blksprs 1.7__tar.gz → 1.8__tar.gz

{blksprs-1.7 → blksprs-1.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: blksprs
-Version: 1.7
+Version: 1.8
 Summary: A lightweight library for operations on blocksparse matrices in PyTorch.
 Author-email: Felix Schön <schoen@kr.tuwien.ac.at>
 Project-URL: Homepage, https://github.com/FelixSchoen/blksprs
@@ -28,13 +28,13 @@ A lightweight and efficient library for operations on block-sparse matrices in P
 
 Currently supported operations (includes gradient calculation):
 
-- Sparse matrix multiplication (_supports any combination of sparse and dense matrices due to support
-  for `sparse = sparse @ sparse` matmul_)
+- Matrix multiplication
 - Softmax
 - Transpose
 - Gather
 - Scatter (_supports either no reduction or summation, gradients are only available for summation_)
 - Repeat (_supports target sparsity layout_)
+- Repeat Interleave (_supports target sparsity layout_)
 - Splitting and merging of matrices along the last dimension
 - Conversion to and from sparse form
 - Conversion to different sparsity layouts and different sparsity block sizes
@@ -51,8 +51,14 @@ These include, e.g.,
 Note that in order to correctly apply element-wise operations between two sparse tensors their sparsity layouts have to
 match.
 
+Further helpful operations (included in the ``bs.misc`` module) that do **not** support gradient calculation include:
+
+- Row-wise sum, max, addition, and subtraction
+- Broadcast addition and subtraction between slices
+
 Furthermore, the library provides a set of utility functions for the creation of sparsity layouts based on existing
-dense tensors.
+dense tensors and for the scatter operation (module ``bs.layout``), as well as utility functions to ensure correct input
+dimensionality (module ``bs.util``).
 
 ## Installation
 

{blksprs-1.7 → blksprs-1.8}/README.md

@@ -9,13 +9,13 @@ A lightweight and efficient library for operations on block-sparse matrices in P
 
 Currently supported operations (includes gradient calculation):
 
-- Sparse matrix multiplication (_supports any combination of sparse and dense matrices due to support
-  for `sparse = sparse @ sparse` matmul_)
+- Matrix multiplication
 - Softmax
 - Transpose
 - Gather
 - Scatter (_supports either no reduction or summation, gradients are only available for summation_)
 - Repeat (_supports target sparsity layout_)
+- Repeat Interleave (_supports target sparsity layout_)
 - Splitting and merging of matrices along the last dimension
 - Conversion to and from sparse form
 - Conversion to different sparsity layouts and different sparsity block sizes
@@ -32,8 +32,14 @@ These include, e.g.,
 Note that in order to correctly apply element-wise operations between two sparse tensors their sparsity layouts have to
 match.
 
+Further helpful operations (included in the ``bs.misc`` module) that do **not** support gradient calculation include:
+
+- Row-wise sum, max, addition, and subtraction
+- Broadcast addition and subtraction between slices
+
 Furthermore, the library provides a set of utility functions for the creation of sparsity layouts based on existing
-dense tensors.
+dense tensors and for the scatter operation (module ``bs.layout``), as well as utility functions to ensure correct input
+dimensionality (module ``bs.util``).
 
 ## Installation
 

{blksprs-1.7 → blksprs-1.8}/blksprs/__init__.py

@@ -1,22 +1,27 @@
-from blksprs.ops.conversion import to_dense, to_sparse
+from blksprs.ops.conversion import to_dense, to_sparse, from_blksprs, to_blksprs
 from blksprs.ops.distribution import gather, scatter, scatter_reduce
-from blksprs.ops.exp import exp
 from blksprs.ops.matmul import matmul
 from blksprs.ops.softmax import softmax
 from blksprs.ops.transpose import transpose
-from blksprs.ops.partitioning import split, merge
+from blksprs.ops.repeat import repeat, repeat_interleave
+from blksprs.misc.partitioning import split, merge
+
 
 class layout:
     from blksprs.layouting.distribution_layout import build_distribution_layout
-    from blksprs.layouting.sparsity_layout import build_sparsity_layout, build_sparsity_layout_adaption, build_sparsity_layout_matmul
+    from blksprs.layouting.sparsity_layout import build_sparsity_layout, build_sparsity_layout_adaption, \
+        build_sparsity_layout_matmul
+
 
 class misc:
     from blksprs.misc.broadcast_ops import broadcast_add, broadcast_sub
-    from blksprs.misc.repeat_interleave import repeat_interleave
+    from blksprs.misc.exp import exp
     from blksprs.misc.row_wise import row_wise_sum, row_wise_max, row_wise_add, row_wise_sub
 
