocated on the device.
There are two types of scalar parameter:
Scaling parameters, such as alpha and beta used for example in
rocsparse_scsrmv()
androcsparse_scoomv()
Scalar results from functions such as
rocsparse_sdoti()
orrocsparse_cdotci()
For scalar parameters such as alpha and beta, memory can be allocated on the host heap or stack, when rocsparse_pointer_mode
is equal to rocsparse_pointer_mode_host
.
The kernel launch is asynchronous, and if the scalar parameter is on the heap, it can be freed after the return from the kernel launch.
When rocsparse_pointer_mode
is equal to rocsparse_pointer_mode_device
, the scalar parameter must not be changed till the kernel completes.
For scalar results, when rocsparse_pointer_mode
is equal to rocsparse_pointer_mode_host
, the function blocks the CPU till the GPU has copied the result back to the host.
Using rocsparse_pointer_mode
equal to rocsparse_pointer_mode_device
, the function will return after the asynchronous launch.
Similarly to vector and matrix results, the scalar result is only available when the kernel has completed execution.
hipSPARSE#
hipSPARSE is a SPARSE marshalling library, with multiple supported backends. It sits between the application and a worker SPARSE library, marshalling inputs into the backend library and marshalling results back to the application. hipSPARSE exports an interface that does not require the client to change, regardless of the chosen backend. hipSPARSE supports rocSPARSE and cuSPARSE as backends.
hipSPARSE focuses on convenience and portability. If performance outweighs these factors, then using rocSPARSE itself is highly recommended. hipSPARSE can be found on GitHub.