aarch64: Add NEON optimizations for 10 and 12 bit vp9 itxfm
authorMartin Storsjö <martin@martin.st>
Tue, 3 Jan 2017 12:35:54 +0000 (14:35 +0200)
committerMartin Storsjö <martin@martin.st>
Tue, 24 Jan 2017 20:36:08 +0000 (22:36 +0200)
This work is sponsored by, and copyright, Google.

Compared to the arm version, on aarch64 we can keep the full 8x8
transform in registers, and for 16x16 and 32x32, we can process
it in slices of 4 pixels instead of 2.

Examples of runtimes vs the 32 bit version, on a Cortex A53:
                                                ARM  AArch64
vp9_inv_adst_adst_4x4_sub4_add_10_neon:       111.0    109.7
vp9_inv_adst_adst_8x8_sub8_add_10_neon:       914.0    733.5
vp9_inv_adst_adst_16x16_sub16_add_10_neon:   5184.0   3745.7
vp9_inv_dct_dct_4x4_sub1_add_10_neon:          65.0     65.7
vp9_inv_dct_dct_4x4_sub4_add_10_neon:         100.0     96.7
vp9_inv_dct_dct_8x8_sub1_add_10_neon:         111.0    119.7
vp9_inv_dct_dct_8x8_sub8_add_10_neon:         618.0    494.7
vp9_inv_dct_dct_16x16_sub1_add_10_neon:       295.1    284.6
vp9_inv_dct_dct_16x16_sub2_add_10_neon:      2303.2   1883.9
vp9_inv_dct_dct_16x16_sub8_add_10_neon:      2984.8   2189.3
vp9_inv_dct_dct_16x16_sub16_add_10_neon:     3890.0   2799.4
vp9_inv_dct_dct_32x32_sub1_add_10_neon:      1044.4   1012.7
vp9_inv_dct_dct_32x32_sub2_add_10_neon:     13333.7   9695.1
vp9_inv_dct_dct_32x32_sub16_add_10_neon:    18531.3  12459.8
vp9_inv_dct_dct_32x32_sub32_add_10_neon:    24470.7  16160.2
vp9_inv_wht_wht_4x4_sub4_add_10_neon:          83.0     79.7

The larger transforms are significantly faster than the corresponding
ARM versions.

The speedup vs C code is smaller than in 32 bit mode, probably
because the 64 bit intermediates in the C code can be expressed
more efficiently in aarch64.

Signed-off-by: Martin Storsjö <martin@martin.st>
libavcodec/aarch64/Makefile
libavcodec/aarch64/vp9dsp_init_16bpp_aarch64_template.c
libavcodec/aarch64/vp9itxfm_16bpp_neon.S [new file with mode: 0644]

index 0766e90..715cc6f 100644 (file)
@@ -42,7 +42,8 @@ NEON-OBJS-$(CONFIG_MPEGAUDIODSP)        += aarch64/mpegaudiodsp_neon.o
 # decoders/encoders
 NEON-OBJS-$(CONFIG_DCA_DECODER)         += aarch64/synth_filter_neon.o
 NEON-OBJS-$(CONFIG_VORBIS_DECODER)      += aarch64/vorbisdsp_neon.o
-NEON-OBJS-$(CONFIG_VP9_DECODER)         += aarch64/vp9itxfm_neon.o             \
+NEON-OBJS-$(CONFIG_VP9_DECODER)         += aarch64/vp9itxfm_16bpp_neon.o       \
+                                           aarch64/vp9itxfm_neon.o             \
                                            aarch64/vp9lpf_neon.o               \
                                            aarch64/vp9mc_16bpp_neon.o          \
                                            aarch64/vp9mc_neon.o
index 4719ea3..0e86b02 100644 (file)
@@ -157,7 +157,54 @@ static av_cold void vp9dsp_mc_init_aarch64(VP9DSPContext *dsp)
     }
 }
 
+#define define_itxfm2(type_a, type_b, sz, bpp)                                     \
+void ff_vp9_##type_a##_##type_b##_##sz##x##sz##_add_##bpp##_neon(uint8_t *_dst,    \
+                                                                 ptrdiff_t stride, \
+                                                                 int16_t *_block, int eob)
+#define define_itxfm(type_a, type_b, sz, bpp) define_itxfm2(type_a, type_b, sz, bpp)
+
+#define define_itxfm_funcs(sz, bpp)      \
+    define_itxfm(idct,  idct,  sz, bpp); \
+    define_itxfm(iadst, idct,  sz, bpp); \
+    define_itxfm(idct,  iadst, sz, bpp); \
+    define_itxfm(iadst, iadst, sz, bpp)
+
+define_itxfm_funcs(4,  BPP);
+define_itxfm_funcs(8,  BPP);
+define_itxfm_funcs(16, BPP);
+define_itxfm(idct, idct, 32, BPP);
+define_itxfm(iwht, iwht, 4,  BPP);
+
+
+static av_cold void vp9dsp_itxfm_init_aarch64(VP9DSPContext *dsp)
+{
+    int cpu_flags = av_get_cpu_flags();
+
+    if (have_neon(cpu_flags)) {
+#define init_itxfm2(tx, sz, bpp)                                               \
+    dsp->itxfm_add[tx][DCT_DCT]   = ff_vp9_idct_idct_##sz##_add_##bpp##_neon;  \
+    dsp->itxfm_add[tx][DCT_ADST]  = ff_vp9_iadst_idct_##sz##_add_##bpp##_neon; \
+    dsp->itxfm_add[tx][ADST_DCT]  = ff_vp9_idct_iadst_##sz##_add_##bpp##_neon; \
+    dsp->itxfm_add[tx][ADST_ADST] = ff_vp9_iadst_iadst_##sz##_add_##bpp##_neon
+#define init_itxfm(tx, sz, bpp) init_itxfm2(tx, sz, bpp)
+
+#define init_idct2(tx, nm, bpp)     \
+    dsp->itxfm_add[tx][DCT_DCT]   = \
+    dsp->itxfm_add[tx][ADST_DCT]  = \
+    dsp->itxfm_add[tx][DCT_ADST]  = \
+    dsp->itxfm_add[tx][ADST_ADST] = ff_vp9_##nm##_add_##bpp##_neon
+#define init_idct(tx, nm, bpp) init_idct2(tx, nm, bpp)
+
+        init_itxfm(TX_4X4,   4x4,   BPP);
+        init_itxfm(TX_8X8,   8x8,   BPP);
+        init_itxfm(TX_16X16, 16x16, BPP);
+        init_idct(TX_32X32, idct_idct_32x32, BPP);
+        init_idct(4,        iwht_iwht_4x4,   BPP);
+    }
+}
+
 av_cold void INIT_FUNC(VP9DSPContext *dsp)
 {
     vp9dsp_mc_init_aarch64(dsp);
+    vp9dsp_itxfm_init_aarch64(dsp);
 }
diff --git a/libavcodec/aarch64/vp9itxfm_16bpp_neon.S b/libavcodec/aarch64/vp9itxfm_16bpp_neon.S
new file mode 100644 (file)
index 0000000..f53e94a
--- /dev/null
@@ -0,0 +1,1517 @@
+/*
+ * Copyright (c) 2017 Google Inc.
+ *
+ * This file is part of FFmpeg.
+ *
+ * FFmpeg 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.
+ *
+ * FFmpeg 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 FFmpeg; 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"
+
+const itxfm4_coeffs, align=4
+        .short  11585, 0, 6270, 15137
+iadst4_coeffs:
+        .short  5283, 15212, 9929, 13377
+endconst
+
+const iadst8_coeffs, align=4
+        .short  16305, 1606, 14449, 7723, 10394, 12665, 4756, 15679
+idct_coeffs:
+        .short  11585, 0, 6270, 15137, 3196, 16069, 13623, 9102
+        .short  1606, 16305, 12665, 10394, 7723, 14449, 15679, 4756
+        .short  804, 16364, 12140, 11003, 7005, 14811, 15426, 5520
+        .short  3981, 15893, 14053, 8423, 9760, 13160, 16207, 2404
+endconst
+
+const iadst16_coeffs, align=4
+        .short  16364, 804, 15893, 3981, 11003, 12140, 8423, 14053
+        .short  14811, 7005, 13160, 9760, 5520, 15426, 2404, 16207
+endconst
+
+.macro transpose_4x4s r0, r1, r2, r3, r4, r5, r6, r7
+        trn1            \r4\().4s,  \r0\().4s,  \r1\().4s
+        trn2            \r5\().4s,  \r0\().4s,  \r1\().4s
+        trn1            \r6\().4s,  \r2\().4s,  \r3\().4s
+        trn2            \r7\().4s,  \r2\().4s,  \r3\().4s
+        trn1            \r0\().2d,  \r4\().2d,  \r6\().2d
+        trn2            \r2\().2d,  \r4\().2d,  \r6\().2d
+        trn1            \r1\().2d,  \r5\().2d,  \r7\().2d
+        trn2            \r3\().2d,  \r5\().2d,  \r7\().2d
+.endm
+
+// Transpose a 8x8 matrix of 32 bit elements, where each row is spread out
+// over two registers.
