diff --git a/src/bitstream.c b/src/bitstream.c
index 212c234d..17829c70 100644
--- a/src/bitstream.c
+++ b/src/bitstream.c
@@ -19,7 +19,9 @@
  
 #include "global.h"
 #include "bitstream.h"
- 
+
+//#define VERBOSE
+
 #ifdef VERBOSE
 void printf_bitstream(char *msg, ...)
 { 
@@ -32,7 +34,7 @@ void printf_bitstream(char *msg, ...)
 } 
 #endif
 
-bitTable *exp_table;
+bitTable *g_exp_table;
 
 //From wikipedia
 //http://en.wikipedia.org/wiki/Binary_logarithm#Algorithm
@@ -52,14 +54,14 @@ void init_exp_golomb(uint32_t len)
     uint32_t code_num;
     uint32_t M;
     uint32_t info;
-    exp_table=(bitTable*)malloc(len*sizeof(bitTable));    
+    g_exp_table=(bitTable*)malloc(len*sizeof(bitTable));    
     
     for(code_num=0;code_num<len;code_num++)
     {
         M=(uint32_t)floorLog2(code_num+1);
         info=code_num+1-(uint32_t)pow(2,M);        
-        exp_table[code_num].len=M*2+1;
-        exp_table[code_num].value=(1<<M)|info;
+        g_exp_table[code_num].len=M*2+1;
+        g_exp_table[code_num].value=(1<<M)|info;
         //printf_cavlc("Len: %i %x\n", M*2+1, (1<<M)|info);
     }
 }
@@ -71,7 +73,6 @@ void bitstream_init(bitstream* stream)
 {
     stream->cur_byte=0;
     stream->cur_bit=0;
-    stream->output = 0;
     memset(stream->data, 0, sizeof(uint32_t)*32); 
 }
 
@@ -97,51 +98,52 @@ void bitstream_alloc(bitstream* stream, uint32_t alloc)
  
 void bitstream_put(bitstream* stream, uint32_t data, uint8_t bits)
 {
-    uint8_t i=0;
-    uint32_t bitsleft=32-stream->cur_bit;
-    #ifdef VERBOSE
-    printf_bitstream("put: ");
-    for(i=0;i<bits;i++)
-    {
-        printf("%i",(data&(1<<(bits-i-1)))?1:0);
-    }
-    printf_bitstream("\n");
-    //printf_bitstream(" count: %i\n",bits);
-    #endif
+  uint32_t bitsleft=32-stream->cur_bit;
+  #ifdef VERBOSE
+  uint8_t i=0;
+  printf_bitstream("put: ");
+  for(i=0;i<bits;i++)
+  {
+      printf("%i",(data&(1<<(bits-i-1)))?1:0);
+  }
+  printf_bitstream("\n");
+  //printf_bitstream(" count: %i\n",bits);
+  #endif
  
-    //Theres space for all the bits
-    if(bits<=bitsleft)
-    {
-        stream->data[stream->cur_byte] |= (data<<((bitsleft-bits)));
-        stream->cur_bit+=bits;
-        bits=0;
-    }
-    //No space for everything, store the bits we can and continue later
-    else
-    {
-        stream->data[stream->cur_byte] |= (data>>(bits-bitsleft));
-        stream->cur_bit=32;
-        bits-=bitsleft;        
-    }
+  //Theres space for all the bits
+  if(bits<=bitsleft)
+  {
+    stream->data[stream->cur_byte] |= (data<<((bitsleft-bits)));
+    stream->cur_bit+=bits;
+    bits=0;
+  }
+  //No space for everything, store the bits we can and continue later
+  else
+  {
+    stream->data[stream->cur_byte] |= (data>>(bits-bitsleft));
+    stream->cur_bit=32;
+    bits-=bitsleft;
+  }
  
-    //Check if the buffer is full
-    if(stream->cur_bit==32)
+  //Check if the buffer is full
+  if(stream->cur_bit==32)
+  {
+    bitsleft=32;
+    stream->cur_byte++;
+    stream->cur_bit = 0;
+    if(stream->cur_byte==32)
     {
-        stream->cur_byte++;
-        bitsleft=32;
-        if(stream->cur_byte==32)
-        {
-            //Flush data out
-            bitstream_flush(stream);
-        }
+        //Flush data out
+        bitstream_flush(stream);
     }
+  }
  
