shithub: dav1d

Download patch

ref: 018e64e714faadcf3840bd6a0a52cf6a8a3acde2
parent: e41a2a1fe0e10a8b2e22f238acb596c35b2f2f7f
author: Martin Storsjö <martin@martin.st>
date: Wed Nov 18 09:20:18 EST 2020

arm32: cdef: Add NEON implementations of CDEF for 16 bpc

Use a shared template file for assembly functions that can be
templated into 8 and 16 bpc forms, just like in the arm64 version.

Checkasm benchmarks:
                          Cortex A7      A8     A53     A72     A73
cdef_dir_16bpc_neon:          975.9   853.2   555.2   378.7   386.9
cdef_filter_4x4_16bpc_neon:   746.9   521.7   481.2   333.0   340.8
cdef_filter_4x8_16bpc_neon:  1300.0   885.5   816.3   582.7   599.5
cdef_filter_8x8_16bpc_neon:  2282.5  1415.0  1417.6  1059.0  1076.3

Corresponding numbers for arm64, for comparison:
                                         Cortex A53     A72     A73
cdef_dir_16bpc_neon:                          418.0   306.7   310.7
cdef_filter_4x4_16bpc_neon:                   453.4   282.9   297.4
cdef_filter_4x8_16bpc_neon:                   807.5   514.2   533.8
cdef_filter_8x8_16bpc_neon:                  1425.2   924.4   942.0

--- a/src/arm/32/cdef.S
+++ b/src/arm/32/cdef.S
@@ -27,6 +27,7 @@
 
 #include "src/arm/asm.S"
 #include "util.S"
+#include "cdef_tmpl.S"
 
 // n1 = s0/d0
 // w1 = d0/q0
@@ -324,231 +325,13 @@
 padding_func_edged 8, 16, d0, 64
 padding_func_edged 4, 8,  s0, 32
 
-.macro dir_table w, stride
-const directions\w
-        .byte           -1 * \stride + 1, -2 * \stride + 2
-        .byte            0 * \stride + 1, -1 * \stride + 2
-        .byte            0 * \stride + 1,  0 * \stride + 2
-        .byte            0 * \stride + 1,  1 * \stride + 2
-        .byte            1 * \stride + 1,  2 * \stride + 2
-        .byte            1 * \stride + 0,  2 * \stride + 1
-        .byte            1 * \stride + 0,  2 * \stride + 0
-        .byte            1 * \stride + 0,  2 * \stride - 1
-// Repeated, to avoid & 7
-        .byte           -1 * \stride + 1, -2 * \stride + 2
-        .byte            0 * \stride + 1, -1 * \stride + 2
-        .byte            0 * \stride + 1,  0 * \stride + 2
-        .byte            0 * \stride + 1,  1 * \stride + 2
-        .byte            1 * \stride + 1,  2 * \stride + 2
-        .byte            1 * \stride + 0,  2 * \stride + 1
-endconst
-.endm
+tables
 
