libvpx: Pull from upstream

source/libvpx/vp9/common/vp9_reconinter.c

Index: source/libvpx/vp9/common/vp9_reconinter.c
===================================================================
--- source/libvpx/vp9/common/vp9_reconinter.c	(revision 240950)
+++ source/libvpx/vp9/common/vp9_reconinter.c	(working copy)
@@ -20,37 +20,56 @@
 #include "vp9/common/vp9_reconinter.h"
 #include "vp9/common/vp9_reconintra.h"
 
-void vp9_setup_interp_filters(MACROBLOCKD *xd,
-                              INTERPOLATION_TYPE mcomp_filter_type,
-                              VP9_COMMON *cm) {
-  if (xd->mi_8x8 && xd->mi_8x8[0]) {
-    MB_MODE_INFO *const mbmi = &xd->mi_8x8[0]->mbmi;
+static void build_mc_border(const uint8_t *src, uint8_t *dst, int stride,
+                             int x, int y, int b_w, int b_h, int w, int h) {
+  // Get a pointer to the start of the real data for this row.
+  const uint8_t *ref_row = src - x - y * stride;
 
-    set_scale_factors(xd, mbmi->ref_frame[0] - LAST_FRAME,
-                          mbmi->ref_frame[1] - LAST_FRAME,
-                          cm->active_ref_scale);
-  } else {
-    set_scale_factors(xd, -1, -1, cm->active_ref_scale);
-  }
+  if (y >= h)
+    ref_row += (h - 1) * stride;
+  else if (y > 0)
+    ref_row += y * stride;
 
-  xd->subpix.filter_x = xd->subpix.filter_y =
-      vp9_get_filter_kernel(mcomp_filter_type == SWITCHABLE ?
-                               EIGHTTAP : mcomp_filter_type);
+  do {
+    int right = 0, copy;
+    int left = x < 0 ? -x : 0;
 
-  assert(((intptr_t)xd->subpix.filter_x & 0xff) == 0);
+    if (left > b_w)
+      left = b_w;
+
+    if (x + b_w > w)
+      right = x + b_w - w;
+
+    if (right > b_w)
+      right = b_w;
+
+    copy = b_w - left - right;
+
+    if (left)
+      memset(dst, ref_row[0], left);
+
+    if (copy)
+      memmove(dst + left, ref_row + x + left, copy);
+
+    if (right)
+      memset(dst + left + copy, ref_row[w - 1], right);
+
+    dst += stride;
+    ++y;
+
+    if (y > 0 && y < h)
+      ref_row += stride;
+  } while (--b_h);
 }
 
 static void inter_predictor(const uint8_t *src, int src_stride,
                             uint8_t *dst, int dst_stride,
-                            const MV32 *mv,
+                            const int subpel_x,
+                            const int subpel_y,
                             const struct scale_factors *scale,
                             int w, int h, int ref,
                             const struct subpix_fn_table *subpix,
                             int xs, int ys) {
-  const int subpel_x = mv->col & SUBPEL_MASK;
-  const int subpel_y = mv->row & SUBPEL_MASK;
-
-  src += (mv->row >> SUBPEL_BITS) * src_stride + (mv->col >> SUBPEL_BITS);
   scale->sfc->predict[subpel_x != 0][subpel_y != 0][ref](
       src, src_stride, dst, dst_stride,
       subpix->filter_x[subpel_x], xs,
@@ -70,9 +89,12 @@
                      is_q4 ? src_mv->col : src_mv->col * 2 };
   const struct scale_factors_common *sfc = scale->sfc;
   const MV32 mv = sfc->scale_mv(&mv_q4, scale);
+  const int subpel_x = mv.col & SUBPEL_MASK;
+  const int subpel_y = mv.row & SUBPEL_MASK;
+  src += (mv.row >> SUBPEL_BITS) * src_stride + (mv.col >> SUBPEL_BITS);
 
-  inter_predictor(src, src_stride, dst, dst_stride, &mv, scale,
-                  w, h, ref, subpix, sfc->x_step_q4, sfc->y_step_q4);
+  inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y,
+                  scale, w, h, ref, subpix, sfc->x_step_q4, sfc->y_step_q4);
 }
 
 static INLINE int round_mv_comp_q4(int value) {
@@ -117,31 +139,18 @@
   return clamped_mv;
 }
 
