设备环境介绍

easily-openJCL 是一个轻量级的 Java 语言下的 GPU 显卡 计算库，它提供了一套简单易用的 API，让用户能够轻松实现 GPU 计算操作。

通过 Java 调用 GPU 计算的一个库，使用非常简单的API就可以轻松应付 Java 数据类型在 GPU 中的计算操作！easily-openJCL 提供了诸多中计算模式，让我们的计算组件更灵活！

我们通过基准测试来对比 GPU 和 CPU 的差距，在这里的基准测试中，使用的硬件设备如下所示

基准测试实战

在这里准备了一份代码，其中我们使用 CPU 和 GPU 分别计算了一个相同的 复杂的数学运算逻辑，并将两个设备的测试结果进行展示。

基准测试代码

导入 maven 依赖

    <dependencies>
        <!-- https://mvnrepository.com/artifact/io.github.BeardedManZhao/easily-openJCL -->
        <dependency>
            <groupId>io.github.BeardedManZhao</groupId>
            <artifactId>easily-openJCL</artifactId>
            <version>1.0.1</version>
        </dependency>
    </dependencies>

package org.example;

import io.github.BeardedManZhao.easilyJopenCL.EasilyOpenJCL;
import io.github.BeardedManZhao.easilyJopenCL.kernel.KernelSource;

public class Main {
    public static void main(String[] args0) {
        // 这个地方的数学表达式 以及下面 for 循环中的 数学表达式 都是相同的，只是为了测试GPU 和 CPU 的计算差异，这个表达式越复杂 计算量就会越大，但数据拷贝量不会有变化，这可能会让 CPU 处于劣势
        final KernelSource kernelSource = new KernelSource(args -> "c[gid] = ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) + ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) + ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) + ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (1024 * a[gid]))))) - ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) + ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) + ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) + ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (1024 * a[gid]))))) + 1;", "double", "MY");
        final EasilyOpenJCL easilyOpenJCL = EasilyOpenJCL.initOpenCLEnvironment(
                kernelSource
        );
        // 这个数组的长度代表的是操作数的数据量 这个数值越大 计算量会成倍增大，同时数据拷贝量也会成倍增加，这将会让 GPU 计算处于劣势
        final double[] a = new double[Integer.MAX_VALUE >> 6];
        for (int i = 0; i < a.length; i++) {
            a[i] = i + 10;
        }
        final double[] b = {2};
        final double[] c = new double[a.length];

        // 热身
        easilyOpenJCL.calculate(a, b, c, kernelSource);

        // 正式计算
        final long l = System.currentTimeMillis();
        easilyOpenJCL.calculate(a, b, c, kernelSource);
        final long l1 = System.currentTimeMillis();
        System.out.println("GPU计算耗时：" + (l1 - l) + " 第一个元素结果 = " + c[0]);

        for (int gid = 0; gid < a.length; gid++) {
            c[gid] = ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) + ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) + ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) + ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (1024 * a[gid]))))) - ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) + ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) + ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) + ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - 1024)))) * ((a[gid] / b[0] * a[0] + (a[gid] / 100 * 33 + 210.5 / 5)) * 2) * 2 / 3 + (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (a[gid] * a[gid] - (1024 * a[gid]))))) + 1;
        }
        final long l2 = System.currentTimeMillis();
        System.out.println("CPU计算耗时：" + (l2 - l1) + " 第一个元素结果 = " + c[0]);
    }
}

测试结果

由于不同的设备性能不同，因此测试结果可能会有些差距！

GPU计算耗时：175 第一个元素结果 = 1069410.6
CPU计算耗时：326 第一个元素结果 = 1069410.6