Merge "Refactor lpf (size 4 and 8) NEON intrinsics optimization"

6 files changed, 435 insertions, 992 deletions

diff --git a/vpx_dsp/arm/loopfilter_16_neon.c b/vpx_dsp/arm/loopfilter_16_neon.c
deleted file mode 100644
index 9607bb240..000000000
--- a/vpx_dsp/arm/loopfilter_16_neon.c
+++ /dev/null
@@ -1,173 +0,0 @@
-/*
- *  Copyright (c) 2014 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 <arm_neon.h>
-
-#include "./vpx_dsp_rtcd.h"
-#include "./vpx_config.h"
-#include "vpx/vpx_integer.h"
-
-static INLINE void loop_filter_neon_16(uint8x16_t qblimit,  // blimit
-                                       uint8x16_t qlimit,   // limit
-                                       uint8x16_t qthresh,  // thresh
-                                       uint8x16_t q3,       // p3
-                                       uint8x16_t q4,       // p2
-                                       uint8x16_t q5,       // p1
-                                       uint8x16_t q6,       // p0
-                                       uint8x16_t q7,       // q0
-                                       uint8x16_t q8,       // q1
-                                       uint8x16_t q9,       // q2
-                                       uint8x16_t q10,      // q3
-                                       uint8x16_t *q5r,     // p1
-                                       uint8x16_t *q6r,     // p0
-                                       uint8x16_t *q7r,     // q0
-                                       uint8x16_t *q8r) {   // q1
-  uint8x16_t q1u8, q2u8, q11u8, q12u8, q13u8, q14u8, q15u8;
-  int16x8_t q2s16, q11s16;
-  uint16x8_t q4u16;
-  int8x16_t q0s8, q1s8, q2s8, q11s8, q12s8, q13s8;
-  int8x8_t d2s8, d3s8;
-
-  q11u8 = vabdq_u8(q3, q4);
-  q12u8 = vabdq_u8(q4, q5);
-  q13u8 = vabdq_u8(q5, q6);
-  q14u8 = vabdq_u8(q8, q7);
-  q3 = vabdq_u8(q9, q8);
-  q4 = vabdq_u8(q10, q9);
-
-  q11u8 = vmaxq_u8(q11u8, q12u8);
-  q12u8 = vmaxq_u8(q13u8, q14u8);
-  q3 = vmaxq_u8(q3, q4);
-  q15u8 = vmaxq_u8(q11u8, q12u8);
-
-  q9 = vabdq_u8(q6, q7);
-
-  // vp8_hevmask
-  q13u8 = vcgtq_u8(q13u8, qthresh);
-  q14u8 = vcgtq_u8(q14u8, qthresh);
-  q15u8 = vmaxq_u8(q15u8, q3);
-
-  q2u8 = vabdq_u8(q5, q8);
-  q9 = vqaddq_u8(q9, q9);
-
-  q15u8 = vcgeq_u8(qlimit, q15u8);
-
-  // vp8_filter() function
-  // convert to signed
-  q10 = vdupq_n_u8(0x80);
-  q8 = veorq_u8(q8, q10);
-  q7 = veorq_u8(q7, q10);
-  q6 = veorq_u8(q6, q10);
-  q5 = veorq_u8(q5, q10);
-
-  q2u8 = vshrq_n_u8(q2u8, 1);
-  q9 = vqaddq_u8(q9, q2u8);
-
-  q2s16 = vsubl_s8(vget_low_s8(vreinterpretq_s8_u8(q7)),
-                   vget_low_s8(vreinterpretq_s8_u8(q6)));
-  q11s16 = vsubl_s8(vget_high_s8(vreinterpretq_s8_u8(q7)),
-                    vget_high_s8(vreinterpretq_s8_u8(q6)));
-
-  q9 = vcgeq_u8(qblimit, q9);
-
-  q1s8 = vqsubq_s8(vreinterpretq_s8_u8(q5), vreinterpretq_s8_u8(q8));
-
-  q14u8 = vorrq_u8(q13u8, q14u8);
-
-  q4u16 = vdupq_n_u16(3);
-  q2s16 = vmulq_s16(q2s16, vreinterpretq_s16_u16(q4u16));
-  q11s16 = vmulq_s16(q11s16, vreinterpretq_s16_u16(q4u16));
-
-  q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q14u8);
-  q15u8 = vandq_u8(q15u8, q9);
-
-  q1s8 = vreinterpretq_s8_u8(q1u8);
-  q2s16 = vaddw_s8(q2s16, vget_low_s8(q1s8));
-  q11s16 = vaddw_s8(q11s16, vget_high_s8(q1s8));
-
-  q4 = vdupq_n_u8(3);
-  q9 = vdupq_n_u8(4);
-  // vp8_filter = clamp(vp8_filter + 3 * ( qs0 - ps0))
-  d2s8 = vqmovn_s16(q2s16);
-  d3s8 = vqmovn_s16(q11s16);
-  q1s8 = vcombine_s8(d2s8, d3s8);
-  q1u8 = vandq_u8(vreinterpretq_u8_s8(q1s8), q15u8);
-  q1s8 = vreinterpretq_s8_u8(q1u8);
-
-  q2s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q4));
-  q1s8 = vqaddq_s8(q1s8, vreinterpretq_s8_u8(q9));
-  q2s8 = vshrq_n_s8(q2s8, 3);
-  q1s8 = vshrq_n_s8(q1s8, 3);
-
-  q11s8 = vqaddq_s8(vreinterpretq_s8_u8(q6), q2s8);
-  q0s8 = vqsubq_s8(vreinterpretq_s8_u8(q7), q1s8);
-
-  q1s8 = vrshrq_n_s8(q1s8, 1);
-  q1s8 = vbicq_s8(q1s8, vreinterpretq_s8_u8(q14u8));
-
-  q13s8 = vqaddq_s8(vreinterpretq_s8_u8(q5), q1s8);
-  q12s8 = vqsubq_s8(vreinterpretq_s8_u8(q8), q1s8);
-
-  *q8r = veorq_u8(vreinterpretq_u8_s8(q12s8), q10);
-  *q7r = veorq_u8(vreinterpretq_u8_s8(q0s8), q10);
-  *q6r = veorq_u8(vreinterpretq_u8_s8(q11s8), q10);
-  *q5r = veorq_u8(vreinterpretq_u8_s8(q13s8), q10);
-  return;
-}
-
-void vpx_lpf_horizontal_4_dual_neon(
-    uint8_t *s, int p /* pitch */, const uint8_t *blimit0,
-    const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1,
-    const uint8_t *limit1, const uint8_t *thresh1) {
-  uint8x8_t dblimit0, dlimit0, dthresh0, dblimit1, dlimit1, dthresh1;
-  uint8x16_t qblimit, qlimit, qthresh;
-  uint8x16_t q3u8, q4u8, q5u8, q6u8, q7u8, q8u8, q9u8, q10u8;
-
-  dblimit0 = vld1_u8(blimit0);
-  dlimit0 = vld1_u8(limit0);
-  dthresh0 = vld1_u8(thresh0);
-  dblimit1 = vld1_u8(blimit1);
-  dlimit1 = vld1_u8(limit1);
-  dthresh1 = vld1_u8(thresh1);
-  qblimit = vcombine_u8(dblimit0, dblimit1);
-  qlimit = vcombine_u8(dlimit0, dlimit1);
-  qthresh = vcombine_u8(dthresh0, dthresh1);
-
-  s -= (p << 2);
-
-  q3u8 = vld1q_u8(s);
-  s += p;
-  q4u8 = vld1q_u8(s);
-  s += p;
-  q5u8 = vld1q_u8(s);
-  s += p;
-  q6u8 = vld1q_u8(s);
-  s += p;
-  q7u8 = vld1q_u8(s);
-  s += p;
-  q8u8 = vld1q_u8(s);
-  s += p;
-  q9u8 = vld1q_u8(s);
-  s += p;
-  q10u8 = vld1q_u8(s);
-
-  loop_filter_neon_16(qblimit, qlimit, qthresh, q3u8, q4u8, q5u8, q6u8, q7u8,
-                      q8u8, q9u8, q10u8, &q5u8, &q6u8, &q7u8, &q8u8);
-
-  s -= (p * 5);
-  vst1q_u8(s, q5u8);
-  s += p;
-  vst1q_u8(s, q6u8);
-  s += p;
-  vst1q_u8(s, q7u8);
-  s += p;
-  vst1q_u8(s, q8u8);
-  return;
-}
diff --git a/vpx_dsp/arm/loopfilter_4_neon.c b/vpx_dsp/arm/loopfilter_4_neon.c
deleted file mode 100644
index 1c1e80e00..000000000
--- a/vpx_dsp/arm/loopfilter_4_neon.c
+++ /dev/null
@@ -1,249 +0,0 @@
-/*
- *  Copyright (c) 2014 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 <arm_neon.h>
-
-#include "./vpx_dsp_rtcd.h"
-
-static INLINE void loop_filter_neon(uint8x8_t dblimit,   // flimit
-                                    uint8x8_t dlimit,    // limit
-                                    uint8x8_t dthresh,   // thresh
-                                    uint8x8_t d3u8,      // p3
-                                    uint8x8_t d4u8,      // p2
-                                    uint8x8_t d5u8,      // p1
-                                    uint8x8_t d6u8,      // p0
-                                    uint8x8_t d7u8,      // q0
-                                    uint8x8_t d16u8,     // q1
-                                    uint8x8_t d17u8,     // q2
-                                    uint8x8_t d18u8,     // q3
-                                    uint8x8_t *d4ru8,    // p1
-                                    uint8x8_t *d5ru8,    // p0
-                                    uint8x8_t *d6ru8,    // q0
-                                    uint8x8_t *d7ru8) {  // q1
-  uint8x8_t d19u8, d20u8, d21u8, d22u8, d23u8, d27u8, d28u8;
-  int16x8_t q12s16;
-  int8x8_t d19s8, d20s8, d21s8, d26s8, d27s8, d28s8;
-
-  d19u8 = vabd_u8(d3u8, d4u8);
-  d20u8 = vabd_u8(d4u8, d5u8);
-  d21u8 = vabd_u8(d5u8, d6u8);
-  d22u8 = vabd_u8(d16u8, d7u8);
-  d3u8 = vabd_u8(d17u8, d16u8);
-  d4u8 = vabd_u8(d18u8, d17u8);
-
-  d19u8 = vmax_u8(d19u8, d20u8);
-  d20u8 = vmax_u8(d21u8, d22u8);
-  d3u8 = vmax_u8(d3u8, d4u8);
-  d23u8 = vmax_u8(d19u8, d20u8);
-
-  d17u8 = vabd_u8(d6u8, d7u8);
-
-  d21u8 = vcgt_u8(d21u8, dthresh);
-  d22u8 = vcgt_u8(d22u8, dthresh);
-  d23u8 = vmax_u8(d23u8, d3u8);
-
-  d28u8 = vabd_u8(d5u8, d16u8);
-  d17u8 = vqadd_u8(d17u8, d17u8);
-
-  d23u8 = vcge_u8(dlimit, d23u8);
-
-  d18u8 = vdup_n_u8(0x80);
-  d5u8 = veor_u8(d5u8, d18u8);
-  d6u8 = veor_u8(d6u8, d18u8);
-  d7u8 = veor_u8(d7u8, d18u8);
-  d16u8 = veor_u8(d16u8, d18u8);
-
-  d28u8 = vshr_n_u8(d28u8, 1);
-  d17u8 = vqadd_u8(d17u8, d28u8);
-
-  d19u8 = vdup_n_u8(3);
-
-  d28s8 = vsub_s8(vreinterpret_s8_u8(d7u8), vreinterpret_s8_u8(d6u8));
-
-  d17u8 = vcge_u8(dblimit, d17u8);
-
-  d27s8 = vqsub_s8(vreinterpret_s8_u8(d5u8), vreinterpret_s8_u8(d16u8));
-
-  d22u8 = vorr_u8(d21u8, d22u8);
-
-  q12s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d19u8));
-
-  d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d22u8);
-  d23u8 = vand_u8(d23u8, d17u8);
-
-  q12s16 = vaddw_s8(q12s16, vreinterpret_s8_u8(d27u8));
-
-  d17u8 = vdup_n_u8(4);
-
-  d27s8 = vqmovn_s16(q12s16);
-  d27u8 = vand_u8(vreinterpret_u8_s8(d27s8), d23u8);
-  d27s8 = vreinterpret_s8_u8(d27u8);
-
-  d28s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d19u8));
-  d27s8 = vqadd_s8(d27s8, vreinterpret_s8_u8(d17u8));
-  d28s8 = vshr_n_s8(d28s8, 3);
-  d27s8 = vshr_n_s8(d27s8, 3);
-
-  d19s8 = vqadd_s8(vreinterpret_s8_u8(d6u8), d28s8);
-  d26s8 = vqsub_s8(vreinterpret_s8_u8(d7u8), d27s8);
-
-  d27s8 = vrshr_n_s8(d27s8, 1);
-  d27s8 = vbic_s8(d27s8, vreinterpret_s8_u8(d22u8));
-
-  d21s8 = vqadd_s8(vreinterpret_s8_u8(d5u8), d27s8);
-  d20s8 = vqsub_s8(vreinterpret_s8_u8(d16u8), d27s8);
-
-  *d4ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d18u8);
-  *d5ru8 = veor_u8(vreinterpret_u8_s8(d19s8), d18u8);
-  *d6ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d18u8);
-  *d7ru8 = veor_u8(vreinterpret_u8_s8(d20s8), d18u8);
-  return;
-}
