aarch64: vp9: Implement NEON loop filters
authorMartin Storsjö <martin@martin.st>
Sun, 13 Nov 2016 22:07:27 +0000 (00:07 +0200)
committerMartin Storsjö <martin@martin.st>
Sun, 13 Nov 2016 22:10:13 +0000 (00:10 +0200)
This work is sponsored by, and copyright, Google.

These are ported from the ARM version; thanks to the larger
amount of registers available, we can do the loop filters with
16 pixels at a time. The implementation is fully templated, with
a single macro which can generate versions for both 8 and
16 pixels wide, for both 4, 8 and 16 pixels loop filters
(and the 4/8 mixed versions as well).

For the 8 pixel wide versions, it is pretty close in speed (the
v_4_8 and v_8_8 filters are the best examples of this; the h_4_8
and h_8_8 filters seem to get some gain in the load/transpose/store
part). For the 16 pixels wide ones, we get a speedup of around
1.2-1.4x compared to the 32 bit version.

Examples of runtimes vs the 32 bit version, on a Cortex A53:
                                       ARM AArch64
vp9_loop_filter_h_4_8_neon:          144.0   127.2
vp9_loop_filter_h_8_8_neon:          207.0   182.5
vp9_loop_filter_h_16_8_neon:         415.0   328.7
vp9_loop_filter_h_16_16_neon:        672.0   558.6
vp9_loop_filter_mix2_h_44_16_neon:   302.0   203.5
vp9_loop_filter_mix2_h_48_16_neon:   365.0   305.2
vp9_loop_filter_mix2_h_84_16_neon:   365.0   305.2
vp9_loop_filter_mix2_h_88_16_neon:   376.0   305.2
vp9_loop_filter_mix2_v_44_16_neon:   193.2   128.2
vp9_loop_filter_mix2_v_48_16_neon:   246.7   218.4
vp9_loop_filter_mix2_v_84_16_neon:   248.0   218.5
vp9_loop_filter_mix2_v_88_16_neon:   302.0   218.2
vp9_loop_filter_v_4_8_neon:           89.0    88.7
vp9_loop_filter_v_8_8_neon:          141.0   137.7
vp9_loop_filter_v_16_8_neon:         295.0   272.7
vp9_loop_filter_v_16_16_neon:        546.0   453.7

The speedup vs C code in checkasm tests is around 2-7x, which is
pretty much the same as for the 32 bit version. Even if these functions
are faster than their 32 bit equivalent, the C version that we compare
to also became around 1.3-1.7x faster than the C version in 32 bit.

Based on START_TIMER/STOP_TIMER wrapping around a few individual
functions, the speedup vs C code is around 4-5x.

Examples of runtimes vs C on a Cortex A57 (for a slightly older version
of the patch):
                         A57 gcc-5.3  neon
loop_filter_h_4_8_neon:        256.6  93.4
loop_filter_h_8_8_neon:        307.3 139.1
loop_filter_h_16_8_neon:       340.1 254.1
loop_filter_h_16_16_neon:      827.0 407.9
loop_filter_mix2_h_44_16_neon: 524.5 155.4
loop_filter_mix2_h_48_16_neon: 644.5 173.3
loop_filter_mix2_h_84_16_neon: 630.5 222.0
loop_filter_mix2_h_88_16_neon: 697.3 222.0
loop_filter_mix2_v_44_16_neon: 598.5 100.6
loop_filter_mix2_v_48_16_neon: 651.5 127.0
loop_filter_mix2_v_84_16_neon: 591.5 167.1
loop_filter_mix2_v_88_16_neon: 855.1 166.7
loop_filter_v_4_8_neon:        271.7  65.3
loop_filter_v_8_8_neon:        312.5 106.9
loop_filter_v_16_8_neon:       473.3 206.5
loop_filter_v_16_16_neon:      976.1 327.8

The speed-up compared to the C functions is 2.5 to 6 and the cortex-a57
is again 30-50% faster than the cortex-a53.

Signed-off-by: Martin Storsjö <martin@martin.st>
libavcodec/aarch64/Makefile
libavcodec/aarch64/vp9dsp_init_aarch64.c
libavcodec/aarch64/vp9lpf_neon.S [new file with mode: 0644]

index 0b06b33..5c1d118 100644 (file)
@@ -45,4 +45,5 @@ NEON-OBJS-$(CONFIG_DCA_DECODER)         += aarch64/dcadsp_neon.o               \
                                            aarch64/synth_filter_neon.o
 NEON-OBJS-$(CONFIG_VORBIS_DECODER)      += aarch64/vorbisdsp_neon.o
 NEON-OBJS-$(CONFIG_VP9_DECODER)         += aarch64/vp9itxfm_neon.o             \
+                                           aarch64/vp9lpf_neon.o               \
                                            aarch64/vp9mc_neon.o
index da6f7f9..486ae1e 100644 (file)
@@ -195,8 +195,53 @@ static av_cold void vp9dsp_itxfm_init_aarch64(VP9DSPContext *dsp)
     }
 }
 
+#define define_loop_filter(dir, wd, len) \
+void ff_vp9_loop_filter_##dir##_##wd##_##len##_neon(uint8_t *dst, ptrdiff_t stride, int E, int I, int H)
+
+#define define_loop_filters(wd, len) \
+    define_loop_filter(h, wd, len);  \
+    define_loop_filter(v, wd, len)
+
+define_loop_filters(4, 8);
+define_loop_filters(8, 8);
+define_loop_filters(16, 8);
+
+define_loop_filters(16, 16);
+
+define_loop_filters(44, 16);
+define_loop_filters(48, 16);
+define_loop_filters(84, 16);
+define_loop_filters(88, 16);
+
+static av_cold void vp9dsp_loopfilter_init_aarch64(VP9DSPContext *dsp)
+{
+    int cpu_flags = av_get_cpu_flags();
+
+    if (have_neon(cpu_flags)) {
+        dsp->loop_filter_8[0][1] = ff_vp9_loop_filter_v_4_8_neon;
+        dsp->loop_filter_8[0][0] = ff_vp9_loop_filter_h_4_8_neon;
+        dsp->loop_filter_8[1][1] = ff_vp9_loop_filter_v_8_8_neon;
+        dsp->loop_filter_8[1][0] = ff_vp9_loop_filter_h_8_8_neon;
+        dsp->loop_filter_8[2][1] = ff_vp9_loop_filter_v_16_8_neon;
+        dsp->loop_filter_8[2][0] = ff_vp9_loop_filter_h_16_8_neon;
+
+        dsp->loop_filter_16[0] = ff_vp9_loop_filter_h_16_16_neon;
+        dsp->loop_filter_16[1] = ff_vp9_loop_filter_v_16_16_neon;
+
+        dsp->loop_filter_mix2[0][0][0] = ff_vp9_loop_filter_h_44_16_neon;
+        dsp->loop_filter_mix2[0][0][1] = ff_vp9_loop_filter_v_44_16_neon;
+        dsp->loop_filter_mix2[0][1][0] = ff_vp9_loop_filter_h_48_16_neon;
+        dsp->loop_filter_mix2[0][1][1] = ff_vp9_loop_filter_v_48_16_neon;
+        dsp->loop_filter_mix2[1][0][0] = ff_vp9_loop_filter_h_84_16_neon;
+        dsp->loop_filter_mix2[1][0][1] = ff_vp9_loop_filter_v_84_16_neon;
+        dsp->loop_filter_mix2[1][1][0] = ff_vp9_loop_filter_h_88_16_neon;
+        dsp->loop_filter_mix2[1][1][1] = ff_vp9_loop_filter_v_88_16_neon;
+    }
+}
+
 av_cold void ff_vp9dsp_init_aarch64(VP9DSPContext *dsp)
 {
     vp9dsp_mc_init_aarch64(dsp);
+    vp9dsp_loopfilter_init_aarch64(dsp);
     vp9dsp_itxfm_init_aarch64(dsp);
 }
diff --git a/libavcodec/aarch64/vp9lpf_neon.S b/libavcodec/aarch64/vp9lpf_neon.S
new file mode 100644 (file)
index 0000000..995a97d
--- /dev/null
@@ -0,0 +1,1355 @@
+/*
+ * Copyright (c) 2016 Google Inc.
+ *
+ * This file is part of Libav.
+ *
+ * Libav is free software; you can redistribute it and/or
+ * modify it under the terms of the GNU Lesser General Public
+ * License as published by the Free Software Foundation; either
+ * version 2.1 of the License, or (at your option) any later version.
+ *
+ * Libav is distributed in the hope that it will be useful,
+ * but WITHOUT ANY WARRANTY; without even the implied warranty of
+ * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+ * Lesser General Public License for more details.
+ *
+ * You should have received a copy of the GNU Lesser General Public
+ * License along with Libav; if not, write to the Free Software
+ * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
+ */
+
+#include "libavutil/aarch64/asm.S"
+#include "neon.S"
+
+
+// The main loop filter macro is templated and can produce filters for
+// vectors of 8 or 16 bytes. The register mapping throughout the filter
+// is close to identical to the arm version (please try to maintain this,
+// if either is changed!). When the arm version uses e.g. d20 for the
+// input variable p3, the aarch64 version uses v20.8b or v20.16b, depending
+// on vector length.
