shithub: libvpx

Download patch

ref: 65e9fb65e824a58ac3827218e3b418f9944e9e85
parent: e54231d613c70c38f857a968fef9d91a2e4bb10a
author: Linfeng Zhang <linfengz@google.com>
date: Tue Mar 14 06:16:35 EDT 2017

Add vpx_highbd_idct32x32_135_add_neon()

BUG=webm:1301

Change-Id: I58c2d65d385080711c3666d6d8f9d241dac7b21a

--- a/test/partial_idct_test.cc
+++ b/test/partial_idct_test.cc
@@ -441,6 +441,15 @@
 #if CONFIG_VP9_HIGHBITDEPTH
   make_tuple(
       &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
+      &highbd_wrapper<vpx_highbd_idct32x32_135_add_neon>, TX_32X32, 135, 8, 2),
+  make_tuple(
+      &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
+      &highbd_wrapper<vpx_highbd_idct32x32_135_add_neon>, TX_32X32, 135, 10, 2),
+  make_tuple(
+      &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
+      &highbd_wrapper<vpx_highbd_idct32x32_135_add_neon>, TX_32X32, 135, 12, 2),
+  make_tuple(
+      &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
       &highbd_wrapper<vpx_highbd_idct32x32_1_add_neon>, TX_32X32, 1, 8, 2),
   make_tuple(
       &vpx_highbd_fdct32x32_c, &highbd_wrapper<vpx_highbd_idct32x32_1024_add_c>,
--- a/vpx_dsp/arm/highbd_idct16x16_add_neon.c
+++ b/vpx_dsp/arm/highbd_idct16x16_add_neon.c
@@ -592,62 +592,6 @@
   vst1q_s32(output + 4, out[15].val[1]);
 }
 
-static INLINE void highbd_idct16x16_add_store(const int32x4x2_t *const out,
-                                              uint16_t *dest, const int stride,
-                                              const int bd) {
-  // Add the result to dest
-  const int16x8_t max = vdupq_n_s16((1 << bd) - 1);
-  int16x8_t o[16];
-  o[0] = vcombine_s16(vrshrn_n_s32(out[0].val[0], 6),
-                      vrshrn_n_s32(out[0].val[1], 6));
-  o[1] = vcombine_s16(vrshrn_n_s32(out[1].val[0], 6),
-                      vrshrn_n_s32(out[1].val[1], 6));
-  o[2] = vcombine_s16(vrshrn_n_s32(out[2].val[0], 6),
-                      vrshrn_n_s32(out[2].val[1], 6));
-  o[3] = vcombine_s16(vrshrn_n_s32(out[3].val[0], 6),
-                      vrshrn_n_s32(out[3].val[1], 6));
-  o[4] = vcombine_s16(vrshrn_n_s32(out[4].val[0], 6),
-                      vrshrn_n_s32(out[4].val[1], 6));
-  o[5] = vcombine_s16(vrshrn_n_s32(out[5].val[0], 6),
-                      vrshrn_n_s32(out[5].val[1], 6));
-  o[6] = vcombine_s16(vrshrn_n_s32(out[6].val[0], 6),
-                      vrshrn_n_s32(out[6].val[1], 6));
-  o[7] = vcombine_s16(vrshrn_n_s32(out[7].val[0], 6),
-                      vrshrn_n_s32(out[7].val[1], 6));
-  o[8] = vcombine_s16(vrshrn_n_s32(out[8].val[0], 6),
-                      vrshrn_n_s32(out[8].val[1], 6));
-  o[9] = vcombine_s16(vrshrn_n_s32(out[9].val[0], 6),
-                      vrshrn_n_s32(out[9].val[1], 6));
-  o[10] = vcombine_s16(vrshrn_n_s32(out[10].val[0], 6),
-                       vrshrn_n_s32(out[10].val[1], 6));
-  o[11] = vcombine_s16(vrshrn_n_s32(out[11].val[0], 6),
-                       vrshrn_n_s32(out[11].val[1], 6));
-  o[12] = vcombine_s16(vrshrn_n_s32(out[12].val[0], 6),
-                       vrshrn_n_s32(out[12].val[1], 6));
-  o[13] = vcombine_s16(vrshrn_n_s32(out[13].val[0], 6),
-                       vrshrn_n_s32(out[13].val[1], 6));
-  o[14] = vcombine_s16(vrshrn_n_s32(out[14].val[0], 6),
-                       vrshrn_n_s32(out[14].val[1], 6));
-  o[15] = vcombine_s16(vrshrn_n_s32(out[15].val[0], 6),
-                       vrshrn_n_s32(out[15].val[1], 6));
-  highbd_idct16x16_add8x1(o[0], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[1], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[2], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[3], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[4], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[5], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[6], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[7], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[8], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[9], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[10], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[11], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[12], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[13], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[14], max, &dest, stride);
-  highbd_idct16x16_add8x1(o[15], max, &dest, stride);
-}
-
 static void highbd_idct16x16_256_add_half1d(const int32_t *input,
                                             int32_t *output, uint16_t *dest,
                                             const int stride, const int bd) {
@@ -900,22 +844,6 @@
   t.val[0] = vmull_lane_s32(vget_low_s32(s), coef, 1);
   t.val[1] = vmull_lane_s32(vget_high_s32(s), coef, 1);
   return highbd_idct16x16_add_wrap_low_4x1(t);
-}
-
-static INLINE int32x4x2_t highbd_idct_add_dual(const int32x4x2_t s0,
-                                               const int32x4x2_t s1) {
-  int32x4x2_t t;
-  t.val[0] = vaddq_s32(s0.val[0], s1.val[0]);
-  t.val[1] = vaddq_s32(s0.val[1], s1.val[1]);
-  return t;
-}
-
-static INLINE int32x4x2_t highbd_idct_sub_dual(const int32x4x2_t s0,
-                                               const int32x4x2_t s1) {
-  int32x4x2_t t;
-  t.val[0] = vsubq_s32(s0.val[0], s1.val[0]);
-  t.val[1] = vsubq_s32(s0.val[1], s1.val[1]);
-  return t;
 }
 
