wp.tile_dot() fails to compile for scalar wp.float64 tiles.

[daedalus5]

Bug Description

wp.tile_dot() fails to compile for scalar wp.float64 tiles.

Repro

import numpy as np
import warp as wp

_TILE = 64


@wp.kernel(enable_backward=False)
def dot_tiled_f64(
    a: wp.array(dtype=wp.float64),
    b: wp.array(dtype=wp.float64),
    partials: wp.array(dtype=wp.float64),
):
    i = wp.tid()
    offset = i * _TILE

    at = wp.tile_load(a, (_TILE,), offset=(offset,))
    bt = wp.tile_load(b, (_TILE,), offset=(offset,))

    d = wp.tile_dot(at, bt)
    partials[i] = d[0]


wp.init()

device = "cuda:0"
n_tiles = 2
n = n_tiles * _TILE

a = wp.array(np.arange(n, dtype=np.float64), dtype=wp.float64, device=device)
b = wp.array(np.arange(n, dtype=np.float64), dtype=wp.float64, device=device)
partials = wp.zeros(n_tiles, dtype=wp.float64, device=device)

wp.launch_tiled(
    dot_tiled_f64,
    dim=[n_tiles],
    inputs=[a, b, partials],
    block_dim=_TILE,
    device=device,
)

Actual Result

CUDA compilation fails:

error: no operator "=" matches these operands
operand types are:
wp::tile_shared_t<wp::float64, ...> = wp::tile_register_t<float, ...>

var_9 = wp::tile_dot(var_5, var_8);

CPU compilation fails similarly with:

error: no viable overloaded '='

Expected Result

wp.tile_dot() should support scalar wp.float64 tiles and return a single-element wp.float64 tile.

Notes

The wp.float32 version of the same kernel compiles and runs correctly on both CPU and CUDA.

The likely issue is a mismatch between Python type inference and the native implementation:

• Python infers the result type as type_scalar_type(a.dtype), so float64 inputs expect a float64 result.
• Native tile_dot() computes its result type with decltype(tensordot(T{}, T{})).
• Scalar tensordot() appears to only have a float overload, so scalar wp::float64 operands resolve to a float result tile.

System Information

No response

[shi-eric]

This is addressed by ebcce32.

The native scalar tensordot() path used by wp.tile_dot() now preserves the operand type, so the generated result tile matches Warp's inferred type instead of narrowing wp.float64 or wp.float16 to float.

Public API behavior:

Kernel scope:

• Existing wp.tile_dot() now compiles for scalar wp.float64 and wp.float16 tiles on CPU and CUDA and returns a one-element tile with the same scalar type.
• The API remains kernel-only and must be called from @wp.kernel or @wp.func.

Test coverage:

• New warp/tests/tile/test_tile_fused_ops.py::test_tile_dot_float, registered on TestTileFusedOps, checks forward values and gradients for wp.float32, wp.float64, and wp.float16 across CPU and CUDA.

The original two-tile float64 reproducer was also verified on CPU and CUDA, including the expected partial values and float64 output.