warp.sparse.BsrMatrix#

class warp.sparse.BsrMatrix[source]#

Untyped base class for BSR and CSR matrices.

Should not be constructed directly but through functions such as bsr_zeros().

nrow#

Number of rows of blocks.

Type:: int

ncol#

Number of columns of blocks.

Type:: int

nnz#

Host-side upper bound on the number of stored block slots used by sparse operations, including for dimensioning launches. Topology-changing operations may leave this larger than the active block count, even for compact matrices. The backing arrays may have additional allocated capacity. Call nnz_sync() before using nnz as an exact count.

Type:: int

offsets#

Array of size at least 1 + nrow such that the start and capacity end indices of row r are offsets[r] and offsets[r+1], respectively.

Type:: Array[int]

row_counts#

Optional array of size at least nrow containing the active block count of each row. Active blocks of row r are stored in offsets[r]:offsets[r] + row_counts[r]. For compact matrices, row_counts is None, in which case all storage in each row is active.

Type:: Array[int] | None

columns#

Array of size at least equal to nnz containing block column indices. Entries outside active row ranges are not part of the matrix and may be uninitialized.

Type:: Array[int]

values#

Array of size at least equal to nnz containing block values. Entries outside active row ranges are not part of the matrix and may be uninitialized. Active entries may also be uninitialized after topology-only or structure-only operations.

Type:: Array[BlockType]

Methods

`clear_status`()	Clear the asynchronous sparse status code.
`copy_nnz_async`()	Start the asynchronous transfer of `offsets[nrow]` from the device offsets array to host.
`nnz_sync`()	Synchronize the stored block upper bound from the device `offsets` array to the host.
`notify_nnz_changed`([nnz, nnz_capacity])	Notify the matrix that sparse storage metadata changed outside `warp.sparse`.
`status_message`()	Return a human-readable message for `status_sync()`.
`status_sync`()	Return the asynchronous sparse status code, synchronizing if needed.
`transpose`()	Return a transposed copy of this matrix.
`uncompress_rows`([out])	Compute the row index for each stored block slot from the compressed row offsets.

Attributes

`block_shape`	Shape of the individual blocks.
`block_size`	Size of the individual blocks, i.e. number of rows per block times number of columns per block.
`device`	Device on which matrix arrays are allocated.
`dtype`	Data type for individual block values.
`requires_grad`	Read-only property indicating whether the matrix participates in adjoint computations.
`scalar_type`	Scalar type for individual block coefficients.
`scalar_values`	Access the `values` array as a 3d scalar array.
`shape`	Shape of the matrix, i.e. number of rows/columns of blocks times number of rows/columns per block.

property scalar_type: Scalar[source]#: Scalar type for individual block coefficients. For CSR matrices, this is the same as the block type.

property block_shape: tuple[int, int][source]#: Shape of the individual blocks.

property block_size: int[source]#: Size of the individual blocks, i.e. number of rows per block times number of columns per block.

property shape: tuple[int, int][source]#: Shape of the matrix, i.e. number of rows/columns of blocks times number of rows/columns per block.

property dtype: type[source]#: Data type for individual block values.

property device: Device[source]#: Device on which matrix arrays are allocated.

property requires_grad: bool[source]#: Read-only property indicating whether the matrix participates in adjoint computations.

property scalar_values: array[source]#: Access the values array as a 3d scalar array.

uncompress_rows(out=None)[source]#

Compute the row index for each stored block slot from the compressed row offsets.

The method writes nnz entries. When out is omitted, it allocates an array of that length. Entries outside active row ranges are -1. When out is provided, entries beyond nnz are unchanged. For a compact matrix, use nnz_sync() before treating the result as COO row data:

nnz = matrix.nnz_sync()
rows = matrix.uncompress_rows()[:nnz]

Parameters:: out (array) – Optional one-dimensional integer output array with at least nnz elements on the matrix’s device.
Returns:: The output array.
Return type:: array

nnz_sync()[source]#

Synchronize the stored block upper bound from the device offsets array to the host.

Ensures that any ongoing transfer of offsets[nrow] from the device offsets array to the host has completed, or, if none has been scheduled yet, starts a new transfer and waits for it to complete.

Then updates the host-side nnz upper bound to match offsets[nrow], and returns it. For compact matrices, this is the active stored block count. For padded matrices, this is the total row-capacity storage size, not necessarily the active block count.