 static void highbd_idct16x16_38_add_half1d(const int32_t *input,
--- /dev/null
+++ b/vpx_dsp/arm/highbd_idct32x32_135_add_neon.c
@@ -1,0 +1,757 @@
+/*
+ *  Copyright (c) 2017 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_config.h"
+#include "./vpx_dsp_rtcd.h"
+#include "vpx_dsp/arm/idct_neon.h"
+#include "vpx_dsp/arm/transpose_neon.h"
+#include "vpx_dsp/txfm_common.h"
+
+static INLINE void load_8x8_s32_dual(
+    const tran_low_t *input, int32x4x2_t *const in0, int32x4x2_t *const in1,
+    int32x4x2_t *const in2, int32x4x2_t *const in3, int32x4x2_t *const in4,
+    int32x4x2_t *const in5, int32x4x2_t *const in6, int32x4x2_t *const in7) {
+  in0->val[0] = vld1q_s32(input);
+  in0->val[1] = vld1q_s32(input + 4);
+  input += 32;
+  in1->val[0] = vld1q_s32(input);
+  in1->val[1] = vld1q_s32(input + 4);
+  input += 32;
+  in2->val[0] = vld1q_s32(input);
+  in2->val[1] = vld1q_s32(input + 4);
+  input += 32;
+  in3->val[0] = vld1q_s32(input);
+  in3->val[1] = vld1q_s32(input + 4);
+  input += 32;
+  in4->val[0] = vld1q_s32(input);
+  in4->val[1] = vld1q_s32(input + 4);
+  input += 32;
+  in5->val[0] = vld1q_s32(input);
+  in5->val[1] = vld1q_s32(input + 4);
+  input += 32;
+  in6->val[0] = vld1q_s32(input);
+  in6->val[1] = vld1q_s32(input + 4);
+  input += 32;
+  in7->val[0] = vld1q_s32(input);
+  in7->val[1] = vld1q_s32(input + 4);
+}
+
+static INLINE void load_4x8_s32_dual(const tran_low_t *input,
+                                     int32x4_t *const in0, int32x4_t *const in1,
+                                     int32x4_t *const in2, int32x4_t *const in3,
+                                     int32x4_t *const in4, int32x4_t *const in5,
+                                     int32x4_t *const in6,
+                                     int32x4_t *const in7) {
+  *in0 = vld1q_s32(input);
+  input += 32;
+  *in1 = vld1q_s32(input);
+  input += 32;
+  *in2 = vld1q_s32(input);
+  input += 32;
+  *in3 = vld1q_s32(input);
+  input += 32;
+  *in4 = vld1q_s32(input);
+  input += 32;
+  *in5 = vld1q_s32(input);
+  input += 32;
+  *in6 = vld1q_s32(input);
+  input += 32;
+  *in7 = vld1q_s32(input);
+}
+
+// Only for the first pass of the  _135_ variant. Since it only uses values from
+// the top left 16x16 it can safely assume all the remaining values are 0 and
+// skip an awful lot of calculations. In fact, only the first 12 columns make
+// the cut. None of the elements in the 13th, 14th, 15th or 16th columns are
+// used so it skips any calls to input[12|13|14|15] too.
+// In C this does a single row of 32 for each call. Here it transposes the top
+// left 12x8 to allow using SIMD.
+
+// vp9/common/vp9_scan.c:vp9_default_iscan_32x32 arranges the first 135 non-zero
+// coefficients as follows:
+//      0   1   2   3   4   5   6   7   8   9  10  11  12  13  14  15
+//  0   0   2   5  10  17  25  38  47  62  83 101 121
+//  1   1   4   8  15  22  30  45  58  74  92 112 133
+//  2   3   7  12  18  28  36  52  64  82 102 118
+//  3   6  11  16  23  31  43  60  73  90 109 126
+//  4   9  14  19  29  37  50  65  78  98 116 134
+//  5  13  20  26  35  44  54  72  85 105 123
+//  6  21  27  33  42  53  63  80  94 113 132
+//  7  24  32  39  48  57  71  88 104 120
+//  8  34  40  46  56  68  81  96 111 130
+//  9  41  49  55  67  77  91 107 124
+// 10  51  59  66  76  89  99 119 131
+// 11  61  69  75  87 100 114 129
+// 12  70  79  86  97 108 122
+// 13  84  93 103 110 125
+// 14  98 106 115 127
+// 15 117 128
+static void highbd_idct32_12_neon(const tran_low_t *const input,
+                                  int32_t *output) {
+  int32x4x2_t in[12], s1[32], s2[32], s3[32], s4[32], s5[32], s6[32], s7[32],
+      s8[32];
+
+  load_8x8_s32_dual(input, &in[0], &in[1], &in[2], &in[3], &in[4], &in[5],
+                    &in[6], &in[7]);
+  transpose_s32_8x8(&in[0], &in[1], &in[2], &in[3], &in[4], &in[5], &in[6],
+                    &in[7]);
+
+  load_4x8_s32_dual(input + 8, &in[8].val[0], &in[8].val[1], &in[9].val[0],
+                    &in[9].val[1], &in[10].val[0], &in[10].val[1],
+                    &in[11].val[0], &in[11].val[1]);
+  transpose_s32_4x8(&in[8].val[0], &in[8].val[1], &in[9].val[0], &in[9].val[1],
+                    &in[10].val[0], &in[10].val[1], &in[11].val[0],
+                    &in[11].val[1]);
+
+  // stage 1
+  s1[16] = multiply_shift_and_narrow_s32_dual(in[1], cospi_31_64);
+  s1[31] = multiply_shift_and_narrow_s32_dual(in[1], cospi_1_64);
+
+  s1[18] = multiply_shift_and_narrow_s32_dual(in[9], cospi_23_64);
+  s1[29] = multiply_shift_and_narrow_s32_dual(in[9], cospi_9_64);
+
+  s1[19] = multiply_shift_and_narrow_s32_dual(in[7], -cospi_25_64);
+  s1[28] = multiply_shift_and_narrow_s32_dual(in[7], cospi_7_64);
+
+  s1[20] = multiply_shift_and_narrow_s32_dual(in[5], cospi_27_64);
+  s1[27] = multiply_shift_and_narrow_s32_dual(in[5], cospi_5_64);
+
+  s1[21] = multiply_shift_and_narrow_s32_dual(in[11], -cospi_21_64);
+  s1[26] = multiply_shift_and_narrow_s32_dual(in[11], cospi_11_64);
+
+  s1[23] = multiply_shift_and_narrow_s32_dual(in[3], -cospi_29_64);
+  s1[24] = multiply_shift_and_narrow_s32_dual(in[3], cospi_3_64);
+
+  // stage 2
+  s2[8] = multiply_shift_and_narrow_s32_dual(in[2], cospi_30_64);
+  s2[15] = multiply_shift_and_narrow_s32_dual(in[2], cospi_2_64);
+
+  s2[10] = multiply_shift_and_narrow_s32_dual(in[10], cospi_22_64);
+  s2[13] = multiply_shift_and_narrow_s32_dual(in[10], cospi_10_64);
+
+  s2[11] = multiply_shift_and_narrow_s32_dual(in[6], -cospi_26_64);
+  s2[12] = multiply_shift_and_narrow_s32_dual(in[6], cospi_6_64);
+
+  s2[18] = highbd_idct_sub_dual(s1[19], s1[18]);
+  s2[19] = highbd_idct_add_dual(s1[18], s1[19]);
+  s2[20] = highbd_idct_add_dual(s1[20], s1[21]);
+  s2[21] = highbd_idct_sub_dual(s1[20], s1[21]);
+  s2[26] = highbd_idct_sub_dual(s1[27], s1[26]);
+  s2[27] = highbd_idct_add_dual(s1[26], s1[27]);
+  s2[28] = highbd_idct_add_dual(s1[28], s1[29]);
+  s2[29] = highbd_idct_sub_dual(s1[28], s1[29]);
+
+  // stage 3
+  s3[4] = multiply_shift_and_narrow_s32_dual(in[4], cospi_28_64);
+  s3[7] = multiply_shift_and_narrow_s32_dual(in[4], cospi_4_64);
+
+  s3[10] = highbd_idct_sub_dual(s2[11], s2[10]);
+  s3[11] = highbd_idct_add_dual(s2[10], s2[11]);
+  s3[12] = highbd_idct_add_dual(s2[12], s2[13]);
+  s3[13] = highbd_idct_sub_dual(s2[12], s2[13]);
+
+  s3[17] = multiply_accumulate_shift_and_narrow_s32_dual(s1[16], -cospi_4_64,
+                                                         s1[31], cospi_28_64);
+  s3[30] = multiply_accumulate_shift_and_narrow_s32_dual(s1[16], cospi_28_64,
+                                                         s1[31], cospi_4_64);
+
+  s3[18] = multiply_accumulate_shift_and_narrow_s32_dual(s2[18], -cospi_28_64,
+                                                         s2[29], -cospi_4_64);
+  s3[29] = multiply_accumulate_shift_and_narrow_s32_dual(s2[18], -cospi_4_64,
+                                                         s2[29], cospi_28_64);
+
+  s3[21] = multiply_accumulate_shift_and_narrow_s32_dual(s2[21], -cospi_20_64,
+                                                         s2[26], cospi_12_64);
+  s3[26] = multiply_accumulate_shift_and_narrow_s32_dual(s2[21], cospi_12_64,
+                                                         s2[26], cospi_20_64);
+
+  s3[22] = multiply_accumulate_shift_and_narrow_s32_dual(s1[23], -cospi_12_64,
+                                                         s1[24], -cospi_20_64);
+  s3[25] = multiply_accumulate_shift_and_narrow_s32_dual(s1[23], -cospi_20_64,
+                                                         s1[24], cospi_12_64);
+
+  // stage 4
+  s4[0] = multiply_shift_and_narrow_s32_dual(in[0], cospi_16_64);
+  s4[2] = multiply_shift_and_narrow_s32_dual(in[8], cospi_24_64);
+  s4[3] = multiply_shift_and_narrow_s32_dual(in[8], cospi_8_64);
+
+  s4[9] = multiply_accumulate_shift_and_narrow_s32_dual(s2[8], -cospi_8_64,
+                                                        s2[15], cospi_24_64);
+  s4[14] = multiply_accumulate_shift_and_narrow_s32_dual(s2[8], cospi_24_64,
+                                                         s2[15], cospi_8_64);
+
+  s4[10] = multiply_accumulate_shift_and_narrow_s32_dual(s3[10], -cospi_24_64,
+                                                         s3[13], -cospi_8_64);
+  s4[13] = multiply_accumulate_shift_and_narrow_s32_dual(s3[10], -cospi_8_64,
+                                                         s3[13], cospi_24_64);
+
