warp-lang 1.2.2__py3-none-manylinux2014_aarch64.whl → 1.3.0__py3-none-manylinux2014_aarch64.whl

warp/sparse.py

@@ -1,3 +1,4 @@
+import ctypes
 from typing import Any, Generic, Optional, Tuple, TypeVar, Union
 
 import warp as wp
@@ -31,7 +32,7 @@ class BsrMatrix(Generic[_BlockType]):
     Attributes:
         nrow (int): Number of rows of blocks
         ncol (int): Number of columns of blocks
-        nnz (int):  Number of non-zero blocks: must be equal to ``offsets[nrow-1]``, cached on host for convenience
+        nnz (int):  Upper bound for the number of non-zero blocks, used for dimensioning launches; the exact number is at ``offsets[nrow-1]``. See also :meth:`nnz_sync`.
         offsets (Array[int]): Array of size at least ``1 + nrows`` such that the start and end indices of the blocks of row ``r`` are ``offsets[r]`` and ``offsets[r+1]``, respectively.
         columns (Array[int]): Array of size at least equal to ``nnz`` containing block column indices
         values (Array[BlockType]): Array of size at least equal to ``nnz`` containing block values
@@ -68,6 +69,111 @@ class BsrMatrix(Generic[_BlockType]):
         """Device on which offsets, columns and values are allocated -- assumed to be the same for all three arrays"""
         return self.values.device
 
+    def nnz_sync(self):
+        """Ensures that any ongoing transfer of the exact nnz number from the device offsets array to the host has completed,
+        and updates the nnz upper bound.
+
+        See also :meth:`copy_nnz_async`
+        """
+
+        if self._is_nnz_transfer_setup():
+            if self.device.is_cuda:
+                wp.synchronize_event(self._nnz_event)
+            self.nnz = int(self._nnz_buf.numpy()[0])
+        return self.nnz
+
+    def copy_nnz_async(self, known_nnz: int = None):
+        """
+        Starts the asynchronous transfer of the exact nnz from the device offsets array to host, and records an event for completion.
+        Needs to be called whenever the offsets array has been modified from outside ``warp.sparse``.
+
+        See also :meth:`nnz_sync`
+        """
+        if known_nnz is not None:
+            self.nnz = int(known_nnz)
+        else:
+            self._setup_nnz_transfer()
+
+        # If a transfer is already ongoing, or if the actual nnz is unknown, schedule a new transfer
+        if self._is_nnz_transfer_setup():
+            stream = wp.get_stream(self.device) if self.device.is_cuda else None
+            wp.copy(src=self.offsets, dest=self._nnz_buf, src_offset=self.nrow, count=1, stream=stream)
+            if self.device.is_cuda:
+                stream.record_event(self._nnz_event)
+
+    def _setup_nnz_transfer(self):
+        if self._is_nnz_transfer_setup():
+            return
+
+        BsrMatrix.__setattr__(
+            self, "_nnz_buf", wp.zeros(dtype=int, shape=(1,), device="cpu", pinned=self.device.is_cuda)
+        )
+        if self.device.is_cuda:
+            BsrMatrix.__setattr__(self, "_nnz_event", wp.Event(self.device))
+
+    def _is_nnz_transfer_setup(self):
+        return hasattr(self, "_nnz_buf")
+
+    def _nnz_transfer_buf_and_event(self):
+        self._setup_nnz_transfer()
+
+        if not self.device.is_cuda:
+            return self._nnz_buf, ctypes.c_void_p(None)
+        return self._nnz_buf, self._nnz_event.cuda_event
+
+    # Overloaded math operators
+    def __add__(self, y):
+        return bsr_axpy(y, bsr_copy(self))
+
+    def __iadd__(self, y):
+        return bsr_axpy(y, self)
+
+    def __radd__(self, x):
+        return bsr_axpy(x, bsr_copy(self))
+
+    def __sub__(self, y):
+        return bsr_axpy(y, bsr_copy(self), alpha=-1.0)
+
+    def __rsub__(self, x):
+        return bsr_axpy(x, bsr_copy(self), beta=-1.0)
+
+    def __isub__(self, y):
+        return bsr_axpy(y, self, alpha=-1.0)
+
+    def __mul__(self, y):
+        return _BsrScalingExpression(self, y)
+
+    def __rmul__(self, x):
+        return _BsrScalingExpression(self, x)
+
+    def __imul__(self, y):
+        return bsr_scale(self, y)
+
+    def __matmul__(self, y):
+        if isinstance(y, wp.array):
+            return bsr_mv(self, y)
+
+        return bsr_mm(self, y)
+
+    def __rmatmul__(self, x):
+        if isinstance(x, wp.array):
+            return bsr_mv(self, x, transpose=True)
+
+        return bsr_mm(x, self)
+
+    def __imatmul__(self, y):
+        return bsr_mm(self, y, self)
+
+    def __truediv__(self, y):
+        return _BsrScalingExpression(self, 1.0 / y)
+
+    def __neg__(self):
+        return _BsrScalingExpression(self, -1.0)
+
+    def transpose(self):
+        """Returns a transposed copy of this matrix"""
+        return bsr_transposed(self)
+
 
 def bsr_matrix_t(dtype: BlockType):
     dtype = wp.types.type_to_warp(dtype)
@@ -83,7 +189,7 @@ def bsr_matrix_t(dtype: BlockType):
         ncol: int
         """Number of columns of blocks"""
         nnz: int
-        """Number of non-zero blocks: equal to offsets[-1], cached on host for convenience"""
+        """Upper bound for the number of non-zeros"""
         offsets: wp.array(dtype=int)
         """Array of size at least 1 + nrows"""
         columns: wp.array(dtype=int)
@@ -130,7 +236,7 @@ def bsr_zeros(
 
     bsr.nrow = int(rows_of_blocks)
     bsr.ncol = int(cols_of_blocks)
-    bsr.nnz = 0
+    bsr.nnz = int(0)
     bsr.columns = wp.empty(shape=(0,), dtype=int, device=device)
     bsr.values = wp.empty(shape=(0,), dtype=block_type, device=device)
     bsr.offsets = wp.zeros(shape=(bsr.nrow + 1,), dtype=int, device=device)
@@ -143,6 +249,9 @@ def _bsr_ensure_fits(bsr: BsrMatrix, nrow: int = None, nnz: int = None):
         nrow = bsr.nrow
     if nnz is None:
         nnz = bsr.nnz
+    else:
+        # update nnz upper bound
+        bsr.nnz = int(nnz)
 
     if bsr.offsets.size < nrow + 1:
         bsr.offsets = wp.empty(shape=(nrow + 1,), dtype=int, device=bsr.offsets.device)
@@ -170,9 +279,10 @@ def bsr_set_zero(
         bsr.nrow = int(rows_of_blocks)
     if cols_of_blocks is not None:
         bsr.ncol = int(cols_of_blocks)
-    bsr.nnz = 0
-    _bsr_ensure_fits(bsr)
+
+    _bsr_ensure_fits(bsr, nnz=0)
     bsr.offsets.zero_()
+    bsr.copy_nnz_async(known_nnz=0)
 
 
 def bsr_set_from_triplets(
@@ -180,11 +290,12 @@ def bsr_set_from_triplets(
     rows: "Array[int]",
     columns: "Array[int]",
     values: "Array[Union[Scalar, BlockType[Rows, Cols, Scalar]]]",
+    prune_numerical_zeros: bool = True,
 ):
     """
     Fills a BSR matrix with values defined by coordinate-oriented (COO) triplets, discarding existing blocks.
 
