Update aom to v1.0.0

Update aom to commit id d14c5bb4f336ef1842046089849dee4a301fbbf0.

1 files changed, 867 insertions, 1259 deletions

diff --git a/third_party/aom/av1/common/convolve.c b/third_party/aom/av1/common/convolve.c
index 5476f59a6..d57f44f8b 100644
--- a/third_party/aom/av1/common/convolve.c
+++ b/third_party/aom/av1/common/convolve.c
@@ -12,76 +12,60 @@
 #include <assert.h>
 #include <string.h>
 
-#include "./aom_dsp_rtcd.h"
-#include "./av1_rtcd.h"
+#include "config/aom_dsp_rtcd.h"
+#include "config/av1_rtcd.h"
+
 #include "av1/common/blockd.h"
 #include "av1/common/convolve.h"
 #include "av1/common/filter.h"
 #include "av1/common/onyxc_int.h"
+#include "av1/common/resize.h"
 #include "aom_dsp/aom_dsp_common.h"
 #include "aom_ports/mem.h"
 
-#define MAX_BLOCK_WIDTH (MAX_SB_SIZE)
-#define MAX_BLOCK_HEIGHT (MAX_SB_SIZE)
-#define MAX_STEP (32)
-
-void av1_convolve_horiz_c(const uint8_t *src, int src_stride, uint8_t *dst,
-                          int dst_stride, int w, int h,
-                          const InterpFilterParams filter_params,
-                          const int subpel_x_q4, int x_step_q4,
-                          ConvolveParams *conv_params) {
-  int x, y;
-  int filter_size = filter_params.taps;
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-  src -= filter_size / 2 - 1;
-  for (y = 0; y < h; ++y) {
-    int x_q4 = subpel_x_q4;
-    for (x = 0; x < w; ++x) {
-      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
-      const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
-          filter_params, x_q4 & SUBPEL_MASK);
-      int k, sum = 0;
-      for (k = 0; k < filter_size; ++k) sum += src_x[k] * x_filter[k];
-
-      sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
-      if (conv_params->do_average)
-        dst[x] = ROUND_POWER_OF_TWO(dst[x] + sum, 1);
-      else
-        dst[x] = sum;
-
-      x_q4 += x_step_q4;
+void av1_convolve_horiz_rs_c(const uint8_t *src, int src_stride, uint8_t *dst,
+                             int dst_stride, int w, int h,
+                             const int16_t *x_filters, int x0_qn,
+                             int x_step_qn) {
+  src -= UPSCALE_NORMATIVE_TAPS / 2 - 1;
+  for (int y = 0; y < h; ++y) {
+    int x_qn = x0_qn;
+    for (int x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_qn >> RS_SCALE_SUBPEL_BITS];
+      const int x_filter_idx =
+          (x_qn & RS_SCALE_SUBPEL_MASK) >> RS_SCALE_EXTRA_BITS;
+      assert(x_filter_idx <= RS_SUBPEL_MASK);
+      const int16_t *const x_filter =
+          &x_filters[x_filter_idx * UPSCALE_NORMATIVE_TAPS];
+      int sum = 0;
+      for (int k = 0; k < UPSCALE_NORMATIVE_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
+      x_qn += x_step_qn;
     }
     src += src_stride;
     dst += dst_stride;
   }
 }
 
-void av1_convolve_horiz_scale(const uint8_t *src, int src_stride, uint8_t *dst,
-                              int dst_stride, int w, int h,
-                              const InterpFilterParams filter_params,
-                              const int subpel_x_qn, int x_step_qn,
-                              ConvolveParams *conv_params) {
-  int x, y;
-  int filter_size = filter_params.taps;
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-  src -= filter_size / 2 - 1;
-  for (y = 0; y < h; ++y) {
-    int x_qn = subpel_x_qn;
-    for (x = 0; x < w; ++x) {
-      const uint8_t *const src_x = &src[x_qn >> SCALE_SUBPEL_BITS];
-      const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
-      assert(x_filter_idx < SUBPEL_SHIFTS);
-      const int16_t *x_filter =
-          av1_get_interp_filter_subpel_kernel(filter_params, x_filter_idx);
-      int k, sum = 0;
-      for (k = 0; k < filter_size; ++k) sum += src_x[k] * x_filter[k];
-
-      sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
-      if (conv_params->do_average)
-        dst[x] = ROUND_POWER_OF_TWO(dst[x] + sum, 1);
-      else
-        dst[x] = sum;
-
+void av1_highbd_convolve_horiz_rs_c(const uint16_t *src, int src_stride,
+                                    uint16_t *dst, int dst_stride, int w, int h,
+                                    const int16_t *x_filters, int x0_qn,
+                                    int x_step_qn, int bd) {
+  src -= UPSCALE_NORMATIVE_TAPS / 2 - 1;
+  for (int y = 0; y < h; ++y) {
+    int x_qn = x0_qn;
+    for (int x = 0; x < w; ++x) {
+      const uint16_t *const src_x = &src[x_qn >> RS_SCALE_SUBPEL_BITS];
+      const int x_filter_idx =
+          (x_qn & RS_SCALE_SUBPEL_MASK) >> RS_SCALE_EXTRA_BITS;
+      assert(x_filter_idx <= RS_SUBPEL_MASK);
+      const int16_t *const x_filter =
+          &x_filters[x_filter_idx * UPSCALE_NORMATIVE_TAPS];
+      int sum = 0;
+      for (int k = 0; k < UPSCALE_NORMATIVE_TAPS; ++k)
+        sum += src_x[k] * x_filter[k];
+      dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
       x_qn += x_step_qn;
     }
     src += src_stride;
@@ -89,417 +73,358 @@ void av1_convolve_horiz_scale(const uint8_t *src, int src_stride, uint8_t *dst,
   }
 }
 
-void av1_convolve_vert_c(const uint8_t *src, int src_stride, uint8_t *dst,
-                         int dst_stride, int w, int h,
-                         const InterpFilterParams filter_params,
-                         const int subpel_y_q4, int y_step_q4,
-                         ConvolveParams *conv_params) {
-  int x, y;
-  int filter_size = filter_params.taps;
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-  src -= src_stride * (filter_size / 2 - 1);
-  for (x = 0; x < w; ++x) {
-    int y_q4 = subpel_y_q4;
-    for (y = 0; y < h; ++y) {
-      const uint8_t *const src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
-      const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
-          filter_params, y_q4 & SUBPEL_MASK);
-      int k, sum = 0;
-      for (k = 0; k < filter_size; ++k)
-        sum += src_y[k * src_stride] * y_filter[k];
-
-      sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
-      if (conv_params->do_average)
-        dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] + sum, 1);
-      else
-        dst[y * dst_stride] = sum;
-
-      y_q4 += y_step_q4;
-    }
-    ++src;
-    ++dst;
-  }
-}
-
-void av1_convolve_vert_scale(const uint8_t *src, int src_stride, uint8_t *dst,
-                             int dst_stride, int w, int h,
-                             const InterpFilterParams filter_params,
-                             const int subpel_y_qn, int y_step_qn,
-                             ConvolveParams *conv_params) {
-  int x, y;
-  int filter_size = filter_params.taps;
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-  src -= src_stride * (filter_size / 2 - 1);
-  for (x = 0; x < w; ++x) {
-    int y_qn = subpel_y_qn;
-    for (y = 0; y < h; ++y) {
-      const uint8_t *const src_y =
-          &src[(y_qn >> SCALE_SUBPEL_BITS) * src_stride];
-      const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
-      assert(y_filter_idx < SUBPEL_SHIFTS);
-      const int16_t *y_filter =
-          av1_get_interp_filter_subpel_kernel(filter_params, y_filter_idx);
-      int k, sum = 0;
-      for (k = 0; k < filter_size; ++k)
-        sum += src_y[k * src_stride] * y_filter[k];
-
-      sum = clip_pixel(ROUND_POWER_OF_TWO(sum, FILTER_BITS));
-      if (conv_params->do_average)
-        dst[y * dst_stride] = ROUND_POWER_OF_TWO(dst[y * dst_stride] + sum, 1);
-      else
-        dst[y * dst_stride] = sum;
-
-      y_qn += y_step_qn;
-    }
-    ++src;
-    ++dst;
-  }
-}
-
-static void convolve_copy(const uint8_t *src, int src_stride, uint8_t *dst,
+void av1_convolve_2d_sr_c(const uint8_t *src, int src_stride, uint8_t *dst,
                           int dst_stride, int w, int h,
+                          InterpFilterParams *filter_params_x,
+                          InterpFilterParams *filter_params_y,
+                          const int subpel_x_q4, const int subpel_y_q4,
                           ConvolveParams *conv_params) {
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-  if (conv_params->do_average == 0) {
-    int r;
-    for (r = 0; r < h; ++r) {
-      memcpy(dst, src, w);
-      src += src_stride;
-      dst += dst_stride;
-    }
-  } else {
-    int r, c;
-    for (r = 0; r < h; ++r) {
-      for (c = 0; c < w; ++c) {
-        dst[c] = clip_pixel(ROUND_POWER_OF_TWO(dst[c] + src[c], 1));
+  int16_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
+  int im_h = h + filter_params_y->taps - 1;
+  int im_stride = w;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const int bd = 8;
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
+
+  // horizontal filter
+  const uint8_t *src_horiz = src - fo_vert * src_stride;
+  const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
+      *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
+  for (int y = 0; y < im_h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t sum = (1 << (bd + FILTER_BITS - 1));
+      for (int k = 0; k < filter_params_x->taps; ++k) {
+        sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
       }
-      src += src_stride;
-      dst += dst_stride;
+      assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
+      im_block[y * im_stride + x] =
+          (int16_t)ROUND_POWER_OF_TWO(sum, conv_params->round_0);
     }
   }
-}
 
-void av1_convolve_horiz_facade(const uint8_t *src, int src_stride, uint8_t *dst,
-                               int dst_stride, int w, int h,
-                               const InterpFilterParams filter_params,
-                               const int subpel_x_q4, int x_step_q4,
-                               ConvolveParams *conv_params) {
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-  if (filter_params.taps == SUBPEL_TAPS) {
-    const int16_t *filter_x =
-        av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
-    if (conv_params->do_average == 0)
-      aom_convolve8_horiz(src, src_stride, dst, dst_stride, filter_x, x_step_q4,
-                          NULL, -1, w, h);
-    else
-      aom_convolve8_avg_horiz(src, src_stride, dst, dst_stride, filter_x,
-                              x_step_q4, NULL, -1, w, h);
-  } else {
-    av1_convolve_horiz(src, src_stride, dst, dst_stride, w, h, filter_params,
-                       subpel_x_q4, x_step_q4, conv_params);
+  // vertical filter
+  int16_t *src_vert = im_block + fo_vert * im_stride;
+  const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+      *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t sum = 1 << offset_bits;
+      for (int k = 0; k < filter_params_y->taps; ++k) {
+        sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
+      }
+      assert(0 <= sum && sum < (1 << (offset_bits + 2)));
+      int16_t res = ROUND_POWER_OF_TWO(sum, conv_params->round_1) -
+                    ((1 << (offset_bits - conv_params->round_1)) +
+                     (1 << (offset_bits - conv_params->round_1 - 1)));
+      dst[y * dst_stride + x] = clip_pixel(ROUND_POWER_OF_TWO(res, bits));
+    }
   }
 }
 
-void av1_convolve_horiz_facade_c(const uint8_t *src, int src_stride,
-                                 uint8_t *dst, int dst_stride, int w, int h,
-                                 const InterpFilterParams filter_params,
-                                 const int subpel_x_q4, int x_step_q4,
-                                 ConvolveParams *conv_params) {
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-  if (filter_params.taps == SUBPEL_TAPS) {
-    const int16_t *filter_x =
-        av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
-    if (conv_params->do_average == 0)
-      aom_convolve8_horiz_c(src, src_stride, dst, dst_stride, filter_x,
-                            x_step_q4, NULL, -1, w, h);
-    else
-      aom_convolve8_avg_horiz_c(src, src_stride, dst, dst_stride, filter_x,
-                                x_step_q4, NULL, -1, w, h);
-  } else {
-    av1_convolve_horiz_c(src, src_stride, dst, dst_stride, w, h, filter_params,
-                         subpel_x_q4, x_step_q4, conv_params);
-  }
-}
+void av1_convolve_y_sr_c(const uint8_t *src, int src_stride, uint8_t *dst,
+                         int dst_stride, int w, int h,
+                         InterpFilterParams *filter_params_x,
+                         InterpFilterParams *filter_params_y,
+                         const int subpel_x_q4, const int subpel_y_q4,
+                         ConvolveParams *conv_params) {
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  (void)filter_params_x;
+  (void)subpel_x_q4;
+  (void)conv_params;
 
