16620437478

Committed 30 Jul 2025 10:46AM UTC coverage: 82.754% (+0.06%) from 82.696%

Build # 16620437478

Build Type

push

github

Committed by

web-flow

Commit Message

varbinview zip kernel (#4054)

Uses the buffer deduplicating builder to construct the zipped array.
This guards against the pathological case where we are zipping two
varbinview arrays with a mask that has lots of contiguous slices. Each
`builder.extend_from_array(input.slice(..))` would duplicate the entire
buffers of `input`, and each slice in the mask would add the same
buffers to the result array over and over again.

---------

Signed-off-by: Onur Satici <onur@spiraldb.com>

Run Details

79 of 82 new or added lines in 2 files covered. (96.34%)

2 existing lines in 1 file now uncovered.

45322 of 54767 relevant lines covered (82.75%)

184684.33 hits per line

Source File
Press 'n' to go to next uncovered line, 'b' for previous

91.89

/vortex-array/src/compute/zip.rs

// SPDX-License-Identifier: Apache-2.0
// SPDX-FileCopyrightText: Copyright the Vortex contributors

use std::sync::LazyLock;

use arcref::ArcRef;
use vortex_dtype::DType;
use vortex_error::{VortexError, VortexResult, vortex_bail, vortex_err};
use vortex_mask::{AllOr, Mask};

use super::{ComputeFnVTable, InvocationArgs, Output, cast};
use crate::builders::{ArrayBuilder, builder_with_capacity};
use crate::compute::{ComputeFn, Kernel};
use crate::vtable::VTable;
use crate::{Array, ArrayRef};

/// Performs element-wise conditional selection between two arrays based on a mask.
///
/// Returns a new array where `result[i] = if_true[i]` when `mask[i]` is true,
/// otherwise `result[i] = if_false[i]`.
pub fn zip(if_true: &dyn Array, if_false: &dyn Array, mask: &Mask) -> VortexResult<ArrayRef> {
    ZIP_FN
        .invoke(&InvocationArgs {
            inputs: &[if_true.into(), if_false.into(), mask.into()],
            options: &(),
        })?
        .unwrap_array()
}

pub static ZIP_FN: LazyLock<ComputeFn> = LazyLock::new(|| {
    let compute = ComputeFn::new("zip".into(), ArcRef::new_ref(&Zip));
    for kernel in inventory::iter::<ZipKernelRef> {
        compute.register_kernel(kernel.0.clone());
    }
    compute
});

struct Zip;

impl ComputeFnVTable for Zip {
    fn invoke(
        &self,
        args: &InvocationArgs,
        kernels: &[ArcRef<dyn Kernel>],
    ) -> VortexResult<Output> {
        let ZipArgs {
            if_true,
            if_false,
            mask,
        } = ZipArgs::try_from(args)?;

        if mask.all_true() {
            return Ok(cast(if_true, &zip_return_dtype(if_true, if_false))?.into());
        }

        if mask.all_false() {
            return Ok(cast(if_false, &zip_return_dtype(if_true, if_false))?.into());
        }

        // check if if_true supports zip directly
        for kernel in kernels {
            if let Some(output) = kernel.invoke(args)? {
                return Ok(output);
            }
        }

        if let Some(output) = if_true.invoke(&ZIP_FN, args)? {
            return Ok(output);
        }

        // TODO(os): add invert_mask opt and check if if_false has a kernel like:
        //           kernel.invoke(Args(if_false, if_true, mask, invert_mask = true))

        Ok(zip_impl(
            if_true.to_canonical()?.as_ref(),
            if_false.to_canonical()?.as_ref(),
            mask,
        )?
        .into())
    }

    fn return_dtype(&self, args: &InvocationArgs) -> VortexResult<DType> {
        let ZipArgs {
            if_true, if_false, ..
        } = ZipArgs::try_from(args)?;

        if !if_true.dtype().eq_ignore_nullability(if_false.dtype()) {
            vortex_bail!("input arrays to zip must have the same dtype");
        }
        Ok(zip_return_dtype(if_true, if_false))
    }

    fn return_len(&self, args: &InvocationArgs) -> VortexResult<usize> {
        let ZipArgs { if_true, mask, .. } = ZipArgs::try_from(args)?;
        // ComputeFn::invoke asserts if_true.len() == if_false.len(), because zip is elementwise
        if if_true.len() != mask.len() {
            vortex_bail!("input arrays must have the same length as the mask");
        }
        Ok(if_true.len())
    }