-
-void vpx_lpf_horizontal_4_neon(uint8_t *src, int pitch, const uint8_t *blimit,
-                               const uint8_t *limit, const uint8_t *thresh) {
-  int i;
-  uint8_t *s, *psrc;
-  uint8x8_t dblimit, dlimit, dthresh;
-  uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
-
-  dblimit = vld1_u8(blimit);
-  dlimit = vld1_u8(limit);
-  dthresh = vld1_u8(thresh);
-
-  psrc = src - (pitch << 2);
-  for (i = 0; i < 1; i++) {
-    s = psrc + i * 8;
-
-    d3u8 = vld1_u8(s);
-    s += pitch;
-    d4u8 = vld1_u8(s);
-    s += pitch;
-    d5u8 = vld1_u8(s);
-    s += pitch;
-    d6u8 = vld1_u8(s);
-    s += pitch;
-    d7u8 = vld1_u8(s);
-    s += pitch;
-    d16u8 = vld1_u8(s);
-    s += pitch;
-    d17u8 = vld1_u8(s);
-    s += pitch;
-    d18u8 = vld1_u8(s);
-
-    loop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
-                     d16u8, d17u8, d18u8, &d4u8, &d5u8, &d6u8, &d7u8);
-
-    s -= (pitch * 5);
-    vst1_u8(s, d4u8);
-    s += pitch;
-    vst1_u8(s, d5u8);
-    s += pitch;
-    vst1_u8(s, d6u8);
-    s += pitch;
-    vst1_u8(s, d7u8);
-  }
-  return;
-}
-
-void vpx_lpf_vertical_4_neon(uint8_t *src, int pitch, const uint8_t *blimit,
-                             const uint8_t *limit, const uint8_t *thresh) {
-  int i, pitch8;
-  uint8_t *s;
-  uint8x8_t dblimit, dlimit, dthresh;
-  uint8x8_t d3u8, d4u8, d5u8, d6u8, d7u8, d16u8, d17u8, d18u8;
-  uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
-  uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
-  uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
-  uint8x8x4_t d4Result;
-
-  dblimit = vld1_u8(blimit);
-  dlimit = vld1_u8(limit);
-  dthresh = vld1_u8(thresh);
-
-  pitch8 = pitch * 8;
-  for (i = 0; i < 1; i++, src += pitch8) {
-    s = src - (i + 1) * 4;
-
-    d3u8 = vld1_u8(s);
-    s += pitch;
-    d4u8 = vld1_u8(s);
-    s += pitch;
-    d5u8 = vld1_u8(s);
-    s += pitch;
-    d6u8 = vld1_u8(s);
-    s += pitch;
-    d7u8 = vld1_u8(s);
-    s += pitch;
-    d16u8 = vld1_u8(s);
-    s += pitch;
-    d17u8 = vld1_u8(s);
-    s += pitch;
-    d18u8 = vld1_u8(s);
-
-    d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8), vreinterpret_u32_u8(d7u8));
-    d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8), vreinterpret_u32_u8(d16u8));
-    d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8), vreinterpret_u32_u8(d17u8));
-    d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8), vreinterpret_u32_u8(d18u8));
-
-    d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
-                      vreinterpret_u16_u32(d2tmp2.val[0]));
-    d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
-                      vreinterpret_u16_u32(d2tmp3.val[0]));
-    d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
-                      vreinterpret_u16_u32(d2tmp2.val[1]));
-    d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
-                      vreinterpret_u16_u32(d2tmp3.val[1]));
-
-    d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
-                     vreinterpret_u8_u16(d2tmp5.val[0]));
-    d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
-                     vreinterpret_u8_u16(d2tmp5.val[1]));
-    d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
-                      vreinterpret_u8_u16(d2tmp7.val[0]));
-    d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
-                      vreinterpret_u8_u16(d2tmp7.val[1]));
-
-    d3u8 = d2tmp8.val[0];
-    d4u8 = d2tmp8.val[1];
-    d5u8 = d2tmp9.val[0];
-    d6u8 = d2tmp9.val[1];
-    d7u8 = d2tmp10.val[0];
-    d16u8 = d2tmp10.val[1];
-    d17u8 = d2tmp11.val[0];
-    d18u8 = d2tmp11.val[1];
-
-    loop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
-                     d16u8, d17u8, d18u8, &d4u8, &d5u8, &d6u8, &d7u8);
-
-    d4Result.val[0] = d4u8;
-    d4Result.val[1] = d5u8;
-    d4Result.val[2] = d6u8;
-    d4Result.val[3] = d7u8;
-
-    src -= 2;
-    vst4_lane_u8(src, d4Result, 0);
-    src += pitch;
-    vst4_lane_u8(src, d4Result, 1);
-    src += pitch;
-    vst4_lane_u8(src, d4Result, 2);
-    src += pitch;
-    vst4_lane_u8(src, d4Result, 3);
-    src += pitch;
-    vst4_lane_u8(src, d4Result, 4);
-    src += pitch;
-    vst4_lane_u8(src, d4Result, 5);
-    src += pitch;
-    vst4_lane_u8(src, d4Result, 6);
-    src += pitch;
-    vst4_lane_u8(src, d4Result, 7);
-  }
-  return;
-}
diff --git a/vpx_dsp/arm/loopfilter_8_neon.c b/vpx_dsp/arm/loopfilter_8_neon.c
deleted file mode 100644
index 8641541b0..000000000
--- a/vpx_dsp/arm/loopfilter_8_neon.c
+++ /dev/null
@@ -1,445 +0,0 @@
-/*
- *  Copyright (c) 2014 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 <arm_neon.h>
-
-#include "./vpx_dsp_rtcd.h"
-
-static INLINE void mbloop_filter_neon(uint8x8_t dblimit,   // mblimit
-                                      uint8x8_t dlimit,    // limit
-                                      uint8x8_t dthresh,   // thresh
-                                      uint8x8_t d3u8,      // p2
-                                      uint8x8_t d4u8,      // p2
-                                      uint8x8_t d5u8,      // p1
-                                      uint8x8_t d6u8,      // p0
-                                      uint8x8_t d7u8,      // q0
-                                      uint8x8_t d16u8,     // q1
-                                      uint8x8_t d17u8,     // q2
-                                      uint8x8_t d18u8,     // q3
-                                      uint8x8_t *d0ru8,    // p1
-                                      uint8x8_t *d1ru8,    // p1
-                                      uint8x8_t *d2ru8,    // p0
-                                      uint8x8_t *d3ru8,    // q0
-                                      uint8x8_t *d4ru8,    // q1
-                                      uint8x8_t *d5ru8) {  // q1
-  uint32_t flat;
-  uint8x8_t d0u8, d1u8, d2u8, d19u8, d20u8, d21u8, d22u8, d23u8, d24u8;
-  uint8x8_t d25u8, d26u8, d27u8, d28u8, d29u8, d30u8, d31u8;
-  int16x8_t q15s16;
-  uint16x8_t q10u16, q14u16;
-  int8x8_t d21s8, d24s8, d25s8, d26s8, d28s8, d29s8, d30s8;
-
-  d19u8 = vabd_u8(d3u8, d4u8);
-  d20u8 = vabd_u8(d4u8, d5u8);
-  d21u8 = vabd_u8(d5u8, d6u8);
-  d22u8 = vabd_u8(d16u8, d7u8);
-  d23u8 = vabd_u8(d17u8, d16u8);
-  d24u8 = vabd_u8(d18u8, d17u8);
-
-  d19u8 = vmax_u8(d19u8, d20u8);
-  d20u8 = vmax_u8(d21u8, d22u8);
-
-  d25u8 = vabd_u8(d6u8, d4u8);
-
-  d23u8 = vmax_u8(d23u8, d24u8);
-
-  d26u8 = vabd_u8(d7u8, d17u8);
-
-  d19u8 = vmax_u8(d19u8, d20u8);
-
-  d24u8 = vabd_u8(d6u8, d7u8);
-  d27u8 = vabd_u8(d3u8, d6u8);
-  d28u8 = vabd_u8(d18u8, d7u8);
-
-  d19u8 = vmax_u8(d19u8, d23u8);
-
-  d23u8 = vabd_u8(d5u8, d16u8);
-  d24u8 = vqadd_u8(d24u8, d24u8);
-
-  d19u8 = vcge_u8(dlimit, d19u8);
-
-  d25u8 = vmax_u8(d25u8, d26u8);
-  d26u8 = vmax_u8(d27u8, d28u8);
-
-  d23u8 = vshr_n_u8(d23u8, 1);
-
-  d25u8 = vmax_u8(d25u8, d26u8);
-
-  d24u8 = vqadd_u8(d24u8, d23u8);
-
-  d20u8 = vmax_u8(d20u8, d25u8);
-
-  d23u8 = vdup_n_u8(1);
-  d24u8 = vcge_u8(dblimit, d24u8);
-
-  d21u8 = vcgt_u8(d21u8, dthresh);
-
-  d20u8 = vcge_u8(d23u8, d20u8);
-
-  d19u8 = vand_u8(d19u8, d24u8);
-
-  d23u8 = vcgt_u8(d22u8, dthresh);
-
-  d20u8 = vand_u8(d20u8, d19u8);
-
-  d22u8 = vdup_n_u8(0x80);
-
-  d23u8 = vorr_u8(d21u8, d23u8);
-
-  q10u16 = vcombine_u16(vreinterpret_u16_u8(d20u8), vreinterpret_u16_u8(d21u8));
-
-  d30u8 = vshrn_n_u16(q10u16, 4);
-  flat = vget_lane_u32(vreinterpret_u32_u8(d30u8), 0);
-
-  if (flat == 0xffffffff) {  // Check for all 1's, power_branch_only
-    d27u8 = vdup_n_u8(3);
-    d21u8 = vdup_n_u8(2);
-    q14u16 = vaddl_u8(d6u8, d7u8);
-    q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
-    q14u16 = vmlal_u8(q14u16, d4u8, d21u8);
-    q14u16 = vaddw_u8(q14u16, d5u8);
-    *d0ru8 = vqrshrn_n_u16(q14u16, 3);
-
-    q14u16 = vsubw_u8(q14u16, d3u8);
-    q14u16 = vsubw_u8(q14u16, d4u8);
-    q14u16 = vaddw_u8(q14u16, d5u8);
-    q14u16 = vaddw_u8(q14u16, d16u8);
-    *d1ru8 = vqrshrn_n_u16(q14u16, 3);
-
-    q14u16 = vsubw_u8(q14u16, d3u8);
-    q14u16 = vsubw_u8(q14u16, d5u8);
-    q14u16 = vaddw_u8(q14u16, d6u8);
-    q14u16 = vaddw_u8(q14u16, d17u8);
-    *d2ru8 = vqrshrn_n_u16(q14u16, 3);
-
-    q14u16 = vsubw_u8(q14u16, d3u8);
-    q14u16 = vsubw_u8(q14u16, d6u8);
-    q14u16 = vaddw_u8(q14u16, d7u8);
-    q14u16 = vaddw_u8(q14u16, d18u8);
-    *d3ru8 = vqrshrn_n_u16(q14u16, 3);
-
-    q14u16 = vsubw_u8(q14u16, d4u8);
-    q14u16 = vsubw_u8(q14u16, d7u8);
-    q14u16 = vaddw_u8(q14u16, d16u8);
-    q14u16 = vaddw_u8(q14u16, d18u8);
-    *d4ru8 = vqrshrn_n_u16(q14u16, 3);
-
-    q14u16 = vsubw_u8(q14u16, d5u8);
-    q14u16 = vsubw_u8(q14u16, d16u8);
-    q14u16 = vaddw_u8(q14u16, d17u8);
-    q14u16 = vaddw_u8(q14u16, d18u8);
-    *d5ru8 = vqrshrn_n_u16(q14u16, 3);
-  } else {
-    d21u8 = veor_u8(d7u8, d22u8);
-    d24u8 = veor_u8(d6u8, d22u8);
-    d25u8 = veor_u8(d5u8, d22u8);
-    d26u8 = veor_u8(d16u8, d22u8);
-
-    d27u8 = vdup_n_u8(3);
-
-    d28s8 = vsub_s8(vreinterpret_s8_u8(d21u8), vreinterpret_s8_u8(d24u8));
-    d29s8 = vqsub_s8(vreinterpret_s8_u8(d25u8), vreinterpret_s8_u8(d26u8));
-
-    q15s16 = vmull_s8(d28s8, vreinterpret_s8_u8(d27u8));
-
-    d29s8 = vand_s8(d29s8, vreinterpret_s8_u8(d23u8));
