Add SSE4 version of new apply_temporal_filter

This adds a preliminary version of vp9_apply_temporal_filter in SSE4.1.
This patch merely adds the function and does not enable it yet.

Speed Up:
         | ss_x=1 | ss_x=0 |
  ss_y=1 | 19.80X | 19.04X |
  ss_y=0 | 21.09X | 20.21X |

BUG=webm:1591

Change-Id: If590f1ccf1d0c6c3b47410541d54f2ce37d8305b

6 files changed, 1117 insertions, 46 deletions

diff --git a/vp9/common/vp9_rtcd_defs.pl b/vp9/common/vp9_rtcd_defs.pl
index 3102b08a7..7e5e3c92c 100644
--- a/vp9/common/vp9_rtcd_defs.pl
+++ b/vp9/common/vp9_rtcd_defs.pl
@@ -188,6 +188,7 @@ add_proto qw/void vp9_temporal_filter_apply/, "const uint8_t *frame1, unsigned i
 specialize qw/vp9_temporal_filter_apply sse4_1/;
 
 add_proto qw/void vp9_apply_temporal_filter/, "const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre, int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src, int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre, int uv_pre_stride, unsigned int block_width, unsigned int block_height, int ss_x, int ss_y, int strength, const int *const blk_fw, int use_32x32, uint32_t *y_accumulator, uint16_t *y_count, uint32_t *u_accumulator, uint16_t *u_count, uint32_t *v_accumulator, uint16_t *v_count";
+specialize qw/vp9_apply_temporal_filter sse4_1/;
 }
 
 if (vpx_config("CONFIG_VP9_HIGHBITDEPTH") eq "yes") {
diff --git a/vp9/encoder/vp9_temporal_filter.c b/vp9/encoder/vp9_temporal_filter.c
index 7d2701355..bc8619bbe 100644
--- a/vp9/encoder/vp9_temporal_filter.c
+++ b/vp9/encoder/vp9_temporal_filter.c
@@ -34,6 +34,9 @@
 #include "vpx_scale/vpx_scale.h"
 
 static int fixed_divide[512];
+static unsigned int index_mult[14] = {
+  0, 0, 0, 0, 49152, 39322, 32768, 28087, 24576, 21846, 19661, 17874, 0, 15124
+};
 
 static void temporal_filter_predictors_mb_c(
     MACROBLOCKD *xd, uint8_t *y_mb_ptr, uint8_t *u_mb_ptr, uint8_t *v_mb_ptr,
@@ -184,7 +187,12 @@ void vp9_temporal_filter_init(void) {
 
 static INLINE int mod_index(int sum_dist, int index, int rounding, int strength,
                             int filter_weight) {
-  int mod = (sum_dist * 3) / index;
+  int mod;
+
+  assert(index >= 0 && index <= 13);
+  assert(index_mult[index] != 0);
+
+  mod = (clamp(sum_dist, 0, UINT16_MAX) * index_mult[index]) >> 16;
   mod += rounding;
   mod >>= strength;
 
