tenferro-gpu 0.3.0

CubeCL-backed CUDA and WebGPU provider backends for tenferro tensors.
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

#![cfg(all(feature = "webgpu", target_os = "macos"))]

use tenferro_gpu::{apple::AppleContext, apple::AppleTransferStats};
use tenferro_tensor::{HostAccessError, Tensor, TensorDot, TypedTensor};

fn apple_context() -> Option<AppleContext> {
    match AppleContext::new() {
        Ok(context) => Some(context),
        Err(error) => {
            eprintln!("skipping Apple context test: {error}");
            None
        }
    }
}

#[test]
fn independent_contexts_reject_foreign_managed_allocations() {
    let (Some(first), Some(second)) = (apple_context(), apple_context()) else {
        return;
    };
    assert_ne!(first.domain_id(), second.domain_id());

    let host = Tensor::from_vec_col_major(vec![2], vec![1.0_f32, 2.0]).unwrap();
    let managed = first.upload_tensor(&host).unwrap();
    let error = second.download_tensor(&managed).unwrap_err();
    assert!(matches!(
        error,
        tenferro_tensor::Error::HostAccess {
            source: HostAccessError::ForeignDomain { .. },
            ..
        }
    ));
}

#[test]
fn managed_upload_maps_without_post_creation_transfers_and_keeps_identity() {
    let Some(context) = apple_context() else {
        return;
    };
    let host = Tensor::from_vec_col_major(vec![2], vec![1.0_f32, 2.0]).unwrap();
    let mut managed = context.upload_tensor(&host).unwrap();
    assert_eq!(
        context.transfer_stats(),
        AppleTransferStats {
            uploaded_bytes: 8,
            downloaded_bytes: 0,
        }
    );

    let Tensor::F32(typed) = &mut managed else {
        panic!("expected f32 tensor")
    };
    assert_eq!(typed.allocation_domain(), Some(context.domain_id()));
    let allocation = typed.allocation_id().unwrap();
    assert_eq!(
        typed.with_host_read(|data| data.to_vec()).unwrap(),
        [1.0, 2.0]
    );
    typed
        .with_host_write(|data| data.copy_from_slice(&[3.0, 2.0]))
        .unwrap();
    assert_eq!(
        typed.with_host_read(|data| data.to_vec()).unwrap(),
        [3.0, 2.0]
    );
    assert_eq!(typed.allocation_id(), Some(allocation));
    assert_eq!(context.transfer_stats().downloaded_bytes, 0);

    let downloaded = context.download_tensor(&managed).unwrap();
    assert_eq!(downloaded.as_slice::<f32>().unwrap(), &[3.0, 2.0]);
    assert_eq!(context.transfer_stats().downloaded_bytes, 8);
}

#[test]
fn cpu_domain_allocator_produces_write_only_managed_outputs_without_transfers() {
    let Some(context) = apple_context() else {
        return;
    };
    let domain = context.cpu_backend().shared_allocation_domain().unwrap();
    let output = domain.allocate(tenferro_tensor::DType::F64, &[2]).unwrap();
    let Tensor::F64(mut output) = output else {
        panic!("expected f64 output")
    };
    assert_eq!(output.allocation_domain(), Some(context.domain_id()));
    output
        .with_host_write(|data| data.copy_from_slice(&[5.0, 8.0]))
        .unwrap();
    assert_eq!(
        output.with_host_read(|data| data.to_vec()).unwrap(),
        [5.0, 8.0]
    );
    assert_eq!(context.transfer_stats(), AppleTransferStats::default());
}

#[test]
fn metal_output_stays_in_the_context_domain_without_host_transfers() {
    let Some(context) = apple_context() else {
        return;
    };
    let lhs = Tensor::F32(TypedTensor::from_vec_col_major(vec![1, 1], vec![2.0_f32]).unwrap());
    let rhs = Tensor::F32(TypedTensor::from_vec_col_major(vec![1, 1], vec![3.0_f32]).unwrap());
    let lhs = context.upload_tensor(&lhs).unwrap();
    let rhs = context.upload_tensor(&rhs).unwrap();
    let Tensor::F32(lhs_typed) = &lhs else {
        panic!("expected f32 lhs")
    };
    let lhs_allocation = lhs_typed.allocation_id().unwrap();
    assert_eq!(
        lhs_typed.with_host_read(|data| data.to_vec()).unwrap(),
        [2.0]
    );
    let before = context.transfer_stats();
    let mut metal = context.metal_backend().clone();
    let output = metal
        .dot_general(
            &lhs,
            &rhs,
            &tenferro_tensor::DotGeneralConfig {
                lhs_contracting_dims: vec![1],
                rhs_contracting_dims: vec![0],
                lhs_batch_dims: vec![],
                rhs_batch_dims: vec![],
            },
        )
        .unwrap();
    metal.synchronize().unwrap();
    assert_eq!(
        lhs_typed.with_host_read(|data| data.to_vec()).unwrap(),
        [2.0]
    );
    assert_eq!(lhs_typed.allocation_id(), Some(lhs_allocation));

    let Tensor::F32(output) = output else {
        panic!("expected f32 output")
    };
    assert_eq!(output.allocation_domain(), Some(context.domain_id()));
    assert_eq!(context.transfer_stats(), before);
    assert_eq!(output.with_host_read(|data| data.to_vec()).unwrap(), [6.0]);
}