Moving quantization and loopfilter data to uncompressed header.

Overall PSNR drop on derf set is -0.024.

Change-Id: I1c05d2ace83488205ca16e3b06cd5f0ebceec8d6

2 files changed, 70 insertions, 96 deletions

diff --git a/vp9/decoder/vp9_decodframe.c b/vp9/decoder/vp9_decodframe.c
index d4b789899..5c03061b5 100644
--- a/vp9/decoder/vp9_decodframe.c
+++ b/vp9/decoder/vp9_decodframe.c
@@ -40,10 +40,6 @@
 int dec_debug = 0;
 #endif
 
-static int read_le16(const uint8_t *p) {
-  return (p[1] << 8) | p[0];
-}
-
 static int read_le32(const uint8_t *p) {
   return (p[3] << 24) | (p[2] << 16) | (p[1] << 8) | p[0];
 }
@@ -186,21 +182,6 @@ static void mb_init_dequantizer(VP9_COMMON *pc, MACROBLOCKD *xd) {
     xd->plane[i].dequant = pc->uv_dequant[xd->q_index];
 }
 
-static INLINE void dequant_add_y(MACROBLOCKD *xd, TX_TYPE tx_type, int idx,
-                                 BLOCK_SIZE_TYPE bsize) {
-  struct macroblockd_plane *const y = &xd->plane[0];
-  uint8_t* const dst = raster_block_offset_uint8(xd, bsize, 0, idx,
-                                                 xd->plane[0].dst.buf,
-                                                 xd->plane[0].dst.stride);
-  if (tx_type != DCT_DCT) {
-    vp9_iht_add_c(tx_type, BLOCK_OFFSET(y->qcoeff, idx, 16),
-                  dst, xd->plane[0].dst.stride, y->eobs[idx]);
-  } else {
-    xd->itxm_add(BLOCK_OFFSET(y->qcoeff, idx, 16),
-                 dst, xd->plane[0].dst.stride, y->eobs[idx]);
-  }
-}
-
 static void decode_block(int plane, int block, BLOCK_SIZE_TYPE bsize,
                          int ss_txfrm_size, void *arg) {
   MACROBLOCKD* const xd = arg;
@@ -383,18 +364,6 @@ static void decode_sb(VP9D_COMP *pbi, MACROBLOCKD *xd, int mi_row, int mi_col,
   }
 }
 
-static int get_delta_q(vp9_reader *r, int *dq) {
-  const int old_value = *dq;
-
-  if (vp9_read_bit(r)) {  // Update bit
-    const int value = vp9_read_literal(r, 4);
-    *dq = vp9_read_and_apply_sign(r, value);
-  }
-
-  // Trigger a quantizer update if the delta-q value has changed
-  return old_value != *dq;
-}
-
 static void set_offsets(VP9D_COMP *pbi, BLOCK_SIZE_TYPE bsize,
                         int mi_row, int mi_col) {
   const int bh = 1 << mi_height_log2(bsize);
@@ -693,48 +662,58 @@ static void setup_pred_probs(VP9_COMMON *pc, vp9_reader *r) {
   }
 }
 
-static void setup_loopfilter(VP9_COMMON *pc, MACROBLOCKD *xd, vp9_reader *r) {
-  pc->filter_level = vp9_read_literal(r, 6);
-  pc->sharpness_level = vp9_read_literal(r, 3);
+static void setup_loopfilter(VP9D_COMP *pbi, struct vp9_read_bit_buffer *rb) {
+  VP9_COMMON *const cm = &pbi->common;
+  MACROBLOCKD *const xd = &pbi->mb;
+
+  cm->filter_level = vp9_rb_read_literal(rb, 6);
+  cm->sharpness_level = vp9_rb_read_literal(rb, 3);
 
   // Read in loop filter deltas applied at the MB level based on mode or ref
   // frame.
   xd->mode_ref_lf_delta_update = 0;
 
-  xd->mode_ref_lf_delta_enabled = vp9_read_bit(r);
+  xd->mode_ref_lf_delta_enabled = vp9_rb_read_bit(rb);
   if (xd->mode_ref_lf_delta_enabled) {
-    xd->mode_ref_lf_delta_update = vp9_read_bit(r);
+    xd->mode_ref_lf_delta_update = vp9_rb_read_bit(rb);
     if (xd->mode_ref_lf_delta_update) {
       int i;
 
       for (i = 0; i < MAX_REF_LF_DELTAS; i++) {
-        if (vp9_read_bit(r)) {
-          const int value = vp9_read_literal(r, 6);
-          xd->ref_lf_deltas[i] = vp9_read_and_apply_sign(r, value);
+        if (vp9_rb_read_bit(rb)) {
+          const int value = vp9_rb_read_literal(rb, 6);
+          xd->ref_lf_deltas[i] = vp9_rb_read_bit(rb) ? -value : value;
         }
       }
 
       for (i = 0; i < MAX_MODE_LF_DELTAS; i++) {
-        if (vp9_read_bit(r)) {
-          const int value = vp9_read_literal(r, 6);
-          xd->mode_lf_deltas[i] = vp9_read_and_apply_sign(r, value);
+        if (vp9_rb_read_bit(rb)) {
+          const int value = vp9_rb_read_literal(rb, 6);
+          xd->mode_lf_deltas[i] = vp9_rb_read_bit(rb) ? -value : value;
         }
       }
     }
   }
 }
 