+
 class util:
     from blksprs.utils.tools import do_shape_blocksparse, undo_shape_blocksparse, disable_validation
 
+
 class experimental:
-    from blksprs.experimental.distribution_mdi import gather_mdi, scatter_reduce_mdi
+    from blksprs.experimental.distribution_mdi import gather_mdi, scatter_reduce_mdi

{blksprs-1.7 → blksprs-1.8}/blksprs/ops/conversion.py

@@ -8,7 +8,12 @@ from triton import language as tl
 from blksprs.layouting.sparsity_layout import build_sparsity_layout_adaption
 from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_contiguous, validate_dimensions, validate_device, \
-    validate_sparsity, validate_sparsity_block_size, validate_triton_block_size
+    validate_sparsity, validate_sparsity_block_size, validate_triton_block_size, validate_sparsity_dense
+
+
+def from_blksprs(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int, fill_value: float = 0,
+                 triton_block_size: int = None) -> Tensor:
+    return to_dense(x, sparsity_layout, sparsity_block_size, fill_value, triton_block_size)
 
 
 def to_dense(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int, fill_value: float = 0,
@@ -144,6 +149,11 @@ class _BlocksparseToDense(torch.autograd.Function):
             tl.store(o + o_idx, blk, o_msk)
 
 
+def to_blksprs(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int,
+               triton_block_size: int = None) -> Tensor:
+    return to_sparse(x, sparsity_layout, sparsity_block_size, triton_block_size)
+
+
 def to_sparse(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int, triton_block_size: int = None) -> Tensor:
     """Converts a block-sparse tensor in regular form to a block-sparse tensor in compressed form based on the given
     sparsity layout.
@@ -163,6 +173,7 @@ def to_sparse(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int, trit
     validate_dimensions(x)
     validate_contiguous(x)
     validate_device(x)
+    validate_sparsity_dense(sparsity_block_size, (x, sparsity_layout))
     validate_sparsity_block_size(sparsity_block_size, x)
     validate_triton_block_size(triton_block_size, sparsity_block_size)
 

{blksprs-1.7 → blksprs-1.8}/blksprs/ops/repeat.py

@@ -11,6 +11,30 @@ from blksprs.utils.validation import validate_dimensions, validate_contiguous, v
 def repeat(x: Tensor, sparsity_layout_x: Tensor, repeats: tuple[int, int, int],
            sparsity_block_size: int, sparsity_layout_output: Tensor = None, triton_block_size: int = None) -> (
         Tensor, Tensor):
+    """Repeats a block-spare tensor in compressed form according to the given repeats.
+    
+    Repeats is a 3-tuple of integers, where each integer represents the number of times the tensor should be repeated in
+        the first, second and third dimension respectively.
+        
+    Note:
+        An output sparsity layout can be provided, in which case only the indicated blocks are filled. This may result
+        in blocks not being present in the output that were present in the input if the output sparsity layout indicates
+        them to be sparse.
+    
+    Args:
+        x (Tensor): A block-sparse tensor in compressed form.
+        sparsity_layout_x (Tensor): The sparsity layout of the block-sparse tensor.
+        repeats (tuple[int, int, int]): The number of times the tensor should be repeated in the first, second and
+            third dimension respectively.
+        sparsity_block_size (int): The size of the sparsity blocks.
+        sparsity_layout_output (Tensor): The desired sparsity layout of the output tensor (default ``None``).
+        triton_block_size (int): The block size to use for the triton kernel (default ``None``).
+
+    Returns:
+        Tensor: A block-sparse tensor in compressed form containing the repeated values.
+        Tensor: The sparsity layout of the resulting output tensor.
+
+    """
     x = x.contiguous()
 
     validate_dimensions(x)
@@ -43,6 +67,64 @@ def repeat(x: Tensor, sparsity_layout_x: Tensor, repeats: tuple[int, int, int],
                                     sparsity_block_size, n_sparse_blocks, triton_block_size), sparsity_layout_o
 