diff --git a/vp9/encoder/vp9_temporal_filter.h b/vp9/encoder/vp9_temporal_filter.h
index f5fa194d1..553a46828 100644
--- a/vp9/encoder/vp9_temporal_filter.h
+++ b/vp9/encoder/vp9_temporal_filter.h
@@ -24,7 +24,7 @@ static const MV kZeroMv = { 0, 0 };
 #define BH_LOG2 5
 #define BW 32
 #define BW_LOG2 5
-#define BLK_PELS 1024  // Pixels in the block
+#define BLK_PELS ((BH) * (BW))  // Pixels in the block
 #define TF_SHIFT 2
 #define TF_ROUND 3
 #define THR_SHIFT 2
diff --git a/vp9/encoder/x86/temporal_filter_constants.h b/vp9/encoder/x86/temporal_filter_constants.h
new file mode 100644
index 000000000..20b7085a3
--- /dev/null
+++ b/vp9/encoder/x86/temporal_filter_constants.h
@@ -0,0 +1,232 @@
+/*
+ *  Copyright (c) 2019 The WebM project authors. All Rights Reserved.
+ *
+ *  Use of this source code is governed by a BSD-style license
+ *  that can be found in the LICENSE file in the root of the source
+ *  tree. An additional intellectual property rights grant can be found
+ *  in the file PATENTS.  All contributing project authors may
+ *  be found in the AUTHORS file in the root of the source tree.
+ */
+
+#include "./vpx_config.h"
+
+// Division using multiplication and shifting. The C implementation does:
+// modifier *= 3;
+// modifier /= index;
+// where 'modifier' is a set of summed values and 'index' is the number of
+// summed values.
+//
+// This equation works out to (m * 3) / i which reduces to:
+// m * 3/4
+// m * 1/2
+// m * 1/3
+//
+// By pairing the multiply with a down shift by 16 (_mm_mulhi_epu16):
+// m * C / 65536
+// we can create a C to replicate the division.
+//
+// m * 49152 / 65536 = m * 3/4
+// m * 32758 / 65536 = m * 1/2
+// m * 21846 / 65536 = m * 0.3333
+//
+// These are loaded using an instruction expecting int16_t values but are used
+// with _mm_mulhi_epu16(), which treats them as unsigned.
+#define NEIGHBOR_CONSTANT_4 (int16_t)49152
+#define NEIGHBOR_CONSTANT_5 (int16_t)39322
+#define NEIGHBOR_CONSTANT_6 (int16_t)32768
+#define NEIGHBOR_CONSTANT_7 (int16_t)28087
+#define NEIGHBOR_CONSTANT_8 (int16_t)24576
+#define NEIGHBOR_CONSTANT_9 (int16_t)21846
+#define NEIGHBOR_CONSTANT_10 (int16_t)19661
+#define NEIGHBOR_CONSTANT_11 (int16_t)17874
+#define NEIGHBOR_CONSTANT_13 (int16_t)15124
+
+DECLARE_ALIGNED(16, static const int16_t, LEFT_CORNER_NEIGHBORS_PLUS_1[8]) = {
+  NEIGHBOR_CONSTANT_5, NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7,
+  NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7,
+  NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7
+};
+
+DECLARE_ALIGNED(16, static const int16_t, RIGHT_CORNER_NEIGHBORS_PLUS_1[8]) = {
+  NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7,
+  NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7,
+  NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_5
+};
+
+DECLARE_ALIGNED(16, static const int16_t, LEFT_EDGE_NEIGHBORS_PLUS_1[8]) = {
+  NEIGHBOR_CONSTANT_7,  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10
+};
+
+DECLARE_ALIGNED(16, static const int16_t, RIGHT_EDGE_NEIGHBORS_PLUS_1[8]) = {
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_7
+};
+
+DECLARE_ALIGNED(16, static const int16_t, MIDDLE_EDGE_NEIGHBORS_PLUS_1[8]) = {
+  NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7,
+  NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7,
+  NEIGHBOR_CONSTANT_7, NEIGHBOR_CONSTANT_7
+};
+
+DECLARE_ALIGNED(16, static const int16_t, MIDDLE_CENTER_NEIGHBORS_PLUS_1[8]) = {
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10
+};
+
+DECLARE_ALIGNED(16, static const int16_t, LEFT_CORNER_NEIGHBORS_PLUS_2[8]) = {
+  NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8,
+  NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8,
+  NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8
+};
+
+DECLARE_ALIGNED(16, static const int16_t, RIGHT_CORNER_NEIGHBORS_PLUS_2[8]) = {
+  NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8,
+  NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8,
+  NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_6
+};
+
+DECLARE_ALIGNED(16, static const int16_t, LEFT_EDGE_NEIGHBORS_PLUS_2[8]) = {
+  NEIGHBOR_CONSTANT_8,  NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11,
+  NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11,
+  NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11
+};
+
+DECLARE_ALIGNED(16, static const int16_t, RIGHT_EDGE_NEIGHBORS_PLUS_2[8]) = {
+  NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11,
+  NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11,
+  NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_8
+};
+
+DECLARE_ALIGNED(16, static const int16_t, MIDDLE_EDGE_NEIGHBORS_PLUS_2[8]) = {
+  NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8,
+  NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8,
+  NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8
+};
+
+DECLARE_ALIGNED(16, static const int16_t, MIDDLE_CENTER_NEIGHBORS_PLUS_2[8]) = {
+  NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11,
+  NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11,
+  NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11
+};
+
+DECLARE_ALIGNED(16, static const int16_t, TWO_CORNER_NEIGHBORS_PLUS_2[8]) = {
+  NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8,
+  NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_8,
+  NEIGHBOR_CONSTANT_8, NEIGHBOR_CONSTANT_6
+};
+
+DECLARE_ALIGNED(16, static const int16_t, TWO_EDGE_NEIGHBORS_PLUS_2[8]) = {
+  NEIGHBOR_CONSTANT_8,  NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11,
+  NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_11,
+  NEIGHBOR_CONSTANT_11, NEIGHBOR_CONSTANT_8
+};
+
+DECLARE_ALIGNED(16, static const int16_t, LEFT_CORNER_NEIGHBORS_PLUS_4[8]) = {
+  NEIGHBOR_CONSTANT_8,  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10
+};
+
+DECLARE_ALIGNED(16, static const int16_t, RIGHT_CORNER_NEIGHBORS_PLUS_4[8]) = {
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_8
+};
+
+DECLARE_ALIGNED(16, static const int16_t, LEFT_EDGE_NEIGHBORS_PLUS_4[8]) = {
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13,
+  NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13,
+  NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13
+};
+
+DECLARE_ALIGNED(16, static const int16_t, RIGHT_EDGE_NEIGHBORS_PLUS_4[8]) = {
+  NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13,
+  NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13,
+  NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_10
+};
+
+DECLARE_ALIGNED(16, static const int16_t, MIDDLE_EDGE_NEIGHBORS_PLUS_4[8]) = {
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10
+};
+
+DECLARE_ALIGNED(16, static const int16_t, MIDDLE_CENTER_NEIGHBORS_PLUS_4[8]) = {
+  NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13,
+  NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13,
+  NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13
+};
+
+DECLARE_ALIGNED(16, static const int16_t, TWO_CORNER_NEIGHBORS_PLUS_4[8]) = {
+  NEIGHBOR_CONSTANT_8,  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_10,
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_8
+};
+
+DECLARE_ALIGNED(16, static const int16_t, TWO_EDGE_NEIGHBORS_PLUS_4[8]) = {
+  NEIGHBOR_CONSTANT_10, NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13,
+  NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_13,
+  NEIGHBOR_CONSTANT_13, NEIGHBOR_CONSTANT_10
+};
+
+static const int16_t *const LUMA_LEFT_COLUMN_NEIGHBORS[2] = {
+  LEFT_CORNER_NEIGHBORS_PLUS_2, LEFT_EDGE_NEIGHBORS_PLUS_2
+};
+
+static const int16_t *const LUMA_MIDDLE_COLUMN_NEIGHBORS[2] = {
+  MIDDLE_EDGE_NEIGHBORS_PLUS_2, MIDDLE_CENTER_NEIGHBORS_PLUS_2
+};
+
+static const int16_t *const LUMA_RIGHT_COLUMN_NEIGHBORS[2] = {
+  RIGHT_CORNER_NEIGHBORS_PLUS_2, RIGHT_EDGE_NEIGHBORS_PLUS_2
+};
+
+static const int16_t *const CHROMA_NO_SS_LEFT_COLUMN_NEIGHBORS[2] = {
+  LEFT_CORNER_NEIGHBORS_PLUS_1, LEFT_EDGE_NEIGHBORS_PLUS_1
+};
+
+static const int16_t *const CHROMA_NO_SS_MIDDLE_COLUMN_NEIGHBORS[2] = {
+  MIDDLE_EDGE_NEIGHBORS_PLUS_1, MIDDLE_CENTER_NEIGHBORS_PLUS_1
+};
+
+static const int16_t *const CHROMA_NO_SS_RIGHT_COLUMN_NEIGHBORS[2] = {
+  RIGHT_CORNER_NEIGHBORS_PLUS_1, RIGHT_EDGE_NEIGHBORS_PLUS_1
+};
+
+static const int16_t *const CHROMA_SINGLE_SS_LEFT_COLUMN_NEIGHBORS[2] = {
+  LEFT_CORNER_NEIGHBORS_PLUS_2, LEFT_EDGE_NEIGHBORS_PLUS_2
+};
+
+static const int16_t *const CHROMA_SINGLE_SS_MIDDLE_COLUMN_NEIGHBORS[2] = {
+  MIDDLE_EDGE_NEIGHBORS_PLUS_2, MIDDLE_CENTER_NEIGHBORS_PLUS_2
+};
+
+static const int16_t *const CHROMA_SINGLE_SS_RIGHT_COLUMN_NEIGHBORS[2] = {
+  RIGHT_CORNER_NEIGHBORS_PLUS_2, RIGHT_EDGE_NEIGHBORS_PLUS_2
+};
+
+static const int16_t *const CHROMA_SINGLE_SS_SINGLE_COLUMN_NEIGHBORS[2] = {
+  TWO_CORNER_NEIGHBORS_PLUS_2, TWO_EDGE_NEIGHBORS_PLUS_2
+};
+
+static const int16_t *const CHROMA_DOUBLE_SS_LEFT_COLUMN_NEIGHBORS[2] = {
+  LEFT_CORNER_NEIGHBORS_PLUS_4, LEFT_EDGE_NEIGHBORS_PLUS_4
+};
+
+static const int16_t *const CHROMA_DOUBLE_SS_MIDDLE_COLUMN_NEIGHBORS[2] = {
+  MIDDLE_EDGE_NEIGHBORS_PLUS_4, MIDDLE_CENTER_NEIGHBORS_PLUS_4
+};
+
+static const int16_t *const CHROMA_DOUBLE_SS_RIGHT_COLUMN_NEIGHBORS[2] = {
+  RIGHT_CORNER_NEIGHBORS_PLUS_4, RIGHT_EDGE_NEIGHBORS_PLUS_4
+};
+
+static const int16_t *const CHROMA_DOUBLE_SS_SINGLE_COLUMN_NEIGHBORS[2] = {
+  TWO_CORNER_NEIGHBORS_PLUS_4, TWO_EDGE_NEIGHBORS_PLUS_4
+};
+
+#define DIST_STRIDE ((BW) + 2)
diff --git a/vp9/encoder/x86/temporal_filter_sse4.c b/vp9/encoder/x86/temporal_filter_sse4.c
index e5860d39c..18c4f02d5 100644
--- a/vp9/encoder/x86/temporal_filter_sse4.c
+++ b/vp9/encoder/x86/temporal_filter_sse4.c
@@ -14,32 +14,9 @@
 #include "./vp9_rtcd.h"
 #include "./vpx_config.h"
 #include "vpx/vpx_integer.h"
-
-// Division using multiplication and shifting. The C implementation does:
-// modifier *= 3;
-// modifier /= index;
-// where 'modifier' is a set of summed values and 'index' is the number of
-// summed values. 'index' may be 4, 6, or 9, representing a block of 9 values
-// which may be bound by the edges of the block being filtered.
-//
-// This equation works out to (m * 3) / i which reduces to:
-// m * 3/4
-// m * 1/2
-// m * 1/3
-//
-// By pairing the multiply with a down shift by 16 (_mm_mulhi_epu16):
-// m * C / 65536
-// we can create a C to replicate the division.
-//
-// m * 49152 / 65536 = m * 3/4
-// m * 32758 / 65536 = m * 1/2
-// m * 21846 / 65536 = m * 0.3333
-//
-// These are loaded using an instruction expecting int16_t values but are used
-// with _mm_mulhi_epu16(), which treats them as unsigned.
-#define NEIGHBOR_CONSTANT_4 (int16_t)49152
-#define NEIGHBOR_CONSTANT_6 (int16_t)32768
-#define NEIGHBOR_CONSTANT_9 (int16_t)21846
+#include "vp9/encoder/vp9_encoder.h"
+#include "vp9/encoder/vp9_temporal_filter.h"
+#include "vp9/encoder/x86/temporal_filter_constants.h"
 
 // Load values from 'a' and 'b'. Compute the difference squared and sum
 // neighboring values such that:
@@ -106,12 +83,62 @@ static void sum_16(const uint8_t *a, const uint8_t *b, __m128i *sum_0,
   *sum_1 = sum_u16;
 }
 