-
-    q15s16 = vaddw_s8(q15s16, d29s8);
-
-    d29u8 = vdup_n_u8(4);
-
-    d28s8 = vqmovn_s16(q15s16);
-
-    d28s8 = vand_s8(d28s8, vreinterpret_s8_u8(d19u8));
-
-    d30s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d27u8));
-    d29s8 = vqadd_s8(d28s8, vreinterpret_s8_u8(d29u8));
-    d30s8 = vshr_n_s8(d30s8, 3);
-    d29s8 = vshr_n_s8(d29s8, 3);
-
-    d24s8 = vqadd_s8(vreinterpret_s8_u8(d24u8), d30s8);
-    d21s8 = vqsub_s8(vreinterpret_s8_u8(d21u8), d29s8);
-
-    d29s8 = vrshr_n_s8(d29s8, 1);
-    d29s8 = vbic_s8(d29s8, vreinterpret_s8_u8(d23u8));
-
-    d25s8 = vqadd_s8(vreinterpret_s8_u8(d25u8), d29s8);
-    d26s8 = vqsub_s8(vreinterpret_s8_u8(d26u8), d29s8);
-
-    if (flat == 0) {  // filter_branch_only
-      *d0ru8 = d4u8;
-      *d1ru8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
-      *d2ru8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
-      *d3ru8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
-      *d4ru8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
-      *d5ru8 = d17u8;
-      return;
-    }
-
-    d21u8 = veor_u8(vreinterpret_u8_s8(d21s8), d22u8);
-    d24u8 = veor_u8(vreinterpret_u8_s8(d24s8), d22u8);
-    d25u8 = veor_u8(vreinterpret_u8_s8(d25s8), d22u8);
-    d26u8 = veor_u8(vreinterpret_u8_s8(d26s8), d22u8);
-
-    d23u8 = vdup_n_u8(2);
-    q14u16 = vaddl_u8(d6u8, d7u8);
-    q14u16 = vmlal_u8(q14u16, d3u8, d27u8);
-    q14u16 = vmlal_u8(q14u16, d4u8, d23u8);
-
-    d0u8 = vbsl_u8(d20u8, dblimit, d4u8);
-
-    q14u16 = vaddw_u8(q14u16, d5u8);
-
-    d1u8 = vbsl_u8(d20u8, dlimit, d25u8);
-
-    d30u8 = vqrshrn_n_u16(q14u16, 3);
-
-    q14u16 = vsubw_u8(q14u16, d3u8);
-    q14u16 = vsubw_u8(q14u16, d4u8);
-    q14u16 = vaddw_u8(q14u16, d5u8);
-    q14u16 = vaddw_u8(q14u16, d16u8);
-
-    d2u8 = vbsl_u8(d20u8, dthresh, d24u8);
-
-    d31u8 = vqrshrn_n_u16(q14u16, 3);
-
-    q14u16 = vsubw_u8(q14u16, d3u8);
-    q14u16 = vsubw_u8(q14u16, d5u8);
-    q14u16 = vaddw_u8(q14u16, d6u8);
-    q14u16 = vaddw_u8(q14u16, d17u8);
-
-    *d0ru8 = vbsl_u8(d20u8, d30u8, d0u8);
-
-    d23u8 = vqrshrn_n_u16(q14u16, 3);
-
-    q14u16 = vsubw_u8(q14u16, d3u8);
-    q14u16 = vsubw_u8(q14u16, d6u8);
-    q14u16 = vaddw_u8(q14u16, d7u8);
-
-    *d1ru8 = vbsl_u8(d20u8, d31u8, d1u8);
-
-    q14u16 = vaddw_u8(q14u16, d18u8);
-
-    *d2ru8 = vbsl_u8(d20u8, d23u8, d2u8);
-
-    d22u8 = vqrshrn_n_u16(q14u16, 3);
-
-    q14u16 = vsubw_u8(q14u16, d4u8);
-    q14u16 = vsubw_u8(q14u16, d7u8);
-    q14u16 = vaddw_u8(q14u16, d16u8);
-
-    d3u8 = vbsl_u8(d20u8, d3u8, d21u8);
-
-    q14u16 = vaddw_u8(q14u16, d18u8);
-
-    d4u8 = vbsl_u8(d20u8, d4u8, d26u8);
-
-    d6u8 = vqrshrn_n_u16(q14u16, 3);
-
-    q14u16 = vsubw_u8(q14u16, d5u8);
-    q14u16 = vsubw_u8(q14u16, d16u8);
-    q14u16 = vaddw_u8(q14u16, d17u8);
-    q14u16 = vaddw_u8(q14u16, d18u8);
-
-    d5u8 = vbsl_u8(d20u8, d5u8, d17u8);
-
-    d7u8 = vqrshrn_n_u16(q14u16, 3);
-
-    *d3ru8 = vbsl_u8(d20u8, d22u8, d3u8);
-    *d4ru8 = vbsl_u8(d20u8, d6u8, d4u8);
-    *d5ru8 = vbsl_u8(d20u8, d7u8, d5u8);
-  }
-  return;
-}
-
-void vpx_lpf_horizontal_8_neon(uint8_t *src, int pitch, const uint8_t *blimit,
-                               const uint8_t *limit, const uint8_t *thresh) {
-  int i;
-  uint8_t *s, *psrc;
-  uint8x8_t dblimit, dlimit, dthresh;
-  uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
-  uint8x8_t d16u8, d17u8, d18u8;
-
-  dblimit = vld1_u8(blimit);
-  dlimit = vld1_u8(limit);
-  dthresh = vld1_u8(thresh);
-
-  psrc = src - (pitch << 2);
-  for (i = 0; i < 1; i++) {
-    s = psrc + i * 8;
-
-    d3u8 = vld1_u8(s);
-    s += pitch;
-    d4u8 = vld1_u8(s);
-    s += pitch;
-    d5u8 = vld1_u8(s);
-    s += pitch;
-    d6u8 = vld1_u8(s);
-    s += pitch;
-    d7u8 = vld1_u8(s);
-    s += pitch;
-    d16u8 = vld1_u8(s);
-    s += pitch;
-    d17u8 = vld1_u8(s);
-    s += pitch;
-    d18u8 = vld1_u8(s);
-
-    mbloop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
-                       d16u8, d17u8, d18u8, &d0u8, &d1u8, &d2u8, &d3u8, &d4u8,
-                       &d5u8);
-
-    s -= (pitch * 6);
-    vst1_u8(s, d0u8);
-    s += pitch;
-    vst1_u8(s, d1u8);
-    s += pitch;
-    vst1_u8(s, d2u8);
-    s += pitch;
-    vst1_u8(s, d3u8);
-    s += pitch;
-    vst1_u8(s, d4u8);
-    s += pitch;
-    vst1_u8(s, d5u8);
-  }
-  return;
-}
-
-void vpx_lpf_horizontal_8_dual_neon(
-    uint8_t *s, int p /* pitch */, const uint8_t *blimit0,
-    const uint8_t *limit0, const uint8_t *thresh0, const uint8_t *blimit1,
-    const uint8_t *limit1, const uint8_t *thresh1) {
-  vpx_lpf_horizontal_8_neon(s, p, blimit0, limit0, thresh0);
-  vpx_lpf_horizontal_8_neon(s + 8, p, blimit1, limit1, thresh1);
-}
-
-void vpx_lpf_vertical_8_neon(uint8_t *src, int pitch, const uint8_t *blimit,
-                             const uint8_t *limit, const uint8_t *thresh) {
-  int i;
-  uint8_t *s;
-  uint8x8_t dblimit, dlimit, dthresh;
-  uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
-  uint8x8_t d16u8, d17u8, d18u8;
-  uint32x2x2_t d2tmp0, d2tmp1, d2tmp2, d2tmp3;
-  uint16x4x2_t d2tmp4, d2tmp5, d2tmp6, d2tmp7;
-  uint8x8x2_t d2tmp8, d2tmp9, d2tmp10, d2tmp11;
-  uint8x8x4_t d4Result;
-  uint8x8x2_t d2Result;
-
-  dblimit = vld1_u8(blimit);
-  dlimit = vld1_u8(limit);
-  dthresh = vld1_u8(thresh);
-
-  for (i = 0; i < 1; i++) {
-    s = src + (i * (pitch << 3)) - 4;
-
-    d3u8 = vld1_u8(s);
-    s += pitch;
-    d4u8 = vld1_u8(s);
-    s += pitch;
-    d5u8 = vld1_u8(s);
-    s += pitch;
-    d6u8 = vld1_u8(s);
-    s += pitch;
-    d7u8 = vld1_u8(s);
-    s += pitch;
-    d16u8 = vld1_u8(s);
-    s += pitch;
-    d17u8 = vld1_u8(s);
-    s += pitch;
-    d18u8 = vld1_u8(s);
-
-    d2tmp0 = vtrn_u32(vreinterpret_u32_u8(d3u8), vreinterpret_u32_u8(d7u8));
-    d2tmp1 = vtrn_u32(vreinterpret_u32_u8(d4u8), vreinterpret_u32_u8(d16u8));
-    d2tmp2 = vtrn_u32(vreinterpret_u32_u8(d5u8), vreinterpret_u32_u8(d17u8));
-    d2tmp3 = vtrn_u32(vreinterpret_u32_u8(d6u8), vreinterpret_u32_u8(d18u8));
-
-    d2tmp4 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[0]),
-                      vreinterpret_u16_u32(d2tmp2.val[0]));
-    d2tmp5 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[0]),
-                      vreinterpret_u16_u32(d2tmp3.val[0]));
-    d2tmp6 = vtrn_u16(vreinterpret_u16_u32(d2tmp0.val[1]),
-                      vreinterpret_u16_u32(d2tmp2.val[1]));
-    d2tmp7 = vtrn_u16(vreinterpret_u16_u32(d2tmp1.val[1]),
-                      vreinterpret_u16_u32(d2tmp3.val[1]));
-
-    d2tmp8 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[0]),
-                     vreinterpret_u8_u16(d2tmp5.val[0]));
-    d2tmp9 = vtrn_u8(vreinterpret_u8_u16(d2tmp4.val[1]),
-                     vreinterpret_u8_u16(d2tmp5.val[1]));
-    d2tmp10 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[0]),
-                      vreinterpret_u8_u16(d2tmp7.val[0]));
-    d2tmp11 = vtrn_u8(vreinterpret_u8_u16(d2tmp6.val[1]),
-                      vreinterpret_u8_u16(d2tmp7.val[1]));
-
-    d3u8 = d2tmp8.val[0];
-    d4u8 = d2tmp8.val[1];
-    d5u8 = d2tmp9.val[0];
-    d6u8 = d2tmp9.val[1];
-    d7u8 = d2tmp10.val[0];
-    d16u8 = d2tmp10.val[1];
-    d17u8 = d2tmp11.val[0];
-    d18u8 = d2tmp11.val[1];
-
-    mbloop_filter_neon(dblimit, dlimit, dthresh, d3u8, d4u8, d5u8, d6u8, d7u8,
-                       d16u8, d17u8, d18u8, &d0u8, &d1u8, &d2u8, &d3u8, &d4u8,
-                       &d5u8);
-
-    d4Result.val[0] = d0u8;
-    d4Result.val[1] = d1u8;
-    d4Result.val[2] = d2u8;
-    d4Result.val[3] = d3u8;
-
-    d2Result.val[0] = d4u8;
-    d2Result.val[1] = d5u8;
-
-    s = src - 3;
-    vst4_lane_u8(s, d4Result, 0);
-    s += pitch;
-    vst4_lane_u8(s, d4Result, 1);
-    s += pitch;
-    vst4_lane_u8(s, d4Result, 2);
-    s += pitch;
-    vst4_lane_u8(s, d4Result, 3);
-    s += pitch;
-    vst4_lane_u8(s, d4Result, 4);
-    s += pitch;
-    vst4_lane_u8(s, d4Result, 5);
-    s += pitch;
-    vst4_lane_u8(s, d4Result, 6);
-    s += pitch;
-    vst4_lane_u8(s, d4Result, 7);
-
-    s = src + 1;
-    vst2_lane_u8(s, d2Result, 0);
-    s += pitch;
-    vst2_lane_u8(s, d2Result, 1);
-    s += pitch;
-    vst2_lane_u8(s, d2Result, 2);
-    s += pitch;
-    vst2_lane_u8(s, d2Result, 3);
-    s += pitch;
-    vst2_lane_u8(s, d2Result, 4);
-    s += pitch;
-    vst2_lane_u8(s, d2Result, 5);
-    s += pitch;
-    vst2_lane_u8(s, d2Result, 6);
-    s += pitch;
-    vst2_lane_u8(s, d2Result, 7);
-  }
-  return;
-}
-
-void vpx_lpf_vertical_8_dual_neon(uint8_t *s, int p, const uint8_t *blimit0,
-                                  const uint8_t *limit0, const uint8_t *thresh0,
-                                  const uint8_t *blimit1, const uint8_t *limit1,
-                                  const uint8_t *thresh1) {
-  vpx_lpf_vertical_8_neon(s, p, blimit0, limit0, thresh0);
-  vpx_lpf_vertical_8_neon(s + 8 * p, p, blimit1, limit1, thresh1);
-}
diff --git a/vpx_dsp/arm/loopfilter_mb_neon.c b/vpx_dsp/arm/loopfilter_mb_neon.c
index aa61220d3..f95267472 100644
--- a/vpx_dsp/arm/loopfilter_mb_neon.c
+++ b/vpx_dsp/arm/loopfilter_mb_neon.c
@@ -31,6 +31,15 @@ FUN_LOAD_THRESH(8, _)    // load_thresh_8
 FUN_LOAD_THRESH(16, q_)  // load_thresh_16
 #undef FUN_LOAD_THRESH
 