    fn is_elementwise(&self) -> bool {
        true
    }
}

struct ZipArgs<'a> {
    if_true: &'a dyn Array,
    if_false: &'a dyn Array,
    mask: &'a Mask,
}

impl<'a> TryFrom<&InvocationArgs<'a>> for ZipArgs<'a> {
    type Error = VortexError;

    fn try_from(value: &InvocationArgs<'a>) -> Result<Self, Self::Error> {
        if value.inputs.len() != 3 {
            vortex_bail!("Expected 3 inputs for zip, found {}", value.inputs.len());
        }
        let if_true = value.inputs[0]
            .array()
            .ok_or_else(|| vortex_err!("Expected input 0 to be an array"))?;

        let if_false = value.inputs[1]
            .array()
            .ok_or_else(|| vortex_err!("Expected input 1 to be an array"))?;

        let mask = value.inputs[2]
            .mask()
            .ok_or_else(|| vortex_err!("Expected input 2 to be a mask"))?;

        Ok(Self {
            if_true,
            if_false,
            mask,
        })
    }
}

pub trait ZipKernel: VTable {
    fn zip(
        &self,
        if_true: &Self::Array,
        if_false: &dyn Array,
        mask: &Mask,
    ) -> VortexResult<Option<ArrayRef>>;
}

pub struct ZipKernelRef(pub ArcRef<dyn Kernel>);
inventory::collect!(ZipKernelRef);

#[derive(Debug)]
pub struct ZipKernelAdapter<V: VTable>(pub V);

impl<V: VTable + ZipKernel> ZipKernelAdapter<V> {
    pub const fn lift(&'static self) -> ZipKernelRef {
        ZipKernelRef(ArcRef::new_ref(self))
    }
}

impl<V: VTable + ZipKernel> Kernel for ZipKernelAdapter<V> {
    fn invoke(&self, args: &InvocationArgs) -> VortexResult<Option<Output>> {
        let ZipArgs {
            if_true,
            if_false,
            mask,
        } = ZipArgs::try_from(args)?;
        let Some(if_true) = if_true.as_opt::<V>() else {
            return Ok(None);
        };
        Ok(V::zip(&self.0, if_true, if_false, mask)?.map(Into::into))
    }
}

pub(crate) fn zip_return_dtype(if_true: &dyn Array, if_false: &dyn Array) -> DType {
    if_true
        .dtype()
        .union_nullability(if_false.dtype().nullability())
}

fn zip_impl(if_true: &dyn Array, if_false: &dyn Array, mask: &Mask) -> VortexResult<ArrayRef> {
    // if_true.len() == if_false.len() from ComputeFn::invoke
    let builder = builder_with_capacity(&zip_return_dtype(if_true, if_false), if_true.len());
    zip_impl_with_builder(if_true, if_false, mask, builder)
}

pub(crate) fn zip_impl_with_builder(
    if_true: &dyn Array,
    if_false: &dyn Array,
    mask: &Mask,
    mut builder: Box<dyn ArrayBuilder>,
) -> VortexResult<ArrayRef> {
    match mask.slices() {
        AllOr::All => Ok(if_true.to_array()),
        AllOr::None => Ok(if_false.to_array()),
        AllOr::Some(slices) => {
            for (start, end) in slices {
                builder.extend_from_array(&if_false.slice(builder.len(), *start)?)?;
                builder.extend_from_array(&if_true.slice(*start, *end)?)?;
            }
            if builder.len() < if_false.len() {
                builder.extend_from_array(&if_false.slice(builder.len(), if_false.len())?)?;
            }
            Ok(builder.finish())
        }
    }
}

#[cfg(test)]
mod tests {
    use vortex_array::arrays::{BoolArray, PrimitiveArray};
    use vortex_array::compute::zip;
    use vortex_array::{IntoArray, ToCanonical};
    use vortex_mask::Mask;

