diff --git a/cs677/hw4/src/transpose.cc b/cs677/hw4/src/transpose.cc
index 361033d..a81bbee 100644
--- a/cs677/hw4/src/transpose.cc
+++ b/cs677/hw4/src/transpose.cc
@@ -1,8 +1,6 @@
 
 #include <iostream>
 #include <sys/time.h>
-#include <unistd.h>     /* gethostname() */
-#include <stdlib.h>     /* rand() */
 #include <mpi.h>
 using namespace std;
 
@@ -56,11 +54,13 @@ int main(int argc, char * argv[])
     }
 #endif
 
-    /* Initialize the matrix */
+    /* Initialize the matrices */
     int matrix[n][n];
-    for (int i = 0; i < n; i++)
-        for (int j = 0; j < n; j++)
+    for (int i = 0; i < n; i++)         /* zero the matrices for prettier */
+        for (int j = 0; j < n; j++)     /* printing */
             matrix[i][j] = 0;
+
+    /* Determine which rows I am responsible for */
     int my_first_row;
     int my_num_rows;
     taskAllocate(n, p, my_rank, &my_first_row, &my_num_rows);
@@ -85,25 +85,21 @@ int main(int argc, char * argv[])
         total += count;
     }
 
-    /* Transpose the matrix with p gather operations */
-    int recvbuf[p];
-    for (int i = 0; i < p; i++)
+    /* Transpose the matrix with n gather operations */
+    for (int i = 0, toproc = 0, proccount = counts[0]; i < n; i++)
     {
-        int col_i_vals_for_proc_p[my_num_rows];
+        int my_col_i_vals[my_num_rows];
         for (int row_offset = 0; row_offset < my_num_rows; row_offset++)
+            my_col_i_vals[row_offset] = matrix[my_first_row + row_offset][i];
+        MPI_Gatherv(&my_col_i_vals[0], my_num_rows, MPI_INT,
+                    &matrix[i][0], &counts[0], &displs[0],
+                    MPI_INT, toproc, MPI_COMM_WORLD);
+        proccount--;
+        if (proccount <= 0)
         {
-            col_i_vals_for_proc_p[row_offset] =
-                matrix[my_first_row + row_offset][i];
+            toproc++;
+            proccount = ccounts[toproc];
         }
-        MPI_Gatherv(&col_i_vals_for_proc_p[0], my_num_rows, MPI_INT,
-                    &recvbuf[0], &counts[0], &displs[0],
-                    MPI_INT, i, MPI_COMM_WORLD);
-    }
-
-    /* Put my received entries into my columns of the matrix */
-    for (int i = 0; i < p; i++)
-    {
-        matrix[i][my_rank] = recvbuf[i];
     }
 
     /* Print out the final matrix */
@@ -114,7 +110,7 @@ int main(int argc, char * argv[])
         {
             for (int j = 0; j < n; j++)
             {
-                cout << matrix[i][j] << " ";
+                cout << recvmatrix[i][j] << " ";
             }
             cout << endl;
         }