ndarrow 0.0.4

Zero-copy bridge between Apache Arrow and ndarray
Documentation 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203

//! Integration tests verifying the zero-copy guarantee.
//!
//! These tests confirm that bridge conversions never allocate or copy data.
//! They compare raw pointer addresses to prove views point directly at Arrow
//! buffers, and ownership transfers preserve the original allocation.

use arrow_array::{
    Float32Array, Float64Array, ListArray, PrimitiveArray,
    types::{Float32Type, Float64Type},
};
use ndarray::{Array1, Array2, ArrayD, IxDyn};
use ndarrow::{
    AsNdarray, IntoArrow, arrayd_to_fixed_shape_tensor, arrays_to_variable_shape_tensor,
    csr_to_extension_array, csr_view_from_extension, fixed_shape_tensor_as_array_viewd,
    variable_shape_tensor_iter,
};

// ─── Inbound: views point to Arrow's buffer ───

#[test]
fn primitive_view_borrows_arrow_buffer_f64() {
    let data = vec![1.0_f64, 2.0, 3.0, 4.0, 5.0];
    let arr = Float64Array::from(data);
    let arrow_ptr = arr.values().as_ref().as_ptr();

    let view = arr.as_ndarray().unwrap();
    let view_ptr = view.as_ptr();

    assert_eq!(arrow_ptr, view_ptr, "ArrayView1 must point directly to Arrow's buffer");
}

#[test]
fn primitive_view_borrows_arrow_buffer_f32() {
    let data = vec![1.0_f32, 2.0, 3.0];
    let arr = Float32Array::from(data);
    let arrow_ptr = arr.values().as_ref().as_ptr();

    let view = arr.as_ndarray().unwrap();
    let view_ptr = view.as_ptr();

    assert_eq!(arrow_ptr, view_ptr, "ArrayView1 must point directly to Arrow's buffer");
}

#[test]
fn fsl_view_borrows_inner_buffer() {
    use std::sync::Arc;

    use arrow_schema::Field;

    let data = vec![1.0_f64, 2.0, 3.0, 4.0, 5.0, 6.0];
    let values = Float64Array::from(data);
    let inner_ptr = values.values().as_ref().as_ptr();

    let field = Arc::new(Field::new("item", arrow_schema::DataType::Float64, false));
    let fsl = arrow_array::FixedSizeListArray::new(field, 3, Arc::new(values), None);

    let view = ndarrow::fixed_size_list_as_array2::<Float64Type>(&fsl).unwrap();
    let view_ptr = view.as_ptr();

    assert_eq!(
        inner_ptr, view_ptr,
        "ArrayView2 must point directly to the inner PrimitiveArray's buffer"
    );
}

#[test]
fn unchecked_view_borrows_same_buffer() {
    let arr = Float64Array::from(vec![1.0, 2.0, 3.0]);
    let arrow_ptr = arr.values().as_ref().as_ptr();

    let view = unsafe { arr.as_ndarray_unchecked() };
    let view_ptr = view.as_ptr();

    assert_eq!(arrow_ptr, view_ptr, "unchecked view must also point to Arrow's buffer");
}

#[test]
fn masked_view_borrows_same_buffer() {
    let arr = Float64Array::from(vec![Some(1.0), None, Some(3.0)]);
    let arrow_ptr = arr.values().as_ref().as_ptr();

    let (view, _mask) = arr.as_ndarray_masked();
    let view_ptr = view.as_ptr();

    assert_eq!(arrow_ptr, view_ptr, "masked view must point to Arrow's buffer");
}

// ─── Outbound: ownership transfer preserves the allocation ───

#[test]
fn array1_transfer_preserves_allocation_f64() {
    let data = vec![10.0_f64, 20.0, 30.0, 40.0, 50.0];
    let original_ptr = data.as_ptr();
    let arr = Array1::from_vec(data);

    let prim: PrimitiveArray<Float64Type> = arr.into_arrow().unwrap();
    let arrow_ptr = prim.values().as_ref().as_ptr();

    assert_eq!(original_ptr, arrow_ptr, "into_arrow must transfer the Vec without copying");
}

#[test]
fn array1_transfer_preserves_allocation_f32() {
    let data = vec![1.0_f32, 2.0, 3.0];
    let original_ptr = data.as_ptr();
    let arr = Array1::from_vec(data);

    let prim: PrimitiveArray<Float32Type> = arr.into_arrow().unwrap();
    let arrow_ptr = prim.values().as_ref().as_ptr();

    assert_eq!(original_ptr, arrow_ptr, "into_arrow must transfer the Vec without copying");
}

