Math Library Benchmark: GPU vs. JBLAS vs. pure Java

MathLibBenchmark.java

package de.jungblut.benchmark;

import java.util.Random;

import com.google.caliper.Param;
import com.google.caliper.Runner;
import com.google.caliper.SimpleBenchmark;

import de.jungblut.math.DoubleMatrix;
import de.jungblut.math.cuda.JCUDAMatrixUtils;
import de.jungblut.math.dense.DenseDoubleMatrix;

public class MathLibBenchmark extends SimpleBenchmark {

  @Param({ "10", "20", "40", "60", "80", "100", "500", "1000", "2000" })
  private int n;

  @Param
  CalcType type;

  private DenseDoubleMatrix mat;
  private DenseDoubleMatrix mat2;

  public enum CalcType {
    GPU, JBLAS, TJ_MATH
  };

  @Override
  protected void setUp() throws Exception {
    mat = new DenseDoubleMatrix(n, n, new Random());
    mat2 = new DenseDoubleMatrix(n, n, new Random());
  }

  public void timeCalculate(int reps) {
    for (int rep = 0; rep < reps; rep++) {
      int sum = 0;
      switch (type) {
        case JBLAS:
          sum = jblas(sum);
          break;
        case TJ_MATH:
          sum = tjmath(sum);
          break;
        case GPU:
          sum = gpu(sum);
          break;
        default:
          break;
      }
      System.out.println(sum);
    }

  }

  private int gpu(int sum) {
    DenseDoubleMatrix multiply2 = JCUDAMatrixUtils.multiply(mat, mat2);
    sum += multiply2.getColumnCount();
    return sum;
  }

  private int tjmath(int sum) {
    DoubleMatrix multiply = mat.multiply(mat2);
    sum += multiply.getRowCount();
    return sum;
  }

  private int jblas(int sum) {
    org.jblas.DoubleMatrix jblasThis = new org.jblas.DoubleMatrix(mat.toArray());
    org.jblas.DoubleMatrix jblasOther = new org.jblas.DoubleMatrix(
        mat2.toArray());
    org.jblas.DoubleMatrix jblasRes = new org.jblas.DoubleMatrix(n, n);
    jblasThis.mmuli(jblasOther, jblasRes);
    sum += jblasRes.columns;
    return sum;
  }

  public static void main(String[] args) {
    Runner.main(MathLibBenchmark.class, args);
  }
}