blksprs 0.2b4__py3-none-any.whl → 1.1__py3-none-any.whl

blksprs/layouting/distribution_layout.py

@@ -0,0 +1,114 @@
+import torch
+import triton
+from torch import Tensor
+from triton import language as tl
+
+from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.validation import validate_triton_block_size, validate_dimensions, validate_device, \
+    validate_contiguous
+
+
+def build_distribution_layout(indices: Tensor, sparsity_layout_indices: Tensor,
+                              size_target: torch.Size,
+                              sparsity_block_size: int, triton_block_size: int = None) -> Tensor:
+    """Builds the sparsity layout of either the source of a gather or the target of a scatter operation.
+
+    Args:
+        indices (Tensor): The block-sparse indices tensor in compressed form used for the gather or scatter operation.
+        sparsity_layout_indices (Tensor): The sparsity layout of the indices block-sparse tensor.
+        size_target (torch.Size): The size of the block-sparse target tensor in regular form.
+        sparsity_block_size (int): The size of the sparsity blocks.
+        triton_block_size (int, optional): The block size to use for the triton kernel (default ``None``).
+
+    Returns:
+        Tensor: The sparsity layout of the source or target tensor.
+
+    """
+    validate_dimensions(indices)
+    validate_contiguous(indices)
+    validate_device(indices)
+
+    sparsity_lut_i = torch.nonzero(sparsity_layout_indices).contiguous()
+
+    output = torch.zeros(size_target[0], size_target[1] // sparsity_block_size, size_target[2] // sparsity_block_size,
+                         device=indices.device, dtype=torch.int32)
+
+    i_b, i_r, i_c = indices.size()
+    i_b_s, i_r_s, i_c_s = indices.stride()
+    s_l_i_b, s_l_i_r, s_l_i_c = sparsity_layout_indices.size()
+    s_l_i_b_s, s_l_i_r_s, s_l_i_c_s = sparsity_layout_indices.stride()
+    s_lut_i_r, s_lut_i_c = sparsity_lut_i.size()
+    s_lut_i_r_s, s_lut_i_c_s = sparsity_lut_i.stride()
+    o_b, o_r, o_c = output.size()
+    o_b_s, o_r_s, o_c_s = output.stride()
+
+    if triton_block_size is None:
+        triton_block_size = get_triton_block_size(sparsity_block_size)
+
+    validate_triton_block_size(triton_block_size, sparsity_block_size)
+
+    triton_grid = lambda meta: [i_b,
+                                triton.cdiv(i_r, meta["TRITON_BLOCK_SIZE"]),
+                                triton.cdiv(i_c, meta["TRITON_BLOCK_SIZE"])]
+
+    (kernel_distribution_layout[triton_grid]
+     (indices,
+      i_b, i_b_s, i_r_s, i_c_s,
+      sparsity_layout_indices,
+      s_l_i_b, s_l_i_b_s, s_l_i_r, s_l_i_r_s, s_l_i_c, s_l_i_c_s,
+      sparsity_lut_i,
+      s_lut_i_r, s_lut_i_r_s, s_lut_i_c, s_lut_i_c_s,
+      output,
+      o_b, o_b_s, o_r, o_r_s, o_c, o_c_s,
+      sparsity_block_size,
+      triton_block_size))
+
+    return output
+
+
+@triton.jit
+def kernel_distribution_layout(i,
