Enable rectangular transforms for UV

Uses an array to map block sizes, y tx sizes, and subsampling
factors to various transform sizes for UV.

Results improve by 0.1-0.2%

Change-Id: Icb58fd96bc7c01a72cbf1332fe2be4d55a0feedc
diff --git a/av1/common/blockd.h b/av1/common/blockd.h
index 8fcaa60..f035bb5 100644
--- a/av1/common/blockd.h
+++ b/av1/common/blockd.h
@@ -684,25 +684,18 @@
 
 void av1_setup_block_planes(MACROBLOCKD *xd, int ss_x, int ss_y);
 
-static INLINE TX_SIZE get_uv_tx_size_impl(TX_SIZE y_tx_size, BLOCK_SIZE bsize,
-                                          int xss, int yss) {
-  if (bsize < BLOCK_8X8) {
-    return TX_4X4;
-  } else {
-    const BLOCK_SIZE plane_bsize = ss_size_lookup[bsize][xss][yss];
-    return AOMMIN(txsize_sqr_map[y_tx_size], max_txsize_lookup[plane_bsize]);
-  }
-}
-
 static INLINE TX_SIZE get_uv_tx_size(const MB_MODE_INFO *mbmi,
                                      const struct macroblockd_plane *pd) {
+  TX_SIZE uv_txsize;
 #if CONFIG_SUPERTX
   if (supertx_enabled(mbmi))
     return uvsupertx_size_lookup[txsize_sqr_map[mbmi->tx_size]]
                                 [pd->subsampling_x][pd->subsampling_y];
 #endif  // CONFIG_SUPERTX
-  return get_uv_tx_size_impl(mbmi->tx_size, mbmi->sb_type, pd->subsampling_x,
-                             pd->subsampling_y);
+  uv_txsize = uv_txsize_lookup[mbmi->sb_type][mbmi->tx_size][pd->subsampling_x]
+                              [pd->subsampling_y];
+  assert(uv_txsize != TX_INVALID);
+  return uv_txsize;
 }
 
 static INLINE BLOCK_SIZE
diff --git a/av1/common/common_data.h b/av1/common/common_data.h
index f86e1f7..450d962 100644
--- a/av1/common/common_data.h
+++ b/av1/common/common_data.h
@@ -518,6 +518,270 @@
 #endif  // CONFIG_EXT_PARTITION
 };
 