 
+def repeat_interleave(x: Tensor, sparsity_layout_x: Tensor, repeats: int,
+                      sparsity_block_size: int, sparsity_layout_output: Tensor = None,
+                      triton_block_size: int = None) -> (
+        Tensor, Tensor):
+    """Repeats and interleaves the block-sparse tensor in compressed form.
+
+    Repeats each matrix contained in the tensors by ``repeats`` amount and places them consecutively in the output
+        tensor.
+
+    Note:
+        In similar fashion to the regular ``repeat`` an output sparsity layout can be provided. In this case only
+        non-sparse blocks will be filled.
+
+    Args:
+        x (Tensor): A block-sparse tensor in compressed form.
+        sparsity_layout_x (Tensor): The sparsity layout of the block-sparse tensor.
+        repeats (int): The number of times to repeat the matrices.
+        sparsity_block_size (int): The size of the sparsity blocks.
+        sparsity_layout_output (Tensor): The desired sparsity layout of the output tensor (default ``None``).
+        triton_block_size (int): The block size to use for the triton kernel (default ``None``).
+
+    Returns:
+        Tensor: A block-sparse tensor in compressed form containing the repeated and interleaved matrices.
+        Tensor: The sparsity layout of the resulting output tensor.
+
+    """
+    x = x.contiguous()
+
+    validate_dimensions(x)
+    validate_contiguous(x)
+    validate_device(x)
+    validate_sparsity(sparsity_block_size, (x, sparsity_layout_x))
+    validate_sparsity_block_size(sparsity_block_size, x)
+    validate_triton_block_size(triton_block_size, sparsity_block_size)
+
+    sparsity_layout_o = torch.repeat_interleave(sparsity_layout_x, repeats, dim=0).contiguous()
+
+    if sparsity_layout_output is not None:
+        sparsity_layout_o = torch.logical_and(sparsity_layout_o, sparsity_layout_output)
+
+    sparsity_lut = torch.nonzero(sparsity_layout_o).contiguous()
+
+    sparsity_layout_flat = sparsity_layout_x.reshape(-1)
+    sparsity_reverse_lut = (((torch.cumsum(sparsity_layout_flat, dim=-1) - 1) *
+                             (sparsity_layout_flat == 1) -
+                             (1 * (sparsity_layout_flat == 0)))
+                            .reshape(sparsity_layout_x.size())
+                            .repeat_interleave(repeats, dim=0)
+                            .reshape(-1).contiguous())
+
+    n_sparse_blocks = torch.sum(sparsity_layout_o.to(torch.int)).item()
+
+    validate_contiguous(sparsity_layout_o, sparsity_lut, sparsity_reverse_lut)
+
+    return _BlocksparseRepeat.apply(x, sparsity_layout_x, sparsity_layout_o, sparsity_lut, sparsity_reverse_lut,
+                                    sparsity_block_size, n_sparse_blocks, triton_block_size), sparsity_layout_o
+
+
 class _BlocksparseRepeat(torch.autograd.Function):
 
     @staticmethod
@@ -215,7 +297,6 @@ def forward_flow(ctx, x: Tensor, sparsity_layout_o: Tensor, sparsity_lut: Tensor
     s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = stride(sparsity_layout_o)
     s_lut_r, s_lut_c = sparsity_lut.size()
     s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
-    asdf = torch.tensor(sparsity_lut).stride()
 
     if triton_block_size is None:
         triton_block_size = get_triton_block_size(sparsity_block_size)

{blksprs-1.7 → blksprs-1.8}/blksprs/ops/softmax.py

@@ -3,7 +3,7 @@ import triton
 from torch import Tensor
 from triton import language as tl
 
-from blksprs.ops.exp import exp
+from blksprs.misc.exp import exp
 from blksprs.misc.row_wise import row_wise_sum, row_wise_max, row_wise_sub
 from blksprs.utils.tools import get_triton_block_size, stride
 from blksprs.utils.validation import validate_contiguous, validate_dimensions, validate_device, \

