Adds high bitdepth convolve, interpred & scaling

Change-Id: Ie51c352a6b250547207cbc1ebba833a01ed053e3

vp9/common/vp9_convolve.c

@@ -282,3 +282,280 @@ void vp9_convolve_avg_c(const uint8_t *src, ptrdiff_t src_stride,
     dst += dst_stride;
   }
 }
+
+#if CONFIG_VP9_HIGHBITDEPTH
+static void high_convolve_horiz(const uint8_t *src8, ptrdiff_t src_stride,
+                                uint8_t *dst8, ptrdiff_t dst_stride,
+                                const InterpKernel *x_filters,
+                                int x0_q4, int x_step_q4,
+                                int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel_high(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void high_convolve_avg_horiz(const uint8_t *src8, ptrdiff_t src_stride,
+                                    uint8_t *dst8, ptrdiff_t dst_stride,
+                                    const InterpKernel *x_filters,
+                                    int x0_q4, int x_step_q4,
+                                    int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] +
+          clip_pixel_high(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd), 1);
+      x_q4 += x_step_q4;
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+static void high_convolve_vert(const uint8_t *src8, ptrdiff_t src_stride,
+                               uint8_t *dst8, ptrdiff_t dst_stride,
+                               const InterpKernel *y_filters,
+                               int y0_q4, int y_step_q4, int w, int h,
+                               int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = clip_pixel_high(
+          ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void high_convolve_avg_vert(const uint8_t *src8, ptrdiff_t src_stride,
+                                   uint8_t *dst8, ptrdiff_t dst_stride,
+                                   const InterpKernel *y_filters,
+                                   int y0_q4, int y_step_q4, int w, int h,
+                                   int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (y = 0; y < h; ++y) {
+      const uint16_t *src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
+      const int16_t *const y_filter = y_filters[y_q4 & SUBPEL_MASK];
+      int k, sum = 0;
+      for (k = 0; k < SUBPEL_TAPS; ++k)
+        sum += src_y[k * src_stride] * y_filter[k];
+      dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] +
+          clip_pixel_high(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd), 1);
+      y_q4 += y_step_q4;
+    }
+    ++src;
+    ++dst;
+  }
+}
+
+static void high_convolve(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const InterpKernel *const x_filters,
+                          int x0_q4, int x_step_q4,
+                          const InterpKernel *const y_filters,
+                          int y0_q4, int y_step_q4,
+                          int w, int h, int bd) {
+  // Note: Fixed size intermediate buffer, temp, places limits on parameters.
+  // 2d filtering proceeds in 2 steps:
+  //   (1) Interpolate horizontally into an intermediate buffer, temp.
+  //   (2) Interpolate temp vertically to derive the sub-pixel result.
+  // Deriving the maximum number of rows in the temp buffer (135):
+  // --Smallest scaling factor is x1/2 ==> y_step_q4 = 32 (Normative).
+  // --Largest block size is 64x64 pixels.
+  // --64 rows in the downscaled frame span a distance of (64 - 1) * 32 in the
+  //   original frame (in 1/16th pixel units).
+  // --Must round-up because block may be located at sub-pixel position.
+  // --Require an additional SUBPEL_TAPS rows for the 8-tap filter tails.
+  // --((64 - 1) * 32 + 15) >> 4 + 8 = 135.
+  uint16_t temp[64 * 135];
+  int intermediate_height = (((h - 1) * y_step_q4 + 15) >> 4) + SUBPEL_TAPS;
+
+  assert(w <= 64);
+  assert(h <= 64);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  if (intermediate_height < h)
+    intermediate_height = h;
+
+  high_convolve_horiz(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                      src_stride, CONVERT_TO_BYTEPTR(temp), 64,
+                      x_filters, x0_q4, x_step_q4, w,
+                      intermediate_height, bd);
+  high_convolve_vert(CONVERT_TO_BYTEPTR(temp) + 64 * (SUBPEL_TAPS / 2 - 1),
+                     64, dst, dst_stride, y_filters, y0_q4, y_step_q4,
+                     w, h, bd);
+}
+
+
+void vp9_high_convolve8_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                                uint8_t *dst, ptrdiff_t dst_stride,
+                                const int16_t *filter_x, int x_step_q4,
+                                const int16_t *filter_y, int y_step_q4,