-    //..still some writing to do
-    if(bits!=0)
-    {
-        stream->data[stream->cur_byte] |= (data<<(bitsleft-bits));
-        stream->cur_bit+=bits;
-    }
+  //..still some writing to do
+  if(bits!=0)
+  {
+    stream->data[stream->cur_byte] |= (data<<(bitsleft-bits));
+    stream->cur_bit+=bits;
+  }
  
  
  
@@ -152,7 +154,7 @@ void bitstream_put(bitstream* stream, uint32_t data, uint8_t bits)
  */
 void bitstream_align(bitstream* stream)
 {  
-  if(stream->cur_bit&7 != 0)
+  if((stream->cur_bit&7) != 0)
   {
     bitstream_put(stream,0, 8-stream->cur_bit&7);
   }
@@ -163,11 +165,12 @@ void bitstream_flush(bitstream* stream)
    /*
     *  SAVE DATA TO OUTPUT
     */
+  int i,j;
   if(stream->output)
   {
     if(stream->cur_byte)
     {
-      fwrite(stream->data, stream->cur_byte*4, 1, stream->output);
+      fwrite(&stream->data[0], stream->cur_byte*4, 1, stream->output);
     }
    
     if(stream->cur_bit>>3)
@@ -175,19 +178,27 @@ void bitstream_flush(bitstream* stream)
       fwrite(&stream->data[stream->cur_byte], stream->cur_bit>>3, 1, stream->output);
     }
   }
+  /* No file open, write to buffer */   
   else
   {
     if(stream->cur_byte)
     {
-      memcpy(&stream->buffer[stream->buffer_pos],&stream->data[0],stream->cur_byte*4);
+      //memcpy((uint8_t*)&stream->buffer[stream->buffer_pos],(uint8_t*)stream->data,stream->cur_byte*4);
+      
+      for(i = 0,j=3; i < stream->cur_byte*4; i++,j--)
+      {
+        if(j == -1) j = 3;
+        stream->buffer[stream->buffer_pos+i] = ((uint8_t*)stream->data)[((i>>2)<<2)+j];
+      }
+      
       stream->buffer_pos += stream->cur_byte*4;
     }
    
-   if(stream->cur_bit>>3)
-   {
-     memcpy(&stream->buffer[stream->buffer_pos],&stream->data[stream->cur_byte],stream->cur_bit>>3);
-     stream->buffer_pos += stream->cur_bit>>3;
-   }    
+    if(stream->cur_bit>>3)
+    {
+      memcpy((uint8_t*)&stream->buffer[stream->buffer_pos],(uint8_t*)&stream->data[stream->cur_byte],stream->cur_bit>>3);
+      stream->buffer_pos += stream->cur_bit>>3;
+    }    
   }
     //Stream flushed, zero out the values
     bitstream_init(stream);
diff --git a/src/bitstream.h b/src/bitstream.h
index 91a304bd..b1def31b 100644
--- a/src/bitstream.h
+++ b/src/bitstream.h
@@ -25,7 +25,7 @@ typedef struct
   uint32_t value;
 }bitTable;
 
-extern bitTable *exp_table;
+extern bitTable *g_exp_table;
 
 int floorLog2(unsigned int n);
  
@@ -36,9 +36,9 @@ void bitstream_put(bitstream* stream, uint32_t data, uint8_t bits);
 void bitstream_put_ue(bitstream* stream, uint32_t data);
 void bitstream_put_se(bitstream* stream, uint32_t data);
 */
-#define bitstream_put_ue(stream, data) { bitstream_put(stream,exp_table[data].value,exp_table[data].len); }
+#define bitstream_put_ue(stream, data) { bitstream_put(stream,g_exp_table[data].value,g_exp_table[data].len); }
 #define bitstream_put_se(stream, data) { uint32_t index=(data<=0)?2*(uint32_t)(-data):2*(uint32_t)(data)-1;    \
-                                         bitstream_put(stream,exp_table[index].value,exp_table[index].len); }
+                                         bitstream_put(stream,g_exp_table[index].value,g_exp_table[index].len); }
 
 void bitstream_align(bitstream* stream); 
 void bitstream_flush(bitstream* stream);
@@ -46,9 +46,9 @@ void init_exp_golomb(uint32_t len);
 