+static INLINE void load_thresh_8_dual(
+    const uint8_t *blimit0, const uint8_t *limit0, const uint8_t *thresh0,
+    const uint8_t *blimit1, const uint8_t *limit1, const uint8_t *thresh1,
+    uint8x16_t *blimit_vec, uint8x16_t *limit_vec, uint8x16_t *thresh_vec) {
+  *blimit_vec = vcombine_u8(vld1_dup_u8(blimit0), vld1_dup_u8(blimit1));
+  *limit_vec = vcombine_u8(vld1_dup_u8(limit0), vld1_dup_u8(limit1));
+  *thresh_vec = vcombine_u8(vld1_dup_u8(thresh0), vld1_dup_u8(thresh1));
+}
+
 // Here flat is 64-bit long, with each 8-bit (or 4-bit) chunk being a mask of a
 // pixel. When used to control filter branches, we only detect whether it is all
 // 0s or all 1s. We pairwise add flat to a 32-bit long number flat_status.
@@ -56,33 +65,51 @@ static INLINE uint32_t calc_flat_status_16(uint8x16_t flat) {
   return calc_flat_status_8(flat_4bit);
 }
 
-#define FUN_FILTER_FLAT_HEV_MASK(w, r)                                        \
-  static INLINE uint8x##w##_t filter_flat_hev_mask_##w(                       \
+#define FUN_FILTER_HEV_MASK4(w, r)                                            \
+  static INLINE uint8x##w##_t filter_hev_mask4_##w(                           \
       const uint8x##w##_t limit, const uint8x##w##_t blimit,                  \
       const uint8x##w##_t thresh, const uint8x##w##_t p3,                     \
       const uint8x##w##_t p2, const uint8x##w##_t p1, const uint8x##w##_t p0, \
       const uint8x##w##_t q0, const uint8x##w##_t q1, const uint8x##w##_t q2, \
-      const uint8x##w##_t q3, uint8x##w##_t *flat, uint32_t *flat_status,     \
-      uint8x##w##_t *hev) {                                                   \
-    uint8x##w##_t t0, t1, mask;                                               \
+      const uint8x##w##_t q3, uint8x##w##_t *hev, uint8x##w##_t *mask) {      \
+    uint8x##w##_t max, t0, t1;                                                \
                                                                               \
-    mask = vabd##r##u8(p1, p0);                                               \
-    mask = vmax##r##u8(mask, vabd##r##u8(q1, q0));                            \
-    *hev = vcgt##r##u8(mask, thresh);                                         \
-    *flat = vmax##r##u8(mask, vabd##r##u8(p2, p0));                           \
-    mask = vmax##r##u8(mask, vabd##r##u8(p3, p2));                            \
-    mask = vmax##r##u8(mask, vabd##r##u8(p2, p1));                            \
-    mask = vmax##r##u8(mask, vabd##r##u8(q2, q1));                            \
-    mask = vmax##r##u8(mask, vabd##r##u8(q3, q2));                            \
+    max = vabd##r##u8(p1, p0);                                                \
+    max = vmax##r##u8(max, vabd##r##u8(q1, q0));                              \
+    *hev = vcgt##r##u8(max, thresh);                                          \
+    *mask = vmax##r##u8(max, vabd##r##u8(p3, p2));                            \
+    *mask = vmax##r##u8(*mask, vabd##r##u8(p2, p1));                          \
+    *mask = vmax##r##u8(*mask, vabd##r##u8(q2, q1));                          \
+    *mask = vmax##r##u8(*mask, vabd##r##u8(q3, q2));                          \
     t0 = vabd##r##u8(p0, q0);                                                 \
     t1 = vabd##r##u8(p1, q1);                                                 \
     t0 = vqadd##r##u8(t0, t0);                                                \
     t1 = vshr##r##n_u8(t1, 1);                                                \
     t0 = vqadd##r##u8(t0, t1);                                                \
-    mask = vcle##r##u8(mask, limit);                                          \
+    *mask = vcle##r##u8(*mask, limit);                                        \
     t0 = vcle##r##u8(t0, blimit);                                             \
-    mask = vand##r##u8(mask, t0);                                             \
+    *mask = vand##r##u8(*mask, t0);                                           \
+                                                                              \
+    return max;                                                               \
+  }
+
+FUN_FILTER_HEV_MASK4(8, _)    // filter_hev_mask4_8
+FUN_FILTER_HEV_MASK4(16, q_)  // filter_hev_mask4_16
+#undef FUN_FILTER_HEV_MASK4
+
+#define FUN_FILTER_FLAT_HEV_MASK(w, r)                                        \
+  static INLINE uint8x##w##_t filter_flat_hev_mask_##w(                       \
+      const uint8x##w##_t limit, const uint8x##w##_t blimit,                  \
+      const uint8x##w##_t thresh, const uint8x##w##_t p3,                     \
+      const uint8x##w##_t p2, const uint8x##w##_t p1, const uint8x##w##_t p0, \
+      const uint8x##w##_t q0, const uint8x##w##_t q1, const uint8x##w##_t q2, \
+      const uint8x##w##_t q3, uint8x##w##_t *flat, uint32_t *flat_status,     \
+      uint8x##w##_t *hev) {                                                   \
+    uint8x##w##_t max, mask;                                                  \
                                                                               \
+    max = filter_hev_mask4_##w(limit, blimit, thresh, p3, p2, p1, p0, q0, q1, \
+                               q2, q3, hev, &mask);                           \
+    *flat = vmax##r##u8(max, vabd##r##u8(p2, p0));                            \
     *flat = vmax##r##u8(*flat, vabd##r##u8(q2, q0));                          \
     *flat = vmax##r##u8(*flat, vabd##r##u8(p3, p0));                          \
     *flat = vmax##r##u8(*flat, vabd##r##u8(q3, q0));                          \
@@ -420,6 +447,33 @@ FUN_FILTER4(8, _)    // filter4_8
 FUN_FILTER4(16, q_)  // filter4_16
 #undef FUN_FILTER4
 