+// Read in 8 pixels from a and b as 8-bit unsigned integers, compute the
+// difference squared, and store as unsigned 16-bit integer to dst.
+static INLINE void store_dist_8(const uint8_t *a, const uint8_t *b,
+                                uint16_t *dst) {
+  const __m128i a_reg = _mm_loadl_epi64((const __m128i *)a);
+  const __m128i b_reg = _mm_loadl_epi64((const __m128i *)b);
+
+  const __m128i a_first = _mm_cvtepu8_epi16(a_reg);
+  const __m128i b_first = _mm_cvtepu8_epi16(b_reg);
+
+  __m128i dist_first;
+
+  dist_first = _mm_sub_epi16(a_first, b_first);
+  dist_first = _mm_mullo_epi16(dist_first, dist_first);
+
+  _mm_storeu_si128((__m128i *)dst, dist_first);
+}
+
+static INLINE void store_dist_16(const uint8_t *a, const uint8_t *b,
+                                 uint16_t *dst) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i a_reg = _mm_loadu_si128((const __m128i *)a);
+  const __m128i b_reg = _mm_loadu_si128((const __m128i *)b);
+
+  const __m128i a_first = _mm_cvtepu8_epi16(a_reg);
+  const __m128i a_second = _mm_unpackhi_epi8(a_reg, zero);
+  const __m128i b_first = _mm_cvtepu8_epi16(b_reg);
+  const __m128i b_second = _mm_unpackhi_epi8(b_reg, zero);
+
+  __m128i dist_first, dist_second;
+
+  dist_first = _mm_sub_epi16(a_first, b_first);
+  dist_second = _mm_sub_epi16(a_second, b_second);
+  dist_first = _mm_mullo_epi16(dist_first, dist_first);
+  dist_second = _mm_mullo_epi16(dist_second, dist_second);
+
+  _mm_storeu_si128((__m128i *)dst, dist_first);
+  _mm_storeu_si128((__m128i *)(dst + 8), dist_second);
+}
+
+static INLINE void read_dist_8(const uint16_t *dist, __m128i *dist_reg) {
+  *dist_reg = _mm_loadu_si128((const __m128i *)dist);
+}
+
+static INLINE void read_dist_16(const uint16_t *dist, __m128i *reg_first,
+                                __m128i *reg_second) {
+  read_dist_8(dist, reg_first);
+  read_dist_8(dist + 8, reg_second);
+}
+
 // Average the value based on the number of values summed (9 for pixels away
 // from the border, 4 for pixels in corners, and 6 for other edge values).
 //
 // Add in the rounding factor and shift, clamp to 16, invert and shift. Multiply
 // by weight.
-static __m128i average_8(__m128i sum, const __m128i mul_constants,
+static __m128i average_8(__m128i sum, const __m128i *mul_constants,
                          const int strength, const int rounding,
                          const int weight) {
   // _mm_srl_epi16 uses the lower 64 bit value for the shift.
@@ -121,7 +148,34 @@ static __m128i average_8(__m128i sum, const __m128i mul_constants,
   const __m128i sixteen = _mm_set1_epi16(16);
 
   // modifier * 3 / index;
-  sum = _mm_mulhi_epu16(sum, mul_constants);
+  sum = _mm_mulhi_epu16(sum, *mul_constants);
+
+  sum = _mm_adds_epu16(sum, rounding_u16);
+  sum = _mm_srl_epi16(sum, strength_u128);
+
+  // The maximum input to this comparison is UINT16_MAX * NEIGHBOR_CONSTANT_4
+  // >> 16 (also NEIGHBOR_CONSTANT_4 -1) which is 49151 / 0xbfff / -16385
+  // So this needs to use the epu16 version which did not come until SSE4.
+  sum = _mm_min_epu16(sum, sixteen);
+
+  sum = _mm_sub_epi16(sixteen, sum);
+
+  return _mm_mullo_epi16(sum, weight_u16);
+}
+
+static __m128i average_4_4(__m128i sum, const __m128i *mul_constants,
+                           const int strength, const int rounding,
+                           const int weight_0, const int weight_1) {
+  // _mm_srl_epi16 uses the lower 64 bit value for the shift.
+  const __m128i strength_u128 = _mm_set_epi32(0, 0, 0, strength);
+  const __m128i rounding_u16 = _mm_set1_epi16(rounding);
+  const __m128i weight_u16 =
+      _mm_setr_epi16(weight_0, weight_0, weight_0, weight_0, weight_1, weight_1,
+                     weight_1, weight_1);
+  const __m128i sixteen = _mm_set1_epi16(16);
+
+  // modifier * 3 / index;
+  sum = _mm_mulhi_epu16(sum, *mul_constants);
 
   sum = _mm_adds_epu16(sum, rounding_u16);
   sum = _mm_srl_epi16(sum, strength_u128);
@@ -136,20 +190,21 @@ static __m128i average_8(__m128i sum, const __m128i mul_constants,
   return _mm_mullo_epi16(sum, weight_u16);
 }
 
-static void average_16(__m128i *sum_0_u16, __m128i *sum_1_u16,
-                       const __m128i mul_constants_0,
-                       const __m128i mul_constants_1, const int strength,
-                       const int rounding, const int weight) {
+static INLINE void average_16(__m128i *sum_0_u16, __m128i *sum_1_u16,
+                              const __m128i *mul_constants_0,
+                              const __m128i *mul_constants_1,
+                              const int strength, const int rounding,
+                              const int weight) {
   const __m128i strength_u128 = _mm_set_epi32(0, 0, 0, strength);
   const __m128i rounding_u16 = _mm_set1_epi16(rounding);
   const __m128i weight_u16 = _mm_set1_epi16(weight);
   const __m128i sixteen = _mm_set1_epi16(16);
   __m128i input_0, input_1;
 
-  input_0 = _mm_mulhi_epu16(*sum_0_u16, mul_constants_0);
+  input_0 = _mm_mulhi_epu16(*sum_0_u16, *mul_constants_0);
   input_0 = _mm_adds_epu16(input_0, rounding_u16);
 
-  input_1 = _mm_mulhi_epu16(*sum_1_u16, mul_constants_1);
+  input_1 = _mm_mulhi_epu16(*sum_1_u16, *mul_constants_1);
   input_1 = _mm_adds_epu16(input_1, rounding_u16);
 
   input_0 = _mm_srl_epi16(input_0, strength_u128);
@@ -192,10 +247,10 @@ static void accumulate_and_store_8(const __m128i sum_u16, const uint8_t *pred,
   _mm_storeu_si128((__m128i *)(accumulator + 4), accum_1_u32);
 }
 
-static void accumulate_and_store_16(const __m128i sum_0_u16,
-                                    const __m128i sum_1_u16,
-                                    const uint8_t *pred, uint16_t *count,
-                                    uint32_t *accumulator) {
+static INLINE void accumulate_and_store_16(const __m128i sum_0_u16,
+                                           const __m128i sum_1_u16,
+                                           const uint8_t *pred, uint16_t *count,
+                                           uint32_t *accumulator) {
   const __m128i pred_u8 = _mm_loadu_si128((const __m128i *)pred);
   const __m128i zero = _mm_setzero_si128();
   __m128i count_0_u16 = _mm_loadu_si128((const __m128i *)count),
@@ -235,6 +290,28 @@ static void accumulate_and_store_16(const __m128i sum_0_u16,
   _mm_storeu_si128((__m128i *)(accumulator + 12), accum_3_u32);
 }
 
+// Read in 8 pixels from y_dist. For each index i, compute y_dist[i-1] +
+// y_dist[i] + y_dist[i+1] and store in sum as 16-bit unsigned int.
+static INLINE void get_sum_8(const uint16_t *y_dist, __m128i *sum) {
+  __m128i dist_reg, dist_left, dist_right;
+
+  dist_reg = _mm_loadu_si128((const __m128i *)y_dist);
+  dist_left = _mm_loadu_si128((const __m128i *)(y_dist - 1));
+  dist_right = _mm_loadu_si128((const __m128i *)(y_dist + 1));
+
+  *sum = _mm_adds_epu16(dist_reg, dist_left);
+  *sum = _mm_adds_epu16(*sum, dist_right);
+}
+
+// Read in 16 pixels from y_dist. For each index i, compute y_dist[i-1] +
+// y_dist[i] + y_dist[i+1]. Store the result for first 8 pixels in sum_first and
+// the rest in sum_second.
+static INLINE void get_sum_16(const uint16_t *y_dist, __m128i *sum_first,
+                              __m128i *sum_second) {
+  get_sum_8(y_dist, sum_first);
+  get_sum_8(y_dist + 8, sum_second);
+}
+
 void vp9_temporal_filter_apply_sse4_1(const uint8_t *a, unsigned int stride,
                                       const uint8_t *b, unsigned int width,
                                       unsigned int height, int strength,
@@ -261,7 +338,8 @@ void vp9_temporal_filter_apply_sse4_1(const uint8_t *a, unsigned int stride,
     sum_8(a, b, &sum_row_a);
     sum_8(a + stride, b + width, &sum_row_b);
     sum_row_c = _mm_adds_epu16(sum_row_a, sum_row_b);
-    sum_row_c = average_8(sum_row_c, mul_constants, strength, rounding, weight);
+    sum_row_c =
+        average_8(sum_row_c, &mul_constants, strength, rounding, weight);
     accumulate_and_store_8(sum_row_c, b, count, accumulator);
 
     a += stride + stride;
@@ -279,7 +357,7 @@ void vp9_temporal_filter_apply_sse4_1(const uint8_t *a, unsigned int stride,
       sum_row_a = _mm_adds_epu16(sum_row_a, sum_row_b);
       sum_row_a = _mm_adds_epu16(sum_row_a, sum_row_c);
       sum_row_a =
-          average_8(sum_row_a, mul_constants, strength, rounding, weight);
+          average_8(sum_row_a, &mul_constants, strength, rounding, weight);
       accumulate_and_store_8(sum_row_a, b, count, accumulator);
 
       a += stride;
@@ -296,7 +374,8 @@ void vp9_temporal_filter_apply_sse4_1(const uint8_t *a, unsigned int stride,
                                    NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_6,
                                    NEIGHBOR_CONSTANT_6, NEIGHBOR_CONSTANT_4);
     sum_row_a = _mm_adds_epu16(sum_row_a, sum_row_b);
-    sum_row_a = average_8(sum_row_a, mul_constants, strength, rounding, weight);
+    sum_row_a =
+        average_8(sum_row_a, &mul_constants, strength, rounding, weight);
     accumulate_and_store_8(sum_row_a, b, count, accumulator);
 