-dir_table 8, 16
-dir_table 4, 8
+filter 8, 8
+filter 4, 8
 
-const pri_taps
-        .byte           4, 2, 3, 3
-endconst
+find_dir 8
 
-.macro load_px d11, d12, d21, d22, w
-.if \w == 8
-        add             r6,  r2,  r9, lsl #1 // x + off
-        sub             r9,  r2,  r9, lsl #1 // x - off
-        vld1.16         {\d11,\d12}, [r6]    // p0
-        vld1.16         {\d21,\d22}, [r9]    // p1
-.else
-        add             r6,  r2,  r9, lsl #1 // x + off
-        sub             r9,  r2,  r9, lsl #1 // x - off
-        vld1.16         {\d11}, [r6]         // p0
-        add             r6,  r6,  #2*8       // += stride
-        vld1.16         {\d21}, [r9]         // p1
-        add             r9,  r9,  #2*8       // += stride
-        vld1.16         {\d12}, [r6]         // p0
-        vld1.16         {\d22}, [r9]         // p1
-.endif
-.endm
-.macro handle_pixel s1, s2, thresh_vec, shift, tap, min
-.if \min
-        vmin.u16        q2,  q2,  \s1
-        vmax.s16        q3,  q3,  \s1
-        vmin.u16        q2,  q2,  \s2
-        vmax.s16        q3,  q3,  \s2
-.endif
-        vabd.u16        q8,  q0,  \s1        // abs(diff)
-        vabd.u16        q11, q0,  \s2        // abs(diff)
-        vshl.u16        q9,  q8,  \shift     // abs(diff) >> shift
-        vshl.u16        q12, q11, \shift     // abs(diff) >> shift
-        vqsub.u16       q9,  \thresh_vec, q9 // clip = imax(0, threshold - (abs(diff) >> shift))
-        vqsub.u16       q12, \thresh_vec, q12// clip = imax(0, threshold - (abs(diff) >> shift))
-        vsub.i16        q10, \s1, q0         // diff = p0 - px
-        vsub.i16        q13, \s2, q0         // diff = p1 - px
-        vneg.s16        q8,  q9              // -clip
-        vneg.s16        q11, q12             // -clip
-        vmin.s16        q10, q10, q9         // imin(diff, clip)
-        vmin.s16        q13, q13, q12        // imin(diff, clip)
-        vdup.16         q9,  \tap            // taps[k]
-        vmax.s16        q10, q10, q8         // constrain() = imax(imin(diff, clip), -clip)
-        vmax.s16        q13, q13, q11        // constrain() = imax(imin(diff, clip), -clip)
-        vmla.i16        q1,  q10, q9         // sum += taps[k] * constrain()
-        vmla.i16        q1,  q13, q9         // sum += taps[k] * constrain()
-.endm
-
-// void dav1d_cdef_filterX_8bpc_neon(pixel *dst, ptrdiff_t dst_stride,
-//                                   const uint16_t *tmp, int pri_strength,
-//                                   int sec_strength, int dir, int damping,
-//                                   int h, size_t edges);
-.macro filter_func w, pri, sec, min, suffix
-function cdef_filter\w\suffix\()_neon
-        cmp             r8,  #0xf
-        beq             cdef_filter\w\suffix\()_edged_neon
-.if \pri
-        movrel_local    r8,  pri_taps
-        and             r9,  r3,  #1
-        add             r8,  r8,  r9, lsl #1
-.endif
-        movrel_local    r9,  directions\w
-        add             r5,  r9,  r5, lsl #1
-        vmov.u16        d17, #15
-        vdup.16         d16, r6              // damping
-
-.if \pri
-        vdup.16         q5,  r3              // threshold
-.endif
-.if \sec
-        vdup.16         q7,  r4              // threshold
-.endif
-        vmov.16         d8[0], r3
-        vmov.16         d8[1], r4
-        vclz.i16        d8,  d8              // clz(threshold)
-        vsub.i16        d8,  d17, d8         // ulog2(threshold)
-        vqsub.u16       d8,  d16, d8         // shift = imax(0, damping - ulog2(threshold))
-        vneg.s16        d8,  d8              // -shift
-.if \sec
-        vdup.16         q6,  d8[1]
-.endif
-.if \pri
-        vdup.16         q4,  d8[0]
-.endif
-
-1:
-.if \w == 8
-        vld1.16         {q0},  [r2, :128]    // px
-.else
-        add             r12, r2,  #2*8
-        vld1.16         {d0},  [r2,  :64]    // px
-        vld1.16         {d1},  [r12, :64]    // px
-.endif
-
-        vmov.u16        q1,  #0              // sum
-.if \min
-        vmov.u16        q2,  q0              // min
-        vmov.u16        q3,  q0              // max
-.endif
-
-        // Instead of loading sec_taps 2, 1 from memory, just set it
-        // to 2 initially and decrease for the second round.
-        // This is also used as loop counter.
-        mov             lr,  #2              // sec_taps[0]
-
-2:
-.if \pri
-        ldrsb           r9,  [r5]            // off1
-
-        load_px         d28, d29, d30, d31, \w
-.endif
-
-.if \sec
-        add             r5,  r5,  #4         // +2*2
-        ldrsb           r9,  [r5]            // off2
-.endif
-
-.if \pri
-        ldrb            r12, [r8]            // *pri_taps
-
-        handle_pixel    q14, q15, q5,  q4,  r12, \min
-.endif
-
-.if \sec
-        load_px         d28, d29, d30, d31, \w
-
-        add             r5,  r5,  #8         // +2*4
-        ldrsb           r9,  [r5]            // off3
-
-        handle_pixel    q14, q15, q7,  q6,  lr, \min
-
-        load_px         d28, d29, d30, d31, \w
-
-        handle_pixel    q14, q15, q7,  q6,  lr, \min
-
-        sub             r5,  r5,  #11        // r5 -= 2*(2+4); r5 += 1;
-.else
-        add             r5,  r5,  #1         // r5 += 1
-.endif
-        subs            lr,  lr,  #1         // sec_tap-- (value)
-.if \pri
-        add             r8,  r8,  #1         // pri_taps++ (pointer)
-.endif
-        bne             2b
-
-        vshr.s16        q14, q1,  #15        // -(sum < 0)
-        vadd.i16        q1,  q1,  q14        // sum - (sum < 0)
-        vrshr.s16       q1,  q1,  #4         // (8 + sum - (sum < 0)) >> 4
-        vadd.i16        q0,  q0,  q1         // px + (8 + sum ...) >> 4
-.if \min
-        vmin.s16        q0,  q0,  q3
-        vmax.s16        q0,  q0,  q2         // iclip(px + .., min, max)
-.endif
-        vmovn.u16       d0,  q0
-.if \w == 8
-        add             r2,  r2,  #2*16      // tmp += tmp_stride
-        subs            r7,  r7,  #1         // h--
-        vst1.8          {d0}, [r0, :64], r1
-.else
-        vst1.32         {d0[0]}, [r0, :32], r1
-        add             r2,  r2,  #2*16      // tmp += 2*tmp_stride
-        subs            r7,  r7,  #2         // h -= 2
-        vst1.32         {d0[1]}, [r0, :32], r1
-.endif
-
-        // Reset pri_taps and directions back to the original point
-        sub             r5,  r5,  #2
-.if \pri
-        sub             r8,  r8,  #2
-.endif
-
-        bgt             1b
-        vpop            {q4-q7}
-        pop             {r4-r9,pc}
-endfunc
-.endm
-
-.macro filter w
-filter_func \w, pri=1, sec=0, min=0, suffix=_pri
-filter_func \w, pri=0, sec=1, min=0, suffix=_sec
-filter_func \w, pri=1, sec=1, min=1, suffix=_pri_sec
-
-function cdef_filter\w\()_8bpc_neon, export=1
-        push            {r4-r9,lr}
-        vpush           {q4-q7}
-        ldrd            r4,  r5,  [sp, #92]
-        ldrd            r6,  r7,  [sp, #100]
-        ldr             r8,  [sp, #108]
-        cmp             r3,  #0 // pri_strength
-        bne             1f
-        b               cdef_filter\w\()_sec_neon // only sec
-1:
-        cmp             r4,  #0 // sec_strength
-        bne             1f
-        b               cdef_filter\w\()_pri_neon // only pri
-1:
-        b               cdef_filter\w\()_pri_sec_neon // both pri and sec
-endfunc
-.endm
-
-filter 8
-filter 4
-
 .macro load_px_8 d11, d12, d21, d22, w
 .if \w == 8
         add             r6,  r2,  r9         // x + off
@@ -753,219 +536,3 @@
 
