Scalable Matrix Extension (SME)Advent Calendar 2024

SME日記その12 OpenBLASのSSCALでSMEが使われているかを検証してみる Part.2

Last updated at 2024-12-12Posted at 2024-12-12

OpenBLASのSSCALでSMEが使われているかを検証すべく，ベンチマークプログラムを作成してみました．

SMEシリーズ

ソースコード

mix new nx_sgemm_bench_openblas

mix.exs

defmodule NxSgemmBenchOpenblas.MixProject do
  use Mix.Project

  def project do
    [
      app: :nx_sgemm_bench_openblas,
      version: "0.1.0",
      elixir: "~> 1.17",
      start_permanent: Mix.env() == :prod,
      deps: deps()
    ]
  end

  # Run "mix help compile.app" to learn about applications.
  def application do
    [
      extra_applications: [:logger]
    ]
  end

  # Run "mix help deps" to learn about dependencies.
  defp deps do
    [
      # {:dep_from_hexpm, "~> 0.3.0"},
      # {:dep_from_git, git: "https://github.com/elixir-lang/my_dep.git", tag: "0.1.0"}
      {:nx_sgemm, github: "zacky1972/nx_sgemm", branch: "main"},
      {:benchee, "~> 1.0", only: :dev}
    ]
  end
end

mix deps.get 
export USE_OPEN_BLAS=true 
mix compile

benchmark.exs

Benchee.run(
  %{
    "Nx" => fn input -> Nx.multiply(input, 2.0) end,
    "OpenBLAS" => fn input -> NxSgemm.multiply(input, 2.0) end
  },
  inputs: %{
    "Small" => Nx.iota({1_000}) |> Nx.multiply(1.0),
    "Medium" => Nx.iota({10_000}) |> Nx.multiply(1.0),
    "Bigger" => Nx.iota({100_000}) |> Nx.multiply(1.0)
  }
)

まず M3 Max で実行してみました．

% mix run -r benchmark.exs
Operating System: macOS
CPU Information: Apple M3 Max
Number of Available Cores: 16
Available memory: 128 GB
Elixir 1.17.3
Erlang 27.1.2
JIT enabled: true

Benchmark suite executing with the following configuration:
warmup: 2 s
time: 5 s
memory time: 0 ns
reduction time: 0 ns
parallel: 1
inputs: Bigger, Medium, Small
Estimated total run time: 42 s

Benchmarking Nx with input Bigger ...
Benchmarking Nx with input Medium ...
Benchmarking Nx with input Small ...
Benchmarking OpenBLAS with input Bigger ...
Benchmarking OpenBLAS with input Medium ...
Benchmarking OpenBLAS with input Small ...
Calculating statistics...
Formatting results...

##### With input Bigger #####
Name               ips        average  deviation         median         99th %
OpenBLAS       10.73 K      0.0932 ms    ±10.50%      0.0924 ms       0.120 ms
Nx              0.21 K        4.75 ms     ±3.68%        4.78 ms        5.31 ms

Comparison: 
OpenBLAS       10.73 K
Nx              0.21 K - 51.02x slower +4.66 ms

##### With input Medium #####
Name               ips        average  deviation         median         99th %
OpenBLAS      652.10 K        1.53 μs   ±868.81%        1.46 μs        1.96 μs
Nx              2.54 K      393.24 μs     ±5.29%      385.63 μs      463.12 μs

Comparison: 
OpenBLAS      652.10 K
Nx              2.54 K - 256.43x slower +391.71 μs

##### With input Small #####
Name               ips        average  deviation         median         99th %
OpenBLAS        3.50 M        0.29 μs  ±6654.78%        0.25 μs           3 μs
Nx            0.0255 M       39.27 μs     ±6.92%       38.38 μs       49.17 μs

Comparison: 
OpenBLAS        3.50 M
Nx            0.0255 M - 137.32x slower +38.99 μs

ips が，Small < Medium なのは想定していましたが，Medium > Bigger なのは，キャッシュメモリから溢れているからかもしれませんね．

次はいよいよM4 Proで実行してみます．

You get articles that match your needs
You can efficiently read back useful information
You can use dark theme

What you can do with signing up