+                                int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  high_convolve_horiz(src, src_stride, dst, dst_stride, filters_x,
+                      x0_q4, x_step_q4, w, h, bd);
+}
+
+void vp9_high_convolve8_avg_horiz_c(const uint8_t *src, ptrdiff_t src_stride,
+                                    uint8_t *dst, ptrdiff_t dst_stride,
+                                    const int16_t *filter_x, int x_step_q4,
+                                    const int16_t *filter_y, int y_step_q4,
+                                    int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+  (void)filter_y;
+  (void)y_step_q4;
+
+  high_convolve_avg_horiz(src, src_stride, dst, dst_stride, filters_x,
+                     x0_q4, x_step_q4, w, h, bd);
+}
+
+void vp9_high_convolve8_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                               uint8_t *dst, ptrdiff_t dst_stride,
+                               const int16_t *filter_x, int x_step_q4,
+                               const int16_t *filter_y, int y_step_q4,
+                               int w, int h, int bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  high_convolve_vert(src, src_stride, dst, dst_stride, filters_y,
+                     y0_q4, y_step_q4, w, h, bd);
+}
+
+void vp9_high_convolve8_avg_vert_c(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const int16_t *filter_x, int x_step_q4,
+                                   const int16_t *filter_y, int y_step_q4,
+                                   int w, int h, int bd) {
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+  (void)filter_x;
+  (void)x_step_q4;
+
+  high_convolve_avg_vert(src, src_stride, dst, dst_stride, filters_y,
+                         y0_q4, y_step_q4, w, h, bd);
+}
+
+void vp9_high_convolve8_c(const uint8_t *src, ptrdiff_t src_stride,
+                          uint8_t *dst, ptrdiff_t dst_stride,
+                          const int16_t *filter_x, int x_step_q4,
+                          const int16_t *filter_y, int y_step_q4,
+                          int w, int h, int bd) {
+  const InterpKernel *const filters_x = get_filter_base(filter_x);
+  const int x0_q4 = get_filter_offset(filter_x, filters_x);
+
+  const InterpKernel *const filters_y = get_filter_base(filter_y);
+  const int y0_q4 = get_filter_offset(filter_y, filters_y);
+
+  high_convolve(src, src_stride, dst, dst_stride,
+                filters_x, x0_q4, x_step_q4,
+                filters_y, y0_q4, y_step_q4, w, h, bd);
+}
+
+void vp9_high_convolve8_avg_c(const uint8_t *src, ptrdiff_t src_stride,
+                              uint8_t *dst, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int x_step_q4,
+                              const int16_t *filter_y, int y_step_q4,
+                              int w, int h, int bd) {
+  // Fixed size intermediate buffer places limits on parameters.
+  DECLARE_ALIGNED_ARRAY(16, uint16_t, temp, 64 * 64);
+  assert(w <= 64);
+  assert(h <= 64);
+
+  vp9_high_convolve8_c(src, src_stride, CONVERT_TO_BYTEPTR(temp), 64,
+                       filter_x, x_step_q4, filter_y, y_step_q4, w, h, bd);
+  vp9_high_convolve_avg_c(CONVERT_TO_BYTEPTR(temp), 64, dst, dst_stride,
+                          NULL, 0, NULL, 0, w, h, bd);
+}
+
+void vp9_high_convolve_copy_c(const uint8_t *src8, ptrdiff_t src_stride,
+                              uint8_t *dst8, ptrdiff_t dst_stride,
+                              const int16_t *filter_x, int filter_x_stride,
+                              const int16_t *filter_y, int filter_y_stride,
+                              int w, int h, int bd) {
+  int r;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+  (void)bd;
+
+  for (r = h; r > 0; --r) {
+    vpx_memcpy(dst, src, w * sizeof(uint16_t));
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+
+void vp9_high_convolve_avg_c(const uint8_t *src8, ptrdiff_t src_stride,
+                             uint8_t *dst8, ptrdiff_t dst_stride,
+                             const int16_t *filter_x, int filter_x_stride,
+                             const int16_t *filter_y, int filter_y_stride,
+                             int w, int h, int bd) {
+  int x, y;
+  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+  (void)filter_x;
+  (void)filter_y;
+  (void)filter_x_stride;
+  (void)filter_y_stride;
+  (void)bd;
+
+  for (y = 0; y < h; ++y) {
+    for (x = 0; x < w; ++x) {
+      dst[x] = ROUND_POWER_OF_TWO(dst[x] + src[x], 1);
+    }
+    src += src_stride;
+    dst += dst_stride;
+  }
+}
+#endif

vp9/common/vp9_convolve.h

@@ -23,6 +23,14 @@ typedef void (*convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride,
                               const int16_t *filter_y, int y_step_q4,
                               int w, int h);
 
