diff --git a/src/bitstream.c b/src/bitstream.c
index 382375d2..2b1c3b85 100644
--- a/src/bitstream.c
+++ b/src/bitstream.c
@@ -35,6 +35,18 @@
 #include <arpa/inet.h>
 #endif
 
+const uint32_t bit_set_mask[] =
+{
+0x00000001,0x00000002,0x00000004,0x00000008,
+0x00000010,0x00000020,0x00000040,0x00000080,
+0x00000100,0x00000200,0x00000400,0x00000800,
+0x00001000,0x00002000,0x00004000,0x00008000,
+0x00010000,0x00020000,0x00040000,0x00080000,
+0x00100000,0x00200000,0x00400000,0x00800000,
+0x01000000,0x02000000,0x04000000,0x08000000,
+0x10000000,0x20000000,0x40000000,0x80000000
+};
+
 
 //#define VERBOSE
 
@@ -93,7 +105,7 @@ int init_exp_golomb(uint32_t len)
 /**
  * \brief Create and initialize a new bitstream
  */
-bitstream *create_bitstream(int32_t width)
+bitstream *create_bitstream()
 {
   bitstream *stream = malloc(sizeof(bitstream));
   if (!stream) {
@@ -101,74 +113,13 @@ bitstream *create_bitstream(int32_t width)
     return stream;
   }
 
-  // Initialize the bitstream
-  bitstream_reinit(stream);
-
   // Initialize buffer-related values
-  stream->output     = NULL;
-  stream->buffer     = NULL;
-  stream->buffer_pos = 0;
-  stream->bufferlen  = 0;
-
-  // Alloc 2kB*width for bitstream buffer (for one coded frame)
-  bitstream_alloc(stream, 1024*2*width);
-  if (!stream->buffer) {
-    fprintf(stderr, "Failed to allocate the bitstream buffer!\n");
-    goto creation_failure;
-  }
-
-  //Clear buffer just to be sure
-  bitstream_clear_buffer(stream);
+  stream->data = 0;
+  stream->cur_bit = 0;
+  stream->zerocount = 0;
 
   // Return the created bitstream
   return stream;
-
-creation_failure:
-  // In case of failure, free whatever was allocated
-  free(stream->buffer);
-  free(stream);
-
-  return NULL;
-}
-
-/**
- * \brief Reinitialize bitstream
- */
-void bitstream_reinit(bitstream *stream)
-{
-  stream->cur_byte = 0;
-  stream->cur_bit = 0;
-  memset(stream->data, 0, sizeof(uint32_t)*32);
-}
-
-/**
- *  \brief Allocate buffer
- *  \param stream pointer bitstream to put the data
- *  \param alloc size to allocate
- */
-void bitstream_alloc(bitstream *stream, uint32_t alloc)
-{
-  stream->buffer = (uint8_t*)malloc(alloc);
-  if (stream->buffer)
-    stream->bufferlen = alloc;
-}
-
-/**
- *  \brief Free bitstream buffer
- */
-void bitstream_free(bitstream *stream)
-{
-  free(stream->buffer);
-  stream->bufferlen = 0;
-}
-
-/**
- *  \brief clear output buffer
- */
-void bitstream_clear_buffer(bitstream *stream)
-{
-  memset(stream->buffer, 0, stream->bufferlen);
-  stream->buffer_pos = 0;
 }
 
 /**
@@ -179,43 +130,29 @@ void bitstream_clear_buffer(bitstream *stream)
  */
 void bitstream_put(bitstream *stream, uint32_t data, uint8_t bits)
 {
-  uint8_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
+  const uint8_t emulation_prevention_three_byte = 0x03;
+  while(bits--) {
+    stream->data <<= 1;
 
-  //There's space for all the bits
-  if (bits <= bitsleft) {
-    stream->data[stream->cur_byte] |= (data<<((bitsleft-bits)));
-    stream->cur_bit += bits;
-    bits = 0;
-  } else { //No space for everything, store the bits we can and continue later
-    stream->data[stream->cur_byte] |= (data>>(bits-bitsleft));
-    stream->cur_bit = 32;
-    bits -= bitsleft;
-  }
-
-  //Check if the buffer is full, and flush to output if it is
-  if (stream->cur_bit == 32) {
-    bitsleft = 32;
-    stream->cur_byte++;
-    stream->cur_bit = 0;
-    if (stream->cur_byte == 32) {
-      //Flush data out
-      bitstream_flush(stream);
+    if (data & bit_set_mask[bits]) {
+      stream->data |= 1;
     }
-  }
+    stream->cur_bit++;
 