+//
+// The number of elements in the vector is passed in via the macro parameter
+// \sz, which is either .8b or .16b. For simple instructions that doesn't
+// lengthen or narrow things, this can easily be templated like this:
+//      uabd            v4\sz,  v20\sz, v21\sz
+//
+// For instructions that lengthen or narrow content, the arm version would
+// have used q registers. For these instructions, we have macros that expand
+// into either a single e.g. uaddl instruction, or into a uaddl + uaddl2
+// pair, depending on the \sz parameter. Wherever the arm version would have
+// used a q register, these macros instead take two v registers, i.e. q3
+// is mapped to v6+v7. For the case with 8 byte input vectors, such a
+// lengthening operation is only stored in v6.8h (what was in q3 in the arm
+// case), while the 16 byte input vectors will use v6.8h + v7.8h.
+// Such a macro invocation would look like this:
+//      uaddl_sz        v8.8h,  v9.8h,  v17, v18, \sz
+//
+// That is, in the 8 byte input vector case, the second register in these
+// register pairs will be unused.
+// Unfortunately, this makes the code quite hard to read. For readability,
+// see the arm version instead.
+
+
+.macro uabdl_sz dst1, dst2, in1, in2, sz
+        uabdl           \dst1,  \in1\().8b,  \in2\().8b
+.ifc \sz, .16b
+        uabdl2          \dst2,  \in1\().16b, \in2\().16b
+.endif
+.endm
+
+.macro add_sz dst1, dst2, in1, in2, in3, in4, sz
+        add             \dst1,  \in1,  \in3
+.ifc \sz, .16b
+        add             \dst2,  \in2,  \in4
+.endif
+.endm
+
+.macro sub_sz dst1, dst2, in1, in2, in3, in4, sz
+        sub             \dst1,  \in1,  \in3
+.ifc \sz, .16b
+        sub             \dst2,  \in2,  \in4
+.endif
+.endm
+
+.macro uaddw_sz dst1, dst2, in1, in2, in3, sz
+        uaddw           \dst1,  \in1, \in3\().8b
+.ifc \sz, .16b
+        uaddw2          \dst2,  \in2, \in3\().16b
+.endif
+.endm
+
+.macro usubw_sz dst1, dst2, in1, in2, in3, sz
+        usubw           \dst1,  \in1, \in3\().8b
+.ifc \sz, .16b
+        usubw2          \dst2,  \in2, \in3\().16b
+.endif
+.endm
+
+.macro cmhs_sz dst1, dst2, in1, in2, in3, in4, sz
+        cmhs            \dst1,  \in1,  \in3
+.ifc \sz, .16b
+        cmhs            \dst2,  \in2,  \in4
+.endif
+.endm
+
+.macro xtn_sz dst, in1, in2, sz
+        xtn             \dst\().8b,  \in1
+.ifc \sz, .16b
+        xtn2            \dst\().16b, \in2
+.endif
+.endm
+
+.macro usubl_sz dst1, dst2, in1, in2, sz
+        usubl           \dst1,  \in1\().8b,  \in2\().8b
+.ifc \sz, .16b
+        usubl2          \dst2,  \in1\().16b, \in2\().16b
+.endif
+.endm
+
+.macro sqxtn_sz dst, in1, in2, sz
+        sqxtn           \dst\().8b,  \in1
+.ifc \sz, .16b
+        sqxtn2          \dst\().16b, \in2
+.endif
+.endm
+
+.macro sqxtun_sz dst, in1, in2, sz
+        sqxtun          \dst\().8b,  \in1
+.ifc \sz, .16b
+        sqxtun2         \dst\().16b, \in2
+.endif
+.endm
+
+.macro mul_sz dst1, dst2, in1, in2, in3, in4, sz
+        mul             \dst1,  \in1,  \in3
+.ifc \sz, .16b
+        mul             \dst2,  \in2,  \in4
+.endif
+.endm
+
+.macro saddw_sz dst1, dst2, in1, in2, in3, sz
+        saddw           \dst1,  \in1, \in3\().8b
+.ifc \sz, .16b
+        saddw2          \dst2,  \in2, \in3\().16b
+.endif
+.endm
+
+.macro ssubw_sz dst1, dst2, in1, in2, in3, sz
+        ssubw           \dst1,  \in1, \in3\().8b
+.ifc \sz, .16b
+        ssubw2          \dst2,  \in2, \in3\().16b
+.endif
+.endm
+
+.macro uxtl_sz dst1, dst2, in, sz
+        uxtl            \dst1,  \in\().8b
+.ifc \sz, .16b
+        uxtl2           \dst2,  \in\().16b
+.endif
+.endm
+
+.macro uaddl_sz dst1, dst2, in1, in2, sz
+        uaddl           \dst1,  \in1\().8b,  \in2\().8b
+.ifc \sz, .16b
+        uaddl2          \dst2,  \in1\().16b, \in2\().16b
+.endif
+.endm
+
+.macro rshrn_sz dst, in1, in2, shift, sz
+        rshrn           \dst\().8b,  \in1, \shift
+.ifc \sz, .16b
+        rshrn2          \dst\().16b, \in2, \shift
+.endif
+.endm
+
+.macro ushll_sz dst1, dst2, in, shift, sz
+        ushll           \dst1,  \in\().8b,  \shift
+.ifc \sz, .16b
+        ushll2          \dst2,  \in\().16b, \shift
+.endif
+.endm
+
+// The input to and output from this macro is in the registers v16-v31,
+// and v0-v7 are used as scratch registers.
+// p7 = v16 .. p3 = v20, p0 = v23, q0 = v24, q3 = v27, q7 = v31
+// Depending on the width of the loop filter, we either use v16-v19
+// and v28-v31 as temp registers, or v8-v15.
+// When comparing to the arm version, tmpq1 == tmp1 + tmp2,
+// tmpq2 == tmp3 + tmp4, etc.
+.macro loop_filter wd, sz, mix, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7, tmp8
+.if \mix == 0
+        dup             v0.8h,  w2        // E
+        dup             v1.8h,  w2        // E
+        dup             v2\sz,  w3        // I
+        dup             v3\sz,  w4        // H
+.else
+        dup             v0.8h,  w2        // E
+        dup             v2.8b,  w3        // I
+        dup             v3.8b,  w4        // H
+        lsr             w6,     w3,  #8
+        lsr             w7,     w4,  #8
+        ushr            v1.8h,  v0.8h, #8 // E
+        dup             v4.8b,  w6        // I
+        bic             v0.8h,  #255, lsl 8 // E
+        dup             v5.8b,  w7        // H
+        trn1            v2.2d,  v2.2d,  v4.2d
+        trn1            v3.2d,  v3.2d,  v5.2d
+.endif
+
+        uabd            v4\sz,  v20\sz, v21\sz        // abs(p3 - p2)
+        uabd            v5\sz,  v21\sz, v22\sz        // abs(p2 - p1)
+        uabd            v6\sz,  v22\sz, v23\sz        // abs(p1 - p0)
+        uabd            v7\sz,  v24\sz, v25\sz        // abs(q0 - q1)
+        uabd            \tmp1\sz,  v25\sz, v26\sz     // abs(q1 - q2)
+        uabd            \tmp2\sz,  v26\sz, v27\sz     // abs(q2 - q3)
+        umax            v4\sz,  v4\sz,  v5\sz
+        umax            v5\sz,  v6\sz,  v7\sz
+        umax            \tmp1\sz, \tmp1\sz, \tmp2\sz
+        uabdl_sz        v6.8h,  v7.8h,  v23, v24, \sz // abs(p0 - q0)
+        umax            v4\sz,  v4\sz,  v5\sz
+        add_sz          v6.8h,  v7.8h,  v6.8h,  v7.8h,  v6.8h,  v7.8h, \sz // abs(p0 - q0) * 2
+        uabd            v5\sz,  v22\sz, v25\sz        // abs(p1 - q1)
+        umax            v4\sz,  v4\sz,  \tmp1\sz      // max(abs(p3 - p2), ..., abs(q2 - q3))
+        ushr            v5\sz,  v5\sz,  #1
+        cmhs            v4\sz,  v2\sz,  v4\sz         // max(abs()) <= I
+        uaddw_sz        v6.8h,  v7.8h,  v6.8h,  v7.8h,  v5, \sz // abs(p0 - q0) * 2 + abs(p1 - q1) >> 1
+        cmhs_sz         v6.8h,  v7.8h,  v0.8h,  v1.8h,  v6.8h,  v7.8h, \sz
+        xtn_sz          v5,     v6.8h,  v7.8h,  \sz
+        and             v4\sz,  v4\sz,  v5\sz         // fm
+
+        mov             x5,  v4.d[0]
+.ifc \sz, .16b
+        mov             x6,  v4.d[1]
+        orr             x5,  x5,  x6
+.endif
+        // If no pixels need filtering, just exit as soon as possible
+        cbz             x5,  9f
+
+.if \wd >= 8
+        movi            v0\sz,  #1
+
+        uabd            v6\sz,  v20\sz, v23\sz    // abs(p3 - p0)
+        uabd            v2\sz,  v21\sz, v23\sz    // abs(p2 - p0)
+        uabd            v1\sz,  v22\sz, v23\sz    // abs(p1 - p0)
+        uabd            \tmp1\sz,  v25\sz, v24\sz // abs(q1 - q0)
+        uabd            \tmp2\sz,  v26\sz, v24\sz // abs(q2 - q0)
+        uabd            \tmp3\sz,  v27\sz, v24\sz // abs(q3 - q0)
+        umax            v6\sz,  v6\sz,  v2\sz
+        umax            v1\sz,  v1\sz,  \tmp1\sz
+        umax            \tmp2\sz,  \tmp2\sz,  \tmp3\sz
+.if \wd == 16
+        uabd            v7\sz,  v16\sz, v23\sz    // abs(p7 - p0)