+#define FUN_FILTER8(w)                                                         \
+  static INLINE void filter8_##w(                                              \
+      const uint8x##w##_t mask, const uint8x##w##_t flat,                      \
+      const uint32_t flat_status, const uint8x##w##_t hev,                     \
+      const uint8x##w##_t p3, const uint8x##w##_t p2, const uint8x##w##_t p1,  \
+      const uint8x##w##_t p0, const uint8x##w##_t q0, const uint8x##w##_t q1,  \
+      const uint8x##w##_t q2, const uint8x##w##_t q3, uint8x##w##_t *op2,      \
+      uint8x##w##_t *op1, uint8x##w##_t *op0, uint8x##w##_t *oq0,              \
+      uint8x##w##_t *oq1, uint8x##w##_t *oq2) {                                \
+    if (flat_status != (uint32_t)-2) {                                         \
+      filter4_##w(mask, hev, p1, p0, q0, q1, op1, op0, oq0, oq1);              \
+      *op2 = p2;                                                               \
+      *oq2 = q2;                                                               \
+      if (flat_status) {                                                       \
+        apply_7_tap_filter_##w(flat, p3, p2, p1, p0, q0, q1, q2, q3, op2, op1, \
+                               op0, oq0, oq1, oq2);                            \
+      }                                                                        \
+    } else {                                                                   \
+      calc_7_tap_filter_##w(p3, p2, p1, p0, q0, q1, q2, q3, op2, op1, op0,     \
+                            oq0, oq1, oq2);                                    \
+    }                                                                          \
+  }
+
+FUN_FILTER8(8)   // filter8_8
+FUN_FILTER8(16)  // filter8_16
+#undef FUN_FILTER8
+
 #define FUN_FILTER16(w)                                                        \
   static INLINE void filter16_##w(                                             \
       const uint8x##w##_t mask, const uint8x##w##_t flat,                      \
@@ -481,6 +535,7 @@ FUN_FILTER16(16)  // filter16_16
     *q3 = vld1##r##u8(s);                                                  \
   }
 
+FUN_LOAD8(8, _)    // load_8x8
 FUN_LOAD8(16, q_)  // load_16x8
 #undef FUN_LOAD8
 
@@ -529,6 +584,71 @@ FUN_LOAD16(8, _)    // load_8x16
 FUN_LOAD16(16, q_)  // load_16x16
 #undef FUN_LOAD16
 
+#define FUN_STORE4(w, r)                                                       \
+  static INLINE void store_##w##x4(                                            \
+      uint8_t *s, const int p, const uint8x##w##_t s0, const uint8x##w##_t s1, \
+      const uint8x##w##_t s2, const uint8x##w##_t s3) {                        \
+    vst1##r##u8(s, s0);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s1);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s2);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s3);                                                        \
+  }
+
+FUN_STORE4(8, _)    // store_8x4
+FUN_STORE4(16, q_)  // store_16x4
+#undef FUN_STORE4
+
+#define FUN_STORE6(w, r)                                                       \
+  static INLINE void store_##w##x6(                                            \
+      uint8_t *s, const int p, const uint8x##w##_t s0, const uint8x##w##_t s1, \
+      const uint8x##w##_t s2, const uint8x##w##_t s3, const uint8x##w##_t s4,  \
+      const uint8x##w##_t s5) {                                                \
+    vst1##r##u8(s, s0);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s1);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s2);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s3);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s4);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s5);                                                        \
+  }
+
+FUN_STORE6(8, _)    // store_8x6
+FUN_STORE6(16, q_)  // store_16x6
+#undef FUN_STORE6
+
+static INLINE void store_4x8(uint8_t *s, const int p, const uint8x8_t p1,
+                             const uint8x8_t p0, const uint8x8_t q0,
+                             const uint8x8_t q1) {
+  uint8x8x4_t o;
+
+  o.val[0] = p1;
+  o.val[1] = p0;
+  o.val[2] = q0;
+  o.val[3] = q1;
+  vst4_lane_u8(s, o, 0);
+  s += p;
+  vst4_lane_u8(s, o, 1);
+  s += p;
+  vst4_lane_u8(s, o, 2);
+  s += p;
+  vst4_lane_u8(s, o, 3);
+  s += p;
+  vst4_lane_u8(s, o, 4);
+  s += p;
+  vst4_lane_u8(s, o, 5);
+  s += p;
+  vst4_lane_u8(s, o, 6);
+  s += p;
+  vst4_lane_u8(s, o, 7);
+}
+
 static INLINE void store_6x8(uint8_t *s, const int p, const uint8x8_t s0,
                              const uint8x8_t s1, const uint8x8_t s2,
                              const uint8x8_t s3, const uint8x8_t s4,
@@ -566,53 +686,64 @@ static INLINE void store_6x8(uint8_t *s, const int p, const uint8x8_t s0,
   vst3_lane_u8(s + 0, o1, 7);
 }
 
