Use explicit block position in foreach_transformed_block

Add the row and column index to the argument list of unit functions
called by foreach_transformed_block wrapper. This avoids the
repeated internal parsing according to the block index.

Change-Id: Ie7508acdac0b498487564639bc5cc6378a8a0df7

6 files changed, 73 insertions, 79 deletions

diff --git a/vp10/common/blockd.c b/vp10/common/blockd.c
index 5394b5e0e..b6f910ff6 100644
--- a/vp10/common/blockd.c
+++ b/vp10/common/blockd.c
@@ -66,7 +66,7 @@ void vp10_foreach_transformed_block_in_plane(
   for (r = 0; r < max_blocks_high; r += (1 << tx_size)) {
     // Skip visiting the sub blocks that are wholly within the UMV.
     for (c = 0; c < max_blocks_wide; c += (1 << tx_size)) {
-      visit(plane, i, plane_bsize, tx_size, arg);
+      visit(plane, i, r, c, plane_bsize, tx_size, arg);
       i += step;
     }
     i += extra_step;
diff --git a/vp10/common/blockd.h b/vp10/common/blockd.h
index 84541545a..b89d79180 100644
--- a/vp10/common/blockd.h
+++ b/vp10/common/blockd.h
@@ -283,6 +283,7 @@ static INLINE void reset_skip_context(MACROBLOCKD *xd, BLOCK_SIZE bsize) {
 }
 
 typedef void (*foreach_transformed_block_visitor)(int plane, int block,
+                                                  int blk_row, int blk_col,
                                                   BLOCK_SIZE plane_bsize,
                                                   TX_SIZE tx_size,
                                                   void *arg);
@@ -296,17 +297,6 @@ void vp10_foreach_transformed_block(
     const MACROBLOCKD* const xd, BLOCK_SIZE bsize,
     foreach_transformed_block_visitor visit, void *arg);
 
-static INLINE void txfrm_block_to_raster_xy(BLOCK_SIZE plane_bsize,
-                                            TX_SIZE tx_size, int block,
-                                            int *x, int *y) {
-  const int bwl = b_width_log2_lookup[plane_bsize];
-  const int tx_cols_log2 = bwl - tx_size;
-  const int tx_cols = 1 << tx_cols_log2;
-  const int raster_mb = block >> (tx_size << 1);
-  *x = (raster_mb & (tx_cols - 1)) << tx_size;
-  *y = (raster_mb >> tx_cols_log2) << tx_size;
-}
-
 void vp10_set_contexts(const MACROBLOCKD *xd, struct macroblockd_plane *pd,
                       BLOCK_SIZE plane_bsize, TX_SIZE tx_size, int has_eob,
                       int aoff, int loff);
diff --git a/vp10/encoder/encodemb.c b/vp10/encoder/encodemb.c
index ff23fee27..ec6b34a61 100644
--- a/vp10/encoder/encodemb.c
+++ b/vp10/encoder/encodemb.c
@@ -324,7 +324,8 @@ static INLINE void highbd_fdct32x32(int rd_transform, const int16_t *src,
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 
 void vp10_xform_quant_fp(MACROBLOCK *x, int plane, int block,
-                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
+                         int blk_row, int blk_col,
+                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
   MACROBLOCKD *const xd = &x->e_mbd;
   const struct macroblock_plane *const p = &x->plane[plane];
   const struct macroblockd_plane *const pd = &xd->plane[plane];
@@ -336,10 +337,8 @@ void vp10_xform_quant_fp(MACROBLOCK *x, int plane, int block,
   tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
   uint16_t *const eob = &p->eobs[block];
   const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
-  int i, j;
   const int16_t *src_diff;
-  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
-  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
+  src_diff = &p->src_diff[4 * (blk_row * diff_stride + blk_col)];
 
 #if CONFIG_VP9_HIGHBITDEPTH
   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
@@ -424,7 +423,8 @@ void vp10_xform_quant_fp(MACROBLOCK *x, int plane, int block,
 }
 
 void vp10_xform_quant_dc(MACROBLOCK *x, int plane, int block,
-                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
+                         int blk_row, int blk_col,
+                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
   MACROBLOCKD *const xd = &x->e_mbd;
   const struct macroblock_plane *const p = &x->plane[plane];
   const struct macroblockd_plane *const pd = &xd->plane[plane];
@@ -433,11 +433,8 @@ void vp10_xform_quant_dc(MACROBLOCK *x, int plane, int block,
   tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
   uint16_t *const eob = &p->eobs[block];
   const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
-  int i, j;
   const int16_t *src_diff;
-
-  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
-  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
+  src_diff = &p->src_diff[4 * (blk_row * diff_stride + blk_col)];
 