-    The first dimension of the three input arrays must match, and determines the number of non-zeros in the constructed matrix.
+    The first dimension of the three input arrays must match and indicates the number of COO triplets.
 
     Args:
         dest: Sparse matrix to populate
@@ -192,6 +303,7 @@ def bsr_set_from_triplets(
         columns: Columns index for each non-zero
         values: Block values for each non-zero. Must be either a one-dimensional array with data type identical
           to the `dest` matrix's block type, or a 3d array with data type equal to the `dest` matrix's scalar type.
+        prune_numerical_zeros: If True, will ignore the zero-valued blocks
     """
 
     if values.device != columns.device or values.device != rows.device or values.device != dest.values.device:
@@ -244,62 +356,477 @@ def bsr_set_from_triplets(
     if not native_func:
         raise NotImplementedError(f"bsr_from_triplets not implemented for scalar type {scalar_type}")
 
-    dest.nnz = native_func(
-        dest.block_shape[0],
-        dest.block_shape[1],
-        dest.nrow,
-        nnz,
-        rows.ptr,
-        columns.ptr,
-        values.ptr,
-        dest.offsets.ptr,
-        dest.columns.ptr,
-        dest.values.ptr,
+    nnz_buf, nnz_event = dest._nnz_transfer_buf_and_event()
+
+    with wp.ScopedDevice(device):
+        native_func(
+            dest.block_shape[0],
+            dest.block_shape[1],
+            dest.nrow,
+            nnz,
+            ctypes.cast(rows.ptr, ctypes.POINTER(ctypes.c_int32)),
+            ctypes.cast(columns.ptr, ctypes.POINTER(ctypes.c_int32)),
+            ctypes.cast(values.ptr, ctypes.c_void_p),
+            prune_numerical_zeros,
+            ctypes.cast(dest.offsets.ptr, ctypes.POINTER(ctypes.c_int32)),
+            ctypes.cast(dest.columns.ptr, ctypes.POINTER(ctypes.c_int32)),
+            ctypes.cast(dest.values.ptr, ctypes.c_void_p),
+            ctypes.cast(nnz_buf.ptr, ctypes.POINTER(ctypes.c_int32)),
+            nnz_event,
+        )
+
+
+class _BsrExpression(Generic[_BlockType]):
+    pass
+
+
+class _BsrScalingExpression(_BsrExpression):
+    def __init__(self, mat, scale):
+        self.mat = mat
+        self.scale = scale
+
+    def eval(self):
+        return bsr_copy(self)
+
+    @property
+    def nrow(self) -> int:
+        return self.mat.nrow
+
+    @property
+    def ncol(self) -> int:
+        return self.mat.ncol
+
+    @property
+    def nnz(self) -> int:
+        return self.mat.nnz
+
+    @property
+    def offsets(self) -> wp.array:
+        return self.mat.offsets
+
+    @property
+    def columns(self) -> wp.array:
+        return self.mat.columns
+
+    @property
+    def scalar_type(self) -> Scalar:
+        return self.mat.scalar_type
+
+    @property
+    def block_shape(self) -> Tuple[int, int]:
+        return self.mat.block_shape
+
+    @property
+    def block_size(self) -> int:
+        return self.mat.block_size
+
+    @property
+    def shape(self) -> Tuple[int, int]:
+        return self.mat.shape
+
+    @property
+    def dtype(self) -> type:
+        return self.mat.dtype
+
+    @property
+    def device(self) -> wp.context.Device:
+        return self.mat.device
+
+    # Overloaded math operators
+    def __add__(self, y):
+        return bsr_axpy(y, bsr_copy(self.mat), alpha=self.scale)
+
+    def __radd__(self, x):
+        return bsr_axpy(x, bsr_copy(self.mat), beta=self.scale)
+
+    def __sub__(self, y):
+        return bsr_axpy(y, bsr_copy(self.mat), alpha=-self.scale)
+
+    def __rsub__(self, x):
+        return bsr_axpy(x, bsr_copy(self.mat), beta=-self.scale)
+
+    def __mul__(self, y):
+        return _BsrScalingExpression(self.mat, y * self.scale)
+
+    def __rmul__(self, x):
+        return _BsrScalingExpression(self.mat, x * self.scale)
+
+    def __matmul__(self, y):
+        if isinstance(y, wp.array):
+            return bsr_mv(self.mat, y, alpha=self.scale)
+
+        return bsr_mm(self.mat, y, alpha=self.scale)
+
+    def __rmatmul__(self, x):
+        if isinstance(x, wp.array):
+            return bsr_mv(self.mat, x, alpha=self.scale, transpose=True)
+
+        return bsr_mm(x, self.mat, alpha=self.scale)
+
+    def __truediv__(self, y):
+        return _BsrScalingExpression(self.mat, self.scale / y)
+
+    def __neg__(self):
+        return _BsrScalingExpression(self.mat, -self.scale)
+
+    def transpose(self):
+        """Returns a transposed copy of this matrix"""
+        return _BsrScalingExpression(self.mat.transpose(), self.scale)
+
+
+BsrMatrixOrExpression = Union[BsrMatrix[_BlockType], _BsrExpression[_BlockType]]
+
+
+def _extract_matrix_and_scale(bsr: BsrMatrixOrExpression):
+    if isinstance(bsr, BsrMatrix):
+        return bsr, 1.0
+    if isinstance(bsr, _BsrScalingExpression):
+        return bsr.mat, bsr.scale
+
+    raise ValueError("Argument cannot be interpreted as a BsrMatrix")
+
+
+@wp.kernel
+def _bsr_assign_split_offsets(
+    row_factor: int,
+    col_factor: int,
+    src_offsets: wp.array(dtype=int),
+    dest_offsets: wp.array(dtype=int),
+):
+    row = wp.tid()
+
+    base_offset = src_offsets[row] * row_factor * col_factor
+    row_count = src_offsets[1 + row] - src_offsets[row]
+
+    for k in range(row_factor):
+        dest_offsets[1 + k + row_factor * row] = base_offset + row_count * col_factor * (k + 1)
+
+    if row == 0:
+        dest_offsets[0] = 0
+
+
+@wp.kernel
+def _bsr_assign_split_blocks(
+    structure_only: wp.bool,
+    scale: Any,
+    row_factor: int,
+    col_factor: int,
+    dest_row_count: int,
+    src_offsets: wp.array(dtype=int),
+    src_columns: wp.array(dtype=int),
+    src_values: wp.array3d(dtype=Any),
+    dest_offsets: wp.array(dtype=int),
+    dest_columns: wp.array(dtype=int),
+    dest_values: wp.array3d(dtype=Any),
+):
+    dest_block = wp.tid()
+
+    if dest_block >= dest_offsets[dest_row_count]:
+        return
+
+    dest_row = wp.lower_bound(dest_offsets, dest_block + 1) - 1
+    src_row = dest_row // row_factor
+
+    dest_col_in_row = dest_block - dest_offsets[dest_row]
+    src_col_in_row = dest_col_in_row // col_factor
+
+    src_block = src_offsets[src_row] + src_col_in_row
+
+    dest_rows_per_block = dest_values.shape[1]
+    dest_cols_per_block = dest_values.shape[2]
+
+    split_row = dest_row - row_factor * src_row
+    split_col = dest_col_in_row - col_factor * src_col_in_row
+
+    dest_columns[dest_block] = src_columns[src_block] * col_factor + split_col
+
+    if not structure_only:
+        src_base_i = split_row * dest_rows_per_block
+        src_base_j = split_col * dest_cols_per_block
+        for i in range(dest_rows_per_block):
+            for j in range(dest_cols_per_block):
+                dest_values[dest_block, i, j] = dest_values.dtype(
+                    scale * src_values[src_block, i + src_base_i, j + src_base_j]
+                )
+
+
+@wp.kernel
+def _bsr_assign_merge_row_col(
+    row_factor: int,
+    col_factor: int,
+    src_row_count: int,
+    src_offsets: wp.array(dtype=int),
+    src_columns: wp.array(dtype=int),
+    dest_rows: wp.array(dtype=int),
+    dest_cols: wp.array(dtype=int),
+):
+    block = wp.tid()
+
+    if block >= src_offsets[src_row_count]:
+        dest_rows[block] = -1  # invalid
+        dest_cols[block] = -1
+    else:
+        row = wp.lower_bound(src_offsets, block + 1) - 1
+        dest_rows[block] = row // row_factor
+        dest_cols[block] = src_columns[block] // col_factor
+
+
+@wp.kernel
+def _bsr_assign_merge_blocks(
+    scale: Any,
+    row_factor: int,
+    col_factor: int,
+    src_row_count: int,
+    src_offsets: wp.array(dtype=int),
+    src_columns: wp.array(dtype=int),
+    src_values: wp.array3d(dtype=Any),
+    dest_offsets: wp.array(dtype=int),
+    dest_columns: wp.array(dtype=int),
+    dest_values: wp.array3d(dtype=Any),
+):
+    src_block = wp.tid()
+
+    if src_block >= src_offsets[src_row_count]:
+        return
+
+    src_row = wp.lower_bound(src_offsets, src_block + 1) - 1
+    src_col = src_columns[src_block]
+
+    dest_row = src_row // row_factor
+    dest_col = src_col // col_factor
+
+    dest_block = wp.lower_bound(dest_columns, dest_offsets[dest_row], dest_offsets[dest_row + 1], dest_col)
+
+    src_rows_per_block = src_values.shape[1]
+    src_cols_per_block = src_values.shape[2]
+
+    split_row = src_row - row_factor * dest_row
+    split_col = src_col - col_factor * dest_col
+
+    dest_base_i = split_row * src_rows_per_block
+    dest_base_j = split_col * src_cols_per_block
+
+    for i in range(src_rows_per_block):
+        for j in range(src_cols_per_block):
+            dest_values[dest_block, i + dest_base_i, j + dest_base_j] = dest_values.dtype(
+                scale * src_values[src_block, i, j]
+            )
+
+
+def _bsr_values_as_3d_array(A: BsrMatrix) -> wp.array:
+    if A.block_shape == (1, 1):
+        return A.values.reshape((A.values.shape[0], 1, 1))
+
+    return wp.array(
+        data=None,
+        ptr=A.values.ptr,
+        capacity=A.values.capacity,
+        device=A.device,
+        dtype=A.scalar_type,
+        shape=(A.values.shape[0], A.block_shape[0], A.block_shape[1]),
     )
 
