diff --git a/av1/encoder/encodeframe.c b/av1/encoder/encodeframe.c
index 0c3cf6d..203f563 100644
--- a/av1/encoder/encodeframe.c
+++ b/av1/encoder/encodeframe.c
@@ -339,7 +339,6 @@
   if (tpl_frame->is_valid == 0) return;
   if (!is_frame_tpl_eligible(gf_group, gf_group->index)) return;
   if (frame_idx >= MAX_TPL_FRAME_IDX) return;
-  if (cpi->superres_mode != AOM_SUPERRES_NONE) return;
   if (cpi->oxcf.q_cfg.aq_mode != NO_AQ) return;
 
   const int is_overlay =
@@ -355,13 +354,21 @@
   int64_t inter_cost[INTER_REFS_PER_FRAME] = { 0 };
   const int step = 1 << block_mis_log2;
   const BLOCK_SIZE sb_size = cm->seq_params.sb_size;
+
   const int mi_row_end =
       AOMMIN(mi_size_high[sb_size] + mi_row, mi_params->mi_rows);
-  const int mi_col_end =
-      AOMMIN(mi_size_wide[sb_size] + mi_col, mi_params->mi_cols);
-
-  for (int row = mi_row; row < mi_row_end; row += step) {
-    for (int col = mi_col; col < mi_col_end; col += step) {
+  const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);
+  const int mi_col_sr =
+      coded_to_superres_mi(mi_col, cm->superres_scale_denominator);
+  const int mi_col_end_sr =
+      AOMMIN(coded_to_superres_mi(mi_col + mi_size_wide[sb_size],
+                                  cm->superres_scale_denominator),
+             mi_cols_sr);
+  const int row_step = step;
+  const int col_step_sr =
+      coded_to_superres_mi(step, cm->superres_scale_denominator);
+  for (int row = mi_row; row < mi_row_end; row += row_step) {
+    for (int col = mi_col_sr; col < mi_col_end_sr; col += col_step_sr) {
       const TplDepStats *this_stats =
           &tpl_stats[av1_tpl_ptr_pos(row, col, tpl_stride, block_mis_log2)];
       int64_t tpl_pred_error[INTER_REFS_PER_FRAME] = { 0 };
diff --git a/av1/encoder/encodeframe_utils.c b/av1/encoder/encodeframe_utils.c
index 7cc3bb7..eb9b3e0 100644
--- a/av1/encoder/encodeframe_utils.c
+++ b/av1/encoder/encodeframe_utils.c
@@ -62,6 +62,26 @@
   aom_clear_system_state();
 }
 
+// Return the end column for the current superblock, in unit of TPL blocks.
+static int get_superblock_tpl_column_end(const AV1_COMMON *const cm, int mi_col,
+                                         int num_mi_w) {
+  // Find the start column of this superblock.
+  const int sb_mi_col_start = (mi_col >> cm->seq_params.mib_size_log2)
+                              << cm->seq_params.mib_size_log2;
+  // Same but in superres upscaled dimension.
+  const int sb_mi_col_start_sr =
+      coded_to_superres_mi(sb_mi_col_start, cm->superres_scale_denominator);
+  // Width of this superblock in mi units.
+  const int sb_mi_width = mi_size_wide[cm->seq_params.sb_size];
+  // Same but in superres upscaled dimension.
+  const int sb_mi_width_sr =
+      coded_to_superres_mi(sb_mi_width, cm->superres_scale_denominator);
+  // Superblock end in mi units.
+  const int sb_mi_end = sb_mi_col_start_sr + sb_mi_width_sr;
+  // Superblock end in TPL units.
+  return (sb_mi_end + num_mi_w - 1) / num_mi_w;
+}
+
 int av1_get_hier_tpl_rdmult(const AV1_COMP *const cpi, MACROBLOCK *const x,
                             const BLOCK_SIZE bsize, const int mi_row,
                             const int mi_col, int orig_rdmult) {
@@ -75,24 +95,34 @@
   if (tpl_frame->is_valid == 0) return deltaq_rdmult;
   if (!is_frame_tpl_eligible(gf_group, gf_group->index)) return deltaq_rdmult;
   if (tpl_idx >= MAX_TPL_FRAME_IDX) return deltaq_rdmult;
-  if (cpi->superres_mode != AOM_SUPERRES_NONE) return deltaq_rdmult;
   if (cpi->oxcf.q_cfg.aq_mode != NO_AQ) return deltaq_rdmult;
 