 #if CONFIG_VP9_HIGHBITDEPTH
   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
@@ -658,7 +655,8 @@ static void highbd_fwd_txfm_32x32(int rd_transform, const int16_t *src_diff,
 #endif  // CONFIG_VP9_HIGHBITDEPTH
 
 void vp10_xform_quant(MACROBLOCK *x, int plane, int block,
-                     BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
+                      int blk_row, int blk_col,
+                      BLOCK_SIZE plane_bsize, TX_SIZE tx_size) {
   MACROBLOCKD *const xd = &x->e_mbd;
   const struct macroblock_plane *const p = &x->plane[plane];
   const struct macroblockd_plane *const pd = &xd->plane[plane];
@@ -670,10 +668,8 @@ void vp10_xform_quant(MACROBLOCK *x, int plane, int block,
   tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
   uint16_t *const eob = &p->eobs[block];
   const int diff_stride = 4 * num_4x4_blocks_wide_lookup[plane_bsize];
-  int i, j;
   const int16_t *src_diff;
-  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
-  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
+  src_diff = &p->src_diff[4 * (blk_row * diff_stride + blk_col)];
 
 #if CONFIG_VP9_HIGHBITDEPTH
   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
@@ -751,7 +747,8 @@ void vp10_xform_quant(MACROBLOCK *x, int plane, int block,
   }
 }
 
-static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
+static void encode_block(int plane, int block, int blk_row, int blk_col,
+                         BLOCK_SIZE plane_bsize,
                          TX_SIZE tx_size, void *arg) {
   struct encode_b_args *const args = arg;
   MACROBLOCK *const x = args->x;
@@ -760,14 +757,12 @@ static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
   struct macroblock_plane *const p = &x->plane[plane];
   struct macroblockd_plane *const pd = &xd->plane[plane];
   tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
-  int i, j;
   uint8_t *dst;
   ENTROPY_CONTEXT *a, *l;
   TX_TYPE tx_type = get_tx_type(pd->plane_type, xd, block);
-  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
-  dst = &pd->dst.buf[4 * j * pd->dst.stride + 4 * i];
-  a = &ctx->ta[plane][i];
-  l = &ctx->tl[plane][j];
+  dst = &pd->dst.buf[4 * blk_row * pd->dst.stride + 4 * blk_col];
+  a = &ctx->ta[plane][blk_col];
+  l = &ctx->tl[plane][blk_row];
 
   // TODO(jingning): per transformed block zero forcing only enabled for
   // luma component. will integrate chroma components as well.
@@ -786,17 +781,20 @@ static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
         *a = *l = 0;
         return;
       } else {
-        vp10_xform_quant_fp(x, plane, block, plane_bsize, tx_size);
+        vp10_xform_quant_fp(x, plane, block, blk_row, blk_col,
+                            plane_bsize, tx_size);
       }
     } else {
       if (max_txsize_lookup[plane_bsize] == tx_size) {
         int txfm_blk_index = (plane << 2) + (block >> (tx_size << 1));
         if (x->skip_txfm[txfm_blk_index] == SKIP_TXFM_NONE) {
           // full forward transform and quantization
-          vp10_xform_quant(x, plane, block, plane_bsize, tx_size);
+          vp10_xform_quant(x, plane, block, blk_row, blk_col,
+                           plane_bsize, tx_size);
         } else if (x->skip_txfm[txfm_blk_index] == SKIP_TXFM_AC_ONLY) {
           // fast path forward transform and quantization
-          vp10_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
+          vp10_xform_quant_dc(x, plane, block, blk_row, blk_col,
+                              plane_bsize, tx_size);
         } else {
           // skip forward transform
           p->eobs[block] = 0;
@@ -804,7 +802,8 @@ static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
           return;
         }
       } else {
-        vp10_xform_quant(x, plane, block, plane_bsize, tx_size);
+        vp10_xform_quant(x, plane, block, blk_row, blk_col,
+                         plane_bsize, tx_size);
       }
     }
   }
@@ -879,19 +878,18 @@ static void encode_block(int plane, int block, BLOCK_SIZE plane_bsize,
   }
 }
 