+#if CONFIG_VP9_HIGHBITDEPTH
+typedef void (*high_convolve_fn_t)(const uint8_t *src, ptrdiff_t src_stride,
+                                   uint8_t *dst, ptrdiff_t dst_stride,
+                                   const int16_t *filter_x, int x_step_q4,
+                                   const int16_t *filter_y, int y_step_q4,
+                                   int w, int h, int bd);
+#endif
+
 #ifdef __cplusplus
 }  // extern "C"
 #endif

vp9/common/vp9_reconinter.c

@@ -63,6 +63,53 @@ static void build_mc_border(const uint8_t *src, int src_stride,
   } while (--b_h);
 }
 
+#if CONFIG_VP9_HIGHBITDEPTH
+static void high_build_mc_border(const uint8_t *src8, int src_stride,
+                                 uint16_t *dst, int dst_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 uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+  const uint16_t *ref_row = src - x - y * src_stride;
+
+  if (y >= h)
+    ref_row += (h - 1) * src_stride;
+  else if (y > 0)
+    ref_row += y * src_stride;
+
+  do {
+    int right = 0, copy;
+    int left = x < 0 ? -x : 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)
+      vpx_memset16(dst, ref_row[0], left);
+
+    if (copy)
+      memcpy(dst + left, ref_row + x + left, copy * sizeof(uint16_t));
+
+    if (right)
+      vpx_memset16(dst + left + copy, ref_row[w - 1], right);
+
+    dst += dst_stride;
+    ++y;
+
+    if (y > 0 && y < h)
+      ref_row += src_stride;
+  } while (--b_h);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
 static void inter_predictor(const uint8_t *src, int src_stride,
                             uint8_t *dst, int dst_stride,
                             const int subpel_x,
@@ -97,6 +144,42 @@ void vp9_build_inter_predictor(const uint8_t *src, int src_stride,
                   sf, w, h, ref, kernel, sf->x_step_q4, sf->y_step_q4);
 }
 
+#if CONFIG_VP9_HIGHBITDEPTH
+static void high_inter_predictor(const uint8_t *src, int src_stride,
+                                 uint8_t *dst, int dst_stride,
+                                 const int subpel_x,
+                                 const int subpel_y,
+                                 const struct scale_factors *sf,
+                                 int w, int h, int ref,
+                                 const InterpKernel *kernel,
+                                 int xs, int ys, int bd) {
+  sf->high_predict[subpel_x != 0][subpel_y != 0][ref](
+      src, src_stride, dst, dst_stride,
+      kernel[subpel_x], xs, kernel[subpel_y], ys, w, h, bd);
+}
+
+void vp9_high_build_inter_predictor(const uint8_t *src, int src_stride,
+                                    uint8_t *dst, int dst_stride,
+                                    const MV *src_mv,
+                                    const struct scale_factors *sf,
+                                    int w, int h, int ref,
+                                    const InterpKernel *kernel,
+                                    enum mv_precision precision,
+                                    int x, int y, int bd) {
+  const int is_q4 = precision == MV_PRECISION_Q4;
+  const MV mv_q4 = { is_q4 ? src_mv->row : src_mv->row * 2,
+                     is_q4 ? src_mv->col : src_mv->col * 2 };
+  MV32 mv = vp9_scale_mv(&mv_q4, x, y, sf);
+  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);
+
+  high_inter_predictor(src, src_stride, dst, dst_stride, subpel_x, subpel_y,
+                       sf, w, h, ref, kernel, sf->x_step_q4, sf->y_step_q4, bd);
+}
+#endif  // CONFIG_VP9_HIGHBITDEPTH
+
 static INLINE int round_mv_comp_q4(int value) {
   return (value < 0 ? value - 2 : value + 2) / 4;
 }
@@ -222,8 +305,19 @@ static void build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
     pre += (scaled_mv.row >> SUBPEL_BITS) * pre_buf->stride
            + (scaled_mv.col >> SUBPEL_BITS);
 