+  s4[16] = highbd_idct_add_dual(s1[16], s2[19]);
+  s4[17] = highbd_idct_add_dual(s3[17], s3[18]);
+  s4[18] = highbd_idct_sub_dual(s3[17], s3[18]);
+  s4[19] = highbd_idct_sub_dual(s1[16], s2[19]);
+  s4[20] = highbd_idct_sub_dual(s1[23], s2[20]);
+  s4[21] = highbd_idct_sub_dual(s3[22], s3[21]);
+  s4[22] = highbd_idct_add_dual(s3[21], s3[22]);
+  s4[23] = highbd_idct_add_dual(s2[20], s1[23]);
+  s4[24] = highbd_idct_add_dual(s1[24], s2[27]);
+  s4[25] = highbd_idct_add_dual(s3[25], s3[26]);
+  s4[26] = highbd_idct_sub_dual(s3[25], s3[26]);
+  s4[27] = highbd_idct_sub_dual(s1[24], s2[27]);
+  s4[28] = highbd_idct_sub_dual(s1[31], s2[28]);
+  s4[29] = highbd_idct_sub_dual(s3[30], s3[29]);
+  s4[30] = highbd_idct_add_dual(s3[29], s3[30]);
+  s4[31] = highbd_idct_add_dual(s2[28], s1[31]);
+
+  // stage 5
+  s5[0] = highbd_idct_add_dual(s4[0], s4[3]);
+  s5[1] = highbd_idct_add_dual(s4[0], s4[2]);
+  s5[2] = highbd_idct_sub_dual(s4[0], s4[2]);
+  s5[3] = highbd_idct_sub_dual(s4[0], s4[3]);
+
+  s5[5] = sub_multiply_shift_and_narrow_s32_dual(s3[7], s3[4], cospi_16_64);
+  s5[6] = add_multiply_shift_and_narrow_s32_dual(s3[4], s3[7], cospi_16_64);
+
+  s5[8] = highbd_idct_add_dual(s2[8], s3[11]);
+  s5[9] = highbd_idct_add_dual(s4[9], s4[10]);
+  s5[10] = highbd_idct_sub_dual(s4[9], s4[10]);
+  s5[11] = highbd_idct_sub_dual(s2[8], s3[11]);
+  s5[12] = highbd_idct_sub_dual(s2[15], s3[12]);
+  s5[13] = highbd_idct_sub_dual(s4[14], s4[13]);
+  s5[14] = highbd_idct_add_dual(s4[13], s4[14]);
+  s5[15] = highbd_idct_add_dual(s2[15], s3[12]);
+
+  s5[18] = multiply_accumulate_shift_and_narrow_s32_dual(s4[18], -cospi_8_64,
+                                                         s4[29], cospi_24_64);
+  s5[29] = multiply_accumulate_shift_and_narrow_s32_dual(s4[18], cospi_24_64,
+                                                         s4[29], cospi_8_64);
+
+  s5[19] = multiply_accumulate_shift_and_narrow_s32_dual(s4[19], -cospi_8_64,
+                                                         s4[28], cospi_24_64);
+  s5[28] = multiply_accumulate_shift_and_narrow_s32_dual(s4[19], cospi_24_64,
+                                                         s4[28], cospi_8_64);
+
+  s5[20] = multiply_accumulate_shift_and_narrow_s32_dual(s4[20], -cospi_24_64,
+                                                         s4[27], -cospi_8_64);
+  s5[27] = multiply_accumulate_shift_and_narrow_s32_dual(s4[20], -cospi_8_64,
+                                                         s4[27], cospi_24_64);
+
+  s5[21] = multiply_accumulate_shift_and_narrow_s32_dual(s4[21], -cospi_24_64,
+                                                         s4[26], -cospi_8_64);
+  s5[26] = multiply_accumulate_shift_and_narrow_s32_dual(s4[21], -cospi_8_64,
+                                                         s4[26], cospi_24_64);
+
+  // stage 6
+  s6[0] = highbd_idct_add_dual(s5[0], s3[7]);
+  s6[1] = highbd_idct_add_dual(s5[1], s5[6]);
+  s6[2] = highbd_idct_add_dual(s5[2], s5[5]);
+  s6[3] = highbd_idct_add_dual(s5[3], s3[4]);
+  s6[4] = highbd_idct_sub_dual(s5[3], s3[4]);
+  s6[5] = highbd_idct_sub_dual(s5[2], s5[5]);
+  s6[6] = highbd_idct_sub_dual(s5[1], s5[6]);
+  s6[7] = highbd_idct_sub_dual(s5[0], s3[7]);
+
+  s6[10] = sub_multiply_shift_and_narrow_s32_dual(s5[13], s5[10], cospi_16_64);
+  s6[13] = add_multiply_shift_and_narrow_s32_dual(s5[10], s5[13], cospi_16_64);
+
+  s6[11] = sub_multiply_shift_and_narrow_s32_dual(s5[12], s5[11], cospi_16_64);
+  s6[12] = add_multiply_shift_and_narrow_s32_dual(s5[11], s5[12], cospi_16_64);
+
+  s6[16] = highbd_idct_add_dual(s4[16], s4[23]);
+  s6[17] = highbd_idct_add_dual(s4[17], s4[22]);
+  s6[18] = highbd_idct_add_dual(s5[18], s5[21]);
+  s6[19] = highbd_idct_add_dual(s5[19], s5[20]);
+  s6[20] = highbd_idct_sub_dual(s5[19], s5[20]);
+  s6[21] = highbd_idct_sub_dual(s5[18], s5[21]);
+  s6[22] = highbd_idct_sub_dual(s4[17], s4[22]);
+  s6[23] = highbd_idct_sub_dual(s4[16], s4[23]);
+
+  s6[24] = highbd_idct_sub_dual(s4[31], s4[24]);
+  s6[25] = highbd_idct_sub_dual(s4[30], s4[25]);
+  s6[26] = highbd_idct_sub_dual(s5[29], s5[26]);
+  s6[27] = highbd_idct_sub_dual(s5[28], s5[27]);
+  s6[28] = highbd_idct_add_dual(s5[27], s5[28]);
+  s6[29] = highbd_idct_add_dual(s5[26], s5[29]);
+  s6[30] = highbd_idct_add_dual(s4[25], s4[30]);
+  s6[31] = highbd_idct_add_dual(s4[24], s4[31]);
+
+  // stage 7
+  s7[0] = highbd_idct_add_dual(s6[0], s5[15]);
+  s7[1] = highbd_idct_add_dual(s6[1], s5[14]);
+  s7[2] = highbd_idct_add_dual(s6[2], s6[13]);
+  s7[3] = highbd_idct_add_dual(s6[3], s6[12]);
+  s7[4] = highbd_idct_add_dual(s6[4], s6[11]);
+  s7[5] = highbd_idct_add_dual(s6[5], s6[10]);
+  s7[6] = highbd_idct_add_dual(s6[6], s5[9]);
+  s7[7] = highbd_idct_add_dual(s6[7], s5[8]);
+  s7[8] = highbd_idct_sub_dual(s6[7], s5[8]);
+  s7[9] = highbd_idct_sub_dual(s6[6], s5[9]);
+  s7[10] = highbd_idct_sub_dual(s6[5], s6[10]);
+  s7[11] = highbd_idct_sub_dual(s6[4], s6[11]);
+  s7[12] = highbd_idct_sub_dual(s6[3], s6[12]);
+  s7[13] = highbd_idct_sub_dual(s6[2], s6[13]);
+  s7[14] = highbd_idct_sub_dual(s6[1], s5[14]);
+  s7[15] = highbd_idct_sub_dual(s6[0], s5[15]);
+
+  s7[20] = sub_multiply_shift_and_narrow_s32_dual(s6[27], s6[20], cospi_16_64);
+  s7[27] = add_multiply_shift_and_narrow_s32_dual(s6[20], s6[27], cospi_16_64);
+
+  s7[21] = sub_multiply_shift_and_narrow_s32_dual(s6[26], s6[21], cospi_16_64);
+  s7[26] = add_multiply_shift_and_narrow_s32_dual(s6[21], s6[26], cospi_16_64);
+
+  s7[22] = sub_multiply_shift_and_narrow_s32_dual(s6[25], s6[22], cospi_16_64);
+  s7[25] = add_multiply_shift_and_narrow_s32_dual(s6[22], s6[25], cospi_16_64);
+
+  s7[23] = sub_multiply_shift_and_narrow_s32_dual(s6[24], s6[23], cospi_16_64);
+  s7[24] = add_multiply_shift_and_narrow_s32_dual(s6[23], s6[24], cospi_16_64);
+
+  // final stage
+  s8[0] = highbd_idct_add_dual(s7[0], s6[31]);
+  s8[1] = highbd_idct_add_dual(s7[1], s6[30]);
+  s8[2] = highbd_idct_add_dual(s7[2], s6[29]);
+  s8[3] = highbd_idct_add_dual(s7[3], s6[28]);
+  s8[4] = highbd_idct_add_dual(s7[4], s7[27]);
+  s8[5] = highbd_idct_add_dual(s7[5], s7[26]);
+  s8[6] = highbd_idct_add_dual(s7[6], s7[25]);
+  s8[7] = highbd_idct_add_dual(s7[7], s7[24]);
+  s8[8] = highbd_idct_add_dual(s7[8], s7[23]);
+  s8[9] = highbd_idct_add_dual(s7[9], s7[22]);
+  s8[10] = highbd_idct_add_dual(s7[10], s7[21]);
+  s8[11] = highbd_idct_add_dual(s7[11], s7[20]);
+  s8[12] = highbd_idct_add_dual(s7[12], s6[19]);
+  s8[13] = highbd_idct_add_dual(s7[13], s6[18]);
+  s8[14] = highbd_idct_add_dual(s7[14], s6[17]);
+  s8[15] = highbd_idct_add_dual(s7[15], s6[16]);
+  s8[16] = highbd_idct_sub_dual(s7[15], s6[16]);
+  s8[17] = highbd_idct_sub_dual(s7[14], s6[17]);
+  s8[18] = highbd_idct_sub_dual(s7[13], s6[18]);
+  s8[19] = highbd_idct_sub_dual(s7[12], s6[19]);
+  s8[20] = highbd_idct_sub_dual(s7[11], s7[20]);
+  s8[21] = highbd_idct_sub_dual(s7[10], s7[21]);
+  s8[22] = highbd_idct_sub_dual(s7[9], s7[22]);
+  s8[23] = highbd_idct_sub_dual(s7[8], s7[23]);
+  s8[24] = highbd_idct_sub_dual(s7[7], s7[24]);
+  s8[25] = highbd_idct_sub_dual(s7[6], s7[25]);
+  s8[26] = highbd_idct_sub_dual(s7[5], s7[26]);
+  s8[27] = highbd_idct_sub_dual(s7[4], s7[27]);
+  s8[28] = highbd_idct_sub_dual(s7[3], s6[28]);
+  s8[29] = highbd_idct_sub_dual(s7[2], s6[29]);
+  s8[30] = highbd_idct_sub_dual(s7[1], s6[30]);
+  s8[31] = highbd_idct_sub_dual(s7[0], s6[31]);
+
+  vst1q_s32(output + 0, s8[0].val[0]);
+  vst1q_s32(output + 4, s8[0].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[1].val[0]);
+  vst1q_s32(output + 4, s8[1].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[2].val[0]);
+  vst1q_s32(output + 4, s8[2].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[3].val[0]);
+  vst1q_s32(output + 4, s8[3].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[4].val[0]);
+  vst1q_s32(output + 4, s8[4].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[5].val[0]);
+  vst1q_s32(output + 4, s8[5].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[6].val[0]);
+  vst1q_s32(output + 4, s8[6].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[7].val[0]);
+  vst1q_s32(output + 4, s8[7].val[1]);
+  output += 16;
+
+  vst1q_s32(output + 0, s8[8].val[0]);
+  vst1q_s32(output + 4, s8[8].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[9].val[0]);
+  vst1q_s32(output + 4, s8[9].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[10].val[0]);
+  vst1q_s32(output + 4, s8[10].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[11].val[0]);
