How to convert (N,3) numpy array into (N) warp vec3?

[cadop]

I am trying to convert a numpy array of integers into a warp array, where the second dim of numpy is captured by the warp dtype.

    faces = mesh.indices.reshape((-1, 3))
    faces = wp.array(faces, shape=(N), dtype=wp.vec3i, device=device)

The kernel signature is:

@wp.kernel
def _matrix(points: wp.array(dtype=wp.vec3),
            faces: wp.array(dtype=wp.vec3i),
            mesh: wp.uint64):
     f = wp.tid()
     face = faces[f]

This was seemingly working when I was in device='cpu' , however changing device to cuda results in:
RuntimeError: Error launching kernel '_matrix', argument 'faces' expects an array with dtype=<class 'warp.types.vec3i'> but passed array has dtype=<class 'warp.types.int32'>.

[cadop]

As a sanity check, this does work:

    faces = mesh.indices.reshape((-1, 3))
    faces = wp.array(faces, shape=(N,3), dtype=wp.int32, device=device)

@wp.kernel
def _matrix(points: wp.array(dtype=wp.vec3),
            faces: wp.array(dtype=wp.int32, ndim=2),
            mesh: wp.uint64):
     f = wp.tid()
     face = faces[f]

[christophercrouzet]

Hi @cadop,

Would you be able to share which version of Warp you've been using and sharing a runnable snippet showcasing the issue?

I didn't manage to repro this issue on my end using the following snippet:

import numpy as np
import warp as wp


POINTS = (
    ( 0.853553, -0.146446,  0.0     ), ( 0.146446, -0.853553,  0.0     ),
    ( 0.353553,  0.353553,  0.707106), (-0.353553, -0.353553,  0.707106),
    (-0.353553, -0.353553, -0.707106), ( 0.353553,  0.353553, -0.707106),
    (-0.853553,  0.146446,  0.0     ), (-0.146446,  0.853553,  0.0     ),
)

FACE_VERTEX_INDICES = (
    0, 3, 1, 0, 2, 3, 4, 7, 5, 4, 6, 7, 6, 2, 7, 6, 3, 2,
    5, 1, 4, 5, 0, 1, 5, 2, 0, 5, 7, 2, 1, 6, 4, 1, 3, 6,
)


@wp.kernel
def _matrix(
    points: wp.array(dtype=wp.vec3),
    faces: wp.array(dtype=wp.vec3i),
):
    f = wp.tid()
    face = faces[f]
    wp.printf("face=(%d, %d, %d)\n", face[0], face[1], face[2])


def run():
    points = wp.array(POINTS, dtype=wp.vec3)

    faces = np.array(FACE_VERTEX_INDICES).reshape((-1, 3))
    faces = wp.array(faces, shape=faces.shape[0], dtype=wp.vec3i)

    wp.launch(_matrix, dim=faces.shape, inputs=(points, faces))
    wp.synchronize()


if __name__ == "__main__":
    run()

Note that these 2 lines:

faces = np.array(FACE_VERTEX_INDICES).reshape((-1, 3))
faces = wp.array(faces, shape=faces.shape[0], dtype=wp.vec3i)

Could be replaced by single call:

faces = wp.array(FACE_VERTEX_INDICES, dtype=wp.vec3i)

[cadop]

Hi Christopher, sorry for the missing context. I opened a discussion instead of an issue because I assumed I just didn't do something right rather than saying it was a bug.

I think the issue I have is because I didn't realize I wasn't using numpy, my array was converted to warp at some point. I was trying to convert a warp array of dtype int to vec3i. I guess my question would be how to do that, or if view works similar to numpy and can be passed to kernels (I wasn't able to view it though).

import numpy as np
import warp as wp

device = wp.get_device("cuda:1")

POINTS = (
    ( 0.853553, -0.146446,  0.0     ), ( 0.146446, -0.853553,  0.0     ),
    ( 0.353553,  0.353553,  0.707106), (-0.353553, -0.353553,  0.707106),
    (-0.353553, -0.353553, -0.707106), ( 0.353553,  0.353553, -0.707106),
    (-0.853553,  0.146446,  0.0     ), (-0.146446,  0.853553,  0.0     ),
)

FACE_VERTEX_INDICES = (
    0, 3, 1, 0, 2, 3, 4, 7, 5, 4, 6, 7, 6, 2, 7, 6, 3, 2,
    5, 1, 4, 5, 0, 1, 5, 2, 0, 5, 7, 2, 1, 6, 4, 1, 3, 6,
)


@wp.kernel
def _matrix(
    points: wp.array(dtype=wp.vec3),
    faces: wp.array(dtype=wp.vec3i),
):
    f = wp.tid()
    face = faces[f]
    wp.printf("face=(%d, %d, %d)\n", face[0], face[1], face[2])


def run():
    points = wp.array(POINTS, dtype=wp.vec3)

    faces = wp.array(FACE_VERTEX_INDICES).reshape((-1, 3))
    faces = wp.array(faces, shape=faces.shape[0], dtype=wp.vec3i)

    wp.launch(_matrix, dim=faces.shape, inputs=(points, faces))
    wp.synchronize()


if __name__ == "__main__":
    run()

[christophercrouzet]

Hi @cadop, thanks for providing this snippet, I've been able to repro the issue and your approach should be valid, so this does look like a bug. Would you be able to open an issue so that we can track it appropriately?

Regarding the view() method, if you look at its implementation, it's basically a shortcut to calling wp.array() while ensuring that no copy is being made, so yes, these array views can be passed to kernels. However, as you've noticed, it's failing here since it expects the new dtype to have the same width as the old one, so directly calling wp.array() as you've done it brings more flexibility here.