    #[test]
    fn test_zip_basic() {
        let mask =
            Mask::try_from(&BoolArray::from_iter([true, false, false, true, false])).unwrap();
        let if_true = PrimitiveArray::from_iter([10, 20, 30, 40, 50]).into_array();
        let if_false = PrimitiveArray::from_iter([1, 2, 3, 4, 5]).into_array();

        let result = zip(&if_true, &if_false, &mask).unwrap();
        let expected = PrimitiveArray::from_iter([10, 2, 3, 40, 5]);

        assert_eq!(
            result.to_primitive().unwrap().as_slice::<i32>(),
            expected.as_slice::<i32>()
        );
    }

    #[test]
    fn test_zip_all_true() {
        let mask = Mask::new_true(4);
        let if_true = PrimitiveArray::from_iter([10, 20, 30, 40]).into_array();
        let if_false =
            PrimitiveArray::from_option_iter([Some(1), Some(2), Some(3), None]).into_array();

        let result = zip(&if_true, &if_false, &mask).unwrap();

        assert_eq!(
            result.to_primitive().unwrap().as_slice::<i32>(),
            if_true.to_primitive().unwrap().as_slice::<i32>()
        );

        // result must be nullable even if_true was not
        assert_eq!(result.dtype(), if_false.dtype())
    }

    #[test]
    #[should_panic]
    fn test_invalid_lengths() {
        let mask = Mask::new_false(4);
        let if_true = PrimitiveArray::from_iter([10, 20, 30]).into_array();
        let if_false = PrimitiveArray::from_iter([1, 2, 3, 4]).into_array();