+                               i_b, i_b_s, i_r_s, i_c_s,
+                               s_l_i,
+                               s_l_i_b, s_l_i_b_s, s_l_i_r, s_l_i_r_s, s_l_i_c, s_l_i_c_s,
+                               s_lut_i,
+                               s_lut_i_r, s_lut_i_r_s, s_lut_i_c, s_lut_i_c_s,
+                               o,
+                               o_b, o_b_s, o_r, o_r_s, o_c, o_c_s,
+                               sparsity_block_size,
+                               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 position of current sparsity block consisting of its batch, row, and column index
+    spa_bat_i_idx = (pid_blk * s_lut_i_r_s + 0 * s_lut_i_c_s)
+    spa_bat_i_msk = (spa_bat_i_idx < s_lut_i_r * s_lut_i_r_s)
+    spa_bat_i = tl.load(s_lut_i + spa_bat_i_idx, mask=spa_bat_i_msk)
+
+    spa_row_i_idx = (pid_blk * s_lut_i_r_s + 1 * s_lut_i_c_s)
+    spa_row_i_msk = (spa_row_i_idx < s_lut_i_r * s_lut_i_r_s)
+    spa_row_i = tl.load(s_lut_i + spa_row_i_idx, mask=spa_row_i_msk)
+
+    blk_i_idx = (pid_blk * i_b_s +
+                 ((pid_row * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * i_r_s)[:, None] +
+                 ((pid_col * TRITON_BLOCK_SIZE + tl.arange(0, TRITON_BLOCK_SIZE)) * i_c_s)[None, :])
+    blk_i_msk = (blk_i_idx < i_b * i_b_s)
+    blk_i = tl.load(i + blk_i_idx, mask=blk_i_msk)
+
+    blk_i = blk_i // sparsity_block_size
+    blk_v = tl.full((TRITON_BLOCK_SIZE, TRITON_BLOCK_SIZE), 1, dtype=tl.int32)
+
+    blk_o_idx = ((spa_bat_i * o_b_s) +
+                 (spa_row_i * o_r_s) +
+                 (blk_i * o_c_s))
+    blk_o_msk = (blk_o_idx < o_b * o_b_s)
+    tl.store(o + blk_o_idx, blk_v, mask=blk_o_msk)
+
+    # if tl.min(blk_x) != 0 or tl.max(blk_x) != 0:
+    #     blk_o_idx = (pid_bat * o_b_s +
+    #                  (((pid_row * TRITON_BLOCK_SIZE) // sparsity_block_size) * o_r_s +
+    #                   ((pid_col * TRITON_BLOCK_SIZE) // sparsity_block_size) * o_c_s))
+    #     blk_o_msk = (blk_o_idx < o_b * o_b_s)
+    #     tl.store(o + blk_o_idx, 1, mask=blk_o_msk)

blksprs/layouting/sparsity_layout.py

@@ -0,0 +1,78 @@
+import torch
+import triton
+from torch import Tensor
+from triton import language as tl
+
+from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.validation import validate_triton_block_size, validate_dimensions, validate_device, \
+    validate_contiguous
+
+
+def build_sparsity_layout(x: Tensor, sparsity_block_size: int, triton_block_size: int = None) -> Tensor:
+    """Builds the sparsity layout of a dense tensor covering its sparse blocks.
+
+    Args:
+        x (Tensor): A block-sparse (or dense) tensor in regular form.
+        sparsity_block_size (int): The size of the sparsity blocks.
+        triton_block_size (int, optional): The block size to use for the triton kernel (default ``None``).
+
+    Returns:
+        Tensor: The sparsity layout of the input block-sparse (or dense) tensor.