-  //Write the last of the bits (if buffer was full and flushed before)
-  if (bits != 0) {
-    stream->data[stream->cur_byte] |= (data<<(bitsleft-bits));
-    stream->cur_bit += bits;
+  // write byte to output
+    if (stream->cur_bit==8) {
+      if((stream->zerocount == 2) && (stream->data < 4)) {
+        fwrite(&emulation_prevention_three_byte, 1, 1, stream->output);
+        stream->zerocount = 0;
+      }
+      if(stream->data == 0) {
+        stream->zerocount++;
+      } else {
+        stream->zerocount = 0;
+      }
+      fwrite(&stream->data, 1, 1, stream->output);
+      stream->cur_bit = 0;
+    }
   }
 }
 
@@ -238,36 +175,4 @@ void bitstream_align_zero(bitstream *stream)
   if ((stream->cur_bit & 7) != 0) {
     bitstream_put(stream, 0, 8 - (stream->cur_bit & 7));
   }
-}
-
-/**
- * \brief Flush bitstream to output
- */
-void bitstream_flush(bitstream *stream)
-{
-  int i;
-  uint32_t correct_endian;
-  //If output open, write to output
-  if (stream->output) {
-    if (stream->cur_byte) fwrite(&stream->data[0], stream->cur_byte * 4, 1, stream->output);
-    if (stream->cur_bit>>3) fwrite(&stream->data[stream->cur_byte], stream->cur_bit>>3, 1, stream->output);
-  } else { //No file open, write to buffer
-    if (stream->cur_byte) {
-      //Handle endianness issue
-      for (i = 0; i < stream->cur_byte; i++) {
-        //"network" is big-endian
-        correct_endian = htonl(stream->data[i]);
-        memcpy((uint8_t*)&stream->buffer[stream->buffer_pos], &correct_endian, 4);
-        stream->buffer_pos += 4;
-      }
-    }
-
-    if (stream->cur_bit>>3) {
-      correct_endian = htonl(stream->data[stream->cur_byte]);
-      memcpy((uint8_t*)&stream->buffer[stream->buffer_pos], &correct_endian, stream->cur_bit>>3);
-      stream->buffer_pos += stream->cur_bit>>3;
-    }
-  }
-  //Stream flushed, zero out the values
-  bitstream_reinit(stream);
-}
+}
\ No newline at end of file
diff --git a/src/bitstream.h b/src/bitstream.h
index 7d4ba5aa..6e5413eb 100644
--- a/src/bitstream.h
+++ b/src/bitstream.h
@@ -29,13 +29,10 @@
 
 typedef struct
 {
-    uint32_t data[32];
-    uint8_t  cur_byte;
-    uint8_t  cur_bit;
-    FILE*    output;
-    uint8_t* buffer;
-    uint32_t buffer_pos;
-    uint32_t bufferlen;
+  uint8_t data;
+  uint8_t cur_bit;
+  uint8_t zerocount;
+  FILE*    output;
 } bitstream;
 
 typedef struct
@@ -48,11 +45,7 @@ extern bit_table *g_exp_table;
 
 int floor_log2(unsigned int n);
 
-bitstream *create_bitstream(int32_t width);
-void bitstream_alloc(bitstream* stream, uint32_t alloc);
-void bitstream_free(bitstream *stream);
-void bitstream_clear_buffer(bitstream* stream);
-void bitstream_reinit(bitstream *stream);
+bitstream *create_bitstream();
 void bitstream_put(bitstream* stream, uint32_t data, uint8_t bits);
 
 /* Use macros to force inlining */
@@ -62,7 +55,6 @@ void bitstream_put(bitstream* stream, uint32_t data, uint8_t bits);
 
 void bitstream_align(bitstream* stream);
 void bitstream_align_zero(bitstream* stream);
-void bitstream_flush(bitstream* stream);
 int init_exp_golomb(uint32_t len);
 
 
diff --git a/src/encmain.c b/src/encmain.c
index 236d5b87..f33428a6 100644
--- a/src/encmain.c
+++ b/src/encmain.c
@@ -208,8 +208,9 @@ int main(int argc, char *argv[])
     return EXIT_FAILURE;
 
   // Set output file
-  encoder->output = output;
 