+        umax            v6\sz,  v6\sz,  v1\sz
+        uabd            v2\sz,  v17\sz, v23\sz    // abs(p6 - p0)
+        umax            v6\sz,  v6\sz,  \tmp2\sz
+        uabd            v1\sz,  v18\sz, v23\sz    // abs(p5 - p0)
+        cmhs            v6\sz,  v0\sz,  v6\sz     // flat8in
+        uabd            v8\sz,  v19\sz, v23\sz    // abs(p4 - p0)
+        and             v6\sz,  v6\sz,  v4\sz     // flat8in && fm
+        uabd            v9\sz,  v28\sz, v24\sz    // abs(q4 - q0)
+        bic             v4\sz,  v4\sz,  v6\sz     // fm && !flat8in
+        uabd            v10\sz, v29\sz, v24\sz    // abs(q5 - q0)
+        uabd            v11\sz, v30\sz, v24\sz    // abs(q6 - q0)
+        uabd            v12\sz, v31\sz, v24\sz    // abs(q7 - q0)
+
+        umax            v7\sz,  v7\sz,  v2\sz
+        umax            v1\sz,  v1\sz,  v8\sz
+        umax            v9\sz,  v9\sz,  v10\sz
+        umax            v11\sz, v11\sz, v12\sz
+        // The rest of the calculation of flat8out is interleaved below
+.else
+        // The rest of the calculation of flat8in is interleaved below
+.endif
+.endif
+
+        // Calculate the normal inner loop filter for 2 or 4 pixels
+        uabd            v5\sz,  v22\sz, v23\sz // abs(p1 - p0)
+.if \wd == 16
+        umax            v7\sz,  v7\sz,  v1\sz
+        umax            v9\sz,  v9\sz,  v11\sz
+.elseif \wd == 8
+        umax            v6\sz,  v6\sz,  v1\sz
+.endif
+        uabd            v1\sz,  v25\sz, v24\sz // abs(q1 - q0)
+.if \wd == 16
+        umax            v7\sz,  v7\sz,  v9\sz
+.elseif \wd == 8
+        umax            v6\sz,  v6\sz,  \tmp2\sz
+.endif
+        usubl_sz        \tmp1\().8h,  \tmp2\().8h,  v22,  v25, \sz // p1 - q1
+        umax            v5\sz,  v5\sz,  v1\sz  // max(abs(p1 - p0), abs(q1 - q0))
+.if \mix != 0
+        mov             v1.d[0], x11
+.endif
+        usubl_sz        \tmp3\().8h,  \tmp4\().8h,  v24,  v23, \sz // q0 - p0
+        movi            \tmp5\().8h,  #3
+.if \wd == 8
+        cmhs            v6\sz,  v0\sz,  v6\sz  // flat8in
+.endif
+.if \mix != 0
+        sxtl            v1.8h,  v1.8b
+.endif
+        cmhi            v5\sz,  v5\sz,  v3\sz  // hev
+.if \wd == 8
+        // If a 4/8 or 8/4 mix is used, clear the relevant half of v6
+.if \mix != 0
+        and             v6\sz,  v6\sz,  v1.16b
+.endif
+        and             v6\sz,  v6\sz,  v4\sz  // flat8in && fm
+.endif
+        sqxtn_sz        \tmp1,        \tmp1\().8h,  \tmp2\().8h, \sz // av_clip_int8(p1 - q1)
+.if \wd == 16
+        cmhs            v7\sz,  v0\sz,  v7\sz  // flat8out
+.elseif \wd == 8
+        bic             v4\sz,  v4\sz,  v6\sz  // fm && !flat8in
+.endif
+        mvn             v5\sz,  v5\sz          // !hev
+.if \wd == 16
+        and             v7\sz,  v7\sz,  v6\sz  // flat8out && flat8in && fm
+.endif
+        and             v5\sz,  v5\sz,  v4\sz  // !hev && fm && !flat8in
+
+        mul_sz          \tmp3\().8h,  \tmp4\().8h,  \tmp3\().8h, \tmp4\().8h,  \tmp5\().8h,  \tmp5\().8h, \sz // 3 * (q0 - p0)
+        bic             \tmp1\sz,  \tmp1\sz,  v5\sz    // if (!hev) av_clip_int8 = 0
+        movi            v2\sz,  #4
+        saddw_sz        \tmp3\().8h,  \tmp4\().8h,  \tmp3\().8h, \tmp4\().8h,  \tmp1, \sz // 3 * (q0 - p0) [+ av_clip_int8(p1 - q1)]
+        movi            v3\sz,  #3
+        sqxtn_sz        \tmp1,        \tmp3\().8h,  \tmp4\().8h, \sz       // f
+.if \wd == 16
+        bic             v6\sz,  v6\sz,  v7\sz  // fm && flat8in && !flat8out
+.endif
+
+        sqadd           \tmp3\sz,  \tmp1\sz,  v2\sz // FFMIN(f + 4, 127)
+        sqadd           \tmp4\sz,  \tmp1\sz,  v3\sz // FFMIN(f + 3, 127)
+        uxtl_sz         v0.8h,  v1.8h,  v23, \sz    // p0
+        sshr            \tmp3\sz,  \tmp3\sz,  #3    // f1
+        sshr            \tmp4\sz,  \tmp4\sz,  #3    // f2
+
+        uxtl_sz         v2.8h,  v3.8h,  v24, \sz    // q0
+        saddw_sz        v0.8h,  v1.8h,  v0.8h,  v1.8h,  \tmp4, \sz // p0 + f2
+        ssubw_sz        v2.8h,  v3.8h,  v2.8h,  v3.8h,  \tmp3, \sz // q0 - f1
+        sqxtun_sz       v0,  v0.8h,  v1.8h,  \sz    // out p0
+        sqxtun_sz       v1,  v2.8h,  v3.8h,  \sz    // out q0
+        srshr           \tmp3\sz, \tmp3\sz, #1      // f = (f1 + 1) >> 1
+        bit             v23\sz, v0\sz,  v4\sz       // if (fm && !flat8in)
+        bit             v24\sz, v1\sz,  v4\sz
+
+        uxtl_sz         v0.8h,  v1.8h,  v22, \sz    // p1
+        uxtl_sz         v2.8h,  v3.8h,  v25, \sz    // q1
+        saddw_sz        v0.8h,  v1.8h,  v0.8h,  v1.8h,  \tmp3, \sz // p1 + f
+        ssubw_sz        v2.8h,  v3.8h,  v2.8h,  v3.8h,  \tmp3, \sz // q1 - f
+        sqxtun_sz       v0,  v0.8h,  v1.8h, \sz     // out p1
+        sqxtun_sz       v2,  v2.8h,  v3.8h, \sz     // out q1
+        bit             v22\sz, v0\sz,  v5\sz       // if (!hev && fm && !flat8in)
+        bit             v25\sz, v2\sz,  v5\sz
+
+.if \wd >= 8
+        mov             x5,  v6.d[0]
+.ifc \sz, .16b
+        mov             x6,  v6.d[1]
+        orr             x5,  x5,  x6
+.endif
+        // If no pixels need flat8in, jump to flat8out
+        // (or to a writeout of the inner 4 pixels, for wd=8)
+        cbz             x5,  6f
+
+        // flat8in
+        uaddl_sz        \tmp1\().8h, \tmp2\().8h,  v20, v21, \sz
+        uaddl_sz        \tmp3\().8h, \tmp4\().8h,  v22, v25, \sz
+        uaddl_sz        \tmp5\().8h, \tmp6\().8h,  v20, v22, \sz
+        uaddl_sz        \tmp7\().8h, \tmp8\().8h,  v23, v26, \sz
+        add_sz          v0.8h,  v1.8h,  \tmp1\().8h, \tmp2\().8h, \tmp1\().8h, \tmp2\().8h, \sz
+        uaddw_sz        v0.8h,  v1.8h,  v0.8h,  v1.8h,  v23, \sz
+        uaddw_sz        v0.8h,  v1.8h,  v0.8h,  v1.8h,  v24, \sz
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  \tmp5\().8h, \tmp6\().8h, \sz
+        sub_sz          \tmp3\().8h, \tmp4\().8h,  \tmp3\().8h, \tmp4\().8h,  \tmp1\().8h, \tmp2\().8h, \sz
+        sub_sz          \tmp7\().8h, \tmp8\().8h,  \tmp7\().8h, \tmp8\().8h,  \tmp5\().8h, \tmp6\().8h, \sz
+        rshrn_sz        v2,  v0.8h,  v1.8h,  #3,  \sz // out p2
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  \tmp3\().8h, \tmp4\().8h, \sz
+        uaddl_sz        \tmp1\().8h, \tmp2\().8h,  v20,  v23, \sz
+        uaddl_sz        \tmp3\().8h, \tmp4\().8h,  v24,  v27, \sz
+        rshrn_sz        v3,  v0.8h,  v1.8h,  #3,  \sz // out p1
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  \tmp7\().8h, \tmp8\().8h, \sz
+        sub_sz          \tmp3\().8h, \tmp4\().8h,  \tmp3\().8h, \tmp4\().8h,  \tmp1\().8h, \tmp2\().8h, \sz
+        uaddl_sz        \tmp5\().8h, \tmp6\().8h,  v21,  v24, \sz
+        uaddl_sz        \tmp7\().8h, \tmp8\().8h,  v25,  v27, \sz
+        rshrn_sz        v4,  v0.8h,  v1.8h,  #3,  \sz // out p0
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  \tmp3\().8h, \tmp4\().8h, \sz
+        sub_sz          \tmp7\().8h, \tmp8\().8h,  \tmp7\().8h, \tmp8\().8h,  \tmp5\().8h, \tmp6\().8h, \sz
+        uaddl_sz        \tmp1\().8h, \tmp2\().8h,  v22,  v25, \sz
+        uaddl_sz        \tmp3\().8h, \tmp4\().8h,  v26,  v27, \sz
+        rshrn_sz        v5,  v0.8h,  v1.8h,  #3,  \sz // out q0
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  \tmp7\().8h, \tmp8\().8h, \sz
+        sub_sz          \tmp3\().8h, \tmp4\().8h,  \tmp3\().8h, \tmp4\().8h,  \tmp1\().8h, \tmp2\().8h, \sz
+        rshrn_sz        \tmp5,  v0.8h,  v1.8h,  #3,  \sz // out q1
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  \tmp3\().8h, \tmp4\().8h, \sz
+        // The output here is written back into the input registers. This doesn't
+        // matter for the flat8part below, since we only update those pixels
+        // which won't be touched below.