{blksprs-1.7 → blksprs-1.8}/blksprs/utils/validation.py

@@ -3,6 +3,7 @@ from torch import Tensor
 
 VALIDATION = True
 
+
 def validate_dimensions(*tensors: Tensor, dims=3) -> None:
     if _check_skip_validation():
         return
@@ -71,10 +72,25 @@ def validate_sparsity(sparsity_block_size: int, *tensor_sparsity_layout_tuples:
             raise ValueError("Mismatch between sparsity layout and blocks")
 
 
+def validate_sparsity_dense(sparsity_block_size: int, *tensor_sparsity_layout_tuples: tuple[Tensor, Tensor]) -> None:
+    if _check_skip_validation():
+        return
+
+    for (tensor, sparsity_layout) in tensor_sparsity_layout_tuples:
+        _validate_sparsity_layout_values(sparsity_layout)
+
+        if not sparsity_layout.dim() == 3:
+            raise ValueError("Sparsity layout must have exactly 3 dimensions")
+        if not (tensor.size(-1) // sparsity_block_size == sparsity_layout.size(-1) and
+                tensor.size(-2) // sparsity_block_size == sparsity_layout.size(-2)):
+            raise ValueError("Tensor not conforming to sparsity layout")
+
+
 def _validate_sparsity_layout_values(sparsity_layout: Tensor):
     if not torch.all(torch.logical_or(sparsity_layout == 0, sparsity_layout == 1)):
         raise ValueError("Sparsity layout values must be either 0 or 1")
 
+
 def validate_sparsity_block_size(sparsity_block_size: int, *tensors):
     if _check_skip_validation():
         return
@@ -86,6 +102,7 @@ def validate_sparsity_block_size(sparsity_block_size: int, *tensors):
         if not (tensor.size(-1) % sparsity_block_size == 0 and tensor.size(-2) % sparsity_block_size == 0):
             raise ValueError("Tensor sizes must be divisible by sparsity block size")
 
+
 def validate_triton_block_size(triton_block_size: int, sparsity_block_size: int):
     if _check_skip_validation():
         return
@@ -99,9 +116,11 @@ def validate_triton_block_size(triton_block_size: int, sparsity_block_size: int)
     if triton_block_size > sparsity_block_size:
         raise ValueError("Triton block size cannot be larger than sparsity block size")
 
+
 def _check_skip_validation():
     return not VALIDATION
 
+
 def _set_skip_validation(skip_validation: bool):
     global VALIDATION
-    VALIDATION = not skip_validation
+    VALIDATION = not skip_validation

{blksprs-1.7 → blksprs-1.8}/blksprs.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: blksprs
-Version: 1.7
+Version: 1.8
 Summary: A lightweight library for operations on blocksparse matrices in PyTorch.
 Author-email: Felix Schön <schoen@kr.tuwien.ac.at>
 Project-URL: Homepage, https://github.com/FelixSchoen/blksprs
@@ -28,13 +28,13 @@ A lightweight and efficient library for operations on block-sparse matrices in P
 
 Currently supported operations (includes gradient calculation):
 
-- Sparse matrix multiplication (_supports any combination of sparse and dense matrices due to support
-  for `sparse = sparse @ sparse` matmul_)
+- Matrix multiplication
 - Softmax
 - Transpose
 - Gather
 - Scatter (_supports either no reduction or summation, gradients are only available for summation_)
 - Repeat (_supports target sparsity layout_)
+- Repeat Interleave (_supports target sparsity layout_)
 - Splitting and merging of matrices along the last dimension
 - Conversion to and from sparse form
 - Conversion to different sparsity layouts and different sparsity block sizes
@@ -51,8 +51,14 @@ These include, e.g.,
 Note that in order to correctly apply element-wise operations between two sparse tensors their sparsity layouts have to
 match.
 