 filter_8 8
 filter_8 4
-
-const div_table, align=4
-        .short         840, 420, 280, 210, 168, 140, 120, 105
-endconst
-
-const alt_fact, align=4
-        .short         420, 210, 140, 105, 105, 105, 105, 105, 140, 210, 420, 0
-endconst
-
-// int dav1d_cdef_find_dir_8bpc_neon(const pixel *img, const ptrdiff_t stride,
-//                                   unsigned *const var)
-function cdef_find_dir_8bpc_neon, export=1
-        push            {lr}
-        vpush           {q4-q7}
-        sub             sp,  sp,  #32          // cost
-        mov             r3,  #8
-        vmov.u16        q1,  #0                // q0-q1   sum_diag[0]
-        vmov.u16        q3,  #0                // q2-q3   sum_diag[1]
-        vmov.u16        q5,  #0                // q4-q5   sum_hv[0-1]
-        vmov.u16        q8,  #0                // q6,d16  sum_alt[0]
-                                               // q7,d17  sum_alt[1]
-        vmov.u16        q9,  #0                // q9,d22  sum_alt[2]
-        vmov.u16        q11, #0
-        vmov.u16        q10, #0                // q10,d23 sum_alt[3]
-
-
-.irpc i, 01234567
-        vld1.8          {d30}, [r0, :64], r1
-        vmov.u8         d31, #128
-        vsubl.u8        q15, d30, d31          // img[x] - 128
-        vmov.u16        q14, #0
-
-.if \i == 0
-        vmov            q0,  q15               // sum_diag[0]
-.else
-        vext.8          q12, q14, q15, #(16-2*\i)
-        vext.8          q13, q15, q14, #(16-2*\i)
-        vadd.i16        q0,  q0,  q12          // sum_diag[0]
-        vadd.i16        q1,  q1,  q13          // sum_diag[0]
-.endif
-        vrev64.16       q13, q15
-        vswp            d26, d27               // [-x]
-.if \i == 0
-        vmov            q2,  q13               // sum_diag[1]
-.else
-        vext.8          q12, q14, q13, #(16-2*\i)
-        vext.8          q13, q13, q14, #(16-2*\i)
-        vadd.i16        q2,  q2,  q12          // sum_diag[1]
-        vadd.i16        q3,  q3,  q13          // sum_diag[1]
-.endif
-
-        vpadd.u16       d26, d30, d31          // [(x >> 1)]
-        vmov.u16        d27, #0
-        vpadd.u16       d24, d26, d28
-        vpadd.u16       d24, d24, d28          // [y]
-        vmov.u16        r12, d24[0]
-        vadd.i16        q5,  q5,  q15          // sum_hv[1]
-.if \i < 4
-        vmov.16         d8[\i],   r12          // sum_hv[0]
-.else
-        vmov.16         d9[\i-4], r12          // sum_hv[0]
-.endif
-
-.if \i == 0
-        vmov.u16        q6,  q13               // sum_alt[0]
-.else
-        vext.8          q12, q14, q13, #(16-2*\i)
-        vext.8          q14, q13, q14, #(16-2*\i)
-        vadd.i16        q6,  q6,  q12          // sum_alt[0]
-        vadd.i16        d16, d16, d28          // sum_alt[0]
-.endif
-        vrev64.16       d26, d26               // [-(x >> 1)]
-        vmov.u16        q14, #0
-.if \i == 0
-        vmov            q7,  q13               // sum_alt[1]
-.else
-        vext.8          q12, q14, q13, #(16-2*\i)
-        vext.8          q13, q13, q14, #(16-2*\i)
-        vadd.i16        q7,  q7,  q12          // sum_alt[1]
-        vadd.i16        d17, d17, d26          // sum_alt[1]
-.endif
-
-.if \i < 6
-        vext.8          q12, q14, q15, #(16-2*(3-(\i/2)))
-        vext.8          q13, q15, q14, #(16-2*(3-(\i/2)))
-        vadd.i16        q9,  q9,  q12          // sum_alt[2]
-        vadd.i16        d22, d22, d26          // sum_alt[2]
-.else
-        vadd.i16        q9,  q9,  q15          // sum_alt[2]
-.endif
-.if \i == 0
-        vmov            q10, q15               // sum_alt[3]
-.elseif \i == 1
-        vadd.i16        q10, q10, q15          // sum_alt[3]
-.else
-        vext.8          q12, q14, q15, #(16-2*(\i/2))
-        vext.8          q13, q15, q14, #(16-2*(\i/2))
-        vadd.i16        q10, q10, q12          // sum_alt[3]
-        vadd.i16        d23, d23, d26          // sum_alt[3]
-.endif
-.endr
-
-        vmov.u32        q15, #105
-
-        vmull.s16       q12, d8,  d8           // sum_hv[0]*sum_hv[0]
-        vmlal.s16       q12, d9,  d9
-        vmull.s16       q13, d10, d10          // sum_hv[1]*sum_hv[1]
-        vmlal.s16       q13, d11, d11
-        vadd.s32        d8,  d24, d25
-        vadd.s32        d9,  d26, d27
-        vpadd.s32       d8,  d8,  d9           // cost[2,6] (s16, s17)
-        vmul.i32        d8,  d8,  d30          // cost[2,6] *= 105
-
-        vrev64.16       q1,  q1
-        vrev64.16       q3,  q3
-        vext.8          q1,  q1,  q1,  #10     // sum_diag[0][14-n]
-        vext.8          q3,  q3,  q3,  #10     // sum_diag[1][14-n]
-
-        vstr            s16, [sp, #2*4]        // cost[2]
-        vstr            s17, [sp, #6*4]        // cost[6]
-
-        movrel_local    r12, div_table
-        vld1.16         {q14}, [r12, :128]
-
-        vmull.s16       q5,  d0,  d0           // sum_diag[0]*sum_diag[0]
-        vmull.s16       q12, d1,  d1
-        vmlal.s16       q5,  d2,  d2
-        vmlal.s16       q12, d3,  d3
-        vmull.s16       q0,  d4,  d4           // sum_diag[1]*sum_diag[1]
-        vmull.s16       q1,  d5,  d5
-        vmlal.s16       q0,  d6,  d6
-        vmlal.s16       q1,  d7,  d7
-        vmovl.u16       q13, d28               // div_table
-        vmovl.u16       q14, d29
-        vmul.i32        q5,  q5,  q13          // cost[0]
-        vmla.i32        q5,  q12, q14
-        vmul.i32        q0,  q0,  q13          // cost[4]
-        vmla.i32        q0,  q1,  q14
-        vadd.i32        d10, d10, d11
-        vadd.i32        d0,  d0,  d1
-        vpadd.i32       d0,  d10, d0           // cost[0,4] = s0,s1
-
-        movrel_local    r12, alt_fact
-        vld1.16         {d29, d30, d31}, [r12, :64] // div_table[2*m+1] + 105
-
-        vstr            s0,  [sp, #0*4]        // cost[0]
-        vstr            s1,  [sp, #4*4]        // cost[4]
-
-        vmovl.u16       q13, d29               // div_table[2*m+1] + 105
-        vmovl.u16       q14, d30
-        vmovl.u16       q15, d31
-
-.macro cost_alt dest, s1, s2, s3, s4, s5, s6
-        vmull.s16       q1,  \s1, \s1          // sum_alt[n]*sum_alt[n]
-        vmull.s16       q2,  \s2, \s2
-        vmull.s16       q3,  \s3, \s3
-        vmull.s16       q5,  \s4, \s4          // sum_alt[n]*sum_alt[n]
-        vmull.s16       q12, \s5, \s5
-        vmull.s16       q6,  \s6, \s6          // q6 overlaps the first \s1-\s2 here
-        vmul.i32        q1,  q1,  q13          // sum_alt[n]^2*fact
-        vmla.i32        q1,  q2,  q14
-        vmla.i32        q1,  q3,  q15
-        vmul.i32        q5,  q5,  q13          // sum_alt[n]^2*fact
-        vmla.i32        q5,  q12, q14
-        vmla.i32        q5,  q6,  q15
-        vadd.i32        d2,  d2,  d3
-        vadd.i32        d3,  d10, d11
-        vpadd.i32       \dest, d2, d3          // *cost_ptr
-.endm
-        cost_alt        d14, d12, d13, d16, d14, d15, d17 // cost[1], cost[3]
-        cost_alt        d15, d18, d19, d22, d20, d21, d23 // cost[5], cost[7]
-        vstr            s28, [sp, #1*4]        // cost[1]
-        vstr            s29, [sp, #3*4]        // cost[3]
-
