tenferro-gpu 0.1.0

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

// Run with: cargo test --features cuda -- --ignored
use crate::config::{PadConfig, ScatterConfig, SliceConfig};
use num_complex::Complex64;
use tenferro_tensor::TensorIndexing;

use super::{
    assert_tensor_close, cpu_backend, diagonal_scatter_config, download, gpu_backend,
    simple_gather_config, tensor_c64, tensor_f64, tensor_i64, upload,
};

#[test]
#[ignore]
fn test_cubecl_slice_dynamic_slice_and_pad_match_cpu() {
    let input = tensor_f64(
        vec![4, 4],
        vec![
            1.0, 2.0, 3.0, 4.0, 5.0, 6.0, 7.0, 8.0, 9.0, 10.0, 11.0, 12.0, 13.0, 14.0, 15.0, 16.0,
        ],
    );
    let slice_config = SliceConfig {
        starts: vec![1, 1],
        limits: vec![4, 4],
        strides: vec![2, 1],
    };
    let pad_config = PadConfig {
        edge_padding_low: vec![1, 1],
        edge_padding_high: vec![1, 0],
        interior_padding: vec![0, 1],
    };
    let starts = tensor_i64(vec![2], vec![2, 3]);

    let mut cpu = cpu_backend();
    let mut gpu = gpu_backend();
    let gpu_input = upload(&gpu, &input);
    let gpu_starts = upload(&gpu, &starts);

    let expected = cpu.slice(&input, &slice_config).unwrap();
    let gpu_out = gpu.slice(&gpu_input, &slice_config).unwrap();
    let actual = download(&gpu, &gpu_out);
    assert_tensor_close(&actual, &expected, 1e-12);

    let expected = cpu.dynamic_slice(&input, &starts, &[2, 2]).unwrap();
    let gpu_out = gpu.dynamic_slice(&gpu_input, &gpu_starts, &[2, 2]).unwrap();
    let actual = download(&gpu, &gpu_out);
    assert_tensor_close(&actual, &expected, 1e-12);

    let expected = cpu.pad(&input, &pad_config).unwrap();
    let gpu_out = gpu.pad(&gpu_input, &pad_config).unwrap();
    let actual = download(&gpu, &gpu_out);
    assert_tensor_close(&actual, &expected, 1e-12);
}

#[test]
#[ignore]
fn test_cubecl_gather_and_scatter_match_cpu() {
    let operand = tensor_f64(vec![5], vec![10.0, 20.0, 30.0, 40.0, 50.0]);
    let start_indices = tensor_i64(vec![3, 1], vec![0, 2, 4]);

    let scatter_operand = tensor_f64(
        vec![3, 3],
        vec![1.0, 0.0, 0.0, 0.0, 2.0, 0.0, 0.0, 0.0, 3.0],
    );
    let scatter_indices = tensor_i64(vec![3, 2], vec![0, 1, 2, 0, 1, 2]);
    let updates = tensor_f64(vec![3], vec![5.0, 6.0, 7.0]);

    let mut cpu = cpu_backend();
    let mut gpu = gpu_backend();

    let gpu_operand = upload(&gpu, &operand);
    let gpu_start_indices = upload(&gpu, &start_indices);
    let expected = cpu
        .gather(&operand, &start_indices, &simple_gather_config())
        .unwrap();
    let gpu_out = gpu
        .gather(&gpu_operand, &gpu_start_indices, &simple_gather_config())
        .unwrap();
    let actual = download(&gpu, &gpu_out);
    assert_tensor_close(&actual, &expected, 1e-12);

    let gpu_scatter_operand = upload(&gpu, &scatter_operand);
    let gpu_scatter_indices = upload(&gpu, &scatter_indices);
    let gpu_updates = upload(&gpu, &updates);
    let expected = cpu
        .scatter(
            &scatter_operand,
            &scatter_indices,
            &updates,
            &diagonal_scatter_config(),
        )
        .unwrap();
    let gpu_out = gpu
        .scatter(
            &gpu_scatter_operand,
            &gpu_scatter_indices,
            &gpu_updates,
            &diagonal_scatter_config(),
        )
        .unwrap();
    let actual = download(&gpu, &gpu_out);
    assert_tensor_close(&actual, &expected, 1e-12);
}