+.macro transpose_8x8s r0, r1, r2, r3, r4, r5, r6, r7, r8, r9, r10, r11, r12, r13, r14, r15, t0, t1, t2, t3
+        transpose_4x4s  \r0,  \r2,  \r4,  \r6,  \t0, \t1, \t2, \t3
+        transpose_4x4s  \r9,  \r11, \r13, \r15, \t0, \t1, \t2, \t3
+
+        // Do 4x4 transposes of r1,r3,r5,r7 and r8,r10,r12,r14
+        // while swapping the two 4x4 matrices between each other
+
+        // First step of the 4x4 transpose of r1-r7, into t0-t3
+        trn1            \t0\().4s,  \r1\().4s,  \r3\().4s
+        trn2            \t1\().4s,  \r1\().4s,  \r3\().4s
+        trn1            \t2\().4s,  \r5\().4s,  \r7\().4s
+        trn2            \t3\().4s,  \r5\().4s,  \r7\().4s
+
+        // First step of the 4x4 transpose of r8-r12, into r1-r7
+        trn1            \r1\().4s,  \r8\().4s,  \r10\().4s
+        trn2            \r3\().4s,  \r8\().4s,  \r10\().4s
+        trn1            \r5\().4s,  \r12\().4s, \r14\().4s
+        trn2            \r7\().4s,  \r12\().4s, \r14\().4s
+
+        // Second step of the 4x4 transpose of r1-r7 (now in t0-r3), into r8-r12
+        trn1            \r8\().2d,  \t0\().2d,  \t2\().2d
+        trn2            \r12\().2d, \t0\().2d,  \t2\().2d
+        trn1            \r10\().2d, \t1\().2d,  \t3\().2d
+        trn2            \r14\().2d, \t1\().2d,  \t3\().2d
+
+        // Second step of the 4x4 transpose of r8-r12 (now in r1-r7), in place as far as possible
+        trn1            \t0\().2d,  \r1\().2d,  \r5\().2d
+        trn2            \r5\().2d,  \r1\().2d,  \r5\().2d
+        trn1            \t1\().2d,  \r3\().2d,  \r7\().2d
+        trn2            \r7\().2d,  \r3\().2d,  \r7\().2d
+
+        // Move the outputs of trn1 back in place
+        mov             \r1\().16b,  \t0\().16b
+        mov             \r3\().16b,  \t1\().16b
+.endm
+
+// out1 = ((in1 + in2) * d0[0] + (1 << 13)) >> 14
+// out2 = ((in1 - in2) * d0[0] + (1 << 13)) >> 14
+// in/out are .4s registers; this can do with 4 temp registers, but is
+// more efficient if 6 temp registers are available.
+.macro dmbutterfly0 out1, out2, in1, in2, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, neg=0
+.if \neg > 0
+        neg             \tmp4\().4s, v0.4s
+.endif
+        add             \tmp1\().4s, \in1\().4s,  \in2\().4s
+        sub             \tmp2\().4s, \in1\().4s,  \in2\().4s
+.if \neg > 0
+        smull           \tmp3\().2d, \tmp1\().2s, \tmp4\().s[0]
+        smull2          \tmp4\().2d, \tmp1\().4s, \tmp4\().s[0]
+.else
+        smull           \tmp3\().2d, \tmp1\().2s, v0.s[0]
+        smull2          \tmp4\().2d, \tmp1\().4s, v0.s[0]
+.endif
+.ifb \tmp5
+        rshrn           \out1\().2s, \tmp3\().2d, #14
+        rshrn2          \out1\().4s, \tmp4\().2d, #14
+        smull           \tmp3\().2d, \tmp2\().2s, v0.s[0]
+        smull2          \tmp4\().2d, \tmp2\().4s, v0.s[0]
+        rshrn           \out2\().2s, \tmp3\().2d, #14
+        rshrn2          \out2\().4s, \tmp4\().2d, #14
+.else
+        smull           \tmp5\().2d, \tmp2\().2s, v0.s[0]
+        smull2          \tmp6\().2d, \tmp2\().4s, v0.s[0]
+        rshrn           \out1\().2s, \tmp3\().2d, #14
+        rshrn2          \out1\().4s, \tmp4\().2d, #14
+        rshrn           \out2\().2s, \tmp5\().2d, #14
+        rshrn2          \out2\().4s, \tmp6\().2d, #14
+.endif
+.endm
+
+// out1,out2 = in1 * coef1 - in2 * coef2
+// out3,out4 = in1 * coef2 + in2 * coef1
+// out are 4 x .2d registers, in are 2 x .4s registers
+.macro dmbutterfly_l out1, out2, out3, out4, in1, in2, coef1, coef2
+        smull           \out1\().2d, \in1\().2s, \coef1
+        smull2          \out2\().2d, \in1\().4s, \coef1
+        smull           \out3\().2d, \in1\().2s, \coef2
+        smull2          \out4\().2d, \in1\().4s, \coef2
+        smlsl           \out1\().2d, \in2\().2s, \coef2
+        smlsl2          \out2\().2d, \in2\().4s, \coef2
+        smlal           \out3\().2d, \in2\().2s, \coef1
+        smlal2          \out4\().2d, \in2\().4s, \coef1
+.endm
+
+// inout1 = (inout1 * coef1 - inout2 * coef2 + (1 << 13)) >> 14
+// inout2 = (inout1 * coef2 + inout2 * coef1 + (1 << 13)) >> 14
+// inout are 2 x .4s registers
+.macro dmbutterfly inout1, inout2, coef1, coef2, tmp1, tmp2, tmp3, tmp4, neg=0
+        dmbutterfly_l   \tmp1, \tmp2, \tmp3, \tmp4, \inout1, \inout2, \coef1, \coef2
+.if \neg > 0
+        neg             \tmp3\().2d, \tmp3\().2d
+        neg             \tmp4\().2d, \tmp4\().2d
+.endif
+        rshrn           \inout1\().2s, \tmp1\().2d,  #14
+        rshrn2          \inout1\().4s, \tmp2\().2d,  #14
+        rshrn           \inout2\().2s, \tmp3\().2d,  #14
+        rshrn2          \inout2\().4s, \tmp4\().2d,  #14
+.endm
+
+// out1 = in1 + in2
+// out2 = in1 - in2
+.macro butterfly_4s out1, out2, in1, in2
+        add             \out1\().4s, \in1\().4s, \in2\().4s
+        sub             \out2\().4s, \in1\().4s, \in2\().4s
+.endm
+
+// out1 = in1 - in2
+// out2 = in1 + in2
+.macro butterfly_4s_r out1, out2, in1, in2
+        sub             \out1\().4s, \in1\().4s, \in2\().4s
+        add             \out2\().4s, \in1\().4s, \in2\().4s
+.endm
+
+// out1 = (in1,in2 + in3,in4 + (1 << 13)) >> 14
+// out2 = (in1,in2 - in3,in4 + (1 << 13)) >> 14
+// out are 2 x .4s registers, in are 4 x .2d registers
+.macro dbutterfly_n out1, out2, in1, in2, in3, in4, tmp1, tmp2, tmp3, tmp4
+        add             \tmp1\().2d, \in1\().2d, \in3\().2d
+        add             \tmp2\().2d, \in2\().2d, \in4\().2d
+        sub             \tmp3\().2d, \in1\().2d, \in3\().2d
+        sub             \tmp4\().2d, \in2\().2d, \in4\().2d
+        rshrn           \out1\().2s, \tmp1\().2d,  #14
+        rshrn2          \out1\().4s, \tmp2\().2d,  #14
+        rshrn           \out2\().2s, \tmp3\().2d,  #14
+        rshrn2          \out2\().4s, \tmp4\().2d,  #14
+.endm
+
+.macro iwht4_10 c0, c1, c2, c3
+        add             \c0\().4s, \c0\().4s, \c1\().4s
+        sub             v17.4s,    \c2\().4s, \c3\().4s
+        sub             v16.4s,    \c0\().4s, v17.4s
+        sshr            v16.4s,    v16.4s,    #1
+        sub             \c2\().4s, v16.4s,    \c1\().4s
+        sub             \c1\().4s, v16.4s,    \c3\().4s
+        add             \c3\().4s, v17.4s,    \c2\().4s
+        sub             \c0\().4s, \c0\().4s, \c1\().4s
+.endm
+
+.macro iwht4_12 c0, c1, c2, c3
+        iwht4_10        \c0, \c1, \c2, \c3
+.endm
+
+.macro idct4_10 c0, c1, c2, c3
+        mul             v22.4s,    \c1\().4s, v0.s[3]
+        mul             v20.4s,    \c1\().4s, v0.s[2]
+        add             v16.4s,    \c0\().4s, \c2\().4s
