// optimized versions of matrix diagonal summing
#include "matvec.h"

// You can write several different versions of your optimized function
// in this file and call one of them in the last function.

int sumdiag_VER1(matrix_t *mat, vector_t *vec) {
  // OPTIMIZED CODE HERE
  // Initialize Variables
  long mrow = mat->rows;
  long lenV = vec->len;
  int minus = mrow-1;
  vector_t vecX = *vec;
                                                             //Checking for bad length of Vector
  if(lenV != (mrow + minus)){
    printf("sumdiag_base: bad sizes\n");
    return 1;
  }
  for(int i=0; i<(lenV); i++){                               // initialize vector of diagonal sums
    VSET(vecX,i,0);                                          // to all 0s
  }
  int dtemp;                                                 //keeps track of current diagonal
  for(int i =0; i<mrow; i++){                                //go through the 2D array sequentially, adds to correct diagnoal each time
    dtemp = minus-i;                                         //begining diagonal decreases each time
    for(int j = 0; j<mrow; j++){ 
      VSET(vecX,dtemp, VGET(vecX, dtemp) + MGET(*mat,i,j));  //gets the previous value in the sum for a diagonal and add the current num to it
      dtemp++;                                               //continue to next diagonal
    }
  }
  return 0; 
}

int sumdiag_VER2(matrix_t *mat, vector_t *vec) {
   // OPTIMIZED CODE HERE
  // Initialize Variables
  long mrow = mat->rows;
  long lenV = vec->len;
  int minus = mrow-1;
  matrix_t matM = *mat;
  vector_t vecX = *vec;
                                                             //Checking for bad length of Vector
  if(lenV != (mrow + minus)){
    printf("sumdiag_base: bad sizes\n");
    return 1;
  }
  for(int i=0; i<(lenV); i++){                               // initialize vector of diagonal sums
   VSET(vecX,i,0);                                          // to all 0s
  }

  int dtemp;                                                 //keeps track of current diagonal
  int j;
  for(int i =0; i<mrow; i++){                                //go through the 2D array sequentially, adds to correct diagnoal each time
    dtemp = minus-i;                                         //begining diagonal decreases each time
    for(j = 0; j<mrow-4; j+=4){ 
      VSET(vecX,dtemp, VGET(vecX, dtemp) + MGET(matM,i,j));  //gets the previous value in the sum for a diagonal and add the current num to it
      dtemp++;                                               //continue to next diagonal
      VSET(vecX,dtemp,VGET(vecX,dtemp) + MGET(matM,i,j+1));  //unrolled to increase pipeline efficency
      dtemp++;
      VSET(vecX, dtemp,VGET(vecX,dtemp) + MGET(matM,i,j+2));
      dtemp++;
      VSET(vecX,dtemp,VGET(vecX,dtemp)+ MGET(matM,i,j+3));
      dtemp++;
    }
    for(;j<mrow;j++){                                        //clean up if the rows/cols are not divisible by 4
      VSET(vecX,dtemp,VGET(vecX,dtemp) + MGET(matM,i,j));
      dtemp++;
    }
  }
  return 0; 
 
}

int sumdiag_OPTM(matrix_t *mat, vector_t *vec) {
  // call your preferred version of the function
  // OPTIMIZED CODE HERE
  // Initialize Variables
  long mrow = mat->rows;
  long lenV = vec->len;
  int minus = mrow-1;
  matrix_t matM = *mat;
  vector_t vecX = *vec;
                                                             //Checking for bad length of Vector
  if(lenV != (mrow + minus)){
    printf("sumdiag_base: bad sizes\n");
    return 1;
  }
  for(int i=0; i<(lenV); i++){                               // initialize vector of diagonal sums
   VSET(vecX,i,0);                                          // to all 0s
  }

  int dtemp;                                                 //keeps track of current diagonal
  int j;
  for(int i =0; i<mrow; i++){                                //go through the 2D array sequentially, adds to correct diagnoal each time
    dtemp = minus-i;                                         //begining diagonal decreases each time
    for(j = 0; j<mrow-4; j+=4){ 
      VSET(vecX,dtemp, VGET(vecX, dtemp) + MGET(matM,i,j));  //gets the previous value in the sum for a diagonal and add the current num to it
      dtemp++;                                               //continue to next diagonal
      VSET(vecX,dtemp,VGET(vecX,dtemp) + MGET(matM,i,j+1));  //unrolled to increase pipeline efficency
      dtemp++;
      VSET(vecX, dtemp,VGET(vecX,dtemp) + MGET(matM,i,j+2));
      dtemp++;
      VSET(vecX,dtemp,VGET(vecX,dtemp)+ MGET(matM,i,j+3));
      dtemp++;
    }
    for(;j<mrow;j++){                                        //clean up if the rows/cols are not divisible by 4
      VSET(vecX,dtemp,VGET(vecX,dtemp) + MGET(matM,i,j));
      dtemp++;
    }
  }
  return 0; 
 
}