   } else {  // width == 16
@@ -318,7 +397,7 @@ void vp9_temporal_filter_apply_sse4_1(const uint8_t *a, unsigned int stride,
     sum_row_c_0 = _mm_adds_epu16(sum_row_a_0, sum_row_b_0);
     sum_row_c_1 = _mm_adds_epu16(sum_row_a_1, sum_row_b_1);
 
-    average_16(&sum_row_c_0, &sum_row_c_1, mul_constants_0, mul_constants_1,
+    average_16(&sum_row_c_0, &sum_row_c_1, &mul_constants_0, &mul_constants_1,
                strength, rounding, weight);
     accumulate_and_store_16(sum_row_c_0, sum_row_c_1, b, count, accumulator);
 
@@ -343,7 +422,7 @@ void vp9_temporal_filter_apply_sse4_1(const uint8_t *a, unsigned int stride,
       sum_row_a_1 = _mm_adds_epu16(sum_row_a_1, sum_row_b_1);
       sum_row_a_1 = _mm_adds_epu16(sum_row_a_1, sum_row_c_1);
 
-      average_16(&sum_row_a_0, &sum_row_a_1, mul_constants_0, mul_constants_1,
+      average_16(&sum_row_a_0, &sum_row_a_1, &mul_constants_0, &mul_constants_1,
                  strength, rounding, weight);
       accumulate_and_store_16(sum_row_a_0, sum_row_a_1, b, count, accumulator);
 
@@ -369,8 +448,757 @@ void vp9_temporal_filter_apply_sse4_1(const uint8_t *a, unsigned int stride,
     sum_row_c_0 = _mm_adds_epu16(sum_row_a_0, sum_row_b_0);
     sum_row_c_1 = _mm_adds_epu16(sum_row_a_1, sum_row_b_1);
 