 
 def bsr_assign(
     dest: BsrMatrix[BlockType[Rows, Cols, Scalar]],
-    src: BsrMatrix[BlockType[Rows, Cols, Any]],
+    src: BsrMatrixOrExpression[BlockType[Any, Any, Any]],
+    structure_only: bool = False,
 ):
-    """Copies the content of the `src` matrix to `dest`, casting the block values if the two matrices use distinct scalar types."""
+    """Copies the content of the `src` BSR matrix to `dest`.
+
+    Args:
+      src: Matrix to be copied
+      dest: Destination matrix. May have a different block shape of scalar type than `src`, in which case the required casting will be performed.
+      structure_only: If ``True``, only the non-zeros indices are copied, and uninitialized value storage is allocated
+        to accommodate at least `src.nnz` blocks. If `structure_only` is ``False``, values are also copied with implicit
+        casting if the two matrices use distinct scalar types.
+    """
+
+    src, src_scale = _extract_matrix_and_scale(src)
 
     if dest.values.device != src.values.device:
         raise ValueError("Source and destination matrices must reside on the same device")
 
-    if dest.block_shape != src.block_shape:
-        raise ValueError("Source and destination matrices must have the same block shape")
+    if dest.block_shape == src.block_shape:
+        dest.nrow = src.nrow
+        dest.ncol = src.ncol
+
+        nnz_alloc = src.nnz
+        _bsr_ensure_fits(dest, nnz=nnz_alloc)
+
+        wp.copy(dest=dest.offsets, src=src.offsets, count=src.nrow + 1)
+        dest.copy_nnz_async()
+
+        if nnz_alloc > 0:
+            wp.copy(dest=dest.columns, src=src.columns, count=nnz_alloc)
 
-    dest.nrow = src.nrow
-    dest.ncol = src.ncol
-    dest.nnz = src.nnz
+            if not structure_only:
+                warp.utils.array_cast(out_array=dest.values, in_array=src.values, count=nnz_alloc)
+                bsr_scale(dest, src_scale)
 
-    _bsr_ensure_fits(dest)
+    elif src.block_shape[0] >= dest.block_shape[0] and src.block_shape[1] >= dest.block_shape[1]:
+        # Split blocks
+
+        row_factor = src.block_shape[0] // dest.block_shape[0]
+        col_factor = src.block_shape[1] // dest.block_shape[1]
+
+        if (
+            row_factor * dest.block_shape[0] != src.block_shape[0]
+            or col_factor * dest.block_shape[1] != src.block_shape[1]
+        ):
+            raise ValueError(
+                f"Dest block shape {dest.block_shape} is not an exact divider of src block shape {src.block_shape}"
+            )
 
-    wp.copy(dest=dest.offsets, src=src.offsets, count=src.nrow + 1)
-    if src.nnz > 0:
-        wp.copy(dest=dest.columns, src=src.columns, count=src.nnz)
-        warp.utils.array_cast(out_array=dest.values, in_array=src.values, count=src.nnz)
+        dest.nrow = src.nrow * row_factor
+        dest.ncol = src.ncol * col_factor
 
+        nnz_alloc = src.nnz * row_factor * col_factor
+        _bsr_ensure_fits(dest, nnz=nnz_alloc)
 