+        sub             v17.4s,    \c0\().4s, \c2\().4s
+        mla             v22.4s,    \c3\().4s, v0.s[2]
+        mul             v18.4s,    v16.4s,    v0.s[0]
+        mul             v24.4s,    v17.4s,    v0.s[0]
+        mls             v20.4s,    \c3\().4s, v0.s[3]
+        srshr           v22.4s,    v22.4s,    #14
+        srshr           v18.4s,    v18.4s,    #14
+        srshr           v24.4s,    v24.4s,    #14
+        srshr           v20.4s,    v20.4s,    #14
+        add             \c0\().4s, v18.4s,    v22.4s
+        sub             \c3\().4s, v18.4s,    v22.4s
+        add             \c1\().4s, v24.4s,    v20.4s
+        sub             \c2\().4s, v24.4s,    v20.4s
+.endm
+
+.macro idct4_12 c0, c1, c2, c3
+        smull           v22.2d,    \c1\().2s, v0.s[3]
+        smull2          v23.2d,    \c1\().4s, v0.s[3]
+        smull           v20.2d,    \c1\().2s, v0.s[2]
+        smull2          v21.2d,    \c1\().4s, v0.s[2]
+        add             v16.4s,    \c0\().4s, \c2\().4s
+        sub             v17.4s,    \c0\().4s, \c2\().4s
+        smlal           v22.2d,    \c3\().2s, v0.s[2]
+        smlal2          v23.2d,    \c3\().4s, v0.s[2]
+        smull           v18.2d,    v16.2s,    v0.s[0]
+        smull2          v19.2d,    v16.4s,    v0.s[0]
+        smull           v24.2d,    v17.2s,    v0.s[0]
+        smull2          v25.2d,    v17.4s,    v0.s[0]
+        smlsl           v20.2d,    \c3\().2s, v0.s[3]
+        smlsl2          v21.2d,    \c3\().4s, v0.s[3]
+        rshrn           v22.2s,    v22.2d,    #14
+        rshrn2          v22.4s,    v23.2d,    #14
+        rshrn           v18.2s,    v18.2d,    #14
+        rshrn2          v18.4s,    v19.2d,    #14
+        rshrn           v24.2s,    v24.2d,    #14
+        rshrn2          v24.4s,    v25.2d,    #14
+        rshrn           v20.2s,    v20.2d,    #14
+        rshrn2          v20.4s,    v21.2d,    #14
+        add             \c0\().4s, v18.4s,    v22.4s
+        sub             \c3\().4s, v18.4s,    v22.4s
+        add             \c1\().4s, v24.4s,    v20.4s
+        sub             \c2\().4s, v24.4s,    v20.4s
+.endm
+
+.macro iadst4_10 c0, c1, c2, c3
+        mul             v16.4s,    \c0\().4s, v1.s[0]
+        mla             v16.4s,    \c2\().4s, v1.s[1]
+        mla             v16.4s,    \c3\().4s, v1.s[2]
+        mul             v18.4s,    \c0\().4s, v1.s[2]
+        mls             v18.4s,    \c2\().4s, v1.s[0]
+        sub             \c0\().4s, \c0\().4s, \c2\().4s
+        mls             v18.4s,    \c3\().4s, v1.s[1]
+        add             \c0\().4s, \c0\().4s, \c3\().4s
+        mul             v22.4s,    \c1\().4s, v1.s[3]
+        mul             v20.4s,    \c0\().4s, v1.s[3]
+        add             v24.4s,    v16.4s,    v22.4s
+        add             v26.4s,    v18.4s,    v22.4s
+        srshr           \c0\().4s, v24.4s,    #14
+        add             v16.4s,    v16.4s,    v18.4s
+        srshr           \c1\().4s, v26.4s,    #14
+        sub             v16.4s,    v16.4s,    v22.4s
+        srshr           \c2\().4s, v20.4s,    #14
+        srshr           \c3\().4s, v16.4s,    #14
+.endm
+
+.macro iadst4_12 c0, c1, c2, c3
+        smull           v16.2d,    \c0\().2s, v1.s[0]
+        smull2          v17.2d,    \c0\().4s, v1.s[0]
+        smlal           v16.2d,    \c2\().2s, v1.s[1]
+        smlal2          v17.2d,    \c2\().4s, v1.s[1]
+        smlal           v16.2d,    \c3\().2s, v1.s[2]
+        smlal2          v17.2d,    \c3\().4s, v1.s[2]
+        smull           v18.2d,    \c0\().2s, v1.s[2]
+        smull2          v19.2d,    \c0\().4s, v1.s[2]
+        smlsl           v18.2d,    \c2\().2s, v1.s[0]
+        smlsl2          v19.2d,    \c2\().4s, v1.s[0]
+        sub             \c0\().4s, \c0\().4s, \c2\().4s
+        smlsl           v18.2d,    \c3\().2s, v1.s[1]
+        smlsl2          v19.2d,    \c3\().4s, v1.s[1]
+        add             \c0\().4s, \c0\().4s, \c3\().4s
+        smull           v22.2d,    \c1\().2s, v1.s[3]
+        smull2          v23.2d,    \c1\().4s, v1.s[3]
+        smull           v20.2d,    \c0\().2s, v1.s[3]
+        smull2          v21.2d,    \c0\().4s, v1.s[3]
+        add             v24.2d,    v16.2d,    v22.2d
+        add             v25.2d,    v17.2d,    v23.2d
+        add             v26.2d,    v18.2d,    v22.2d
+        add             v27.2d,    v19.2d,    v23.2d
+        rshrn           \c0\().2s, v24.2d,    #14
+        rshrn2          \c0\().4s, v25.2d,    #14
+        add             v16.2d,    v16.2d,    v18.2d
+        add             v17.2d,    v17.2d,    v19.2d
+        rshrn           \c1\().2s, v26.2d,    #14
+        rshrn2          \c1\().4s, v27.2d,    #14
+        sub             v16.2d,    v16.2d,    v22.2d
+        sub             v17.2d,    v17.2d,    v23.2d
+        rshrn           \c2\().2s, v20.2d,    #14
+        rshrn2          \c2\().4s, v21.2d,    #14
+        rshrn           \c3\().2s, v16.2d,    #14
+        rshrn2          \c3\().4s, v17.2d,    #14
+.endm
+
+// The public functions in this file have got the following signature:
+// void itxfm_add(uint8_t *dst, ptrdiff_t stride, int16_t *block, int eob);
+
+.macro itxfm_func4x4 txfm1, txfm2, bpp
+function ff_vp9_\txfm1\()_\txfm2\()_4x4_add_\bpp\()_neon, export=1
+.ifc \txfm1,\txfm2
+.ifc \txfm1,idct
+        movrel          x4,  itxfm4_coeffs
+        ld1             {v0.4h}, [x4]
+        sxtl            v0.4s,  v0.4h
+.endif
+.ifc \txfm1,iadst
+        movrel          x4,  iadst4_coeffs
+        ld1             {v0.d}[1], [x4]
+        sxtl2           v1.4s,  v0.8h
+.endif
+.else
+        movrel          x4,  itxfm4_coeffs
+        ld1             {v0.8h}, [x4]
+        sxtl2           v1.4s,  v0.8h
+        sxtl            v0.4s,  v0.4h
+.endif
+
+        movi            v30.4s, #0
+        movi            v31.4s, #0
+.ifc \txfm1\()_\txfm2,idct_idct
+        cmp             w3,  #1
+        b.ne            1f
+        // DC-only for idct/idct
+        ld1             {v2.s}[0],  [x2]
+        smull           v2.2d,  v2.2s, v0.s[0]
+        rshrn           v2.2s,  v2.2d, #14
+        smull           v2.2d,  v2.2s, v0.s[0]
+        rshrn           v2.2s,  v2.2d, #14
+        st1             {v31.s}[0], [x2]
+        dup             v4.4s,  v2.s[0]
+        mov             v5.16b, v4.16b
+        mov             v6.16b, v4.16b
+        mov             v7.16b, v4.16b
+        b               2f
+.endif
+
+1:
+        ld1             {v4.4s,v5.4s,v6.4s,v7.4s},  [x2]
+        st1             {v30.4s,v31.4s}, [x2], #32
+
+.ifc \txfm1,iwht