-static INLINE void store_4x8(uint8_t *s, const int p, const uint8x8_t p1,
-                             const uint8x8_t p0, const uint8x8_t q0,
-                             const uint8x8_t q1) {
-  uint8x8x4_t o;
+#define FUN_STORE8(w, r)                                                       \
+  static INLINE void store_##w##x8(                                            \
+      uint8_t *s, const int p, const uint8x##w##_t s0, const uint8x##w##_t s1, \
+      const uint8x##w##_t s2, const uint8x##w##_t s3, const uint8x##w##_t s4,  \
+      const uint8x##w##_t s5, const uint8x##w##_t s6,                          \
+      const uint8x##w##_t s7) {                                                \
+    vst1##r##u8(s, s0);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s1);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s2);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s3);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s4);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s5);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s6);                                                        \
+    s += p;                                                                    \
+    vst1##r##u8(s, s7);                                                        \
+  }
 
-  o.val[0] = p1;
-  o.val[1] = p0;
-  o.val[2] = q0;
-  o.val[3] = q1;
-  vst4_lane_u8(s, o, 0);
-  s += p;
-  vst4_lane_u8(s, o, 1);
-  s += p;
-  vst4_lane_u8(s, o, 2);
-  s += p;
-  vst4_lane_u8(s, o, 3);
-  s += p;
-  vst4_lane_u8(s, o, 4);
-  s += p;
-  vst4_lane_u8(s, o, 5);
-  s += p;
-  vst4_lane_u8(s, o, 6);
-  s += p;
-  vst4_lane_u8(s, o, 7);
-}
+FUN_STORE8(8, _)    // store_8x8
+FUN_STORE8(16, q_)  // store_16x8
+#undef FUN_STORE8
 
-static INLINE void store_16x8(uint8_t *s, const int p, const uint8x16_t s0,
-                              const uint8x16_t s1, const uint8x16_t s2,
-                              const uint8x16_t s3, const uint8x16_t s4,
-                              const uint8x16_t s5, const uint8x16_t s6,
-                              const uint8x16_t s7) {
-  vst1q_u8(s, s0);
-  s += p;
-  vst1q_u8(s, s1);
-  s += p;
-  vst1q_u8(s, s2);
-  s += p;
-  vst1q_u8(s, s3);
-  s += p;
-  vst1q_u8(s, s4);
-  s += p;
-  vst1q_u8(s, s5);
-  s += p;
-  vst1q_u8(s, s6);
-  s += p;
-  vst1q_u8(s, s7);
-}
+#define FUN_STORE14(w, r)                                                      \
+  static INLINE void store_##w##x14(                                           \
+      uint8_t *s, const int p, const uint8x##w##_t p6, const uint8x##w##_t p5, \
+      const uint8x##w##_t p4, const uint8x##w##_t p3, const uint8x##w##_t p2,  \
+      const uint8x##w##_t p1, const uint8x##w##_t p0, const uint8x##w##_t q0,  \
+      const uint8x##w##_t q1, const uint8x##w##_t q2, const uint8x##w##_t q3,  \
+      const uint8x##w##_t q4, const uint8x##w##_t q5, const uint8x##w##_t q6,  \
+      const uint32_t flat_status, const uint32_t flat2_status) {               \
+    if (flat_status) {                                                         \
+      if (flat2_status) {                                                      \
+        vst1##r##u8(s - 7 * p, p6);                                            \
+        vst1##r##u8(s - 6 * p, p5);                                            \
+        vst1##r##u8(s - 5 * p, p4);                                            \
+        vst1##r##u8(s - 4 * p, p3);                                            \
+        vst1##r##u8(s + 3 * p, q3);                                            \
+        vst1##r##u8(s + 4 * p, q4);                                            \
+        vst1##r##u8(s + 5 * p, q5);                                            \
+        vst1##r##u8(s + 6 * p, q6);                                            \
+      }                                                                        \
+      vst1##r##u8(s - 3 * p, p2);                                              \
+      vst1##r##u8(s + 2 * p, q2);                                              \
+    }                                                                          \
+    vst1##r##u8(s - 2 * p, p1);                                                \
+    vst1##r##u8(s - 1 * p, p0);                                                \
+    vst1##r##u8(s + 0 * p, q0);                                                \
+    vst1##r##u8(s + 1 * p, q1);                                                \
+  }
+
+FUN_STORE14(8, _)    // store_8x14
+FUN_STORE14(16, q_)  // store_16x14
+#undef FUN_STORE14
 
 static INLINE void store_16x16(uint8_t *s, const int p, const uint8x16_t s0,
                                const uint8x16_t s1, const uint8x16_t s2,
@@ -656,37 +787,160 @@ static INLINE void store_16x16(uint8_t *s, const int p, const uint8x16_t s0,
   vst1q_u8(s, s15);
 }
 
-#define FUN_STORE14(w, r)                                                      \
-  static INLINE void store_##w##x14(                                           \
-      uint8_t *s, const int p, const uint8x##w##_t p6, const uint8x##w##_t p5, \
-      const uint8x##w##_t p4, const uint8x##w##_t p3, const uint8x##w##_t p2,  \
-      const uint8x##w##_t p1, const uint8x##w##_t p0, const uint8x##w##_t q0,  \
-      const uint8x##w##_t q1, const uint8x##w##_t q2, const uint8x##w##_t q3,  \
-      const uint8x##w##_t q4, const uint8x##w##_t q5, const uint8x##w##_t q6,  \
-      const uint32_t flat_status, const uint32_t flat2_status) {               \
-    if (flat_status) {                                                         \
-      if (flat2_status) {                                                      \
-        vst1##r##u8(s - 7 * p, p6);                                            \
-        vst1##r##u8(s - 6 * p, p5);                                            \
-        vst1##r##u8(s - 5 * p, p4);                                            \
-        vst1##r##u8(s - 4 * p, p3);                                            \
-        vst1##r##u8(s + 3 * p, q3);                                            \
-        vst1##r##u8(s + 4 * p, q4);                                            \
-        vst1##r##u8(s + 5 * p, q5);                                            \
-        vst1##r##u8(s + 6 * p, q6);                                            \
-      }                                                                        \
-      vst1##r##u8(s - 3 * p, p2);                                              \
-      vst1##r##u8(s + 2 * p, q2);                                              \
-    }                                                                          \
-    vst1##r##u8(s - 2 * p, p1);                                                \
-    vst1##r##u8(s - 1 * p, p0);                                                \
-    vst1##r##u8(s + 0 * p, q0);                                                \
-    vst1##r##u8(s + 1 * p, q1);                                                \
+#define FUN_HOR_4_KERNEL(name, w)                                           \
+  static INLINE void lpf_horizontal_4##name##kernel(                        \
+      uint8_t *s, const int p, const uint8x##w##_t blimit,                  \
+      const uint8x##w##_t limit, const uint8x##w##_t thresh) {              \
+    uint8x##w##_t p3, p2, p1, p0, q0, q1, q2, q3, mask, hev;                \
+                                                                            \
+    load_##w##x8(s - 4 * p, p, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);     \
+    filter_hev_mask4_##w(limit, blimit, thresh, p3, p2, p1, p0, q0, q1, q2, \
+                         q3, &hev, &mask);                                  \
+    filter4_##w(mask, hev, p1, p0, q0, q1, &p1, &p0, &q0, &q1);             \
+    store_##w##x4(s - 2 * p, p, p1, p0, q0, q1);                            \
   }
 
