[REQ] Add support for tile slicing syntax, including fancy indexing

Feature Request: NumPy-Style Slicing for Warp Tiles

Problem

Warp tiles currently lack slicing syntax support. While tile_view() exists, it requires explicit offset and shape parameters rather than the familiar NumPy slicing notation that users expect.

Current approach with tile_view():

@wp.kernel
def process_tile(data: wp.array2d(dtype=float)):
    t = wp.tile_load(data, shape=(64, 64))
    
    # Extract rows 10-20
    subset = wp.tile_view(t, offset=(10, 0), shape=(10, 64))
    
    # Extract a single row — requires knowing the full shape
    row = wp.tile_view(t, offset=(5, 0), shape=(1, 64))

What we want:

@wp.kernel
def process_tile(data: wp.array2d(dtype=float)):
    t = wp.tile_load(data, shape=(64, 64))
    
    # NumPy-style slicing
    subset = t[10:20, :]
    row = t[5, :]
    cols = t[:, 0:32]
    strided = t[::2, ::2]

    # Slice assignment
    t[0:8, :] = wp.tile_ones(shape=(8, 64), dtype=float)

Additionally, there's no support for index array slicing (fancy indexing), which is essential for gathering non-contiguous elements:

@wp.kernel
def gather_elements(data: wp.array2d(dtype=float)):
    t = wp.tile_load(data, shape=(64, 64))
    
    # Gather specific rows by index
    indices = wp.tile_arange(0, 8, dtype=int) * 8  # [0, 8, 16, 24, 32, 40, 48, 56]
    selected_rows = t[indices, :]  # Not currently possible

Implementation Notes

Relationship to Existing APIs

tile_view(t, offset, shape) — Should remain as the explicit API; basic slice syntax (e.g., t[0:8, :]) compiles down to tile_view calls
tile_slice_indexed (new primitive) — Index array slicing (e.g., t[indices, :]) would require a new gather operation
tile_extract(t, i, j, ...) — Unchanged; extracts scalar elements
t[i] — Existing integer indexing unchanged; reduces dimensionality

Compile-Time vs Runtime

Slice bounds should be compile-time constants for optimal code generation, consistent with existing tile shape requirements. The compiler can verify bounds at compile time when constants are used.

Add NumPy-style tile slicing and fancy indexing (NVIDIAGH-1176)

Fix correctness issues found in tile slicing review (NVIDIAGH-1176)

Address tile slicing review (NVIDIAGH-1176)

Address GitHub review feedback (NVIDIAGH-1176)

Add NumPy-style tile slicing and fancy indexing (NVIDIAGH-1176)

Fix correctness issues found in tile slicing review (NVIDIAGH-1176)

Address review feedback on tile slicing tests (NVIDIAGH-1176)

Promote tile indexing bounds warnings to caution directives (NVIDIAGH…

Address review feedback on tile slicing (NVIDIAGH-1176)

Fix chained tile row gather (NVIDIAGH-1176)

Address tile slicing review (NVIDIAGH-1176)

Address GitHub review feedback (NVIDIAGH-1176)

Guard tile type probe symbol copy (NVIDIAGH-1176)

Add NumPy-style tile slicing and fancy indexing (NVIDIAGH-1176)

This is addressed by c07653a.

The change adds NumPy-style tile indexing in kernel code. Basic, strided, and reversed slices and dimension-collapsing integer indices now lower to tile views; slice assignment is supported; and 1D integer index tiles gather through the new differentiable wp.tile_slice_indexed() primitive. Existing explicit wp.tile_view() and scalar wp.tile_extract() behavior remains available.

Public API behavior:

Kernel scope:

@wp.kernel
def process_tile(data: wp.array2d[float]):
    t = wp.tile_load(data, shape=(64, 64))

    block = t[2:6, :]           # Select rows 2 through 5
    columns = t[:, ::2]         # Select every other column
    reversed_rows = t[::-1, :]  # Reverse the row order
    row = t[5, :]               # Select one row and collapse that dimension

    # Gather non-contiguous rows using a 1D integer index tile
    indices = wp.tile_arange(0, 8, dtype=int) * 8
    selected_rows = t[indices, :]

    # Assign a tile to a view in place
    t[0:8, :] = wp.tile_ones(shape=(8, 64), dtype=float)

Slice bounds must be compile-time constants. Runtime integer indices are supported, including negative indices, but out-of-range runtime values are checked only in debug builds.

Advanced-index assignment, compound assignment to views, scalar broadcasting into slices, and zero-length tile views remain unsupported.

Test coverage:

warp/tests/tile/test_tile_view.py adds 39 tests covering basic, strided, reversed, dimension-collapsing, and assigned slices, including test_tile_slice_rows, test_tile_slice_strided, test_tile_slice_reverse, and test_tile_slice_assign.
Advanced indexing and autograd coverage includes test_tile_advanced_index, test_tile_advanced_index_duplicate_grad, test_tile_slice_assign_grad, and test_tile_gather_neg_index.
Runtime and negative indices, chained expressions, explicit slice offsets, overlapping assignment, bounds validation, and unsupported inputs are also covered.

An independent probe matching the original 64×64 examples passed on CPU and CUDA. No remaining requirement from this issue was identified.

Add NumPy-style tile slicing and fancy indexing (NVIDIAGH-1176)

Navigation Menu

Repository navigation

[REQ] Add support for tile slicing syntax, including fancy indexing #1176

Feature Request: NumPy-Style Slicing for Warp Tiles

Problem

Implementation Notes

Relationship to Existing APIs

Compile-Time vs Runtime

4 remaining items

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

[REQ] Add support for tile slicing syntax, including fancy indexing #1176

Description

Feature Request: NumPy-Style Slicing for Warp Tiles

Problem

Implementation Notes

Relationship to Existing APIs

Compile-Time vs Runtime

Activity

4 remaining items

shi-eric commented on Jul 23, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions