libvpx: Pull from upstream

source/libvpx/vp9/encoder/vp9_denoiser.c

Index: source/libvpx/vp9/encoder/vp9_denoiser.c
===================================================================
--- source/libvpx/vp9/encoder/vp9_denoiser.c	(revision 293588)
+++ source/libvpx/vp9/encoder/vp9_denoiser.c	(working copy)
@@ -49,9 +49,7 @@
 }
 
 static unsigned int sse_thresh(BLOCK_SIZE bs, int increase_denoising) {
-  return (4 << b_width_log2_lookup[bs]) *
-         (4 << b_height_log2_lookup[bs]) *
-         (increase_denoising ? 60 : 40);
+  return (1 << num_pels_log2_lookup[bs]) * (increase_denoising ? 60 : 40);
 }
 
 static int sse_diff_thresh(BLOCK_SIZE bs, int increase_denoising,
@@ -60,19 +58,16 @@
       noise_motion_thresh(bs, increase_denoising)) {
     return 0;
   } else {
-    return (4 << b_width_log2_lookup[bs]) *
-           (4 << b_height_log2_lookup[bs]) * 20;
+    return (1 << num_pels_log2_lookup[bs]) * 20;
   }
 }
 
 int total_adj_strong_thresh(BLOCK_SIZE bs, int increase_denoising) {
-  return (4 << b_width_log2_lookup[bs]) *
-         (4 << b_height_log2_lookup[bs]) * (increase_denoising ? 3 : 2);
+  return (1 << num_pels_log2_lookup[bs]) * (increase_denoising ? 3 : 2);
 }
 
 static int total_adj_weak_thresh(BLOCK_SIZE bs, int increase_denoising) {
-  return (4 << b_width_log2_lookup[bs]) *
-         (4 << b_height_log2_lookup[bs]) * (increase_denoising ? 3 : 2);
+  return (1 << num_pels_log2_lookup[bs]) * (increase_denoising ? 3 : 2);
 }
 
 // TODO(jackychen): If increase_denoising is enabled in the future,
@@ -195,16 +190,6 @@
   return framebuf + (stride * mi_row * 8) + (mi_col * 8);
 }
 
-static void copy_block(uint8_t *dest, int dest_stride,
-                       const uint8_t *src, int src_stride, BLOCK_SIZE bs) {
-  int r;
-  for (r = 0; r < (4 << b_height_log2_lookup[bs]); ++r) {
-    vpx_memcpy(dest, src, (4 << b_width_log2_lookup[bs]));
-    dest += dest_stride;
-    src += src_stride;
-  }
-}
-
 static VP9_DENOISER_DECISION perform_motion_compensation(VP9_DENOISER *denoiser,
                                                          MACROBLOCK *mb,
                                                          BLOCK_SIZE bs,
@@ -219,28 +204,18 @@
   MV_REFERENCE_FRAME frame;
   MACROBLOCKD *filter_mbd = &mb->e_mbd;
   MB_MODE_INFO *mbmi = &filter_mbd->mi[0].src_mi->mbmi;
-
   MB_MODE_INFO saved_mbmi;
   int i, j;
   struct buf_2d saved_dst[MAX_MB_PLANE];
   struct buf_2d saved_pre[MAX_MB_PLANE][2];  // 2 pre buffers
 
-  // We will restore these after motion compensation.
-  saved_mbmi = *mbmi;
-  for (i = 0; i < MAX_MB_PLANE; ++i) {
-    for (j = 0; j < 2; ++j) {
-      saved_pre[i][j] = filter_mbd->plane[i].pre[j];
-    }
-    saved_dst[i] = filter_mbd->plane[i].dst;
-  }
-
   mv_col = ctx->best_sse_mv.as_mv.col;
   mv_row = ctx->best_sse_mv.as_mv.row;
-
   *motion_magnitude = mv_row * mv_row + mv_col * mv_col;
-
   frame = ctx->best_reference_frame;
 