+        sshr            v4.4s,  v4.4s,  #2
+        sshr            v5.4s,  v5.4s,  #2
+        sshr            v6.4s,  v6.4s,  #2
+        sshr            v7.4s,  v7.4s,  #2
+.endif
+
+        \txfm1\()4_\bpp v4,  v5,  v6,  v7
+
+        st1             {v30.4s,v31.4s}, [x2], #32
+        // Transpose 4x4 with 32 bit elements
+        transpose_4x4s  v4,  v5,  v6,  v7,  v16, v17, v18, v19
+
+        \txfm2\()4_\bpp v4,  v5,  v6,  v7
+2:
+        mvni            v31.8h, #((0xff << (\bpp - 8)) & 0xff), lsl #8
+        ld1             {v0.4h},   [x0], x1
+        ld1             {v1.4h},   [x0], x1
+.ifnc \txfm1,iwht
+        srshr           v4.4s,  v4.4s,  #4
+        srshr           v5.4s,  v5.4s,  #4
+        srshr           v6.4s,  v6.4s,  #4
+        srshr           v7.4s,  v7.4s,  #4
+.endif
+        uaddw           v4.4s,  v4.4s,  v0.4h
+        uaddw           v5.4s,  v5.4s,  v1.4h
+        ld1             {v2.4h},   [x0], x1
+        ld1             {v3.4h},   [x0], x1
+        sqxtun          v0.4h,  v4.4s
+        sqxtun2         v0.8h,  v5.4s
+        sub             x0,  x0,  x1, lsl #2
+
+        uaddw           v6.4s,  v6.4s,  v2.4h
+        umin            v0.8h,  v0.8h,  v31.8h
+        uaddw           v7.4s,  v7.4s,  v3.4h
+        st1             {v0.4h},   [x0], x1
+        sqxtun          v2.4h,  v6.4s
+        sqxtun2         v2.8h,  v7.4s
+        umin            v2.8h,  v2.8h,  v31.8h
+
+        st1             {v0.d}[1], [x0], x1
+        st1             {v2.4h},   [x0], x1
+        st1             {v2.d}[1], [x0], x1
+
+        ret
+endfunc
+.endm
+
+.macro itxfm_funcs4x4 bpp
+itxfm_func4x4 idct,  idct,  \bpp
+itxfm_func4x4 iadst, idct,  \bpp
+itxfm_func4x4 idct,  iadst, \bpp
+itxfm_func4x4 iadst, iadst, \bpp
+itxfm_func4x4 iwht,  iwht,  \bpp
+.endm
+
+itxfm_funcs4x4 10
+itxfm_funcs4x4 12
+
+function idct8x8_dc_add_neon
+        movrel          x4,  idct_coeffs
+        ld1             {v0.4h}, [x4]
+
+        movi            v1.4h,  #0
+        sxtl            v0.4s,  v0.4h
+
+        ld1             {v2.s}[0],  [x2]
+        smull           v2.2d,  v2.2s,  v0.s[0]
+        rshrn           v2.2s,  v2.2d,  #14
+        smull           v2.2d,  v2.2s,  v0.s[0]
+        rshrn           v2.2s,  v2.2d,  #14
+        st1             {v1.s}[0],  [x2]
+        dup             v2.4s,  v2.s[0]
+
+        srshr           v2.4s,  v2.4s,  #5
+
+        mov             x4,  #8
+        mov             x3,  x0
+        dup             v31.8h, w5
+1:
+        // Loop to add the constant from v2 into all 8x8 outputs
+        subs            x4,  x4,  #2
+        ld1             {v3.8h},  [x0], x1
+        ld1             {v4.8h},  [x0], x1
+        uaddw           v16.4s, v2.4s,  v3.4h
+        uaddw2          v17.4s, v2.4s,  v3.8h
+        uaddw           v18.4s, v2.4s,  v4.4h
+        uaddw2          v19.4s, v2.4s,  v4.8h
+        sqxtun          v3.4h,  v16.4s
+        sqxtun2         v3.8h,  v17.4s
+        sqxtun          v4.4h,  v18.4s
+        sqxtun2         v4.8h,  v19.4s
+        umin            v3.8h,  v3.8h,  v31.8h
+        umin            v4.8h,  v4.8h,  v31.8h
+        st1             {v3.8h},  [x3], x1
+        st1             {v4.8h},  [x3], x1
+        b.ne            1b
+
+        ret
+endfunc
+
+.macro idct8 r0, r1, r2, r3, r4, r5, r6, r7, t0, t1, t2, t3, t4, t5
+        dmbutterfly0    \r0, \r4, \r0, \r4, \t0, \t1, \t2, \t3, \t4, \t5 // r0 = t0a, r4 = t1a
+        dmbutterfly     \r2, \r6, v0.s[2], v0.s[3], \t0, \t1, \t2, \t3   // r2 = t2a, r6 = t3a
+        dmbutterfly     \r1, \r7, v1.s[0], v1.s[1], \t0, \t1, \t2, \t3   // r1 = t4a, r7 = t7a
+        dmbutterfly     \r5, \r3, v1.s[2], v1.s[3], \t0, \t1, \t2, \t3   // r5 = t5a, r3 = t6a
+
+        butterfly_4s    \t0, \t1, \r0, \r6 // t0 = t0, t1 = t3
+        butterfly_4s    \t2, \r5, \r1, \r5 // t2 = t4, r5 = t5a
+        butterfly_4s    \t3, \r6, \r7, \r3 // t3 = t7, r6 = t6a
+        butterfly_4s    \r7, \r4, \r4, \r2 // r7 = t1, r4 = t2
+
+        dmbutterfly0    \r6, \r5, \r6, \r5, \r0, \r1, \r2, \r3, \t4, \t5 // r6 = t6, r5 = t5
+
+        butterfly_4s    \r1, \r6, \r7, \r6 // r1 = out[1], r6 = out[6]
+        butterfly_4s    \r0, \r7, \t0, \t3 // r0 = out[0], r7 = out[7]
+        butterfly_4s    \r2, \r5, \r4, \r5 // r2 = out[2], r5 = out[5]
+        butterfly_4s    \r3, \r4, \t1, \t2 // r3 = out[3], r4 = out[4]
+.endm
+
+.macro iadst8 r0, r1, r2, r3, r4, r5, r6, r7, t0, t1, t2, t3, t4, t5
+        dmbutterfly_l   \t2, \t3, \t0, \t1, \r7, \r0, v2.s[1], v2.s[0]   // t2,t3 = t1a, t0,t1 = t0a
+        dmbutterfly_l   \r0, \r7, \t4, \t5, \r3, \r4, v3.s[1], v3.s[0]   // r0,r7 = t5a, t4,t5 = t4a
+
+        dbutterfly_n    \r3, \t0, \t0, \t1, \t4, \t5, \r3, \r4, \t0, \t1 // r3 = t0, t0 = t4
+        dbutterfly_n    \r4, \t1, \t2, \t3, \r0, \r7, \r4, \t1, \t4, \t5 // r4 = t1, t1 = t5
+
+        dmbutterfly_l   \t4, \t5, \t2, \t3, \r5, \r2, v2.s[3], v2.s[2]   // t4,t5 = t3a, t2,t3 = t2a
+        dmbutterfly_l   \r2, \r5, \r0, \r7, \r1, \r6, v3.s[3], v3.s[2]   // r2,r5 = t7a, r0,r7 = t6a
+
+        dbutterfly_n    \r1, \t2, \t2, \t3, \r0, \r7, \r1, \r6, \t2, \t3 // r1 = t2, t2 = t6
+        dbutterfly_n    \r0, \t4, \t4, \t5, \r2, \r5, \r0, \r7, \t4, \t5 // r0 = t3, t4 = t7
+
+        butterfly_4s    \r7, \r4, \r4, \r0   // r7 = -out[7], r4 = t3
+        neg             \r7\().4s, \r7\().4s // r7 = out[7]
+        butterfly_4s    \r0, \r1, \r3, \r1   // r0 = out[0],  r1 = t2
+
+        dmbutterfly_l   \r2, \r3, \t3, \t5, \t0, \t1, v0.s[2], v0.s[3]   // r2,r3 = t5a, t3,t5 = t4a
+        dmbutterfly_l   \t0, \t1, \r5, \r6, \t4, \t2, v0.s[3], v0.s[2]   // t0,t1 = t6a, r5,r6 = t7a
+
+        dbutterfly_n    \r6, \t2, \r2, \r3, \r5, \r6, \t2, \t4, \r2, \r3 // r6 = out[6],  t2 = t7
+
+        dmbutterfly0    \r3, \r4, \r1, \r4, \t4, \r5, \r1, \r2           // r3 = -out[3], r4 = out[4]
+        neg             \r3\().4s, \r3\().4s  // r3 = out[3]
+
+        dbutterfly_n    \r1, \t0, \t3, \t5, \t0, \t1, \r1, \r2, \t0, \t1 // r1 = -out[1], t0 = t6
+        neg             \r1\().4s, \r1\().4s  // r1 = out[1]
+
+        dmbutterfly0    \r2, \r5, \t0, \t2, \t1, \t3, \t4, \t5           // r2 = out[2],  r5 = -out[5]
+        neg             \r5\().4s, \r5\().4s  // r5 = out[5]
+.endm
+
+
+.macro itxfm_func8x8 txfm1, txfm2
+function vp9_\txfm1\()_\txfm2\()_8x8_add_16_neon
+.ifc \txfm1\()_\txfm2,idct_idct
+        cmp             w3,  #1
+        b.eq            idct8x8_dc_add_neon
+.endif
+        // The iadst also uses a few coefficients from
+        // idct, so those always need to be loaded.