-void av1_convolve_horiz_facade_scale(const uint8_t *src, int src_stride,
-                                     uint8_t *dst, int dst_stride, int w, int h,
-                                     const InterpFilterParams filter_params,
-                                     const int subpel_x_qn, int x_step_qn,
-                                     ConvolveParams *conv_params) {
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-  if (filter_params.taps == SUBPEL_TAPS) {
-    const int16_t *filter_x = av1_get_interp_filter_subpel_kernel(
-        filter_params, subpel_x_qn >> SCALE_EXTRA_BITS);
-    if (conv_params->do_average == 0)
-      aom_convolve8_horiz_scale(src, src_stride, dst, dst_stride, filter_x,
-                                subpel_x_qn, x_step_qn, NULL, 0, -1, w, h);
-    else
-      aom_convolve8_avg_horiz_scale(src, src_stride, dst, dst_stride, filter_x,
-                                    subpel_x_qn, x_step_qn, NULL, 0, -1, w, h);
-  } else {
-    av1_convolve_horiz_scale(src, src_stride, dst, dst_stride, w, h,
-                             filter_params, subpel_x_qn, x_step_qn,
-                             conv_params);
-  }
-}
+  assert(conv_params->round_0 <= FILTER_BITS);
+  assert(((conv_params->round_0 + conv_params->round_1) <= (FILTER_BITS + 1)) ||
+         ((conv_params->round_0 + conv_params->round_1) == (2 * FILTER_BITS)));
 
-void av1_convolve_vert_facade(const uint8_t *src, int src_stride, uint8_t *dst,
-                              int dst_stride, int w, int h,
-                              const InterpFilterParams filter_params,
-                              const int subpel_y_q4, int y_step_q4,
-                              ConvolveParams *conv_params) {
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-  if (filter_params.taps == SUBPEL_TAPS) {
-    const int16_t *filter_y =
-        av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
-    if (conv_params->do_average == 0) {
-      aom_convolve8_vert(src, src_stride, dst, dst_stride, NULL, -1, filter_y,
-                         y_step_q4, w, h);
-    } else {
-      aom_convolve8_avg_vert(src, src_stride, dst, dst_stride, NULL, -1,
-                             filter_y, y_step_q4, w, h);
+  // vertical filter
+  const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+      *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t res = 0;
+      for (int k = 0; k < filter_params_y->taps; ++k) {
+        res += y_filter[k] * src[(y - fo_vert + k) * src_stride + x];
+      }
+      dst[y * dst_stride + x] =
+          clip_pixel(ROUND_POWER_OF_TWO(res, FILTER_BITS));
     }
-  } else {
-    av1_convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params,
-                      subpel_y_q4, y_step_q4, conv_params);
   }
 }
 
-void av1_convolve_vert_facade_c(const uint8_t *src, int src_stride,
-                                uint8_t *dst, int dst_stride, int w, int h,
-                                const InterpFilterParams filter_params,
-                                const int subpel_y_q4, int y_step_q4,
-                                ConvolveParams *conv_params) {
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-  if (filter_params.taps == SUBPEL_TAPS) {
-    const int16_t *filter_y =
-        av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
-    if (conv_params->do_average == 0) {
-      aom_convolve8_vert_c(src, src_stride, dst, dst_stride, NULL, -1, filter_y,
-                           y_step_q4, w, h);
-    } else {
-      aom_convolve8_avg_vert_c(src, src_stride, dst, dst_stride, NULL, -1,
-                               filter_y, y_step_q4, w, h);
-    }
-  } else {
-    av1_convolve_vert_c(src, src_stride, dst, dst_stride, w, h, filter_params,
-                        subpel_y_q4, y_step_q4, conv_params);
-  }
-}
+void av1_convolve_x_sr_c(const uint8_t *src, int src_stride, uint8_t *dst,
+                         int dst_stride, int w, int h,
+                         InterpFilterParams *filter_params_x,
+                         InterpFilterParams *filter_params_y,
+                         const int subpel_x_q4, const int subpel_y_q4,
+                         ConvolveParams *conv_params) {
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const int bits = FILTER_BITS - conv_params->round_0;
+  (void)filter_params_y;
+  (void)subpel_y_q4;
+  (void)conv_params;
+
+  assert(bits >= 0);
+  assert((FILTER_BITS - conv_params->round_1) >= 0 ||
+         ((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS));
 
-void av1_convolve_vert_facade_scale(const uint8_t *src, int src_stride,
-                                    uint8_t *dst, int dst_stride, int w, int h,
-                                    const InterpFilterParams filter_params,
-                                    const int subpel_y_qn, int y_step_qn,
-                                    ConvolveParams *conv_params) {
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-  if (filter_params.taps == SUBPEL_TAPS) {
-    const int16_t *filter_y = av1_get_interp_filter_subpel_kernel(
-        filter_params, subpel_y_qn >> SCALE_EXTRA_BITS);
-    if (conv_params->do_average == 0) {
-      aom_convolve8_vert_scale(src, src_stride, dst, dst_stride, NULL, 0, -1,
-                               filter_y, subpel_y_qn, y_step_qn, w, h);
-    } else {
-      aom_convolve8_avg_vert_scale(src, src_stride, dst, dst_stride, NULL, 0,
-                                   -1, filter_y, subpel_y_qn, y_step_qn, w, h);
+  // horizontal filter
+  const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
+      *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t res = 0;
+      for (int k = 0; k < filter_params_x->taps; ++k) {
+        res += x_filter[k] * src[y * src_stride + x - fo_horiz + k];
+      }
+      res = ROUND_POWER_OF_TWO(res, conv_params->round_0);
+      dst[y * dst_stride + x] = clip_pixel(ROUND_POWER_OF_TWO(res, bits));
     }
-  } else {
-    av1_convolve_vert_scale(src, src_stride, dst, dst_stride, w, h,
-                            filter_params, subpel_y_qn, y_step_qn, conv_params);
   }
 }
 
-#if CONFIG_CONVOLVE_ROUND
-void av1_convolve_rounding_c(const int32_t *src, int src_stride, uint8_t *dst,
-                             int dst_stride, int w, int h, int bits) {
-  int r, c;
-  for (r = 0; r < h; ++r) {
-    for (c = 0; c < w; ++c) {
-      dst[r * dst_stride + c] =
-          clip_pixel(ROUND_POWER_OF_TWO(src[r * src_stride + c], bits));
-    }
+void av1_convolve_2d_copy_sr_c(const uint8_t *src, int src_stride, uint8_t *dst,
+                               int dst_stride, int w, int h,
+                               InterpFilterParams *filter_params_x,
+                               InterpFilterParams *filter_params_y,
+                               const int subpel_x_q4, const int subpel_y_q4,
+                               ConvolveParams *conv_params) {
+  (void)filter_params_x;
+  (void)filter_params_y;
+  (void)subpel_x_q4;
+  (void)subpel_y_q4;
+  (void)conv_params;
+
+  for (int y = 0; y < h; ++y) {
+    memcpy(dst + y * dst_stride, src + y * src_stride, w * sizeof(src[0]));
   }
 }
 
-#if CONFIG_COMPOUND_ROUND
-void av1_convolve_2d_c(const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst,
-                       int dst_stride, int w, int h,
-                       InterpFilterParams *filter_params_x,
-                       InterpFilterParams *filter_params_y,
-                       const int subpel_x_q4, const int subpel_y_q4,
-                       ConvolveParams *conv_params) {
-  int x, y, k;
-  uint8_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
+void av1_jnt_convolve_2d_c(const uint8_t *src, int src_stride, uint8_t *dst8,
+                           int dst8_stride, int w, int h,
+                           InterpFilterParams *filter_params_x,
+                           InterpFilterParams *filter_params_y,
+                           const int subpel_x_q4, const int subpel_y_q4,
+                           ConvolveParams *conv_params) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  int16_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
   int im_h = h + filter_params_y->taps - 1;
   int im_stride = w;
   const int fo_vert = filter_params_y->taps / 2 - 1;
   const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const int bd = 8;
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
 
   // horizontal filter
   const uint8_t *src_horiz = src - fo_vert * src_stride;
   const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
       *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
-  for (y = 0; y < im_h; ++y) {
-    for (x = 0; x < w; ++x) {
-      int32_t sum = 0;
-      for (k = 0; k < filter_params_x->taps; ++k) {
+  for (int y = 0; y < im_h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t sum = (1 << (bd + FILTER_BITS - 1));
+      for (int k = 0; k < filter_params_x->taps; ++k) {
         sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
       }
+      assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
       im_block[y * im_stride + x] =
-          clip_pixel(ROUND_POWER_OF_TWO(sum, conv_params->round_0));
+          (int16_t)ROUND_POWER_OF_TWO(sum, conv_params->round_0);
     }
   }
 
   // vertical filter
-  uint8_t *src_vert = im_block + fo_vert * im_stride;
+  int16_t *src_vert = im_block + fo_vert * im_stride;
   const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
       *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
-  for (y = 0; y < h; ++y) {
-    for (x = 0; x < w; ++x) {
-      CONV_BUF_TYPE sum = 0;
-      for (k = 0; k < filter_params_y->taps; ++k) {
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t sum = 1 << offset_bits;
+      for (int k = 0; k < filter_params_y->taps; ++k) {
         sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
       }
+      assert(0 <= sum && sum < (1 << (offset_bits + 2)));
       CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
-      if (conv_params->do_average)
-        dst[y * dst_stride + x] += res;
-      else
+      if (conv_params->do_average) {
+        int32_t tmp = dst[y * dst_stride + x];
+        if (conv_params->use_jnt_comp_avg) {
+          tmp = tmp * conv_params->fwd_offset + res * conv_params->bck_offset;
+          tmp = tmp >> DIST_PRECISION_BITS;
+        } else {
+          tmp += res;
+          tmp = tmp >> 1;
+        }
+        tmp -= (1 << (offset_bits - conv_params->round_1)) +
+               (1 << (offset_bits - conv_params->round_1 - 1));
+        dst8[y * dst8_stride + x] =
+            clip_pixel(ROUND_POWER_OF_TWO(tmp, round_bits));
+      } else {
         dst[y * dst_stride + x] = res;
+      }
     }
   }
 }
 
-void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride,
-                             CONV_BUF_TYPE *dst, int dst_stride, int w, int h,
-                             InterpFilterParams *filter_params_x,
-                             InterpFilterParams *filter_params_y,
-                             const int subpel_x_qn, const int x_step_qn,
-                             const int subpel_y_qn, const int y_step_qn,
-                             ConvolveParams *conv_params) {
-  int x, y, k;
-  uint8_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
-  int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
-             filter_params_y->taps;
-  int im_stride = w;
+void av1_jnt_convolve_y_c(const uint8_t *src, int src_stride, uint8_t *dst8,
+                          int dst8_stride, int w, int h,
+                          InterpFilterParams *filter_params_x,
+                          InterpFilterParams *filter_params_y,
+                          const int subpel_x_q4, const int subpel_y_q4,
+                          ConvolveParams *conv_params) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
   const int fo_vert = filter_params_y->taps / 2 - 1;
-  const int fo_horiz = filter_params_x->taps / 2 - 1;
-
-  // horizontal filter
-  const uint8_t *src_horiz = src - fo_vert * src_stride;
-  for (y = 0; y < im_h; ++y) {
-    int x_qn = subpel_x_qn;
-    for (x = 0; x < w; ++x, x_qn += x_step_qn) {
-      const uint8_t *const src_x = &src_horiz[(x_qn >> SCALE_SUBPEL_BITS)];
-      const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
-      assert(x_filter_idx < SUBPEL_SHIFTS);
-      const int16_t *x_filter =
-          av1_get_interp_filter_subpel_kernel(*filter_params_x, x_filter_idx);
-      int sum = 0;
-      for (k = 0; k < filter_params_x->taps; ++k)
-        sum += x_filter[k] * src_x[k - fo_horiz];
-      im_block[y * im_stride + x] =
-          clip_pixel(ROUND_POWER_OF_TWO(sum, conv_params->round_0));
-    }
-    src_horiz += src_stride;
-  }
+  const int bits = FILTER_BITS - conv_params->round_0;
+  const int bd = 8;
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  const int round_offset = (1 << (offset_bits - conv_params->round_1)) +
+                           (1 << (offset_bits - conv_params->round_1 - 1));
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  (void)filter_params_x;
+  (void)subpel_x_q4;
 
   // vertical filter
-  const uint8_t *src_vert = im_block + fo_vert * im_stride;
-  for (x = 0; x < w; ++x) {
-    int y_qn = subpel_y_qn;
-    for (y = 0; y < h; ++y, y_qn += y_step_qn) {
-      const uint8_t *const src_y =
-          &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
-      const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
-      assert(y_filter_idx < SUBPEL_SHIFTS);
-      const int16_t *y_filter =
-          av1_get_interp_filter_subpel_kernel(*filter_params_y, y_filter_idx);
-      CONV_BUF_TYPE sum = 0;
-      for (k = 0; k < filter_params_y->taps; ++k) {
-        sum += y_filter[k] * src_y[(k - fo_vert) * im_stride];
+  const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+      *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t res = 0;
+      for (int k = 0; k < filter_params_y->taps; ++k) {
+        res += y_filter[k] * src[(y - fo_vert + k) * src_stride + x];
       }
-      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
-      if (conv_params->do_average)
-        dst[y * dst_stride + x] += res;
-      else
+      res *= (1 << bits);
+      res = ROUND_POWER_OF_TWO(res, conv_params->round_1) + round_offset;
+
+      if (conv_params->do_average) {
+        int32_t tmp = dst[y * dst_stride + x];
+        if (conv_params->use_jnt_comp_avg) {
+          tmp = tmp * conv_params->fwd_offset + res * conv_params->bck_offset;
+          tmp = tmp >> DIST_PRECISION_BITS;
+        } else {
+          tmp += res;
+          tmp = tmp >> 1;
+        }
+        tmp -= round_offset;
+        dst8[y * dst8_stride + x] =
+            clip_pixel(ROUND_POWER_OF_TWO(tmp, round_bits));
+      } else {
         dst[y * dst_stride + x] = res;
+      }
     }
-    src_vert++;
   }
 }
 