+  saved_mbmi = *mbmi;
+
   // If the best reference frame uses inter-prediction and there is enough of a
   // difference in sum-squared-error, use it.
   if (frame != INTRA_FRAME &&
@@ -261,6 +236,26 @@
     ctx->newmv_sse = ctx->zeromv_sse;
   }
 
+  if (ctx->newmv_sse > sse_thresh(bs, increase_denoising)) {
+    // Restore everything to its original state
+    *mbmi = saved_mbmi;
+    return COPY_BLOCK;
+  }
+  if (mv_row * mv_row + mv_col * mv_col >
+      8 * noise_motion_thresh(bs, increase_denoising)) {
+    // Restore everything to its original state
+    *mbmi = saved_mbmi;
+    return COPY_BLOCK;
+  }
+
+  // We will restore these after motion compensation.
+  for (i = 0; i < MAX_MB_PLANE; ++i) {
+    for (j = 0; j < 2; ++j) {
+      saved_pre[i][j] = filter_mbd->plane[i].pre[j];
+    }
+    saved_dst[i] = filter_mbd->plane[i].dst;
+  }
+
   // Set the pointers in the MACROBLOCKD to point to the buffers in the denoiser
   // struct.
   for (j = 0; j < 2; ++j) {
@@ -313,13 +308,6 @@
   mv_row = ctx->best_sse_mv.as_mv.row;
   mv_col = ctx->best_sse_mv.as_mv.col;
 
-  if (ctx->newmv_sse > sse_thresh(bs, increase_denoising)) {
-    return COPY_BLOCK;
-  }
-  if (mv_row * mv_row + mv_col * mv_col >
-      8 * noise_motion_thresh(bs, increase_denoising)) {
-    return COPY_BLOCK;
-  }
   return FILTER_BLOCK;
 }
 
@@ -348,9 +336,15 @@
   }
 
   if (decision == FILTER_BLOCK) {
-    copy_block(src.buf, src.stride, avg_start, avg.y_stride, bs);
+    vp9_convolve_copy(avg_start, avg.y_stride, src.buf, src.stride,
+                      NULL, 0, NULL, 0,
+                      num_4x4_blocks_wide_lookup[bs] << 2,
+                      num_4x4_blocks_high_lookup[bs] << 2);
   } else {  // COPY_BLOCK
-    copy_block(avg_start, avg.y_stride, src.buf, src.stride, bs);
+    vp9_convolve_copy(src.buf, src.stride, avg_start, avg.y_stride,
+                      NULL, 0, NULL, 0,
+                      num_4x4_blocks_wide_lookup[bs] << 2,
+                      num_4x4_blocks_high_lookup[bs] << 2);
   }
 }
 