        zip(&if_true, &if_false, &mask).unwrap();
    }
}

1	// SPDX-License-Identifier: Apache-2.0
2	// SPDX-FileCopyrightText: Copyright the Vortex contributors
3
4	use std::sync::LazyLock;
5
6	use arcref::ArcRef;
7	use vortex_dtype::DType;
8	use vortex_error::{VortexError, VortexResult, vortex_bail, vortex_err};
9	use vortex_mask::{AllOr, Mask};
10
11	use super::{ComputeFnVTable, InvocationArgs, Output, cast};
12	use crate::builders::{ArrayBuilder, builder_with_capacity};
13	use crate::compute::{ComputeFn, Kernel};
14	use crate::vtable::VTable;
15	use crate::{Array, ArrayRef};
16
17	/// Performs element-wise conditional selection between two arrays based on a mask.
18	///
19	/// Returns a new array where `result[i] = if_true[i]` when `mask[i]` is true,
20	/// otherwise `result[i] = if_false[i]`.
21	pub fn zip(if_true: &dyn Array, if_false: &dyn Array, mask: &Mask) -> VortexResult<ArrayRef> {	115✔
22	ZIP_FN	115✔
23	.invoke(&InvocationArgs {	115✔
24	inputs: &[if_true.into(), if_false.into(), mask.into()],	115✔
25	options: &(),	115✔
26	})?	115✔
27	.unwrap_array()	77✔
28	}	115✔
29
30	pub static ZIP_FN: LazyLock<ComputeFn> = LazyLock::new(\|\| {	115✔
31	let compute = ComputeFn::new("zip".into(), ArcRef::new_ref(&Zip));	115✔
32	for kernel in inventory::iter::<ZipKernelRef> {	230✔
33	compute.register_kernel(kernel.0.clone());	115✔
34	}	115✔
35	compute	115✔
36	});	115✔
37
38	struct Zip;
39
40	impl ComputeFnVTable for Zip {
41	fn invoke(	77✔
42	&self,	77✔
43	args: &InvocationArgs,	77✔
44	kernels: &[ArcRef<dyn Kernel>],	77✔
45	) -> VortexResult<Output> {	77✔
46	let ZipArgs {
47	if_true,	77✔
48	if_false,	77✔
49	mask,	77✔
50	} = ZipArgs::try_from(args)?;	77✔
51
52	if mask.all_true() {	77✔
53	return Ok(cast(if_true, &zip_return_dtype(if_true, if_false))?.into());	38✔
54	}	39✔
55
56	if mask.all_false() {	39✔
57	return Ok(cast(if_false, &zip_return_dtype(if_true, if_false))?.into());	×
58	}	39✔
59
60	// check if if_true supports zip directly
61	for kernel in kernels {	77✔
62	if let Some(output) = kernel.invoke(args)? {	39✔
63	return Ok(output);	1✔
64	}	38✔
65	}
66
67	if let Some(output) = if_true.invoke(&ZIP_FN, args)? {	38✔
68	return Ok(output);	×
69	}	38✔
70
71	// TODO(os): add invert_mask opt and check if if_false has a kernel like:
72	// kernel.invoke(Args(if_false, if_true, mask, invert_mask = true))
73
74	Ok(zip_impl(	38✔
75	if_true.to_canonical()?.as_ref(),	38✔
76	if_false.to_canonical()?.as_ref(),	38✔
77	mask,	38✔
78	)?	×
79	.into())	38✔
80	}	77✔
81
82	fn return_dtype(&self, args: &InvocationArgs) -> VortexResult<DType> {	77✔
83	let ZipArgs {
84	if_true, if_false, ..	77✔
85	} = ZipArgs::try_from(args)?;	77✔
86
87	if !if_true.dtype().eq_ignore_nullability(if_false.dtype()) {	77✔
88	vortex_bail!("input arrays to zip must have the same dtype");	×
89	}	77✔
90	Ok(zip_return_dtype(if_true, if_false))	77✔
91	}	77✔
92
93	fn return_len(&self, args: &InvocationArgs) -> VortexResult<usize> {	77✔
94	let ZipArgs { if_true, mask, .. } = ZipArgs::try_from(args)?;	77✔
95	// ComputeFn::invoke asserts if_true.len() == if_false.len(), because zip is elementwise
96	if if_true.len() != mask.len() {	77✔
NEW 97	vortex_bail!("input arrays must have the same length as the mask");	×
98	}	77✔
99	Ok(if_true.len())	77✔
100	}	77✔
101
102	fn is_elementwise(&self) -> bool {	115✔
103	true	115✔
104	}	115✔
105	}
106
107	struct ZipArgs<'a> {
108	if_true: &'a dyn Array,
109	if_false: &'a dyn Array,
110	mask: &'a Mask,
111	}
112
113	impl<'a> TryFrom<&InvocationArgs<'a>> for ZipArgs<'a> {
114	type Error = VortexError;
115
116	fn try_from(value: &InvocationArgs<'a>) -> Result<Self, Self::Error> {	270✔
117	if value.inputs.len() != 3 {	270✔
118	vortex_bail!("Expected 3 inputs for zip, found {}", value.inputs.len());	×
119	}	270✔
120	let if_true = value.inputs[0]	270✔
121	.array()	270✔
122	.ok_or_else(\|\| vortex_err!("Expected input 0 to be an array"))?;	270✔
123
124	let if_false = value.inputs[1]	270✔
125	.array()	270✔
126	.ok_or_else(\|\| vortex_err!("Expected input 1 to be an array"))?;	270✔
127
128	let mask = value.inputs[2]	270✔
129	.mask()	270✔
130	.ok_or_else(\|\| vortex_err!("Expected input 2 to be a mask"))?;	270✔
131
132	Ok(Self {	270✔
133	if_true,	270✔