+
+    """
+    validate_dimensions(x)
+    validate_contiguous(x)
+    validate_device(x)
+
+    output = torch.zeros(x.size(0), x.size(1) // sparsity_block_size, x.size(2) // sparsity_block_size,
+                         device=x.device, dtype=torch.int32)
+
+    x_b, x_r, x_c = x.size()
+    x_b_s, x_r_s, x_c_s = x.stride()
+    o_b, o_r, o_c = output.size()
+    o_b_s, o_r_s, o_c_s = output.stride()
+
+    if triton_block_size is None:
+        triton_block_size = get_triton_block_size(sparsity_block_size)
+
+    validate_triton_block_size(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_sparsity_layout[triton_grid]
+     (x,
+      x_b, x_b_s, x_r_s, x_c_s,
+      output,
+      o_b, o_b_s, o_r_s, o_c_s,
+      sparsity_block_size,
+      triton_block_size))
+
+    return output
+
+
+@triton.jit
+def kernel_sparsity_layout(x,
+                           x_b, x_b_s, x_r_s, x_c_s,
+                           o,
+                           o_b, o_b_s, o_r_s, o_c_s,
+                           sparsity_block_size,
+                           TRITON_BLOCK_SIZE: tl.constexpr) -> None:
+    # Get triton block indices
+    pid_bat = tl.program_id(axis=0)
+    pid_row = tl.program_id(axis=1)
+    pid_col = tl.program_id(axis=2)
+
+    blk_x_idx = (pid_bat * 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)
+
+    if tl.min(blk_x) != 0 or tl.max(blk_x) != 0:
+        blk_o_idx = (pid_bat * o_b_s +
+                     (((pid_row * TRITON_BLOCK_SIZE) // sparsity_block_size) * o_r_s +
+                      ((pid_col * TRITON_BLOCK_SIZE) // sparsity_block_size) * o_c_s))
+        blk_o_msk = (blk_o_idx < o_b * o_b_s)
+        tl.store(o + blk_o_idx, 1, mask=blk_o_msk)

blksprs/misc/broadcast_addition.py

@@ -0,0 +1,132 @@
+import torch
+import triton
+from torch import Tensor
+from triton import language as tl
+
+from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.validation import validate_contiguous, validate_device, \
+    validate_sparsity_block_size, validate_triton_block_size
+
+
+def broadcast_addition(x: Tensor, y: Tensor, sparsity_layout_output: Tensor,
+                       sparsity_block_size: int, triton_block_size: int = None) -> Tensor:
+    """Performs a broadcast and subsequent addition of two dense tensors x and y. Returns a block-sparse tensor in
+        compressed form.
+
+    Args:
+        x (Tensor): A dense input tensor.
+        y (Tensor): A dense input tensor.
+        sparsity_layout_output (Tensor): The sparsity layout of the output tensor.
+        sparsity_block_size (int): The size of the sparsity blocks.
+        triton_block_size (int, optional): The block size to use for the triton kernel (default ``None``).
+
+    Returns:
+        Tensor: The result of the operation as a block-sparse tensor in compressed form. Each element o(i, j) of the
+            output tensor corresponds to x(i) + y(j).
+
+    """
+    validate_device(x, y)
+    validate_contiguous(x, y)
+    if x.size(-1) != y.size(-1):
+        raise ValueError("Dimensions of tensors must match")
+    validate_sparsity_block_size(sparsity_block_size)
+    validate_triton_block_size(triton_block_size, sparsity_block_size)
+
+    sparsity_lut_o = torch.nonzero(sparsity_layout_output).contiguous()
+
+    n_sparse_blocks = torch.sum(sparsity_layout_output.to(torch.int)).item()
+
+    validate_contiguous(sparsity_layout_output, sparsity_lut_o)
+
+    output = torch.zeros(n_sparse_blocks, sparsity_block_size, sparsity_block_size, device=x.device)
+
+    x_b, x_c = x.size()
+    x_b_s, x_c_s = x.stride()
+    y_b, y_c = y.size()
+    y_b_s, y_c_s = y.stride()
+    o_b, o_r, o_c = output.size()
+    o_b_s, o_r_s, o_c_s = output.stride()
+    s_lut_o_r, s_lut_o_c = sparsity_lut_o.size()
+    s_lut_o_r_s, s_lut_o_c_s = sparsity_lut_o.stride()
+
+    if triton_block_size is None:
+        triton_block_size = get_triton_block_size(sparsity_block_size)
+
+    triton_grid = lambda meta: [o_b,
+                                triton.cdiv(o_r, meta["TRITON_BLOCK_SIZE"]),
+                                triton.cdiv(o_c, meta["TRITON_BLOCK_SIZE"])]
+
+    (kernel_broadcast_addition[triton_grid]
+     (x,
+      x_b, x_b_s, x_c_s,
+      y,
+      y_b, y_b_s, y_c_s,
+      output,
+      o_b, o_b_s, o_r_s, o_c_s,
+      sparsity_lut_o, s_lut_o_r, s_lut_o_r_s, s_lut_o_c_s,
+      sparsity_block_size,
+      triton_block_size))
+
+    return output
+
+
+def broadcast_subtraction(x: Tensor, y: Tensor, sparsity_layout_output: Tensor,
+                          sparsity_block_size: int, triton_block_size: int = None) -> Tensor:
+    """Wrapper for ``broadcast_addition`` with negated y.