-#else
-
-/* When convolve-round is enabled and compound-round is disabled, we use a
-   high-precision convolve filter.
-   Note: For notes on hardware implementations, including the required
-   bit widths for various intermediate values, see the comments above
-   av1_warp_affine_c.
-*/
-void av1_convolve_2d_c(const uint8_t *src, int src_stride, CONV_BUF_TYPE *dst,
-                       int dst_stride, int w, int h,
-                       InterpFilterParams *filter_params_x,
-                       InterpFilterParams *filter_params_y,
-                       const int subpel_x_q4, const int subpel_y_q4,
-                       ConvolveParams *conv_params) {
-  int x, y, k;
-  int32_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
-  int im_h = h + filter_params_y->taps - 1;
-  int im_stride = w;
-  const int fo_vert = filter_params_y->taps / 2 - 1;
+void av1_jnt_convolve_x_c(const uint8_t *src, int src_stride, uint8_t *dst8,
+                          int dst8_stride, int w, int h,
+                          InterpFilterParams *filter_params_x,
+                          InterpFilterParams *filter_params_y,
+                          const int subpel_x_q4, const int subpel_y_q4,
+                          ConvolveParams *conv_params) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
   const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const int bits = FILTER_BITS - conv_params->round_1;
   const int bd = 8;
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  const int round_offset = (1 << (offset_bits - conv_params->round_1)) +
+                           (1 << (offset_bits - conv_params->round_1 - 1));
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  (void)filter_params_y;
+  (void)subpel_y_q4;
 
   // horizontal filter
-  const uint8_t *src_horiz = src - fo_vert * src_stride;
   const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
       *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
-  for (y = 0; y < im_h; ++y) {
-    for (x = 0; x < w; ++x) {
-      int32_t sum = (1 << (bd + FILTER_BITS - 1));
-      for (k = 0; k < filter_params_x->taps; ++k) {
-        sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t res = 0;
+      for (int k = 0; k < filter_params_x->taps; ++k) {
+        res += x_filter[k] * src[y * src_stride + x - fo_horiz + k];
+      }
+      res = (1 << bits) * ROUND_POWER_OF_TWO(res, conv_params->round_0);
+      res += round_offset;
+
+      if (conv_params->do_average) {
+        int32_t tmp = dst[y * dst_stride + x];
+        if (conv_params->use_jnt_comp_avg) {
+          tmp = tmp * conv_params->fwd_offset + res * conv_params->bck_offset;
+          tmp = tmp >> DIST_PRECISION_BITS;
+        } else {
+          tmp += res;
+          tmp = tmp >> 1;
+        }
+        tmp -= round_offset;
+        dst8[y * dst8_stride + x] =
+            clip_pixel(ROUND_POWER_OF_TWO(tmp, round_bits));
+      } else {
+        dst[y * dst_stride + x] = res;
       }
-      assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
-      im_block[y * im_stride + x] =
-          ROUND_POWER_OF_TWO(sum, conv_params->round_0);
     }
   }
+}
 
-  // vertical filter
-  int32_t *src_vert = im_block + fo_vert * im_stride;
-  const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
-      *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+void av1_jnt_convolve_2d_copy_c(const uint8_t *src, int src_stride,
+                                uint8_t *dst8, int dst8_stride, int w, int h,
+                                InterpFilterParams *filter_params_x,
+                                InterpFilterParams *filter_params_y,
+                                const int subpel_x_q4, const int subpel_y_q4,
+                                ConvolveParams *conv_params) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_1 - conv_params->round_0;
+  const int bd = 8;
   const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
-  for (y = 0; y < h; ++y) {
-    for (x = 0; x < w; ++x) {
-      CONV_BUF_TYPE sum = 1 << offset_bits;
-      for (k = 0; k < filter_params_y->taps; ++k) {
-        sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
-      }
-      assert(0 <= sum && sum < (1 << (offset_bits + 2)));
-      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1) -
-                          ((1 << (offset_bits - conv_params->round_1)) +
-                           (1 << (offset_bits - conv_params->round_1 - 1)));
-      if (conv_params->do_average)
-        dst[y * dst_stride + x] += res;
-      else
+  const int round_offset = (1 << (offset_bits - conv_params->round_1)) +
+                           (1 << (offset_bits - conv_params->round_1 - 1));
+  (void)filter_params_x;
+  (void)filter_params_y;
+  (void)subpel_x_q4;
+  (void)subpel_y_q4;
+
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      CONV_BUF_TYPE res = src[y * src_stride + x] << bits;
+      res += round_offset;
+
+      if (conv_params->do_average) {
+        int32_t tmp = dst[y * dst_stride + x];
+        if (conv_params->use_jnt_comp_avg) {
+          tmp = tmp * conv_params->fwd_offset + res * conv_params->bck_offset;
+          tmp = tmp >> DIST_PRECISION_BITS;
+        } else {
+          tmp += res;
+          tmp = tmp >> 1;
+        }
+        tmp -= round_offset;
+        dst8[y * dst8_stride + x] = clip_pixel(ROUND_POWER_OF_TWO(tmp, bits));
+      } else {
         dst[y * dst_stride + x] = res;
+      }
     }
   }
 }
 
-void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride,
-                             CONV_BUF_TYPE *dst, int dst_stride, int w, int h,
+void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride, uint8_t *dst8,
+                             int dst8_stride, int w, int h,
                              InterpFilterParams *filter_params_x,
                              InterpFilterParams *filter_params_y,
                              const int subpel_x_qn, const int x_step_qn,
                              const int subpel_y_qn, const int y_step_qn,
                              ConvolveParams *conv_params) {
-  int x, y, k;
-  int32_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
+  int16_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
   int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
              filter_params_y->taps;
+  CONV_BUF_TYPE *dst16 = conv_params->dst;
+  const int dst16_stride = conv_params->dst_stride;
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
+  assert(bits >= 0);
   int im_stride = w;
   const int fo_vert = filter_params_y->taps / 2 - 1;
   const int fo_horiz = filter_params_x->taps / 2 - 1;
@@ -507,245 +432,255 @@ void av1_convolve_2d_scale_c(const uint8_t *src, int src_stride,
 
   // horizontal filter
   const uint8_t *src_horiz = src - fo_vert * src_stride;
-  for (y = 0; y < im_h; ++y) {
+  for (int y = 0; y < im_h; ++y) {
     int x_qn = subpel_x_qn;
-    for (x = 0; x < w; ++x, x_qn += x_step_qn) {
+    for (int x = 0; x < w; ++x, x_qn += x_step_qn) {
       const uint8_t *const src_x = &src_horiz[(x_qn >> SCALE_SUBPEL_BITS)];
       const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
       assert(x_filter_idx < SUBPEL_SHIFTS);
       const int16_t *x_filter =
           av1_get_interp_filter_subpel_kernel(*filter_params_x, x_filter_idx);
       int32_t sum = (1 << (bd + FILTER_BITS - 1));
-      for (k = 0; k < filter_params_x->taps; ++k) {
+      for (int k = 0; k < filter_params_x->taps; ++k) {
         sum += x_filter[k] * src_x[k - fo_horiz];
       }
       assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
       im_block[y * im_stride + x] =
-          ROUND_POWER_OF_TWO(sum, conv_params->round_0);
+          (int16_t)ROUND_POWER_OF_TWO(sum, conv_params->round_0);
     }
     src_horiz += src_stride;
   }
 
   // vertical filter
-  int32_t *src_vert = im_block + fo_vert * im_stride;
+  int16_t *src_vert = im_block + fo_vert * im_stride;
   const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
-  for (x = 0; x < w; ++x) {
+  for (int x = 0; x < w; ++x) {
     int y_qn = subpel_y_qn;
-    for (y = 0; y < h; ++y, y_qn += y_step_qn) {
-      const int32_t *src_y = &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
+    for (int y = 0; y < h; ++y, y_qn += y_step_qn) {
+      const int16_t *src_y = &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
       const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
       assert(y_filter_idx < SUBPEL_SHIFTS);
       const int16_t *y_filter =
           av1_get_interp_filter_subpel_kernel(*filter_params_y, y_filter_idx);
-      CONV_BUF_TYPE sum = 1 << offset_bits;
-      for (k = 0; k < filter_params_y->taps; ++k) {
+      int32_t sum = 1 << offset_bits;
+      for (int k = 0; k < filter_params_y->taps; ++k) {
         sum += y_filter[k] * src_y[(k - fo_vert) * im_stride];
       }
       assert(0 <= sum && sum < (1 << (offset_bits + 2)));
-      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1) -
-                          ((1 << (offset_bits - conv_params->round_1)) +
-                           (1 << (offset_bits - conv_params->round_1 - 1)));
-      if (conv_params->do_average)
-        dst[y * dst_stride + x] += res;
-      else
-        dst[y * dst_stride + x] = res;
+      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
+      if (conv_params->is_compound) {
+        if (conv_params->do_average) {
+          int32_t tmp = dst16[y * dst16_stride + x];
+          if (conv_params->use_jnt_comp_avg) {
+            tmp = tmp * conv_params->fwd_offset + res * conv_params->bck_offset;
+            tmp = tmp >> DIST_PRECISION_BITS;
+          } else {
+            tmp += res;
+            tmp = tmp >> 1;
+          }
+          /* Subtract round offset and convolve round */
+          tmp = tmp - ((1 << (offset_bits - conv_params->round_1)) +
+                       (1 << (offset_bits - conv_params->round_1 - 1)));
+          dst8[y * dst8_stride + x] = clip_pixel(ROUND_POWER_OF_TWO(tmp, bits));
+        } else {
+          dst16[y * dst16_stride + x] = res;
+        }
+      } else {
+        /* Subtract round offset and convolve round */
+        int32_t tmp = res - ((1 << (offset_bits - conv_params->round_1)) +
+                             (1 << (offset_bits - conv_params->round_1 - 1)));
+        dst8[y * dst8_stride + x] = clip_pixel(ROUND_POWER_OF_TWO(tmp, bits));
+      }
     }
     src_vert++;
   }
 }
-#endif  // CONFIG_COMPOUND_ROUND
+
+static void convolve_2d_scale_wrapper(
+    const uint8_t *src, int src_stride, uint8_t *dst, int dst_stride, int w,
+    int h, InterpFilterParams *filter_params_x,
+    InterpFilterParams *filter_params_y, const int subpel_x_qn,
+    const int x_step_qn, const int subpel_y_qn, const int y_step_qn,
+    ConvolveParams *conv_params) {
+  if (conv_params->is_compound) {
+    assert(conv_params->dst != NULL);
+  }
+  av1_convolve_2d_scale(src, src_stride, dst, dst_stride, w, h, filter_params_x,
+                        filter_params_y, subpel_x_qn, x_step_qn, subpel_y_qn,
+                        y_step_qn, conv_params);
+}
 
 void av1_convolve_2d_facade(const uint8_t *src, int src_stride, uint8_t *dst,
                             int dst_stride, int w, int h,
                             InterpFilters interp_filters, const int subpel_x_q4,
                             int x_step_q4, const int subpel_y_q4, int y_step_q4,
-                            int scaled, ConvolveParams *conv_params) {
+                            int scaled, ConvolveParams *conv_params,
+                            const struct scale_factors *sf) {
   (void)x_step_q4;
   (void)y_step_q4;
   (void)dst;
   (void)dst_stride;
 
   InterpFilterParams filter_params_x, filter_params_y;
-  av1_get_convolve_filter_params(interp_filters, 1, &filter_params_x,
-                                 &filter_params_y);
-
-  if (filter_params_y.taps < filter_params_x.taps) {
-    uint8_t tr_src[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) *
-                   (MAX_SB_SIZE + MAX_FILTER_TAP - 1)];
-    int tr_src_stride = MAX_SB_SIZE + MAX_FILTER_TAP - 1;
-    CONV_BUF_TYPE tr_dst[MAX_SB_SIZE * MAX_SB_SIZE];
-    int tr_dst_stride = MAX_SB_SIZE;
-    int fo_vert = filter_params_y.taps / 2 - 1;
-    int fo_horiz = filter_params_x.taps / 2 - 1;
-
-    transpose_uint8(tr_src, tr_src_stride,
-                    src - fo_vert * src_stride - fo_horiz, src_stride,
-                    w + filter_params_x.taps - 1, h + filter_params_y.taps - 1);
-    transpose_int32(tr_dst, tr_dst_stride, conv_params->dst,
-                    conv_params->dst_stride, w, h);
-
-    // horizontal and vertical parameters are swapped because of the transpose
-    if (scaled)
-      av1_convolve_2d_scale(tr_src + fo_horiz * tr_src_stride + fo_vert,
-                            tr_src_stride, tr_dst, tr_dst_stride, h, w,
-                            &filter_params_y, &filter_params_x, subpel_y_q4,
-                            y_step_q4, subpel_x_q4, x_step_q4, conv_params);
-    else
-      av1_convolve_2d(tr_src + fo_horiz * tr_src_stride + fo_vert,
-                      tr_src_stride, tr_dst, tr_dst_stride, h, w,
-                      &filter_params_y, &filter_params_x, subpel_y_q4,
-                      subpel_x_q4, conv_params);
-    transpose_int32(conv_params->dst, conv_params->dst_stride, tr_dst,
-                    tr_dst_stride, h, w);
-  } else {
-    if (scaled)
-      av1_convolve_2d_scale(src, src_stride, conv_params->dst,
-                            conv_params->dst_stride, w, h, &filter_params_x,
-                            &filter_params_y, subpel_x_q4, x_step_q4,
-                            subpel_y_q4, y_step_q4, conv_params);
-    else
-      av1_convolve_2d(src, src_stride, conv_params->dst,
-                      conv_params->dst_stride, w, h, &filter_params_x,
-                      &filter_params_y, subpel_x_q4, subpel_y_q4, conv_params);
-  }
+  av1_get_convolve_filter_params(interp_filters, &filter_params_x,
+                                 &filter_params_y, w, h);
+
+  if (scaled)
+    convolve_2d_scale_wrapper(src, src_stride, dst, dst_stride, w, h,
+                              &filter_params_x, &filter_params_y, subpel_x_q4,
+                              x_step_q4, subpel_y_q4, y_step_q4, conv_params);
+  else
+    sf->convolve[subpel_x_q4 != 0][subpel_y_q4 != 0][conv_params->is_compound](
+        src, src_stride, dst, dst_stride, w, h, &filter_params_x,
+        &filter_params_y, subpel_x_q4, subpel_y_q4, conv_params);
 }
 
