Decrease AC chroma on 4:2:2 content with `tune=ssimulacra2`

Increase U and V chroma qindex (i.e. lower chroma quality) on AC
coefficients when encoding content that uses 4:2:2 chroma subsampling.

SSIMULACRA 2 appears to have trouble correctly scoring 4:2:2, so
the chroma delta-q value (6) was chosen so encoded chroma channels
are roughly the same size of their 4:2:0 and 4:4:4 encoded
counterparts, with a subjective validation check to ensure balanced
luma/chroma allocation.

Approximate BD-Rate gains over no AC chroma decrease - cpu-used=6
(Daala's subset1, for completeness):
- SSIMULACRA2 60: -0.3%
- SSIMULACRA2 70: -0.3%
- SSIMULACRA2 80: -0.19%
- SSIMULACRA2 90: N/A (unreacheable)

Bug: aomedia:375221136
Change-Id: I61e7d3d5381b794697db5c1c2ad267f192adc1d6

av1/encoder/av1_quantize.c

diff --git a/av1/encoder/av1_quantize.c b/av1/encoder/av1_quantize.c
index b572923..3ddedce 100644
--- a/av1/encoder/av1_quantize.c
+++ b/av1/encoder/av1_quantize.c
@@ -891,6 +891,25 @@
         // finding a linear equation that fits the convex hull.
         chroma_dc_delta_q = -clamp((quant_params->base_qindex / 2) - 14, 0, 16);
         chroma_ac_delta_q = chroma_dc_delta_q;
+      } else if (cm->seq_params->subsampling_x == 1 &&
+                 cm->seq_params->subsampling_y == 0) {
+        // 4:2:2 subsampling: Constant chroma AC delta_q increase (i.e. improved
+        // luma quality relative to chroma) with gradual ramp-down for very low
+        // qindexes.
+        // SSIMULACRA 2 appears to have some issues correctly scoring 4:2:2
+        // material. Solely optimizing for maximum scores suggests a chroma AC
+        // delta_q of 12 is the most efficient. However, visual inspection on
+        // difficult-to-encode material resulted in chroma quality degrading too
+        // much relative to luma, and chroma channels ending up being too small
+        // compared to equivalent 4:4:4 or 4:2:0 encodes.
+        // A chroma AC delta_q of 6 was selected because encoded chroma channels
+        // have a much closer size to 4:4:4 and 4:2:0 encodes, and have more
+        // favorable visual quality characteristics.
+        // The ramp-down of chroma decrease was put into place to match 4:2:0
+        // and 4:4:4 behavior. There were no special considerations on
+        // SSIMULACRA 2 scores.
+        chroma_dc_delta_q = 0;
+        chroma_ac_delta_q = clamp((quant_params->base_qindex / 2), 0, 6);
       } else if (cm->seq_params->subsampling_x == 0 &&
                  cm->seq_params->subsampling_y == 0) {
         // 4:4:4 subsampling: Constant chroma AC delta_q increase (i.e. improved