+
+    """
+    return broadcast_addition(x, torch.neg(y), sparsity_layout_output, sparsity_block_size, triton_block_size)
+
+
+@triton.jit
+def kernel_broadcast_addition(x,
+                              x_b, x_b_s, x_c_s,
+                              y,
+                              y_b, y_b_s, y_c_s,
+                              o,
+                              o_b, o_b_s, o_r_s, o_c_s,
+                              s_lut_o, s_lut_o_r, s_lut_o_r_s, s_lut_o_c_s,
+                              sparsity_block_size,
+                              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 position of current sparsity block consisting of its batch, row, and column index
+    spa_bat_o_idx = (pid_blk * s_lut_o_r_s + 0 * s_lut_o_c_s)
+    spa_bat_o_msk = (spa_bat_o_idx < s_lut_o_r * s_lut_o_r_s)
+    spa_bat_o = tl.load(s_lut_o + spa_bat_o_idx, mask=spa_bat_o_msk)
+
+    spa_row_o_idx = (pid_blk * s_lut_o_r_s + 1 * s_lut_o_c_s)
+    spa_row_o_msk = (spa_row_o_idx < s_lut_o_r * s_lut_o_r_s)
+    spa_row_o = tl.load(s_lut_o + spa_row_o_idx, mask=spa_row_o_msk)
+
+    spa_col_o_idx = (pid_blk * s_lut_o_r_s + 2 * s_lut_o_c_s)
+    spa_col_o_msk = (spa_col_o_idx < s_lut_o_r * s_lut_o_r_s)
+    spa_col_o = tl.load(s_lut_o + spa_col_o_idx, mask=spa_col_o_msk)
+
+    # Load x block
+    blk_x_idx = (spa_bat_o * x_b_s +
+                 ((spa_row_o * sparsity_block_size + pid_row * 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)
+
+    # Load y block
+    blk_y_idx = (spa_bat_o * y_b_s +
+                 ((spa_col_o * sparsity_block_size + pid_col * TRITON_BLOCK_SIZE +
+                   tl.arange(0, TRITON_BLOCK_SIZE)) * y_c_s)[None, :])
+    blk_y_msk = (blk_y_idx < y_b * y_b_s)
+    blk_y = tl.load(y + blk_y_idx, mask=blk_y_msk)
+
+    # Compute sum
+    blk_x, blk_y = tl.broadcast(tl.trans(blk_x), blk_y)
+    buf = blk_x + blk_y
+
+    # Store result
+    blk_o_idx = ((pid_blk * 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, buf, mask=blk_o_msk)

blksprs/ops/conversion.py

@@ -0,0 +1,256 @@
+import torch
+import triton
+from torch import Tensor
+from triton import language as tl
+
+from blksprs.utils.tools import get_triton_block_size
+from blksprs.utils.validation import validate_contiguous, validate_dimensions, validate_device, \
+    validate_sparsity, validate_sparsity_block_size, validate_triton_block_size
+
+
+def to_dense(x: Tensor, sparsity_layout: Tensor, sparsity_block_size: int, fill_value: float = 0,
+             triton_block_size: int = None) -> Tensor:
+    """Converts a block-sparse tensor in compressed form to a block-sparse tensor in regular form based on the given
+        sparsity layout.
+
+    Args:
+        x (Tensor): A block-sparse tensor in compressed form.
+        sparsity_layout (Tensor): The sparsity layout of the block-sparse tensor.
+        sparsity_block_size (int): The size of the sparsity blocks.
+        fill_value (float): The value to fill the resulting dense tensor with where the block-sparse tensor is not
+            present (default ``0``).
+        triton_block_size (int): The block size to use for the triton kernel (default ``None``).
+
+    Returns:
+        Tensor: The block-sparse tensor converted to regular form.
