tenferro-cpu-basic 0.9.0

Shared CPU execution resources and host strided helpers for tenferro.
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

use super::{checked_compact_strides, PooledUninitOutput};
use crate::buffer_pool::BufferPool;

fn recycled(pool: &mut BufferPool) -> tenferro_tensor::TypedTensor<f64> {
    let mut output = PooledUninitOutput::<f64>::new(pool, vec![4]).unwrap();
    for (i, value) in output.as_uninit_slice_mut().iter_mut().enumerate() {
        value.write(i as f64);
    }
    // SAFETY: every logical element was written above.
    unsafe { output.assume_init_recycled() }.unwrap()
}

#[test]
fn recycled_output_returns_only_after_final_group_owner_drops() {
    let mut pool = BufferPool::new();
    let tensor = recycled(&mut pool);
    let pointer = tensor.as_slice().unwrap().as_ptr();
    let (group, slots) = tenferro_tensor::AllocationGroup::from_tensors(vec![
        tenferro_tensor::Tensor::from_typed::<f64>(tensor),
    ])
    .unwrap();
    assert!(pool.is_empty());
    {
        let reads = group.read_views(&slots).unwrap();
        assert_eq!(reads[0].shape(), &[4]);
        assert!(pool.is_empty());
    }
    drop(group);
    assert_eq!(pool.len(), 1);
    let tensor = recycled(&mut pool);
    assert_eq!(tensor.as_slice().unwrap().as_ptr(), pointer);
    drop(tensor);
    assert_eq!(pool.len(), 1);
}

#[test]
fn recycled_drop_during_another_checkout_does_not_reenter_the_execution_lock() {
    let mut pool = BufferPool::new();
    let tensor = recycled(&mut pool);
    let output = PooledUninitOutput::<f64>::new(&pool, vec![4]).unwrap();
    drop(tensor);
    drop(output);
    // The recycled tensor and the abandoned checkout both return their storage, and neither
    // return needs the pool's execution lock while the other is held.
    assert_eq!(pool.len(), 2);
}

#[test]
fn output_and_scratch_checkouts_coexist_without_borrowing_the_pool() {
    let mut pool = BufferPool::new();
    let mut output = PooledUninitOutput::<f64>::new(&pool, vec![2]).unwrap();
    let mut scratch = PooledUninitOutput::<f64>::new(&pool, vec![2]).unwrap();
    output.as_uninit_slice_mut()[0].write(2.0);
    output.as_uninit_slice_mut()[1].write(3.0);
    scratch.as_uninit_slice_mut()[0].write(5.0);
    scratch.as_uninit_slice_mut()[1].write(7.0);
    // An active lease must not prevent using the pool for other resources.
    let zeroed = pool.acquire_zeroed::<f64>(1);
    assert_eq!(zeroed, vec![0.0]);
    // SAFETY: both elements of each lease were initialized above.
    let result = unsafe { output.assume_init_recycled() }.unwrap();
    let temporary = unsafe { scratch.assume_init_recycled() }.unwrap();
    assert_eq!(result.as_slice().unwrap(), &[2.0, 3.0]);
    assert_eq!(temporary.as_slice().unwrap(), &[5.0, 7.0]);
    drop(result);
    drop(temporary);
    assert_eq!(pool.len(), 2);
}

#[test]
fn ranked_recycled_handoff_returns_once_and_rank_error_cancels_checkout() {
    use tenferro_tensor::Rank;

    let pool = BufferPool::new();
    let mut wrong = PooledUninitOutput::<f64>::new(&pool, vec![2]).unwrap();
    wrong.as_uninit_slice_mut()[0].write(1.0);
    wrong.as_uninit_slice_mut()[1].write(2.0);
    // SAFETY: both elements have been initialized, although the selected rank is wrong.
    assert!(unsafe { wrong.assume_init_as_recycled::<Rank<2>>() }.is_err());
    // A handoff that fails on its rank leaves the storage the pool's, so it comes straight back.
    assert_eq!(pool.len(), 1);

    let mut output = PooledUninitOutput::<f64>::new(&pool, vec![2]).unwrap();
    output.as_uninit_slice_mut()[0].write(3.0);
    output.as_uninit_slice_mut()[1].write(4.0);
    // SAFETY: both elements have been initialized.
    let owner = unsafe { output.assume_init_as_recycled::<Rank<1>>() }.unwrap();
    assert_eq!(owner.as_slice().unwrap(), &[3.0, 4.0]);
    assert!(pool.is_empty());
    drop(owner);
    assert_eq!(pool.len(), 1);
}

