cubek-interpolate 0.3.0

CubeK: Interpolation
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

use cubecl::{
    benchmark::{Benchmark, ProfileDuration, TimingMethod},
    client::Client,
    future,
    prelude::*,
    std::tensor::TensorHandle,
    zspace::Shape,
};
use cubek_test_utils::{RunSamples, TestInput};

use crate::{definition::InterpolateProblem, interpolate, interpolate_backward};

use crate::InterpolateStrategy;

pub fn bench(
    strategy: &InterpolateStrategy,
    problem: &InterpolateProblem,
    num_samples: usize,
) -> Result<RunSamples, String> {
    let device = cubecl::test_device();
    let client = device.client();

    let dtype = f32::elem_type_native();

    let bench = InterpolateBench {
        problem: problem.clone(),
        strategy: *strategy,
        device,
        client: client.clone(),
        dtype,
        samples: num_samples,
    };

    let durations = bench
        .run(cubek_test_utils::timing_method(TimingMethod::Device))
        .map_err(|e| format!("benchmark failed: {e}"))?
        .durations;

    Ok(RunSamples::new(durations))
}

struct InterpolateBench {
    problem: InterpolateProblem,
    strategy: InterpolateStrategy,
    device: cubecl::Device,
    client: Client,
    dtype: ElemType,
    samples: usize,
}

impl Benchmark for InterpolateBench {
    type Input = TensorHandle;
    type Output = TensorHandle;

    fn prepare(&self) -> Self::Input {
        let shape = match &self.problem {
            InterpolateProblem::Forward(prob) => prob.input_shape(),
            InterpolateProblem::Backward(prob) => Shape::new(prob.out_grad_shape),
        };
        TestInput::builder(self.client.clone(), shape)
            .dtype(self.dtype)
            .uniform(0, -1., 1.)
            .generate_without_host_data()
    }

    fn execute(&self, input: Self::Input) -> Result<TensorHandle, String> {
        match &self.problem {
            InterpolateProblem::Forward(prob) => {
                let output = TensorHandle::empty(&self.client, prob.output_shape(), self.dtype);

                interpolate(
                    &self.client,
                    input.binding(),
                    output.clone().binding(),
                    prob.options,
                    self.strategy,
                    self.dtype,
                )
                .map_err(|err| format!("{err}"))?;

                Ok(output)
            }
            InterpolateProblem::Backward(prob) => {
                let [n, _, _, c] = prob.out_grad_shape;
                let input_grad_shape = vec![n, prob.input_size[0], prob.input_size[1], c];

                // Random input tensor for the backward pass. The actual values don't matter
                // for benchmarking, so we just fill it with random data.
                let backward_input =
                    TestInput::builder(self.client.clone(), input_grad_shape.clone())
                        .dtype(self.dtype)
                        .uniform(0, -1., 1.)
                        .generate_without_host_data();

                let input_grad = TensorHandle::empty(&self.client, input_grad_shape, self.dtype);

                interpolate_backward(
                    &self.client,
                    backward_input.binding(),
                    input.clone().binding(),
                    input_grad.clone().binding(),
                    prob.options,
                    self.dtype,
                )
                .map_err(|err| format!("{err}"))?;

                Ok(input_grad)
            }
        }
    }

    fn num_samples(&self) -> usize {
        self.samples
    }

    fn name(&self) -> String {
        match &self.problem {
            InterpolateProblem::Forward(prob) => format!(
                "interpolate-{:?}-{:?}-{:?}-{:?}-{:?}",
                self.strategy,
                self.dtype,
                prob.options.mode,
                self.device,
                prob.input_shape(),
            )
            .to_lowercase(),
            InterpolateProblem::Backward(prob) => format!(
                "interpolate-backward-{:?}-{:?}-{:?}-{:?}-{:?}",
                self.strategy, self.dtype, prob.options.mode, self.device, prob.out_grad_shape,
            )
            .to_lowercase(),
        }
    }

    fn sync(&self) {
        future::block_on(self.client.sync()).unwrap()
    }

    /// Measure with device timestamps around the launch, so the reported duration is the
    /// kernel's rather than the host's view of launch, output allocation and sync.
    fn profile(&self, args: Self::Input) -> Result<ProfileDuration, String> {
        cubek_test_utils::profile_launch(&self.client, "interpolate-bench", || self.execute(args))
    }
}