@@ -368,6 +362,15 @@
   }
 }
 
+static void swap_frame_buffer(YV12_BUFFER_CONFIG dest,
+                              YV12_BUFFER_CONFIG src) {
+  uint8_t *tmp_buf = dest.y_buffer;
+  assert(dest.y_width == src.y_width);
+  assert(dest.y_height == src.y_height);
+  dest.y_buffer = src.y_buffer;
+  src.y_buffer = tmp_buf;
+}
+
 void vp9_denoiser_update_frame_info(VP9_DENOISER *denoiser,
                                     YV12_BUFFER_CONFIG src,
                                     FRAME_TYPE frame_type,
@@ -377,28 +380,32 @@
   if (frame_type == KEY_FRAME) {
     int i;
     // Start at 1 so as not to overwrite the INTRA_FRAME
-    for (i = 1; i < MAX_REF_FRAMES; ++i) {
+    for (i = 1; i < MAX_REF_FRAMES; ++i)
       copy_frame(denoiser->running_avg_y[i], src);
-    }
-  } else {  /* For non key frames */
-    if (refresh_alt_ref_frame) {
-      copy_frame(denoiser->running_avg_y[ALTREF_FRAME],
-                 denoiser->running_avg_y[INTRA_FRAME]);
-    }
-    if (refresh_golden_frame) {
-      copy_frame(denoiser->running_avg_y[GOLDEN_FRAME],
-                 denoiser->running_avg_y[INTRA_FRAME]);
-    }
-    if (refresh_last_frame) {
-      copy_frame(denoiser->running_avg_y[LAST_FRAME],
-                 denoiser->running_avg_y[INTRA_FRAME]);
-    }
+    return;
   }
+
+  /* For non key frames */
+  if (refresh_alt_ref_frame) {
+    swap_frame_buffer(denoiser->running_avg_y[ALTREF_FRAME],
+                      denoiser->running_avg_y[INTRA_FRAME]);
+  }
+  if (refresh_golden_frame) {
+    swap_frame_buffer(denoiser->running_avg_y[GOLDEN_FRAME],
+                      denoiser->running_avg_y[INTRA_FRAME]);
+  }
+  if (refresh_last_frame) {
+    swap_frame_buffer(denoiser->running_avg_y[LAST_FRAME],
+                      denoiser->running_avg_y[INTRA_FRAME]);
+  }
 }
 
 void vp9_denoiser_reset_frame_stats(PICK_MODE_CONTEXT *ctx) {
   ctx->zeromv_sse = UINT_MAX;
-  ctx->newmv_sse = UINT_MAX;
+  // This should be initialized as zero since mode search stage might skip
+  // NEWMV mode if inferred motion vector modes provide sufficiently good
+  // prediction quality.
+  ctx->newmv_sse = 0;
 }
 
 void vp9_denoiser_update_frame_stats(MB_MODE_INFO *mbmi, unsigned int sse,
@@ -425,6 +432,7 @@
 #endif
                        int border) {
   int i, fail;
+  const int legacy_byte_alignment = 0;
   assert(denoiser != NULL);
 
   for (i = 0; i < MAX_REF_FRAMES; ++i) {
@@ -433,7 +441,7 @@
 #if CONFIG_VP9_HIGHBITDEPTH
                                   use_highbitdepth,
 #endif
-                                  border);
+                                  border, legacy_byte_alignment);
     if (fail) {
       vp9_denoiser_free(denoiser);
       return 1;
@@ -448,7 +456,7 @@
 #if CONFIG_VP9_HIGHBITDEPTH
                                 use_highbitdepth,
 #endif
-                                border);
+                                border, legacy_byte_alignment);
   if (fail) {
     vp9_denoiser_free(denoiser);
     return 1;
@@ -457,6 +465,7 @@
   make_grayscale(&denoiser->running_avg_y[i]);
 #endif
   denoiser->increase_denoising = 0;
+  denoiser->frame_buffer_initialized = 1;
 
   return 0;
 }
@@ -463,6 +472,7 @@
 
 void vp9_denoiser_free(VP9_DENOISER *denoiser) {
   int i;
+  denoiser->frame_buffer_initialized = 0;
   if (denoiser == NULL) {
     return;
   }
@@ -482,15 +492,13 @@
   uint8_t *u = yuv->u_buffer;
   uint8_t *v = yuv->v_buffer;
 
-  // The '/2's are there because we have a 440 buffer, but we want to output
-  // 420.
-  for (r = 0; r < yuv->uv_height / 2; ++r) {
-    for (c = 0; c < yuv->uv_width / 2; ++c) {
+  for (r = 0; r < yuv->uv_height; ++r) {
+    for (c = 0; c < yuv->uv_width; ++c) {
       u[c] = UINT8_MAX / 2;
       v[c] = UINT8_MAX / 2;
     }
-    u += yuv->uv_stride + yuv->uv_width / 2;
-    v += yuv->uv_stride + yuv->uv_width / 2;
+    u += yuv->uv_stride;
+    v += yuv->uv_stride;
   }
 }
 #endif