 #ifdef _DEBUG
 static int WRITE_VALUE = 0;
-#define WRITE_U(stream, data, bits, name) { printf("%8d %40s u(%d) : %d\n",WRITE_VALUE, name,bits,data); bitstream_put(stream,data,bits); WRITE_VALUE++;}
-#define WRITE_UE(stream, data, name) { printf("%8d %40s ue(v): %d\n",WRITE_VALUE, name,data); bitstream_put_ue(stream,data); WRITE_VALUE++;}
-#define WRITE_SE(stream, data, name) { printf("%8d %40s se(v): %d\n",WRITE_VALUE, name,data); bitstream_put_se(stream,data); WRITE_VALUE++;}
+#define WRITE_U(stream, data, bits, name) { printf("%8d  %-40s u(%d) : %d\n",WRITE_VALUE, name,bits,data); bitstream_put(stream,data,bits); WRITE_VALUE++;}
+#define WRITE_UE(stream, data, name) { printf("%8d  %-40s ue(v): %d\n",WRITE_VALUE, name,data); bitstream_put_ue(stream,data); WRITE_VALUE++;}
+#define WRITE_SE(stream, data, name) { printf("%8d  %-40s se(v): %d\n",WRITE_VALUE, name,data); bitstream_put_se(stream,data); WRITE_VALUE++;}
 #else
 #define WRITE_U(stream, data, bits, name) { bitstream_put(stream,data,bits); }
 #define WRITE_UE(stream, data, name) { bitstream_put_ue(stream,data); }
diff --git a/src/encmain.c b/src/encmain.c
index c2f7f697..d35fe396 100644
--- a/src/encmain.c
+++ b/src/encmain.c
@@ -80,14 +80,14 @@
       return EXIT_FAILURE;
     }
 
-	  printf("Input: %s, output: %s\r\n", cfg->input, cfg->output);
-    printf("  Video size: %dx%d\r\n", cfg->width, cfg->height);
+	  printf("Input: %s, output: %s\n", cfg->input, cfg->output);
+    printf("  Video size: %dx%d\n", cfg->width, cfg->height);
 
     /* Open input file and check that it was opened correctly */
     input = fopen(cfg->input, "rb");
     if(input == NULL)
     {
-      fprintf(stderr, "Could not open input file, shutting down!\r\n");
+      fprintf(stderr, "Could not open input file, shutting down!\n");
       config_destroy(cfg);
       return EXIT_FAILURE;
     }
@@ -96,7 +96,7 @@
     output = fopen(cfg->output, "wb");
     if(output == NULL)
     {
-      fprintf(stderr, "Could not open output file, shutting down!\r\n");
+      fprintf(stderr, "Could not open output file, shutting down!\n");
       config_destroy(cfg);
       return EXIT_FAILURE;
     }
@@ -109,6 +109,7 @@
     /* Init bitstream */
     bitstream_init(encoder->stream);
     encoder->stream->buffer_pos = 0;
+    encoder->stream->output = 0;
     bitstream_alloc(encoder->stream, 1024*1024);
 
     /* Config pointer to encoder struct */
@@ -134,7 +135,7 @@
     }
     /* Coding finished */
 
-
+    printf(" Processed %d frames\n", encoder->frame-1);
 
     fclose(input);
     fclose(output);
diff --git a/src/encoder.c b/src/encoder.c
index 021561de..125ca558 100644
--- a/src/encoder.c
+++ b/src/encoder.c
@@ -56,7 +56,7 @@ void encode_one_frame(encoder_control* encoder)
     bitstream_align(encoder->stream);
     bitstream_flush(encoder->stream);
     nal_write(encoder->output, encoder->stream->buffer, encoder->stream->buffer_pos, 1, NAL_PIC_PARAMETER_SET, 0);
-    
+
   }
 
 }
diff --git a/src/nal.c b/src/nal.c
index 7ffb4ee9..5d25637c 100644
--- a/src/nal.c
+++ b/src/nal.c
@@ -62,7 +62,7 @@ void nal_write(FILE* output, uint8_t* buffer, uint32_t buffer_len, uint8_t nal_r
     fwrite(&buffer[i], 1, 1, output);
   }
   //If last byte was 0, add emulation_prevention_three_byte
-  if(buffer[buffer_len-1] == 0)
-    fwrite(&emulation_prevention_three_byte, 1, 1, output);
+  //if(buffer[buffer_len-1] == 0)
+  //  fwrite(&emulation_prevention_three_byte, 1, 1, output);
 
 }
\ No newline at end of file