-    average_16(&sum_row_c_0, &sum_row_c_1, mul_constants_0, mul_constants_1,
+    average_16(&sum_row_c_0, &sum_row_c_1, &mul_constants_0, &mul_constants_1,
                strength, rounding, weight);
     accumulate_and_store_16(sum_row_c_0, sum_row_c_1, b, count, accumulator);
   }
 }
+
+// Read in a row of chroma values corresponds to a row of 16 luma values.
+static INLINE void read_chroma_dist_row_16(int ss_x, const uint16_t *u_dist,
+                                           const uint16_t *v_dist,
+                                           __m128i *u_first, __m128i *u_second,
+                                           __m128i *v_first,
+                                           __m128i *v_second) {
+  if (!ss_x) {
+    // If there is no chroma subsampling in the horizaontal direction, then we
+    // need to load 16 entries from chroma.
+    read_dist_16(u_dist, u_first, u_second);
+    read_dist_16(v_dist, v_first, v_second);
+  } else {  // ss_x == 1
+    // Otherwise, we only need to load 8 entries
+    __m128i u_reg, v_reg;
+
+    read_dist_8(u_dist, &u_reg);
+
+    *u_first = _mm_unpacklo_epi16(u_reg, u_reg);
+    *u_second = _mm_unpackhi_epi16(u_reg, u_reg);
+
+    read_dist_8(v_dist, &v_reg);
+
+    *v_first = _mm_unpacklo_epi16(v_reg, v_reg);
+    *v_second = _mm_unpackhi_epi16(v_reg, v_reg);
+  }
+}
+
+// Horizonta add unsigned 16-bit ints in src and store them as signed 32-bit int
+// in dst.
+static INLINE void hadd_epu16(__m128i *src, __m128i *dst) {
+  const __m128i zero = _mm_setzero_si128();
+  const __m128i shift_right = _mm_srli_si128(*src, 2);
+
+  const __m128i odd = _mm_blend_epi16(shift_right, zero, 170);
+  const __m128i even = _mm_blend_epi16(*src, zero, 170);
+
+  *dst = _mm_add_epi32(even, odd);
+}
+
+// Add a row of luma distortion to 8 corresponding chroma mods.
+static INLINE void add_luma_dist_to_8_chroma_mod(const uint16_t *y_dist,
+                                                 int ss_x, int ss_y,
+                                                 __m128i *u_mod,
+                                                 __m128i *v_mod) {
+  __m128i y_reg;
+  if (!ss_x) {
+    read_dist_8(y_dist, &y_reg);
+    if (ss_y == 1) {
+      __m128i y_tmp;
+      read_dist_8(y_dist + DIST_STRIDE, &y_tmp);
+
+      y_reg = _mm_adds_epu16(y_reg, y_tmp);
+    }
+  } else {
+    __m128i y_first, y_second;
+    read_dist_16(y_dist, &y_first, &y_second);
+    if (ss_y == 1) {
+      __m128i y_tmp_0, y_tmp_1;
+      read_dist_16(y_dist + DIST_STRIDE, &y_tmp_0, &y_tmp_1);
+
+      y_first = _mm_adds_epu16(y_first, y_tmp_0);
+      y_second = _mm_adds_epu16(y_second, y_tmp_1);
+    }
+
+    hadd_epu16(&y_first, &y_first);
+    hadd_epu16(&y_second, &y_second);
+
+    y_reg = _mm_packus_epi32(y_first, y_second);
+  }
+
+  *u_mod = _mm_adds_epu16(*u_mod, y_reg);
+  *v_mod = _mm_adds_epu16(*v_mod, y_reg);
+}
+
+// Apply temporal filter to the luma components. This performs temporal
+// filtering on a luma block of 16 X block_height. Use blk_fw as an array of
+// size 4for the weights for each of the 4 subblocks if blk_fw is not NULL,
+// else use top_weight for top half, and bottom weight for bottom half.
+static void vp9_apply_temporal_filter_luma_16(
+    const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre,
+    int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src,
+    int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre,
+    int uv_pre_stride, unsigned int block_width, unsigned int block_height,
+    int ss_x, int ss_y, int strength, int use_whole_blk, uint32_t *y_accum,
+    uint16_t *y_count, const uint16_t *y_dist, const uint16_t *u_dist,
+    const uint16_t *v_dist, const int16_t *const *neighbors_first,
+    const int16_t *const *neighbors_second, int top_weight, int bottom_weight,
+    const int *blk_fw) {
+  const int rounding = (1 << strength) >> 1;
+  int weight = top_weight;
+
+  __m128i mul_first, mul_second;
+
+  __m128i sum_row_1_first, sum_row_1_second;
+  __m128i sum_row_2_first, sum_row_2_second;
+  __m128i sum_row_3_first, sum_row_3_second;
+
+  __m128i u_first, u_second;
+  __m128i v_first, v_second;
+
+  __m128i sum_row_first;
+  __m128i sum_row_second;
+
+  assert(strength >= 0);
+  assert(strength <= 6);
+
+  assert(block_width == 16);
+
+  (void)block_width;
+
+  // First row
+  mul_first = _mm_loadu_si128((const __m128i *)neighbors_first[0]);
+  mul_second = _mm_loadu_si128((const __m128i *)neighbors_second[0]);
+
+  // Add luma values
+  get_sum_16(y_dist, &sum_row_2_first, &sum_row_2_second);
+  get_sum_16(y_dist + DIST_STRIDE, &sum_row_3_first, &sum_row_3_second);
+
+  sum_row_first = _mm_adds_epu16(sum_row_2_first, sum_row_3_first);
+  sum_row_second = _mm_adds_epu16(sum_row_2_second, sum_row_3_second);
+
+  // Add chroma values
+  read_chroma_dist_row_16(ss_x, u_dist, v_dist, &u_first, &u_second, &v_first,
+                          &v_second);
+
+  sum_row_first = _mm_adds_epu16(sum_row_first, u_first);
+  sum_row_second = _mm_adds_epu16(sum_row_second, u_second);
+
+  sum_row_first = _mm_adds_epu16(sum_row_first, v_first);
+  sum_row_second = _mm_adds_epu16(sum_row_second, v_second);
+
+  // Get modifier and store result
+  if (blk_fw) {
+    sum_row_first =
+        average_8(sum_row_first, &mul_first, strength, rounding, blk_fw[0]);
+    sum_row_second =
+        average_8(sum_row_second, &mul_second, strength, rounding, blk_fw[1]);
+  } else {
+    average_16(&sum_row_first, &sum_row_second, &mul_first, &mul_second,
+               strength, rounding, weight);
+  }
+  accumulate_and_store_16(sum_row_first, sum_row_second, y_pre, y_count,
+                          y_accum);
+
+  y_src += y_src_stride;
+  y_pre += y_pre_stride;
+  y_count += y_pre_stride;
+  y_accum += y_pre_stride;
+  y_dist += DIST_STRIDE;
+
+  u_src += uv_src_stride;
+  u_pre += uv_pre_stride;
+  u_dist += DIST_STRIDE;
+  v_src += uv_src_stride;
+  v_pre += uv_pre_stride;
+  v_dist += DIST_STRIDE;
+
+  // Then all the rows except the last one
+  mul_first = _mm_loadu_si128((const __m128i *)neighbors_first[1]);
+  mul_second = _mm_loadu_si128((const __m128i *)neighbors_second[1]);
+
+  for (unsigned int h = 1; h < block_height - 1; ++h) {
+    // Move the weight to bottom half
+    if (!use_whole_blk && h == block_height / 2) {
+      if (blk_fw) {
+        blk_fw += 2;
+      } else {
+        weight = bottom_weight;
+      }
+    }
+    // Shift the rows up
+    sum_row_1_first = sum_row_2_first;
+    sum_row_1_second = sum_row_2_second;
+    sum_row_2_first = sum_row_3_first;
+    sum_row_2_second = sum_row_3_second;
+
+    // Add luma values to the modifier
+    sum_row_first = _mm_adds_epu16(sum_row_1_first, sum_row_2_first);
+    sum_row_second = _mm_adds_epu16(sum_row_1_second, sum_row_2_second);
+
+    get_sum_16(y_dist + DIST_STRIDE, &sum_row_3_first, &sum_row_3_second);
+
+    sum_row_first = _mm_adds_epu16(sum_row_first, sum_row_3_first);
+    sum_row_second = _mm_adds_epu16(sum_row_second, sum_row_3_second);
+
+    // Add chroma values to the modifier
+    if (ss_y == 0 || h % 2 == 0) {
+      // Only calculate the new chroma distortion if we are at a pixel that
+      // corresponds to a new chroma row
+      read_chroma_dist_row_16(ss_x, u_dist, v_dist, &u_first, &u_second,
+                              &v_first, &v_second);
+
+      u_src += uv_src_stride;
+      u_pre += uv_pre_stride;
+      u_dist += DIST_STRIDE;
+      v_src += uv_src_stride;
+      v_pre += uv_pre_stride;
+      v_dist += DIST_STRIDE;
+    }
+
+    sum_row_first = _mm_adds_epu16(sum_row_first, u_first);
+    sum_row_second = _mm_adds_epu16(sum_row_second, u_second);
+    sum_row_first = _mm_adds_epu16(sum_row_first, v_first);
+    sum_row_second = _mm_adds_epu16(sum_row_second, v_second);
+
+    // Get modifier and store result
+    if (blk_fw) {
+      sum_row_first =
+          average_8(sum_row_first, &mul_first, strength, rounding, blk_fw[0]);
+      sum_row_second =
+          average_8(sum_row_second, &mul_second, strength, rounding, blk_fw[1]);
+    } else {
+      average_16(&sum_row_first, &sum_row_second, &mul_first, &mul_second,
+                 strength, rounding, weight);
+    }
+    accumulate_and_store_16(sum_row_first, sum_row_second, y_pre, y_count,
+                            y_accum);
+
+    y_src += y_src_stride;
+    y_pre += y_pre_stride;
+    y_count += y_pre_stride;
+    y_accum += y_pre_stride;
+    y_dist += DIST_STRIDE;
+  }
+
+  // The last row
+  mul_first = _mm_loadu_si128((const __m128i *)neighbors_first[0]);
+  mul_second = _mm_loadu_si128((const __m128i *)neighbors_second[0]);
+
+  // Shift the rows up
+  sum_row_1_first = sum_row_2_first;
+  sum_row_1_second = sum_row_2_second;
+  sum_row_2_first = sum_row_3_first;
+  sum_row_2_second = sum_row_3_second;
+
+  // Add luma values to the modifier
+  sum_row_first = _mm_adds_epu16(sum_row_1_first, sum_row_2_first);
+  sum_row_second = _mm_adds_epu16(sum_row_1_second, sum_row_2_second);
+
+  // Add chroma values to the modifier
+  if (ss_y == 0) {
+    // Only calculate the new chroma distortion if we are at a pixel that
+    // corresponds to a new chroma row
+    read_chroma_dist_row_16(ss_x, u_dist, v_dist, &u_first, &u_second, &v_first,
+                            &v_second);
+  }
+
+  sum_row_first = _mm_adds_epu16(sum_row_first, u_first);
+  sum_row_second = _mm_adds_epu16(sum_row_second, u_second);
+  sum_row_first = _mm_adds_epu16(sum_row_first, v_first);
+  sum_row_second = _mm_adds_epu16(sum_row_second, v_second);
+
+  // Get modifier and store result
+  if (blk_fw) {
+    sum_row_first =
+        average_8(sum_row_first, &mul_first, strength, rounding, blk_fw[0]);
+    sum_row_second =
+        average_8(sum_row_second, &mul_second, strength, rounding, blk_fw[1]);
+  } else {
+    average_16(&sum_row_first, &sum_row_second, &mul_first, &mul_second,
+               strength, rounding, weight);
+  }
+  accumulate_and_store_16(sum_row_first, sum_row_second, y_pre, y_count,
+                          y_accum);
+}
+
+// Perform temporal filter for the luma component.
+static void vp9_apply_temporal_filter_luma(