#[test]
#[ignore]
fn test_cubecl_scatter_skips_invalid_windows_like_cpu() {
    let operand = tensor_f64(vec![4], vec![9.0, 8.0, 7.0, 6.0]);
    let scatter_indices = tensor_i64(vec![3, 1], vec![-1, 2, 4]);
    let updates = tensor_f64(vec![3], vec![5.0, 6.0, 7.0]);
    let config = ScatterConfig {
        update_window_dims: vec![],
        inserted_window_dims: vec![0],
        scatter_dims_to_operand_dims: vec![0],
        index_vector_dim: 1,
    };

    let mut cpu = cpu_backend();
    let mut gpu = gpu_backend();
    let gpu_operand = upload(&gpu, &operand);
    let gpu_indices = upload(&gpu, &scatter_indices);
    let gpu_updates = upload(&gpu, &updates);

    let expected = cpu
        .scatter(&operand, &scatter_indices, &updates, &config)
        .unwrap();
    let gpu_out = gpu
        .scatter(&gpu_operand, &gpu_indices, &gpu_updates, &config)
        .unwrap();
    let actual = download(&gpu, &gpu_out);
    assert_tensor_close(&actual, &expected, 1e-12);
}

#[test]
#[ignore]
fn test_cubecl_scatter_accumulates_overlapping_updates_like_cpu() {
    let operand = tensor_f64(vec![4], vec![1.0, 10.0, 100.0, 1000.0]);
    let scatter_indices = tensor_i64(vec![4, 1], vec![1, 1, -1, 3]);
    let updates = tensor_f64(vec![4], vec![2.0, 3.0, 99.0, 5.0]);
    let config = ScatterConfig {
        update_window_dims: vec![],
        inserted_window_dims: vec![0],
        scatter_dims_to_operand_dims: vec![0],
        index_vector_dim: 1,
    };

    let mut cpu = cpu_backend();
    let mut gpu = gpu_backend();
    let gpu_operand = upload(&gpu, &operand);
    let gpu_indices = upload(&gpu, &scatter_indices);
    let gpu_updates = upload(&gpu, &updates);

    let expected = cpu
        .scatter(&operand, &scatter_indices, &updates, &config)
        .unwrap();
    let gpu_out = gpu
        .scatter(&gpu_operand, &gpu_indices, &gpu_updates, &config)
        .unwrap();
    let actual = download(&gpu, &gpu_out);
    assert_tensor_close(&actual, &expected, 1e-12);
}

#[test]
#[ignore]
fn test_cubecl_complex_scatter_accumulates_overlapping_updates_like_cpu() {
    let operand = tensor_c64(
        vec![3],
        vec![
            Complex64::new(1.0, 10.0),
            Complex64::new(2.0, 20.0),
            Complex64::new(3.0, 30.0),
        ],
    );
    let scatter_indices = tensor_i64(vec![3, 1], vec![0, 0, 2]);
    let updates = tensor_c64(
        vec![3],
        vec![
            Complex64::new(4.0, 1.0),
            Complex64::new(5.0, 2.0),
            Complex64::new(6.0, 3.0),
        ],
    );
    let config = ScatterConfig {
        update_window_dims: vec![],
        inserted_window_dims: vec![0],
        scatter_dims_to_operand_dims: vec![0],
        index_vector_dim: 1,
    };

    let mut cpu = cpu_backend();
    let mut gpu = gpu_backend();
    let gpu_operand = upload(&gpu, &operand);
    let gpu_indices = upload(&gpu, &scatter_indices);
    let gpu_updates = upload(&gpu, &updates);

    let expected = cpu
        .scatter(&operand, &scatter_indices, &updates, &config)
        .unwrap();
    let gpu_out = gpu
        .scatter(&gpu_operand, &gpu_indices, &gpu_updates, &config)
        .unwrap();
    let actual = download(&gpu, &gpu_out);
    assert_tensor_close(&actual, &expected, 1e-12);
}