-#if CONFIG_HIGHBITDEPTH
-void av1_highbd_convolve_rounding_c(const int32_t *src, int src_stride,
-                                    uint8_t *dst8, int dst_stride, int w, int h,
-                                    int bits, int bd) {
-  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
-  int r, c;
-  for (r = 0; r < h; ++r) {
-    for (c = 0; c < w; ++c) {
-      dst[r * dst_stride + c] = clip_pixel_highbd(
-          ROUND_POWER_OF_TWO(src[r * src_stride + c], bits), bd);
-    }
+void av1_highbd_convolve_2d_copy_sr_c(
+    const uint16_t *src, int src_stride, uint16_t *dst, int dst_stride, int w,
+    int h, InterpFilterParams *filter_params_x,
+    InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  (void)filter_params_x;
+  (void)filter_params_y;
+  (void)subpel_x_q4;
+  (void)subpel_y_q4;
+  (void)conv_params;
+  (void)bd;
+
+  for (int y = 0; y < h; ++y) {
+    memcpy(dst + y * dst_stride, src + y * src_stride, w * sizeof(src[0]));
   }
 }
 
-#if CONFIG_COMPOUND_ROUND
-void av1_highbd_convolve_2d_c(const uint16_t *src, int src_stride,
-                              CONV_BUF_TYPE *dst, int dst_stride, int w, int h,
-                              InterpFilterParams *filter_params_x,
-                              InterpFilterParams *filter_params_y,
-                              const int subpel_x_q4, const int subpel_y_q4,
-                              ConvolveParams *conv_params, int bd) {
-  int x, y, k;
-  uint16_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
-  int im_h = h + filter_params_y->taps - 1;
-  int im_stride = w;
-  const int fo_vert = filter_params_y->taps / 2 - 1;
+void av1_highbd_convolve_x_sr_c(const uint16_t *src, int src_stride,
+                                uint16_t *dst, int dst_stride, int w, int h,
+                                InterpFilterParams *filter_params_x,
+                                InterpFilterParams *filter_params_y,
+                                const int subpel_x_q4, const int subpel_y_q4,
+                                ConvolveParams *conv_params, int bd) {
   const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const int bits = FILTER_BITS - conv_params->round_0;
+  (void)filter_params_y;
+  (void)subpel_y_q4;
+
+  assert(bits >= 0);
+  assert((FILTER_BITS - conv_params->round_1) >= 0 ||
+         ((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS));
 
   // horizontal filter
-  const uint16_t *src_horiz = src - fo_vert * src_stride;
   const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
       *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
-  for (y = 0; y < im_h; ++y) {
-    for (x = 0; x < w; ++x) {
-      int32_t sum = 0;
-      for (k = 0; k < filter_params_x->taps; ++k) {
-        sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t res = 0;
+      for (int k = 0; k < filter_params_x->taps; ++k) {
+        res += x_filter[k] * src[y * src_stride + x - fo_horiz + k];
       }
-      im_block[y * im_stride + x] =
-          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, conv_params->round_0), bd);
+      res = ROUND_POWER_OF_TWO(res, conv_params->round_0);
+      dst[y * dst_stride + x] =
+          clip_pixel_highbd(ROUND_POWER_OF_TWO(res, bits), bd);
     }
   }
+}
 
+void av1_highbd_convolve_y_sr_c(const uint16_t *src, int src_stride,
+                                uint16_t *dst, int dst_stride, int w, int h,
+                                InterpFilterParams *filter_params_x,
+                                InterpFilterParams *filter_params_y,
+                                const int subpel_x_q4, const int subpel_y_q4,
+                                ConvolveParams *conv_params, int bd) {
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  (void)filter_params_x;
+  (void)subpel_x_q4;
+  (void)conv_params;
+
+  assert(conv_params->round_0 <= FILTER_BITS);
+  assert(((conv_params->round_0 + conv_params->round_1) <= (FILTER_BITS + 1)) ||
+         ((conv_params->round_0 + conv_params->round_1) == (2 * FILTER_BITS)));
   // vertical filter
-  uint16_t *src_vert = im_block + fo_vert * im_stride;
   const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
       *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
-  for (y = 0; y < h; ++y) {
-    for (x = 0; x < w; ++x) {
-      CONV_BUF_TYPE sum = 0;
-      for (k = 0; k < filter_params_y->taps; ++k) {
-        sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t res = 0;
+      for (int k = 0; k < filter_params_y->taps; ++k) {
+        res += y_filter[k] * src[(y - fo_vert + k) * src_stride + x];
       }
-      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
-      if (conv_params->do_average)
-        dst[y * dst_stride + x] += res;
-      else
-        dst[y * dst_stride + x] = res;
+      dst[y * dst_stride + x] =
+          clip_pixel_highbd(ROUND_POWER_OF_TWO(res, FILTER_BITS), bd);
     }
   }
 }
 
-void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride,
-                                    CONV_BUF_TYPE *dst, int dst_stride, int w,
-                                    int h, InterpFilterParams *filter_params_x,
-                                    InterpFilterParams *filter_params_y,
-                                    const int subpel_x_qn, const int x_step_qn,
-                                    const int subpel_y_qn, const int y_step_qn,
-                                    ConvolveParams *conv_params, int bd) {
-  int x, y, k;
-  uint16_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
-  int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
-             filter_params_y->taps;
+void av1_highbd_convolve_2d_sr_c(const uint16_t *src, int src_stride,
+                                 uint16_t *dst, int dst_stride, int w, int h,
+                                 InterpFilterParams *filter_params_x,
+                                 InterpFilterParams *filter_params_y,
+                                 const int subpel_x_q4, const int subpel_y_q4,
+                                 ConvolveParams *conv_params, int bd) {
+  int16_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
+  int im_h = h + filter_params_y->taps - 1;
   int im_stride = w;
   const int fo_vert = filter_params_y->taps / 2 - 1;
   const int fo_horiz = filter_params_x->taps / 2 - 1;
-  (void)bd;
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
+  assert(bits >= 0);
 
   // horizontal filter
   const uint16_t *src_horiz = src - fo_vert * src_stride;
-  for (y = 0; y < im_h; ++y) {
-    int x_qn = subpel_x_qn;
-    for (x = 0; x < w; ++x, x_qn += x_step_qn) {
-      const uint16_t *const src_x = &src_horiz[(x_qn >> SCALE_SUBPEL_BITS)];
-      const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
-      assert(x_filter_idx < SUBPEL_SHIFTS);
-      const int16_t *x_filter =
-          av1_get_interp_filter_subpel_kernel(*filter_params_x, x_filter_idx);
-      int sum = 0;
-      for (k = 0; k < filter_params_x->taps; ++k)
-        sum += x_filter[k] * src_x[k - fo_horiz];
+  const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
+      *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
+  for (int y = 0; y < im_h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t sum = (1 << (bd + FILTER_BITS - 1));
+      for (int k = 0; k < filter_params_x->taps; ++k) {
+        sum += x_filter[k] * src_horiz[y * src_stride + x - fo_horiz + k];
+      }
+      assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
       im_block[y * im_stride + x] =
-          clip_pixel(ROUND_POWER_OF_TWO(sum, conv_params->round_0));
+          ROUND_POWER_OF_TWO(sum, conv_params->round_0);
     }
-    src_horiz += src_stride;
   }
 
   // vertical filter
-  uint16_t *src_vert = im_block + fo_vert * im_stride;
-  for (x = 0; x < w; ++x) {
-    int y_qn = subpel_y_qn;
-    for (y = 0; y < h; ++y, y_qn += y_step_qn) {
-      const uint16_t *const src_y =
-          &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
-      const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
-      assert(y_filter_idx < SUBPEL_SHIFTS);
-      const int16_t *y_filter =
-          av1_get_interp_filter_subpel_kernel(*filter_params_y, y_filter_idx);
-      CONV_BUF_TYPE sum = 0;
-      for (k = 0; k < filter_params_y->taps; ++k) {
-        sum += y_filter[k] * src_y[(k - fo_vert) * im_stride];
+  int16_t *src_vert = im_block + fo_vert * im_stride;
+  const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+      *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t sum = 1 << offset_bits;
+      for (int k = 0; k < filter_params_y->taps; ++k) {
+        sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
       }
-      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
-      if (conv_params->do_average)
-        dst[y * dst_stride + x] += res;
-      else
-        dst[y * dst_stride + x] = res;
+      assert(0 <= sum && sum < (1 << (offset_bits + 2)));
+      int32_t res = ROUND_POWER_OF_TWO(sum, conv_params->round_1) -
+                    ((1 << (offset_bits - conv_params->round_1)) +
+                     (1 << (offset_bits - conv_params->round_1 - 1)));
+      dst[y * dst_stride + x] =
+          clip_pixel_highbd(ROUND_POWER_OF_TWO(res, bits), bd);
     }
-    src_vert++;
   }
 }
 
-#else
-
-void av1_highbd_convolve_2d_c(const uint16_t *src, int src_stride,
-                              CONV_BUF_TYPE *dst, int dst_stride, int w, int h,
-                              InterpFilterParams *filter_params_x,
-                              InterpFilterParams *filter_params_y,
-                              const int subpel_x_q4, const int subpel_y_q4,
-                              ConvolveParams *conv_params, int bd) {
+void av1_highbd_jnt_convolve_2d_c(const uint16_t *src, int src_stride,
+                                  uint16_t *dst16, int dst16_stride, int w,
+                                  int h, InterpFilterParams *filter_params_x,
+                                  InterpFilterParams *filter_params_y,
+                                  const int subpel_x_q4, const int subpel_y_q4,
+                                  ConvolveParams *conv_params, int bd) {
   int x, y, k;
-  int32_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
+  int16_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
   int im_h = h + filter_params_y->taps - 1;
   int im_stride = w;
   const int fo_vert = filter_params_y->taps / 2 - 1;
   const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  assert(round_bits >= 0);
 
   // horizontal filter
   const uint16_t *src_horiz = src - fo_vert * src_stride;
@@ -760,439 +695,367 @@ void av1_highbd_convolve_2d_c(const uint16_t *src, int src_stride,
       assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
       (void)bd;
       im_block[y * im_stride + x] =
-          ROUND_POWER_OF_TWO(sum, conv_params->round_0);
+          (int16_t)ROUND_POWER_OF_TWO(sum, conv_params->round_0);
     }
   }
 