+    const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre,
+    int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src,
+    int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre,
+    int uv_pre_stride, unsigned int block_width, unsigned int block_height,
+    int ss_x, int ss_y, int strength, const int *blk_fw, int use_whole_blk,
+    uint32_t *y_accum, uint16_t *y_count, const uint16_t *y_dist,
+    const uint16_t *u_dist, const uint16_t *v_dist) {
+  unsigned int blk_col = 0, uv_blk_col = 0;
+  const unsigned int blk_col_step = 16, uv_blk_col_step = 16 >> ss_x;
+  const unsigned int mid_width = block_width >> 1,
+                     last_width = block_width - blk_col_step;
+  int top_weight = blk_fw[0],
+      bottom_weight = use_whole_blk ? blk_fw[0] : blk_fw[2];
+  const int16_t *const *neighbors_first;
+  const int16_t *const *neighbors_second;
+
+  if (block_width == 16) {
+    // Special Case: The blockwidth is 16 and we are operating on a row of 16
+    // chroma pixels. In this case, we can't use the usualy left-midle-right
+    // pattern. We also don't support splitting now.
+    neighbors_first = LUMA_LEFT_COLUMN_NEIGHBORS;
+    neighbors_second = LUMA_RIGHT_COLUMN_NEIGHBORS;
+    if (use_whole_blk) {
+      vp9_apply_temporal_filter_luma_16(
+          y_src + blk_col, y_src_stride, y_pre + blk_col, y_pre_stride,
+          u_src + uv_blk_col, v_src + uv_blk_col, uv_src_stride,
+          u_pre + uv_blk_col, v_pre + uv_blk_col, uv_pre_stride, 16,
+          block_height, ss_x, ss_y, strength, use_whole_blk, y_accum + blk_col,
+          y_count + blk_col, y_dist + blk_col, u_dist + uv_blk_col,
+          v_dist + uv_blk_col, neighbors_first, neighbors_second, top_weight,
+          bottom_weight, NULL);
+    } else {
+      vp9_apply_temporal_filter_luma_16(
+          y_src + blk_col, y_src_stride, y_pre + blk_col, y_pre_stride,
+          u_src + uv_blk_col, v_src + uv_blk_col, uv_src_stride,
+          u_pre + uv_blk_col, v_pre + uv_blk_col, uv_pre_stride, 16,
+          block_height, ss_x, ss_y, strength, use_whole_blk, y_accum + blk_col,
+          y_count + blk_col, y_dist + blk_col, u_dist + uv_blk_col,
+          v_dist + uv_blk_col, neighbors_first, neighbors_second, 0, 0, blk_fw);
+    }
+
+    return;
+  }
+
+  // Left
+  neighbors_first = LUMA_LEFT_COLUMN_NEIGHBORS;
+  neighbors_second = LUMA_MIDDLE_COLUMN_NEIGHBORS;
+  vp9_apply_temporal_filter_luma_16(
+      y_src + blk_col, y_src_stride, y_pre + blk_col, y_pre_stride,
+      u_src + uv_blk_col, v_src + uv_blk_col, uv_src_stride, u_pre + uv_blk_col,
+      v_pre + uv_blk_col, uv_pre_stride, 16, block_height, ss_x, ss_y, strength,
+      use_whole_blk, y_accum + blk_col, y_count + blk_col, y_dist + blk_col,
+      u_dist + uv_blk_col, v_dist + uv_blk_col, neighbors_first,
+      neighbors_second, top_weight, bottom_weight, NULL);
+
+  blk_col += blk_col_step;
+  uv_blk_col += uv_blk_col_step;
+
+  // Middle First
+  neighbors_first = LUMA_MIDDLE_COLUMN_NEIGHBORS;
+  for (; blk_col < mid_width;
+       blk_col += blk_col_step, uv_blk_col += uv_blk_col_step) {
+    vp9_apply_temporal_filter_luma_16(
+        y_src + blk_col, y_src_stride, y_pre + blk_col, y_pre_stride,
+        u_src + uv_blk_col, v_src + uv_blk_col, uv_src_stride,
+        u_pre + uv_blk_col, v_pre + uv_blk_col, uv_pre_stride, 16, block_height,
+        ss_x, ss_y, strength, use_whole_blk, y_accum + blk_col,
+        y_count + blk_col, y_dist + blk_col, u_dist + uv_blk_col,
+        v_dist + uv_blk_col, neighbors_first, neighbors_second, top_weight,
+        bottom_weight, NULL);
+  }
+
+  if (!use_whole_blk) {
+    top_weight = blk_fw[1];
+    bottom_weight = blk_fw[3];
+  }
+
+  // Middle Second
+  for (; blk_col < last_width;
+       blk_col += blk_col_step, uv_blk_col += uv_blk_col_step) {
+    vp9_apply_temporal_filter_luma_16(
+        y_src + blk_col, y_src_stride, y_pre + blk_col, y_pre_stride,
+        u_src + uv_blk_col, v_src + uv_blk_col, uv_src_stride,
+        u_pre + uv_blk_col, v_pre + uv_blk_col, uv_pre_stride, 16, block_height,
+        ss_x, ss_y, strength, use_whole_blk, y_accum + blk_col,
+        y_count + blk_col, y_dist + blk_col, u_dist + uv_blk_col,
+        v_dist + uv_blk_col, neighbors_first, neighbors_second, top_weight,
+        bottom_weight, NULL);
+  }
+
+  // Right
+  neighbors_second = LUMA_RIGHT_COLUMN_NEIGHBORS;
+  vp9_apply_temporal_filter_luma_16(
+      y_src + blk_col, y_src_stride, y_pre + blk_col, y_pre_stride,
+      u_src + uv_blk_col, v_src + uv_blk_col, uv_src_stride, u_pre + uv_blk_col,
+      v_pre + uv_blk_col, uv_pre_stride, 16, block_height, ss_x, ss_y, strength,
+      use_whole_blk, y_accum + blk_col, y_count + blk_col, y_dist + blk_col,
+      u_dist + uv_blk_col, v_dist + uv_blk_col, neighbors_first,
+      neighbors_second, top_weight, bottom_weight, NULL);
+}
+
+// Apply temporal filter to the chroma components. This performs temporal
+// filtering on a chroma block of 8 X uv_height. If blk_fw is not NULL, use
+// blk_fw as an array of size 4 for the weights for each of the 4 subblocks,
+// else use top_weight for top half, and bottom weight for bottom half.
+static void vp9_apply_temporal_filter_chroma_8(
+    const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre,
+    int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src,
+    int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre,
+    int uv_pre_stride, unsigned int uv_block_width,
+    unsigned int uv_block_height, int ss_x, int ss_y, int strength,
+    uint32_t *u_accum, uint16_t *u_count, uint32_t *v_accum, uint16_t *v_count,
+    const uint16_t *y_dist, const uint16_t *u_dist, const uint16_t *v_dist,
+    const int16_t *const *neighbors, int top_weight, int bottom_weight,
+    const int *blk_fw) {
+  const int rounding = (1 << strength) >> 1;
+  int weight = top_weight;
+
+  __m128i mul;
+
+  __m128i u_sum_row_1, u_sum_row_2, u_sum_row_3;
+  __m128i v_sum_row_1, v_sum_row_2, v_sum_row_3;
+
+  __m128i u_sum_row, v_sum_row;
+
+  (void)uv_block_width;
+
+  // First row
+  mul = _mm_loadu_si128((const __m128i *)neighbors[0]);
+
+  // Add chroma values
+  get_sum_8(u_dist, &u_sum_row_2);
+  get_sum_8(u_dist + DIST_STRIDE, &u_sum_row_3);
+
+  u_sum_row = _mm_adds_epu16(u_sum_row_2, u_sum_row_3);
+
+  get_sum_8(v_dist, &v_sum_row_2);
+  get_sum_8(v_dist + DIST_STRIDE, &v_sum_row_3);
+
+  v_sum_row = _mm_adds_epu16(v_sum_row_2, v_sum_row_3);
+
+  // Add luma values
+  add_luma_dist_to_8_chroma_mod(y_dist, ss_x, ss_y, &u_sum_row, &v_sum_row);
+
+  // Get modifier and store result
+  if (blk_fw) {
+    u_sum_row =
+        average_4_4(u_sum_row, &mul, strength, rounding, blk_fw[0], blk_fw[1]);
+    v_sum_row =
+        average_4_4(v_sum_row, &mul, strength, rounding, blk_fw[0], blk_fw[1]);
+  } else {
+    u_sum_row = average_8(u_sum_row, &mul, strength, rounding, weight);
+    v_sum_row = average_8(v_sum_row, &mul, strength, rounding, weight);
+  }
+  accumulate_and_store_8(u_sum_row, u_pre, u_count, u_accum);
+  accumulate_and_store_8(v_sum_row, v_pre, v_count, v_accum);
+
+  u_src += uv_src_stride;
+  u_pre += uv_pre_stride;
+  u_dist += DIST_STRIDE;
+  v_src += uv_src_stride;
+  v_pre += uv_pre_stride;
+  v_dist += DIST_STRIDE;
+  u_count += uv_pre_stride;
+  u_accum += uv_pre_stride;
+  v_count += uv_pre_stride;
+  v_accum += uv_pre_stride;
+
+  y_src += y_src_stride * (1 + ss_y);
+  y_pre += y_pre_stride * (1 + ss_y);
+  y_dist += DIST_STRIDE * (1 + ss_y);
+
+  // Then all the rows except the last one
+  mul = _mm_loadu_si128((const __m128i *)neighbors[1]);
+
+  for (unsigned int h = 1; h < uv_block_height - 1; ++h) {
+    // Move the weight pointer to the bottom half of the blocks
+    if (h == uv_block_height / 2) {
+      if (blk_fw) {
+        blk_fw += 2;
+      } else {
+        weight = bottom_weight;
+      }
+    }
+
+    // Shift the rows up
+    u_sum_row_1 = u_sum_row_2;
+    u_sum_row_2 = u_sum_row_3;
+
+    v_sum_row_1 = v_sum_row_2;
+    v_sum_row_2 = v_sum_row_3;
+
+    // Add chroma values
+    u_sum_row = _mm_adds_epu16(u_sum_row_1, u_sum_row_2);
+    get_sum_8(u_dist + DIST_STRIDE, &u_sum_row_3);
+    u_sum_row = _mm_adds_epu16(u_sum_row, u_sum_row_3);
+
+    v_sum_row = _mm_adds_epu16(v_sum_row_1, v_sum_row_2);
+    get_sum_8(v_dist + DIST_STRIDE, &v_sum_row_3);
+    v_sum_row = _mm_adds_epu16(v_sum_row, v_sum_row_3);
+
+    // Add luma values
+    add_luma_dist_to_8_chroma_mod(y_dist, ss_x, ss_y, &u_sum_row, &v_sum_row);
+
+    // Get modifier and store result
+    if (blk_fw) {
+      u_sum_row = average_4_4(u_sum_row, &mul, strength, rounding, blk_fw[0],
+                              blk_fw[1]);
+      v_sum_row = average_4_4(v_sum_row, &mul, strength, rounding, blk_fw[0],
+                              blk_fw[1]);
+    } else {
+      u_sum_row = average_8(u_sum_row, &mul, strength, rounding, weight);
+      v_sum_row = average_8(v_sum_row, &mul, strength, rounding, weight);
+    }
+
+    accumulate_and_store_8(u_sum_row, u_pre, u_count, u_accum);
+    accumulate_and_store_8(v_sum_row, v_pre, v_count, v_accum);
+
+    u_src += uv_src_stride;
+    u_pre += uv_pre_stride;
+    u_dist += DIST_STRIDE;
+    v_src += uv_src_stride;
+    v_pre += uv_pre_stride;
+    v_dist += DIST_STRIDE;
+    u_count += uv_pre_stride;
+    u_accum += uv_pre_stride;
+    v_count += uv_pre_stride;
+    v_accum += uv_pre_stride;
+
+    y_src += y_src_stride * (1 + ss_y);
+    y_pre += y_pre_stride * (1 + ss_y);
+    y_dist += DIST_STRIDE * (1 + ss_y);
+  }
+
+  // The last row
+  mul = _mm_loadu_si128((const __m128i *)neighbors[0]);
+
+  // Shift the rows up
+  u_sum_row_1 = u_sum_row_2;
+  u_sum_row_2 = u_sum_row_3;
+
+  v_sum_row_1 = v_sum_row_2;
+  v_sum_row_2 = v_sum_row_3;
+
+  // Add chroma values
+  u_sum_row = _mm_adds_epu8(u_sum_row_1, u_sum_row_2);
+  v_sum_row = _mm_adds_epu8(v_sum_row_1, v_sum_row_2);
+
+  // Add luma values
+  add_luma_dist_to_8_chroma_mod(y_dist, ss_x, ss_y, &u_sum_row, &v_sum_row);
+
+  // Get modifier and store result
+  if (blk_fw) {
+    u_sum_row =