+  vst1q_s32(output + 4, s8[11].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[12].val[0]);
+  vst1q_s32(output + 4, s8[12].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[13].val[0]);
+  vst1q_s32(output + 4, s8[13].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[14].val[0]);
+  vst1q_s32(output + 4, s8[14].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[15].val[0]);
+  vst1q_s32(output + 4, s8[15].val[1]);
+  output += 16;
+
+  vst1q_s32(output + 0, s8[16].val[0]);
+  vst1q_s32(output + 4, s8[16].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[17].val[0]);
+  vst1q_s32(output + 4, s8[17].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[18].val[0]);
+  vst1q_s32(output + 4, s8[18].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[19].val[0]);
+  vst1q_s32(output + 4, s8[19].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[20].val[0]);
+  vst1q_s32(output + 4, s8[20].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[21].val[0]);
+  vst1q_s32(output + 4, s8[21].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[22].val[0]);
+  vst1q_s32(output + 4, s8[22].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[23].val[0]);
+  vst1q_s32(output + 4, s8[23].val[1]);
+  output += 16;
+
+  vst1q_s32(output + 0, s8[24].val[0]);
+  vst1q_s32(output + 4, s8[24].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[25].val[0]);
+  vst1q_s32(output + 4, s8[25].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[26].val[0]);
+  vst1q_s32(output + 4, s8[26].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[27].val[0]);
+  vst1q_s32(output + 4, s8[27].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[28].val[0]);
+  vst1q_s32(output + 4, s8[28].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[29].val[0]);
+  vst1q_s32(output + 4, s8[29].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[30].val[0]);
+  vst1q_s32(output + 4, s8[30].val[1]);
+  output += 16;
+  vst1q_s32(output + 0, s8[31].val[0]);
+  vst1q_s32(output + 4, s8[31].val[1]);
+}
+
+static void highbd_idct32_16_neon(const int32_t *const input,
+                                  uint16_t *const output, const int stride,
+                                  const int bd) {
+  int32x4x2_t in[16], s1[32], s2[32], s3[32], s4[32], s5[32], s6[32], s7[32],
+      out[32];
+
+  load_and_transpose_s32_8x8(input, 16, &in[0], &in[1], &in[2], &in[3], &in[4],
+                             &in[5], &in[6], &in[7]);
+
+  load_and_transpose_s32_8x8(input + 8, 16, &in[8], &in[9], &in[10], &in[11],
+                             &in[12], &in[13], &in[14], &in[15]);
+
+  // stage 1
+  s1[16] = multiply_shift_and_narrow_s32_dual(in[1], cospi_31_64);
+  s1[31] = multiply_shift_and_narrow_s32_dual(in[1], cospi_1_64);
+
+  s1[17] = multiply_shift_and_narrow_s32_dual(in[15], -cospi_17_64);
+  s1[30] = multiply_shift_and_narrow_s32_dual(in[15], cospi_15_64);
+
+  s1[18] = multiply_shift_and_narrow_s32_dual(in[9], cospi_23_64);
+  s1[29] = multiply_shift_and_narrow_s32_dual(in[9], cospi_9_64);
+
+  s1[19] = multiply_shift_and_narrow_s32_dual(in[7], -cospi_25_64);
+  s1[28] = multiply_shift_and_narrow_s32_dual(in[7], cospi_7_64);
+
+  s1[20] = multiply_shift_and_narrow_s32_dual(in[5], cospi_27_64);
+  s1[27] = multiply_shift_and_narrow_s32_dual(in[5], cospi_5_64);
+
+  s1[21] = multiply_shift_and_narrow_s32_dual(in[11], -cospi_21_64);
+  s1[26] = multiply_shift_and_narrow_s32_dual(in[11], cospi_11_64);
+
+  s1[22] = multiply_shift_and_narrow_s32_dual(in[13], cospi_19_64);
+  s1[25] = multiply_shift_and_narrow_s32_dual(in[13], cospi_13_64);
+
+  s1[23] = multiply_shift_and_narrow_s32_dual(in[3], -cospi_29_64);
+  s1[24] = multiply_shift_and_narrow_s32_dual(in[3], cospi_3_64);
+
+  // stage 2
+  s2[8] = multiply_shift_and_narrow_s32_dual(in[2], cospi_30_64);
+  s2[15] = multiply_shift_and_narrow_s32_dual(in[2], cospi_2_64);
+
+  s2[9] = multiply_shift_and_narrow_s32_dual(in[14], -cospi_18_64);
+  s2[14] = multiply_shift_and_narrow_s32_dual(in[14], cospi_14_64);
+
+  s2[10] = multiply_shift_and_narrow_s32_dual(in[10], cospi_22_64);
+  s2[13] = multiply_shift_and_narrow_s32_dual(in[10], cospi_10_64);
+
+  s2[11] = multiply_shift_and_narrow_s32_dual(in[6], -cospi_26_64);
+  s2[12] = multiply_shift_and_narrow_s32_dual(in[6], cospi_6_64);
+
+  s2[16] = highbd_idct_add_dual(s1[16], s1[17]);
+  s2[17] = highbd_idct_sub_dual(s1[16], s1[17]);
+  s2[18] = highbd_idct_sub_dual(s1[19], s1[18]);
+  s2[19] = highbd_idct_add_dual(s1[18], s1[19]);
+  s2[20] = highbd_idct_add_dual(s1[20], s1[21]);
+  s2[21] = highbd_idct_sub_dual(s1[20], s1[21]);
+  s2[22] = highbd_idct_sub_dual(s1[23], s1[22]);
+  s2[23] = highbd_idct_add_dual(s1[22], s1[23]);
+  s2[24] = highbd_idct_add_dual(s1[24], s1[25]);
+  s2[25] = highbd_idct_sub_dual(s1[24], s1[25]);
+  s2[26] = highbd_idct_sub_dual(s1[27], s1[26]);
+  s2[27] = highbd_idct_add_dual(s1[26], s1[27]);
+  s2[28] = highbd_idct_add_dual(s1[28], s1[29]);
+  s2[29] = highbd_idct_sub_dual(s1[28], s1[29]);
+  s2[30] = highbd_idct_sub_dual(s1[31], s1[30]);
+  s2[31] = highbd_idct_add_dual(s1[30], s1[31]);
+
+  // stage 3
+  s3[4] = multiply_shift_and_narrow_s32_dual(in[4], cospi_28_64);
+  s3[7] = multiply_shift_and_narrow_s32_dual(in[4], cospi_4_64);
+
+  s3[5] = multiply_shift_and_narrow_s32_dual(in[12], -cospi_20_64);
+  s3[6] = multiply_shift_and_narrow_s32_dual(in[12], cospi_12_64);
+
+  s3[8] = highbd_idct_add_dual(s2[8], s2[9]);
+  s3[9] = highbd_idct_sub_dual(s2[8], s2[9]);
+  s3[10] = highbd_idct_sub_dual(s2[11], s2[10]);
+  s3[11] = highbd_idct_add_dual(s2[10], s2[11]);
+  s3[12] = highbd_idct_add_dual(s2[12], s2[13]);
+  s3[13] = highbd_idct_sub_dual(s2[12], s2[13]);
+  s3[14] = highbd_idct_sub_dual(s2[15], s2[14]);
+  s3[15] = highbd_idct_add_dual(s2[14], s2[15]);
+
+  s3[17] = multiply_accumulate_shift_and_narrow_s32_dual(s2[17], -cospi_4_64,
+                                                         s2[30], cospi_28_64);
+  s3[30] = multiply_accumulate_shift_and_narrow_s32_dual(s2[17], cospi_28_64,
+                                                         s2[30], cospi_4_64);
+
+  s3[18] = multiply_accumulate_shift_and_narrow_s32_dual(s2[18], -cospi_28_64,
+                                                         s2[29], -cospi_4_64);
+  s3[29] = multiply_accumulate_shift_and_narrow_s32_dual(s2[18], -cospi_4_64,
+                                                         s2[29], cospi_28_64);
+
+  s3[21] = multiply_accumulate_shift_and_narrow_s32_dual(s2[21], -cospi_20_64,
+                                                         s2[26], cospi_12_64);
+  s3[26] = multiply_accumulate_shift_and_narrow_s32_dual(s2[21], cospi_12_64,
+                                                         s2[26], cospi_20_64);
+
+  s3[22] = multiply_accumulate_shift_and_narrow_s32_dual(s2[22], -cospi_12_64,
+                                                         s2[25], -cospi_20_64);
+  s3[25] = multiply_accumulate_shift_and_narrow_s32_dual(s2[22], -cospi_20_64,
+                                                         s2[25], cospi_12_64);
+
+  // stage 4
+  s4[0] = multiply_shift_and_narrow_s32_dual(in[0], cospi_16_64);
+  s4[2] = multiply_shift_and_narrow_s32_dual(in[8], cospi_24_64);
+  s4[3] = multiply_shift_and_narrow_s32_dual(in[8], cospi_8_64);
+
+  s4[4] = highbd_idct_add_dual(s3[4], s3[5]);
+  s4[5] = highbd_idct_sub_dual(s3[4], s3[5]);
+  s4[6] = highbd_idct_sub_dual(s3[7], s3[6]);
+  s4[7] = highbd_idct_add_dual(s3[6], s3[7]);
+
+  s4[9] = multiply_accumulate_shift_and_narrow_s32_dual(s3[9], -cospi_8_64,
+                                                        s3[14], cospi_24_64);
+  s4[14] = multiply_accumulate_shift_and_narrow_s32_dual(s3[9], cospi_24_64,
+                                                         s3[14], cospi_8_64);
+
+  s4[10] = multiply_accumulate_shift_and_narrow_s32_dual(s3[10], -cospi_24_64,
+                                                         s3[13], -cospi_8_64);
+  s4[13] = multiply_accumulate_shift_and_narrow_s32_dual(s3[10], -cospi_8_64,
+                                                         s3[13], cospi_24_64);
+
+  s4[16] = highbd_idct_add_dual(s2[16], s2[19]);
+  s4[17] = highbd_idct_add_dual(s3[17], s3[18]);
+  s4[18] = highbd_idct_sub_dual(s3[17], s3[18]);
+  s4[19] = highbd_idct_sub_dual(s2[16], s2[19]);
+  s4[20] = highbd_idct_sub_dual(s2[23], s2[20]);
+  s4[21] = highbd_idct_sub_dual(s3[22], s3[21]);