#[test]
fn recycled_output_extraction_disarms_return_and_pool_may_drop_first() {
    let mut pool = BufferPool::new();
    let tensor = recycled(&mut pool);
    let pointer = tensor.as_slice().unwrap().as_ptr();
    let data = tensor.into_host_vec().unwrap();
    assert_eq!(data.as_ptr(), pointer);
    assert!(pool.is_empty());
    drop(data);
    assert!(pool.is_empty());
    let tensor = recycled(&mut pool);
    drop(pool);
    assert_eq!(tensor.as_slice().unwrap(), &[0., 1., 2., 3.]);
    drop(tensor);
}

#[test]
fn partial_output_returns_to_the_pool_on_error_and_unwind() {
    let mut pool = BufferPool::new();
    for unwind in [false, true] {
        drop(recycled(&mut pool));
        assert_eq!(pool.len(), 1);
        let outcome = std::panic::catch_unwind(std::panic::AssertUnwindSafe(|| {
            let mut output = PooledUninitOutput::<f64>::new(&pool, vec![4]).unwrap();
            output.as_uninit_slice_mut()[0].write(f64::NAN);
            if unwind {
                panic!("kernel stopped after a partial write");
            }
            Err::<(), _>("kernel returned an error after a partial write")
        }));
        if unwind {
            assert!(outcome.is_err());
        } else {
            assert!(outcome.unwrap().is_err());
        }
        // An aborted destination never became a tensor, so its storage is still the pool's and
        // goes back for the next call rather than to the allocator.
        assert_eq!(
            pool.len(),
            1,
            "an aborted checkout returns its buffer to the pool"
        );
        assert_eq!(
            pool.retained_capacity_bytes(),
            4 * std::mem::size_of::<f64>()
        );
        let tensor = recycled(&mut pool);
        assert_eq!(tensor.as_slice().unwrap(), &[0., 1., 2., 3.]);
        drop(tensor);
        pool.clear();
    }
}

#[test]
fn recycled_output_respects_current_limit_and_clear() {
    let mut pool = BufferPool::new();
    let tensor = recycled(&mut pool);
    pool.set_max_retained_capacity_bytes(0);
    drop(tensor);
    assert!(pool.is_empty());
    pool.set_max_retained_capacity_bytes(1024);
    drop(recycled(&mut pool));
    assert_eq!(pool.len(), 1);
    pool.clear();
    assert!(pool.is_empty());
    assert_eq!(pool.retained_capacity_bytes(), 0);
}

#[test]
fn compact_stride_validation_reports_dimension_and_product_overflow() {
    let dimension_error = checked_compact_strides(&[isize::MAX as usize + 1]).unwrap_err();
    assert!(dimension_error
        .to_string()
        .contains("dimension exceeds isize"));

    let stride_error = checked_compact_strides(&[isize::MAX as usize, 2]).unwrap_err();
    assert!(stride_error.to_string().contains("compact stride overflow"));
}

#[test]
fn pooled_uninit_output_public_contract_covers_zero_length_handoff() {
    let pool = BufferPool::new();
    let mut output = PooledUninitOutput::<i32>::new(&pool, vec![0]).unwrap();
    assert!(output.as_uninit_slice_mut().is_empty());
    assert!(output.as_uninit_bytes_mut().is_empty());
    assert_eq!(output.as_uninit_view_mut().unwrap().dims(), &[0]);

    let tensor = unsafe { output.assume_init_recycled() }.unwrap();
    assert_eq!(tensor.shape(), &[0]);
    assert!(pool.is_empty());
}