+        average_4_4(u_sum_row, &mul, strength, rounding, blk_fw[0], blk_fw[1]);
+    v_sum_row =
+        average_4_4(v_sum_row, &mul, strength, rounding, blk_fw[0], blk_fw[1]);
+  } else {
+    u_sum_row = average_8(u_sum_row, &mul, strength, rounding, weight);
+    v_sum_row = average_8(v_sum_row, &mul, strength, rounding, weight);
+  }
+
+  accumulate_and_store_8(u_sum_row, u_pre, u_count, u_accum);
+  accumulate_and_store_8(v_sum_row, v_pre, v_count, v_accum);
+}
+
+// Perform temporal filter for the chroma components.
+static void vp9_apply_temporal_filter_chroma(
+    const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre,
+    int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src,
+    int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre,
+    int uv_pre_stride, unsigned int block_width, unsigned int block_height,
+    int ss_x, int ss_y, int strength, const int *blk_fw, int use_whole_blk,
+    uint32_t *u_accum, uint16_t *u_count, uint32_t *v_accum, uint16_t *v_count,
+    const uint16_t *y_dist, const uint16_t *u_dist, const uint16_t *v_dist) {
+  const unsigned int uv_width = block_width >> ss_x,
+                     uv_height = block_height >> ss_y;
+
+  unsigned int blk_col = 0, uv_blk_col = 0;
+  const unsigned int uv_blk_col_step = 8, blk_col_step = 8 << ss_x;
+  const unsigned int uv_mid_width = uv_width >> 1,
+                     uv_last_width = uv_width - uv_blk_col_step;
+  int top_weight = blk_fw[0],
+      bottom_weight = use_whole_blk ? blk_fw[0] : blk_fw[2];
+  const int16_t *const *neighbors;
+
+  if (uv_width == 8) {
+    // Special Case: We are subsampling in x direction on a 16x16 block. Since
+    // we are operating on a row of 8 chroma pixels, we can't use the usual
+    // left-middle-right pattern.
+    assert(ss_x);
+
+    if (ss_y) {
+      neighbors = CHROMA_DOUBLE_SS_SINGLE_COLUMN_NEIGHBORS;
+    } else {
+      neighbors = CHROMA_SINGLE_SS_SINGLE_COLUMN_NEIGHBORS;
+    }
+
+    if (use_whole_blk) {
+      vp9_apply_temporal_filter_chroma_8(
+          y_src + blk_col, y_src_stride, y_pre + blk_col, y_pre_stride,
+          u_src + uv_blk_col, v_src + uv_blk_col, uv_src_stride,
+          u_pre + uv_blk_col, v_pre + uv_blk_col, uv_pre_stride, uv_width,
+          uv_height, ss_x, ss_y, strength, u_accum + uv_blk_col,
+          u_count + uv_blk_col, v_accum + uv_blk_col, v_count + uv_blk_col,
+          y_dist + blk_col, u_dist + uv_blk_col, v_dist + uv_blk_col, neighbors,
+          top_weight, bottom_weight, NULL);
+    } else {
+      vp9_apply_temporal_filter_chroma_8(
+          y_src + blk_col, y_src_stride, y_pre + blk_col, y_pre_stride,
+          u_src + uv_blk_col, v_src + uv_blk_col, uv_src_stride,
+          u_pre + uv_blk_col, v_pre + uv_blk_col, uv_pre_stride, uv_width,
+          uv_height, ss_x, ss_y, strength, u_accum + uv_blk_col,
+          u_count + uv_blk_col, v_accum + uv_blk_col, v_count + uv_blk_col,
+          y_dist + blk_col, u_dist + uv_blk_col, v_dist + uv_blk_col, neighbors,
+          0, 0, blk_fw);
+    }
+
+    return;
+  }
+
+  // Left
+  if (ss_x && ss_y) {
+    neighbors = CHROMA_DOUBLE_SS_LEFT_COLUMN_NEIGHBORS;
+  } else if (ss_x || ss_y) {
+    neighbors = CHROMA_SINGLE_SS_LEFT_COLUMN_NEIGHBORS;
+  } else {
+    neighbors = CHROMA_NO_SS_LEFT_COLUMN_NEIGHBORS;
+  }
+
+  vp9_apply_temporal_filter_chroma_8(
+      y_src + blk_col, y_src_stride, y_pre + blk_col, y_pre_stride,
+      u_src + uv_blk_col, v_src + uv_blk_col, uv_src_stride, u_pre + uv_blk_col,
+      v_pre + uv_blk_col, uv_pre_stride, uv_width, uv_height, ss_x, ss_y,
+      strength, u_accum + uv_blk_col, u_count + uv_blk_col,
+      v_accum + uv_blk_col, v_count + uv_blk_col, y_dist + blk_col,
+      u_dist + uv_blk_col, v_dist + uv_blk_col, neighbors, top_weight,
+      bottom_weight, NULL);
+
+  blk_col += blk_col_step;
+  uv_blk_col += uv_blk_col_step;
+
+  // Middle First
+  if (ss_x && ss_y) {
+    neighbors = CHROMA_DOUBLE_SS_MIDDLE_COLUMN_NEIGHBORS;
+  } else if (ss_x || ss_y) {
+    neighbors = CHROMA_SINGLE_SS_MIDDLE_COLUMN_NEIGHBORS;
+  } else {
+    neighbors = CHROMA_NO_SS_MIDDLE_COLUMN_NEIGHBORS;
+  }
+
+  for (; uv_blk_col < uv_mid_width;
+       blk_col += blk_col_step, uv_blk_col += uv_blk_col_step) {
+    vp9_apply_temporal_filter_chroma_8(
+        y_src + blk_col, y_src_stride, y_pre + blk_col, y_pre_stride,
+        u_src + uv_blk_col, v_src + uv_blk_col, uv_src_stride,
+        u_pre + uv_blk_col, v_pre + uv_blk_col, uv_pre_stride, uv_width,
+        uv_height, ss_x, ss_y, strength, u_accum + uv_blk_col,
+        u_count + uv_blk_col, v_accum + uv_blk_col, v_count + uv_blk_col,
+        y_dist + blk_col, u_dist + uv_blk_col, v_dist + uv_blk_col, neighbors,
+        top_weight, bottom_weight, NULL);
+  }
+
+  if (!use_whole_blk) {
+    top_weight = blk_fw[1];
+    bottom_weight = blk_fw[3];
+  }
+
+  // Middle Second
+  for (; uv_blk_col < uv_last_width;
+       blk_col += blk_col_step, uv_blk_col += uv_blk_col_step) {
+    vp9_apply_temporal_filter_chroma_8(
+        y_src + blk_col, y_src_stride, y_pre + blk_col, y_pre_stride,
+        u_src + uv_blk_col, v_src + uv_blk_col, uv_src_stride,
+        u_pre + uv_blk_col, v_pre + uv_blk_col, uv_pre_stride, uv_width,
+        uv_height, ss_x, ss_y, strength, u_accum + uv_blk_col,
+        u_count + uv_blk_col, v_accum + uv_blk_col, v_count + uv_blk_col,
+        y_dist + blk_col, u_dist + uv_blk_col, v_dist + uv_blk_col, neighbors,
+        top_weight, bottom_weight, NULL);
+  }
+
+  // Right
+  if (ss_x && ss_y) {
+    neighbors = CHROMA_DOUBLE_SS_RIGHT_COLUMN_NEIGHBORS;
+  } else if (ss_x || ss_y) {
+    neighbors = CHROMA_SINGLE_SS_RIGHT_COLUMN_NEIGHBORS;
+  } else {
+    neighbors = CHROMA_NO_SS_RIGHT_COLUMN_NEIGHBORS;
+  }
+
+  vp9_apply_temporal_filter_chroma_8(
+      y_src + blk_col, y_src_stride, y_pre + blk_col, y_pre_stride,
+      u_src + uv_blk_col, v_src + uv_blk_col, uv_src_stride, u_pre + uv_blk_col,
+      v_pre + uv_blk_col, uv_pre_stride, uv_width, uv_height, ss_x, ss_y,
+      strength, u_accum + uv_blk_col, u_count + uv_blk_col,
+      v_accum + uv_blk_col, v_count + uv_blk_col, y_dist + blk_col,
+      u_dist + uv_blk_col, v_dist + uv_blk_col, neighbors, top_weight,
+      bottom_weight, NULL);
+}
+
+void vp9_apply_temporal_filter_sse4_1(
+    const uint8_t *y_src, int y_src_stride, const uint8_t *y_pre,
+    int y_pre_stride, const uint8_t *u_src, const uint8_t *v_src,
+    int uv_src_stride, const uint8_t *u_pre, const uint8_t *v_pre,
+    int uv_pre_stride, unsigned int block_width, unsigned int block_height,
+    int ss_x, int ss_y, int strength, const int *const blk_fw,
+    int use_whole_blk, uint32_t *y_accum, uint16_t *y_count, uint32_t *u_accum,
+    uint16_t *u_count, uint32_t *v_accum, uint16_t *v_count) {
+  const unsigned int chroma_height = block_height >> ss_y,
+                     chroma_width = block_width >> ss_x;
+
+  DECLARE_ALIGNED(16, uint16_t, y_dist[BH * DIST_STRIDE]) = { 0 };
+  DECLARE_ALIGNED(16, uint16_t, u_dist[BH * DIST_STRIDE]) = { 0 };
+  DECLARE_ALIGNED(16, uint16_t, v_dist[BH * DIST_STRIDE]) = { 0 };
+  const int *blk_fw_ptr = blk_fw;
+
+  uint16_t *y_dist_ptr = y_dist + 1, *u_dist_ptr = u_dist + 1,
+           *v_dist_ptr = v_dist + 1;
+  const uint8_t *y_src_ptr = y_src, *u_src_ptr = u_src, *v_src_ptr = v_src;
+  const uint8_t *y_pre_ptr = y_pre, *u_pre_ptr = u_pre, *v_pre_ptr = v_pre;
+
+  assert(block_width <= BW && "block width too large");
+  assert(block_height <= BH && "block height too large");
+  assert(block_width % 16 == 0 && "block width must be multiple of 16");
+  assert(block_height % 2 == 0 && "block height must be even");
+  assert((ss_x == 0 || ss_x == 1) && (ss_y == 0 || ss_y == 1) &&
+         "invalid chroma subsampling");
+  assert(strength >= 0 && strength <= 6 && "invalid temporal filter strength");
+  assert(blk_fw[0] >= 0 && "filter weight must be positive");
+  assert(
+      (use_whole_blk || (blk_fw[1] >= 0 && blk_fw[2] >= 0 && blk_fw[3] >= 0)) &&
+      "subblock filter weight must be positive");
+  assert(blk_fw[0] <= 2 && "sublock filter weight must be less than 2");
+  assert(
+      (use_whole_blk || (blk_fw[1] <= 2 && blk_fw[2] <= 2 && blk_fw[3] <= 2)) &&
+      "subblock filter weight must be less than 2");
+
+  // Precompute the difference sqaured
+  for (unsigned int row = 0; row < block_height; row++) {
+    for (unsigned int blk_col = 0; blk_col < block_width; blk_col += 16) {
+      store_dist_16(y_src_ptr + blk_col, y_pre_ptr + blk_col,
+                    y_dist_ptr + blk_col);
+    }
+    y_src_ptr += y_src_stride;
+    y_pre_ptr += y_pre_stride;
+    y_dist_ptr += DIST_STRIDE;
+  }
+
+  for (unsigned int row = 0; row < chroma_height; row++) {
+    for (unsigned int blk_col = 0; blk_col < chroma_width; blk_col += 8) {
+      store_dist_8(u_src_ptr + blk_col, u_pre_ptr + blk_col,
+                   u_dist_ptr + blk_col);
+      store_dist_8(v_src_ptr + blk_col, v_pre_ptr + blk_col,
+                   v_dist_ptr + blk_col);
+    }
+
+    u_src_ptr += uv_src_stride;
+    u_pre_ptr += uv_pre_stride;
+    u_dist_ptr += DIST_STRIDE;
+    v_src_ptr += uv_src_stride;
+    v_pre_ptr += uv_pre_stride;
+    v_dist_ptr += DIST_STRIDE;
+  }
+
+  y_dist_ptr = y_dist + 1;
+  u_dist_ptr = u_dist + 1;
+  v_dist_ptr = v_dist + 1;
+
+  vp9_apply_temporal_filter_luma(
+      y_src, y_src_stride, y_pre, y_pre_stride, u_src, v_src, uv_src_stride,
+      u_pre, v_pre, uv_pre_stride, block_width, block_height, ss_x, ss_y,
+      strength, blk_fw_ptr, use_whole_blk, y_accum, y_count, y_dist_ptr,
+      u_dist_ptr, v_dist_ptr);
+
+  vp9_apply_temporal_filter_chroma(
+      y_src, y_src_stride, y_pre, y_pre_stride, u_src, v_src, uv_src_stride,
+      u_pre, v_pre, uv_pre_stride, block_width, block_height, ss_x, ss_y,
+      strength, blk_fw_ptr, use_whole_blk, u_accum, u_count, v_accum, v_count,
+      y_dist_ptr, u_dist_ptr, v_dist_ptr);
+}
diff --git a/vp9/vp9cx.mk b/vp9/vp9cx.mk
index 05981d689..e16152a6d 100644
--- a/vp9/vp9cx.mk
+++ b/vp9/vp9cx.mk
@@ -103,6 +103,7 @@ VP9_CX_SRCS-yes += encoder/vp9_mbgraph.c
 VP9_CX_SRCS-yes += encoder/vp9_mbgraph.h
 