-def bsr_copy(A: BsrMatrix, scalar_type: Optional[Scalar] = None):
+        wp.launch(
+            _bsr_assign_split_offsets,
+            dim=src.nrow,
+            device=dest.device,
+            inputs=[row_factor, col_factor, src.offsets, dest.offsets],
+        )
+        wp.launch(
+            _bsr_assign_split_blocks,
+            dim=dest.nnz,
+            device=dest.device,
+            inputs=[
+                wp.bool(structure_only),
+                src.scalar_type(src_scale),
+                row_factor,
+                col_factor,
+                dest.nrow,
+                src.offsets,
+                src.columns,
+                _bsr_values_as_3d_array(src),
+                dest.offsets,
+                dest.columns,
+                _bsr_values_as_3d_array(dest),
+            ],
+        )
+
+    elif src.block_shape[0] <= dest.block_shape[0] and src.block_shape[1] <= dest.block_shape[1]:
+        # Merge blocks
+
+        row_factor = dest.block_shape[0] // src.block_shape[0]
+        col_factor = dest.block_shape[1] // src.block_shape[1]
+
+        if (
+            row_factor * src.block_shape[0] != dest.block_shape[0]
+            or col_factor * src.block_shape[1] != dest.block_shape[1]
+        ):
+            raise ValueError(
+                f"Dest block shape {dest.block_shape} is not an exact multiple of src block shape {src.block_shape}"
+            )
+
+        if src.nrow % row_factor != 0 or src.ncol % col_factor != 0:
+            raise ValueError(
+                "The total rows and columns of the src matrix cannot be evenly divided using the requested block shape"
+            )
+
+        dest.nrow = src.nrow // row_factor
+        dest.ncol = src.ncol // col_factor
+
+        nnz_alloc = src.nnz  # Conservative, in case all nnz in src belong to distinct merged blocks
+        _bsr_ensure_fits(dest, nnz=nnz_alloc)
+
+        # Compute destination rows and columns
+        dest_rows = wp.empty_like(src.columns)
+        dest_cols = wp.empty_like(src.columns)
+        wp.launch(
+            _bsr_assign_merge_row_col,
+            dim=src.nnz,
+            device=dest.device,
+            inputs=[row_factor, col_factor, src.nrow, src.offsets, src.columns, dest_rows, dest_cols],
+        )
+
+        # Compute destination offsets from triplets
+        from warp.context import runtime
+
+        if dest.device.is_cpu:
+            native_func = runtime.core.bsr_matrix_from_triplets_float_host
+        else:
+            native_func = runtime.core.bsr_matrix_from_triplets_float_device
+
+        nnz_buf, nnz_event = dest._nnz_transfer_buf_and_event()
+        with wp.ScopedDevice(dest.device):
+            native_func(
+                dest.block_shape[0],
+                dest.block_shape[1],
+                dest.nrow,
+                dest.nnz,
+                ctypes.cast(dest_rows.ptr, ctypes.POINTER(ctypes.c_int32)),
+                ctypes.cast(dest_cols.ptr, ctypes.POINTER(ctypes.c_int32)),
+                0,
+                False,
+                ctypes.cast(dest.offsets.ptr, ctypes.POINTER(ctypes.c_int32)),
+                ctypes.cast(dest.columns.ptr, ctypes.POINTER(ctypes.c_int32)),
+                0,
+                ctypes.cast(nnz_buf.ptr, ctypes.POINTER(ctypes.c_int32)),
+                nnz_event,
+            )
+
+        # merge block values
+        if not structure_only:
+            dest.values.zero_()
+            wp.launch(
+                _bsr_assign_merge_blocks,
+                dim=src.nnz,
+                device=dest.device,
+                inputs=[
+                    src.scalar_type(src_scale),
+                    row_factor,
+                    col_factor,
+                    src.nrow,
+                    src.offsets,
+                    src.columns,
+                    _bsr_values_as_3d_array(src),
+                    dest.offsets,
+                    dest.columns,
+                    _bsr_values_as_3d_array(dest),
+                ],
+            )
+
+    else:
+        raise ValueError("Incompatible dest and src block shapes")
+
+
+def bsr_copy(
+    A: BsrMatrixOrExpression,
+    scalar_type: Optional[Scalar] = None,
+    block_shape: Optional[Tuple[int, int]] = None,
+    structure_only: bool = False,
+):
     """Returns a copy of matrix ``A``, possibly changing its scalar type.
 
     Args:
+       A: Matrix to be copied
        scalar_type: If provided, the returned matrix will use this scalar type instead of the one from `A`.
+       block_shape: If provided, the returned matrix will use blocks of this shape instead of the one from `A`.
+         Both dimensions of `block_shape` must be either a multiple or an exact divider of the ones from `A`.
+       structure_only: If ``True``, only the non-zeros indices are copied, and uninitialized value storage is allocated
+         to accommodate at least `src.nnz` blocks. If `structure_only` is ``False``, values are also copied with implicit
+         casting if the two matrices use distinct scalar types.
     """
     if scalar_type is None:
-        block_type = A.values.dtype
-    elif A.block_shape == (1, 1):
+        scalar_type = A.scalar_type
+    if block_shape is None:
+        block_shape = A.block_shape
+
+    if block_shape == (1, 1):
         block_type = scalar_type
     else:
-        block_type = wp.types.matrix(shape=A.block_shape, dtype=scalar_type)
+        block_type = wp.types.matrix(shape=block_shape, dtype=scalar_type)
 
     copy = bsr_zeros(
         rows_of_blocks=A.nrow,
         cols_of_blocks=A.ncol,
         block_type=block_type,
-        device=A.values.device,
+        device=A.device,
     )
     bsr_assign(dest=copy, src=A)
     return copy
@@ -307,10 +834,12 @@ def bsr_copy(A: BsrMatrix, scalar_type: Optional[Scalar] = None):
 
 def bsr_set_transpose(
     dest: BsrMatrix[BlockType[Cols, Rows, Scalar]],
-    src: BsrMatrix[BlockType[Rows, Cols, Scalar]],
+    src: BsrMatrixOrExpression[BlockType[Rows, Cols, Scalar]],
 ):
     """Assigns the transposed matrix `src` to matrix `dest`"""
 
+    src, src_scale = _extract_matrix_and_scale(src)
+
     if dest.values.device != src.values.device:
         raise ValueError("All arguments must reside on the same device")
 
@@ -322,15 +851,16 @@ def bsr_set_transpose(
     if dest.block_shape != transpose_block_shape:
         raise ValueError(f"Destination block shape must be {transpose_block_shape}")
 
+    nnz = src.nnz
     dest.nrow = src.ncol
     dest.ncol = src.nrow
-    dest.nnz = src.nnz
 
-    if src.nnz == 0:
+    if nnz == 0:
+        bsr_set_zero(dest)
         return
 
     # Increase dest array sizes if needed
-    _bsr_ensure_fits(dest)
+    _bsr_ensure_fits(dest, nnz=nnz)
 
     from warp.context import runtime
 
@@ -348,22 +878,26 @@ def bsr_set_transpose(
     if not native_func:
         raise NotImplementedError(f"bsr_set_transpose not implemented for scalar type {dest.scalar_type}")
 
-    native_func(
-        src.block_shape[0],
-        src.block_shape[1],
-        src.nrow,
-        src.ncol,
-        src.nnz,
-        src.offsets.ptr,
-        src.columns.ptr,
-        src.values.ptr,
-        dest.offsets.ptr,
-        dest.columns.ptr,
-        dest.values.ptr,
-    )
+    with wp.ScopedDevice(dest.device):
+        native_func(
+            src.block_shape[0],
+            src.block_shape[1],
+            src.nrow,
+            src.ncol,
+            nnz,
+            ctypes.cast(src.offsets.ptr, ctypes.POINTER(ctypes.c_int32)),
+            ctypes.cast(src.columns.ptr, ctypes.POINTER(ctypes.c_int32)),
+            ctypes.cast(src.values.ptr, ctypes.c_void_p),
+            ctypes.cast(dest.offsets.ptr, ctypes.POINTER(ctypes.c_int32)),
+            ctypes.cast(dest.columns.ptr, ctypes.POINTER(ctypes.c_int32)),
+            ctypes.cast(dest.values.ptr, ctypes.c_void_p),
+        )
 
+    dest.copy_nnz_async()
+    bsr_scale(dest, src_scale)
 
-def bsr_transposed(A: BsrMatrix):
+
+def bsr_transposed(A: BsrMatrixOrExpression):
     """Returns a copy of the transposed matrix `A`"""
 
     if A.block_shape == (1, 1):
@@ -375,7 +909,7 @@ def bsr_transposed(A: BsrMatrix):
         rows_of_blocks=A.ncol,
         cols_of_blocks=A.nrow,
         block_type=block_type,
-        device=A.values.device,
+        device=A.device,
     )
     bsr_set_transpose(dest=transposed, src=A)
     return transposed
@@ -383,6 +917,7 @@ def bsr_transposed(A: BsrMatrix):
 
 @wp.kernel
 def _bsr_get_diag_kernel(
+    scale: Any,
     A_offsets: wp.array(dtype=int),
     A_columns: wp.array(dtype=int),
     A_values: wp.array(dtype=Any),
@@ -395,10 +930,10 @@ def _bsr_get_diag_kernel(
     diag = wp.lower_bound(A_columns, beg, end, row)
     if diag < end:
         if A_columns[diag] == row:
-            out[row] = A_values[diag]
+            out[row] = scale * A_values[diag]
 
 
-def bsr_get_diag(A: BsrMatrix[_BlockType], out: "Optional[Array[BlockType]]" = None) -> "Array[BlockType]":
+def bsr_get_diag(A: BsrMatrixOrExpression[BlockType], out: "Optional[Array[BlockType]]" = None) -> "Array[BlockType]":
     """Returns the array of blocks that constitute the diagonal of a sparse matrix.
 