   // vertical filter
-  int32_t *src_vert = im_block + fo_vert * im_stride;
+  int16_t *src_vert = im_block + fo_vert * im_stride;
   const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
   const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
       *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
   for (y = 0; y < h; ++y) {
     for (x = 0; x < w; ++x) {
-      CONV_BUF_TYPE sum = 1 << offset_bits;
+      int32_t sum = 1 << offset_bits;
       for (k = 0; k < filter_params_y->taps; ++k) {
         sum += y_filter[k] * src_vert[(y - fo_vert + k) * im_stride + x];
       }
       assert(0 <= sum && sum < (1 << (offset_bits + 2)));
-      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1) -
-                          ((1 << (offset_bits - conv_params->round_1)) +
-                           (1 << (offset_bits - conv_params->round_1 - 1)));
-      if (conv_params->do_average)
-        dst[y * dst_stride + x] += res;
-      else
+      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
+      if (conv_params->do_average) {
+        int32_t tmp = dst[y * dst_stride + x];
+        if (conv_params->use_jnt_comp_avg) {
+          tmp = tmp * conv_params->fwd_offset + res * conv_params->bck_offset;
+          tmp = tmp >> DIST_PRECISION_BITS;
+        } else {
+          tmp += res;
+          tmp = tmp >> 1;
+        }
+        tmp -= (1 << (offset_bits - conv_params->round_1)) +
+               (1 << (offset_bits - conv_params->round_1 - 1));
+        dst16[y * dst16_stride + x] =
+            clip_pixel_highbd(ROUND_POWER_OF_TWO(tmp, round_bits), bd);
+      } else {
+        dst[y * dst_stride + x] = res;
+      }
+    }
+  }
+}
+
+void av1_highbd_jnt_convolve_x_c(const uint16_t *src, int src_stride,
+                                 uint16_t *dst16, int dst16_stride, int w,
+                                 int h, InterpFilterParams *filter_params_x,
+                                 InterpFilterParams *filter_params_y,
+                                 const int subpel_x_q4, const int subpel_y_q4,
+                                 ConvolveParams *conv_params, int bd) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  const int fo_horiz = filter_params_x->taps / 2 - 1;
+  const int bits = FILTER_BITS - conv_params->round_1;
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  const int round_offset = (1 << (offset_bits - conv_params->round_1)) +
+                           (1 << (offset_bits - conv_params->round_1 - 1));
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  assert(round_bits >= 0);
+  (void)filter_params_y;
+  (void)subpel_y_q4;
+  assert(bits >= 0);
+  // horizontal filter
+  const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
+      *filter_params_x, subpel_x_q4 & SUBPEL_MASK);
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t res = 0;
+      for (int k = 0; k < filter_params_x->taps; ++k) {
+        res += x_filter[k] * src[y * src_stride + x - fo_horiz + k];
+      }
+      res = (1 << bits) * ROUND_POWER_OF_TWO(res, conv_params->round_0);
+      res += round_offset;
+
+      if (conv_params->do_average) {
+        int32_t tmp = dst[y * dst_stride + x];
+        if (conv_params->use_jnt_comp_avg) {
+          tmp = tmp * conv_params->fwd_offset + res * conv_params->bck_offset;
+          tmp = tmp >> DIST_PRECISION_BITS;
+        } else {
+          tmp += res;
+          tmp = tmp >> 1;
+        }
+        tmp -= round_offset;
+        dst16[y * dst16_stride + x] =
+            clip_pixel_highbd(ROUND_POWER_OF_TWO(tmp, round_bits), bd);
+      } else {
+        dst[y * dst_stride + x] = res;
+      }
+    }
+  }
+}
+
+void av1_highbd_jnt_convolve_y_c(const uint16_t *src, int src_stride,
+                                 uint16_t *dst16, int dst16_stride, int w,
+                                 int h, InterpFilterParams *filter_params_x,
+                                 InterpFilterParams *filter_params_y,
+                                 const int subpel_x_q4, const int subpel_y_q4,
+                                 ConvolveParams *conv_params, int bd) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  const int fo_vert = filter_params_y->taps / 2 - 1;
+  const int bits = FILTER_BITS - conv_params->round_0;
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  const int round_offset = (1 << (offset_bits - conv_params->round_1)) +
+                           (1 << (offset_bits - conv_params->round_1 - 1));
+  const int round_bits =
+      2 * FILTER_BITS - conv_params->round_0 - conv_params->round_1;
+  assert(round_bits >= 0);
+  (void)filter_params_x;
+  (void)subpel_x_q4;
+  assert(bits >= 0);
+  // vertical filter
+  const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
+      *filter_params_y, subpel_y_q4 & SUBPEL_MASK);
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      int32_t res = 0;
+      for (int k = 0; k < filter_params_y->taps; ++k) {
+        res += y_filter[k] * src[(y - fo_vert + k) * src_stride + x];
+      }
+      res *= (1 << bits);
+      res = ROUND_POWER_OF_TWO(res, conv_params->round_1) + round_offset;
+
+      if (conv_params->do_average) {
+        int32_t tmp = dst[y * dst_stride + x];
+        if (conv_params->use_jnt_comp_avg) {
+          tmp = tmp * conv_params->fwd_offset + res * conv_params->bck_offset;
+          tmp = tmp >> DIST_PRECISION_BITS;
+        } else {
+          tmp += res;
+          tmp = tmp >> 1;
+        }
+        tmp -= round_offset;
+        dst16[y * dst16_stride + x] =
+            clip_pixel_highbd(ROUND_POWER_OF_TWO(tmp, round_bits), bd);
+      } else {
         dst[y * dst_stride + x] = res;
+      }
+    }
+  }
+}
+
+void av1_highbd_jnt_convolve_2d_copy_c(
+    const uint16_t *src, int src_stride, uint16_t *dst16, int dst16_stride,
+    int w, int h, InterpFilterParams *filter_params_x,
+    InterpFilterParams *filter_params_y, const int subpel_x_q4,
+    const int subpel_y_q4, ConvolveParams *conv_params, int bd) {
+  CONV_BUF_TYPE *dst = conv_params->dst;
+  int dst_stride = conv_params->dst_stride;
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_1 - conv_params->round_0;
+  const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
+  const int round_offset = (1 << (offset_bits - conv_params->round_1)) +
+                           (1 << (offset_bits - conv_params->round_1 - 1));
+  assert(bits >= 0);
+  (void)filter_params_x;
+  (void)filter_params_y;
+  (void)subpel_x_q4;
+  (void)subpel_y_q4;
+
+  for (int y = 0; y < h; ++y) {
+    for (int x = 0; x < w; ++x) {
+      CONV_BUF_TYPE res = src[y * src_stride + x] << bits;
+      res += round_offset;
+      if (conv_params->do_average) {
+        int32_t tmp = dst[y * dst_stride + x];
+        if (conv_params->use_jnt_comp_avg) {
+          tmp = tmp * conv_params->fwd_offset + res * conv_params->bck_offset;
+          tmp = tmp >> DIST_PRECISION_BITS;
+        } else {
+          tmp += res;
+          tmp = tmp >> 1;
+        }
+        tmp -= round_offset;
+        dst16[y * dst16_stride + x] =
+            clip_pixel_highbd(ROUND_POWER_OF_TWO(tmp, bits), bd);
+      } else {
+        dst[y * dst_stride + x] = res;
+      }
     }
   }
 }
 
 void av1_highbd_convolve_2d_scale_c(const uint16_t *src, int src_stride,
-                                    CONV_BUF_TYPE *dst, int dst_stride, int w,
-                                    int h, InterpFilterParams *filter_params_x,
+                                    uint16_t *dst, int dst_stride, int w, int h,
+                                    InterpFilterParams *filter_params_x,
                                     InterpFilterParams *filter_params_y,
                                     const int subpel_x_qn, const int x_step_qn,
                                     const int subpel_y_qn, const int y_step_qn,
                                     ConvolveParams *conv_params, int bd) {
-  int x, y, k;
-  int32_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
+  int16_t im_block[(2 * MAX_SB_SIZE + MAX_FILTER_TAP) * MAX_SB_SIZE];
   int im_h = (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
              filter_params_y->taps;
   int im_stride = w;
   const int fo_vert = filter_params_y->taps / 2 - 1;
   const int fo_horiz = filter_params_x->taps / 2 - 1;
-
+  CONV_BUF_TYPE *dst16 = conv_params->dst;
+  const int dst16_stride = conv_params->dst_stride;
+  const int bits =
+      FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
+  assert(bits >= 0);
   // horizontal filter
   const uint16_t *src_horiz = src - fo_vert * src_stride;
-  for (y = 0; y < im_h; ++y) {
+  for (int y = 0; y < im_h; ++y) {
     int x_qn = subpel_x_qn;
-    for (x = 0; x < w; ++x, x_qn += x_step_qn) {
+    for (int x = 0; x < w; ++x, x_qn += x_step_qn) {
       const uint16_t *const src_x = &src_horiz[(x_qn >> SCALE_SUBPEL_BITS)];
       const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
       assert(x_filter_idx < SUBPEL_SHIFTS);
       const int16_t *x_filter =
           av1_get_interp_filter_subpel_kernel(*filter_params_x, x_filter_idx);
       int32_t sum = (1 << (bd + FILTER_BITS - 1));
-      for (k = 0; k < filter_params_x->taps; ++k) {
+      for (int k = 0; k < filter_params_x->taps; ++k) {
         sum += x_filter[k] * src_x[k - fo_horiz];
       }
       assert(0 <= sum && sum < (1 << (bd + FILTER_BITS + 1)));
       im_block[y * im_stride + x] =
-          ROUND_POWER_OF_TWO(sum, conv_params->round_0);
+          (int16_t)ROUND_POWER_OF_TWO(sum, conv_params->round_0);
     }
     src_horiz += src_stride;
   }
 
   // vertical filter
-  int32_t *src_vert = im_block + fo_vert * im_stride;
+  int16_t *src_vert = im_block + fo_vert * im_stride;
   const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
-  for (x = 0; x < w; ++x) {
+  for (int x = 0; x < w; ++x) {
     int y_qn = subpel_y_qn;
-    for (y = 0; y < h; ++y, y_qn += y_step_qn) {
-      const int32_t *src_y = &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
+    for (int y = 0; y < h; ++y, y_qn += y_step_qn) {
+      const int16_t *src_y = &src_vert[(y_qn >> SCALE_SUBPEL_BITS) * im_stride];
       const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
       assert(y_filter_idx < SUBPEL_SHIFTS);
       const int16_t *y_filter =
           av1_get_interp_filter_subpel_kernel(*filter_params_y, y_filter_idx);
-      CONV_BUF_TYPE sum = 1 << offset_bits;
-      for (k = 0; k < filter_params_y->taps; ++k) {
+      int32_t sum = 1 << offset_bits;
+      for (int k = 0; k < filter_params_y->taps; ++k) {
         sum += y_filter[k] * src_y[(k - fo_vert) * im_stride];
       }
       assert(0 <= sum && sum < (1 << (offset_bits + 2)));
-      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1) -
-                          ((1 << (offset_bits - conv_params->round_1)) +
-                           (1 << (offset_bits - conv_params->round_1 - 1)));
-      if (conv_params->do_average)
-        dst[y * dst_stride + x] += res;
-      else
-        dst[y * dst_stride + x] = res;
+      CONV_BUF_TYPE res = ROUND_POWER_OF_TWO(sum, conv_params->round_1);
+      if (conv_params->is_compound) {
+        if (conv_params->do_average) {
+          int32_t tmp = dst16[y * dst16_stride + x];
+          if (conv_params->use_jnt_comp_avg) {
+            tmp = tmp * conv_params->fwd_offset + res * conv_params->bck_offset;
+            tmp = tmp >> DIST_PRECISION_BITS;
+          } else {
+            tmp += res;
+            tmp = tmp >> 1;
+          }
+          /* Subtract round offset and convolve round */
+          tmp = tmp - ((1 << (offset_bits - conv_params->round_1)) +
+                       (1 << (offset_bits - conv_params->round_1 - 1)));
+          dst[y * dst_stride + x] =
+              clip_pixel_highbd(ROUND_POWER_OF_TWO(tmp, bits), bd);
+        } else {
+          dst16[y * dst16_stride + x] = res;
+        }
+      } else {
+        /* Subtract round offset and convolve round */
+        int32_t tmp = res - ((1 << (offset_bits - conv_params->round_1)) +
+                             (1 << (offset_bits - conv_params->round_1 - 1)));
+        dst[y * dst_stride + x] =
+            clip_pixel_highbd(ROUND_POWER_OF_TWO(tmp, bits), bd);
+      }
     }
     src_vert++;
   }
 }
-#endif  // CONFIG_COMPOUND_ROUND
 
 void av1_highbd_convolve_2d_facade(const uint8_t *src8, int src_stride,
-                                   uint8_t *dst, int dst_stride, int w, int h,
+                                   uint8_t *dst8, int dst_stride, int w, int h,
                                    InterpFilters interp_filters,
                                    const int subpel_x_q4, int x_step_q4,
                                    const int subpel_y_q4, int y_step_q4,
                                    int scaled, ConvolveParams *conv_params,
-                                   int bd) {
+                                   const struct scale_factors *sf, int bd) {
   (void)x_step_q4;
   (void)y_step_q4;
-  (void)dst;
   (void)dst_stride;
 