+        bit             v21\sz, v2\sz,  v6\sz
+        bit             v22\sz, v3\sz,  v6\sz
+        bit             v23\sz, v4\sz,  v6\sz
+        rshrn_sz        \tmp6,  v0.8h,  v1.8h,  #3,  \sz // out q2
+        bit             v24\sz, v5\sz,  v6\sz
+        bit             v25\sz, \tmp5\sz,  v6\sz
+        bit             v26\sz, \tmp6\sz,  v6\sz
+.endif
+.if \wd == 16
+6:
+        orr             v2\sz,  v6\sz,  v7\sz
+        mov             x5,  v2.d[0]
+.ifc \sz, .16b
+        mov             x6,  v2.d[1]
+        orr             x5,  x5,  x6
+.endif
+        // If no pixels needed flat8in nor flat8out, jump to a
+        // writeout of the inner 4 pixels
+        cbz             x5,  7f
+        mov             x5,  v7.d[0]
+.ifc \sz, .16b
+        mov             x6,  v2.d[1]
+        orr             x5,  x5,  x6
+.endif
+        // If no pixels need flat8out, jump to a writeout of the inner 6 pixels
+        cbz             x5,  8f
+
+        // flat8out
+        // This writes all outputs into v2-v17 (skipping v6 and v16).
+        // If this part is skipped, the output is read from v21-v26 (which is the input
+        // to this section).
+        ushll_sz        v0.8h,  v1.8h,  v16,  #3,  \sz           // 8 * v16
+        usubw_sz        v0.8h,  v1.8h,  v0.8h,  v1.8h,  v16, \sz // 7 * v16
+        uaddw_sz        v0.8h,  v1.8h,  v0.8h,  v1.8h,  v17, \sz
+        uaddl_sz        v8.8h,  v9.8h,  v17, v18, \sz
+        uaddl_sz        v10.8h, v11.8h, v19, v20, \sz
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v8.8h,  v9.8h,  \sz
+        uaddl_sz        v8.8h,  v9.8h,  v16, v17, \sz
+        uaddl_sz        v12.8h, v13.8h, v21, v22, \sz
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v10.8h, v11.8h, \sz
+        uaddl_sz        v10.8h, v11.8h, v18, v25, \sz
+        uaddl_sz        v14.8h, v15.8h, v23, v24, \sz
+        sub_sz          v10.8h, v11.8h, v10.8h, v11.8h, v8.8h,  v9.8h,  \sz
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v12.8h, v13.8h, \sz
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v14.8h, v15.8h, \sz
+        uaddl_sz        v12.8h, v13.8h, v16, v18, \sz
+        uaddl_sz        v14.8h, v15.8h, v19, v26, \sz
+        rshrn_sz        v2,  v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v10.8h, v11.8h, \sz
+        uaddl_sz        v8.8h,  v9.8h,  v16, v19, \sz
+        uaddl_sz        v10.8h, v11.8h, v20, v27, \sz
+        sub_sz          v14.8h, v15.8h, v14.8h, v15.8h, v12.8h, v13.8h, \sz
+        bif             v2\sz,  v17\sz, v7\sz
+        rshrn_sz        v3,  v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v14.8h, v15.8h, \sz
+        uaddl_sz        v12.8h, v13.8h, v16, v20, \sz
+        uaddl_sz        v14.8h, v15.8h, v21, v28, \sz
+        sub_sz          v10.8h, v11.8h, v10.8h, v11.8h, v8.8h,  v9.8h,  \sz
+        bif             v3\sz,  v18\sz, v7\sz
+        rshrn_sz        v4,  v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v10.8h, v11.8h, \sz
+        uaddl_sz        v8.8h,  v9.8h,  v16, v21, \sz
+        uaddl_sz        v10.8h, v11.8h, v22, v29, \sz
+        sub_sz          v14.8h, v15.8h, v14.8h, v15.8h, v12.8h, v13.8h, \sz
+        bif             v4\sz,  v19\sz, v7\sz
+        rshrn_sz        v5,  v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v14.8h, v15.8h, \sz
+        uaddl_sz        v12.8h, v13.8h, v16, v22, \sz
+        uaddl_sz        v14.8h, v15.8h, v23, v30, \sz
+        sub_sz          v10.8h, v11.8h, v10.8h, v11.8h, v8.8h,  v9.8h,  \sz
+        bif             v5\sz,  v20\sz, v7\sz
+        rshrn_sz        v6,  v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v10.8h, v11.8h, \sz
+        uaddl_sz        v10.8h, v11.8h, v16, v23, \sz
+        sub_sz          v14.8h, v15.8h, v14.8h, v15.8h, v12.8h, v13.8h, \sz
+        uaddl_sz        v12.8h, v13.8h, v24, v31, \sz
+        bif             v6\sz,  v21\sz, v7\sz
+        rshrn_sz        v8,  v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v14.8h, v15.8h, \sz
+        sub_sz          v10.8h, v11.8h, v12.8h, v13.8h, v10.8h, v11.8h, \sz
+        uaddl_sz        v12.8h, v13.8h, v17, v24, \sz
+        uaddl_sz        v14.8h, v15.8h, v25, v31, \sz
+        bif             v8\sz,  v22\sz, v7\sz
+        rshrn_sz        v9,  v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v10.8h, v11.8h, \sz
+        sub_sz          v14.8h, v15.8h, v14.8h, v15.8h, v12.8h, v13.8h, \sz
+        uaddl_sz        v12.8h, v13.8h, v26, v31, \sz
+        bif             v9\sz,  v23\sz, v7\sz
+        rshrn_sz        v10, v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v14.8h, v15.8h, \sz
+        uaddl_sz        v14.8h, v15.8h, v18, v25, \sz
+        uaddl_sz        v18.8h, v19.8h, v19, v26, \sz
+        sub_sz          v12.8h, v13.8h, v12.8h, v13.8h, v14.8h, v15.8h, \sz
+        uaddl_sz        v14.8h, v15.8h, v27, v31, \sz
+        bif             v10\sz, v24\sz, v7\sz
+        rshrn_sz        v11, v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v12.8h, v13.8h, \sz
+        uaddl_sz        v12.8h, v13.8h, v20, v27, \sz
+        sub_sz          v14.8h, v15.8h, v14.8h, v15.8h, v18.8h, v19.8h, \sz
+        uaddl_sz        v18.8h, v19.8h, v28, v31, \sz
+        bif             v11\sz, v25\sz, v7\sz
+        sub_sz          v18.8h, v19.8h, v18.8h, v19.8h, v12.8h, v13.8h, \sz
+        rshrn_sz        v12, v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v14.8h, v15.8h, \sz
+        uaddl_sz        v14.8h, v15.8h, v21, v28, \sz
+        uaddl_sz        v20.8h, v21.8h, v29, v31, \sz
+        bif             v12\sz, v26\sz, v7\sz
+        rshrn_sz        v13, v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v18.8h, v19.8h, \sz
+        sub_sz          v20.8h, v21.8h, v20.8h, v21.8h, v14.8h, v15.8h, \sz
+        uaddl_sz        v18.8h, v19.8h, v22, v29, \sz
+        uaddl_sz        v22.8h, v23.8h, v30, v31, \sz
+        bif             v13\sz, v27\sz, v7\sz
+        rshrn_sz        v14, v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v20.8h, v21.8h, \sz
+        sub_sz          v22.8h, v23.8h, v22.8h, v23.8h, v18.8h, v19.8h, \sz
+        bif             v14\sz, v28\sz, v7\sz
+        rshrn_sz        v15, v0.8h,  v1.8h,  #4,  \sz
+
+        add_sz          v0.8h,  v1.8h,  v0.8h,  v1.8h,  v22.8h, v23.8h, \sz
+        bif             v15\sz, v29\sz, v7\sz
+        rshrn_sz        v17, v0.8h,  v1.8h,  #4,  \sz
+        bif             v17\sz, v30\sz, v7\sz
+.endif
+.endm
+
+// For wd <= 8, we use v16-v19 and v28-v31 for temp registers,
+// while we need those for inputs/outputs in wd=16 and use v8-v15
+// for temp registers there instead.