-struct build_inter_predictors_args {
-  MACROBLOCKD *xd;
-  int x, y;
-};
-
-static void build_inter_predictors(int plane, int block, BLOCK_SIZE bsize,
-                                   int pred_w, int pred_h,
-                                   void *argv) {
-  const struct build_inter_predictors_args* const arg = argv;
-  MACROBLOCKD *const xd = arg->xd;
+// TODO(jkoleszar): In principle, pred_w, pred_h are unnecessary, as we could
+// calculate the subsampled BLOCK_SIZE, but that type isn't defined for
+// sizes smaller than 16x16 yet.
+static void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
+                                   int bw, int bh,
+                                   int x, int y, int w, int h,
+                                   int mi_x, int mi_y) {
   struct macroblockd_plane *const pd = &xd->plane[plane];
-  const int bwl = b_width_log2(bsize) - pd->subsampling_x;
-  const int bw = 4 << bwl;
-  const int bh = plane_block_height(bsize, pd);
-  const int x = 4 * (block & ((1 << bwl) - 1));
-  const int y = 4 * (block >> bwl);
   const MODE_INFO *mi = xd->mi_8x8[0];
   const int is_compound = has_second_ref(&mi->mbmi);
   int ref;
 
-  assert(x < bw);
-  assert(y < bh);
-  assert(mi->mbmi.sb_type < BLOCK_8X8 || 4 << pred_w == bw);
-  assert(mi->mbmi.sb_type < BLOCK_8X8 || 4 << pred_h == bh);
-
   for (ref = 0; ref < 1 + is_compound; ++ref) {
     struct scale_factors *const scale = &xd->scale_factor[ref];
     struct buf_2d *const pre_buf = &pd->pre[ref];
@@ -168,11 +177,11 @@
 
     uint8_t *pre;
     MV32 scaled_mv;
-    int xs, ys;
+    int xs, ys, subpel_x, subpel_y;
 
     if (vp9_is_scaled(scale->sfc)) {
       pre = pre_buf->buf + scaled_buffer_offset(x, y, pre_buf->stride, scale);
-      scale->sfc->set_scaled_offsets(scale, arg->y + y, arg->x + x);
+      scale->sfc->set_scaled_offsets(scale, mi_y + y, mi_x + x);
       scaled_mv = scale->sfc->scale_mv(&mv_q4, scale);
       xs = scale->sfc->x_step_q4;
       ys = scale->sfc->y_step_q4;
@@ -182,48 +191,41 @@
       scaled_mv.col = mv_q4.col;
       xs = ys = 16;
     }
+    subpel_x = scaled_mv.col & SUBPEL_MASK;
+    subpel_y = scaled_mv.row & SUBPEL_MASK;
+    pre += (scaled_mv.row >> SUBPEL_BITS) * pre_buf->stride
+           + (scaled_mv.col >> SUBPEL_BITS);
 
     inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
-                    &scaled_mv, scale,
-                    4 << pred_w, 4 << pred_h, ref,
-                    &xd->subpix, xs, ys);
+                    subpel_x, subpel_y, scale, w, h, ref, &xd->subpix, xs, ys);
   }
 }
 
-// TODO(jkoleszar): In principle, pred_w, pred_h are unnecessary, as we could
-// calculate the subsampled BLOCK_SIZE, but that type isn't defined for
-// sizes smaller than 16x16 yet.
-typedef void (*foreach_predicted_block_visitor)(int plane, int block,
-                                                BLOCK_SIZE bsize,
-                                                int pred_w, int pred_h,
-                                                void *arg);
-static INLINE void foreach_predicted_block_in_plane(
-    const MACROBLOCKD* const xd, BLOCK_SIZE bsize, int plane,
-    foreach_predicted_block_visitor visit, void *arg) {
-  const int bwl = b_width_log2(bsize) - xd->plane[plane].subsampling_x;
-  const int bhl = b_height_log2(bsize) - xd->plane[plane].subsampling_y;
-
-  if (xd->mi_8x8[0]->mbmi.sb_type < BLOCK_8X8) {
-    int i = 0, x, y;
-    assert(bsize == BLOCK_8X8);
-    for (y = 0; y < 1 << bhl; ++y)
-      for (x = 0; x < 1 << bwl; ++x)
-        visit(plane, i++, bsize, 0, 0, arg);
-  } else {
-    visit(plane, 0, bsize, bwl, bhl, arg);
-  }
-}
-
 static void build_inter_predictors_for_planes(MACROBLOCKD *xd, BLOCK_SIZE bsize,
                                               int mi_row, int mi_col,
                                               int plane_from, int plane_to) {
   int plane;
+  const int mi_x = mi_col * MI_SIZE;
+  const int mi_y = mi_row * MI_SIZE;
   for (plane = plane_from; plane <= plane_to; ++plane) {
-    struct build_inter_predictors_args args = {
-      xd, mi_col * MI_SIZE, mi_row * MI_SIZE,
-    };
-    foreach_predicted_block_in_plane(xd, bsize, plane, build_inter_predictors,
-                                     &args);
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize,
+                                                        &xd->plane[plane]);
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+    const int bw = 4 * num_4x4_w;
+    const int bh = 4 * num_4x4_h;
+
+    if (xd->mi_8x8[0]->mbmi.sb_type < BLOCK_8X8) {
+      int i = 0, x, y;
+      assert(bsize == BLOCK_8X8);
+      for (y = 0; y < num_4x4_h; ++y)
+        for (x = 0; x < num_4x4_w; ++x)
+           build_inter_predictors(xd, plane, i++, bw, bh,
+                                  4 * x, 4 * y, 4, 4, mi_x, mi_y);
+    } else {
+      build_inter_predictors(xd, plane, 0, bw, bh,
+                             0, 0, bw, bh, mi_x, mi_y);
+    }
   }
 }
 
