vp9_reconintra_neon: add d45 16x16

~90% faster over 20M pixels

Change-Id: I92d80f66e91e0a870a672cfb5dd29bf1a17cb11a

test/test_intra_pred_speed.cc

@@ -316,8 +316,8 @@ INTRA_PRED_TEST(NEON, TestIntraPred16, vp9_dc_predictor_16x16_neon,
                 vp9_dc_left_predictor_16x16_neon,
                 vp9_dc_top_predictor_16x16_neon,
                 vp9_dc_128_predictor_16x16_neon, vp9_v_predictor_16x16_neon,
-                vp9_h_predictor_16x16_neon, NULL, NULL, NULL, NULL, NULL, NULL,
-                vp9_tm_predictor_16x16_neon)
+                vp9_h_predictor_16x16_neon, vp9_d45_predictor_16x16_neon, NULL,
+                NULL, NULL, NULL, NULL, vp9_tm_predictor_16x16_neon)
 #endif  // HAVE_NEON
 
 #if HAVE_MSA

vp9/common/arm/neon/vp9_reconintra_neon.c

@@ -358,6 +358,23 @@ void vp9_d45_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
   vst1_u8(dst + i * stride, row);
 }
 
+void vp9_d45_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
+                                  const uint8_t *above, const uint8_t *left) {
+  const uint8x16_t A0 = vld1q_u8(above);  // top row
+  const uint8x16_t above_right = vld1q_dup_u8(above + 15);
+  const uint8x16_t A1 = vextq_u8(A0, above_right, 1);
+  const uint8x16_t A2 = vextq_u8(A0, above_right, 2);
+  const uint8x16_t avg1 = vhaddq_u8(A0, A2);
+  uint8x16_t row = vrhaddq_u8(avg1, A1);
+  int i;
+  (void)left;
+  for (i = 0; i < 15; ++i) {
+    vst1q_u8(dst + i * stride, row);
+    row = vextq_u8(row, above_right, 1);
+  }
+  vst1q_u8(dst + i * stride, row);
+}
+
 // -----------------------------------------------------------------------------
 
 void vp9_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,

vp9/common/vp9_rtcd_defs.pl

@@ -138,7 +138,7 @@ add_proto qw/void vp9_d207_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride,
 specialize qw/vp9_d207_predictor_16x16/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d45_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_d45_predictor_16x16/, "$ssse3_x86inc";
+specialize qw/vp9_d45_predictor_16x16 neon/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d63_predictor_16x16/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d63_predictor_16x16/, "$ssse3_x86inc";