+function vp9_loop_filter_4
+        loop_filter     4,  .8b,  0,    v16, v17, v18, v19, v28, v29, v30, v31
+        ret
+9:
+        br              x10
+endfunc
+
+function vp9_loop_filter_4_16b_mix_44
+        loop_filter     4,  .16b, 44,   v16, v17, v18, v19, v28, v29, v30, v31
+        ret
+9:
+        br              x10
+endfunc
+
+function vp9_loop_filter_8
+        loop_filter     8,  .8b,  0,    v16, v17, v18, v19, v28, v29, v30, v31
+        mov             x5,  #0
+        ret
+6:
+        mov             x5,  #6
+        ret
+9:
+        br              x10
+endfunc
+
+function vp9_loop_filter_8_16b_mix
+        loop_filter     8,  .16b, 88,   v16, v17, v18, v19, v28, v29, v30, v31
+        mov             x5,  #0
+        ret
+6:
+        mov             x5,  #6
+        ret
+9:
+        br              x10
+endfunc
+
+function vp9_loop_filter_16
+        loop_filter     16, .8b,  0,    v8,  v9,  v10, v11, v12, v13, v14, v15
+        mov             x5,  #0
+        ret
+7:
+        mov             x5,  #7
+        ret
+8:
+        mov             x5,  #8
+        ret
+9:
+        ldp             d8,  d9,  [sp], 0x10
+        ldp             d10, d11, [sp], 0x10
+        ldp             d12, d13, [sp], 0x10
+        ldp             d14, d15, [sp], 0x10
+        br              x10
+endfunc
+
+function vp9_loop_filter_16_16b
+        loop_filter     16, .16b, 0,    v8,  v9,  v10, v11, v12, v13, v14, v15
+        mov             x5,  #0
+        ret
+7:
+        mov             x5,  #7
+        ret
+8:
+        mov             x5,  #8
+        ret
+9:
+        ldp             d8,  d9,  [sp], 0x10
+        ldp             d10, d11, [sp], 0x10
+        ldp             d12, d13, [sp], 0x10
+        ldp             d14, d15, [sp], 0x10
+        br              x10
+endfunc
+
+.macro loop_filter_4
+        bl              vp9_loop_filter_4
+.endm
+
+.macro loop_filter_4_16b_mix mix
+        bl              vp9_loop_filter_4_16b_mix_\mix
+.endm
+
+.macro loop_filter_8
+        bl              vp9_loop_filter_8
+        cbnz            x5,  6f
+.endm
+
+.macro loop_filter_8_16b_mix mix
+.if \mix == 48
+        mov             x11, #0xffffffff00000000
+.elseif \mix == 84
+        mov             x11, #0x00000000ffffffff
+.else
+        mov             x11, #0xffffffffffffffff
+.endif
+        bl              vp9_loop_filter_8_16b_mix
+        cbnz            x5,  6f
+.endm
+
+.macro loop_filter_16
+        bl              vp9_loop_filter_16
+        cmp             x5,  7
+        b.gt            8f
+        b.eq            7f
+.endm
+
+.macro loop_filter_16_16b
+        bl              vp9_loop_filter_16_16b
+        cmp             x5,  7
+        b.gt            8f
+        b.eq            7f
+.endm
+
+
+// The public functions in this file have got the following signature:
+// void loop_filter(uint8_t *dst, ptrdiff_t stride, int mb_lim, int lim, int hev_thr);
+
+function ff_vp9_loop_filter_v_4_8_neon, export=1
+        mov             x10, x30
+        sub             x9,  x0,  x1, lsl #2
+        ld1             {v20.8b}, [x9], x1 // p3
+        ld1             {v24.8b}, [x0], x1 // q0
+        ld1             {v21.8b}, [x9], x1 // p2
+        ld1             {v25.8b}, [x0], x1 // q1
+        ld1             {v22.8b}, [x9], x1 // p1
+        ld1             {v26.8b}, [x0], x1 // q2
+        ld1             {v23.8b}, [x9], x1 // p0
+        ld1             {v27.8b}, [x0], x1 // q3
+        sub             x0,  x0,  x1, lsl #2
+        sub             x9,  x9,  x1, lsl #1
+
+        loop_filter_4
+
+        st1             {v22.8b}, [x9], x1
+        st1             {v24.8b}, [x0], x1
+        st1             {v23.8b}, [x9], x1
+        st1             {v25.8b}, [x0], x1
+
+        br              x10
+endfunc
+
+function ff_vp9_loop_filter_v_44_16_neon, export=1
+        mov             x10, x30
+        sub             x9,  x0,  x1, lsl #2
+        ld1             {v20.16b}, [x9], x1 // p3
+        ld1             {v24.16b}, [x0], x1 // q0
+        ld1             {v21.16b}, [x9], x1 // p2
+        ld1             {v25.16b}, [x0], x1 // q1
+        ld1             {v22.16b}, [x9], x1 // p1
+        ld1             {v26.16b}, [x0], x1 // q2
+        ld1             {v23.16b}, [x9], x1 // p0
+        ld1             {v27.16b}, [x0], x1 // q3
+        sub             x0,  x0,  x1, lsl #2
+        sub             x9,  x9,  x1, lsl #1
+
+        loop_filter_4_16b_mix 44
+
+        st1             {v22.16b}, [x9], x1
+        st1             {v24.16b}, [x0], x1
+        st1             {v23.16b}, [x9], x1
+        st1             {v25.16b}, [x0], x1
+
+        br              x10
+endfunc
+
+function ff_vp9_loop_filter_h_4_8_neon, export=1
+        mov             x10, x30
+        sub             x9,  x0,  #4
+        add             x0,  x9,  x1, lsl #2
+        ld1             {v20.8b}, [x9], x1
+        ld1             {v24.8b}, [x0], x1
+        ld1             {v21.8b}, [x9], x1
+        ld1             {v25.8b}, [x0], x1
+        ld1             {v22.8b}, [x9], x1
+        ld1             {v26.8b}, [x0], x1
+        ld1             {v23.8b}, [x9], x1
+        ld1             {v27.8b}, [x0], x1
+
+        sub             x9,  x9,  x1, lsl #2
+        sub             x0,  x0,  x1, lsl #2
+        // Move x0/x9 forward by 2 pixels; we don't need to rewrite the
+        // outermost 2 pixels since they aren't changed.
+        add             x9,  x9,  #2
+        add             x0,  x0,  #2
+
+        transpose_8x8B  v20, v21, v22, v23, v24, v25, v26, v27, v28, v29
+
+        loop_filter_4
+
+        // We only will write the mid 4 pixels back; after the loop filter,
+        // these are in v22, v23, v24, v25, ordered as rows (8x4 pixels).
+        // We need to transpose them to columns, done with a 4x8 transpose
+        // (which in practice is two 4x4 transposes of the two 4x4 halves
+        // of the 8x4 pixels; into 4x8 pixels).