+.ifc \txfm1\()_\txfm2,idct_idct
+        movrel          x4,  idct_coeffs
+.else
+        movrel          x4,  iadst8_coeffs
+        ld1             {v1.8h}, [x4], #16
+        stp             d8,  d9,  [sp, #-0x10]!
+        sxtl2           v3.4s,  v1.8h
+        sxtl            v2.4s,  v1.4h
+.endif
+        ld1             {v0.8h}, [x4]
+        sxtl2           v1.4s,  v0.8h
+        sxtl            v0.4s,  v0.4h
+
+        movi            v4.4s, #0
+        movi            v5.4s, #0
+        movi            v6.4s, #0
+        movi            v7.4s, #0
+
+1:
+        ld1             {v16.4s,v17.4s,v18.4s,v19.4s},  [x2], #64
+        ld1             {v20.4s,v21.4s,v22.4s,v23.4s},  [x2], #64
+        ld1             {v24.4s,v25.4s,v26.4s,v27.4s},  [x2], #64
+        ld1             {v28.4s,v29.4s,v30.4s,v31.4s},  [x2], #64
+        sub             x2,  x2,  #256
+        st1             {v4.4s,v5.4s,v6.4s,v7.4s},      [x2], #64
+        st1             {v4.4s,v5.4s,v6.4s,v7.4s},      [x2], #64
+        st1             {v4.4s,v5.4s,v6.4s,v7.4s},      [x2], #64
+        st1             {v4.4s,v5.4s,v6.4s,v7.4s},      [x2], #64
+
+.ifc \txfm1\()_\txfm2,idct_idct
+        idct8           v16, v18, v20, v22, v24, v26, v28, v30, v2,  v3,  v4,  v5,  v6,  v7
+        idct8           v17, v19, v21, v23, v25, v27, v29, v31, v2,  v3,  v4,  v5,  v6,  v7
+.else
+        \txfm1\()8      v16, v18, v20, v22, v24, v26, v28, v30, v4,  v5,  v6,  v7,  v8,  v9
+        \txfm1\()8      v17, v19, v21, v23, v25, v27, v29, v31, v4,  v5,  v6,  v7,  v8,  v9
+.endif
+
+        // Transpose 8x8 with 16 bit elements
+        transpose_8x8s  v16, v17, v18, v19, v20, v21, v22, v23, v24, v25, v26, v27, v28, v29, v30, v31, v4, v5, v6, v7
+
+.ifc \txfm1\()_\txfm2,idct_idct
+        idct8           v16, v18, v20, v22, v24, v26, v28, v30, v2,  v3,  v4,  v5,  v6,  v7
+        idct8           v17, v19, v21, v23, v25, v27, v29, v31, v2,  v3,  v4,  v5,  v6,  v7
+.else
+        \txfm2\()8      v16, v18, v20, v22, v24, v26, v28, v30, v4,  v5,  v6,  v7,  v8,  v9
+        \txfm2\()8      v17, v19, v21, v23, v25, v27, v29, v31, v4,  v5,  v6,  v7,  v8,  v9
+.endif
+2:
+        mov             x3,  x0
+        // Add into the destination
+        ld1             {v0.8h},  [x0], x1
+        srshr           v16.4s, v16.4s, #5
+        srshr           v17.4s, v17.4s, #5
+        ld1             {v1.8h},  [x0], x1
+        srshr           v18.4s, v18.4s, #5
+        srshr           v19.4s, v19.4s, #5
+        ld1             {v2.8h},  [x0], x1
+        srshr           v20.4s, v20.4s, #5
+        srshr           v21.4s, v21.4s, #5
+        uaddw           v16.4s, v16.4s, v0.4h
+        uaddw2          v17.4s, v17.4s, v0.8h
+        ld1             {v3.8h},  [x0], x1
+        srshr           v22.4s, v22.4s, #5
+        srshr           v23.4s, v23.4s, #5
+        uaddw           v18.4s, v18.4s, v1.4h
+        uaddw2          v19.4s, v19.4s, v1.8h
+        ld1             {v4.8h},  [x0], x1
+        srshr           v24.4s, v24.4s, #5
+        srshr           v25.4s, v25.4s, #5
+        uaddw           v20.4s, v20.4s, v2.4h
+        uaddw2          v21.4s, v21.4s, v2.8h
+        sqxtun          v0.4h,  v16.4s
+        sqxtun2         v0.8h,  v17.4s
+        dup             v16.8h, w5
+        ld1             {v5.8h},  [x0], x1
+        srshr           v26.4s, v26.4s, #5
+        srshr           v27.4s, v27.4s, #5
+        uaddw           v22.4s, v22.4s, v3.4h
+        uaddw2          v23.4s, v23.4s, v3.8h
+        sqxtun          v1.4h,  v18.4s
+        sqxtun2         v1.8h,  v19.4s
+        umin            v0.8h,  v0.8h,  v16.8h
+        ld1             {v6.8h},  [x0], x1
+        srshr           v28.4s, v28.4s, #5
+        srshr           v29.4s, v29.4s, #5
+        uaddw           v24.4s, v24.4s, v4.4h
+        uaddw2          v25.4s, v25.4s, v4.8h
+        sqxtun          v2.4h,  v20.4s
+        sqxtun2         v2.8h,  v21.4s
+        umin            v1.8h,  v1.8h,  v16.8h
+        ld1             {v7.8h},  [x0], x1
+        srshr           v30.4s, v30.4s, #5
+        srshr           v31.4s, v31.4s, #5
+        uaddw           v26.4s, v26.4s, v5.4h
+        uaddw2          v27.4s, v27.4s, v5.8h
+        sqxtun          v3.4h,  v22.4s
+        sqxtun2         v3.8h,  v23.4s
+        umin            v2.8h,  v2.8h,  v16.8h
+
+        st1             {v0.8h},  [x3], x1
+        uaddw           v28.4s, v28.4s, v6.4h
+        uaddw2          v29.4s, v29.4s, v6.8h
+        st1             {v1.8h},  [x3], x1
+        sqxtun          v4.4h,  v24.4s
+        sqxtun2         v4.8h,  v25.4s
+        umin            v3.8h,  v3.8h,  v16.8h
+        st1             {v2.8h},  [x3], x1
+        uaddw           v30.4s, v30.4s, v7.4h
+        uaddw2          v31.4s, v31.4s, v7.8h
+        st1             {v3.8h},  [x3], x1
+        sqxtun          v5.4h,  v26.4s
+        sqxtun2         v5.8h,  v27.4s
+        umin            v4.8h,  v4.8h,  v16.8h
+        st1             {v4.8h},  [x3], x1
+        sqxtun          v6.4h,  v28.4s
+        sqxtun2         v6.8h,  v29.4s
+        umin            v5.8h,  v5.8h,  v16.8h
+        st1             {v5.8h},  [x3], x1
+        sqxtun          v7.4h,  v30.4s
+        sqxtun2         v7.8h,  v31.4s
+        umin            v6.8h,  v6.8h,  v16.8h
+
+        st1             {v6.8h},  [x3], x1
+        umin            v7.8h,  v7.8h,  v16.8h
+        st1             {v7.8h},  [x3], x1
+
+.ifnc \txfm1\()_\txfm2,idct_idct
+        ldp             d8,  d9,  [sp], 0x10
+.endif
+        ret
+endfunc
+
+function ff_vp9_\txfm1\()_\txfm2\()_8x8_add_10_neon, export=1
+        mov             x5,  #0x03ff
+        b               vp9_\txfm1\()_\txfm2\()_8x8_add_16_neon
+endfunc
+
+function ff_vp9_\txfm1\()_\txfm2\()_8x8_add_12_neon, export=1
+        mov             x5,  #0x0fff
+        b               vp9_\txfm1\()_\txfm2\()_8x8_add_16_neon
