vp9_reconintra_neon: add DC 32x32 predictors

~84-85% faster over 20M pixels

Change-Id: Ia67a7f4a342bf7b0a9280e05c25d81a774d90469

diff --git a/test/test_intra_pred_speed.cc b/test/test_intra_pred_speed.cc
index 7e6d326f1b328728d22345ef92d54406d753e071..e387857ae4541c8ac34dfd3d919929adb7b4ad8f 100644 (file)
--- a/test/test_intra_pred_speed.cc
+++ b/test/test_intra_pred_speed.cc
@@ -362,9 +362,12 @@ INTRA_PRED_TEST(SSSE3, TestIntraPred32, NULL, NULL, NULL, NULL, NULL,
 #endif  // HAVE_SSSE3
 
 #if HAVE_NEON
-INTRA_PRED_TEST(NEON, TestIntraPred32, NULL, NULL, NULL, NULL,
-                vp9_v_predictor_32x32_neon, vp9_h_predictor_32x32_neon, NULL,
-                NULL, NULL, NULL, NULL, NULL, vp9_tm_predictor_32x32_neon)
+INTRA_PRED_TEST(NEON, TestIntraPred32, vp9_dc_predictor_32x32_neon,
+                vp9_dc_left_predictor_32x32_neon,
+                vp9_dc_top_predictor_32x32_neon,
+                vp9_dc_128_predictor_32x32_neon, vp9_v_predictor_32x32_neon,
+                vp9_h_predictor_32x32_neon, NULL, NULL, NULL, NULL, NULL, NULL,
+                vp9_tm_predictor_32x32_neon)
 #endif  // HAVE_NEON
 
 #if HAVE_MSA

diff --git a/vp9/common/arm/neon/vp9_reconintra_neon.c b/vp9/common/arm/neon/vp9_reconintra_neon.c
index 499c42ac3aca9145b3a2921000b0cf17b4a91057..a053e4f10a69271ffdd94903a9f8c1f0b7203d00 100644 (file)
--- a/vp9/common/arm/neon/vp9_reconintra_neon.c
+++ b/vp9/common/arm/neon/vp9_reconintra_neon.c
@@ -161,6 +161,89 @@ void vp9_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
   dc_16x16(dst, stride, NULL, NULL, 0, 0);
 }
 
+//------------------------------------------------------------------------------
+// DC 32x32
+
+// 'do_above' and 'do_left' facilitate branch removal when inlined.
+static INLINE void dc_32x32(uint8_t *dst, ptrdiff_t stride,
+                            const uint8_t *above, const uint8_t *left,
+                            int do_above, int do_left) {
+  uint16x8_t sum_top;
+  uint16x8_t sum_left;
+  uint8x8_t dc0;
+
+  if (do_above) {
+    const uint8x16_t A0 = vld1q_u8(above);  // top row
+    const uint8x16_t A1 = vld1q_u8(above + 16);
+    const uint16x8_t p0 = vpaddlq_u8(A0);  // cascading summation of the top
+    const uint16x8_t p1 = vpaddlq_u8(A1);
+    const uint16x8_t p2 = vaddq_u16(p0, p1);
+    const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
+    const uint16x4_t p4 = vpadd_u16(p3, p3);
+    const uint16x4_t p5 = vpadd_u16(p4, p4);
+    sum_top = vcombine_u16(p5, p5);
+  }
+
+  if (do_left) {
+    const uint8x16_t L0 = vld1q_u8(left);  // left row
+    const uint8x16_t L1 = vld1q_u8(left + 16);
+    const uint16x8_t p0 = vpaddlq_u8(L0);  // cascading summation of the left
+    const uint16x8_t p1 = vpaddlq_u8(L1);
+    const uint16x8_t p2 = vaddq_u16(p0, p1);
+    const uint16x4_t p3 = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
+    const uint16x4_t p4 = vpadd_u16(p3, p3);
+    const uint16x4_t p5 = vpadd_u16(p4, p4);
+    sum_left = vcombine_u16(p5, p5);
+  }
+
+  if (do_above && do_left) {
+    const uint16x8_t sum = vaddq_u16(sum_left, sum_top);
+    dc0 = vrshrn_n_u16(sum, 6);
+  } else if (do_above) {
+    dc0 = vrshrn_n_u16(sum_top, 5);
+  } else if (do_left) {
+    dc0 = vrshrn_n_u16(sum_left, 5);
+  } else {
+    dc0 = vdup_n_u8(0x80);
+  }
+
+  {
+    const uint8x16_t dc = vdupq_lane_u8(dc0, 0);
+    int i;
+    for (i = 0; i < 32; ++i) {
+      vst1q_u8(dst + i * stride, dc);
+      vst1q_u8(dst + i * stride + 16, dc);
+    }
+  }
+}
+
+void vp9_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                 const uint8_t *above, const uint8_t *left) {
+  dc_32x32(dst, stride, above, left, 1, 1);
+}
+
+void vp9_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                      const uint8_t *above,
+                                      const uint8_t *left) {
+  (void)above;
+  dc_32x32(dst, stride, NULL, left, 0, 1);
+}
+
+void vp9_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)left;
+  dc_32x32(dst, stride, above, NULL, 1, 0);
+}
+
+void vp9_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
+                                     const uint8_t *above,
+                                     const uint8_t *left) {
+  (void)above;
+  (void)left;
+  dc_32x32(dst, stride, NULL, NULL, 0, 0);
+}
+
 #if !HAVE_NEON_ASM
 
 void vp9_v_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 5326087ea4094fa6b8e455aaea4e89d28a43d0d3..bc1a3c913f7763281acb9d8c5e4d587eacce82a7 100644 (file)
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -201,16 +201,16 @@ add_proto qw/void vp9_tm_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, co
 specialize qw/vp9_tm_predictor_32x32 neon msa/, "$sse2_x86_64";
 
 add_proto qw/void vp9_dc_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_predictor_32x32 msa/, "$sse2_x86inc";
+specialize qw/vp9_dc_predictor_32x32 msa neon/, "$sse2_x86inc";
 
 add_proto qw/void vp9_dc_top_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_top_predictor_32x32 msa/, "$sse2_x86inc";
+specialize qw/vp9_dc_top_predictor_32x32 msa neon/, "$sse2_x86inc";
 
 add_proto qw/void vp9_dc_left_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_left_predictor_32x32 msa/, "$sse2_x86inc";
+specialize qw/vp9_dc_left_predictor_32x32 msa neon/, "$sse2_x86inc";
 
 add_proto qw/void vp9_dc_128_predictor_32x32/, "uint8_t *dst, ptrdiff_t y_stride, const uint8_t *above, const uint8_t *left";
-specialize qw/vp9_dc_128_predictor_32x32 msa/, "$sse2_x86inc";
+specialize qw/vp9_dc_128_predictor_32x32 msa neon/, "$sse2_x86inc";
 
 #
 # Loopfilter