+  const int mi_col_sr =
+      coded_to_superres_mi(mi_col, cm->superres_scale_denominator);
+  const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);
+  const int block_mi_width_sr =
+      coded_to_superres_mi(mi_size_wide[bsize], cm->superres_scale_denominator);
+
   const int bsize_base = BLOCK_16X16;
   const int num_mi_w = mi_size_wide[bsize_base];
   const int num_mi_h = mi_size_high[bsize_base];
-  const int num_cols = (cm->mi_params.mi_cols + num_mi_w - 1) / num_mi_w;
+  const int num_cols = (mi_cols_sr + num_mi_w - 1) / num_mi_w;
   const int num_rows = (cm->mi_params.mi_rows + num_mi_h - 1) / num_mi_h;
-  const int num_bcols = (mi_size_wide[bsize] + num_mi_w - 1) / num_mi_w;
+  const int num_bcols = (block_mi_width_sr + num_mi_w - 1) / num_mi_w;
   const int num_brows = (mi_size_high[bsize] + num_mi_h - 1) / num_mi_h;
+  // This is required because the end col of superblock may be off by 1 in case
+  // of superres.
+  const int sb_bcol_end = get_superblock_tpl_column_end(cm, mi_col, num_mi_w);
   int row, col;
   double base_block_count = 0.0;
   double geom_mean_of_scale = 0.0;
   aom_clear_system_state();
   for (row = mi_row / num_mi_w;
        row < num_rows && row < mi_row / num_mi_w + num_brows; ++row) {
-    for (col = mi_col / num_mi_h;
-         col < num_cols && col < mi_col / num_mi_h + num_bcols; ++col) {
+    for (col = mi_col_sr / num_mi_h;
+         col < num_cols && col < mi_col_sr / num_mi_h + num_bcols &&
+         col < sb_bcol_end;
+         ++col) {
       const int index = row * num_cols + col;
       geom_mean_of_scale += log(cpi->tpl_sb_rdmult_scaling_factors[index]);
       base_block_count += 1.0;
@@ -677,8 +707,11 @@
       coded_to_superres_mi(mi_col + mi_wide, cm->superres_scale_denominator);
   const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);
   const int step = 1 << block_mis_log2;
-  for (int row = mi_row; row < mi_row + mi_high; row += step) {
-    for (int col = mi_col_sr; col < mi_col_end_sr; col += step) {
+  const int row_step = step;
+  const int col_step_sr =
+      coded_to_superres_mi(step, cm->superres_scale_denominator);
+  for (int row = mi_row; row < mi_row + mi_high; row += row_step) {
+    for (int col = mi_col_sr; col < mi_col_end_sr; col += col_step_sr) {
       if (row >= cm->mi_params.mi_rows || col >= mi_cols_sr) continue;
       TplDepStats *this_stats =
           &tpl_stats[av1_tpl_ptr_pos(row, col, tpl_stride, block_mis_log2)];
@@ -690,6 +723,7 @@
       mi_count++;
     }
   }
+  assert(mi_count <= MAX_TPL_BLK_IN_SB * MAX_TPL_BLK_IN_SB);
 
   aom_clear_system_state();
 
@@ -747,7 +781,6 @@
   sb_enc->tpl_data_count = 0;
 
   if (!cpi->oxcf.algo_cfg.enable_tpl_model) return;
-  if (cpi->superres_mode != AOM_SUPERRES_NONE) return;
   if (cpi->common.current_frame.frame_type == KEY_FRAME) return;
   const FRAME_UPDATE_TYPE update_type = get_frame_update_type(&cpi->gf_group);
   if (update_type == INTNL_OVERLAY_UPDATE || update_type == OVERLAY_UPDATE ||
@@ -781,15 +814,18 @@
   // Here always use motion estimation size to avoid getting repetitive inter/
   // intra cost.
   const BLOCK_SIZE tpl_bsize = convert_length_to_bsize(tpl_data->tpl_bsize_1d);
-  const int step = mi_size_wide[tpl_bsize];
   assert(mi_size_wide[tpl_bsize] == mi_size_high[tpl_bsize]);
+  const int row_step = mi_size_high[tpl_bsize];
+  const int col_step_sr = coded_to_superres_mi(mi_size_wide[tpl_bsize],
+                                               cm->superres_scale_denominator);
 
   // Stride is only based on SB size, and we fill in values for every 16x16
   // block in a SB.
-  sb_enc->tpl_stride = (mi_col_end_sr - mi_col_sr) / step;
+  sb_enc->tpl_stride = (mi_col_end_sr - mi_col_sr) / col_step_sr;
 
-  for (int row = mi_row; row < mi_row + mi_high; row += step) {
-    for (int col = mi_col_sr; col < mi_col_end_sr; col += step) {
+  for (int row = mi_row; row < mi_row + mi_high; row += row_step) {
+    for (int col = mi_col_sr; col < mi_col_end_sr; col += col_step_sr) {
+      assert(count < MAX_TPL_BLK_IN_SB * MAX_TPL_BLK_IN_SB);
       // Handle partial SB, so that no invalid values are used later.
       if (row >= cm->mi_params.mi_rows || col >= mi_cols_sr) {
         sb_enc->tpl_inter_cost[count] = INT64_MAX;
@@ -811,6 +847,7 @@
     }
   }
 