-static void encode_block_pass1(int plane, int block, BLOCK_SIZE plane_bsize,
+static void encode_block_pass1(int plane, int block, int blk_row, int blk_col,
+                               BLOCK_SIZE plane_bsize,
                                TX_SIZE tx_size, void *arg) {
   MACROBLOCK *const x = (MACROBLOCK *)arg;
   MACROBLOCKD *const xd = &x->e_mbd;
   struct macroblock_plane *const p = &x->plane[plane];
   struct macroblockd_plane *const pd = &xd->plane[plane];
   tran_low_t *const dqcoeff = BLOCK_OFFSET(pd->dqcoeff, block);
-  int i, j;
   uint8_t *dst;
-  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
-  dst = &pd->dst.buf[4 * j * pd->dst.stride + 4 * i];
+  dst = &pd->dst.buf[4 * blk_row * pd->dst.stride + 4 * blk_col];
 
-  vp10_xform_quant(x, plane, block, plane_bsize, tx_size);
+  vp10_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
 
   if (p->eobs[block] > 0) {
 #if CONFIG_VP9_HIGHBITDEPTH
@@ -948,8 +946,9 @@ void vp10_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize) {
   }
 }
 
-void vp10_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
-                            TX_SIZE tx_size, void *arg) {
+void vp10_encode_block_intra(int plane, int block, int blk_row, int blk_col,
+                             BLOCK_SIZE plane_bsize,
+                             TX_SIZE tx_size, void *arg) {
   struct encode_b_args* const args = arg;
   MACROBLOCK *const x = args->x;
   MACROBLOCKD *const xd = &x->e_mbd;
@@ -971,15 +970,13 @@ void vp10_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
   uint16_t *eob = &p->eobs[block];
   const int src_stride = p->src.stride;
   const int dst_stride = pd->dst.stride;
-  int i, j;
-  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &i, &j);
-  dst = &pd->dst.buf[4 * (j * dst_stride + i)];
-  src = &p->src.buf[4 * (j * src_stride + i)];
-  src_diff = &p->src_diff[4 * (j * diff_stride + i)];
+  dst = &pd->dst.buf[4 * (blk_row * dst_stride + blk_col)];
+  src = &p->src.buf[4 * (blk_row * src_stride + blk_col)];
+  src_diff = &p->src_diff[4 * (blk_row * diff_stride + blk_col)];
 
   mode = plane == 0 ? get_y_mode(xd->mi[0], block) : mbmi->uv_mode;
   vp10_predict_intra_block(xd, bwl, bhl, tx_size, mode, dst, dst_stride,
-                          dst, dst_stride, i, j, plane);
+                          dst, dst_stride, blk_col, blk_row, plane);
 
 #if CONFIG_VP9_HIGHBITDEPTH
   if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
@@ -1130,5 +1127,5 @@ void vp10_encode_intra_block_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
   struct encode_b_args arg = {x, NULL, &xd->mi[0]->mbmi.skip};
 
   vp10_foreach_transformed_block_in_plane(xd, bsize, plane,
-                                         vp10_encode_block_intra, &arg);
+                                          vp10_encode_block_intra, &arg);
 }
diff --git a/vp10/encoder/encodemb.h b/vp10/encoder/encodemb.h
index 62a7db4a2..2e6516e0b 100644
--- a/vp10/encoder/encodemb.h
+++ b/vp10/encoder/encodemb.h
@@ -26,16 +26,20 @@ struct encode_b_args {
 void vp10_encode_sb(MACROBLOCK *x, BLOCK_SIZE bsize);
 void vp10_encode_sby_pass1(MACROBLOCK *x, BLOCK_SIZE bsize);
 void vp10_xform_quant_fp(MACROBLOCK *x, int plane, int block,
-                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
+                         int blk_row, int blk_col,
+                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
 void vp10_xform_quant_dc(MACROBLOCK *x, int plane, int block,
-                        BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
+                         int blk_row, int blk_col,
+                         BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
 void vp10_xform_quant(MACROBLOCK *x, int plane, int block,
-                     BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
+                      int blk_row, int blk_col,
+                      BLOCK_SIZE plane_bsize, TX_SIZE tx_size);
 