-        mov             r0,  #0                // best_dir
-        vmov.32         r1,  d0[0]             // best_cost
-        mov             r3,  #1                // n
-
-        vstr            s30, [sp, #5*4]        // cost[5]
-        vstr            s31, [sp, #7*4]        // cost[7]
-
-        vmov.32         r12, d14[0]
-
-.macro find_best s1, s2, s3
-.ifnb \s2
-        vmov.32         lr,  \s2
-.endif
-        cmp             r12, r1                // cost[n] > best_cost
-        itt             gt
-        movgt           r0,  r3                // best_dir = n
-        movgt           r1,  r12               // best_cost = cost[n]
-.ifnb \s2
-        add             r3,  r3,  #1           // n++
-        cmp             lr,  r1                // cost[n] > best_cost
-        vmov.32         r12, \s3
-        itt             gt
-        movgt           r0,  r3                // best_dir = n
-        movgt           r1,  lr                // best_cost = cost[n]
-        add             r3,  r3,  #1           // n++
-.endif
-.endm
-        find_best       d14[0], d8[0], d14[1]
-        find_best       d14[1], d0[1], d15[0]
-        find_best       d15[0], d8[1], d15[1]
-        find_best       d15[1]
-
-        eor             r3,  r0,  #4           // best_dir ^4
-        ldr             r12, [sp, r3, lsl #2]
-        sub             r1,  r1,  r12          // best_cost - cost[best_dir ^ 4]
-        lsr             r1,  r1,  #10
-        str             r1,  [r2]              // *var
-
-        add             sp,  sp,  #32
-        vpop            {q4-q7}
-        pop             {pc}
-endfunc
--- /dev/null
+++ b/src/arm/32/cdef16.S
@@ -1,0 +1,232 @@
+/*
+ * Copyright © 2018, VideoLAN and dav1d authors
+ * Copyright © 2020, Martin Storsjo
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice, this
+ *    list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ *    this list of conditions and the following disclaimer in the documentation
+ *    and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+ * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "src/arm/asm.S"
+#include "util.S"
+#include "cdef_tmpl.S"
+
+// r1 = d0/q0
+// r2 = d2/q1
+.macro pad_top_bot_16 s1, s2, w, stride, r1, r2, align, ret
+        tst             r6,  #1 // CDEF_HAVE_LEFT
+        beq             2f
+        // CDEF_HAVE_LEFT
+        tst             r6,  #2 // CDEF_HAVE_RIGHT
+        beq             1f
+        // CDEF_HAVE_LEFT+CDEF_HAVE_RIGHT
+        vldr            s8,  [\s1, #-4]
+        vld1.16         {\r1}, [\s1, :\align]
+        vldr            s9,  [\s1, #2*\w]
+        vldr            s10, [\s2, #-4]
+        vld1.16         {\r2}, [\s2, :\align]
+        vldr            s11, [\s2, #2*\w]
+        vstr            s8,  [r0, #-4]
+        vst1.16         {\r1}, [r0, :\align]
+        vstr            s9,  [r0, #2*\w]
+        add             r0,  r0,  #2*\stride
+        vstr            s10, [r0, #-4]
+        vst1.16         {\r2}, [r0, :\align]
+        vstr            s11, [r0, #2*\w]
+.if \ret
+        pop             {r4-r7,pc}
+.else
+        add             r0,  r0,  #2*\stride
+        b               3f
+.endif
+
+1:
+        // CDEF_HAVE_LEFT+!CDEF_HAVE_RIGHT
+        vldr            s8,  [\s1, #-4]
+        vld1.16         {\r1}, [\s1, :\align]
+        vldr            s9,  [\s2, #-4]
+        vld1.16         {\r2}, [\s2, :\align]
+        vstr            s8,  [r0, #-4]
+        vst1.16         {\r1}, [r0, :\align]
+        vstr            s12, [r0, #2*\w]
+        add             r0,  r0,  #2*\stride
+        vstr            s9,  [r0, #-4]
+        vst1.16         {\r2}, [r0, :\align]
+        vstr            s12, [r0, #2*\w]
+.if \ret
+        pop             {r4-r7,pc}
+.else
+        add             r0,  r0,  #2*\stride
+        b               3f
+.endif
+
+2:
+        // !CDEF_HAVE_LEFT
+        tst             r6,  #2 // CDEF_HAVE_RIGHT
+        beq             1f
+        // !CDEF_HAVE_LEFT+CDEF_HAVE_RIGHT
+        vld1.16         {\r1}, [\s1, :\align]
+        vldr            s8,  [\s1, #2*\w]
+        vld1.16         {\r2}, [\s2, :\align]
+        vldr            s9,  [\s2, #2*\w]
+        vstr            s12, [r0, #-4]
+        vst1.16         {\r1}, [r0, :\align]
+        vstr            s8,  [r0, #2*\w]
+        add             r0,  r0,  #2*\stride
+        vstr            s12, [r0, #-4]
+        vst1.16         {\r2}, [r0, :\align]
+        vstr            s9,  [r0, #2*\w]
+.if \ret
+        pop             {r4-r7,pc}
+.else
+        add             r0,  r0,  #2*\stride
+        b               3f
+.endif
+
+1:
+        // !CDEF_HAVE_LEFT+!CDEF_HAVE_RIGHT
+        vld1.16         {\r1}, [\s1, :\align]
+        vld1.16         {\r2}, [\s2, :\align]
+        vstr            s12, [r0, #-4]
+        vst1.16         {\r1}, [r0, :\align]
+        vstr            s12, [r0, #2*\w]
+        add             r0,  r0,  #2*\stride
+        vstr            s12, [r0, #-4]
+        vst1.16         {\r2}, [r0, :\align]
+        vstr            s12, [r0, #2*\w]
+.if \ret
+        pop             {r4-r7,pc}
+.else
+        add             r0,  r0,  #2*\stride
+.endif
+3:
+.endm
+
+// void dav1d_cdef_paddingX_16bpc_neon(uint16_t *tmp, const pixel *src,
+//                                     ptrdiff_t src_stride, const pixel (*left)[2],
+//                                     const pixel *const top, int h,
+//                                     enum CdefEdgeFlags edges);
+
+// r1 = d0/q0
+// r2 = d2/q1
+.macro padding_func_16 w, stride, r1, r2, align
+function cdef_padding\w\()_16bpc_neon, export=1
+        push            {r4-r7,lr}
+        ldrd            r4,  r5,  [sp, #20]
+        ldr             r6,  [sp, #28]
+        vmov.i16        q3,  #0x8000
+        tst             r6,  #4 // CDEF_HAVE_TOP
+        bne             1f
+        // !CDEF_HAVE_TOP
+        sub             r12, r0,  #2*(2*\stride+2)
+        vmov.i16        q2,  #0x8000
+        vst1.16         {q2,q3}, [r12]!
+.if \w == 8
+        vst1.16         {q2,q3}, [r12]!
+.endif
+        b               3f
+1:
+        // CDEF_HAVE_TOP
+        add             r7,  r4,  r2
+        sub             r0,  r0,  #2*(2*\stride)
+        pad_top_bot_16  r4,  r7,  \w, \stride, \r1, \r2, \align, 0
+
+        // Middle section
+3:
+        tst             r6,  #1 // CDEF_HAVE_LEFT
+        beq             2f
+        // CDEF_HAVE_LEFT
+        tst             r6,  #2 // CDEF_HAVE_RIGHT
+        beq             1f
+        // CDEF_HAVE_LEFT+CDEF_HAVE_RIGHT
+0:
+        vld1.32         {d2[]}, [r3, :32]!
+        vldr            s5,  [r1, #2*\w]
+        vld1.16         {\r1}, [r1, :\align], r2
+        subs            r5,  r5,  #1
+        vstr            s4,  [r0, #-4]
+        vst1.16         {\r1}, [r0, :\align]
+        vstr            s5,  [r0, #2*\w]
+        add             r0,  r0,  #2*\stride
+        bgt             0b
+        b               3f
+1:
+        // CDEF_HAVE_LEFT+!CDEF_HAVE_RIGHT
+        vld1.32         {d2[]}, [r3, :32]!
+        vld1.16         {\r1}, [r1, :\align], r2
+        subs            r5,  r5,  #1