@@ -242,12 +244,160 @@
                                     MAX_MB_PLANE - 1);
 }
 
+// TODO(jingning): This function serves as a placeholder for decoder prediction
+// using on demand border extension. It should be moved to /decoder/ directory.
+static void dec_build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
+                                       int bw, int bh,
+                                       int x, int y, int w, int h,
+                                       int mi_x, int mi_y) {
+  struct macroblockd_plane *const pd = &xd->plane[plane];
+  const MODE_INFO *mi = xd->mi_8x8[0];
+  const int is_compound = has_second_ref(&mi->mbmi);
+  int ref;
+
+  for (ref = 0; ref < 1 + is_compound; ++ref) {
+    struct scale_factors *const scale = &xd->scale_factor[ref];
+    struct buf_2d *const pre_buf = &pd->pre[ref];
+    struct buf_2d *const dst_buf = &pd->dst;
+    uint8_t *const dst = dst_buf->buf + dst_buf->stride * y + x;
+
+    // TODO(jkoleszar): All chroma MVs in SPLITMV mode are taken as the
+    // same MV (the average of the 4 luma MVs) but we could do something
+    // smarter for non-4:2:0. Just punt for now, pending the changes to get
+    // rid of SPLITMV mode entirely.
+    const MV mv = mi->mbmi.sb_type < BLOCK_8X8
+               ? (plane == 0 ? mi->bmi[block].as_mv[ref].as_mv
+                             : mi_mv_pred_q4(mi, ref))
+               : mi->mbmi.mv[ref].as_mv;
+
+    // TODO(jkoleszar): This clamping is done in the incorrect place for the
+    // scaling case. It needs to be done on the scaled MV, not the pre-scaling
+    // MV. Note however that it performs the subsampling aware scaling so
+    // that the result is always q4.
+    // mv_precision precision is MV_PRECISION_Q4.
+    const MV mv_q4 = clamp_mv_to_umv_border_sb(xd, &mv, bw, bh,
+                                               pd->subsampling_x,
+                                               pd->subsampling_y);
+
+    MV32 scaled_mv;
+    int xs, ys, x0, y0, x0_16, y0_16, x1, y1, frame_width,
+        frame_height, subpel_x, subpel_y;
+    uint8_t *ref_frame, *buf_ptr;
+    const YV12_BUFFER_CONFIG *ref_buf = xd->ref_buf[ref];
+
+    // Get reference frame pointer, width and height.
+    if (plane == 0) {
+      frame_width = ref_buf->y_crop_width;
+      frame_height = ref_buf->y_crop_height;
+      ref_frame = ref_buf->y_buffer;
+    } else {
+      frame_width = ref_buf->uv_crop_width;
+      frame_height = ref_buf->uv_crop_height;
+      ref_frame = plane == 1 ? ref_buf->u_buffer : ref_buf->v_buffer;
+    }
+
+    // Get block position in current frame.
+    x0 = (-xd->mb_to_left_edge >> (3 + pd->subsampling_x)) + x;
+    y0 = (-xd->mb_to_top_edge >> (3 + pd->subsampling_y)) + y;
+
+    // Precision of x0_16 and y0_16 is 1/16th pixel.
+    x0_16 = x0 << SUBPEL_BITS;
+    y0_16 = y0 << SUBPEL_BITS;
+
+    if (vp9_is_scaled(scale->sfc)) {
+      scale->sfc->set_scaled_offsets(scale, mi_y + y, mi_x + x);
+      scaled_mv = scale->sfc->scale_mv(&mv_q4, scale);
+      xs = scale->sfc->x_step_q4;
+      ys = scale->sfc->y_step_q4;
+      // Get block position in the scaled reference frame.
+      x0 = scale->sfc->scale_value_x(x0, scale->sfc);
+      y0 = scale->sfc->scale_value_y(y0, scale->sfc);
+      x0_16 = scale->sfc->scale_value_x(x0_16, scale->sfc);
+      y0_16 = scale->sfc->scale_value_y(y0_16, scale->sfc);
+    } else {
+      scaled_mv.row = mv_q4.row;
+      scaled_mv.col = mv_q4.col;
