Treat color space changes as config changes.

It turns out some decoding pipelines don't reliably support just
evicting picture buffers. We need to actual go through a full
reset of the underlying decoder to avoid issues.

This removes the kColorSpaceChange constant and adds a feature
flag controlling the usage of color space changes (enabled by
default).

R=andrescj, sandersd

Bug: 1474191
Change-Id: I36e1e8beb02274776043d0ebc115d1048993102f
Reviewed-on: https://chromium-review.googlesource.com/c/chromium/src/+/4919595
Commit-Queue: Dale Curtis <dalecurtis@chromium.org>
Reviewed-by: Andres Calderon Jaramillo <andrescj@chromium.org>
Reviewed-by: Dan Sanders <sandersd@chromium.org>
Cr-Commit-Position: refs/heads/main@{#1207936}

diff --git a/media/gpu/vp9_decoder.cc b/media/gpu/vp9_decoder.cc
index 8dfb1a85..9f71d15 100644
--- a/media/gpu/vp9_decoder.cc
+++ b/media/gpu/vp9_decoder.cc

@@ -291,6 +291,12 @@
 
     DCHECK(!new_pic_size.IsEmpty());
 
+    bool is_color_space_change = false;
+    if (base::FeatureList::IsEnabled(kAVDColorSpaceChanges)) {
+      is_color_space_change = new_color_space.IsSpecified() &&
+                              new_color_space != picture_color_space_;
+    }
+
     const bool is_pic_size_different = new_pic_size != pic_size_;
     const bool is_pic_size_larger = new_pic_size.width() > pic_size_.width() ||
                                     new_pic_size.height() > pic_size_.height();
@@ -298,13 +304,15 @@
         (ignore_resolution_changes_to_smaller_for_testing_
              ? is_pic_size_larger
              : is_pic_size_different) ||
-        new_profile != profile_ || curr_frame_hdr_->bit_depth != bit_depth_;
+        new_profile != profile_ || curr_frame_hdr_->bit_depth != bit_depth_ ||
+        is_color_space_change;
 
     if (is_new_configuration_different_enough) {
       DVLOG(1) << "New profile: " << GetProfileName(new_profile)
                << ", new resolution: " << new_pic_size.ToString()
                << ", new bit depth: "
-               << base::strict_cast<int>(curr_frame_hdr_->bit_depth);
+               << base::strict_cast<int>(curr_frame_hdr_->bit_depth)
+               << ", new color space: " << new_color_space.ToString();
 
       if (!curr_frame_hdr_->IsKeyframe() &&
           !(curr_frame_hdr_->IsIntra() && pic_size_.IsEmpty())) {
@@ -339,19 +347,6 @@
       return kConfigChange;
     }
 
-    if (new_color_space.IsSpecified() &&
-        new_color_space != picture_color_space_ &&
-        curr_frame_hdr_->IsKeyframe()) {
-      // TODO(posciak): This requires us to be on a keyframe (see above) and is
-      // required, because VDA clients expect all surfaces to be returned before
-      // they can cycle surface sets after receiving kConfigChange.
-      // This is only an implementation detail of VDAs and can be improved.
-      ref_frames_.Clear();
-
-      picture_color_space_ = new_color_space;
-      return kColorSpaceChange;
-    }
-
     scoped_refptr<VP9Picture> pic = accelerator_->CreateVP9Picture();
     if (!pic) {
       return kRanOutOfSurfaces;