From f1ab745063ac6e7a34c66763fdacb01773e90e65 Mon Sep 17 00:00:00 2001
From: anna_schlittenhardt <anna.schlittenhardt@stud.uni-due.de>
Date: Sat, 19 Dec 2020 12:39:46 +0100
Subject: [PATCH] 32x32 matrix now with <300 misses

---
 src/trans.c | 33 +++++++++++++++++++++++++++++++--
 1 file changed, 31 insertions(+), 2 deletions(-)

diff --git a/src/trans.c b/src/trans.c
index 75ce9a0..ecbe6e2 100644
--- a/src/trans.c
+++ b/src/trans.c
@@ -29,14 +29,23 @@ void transpose_submit(int M, int N, int A[N][M], int B[M][N])
 {   
     //343 misses (goal: <300; 0 points if >600)
     if(N == 32){
-        int n, m, nn, mm;
+        int n, m, nn, mm, diag_pos, diag_value;
         
         for (n=0; n<N;n+=8){
             for (m=0;m<M; m+=8){
 
                 for(nn=n; nn<n+8;nn++){
                     for(mm=m; mm<m+8;mm++){
-                       B[mm][nn] = A[nn][mm]; 
+                       
+                        if(nn!=mm){
+                            B[mm][nn] = A[nn][mm]; 
+                        }else{
+                            diag_pos = nn;
+						    diag_value = A[nn][mm];
+                        }
+                    }
+                    if(n==m){
+                        B[diag_pos][diag_pos]= diag_value;
                     }
                 }
             }
@@ -44,7 +53,27 @@ void transpose_submit(int M, int N, int A[N][M], int B[M][N])
         
     }
     if(N == 64){
+        int n, m, nn, mm, diag_pos, diag_value;
+        
+        for (n=0; n<N;n+=4){
+            for (m=0;m<M; m+=4){
 
+                for(nn=n; nn<n+4;nn++){
+                    for(mm=m; mm<m+4;mm++){
+                       
+                        if(nn!=mm){
+                            B[mm][nn] = A[nn][mm]; 
+                        }else{
+                            diag_pos = nn;
+						    diag_value = A[nn][mm];
+                        }
+                    }
+                    if(n==m){
+                        B[diag_pos][diag_pos]= diag_value;
+                    }
+                }
+            }
+        }
     }
 }