+endfunc
+.endm
+
+itxfm_func8x8 idct,  idct
+itxfm_func8x8 iadst, idct
+itxfm_func8x8 idct,  iadst
+itxfm_func8x8 iadst, iadst
+
+
+function idct16x16_dc_add_neon
+        movrel          x4,  idct_coeffs
+        ld1             {v0.4h}, [x4]
+        sxtl            v0.4s,  v0.4h
+
+        movi            v1.4h,  #0
+
+        ld1             {v2.s}[0],  [x2]
+        smull           v2.2d,  v2.2s,  v0.s[0]
+        rshrn           v2.2s,  v2.2d,  #14
+        smull           v2.2d,  v2.2s,  v0.s[0]
+        rshrn           v2.2s,  v2.2d,  #14
+        st1             {v1.s}[0],  [x2]
+        dup             v2.4s,  v2.s[0]
+
+        srshr           v0.4s,  v2.4s,  #6
+
+        mov             x3, x0
+        mov             x4, #16
+        dup             v31.8h, w13
+1:
+        // Loop to add the constant from v2 into all 16x16 outputs
+        subs            x4,  x4,  #2
+        ld1             {v1.8h,v2.8h},  [x0], x1
+        uaddw           v16.4s, v0.4s,  v1.4h
+        uaddw2          v17.4s, v0.4s,  v1.8h
+        ld1             {v3.8h,v4.8h},  [x0], x1
+        uaddw           v18.4s, v0.4s,  v2.4h
+        uaddw2          v19.4s, v0.4s,  v2.8h
+        uaddw           v20.4s, v0.4s,  v3.4h
+        uaddw2          v21.4s, v0.4s,  v3.8h
+        uaddw           v22.4s, v0.4s,  v4.4h
+        uaddw2          v23.4s, v0.4s,  v4.8h
+        sqxtun          v1.4h,  v16.4s
+        sqxtun2         v1.8h,  v17.4s
+        sqxtun          v2.4h,  v18.4s
+        sqxtun2         v2.8h,  v19.4s
+        sqxtun          v3.4h,  v20.4s
+        sqxtun2         v3.8h,  v21.4s
+        sqxtun          v4.4h,  v22.4s
+        sqxtun2         v4.8h,  v23.4s
+        umin            v1.8h,  v1.8h,  v31.8h
+        umin            v2.8h,  v2.8h,  v31.8h
+        st1             {v1.8h,v2.8h},  [x3], x1
+        umin            v3.8h,  v3.8h,  v31.8h
+        umin            v4.8h,  v4.8h,  v31.8h
+        st1             {v3.8h,v4.8h},  [x3], x1
+        b.ne            1b
+
+        ret
+endfunc
+
+.macro idct16
+        dmbutterfly0    v16, v24, v16, v24, v4, v5, v6, v7, v8, v9 // v16 = t0a,  v24 = t1a
+        dmbutterfly     v20, v28, v0.s[2], v0.s[3], v4, v5, v6, v7 // v20 = t2a,  v28 = t3a
+        dmbutterfly     v18, v30, v1.s[0], v1.s[1], v4, v5, v6, v7 // v18 = t4a,  v30 = t7a
+        dmbutterfly     v26, v22, v1.s[2], v1.s[3], v4, v5, v6, v7 // v26 = t5a,  v22 = t6a
+        dmbutterfly     v17, v31, v2.s[0], v2.s[1], v4, v5, v6, v7 // v17 = t8a,  v31 = t15a
+        dmbutterfly     v25, v23, v2.s[2], v2.s[3], v4, v5, v6, v7 // v25 = t9a,  v23 = t14a
+        dmbutterfly     v21, v27, v3.s[0], v3.s[1], v4, v5, v6, v7 // v21 = t10a, v27 = t13a
+        dmbutterfly     v29, v19, v3.s[2], v3.s[3], v4, v5, v6, v7 // v29 = t11a, v19 = t12a
+
+        butterfly_4s    v4,  v28, v16, v28               // v4  = t0,   v28 = t3
+        butterfly_4s    v5,  v20, v24, v20               // v5  = t1,   v20 = t2
+        butterfly_4s    v6,  v26, v18, v26               // v6  = t4,   v26 = t5
+        butterfly_4s    v7,  v22, v30, v22               // v7  = t7,   v22 = t6
+        butterfly_4s    v16, v25, v17, v25               // v16 = t8,   v25 = t9
+        butterfly_4s    v24, v21, v29, v21               // v24 = t11,  v21 = t10
+        butterfly_4s    v17, v27, v19, v27               // v17 = t12,  v27 = t13
+        butterfly_4s    v29, v23, v31, v23               // v29 = t15,  v23 = t14
+
+        dmbutterfly0    v22, v26, v22, v26, v8, v9, v18, v19, v30, v31        // v22 = t6a,  v26 = t5a
+        dmbutterfly     v23, v25, v0.s[2], v0.s[3], v18, v19, v30, v31        // v23 = t9a,  v25 = t14a
+        dmbutterfly     v27, v21, v0.s[2], v0.s[3], v18, v19, v30, v31, neg=1 // v27 = t13a, v21 = t10a
+
+        butterfly_4s    v18, v7,  v4,  v7                // v18 = t0a,  v7  = t7a
+        butterfly_4s    v19, v22, v5,  v22               // v19 = t1a,  v22 = t6
+        butterfly_4s    v4,  v26, v20, v26               // v4  = t2a,  v26 = t5
+        butterfly_4s    v5,  v6,  v28, v6                // v5  = t3a,  v6  = t4
+        butterfly_4s    v20, v28, v16, v24               // v20 = t8a,  v28 = t11a
+        butterfly_4s    v24, v21, v23, v21               // v24 = t9,   v21 = t10
+        butterfly_4s    v23, v27, v25, v27               // v23 = t14,  v27 = t13
+        butterfly_4s    v25, v29, v29, v17               // v25 = t15a, v29 = t12a
+
+        dmbutterfly0    v8,  v9,  v27, v21, v8,  v9,  v16, v17, v30, v31 // v8  = t13a, v9  = t10a
+        dmbutterfly0    v28, v27, v29, v28, v21, v29, v16, v17, v30, v31 // v28 = t12,  v27 = t11
+
+        butterfly_4s    v16, v31, v18, v25               // v16 = out[0], v31 = out[15]
+        butterfly_4s    v17, v30, v19, v23               // v17 = out[1], v30 = out[14]
+        butterfly_4s_r  v25, v22, v22, v24               // v25 = out[9], v22 = out[6]
+        butterfly_4s    v23, v24, v7,  v20               // v23 = out[7], v24 = out[8]
+        butterfly_4s    v18, v29, v4,  v8                // v18 = out[2], v29 = out[13]
+        butterfly_4s    v19, v28, v5,  v28               // v19 = out[3], v28 = out[12]
+        butterfly_4s    v20, v27, v6,  v27               // v20 = out[4], v27 = out[11]
+        butterfly_4s    v21, v26, v26, v9                // v21 = out[5], v26 = out[10]
+.endm
+
+.macro iadst16
+        ld1             {v0.8h,v1.8h}, [x11]
+        sxtl            v2.4s,  v1.4h
+        sxtl2           v3.4s,  v1.8h
+        sxtl2           v1.4s,  v0.8h
+        sxtl            v0.4s,  v0.4h
+
+        dmbutterfly_l   v6,  v7,  v4,  v5,  v31, v16, v0.s[1], v0.s[0]   // v6,v7   = t1,   v4,v5   = t0
+        dmbutterfly_l   v10, v11, v8,  v9,  v23, v24, v1.s[1], v1.s[0]   // v10,v11 = t9,   v8,v9   = t8
+        dbutterfly_n    v31, v24, v6,  v7,  v10, v11, v12, v13, v10, v11 // v31     = t1a,  v24     = t9a