-FUN_STORE14(8, _)    // store_8x14
-FUN_STORE14(16, q_)  // store_16x14
-#undef FUN_STORE14
+FUN_HOR_4_KERNEL(_, 8)        // lpf_horizontal_4_kernel
+FUN_HOR_4_KERNEL(_dual_, 16)  // lpf_horizontal_4_dual_kernel
+#undef FUN_HOR_4_KERNEL
+
+void vpx_lpf_horizontal_4_neon(uint8_t *s, int p, const uint8_t *blimit,
+                               const uint8_t *limit, const uint8_t *thresh) {
+  uint8x8_t blimit_vec, limit_vec, thresh_vec;
+  load_thresh_8(blimit, limit, thresh, &blimit_vec, &limit_vec, &thresh_vec);
+  lpf_horizontal_4_kernel(s, p, blimit_vec, limit_vec, thresh_vec);
+}
+
+void vpx_lpf_horizontal_4_dual_neon(uint8_t *s, int p, const uint8_t *blimit0,
+                                    const uint8_t *limit0,
+                                    const uint8_t *thresh0,
+                                    const uint8_t *blimit1,
+                                    const uint8_t *limit1,
+                                    const uint8_t *thresh1) {
+  uint8x16_t blimit_vec, limit_vec, thresh_vec;
+  load_thresh_8_dual(blimit0, limit0, thresh0, blimit1, limit1, thresh1,
+                     &blimit_vec, &limit_vec, &thresh_vec);
+  lpf_horizontal_4_dual_kernel(s, p, blimit_vec, limit_vec, thresh_vec);
+}
+
+void vpx_lpf_vertical_4_neon(uint8_t *s, int p, const uint8_t *blimit,
+                             const uint8_t *limit, const uint8_t *thresh) {
+  uint8x8_t blimit_vec, limit_vec, thresh_vec, p3, p2, p1, p0, q0, q1, q2, q3,
+      mask, hev;
+  load_thresh_8(blimit, limit, thresh, &blimit_vec, &limit_vec, &thresh_vec);
+  load_8x8(s - 4, p, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  transpose_u8_8x8(&p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  filter_hev_mask4_8(limit_vec, blimit_vec, thresh_vec, p3, p2, p1, p0, q0, q1,
+                     q2, q3, &hev, &mask);
+  filter4_8(mask, hev, p1, p0, q0, q1, &p1, &p0, &q0, &q1);
+  store_4x8(s - 2, p, p1, p0, q0, q1);
+}
+
+void vpx_lpf_vertical_4_dual_neon(uint8_t *s, int p, const uint8_t *blimit0,
+                                  const uint8_t *limit0, const uint8_t *thresh0,
+                                  const uint8_t *blimit1, const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  uint8x16_t blimit_vec, limit_vec, thresh_vec, p3, p2, p1, p0, q0, q1, q2, q3,
+      mask, hev;
+  uint8x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, s14,
+      s15;
+
+  load_thresh_8_dual(blimit0, limit0, thresh0, blimit1, limit1, thresh1,
+                     &blimit_vec, &limit_vec, &thresh_vec);
+  load_8x16(s - 4, p, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7, &s8, &s9, &s10,
+            &s11, &s12, &s13, &s14, &s15);
+  transpose_u8_8x16(s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13,
+                    s14, s15, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  filter_hev_mask4_16(limit_vec, blimit_vec, thresh_vec, p3, p2, p1, p0, q0, q1,
+                      q2, q3, &hev, &mask);
+  filter4_16(mask, hev, p1, p0, q0, q1, &p1, &p0, &q0, &q1);
+  s -= 2;
+  store_4x8(s, p, vget_low_u8(p1), vget_low_u8(p0), vget_low_u8(q0),
+            vget_low_u8(q1));
+  store_4x8(s + 8 * p, p, vget_high_u8(p1), vget_high_u8(p0), vget_high_u8(q0),
+            vget_high_u8(q1));
+}
+
+void vpx_lpf_horizontal_8_neon(uint8_t *s, int p, const uint8_t *blimit,
+                               const uint8_t *limit, const uint8_t *thresh) {
+  uint8x8_t blimit_vec, limit_vec, thresh_vec, p3, p2, p1, p0, q0, q1, q2, q3,
+      op2, op1, op0, oq0, oq1, oq2, mask, flat, hev;
+  uint32_t flat_status;
+
+  load_thresh_8(blimit, limit, thresh, &blimit_vec, &limit_vec, &thresh_vec);
+  load_8x8(s - 4 * p, p, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  mask = filter_flat_hev_mask_8(limit_vec, blimit_vec, thresh_vec, p3, p2, p1,
+                                p0, q0, q1, q2, q3, &flat, &flat_status, &hev);
+  filter8_8(mask, flat, flat_status, hev, p3, p2, p1, p0, q0, q1, q2, q3, &op2,
+            &op1, &op0, &oq0, &oq1, &oq2);
+  store_8x6(s - 3 * p, p, op2, op1, op0, oq0, oq1, oq2);
+}
+
+void vpx_lpf_horizontal_8_dual_neon(uint8_t *s, int p, const uint8_t *blimit0,
+                                    const uint8_t *limit0,
+                                    const uint8_t *thresh0,
+                                    const uint8_t *blimit1,
+                                    const uint8_t *limit1,
+                                    const uint8_t *thresh1) {
+  uint8x16_t blimit_vec, limit_vec, thresh_vec, p3, p2, p1, p0, q0, q1, q2, q3,
+      op2, op1, op0, oq0, oq1, oq2, mask, flat, hev;
+  uint32_t flat_status;
+
+  load_thresh_8_dual(blimit0, limit0, thresh0, blimit1, limit1, thresh1,
+                     &blimit_vec, &limit_vec, &thresh_vec);
+  load_16x8(s - 4 * p, p, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  mask = filter_flat_hev_mask_16(limit_vec, blimit_vec, thresh_vec, p3, p2, p1,
+                                 p0, q0, q1, q2, q3, &flat, &flat_status, &hev);
+  filter8_16(mask, flat, flat_status, hev, p3, p2, p1, p0, q0, q1, q2, q3, &op2,
+             &op1, &op0, &oq0, &oq1, &oq2);
+  store_16x6(s - 3 * p, p, op2, op1, op0, oq0, oq1, oq2);
+}
+
+void vpx_lpf_vertical_8_neon(uint8_t *s, int p, const uint8_t *blimit,
+                             const uint8_t *limit, const uint8_t *thresh) {
+  uint8x8_t blimit_vec, limit_vec, thresh_vec, p3, p2, p1, p0, q0, q1, q2, q3,
+      op2, op1, op0, oq0, oq1, oq2, mask, flat, hev;
+  uint32_t flat_status;
+
+  load_thresh_8(blimit, limit, thresh, &blimit_vec, &limit_vec, &thresh_vec);
+  load_8x8(s - 4, p, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  transpose_u8_8x8(&p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  mask = filter_flat_hev_mask_8(limit_vec, blimit_vec, thresh_vec, p3, p2, p1,
+                                p0, q0, q1, q2, q3, &flat, &flat_status, &hev);
+  filter8_8(mask, flat, flat_status, hev, p3, p2, p1, p0, q0, q1, q2, q3, &op2,
+            &op1, &op0, &oq0, &oq1, &oq2);
+  // Note: tranpose + store_8x8() is faster than store_6x8().
+  transpose_u8_8x8(&p3, &op2, &op1, &op0, &oq0, &oq1, &oq2, &q3);
+  store_8x8(s - 4, p, p3, op2, op1, op0, oq0, oq1, oq2, q3);
+}
+
+void vpx_lpf_vertical_8_dual_neon(uint8_t *s, int p, const uint8_t *blimit0,
+                                  const uint8_t *limit0, const uint8_t *thresh0,
+                                  const uint8_t *blimit1, const uint8_t *limit1,
+                                  const uint8_t *thresh1) {
+  uint8x16_t blimit_vec, limit_vec, thresh_vec, p3, p2, p1, p0, q0, q1, q2, q3,
+      op2, op1, op0, oq0, oq1, oq2, mask, flat, hev;
+  uint8x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, s14,
+      s15;
+  uint32_t flat_status;
+
+  load_thresh_8_dual(blimit0, limit0, thresh0, blimit1, limit1, thresh1,
+                     &blimit_vec, &limit_vec, &thresh_vec);
+  load_8x16(s - 4, p, &s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7, &s8, &s9, &s10,
+            &s11, &s12, &s13, &s14, &s15);
+  transpose_u8_8x16(s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13,
+                    s14, s15, &p3, &p2, &p1, &p0, &q0, &q1, &q2, &q3);
+  mask = filter_flat_hev_mask_16(limit_vec, blimit_vec, thresh_vec, p3, p2, p1,
+                                 p0, q0, q1, q2, q3, &flat, &flat_status, &hev);
+  filter8_16(mask, flat, flat_status, hev, p3, p2, p1, p0, q0, q1, q2, q3, &op2,
+             &op1, &op0, &oq0, &oq1, &oq2);
+  // Note: store_6x8() twice is faster than tranpose + store_8x16().
+  store_6x8(s, p, vget_low_u8(op2), vget_low_u8(op1), vget_low_u8(op0),
+            vget_low_u8(oq0), vget_low_u8(oq1), vget_low_u8(oq2));
+  store_6x8(s + 8 * p, p, vget_high_u8(op2), vget_high_u8(op1),
+            vget_high_u8(op0), vget_high_u8(oq0), vget_high_u8(oq1),
+            vget_high_u8(oq2));
+}
 
 #define FUN_LPF_16_KERNEL(name, w)                                             \
   static INLINE void lpf_16##name##kernel(                                     \
@@ -784,7 +1038,9 @@ void vpx_lpf_vertical_16_neon(uint8_t *s, int p, const uint8_t *blimit,
                         &s6, &s7);
       store_16x8(s, p, s0, s1, s2, s3, s4, s5, s6, s7);
     } else {
-      store_6x8(s + 8, p, op2, op1, op0, oq0, oq1, oq2);
+      // Note: tranpose + store_8x8() is faster than store_6x8().
+      transpose_u8_8x8(&p3, &op2, &op1, &op0, &oq0, &oq1, &oq2, &q3);
+      store_8x8(s + 4, p, p3, op2, op1, op0, oq0, oq1, oq2, q3);
     }
   } else {
     store_4x8(s + 6, p, op1, op0, oq0, oq1);
@@ -819,6 +1075,7 @@ void vpx_lpf_vertical_16_dual_neon(uint8_t *s, int p, const uint8_t *blimit,
       store_16x16(s, p, s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12,
                   s13, s14, s15);
     } else {
+      // Note: store_6x8() twice is faster than tranpose + store_8x16().
       s += 8;
       store_6x8(s, p, vget_low_u8(op2), vget_low_u8(op1), vget_low_u8(op0),
                 vget_low_u8(oq0), vget_low_u8(oq1), vget_low_u8(oq2));
diff --git a/vpx_dsp/arm/transpose_neon.h b/vpx_dsp/arm/transpose_neon.h
index 2318fb44b..3d0b41f93 100644
--- a/vpx_dsp/arm/transpose_neon.h
+++ b/vpx_dsp/arm/transpose_neon.h
@@ -39,28 +39,84 @@ static INLINE uint8x16x2_t vpx_vtrnq_u64(uint32x4_t a0, uint32x4_t a1) {
   return b0;
 }
 