+        transpose_4x8B  v22, v23, v24, v25, v26, v27, v28, v29
+        st1             {v22.s}[0], [x9], x1
+        st1             {v22.s}[1], [x0], x1
+        st1             {v23.s}[0], [x9], x1
+        st1             {v23.s}[1], [x0], x1
+        st1             {v24.s}[0], [x9], x1
+        st1             {v24.s}[1], [x0], x1
+        st1             {v25.s}[0], [x9], x1
+        st1             {v25.s}[1], [x0], x1
+
+        br              x10
+endfunc
+
+function ff_vp9_loop_filter_h_44_16_neon, export=1
+        mov             x10, x30
+        sub             x9,  x0,  #4
+        add             x0,  x9,  x1, lsl #3
+        ld1             {v20.8b},   [x9], x1
+        ld1             {v20.d}[1], [x0], x1
+        ld1             {v21.8b},   [x9], x1
+        ld1             {v21.d}[1], [x0], x1
+        ld1             {v22.8b},   [x9], x1
+        ld1             {v22.d}[1], [x0], x1
+        ld1             {v23.8b},   [x9], x1
+        ld1             {v23.d}[1], [x0], x1
+        ld1             {v24.8b},   [x9], x1
+        ld1             {v24.d}[1], [x0], x1
+        ld1             {v25.8b},   [x9], x1
+        ld1             {v25.d}[1], [x0], x1
+        ld1             {v26.8b},   [x9], x1
+        ld1             {v26.d}[1], [x0], x1
+        ld1             {v27.8b},   [x9], x1
+        ld1             {v27.d}[1], [x0], x1
+
+        sub             x9,  x9,  x1, lsl #3
+        sub             x0,  x0,  x1, lsl #3
+        add             x9,  x9,  #2
+        add             x0,  x0,  #2
+
+        transpose_8x16B v20, v21, v22, v23, v24, v25, v26, v27, v28, v29
+
+        loop_filter_4_16b_mix 44
+
+        transpose_4x16B v22, v23, v24, v25, v26, v27, v28, v29
+
+        st1             {v22.s}[0], [x9], x1
+        st1             {v22.s}[2], [x0], x1
+        st1             {v23.s}[0], [x9], x1
+        st1             {v23.s}[2], [x0], x1
+        st1             {v24.s}[0], [x9], x1
+        st1             {v24.s}[2], [x0], x1
+        st1             {v25.s}[0], [x9], x1
+        st1             {v25.s}[2], [x0], x1
+        st1             {v22.s}[1], [x9], x1
+        st1             {v22.s}[3], [x0], x1
+        st1             {v23.s}[1], [x9], x1
+        st1             {v23.s}[3], [x0], x1
+        st1             {v24.s}[1], [x9], x1
+        st1             {v24.s}[3], [x0], x1
+        st1             {v25.s}[1], [x9], x1
+        st1             {v25.s}[3], [x0], x1
+
+        br              x10
+endfunc
+
+function ff_vp9_loop_filter_v_8_8_neon, export=1
+        mov             x10, x30
+        sub             x9,  x0,  x1, lsl #2
+        ld1             {v20.8b}, [x9], x1 // p3
+        ld1             {v24.8b}, [x0], x1 // q0
+        ld1             {v21.8b}, [x9], x1 // p2
+        ld1             {v25.8b}, [x0], x1 // q1
+        ld1             {v22.8b}, [x9], x1 // p1
+        ld1             {v26.8b}, [x0], x1 // q2
+        ld1             {v23.8b}, [x9], x1 // p0
+        ld1             {v27.8b}, [x0], x1 // q3
+        sub             x9,  x9,  x1, lsl #2
+        sub             x0,  x0,  x1, lsl #2
+        add             x9,  x9,  x1
+
+        loop_filter_8
+
+        st1             {v21.8b}, [x9], x1
+        st1             {v24.8b}, [x0], x1
+        st1             {v22.8b}, [x9], x1
+        st1             {v25.8b}, [x0], x1
+        st1             {v23.8b}, [x9], x1
+        st1             {v26.8b}, [x0], x1
+
+        br              x10
+6:
+        sub             x9,  x0,  x1, lsl #1
+        st1             {v22.8b}, [x9], x1
+        st1             {v24.8b}, [x0], x1
+        st1             {v23.8b}, [x9], x1
+        st1             {v25.8b}, [x0], x1
+        br              x10
+endfunc
+
+.macro mix_v_16 mix
+function ff_vp9_loop_filter_v_\mix\()_16_neon, export=1
+        mov             x10, x30
+        sub             x9,  x0,  x1, lsl #2
+        ld1             {v20.16b}, [x9], x1 // p3
+        ld1             {v24.16b}, [x0], x1 // q0
+        ld1             {v21.16b}, [x9], x1 // p2
+        ld1             {v25.16b}, [x0], x1 // q1
+        ld1             {v22.16b}, [x9], x1 // p1
+        ld1             {v26.16b}, [x0], x1 // q2
+        ld1             {v23.16b}, [x9], x1 // p0
+        ld1             {v27.16b}, [x0], x1 // q3
+        sub             x9,  x9,  x1, lsl #2
+        sub             x0,  x0,  x1, lsl #2
+        add             x9,  x9,  x1
+
+        loop_filter_8_16b_mix \mix
+
+        st1             {v21.16b}, [x9], x1
+        st1             {v24.16b}, [x0], x1
+        st1             {v22.16b}, [x9], x1
+        st1             {v25.16b}, [x0], x1
+        st1             {v23.16b}, [x9], x1
+        st1             {v26.16b}, [x0], x1
+
+        br              x10
+6:
+        sub             x9,  x0,  x1, lsl #1
+        st1             {v22.16b}, [x9], x1
+        st1             {v24.16b}, [x0], x1
+        st1             {v23.16b}, [x9], x1
+        st1             {v25.16b}, [x0], x1
+        br              x10
+endfunc
+.endm
+
+mix_v_16 48
+mix_v_16 84
+mix_v_16 88
+
+function ff_vp9_loop_filter_h_8_8_neon, export=1
+        mov             x10, x30
+        sub             x9,  x0,  #4
+        add             x0,  x9,  x1, lsl #2
+        ld1             {v20.8b}, [x9], x1
+        ld1             {v24.8b}, [x0], x1
+        ld1             {v21.8b}, [x9], x1
+        ld1             {v25.8b}, [x0], x1
+        ld1             {v22.8b}, [x9], x1
+        ld1             {v26.8b}, [x0], x1
+        ld1             {v23.8b}, [x9], x1
+        ld1             {v27.8b}, [x0], x1
+
+        sub             x9,  x9,  x1, lsl #2
+        sub             x0,  x0,  x1, lsl #2
+
+        transpose_8x8B  v20, v21, v22, v23, v24, v25, v26, v27, v28, v29
+
+        loop_filter_8
+
+        // Even though only 6 pixels per row have been changed, we write the
+        // full 8 pixel registers.
+        transpose_8x8B  v20, v21, v22, v23, v24, v25, v26, v27, v28, v29
+
+        st1             {v20.8b}, [x9], x1
+        st1             {v24.8b}, [x0], x1
+        st1             {v21.8b}, [x9], x1
+        st1             {v25.8b}, [x0], x1
+        st1             {v22.8b}, [x9], x1
+        st1             {v26.8b}, [x0], x1
+        st1             {v23.8b}, [x9], x1
+        st1             {v27.8b}, [x0], x1
+
+        br              x10
+6:
+        // If we didn't need to do the flat8in part, we use the same writeback
+        // as in loop_filter_h_4_8.
+        add             x9,  x9,  #2
+        add             x0,  x0,  #2
+        transpose_4x8B  v22, v23, v24, v25, v26, v27, v28, v29
+        st1             {v22.s}[0], [x9], x1
+        st1             {v22.s}[1], [x0], x1
+        st1             {v23.s}[0], [x9], x1
+        st1             {v23.s}[1], [x0], x1
+        st1             {v24.s}[0], [x9], x1
+        st1             {v24.s}[1], [x0], x1
+        st1             {v25.s}[0], [x9], x1
+        st1             {v25.s}[1], [x0], x1
+        br              x10
+endfunc
+
+.macro mix_h_16 mix
+function ff_vp9_loop_filter_h_\mix\()_16_neon, export=1
+        mov             x10, x30
+        sub             x9,  x0,  #4
+        add             x0,  x9,  x1, lsl #3
+        ld1             {v20.8b},   [x9], x1
+        ld1             {v20.d}[1], [x0], x1
+        ld1             {v21.8b},   [x9], x1
+        ld1             {v21.d}[1], [x0], x1
+        ld1             {v22.8b},   [x9], x1
+        ld1             {v22.d}[1], [x0], x1
+        ld1             {v23.8b},   [x9], x1
+        ld1             {v23.d}[1], [x0], x1
+        ld1             {v24.8b},   [x9], x1
+        ld1             {v24.d}[1], [x0], x1
+        ld1             {v25.8b},   [x9], x1
+        ld1             {v25.d}[1], [x0], x1
+        ld1             {v26.8b},   [x9], x1
+        ld1             {v26.d}[1], [x0], x1
+        ld1             {v27.8b},   [x9], x1
+        ld1             {v27.d}[1], [x0], x1
+
+        sub             x9,  x9,  x1, lsl #3
+        sub             x0,  x0,  x1, lsl #3
+
+        transpose_8x16B v20, v21, v22, v23, v24, v25, v26, v27, v28, v29
+
+        loop_filter_8_16b_mix \mix
+
+        transpose_8x16B v20, v21, v22, v23, v24, v25, v26, v27, v28, v29
+
+        st1             {v20.8b},   [x9], x1
+        st1             {v20.d}[1], [x0], x1
+        st1             {v21.8b},   [x9], x1
+        st1             {v21.d}[1], [x0], x1
+        st1             {v22.8b},   [x9], x1
+        st1             {v22.d}[1], [x0], x1
+        st1             {v23.8b},   [x9], x1
+        st1             {v23.d}[1], [x0], x1
+        st1             {v24.8b},   [x9], x1
+        st1             {v24.d}[1], [x0], x1
+        st1             {v25.8b},   [x9], x1
+        st1             {v25.d}[1], [x0], x1
+        st1             {v26.8b},   [x9], x1
+        st1             {v26.d}[1], [x0], x1
+        st1             {v27.8b},   [x9], x1
+        st1             {v27.d}[1], [x0], x1
+
+        br              x10
+6:
+        add             x9,  x9,  #2
+        add             x0,  x0,  #2
+        transpose_4x16B v22, v23, v24, v25, v26, v27, v28, v29
+        st1             {v22.s}[0], [x9], x1
+        st1             {v22.s}[2], [x0], x1
+        st1             {v23.s}[0], [x9], x1
+        st1             {v23.s}[2], [x0], x1
+        st1             {v24.s}[0], [x9], x1
+        st1             {v24.s}[2], [x0], x1
+        st1             {v25.s}[0], [x9], x1
+        st1             {v25.s}[2], [x0], x1
+        st1             {v22.s}[1], [x9], x1
+        st1             {v22.s}[3], [x0], x1
+        st1             {v23.s}[1], [x9], x1
+        st1             {v23.s}[3], [x0], x1
+        st1             {v24.s}[1], [x9], x1
+        st1             {v24.s}[3], [x0], x1
+        st1             {v25.s}[1], [x9], x1
+        st1             {v25.s}[3], [x0], x1
+        br              x10
+endfunc
+.endm
+
+mix_h_16 48
+mix_h_16 84
+mix_h_16 88
+
+function ff_vp9_loop_filter_v_16_8_neon, export=1
+        mov             x10, x30
+        stp             d14, d15, [sp, #-0x10]!