+  assert(mi_count <= MAX_TPL_BLK_IN_SB * MAX_TPL_BLK_IN_SB);
   sb_enc->tpl_data_count = mi_count;
 }
 
@@ -849,8 +886,11 @@
       coded_to_superres_mi(mi_col + mi_wide, cm->superres_scale_denominator);
   const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);
   const int step = 1 << block_mis_log2;
-  for (int row = mi_row; row < mi_row + mi_high; row += step) {
-    for (int col = mi_col_sr; col < mi_col_end_sr; col += step) {
+  const int row_step = step;
+  const int col_step_sr =
+      coded_to_superres_mi(step, cm->superres_scale_denominator);
+  for (int row = mi_row; row < mi_row + mi_high; row += row_step) {
+    for (int col = mi_col_sr; col < mi_col_end_sr; col += col_step_sr) {
       if (row >= cm->mi_params.mi_rows || col >= mi_cols_sr) continue;
       TplDepStats *this_stats =
           &tpl_stats[av1_tpl_ptr_pos(row, col, tpl_stride, block_mis_log2)];
@@ -862,6 +902,7 @@
       mi_count++;
     }
   }
+  assert(mi_count <= MAX_TPL_BLK_IN_SB * MAX_TPL_BLK_IN_SB);
 
   aom_clear_system_state();
 
diff --git a/av1/encoder/encoder_utils.c b/av1/encoder/encoder_utils.c
index 754b4a6..17f72ff 100644
--- a/av1/encoder/encoder_utils.c
+++ b/av1/encoder/encoder_utils.c
@@ -370,10 +370,13 @@
     int64_t intra_cost_base = 0;
     int64_t mc_dep_cost_base = 0;
     const int step = 1 << tpl_data->tpl_stats_block_mis_log2;
+    const int row_step = step;
+    const int col_step_sr =
+        coded_to_superres_mi(step, cm->superres_scale_denominator);
     const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);
 
-    for (int row = 0; row < cm->mi_params.mi_rows; row += step) {
-      for (int col = 0; col < mi_cols_sr; col += step) {
+    for (int row = 0; row < cm->mi_params.mi_rows; row += row_step) {
+      for (int col = 0; col < mi_cols_sr; col += col_step_sr) {
         TplDepStats *this_stats = &tpl_stats[av1_tpl_ptr_pos(
             row, col, tpl_stride, tpl_data->tpl_stats_block_mis_log2)];
         int64_t mc_dep_delta =
diff --git a/av1/encoder/rdopt.c b/av1/encoder/rdopt.c
index c477af4..7c21f4d 100644
--- a/av1/encoder/rdopt.c
+++ b/av1/encoder/rdopt.c
@@ -2171,10 +2171,13 @@
       coded_to_superres_mi(mi_col + mi_wide, cm->superres_scale_denominator);
   const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);
 
+  const int row_step = step;
+  const int col_step_sr =
+      coded_to_superres_mi(step, cm->superres_scale_denominator);
   for (int row = mi_row; row < AOMMIN(mi_row + mi_high, cm->mi_params.mi_rows);
-       row += step) {
+       row += row_step) {
     for (int col = mi_col_sr; col < AOMMIN(mi_col_end_sr, mi_cols_sr);
-         col += step) {
+         col += col_step_sr) {
       const TplDepStats *this_stats = &tpl_stats[av1_tpl_ptr_pos(
           row, col, tpl_stride, tpl_data->tpl_stats_block_mis_log2)];
 
diff --git a/av1/encoder/rdopt.h b/av1/encoder/rdopt.h
index 6b4b8a2..8717024 100644
--- a/av1/encoder/rdopt.h
+++ b/av1/encoder/rdopt.h
@@ -205,11 +205,9 @@
 void av1_inter_mode_data_init(struct TileDataEnc *tile_data);
 void av1_inter_mode_data_fit(TileDataEnc *tile_data, int rdmult);
 
-#if !CONFIG_REALTIME_ONLY
 static INLINE int coded_to_superres_mi(int mi_col, int denom) {
   return (mi_col * denom + SCALE_NUMERATOR / 2) / SCALE_NUMERATOR;
 }
-#endif
 
 static INLINE int av1_encoder_get_relative_dist(int a, int b) {
   assert(a >= 0 && b >= 0);
diff --git a/av1/encoder/tpl_model.c b/av1/encoder/tpl_model.c
index c5b960d..636b57d 100644
--- a/av1/encoder/tpl_model.c
+++ b/av1/encoder/tpl_model.c
@@ -1172,7 +1172,10 @@
   EncodeFrameParams this_frame_params = *frame_params;
   TplParams *const tpl_data = &cpi->tpl_data;
 