#[test]
fn array2_transfer_preserves_allocation() {
    let data = vec![1.0_f64, 2.0, 3.0, 4.0, 5.0, 6.0];
    let original_ptr = data.as_ptr();
    let arr = Array2::from_shape_vec((2, 3), data).unwrap();

    let fsl = arr.into_arrow().unwrap();
    let inner = fsl.values().as_any().downcast_ref::<PrimitiveArray<Float64Type>>().unwrap();
    let arrow_ptr = inner.values().as_ref().as_ptr();

    assert_eq!(
        original_ptr, arrow_ptr,
        "into_arrow for Array2 must transfer the Vec without copying"
    );
}

// ─── Full pipeline: only the computation allocates ───

#[test]
fn pipeline_only_computation_allocates() {
    // Create Arrow data
    let input_data = vec![1.0_f64, 2.0, 3.0, 4.0];
    let arrow_input = Float64Array::from(input_data);
    let input_ptr = arrow_input.values().as_ref().as_ptr();

    // Step 1: Arrow → ndarray view (zero-copy)
    let view = arrow_input.as_ndarray().unwrap();
    assert_eq!(view.as_ptr(), input_ptr, "inbound view must share input buffer");

    // Step 2: Computation (this IS the allocation — it's expected)
    let result = &view * 2.0;

    // Step 3: ndarray → Arrow (zero-copy ownership transfer)
    let result_ptr = result.as_ptr();
    let arrow_output: Float64Array = result.into_arrow().unwrap();
    let output_ptr = arrow_output.values().as_ref().as_ptr();

    assert_eq!(result_ptr, output_ptr, "outbound transfer must preserve the computation buffer");

    // The only allocation in the pipeline is step 2 (the computation).
    // Steps 1 and 3 are O(1) pointer operations.
}

// ─── Sparse/tensor zero-copy checks ───

#[test]
fn csr_extension_view_borrows_arrow_buffers() {
    let row_ptrs = vec![0_i32, 2, 3];
    let col_indices = vec![0_u32, 2, 1];
    let values = vec![1.0_f64, 5.0, 2.0];
    let (field, array) = csr_to_extension_array("csr", 4, row_ptrs, col_indices, values).unwrap();

    let indices = array.column(0).as_any().downcast_ref::<ListArray>().unwrap();
    let indices_values =
        indices.values().as_any().downcast_ref::<arrow_array::UInt32Array>().unwrap();
    let values_col = array.column(1).as_any().downcast_ref::<ListArray>().unwrap();
    let values_values = values_col.values().as_any().downcast_ref::<Float64Array>().unwrap();

    let view = csr_view_from_extension::<Float64Type>(&field, &array).unwrap();
    assert_eq!(view.col_indices.as_ptr(), indices_values.values().as_ref().as_ptr());
    assert_eq!(view.values.as_ptr(), values_values.values().as_ref().as_ptr());
}

#[test]
fn fixed_shape_tensor_view_borrows_arrow_buffer() {
    let data = vec![1.0_f32, 2.0, 3.0, 4.0, 5.0, 6.0];
    let tensor = ArrayD::from_shape_vec(IxDyn(&[2, 3]), data).unwrap();
    let (field, fsl) = arrayd_to_fixed_shape_tensor("tensor", tensor).unwrap();

    let inner = fsl.values().as_any().downcast_ref::<Float32Array>().unwrap();
    let arrow_ptr = inner.values().as_ref().as_ptr();
    let view = fixed_shape_tensor_as_array_viewd::<Float32Type>(&field, &fsl).unwrap();
    assert_eq!(view.as_ptr(), arrow_ptr);
}

#[test]
fn variable_shape_tensor_iter_borrows_arrow_buffer() {
    let a = ArrayD::from_shape_vec(IxDyn(&[2, 2]), vec![1.0_f32, 2.0, 3.0, 4.0]).unwrap();
    let b = ArrayD::from_shape_vec(IxDyn(&[1, 2]), vec![5.0_f32, 6.0]).unwrap();
    let (field, array) = arrays_to_variable_shape_tensor("ragged", vec![a, b], None).unwrap();

    let data_col =
        array.column_by_name("data").unwrap().as_any().downcast_ref::<ListArray>().unwrap();
    let data_values = data_col.values().as_any().downcast_ref::<Float32Array>().unwrap();
    let arrow_ptr = data_values.values().as_ref().as_ptr();

    let mut iter = variable_shape_tensor_iter::<Float32Type>(&field, &array).unwrap();
    let (_row, view) = iter.next().unwrap().unwrap();
    assert_eq!(view.as_ptr(), arrow_ptr);
}