cubek-convolution 0.3.0

CubeK: Convolution Kernels
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

use cubecl::prelude::Scalar;
use cubek_matmul::{
    definition::{MatmulElems, MatmulGlobalElems},
    definition::{MatmulPrecision, MatrixPrecision},
};
use cubek_test_utils::{HostData, Progress};

use crate::{
    Strategy,
    eval::{
        benchmarks::problem::Conv2dProblem,
        cpu_reference::{ConvSpec, cpu_reference_result, strategy_result},
    },
};

type LhsG<MP> = <<MP as MatmulPrecision>::Lhs as MatrixPrecision>::Global;
type RhsG<MP> = <<MP as MatmulPrecision>::Rhs as MatrixPrecision>::Global;
type AccG<MP> = <<MP as MatmulPrecision>::Acc as MatrixPrecision>::Global;

pub struct Conv2dCorrectness;

impl cubek_test_utils::Correctness for Conv2dCorrectness {
    type Problem = Conv2dProblem;
    type Strategy = Strategy;

    fn kernel_result(
        &self,
        strategy: &Strategy,
        problem: &Conv2dProblem,
        seeds: &[u64],
    ) -> Result<HostData, String> {
        let device = cubecl::test_device();
        let client = device.client();
        let (spec, dtypes) = build_spec_and_dtypes::<half::f16>(problem);
        strategy_result(client, spec, strategy.clone(), dtypes, seeds[0], seeds[1])
    }

    fn reference_result(
        &self,
        problem: &Conv2dProblem,
        seeds: &[u64],
        progress: Option<&Progress>,
    ) -> Result<HostData, String> {
        let device = cubecl::test_device();
        let client = device.client();
        let (spec, dtypes) = build_spec_and_dtypes::<half::f16>(problem);
        cpu_reference_result(client, spec, dtypes, seeds[0], seeds[1], progress)
    }
}

fn build_spec_and_dtypes<MP: MatmulPrecision>(p: &Conv2dProblem) -> (ConvSpec, MatmulElems) {
    let [n, c_in, h_in, w_in] = p.input_shape;
    let [c_out, _, k_h, k_w] = p.weight_shape;
    let dtypes = MatmulElems::from_globals(&MatmulGlobalElems {
        lhs: LhsG::<MP>::elem_type_native(),
        rhs: RhsG::<MP>::elem_type_native(),
        out: AccG::<MP>::elem_type_native(),
    });
    let spec = ConvSpec {
        batches: n,
        in_h: h_in,
        in_w: w_in,
        channels: c_in,
        out_channels: c_out,
        args: p.args.clone(),
        kernel_size: [k_h, k_w],
    };
    (spec, dtypes)
}