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

use num_complex::{Complex32, Complex64};

use crate::config::{GatherConfig, ScatterConfig};
use crate::cubecl::{download_tensor, upload_tensor, CudaBackend};
use crate::{Tensor, TypedTensor};
use tenferro_cpu::CpuBackend;

mod elementwise_tests;
mod fusion_tests;
mod gemm_tests;
mod indexing_tests;
mod metadata_tests;
mod reduction_tests;
mod runtime_tests;
mod structural_tests;

fn cpu_backend() -> CpuBackend {
    CpuBackend::new()
}

fn gpu_backend() -> CudaBackend {
    CudaBackend::new(0).unwrap()
}

fn upload(backend: &CudaBackend, tensor: &Tensor) -> Tensor {
    upload_tensor(backend.runtime(), tensor).unwrap()
}

fn download(backend: &CudaBackend, tensor: &Tensor) -> Tensor {
    download_tensor(backend.runtime(), tensor).unwrap()
}

fn tensor_f32(shape: Vec<usize>, data: Vec<f32>) -> Tensor {
    Tensor::F32(TypedTensor::from_vec_col_major(shape, data).unwrap())
}

fn tensor_f64(shape: Vec<usize>, data: Vec<f64>) -> Tensor {
    Tensor::F64(TypedTensor::from_vec_col_major(shape, data).unwrap())
}

fn tensor_i64(shape: Vec<usize>, data: Vec<i64>) -> Tensor {
    Tensor::I64(TypedTensor::from_vec_col_major(shape, data).unwrap())
}

fn tensor_i32(shape: Vec<usize>, data: Vec<i32>) -> Tensor {
    Tensor::I32(TypedTensor::from_vec_col_major(shape, data).unwrap())
}

fn tensor_bool(shape: Vec<usize>, data: Vec<bool>) -> Tensor {
    Tensor::Bool(TypedTensor::from_vec_col_major(shape, data).unwrap())
}

fn tensor_c32(shape: Vec<usize>, data: Vec<Complex32>) -> Tensor {
    Tensor::C32(TypedTensor::from_vec_col_major(shape, data).unwrap())
}

fn tensor_c64(shape: Vec<usize>, data: Vec<Complex64>) -> Tensor {
    Tensor::C64(TypedTensor::from_vec_col_major(shape, data).unwrap())
}

fn assert_tensor_close(actual: &Tensor, expected: &Tensor, tol: f64) {
    assert_eq!(actual.shape(), expected.shape());
    match (actual, expected) {
        (Tensor::F32(_), Tensor::F32(_)) => {
            let actual = actual.as_slice::<f32>().unwrap();
            let expected = expected.as_slice::<f32>().unwrap();
            for (lhs, rhs) in actual.iter().zip(expected.iter()) {
                let diff = (*lhs as f64 - *rhs as f64).abs();
                assert!(
                    diff <= tol,
                    "f32 tensors differ: lhs={lhs:?} rhs={rhs:?} diff={diff}"
                );
            }
        }
        (Tensor::F64(_), Tensor::F64(_)) => {
            let actual = actual.as_slice::<f64>().unwrap();
            let expected = expected.as_slice::<f64>().unwrap();
            for (lhs, rhs) in actual.iter().zip(expected.iter()) {
                let diff = (*lhs - *rhs).abs();
                assert!(
                    diff <= tol,
                    "f64 tensors differ: lhs={lhs:?} rhs={rhs:?} diff={diff}"
                );
            }
        }
        (Tensor::I64(_), Tensor::I64(_)) => {
            let actual = actual.as_slice::<i64>().unwrap();
            let expected = expected.as_slice::<i64>().unwrap();
            assert_eq!(actual, expected);
        }
        (Tensor::I32(_), Tensor::I32(_)) => {
            let actual = actual.as_slice::<i32>().unwrap();
            let expected = expected.as_slice::<i32>().unwrap();
            assert_eq!(actual, expected);
        }
        (Tensor::Bool(_), Tensor::Bool(_)) => {
            let actual = actual.as_slice::<bool>().unwrap();
            let expected = expected.as_slice::<bool>().unwrap();
            assert_eq!(actual, expected);
        }
        (Tensor::C32(_), Tensor::C32(_)) => {
            let actual = actual.as_slice::<Complex32>().unwrap();
            let expected = expected.as_slice::<Complex32>().unwrap();
            for (lhs, rhs) in actual.iter().zip(expected.iter()) {
                let real_diff = (lhs.re as f64 - rhs.re as f64).abs();
                let imag_diff = (lhs.im as f64 - rhs.im as f64).abs();
                assert!(
                    real_diff <= tol && imag_diff <= tol,
                    "c32 tensors differ: lhs={lhs:?} rhs={rhs:?}"
                );
            }
        }
        (Tensor::C64(_), Tensor::C64(_)) => {
            let actual = actual.as_slice::<Complex64>().unwrap();
            let expected = expected.as_slice::<Complex64>().unwrap();
            for (lhs, rhs) in actual.iter().zip(expected.iter()) {
                let real_diff = (lhs.re - rhs.re).abs();
                let imag_diff = (lhs.im - rhs.im).abs();
                assert!(
                    real_diff <= tol && imag_diff <= tol,
                    "c64 tensors differ: lhs={lhs:?} rhs={rhs:?}"
                );
            }
        }
        _ => panic!(
            "dtype mismatch actual={:?} expected={:?}",
            actual.dtype(),
            expected.dtype()
        ),
    }
}

fn simple_gather_config() -> GatherConfig {
    GatherConfig {
        offset_dims: vec![],
        collapsed_slice_dims: vec![0],
        start_index_map: vec![0],
        index_vector_dim: 1,
        slice_sizes: vec![1],
    }
}

fn diagonal_scatter_config() -> ScatterConfig {
    ScatterConfig {
        update_window_dims: vec![],
        inserted_window_dims: vec![0, 1],
        scatter_dims_to_operand_dims: vec![0, 1],
        index_vector_dim: 1,
    }
}