+
+    """
+    validate_dimensions(x)
+    validate_contiguous(x, sparsity_layout)
+    validate_device(x)
+    validate_sparsity(sparsity_block_size, (x, sparsity_layout))
+    validate_sparsity_block_size(sparsity_block_size, x)
+    validate_triton_block_size(triton_block_size, sparsity_block_size)
+
+    sparsity_layout_flat = sparsity_layout.reshape(-1)
+    sparsity_reverse_lut = ((torch.cumsum(sparsity_layout_flat, dim=-1) - 1) *
+                            (sparsity_layout_flat == 1) -
+                            (1 * (sparsity_layout_flat == 0)))
+
+    validate_contiguous(sparsity_reverse_lut)
+
+    return _BlocksparseToDense.apply(x,
+                                     sparsity_layout, sparsity_reverse_lut,
+                                     sparsity_block_size, fill_value,
+                                     triton_block_size)
+
+
+class _BlocksparseToDense(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, x: Tensor,
+                sparsity_layout: Tensor, sparsity_reverse_lut: Tensor,
+                sparsity_block_size: int, fill_value: float,
+                triton_block_size: int) -> Tensor:
+        output = torch.full(size=(sparsity_layout.size(0), sparsity_layout.size(1) * sparsity_block_size,
+                                  sparsity_layout.size(2) * sparsity_block_size), fill_value=fill_value,
+                            dtype=x.dtype, device=x.device)
+
+        x_b, x_r, x_c = x.shape
+        x_b_s, x_r_s, x_c_s = x.stride()
+        s_l_b, s_l_r, s_l_c = sparsity_layout.size()
+        s_l_b_s, s_l_r_s, s_l_c_s = sparsity_layout.stride()
+        o_b, o_r, o_c = output.size()
+        o_b_s, o_r_s, o_c_s = output.stride()
+
+        if triton_block_size is None:
+            triton_block_size = get_triton_block_size(sparsity_block_size)
+
+        triton_grid = lambda meta: [o_b,
+                                    triton.cdiv(o_r, meta["TRITON_BLOCK_SIZE"]),
+                                    triton.cdiv(o_c, meta["TRITON_BLOCK_SIZE"])]
+
+        (_BlocksparseToDense.kernel_blocksparse_to_dense[triton_grid]
+         (x,
+          x_b, x_b_s, x_r_s, x_c_s,
+          s_l_b, s_l_b_s, s_l_r_s, s_l_c_s,
+          sparsity_reverse_lut,
+          output,
+          o_b, o_b_s, o_r_s, o_c_s,
+          sparsity_block_size,
+          triton_block_size))
+
+        ctx.save_for_backward(sparsity_layout)
+        ctx.sparsity_block_size = sparsity_block_size
+        ctx.triton_block_size = triton_block_size
+
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        sparsity_layout = ctx.saved_tensors[0]
+        sparsity_block_size = ctx.sparsity_block_size
+        triton_block_size = ctx.triton_block_size
+
+        return to_sparse(grad_output, sparsity_layout, sparsity_block_size,
+                         triton_block_size), None, None, None, None, None
+
+    @staticmethod
+    @triton.jit
+    def kernel_blocksparse_to_dense(x,
+                                    x_b, x_b_s, x_r_s, x_c_s,
+                                    s_l_b, s_l_b_s, s_l_r_s, s_l_c_s,
+                                    sparsity_reverse_lut,
+                                    o,
+                                    o_b, o_b_s, o_r_s, o_c_s,
+                                    sparsity_block_size,
+                                    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 block
+        spa_row = (pid_row * TRITON_BLOCK_SIZE) // sparsity_block_size
+        spa_col = (pid_col * TRITON_BLOCK_SIZE) // sparsity_block_size
+
+        # Get reverse sparsity index for current block
+        rev_idx_spa_idx = (pid_blk * s_l_b_s + spa_row * s_l_r_s + spa_col * s_l_c_s)
+        rev_idx_spa_msk = (rev_idx_spa_idx < s_l_b * s_l_b_s)
+        rev_idx_spa = tl.load(sparsity_reverse_lut + rev_idx_spa_idx, mask=rev_idx_spa_msk).to(tl.int32)
+
+        # If block is present commence operations
+        if rev_idx_spa >= 0:
+            blk_idx = (rev_idx_spa * x_b_s +
+                       (((pid_row % (sparsity_block_size // TRITON_BLOCK_SIZE)) * TRITON_BLOCK_SIZE +
+                         tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
+                       (((pid_col % (sparsity_block_size // TRITON_BLOCK_SIZE)) * TRITON_BLOCK_SIZE +
+                         tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
+            blk_msk = (blk_idx < x_b * x_b_s)
+            blk = tl.load(x + blk_idx, mask=blk_msk)
+
+            o_idx = (pid_blk * 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, :])
+            o_msk = (o_idx < o_b * o_b_s)
+            tl.store(o + o_idx, blk, o_msk)
+
+
+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.