+  s4[22] = highbd_idct_add_dual(s3[21], s3[22]);
+  s4[23] = highbd_idct_add_dual(s2[20], s2[23]);
+  s4[24] = highbd_idct_add_dual(s2[24], s2[27]);
+  s4[25] = highbd_idct_add_dual(s3[25], s3[26]);
+  s4[26] = highbd_idct_sub_dual(s3[25], s3[26]);
+  s4[27] = highbd_idct_sub_dual(s2[24], s2[27]);
+  s4[28] = highbd_idct_sub_dual(s2[31], s2[28]);
+  s4[29] = highbd_idct_sub_dual(s3[30], s3[29]);
+  s4[30] = highbd_idct_add_dual(s3[29], s3[30]);
+  s4[31] = highbd_idct_add_dual(s2[28], s2[31]);
+
+  // stage 5
+  s5[0] = highbd_idct_add_dual(s4[0], s4[3]);
+  s5[1] = highbd_idct_add_dual(s4[0], s4[2]);
+  s5[2] = highbd_idct_sub_dual(s4[0], s4[2]);
+  s5[3] = highbd_idct_sub_dual(s4[0], s4[3]);
+
+  s5[5] = sub_multiply_shift_and_narrow_s32_dual(s4[6], s4[5], cospi_16_64);
+  s5[6] = add_multiply_shift_and_narrow_s32_dual(s4[5], s4[6], cospi_16_64);
+
+  s5[8] = highbd_idct_add_dual(s3[8], s3[11]);
+  s5[9] = highbd_idct_add_dual(s4[9], s4[10]);
+  s5[10] = highbd_idct_sub_dual(s4[9], s4[10]);
+  s5[11] = highbd_idct_sub_dual(s3[8], s3[11]);
+  s5[12] = highbd_idct_sub_dual(s3[15], s3[12]);
+  s5[13] = highbd_idct_sub_dual(s4[14], s4[13]);
+  s5[14] = highbd_idct_add_dual(s4[13], s4[14]);
+  s5[15] = highbd_idct_add_dual(s3[15], s3[12]);
+
+  s5[18] = multiply_accumulate_shift_and_narrow_s32_dual(s4[18], -cospi_8_64,
+                                                         s4[29], cospi_24_64);
+  s5[29] = multiply_accumulate_shift_and_narrow_s32_dual(s4[18], cospi_24_64,
+                                                         s4[29], cospi_8_64);
+
+  s5[19] = multiply_accumulate_shift_and_narrow_s32_dual(s4[19], -cospi_8_64,
+                                                         s4[28], cospi_24_64);
+  s5[28] = multiply_accumulate_shift_and_narrow_s32_dual(s4[19], cospi_24_64,
+                                                         s4[28], cospi_8_64);
+
+  s5[20] = multiply_accumulate_shift_and_narrow_s32_dual(s4[20], -cospi_24_64,
+                                                         s4[27], -cospi_8_64);
+  s5[27] = multiply_accumulate_shift_and_narrow_s32_dual(s4[20], -cospi_8_64,
+                                                         s4[27], cospi_24_64);
+
+  s5[21] = multiply_accumulate_shift_and_narrow_s32_dual(s4[21], -cospi_24_64,
+                                                         s4[26], -cospi_8_64);
+  s5[26] = multiply_accumulate_shift_and_narrow_s32_dual(s4[21], -cospi_8_64,
+                                                         s4[26], cospi_24_64);
+
+  // stage 6
+  s6[0] = highbd_idct_add_dual(s5[0], s4[7]);
+  s6[1] = highbd_idct_add_dual(s5[1], s5[6]);
+  s6[2] = highbd_idct_add_dual(s5[2], s5[5]);
+  s6[3] = highbd_idct_add_dual(s5[3], s4[4]);
+  s6[4] = highbd_idct_sub_dual(s5[3], s4[4]);
+  s6[5] = highbd_idct_sub_dual(s5[2], s5[5]);
+  s6[6] = highbd_idct_sub_dual(s5[1], s5[6]);
+  s6[7] = highbd_idct_sub_dual(s5[0], s4[7]);
+
+  s6[10] = sub_multiply_shift_and_narrow_s32_dual(s5[13], s5[10], cospi_16_64);
+  s6[13] = add_multiply_shift_and_narrow_s32_dual(s5[10], s5[13], cospi_16_64);
+
+  s6[11] = sub_multiply_shift_and_narrow_s32_dual(s5[12], s5[11], cospi_16_64);
+  s6[12] = add_multiply_shift_and_narrow_s32_dual(s5[11], s5[12], cospi_16_64);
+
+  s6[16] = highbd_idct_add_dual(s4[16], s4[23]);
+  s6[17] = highbd_idct_add_dual(s4[17], s4[22]);
+  s6[18] = highbd_idct_add_dual(s5[18], s5[21]);
+  s6[19] = highbd_idct_add_dual(s5[19], s5[20]);
+  s6[20] = highbd_idct_sub_dual(s5[19], s5[20]);
+  s6[21] = highbd_idct_sub_dual(s5[18], s5[21]);
+  s6[22] = highbd_idct_sub_dual(s4[17], s4[22]);
+  s6[23] = highbd_idct_sub_dual(s4[16], s4[23]);
+  s6[24] = highbd_idct_sub_dual(s4[31], s4[24]);
+  s6[25] = highbd_idct_sub_dual(s4[30], s4[25]);
+  s6[26] = highbd_idct_sub_dual(s5[29], s5[26]);
+  s6[27] = highbd_idct_sub_dual(s5[28], s5[27]);
+  s6[28] = highbd_idct_add_dual(s5[27], s5[28]);
+  s6[29] = highbd_idct_add_dual(s5[26], s5[29]);
+  s6[30] = highbd_idct_add_dual(s4[25], s4[30]);
+  s6[31] = highbd_idct_add_dual(s4[24], s4[31]);
+
+  // stage 7
+  s7[0] = highbd_idct_add_dual(s6[0], s5[15]);
+  s7[1] = highbd_idct_add_dual(s6[1], s5[14]);
+  s7[2] = highbd_idct_add_dual(s6[2], s6[13]);
+  s7[3] = highbd_idct_add_dual(s6[3], s6[12]);
+  s7[4] = highbd_idct_add_dual(s6[4], s6[11]);
+  s7[5] = highbd_idct_add_dual(s6[5], s6[10]);
+  s7[6] = highbd_idct_add_dual(s6[6], s5[9]);
+  s7[7] = highbd_idct_add_dual(s6[7], s5[8]);
+  s7[8] = highbd_idct_sub_dual(s6[7], s5[8]);
+  s7[9] = highbd_idct_sub_dual(s6[6], s5[9]);
+  s7[10] = highbd_idct_sub_dual(s6[5], s6[10]);
+  s7[11] = highbd_idct_sub_dual(s6[4], s6[11]);
+  s7[12] = highbd_idct_sub_dual(s6[3], s6[12]);
+  s7[13] = highbd_idct_sub_dual(s6[2], s6[13]);
+  s7[14] = highbd_idct_sub_dual(s6[1], s5[14]);
+  s7[15] = highbd_idct_sub_dual(s6[0], s5[15]);
+
+  s7[20] = sub_multiply_shift_and_narrow_s32_dual(s6[27], s6[20], cospi_16_64);
+  s7[27] = add_multiply_shift_and_narrow_s32_dual(s6[20], s6[27], cospi_16_64);
+
+  s7[21] = sub_multiply_shift_and_narrow_s32_dual(s6[26], s6[21], cospi_16_64);
+  s7[26] = add_multiply_shift_and_narrow_s32_dual(s6[21], s6[26], cospi_16_64);
+
+  s7[22] = sub_multiply_shift_and_narrow_s32_dual(s6[25], s6[22], cospi_16_64);
+  s7[25] = add_multiply_shift_and_narrow_s32_dual(s6[22], s6[25], cospi_16_64);
+
+  s7[23] = sub_multiply_shift_and_narrow_s32_dual(s6[24], s6[23], cospi_16_64);
+  s7[24] = add_multiply_shift_and_narrow_s32_dual(s6[23], s6[24], cospi_16_64);
+
+  // final stage
+  out[0] = highbd_idct_add_dual(s7[0], s6[31]);
+  out[1] = highbd_idct_add_dual(s7[1], s6[30]);
+  out[2] = highbd_idct_add_dual(s7[2], s6[29]);
+  out[3] = highbd_idct_add_dual(s7[3], s6[28]);
+  out[4] = highbd_idct_add_dual(s7[4], s7[27]);
+  out[5] = highbd_idct_add_dual(s7[5], s7[26]);
+  out[6] = highbd_idct_add_dual(s7[6], s7[25]);
+  out[7] = highbd_idct_add_dual(s7[7], s7[24]);
+  out[8] = highbd_idct_add_dual(s7[8], s7[23]);
+  out[9] = highbd_idct_add_dual(s7[9], s7[22]);
+  out[10] = highbd_idct_add_dual(s7[10], s7[21]);
+  out[11] = highbd_idct_add_dual(s7[11], s7[20]);
+  out[12] = highbd_idct_add_dual(s7[12], s6[19]);
+  out[13] = highbd_idct_add_dual(s7[13], s6[18]);
+  out[14] = highbd_idct_add_dual(s7[14], s6[17]);
+  out[15] = highbd_idct_add_dual(s7[15], s6[16]);
+  out[16] = highbd_idct_sub_dual(s7[15], s6[16]);
+  out[17] = highbd_idct_sub_dual(s7[14], s6[17]);
+  out[18] = highbd_idct_sub_dual(s7[13], s6[18]);
+  out[19] = highbd_idct_sub_dual(s7[12], s6[19]);
+  out[20] = highbd_idct_sub_dual(s7[11], s7[20]);
+  out[21] = highbd_idct_sub_dual(s7[10], s7[21]);
+  out[22] = highbd_idct_sub_dual(s7[9], s7[22]);
+  out[23] = highbd_idct_sub_dual(s7[8], s7[23]);
+  out[24] = highbd_idct_sub_dual(s7[7], s7[24]);
+  out[25] = highbd_idct_sub_dual(s7[6], s7[25]);
+  out[26] = highbd_idct_sub_dual(s7[5], s7[26]);
+  out[27] = highbd_idct_sub_dual(s7[4], s7[27]);
+  out[28] = highbd_idct_sub_dual(s7[3], s6[28]);
+  out[29] = highbd_idct_sub_dual(s7[2], s6[29]);
+  out[30] = highbd_idct_sub_dual(s7[1], s6[30]);
+  out[31] = highbd_idct_sub_dual(s7[0], s6[31]);
+
+  highbd_idct16x16_add_store(out, output, stride, bd);
+  highbd_idct16x16_add_store(out + 16, output + 16 * stride, stride, bd);
+}
+
+void vpx_highbd_idct32x32_135_add_neon(const tran_low_t *input, uint8_t *dest,
+                                       int stride, int bd) {
+  int i;
+
+  if (bd == 8) {
+    int16_t temp[32 * 16];
+    int16_t *t = temp;
+    idct32_12_neon(input, temp);
+    idct32_12_neon(input + 32 * 8, temp + 8);
+
+    for (i = 0; i < 32; i += 8) {
+      idct32_16_neon(t, dest, stride, 1);
+      t += (16 * 8);
+      dest += 8;
+    }
+  } else {
+    uint16_t *dst = CONVERT_TO_SHORTPTR(dest);
+    int32_t temp[32 * 16];
+    int32_t *t = temp;
+    highbd_idct32_12_neon(input, temp);
+    highbd_idct32_12_neon(input + 32 * 8, temp + 8);
+
+    for (i = 0; i < 32; i += 8) {
+      highbd_idct32_16_neon(t, dst, stride, bd);
+      t += (16 * 8);
+      dst += 8;
+    }
+  }
+}
--- a/vpx_dsp/arm/idct32x32_135_add_neon.c
+++ b/vpx_dsp/arm/idct32x32_135_add_neon.c
@@ -87,7 +87,7 @@
 // 13  84  93 103 110 125
 // 14  98 106 115 127
 // 15 117 128
-static void idct32_12_neon(const tran_low_t *const input, int16_t *output) {
+void idct32_12_neon(const tran_low_t *const input, int16_t *output) {
   int16x4_t tmp[8];
   int16x8_t in[12], s1[32], s2[32], s3[32], s4[32], s5[32], s6[32], s7[32];
 