+  encoder->output = output;
+  encoder->stream->output = output;
   // input init (TODO: read from commandline / config)
   encoder->bitdepth = 8;
   encoder->frame    = 0;
@@ -386,7 +387,6 @@ int main(int argc, char *argv[])
   picture_list_destroy(encoder->ref);
   picture_destroy(encoder->in.cur_pic);
   FREE_POINTER(encoder->in.cur_pic);
-  bitstream_free(encoder->stream);
   FREE_POINTER(encoder->stream);
   free(encoder);
   free_tables();
diff --git a/src/encoder.c b/src/encoder.c
index 003387a9..99d7faac 100644
--- a/src/encoder.c
+++ b/src/encoder.c
@@ -279,7 +279,7 @@ encoder_control *init_encoder_control(config *cfg)
   enc_c->sao_enable = 1;
 
   // Allocate the bitstream struct
-  stream = create_bitstream(enc_c->cfg->width);
+  stream = create_bitstream();
   if (!stream) {
     fprintf(stderr, "Failed to allocate the bitstream object!\n");
     goto init_failure;
@@ -377,11 +377,8 @@ void init_encoder_input(encoder_input *input, FILE *inputfile,
 static void write_aud(encoder_control* encoder)
 {
   encode_access_unit_delimiter(encoder);
+  nal_write(encoder->output, 0, AUD_NUT, 0, 1);
   bitstream_align(encoder->stream);
-  bitstream_flush(encoder->stream);
-  nal_write(encoder->output, encoder->stream->buffer,
-            encoder->stream->buffer_pos, 0, AUD_NUT, 0, 1);
-  bitstream_clear_buffer(encoder->stream);
 }
 
 void encode_one_frame(encoder_control* encoder)
@@ -424,37 +421,25 @@ void encode_one_frame(encoder_control* encoder)
       write_aud(encoder);
 
     // Video Parameter Set (VPS)
+    nal_write(encoder->output, 0, NAL_VPS_NUT, 0, 1);
     encode_vid_parameter_set(encoder);
     bitstream_align(encoder->stream);
-    bitstream_flush(encoder->stream);
-    nal_write(encoder->output, encoder->stream->buffer,
-              encoder->stream->buffer_pos, 0, NAL_VPS_NUT, 0, 1);
-    bitstream_clear_buffer(encoder->stream);
 
     // Sequence Parameter Set (SPS)
+    nal_write(encoder->output, 0, NAL_SPS_NUT, 0, 1);
     encode_seq_parameter_set(encoder);
     bitstream_align(encoder->stream);
-    bitstream_flush(encoder->stream);
-    nal_write(encoder->output, encoder->stream->buffer,
-              encoder->stream->buffer_pos, 0, NAL_SPS_NUT, 0, 1);
-    bitstream_clear_buffer(encoder->stream);
 
     // Picture Parameter Set (PPS)
+    nal_write(encoder->output, 0, NAL_PPS_NUT, 0, 1);
     encode_pic_parameter_set(encoder);
     bitstream_align(encoder->stream);
-    bitstream_flush(encoder->stream);
-    nal_write(encoder->output, encoder->stream->buffer,
-              encoder->stream->buffer_pos, 0, NAL_PPS_NUT, 0, 1);
-    bitstream_clear_buffer(encoder->stream);
 
     if (encoder->frame == 0) {
       // Prefix SEI
+      nal_write(encoder->output, 0, PREFIX_SEI_NUT, 0, 0);
       encode_prefix_sei_version(encoder);
       bitstream_align(encoder->stream);
-      bitstream_flush(encoder->stream);
-      nal_write(encoder->output, encoder->stream->buffer,
-                encoder->stream->buffer_pos, 0, PREFIX_SEI_NUT, 0, 0);
-      bitstream_clear_buffer(encoder->stream);
     }
   } else {
     // When intra period == 1, all pictures are intra
@@ -466,6 +451,15 @@ void encode_one_frame(encoder_control* encoder)
       write_aud(encoder);
   }
 