+        dmbutterfly_l   v14, v15, v12, v13, v29, v18, v0.s[3], v0.s[2]   // v14,v15 = t3,   v12,v13 = t2
+        dbutterfly_n    v16, v23, v4,  v5,  v8,  v9,  v6,  v7,  v8,  v9  // v16     = t0a,  v23     = t8a
+
+        dmbutterfly_l   v6,  v7,  v4,  v5,  v21, v26, v1.s[3], v1.s[2]   // v6,v7   = t11,  v4,v5   = t10
+        dbutterfly_n    v29, v26, v14, v15, v6,  v7,  v8,  v9,  v6,  v7  // v29     = t3a,  v26     = t11a
+        dmbutterfly_l   v10, v11, v8,  v9,  v27, v20, v2.s[1], v2.s[0]   // v10,v11 = t5,   v8,v9   = t4
+        dbutterfly_n    v18, v21, v12, v13, v4,  v5,  v6,  v7,  v4,  v5  // v18     = t2a,  v21     = t10a
+
+        dmbutterfly_l   v14, v15, v12, v13, v19, v28, v3.s[1], v3.s[0]   // v14,v15 = t13,  v12,v13 = t12
+        dbutterfly_n    v20, v28, v10, v11, v14, v15, v4,  v5,  v14, v15 // v20     = t5a,  v28     = t13a
+        dmbutterfly_l   v6,  v7,  v4,  v5,  v25, v22, v2.s[3], v2.s[2]   // v6,v7   = t7,   v4,v5   = t6
+        dbutterfly_n    v27, v19, v8,  v9,  v12, v13, v10, v11, v12, v13 // v27     = t4a,  v19     = t12a
+
+        dmbutterfly_l   v10, v11, v8,  v9,  v17, v30, v3.s[3], v3.s[2]   // v10,v11 = t15,  v8,v9   = t14
+        ld1             {v0.8h}, [x10]
+        dbutterfly_n    v22, v30, v6,  v7,  v10, v11, v12, v13, v10, v11 // v22     = t7a,  v30     = t15a
+        sxtl2           v1.4s,  v0.8h
+        sxtl            v0.4s,  v0.4h
+        dmbutterfly_l   v14, v15, v12, v13, v23, v24, v1.s[0], v1.s[1]   // v14,v15 = t9,   v12,v13 = t8
+        dbutterfly_n    v25, v17, v4,  v5,  v8,  v9,  v6,  v7,  v8,  v9  // v25     = t6a,  v17     = t14a
+
+        dmbutterfly_l   v4,  v5,  v6,  v7,  v28, v19, v1.s[1], v1.s[0]   // v4,v5   = t12,  v6,v7   = t13
+        dbutterfly_n    v23, v19, v12, v13, v4,  v5,  v8,  v9,  v4,  v5  // v23     = t8a,  v19     = t12a
+        dmbutterfly_l   v10, v11, v8,  v9,  v21, v26, v1.s[2], v1.s[3]   // v10,v11 = t11,  v8,v9   = t10
+        butterfly_4s_r  v4,  v27, v16, v27               // v4  = t4,   v27 = t0
+        dbutterfly_n    v24, v28, v14, v15, v6,  v7,  v12, v13, v6,  v7  // v24     = t9a,  v28     = t13a
+
+        dmbutterfly_l   v12, v13, v14, v15, v30, v17, v1.s[3], v1.s[2]   // v12,v13 = t14,  v14,v15 = t15
+        butterfly_4s_r  v5,  v20, v31, v20               // v5  = t5, v20 = t1
+        dbutterfly_n    v21, v17, v8,  v9,  v12, v13, v6,  v7,  v12, v13 // v21     = t10a, v17     = t14a
+        dbutterfly_n    v26, v30, v10, v11, v14, v15, v8,  v9,  v14, v15 // v26     = t11a, v30     = t15a
+
+        butterfly_4s_r  v6,  v25, v18, v25               // v6  = t6, v25 = t2
+        butterfly_4s_r  v7,  v22, v29, v22               // v7  = t7, v22 = t3
+
+        dmbutterfly_l   v10, v11, v8,  v9,  v19, v28, v0.s[2], v0.s[3]   // v10,v11 = t13,  v8,v9   = t12
+        dmbutterfly_l   v12, v13, v14, v15, v30, v17, v0.s[3], v0.s[2]   // v12,v13 = t14,  v14,v15 = t15
+
+        dbutterfly_n    v18, v30, v8,  v9,  v12, v13, v16, v17, v12, v13 // v18   = out[2], v30     = t14a
+        dbutterfly_n    v29, v17, v10, v11, v14, v15, v12, v13, v14, v15 // v29 = -out[13], v17     = t15a
+        neg             v29.4s, v29.4s                   // v29 = out[13]
+
+        dmbutterfly_l   v10, v11, v8,  v9,  v4,  v5,  v0.s[2], v0.s[3]   // v10,v11 = t5a,  v8,v9   = t4a
+        dmbutterfly_l   v12, v13, v14, v15, v7,  v6,  v0.s[3], v0.s[2]   // v12,v13 = t6a,  v14,v15 = t7a
+
+        butterfly_4s    v2,  v6,  v27, v25               // v2 = out[0], v6 = t2a
+        butterfly_4s    v3,  v7,  v23, v21               // v3 =-out[1], v7 = t10
+
+        dbutterfly_n    v19, v31, v8,  v9,  v12, v13, v4,  v5,  v8,  v9  // v19 = -out[3],  v31 = t6
+        neg             v19.4s, v19.4s                   // v19 = out[3]
+        dbutterfly_n    v28, v16, v10, v11, v14, v15, v4,  v5,  v10, v11 // v28 = out[12],  v16 = t7
+
+        butterfly_4s    v5,  v8,  v20, v22               // v5 =-out[15],v8 = t3a
+        butterfly_4s    v4,  v9,  v24, v26               // v4 = out[14],v9 = t11
+
+        dmbutterfly0    v23, v24, v6,  v8,  v10, v11, v12, v13, v14, v15, 1 // v23 = out[7], v24 = out[8]
+        dmbutterfly0    v21, v26, v30, v17, v10, v11, v12, v13, v14, v15, 1 // v21 = out[5], v26 = out[10]
+        dmbutterfly0    v20, v27, v16, v31, v10, v11, v12, v13, v14, v15    // v20 = out[4], v27 = out[11]
+        dmbutterfly0    v22, v25, v9,  v7,  v10, v11, v12, v13, v14, v15    // v22 = out[6], v25 = out[9]
+
+        neg             v31.4s,  v5.4s                    // v31 = out[15]
+        neg             v17.4s,  v3.4s                    // v17 = out[1]
+
+        mov             v16.16b, v2.16b
+        mov             v30.16b, v4.16b
+.endm
+
+// Helper macros; we can't use these expressions directly within
+// e.g. .irp due to the extra concatenation \(). Therefore wrap
+// them in macros to allow using .irp below.
+.macro load i, src, inc
+        ld1             {v\i\().4s},  [\src], \inc
+.endm
+.macro store i, dst, inc
+        st1             {v\i\().4s},  [\dst], \inc
+.endm
+.macro movi_v i, size, imm
+        movi            v\i\()\size,  \imm
+.endm
+.macro load_clear i, src, inc
+        ld1             {v\i\().4s}, [\src]
+        st1             {v4.4s},  [\src], \inc
+.endm
+
+// Read a vertical 4x16 slice out of a 16x16 matrix, do a transform on it,
+// transpose into a horizontal 16x4 slice and store.
+// x0 = dst (temp buffer)
+// x1 = slice offset
+// x2 = src
+// x9 = input stride
+.macro itxfm16_1d_funcs txfm
+function \txfm\()16_1d_4x16_pass1_neon
+        movi            v4.4s, #0
+.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
+        load_clear      \i,  x2,  x9
+.endr
+
+        \txfm\()16
+
+        // Do four 4x4 transposes. Originally, v16-v31 contain the
+        // 16 rows. Afterwards, v16-v19, v20-v23, v24-v27 and v28-v31
+        // contain the four transposed 4x4 blocks.