-  if (cpi->superres_mode != AOM_SUPERRES_NONE) return 0;
+  if (cpi->superres_mode != AOM_SUPERRES_NONE) {
+    assert(cpi->superres_mode != AOM_SUPERRES_AUTO);
+    return 0;
+  }
 
   cm->current_frame.frame_type = frame_params->frame_type;
   for (int gf_index = gf_group->index; gf_index < gf_group->size; ++gf_index) {
@@ -1258,10 +1261,13 @@
     int64_t intra_cost_base = 0;
     int64_t mc_dep_cost_base = 0;
     const int step = 1 << tpl_data->tpl_stats_block_mis_log2;
+    const int row_step = step;
+    const int col_step_sr =
+        coded_to_superres_mi(step, cm->superres_scale_denominator);
     const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);
 
-    for (int row = 0; row < cm->mi_params.mi_rows; row += step) {
-      for (int col = 0; col < mi_cols_sr; col += step) {
+    for (int row = 0; row < cm->mi_params.mi_rows; row += row_step) {
+      for (int col = 0; col < mi_cols_sr; col += col_step_sr) {
         TplDepStats *this_stats = &tpl_stats[av1_tpl_ptr_pos(
             row, col, tpl_stride, tpl_data->tpl_stats_block_mis_log2)];
         int64_t mc_dep_delta =
@@ -1293,7 +1299,6 @@
   const TplDepFrame *const tpl_frame = &tpl_data->tpl_frame[tpl_idx];
 
   if (!tpl_frame->is_valid) return;
-  if (cpi->superres_mode != AOM_SUPERRES_NONE) return;
 
   const TplDepStats *const tpl_stats = tpl_frame->tpl_stats_ptr;
   const int tpl_stride = tpl_frame->stride;
@@ -1349,15 +1354,20 @@
   if (tpl_frame->is_valid == 0) return;
   if (!is_frame_tpl_eligible(gf_group, gf_group->index)) return;
   if (tpl_idx >= MAX_TPL_FRAME_IDX) return;
-  if (cpi->superres_mode != AOM_SUPERRES_NONE) return;
   if (cpi->oxcf.q_cfg.aq_mode != NO_AQ) return;
 
+  const int mi_col_sr =
+      coded_to_superres_mi(mi_col, cm->superres_scale_denominator);
+  const int mi_cols_sr = av1_pixels_to_mi(cm->superres_upscaled_width);
+  const int sb_mi_width_sr = coded_to_superres_mi(
+      mi_size_wide[sb_size], cm->superres_scale_denominator);
+
   const int bsize_base = BLOCK_16X16;
   const int num_mi_w = mi_size_wide[bsize_base];
   const int num_mi_h = mi_size_high[bsize_base];
-  const int num_cols = (cm->mi_params.mi_cols + num_mi_w - 1) / num_mi_w;
+  const int num_cols = (mi_cols_sr + num_mi_w - 1) / num_mi_w;
   const int num_rows = (cm->mi_params.mi_rows + num_mi_h - 1) / num_mi_h;
-  const int num_bcols = (mi_size_wide[sb_size] + num_mi_w - 1) / num_mi_w;
+  const int num_bcols = (sb_mi_width_sr + num_mi_w - 1) / num_mi_w;
   const int num_brows = (mi_size_high[sb_size] + num_mi_h - 1) / num_mi_h;
   int row, col;
 
@@ -1367,8 +1377,8 @@
   aom_clear_system_state();
   for (row = mi_row / num_mi_w;
        row < num_rows && row < mi_row / num_mi_w + num_brows; ++row) {
-    for (col = mi_col / num_mi_h;
-         col < num_cols && col < mi_col / num_mi_h + num_bcols; ++col) {
+    for (col = mi_col_sr / num_mi_h;
+         col < num_cols && col < mi_col_sr / num_mi_h + num_bcols; ++col) {
       const int index = row * num_cols + col;
       log_sum += log(cpi->tpl_rdmult_scaling_factors[index]);
       base_block_count += 1.0;
@@ -1388,8 +1398,8 @@
 
   for (row = mi_row / num_mi_w;
        row < num_rows && row < mi_row / num_mi_w + num_brows; ++row) {
-    for (col = mi_col / num_mi_h;
-         col < num_cols && col < mi_col / num_mi_h + num_bcols; ++col) {
+    for (col = mi_col_sr / num_mi_h;
+         col < num_cols && col < mi_col_sr / num_mi_h + num_bcols; ++col) {
       const int index = row * num_cols + col;
       cpi->tpl_sb_rdmult_scaling_factors[index] =
           scale_adj * cpi->tpl_rdmult_scaling_factors[index];