@@ -371,10 +371,10 @@
   vst1q_s16(output, vsubq_s16(s7[0], s6[31]));
 }
 
-static void idct32_16_neon(const int16_t *const input, uint8_t *const output,
-                           const int stride) {
+void idct32_16_neon(const int16_t *const input, uint8_t *const output,
+                    const int stride, const int highbd_flag) {
   int16x8_t in[16], s1[32], s2[32], s3[32], s4[32], s5[32], s6[32], s7[32],
-      out[8];
+      out[32];
 
   load_and_transpose_s16_8x8(input, 16, &in[0], &in[1], &in[2], &in[3], &in[4],
                              &in[5], &in[6], &in[7]);
@@ -620,45 +620,44 @@
   out[5] = final_add(s7[5], s7[26]);
   out[6] = final_add(s7[6], s7[25]);
   out[7] = final_add(s7[7], s7[24]);
+  out[8] = final_add(s7[8], s7[23]);
+  out[9] = final_add(s7[9], s7[22]);
+  out[10] = final_add(s7[10], s7[21]);
+  out[11] = final_add(s7[11], s7[20]);
+  out[12] = final_add(s7[12], s6[19]);
+  out[13] = final_add(s7[13], s6[18]);
+  out[14] = final_add(s7[14], s6[17]);
+  out[15] = final_add(s7[15], s6[16]);
+  out[16] = final_sub(s7[15], s6[16]);
+  out[17] = final_sub(s7[14], s6[17]);
+  out[18] = final_sub(s7[13], s6[18]);
+  out[19] = final_sub(s7[12], s6[19]);
+  out[20] = final_sub(s7[11], s7[20]);
+  out[21] = final_sub(s7[10], s7[21]);
+  out[22] = final_sub(s7[9], s7[22]);
+  out[23] = final_sub(s7[8], s7[23]);
+  out[24] = final_sub(s7[7], s7[24]);
+  out[25] = final_sub(s7[6], s7[25]);
+  out[26] = final_sub(s7[5], s7[26]);
+  out[27] = final_sub(s7[4], s7[27]);
+  out[28] = final_sub(s7[3], s6[28]);
+  out[29] = final_sub(s7[2], s6[29]);
+  out[30] = final_sub(s7[1], s6[30]);
+  out[31] = final_sub(s7[0], s6[31]);
 