134	if_false,	270✔
135	mask,	270✔
136	})	270✔
137	}	270✔
138	}
139
140	pub trait ZipKernel: VTable {
141	fn zip(
142	&self,
143	if_true: &Self::Array,
144	if_false: &dyn Array,
145	mask: &Mask,
146	) -> VortexResult<Option<ArrayRef>>;
147	}
148
149	pub struct ZipKernelRef(pub ArcRef<dyn Kernel>);
150	inventory::collect!(ZipKernelRef);
151
152	#[derive(Debug)]
153	pub struct ZipKernelAdapter<V: VTable>(pub V);
154
155	impl<V: VTable + ZipKernel> ZipKernelAdapter<V> {
156	pub const fn lift(&'static self) -> ZipKernelRef {	×
157	ZipKernelRef(ArcRef::new_ref(self))	×
158	}	×
159	}
160
161	impl<V: VTable + ZipKernel> Kernel for ZipKernelAdapter<V> {
162	fn invoke(&self, args: &InvocationArgs) -> VortexResult<Option<Output>> {	39✔
163	let ZipArgs {
164	if_true,	39✔
165	if_false,	39✔
166	mask,	39✔
167	} = ZipArgs::try_from(args)?;	39✔
168	let Some(if_true) = if_true.as_opt::<V>() else {	39✔
169	return Ok(None);	38✔
170	};
171	Ok(V::zip(&self.0, if_true, if_false, mask)?.map(Into::into))	1✔
172	}	39✔
173	}
174
175	pub(crate) fn zip_return_dtype(if_true: &dyn Array, if_false: &dyn Array) -> DType {	154✔
176	if_true	154✔
177	.dtype()	154✔
178	.union_nullability(if_false.dtype().nullability())	154✔
179	}	154✔
180
181	fn zip_impl(if_true: &dyn Array, if_false: &dyn Array, mask: &Mask) -> VortexResult<ArrayRef> {	38✔
182	// if_true.len() == if_false.len() from ComputeFn::invoke
183	let builder = builder_with_capacity(&zip_return_dtype(if_true, if_false), if_true.len());	38✔
184	zip_impl_with_builder(if_true, if_false, mask, builder)	38✔
185	}	38✔
186
187	pub(crate) fn zip_impl_with_builder(	39✔
188	if_true: &dyn Array,	39✔
189	if_false: &dyn Array,	39✔
190	mask: &Mask,	39✔
191	mut builder: Box<dyn ArrayBuilder>,	39✔
192	) -> VortexResult<ArrayRef> {	39✔
193	match mask.slices() {	39✔
194	AllOr::All => Ok(if_true.to_array()),	×
195	AllOr::None => Ok(if_false.to_array()),	×
196	AllOr::Some(slices) => {	39✔
197	for (start, end) in slices {	148✔
198	builder.extend_from_array(&if_false.slice(builder.len(), *start)?)?;	109✔
199	builder.extend_from_array(&if_true.slice(start, end)?)?;	109✔
200	}
201	if builder.len() < if_false.len() {	39✔
202	builder.extend_from_array(&if_false.slice(builder.len(), if_false.len())?)?;	39✔
203	}	×
204	Ok(builder.finish())	39✔
205	}
206	}
207	}	39✔
208
209	#[cfg(test)]
210	mod tests {
211	use vortex_array::arrays::{BoolArray, PrimitiveArray};
212	use vortex_array::compute::zip;
213	use vortex_array::{IntoArray, ToCanonical};
214	use vortex_mask::Mask;
215
216	#[test]
217	fn test_zip_basic() {	1✔
218	let mask =	1✔
219	Mask::try_from(&BoolArray::from_iter([true, false, false, true, false])).unwrap();	1✔
220	let if_true = PrimitiveArray::from_iter([10, 20, 30, 40, 50]).into_array();	1✔
221	let if_false = PrimitiveArray::from_iter([1, 2, 3, 4, 5]).into_array();	1✔
222
223	let result = zip(&if_true, &if_false, &mask).unwrap();	1✔
224	let expected = PrimitiveArray::from_iter([10, 2, 3, 40, 5]);	1✔
225
226	assert_eq!(	1✔
227	result.to_primitive().unwrap().as_slice::<i32>(),	1✔
228	expected.as_slice::<i32>()	1✔
229	);
230	}	1✔
231
232	#[test]
233	fn test_zip_all_true() {	1✔
234	let mask = Mask::new_true(4);	1✔
235	let if_true = PrimitiveArray::from_iter([10, 20, 30, 40]).into_array();	1✔
236	let if_false =	1✔
237	PrimitiveArray::from_option_iter([Some(1), Some(2), Some(3), None]).into_array();	1✔
238
239	let result = zip(&if_true, &if_false, &mask).unwrap();	1✔
240
241	assert_eq!(	1✔
242	result.to_primitive().unwrap().as_slice::<i32>(),	1✔
243	if_true.to_primitive().unwrap().as_slice::<i32>()	1✔
244	);
245
246	// result must be nullable even if_true was not
247	assert_eq!(result.dtype(), if_false.dtype())	1✔
248	}	1✔
249
250	#[test]
251	#[should_panic]
252	fn test_invalid_lengths() {	1✔
253	let mask = Mask::new_false(4);	1✔
254	let if_true = PrimitiveArray::from_iter([10, 20, 30]).into_array();	1✔
255	let if_false = PrimitiveArray::from_iter([1, 2, 3, 4]).into_array();	1✔
256
257	zip(&if_true, &if_false, &mask).unwrap();	1✔
258	}	1✔
259	}

vortex-data / vortex / 16620437478

Source File Press 'n' to go to next uncovered line, 'b' for previous

Source File
Press 'n' to go to next uncovered line, 'b' for previous