-static void setup_quantization(VP9D_COMP *pbi, vp9_reader *r) {
-  // Read the default quantizers
-  VP9_COMMON *const pc = &pbi->common;
-
-  pc->base_qindex = vp9_read_literal(r, QINDEX_BITS);
-  if (get_delta_q(r, &pc->y_dc_delta_q) |
-      get_delta_q(r, &pc->uv_dc_delta_q) |
-      get_delta_q(r, &pc->uv_ac_delta_q))
-    vp9_init_dequantizer(pc);
+static int read_delta_q(struct vp9_read_bit_buffer *rb, int *delta_q) {
+  const int old = *delta_q;
+  if (vp9_rb_read_bit(rb)) {
+    const int value = vp9_rb_read_literal(rb, 4);
+    *delta_q = vp9_rb_read_bit(rb) ? -value : value;
+  }
+  return old != *delta_q;
+}
 
-  mb_init_dequantizer(pc, &pbi->mb);  // MB level dequantizer setup
+static void setup_quantization(VP9D_COMP *pbi, struct vp9_read_bit_buffer *rb) {
+  VP9_COMMON *const cm = &pbi->common;
+  int update = 0;
+  cm->base_qindex = vp9_rb_read_literal(rb, QINDEX_BITS);
+  update |= read_delta_q(rb, &cm->y_dc_delta_q);
+  update |= read_delta_q(rb, &cm->uv_dc_delta_q);
+  update |= read_delta_q(rb, &cm->uv_ac_delta_q);
+  if (update)
+    vp9_init_dequantizer(cm);
 }
 
 static INTERPOLATIONFILTERTYPE read_mcomp_filter_type(vp9_reader *r) {
@@ -966,6 +945,9 @@ size_t read_uncompressed_header(VP9D_COMP *pbi,
     cm->frame_parallel_decoding_mode = 1;
   }
 
+  setup_loopfilter(pbi, rb);
+  setup_quantization(pbi, rb);
+
   return vp9_rb_read_literal(rb, 16);
 }
 
@@ -1008,9 +990,7 @@ int vp9_decode_frame(VP9D_COMP *pbi, const uint8_t **p_data_end) {
     vpx_internal_error(&pc->error, VPX_CODEC_MEM_ERROR,
                        "Failed to allocate bool decoder 0");
 
-  setup_loopfilter(pc, xd, &header_bc);
-
-  setup_quantization(pbi, &header_bc);
+  mb_init_dequantizer(pc, &pbi->mb);  // MB level dequantizer setup
 
   xd->lossless = pc->base_qindex == 0 &&
                  pc->y_dc_delta_q == 0 &&
diff --git a/vp9/encoder/vp9_bitstream.c b/vp9/encoder/vp9_bitstream.c
index 95e9a7df0..043ac6ad2 100644
--- a/vp9/encoder/vp9_bitstream.c
+++ b/vp9/encoder/vp9_bitstream.c
@@ -1205,20 +1205,21 @@ static void segment_reference_frames(VP9_COMP *cpi) {
   }
 }
 
-static void encode_loopfilter(VP9_COMMON *pc, MACROBLOCKD *xd, vp9_writer *w) {
+static void encode_loopfilter(VP9_COMMON *pc, MACROBLOCKD *xd,
+                              struct vp9_write_bit_buffer *wb) {
   int i;
 
   // Encode the loop filter level and type
-  vp9_write_literal(w, pc->filter_level, 6);
-  vp9_write_literal(w, pc->sharpness_level, 3);
+  vp9_wb_write_literal(wb, pc->filter_level, 6);
+  vp9_wb_write_literal(wb, pc->sharpness_level, 3);
 
   // Write out loop filter deltas applied at the MB level based on mode or
   // ref frame (if they are enabled).
-  vp9_write_bit(w, xd->mode_ref_lf_delta_enabled);
+  vp9_wb_write_bit(wb, xd->mode_ref_lf_delta_enabled);
 
   if (xd->mode_ref_lf_delta_enabled) {
     // Do the deltas need to be updated
-    vp9_write_bit(w, xd->mode_ref_lf_delta_update);
+    vp9_wb_write_bit(wb, xd->mode_ref_lf_delta_update);
     if (xd->mode_ref_lf_delta_update) {
       // Send update
       for (i = 0; i < MAX_REF_LF_DELTAS; i++) {
@@ -1227,18 +1228,13 @@ static void encode_loopfilter(VP9_COMMON *pc, MACROBLOCKD *xd, vp9_writer *w) {
         // Frame level data
         if (delta != xd->last_ref_lf_deltas[i]) {
           xd->last_ref_lf_deltas[i] = delta;
-          vp9_write_bit(w, 1);
+          vp9_wb_write_bit(wb, 1);
 