+#define USE_UV_RECT_TX 1
+static const TX_SIZE uv_txsize_lookup[BLOCK_SIZES][TX_SIZES_ALL][2][2] = {
+  //  ss_x == 0    ss_x == 0        ss_x == 1      ss_x == 1
+  //  ss_y == 0    ss_y == 1        ss_y == 0      ss_y == 1
+  {
+      // BLOCK_4X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#if CONFIG_EXT_TX
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_4X8
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#if CONFIG_EXT_TX
+#if CONFIG_RECT_TX && USE_UV_RECT_TX
+      { { TX_4X8, TX_4X4 }, { TX_4X4, TX_4X4 } },  // used
+#else
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },      // used
+#endif  // CONFIG_RECT_TX && USE_UV_RECT_TX
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_8X4
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+#if CONFIG_EXT_TX
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+#if CONFIG_RECT_TX && USE_UV_RECT_TX
+      { { TX_8X4, TX_4X4 }, { TX_4X4, TX_4X4 } },  // used
+#else
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },      // used
+#endif  // CONFIG_RECT_TX && USE_UV_RECT_TX
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_8X8
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_4X4, TX_4X4 } },
+#if CONFIG_EXT_TX
+      { { TX_4X8, TX_4X4 }, { TX_4X8, TX_4X4 } },
+      { { TX_8X4, TX_8X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X4 }, { TX_4X8, TX_4X4 } },
+      { { TX_8X8, TX_8X4 }, { TX_4X8, TX_4X4 } },
+      { { TX_8X8, TX_8X4 }, { TX_4X8, TX_4X4 } },
+      { { TX_8X8, TX_8X4 }, { TX_4X8, TX_4X4 } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_8X16
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_4X4, TX_4X4 } },
+#if CONFIG_EXT_TX
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_4X4, TX_4X4 } },
+#if CONFIG_RECT_TX && USE_UV_RECT_TX
+      { { TX_8X16, TX_8X8 }, { TX_4X8, TX_4X8 } },  // used
+#else
+      { { TX_8X8, TX_8X8 }, { TX_4X4, TX_4X4 } },      // used
+#endif  // CONFIG_RECT_TX && USE_UV_RECT_TX
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_16X8
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_4X4 }, { TX_8X8, TX_4X4 } },
+      { { TX_8X8, TX_4X4 }, { TX_8X8, TX_8X8 } },
+      { { TX_8X8, TX_4X4 }, { TX_8X8, TX_8X8 } },
+#if CONFIG_EXT_TX
+      { { TX_4X8, TX_4X4 }, { TX_4X8, TX_4X4 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+#if CONFIG_RECT_TX && USE_UV_RECT_TX
+      { { TX_16X8, TX_8X4 }, { TX_8X8, TX_8X4 } },  // used
+#else
+      { { TX_8X8, TX_4X4 }, { TX_8X8, TX_4X4 } },      // used
+#endif  // CONFIG_RECT_TX && USE_UV_RECT_TX
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_16X16
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_8X8 }, { TX_8X8, TX_8X8 } },
+#if CONFIG_EXT_TX
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X8 }, { TX_8X16, TX_8X8 } },
+      { { TX_16X8, TX_16X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X8 }, { TX_8X16, TX_8X8 } },
+      { { TX_16X16, TX_16X8 }, { TX_8X16, TX_8X8 } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_16X32
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_8X8, TX_8X8 } },
+#if CONFIG_EXT_TX
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_8X8, TX_8X8 } },
+#if CONFIG_RECT_TX && USE_UV_RECT_TX
+      { { TX_16X32, TX_16X16 }, { TX_8X16, TX_8X16 } },  // used
+#else
+      { { TX_16X16, TX_16X16 }, { TX_8X8, TX_8X8 } },  // used
+#endif  // CONFIG_RECT_TX && USE_UV_RECT_TX
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_32X16
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_8X8 }, { TX_16X16, TX_8X8 } },
+      { { TX_16X16, TX_8X8 }, { TX_16X16, TX_8X8 } },
+#if CONFIG_EXT_TX
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X8 }, { TX_8X16, TX_8X8 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_INVALID, TX_INVALID }, { TX_INVALID, TX_INVALID } },
+#if CONFIG_RECT_TX && USE_UV_RECT_TX
+      { { TX_32X16, TX_16X8 }, { TX_16X16, TX_16X8 } },  // used
+#else
+      { { TX_16X16, TX_8X8 }, { TX_16X16, TX_8X8 } },  // used
+#endif  // CONFIG_RECT_TX && USE_UV_RECT_TX
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_32X32
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_16X16 }, { TX_16X16, TX_16X16 } },
+#if CONFIG_EXT_TX
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X16 }, { TX_16X32, TX_16X16 } },
+      { { TX_32X16, TX_32X16 }, { TX_16X16, TX_16X16 } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_32X64
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_32X32 }, { TX_16X16, TX_16X16 } },
+#if CONFIG_EXT_TX
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X32 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X16, TX_32X16 }, { TX_16X16, TX_16X16 } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_64X32
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_16X16 }, { TX_32X32, TX_16X16 } },
+#if CONFIG_EXT_TX
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X16 }, { TX_16X32, TX_16X16 } },
+      { { TX_32X16, TX_16X16 }, { TX_32X16, TX_16X16 } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_64X64
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_32X32 }, { TX_32X32, TX_32X32 } },
+#if CONFIG_EXT_PARTITION
+  },
+  {
+      // BLOCK_64X128
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_32X32 }, { TX_32X32, TX_32X32 } },
+#if CONFIG_EXT_TX
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X32 }, { TX_16X32, TX_16X32 } },
+      { { TX_32X16, TX_32X16 }, { TX_32X16, TX_32X16 } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_128X64
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_32X32 }, { TX_32X32, TX_32X32 } },
+#if CONFIG_EXT_TX
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X32 }, { TX_16X32, TX_16X32 } },
+      { { TX_32X16, TX_32X16 }, { TX_32X16, TX_32X16 } },
+#endif  // CONFIG_EXT_TX
+  },
+  {
+      // BLOCK_128X128
+      { { TX_4X4, TX_4X4 }, { TX_4X4, TX_4X4 } },
+      { { TX_8X8, TX_8X8 }, { TX_8X8, TX_8X8 } },
+      { { TX_16X16, TX_16X16 }, { TX_16X16, TX_16X16 } },
+      { { TX_32X32, TX_32X32 }, { TX_32X32, TX_32X32 } },
+#if CONFIG_EXT_TX
+      { { TX_4X8, TX_4X8 }, { TX_4X8, TX_4X8 } },
+      { { TX_8X4, TX_8X4 }, { TX_8X4, TX_8X4 } },
+      { { TX_8X16, TX_8X16 }, { TX_8X16, TX_8X16 } },
+      { { TX_16X8, TX_16X8 }, { TX_16X8, TX_16X8 } },
+      { { TX_16X32, TX_16X32 }, { TX_16X32, TX_16X32 } },
+      { { TX_32X16, TX_32X16 }, { TX_32X16, TX_32X16 } },
+#endif  // CONFIG_EXT_TX
+#endif  // CONFIG_EXT_PARTITION
+  },
+};
+
 // Generates 4 bit field in which each bit set to 1 represents
 // a blocksize partition  1111 means we split 64x64, 32x32, 16x16
 // and 8x8.  1000 means we just split the 64x64 to 32x32
diff --git a/av1/common/enums.h b/av1/common/enums.h
index 9f1bbaa..b1ac2a0 100644
--- a/av1/common/enums.h
+++ b/av1/common/enums.h
@@ -137,6 +137,7 @@
 #define TX_16X16 ((TX_SIZE)2)  // 16x16 transform
 #define TX_32X32 ((TX_SIZE)3)  // 32x32 transform
 #define TX_SIZES ((TX_SIZE)4)
+#define TX_INVALID ((TX_SIZE)255)  // Invalid transform size
 