-  add_and_store_u8_s16(out[0], out[1], out[2], out[3], out[4], out[5], out[6],
-                       out[7], output, stride);
-
-  out[0] = final_add(s7[8], s7[23]);
-  out[1] = final_add(s7[9], s7[22]);
-  out[2] = final_add(s7[10], s7[21]);
-  out[3] = final_add(s7[11], s7[20]);
-  out[4] = final_add(s7[12], s6[19]);
-  out[5] = final_add(s7[13], s6[18]);
-  out[6] = final_add(s7[14], s6[17]);
-  out[7] = final_add(s7[15], s6[16]);
-
-  add_and_store_u8_s16(out[0], out[1], out[2], out[3], out[4], out[5], out[6],
-                       out[7], output + (8 * stride), stride);
-
-  out[0] = final_sub(s7[15], s6[16]);
-  out[1] = final_sub(s7[14], s6[17]);
-  out[2] = final_sub(s7[13], s6[18]);
-  out[3] = final_sub(s7[12], s6[19]);
-  out[4] = final_sub(s7[11], s7[20]);
-  out[5] = final_sub(s7[10], s7[21]);
-  out[6] = final_sub(s7[9], s7[22]);
-  out[7] = final_sub(s7[8], s7[23]);
-
-  add_and_store_u8_s16(out[0], out[1], out[2], out[3], out[4], out[5], out[6],
-                       out[7], output + (16 * stride), stride);
-
-  out[0] = final_sub(s7[7], s7[24]);
-  out[1] = final_sub(s7[6], s7[25]);
-  out[2] = final_sub(s7[5], s7[26]);
-  out[3] = final_sub(s7[4], s7[27]);
-  out[4] = final_sub(s7[3], s6[28]);
-  out[5] = final_sub(s7[2], s6[29]);
-  out[6] = final_sub(s7[1], s6[30]);
-  out[7] = final_sub(s7[0], s6[31]);
-
-  add_and_store_u8_s16(out[0], out[1], out[2], out[3], out[4], out[5], out[6],
-                       out[7], output + (24 * stride), stride);
+  if (highbd_flag) {
+    uint16_t *const outputT = CONVERT_TO_SHORTPTR(output);
+    highbd_add_and_store_bd8(out, outputT, stride);
+  } else {
+    add_and_store_u8_s16(out[0], out[1], out[2], out[3], out[4], out[5], out[6],
+                         out[7], output, stride);
+    add_and_store_u8_s16(out[8], out[9], out[10], out[11], out[12], out[13],
+                         out[14], out[15], output + (8 * stride), stride);
+    add_and_store_u8_s16(out[16], out[17], out[18], out[19], out[20], out[21],
+                         out[22], out[23], output + (16 * stride), stride);
+    add_and_store_u8_s16(out[24], out[25], out[26], out[27], out[28], out[29],
+                         out[30], out[31], output + (24 * stride), stride);
+  }
 }
 
 void vpx_idct32x32_135_add_neon(const tran_low_t *input, uint8_t *dest,
@@ -671,7 +670,7 @@
   idct32_12_neon(input + 32 * 8, temp + 8);
 
   for (i = 0; i < 32; i += 8) {
-    idct32_16_neon(t, dest, stride);
+    idct32_16_neon(t, dest, stride, 0);
     t += (16 * 8);
     dest += 8;
   }
--- a/vpx_dsp/arm/idct_neon.h
+++ b/vpx_dsp/arm/idct_neon.h
@@ -112,6 +112,24 @@
 
 //------------------------------------------------------------------------------
 
+static INLINE int32x4x2_t highbd_idct_add_dual(const int32x4x2_t s0,
+                                               const int32x4x2_t s1) {
+  int32x4x2_t t;
+  t.val[0] = vaddq_s32(s0.val[0], s1.val[0]);
+  t.val[1] = vaddq_s32(s0.val[1], s1.val[1]);
+  return t;
+}
+
+static INLINE int32x4x2_t highbd_idct_sub_dual(const int32x4x2_t s0,
+                                               const int32x4x2_t s1) {
+  int32x4x2_t t;
+  t.val[0] = vsubq_s32(s0.val[0], s1.val[0]);
+  t.val[1] = vsubq_s32(s0.val[1], s1.val[1]);
+  return t;
+}
+
+//------------------------------------------------------------------------------
+
 // Multiply a by a_const. Saturate, shift and narrow by DCT_CONST_BITS.
 static INLINE int16x8_t multiply_shift_and_narrow_s16(const int16x8_t a,
                                                       const int16_t a_const) {
@@ -169,6 +187,87 @@
                       vrshrn_n_s32(temp_high, DCT_CONST_BITS));
 }
 
+//------------------------------------------------------------------------------
+
+// Note: The following 4 functions could use 32-bit operations for bit-depth 10.
+//       However, although it's 20% faster with gcc, it's 20% slower with clang.
+//       Use 64-bit operations for now.
+
+// Multiply a by a_const. Saturate, shift and narrow by DCT_CONST_BITS.
+static INLINE int32x4x2_t
+multiply_shift_and_narrow_s32_dual(const int32x4x2_t a, const int32_t a_const) {
+  int64x2_t b[4];
+  int32x4x2_t c;
+  b[0] = vmull_n_s32(vget_low_s32(a.val[0]), a_const);
+  b[1] = vmull_n_s32(vget_high_s32(a.val[0]), a_const);
+  b[2] = vmull_n_s32(vget_low_s32(a.val[1]), a_const);
+  b[3] = vmull_n_s32(vget_high_s32(a.val[1]), a_const);
+  c.val[0] = vcombine_s32(vrshrn_n_s64(b[0], DCT_CONST_BITS),
+                          vrshrn_n_s64(b[1], DCT_CONST_BITS));
+  c.val[1] = vcombine_s32(vrshrn_n_s64(b[2], DCT_CONST_BITS),
+                          vrshrn_n_s64(b[3], DCT_CONST_BITS));
+  return c;
+}
+
+// Add a and b, then multiply by ab_const. Shift and narrow by DCT_CONST_BITS.
+static INLINE int32x4x2_t add_multiply_shift_and_narrow_s32_dual(
+    const int32x4x2_t a, const int32x4x2_t b, const int32_t ab_const) {
+  const int32x4_t temp_low = vaddq_s32(a.val[0], b.val[0]);
+  const int32x4_t temp_high = vaddq_s32(a.val[1], b.val[1]);
+  int64x2_t c[4];
+  int32x4x2_t d;
+  c[0] = vmull_n_s32(vget_low_s32(temp_low), ab_const);
+  c[1] = vmull_n_s32(vget_high_s32(temp_low), ab_const);
+  c[2] = vmull_n_s32(vget_low_s32(temp_high), ab_const);
+  c[3] = vmull_n_s32(vget_high_s32(temp_high), ab_const);
+  d.val[0] = vcombine_s32(vrshrn_n_s64(c[0], DCT_CONST_BITS),
+                          vrshrn_n_s64(c[1], DCT_CONST_BITS));
+  d.val[1] = vcombine_s32(vrshrn_n_s64(c[2], DCT_CONST_BITS),
+                          vrshrn_n_s64(c[3], DCT_CONST_BITS));
+  return d;
+}
+
+// Subtract b from a, then multiply by ab_const. Shift and narrow by
+// DCT_CONST_BITS.
+static INLINE int32x4x2_t sub_multiply_shift_and_narrow_s32_dual(
+    const int32x4x2_t a, const int32x4x2_t b, const int32_t ab_const) {
+  const int32x4_t temp_low = vsubq_s32(a.val[0], b.val[0]);
+  const int32x4_t temp_high = vsubq_s32(a.val[1], b.val[1]);
+  int64x2_t c[4];
+  int32x4x2_t d;
+  c[0] = vmull_n_s32(vget_low_s32(temp_low), ab_const);
+  c[1] = vmull_n_s32(vget_high_s32(temp_low), ab_const);
+  c[2] = vmull_n_s32(vget_low_s32(temp_high), ab_const);
+  c[3] = vmull_n_s32(vget_high_s32(temp_high), ab_const);
+  d.val[0] = vcombine_s32(vrshrn_n_s64(c[0], DCT_CONST_BITS),
+                          vrshrn_n_s64(c[1], DCT_CONST_BITS));
+  d.val[1] = vcombine_s32(vrshrn_n_s64(c[2], DCT_CONST_BITS),
+                          vrshrn_n_s64(c[3], DCT_CONST_BITS));
+  return d;
+}
+
+// Multiply a by a_const and b by b_const, then accumulate. Shift and narrow by
+// DCT_CONST_BITS.
+static INLINE int32x4x2_t multiply_accumulate_shift_and_narrow_s32_dual(
+    const int32x4x2_t a, const int32_t a_const, const int32x4x2_t b,
+    const int32_t b_const) {
+  int64x2_t c[4];
+  int32x4x2_t d;
+  c[0] = vmull_n_s32(vget_low_s32(a.val[0]), a_const);
+  c[1] = vmull_n_s32(vget_high_s32(a.val[0]), a_const);
+  c[2] = vmull_n_s32(vget_low_s32(a.val[1]), a_const);
+  c[3] = vmull_n_s32(vget_high_s32(a.val[1]), a_const);
+  c[0] = vmlal_n_s32(c[0], vget_low_s32(b.val[0]), b_const);
+  c[1] = vmlal_n_s32(c[1], vget_high_s32(b.val[0]), b_const);
+  c[2] = vmlal_n_s32(c[2], vget_low_s32(b.val[1]), b_const);
+  c[3] = vmlal_n_s32(c[3], vget_high_s32(b.val[1]), b_const);
+  d.val[0] = vcombine_s32(vrshrn_n_s64(c[0], DCT_CONST_BITS),
+                          vrshrn_n_s64(c[1], DCT_CONST_BITS));
+  d.val[1] = vcombine_s32(vrshrn_n_s64(c[2], DCT_CONST_BITS),
+                          vrshrn_n_s64(c[3], DCT_CONST_BITS));
+  return d;
+}
+
 // Shift the output down by 6 and add it to the destination buffer.
 static INLINE void add_and_store_u8_s16(const int16x8_t a0, const int16x8_t a1,
                                         const int16x8_t a2, const int16x8_t a3,
@@ -762,6 +861,108 @@
   *dest += stride;
 }
 