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      high_inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
+                           subpel_x, subpel_y, sf, w, h, ref, kernel, xs, ys,
+                           xd->bd);
+    } else {
+      inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
+                      subpel_x, subpel_y, sf, w, h, ref, kernel, xs, ys);
+    }
+#else
     inter_predictor(pre, pre_buf->stride, dst, dst_buf->stride,
                     subpel_x, subpel_y, sf, w, h, ref, kernel, xs, ys);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
   }
 }
 
@@ -393,16 +487,64 @@ static void dec_build_inter_predictors(MACROBLOCKD *xd, int plane, int block,
           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, pre_buf->stride, xd->mc_buf, x1 - x0 + 1,
-                        x0, y0, x1 - x0 + 1, y1 - y0 + 1, frame_width,
+#if CONFIG_VP9_HIGHBITDEPTH
+        if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+          high_build_mc_border(buf_ptr1,
+                               pre_buf->stride,
+                               xd->mc_buf_high,
+                               x1 - x0 + 1,
+                               x0,
+                               y0,
+                               x1 - x0 + 1,
+                               y1 - y0 + 1,
+                               frame_width,
+                               frame_height);
+          buf_stride = x1 - x0 + 1;
+          buf_ptr = CONVERT_TO_BYTEPTR(xd->mc_buf_high) +
+              y_pad * 3 * buf_stride + x_pad * 3;
+        } else {
+          build_mc_border(buf_ptr1,
+                          pre_buf->stride,
+                          xd->mc_buf,
+                          x1 - x0 + 1,
+                          x0,
+                          y0,
+                          x1 - x0 + 1,
+                          y1 - y0 + 1,
+                          frame_width,
+                          frame_height);
+          buf_stride = x1 - x0 + 1;
+          buf_ptr = xd->mc_buf + y_pad * 3 * buf_stride + x_pad * 3;
+        }
+#else
+        build_mc_border(buf_ptr1,
+                        pre_buf->stride,
+                        xd->mc_buf,
+                        x1 - x0 + 1,
+                        x0,
+                        y0,
+                        x1 - x0 + 1,
+                        y1 - y0 + 1,
+                        frame_width,
                         frame_height);
         buf_stride = x1 - x0 + 1;
         buf_ptr = xd->mc_buf + y_pad * 3 * buf_stride + x_pad * 3;
+#endif  // CONFIG_VP9_HIGHBITDEPTH
       }
     }
 
+#if CONFIG_VP9_HIGHBITDEPTH
+    if (xd->cur_buf->flags & YV12_FLAG_HIGHBITDEPTH) {
+      high_inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+                           subpel_y, sf, w, h, ref, kernel, xs, ys, xd->bd);
+    } else {
+      inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
+                    subpel_y, sf, w, h, ref, kernel, xs, ys);
+    }
+#else
     inter_predictor(buf_ptr, buf_stride, dst, dst_buf->stride, subpel_x,
                     subpel_y, sf, w, h, ref, kernel, xs, ys);
+#endif  // CONFIG_VP9_HIGHBITDEPTH
   }
 }
 

vp9/common/vp9_reconinter.h

@@ -39,6 +39,17 @@ void vp9_build_inter_predictor(const uint8_t *src, int src_stride,
                                enum mv_precision precision,
                                int x, int y);
 
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_high_build_inter_predictor(const uint8_t *src, int src_stride,
+                                    uint8_t *dst, int dst_stride,
+                                    const MV *mv_q3,
+                                    const struct scale_factors *sf,
+                                    int w, int h, int do_avg,
+                                    const InterpKernel *kernel,
+                                    enum mv_precision precision,
+                                    int x, int y, int bd);
+#endif
+
 static INLINE int scaled_buffer_offset(int x_offset, int y_offset, int stride,
                                        const struct scale_factors *sf) {
   const int x = sf ? sf->scale_value_x(x_offset, sf) : x_offset;

vp9/common/vp9_rtcd_defs.pl

@@ -606,6 +606,33 @@ if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
   add_proto qw/void vp9_high_dc_128_predictor_32x32/, "uint16_t *dst, ptrdiff_t y_stride, const uint16_t *above, const uint16_t *left, int bps";
   specialize qw/vp9_high_dc_128_predictor_32x32/;
 