-  InterpFilterParams filter_params_x, filter_params_y;
-  av1_get_convolve_filter_params(interp_filters, 1, &filter_params_x,
-                                 &filter_params_y);
-
   const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
-  if (filter_params_y.taps < filter_params_x.taps) {
-    uint16_t tr_src[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) *
-                    (MAX_SB_SIZE + MAX_FILTER_TAP - 1)];
-    int tr_src_stride = MAX_SB_SIZE + MAX_FILTER_TAP - 1;
-    CONV_BUF_TYPE tr_dst[MAX_SB_SIZE * MAX_SB_SIZE];
-    int tr_dst_stride = MAX_SB_SIZE;
-    int fo_vert = filter_params_y.taps / 2 - 1;
-    int fo_horiz = filter_params_x.taps / 2 - 1;
-
-    transpose_uint16(
-        tr_src, tr_src_stride, src - fo_vert * src_stride - fo_horiz,
-        src_stride, w + filter_params_x.taps - 1, h + filter_params_y.taps - 1);
-    transpose_int32(tr_dst, tr_dst_stride, conv_params->dst,
-                    conv_params->dst_stride, w, h);
-
-    // horizontal and vertical parameters are swapped because of the transpose
-    if (scaled)
-      av1_highbd_convolve_2d_scale(
-          tr_src + fo_horiz * tr_src_stride + fo_vert, tr_src_stride, tr_dst,
-          tr_dst_stride, h, w, &filter_params_y, &filter_params_x, subpel_y_q4,
-          y_step_q4, subpel_x_q4, x_step_q4, conv_params, bd);
-    else
-      av1_highbd_convolve_2d(tr_src + fo_horiz * tr_src_stride + fo_vert,
-                             tr_src_stride, tr_dst, tr_dst_stride, h, w,
-                             &filter_params_y, &filter_params_x, subpel_y_q4,
-                             subpel_x_q4, conv_params, bd);
-    transpose_int32(conv_params->dst, conv_params->dst_stride, tr_dst,
-                    tr_dst_stride, h, w);
-  } else {
-    if (scaled)
-      av1_highbd_convolve_2d_scale(
-          src, src_stride, conv_params->dst, conv_params->dst_stride, w, h,
-          &filter_params_x, &filter_params_y, subpel_x_q4, x_step_q4,
-          subpel_y_q4, y_step_q4, conv_params, bd);
-    else
-      av1_highbd_convolve_2d(src, src_stride, conv_params->dst,
-                             conv_params->dst_stride, w, h, &filter_params_x,
-                             &filter_params_y, subpel_x_q4, subpel_y_q4,
-                             conv_params, bd);
-  }
-}
-#endif  // CONFIG_HIGHBITDEPTH
-
-#endif  // CONFIG_CONVOLVE_ROUND
-
-typedef void (*ConvolveFunc)(const uint8_t *src, int src_stride, uint8_t *dst,
-                             int dst_stride, int w, int h,
-                             const InterpFilterParams filter_params,
-                             const int subpel_q4, int step_q4,
-                             ConvolveParams *conv_params);
-
-static void convolve_helper(const uint8_t *src, int src_stride, uint8_t *dst,
-                            int dst_stride, int w, int h,
-                            const InterpFilters interp_filters,
-                            const int subpel_x_q4, int x_step_q4,
-                            const int subpel_y_q4, int y_step_q4,
-                            ConvolveParams *conv_params,
-                            ConvolveFunc convolve_horiz,
-                            ConvolveFunc convolve_vert) {
-  int ignore_horiz = x_step_q4 == SUBPEL_SHIFTS && subpel_x_q4 == 0;
-  int ignore_vert = y_step_q4 == SUBPEL_SHIFTS && subpel_y_q4 == 0;
-
   InterpFilterParams filter_params_x, filter_params_y;
-  av1_get_convolve_filter_params(interp_filters, 0, &filter_params_x,
-                                 &filter_params_y);
-
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-
-  assert(w <= MAX_BLOCK_WIDTH);
-  assert(h <= MAX_BLOCK_HEIGHT);
-  assert(y_step_q4 <= MAX_STEP);
-  assert(x_step_q4 <= MAX_STEP);
-
-  if (ignore_horiz && ignore_vert) {
-    convolve_copy(src, src_stride, dst, dst_stride, w, h, conv_params);
-  } else if (ignore_vert) {
-    assert(filter_params_x.taps <= MAX_FILTER_TAP);
-    convolve_horiz(src, src_stride, dst, dst_stride, w, h, filter_params_x,
-                   subpel_x_q4, x_step_q4, conv_params);
-  } else if (ignore_horiz) {
-    assert(filter_params_y.taps <= MAX_FILTER_TAP);
-    convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params_y,
-                  subpel_y_q4, y_step_q4, conv_params);
-  } else {
-    // temp's size is set to a 256 aligned value to facilitate SIMD
-    // implementation. The value is greater than (maximum possible intermediate
-    // height or width) * MAX_SB_SIZE
-    DECLARE_ALIGNED(16, uint8_t,
-                    temp[((MAX_SB_SIZE * 2 + 16) + 16) * MAX_SB_SIZE]);
-    int max_intermediate_size = ((MAX_SB_SIZE * 2 + 16) + 16);
-    int filter_size;
-#if CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
-    av1_convolve_filter_params_fixup_1212(&filter_params_x, &filter_params_y);
-
-    // we do filter with fewer taps first to reduce hardware implementation
-    // complexity
-    if (filter_params_y.taps < filter_params_x.taps) {
-      int intermediate_width;
-      int temp_stride = max_intermediate_size;
-      ConvolveParams temp_conv_params;
-      temp_conv_params.ref = 0;
-      temp_conv_params.do_average = 0;
-      temp_conv_params.round = CONVOLVE_OPT_ROUND;
-      filter_size = filter_params_x.taps;
-      intermediate_width =
-          (((w - 1) * x_step_q4 + subpel_x_q4) >> SUBPEL_BITS) + filter_size;
-      assert(intermediate_width <= max_intermediate_size);
-
-      assert(filter_params_y.taps <= MAX_FILTER_TAP);
-
-      convolve_vert(src - (filter_size / 2 - 1), src_stride, temp, temp_stride,
-                    intermediate_width, h, filter_params_y, subpel_y_q4,
-                    y_step_q4, &temp_conv_params);
-
-      assert(filter_params_x.taps <= MAX_FILTER_TAP);
-      convolve_horiz(temp + (filter_size / 2 - 1), temp_stride, dst, dst_stride,
-                     w, h, filter_params_x, subpel_x_q4, x_step_q4,
-                     conv_params);
-    } else
-#endif  // CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
-    {
-      int intermediate_height;
-      int temp_stride = MAX_SB_SIZE;
-      ConvolveParams temp_conv_params;
-      temp_conv_params.ref = 0;
-      temp_conv_params.do_average = 0;
-      temp_conv_params.round = CONVOLVE_OPT_ROUND;
-      filter_size = filter_params_y.taps;
-      intermediate_height =
-          (((h - 1) * y_step_q4 + subpel_y_q4) >> SUBPEL_BITS) + filter_size;
-      assert(intermediate_height <= max_intermediate_size);
-      (void)max_intermediate_size;
-
-      assert(filter_params_x.taps <= MAX_FILTER_TAP);
-
-      convolve_horiz(src - src_stride * (filter_size / 2 - 1), src_stride, temp,
-                     temp_stride, w, intermediate_height, filter_params_x,
-                     subpel_x_q4, x_step_q4, &temp_conv_params);
-
-      assert(filter_params_y.taps <= MAX_FILTER_TAP);
-
-      convolve_vert(temp + temp_stride * (filter_size / 2 - 1), temp_stride,
-                    dst, dst_stride, w, h, filter_params_y, subpel_y_q4,
-                    y_step_q4, conv_params);
-    }
-  }
-}
+  av1_get_convolve_filter_params(interp_filters, &filter_params_x,
+                                 &filter_params_y, w, h);
 
-static void convolve_scale_helper(const uint8_t *src, int src_stride,
-                                  uint8_t *dst, int dst_stride, int w, int h,
-                                  const InterpFilters interp_filters,
-                                  const int subpel_x_qn, int x_step_qn,
-                                  const int subpel_y_qn, int y_step_qn,
-                                  ConvolveParams *conv_params,
-                                  ConvolveFunc convolve_horiz,
-                                  ConvolveFunc convolve_vert) {
-  int ignore_horiz = x_step_qn == SCALE_SUBPEL_SHIFTS && subpel_x_qn == 0;
-  int ignore_vert = y_step_qn == SCALE_SUBPEL_SHIFTS && subpel_y_qn == 0;
-
-  InterpFilterParams filter_params_x, filter_params_y;
-  av1_get_convolve_filter_params(interp_filters, 0, &filter_params_x,
-                                 &filter_params_y);
-
-  assert(conv_params->round == CONVOLVE_OPT_ROUND);
-
-  assert(w <= MAX_BLOCK_WIDTH);
-  assert(h <= MAX_BLOCK_HEIGHT);
-  assert(y_step_qn <= (MAX_STEP << SCALE_EXTRA_BITS));
-  assert(x_step_qn <= (MAX_STEP << SCALE_EXTRA_BITS));
-
-  if (ignore_horiz && ignore_vert) {
-    convolve_copy(src, src_stride, dst, dst_stride, w, h, conv_params);
-  } else if (ignore_vert) {
-    assert(filter_params_x.taps <= MAX_FILTER_TAP);
-    convolve_horiz(src, src_stride, dst, dst_stride, w, h, filter_params_x,
-                   subpel_x_qn, x_step_qn, conv_params);
-  } else if (ignore_horiz) {
-    assert(filter_params_y.taps <= MAX_FILTER_TAP);
-    convolve_vert(src, src_stride, dst, dst_stride, w, h, filter_params_y,
-                  subpel_y_qn, y_step_qn, conv_params);
-  } else {
-    // temp's size is set to a 256 aligned value to facilitate SIMD
-    // implementation. The value is greater than (maximum possible intermediate
-    // height or width) * MAX_SB_SIZE
-    DECLARE_ALIGNED(16, uint8_t,
-                    temp[((MAX_SB_SIZE * 2 + 16) + 16) * MAX_SB_SIZE]);
-    int max_intermediate_size = ((MAX_SB_SIZE * 2 + 16) + 16);
-    int filter_size;
-#if CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
-    av1_convolve_filter_params_fixup_1212(&filter_params_x, &filter_params_y);
-
-    // we do filter with fewer taps first to reduce hardware implementation
-    // complexity
-    if (filter_params_y.taps < filter_params_x.taps) {
-      int intermediate_width;
-      int temp_stride = max_intermediate_size;
-      ConvolveParams temp_conv_params;
-      temp_conv_params.ref = 0;
-      temp_conv_params.do_average = 0;
-      temp_conv_params.round = CONVOLVE_OPT_ROUND;
-      filter_size = filter_params_x.taps;
-      intermediate_width =
-          (((w - 1) * x_step_qn + subpel_x_qn) >> SCALE_SUBPEL_BITS) +
-          filter_size;
-      assert(intermediate_width <= max_intermediate_size);
-
-      assert(filter_params_y.taps <= MAX_FILTER_TAP);
-
-      convolve_vert(src - (filter_size / 2 - 1), src_stride, temp, temp_stride,
-                    intermediate_width, h, filter_params_y, subpel_y_qn,
-                    y_step_qn, &temp_conv_params);
-
-      assert(filter_params_x.taps <= MAX_FILTER_TAP);
-      convolve_horiz(temp + (filter_size / 2 - 1), temp_stride, dst, dst_stride,
-                     w, h, filter_params_x, subpel_x_qn, x_step_qn,
-                     conv_params);
-    } else {
-#endif  // CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
-      int intermediate_height;
-      int temp_stride = MAX_SB_SIZE;
-      ConvolveParams temp_conv_params;
-      temp_conv_params.ref = 0;
-      temp_conv_params.do_average = 0;
-      temp_conv_params.round = CONVOLVE_OPT_ROUND;
-      filter_size = filter_params_y.taps;
-      intermediate_height =
-          (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
-          filter_size;
-      assert(intermediate_height <= max_intermediate_size);
-      (void)max_intermediate_size;
-
-      assert(filter_params_x.taps <= MAX_FILTER_TAP);
-
-      convolve_horiz(src - src_stride * (filter_size / 2 - 1), src_stride, temp,
-                     temp_stride, w, intermediate_height, filter_params_x,
-                     subpel_x_qn, x_step_qn, &temp_conv_params);
-
-      assert(filter_params_y.taps <= MAX_FILTER_TAP);
-
-      convolve_vert(temp + temp_stride * (filter_size / 2 - 1), temp_stride,
-                    dst, dst_stride, w, h, filter_params_y, subpel_y_qn,
-                    y_step_qn, conv_params);
-#if CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
+  if (scaled) {
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
+    if (conv_params->is_compound) {
+      assert(conv_params->dst != NULL);
     }
-#endif  // CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
-  }
-}
+    av1_highbd_convolve_2d_scale(src, src_stride, dst, dst_stride, w, h,
+                                 &filter_params_x, &filter_params_y,
+                                 subpel_x_q4, x_step_q4, subpel_y_q4, y_step_q4,
+                                 conv_params, bd);
+  } else {
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
 
-void av1_convolve(const uint8_t *src, int src_stride, uint8_t *dst,
-                  int dst_stride, int w, int h, InterpFilters interp_filters,
-                  const int subpel_x_q4, int x_step_q4, const int subpel_y_q4,
-                  int y_step_q4, ConvolveParams *conv_params) {
-  convolve_helper(src, src_stride, dst, dst_stride, w, h, interp_filters,
-                  subpel_x_q4, x_step_q4, subpel_y_q4, y_step_q4, conv_params,
-                  av1_convolve_horiz_facade, av1_convolve_vert_facade);
+    sf->highbd_convolve[subpel_x_q4 != 0][subpel_y_q4 !=
+                                          0][conv_params->is_compound](
+        src, src_stride, dst, dst_stride, w, h, &filter_params_x,
+        &filter_params_y, subpel_x_q4, subpel_y_q4, conv_params, bd);
+  }
 }
 
-void av1_convolve_c(const uint8_t *src, int src_stride, uint8_t *dst,
-                    int dst_stride, int w, int h, InterpFilters interp_filters,
-                    const int subpel_x_q4, int x_step_q4, const int subpel_y_q4,
-                    int y_step_q4, ConvolveParams *conv_params) {
-  convolve_helper(src, src_stride, dst, dst_stride, w, h, interp_filters,
-                  subpel_x_q4, x_step_q4, subpel_y_q4, y_step_q4, conv_params,
-                  av1_convolve_horiz_facade_c, av1_convolve_vert_facade_c);
+// Note: Fixed size intermediate buffers, place limits on parameters
+// of some functions. 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 128x128 pixels.
+// --128 rows in the downscaled frame span a distance of (128 - 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.
+// --((128 - 1) * 32 + 15) >> 4 + 8 = 263.
+#define WIENER_MAX_EXT_SIZE 263
+
+static INLINE int horz_scalar_product(const uint8_t *a, const int16_t *b) {
+  int sum = 0;
+  for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k] * b[k];
+  return sum;
 }
 
