From 99ddc209fc6e118b49432ffc3588411c2951baa7 Mon Sep 17 00:00:00 2001
From: Joose Sainio <joose.sainio@tuni.fi>
Date: Fri, 25 Mar 2022 10:18:53 +0200
Subject: [PATCH] Fix cclm when tiles are enabled

---
 src/intra.c  | 9 +++++----
 src/search.c | 7 ++++---
 2 files changed, 9 insertions(+), 7 deletions(-)

diff --git a/src/intra.c b/src/intra.c
index b955302e..f5f255f6 100644
--- a/src/intra.c
+++ b/src/intra.c
@@ -498,6 +498,7 @@ void kvz_predict_cclm(
 
 
   kvz_pixel *y_rec = lcu->rec.y + x_scu + y_scu * LCU_WIDTH;
+  const int stride2 = (((state->tile->frame->width + 7) & ~7) + FRAME_PADDING_LUMA);
 
   // Essentially what this does is that it uses 6-tap filtering to downsample
   // the luma intra references down to match the resolution of the chroma channel.
@@ -513,7 +514,7 @@ void kvz_predict_cclm(
     }
     if(y_scu == 0) {
       if(!state->encoder_control->cfg.wpp) available_above_right = MIN(width / 2, (state->tile->frame->width - x0 - width * 2) / 4);
-      memcpy(sampled_luma_ref.top, &state->tile->frame->cclm_luma_rec_top_line[x0 / 2 + (y0 / 64 - 1) * (stride / 2)], sizeof(kvz_pixel) * (width + available_above_right * 2));
+      memcpy(sampled_luma_ref.top, &state->tile->frame->cclm_luma_rec_top_line[x0 / 2 + (y0 / 64 - 1) * (stride2 / 2)], sizeof(kvz_pixel) * (width + available_above_right * 2));
     }
     else {
       for (int x = 0; x < width * (available_above_right ? 4 : 2); x += 2) {
@@ -538,11 +539,11 @@ void kvz_predict_cclm(
       if(x_scu == 32 && y_scu == 0 && pu->depth == 0) break;
     }
     for(int i = 0; i < height + available_left_below * 2; i++) {
-      sampled_luma_ref.left[i] = state->tile->frame->cclm_luma_rec[(y0/2 + i) * (stride/2) + x0 / 2 - 1];
+      sampled_luma_ref.left[i] = state->tile->frame->cclm_luma_rec[(y0/2 + i) * (stride2/2) + x0 / 2 - 1];
     }    
   }
 
-  kvz_pixels_blit(&state->tile->frame->cclm_luma_rec[x0 / 2 + (y0 * stride) / 4], sampled_luma, width, height, stride / 2, width);
+  kvz_pixels_blit(&state->tile->frame->cclm_luma_rec[x0 / 2 + (y0 * stride2) / 4], sampled_luma, width, height, stride2 / 2, width);
 
   int16_t a, b, shift;
   get_cclm_parameters(state, width, height, mode,x0, y0, available_above_right, available_left_below, &sampled_luma_ref, chroma_ref, &a, &b, &shift);
@@ -1402,7 +1403,7 @@ static void intra_recon_tb_leaf(
   kvz_intra_build_reference(log2width, color, &luma_px, &pic_px, lcu, &refs, cfg->wpp, extra_refs, multi_ref_index);
 
   kvz_pixel pred[32 * 32];
-  int stride = state->tile->frame->source->stride;
+  const int stride = (((state->tile->frame->width + 7) & ~7) + FRAME_PADDING_LUMA);
   const bool filter_boundary = color == COLOR_Y && !(cfg->lossless && cfg->implicit_rdpcm);
   bool use_mip = false;
   int8_t intra_mode = color == COLOR_Y ? intra_paramas->luma_mode : intra_paramas->chroma_mode;
diff --git a/src/search.c b/src/search.c
index 5c409b3f..244a2b7e 100644
--- a/src/search.c
+++ b/src/search.c
@@ -244,7 +244,8 @@ static void downsample_cclm_rec(encoder_state_t *state, int x, int y, int width,
   int x_scu = SUB_SCU(x);
   int y_scu = SUB_SCU(y);
   y_rec += x_scu + y_scu * LCU_WIDTH;
-  int stride = state->tile->frame->source->stride;
+  const int stride = state->tile->frame->rec->stride;
+  const int stride2 = (((state->tile->frame->width + 7) & ~7) + FRAME_PADDING_LUMA);
 
   for (int y_ = 0; y_ < height && y_ * 2 + y < state->encoder_control->cfg.height; y_++) {
     for (int x_ = 0; x_ < width; x_++) {
@@ -258,13 +259,13 @@ static void downsample_cclm_rec(encoder_state_t *state, int x, int y, int width,
       s += y_rec[2 * x_ + LCU_WIDTH] * 2;
       s += y_rec[2 * x_ + 1 + LCU_WIDTH];
       s += !x_scu && !x_ && x ? state->tile->frame->rec->y[x - 1 + (y + y_ * 2 + 1) * stride] : y_rec[2 * x_ - ((x_ + x) > 0) + LCU_WIDTH];
-      int index = x / 2 + x_ + (y / 2 + y_ )* stride / 2;
+      int index = x / 2 + x_ + (y / 2 + y_ )* stride2 / 2;
       state->tile->frame->cclm_luma_rec[index] = s >> 3;
     }
     y_rec += LCU_WIDTH * 2;
   }
   if((y + height * 2) % 64 == 0) {
-    int line = y / 64 * stride / 2;
+    int line = y / 64 * stride2 / 2;
     y_rec -= LCU_WIDTH;
     for (int i = 0; i < width; ++i) {
       int s = 2;