+  {
+    // Not quite sure if this is correct, but it seems to have worked so far
+    // so I tried to not change it's behavior.
+    int long_start_code = is_radl_frame || encoder->aud_enable ? 0 : 1;
+
+    nal_write(encoder->output,  0,
+              is_radl_frame ? NAL_IDR_W_RADL : NAL_TRAIL_R, 0, long_start_code);
+  }
+
   cabac_start(&cabac);
   init_contexts(encoder, encoder->in.cur_pic->slicetype);
   scalinglist_process();
@@ -561,19 +555,6 @@ void encode_one_frame(encoder_control* encoder)
 
   cabac_flush(&cabac);
   bitstream_align(encoder->stream);
-  bitstream_flush(encoder->stream);
-
-  {
-    // Not quite sure if this is correct, but it seems to have worked so far
-    // so I tried to not change it's behavior.
-    int long_start_code = is_radl_frame || encoder->aud_enable ? 0 : 1;
-
-    nal_write(encoder->output, encoder->stream->buffer,
-              encoder->stream->buffer_pos, 0,
-              is_radl_frame ? NAL_IDR_W_RADL : NAL_TRAIL_R, 0, long_start_code);
-  }
-
-  bitstream_clear_buffer(encoder->stream);
 
   if (encoder->sao_enable) {
     sao_reconstruct_frame(encoder);
@@ -679,6 +660,8 @@ static void add_checksum(encoder_control* encoder)
   uint32_t checksum_val;
   unsigned int i;
 
+  nal_write(encoder->output, 0, NAL_SUFFIT_SEI_NUT, 0, 0);
+
   picture_checksum(encoder->in.cur_pic, checksum);
 
   WRITE_U(encoder->stream, 132, 8, "sei_type");
@@ -694,10 +677,6 @@ static void add_checksum(encoder_control* encoder)
   }
 
   bitstream_align(encoder->stream);
-  bitstream_flush(encoder->stream);
-  nal_write(encoder->output, encoder->stream->buffer,
-            encoder->stream->buffer_pos, 0, NAL_SUFFIT_SEI_NUT, 0, 0);
-  bitstream_clear_buffer(encoder->stream);
 }
 
 void encode_access_unit_delimiter(encoder_control* encoder)
diff --git a/src/nal.c b/src/nal.c
index fb0fe02f..24c57d56 100644
--- a/src/nal.c
+++ b/src/nal.c
@@ -35,16 +35,12 @@
 /**
  * \brief Write a Network Abstraction Layer (NAL) packet to the output.
  */
-void nal_write(FILE *output, uint8_t *buffer, uint32_t buffer_len,
-               uint8_t nal_ref, uint8_t nal_type, uint8_t temporal_id,
-               int long_start_code)
+void nal_write(FILE *output,uint8_t nal_ref, uint8_t nal_type,
+               uint8_t temporal_id,int long_start_code)
 {
   uint8_t byte;
-  uint32_t i;
-  uint8_t zerocount = 0;
 
   // Some useful constants
-  const uint8_t emulation_prevention_three_byte = 0x03;
   const uint8_t start_code_prefix_one_3bytes = 0x01;
   const uint8_t zero = 0x00;
 
@@ -70,29 +66,6 @@ void nal_write(FILE *output, uint8_t *buffer, uint32_t buffer_len,
   // 5bits of nuh_layer_id + nuh_temporal_id_plus1(3)
   byte = (temporal_id + 1) & 7;
   fwrite(&byte, 1, 1, output);
-
-  // Write out bytes and add emulation_prevention_three_byte when needed
-  for (i = 0; i < buffer_len; ++i) {
-    // Prevent 0x0000 + 00/01/02 byte sequences from occurring by prefixing
-    // the last byte with 0x03. Do the same for 0x03.
-    if (zerocount == 2 && buffer[i] < 4) {
-      fwrite(&emulation_prevention_three_byte, 1, 1, output);
-      zerocount = 0;
-    }
-    if(buffer[i] == 0) {
-      zerocount++;
-    } else {
-      zerocount = 0;
-    }
-
-    // Write the actual data
-    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);
-  }
 }
 
 
diff --git a/src/nal.h b/src/nal.h
index 2f5fc140..dc5d4648 100644
--- a/src/nal.h
+++ b/src/nal.h
@@ -92,9 +92,8 @@ enum {
 
 //////////////////////////////////////////////////////////////////////////
 // FUNCTIONS
-void nal_write(FILE *output, uint8_t *buffer, uint32_t buffer_len,
-               uint8_t nal_ref, uint8_t nal_type, uint8_t temporal_id,
-               int long_start_code);
+void nal_write(FILE *output,uint8_t nal_ref, uint8_t nal_type,
+               uint8_t temporal_id,int long_start_code);
 void picture_checksum(const picture *pic,
                       unsigned char checksum_out[][SEI_HASH_MAX_LENGTH]);