+static INLINE void highbd_idct16x16_add8x1_bd8(int16x8_t res, uint16_t **dest,
+                                               const int stride) {
+  uint16x8_t d = vld1q_u16(*dest);
+
+  res = vrsraq_n_s16(vreinterpretq_s16_u16(d), res, 6);
+  d = vmovl_u8(vqmovun_s16(res));
+  vst1q_u16(*dest, d);
+  *dest += stride;
+}
+
+static INLINE void highbd_add_and_store_bd8(const int16x8_t *const a,
+                                            uint16_t *out, const int b_stride) {
+  highbd_idct16x16_add8x1_bd8(a[0], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[1], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[2], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[3], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[4], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[5], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[6], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[7], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[8], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[9], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[10], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[11], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[12], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[13], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[14], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[15], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[16], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[17], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[18], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[19], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[20], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[21], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[22], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[23], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[24], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[25], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[26], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[27], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[28], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[29], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[30], &out, b_stride);
+  highbd_idct16x16_add8x1_bd8(a[31], &out, b_stride);
+}
+
+static INLINE void highbd_idct16x16_add_store(const int32x4x2_t *const out,
+                                              uint16_t *dest, const int stride,
+                                              const int bd) {
+  // Add the result to dest
+  const int16x8_t max = vdupq_n_s16((1 << bd) - 1);
+  int16x8_t o[16];
+  o[0] = vcombine_s16(vrshrn_n_s32(out[0].val[0], 6),
+                      vrshrn_n_s32(out[0].val[1], 6));
+  o[1] = vcombine_s16(vrshrn_n_s32(out[1].val[0], 6),
+                      vrshrn_n_s32(out[1].val[1], 6));
+  o[2] = vcombine_s16(vrshrn_n_s32(out[2].val[0], 6),
+                      vrshrn_n_s32(out[2].val[1], 6));
+  o[3] = vcombine_s16(vrshrn_n_s32(out[3].val[0], 6),
+                      vrshrn_n_s32(out[3].val[1], 6));
+  o[4] = vcombine_s16(vrshrn_n_s32(out[4].val[0], 6),
+                      vrshrn_n_s32(out[4].val[1], 6));
+  o[5] = vcombine_s16(vrshrn_n_s32(out[5].val[0], 6),
+                      vrshrn_n_s32(out[5].val[1], 6));
+  o[6] = vcombine_s16(vrshrn_n_s32(out[6].val[0], 6),
+                      vrshrn_n_s32(out[6].val[1], 6));
+  o[7] = vcombine_s16(vrshrn_n_s32(out[7].val[0], 6),
+                      vrshrn_n_s32(out[7].val[1], 6));
+  o[8] = vcombine_s16(vrshrn_n_s32(out[8].val[0], 6),
+                      vrshrn_n_s32(out[8].val[1], 6));
+  o[9] = vcombine_s16(vrshrn_n_s32(out[9].val[0], 6),
+                      vrshrn_n_s32(out[9].val[1], 6));
+  o[10] = vcombine_s16(vrshrn_n_s32(out[10].val[0], 6),
+                       vrshrn_n_s32(out[10].val[1], 6));
+  o[11] = vcombine_s16(vrshrn_n_s32(out[11].val[0], 6),
+                       vrshrn_n_s32(out[11].val[1], 6));
+  o[12] = vcombine_s16(vrshrn_n_s32(out[12].val[0], 6),
+                       vrshrn_n_s32(out[12].val[1], 6));
+  o[13] = vcombine_s16(vrshrn_n_s32(out[13].val[0], 6),
+                       vrshrn_n_s32(out[13].val[1], 6));
+  o[14] = vcombine_s16(vrshrn_n_s32(out[14].val[0], 6),
+                       vrshrn_n_s32(out[14].val[1], 6));
+  o[15] = vcombine_s16(vrshrn_n_s32(out[15].val[0], 6),
+                       vrshrn_n_s32(out[15].val[1], 6));
+  highbd_idct16x16_add8x1(o[0], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[1], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[2], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[3], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[4], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[5], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[6], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[7], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[8], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[9], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[10], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[11], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[12], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[13], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[14], max, &dest, stride);
+  highbd_idct16x16_add8x1(o[15], max, &dest, stride);
+}
+
 void idct16x16_256_add_half1d(const void *const input, int16_t *output,
                               void *const dest, const int stride,
                               const int highbd_flag);
@@ -775,5 +976,9 @@
 void idct16x16_10_add_half1d_pass2(const int16_t *input, int16_t *const output,
                                    void *const dest, const int stride,
                                    const int highbd_flag);
+
+void idct32_12_neon(const tran_low_t *const input, int16_t *output);
+void idct32_16_neon(const int16_t *const input, uint8_t *const output,
+                    const int stride, const int highbd_flag);
 
 #endif  // VPX_DSP_ARM_IDCT_NEON_H_
--- a/vpx_dsp/arm/transpose_neon.h
+++ b/vpx_dsp/arm/transpose_neon.h
@@ -1281,4 +1281,36 @@
 
   transpose_s16_8x8(a0, a1, a2, a3, a4, a5, a6, a7);
 }
+
+static INLINE void load_and_transpose_s32_8x8(
+    const int32_t *a, const int a_stride, int32x4x2_t *const a0,
+    int32x4x2_t *const a1, int32x4x2_t *const a2, int32x4x2_t *const a3,
+    int32x4x2_t *const a4, int32x4x2_t *const a5, int32x4x2_t *const a6,
+    int32x4x2_t *const a7) {
+  a0->val[0] = vld1q_s32(a);
+  a0->val[1] = vld1q_s32(a + 4);
+  a += a_stride;
+  a1->val[0] = vld1q_s32(a);
+  a1->val[1] = vld1q_s32(a + 4);
+  a += a_stride;
+  a2->val[0] = vld1q_s32(a);
+  a2->val[1] = vld1q_s32(a + 4);
+  a += a_stride;
+  a3->val[0] = vld1q_s32(a);
+  a3->val[1] = vld1q_s32(a + 4);
+  a += a_stride;
+  a4->val[0] = vld1q_s32(a);
+  a4->val[1] = vld1q_s32(a + 4);
+  a += a_stride;
+  a5->val[0] = vld1q_s32(a);
+  a5->val[1] = vld1q_s32(a + 4);
+  a += a_stride;
+  a6->val[0] = vld1q_s32(a);
+  a6->val[1] = vld1q_s32(a + 4);
+  a += a_stride;
+  a7->val[0] = vld1q_s32(a);
+  a7->val[1] = vld1q_s32(a + 4);
+
+  transpose_s32_8x8(a0, a1, a2, a3, a4, a5, a6, a7);
+}
 #endif  // VPX_DSP_ARM_TRANSPOSE_NEON_H_
--- a/vpx_dsp/vpx_dsp.mk
+++ b/vpx_dsp/vpx_dsp.mk
@@ -224,6 +224,7 @@
 DSP_SRCS-$(HAVE_NEON)  += arm/highbd_idct8x8_add_neon.c
 DSP_SRCS-$(HAVE_NEON)  += arm/highbd_idct16x16_add_neon.c
 DSP_SRCS-$(HAVE_NEON)  += arm/highbd_idct32x32_add_neon.c
+DSP_SRCS-$(HAVE_NEON)  += arm/highbd_idct32x32_135_add_neon.c
 endif  # !CONFIG_VP9_HIGHBITDEPTH
 
 ifeq ($(HAVE_NEON_ASM),yes)
--- a/vpx_dsp/vpx_dsp_rtcd_defs.pl
+++ b/vpx_dsp/vpx_dsp_rtcd_defs.pl
@@ -628,8 +628,6 @@
 
   add_proto qw/void vpx_highbd_idct32x32_1024_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
 
-  add_proto qw/void vpx_highbd_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
-
   add_proto qw/void vpx_highbd_idct32x32_34_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
 
   add_proto qw/void vpx_highbd_idct32x32_1_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
@@ -678,6 +676,8 @@
     add_proto qw/void vpx_highbd_idct16x16_38_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
 
     add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
+
+    add_proto qw/void vpx_highbd_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
   } else {
     add_proto qw/void vpx_idct4x4_16_add/, "const tran_low_t *input, uint8_t *dest, int stride";
     specialize qw/vpx_idct4x4_16_add neon sse2/;
@@ -739,6 +739,9 @@
 
     add_proto qw/void vpx_highbd_idct16x16_10_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
     specialize qw/vpx_highbd_idct16x16_10_add neon sse2/;
+
+    add_proto qw/void vpx_highbd_idct32x32_135_add/, "const tran_low_t *input, uint8_t *dest, int stride, int bd";
+    specialize qw/vpx_highbd_idct32x32_135_add neon/;
   }  # CONFIG_EMULATE_HARDWARE
 } else {
   # Force C versions if CONFIG_EMULATE_HARDWARE is 1