+
+        Args:
+        x (Tensor): A block-sparse tensor in regular form.
+        sparsity_layout (Tensor): The sparsity layout of the block-sparse tensor.
+        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: The block-sparse tensor converted to compressed form.
+
+    """
+    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_lut = torch.nonzero(sparsity_layout).contiguous()
+    n_sparse_blocks = torch.sum(sparsity_layout.to(torch.int)).item()
+
+    validate_contiguous(sparsity_layout, sparsity_lut)
+
+    return _BlocksparseToSparse.apply(x,
+                                      sparsity_layout, sparsity_lut,
+                                      sparsity_block_size, n_sparse_blocks,
+                                      triton_block_size)
+
+
+class _BlocksparseToSparse(torch.autograd.Function):
+
+    @staticmethod
+    def forward(ctx, x: Tensor,
+                sparsity_layout: Tensor, sparsity_lut: Tensor,
+                sparsity_block_size: int, n_sparse_blocks: int, triton_block_size: int) -> Tensor:
+        output = torch.empty(size=(n_sparse_blocks, 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 = x.stride()
+        o_b, o_r, o_c = output.size()
+        o_b_s, o_r_s, o_c_s = output.stride()
+        s_lut_r, s_lut_c = sparsity_lut.size()
+        s_lut_r_s, s_lut_c_s = sparsity_lut.stride()
+
+        if triton_block_size is None:
+            triton_block_size = get_triton_block_size(sparsity_block_size)
+
+        triton_grid = lambda meta: [o_b,
+                                    triton.cdiv(o_r, meta["TRITON_BLOCK_SIZE"]),
+                                    triton.cdiv(o_c, meta["TRITON_BLOCK_SIZE"])]
+
+        (_BlocksparseToSparse.kernel_blocksparse_to_sparse[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_s, o_r_s, o_c_s,
+          sparsity_block_size,
+          triton_block_size))
+
+        ctx.save_for_backward(sparsity_layout)
+        ctx.sparsity_block_size = sparsity_block_size
+        ctx.triton_block_size = triton_block_size
+
+        return output
+
+    @staticmethod
+    def backward(ctx, grad_output):
+        sparsity_layout = ctx.saved_tensors[0]
+        sparsity_block_size = ctx.sparsity_block_size
+        triton_block_size = ctx.triton_block_size
+
+        return to_dense(grad_output, sparsity_layout, sparsity_block_size,
+                        triton_block_size=triton_block_size), None, None, None, None, None
+
+    @staticmethod
+    @triton.jit
+    def kernel_blocksparse_to_sparse(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_s, o_r_s, o_c_s,
+                                     sparsity_block_size,
+                                     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 from dense tensor
+        blk_d_idx = (spa_bat * x_b_s +
+                     ((spa_row * sparsity_block_size + pid_row * TRITON_BLOCK_SIZE +
+                       tl.arange(0, TRITON_BLOCK_SIZE)) * x_r_s)[:, None] +
+                     ((spa_col * sparsity_block_size + pid_col * TRITON_BLOCK_SIZE +
+                       tl.arange(0, TRITON_BLOCK_SIZE)) * x_c_s)[None, :])
+        blk_d_msk = (blk_d_idx < x_b * x_b_s)
+        blk_d = tl.load(x + blk_d_idx, mask=blk_d_msk)
+
+        # Store block in sparse tensor
+        blk_o_idx = ((pid_blk * 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 < (pid_blk + 1) * o_b_s)
+        tl.store(o + blk_o_idx, blk_d, mask=blk_o_msk)