 void vp10_subtract_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
 
-void vp10_encode_block_intra(int plane, int block, BLOCK_SIZE plane_bsize,
-                            TX_SIZE tx_size, void *arg);
+void vp10_encode_block_intra(int plane, int block, int blk_row, int blk_col,
+                             BLOCK_SIZE plane_bsize,
+                             TX_SIZE tx_size, void *arg);
 
 void vp10_encode_intra_block_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane);
 
diff --git a/vp10/encoder/rdopt.c b/vp10/encoder/rdopt.c
index 2e000af65..cecc59c8f 100644
--- a/vp10/encoder/rdopt.c
+++ b/vp10/encoder/rdopt.c
@@ -451,18 +451,16 @@ static void dist_block(MACROBLOCK *x, int plane, int block, TX_SIZE tx_size,
   *out_sse = this_sse >> shift;
 }
 
-static int rate_block(int plane, int block, BLOCK_SIZE plane_bsize,
+static int rate_block(int plane, int block, int blk_row, int blk_col,
                       TX_SIZE tx_size, struct rdcost_block_args* args) {
-  int x_idx, y_idx;
-  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &x_idx, &y_idx);
-
-  return cost_coeffs(args->x, plane, block, args->t_above + x_idx,
-                     args->t_left + y_idx, tx_size,
+  return cost_coeffs(args->x, plane, block, args->t_above + blk_col,
+                     args->t_left + blk_row, tx_size,
                      args->so->scan, args->so->neighbors,
                      args->use_fast_coef_costing);
 }
 
-static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
+static void block_rd_txfm(int plane, int block, int blk_row, int blk_col,
+                          BLOCK_SIZE plane_bsize,
                           TX_SIZE tx_size, void *arg) {
   struct rdcost_block_args *args = arg;
   MACROBLOCK *const x = args->x;
@@ -478,20 +476,23 @@ static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
 
   if (!is_inter_block(mbmi)) {
     struct encode_b_args arg = {x, NULL, &mbmi->skip};
-    vp10_encode_block_intra(plane, block, plane_bsize, tx_size, &arg);
+    vp10_encode_block_intra(plane, block, blk_row, blk_col,
+                            plane_bsize, tx_size, &arg);
     dist_block(x, plane, block, tx_size, &dist, &sse);
   } else if (max_txsize_lookup[plane_bsize] == tx_size) {
     if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
         SKIP_TXFM_NONE) {
       // full forward transform and quantization
-      vp10_xform_quant(x, plane, block, plane_bsize, tx_size);
+      vp10_xform_quant(x, plane, block, blk_row, blk_col,
+                       plane_bsize, tx_size);
       dist_block(x, plane, block, tx_size, &dist, &sse);
     } else if (x->skip_txfm[(plane << 2) + (block >> (tx_size << 1))] ==
                SKIP_TXFM_AC_ONLY) {
       // compute DC coefficient
       tran_low_t *const coeff   = BLOCK_OFFSET(x->plane[plane].coeff, block);
       tran_low_t *const dqcoeff = BLOCK_OFFSET(xd->plane[plane].dqcoeff, block);
-      vp10_xform_quant_dc(x, plane, block, plane_bsize, tx_size);
+      vp10_xform_quant_dc(x, plane, block, blk_row, blk_col,
+                          plane_bsize, tx_size);
       sse  = x->bsse[(plane << 2) + (block >> (tx_size << 1))] << 4;
       dist = sse;
       if (x->plane[plane].eobs[block]) {
@@ -515,7 +516,7 @@ static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
     }
   } else {
     // full forward transform and quantization
-    vp10_xform_quant(x, plane, block, plane_bsize, tx_size);
+    vp10_xform_quant(x, plane, block, blk_row, blk_col, plane_bsize, tx_size);
     dist_block(x, plane, block, tx_size, &dist, &sse);
   }
 
@@ -525,7 +526,7 @@ static void block_rd_txfm(int plane, int block, BLOCK_SIZE plane_bsize,
     return;
   }
 