+        stp             d12, d13, [sp, #-0x10]!
+        stp             d10, d11, [sp, #-0x10]!
+        stp             d8,  d9,  [sp, #-0x10]!
+        sub             x9,  x0,  x1, lsl #3
+        ld1             {v16.8b}, [x9], x1 // p7
+        ld1             {v24.8b}, [x0], x1 // q0
+        ld1             {v17.8b}, [x9], x1 // p6
+        ld1             {v25.8b}, [x0], x1 // q1
+        ld1             {v18.8b}, [x9], x1 // p5
+        ld1             {v26.8b}, [x0], x1 // q2
+        ld1             {v19.8b}, [x9], x1 // p4
+        ld1             {v27.8b}, [x0], x1 // q3
+        ld1             {v20.8b}, [x9], x1 // p3
+        ld1             {v28.8b}, [x0], x1 // q4
+        ld1             {v21.8b}, [x9], x1 // p2
+        ld1             {v29.8b}, [x0], x1 // q5
+        ld1             {v22.8b}, [x9], x1 // p1
+        ld1             {v30.8b}, [x0], x1 // q6
+        ld1             {v23.8b}, [x9], x1 // p0
+        ld1             {v31.8b}, [x0], x1 // q7
+        sub             x9,  x9,  x1, lsl #3
+        sub             x0,  x0,  x1, lsl #3
+        add             x9,  x9,  x1
+
+        loop_filter_16
+
+        // If we did the flat8out part, we get the output in
+        // v2-v17 (skipping v7 and v16). x9 points to x0 - 7 * stride,
+        // store v2-v9 there, and v10-v17 into x0.
+        st1             {v2.8b},  [x9], x1
+        st1             {v10.8b}, [x0], x1
+        st1             {v3.8b},  [x9], x1
+        st1             {v11.8b}, [x0], x1
+        st1             {v4.8b},  [x9], x1
+        st1             {v12.8b}, [x0], x1
+        st1             {v5.8b},  [x9], x1
+        st1             {v13.8b}, [x0], x1
+        st1             {v6.8b},  [x9], x1
+        st1             {v14.8b}, [x0], x1
+        st1             {v8.8b},  [x9], x1
+        st1             {v15.8b}, [x0], x1
+        st1             {v9.8b},  [x9], x1
+        st1             {v17.8b}, [x0], x1
+9:
+        ldp             d8,  d9,  [sp], 0x10
+        ldp             d10, d11, [sp], 0x10
+        ldp             d12, d13, [sp], 0x10
+        ldp             d14, d15, [sp], 0x10
+        br              x10
+8:
+        add             x9,  x9,  x1, lsl #2
+        // If we didn't do the flat8out part, the output is left in the
+        // input registers.
+        st1             {v21.8b}, [x9], x1
+        st1             {v24.8b}, [x0], x1
+        st1             {v22.8b}, [x9], x1
+        st1             {v25.8b}, [x0], x1
+        st1             {v23.8b}, [x9], x1
+        st1             {v26.8b}, [x0], x1
+        b               9b
+7:
+        sub             x9,  x0,  x1, lsl #1
+        st1             {v22.8b}, [x9], x1
+        st1             {v24.8b}, [x0], x1
+        st1             {v23.8b}, [x9], x1
+        st1             {v25.8b}, [x0], x1
+        b               9b
+endfunc
+
+function ff_vp9_loop_filter_v_16_16_neon, export=1
+        mov             x10, x30
+        stp             d14, d15, [sp, #-0x10]!
+        stp             d12, d13, [sp, #-0x10]!
+        stp             d10, d11, [sp, #-0x10]!
+        stp             d8,  d9,  [sp, #-0x10]!
+        sub             x9,  x0,  x1, lsl #3
+        ld1             {v16.16b}, [x9], x1 // p7
+        ld1             {v24.16b}, [x0], x1 // q0
+        ld1             {v17.16b}, [x9], x1 // p6
+        ld1             {v25.16b}, [x0], x1 // q1
+        ld1             {v18.16b}, [x9], x1 // p5
+        ld1             {v26.16b}, [x0], x1 // q2
+        ld1             {v19.16b}, [x9], x1 // p4
+        ld1             {v27.16b}, [x0], x1 // q3
+        ld1             {v20.16b}, [x9], x1 // p3
+        ld1             {v28.16b}, [x0], x1 // q4
+        ld1             {v21.16b}, [x9], x1 // p2
+        ld1             {v29.16b}, [x0], x1 // q5
+        ld1             {v22.16b}, [x9], x1 // p1
+        ld1             {v30.16b}, [x0], x1 // q6
+        ld1             {v23.16b}, [x9], x1 // p0
+        ld1             {v31.16b}, [x0], x1 // q7
+        sub             x9,  x9,  x1, lsl #3
+        sub             x0,  x0,  x1, lsl #3
+        add             x9,  x9,  x1
+
+        loop_filter_16_16b
+
+        st1             {v2.16b},  [x9], x1
+        st1             {v10.16b}, [x0], x1
+        st1             {v3.16b},  [x9], x1
+        st1             {v11.16b}, [x0], x1
+        st1             {v4.16b},  [x9], x1
+        st1             {v12.16b}, [x0], x1
+        st1             {v5.16b},  [x9], x1
+        st1             {v13.16b}, [x0], x1
+        st1             {v6.16b},  [x9], x1
+        st1             {v14.16b}, [x0], x1
+        st1             {v8.16b},  [x9], x1
+        st1             {v15.16b}, [x0], x1
+        st1             {v9.16b},  [x9], x1
+        st1             {v17.16b}, [x0], x1
+9:
+        ldp             d8,  d9,  [sp], 0x10
+        ldp             d10, d11, [sp], 0x10
+        ldp             d12, d13, [sp], 0x10
+        ldp             d14, d15, [sp], 0x10
+        br              x10
+8:
+        add             x9,  x9,  x1, lsl #2
+        st1             {v21.16b}, [x9], x1
+        st1             {v24.16b}, [x0], x1
+        st1             {v22.16b}, [x9], x1
+        st1             {v25.16b}, [x0], x1
+        st1             {v23.16b}, [x9], x1
+        st1             {v26.16b}, [x0], x1
+        b               9b
+7:
+        sub             x9,  x0,  x1, lsl #1
+        st1             {v22.16b}, [x9], x1
+        st1             {v24.16b}, [x0], x1
+        st1             {v23.16b}, [x9], x1
+        st1             {v25.16b}, [x0], x1
+        b               9b
+endfunc
+
+function ff_vp9_loop_filter_h_16_8_neon, export=1
+        mov             x10, x30
+        stp             d14, d15, [sp, #-0x10]!
+        stp             d12, d13, [sp, #-0x10]!
+        stp             d10, d11, [sp, #-0x10]!
+        stp             d8,  d9,  [sp, #-0x10]!
+        sub             x9,  x0,  #8
+        ld1             {v16.8b}, [x9], x1
+        ld1             {v24.8b}, [x0], x1
+        ld1             {v17.8b}, [x9], x1
+        ld1             {v25.8b}, [x0], x1
+        ld1             {v18.8b}, [x9], x1
+        ld1             {v26.8b}, [x0], x1
+        ld1             {v19.8b}, [x9], x1
+        ld1             {v27.8b}, [x0], x1
+        ld1             {v20.8b}, [x9], x1
+        ld1             {v28.8b}, [x0], x1
+        ld1             {v21.8b}, [x9], x1
+        ld1             {v29.8b}, [x0], x1
+        ld1             {v22.8b}, [x9], x1
+        ld1             {v30.8b}, [x0], x1
+        ld1             {v23.8b}, [x9], x1
+        ld1             {v31.8b}, [x0], x1
+        sub             x0,  x0,  x1, lsl #3
+        sub             x9,  x9,  x1, lsl #3
+
+        // The 16x8 pixels read above is in two 8x8 blocks; the left
+        // half in v16-v23, and the right half in v24-v31. Do two 8x8 transposes
+        // of this, to get one column per register.