+Further helpful operations (included in the ``bs.misc`` module) that do **not** support gradient calculation include:
+
+- Row-wise sum, max, addition, and subtraction
+- Broadcast addition and subtraction between slices
+
 Furthermore, the library provides a set of utility functions for the creation of sparsity layouts based on existing
-dense tensors.
+dense tensors and for the scatter operation (module ``bs.layout``), as well as utility functions to ensure correct input
+dimensionality (module ``bs.util``).
 
 ## Installation
 

{blksprs-1.7 → blksprs-1.8}/blksprs.egg-info/SOURCES.txt

@@ -10,13 +10,12 @@ blksprs/experimental/distribution_mdi.py
 blksprs/layouting/distribution_layout.py
 blksprs/layouting/sparsity_layout.py
 blksprs/misc/broadcast_ops.py
-blksprs/misc/repeat_interleave.py
+blksprs/misc/exp.py
+blksprs/misc/partitioning.py
 blksprs/misc/row_wise.py
 blksprs/ops/conversion.py
 blksprs/ops/distribution.py
-blksprs/ops/exp.py
 blksprs/ops/matmul.py
-blksprs/ops/partitioning.py
 blksprs/ops/repeat.py
 blksprs/ops/softmax.py
 blksprs/ops/transpose.py

{blksprs-1.7 → blksprs-1.8}/pyproject.toml

@@ -1,6 +1,6 @@
 [project]
 name = "blksprs"
-version = "1.7"
+version = "1.8"
 authors = [{ name = "Felix Schön", email = "schoen@kr.tuwien.ac.at" }]
 description = "A lightweight library for operations on blocksparse matrices in PyTorch."
 readme = "README.md"