+        vstr            s4,  [r0, #-4]
+        vst1.16         {\r1}, [r0, :\align]
+        vstr            s12, [r0, #2*\w]
+        add             r0,  r0,  #2*\stride
+        bgt             1b
+        b               3f
+2:
+        tst             r6,  #2 // CDEF_HAVE_RIGHT
+        beq             1f
+        // !CDEF_HAVE_LEFT+CDEF_HAVE_RIGHT
+0:
+        vldr            s4,  [r1, #2*\w]
+        vld1.16         {\r1}, [r1, :\align], r2
+        subs            r5,  r5,  #1
+        vstr            s12, [r0, #-4]
+        vst1.16         {\r1}, [r0, :\align]
+        vstr            s4,  [r0, #2*\w]
+        add             r0,  r0,  #2*\stride
+        bgt             0b
+        b               3f
+1:
+        // !CDEF_HAVE_LEFT+!CDEF_HAVE_RIGHT
+        vld1.16         {\r1}, [r1, :\align], r2
+        subs            r5,  r5,  #1
+        vstr            s12, [r0, #-4]
+        vst1.16         {\r1}, [r0, :\align]
+        vstr            s12, [r0, #2*\w]
+        add             r0,  r0,  #2*\stride
+        bgt             1b
+
+3:
+        tst             r6,  #8 // CDEF_HAVE_BOTTOM
+        bne             1f
+        // !CDEF_HAVE_BOTTOM
+        sub             r12, r0,  #4
+        vmov.i16        q2,  #0x8000
+        vst1.16         {q2,q3}, [r12]!
+.if \w == 8
+        vst1.16         {q2,q3}, [r12]!
+.endif
+        pop             {r4-r7,pc}
+1:
+        // CDEF_HAVE_BOTTOM
+        add             r7,  r1,  r2
+        pad_top_bot_16  r1,  r7,  \w, \stride, \r1, \r2, \align, 1
+endfunc
+.endm
+
+padding_func_16 8, 16, q0, q1, 128
+padding_func_16 4, 8,  d0, d2, 64
+
+tables
+
+filter 8, 16
+filter 4, 16
+
+find_dir 16
--- /dev/null
+++ b/src/arm/32/cdef_tmpl.S
@@ -1,0 +1,515 @@
+/*
+ * Copyright © 2018, VideoLAN and dav1d authors
+ * Copyright © 2020, Martin Storsjo
+ * All rights reserved.
+ *
+ * Redistribution and use in source and binary forms, with or without
+ * modification, are permitted provided that the following conditions are met:
+ *
+ * 1. Redistributions of source code must retain the above copyright notice, this
+ *    list of conditions and the following disclaimer.
+ *
+ * 2. Redistributions in binary form must reproduce the above copyright notice,
+ *    this list of conditions and the following disclaimer in the documentation
+ *    and/or other materials provided with the distribution.
+ *
+ * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
+ * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
+ * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
+ * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
+ * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
+ * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
+ * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
+ * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
+ * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
+ * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+ */
+
+#include "src/arm/asm.S"
+#include "util.S"
+
+.macro dir_table w, stride
+const directions\w
+        .byte           -1 * \stride + 1, -2 * \stride + 2
+        .byte            0 * \stride + 1, -1 * \stride + 2
+        .byte            0 * \stride + 1,  0 * \stride + 2
+        .byte            0 * \stride + 1,  1 * \stride + 2
+        .byte            1 * \stride + 1,  2 * \stride + 2
+        .byte            1 * \stride + 0,  2 * \stride + 1
+        .byte            1 * \stride + 0,  2 * \stride + 0
+        .byte            1 * \stride + 0,  2 * \stride - 1
+// Repeated, to avoid & 7
+        .byte           -1 * \stride + 1, -2 * \stride + 2
+        .byte            0 * \stride + 1, -1 * \stride + 2
+        .byte            0 * \stride + 1,  0 * \stride + 2
+        .byte            0 * \stride + 1,  1 * \stride + 2
+        .byte            1 * \stride + 1,  2 * \stride + 2
+        .byte            1 * \stride + 0,  2 * \stride + 1
+endconst
+.endm
+
+.macro tables
+dir_table 8, 16
+dir_table 4, 8
+
+const pri_taps
+        .byte           4, 2, 3, 3
+endconst
+.endm
+
+.macro load_px d11, d12, d21, d22, w
+.if \w == 8
+        add             r6,  r2,  r9, lsl #1 // x + off
+        sub             r9,  r2,  r9, lsl #1 // x - off
+        vld1.16         {\d11,\d12}, [r6]    // p0
+        vld1.16         {\d21,\d22}, [r9]    // p1
+.else
+        add             r6,  r2,  r9, lsl #1 // x + off
+        sub             r9,  r2,  r9, lsl #1 // x - off
+        vld1.16         {\d11}, [r6]         // p0
+        add             r6,  r6,  #2*8       // += stride
+        vld1.16         {\d21}, [r9]         // p1
+        add             r9,  r9,  #2*8       // += stride
+        vld1.16         {\d12}, [r6]         // p0
+        vld1.16         {\d22}, [r9]         // p1
+.endif
+.endm
+.macro handle_pixel s1, s2, thresh_vec, shift, tap, min
+.if \min
+        vmin.u16        q2,  q2,  \s1
+        vmax.s16        q3,  q3,  \s1
+        vmin.u16        q2,  q2,  \s2
+        vmax.s16        q3,  q3,  \s2
+.endif
+        vabd.u16        q8,  q0,  \s1        // abs(diff)
+        vabd.u16        q11, q0,  \s2        // abs(diff)
+        vshl.u16        q9,  q8,  \shift     // abs(diff) >> shift
+        vshl.u16        q12, q11, \shift     // abs(diff) >> shift
+        vqsub.u16       q9,  \thresh_vec, q9 // clip = imax(0, threshold - (abs(diff) >> shift))
+        vqsub.u16       q12, \thresh_vec, q12// clip = imax(0, threshold - (abs(diff) >> shift))
+        vsub.i16        q10, \s1, q0         // diff = p0 - px
+        vsub.i16        q13, \s2, q0         // diff = p1 - px
+        vneg.s16        q8,  q9              // -clip
+        vneg.s16        q11, q12             // -clip
+        vmin.s16        q10, q10, q9         // imin(diff, clip)
+        vmin.s16        q13, q13, q12        // imin(diff, clip)
+        vdup.16         q9,  \tap            // taps[k]
+        vmax.s16        q10, q10, q8         // constrain() = imax(imin(diff, clip), -clip)
+        vmax.s16        q13, q13, q11        // constrain() = imax(imin(diff, clip), -clip)
+        vmla.i16        q1,  q10, q9         // sum += taps[k] * constrain()
+        vmla.i16        q1,  q13, q9         // sum += taps[k] * constrain()
+.endm
+
+// void dav1d_cdef_filterX_Ybpc_neon(pixel *dst, ptrdiff_t dst_stride,
+//                                   const uint16_t *tmp, int pri_strength,
+//                                   int sec_strength, int dir, int damping,
+//                                   int h, size_t edges);
+.macro filter_func w, bpc, pri, sec, min, suffix
+function cdef_filter\w\suffix\()_\bpc\()bpc_neon
+.if \bpc == 8
+        cmp             r8,  #0xf
+        beq             cdef_filter\w\suffix\()_edged_neon
+.endif
+.if \pri
+.if \bpc == 16
+        clz             r9,  r9
+        sub             r9,  r9,  #24        // -bitdepth_min_8
+        neg             r9,  r9              // bitdepth_min_8
+.endif
+        movrel_local    r8,  pri_taps
+.if \bpc == 16
+        lsr             r9,  r3,  r9         // pri_strength >> bitdepth_min_8
+        and             r9,  r9,  #1         // (pri_strength >> bitdepth_min_8) & 1
