vp9_reconintra_neon: add d45 8x8

based on ssse3 implementation

~91% faster over 20M pixels

Change-Id: I6d743a53352c2d6de0efe7899d7996e8b0f7fa29

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index 352cde2..7b4c435 100644 (file)
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -264,8 +264,8 @@ INTRA_PRED_TEST(DSPR2, TestIntraPred8, vp9_dc_predictor_8x8_dspr2, NULL, NULL,
 INTRA_PRED_TEST(NEON, TestIntraPred8, vp9_dc_predictor_8x8_neon,
                 vp9_dc_left_predictor_8x8_neon, vp9_dc_top_predictor_8x8_neon,
                 vp9_dc_128_predictor_8x8_neon, vp9_v_predictor_8x8_neon,
-                vp9_h_predictor_8x8_neon, NULL, NULL, NULL, NULL, NULL, NULL,
-                vp9_tm_predictor_8x8_neon)
+                vp9_h_predictor_8x8_neon, vp9_d45_predictor_8x8_neon, NULL,
+                NULL, NULL, NULL, NULL, vp9_tm_predictor_8x8_neon)
 
 #endif  // HAVE_NEON
 

diff --git a/vp9/common/arm/neon/vp9_reconintra_neon.c b/vp9/common/arm/neon/vp9_reconintra_neon.c
index 13c46a5..cfd5905 100644 (file)
--- a/vp9/common/arm/neon/vp9_reconintra_neon.c
+++ b/vp9/common/arm/neon/vp9_reconintra_neon.c
@@ -338,6 +338,26 @@ void vp9_d45_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
   dst[3 * stride + 3] = above[7];
 }
 
+void vp9_d45_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
+                                const uint8_t *above, const uint8_t *left) {
+  static const uint8_t shuffle1[8] = { 1, 2, 3, 4, 5, 6, 7, 7 };
+  static const uint8_t shuffle2[8] = { 2, 3, 4, 5, 6, 7, 7, 7 };
+  const uint8x8_t sh_12345677 = vld1_u8(shuffle1);
+  const uint8x8_t sh_23456777 = vld1_u8(shuffle2);
+  const uint8x8_t A0 = vld1_u8(above);  // top row
+  const uint8x8_t A1 = vtbl1_u8(A0, sh_12345677);
+  const uint8x8_t A2 = vtbl1_u8(A0, sh_23456777);
+  const uint8x8_t avg1 = vhadd_u8(A0, A2);
+  uint8x8_t row = vrhadd_u8(avg1, A1);
+  int i;
+  (void)left;
+  for (i = 0; i < 7; ++i) {
+    vst1_u8(dst + i * stride, row);
+    row = vtbl1_u8(row, sh_12345677);
+  }
+  vst1_u8(dst + i * stride, row);
+}
+
 // -----------------------------------------------------------------------------
 
 void vp9_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,

diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 27cd3d0..1f265fe 100644 (file)
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -99,7 +99,7 @@ add_proto qw/void vp9_d207_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, co
 specialize qw/vp9_d207_predictor_8x8/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d45_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_d45_predictor_8x8/, "$ssse3_x86inc";
+specialize qw/vp9_d45_predictor_8x8 neon/, "$ssse3_x86inc";
 
 add_proto qw/void vp9_d63_predictor_8x8/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
 specialize qw/vp9_d63_predictor_8x8/, "$ssse3_x86inc";