+// Note: Using 'd' registers or 'q' registers has almost identical speed. We use
+// 'q' registers here to save some instructions.
+static INLINE void transpose_u8_8x8(uint8x8_t *a0, uint8x8_t *a1, uint8x8_t *a2,
+                                    uint8x8_t *a3, uint8x8_t *a4, uint8x8_t *a5,
+                                    uint8x8_t *a6, uint8x8_t *a7) {
+  // Swap 8 bit elements. Goes from:
+  // a0: 00 01 02 03 04 05 06 07
+  // a1: 10 11 12 13 14 15 16 17
+  // a2: 20 21 22 23 24 25 26 27
+  // a3: 30 31 32 33 34 35 36 37
+  // a4: 40 41 42 43 44 45 46 47
+  // a5: 50 51 52 53 54 55 56 57
+  // a6: 60 61 62 63 64 65 66 67
+  // a7: 70 71 72 73 74 75 76 77
+  // to:
+  // b0.val[0]: 00 10 02 12 04 14 06 16  40 50 42 52 44 54 46 56
+  // b0.val[1]: 01 11 03 13 05 15 07 17  41 51 43 53 45 55 47 57
+  // b1.val[0]: 20 30 22 32 24 34 26 36  60 70 62 72 64 74 66 76
+  // b1.val[1]: 21 31 23 33 25 35 27 37  61 71 63 73 65 75 67 77
+
+  const uint8x16x2_t b0 =
+      vtrnq_u8(vcombine_u8(*a0, *a4), vcombine_u8(*a1, *a5));
+  const uint8x16x2_t b1 =
+      vtrnq_u8(vcombine_u8(*a2, *a6), vcombine_u8(*a3, *a7));
+
+  // Swap 16 bit elements resulting in:
+  // c0.val[0]: 00 10 20 30 04 14 24 34  40 50 60 70 44 54 64 74
+  // c0.val[1]: 02 12 22 32 06 16 26 36  42 52 62 72 46 56 66 76
+  // c1.val[0]: 01 11 21 31 05 15 25 35  41 51 61 71 45 55 65 75
+  // c1.val[1]: 03 13 23 33 07 17 27 37  43 53 63 73 47 57 67 77
+
+  const uint16x8x2_t c0 = vtrnq_u16(vreinterpretq_u16_u8(b0.val[0]),
+                                    vreinterpretq_u16_u8(b1.val[0]));
+  const uint16x8x2_t c1 = vtrnq_u16(vreinterpretq_u16_u8(b0.val[1]),
+                                    vreinterpretq_u16_u8(b1.val[1]));
+
+  // Unzip 32 bit elements resulting in:
+  // d0.val[0]: 00 10 20 30 40 50 60 70  01 11 21 31 41 51 61 71
+  // d0.val[1]: 04 14 24 34 44 54 64 74  05 15 25 35 45 55 65 75
+  // d1.val[0]: 02 12 22 32 42 52 62 72  03 13 23 33 43 53 63 73
+  // d1.val[1]: 06 16 26 36 46 56 66 76  07 17 27 37 47 57 67 77
+  const uint32x4x2_t d0 = vuzpq_u32(vreinterpretq_u32_u16(c0.val[0]),
+                                    vreinterpretq_u32_u16(c1.val[0]));
+  const uint32x4x2_t d1 = vuzpq_u32(vreinterpretq_u32_u16(c0.val[1]),
+                                    vreinterpretq_u32_u16(c1.val[1]));
+
+  *a0 = vreinterpret_u8_u32(vget_low_u32(d0.val[0]));
+  *a1 = vreinterpret_u8_u32(vget_high_u32(d0.val[0]));
+  *a2 = vreinterpret_u8_u32(vget_low_u32(d1.val[0]));
+  *a3 = vreinterpret_u8_u32(vget_high_u32(d1.val[0]));
+  *a4 = vreinterpret_u8_u32(vget_low_u32(d0.val[1]));
+  *a5 = vreinterpret_u8_u32(vget_high_u32(d0.val[1]));
+  *a6 = vreinterpret_u8_u32(vget_low_u32(d1.val[1]));
+  *a7 = vreinterpret_u8_u32(vget_high_u32(d1.val[1]));
+}
+
 static INLINE void transpose_s16_8x8(int16x8_t *a0, int16x8_t *a1,
                                      int16x8_t *a2, int16x8_t *a3,
                                      int16x8_t *a4, int16x8_t *a5,
                                      int16x8_t *a6, int16x8_t *a7) {
   // Swap 16 bit elements. Goes from:
   // a0: 00 01 02 03 04 05 06 07
-  // a1: 08 09 10 11 12 13 14 15
-  // a2: 16 17 18 19 20 21 22 23
-  // a3: 24 25 26 27 28 29 30 31
-  // a4: 32 33 34 35 36 37 38 39
-  // a5: 40 41 42 43 44 45 46 47
-  // a6: 48 49 50 51 52 53 54 55
-  // a7: 56 57 58 59 60 61 62 63
+  // a1: 10 11 12 13 14 15 16 17
+  // a2: 20 21 22 23 24 25 26 27
+  // a3: 30 31 32 33 34 35 36 37
+  // a4: 40 41 42 43 44 45 46 47
+  // a5: 50 51 52 53 54 55 56 57
+  // a6: 60 61 62 63 64 65 66 67
+  // a7: 70 71 72 73 74 75 76 77
   // to:
-  // b0.val[0]: 00 08 02 10 04 12 06 14
-  // b0.val[1]: 01 09 03 11 05 13 07 15
-  // b1.val[0]: 16 24 18 26 20 28 22 30
-  // b1.val[1]: 17 25 19 27 21 29 23 31
-  // b2.val[0]: 32 40 34 42 36 44 38 46
-  // b2.val[1]: 33 41 35 43 37 45 39 47
-  // b3.val[0]: 48 56 50 58 52 60 54 62
-  // b3.val[1]: 49 57 51 59 53 61 55 63
+  // b0.val[0]: 00 10 02 12 04 14 06 16
+  // b0.val[1]: 01 11 03 13 05 15 07 17
+  // b1.val[0]: 20 30 22 32 24 34 26 36
+  // b1.val[1]: 21 31 23 33 25 35 27 37
+  // b2.val[0]: 40 50 42 52 44 54 46 56
+  // b2.val[1]: 41 51 43 53 45 55 47 57
+  // b3.val[0]: 60 70 62 72 64 74 66 76
+  // b3.val[1]: 61 71 63 73 65 75 67 77
 
   const int16x8x2_t b0 = vtrnq_s16(*a0, *a1);
   const int16x8x2_t b1 = vtrnq_s16(*a2, *a3);
@@ -68,14 +124,14 @@ static INLINE void transpose_s16_8x8(int16x8_t *a0, int16x8_t *a1,
   const int16x8x2_t b3 = vtrnq_s16(*a6, *a7);
 
   // Swap 32 bit elements resulting in:
-  // c0.val[0]: 00 08 16 24 04 12 20 28
-  // c0.val[1]: 02 10 18 26 06 14 22 30
-  // c1.val[0]: 01 09 17 25 05 13 21 29
-  // c1.val[1]: 03 11 19 27 07 15 23 31
-  // c2.val[0]: 32 40 48 56 36 44 52 60
-  // c2.val[1]: 34 42 50 58 38 46 54 62
-  // c3.val[0]: 33 41 49 57 37 45 53 61
-  // c3.val[1]: 35 43 51 59 39 47 55 63
+  // c0.val[0]: 00 10 20 30 04 14 24 34
+  // c0.val[1]: 02 12 22 32 06 16 26 36
+  // c1.val[0]: 01 11 21 31 05 15 25 35
+  // c1.val[1]: 03 13 23 33 07 17 27 37
+  // c2.val[0]: 40 50 60 70 44 54 64 74
+  // c2.val[1]: 42 52 62 72 46 56 66 76
+  // c3.val[0]: 41 51 61 71 45 55 65 75
+  // c3.val[1]: 43 53 63 73 47 57 67 77
 
   const int32x4x2_t c0 = vtrnq_s32(vreinterpretq_s32_s16(b0.val[0]),
                                    vreinterpretq_s32_s16(b1.val[0]));
@@ -87,14 +143,14 @@ static INLINE void transpose_s16_8x8(int16x8_t *a0, int16x8_t *a1,
                                    vreinterpretq_s32_s16(b3.val[1]));
 
   // Swap 64 bit elements resulting in:
-  // d0.val[0]: 00 08 16 24 32 40 48 56
-  // d0.val[1]: 04 12 20 28 36 44 52 60
-  // d1.val[0]: 01 09 17 25 33 41 49 57
-  // d1.val[1]: 05 13 21 29 37 45 53 61
-  // d2.val[0]: 02 10 18 26 34 42 50 58
-  // d2.val[1]: 06 14 22 30 38 46 54 62
-  // d3.val[0]: 03 11 19 27 35 43 51 59
-  // d3.val[1]: 07 15 23 31 39 47 55 63
+  // d0.val[0]: 00 10 20 30 40 50 60 70
+  // d0.val[1]: 04 14 24 34 44 54 64 74
+  // d1.val[0]: 01 11 21 31 41 51 61 71
+  // d1.val[1]: 05 15 25 35 45 55 65 75
+  // d2.val[0]: 02 12 22 32 42 52 62 72
+  // d2.val[1]: 06 16 26 36 46 56 66 76
+  // d3.val[0]: 03 13 23 33 43 53 63 73
+  // d3.val[1]: 07 17 27 37 47 57 67 77
   const int16x8x2_t d0 = vpx_vtrnq_s64(c0.val[0], c2.val[0]);
   const int16x8x2_t d1 = vpx_vtrnq_s64(c1.val[0], c3.val[0]);
   const int16x8x2_t d2 = vpx_vtrnq_s64(c0.val[1], c2.val[1]);
diff --git a/vpx_dsp/vpx_dsp.mk b/vpx_dsp/vpx_dsp.mk
index def9c8e1b..a78041ce7 100644
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -136,8 +136,8 @@ DSP_SRCS-yes += loopfilter.c
 DSP_SRCS-$(ARCH_X86)$(ARCH_X86_64)   += x86/loopfilter_sse2.c
 DSP_SRCS-$(HAVE_AVX2)                += x86/loopfilter_avx2.c
 
-DSP_SRCS-$(HAVE_NEON)   += arm/loopfilter_neon.c
 ifeq ($(HAVE_NEON_ASM),yes)
+DSP_SRCS-yes  += arm/loopfilter_neon.c
 DSP_SRCS-yes  += arm/loopfilter_mb_neon$(ASM)
 DSP_SRCS-yes  += arm/loopfilter_16_neon$(ASM)
 DSP_SRCS-yes  += arm/loopfilter_8_neon$(ASM)
@@ -145,9 +145,6 @@ DSP_SRCS-yes  += arm/loopfilter_4_neon$(ASM)
 else
 ifeq ($(HAVE_NEON),yes)
 DSP_SRCS-yes   += arm/loopfilter_mb_neon.c
-DSP_SRCS-yes   += arm/loopfilter_16_neon.c
-DSP_SRCS-yes   += arm/loopfilter_8_neon.c
-DSP_SRCS-yes   += arm/loopfilter_4_neon.c
 endif  # HAVE_NEON
 endif  # HAVE_NEON_ASM