     Args:
@@ -406,6 +941,8 @@ def bsr_get_diag(A: BsrMatrix[_BlockType], out: "Optional[Array[BlockType]]" = N
         out: if provided, the array into which to store the diagonal blocks
     """
 
+    A, scale = _extract_matrix_and_scale(A)
+
     dim = min(A.nrow, A.ncol)
 
     if out is None:
@@ -422,7 +959,7 @@ def bsr_get_diag(A: BsrMatrix[_BlockType], out: "Optional[Array[BlockType]]" = N
         kernel=_bsr_get_diag_kernel,
         dim=dim,
         device=A.values.device,
-        inputs=[A.offsets, A.columns, A.values, out],
+        inputs=[A.scalar_type(scale), A.offsets, A.columns, A.values, out],
     )
 
     return out
@@ -495,13 +1032,13 @@ def bsr_set_diag(
         A.nrow = rows_of_blocks
         A.ncol = cols_of_blocks
 
-    A.nnz = min(A.nrow, A.ncol)
-    _bsr_ensure_fits(A)
+    nnz = min(A.nrow, A.ncol)
+    _bsr_ensure_fits(A, nnz=nnz)
 
     if warp.types.is_array(diag):
         wp.launch(
             kernel=_bsr_set_diag_kernel,
-            dim=A.nnz,
+            dim=nnz,
             device=A.values.device,
             inputs=[diag, A.offsets, A.columns, A.values],
         )
@@ -511,11 +1048,13 @@ def bsr_set_diag(
             diag = A.values.dtype(diag)
         wp.launch(
             kernel=_bsr_set_diag_constant_kernel,
-            dim=A.nnz,
+            dim=nnz,
             device=A.values.device,
             inputs=[diag, A.offsets, A.columns, A.values],
         )
 
+    A.copy_nnz_async(known_nnz=nnz)
+
 
 def bsr_diag(
     diag: "Union[BlockType, Array[BlockType]]",
@@ -619,11 +1158,14 @@ def _bsr_scale_kernel(
     values[wp.tid()] = alpha * values[wp.tid()]
 
 
-def bsr_scale(x: BsrMatrix, alpha: Scalar) -> BsrMatrix:
+def bsr_scale(x: BsrMatrixOrExpression, alpha: Scalar) -> BsrMatrix:
     """
     Performs the operation ``x := alpha * x`` on BSR matrix `x` and returns `x`
     """
 
+    x, scale = _extract_matrix_and_scale(x)
+    alpha *= scale
+
     if alpha != 1.0 and x.nnz > 0:
         if alpha == 0.0:
             bsr_set_zero(x)
@@ -642,11 +1184,14 @@ def bsr_scale(x: BsrMatrix, alpha: Scalar) -> BsrMatrix:
 
 
 @wp.kernel
-def _bsr_get_block_row(dest_offset: int, bsr_offsets: wp.array(dtype=int), rows: wp.array(dtype=int)):
+def _bsr_get_block_row(dest_offset: int, row_count: int, bsr_offsets: wp.array(dtype=int), rows: wp.array(dtype=int)):
     i = wp.tid()
 
-    row = wp.lower_bound(bsr_offsets, i + 1) - 1
-    rows[dest_offset + i] = row
+    if i >= bsr_offsets[row_count]:
+        rows[dest_offset + i] = -1  # invalid
+    else:
+        row = wp.lower_bound(bsr_offsets, i + 1) - 1
+        rows[dest_offset + i] = row
 
 
 @wp.kernel
@@ -662,6 +1207,10 @@ def _bsr_axpy_add_block(
 ):
     i = wp.tid()
     row = rows[i + src_offset]
+
+    if row < 0:
+        return
+
     col = cols[i + src_offset]
     beg = dst_offsets[row]
     end = dst_offsets[row + 1]
@@ -694,11 +1243,11 @@ class bsr_axpy_work_arrays:
             self._sum_cols = wp.empty(shape=(sum_nnz), dtype=int, device=self.device)
 
         if self._old_y_values is None or self._old_y_values.size < y.nnz:
-            self._old_y_values = wp.empty(shape=(y.nnz), dtype=y.values.dtype, device=self.device)
+            self._old_y_values = wp.empty(shape=(y.nnz,), dtype=y.values.dtype, device=self.device)
 
 
 def bsr_axpy(
-    x: BsrMatrix[BlockType[Rows, Cols, Scalar]],
+    x: BsrMatrixOrExpression,
     y: Optional[BsrMatrix[BlockType[Rows, Cols, Scalar]]] = None,
     alpha: Scalar = 1.0,
     beta: Scalar = 1.0,
@@ -717,17 +1266,23 @@ def bsr_axpy(
         work_arrays: In most cases this function will require the use of temporary storage; this storage can be reused across calls by passing an instance of :class:`bsr_axpy_work_arrays` in `work_arrays`.
     """
 
+    x, x_scale = _extract_matrix_and_scale(x)
+    alpha *= x_scale
+
     if y is None:
         # If not output matrix is provided, allocate it for convenience
         y = bsr_zeros(x.nrow, x.ncol, block_type=x.values.dtype, device=x.values.device)
         beta = 0.0
 
+    x_nnz = x.nnz
+    y_nnz = y.nnz
+
     # Handle easy cases first
-    if beta == 0.0 or y.nnz == 0:
+    if beta == 0.0 or y_nnz == 0:
         bsr_assign(src=x, dest=y)
         return bsr_scale(y, alpha=alpha)
 
-    if alpha == 0.0 or x.nnz == 0:
+    if alpha == 0.0 or x_nnz == 0:
         return bsr_scale(y, alpha=beta)
 
     if not isinstance(alpha, y.scalar_type):
@@ -753,28 +1308,28 @@ def bsr_axpy(
     if work_arrays is None:
         work_arrays = bsr_axpy_work_arrays()
 
-    sum_nnz = x.nnz + y.nnz
+    sum_nnz = x_nnz + y_nnz
     device = y.values.device
     work_arrays._allocate(device, y, sum_nnz)
 
-    wp.copy(work_arrays._sum_cols, y.columns, 0, 0, y.nnz)
+    wp.copy(work_arrays._sum_cols, y.columns, 0, 0, y_nnz)
     wp.launch(
         kernel=_bsr_get_block_row,
         device=device,
-        dim=y.nnz,
-        inputs=[0, y.offsets, work_arrays._sum_rows],
+        dim=y_nnz,
+        inputs=[0, y.nrow, y.offsets, work_arrays._sum_rows],
     )
 