-void av1_convolve_scale(const uint8_t *src, int src_stride, uint8_t *dst,
-                        int dst_stride, int w, int h,
-                        InterpFilters interp_filters, const int subpel_x_qn,
-                        int x_step_qn, const int subpel_y_qn, int y_step_qn,
-                        ConvolveParams *conv_params) {
-  convolve_scale_helper(src, src_stride, dst, dst_stride, w, h, interp_filters,
-                        subpel_x_qn, x_step_qn, subpel_y_qn, y_step_qn,
-                        conv_params, av1_convolve_horiz_facade_scale,
-                        av1_convolve_vert_facade_scale);
+static INLINE int highbd_horz_scalar_product(const uint16_t *a,
+                                             const int16_t *b) {
+  int sum = 0;
+  for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k] * b[k];
+  return sum;
 }
 
-void av1_lowbd_convolve_init_c(void) {
-  // A placeholder for SIMD initialization
-  return;
+static INLINE int highbd_vert_scalar_product(const uint16_t *a,
+                                             ptrdiff_t a_stride,
+                                             const int16_t *b) {
+  int sum = 0;
+  for (int k = 0; k < SUBPEL_TAPS; ++k) sum += a[k * a_stride] * b[k];
+  return sum;
 }
 
-void av1_highbd_convolve_init_c(void) {
-  // A placeholder for SIMD initialization
-  return;
+static const InterpKernel *get_filter_base(const int16_t *filter) {
+  // NOTE: This assumes that the filter table is 256-byte aligned.
+  // TODO(agrange) Modify to make independent of table alignment.
+  return (const InterpKernel *)(((intptr_t)filter) & ~((intptr_t)0xFF));
 }
 
-void av1_convolve_init(AV1_COMMON *cm) {
-#if CONFIG_HIGHBITDEPTH
-  if (cm->use_highbitdepth)
-    av1_highbd_convolve_init();
-  else
-    av1_lowbd_convolve_init();
-#else
-  (void)cm;
-  av1_lowbd_convolve_init();
-#endif
-  return;
+static int get_filter_offset(const int16_t *f, const InterpKernel *base) {
+  return (int)((const InterpKernel *)(intptr_t)f - base);
 }
 
-#if CONFIG_HIGHBITDEPTH
-void av1_highbd_convolve_horiz_c(const uint16_t *src, int src_stride,
-                                 uint16_t *dst, int dst_stride, int w, int h,
-                                 const InterpFilterParams filter_params,
-                                 const int subpel_x_q4, int x_step_q4, int avg,
-                                 int bd) {
-  int x, y;
-  int filter_size = filter_params.taps;
-  src -= filter_size / 2 - 1;
-  for (y = 0; y < h; ++y) {
-    int x_q4 = subpel_x_q4;
-    for (x = 0; x < w; ++x) {
-      const uint16_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
-      const int16_t *x_filter = av1_get_interp_filter_subpel_kernel(
-          filter_params, x_q4 & SUBPEL_MASK);
-      int k, sum = 0;
-      for (k = 0; k < filter_size; ++k) sum += src_x[k] * x_filter[k];
-      if (avg)
-        dst[x] = ROUND_POWER_OF_TWO(
-            dst[x] +
-                clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
-            1);
-      else
-        dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
+static void convolve_add_src_horiz_hip(const uint8_t *src, ptrdiff_t src_stride,
+                                       uint16_t *dst, ptrdiff_t dst_stride,
+                                       const InterpKernel *x_filters, int x0_q4,
+                                       int x_step_q4, int w, int h,
+                                       int round0_bits) {
+  const int bd = 8;
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (int y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (int x = 0; x < w; ++x) {
+      const uint8_t *const src_x = &src[x_q4 >> SUBPEL_BITS];
+      const int16_t *const x_filter = x_filters[x_q4 & SUBPEL_MASK];
+      const int rounding = ((int)src_x[SUBPEL_TAPS / 2 - 1] << FILTER_BITS) +
+                           (1 << (bd + FILTER_BITS - 1));
+      const int sum = horz_scalar_product(src_x, x_filter) + rounding;
+      dst[x] = (uint16_t)clamp(ROUND_POWER_OF_TWO(sum, round0_bits), 0,
+                               WIENER_CLAMP_LIMIT(round0_bits, bd) - 1);
       x_q4 += x_step_q4;
     }
     src += src_stride;
@@ -1200,66 +1063,25 @@ void av1_highbd_convolve_horiz_c(const uint16_t *src, int src_stride,
   }
 }
 
-void av1_highbd_convolve_horiz_scale(const uint16_t *src, int src_stride,
-                                     uint16_t *dst, int dst_stride, int w,
-                                     int h,
-                                     const InterpFilterParams filter_params,
-                                     const int subpel_x_qn, int x_step_qn,
-                                     int avg, int bd) {
-  int x, y;
-  int filter_size = filter_params.taps;
-  src -= filter_size / 2 - 1;
-  for (y = 0; y < h; ++y) {
-    int x_qn = subpel_x_qn;
-    for (x = 0; x < w; ++x) {
-      const uint16_t *const src_x = &src[x_qn >> SCALE_SUBPEL_BITS];
-      const int x_filter_idx = (x_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
-      assert(x_filter_idx < SUBPEL_SHIFTS);
-      const int16_t *x_filter =
-          av1_get_interp_filter_subpel_kernel(filter_params, x_filter_idx);
-      int k, sum = 0;
-      for (k = 0; k < filter_size; ++k) sum += src_x[k] * x_filter[k];
-      if (avg)
-        dst[x] = ROUND_POWER_OF_TWO(
-            dst[x] +
-                clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
-            1);
-      else
-        dst[x] = clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
-      x_qn += x_step_qn;
-    }
-    src += src_stride;
-    dst += dst_stride;
-  }
-}
-
-void av1_highbd_convolve_vert_c(const uint16_t *src, int src_stride,
-                                uint16_t *dst, int dst_stride, int w, int h,
-                                const InterpFilterParams filter_params,
-                                const int subpel_y_q4, int y_step_q4, int avg,
-                                int bd) {
-  int x, y;
-  int filter_size = filter_params.taps;
-  src -= src_stride * (filter_size / 2 - 1);
-
-  for (x = 0; x < w; ++x) {
-    int y_q4 = subpel_y_q4;
-    for (y = 0; y < h; ++y) {
-      const uint16_t *const src_y = &src[(y_q4 >> SUBPEL_BITS) * src_stride];
-      const int16_t *y_filter = av1_get_interp_filter_subpel_kernel(
-          filter_params, y_q4 & SUBPEL_MASK);
-      int k, sum = 0;
-      for (k = 0; k < filter_size; ++k)
-        sum += src_y[k * src_stride] * y_filter[k];
-      if (avg) {
-        dst[y * dst_stride] = ROUND_POWER_OF_TWO(
-            dst[y * dst_stride] +
-                clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
-            1);
-      } else {
-        dst[y * dst_stride] =
-            clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
-      }
+static void convolve_add_src_vert_hip(const uint16_t *src, ptrdiff_t src_stride,
+                                      uint8_t *dst, ptrdiff_t dst_stride,
+                                      const InterpKernel *y_filters, int y0_q4,
+                                      int y_step_q4, int w, int h,
+                                      int round1_bits) {
+  const int bd = 8;
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+
+  for (int x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (int 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];
+      const int rounding =
+          ((int)src_y[(SUBPEL_TAPS / 2 - 1) * src_stride] << FILTER_BITS) -
+          (1 << (bd + round1_bits - 1));
+      const int sum =
+          highbd_vert_scalar_product(src_y, src_stride, y_filter) + rounding;
+      dst[y * dst_stride] = clip_pixel(ROUND_POWER_OF_TWO(sum, round1_bits));
       y_q4 += y_step_q4;
     }
     ++src;
@@ -1267,325 +1089,111 @@ void av1_highbd_convolve_vert_c(const uint16_t *src, int src_stride,
   }
 }
 
-void av1_highbd_convolve_vert_scale(const uint16_t *src, int src_stride,
-                                    uint16_t *dst, int dst_stride, int w, int h,
-                                    const InterpFilterParams filter_params,
-                                    const int subpel_y_qn, int y_step_qn,
-                                    int avg, int bd) {
-  int x, y;
-  int filter_size = filter_params.taps;
-  src -= src_stride * (filter_size / 2 - 1);
-
-  for (x = 0; x < w; ++x) {
-    int y_qn = subpel_y_qn;
-    for (y = 0; y < h; ++y) {
-      const uint16_t *const src_y =
-          &src[(y_qn >> SCALE_SUBPEL_BITS) * src_stride];
-      const int y_filter_idx = (y_qn & SCALE_SUBPEL_MASK) >> SCALE_EXTRA_BITS;
-      assert(y_filter_idx < SUBPEL_SHIFTS);
-      const int16_t *y_filter =
-          av1_get_interp_filter_subpel_kernel(filter_params, y_filter_idx);
-      int k, sum = 0;
-      for (k = 0; k < filter_size; ++k)
-        sum += src_y[k * src_stride] * y_filter[k];
-      if (avg) {
-        dst[y * dst_stride] = ROUND_POWER_OF_TWO(
-            dst[y * dst_stride] +
-                clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd),
-            1);
-      } else {
-        dst[y * dst_stride] =
-            clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, FILTER_BITS), bd);
-      }
-      y_qn += y_step_qn;
-    }
-    ++src;
-    ++dst;
-  }
-}
-
-static void highbd_convolve_copy(const uint16_t *src, int src_stride,
-                                 uint16_t *dst, int dst_stride, int w, int h,
-                                 int avg, int bd) {
-  if (avg == 0) {
-    int r;
-    for (r = 0; r < h; ++r) {
-      memcpy(dst, src, w * sizeof(*src));
-      src += src_stride;
-      dst += dst_stride;
-    }
-  } else {
-    int r, c;
-    for (r = 0; r < h; ++r) {
-      for (c = 0; c < w; ++c) {
-        dst[c] = clip_pixel_highbd(ROUND_POWER_OF_TWO(dst[c] + src[c], 1), bd);
-      }
-      src += src_stride;
-      dst += dst_stride;
-    }
-  }
+void av1_wiener_convolve_add_src_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,
+                                   const ConvolveParams *conv_params) {
+  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);
+
+  uint16_t temp[WIENER_MAX_EXT_SIZE * MAX_SB_SIZE];
+  const int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+
+  convolve_add_src_horiz_hip(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                             src_stride, temp, MAX_SB_SIZE, filters_x, x0_q4,
+                             x_step_q4, w, intermediate_height,
+                             conv_params->round_0);
+  convolve_add_src_vert_hip(temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1),
+                            MAX_SB_SIZE, dst, dst_stride, filters_y, y0_q4,
+                            y_step_q4, w, h, conv_params->round_1);
 }
 
-void av1_highbd_convolve_horiz_facade(const uint8_t *src8, int src_stride,
-                                      uint8_t *dst8, int dst_stride, int w,
-                                      int h,
-                                      const InterpFilterParams filter_params,
-                                      const int subpel_x_q4, int x_step_q4,
-                                      int avg, int bd) {
+static void highbd_convolve_add_src_horiz_hip(
+    const uint8_t *src8, ptrdiff_t src_stride, uint16_t *dst,
+    ptrdiff_t dst_stride, const InterpKernel *x_filters, int x0_q4,
+    int x_step_q4, int w, int h, int round0_bits, int bd) {
+  const int extraprec_clamp_limit = WIENER_CLAMP_LIMIT(round0_bits, bd);
   uint16_t *src = CONVERT_TO_SHORTPTR(src8);
-  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
-  if (filter_params.taps == SUBPEL_TAPS) {
-    const int16_t *filter_x =
-        av1_get_interp_filter_subpel_kernel(filter_params, subpel_x_q4);
-    if (avg == 0)
-      aom_highbd_convolve8_horiz(src8, src_stride, dst8, dst_stride, filter_x,
-                                 x_step_q4, NULL, -1, w, h, bd);
-    else
-      aom_highbd_convolve8_avg_horiz(src8, src_stride, dst8, dst_stride,
-                                     filter_x, x_step_q4, NULL, -1, w, h, bd);
-  } else {
-    av1_highbd_convolve_horiz(src, src_stride, dst, dst_stride, w, h,
-                              filter_params, subpel_x_q4, x_step_q4, avg, bd);
-  }
-}
-
-void av1_highbd_convolve_horiz_facade_scale(
-    const uint8_t *src8, int src_stride, uint8_t *dst8, int dst_stride, int w,
-    int h, const InterpFilterParams filter_params, const int subpel_x_qn,
-    int x_step_qn, int avg, int bd) {
-  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
-  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
-  // TODO(debargha): Add special functions for filter_params.taps == SUBPEL_TAPS
-  // as in the function above.
-  av1_highbd_convolve_horiz_scale(src, src_stride, dst, dst_stride, w, h,
-                                  filter_params, subpel_x_qn, x_step_qn, avg,
-                                  bd);
-}
-
-void av1_highbd_convolve_vert_facade(const uint8_t *src8, int src_stride,
-                                     uint8_t *dst8, int dst_stride, int w,
-                                     int h,
-                                     const InterpFilterParams filter_params,
-                                     const int subpel_y_q4, int y_step_q4,
-                                     int avg, int bd) {
-  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
-  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
-
-  if (filter_params.taps == SUBPEL_TAPS) {
-    const int16_t *filter_y =
-        av1_get_interp_filter_subpel_kernel(filter_params, subpel_y_q4);
-    if (avg == 0) {
-      aom_highbd_convolve8_vert(src8, src_stride, dst8, dst_stride, NULL, -1,
-                                filter_y, y_step_q4, w, h, bd);
-    } else {
-      aom_highbd_convolve8_avg_vert(src8, src_stride, dst8, dst_stride, NULL,
-                                    -1, filter_y, y_step_q4, w, h, bd);
+  src -= SUBPEL_TAPS / 2 - 1;
+  for (int y = 0; y < h; ++y) {
+    int x_q4 = x0_q4;
+    for (int 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];
+      const int rounding = ((int)src_x[SUBPEL_TAPS / 2 - 1] << FILTER_BITS) +
+                           (1 << (bd + FILTER_BITS - 1));
+      const int sum = highbd_horz_scalar_product(src_x, x_filter) + rounding;
+      dst[x] = (uint16_t)clamp(ROUND_POWER_OF_TWO(sum, round0_bits), 0,
+                               extraprec_clamp_limit - 1);
+      x_q4 += x_step_q4;
     }
-  } else {
-    av1_highbd_convolve_vert(src, src_stride, dst, dst_stride, w, h,
-                             filter_params, subpel_y_q4, y_step_q4, avg, bd);
+    src += src_stride;
+    dst += dst_stride;
   }
 }
 
