Add 2x2 adst and dct units

Change-Id: I6d3e97385a9d1c82fd3b7cbc8fcc02b3f682e955

aom_dsp/txfm_common.h

@@ -58,11 +58,16 @@ static const tran_high_t cospi_29_64 = 2404;
 static const tran_high_t cospi_30_64 = 1606;
 static const tran_high_t cospi_31_64 = 804;
 
-//  16384 * sqrt(2) * sin(kPi/9) * 2 / 3
+//  16384 * sqrt(2) * sin(k * Pi / 9) * 2 / 3
 static const tran_high_t sinpi_1_9 = 5283;
 static const tran_high_t sinpi_2_9 = 9929;
 static const tran_high_t sinpi_3_9 = 13377;
 static const tran_high_t sinpi_4_9 = 15212;
+#if CONFIG_CB4X4
+// 16384 * sqrt(2/5) * 2 * sin(k * Pi / 5)
+static const tran_high_t sinpi_1_5 = 12181;
+static const tran_high_t sinpi_2_5 = 19710;
+#endif
 
 // 16384 * sqrt(2)
 static const tran_high_t Sqrt2 = 23170;

av1/common/av1_inv_txfm.c

@@ -242,6 +242,20 @@ void av1_idct8x8_1_add_c(const tran_low_t *input, uint8_t *dest, int stride) {
   }
 }
 
+#if CONFIG_CB4X4
+void av1_iadst2_c(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1;
+  const tran_low_t x0 = input[0];
+  const tran_low_t x1 = input[1];
+
+  s0 = sinpi_1_5 * x0 + sinpi_2_5 * x1;
+  s1 = sinpi_2_5 * x0 - sinpi_1_5 * x1;
+
+  output[0] = WRAPLOW(dct_const_round_shift(s0), 8);
+  output[1] = WRAPLOW(dct_const_round_shift(s1), 8);
+}
+#endif
+
 void av1_iadst4_c(const tran_low_t *input, tran_low_t *output) {
   tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
 

av1/encoder/dct.c

@@ -38,6 +38,20 @@ static INLINE void range_check(const tran_low_t *input, const int size,
 #endif
 }
 
+#if CONFIG_CB4X4
+static void fdct2(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1;
+  const tran_low_t x0 = input[0];
+  const tran_low_t x1 = input[1];
+
+  s0 = (tran_high_t)x0 + x1;
+  s1 = (tran_high_t)x0 - x1;
+
+  output[0] = (tran_low_t)s0;
+  output[1] = (tran_low_t)s1;
+}
+#endif
+
 static void fdct4(const tran_low_t *input, tran_low_t *output) {
   tran_high_t temp;
   tran_low_t step[4];
@@ -694,6 +708,22 @@ static void fdct32(const tran_low_t *input, tran_low_t *output) {
 }
 #endif  // CONFIG_EXT_TX
 
+#if CONFIG_CB4X4
+static void fadst2(const tran_low_t *input, tran_low_t *output) {
+  tran_high_t s0, s1;
+  tran_low_t x0, x1;
+
+  x0 = input[0];
+  x1 = input[1];
+
+  s0 = sinpi_1_5 * x0 + sinpi_2_5 * x1;
+  s1 = sinpi_2_5 * x0 - sinpi_1_5 * x1;
+
+  output[0] = (tran_low_t)fdct_round_shift(s0);
+  output[1] = (tran_low_t)fdct_round_shift(s1);
+}
+#endif
+
 static void fadst4(const tran_low_t *input, tran_low_t *output) {
   tran_high_t x0, x1, x2, x3;
   tran_high_t s0, s1, s2, s3, s4, s5, s6, s7;
@@ -1101,6 +1131,12 @@ static void maybe_flip_input(const int16_t **src, int *src_stride, int l,
 }
 #endif  // CONFIG_EXT_TX
 
+#if CONFIG_CB4X4
+static const transform_2d FHT_2[] = {
+  { fdct2, fdct2 }, { fadst2, fdct2 }, { fdct2, fadst2 }, { fadst2, fadst2 },
+};
+#endif
+
 static const transform_2d FHT_4[] = {
   { fdct4, fdct4 },    // DCT_DCT  = 0
   { fadst4, fdct4 },   // ADST_DCT = 1