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

//! Benchmarks the exploration on CUDA gpus.
use std::io::{self, Write};

use telamon::device::{ArgMap, Context};
use telamon::explorer::config::Config;
use telamon::helper::MemInit;
use telamon_cli::{
    Bench, CommonOpt, ContextBuilder, CublasHandle, KernelParam, Reference,
};
use telamon_kernels::statistics::estimate_mean;
use telamon_kernels::{linalg, Kernel};

use structopt::StructOpt;

/// The number of times to run the generated code to evaluate its performance.
const NUM_CODE_RUNS: usize = 40;
/// Search timeout in minutes.
const TIMEOUT: u64 = 240;

/// Benchmarks a kernel against a reference implementation.
fn benchmark<'a, K, REF, CB>(
    mut config: Config,
    params: K::Parameters,
    executor: CB,
    reference: &REF,
) where
    K: Kernel<'a>,
    CB: ContextBuilder<'a>,
    REF: Reference<'a, K, Context = CB::Context>,
{
    config.timeout.get_or_insert(TIMEOUT);

    let mut context = executor.build_context();
    let runtime = K::benchmark(
        &config,
        params.clone(),
        NUM_CODE_RUNS,
        MemInit::RandomFill,
        &mut context,
    );
    let ref_runtime = Bench::default()
        .warmup(4)
        .runs(NUM_CODE_RUNS)
        .benchmark_fn(|| reference.eval_reference(&params, &context));
    let mut f =
        std::fs::File::create(config.output_path("benchmark.txt").unwrap()).unwrap();
    writeln!(f, "runtimes: {:?}", runtime).unwrap();
    let mean = estimate_mean(runtime, 0.95, "ns");
    let ref_mean = estimate_mean(ref_runtime, 0.95, "ns");
    writeln!(
        f,
        "{}: {}, reference: {}, speedup: {:.2}",
        K::name(),
        mean,
        ref_mean,
        ref_mean.value / mean.value
    )
    .unwrap();
}

#[derive(StructOpt)]
struct Opt {
    #[structopt(flatten)]
    common: CommonOpt,

    #[structopt(short = "r", long = "repeat", default_value = "10")]
    repeat: usize,

    #[structopt(short = "k", long = "kernel")]
    kernels: Vec<KernelParam>,
}

trait BenchRun<'a, B, R> {
    fn run(self, config: &Config, builder: B, reference: &R);
}

struct Benchmark<'a, K>
where
    K: Kernel<'a>,
{
    params: K::Parameters,
    name: String,
    iteration: usize,
}

impl<'a, K> Benchmark<'a, K>
where
    K: Kernel<'a>,
{
    fn new(params: K::Parameters, name: String, iteration: usize) -> Self {
        Benchmark {
            params,
            name,
            iteration,
        }
    }

    fn output_dir(&self) -> String {
        format!("{}/{}", self.name, self.iteration)
    }
}

impl<'a, K, B, R> BenchRun<'a, B, R> for Benchmark<'a, K>
where
    K: Kernel<'a>,
    B: ContextBuilder<'a>,
    R: Reference<'a, K, Context = B::Context>,
{
    fn run(self, config: &Config, builder: B, reference: &R) {
        let mut config = config.clone();
        config.output_dir = std::path::Path::new(&config.output_dir)
            .join(self.output_dir())
            .to_str()
            .unwrap()
            .to_string();
        benchmark::<K, _, _>(config.clone(), self.params, builder, reference)
    }
}

macro_rules! benchmark {
    (Sgemm($m:expr, $n:expr, $k:expr)[$iter:expr]) => {{
        self::Benchmark::<'_, linalg::FusedMM<'_, f32>>::new(
            linalg::FusedMMP::new($m, $n, $k),
            format!("Sgemm_{}_{}_{}", $m, $n, $k),
            $iter,
        )
    }};

    (BatchMM($b:expr, $m:expr, $n:expr, $k:expr)[$iter:expr]) => {{
        self::Benchmark::<'_, linalg::BatchMM<'_, f32>>::new(
            linalg::BatchMMP::new($b, $m, $n, $k),
            format!("BatchMM_{}_{}_{}_{}", $b, $m, $n, $k),
            $iter,
        )
    }};
}

fn main() {
    env_logger::init();
    let args = Opt::from_args();

    let executor = telamon_cuda::Executor::init();
    let reference = CublasHandle::new();

    let config = args.common.config().unwrap();

    for idx in 0..args.repeat {
        for kernel in &args.kernels {
            use KernelParam::*;

            match *kernel {
                Axpy { n } => Benchmark::<'_, linalg::Axpy<f32>>::new(
                    (n, true),
                    format!("Axpy_{}", n),
                    idx,
                )
                .run(&config, &executor, &reference),
                MatVec { m, n } => Benchmark::<'_, linalg::MatVec<f32>>::new(
                    (m, n, true),
                    format!("Sgemv_{}_{}", m, n),
                    idx,
                )
                .run(&config, &executor, &reference),
                Gesummv { m, n } => Benchmark::<'_, linalg::Gesummv<f32>>::new(
                    (m, n, true),
                    format!("Gesummv_{}_{}", m, n),
                    idx,
                )
                .run(&config, &executor, &reference),
                Gemm { m, n, k } => Benchmark::<'_, linalg::FusedMM<'_, f32>>::new(
                    linalg::FusedMMP::new(m, n, k),
                    format!("Sgemm_{}_{}_{}", m, n, k),
                    idx,
                )
                .run(&config, &executor, &reference),
                BatchMM { b, m, n, k } => Benchmark::<'_, linalg::BatchMM<'_, f32>>::new(
                    linalg::BatchMMP::new(b, m, n, k),
                    format!("BatchMM_{}_{}_{}_{}", b, m, n, k),
                    idx,
                )
                .run(&config, &executor, &reference),
            }
        }
    }
}