-    wp.copy(work_arrays._sum_cols, x.columns, y.nnz, 0, x.nnz)
+    wp.copy(work_arrays._sum_cols, x.columns, y_nnz, 0, x_nnz)
     wp.launch(
         kernel=_bsr_get_block_row,
         device=device,
-        dim=x.nnz,
-        inputs=[y.nnz, x.offsets, work_arrays._sum_rows],
+        dim=x_nnz,
+        inputs=[y_nnz, x.nrow, x.offsets, work_arrays._sum_rows],
     )
 
     # Save old y values before overwriting matrix
-    wp.copy(dest=work_arrays._old_y_values, src=y.values, count=y.nnz)
+    wp.copy(dest=work_arrays._old_y_values, src=y.values, count=y_nnz)
 
     # Increase dest array sizes if needed
     if y.columns.shape[0] < sum_nnz:
@@ -787,21 +1342,28 @@ def bsr_axpy(
     else:
         native_func = runtime.core.bsr_matrix_from_triplets_float_device
 
-    old_y_nnz = y.nnz
-    y.nnz = native_func(
-        y.block_shape[0],
-        y.block_shape[1],
-        y.nrow,
-        sum_nnz,
-        work_arrays._sum_rows.ptr,
-        work_arrays._sum_cols.ptr,
-        0,
-        y.offsets.ptr,
-        y.columns.ptr,
-        0,
-    )
+    old_y_nnz = y_nnz
+    nnz_buf, nnz_event = y._nnz_transfer_buf_and_event()
+
+    with wp.ScopedDevice(y.device):
+        native_func(
+            y.block_shape[0],
+            y.block_shape[1],
+            y.nrow,
+            sum_nnz,
+            ctypes.cast(work_arrays._sum_rows.ptr, ctypes.POINTER(ctypes.c_int32)),
+            ctypes.cast(work_arrays._sum_cols.ptr, ctypes.POINTER(ctypes.c_int32)),
+            0,
+            False,
+            ctypes.cast(y.offsets.ptr, ctypes.POINTER(ctypes.c_int32)),
+            ctypes.cast(y.columns.ptr, ctypes.POINTER(ctypes.c_int32)),
+            0,
+            ctypes.cast(nnz_buf.ptr, ctypes.POINTER(ctypes.c_int32)),
+            nnz_event,
+        )
+
+    _bsr_ensure_fits(y, nnz=sum_nnz)
 
-    _bsr_ensure_fits(y)
     y.values.zero_()
 
     wp.launch(
@@ -823,7 +1385,7 @@ def bsr_axpy(
     wp.launch(
         kernel=_bsr_axpy_add_block,
         device=device,
-        dim=x.nnz,
+        dim=x_nnz,
         inputs=[
             old_y_nnz,
             alpha,
@@ -918,8 +1480,9 @@ def _bsr_mm_compute_values(
         y_end = y_offsets[x_col + 1]
 
         y_block = wp.lower_bound(y_columns, y_beg, y_end, col)
-        if y_block < y_end and y_columns[y_block] == col:
-            mm_val += x_values[x_block] * y_values[y_block]
+        if y_block < y_end:
+            if y_columns[y_block] == col:
+                mm_val += x_values[x_block] * y_values[y_block]
 
     mm_values[mm_block] += alpha * mm_val
 
@@ -932,38 +1495,38 @@ class bsr_mm_work_arrays:
 
     def _reset(self, device):
         self.device = device
-        self._pinned_count_buffer = None
         self._mm_row_counts = None
         self._mm_rows = None
         self._mm_cols = None
         self._old_z_values = None
         self._old_z_offsets = None
         self._old_z_columns = None
+        self._mm_nnz = 0
 
-    def _allocate_stage_1(self, device, z: BsrMatrix, copied_z_nnz: int, z_aliasing: bool):
+    def _allocate_stage_1(self, device, z: BsrMatrix, beta: float, z_aliasing: bool):
         if self.device != device:
             self._reset(device)
 
         # Allocations that do not depend on any computation
-        if self.device.is_cuda:
-            if self._pinned_count_buffer is None:
-                self._pinned_count_buffer = wp.empty(shape=(1,), dtype=int, pinned=True, device="cpu")
+        z_nnz = z.nnz_sync()
+        self._copied_z_nnz = z_nnz if beta != 0.0 or z_aliasing else 0
 
         if self._mm_row_counts is None or self._mm_row_counts.size < z.nrow + 1:
             self._mm_row_counts = wp.empty(shape=(z.nrow + 1,), dtype=int, device=self.device)
 
-        if copied_z_nnz > 0:
-            if self._old_z_values is None or self._old_z_values.size < copied_z_nnz:
-                self._old_z_values = wp.empty(shape=(copied_z_nnz,), dtype=z.values.dtype, device=self.device)
+        if self._copied_z_nnz > 0:
+            if self._old_z_values is None or self._old_z_values.size < self._copied_z_nnz:
+                self._old_z_values = wp.empty(shape=(self._copied_z_nnz,), dtype=z.values.dtype, device=self.device)
 
         if z_aliasing:
-            if self._old_z_columns is None or self._old_z_columns.size < z.nnz:
-                self._old_z_columns = wp.empty(shape=(z.nnz,), dtype=z.columns.dtype, device=self.device)
+            if self._old_z_columns is None or self._old_z_columns.size < z_nnz:
+                self._old_z_columns = wp.empty(shape=(z_nnz,), dtype=z.columns.dtype, device=self.device)
             if self._old_z_offsets is None or self._old_z_offsets.size < z.nrow + 1:
                 self._old_z_offsets = wp.empty(shape=(z.nrow + 1,), dtype=z.offsets.dtype, device=self.device)
 
     def _allocate_stage_2(self, mm_nnz: int):
         # Allocations that depend on unmerged nnz estimate
+        self._mm_nnz = mm_nnz
         if self._mm_rows is None or self._mm_rows.size < mm_nnz:
             self._mm_rows = wp.empty(shape=(mm_nnz,), dtype=int, device=self.device)
         if self._mm_cols is None or self._mm_cols.size < mm_nnz:
@@ -971,12 +1534,13 @@ class bsr_mm_work_arrays:
 
 
 def bsr_mm(
-    x: BsrMatrix[BlockType[Rows, Any, Scalar]],
-    y: BsrMatrix[BlockType[Any, Cols, Scalar]],
+    x: BsrMatrixOrExpression[BlockType[Rows, Any, Scalar]],
+    y: BsrMatrixOrExpression[BlockType[Any, Cols, Scalar]],
     z: Optional[BsrMatrix[BlockType[Rows, Cols, Scalar]]] = None,
     alpha: Scalar = 1.0,
     beta: Scalar = 0.0,
     work_arrays: Optional[bsr_mm_work_arrays] = None,
+    reuse_topology: bool = False,
 ) -> BsrMatrix[BlockType[Rows, Cols, Scalar]]:
     """
     Performs the sparse matrix-matrix multiplication ``z := alpha * x * y + beta * z`` on BSR matrices `x`, `y` and `z`, and returns `z`.
@@ -991,8 +1555,16 @@ def bsr_mm(
         alpha: Uniform scaling factor for the ``x * y`` product
         beta: Uniform scaling factor for `z`
         work_arrays: In most cases this function will require the use of temporary storage; this storage can be reused across calls by passing an instance of :class:`bsr_mm_work_arrays` in `work_arrays`.
+        reuse_topology: If True, reuse the product topology information stored in `work_arrays` rather than recompute it from scratch.
+            The matrices x, y and z must be structurally similar to the previous call in which `work_arrays` were populated.
+            This is necessary for `bsr_mm` to be captured in a CUDA graph.
     """
 