+.else
+        and             r9,  r3,  #1
+.endif
+        add             r8,  r8,  r9, lsl #1
+.endif
+        movrel_local    r9,  directions\w
+        add             r5,  r9,  r5, lsl #1
+        vmov.u16        d17, #15
+        vdup.16         d16, r6              // damping
+
+.if \pri
+        vdup.16         q5,  r3              // threshold
+.endif
+.if \sec
+        vdup.16         q7,  r4              // threshold
+.endif
+        vmov.16         d8[0], r3
+        vmov.16         d8[1], r4
+        vclz.i16        d8,  d8              // clz(threshold)
+        vsub.i16        d8,  d17, d8         // ulog2(threshold)
+        vqsub.u16       d8,  d16, d8         // shift = imax(0, damping - ulog2(threshold))
+        vneg.s16        d8,  d8              // -shift
+.if \sec
+        vdup.16         q6,  d8[1]
+.endif
+.if \pri
+        vdup.16         q4,  d8[0]
+.endif
+
+1:
+.if \w == 8
+        vld1.16         {q0},  [r2, :128]    // px
+.else
+        add             r12, r2,  #2*8
+        vld1.16         {d0},  [r2,  :64]    // px
+        vld1.16         {d1},  [r12, :64]    // px
+.endif
+
+        vmov.u16        q1,  #0              // sum
+.if \min
+        vmov.u16        q2,  q0              // min
+        vmov.u16        q3,  q0              // max
+.endif
+
+        // Instead of loading sec_taps 2, 1 from memory, just set it
+        // to 2 initially and decrease for the second round.
+        // This is also used as loop counter.
+        mov             lr,  #2              // sec_taps[0]
+
+2:
+.if \pri
+        ldrsb           r9,  [r5]            // off1
+
+        load_px         d28, d29, d30, d31, \w
+.endif
+
+.if \sec
+        add             r5,  r5,  #4         // +2*2
+        ldrsb           r9,  [r5]            // off2
+.endif
+
+.if \pri
+        ldrb            r12, [r8]            // *pri_taps
+
+        handle_pixel    q14, q15, q5,  q4,  r12, \min
+.endif
+
+.if \sec
+        load_px         d28, d29, d30, d31, \w
+
+        add             r5,  r5,  #8         // +2*4
+        ldrsb           r9,  [r5]            // off3
+
+        handle_pixel    q14, q15, q7,  q6,  lr, \min
+
+        load_px         d28, d29, d30, d31, \w
+
+        handle_pixel    q14, q15, q7,  q6,  lr, \min
+
+        sub             r5,  r5,  #11        // r5 -= 2*(2+4); r5 += 1;
+.else
+        add             r5,  r5,  #1         // r5 += 1
+.endif
+        subs            lr,  lr,  #1         // sec_tap-- (value)
+.if \pri
+        add             r8,  r8,  #1         // pri_taps++ (pointer)
+.endif
+        bne             2b
+
+        vshr.s16        q14, q1,  #15        // -(sum < 0)
+        vadd.i16        q1,  q1,  q14        // sum - (sum < 0)
+        vrshr.s16       q1,  q1,  #4         // (8 + sum - (sum < 0)) >> 4
+        vadd.i16        q0,  q0,  q1         // px + (8 + sum ...) >> 4
+.if \min
+        vmin.s16        q0,  q0,  q3
+        vmax.s16        q0,  q0,  q2         // iclip(px + .., min, max)
+.endif
+.if \bpc == 8
+        vmovn.u16       d0,  q0
+.endif
+.if \w == 8
+        add             r2,  r2,  #2*16      // tmp += tmp_stride
+        subs            r7,  r7,  #1         // h--
+.if \bpc == 8
+        vst1.8          {d0}, [r0, :64], r1
+.else
+        vst1.16         {q0}, [r0, :128], r1
+.endif
+.else
+.if \bpc == 8
+        vst1.32         {d0[0]}, [r0, :32], r1
+.else
+        vst1.16         {d0},    [r0, :64], r1
+.endif
+        add             r2,  r2,  #2*16      // tmp += 2*tmp_stride
+        subs            r7,  r7,  #2         // h -= 2
+.if \bpc == 8
+        vst1.32         {d0[1]}, [r0, :32], r1
+.else
+        vst1.16         {d1},    [r0, :64], r1
+.endif
+.endif
+
+        // Reset pri_taps and directions back to the original point
+        sub             r5,  r5,  #2
+.if \pri
+        sub             r8,  r8,  #2
+.endif
+
+        bgt             1b
+        vpop            {q4-q7}
+        pop             {r4-r9,pc}
+endfunc
+.endm
+
+.macro filter w, bpc
+filter_func \w, \bpc, pri=1, sec=0, min=0, suffix=_pri
+filter_func \w, \bpc, pri=0, sec=1, min=0, suffix=_sec
+filter_func \w, \bpc, pri=1, sec=1, min=1, suffix=_pri_sec
+
+function cdef_filter\w\()_\bpc\()bpc_neon, export=1
+        push            {r4-r9,lr}
+        vpush           {q4-q7}
+        ldrd            r4,  r5,  [sp, #92]
+        ldrd            r6,  r7,  [sp, #100]
+.if \bpc == 16
+        ldrd            r8,  r9,  [sp, #108]
+.else
+        ldr             r8,  [sp, #108]
+.endif
+        cmp             r3,  #0 // pri_strength
+        bne             1f
+        b               cdef_filter\w\()_sec_\bpc\()bpc_neon // only sec
+1:
+        cmp             r4,  #0 // sec_strength
+        bne             1f
+        b               cdef_filter\w\()_pri_\bpc\()bpc_neon // only pri
+1:
+        b               cdef_filter\w\()_pri_sec_\bpc\()bpc_neon // both pri and sec
+endfunc
+.endm
+
+const div_table, align=4
+        .short         840, 420, 280, 210, 168, 140, 120, 105
+endconst
+
+const alt_fact, align=4
+        .short         420, 210, 140, 105, 105, 105, 105, 105, 140, 210, 420, 0
+endconst
+
+.macro cost_alt dest, s1, s2, s3, s4, s5, s6
+        vmull.s16       q1,  \s1, \s1          // sum_alt[n]*sum_alt[n]
+        vmull.s16       q2,  \s2, \s2
+        vmull.s16       q3,  \s3, \s3
+        vmull.s16       q5,  \s4, \s4          // sum_alt[n]*sum_alt[n]
+        vmull.s16       q12, \s5, \s5
+        vmull.s16       q6,  \s6, \s6          // q6 overlaps the first \s1-\s2 here
+        vmul.i32        q1,  q1,  q13          // sum_alt[n]^2*fact
+        vmla.i32        q1,  q2,  q14
+        vmla.i32        q1,  q3,  q15
+        vmul.i32        q5,  q5,  q13          // sum_alt[n]^2*fact
+        vmla.i32        q5,  q12, q14
+        vmla.i32        q5,  q6,  q15
+        vadd.i32        d2,  d2,  d3
+        vadd.i32        d3,  d10, d11
+        vpadd.i32       \dest, d2, d3          // *cost_ptr
+.endm
+
+.macro find_best s1, s2, s3
+.ifnb \s2
+        vmov.32         lr,  \s2
+.endif
+        cmp             r12, r1                // cost[n] > best_cost
+        itt             gt
+        movgt           r0,  r3                // best_dir = n
+        movgt           r1,  r12               // best_cost = cost[n]
+.ifnb \s2
+        add             r3,  r3,  #1           // n++
+        cmp             lr,  r1                // cost[n] > best_cost
+        vmov.32         r12, \s3
+        itt             gt
+        movgt           r0,  r3                // best_dir = n
+        movgt           r1,  lr                // best_cost = cost[n]
+        add             r3,  r3,  #1           // n++
+.endif
+.endm
+
+// int dav1d_cdef_find_dir_Xbpc_neon(const pixel *img, const ptrdiff_t stride,
+//                                   unsigned *const var)
+.macro find_dir bpc
+function cdef_find_dir_\bpc\()bpc_neon, export=1
+        push            {lr}
+        vpush           {q4-q7}
+.if \bpc == 16
+        clz             r3,  r3                // clz(bitdepth_max)
+        sub             lr,  r3,  #24          // -bitdepth_min_8
+.endif