+        transpose_8x8B  v16, v17, v18, v19, v20, v21, v22, v23, v0, v1
+        transpose_8x8B  v24, v25, v26, v27, v28, v29, v30, v31, v0, v1
+
+        loop_filter_16
+
+        transpose_8x8B  v16, v2,  v3,  v4,  v5,  v6,  v8,  v9,  v0, v1
+        transpose_8x8B  v10, v11, v12, v13, v14, v15, v17, v31, v0, v1
+
+        st1             {v16.8b}, [x9], x1
+        st1             {v10.8b}, [x0], x1
+        st1             {v2.8b},  [x9], x1
+        st1             {v11.8b}, [x0], x1
+        st1             {v3.8b},  [x9], x1
+        st1             {v12.8b}, [x0], x1
+        st1             {v4.8b},  [x9], x1
+        st1             {v13.8b}, [x0], x1
+        st1             {v5.8b},  [x9], x1
+        st1             {v14.8b}, [x0], x1
+        st1             {v6.8b},  [x9], x1
+        st1             {v15.8b}, [x0], x1
+        st1             {v8.8b},  [x9], x1
+        st1             {v17.8b}, [x0], x1
+        st1             {v9.8b},  [x9], x1
+        st1             {v31.8b}, [x0], x1
+9:
+        ldp             d8,  d9,  [sp], 0x10
+        ldp             d10, d11, [sp], 0x10
+        ldp             d12, d13, [sp], 0x10
+        ldp             d14, d15, [sp], 0x10
+        br              x10
+8:
+        // The same writeback as in loop_filter_h_8_8
+        sub             x9,  x0,  #4
+        add             x0,  x9,  x1, lsl #2
+        transpose_8x8B  v20, v21, v22, v23, v24, v25, v26, v27, v28, v29
+
+        st1             {v20.8b}, [x9], x1
+        st1             {v24.8b}, [x0], x1
+        st1             {v21.8b}, [x9], x1
+        st1             {v25.8b}, [x0], x1
+        st1             {v22.8b}, [x9], x1
+        st1             {v26.8b}, [x0], x1
+        st1             {v23.8b}, [x9], x1
+        st1             {v27.8b}, [x0], x1
+        b               9b
+7:
+        // The same writeback as in loop_filter_h_4_8
+        sub             x9,  x0,  #2
+        add             x0,  x9,  x1, lsl #2
+        transpose_4x8B  v22, v23, v24, v25, v26, v27, v28, v29
+        st1             {v22.s}[0], [x9], x1
+        st1             {v22.s}[1], [x0], x1
+        st1             {v23.s}[0], [x9], x1
+        st1             {v23.s}[1], [x0], x1
+        st1             {v24.s}[0], [x9], x1
+        st1             {v24.s}[1], [x0], x1
+        st1             {v25.s}[0], [x9], x1
+        st1             {v25.s}[1], [x0], x1
+        b               9b
+endfunc
+
+function ff_vp9_loop_filter_h_16_16_neon, export=1
+        mov             x10, x30
+        stp             d14, d15, [sp, #-0x10]!
+        stp             d12, d13, [sp, #-0x10]!
+        stp             d10, d11, [sp, #-0x10]!
+        stp             d8,  d9,  [sp, #-0x10]!
+        sub             x9,  x0,  #8
+        ld1             {v16.8b},   [x9], x1
+        ld1             {v24.8b},   [x0], x1
+        ld1             {v17.8b},   [x9], x1
+        ld1             {v25.8b},   [x0], x1
+        ld1             {v18.8b},   [x9], x1
+        ld1             {v26.8b},   [x0], x1
+        ld1             {v19.8b},   [x9], x1
+        ld1             {v27.8b},   [x0], x1
+        ld1             {v20.8b},   [x9], x1
+        ld1             {v28.8b},   [x0], x1
+        ld1             {v21.8b},   [x9], x1
+        ld1             {v29.8b},   [x0], x1
+        ld1             {v22.8b},   [x9], x1
+        ld1             {v30.8b},   [x0], x1
+        ld1             {v23.8b},   [x9], x1
+        ld1             {v31.8b},   [x0], x1
+        ld1             {v16.d}[1], [x9], x1
+        ld1             {v24.d}[1], [x0], x1
+        ld1             {v17.d}[1], [x9], x1
+        ld1             {v25.d}[1], [x0], x1
+        ld1             {v18.d}[1], [x9], x1
+        ld1             {v26.d}[1], [x0], x1
+        ld1             {v19.d}[1], [x9], x1
+        ld1             {v27.d}[1], [x0], x1
+        ld1             {v20.d}[1], [x9], x1
+        ld1             {v28.d}[1], [x0], x1
+        ld1             {v21.d}[1], [x9], x1
+        ld1             {v29.d}[1], [x0], x1
+        ld1             {v22.d}[1], [x9], x1
+        ld1             {v30.d}[1], [x0], x1
+        ld1             {v23.d}[1], [x9], x1
+        ld1             {v31.d}[1], [x0], x1
+        sub             x0,  x0,  x1, lsl #4
+        sub             x9,  x9,  x1, lsl #4
+
+        transpose_8x16B v16, v17, v18, v19, v20, v21, v22, v23, v0, v1
+        transpose_8x16B v24, v25, v26, v27, v28, v29, v30, v31, v0, v1
+
+        loop_filter_16_16b
+
+        transpose_8x16B v16, v2,  v3,  v4,  v5,  v6,  v8,  v9,  v0, v1
+        transpose_8x16B v10, v11, v12, v13, v14, v15, v17, v31, v0, v1
+
+        st1             {v16.8b},   [x9], x1
+        st1             {v10.8b},   [x0], x1
+        st1             {v2.8b},    [x9], x1
+        st1             {v11.8b},   [x0], x1
+        st1             {v3.8b},    [x9], x1
+        st1             {v12.8b},   [x0], x1
+        st1             {v4.8b},    [x9], x1
+        st1             {v13.8b},   [x0], x1
+        st1             {v5.8b},    [x9], x1
+        st1             {v14.8b},   [x0], x1
+        st1             {v6.8b},    [x9], x1
+        st1             {v15.8b},   [x0], x1
+        st1             {v8.8b},    [x9], x1
+        st1             {v17.8b},   [x0], x1
+        st1             {v9.8b},    [x9], x1
+        st1             {v31.8b},   [x0], x1
+        st1             {v16.d}[1], [x9], x1
+        st1             {v10.d}[1], [x0], x1
+        st1             {v2.d}[1],  [x9], x1
+        st1             {v11.d}[1], [x0], x1
+        st1             {v3.d}[1],  [x9], x1
+        st1             {v12.d}[1], [x0], x1
+        st1             {v4.d}[1],  [x9], x1
+        st1             {v13.d}[1], [x0], x1
+        st1             {v5.d}[1],  [x9], x1
+        st1             {v14.d}[1], [x0], x1
+        st1             {v6.d}[1],  [x9], x1
+        st1             {v15.d}[1], [x0], x1
+        st1             {v8.d}[1],  [x9], x1
+        st1             {v17.d}[1], [x0], x1
+        st1             {v9.d}[1],  [x9], x1
+        st1             {v31.d}[1], [x0], x1
+9:
+        ldp             d8,  d9,  [sp], 0x10
+        ldp             d10, d11, [sp], 0x10
+        ldp             d12, d13, [sp], 0x10
+        ldp             d14, d15, [sp], 0x10
+        br              x10
+8:
+        sub             x9,  x0,  #4
+        add             x0,  x9,  x1, lsl #3
+        transpose_8x16B v20, v21, v22, v23, v24, v25, v26, v27, v28, v29
+
+        st1             {v20.8b},   [x9], x1
+        st1             {v20.d}[1], [x0], x1
+        st1             {v21.8b},   [x9], x1
+        st1             {v21.d}[1], [x0], x1
+        st1             {v22.8b},   [x9], x1
+        st1             {v22.d}[1], [x0], x1
+        st1             {v23.8b},   [x9], x1
+        st1             {v23.d}[1], [x0], x1
+        st1             {v24.8b},   [x9], x1
+        st1             {v24.d}[1], [x0], x1
+        st1             {v25.8b},   [x9], x1
+        st1             {v25.d}[1], [x0], x1
+        st1             {v26.8b},   [x9], x1
+        st1             {v26.d}[1], [x0], x1
+        st1             {v27.8b},   [x9], x1
+        st1             {v27.d}[1], [x0], x1
+        b               9b
+7:
+        sub             x9,  x0,  #2
+        add             x0,  x9,  x1, lsl #3
+        transpose_4x16B v22, v23, v24, v25, v26, v27, v28, v29
+        st1             {v22.s}[0], [x9], x1
+        st1             {v22.s}[2], [x0], x1
+        st1             {v23.s}[0], [x9], x1
+        st1             {v23.s}[2], [x0], x1
+        st1             {v24.s}[0], [x9], x1
+        st1             {v24.s}[2], [x0], x1
+        st1             {v25.s}[0], [x9], x1
+        st1             {v25.s}[2], [x0], x1
+        st1             {v22.s}[1], [x9], x1
+        st1             {v22.s}[3], [x0], x1
+        st1             {v23.s}[1], [x9], x1
+        st1             {v23.s}[3], [x0], x1
+        st1             {v24.s}[1], [x9], x1
+        st1             {v24.s}[3], [x0], x1
+        st1             {v25.s}[1], [x9], x1
+        st1             {v25.s}[3], [x0], x1
+        b               9b
+endfunc