blksprs-1.7/blksprs/misc/repeat_interleave.py

@@ -1,132 +0,0 @@
-import torch
-import triton
-from torch import Tensor
-from triton import language as tl
-
-from blksprs.utils.tools import get_triton_block_size, stride
-from blksprs.utils.validation import validate_contiguous, validate_device, \
-    validate_sparsity_block_size, validate_triton_block_size, validate_dimensions
-
-
-def repeat_interleave(x: Tensor, sparsity_layout: Tensor, repeats: int,
-                      sparsity_block_size: int, triton_block_size: int = None) -> tuple[Tensor, Tensor]:
-    """Repeats and interleaves the block-sparse tensor in compressed form.
-
-    Repeats each matrix contained in the tensors by ``repeats`` amount and places them consecutively in the output
-        tensor.
-
-    Args:
-        x (Tensor): A block-sparse tensor in compressed form.
-        sparsity_layout (Tensor): The sparsity layout of the block-sparse tensor.
-        repeats (int): The number of times to repeat the matrices.
-        sparsity_block_size (int): The size of the sparsity blocks.
-        triton_block_size (int): The block size to use for the triton kernel (default ``None``).
-
-    Returns:
-        Tensor: A block-sparse tensor in compressed form containing the repeated and interleaved matrices.
-        Tensor: The sparsity layout of the resulting output tensor.
-
-    """
-    x = x.contiguous()
-
-    validate_dimensions(x)
-    validate_contiguous(x)
-    validate_device(x)
-    validate_sparsity_block_size(sparsity_block_size, x)
-    validate_triton_block_size(triton_block_size, sparsity_block_size)
-
-    sparsity_layout_output = torch.repeat_interleave(sparsity_layout, repeats, dim=0).contiguous()
-
-    sparsity_lut = torch.nonzero(sparsity_layout).contiguous()
-
-    sparsity_layout_output_flat = sparsity_layout_output.reshape(-1)
-    sparsity_output_reverse_lut = ((torch.cumsum(sparsity_layout_output_flat, dim=-1) - 1) *
-                                   (sparsity_layout_output_flat == 1) -
-                                   (1 * (sparsity_layout_output_flat == 0)))
-
-    n_sparse_blocks = torch.sum(sparsity_layout.to(torch.int)).item()
-
-    validate_contiguous(sparsity_layout, sparsity_lut, sparsity_layout_output, sparsity_output_reverse_lut)
-
-    output = torch.empty(n_sparse_blocks * repeats, sparsity_block_size, sparsity_block_size,
-                         dtype=x.dtype, device=x.device)
-
-    x_b, x_r, x_c = x.size()
-    x_b_s, x_r_s, x_c_s = stride(x)
-    s_lut_r, s_lut_c = sparsity_lut.size()
-    s_lut_r_s, s_lut_c_s = stride(sparsity_lut)
-    o_b, o_r, o_c = output.size()
-    o_b_s, o_r_s, o_c_s = stride(output)
-    s_l_o_b, s_l_o_r, s_l_o_c = sparsity_layout_output.size()
-    s_l_o_b_s, s_l_o_r_s, s_l_o_c_s = stride(sparsity_layout_output)
-
-    if triton_block_size is None:
-        triton_block_size = get_triton_block_size(sparsity_block_size)
-
-    triton_grid = lambda meta: [x_b,
-                                triton.cdiv(x_r, meta["TRITON_BLOCK_SIZE"]),
-                                triton.cdiv(x_c, meta["TRITON_BLOCK_SIZE"])]
-
-    (kernel_repeat_interleave[triton_grid]
-     (x,
-      x_b, x_b_s, x_r_s, x_c_s,
-      sparsity_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
-      output,
-      o_b, o_b_s, o_r_s, o_c_s,
-      s_l_o_b, s_l_o_b_s, s_l_o_r_s, s_l_o_c_s,
-      sparsity_output_reverse_lut,
-      repeats,
-      triton_block_size))
-
-    return output, sparsity_layout_output
-
-
-@triton.jit
-def kernel_repeat_interleave(x,
-                             x_b, x_b_s, x_r_s, x_c_s,
-                             s_lut, s_lut_r, s_lut_r_s, s_lut_c_s,
-                             o,
-                             o_b, o_b_s, o_r_s, o_c_s,
-                             s_l_o_b, s_l_o_b_s, s_l_o_r_s, s_l_o_c_s,
-                             r_lut_o,
-                             repeats,
-                             TRITON_BLOCK_SIZE: tl.constexpr) -> None:
-    # Get triton block indices
-    pid_blk = tl.program_id(axis=0)
-    pid_row = tl.program_id(axis=1)
-    pid_col = tl.program_id(axis=2)
-
-    # Get sparsity index of current output block consisting of its batch, row, and column index
-    spa_bat_idx = (pid_blk * s_lut_r_s + 0 * s_lut_c_s)
-    spa_bat_msk = (spa_bat_idx < s_lut_r * s_lut_r_s)
-    spa_bat = tl.load(s_lut + spa_bat_idx, mask=spa_bat_msk)
-
-    spa_row_idx = (pid_blk * s_lut_r_s + 1 * s_lut_c_s)
-    spa_row_msk = (spa_row_idx < s_lut_r * s_lut_r_s)
-    spa_row = tl.load(s_lut + spa_row_idx, mask=spa_row_msk)
-
-    spa_col_idx = (pid_blk * s_lut_r_s + 2 * s_lut_c_s)
-    spa_col_msk = (spa_col_idx < s_lut_r * s_lut_r_s)
-    spa_col = tl.load(s_lut + spa_col_idx, mask=spa_col_msk)
-
-    # Load block
-    blk_x_idx = ((pid_blk * x_b_s) +
-                 ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
-                 ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
-    blk_x_msk = (blk_x_idx < x_b * x_b_s)
-    blk_x = tl.load(x + blk_x_idx, mask=blk_x_msk)
-
-    for repeat in range(repeats):
-        # Get reverse sparsity index
-        rev_idx_spa_idx = ((spa_bat * repeats + repeat) * s_l_o_b_s +
-                           spa_row * s_l_o_r_s +
-                           spa_col * s_l_o_c_s)
-        rev_idx_spa_msk = (rev_idx_spa_idx < s_l_o_b * s_l_o_b_s)
-        rev_idx_spa = tl.load(r_lut_o + rev_idx_spa_idx, mask=rev_idx_spa_msk).to(tl.int32)
-
-        # Store block
-        blk_o_idx = ((rev_idx_spa * o_b_s) +
-                     ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * o_r_s)[:, None] +
-                     ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * o_c_s)[None, :])
-        blk_o_msk = (blk_o_idx < o_b * o_b_s)
-        tl.store(o + blk_o_idx, blk_x, mask=blk_o_msk)