Split fine directional prediction search into 2 passes

1. Search matching deltas for luma and chroma
2. Refine chroma delta for best luma delta

Iterations are the sum of luma/chroma deltas rather than the product.

On thaloundeskmtg360p_60f.y4m at speed 5 and quantizer 100, this yields
37% overall speed-up. This clip has 2 keyframes out of 60 frames and
they are ~40% directionally predicted.
Expected average speedup on objective-1-fast is ~16%.

Author: barrbrain

src/rdo.rs

@@ -1308,30 +1308,16 @@ fn intra_frame_rdo_mode_decision<T: Pixel>(
     && bsize >= BlockSize::BLOCK_8X8
   {
     // Find the best angle delta for the current best prediction mode
-    let luma_angle_delta_count = best.pred_mode_luma.angle_delta_count();
-    let chroma_angle_delta_count = best.pred_mode_chroma.angle_delta_count();
-
-    'luma_loop: for i in 0..luma_angle_delta_count {
-      for j in 0..chroma_angle_delta_count {
-        let mvs = [MotionVector::default(); 2];
-        let ref_frames = [INTRA_FRAME, NONE_FRAME];
-        let mut mode_set_chroma = ArrayVec::<[_; 2]>::new();
-        mode_set_chroma.push(best.pred_mode_chroma);
-
-        let angle_delta_y: i8 = if luma_angle_delta_count == 1 {
-          0
-        } else {
-          i - MAX_ANGLE_DELTA as i8
-        };
-        let angle_delta_uv: i8 = if chroma_angle_delta_count == 1 {
-          0
-        } else {
-          j - MAX_ANGLE_DELTA as i8
-        };
-        if luma_angle_delta_count == 1 && chroma_angle_delta_count == 1 {
-          break 'luma_loop;
-        }
+    let luma_deltas = best.pred_mode_luma.angle_delta_count();
+    let chroma_deltas = best.pred_mode_chroma.angle_delta_count();
 
+    let mvs = [MotionVector::default(); 2];
+    let ref_frames = [INTRA_FRAME, NONE_FRAME];
+    let mode_set_chroma = [best.pred_mode_chroma];
+    let mv_stack = ArrayVec::<[_; 9]>::new();
+    let mut best_angle_delta_y = best.angle_delta.y;
+    let mut angle_delta_rdo = |y, uv| -> i8 {
+      if best.angle_delta.y != y || best.angle_delta.uv != uv {
         luma_chroma_mode_rdo(
           best.pred_mode_luma,
           fi,
@@ -1347,10 +1333,23 @@ fn intra_frame_rdo_mode_decision<T: Pixel>(
           &mode_set_chroma,
           true,
           0,
-          &ArrayVec::<[CandidateMV; 9]>::new(),
-          AngleDelta { y: angle_delta_y, uv: angle_delta_uv },
+          &mv_stack,
+          AngleDelta { y, uv },
         );
       }
+      best.angle_delta.y
+    };
+
+    for i in 0..luma_deltas {
+      let angle_delta_y =
+        if luma_deltas == 1 { 0 } else { i - MAX_ANGLE_DELTA as i8 };
+      let angle_delta_uv = if chroma_deltas == 1 { 0 } else { angle_delta_y };
+      best_angle_delta_y = angle_delta_rdo(angle_delta_y, angle_delta_uv);
+    }
+    for j in 0..chroma_deltas {
+      let angle_delta_uv =
+        if chroma_deltas == 1 { 0 } else { j - MAX_ANGLE_DELTA as i8 };
+      angle_delta_rdo(best_angle_delta_y, angle_delta_uv);
     }
   }