Question

I was reading the paper, http://www.cs.utexas.edu/~lin/papers/hpca01.pdf, on Dynamic Branch Prediction with Perceptrons. I was wondering how to implement the perceptron branch predictor in C if given a list of 1000 PC addresses (word addresses) and 1000 number of actual outcome of the branch which are recorded in a trace line. Essentially, I want to use these traces to measure the accuracy of various predictors. The branch outcomes from the trace file should be used to train your predictors. Any suggestions?

Était-ce utile?

La solution

I think its fairly simple. Section 3.2 and 3.3 is all you really have to understand.

Section 3.2 says output percepatron is sum of past histories multipled by their wieghting factors:

#define SIZE_N 62 //or whatever see section 5.3
float history[n] = {0}; //Put branch history here, -1 not taken, 1 taken.
float weight[n] = {0};  //storage for weights

float percepatron(void )
{
    int i;
    float y=0;
    for (i=0;i<SIZE_N;i++) { y+= weight[i] * history[i];}
    return y;
}

Then in 3.3 the weighting factors come from training, which is simply train each one past on past result comparison:

void train(float result, float y, float theta) //passed result of last branch (-1 not taken, 1 taken), and perceptron value
{
    int i;
    if ((y<0) != (result<0)) || (abs(y) < theta))
    {
     for (i=0;i<SIZE_N;i++;) {
          weight[i] = weight[i] + result*history[i];
       }
    }
}

So all thats left is theta, which they tell you:

float theta = (1.93 * SIZE_N) + 14;

So the usage is:

y = percepatron();
//make prediction:
if (y < 0) predict_not_taken();
else predict_taken();
//get actual result
result = get_actual_branch_taken_result();//must return -1 not taken, 1 taken
//train for future predictions
train(y,result,theta);

//Then you need to shift everything down....
for (i=1;i<SIZE_N;i++)
{
  history[i] = history[i-1];
  //weight[i] = history[i-1]; //toggle this and see what happens :-)
}
history[0] = 1; //weighting - see section 3.2
Licencié sous: CC-BY-SA avec attribution
Non affilié à StackOverflow
scroll top