Graviton4 向けの C++ ビルド環境を作成して OpenBLAS で行列計算を試してみた

sudo dnf update -y

# Development Tools グループのインストール
sudo dnf groupinstall -y "Development Tools"

# C++ コンパイラとビルドツール
sudo dnf install -y gcc gcc-c++ cmake make git

# OpenBLAS ライブラリ
sudo dnf install -y openblas-devel lapack-devel

$ gcc --version
gcc (GCC) 11.5.0 20240719 (Red Hat 11.5.0-5)
Copyright (C) 2021 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

$ g++ --version
g++ (GCC) 11.5.0 20240719 (Red Hat 11.5.0-5)
Copyright (C) 2021 Free Software Foundation, Inc.
This is free software; see the source for copying conditions.  There is NO
warranty; not even for MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.

$ cmake --version
cmake version 3.22.2

CMake suite maintained and supported by Kitware (kitware.com/cmake).

$ make --version
GNU Make 4.3
Built for aarch64-amazon-linux-gnu
Copyright (C) 1988-2020 Free Software Foundation, Inc.
License GPLv3+: GNU GPL version 3 or later <http://gnu.org/licenses/gpl.html>
This is free software: you are free to change and redistribute it.
There is NO WARRANTY, to the extent permitted by law.
[ec2-user@ip-10-0-0-195 ~]$

# OpenBLAS パッケージ情報
$ rpm -q openblas openblas-devel
openblas-0.3.18-1.amzn2023.0.3.aarch64
openblas-devel-0.3.18-1.amzn2023.0.3.aarch64

$ cat /proc/cpuinfo | grep Features
Features	: fp asimd evtstrm aes pmull sha1 sha2 crc32 atomics fphp asimdhp cpuid asimdrdm jscvt fcma lrcpc dcpop sha3 asimddp sha512 sve asimdfhm dit uscat ilrcpc flagm sb paca pacg dcpodp sve2 sveaes svepmull svebitperm svesha3 flagm2 frint svei8mm svebf16 i8mm bf16 dgh rng bti

# コアごとに同じ結果が出力されるため省略

$ cat /proc/cpuinfo | grep -E "CPU (implementer|architecture|part)"
CPU implementer	: 0x41
CPU architecture: 8
CPU part	: 0xd4f
CPU implementer	: 0x41
CPU architecture: 8
CPU part	: 0xd4f

#include <stdio.h>

extern "C" {
    char* openblas_get_config(void);
    char* openblas_get_corename(void);
    int openblas_get_parallel(void);
    int openblas_get_num_threads(void);
}

int main() {
    printf("=== OpenBLAS Configuration ===\n");
    printf("%s\n", openblas_get_config());
    printf("\n=== Runtime Information ===\n");
    printf("Core Name: %s\n", openblas_get_corename());
    printf("Parallel: %d\n", openblas_get_parallel());
    printf("Threads: %d\n", openblas_get_num_threads());
    return 0;
}

# コンパイル
g++ openblas_info.cpp -o openblas_info -lopenblas

# 実行
./openblas_info

$ ./openblas_info
=== OpenBLAS Configuration ===
OpenBLAS 0.3.18 DYNAMIC_ARCH NO_AFFINITY armv8 SINGLE_THREADED

=== Runtime Information ===
Core Name: armv8
Parallel: 0
Threads: 1

#include <cblas.h>
#include <chrono>
#include <iostream>
#include <vector>
#include <unistd.h>

void run_benchmark(const char* label) {
    const int N = 10000;
    std::vector<double> A(N*N, 1.0);
    std::vector<double> B(N*N, 2.0);
    std::vector<double> C(N*N, 0.0);

    auto start = std::chrono::high_resolution_clock::now();

    cblas_dgemm(CblasRowMajor, CblasNoTrans, CblasNoTrans,
                N, N, N, 1.0, A.data(), N, B.data(), N, 0.0, C.data(), N);

    auto end = std::chrono::high_resolution_clock::now();
    auto duration = std::chrono::duration_cast<std::chrono::milliseconds>(end - start);

    std::cout << label << ": " << duration.count() << " ms" << std::endl;
}

int main() {
    std::cout << "Available CPUs: " << sysconf(_SC_NPROCESSORS_ONLN) << std::endl;

    const char* num_threads = getenv("OPENBLAS_NUM_THREADS");
    std::cout << "OPENBLAS_NUM_THREADS: "
              << (num_threads ? num_threads : "not set") << std::endl;

    std::cout << "\nRunning 10000x10000 matrix multiplication...\n" << std::endl;

    run_benchmark("Test 1");
    run_benchmark("Test 2");
    run_benchmark("Test 3");

    return 0;
}

# Serial 版のコンパイル
g++ -O3 performance_test.cpp -o test_serial -lopenblas

# OpenMP 版のコンパイル、OpenMP 有効化のフラグ付き
g++ -O3 performance_test.cpp -o test_parallel -lopenblasp -fopenmp

# Serial 版テスト（1 コアを明示的に指定）
$ export OPENBLAS_NUM_THREADS=1
$ ./test_serial

Available CPUs: 2
OPENBLAS_NUM_THREADS: 1

Running 10000x10000 matrix multiplication...

Test 1: 86053 ms
Test 2: 86054 ms
Test 3: 86072 ms

# OpenMP 版テスト（2 コアを明示的に指定）
$  export OPENBLAS_NUM_THREADS=2
$  ./test_parallel

Available CPUs: 2
OPENBLAS_NUM_THREADS: 2

Running 10000x10000 matrix multiplication...

Test 1: 48214 ms
Test 2: 48250 ms
Test 3: 48217 ms

$ sudo dnf install -y clang19
$ clang++-19 --version
clang version 19.1.7 (AWS 19.1.7-13.amzn2023.0.1)
Target: aarch64-amazon-linux-gnu

clang++-19 -O3 -mcpu=neoverse-v2 -moutline-atomics performance_test.cpp -o test_g4_optimized -lopenblasp -fopenmp

export OPENBLAS_NUM_THREADS=2
./test_g4_optimized

Available CPUs: 2
OPENBLAS_NUM_THREADS: 2

Running 10000x10000 matrix multiplication...

Test 1: 48238 ms
Test 2: 48254 ms
Test 3: 48215 ms