-  rate = rate_block(plane, block, plane_bsize, tx_size, args);
+  rate = rate_block(plane, block, blk_row, blk_col, tx_size, args);
   rd1 = RDCOST(x->rdmult, x->rddiv, rate, dist);
   rd2 = RDCOST(x->rdmult, x->rddiv, 0, sse);
 
diff --git a/vp10/encoder/tokenize.c b/vp10/encoder/tokenize.c
index 2c9998ba0..e568c0ba5 100644
--- a/vp10/encoder/tokenize.c
+++ b/vp10/encoder/tokenize.c
@@ -443,7 +443,9 @@ struct tokenize_b_args {
   TOKENEXTRA **tp;
 };
 
-static void set_entropy_context_b(int plane, int block, BLOCK_SIZE plane_bsize,
+static void set_entropy_context_b(int plane, int block,
+                                  int blk_row, int blk_col,
+                                  BLOCK_SIZE plane_bsize,
                                   TX_SIZE tx_size, void *arg) {
   struct tokenize_b_args* const args = arg;
   ThreadData *const td = args->td;
@@ -451,10 +453,8 @@ static void set_entropy_context_b(int plane, int block, BLOCK_SIZE plane_bsize,
   MACROBLOCKD *const xd = &x->e_mbd;
   struct macroblock_plane *p = &x->plane[plane];
   struct macroblockd_plane *pd = &xd->plane[plane];
-  int aoff, loff;
-  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &aoff, &loff);
   vp10_set_contexts(xd, pd, plane_bsize, tx_size, p->eobs[block] > 0,
-                   aoff, loff);
+                    blk_col, blk_row);
 }
 
 static INLINE void add_token(TOKENEXTRA **t, const vpx_prob *context_tree,
@@ -520,7 +520,8 @@ void vp10_tokenize_palette_sb(struct ThreadData *const td,
   }
 }
 
-static void tokenize_b(int plane, int block, BLOCK_SIZE plane_bsize,
+static void tokenize_b(int plane, int block, int blk_row, int blk_col,
+                       BLOCK_SIZE plane_bsize,
                        TX_SIZE tx_size, void *arg) {
   struct tokenize_b_args* const args = arg;
   VP10_COMP *cpi = args->cpi;
@@ -553,11 +554,8 @@ static void tokenize_b(int plane, int block, BLOCK_SIZE plane_bsize,
   const int seg_eob = get_tx_eob(&cpi->common.seg, segment_id, tx_size);
   int16_t token;
   EXTRABIT extra;
-  int aoff, loff;
-  txfrm_block_to_raster_xy(plane_bsize, tx_size, block, &aoff, &loff);
-
-  pt = get_entropy_context(tx_size, pd->above_context + aoff,
-                           pd->left_context + loff);
+  pt = get_entropy_context(tx_size, pd->above_context + blk_col,
+                           pd->left_context + blk_row);
   scan = so->scan;
   nb = so->neighbors;
   c = 0;
@@ -597,20 +595,22 @@ static void tokenize_b(int plane, int block, BLOCK_SIZE plane_bsize,
 
   *tp = t;
 
-  vp10_set_contexts(xd, pd, plane_bsize, tx_size, c > 0, aoff, loff);
+  vp10_set_contexts(xd, pd, plane_bsize, tx_size, c > 0, blk_col, blk_row);
 }
 
 struct is_skippable_args {
   uint16_t *eobs;
   int *skippable;
 };
-static void is_skippable(int plane, int block,
+static void is_skippable(int plane, int block, int blk_row, int blk_col,
                          BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                          void *argv) {
   struct is_skippable_args *args = argv;
   (void)plane;
   (void)plane_bsize;
   (void)tx_size;
+  (void)blk_row;
+  (void)blk_col;
   args->skippable[0] &= (!args->eobs[block]);
 }
 
@@ -624,13 +624,15 @@ int vp10_is_skippable_in_plane(MACROBLOCK *x, BLOCK_SIZE bsize, int plane) {
   return result;
 }
 
-static void has_high_freq_coeff(int plane, int block,
+static void has_high_freq_coeff(int plane, int block, int blk_row, int blk_col,
                                 BLOCK_SIZE plane_bsize, TX_SIZE tx_size,
                                 void *argv) {
   struct is_skippable_args *args = argv;
   int eobs = (tx_size == TX_4X4) ? 3 : 10;
   (void) plane;
   (void) plane_bsize;
+  (void) blk_row;
+  (void) blk_col;
 
   *(args->skippable) |= (args->eobs[block] > eobs);
 }