-void av1_highbd_convolve_vert_facade_scale(
-    const uint8_t *src8, int src_stride, uint8_t *dst8, int dst_stride, int w,
-    int h, const InterpFilterParams filter_params, const int subpel_y_qn,
-    int y_step_qn, int avg, int bd) {
-  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
-  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
-  // TODO(debargha): Add special functions for filter_params.taps == SUBPEL_TAPS
-  // as in the function above.
-  av1_highbd_convolve_vert_scale(src, src_stride, dst, dst_stride, w, h,
-                                 filter_params, subpel_y_qn, y_step_qn, avg,
-                                 bd);
-}
-
-void av1_highbd_convolve(const uint8_t *src8, int src_stride, uint8_t *dst8,
-                         int dst_stride, int w, int h,
-                         InterpFilters interp_filters, const int subpel_x_q4,
-                         int x_step_q4, const int subpel_y_q4, int y_step_q4,
-                         int ref_idx, int bd) {
-  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
+static void highbd_convolve_add_src_vert_hip(
+    const uint16_t *src, 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 round1_bits, int bd) {
   uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
-  int ignore_horiz = x_step_q4 == SUBPEL_SHIFTS && subpel_x_q4 == 0;
-  int ignore_vert = y_step_q4 == SUBPEL_SHIFTS && subpel_y_q4 == 0;
-
-  assert(w <= MAX_BLOCK_WIDTH);
-  assert(h <= MAX_BLOCK_HEIGHT);
-  assert(y_step_q4 <= MAX_STEP);
-  assert(x_step_q4 <= MAX_STEP);
-
-  if (ignore_horiz && ignore_vert) {
-    highbd_convolve_copy(src, src_stride, dst, dst_stride, w, h, ref_idx, bd);
-    return;
-  }
-
-  InterpFilterParams filter_params_x, filter_params_y;
-  av1_get_convolve_filter_params(interp_filters, 0, &filter_params_x,
-                                 &filter_params_y);
-
-  if (ignore_vert) {
-    av1_highbd_convolve_horiz_facade(src8, src_stride, dst8, dst_stride, w, h,
-                                     filter_params_x, subpel_x_q4, x_step_q4,
-                                     ref_idx, bd);
-  } else if (ignore_horiz) {
-    av1_highbd_convolve_vert_facade(src8, src_stride, dst8, dst_stride, w, h,
-                                    filter_params_y, subpel_y_q4, y_step_q4,
-                                    ref_idx, bd);
-  } else {
-    // temp's size is set to a 256 aligned value to facilitate SIMD
-    // implementation. The value is greater than (maximum possible intermediate
-    // height or width) * MAX_SB_SIZE
-    DECLARE_ALIGNED(16, uint16_t,
-                    temp[((MAX_SB_SIZE * 2 + 16) + 16) * MAX_SB_SIZE]);
-    uint8_t *temp8 = CONVERT_TO_BYTEPTR(temp);
-    int max_intermediate_size = ((MAX_SB_SIZE * 2 + 16) + 16);
-    int filter_size;
-
-#if CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
-    av1_convolve_filter_params_fixup_1212(&filter_params_x, &filter_params_y);
-
-    if (filter_params_y.taps < filter_params_x.taps) {
-      int intermediate_width;
-      int temp_stride = max_intermediate_size;
-      filter_size = filter_params_x.taps;
-      intermediate_width =
-          (((w - 1) * x_step_q4 + subpel_x_q4) >> SUBPEL_BITS) + filter_size;
-      assert(intermediate_width <= max_intermediate_size);
-
-      assert(filter_params_y.taps <= MAX_FILTER_TAP);
-
-      av1_highbd_convolve_vert_facade(src8 - (filter_size / 2 - 1), src_stride,
-                                      temp8, temp_stride, intermediate_width, h,
-                                      filter_params_y, subpel_y_q4, y_step_q4,
-                                      0, bd);
-
-      assert(filter_params_x.taps <= MAX_FILTER_TAP);
-
-      av1_highbd_convolve_horiz_facade(
-          temp8 + (filter_size / 2 - 1), temp_stride, dst8, dst_stride, w, h,
-          filter_params_x, subpel_x_q4, x_step_q4, ref_idx, bd);
-    } else
-#endif  // CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
-    {
-      int intermediate_height;
-      int temp_stride = MAX_SB_SIZE;
-      filter_size = filter_params_y.taps;
-
-      intermediate_height =
-          (((h - 1) * y_step_q4 + subpel_y_q4) >> SUBPEL_BITS) + filter_size;
-      assert(intermediate_height <= max_intermediate_size);
-      (void)max_intermediate_size;
-
-      av1_highbd_convolve_horiz_facade(
-          src8 - src_stride * (filter_size / 2 - 1), src_stride, temp8,
-          temp_stride, w, intermediate_height, filter_params_x, subpel_x_q4,
-          x_step_q4, 0, bd);
-
-      filter_size = filter_params_y.taps;
-      assert(filter_params_y.taps <= MAX_FILTER_TAP);
-
-      av1_highbd_convolve_vert_facade(
-          temp8 + temp_stride * (filter_size / 2 - 1), temp_stride, dst8,
-          dst_stride, w, h, filter_params_y, subpel_y_q4, y_step_q4, ref_idx,
-          bd);
+  src -= src_stride * (SUBPEL_TAPS / 2 - 1);
+  for (int x = 0; x < w; ++x) {
+    int y_q4 = y0_q4;
+    for (int 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];
+      const int rounding =
+          ((int)src_y[(SUBPEL_TAPS / 2 - 1) * src_stride] << FILTER_BITS) -
+          (1 << (bd + round1_bits - 1));
+      const int sum =
+          highbd_vert_scalar_product(src_y, src_stride, y_filter) + rounding;
+      dst[y * dst_stride] =
+          clip_pixel_highbd(ROUND_POWER_OF_TWO(sum, round1_bits), bd);
+      y_q4 += y_step_q4;
     }
+    ++src;
+    ++dst;
   }
 }
 
-void av1_highbd_convolve_scale(const uint8_t *src8, int src_stride,
-                               uint8_t *dst8, int dst_stride, int w, int h,
-                               InterpFilters interp_filters,
-                               const int subpel_x_qn, int x_step_qn,
-                               const int subpel_y_qn, int y_step_qn,
-                               int ref_idx, int bd) {
-  uint16_t *src = CONVERT_TO_SHORTPTR(src8);
-  uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
-  int ignore_horiz = x_step_qn == SCALE_SUBPEL_SHIFTS && subpel_x_qn == 0;
-  int ignore_vert = y_step_qn == SCALE_SUBPEL_SHIFTS && subpel_y_qn == 0;
-
-  assert(w <= MAX_BLOCK_WIDTH);
-  assert(h <= MAX_BLOCK_HEIGHT);
-  assert(y_step_qn <= (MAX_STEP << SCALE_EXTRA_BITS));
-  assert(x_step_qn <= (MAX_STEP << SCALE_EXTRA_BITS));
-
-  if (ignore_horiz && ignore_vert) {
-    highbd_convolve_copy(src, src_stride, dst, dst_stride, w, h, ref_idx, bd);
-    return;
-  }
-
-  InterpFilterParams filter_params_x, filter_params_y;
-  av1_get_convolve_filter_params(interp_filters, 0, &filter_params_x,
-                                 &filter_params_y);
-
-  if (ignore_vert) {
-    av1_highbd_convolve_horiz_facade_scale(src8, src_stride, dst8, dst_stride,
-                                           w, h, filter_params_x, subpel_x_qn,
-                                           x_step_qn, ref_idx, bd);
-  } else if (ignore_horiz) {
-    av1_highbd_convolve_vert_facade_scale(src8, src_stride, dst8, dst_stride, w,
-                                          h, filter_params_y, subpel_y_qn,
-                                          y_step_qn, ref_idx, bd);
-  } else {
-    // temp's size is set to a 256 aligned value to facilitate SIMD
-    // implementation. The value is greater than (maximum possible intermediate
-    // height or width) * MAX_SB_SIZE
-    DECLARE_ALIGNED(16, uint16_t,
-                    temp[((MAX_SB_SIZE * 2 + 16) + 16) * MAX_SB_SIZE]);
-    uint8_t *temp8 = CONVERT_TO_BYTEPTR(temp);
-    int max_intermediate_size = ((MAX_SB_SIZE * 2 + 16) + 16);
-    int filter_size;
-
-#if CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
-    av1_convolve_filter_params_fixup_1212(&filter_params_x, &filter_params_y);
-
-    if (filter_params_y.taps < filter_params_x.taps) {
-      int intermediate_width;
-      int temp_stride = max_intermediate_size;
-      filter_size = filter_params_x.taps;
-      intermediate_width =
-          (((w - 1) * x_step_qn + subpel_x_qn) >> SCALE_SUBPEL_BITS) +
-          filter_size;
-      assert(intermediate_width <= max_intermediate_size);
-
-      assert(filter_params_y.taps <= MAX_FILTER_TAP);
-
-      av1_highbd_convolve_vert_facade_scale(
-          src8 - (filter_size / 2 - 1), src_stride, temp8, temp_stride,
-          intermediate_width, h, filter_params_y, subpel_y_qn, y_step_qn, 0,
-          bd);
-
-      assert(filter_params_x.taps <= MAX_FILTER_TAP);
-
-      av1_highbd_convolve_horiz_facade_scale(
-          temp8 + (filter_size / 2 - 1), temp_stride, dst8, dst_stride, w, h,
-          filter_params_x, subpel_x_qn, x_step_qn, ref_idx, bd);
-    } else {
-#endif  // CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
-      int intermediate_height;
-      int temp_stride = MAX_SB_SIZE;
-      filter_size = filter_params_y.taps;
-      intermediate_height =
-          (((h - 1) * y_step_qn + subpel_y_qn) >> SCALE_SUBPEL_BITS) +
-          filter_size;
-      assert(intermediate_height <= max_intermediate_size);
-      (void)max_intermediate_size;
-
-      av1_highbd_convolve_horiz_facade_scale(
-          src8 - src_stride * (filter_size / 2 - 1), src_stride, temp8,
-          temp_stride, w, intermediate_height, filter_params_x, subpel_x_qn,
-          x_step_qn, 0, bd);
-
-      filter_size = filter_params_y.taps;
-      assert(filter_params_y.taps <= MAX_FILTER_TAP);
-
-      av1_highbd_convolve_vert_facade_scale(
-          temp8 + temp_stride * (filter_size / 2 - 1), temp_stride, dst8,
-          dst_stride, w, h, filter_params_y, subpel_y_qn, y_step_qn, ref_idx,
-          bd);
-#if CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
-    }
-#endif  // CONFIG_DUAL_FILTER && USE_EXTRA_FILTER
-  }
+void av1_highbd_wiener_convolve_add_src_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,
+    const ConvolveParams *conv_params, 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);
+
+  uint16_t temp[WIENER_MAX_EXT_SIZE * MAX_SB_SIZE];
+  const int intermediate_height =
+      (((h - 1) * y_step_q4 + y0_q4) >> SUBPEL_BITS) + SUBPEL_TAPS;
+
+  assert(w <= MAX_SB_SIZE);
+  assert(h <= MAX_SB_SIZE);
+  assert(y_step_q4 <= 32);
+  assert(x_step_q4 <= 32);
+  assert(bd + FILTER_BITS - conv_params->round_0 + 2 <= 16);
+
+  highbd_convolve_add_src_horiz_hip(src - src_stride * (SUBPEL_TAPS / 2 - 1),
+                                    src_stride, temp, MAX_SB_SIZE, filters_x,
+                                    x0_q4, x_step_q4, w, intermediate_height,
+                                    conv_params->round_0, bd);
+  highbd_convolve_add_src_vert_hip(
+      temp + MAX_SB_SIZE * (SUBPEL_TAPS / 2 - 1), MAX_SB_SIZE, dst, dst_stride,
+      filters_y, y0_q4, y_step_q4, w, h, conv_params->round_1, bd);
 }
-#endif  // CONFIG_HIGHBITDEPTH