+      xs = ys = 16;
+    }
+    subpel_x = scaled_mv.col & SUBPEL_MASK;
+    subpel_y = scaled_mv.row & SUBPEL_MASK;
+
+    // Get reference block top left coordinate.
+    x0 += scaled_mv.col >> SUBPEL_BITS;
+    y0 += scaled_mv.row >> SUBPEL_BITS;
+    x0_16 += scaled_mv.col;
+    y0_16 += scaled_mv.row;
+
+    // Get reference block bottom right coordinate.
+    x1 = ((x0_16 + (w - 1) * xs) >> SUBPEL_BITS) + 1;
+    y1 = ((y0_16 + (h - 1) * xs) >> SUBPEL_BITS) + 1;
+
+    // Get reference block pointer.
+    buf_ptr = ref_frame + y0 * pre_buf->stride + x0;
+
+    // Do border extension if there is motion or
+    // width/height is not a multiple of 8 pixels.
+    if (scaled_mv.col || scaled_mv.row ||
+        (frame_width & 0x7) || (frame_height & 0x7)) {
+
+      if (subpel_x) {
+        x0 -= VP9_INTERP_EXTEND - 1;
+        x1 += VP9_INTERP_EXTEND;
+      }
+
+      if (subpel_y) {
+        y0 -= VP9_INTERP_EXTEND - 1;
+        y1 += VP9_INTERP_EXTEND;
+      }
+
+      // Skip border extension if block is inside the frame.
+      if (x0 < 0 || x0 > frame_width - 1 || x1 < 0 || x1 > frame_width ||
+          y0 < 0 || y0 > frame_height - 1 || y1 < 0 || y1 > frame_height - 1) {
+        uint8_t *buf_ptr1 = ref_frame + y0 * pre_buf->stride + x0;
+        // Extend the border.
+        build_mc_border(buf_ptr1, buf_ptr1, pre_buf->stride, x0, y0, x1 - x0,
+                        y1 - y0, frame_width, frame_height);
+      }
+    }
+
+    inter_predictor(buf_ptr, pre_buf->stride, dst, dst_buf->stride, subpel_x,
+                    subpel_y, scale, w, h, ref, &xd->subpix, xs, ys);
+  }
+}
+
+void vp9_dec_build_inter_predictors_sb(MACROBLOCKD *xd, int mi_row, int mi_col,
+                                       BLOCK_SIZE bsize) {
+  int plane;
+  const int mi_x = mi_col * MI_SIZE;
+  const int mi_y = mi_row * MI_SIZE;
+  for (plane = 0; plane < MAX_MB_PLANE; ++plane) {
+    const BLOCK_SIZE plane_bsize = get_plane_block_size(bsize,
+                                                        &xd->plane[plane]);
+    const int num_4x4_w = num_4x4_blocks_wide_lookup[plane_bsize];
+    const int num_4x4_h = num_4x4_blocks_high_lookup[plane_bsize];
+    const int bw = 4 * num_4x4_w;
+    const int bh = 4 * num_4x4_h;
+
+    if (xd->mi_8x8[0]->mbmi.sb_type < BLOCK_8X8) {
+      int i = 0, x, y;
+      assert(bsize == BLOCK_8X8);
+      for (y = 0; y < num_4x4_h; ++y)
+        for (x = 0; x < num_4x4_w; ++x)
+          dec_build_inter_predictors(xd, plane, i++, bw, bh,
+                                     4 * x, 4 * y, 4, 4, mi_x, mi_y);
+    } else {
+      dec_build_inter_predictors(xd, plane, 0, bw, bh,
+                                 0, 0, bw, bh, mi_x, mi_y);
+    }
+  }
+}
+
 // TODO(dkovalev: find better place for this function)
 void vp9_setup_scale_factors(VP9_COMMON *cm, int i) {
   const int ref = cm->active_ref_idx[i];
   struct scale_factors *const sf = &cm->active_ref_scale[i];
   struct scale_factors_common *const sfc = &cm->active_ref_scale_comm[i];
-  if (ref >= NUM_YV12_BUFFERS) {
+  if (ref >= cm->fb_count) {
     vp9_zero(*sf);
     vp9_zero(*sfc);
   } else {
@@ -255,9 +405,6 @@
     vp9_setup_scale_factors_for_frame(sf, sfc,
                                       fb->y_crop_width, fb->y_crop_height,
                                       cm->width, cm->height);
-
-    if (vp9_is_scaled(sfc))
-      vp9_extend_frame_borders(fb, cm->subsampling_x, cm->subsampling_y);
   }
 }