+        sub             sp,  sp,  #32          // cost
+        mov             r3,  #8
+        vmov.u16        q1,  #0                // q0-q1   sum_diag[0]
+        vmov.u16        q3,  #0                // q2-q3   sum_diag[1]
+        vmov.u16        q5,  #0                // q4-q5   sum_hv[0-1]
+        vmov.u16        q8,  #0                // q6,d16  sum_alt[0]
+                                               // q7,d17  sum_alt[1]
+        vmov.u16        q9,  #0                // q9,d22  sum_alt[2]
+        vmov.u16        q11, #0
+        vmov.u16        q10, #0                // q10,d23 sum_alt[3]
+
+
+.irpc i, 01234567
+.if \bpc == 8
+        vld1.8          {d30}, [r0, :64], r1
+        vmov.u8         d31, #128
+        vsubl.u8        q15, d30, d31          // img[x] - 128
+.else
+        vld1.16         {q15}, [r0, :128], r1
+        vdup.16         q14, lr                // -bitdepth_min_8
+        vshl.u16        q15, q15, q14
+        vmov.u16        q14, #128
+        vsub.i16        q15, q15, q14          // img[x] - 128
+.endif
+        vmov.u16        q14, #0
+
+.if \i == 0
+        vmov            q0,  q15               // sum_diag[0]
+.else
+        vext.8          q12, q14, q15, #(16-2*\i)
+        vext.8          q13, q15, q14, #(16-2*\i)
+        vadd.i16        q0,  q0,  q12          // sum_diag[0]
+        vadd.i16        q1,  q1,  q13          // sum_diag[0]
+.endif
+        vrev64.16       q13, q15
+        vswp            d26, d27               // [-x]
+.if \i == 0
+        vmov            q2,  q13               // sum_diag[1]
+.else
+        vext.8          q12, q14, q13, #(16-2*\i)
+        vext.8          q13, q13, q14, #(16-2*\i)
+        vadd.i16        q2,  q2,  q12          // sum_diag[1]
+        vadd.i16        q3,  q3,  q13          // sum_diag[1]
+.endif
+
+        vpadd.u16       d26, d30, d31          // [(x >> 1)]
+        vmov.u16        d27, #0
+        vpadd.u16       d24, d26, d28
+        vpadd.u16       d24, d24, d28          // [y]
+        vmov.u16        r12, d24[0]
+        vadd.i16        q5,  q5,  q15          // sum_hv[1]
+.if \i < 4
+        vmov.16         d8[\i],   r12          // sum_hv[0]
+.else
+        vmov.16         d9[\i-4], r12          // sum_hv[0]
+.endif
+
+.if \i == 0
+        vmov.u16        q6,  q13               // sum_alt[0]
+.else
+        vext.8          q12, q14, q13, #(16-2*\i)
+        vext.8          q14, q13, q14, #(16-2*\i)
+        vadd.i16        q6,  q6,  q12          // sum_alt[0]
+        vadd.i16        d16, d16, d28          // sum_alt[0]
+.endif
+        vrev64.16       d26, d26               // [-(x >> 1)]
+        vmov.u16        q14, #0
+.if \i == 0
+        vmov            q7,  q13               // sum_alt[1]
+.else
+        vext.8          q12, q14, q13, #(16-2*\i)
+        vext.8          q13, q13, q14, #(16-2*\i)
+        vadd.i16        q7,  q7,  q12          // sum_alt[1]
+        vadd.i16        d17, d17, d26          // sum_alt[1]
+.endif
+
+.if \i < 6
+        vext.8          q12, q14, q15, #(16-2*(3-(\i/2)))
+        vext.8          q13, q15, q14, #(16-2*(3-(\i/2)))
+        vadd.i16        q9,  q9,  q12          // sum_alt[2]
+        vadd.i16        d22, d22, d26          // sum_alt[2]
+.else
+        vadd.i16        q9,  q9,  q15          // sum_alt[2]
+.endif
+.if \i == 0
+        vmov            q10, q15               // sum_alt[3]
+.elseif \i == 1
+        vadd.i16        q10, q10, q15          // sum_alt[3]
+.else
+        vext.8          q12, q14, q15, #(16-2*(\i/2))
+        vext.8          q13, q15, q14, #(16-2*(\i/2))
+        vadd.i16        q10, q10, q12          // sum_alt[3]
+        vadd.i16        d23, d23, d26          // sum_alt[3]
+.endif
+.endr
+
+        vmov.u32        q15, #105
+
+        vmull.s16       q12, d8,  d8           // sum_hv[0]*sum_hv[0]
+        vmlal.s16       q12, d9,  d9
+        vmull.s16       q13, d10, d10          // sum_hv[1]*sum_hv[1]
+        vmlal.s16       q13, d11, d11
+        vadd.s32        d8,  d24, d25
+        vadd.s32        d9,  d26, d27
+        vpadd.s32       d8,  d8,  d9           // cost[2,6] (s16, s17)
+        vmul.i32        d8,  d8,  d30          // cost[2,6] *= 105
+
+        vrev64.16       q1,  q1
+        vrev64.16       q3,  q3
+        vext.8          q1,  q1,  q1,  #10     // sum_diag[0][14-n]
+        vext.8          q3,  q3,  q3,  #10     // sum_diag[1][14-n]
+
+        vstr            s16, [sp, #2*4]        // cost[2]
+        vstr            s17, [sp, #6*4]        // cost[6]
+
+        movrel_local    r12, div_table
+        vld1.16         {q14}, [r12, :128]
+
+        vmull.s16       q5,  d0,  d0           // sum_diag[0]*sum_diag[0]
+        vmull.s16       q12, d1,  d1
+        vmlal.s16       q5,  d2,  d2
+        vmlal.s16       q12, d3,  d3
+        vmull.s16       q0,  d4,  d4           // sum_diag[1]*sum_diag[1]
+        vmull.s16       q1,  d5,  d5
+        vmlal.s16       q0,  d6,  d6
+        vmlal.s16       q1,  d7,  d7
+        vmovl.u16       q13, d28               // div_table
+        vmovl.u16       q14, d29
+        vmul.i32        q5,  q5,  q13          // cost[0]
+        vmla.i32        q5,  q12, q14
+        vmul.i32        q0,  q0,  q13          // cost[4]
+        vmla.i32        q0,  q1,  q14
+        vadd.i32        d10, d10, d11
+        vadd.i32        d0,  d0,  d1
+        vpadd.i32       d0,  d10, d0           // cost[0,4] = s0,s1
+
+        movrel_local    r12, alt_fact
+        vld1.16         {d29, d30, d31}, [r12, :64] // div_table[2*m+1] + 105
+
+        vstr            s0,  [sp, #0*4]        // cost[0]
+        vstr            s1,  [sp, #4*4]        // cost[4]
+
+        vmovl.u16       q13, d29               // div_table[2*m+1] + 105
+        vmovl.u16       q14, d30
+        vmovl.u16       q15, d31
+
+        cost_alt        d14, d12, d13, d16, d14, d15, d17 // cost[1], cost[3]
+        cost_alt        d15, d18, d19, d22, d20, d21, d23 // cost[5], cost[7]
+        vstr            s28, [sp, #1*4]        // cost[1]
+        vstr            s29, [sp, #3*4]        // cost[3]
+
+        mov             r0,  #0                // best_dir
+        vmov.32         r1,  d0[0]             // best_cost
+        mov             r3,  #1                // n
+
+        vstr            s30, [sp, #5*4]        // cost[5]
+        vstr            s31, [sp, #7*4]        // cost[7]
+
+        vmov.32         r12, d14[0]
+
+        find_best       d14[0], d8[0], d14[1]
+        find_best       d14[1], d0[1], d15[0]
+        find_best       d15[0], d8[1], d15[1]
+        find_best       d15[1]
+
+        eor             r3,  r0,  #4           // best_dir ^4
+        ldr             r12, [sp, r3, lsl #2]
+        sub             r1,  r1,  r12          // best_cost - cost[best_dir ^ 4]
+        lsr             r1,  r1,  #10
+        str             r1,  [r2]              // *var
+
+        add             sp,  sp,  #32
+        vpop            {q4-q7}
+        pop             {pc}
+endfunc
+.endm
--- a/src/arm/cdef_init_tmpl.c
+++ b/src/arm/cdef_init_tmpl.c
@@ -27,7 +27,6 @@
 #include "src/cpu.h"
 #include "src/cdef.h"
 