+  #
+  # Sub Pixel Filters
+  #
+  add_proto qw/void vp9_high_convolve_copy/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_high_convolve_copy/;
+
+  add_proto qw/void vp9_high_convolve_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_high_convolve_avg/;
+
+  add_proto qw/void vp9_high_convolve8/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_high_convolve8/, "$sse2_x86_64";
+
+  add_proto qw/void vp9_high_convolve8_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_high_convolve8_horiz/, "$sse2_x86_64";
+
+  add_proto qw/void vp9_high_convolve8_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_high_convolve8_vert/, "$sse2_x86_64";
+
+  add_proto qw/void vp9_high_convolve8_avg/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_high_convolve8_avg/, "$sse2_x86_64";
+
+  add_proto qw/void vp9_high_convolve8_avg_horiz/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_high_convolve8_avg_horiz/, "$sse2_x86_64";
+
+  add_proto qw/void vp9_high_convolve8_avg_vert/, "const uint8_t *src, ptrdiff_t src_stride, uint8_t *dst, ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4, const int16_t *filter_y, int y_step_q4, int w, int h, int bps";
+  specialize qw/vp9_high_convolve8_avg_vert/, "$sse2_x86_64";
+
   #
   # dct
   #

vp9/common/vp9_scale.c

@@ -43,9 +43,16 @@ MV32 vp9_scale_mv(const MV *mv, int x, int y, const struct scale_factors *sf) {
   return res;
 }
 
+#if CONFIG_VP9_HIGHBITDEPTH
+void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
+                                       int other_w, int other_h,
+                                       int this_w, int this_h,
+                                       int use_high) {
+#else
 void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
                                        int other_w, int other_h,
                                        int this_w, int this_h) {
+#endif
   if (!valid_ref_frame_size(other_w, other_h, this_w, this_h)) {
     sf->x_scale_fp = REF_INVALID_SCALE;
     sf->y_scale_fp = REF_INVALID_SCALE;
@@ -111,4 +118,48 @@ void vp9_setup_scale_factors_for_frame(struct scale_factors *sf,
   // 2D subpel motion always gets filtered in both directions
   sf->predict[1][1][0] = vp9_convolve8;
   sf->predict[1][1][1] = vp9_convolve8_avg;
+#if CONFIG_VP9_HIGHBITDEPTH
+  if (use_high) {
+    if (sf->x_step_q4 == 16) {
+      if (sf->y_step_q4 == 16) {
+        // No scaling in either direction.
+        sf->high_predict[0][0][0] = vp9_high_convolve_copy;
+        sf->high_predict[0][0][1] = vp9_high_convolve_avg;
+        sf->high_predict[0][1][0] = vp9_high_convolve8_vert;
+        sf->high_predict[0][1][1] = vp9_high_convolve8_avg_vert;
+        sf->high_predict[1][0][0] = vp9_high_convolve8_horiz;
+        sf->high_predict[1][0][1] = vp9_high_convolve8_avg_horiz;
+      } else {
+        // No scaling in x direction. Must always scale in the y direction.
+        sf->high_predict[0][0][0] = vp9_high_convolve8_vert;
+        sf->high_predict[0][0][1] = vp9_high_convolve8_avg_vert;
+        sf->high_predict[0][1][0] = vp9_high_convolve8_vert;
+        sf->high_predict[0][1][1] = vp9_high_convolve8_avg_vert;
+        sf->high_predict[1][0][0] = vp9_high_convolve8;
+        sf->high_predict[1][0][1] = vp9_high_convolve8_avg;
+      }
+    } else {
+      if (sf->y_step_q4 == 16) {
+        // No scaling in the y direction. Must always scale in the x direction.
+        sf->high_predict[0][0][0] = vp9_high_convolve8_horiz;
+        sf->high_predict[0][0][1] = vp9_high_convolve8_avg_horiz;
+        sf->high_predict[0][1][0] = vp9_high_convolve8;
+        sf->high_predict[0][1][1] = vp9_high_convolve8_avg;
+        sf->high_predict[1][0][0] = vp9_high_convolve8_horiz;
+        sf->high_predict[1][0][1] = vp9_high_convolve8_avg_horiz;
+      } else {
+        // Must always scale in both directions.
+        sf->high_predict[0][0][0] = vp9_high_convolve8;
+        sf->high_predict[0][0][1] = vp9_high_convolve8_avg;