+    x, x_scale = _extract_matrix_and_scale(x)
+    alpha *= x_scale
+    y, y_scale = _extract_matrix_and_scale(y)
+    alpha *= y_scale
+
     if z is None:
         # If not output matrix is provided, allocate it for convenience
         z_block_shape = (x.block_shape[0], y.block_shape[1])
@@ -1030,76 +1602,84 @@ def bsr_mm(
     if not isinstance(beta, z.scalar_type):
         beta = z.scalar_type(beta)
 
-    if work_arrays is None:
-        work_arrays = bsr_mm_work_arrays()
-
     z_aliasing = z == x or z == y
-    copied_z_nnz = z.nnz if beta != 0.0 or z_aliasing else 0
 
-    work_arrays._allocate_stage_1(device, z, copied_z_nnz, z_aliasing)
+    if reuse_topology:
+        if work_arrays is None:
+            raise ValueError("`work_arrays` must not be ``None`` in order to reuse matrix-matrix product topology")
 
-    # Prefix sum of number of (unmerged) mm blocks per row
-    wp.launch(
-        kernel=_bsr_mm_count_coeffs,
-        device=device,
-        dim=z.nrow,
-        inputs=[
-            copied_z_nnz,
-            x.offsets,
-            x.columns,
-            y.offsets,
-            work_arrays._mm_row_counts,
-        ],
-    )
-    warp.utils.array_scan(work_arrays._mm_row_counts, work_arrays._mm_row_counts)
-
-    # Get back total counts on host
-    if device.is_cuda:
-        wp.copy(
-            dest=work_arrays._pinned_count_buffer,
-            src=work_arrays._mm_row_counts,
-            src_offset=z.nrow,
-            count=1,
-        )
-        wp.synchronize_stream(wp.get_stream(device))
-        mm_nnz = int(work_arrays._pinned_count_buffer.numpy()[0])
+        copied_z_nnz = work_arrays._copied_z_nnz
+        mm_nnz = work_arrays._mm_nnz
     else:
-        mm_nnz = int(work_arrays._mm_row_counts.numpy()[z.nrow])
+        if device.is_capturing:
+            raise RuntimeError("`bsr_mm` requires `reuse_topology=True` for use in graph capture")
 
-    work_arrays._allocate_stage_2(mm_nnz)
+        if work_arrays is None:
+            work_arrays = bsr_mm_work_arrays()
 
-    # If z has a non-zero scale, save current data before overwriting it
-    if copied_z_nnz > 0:
-        # Copy z row and column indices
-        wp.copy(dest=work_arrays._mm_cols, src=z.columns, count=copied_z_nnz)
+        work_arrays._allocate_stage_1(device, z, beta, z_aliasing)
+        copied_z_nnz = work_arrays._copied_z_nnz
+
+        # Prefix sum of number of (unmerged) mm blocks per row
         wp.launch(
-            kernel=_bsr_get_block_row,
+            kernel=_bsr_mm_count_coeffs,
             device=device,
-            dim=copied_z_nnz,
-            inputs=[0, z.offsets, work_arrays._mm_rows],
+            dim=z.nrow,
+            inputs=[
+                copied_z_nnz,
+                x.offsets,
+                x.columns,
+                y.offsets,
+                work_arrays._mm_row_counts,
+            ],
+        )
+        warp.utils.array_scan(work_arrays._mm_row_counts, work_arrays._mm_row_counts)
+
+        # Get back total counts on host -- we need a synchronization here
+        # Use pinned buffer from z, we are going to need it later anyway
+        nnz_buf, _ = z._nnz_transfer_buf_and_event()
+        stream = wp.get_stream(device) if device.is_cuda else None
+        wp.copy(dest=nnz_buf, src=work_arrays._mm_row_counts, src_offset=z.nrow, count=1, stream=stream)
+        if device.is_cuda:
+            wp.synchronize_stream(stream)
+        mm_nnz = int(nnz_buf.numpy()[0])
+
+        work_arrays._allocate_stage_2(mm_nnz)
+
+        # If z has a non-zero scale, save current data before overwriting it
+        if copied_z_nnz > 0:
+            # Copy z row and column indices
+            wp.copy(dest=work_arrays._mm_cols, src=z.columns, count=copied_z_nnz)
+            wp.launch(
+                kernel=_bsr_get_block_row,
+                device=device,
+                dim=copied_z_nnz,
+                inputs=[0, z.nrow, z.offsets, work_arrays._mm_rows],
+            )
+            if z_aliasing:
+                # If z is aliasing with x or y, need to save topology as well
+                wp.copy(src=z.columns, dest=work_arrays._old_z_columns, count=copied_z_nnz)
+                wp.copy(src=z.offsets, dest=work_arrays._old_z_offsets, count=z.nrow + 1)
+
+        # Fill unmerged mm blocks rows and columns
+        wp.launch(
+            kernel=_bsr_mm_list_coeffs,
+            device=device,
+            dim=z.nrow,
+            inputs=[
+                x.offsets,
+                x.columns,
+                y.offsets,
+                y.columns,
+                work_arrays._mm_row_counts,
+                work_arrays._mm_rows,
+                work_arrays._mm_cols,
+            ],
         )
+
+    if copied_z_nnz > 0:
         # Save current z values in temporary buffer
         wp.copy(src=z.values, dest=work_arrays._old_z_values, count=copied_z_nnz)
-        if z_aliasing:
-            # If z is aliasing with x or y, need to save topology as well
-            wp.copy(src=z.columns, dest=work_arrays._old_z_columns, count=copied_z_nnz)
-            wp.copy(src=z.offsets, dest=work_arrays._old_z_offsets, count=z.nrow + 1)
-
-    # Fill unmerged mm blocks rows and columns
-    wp.launch(
-        kernel=_bsr_mm_list_coeffs,
-        device=device,
-        dim=z.nrow,
-        inputs=[
-            x.offsets,
-            x.columns,
-            y.offsets,
-            y.columns,
-            work_arrays._mm_row_counts,
-            work_arrays._mm_rows,
-            work_arrays._mm_cols,
-        ],
-    )
 
     # Increase dest array size if needed
     if z.columns.shape[0] < mm_nnz:
@@ -1112,20 +1692,31 @@ def bsr_mm(
     else:
         native_func = runtime.core.bsr_matrix_from_triplets_float_device
 
-    z.nnz = native_func(
-        z.block_shape[0],
-        z.block_shape[1],
-        z.nrow,
-        mm_nnz,
-        work_arrays._mm_rows.ptr,
-        work_arrays._mm_cols.ptr,
-        0,
-        z.offsets.ptr,
-        z.columns.ptr,
-        0,
-    )
+    nnz_buf, nnz_event = z._nnz_transfer_buf_and_event()
+
+    with wp.ScopedDevice(z.device):
+        native_func(
+            z.block_shape[0],
+            z.block_shape[1],
+            z.nrow,
+            mm_nnz,
+            ctypes.cast(work_arrays._mm_rows.ptr, ctypes.POINTER(ctypes.c_int32)),
+            ctypes.cast(work_arrays._mm_cols.ptr, ctypes.POINTER(ctypes.c_int32)),
+            0,
+            False,
+            ctypes.cast(z.offsets.ptr, ctypes.POINTER(ctypes.c_int32)),
+            ctypes.cast(z.columns.ptr, ctypes.POINTER(ctypes.c_int32)),
+            0,
+            ctypes.cast(nnz_buf.ptr, ctypes.POINTER(ctypes.c_int32)),
+            nnz_event,
+        )
+
+    # Resize z to fit mm result if necessary
+    # If we are not reusing the product topology, this needs another synchronization
+    if not reuse_topology:
+        work_arrays.result_nnz = z.nnz_sync()
+    _bsr_ensure_fits(z, nnz=work_arrays.result_nnz)
 