-          if (delta > 0) {
-            vp9_write_literal(w, delta & 0x3F, 6);
-            vp9_write_bit(w, 0);  // sign
-          } else {
-            assert(delta < 0);
-            vp9_write_literal(w, (-delta) & 0x3F, 6);
-            vp9_write_bit(w, 1);  // sign
-          }
+          assert(delta != 0);
+          vp9_wb_write_literal(wb, abs(delta) & 0x3F, 6);
+          vp9_wb_write_bit(wb, delta < 0);
         } else {
-          vp9_write_bit(w, 0);
+          vp9_wb_write_bit(wb, 0);
         }
       }
 
@@ -1247,39 +1243,35 @@ static void encode_loopfilter(VP9_COMMON *pc, MACROBLOCKD *xd, vp9_writer *w) {
         const int delta = xd->mode_lf_deltas[i];
         if (delta != xd->last_mode_lf_deltas[i]) {
           xd->last_mode_lf_deltas[i] = delta;
-          vp9_write_bit(w, 1);
+          vp9_wb_write_bit(wb, 1);
 
-          if (delta > 0) {
-            vp9_write_literal(w, delta & 0x3F, 6);
-            vp9_write_bit(w, 0);  // sign
-          } else {
-            assert(delta < 0);
-            vp9_write_literal(w, (-delta) & 0x3F, 6);
-            vp9_write_bit(w, 1);  // sign
-          }
+          assert(delta != 0);
+          vp9_wb_write_literal(wb, abs(delta) & 0x3F, 6);
+          vp9_wb_write_bit(wb, delta < 0);
         } else {
-          vp9_write_bit(w, 0);
+          vp9_wb_write_bit(wb, 0);
         }
       }
     }
   }
 }
 
-static void put_delta_q(vp9_writer *bc, int delta_q) {
+static void write_delta_q(struct vp9_write_bit_buffer *wb, int delta_q) {
   if (delta_q != 0) {
-    vp9_write_bit(bc, 1);
-    vp9_write_literal(bc, abs(delta_q), 4);
-    vp9_write_bit(bc, delta_q < 0);
+    vp9_wb_write_bit(wb, 1);
+    vp9_wb_write_literal(wb, abs(delta_q), 4);
+    vp9_wb_write_bit(wb, delta_q < 0);
   } else {
-    vp9_write_bit(bc, 0);
+    vp9_wb_write_bit(wb, 0);
   }
 }
 
-static void encode_quantization(VP9_COMMON *pc, vp9_writer *w) {
-  vp9_write_literal(w, pc->base_qindex, QINDEX_BITS);
-  put_delta_q(w, pc->y_dc_delta_q);
-  put_delta_q(w, pc->uv_dc_delta_q);
-  put_delta_q(w, pc->uv_ac_delta_q);
+static void encode_quantization(VP9_COMMON *cm,
+                                struct vp9_write_bit_buffer *wb) {
+  vp9_wb_write_literal(wb, cm->base_qindex, QINDEX_BITS);
+  write_delta_q(wb, cm->y_dc_delta_q);
+  write_delta_q(wb, cm->uv_dc_delta_q);
+  write_delta_q(wb, cm->uv_ac_delta_q);
 }
 
 
@@ -1399,8 +1391,11 @@ static void encode_txfm(VP9_COMP *cpi, vp9_writer *w) {
   }
 }
 
-void write_uncompressed_header(VP9_COMMON *cm,
+void write_uncompressed_header(VP9_COMP *cpi,
                                struct vp9_write_bit_buffer *wb) {
+  VP9_COMMON *const cm = &cpi->common;
+  MACROBLOCKD *const xd = &cpi->mb.e_mbd;
+
   const int scaling_active = cm->width != cm->display_width ||
                              cm->height != cm->display_height;
 
@@ -1438,6 +1433,9 @@ void write_uncompressed_header(VP9_COMMON *cm,
     vp9_wb_write_bit(wb, cm->refresh_frame_context);
     vp9_wb_write_bit(wb, cm->frame_parallel_decoding_mode);
   }
+
+  encode_loopfilter(cm, xd, wb);
+  encode_quantization(cm, wb);
 }
 
 void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, unsigned long *size) {
@@ -1450,7 +1448,7 @@ void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, unsigned long *size) {
   struct vp9_write_bit_buffer wb = {dest, 0};
   struct vp9_write_bit_buffer first_partition_size_wb;
 
-  write_uncompressed_header(pc, &wb);
+  write_uncompressed_header(cpi, &wb);
   first_partition_size_wb = wb;
   vp9_wb_write_literal(&wb, 0, 16);  // don't know in advance first part. size
 
@@ -1461,10 +1459,6 @@ void vp9_pack_bitstream(VP9_COMP *cpi, uint8_t *dest, unsigned long *size) {
 
   vp9_start_encode(&header_bc, cx_data);
 
-  encode_loopfilter(pc, xd, &header_bc);
-
-  encode_quantization(pc, &header_bc);
-
   // When there is a key frame all reference buffers are updated using the new key frame
   if (pc->frame_type != KEY_FRAME) {
     int refresh_mask;