-#if BITDEPTH == 8 || ARCH_AARCH64
 decl_cdef_dir_fn(BF(dav1d_cdef_find_dir, neon));
 
 void BF(dav1d_cdef_padding4, neon)(uint16_t *tmp, const pixel *src,
@@ -72,7 +71,6 @@
 DEFINE_FILTER(8, 8, 16)
 DEFINE_FILTER(4, 8, 8)
 DEFINE_FILTER(4, 4, 8)
-#endif
 
 
 COLD void bitfn(dav1d_cdef_dsp_init_arm)(Dav1dCdefDSPContext *const c) {
@@ -80,10 +78,8 @@
 
     if (!(flags & DAV1D_ARM_CPU_FLAG_NEON)) return;
 
-#if BITDEPTH == 8 || ARCH_AARCH64
     c->dir = BF(dav1d_cdef_find_dir, neon);
     c->fb[0] = cdef_filter_8x8_neon;
     c->fb[1] = cdef_filter_4x8_neon;
     c->fb[2] = cdef_filter_4x4_neon;
-#endif
 }
--- a/src/meson.build
+++ b/src/meson.build
@@ -147,6 +147,7 @@
 
             if dav1d_bitdepths.contains('16')
                 libdav1d_sources_asm += files(
+                    'arm/32/cdef16.S',
                     'arm/32/looprestoration16.S',
                     'arm/32/mc16.S',
                 )