 VP9_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/temporal_filter_sse4.c
+VP9_CX_SRCS-$(HAVE_SSE4_1) += encoder/x86/temporal_filter_constants.h
 
 VP9_CX_SRCS-$(HAVE_SSE2) += encoder/x86/vp9_quantize_sse2.c
 VP9_CX_SRCS-$(HAVE_AVX2) += encoder/x86/vp9_quantize_avx2.c
@@ -149,5 +150,6 @@ VP9_CX_SRCS_REMOVE-$(CONFIG_REALTIME_ONLY) += encoder/vp9_firstpass.c
 VP9_CX_SRCS_REMOVE-$(CONFIG_REALTIME_ONLY) += encoder/vp9_mbgraph.c
 VP9_CX_SRCS_REMOVE-$(CONFIG_REALTIME_ONLY) += encoder/vp9_temporal_filter.c
 VP9_CX_SRCS_REMOVE-$(CONFIG_REALTIME_ONLY) += encoder/x86/temporal_filter_sse4.c
+VP9_CX_SRCS_REMOVE-$(CONFIG_REALTIME_ONLY) += encoder/x86/temporal_filter_constants.h
 
 VP9_CX_SRCS-yes := $(filter-out $(VP9_CX_SRCS_REMOVE-yes),$(VP9_CX_SRCS-yes))