-    _bsr_ensure_fits(z)
     z.values.zero_()
 
     if copied_z_nnz > 0:
@@ -1206,12 +1797,57 @@ def _bsr_mv_kernel(
     y[row] = v
 
 
+@wp.kernel
+def _bsr_mv_transpose_kernel(
+    alpha: Any,
+    A_offsets: wp.array(dtype=int),
+    A_columns: wp.array(dtype=int),
+    A_values: wp.array(dtype=Any),
+    x: wp.array(dtype=Any),
+    y: wp.array(dtype=Any),
+):
+    row = wp.tid()
+    beg = A_offsets[row]
+    end = A_offsets[row + 1]
+    xr = alpha * x[row]
+    for block in range(beg, end):
+        v = wp.transpose(A_values[block]) * xr
+        wp.atomic_add(y, A_columns[block], v)
+
+
+def _bsr_mv_as_vec_array(array: wp.array) -> wp.array:
+    if array.ndim == 1:
+        return array
+
+    if array.ndim > 2:
+        raise ValueError(f"Incompatible array number of dimensions {array.ndim}")
+
+    if not array.is_contiguous:
+        raise ValueError("2d array must be contiguous")
+
+    def vec_view(array):
+        return wp.array(
+            data=None,
+            ptr=array.ptr,
+            capacity=array.capacity,
+            device=array.device,
+            dtype=wp.vec(length=array.shape[1], dtype=array.dtype),
+            shape=array.shape[0],
+            grad=None if array.grad is None else vec_view(array.grad),
+        )
+
+    view = vec_view(array)
+    view._ref = array
+    return view
+
+
 def bsr_mv(
-    A: BsrMatrix[BlockType[Rows, Cols, Scalar]],
+    A: BsrMatrixOrExpression[BlockType[Rows, Cols, Scalar]],
     x: "Array[Vector[Cols, Scalar] | Scalar]",
     y: Optional["Array[Vector[Rows, Scalar] | Scalar]"] = None,
     alpha: Scalar = 1.0,
     beta: Scalar = 0.0,
+    transpose: bool = False,
     work_buffer: Optional["Array[Vector[Rows, Scalar] | Scalar]"] = None,
 ) -> "Array[Vector[Rows, Scalar] | Scalar]":
     """
@@ -1225,16 +1861,26 @@ def bsr_mv(
         y: Mutable left-hand-side. If `y` is not provided, it will be allocated and treated as zero.
         alpha: Uniform scaling factor for `x`. If zero, `x` will not be read and may be left uninitialized.
         beta: Uniform scaling factor for `y`. If zero, `y` will not be read and may be left uninitialized.
+        transpose: If ``True``, use the transpose of the matrix `A`. In this case the result is **non-deterministic**.
         work_buffer: Temporary storage is required if and only if `x` and `y` are the same vector. If provided the `work_buffer` array
             will be used for this purpose, otherwise a temporary allocation will be performed.
     """
 
+    A, A_scale = _extract_matrix_and_scale(A)
+    alpha *= A_scale
+
+    if transpose:
+        block_shape = A.block_shape[1], A.block_shape[0]
+        nrow, ncol = A.ncol, A.nrow
+    else:
+        block_shape = A.block_shape
+        nrow, ncol = A.nrow, A.ncol
+
     if y is None:
         # If no output array is provided, allocate one for convenience
-        y_vec_len = A.block_shape[0]
+        y_vec_len = block_shape[0]
         y_dtype = A.scalar_type if y_vec_len == 1 else wp.vec(length=y_vec_len, dtype=A.scalar_type)
-        y = wp.empty(shape=(A.nrow,), device=A.values.device, dtype=y_dtype)
-        y.zero_()
+        y = wp.empty(shape=(nrow,), device=A.values.device, dtype=y_dtype)
         beta = 0.0
 
     if not isinstance(alpha, A.scalar_type):
@@ -1245,12 +1891,16 @@ def bsr_mv(
     if A.values.device != x.device or A.values.device != y.device:
         raise ValueError("A, x and y must reside on the same device")
 
-    if x.shape[0] != A.ncol:
+    if x.shape[0] != ncol:
         raise ValueError("Number of columns of A must match number of rows of x")
-    if y.shape[0] != A.nrow:
+    if y.shape[0] != nrow:
         raise ValueError("Number of rows of A must match number of rows of y")
 
-    if x == y:
+    # View 2d arrays as arrays of vecs
+    x = _bsr_mv_as_vec_array(x)
+    y = _bsr_mv_as_vec_array(y)
+
+    if x.ptr == y.ptr:
         # Aliasing case, need temporary storage
         if work_buffer is None:
             work_buffer = wp.empty_like(y)
@@ -1265,25 +1915,39 @@ def bsr_mv(
 
     # Promote scalar vectors to length-1 vecs and conversely
     if warp.types.type_is_matrix(A.values.dtype):
-        if A.block_shape[0] == 1:
-            if y.dtype == A.scalar_type:
-                y = y.view(dtype=wp.vec(length=1, dtype=A.scalar_type))
-        if A.block_shape[1] == 1:
-            if x.dtype == A.scalar_type:
-                x = x.view(dtype=wp.vec(length=1, dtype=A.scalar_type))
+        if block_shape[0] == 1 and y.dtype == A.scalar_type:
+            y = y.view(dtype=wp.vec(length=1, dtype=A.scalar_type))
+        if block_shape[1] == 1 and x.dtype == A.scalar_type:
+            x = x.view(dtype=wp.vec(length=1, dtype=A.scalar_type))
     else:
-        if A.block_shape[0] == 1:
-            if y.dtype != A.scalar_type:
-                y = y.view(dtype=A.scalar_type)
-        if A.block_shape[1] == 1:
-            if x.dtype != A.scalar_type:
-                x = x.view(dtype=A.scalar_type)
-
-    wp.launch(
-        kernel=_bsr_mv_kernel,
-        device=A.values.device,
-        dim=A.nrow,
-        inputs=[alpha, A.offsets, A.columns, A.values, x, beta, y],
-    )
+        if block_shape[0] == 1 and y.dtype != A.scalar_type:
+            y = y.view(dtype=A.scalar_type)
+        if block_shape[1] == 1 and x.dtype != A.scalar_type:
+            x = x.view(dtype=A.scalar_type)
+
+    if transpose:
+        if beta.value == 0.0:
+            y.zero_()
+        elif beta.value != 1.0:
+            wp.launch(
+                kernel=_bsr_scale_kernel,
+                device=y.device,
+                dim=y.shape[0],
+                inputs=[beta, y],
+            )
+        if alpha.value != 0.0:
+            wp.launch(
+                kernel=_bsr_mv_transpose_kernel,
+                device=A.values.device,
+                dim=ncol,
+                inputs=[alpha, A.offsets, A.columns, A.values, x, y],
+            )
+    else:
+        wp.launch(
+            kernel=_bsr_mv_kernel,
+            device=A.values.device,
+            dim=nrow,
+            inputs=[alpha, A.offsets, A.columns, A.values, x, beta, y],
+        )
 
     return y