目前正在开发用于矩阵运算和机器学习的 Java 库。 我使用 Aparapi 来利用 GPU。
我编写了这段代码来将两个矩阵相乘:
public static NDmatrix matMul(float[][] a, float[][] b) {
int[] aDim = new int[]{a.length, a[0].length};
int[] bDim = new int[]{b.length, b[0].length};
if(aDim[1] == bDim[0]){
int[] Dim = new int[]{aDim[0], bDim[1]};
int aVSize = aDim[0] * aDim[1];
float[] aVector = new float[aVSize];
for(int i = 0; i < aDim[0]; i++)
System.arraycopy(a[i], 0, aVector, i * aDim[1], aDim[1]);
int bVSize = bDim[0] * bDim[1];
float[] bVector = new float[bVSize];
for(int i = 0; i < bDim[0]; i++)
System.arraycopy(b[i], 0, bVector, i * bDim[1], bDim[1]);
int resVSize = Dim[0] * Dim[1];
float[] resVector = new float[resVSize];
int d[] = new int[]{aDim[1]};
Kernel mKernel = new Kernel() {
final int ht = Dim[0];
final int wt = Dim[1];
final int dpt = d[0];
public void run() {
int c = getGlobalId(0);
int r = getGlobalId(1);
int l = getGlobalId(2);
localBarrier();
//for(int l = 0; l < dpt; l++)
resVector[r * wt + c] = resVector[r * wt + c] + aVector[r * dpt + l] * bVector[l * wt + c];
}
};
mKernel.setExplicit(true);
mKernel.put(aVector);
mKernel.put(bVector);
mKernel.put(resVector);
mKernel.put(Dim);
mKernel.put(d);
mKernel.execute(Range.create3D(Dim[1], Dim[0], d[0]));
//mKernel.execute(Range.create2D(Dim[1], Dim[0]));
mKernel.get(resVector);
mKernel.dispose();
return new NDmatrix(Dim, resVector, null);
}
System.out.println("The number of columns in left matrix and number of rows in right matrix do not match.");
System.out.println();
return null;
}
但是,看起来 resVector[some_index] 只更新一次。 相反,如果我使用 2D 范围和循环(代码中的注释位),则它可以正常工作。 这种行为的原因可能是什么?我如何强制它完全并行工作?
有趣的是,我尝试了一件“有效”的事情 - 更新 resVector[some_index] 后,我调用了 this.put(resVector)。 然而,它无法在 OpenCL 中编译,最终改用 Java 的多线程,最终得到了正确的结果。
好吧,我想我找到了解决方案,虽然有点疯狂:
public static NDmatrix matMul(float[][] a, float[][] b) {
int[] aDim = new int[]{a.length, a[0].length};
int[] bDim = new int[]{b.length, b[0].length};
if(aDim[1] == bDim[0]){
int[] Dim = new int[]{aDim[0], bDim[1]};
int aVSize = aDim[0] * aDim[1];
float[] aVector = new float[aVSize];
for(int i = 0; i < aDim[0]; i++)
System.arraycopy(a[i], 0, aVector, i * aDim[1], aDim[1]);
int bVSize = bDim[0] * bDim[1];
float[] bVector = new float[bVSize];
for(int i = 0; i < bDim[0]; i++)
System.arraycopy(b[i], 0, bVector, i * bDim[1], bDim[1]);
int resVSize = Dim[0] * Dim[1];
float[] resVector = new float[resVSize];
int d[] = new int[]{aDim[1]};
return getMult(Dim, aVector, bVector, resVector, d);
}
System.out.println("The number of columns in left matrix and number of rows in right matrix do not match.");
System.out.println();
return null;
}
@NotNull
private static NDmatrix getMult(int[] Dim, float[] aVector, float[] bVector, float[] resVector, int[] d) {
Kernel mKernel = new Kernel() {
final int ht = Dim[0];
final int wt = Dim[1];
final int dpt = d[0];
public void run() {
int c = getGlobalId(0);
int r = getGlobalId(1);
int l = getGlobalId(2);
localBarrier();
for(int i = 0; i < dpt; i++) // really?!
if(i == l) // why, oh why do I have to do this...
resVector[r * wt + c] += aVector[r * dpt + i] * bVector[i * wt + c];
}
};
mKernel.setExplicit(true);
mKernel.put(aVector);
mKernel.put(bVector);
mKernel.put(resVector);
mKernel.put(Dim);
mKernel.put(d);
mKernel.execute(Range.create3D(Dim[1], Dim[0], d[0]));
mKernel.get(resVector);
mKernel.dispose();
return new NDmatrix(Dim, resVector, null);
}
不敢相信这样一个奇怪的循环会以某种方式迫使它做它的事情......