 #if CONFIG_EXT_TX
 #define TX_4X8 ((TX_SIZE)4)         // 4x8 transform
diff --git a/av1/common/loopfilter.c b/av1/common/loopfilter.c
index 726271e..7086be8 100644
--- a/av1/common/loopfilter.c
+++ b/av1/common/loopfilter.c
@@ -713,7 +713,7 @@
   // rectangular transfroms are used with the EXT_TX expt.
   const TX_SIZE tx_size_y = txsize_sqr_up_map[mbmi->tx_size];
   const TX_SIZE tx_size_uv =
-      get_uv_tx_size_impl(mbmi->tx_size, block_size, 1, 1);
+      txsize_sqr_up_map[uv_txsize_lookup[block_size][mbmi->tx_size][1][1]];
   const int filter_level = get_filter_level(lfi_n, mbmi);
   uint64_t *const left_y = &lfm->left_y[tx_size_y];
   uint64_t *const above_y = &lfm->above_y[tx_size_y];
@@ -1238,9 +1238,9 @@
       // Filter level can vary per MI
       if (!(lfl[r][c >> ss_x] = get_filter_level(&cm->lf_info, mbmi))) continue;
 
-      if (tx_size == TX_32X32)
+      if (txsize_sqr_up_map[tx_size] == TX_32X32)
         tx_size_mask = 3;
-      else if (tx_size == TX_16X16)
+      else if (txsize_sqr_up_map[tx_size] == TX_16X16)
         tx_size_mask = 1;
       else
         tx_size_mask = 0;
@@ -1249,8 +1249,8 @@
       if (is_inter_block(mbmi) && !mbmi->skip)
         tx_size =
             (plane->plane_type == PLANE_TYPE_UV)
-                ? get_uv_tx_size_impl(mbmi->inter_tx_size[blk_row][blk_col],
-                                      sb_type, ss_x, ss_y)
+                ? uv_txsize_lookup[sb_type][mbmi->inter_tx_size
+                                                [blk_row][blk_col]][ss_x][ss_y]
                 : mbmi->inter_tx_size[blk_row][blk_col];
 
 #if CONFIG_EXT_TX && CONFIG_RECT_TX