+        transpose_4x4s  v16, v17, v18, v19, v4, v5, v6, v7
+        transpose_4x4s  v20, v21, v22, v23, v4, v5, v6, v7
+        transpose_4x4s  v24, v25, v26, v27, v4, v5, v6, v7
+        transpose_4x4s  v28, v29, v30, v31, v4, v5, v6, v7
+
+        // Store the transposed 8x8 blocks horizontally.
+        cmp             x1,  #12
+        b.eq            1f
+.irp i, 16, 20, 24, 28, 17, 21, 25, 29, 18, 22, 26, 30, 19, 23, 27, 31
+        store           \i,  x0,  #16
+.endr
+        ret
+1:
+        // Special case: For the last input column (x1 == 12),
+        // which would be stored as the last row in the temp buffer,
+        // don't store the first 4x4 block, but keep it in registers
+        // for the first slice of the second pass (where it is the
+        // last 4x4 block).
+        add             x0,  x0,  #16
+.irp i, 20, 24, 28
+        store           \i,  x0,  #16
+.endr
+        add             x0,  x0,  #16
+.irp i, 21, 25, 29
+        store           \i,  x0,  #16
+.endr
+        add             x0,  x0,  #16
+.irp i, 22, 26, 30
+        store           \i,  x0,  #16
+.endr
+        add             x0,  x0,  #16
+.irp i, 23, 27, 31
+        store           \i,  x0,  #16
+.endr
+
+        mov             v28.16b, v16.16b
+        mov             v29.16b, v17.16b
+        mov             v30.16b, v18.16b
+        mov             v31.16b, v19.16b
+        ret
+endfunc
+
+// Read a vertical 4x16 slice out of a 16x16 matrix, do a transform on it,
+// load the destination pixels (from a similar 4x16 slice), add and store back.
+// x0 = dst
+// x1 = dst stride
+// x2 = src (temp buffer)
+// x3 = slice offset
+// x9 = temp buffer stride
+function \txfm\()16_1d_4x16_pass2_neon
+.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27
+        load            \i,  x2,  x9
+.endr
+        cbz             x3,  1f
+.irp i, 28, 29, 30, 31
+        load            \i,  x2,  x9
+.endr
+1:
+
+        add             x3,  x0,  x1
+        lsl             x1,  x1,  #1
+        \txfm\()16
+
+        dup             v8.8h, w13
+.macro load_add_store coef0, coef1, coef2, coef3, coef4, coef5, coef6, coef7
+        srshr           \coef0, \coef0, #6
+        ld1             {v4.4h},   [x0], x1
+        srshr           \coef1, \coef1, #6
+        ld1             {v4.d}[1], [x3], x1
+        srshr           \coef2, \coef2, #6
+        ld1             {v5.4h},   [x0], x1
+        srshr           \coef3, \coef3, #6
+        uaddw           \coef0, \coef0, v4.4h
+        ld1             {v5.d}[1], [x3], x1
+        srshr           \coef4, \coef4, #6
+        uaddw2          \coef1, \coef1, v4.8h
+        ld1             {v6.4h},   [x0], x1
+        srshr           \coef5, \coef5, #6
+        uaddw           \coef2, \coef2, v5.4h
+        ld1             {v6.d}[1], [x3], x1
+        sqxtun          v4.4h,  \coef0
+        srshr           \coef6, \coef6, #6
+        uaddw2          \coef3, \coef3, v5.8h
+        ld1             {v7.4h},   [x0], x1
+        sqxtun2         v4.8h,  \coef1
+        srshr           \coef7, \coef7, #6
+        uaddw           \coef4, \coef4, v6.4h
+        ld1             {v7.d}[1], [x3], x1
+        umin            v4.8h,  v4.8h,  v8.8h
+        sub             x0,  x0,  x1, lsl #2
+        sub             x3,  x3,  x1, lsl #2
+        sqxtun          v5.4h,  \coef2
+        uaddw2          \coef5, \coef5, v6.8h
+        st1             {v4.4h},   [x0], x1
+        sqxtun2         v5.8h,  \coef3
+        uaddw           \coef6, \coef6, v7.4h
+        st1             {v4.d}[1], [x3], x1
+        umin            v5.8h,  v5.8h,  v8.8h
+        sqxtun          v6.4h,  \coef4
+        uaddw2          \coef7, \coef7, v7.8h
+        st1             {v5.4h},   [x0], x1
+        sqxtun2         v6.8h,  \coef5
+        st1             {v5.d}[1], [x3], x1
+        umin            v6.8h,  v6.8h,  v8.8h
+        sqxtun          v7.4h,  \coef6
+        st1             {v6.4h},   [x0], x1
+        sqxtun2         v7.8h,  \coef7
+        st1             {v6.d}[1], [x3], x1
+        umin            v7.8h,  v7.8h,  v8.8h
+        st1             {v7.4h},   [x0], x1
+        st1             {v7.d}[1], [x3], x1
+.endm
+        load_add_store  v16.4s, v17.4s, v18.4s, v19.4s, v20.4s, v21.4s, v22.4s, v23.4s
+        load_add_store  v24.4s, v25.4s, v26.4s, v27.4s, v28.4s, v29.4s, v30.4s, v31.4s
+.purgem load_add_store
+
+        ret
+endfunc
+.endm
+
+itxfm16_1d_funcs idct
+itxfm16_1d_funcs iadst
+
+// This is the minimum eob value for each subpartition, in increments of 4
+const min_eob_idct_idct_16, align=4
+        .short  0, 10, 38, 89
+endconst
+
+.macro itxfm_func16x16 txfm1, txfm2
+function vp9_\txfm1\()_\txfm2\()_16x16_add_16_neon
+.ifc \txfm1\()_\txfm2,idct_idct
+        cmp             w3,  #1
+        b.eq            idct16x16_dc_add_neon
+.endif
+        mov             x15, x30
+        // iadst16 requires clobbering v8-v15, idct16 only clobbers v8-v9.
+.ifnc \txfm1\()_\txfm2,idct_idct
+        stp             d14, d15, [sp, #-0x10]!
+        stp             d12, d13, [sp, #-0x10]!
+        stp             d10, d11, [sp, #-0x10]!
+.endif
+        stp             d8,  d9,  [sp, #-0x10]!
+
+        sub             sp,  sp,  #1024
+
+        mov             x4,  x0
+        mov             x5,  x1
+        mov             x6,  x2
+
+        movrel          x10, idct_coeffs
+.ifnc \txfm1\()_\txfm2,idct_idct
+        movrel          x11, iadst16_coeffs
+.endif
+        movrel          x12, min_eob_idct_idct_16, 2
+.ifc \txfm1,idct
+        ld1             {v0.8h,v1.8h}, [x10]
+        sxtl            v2.4s,  v1.4h
+        sxtl2           v3.4s,  v1.8h
+        sxtl2           v1.4s,  v0.8h
+        sxtl            v0.4s,  v0.4h
+.endif
+        mov             x9,  #64
+
+.irp i, 0, 4, 8, 12
+        add             x0,  sp,  #(\i*64)
+.ifc \txfm1\()_\txfm2,idct_idct
+.if \i > 0
+        ldrh            w1,  [x12], #2
+        cmp             w3,  w1
+        mov             x1,  #(16 - \i)/4
+        b.le            1f
+.endif
+.endif
+        mov             x1,  #\i
+        add             x2,  x6,  #(\i*4)
+        bl              \txfm1\()16_1d_4x16_pass1_neon
+.endr
+.ifc \txfm1\()_\txfm2,iadst_idct
+        ld1             {v0.8h,v1.8h}, [x10]
+        sxtl            v2.4s,  v1.4h
+        sxtl2           v3.4s,  v1.8h
+        sxtl2           v1.4s,  v0.8h
+        sxtl            v0.4s,  v0.4h
+.endif
+
+.ifc \txfm1\()_\txfm2,idct_idct
+        b               3f
+1:
+        // Set v28-v31 to zero, for the in-register passthrough of
+        // coefficients to pass 2.
+        movi            v28.4s,  #0
+        movi            v29.4s,  #0
+        movi            v30.4s,  #0
+        movi            v31.4s,  #0
+2:
+        subs            x1,  x1,  #1
+.rept 4
+        st1             {v28.4s,v29.4s,v30.4s,v31.4s}, [x0], x9
+.endr
+        b.ne            2b
+3:
+.endif
+
+.irp i, 0, 4, 8, 12
+        add             x0,  x4,  #(\i*2)
+        mov             x1,  x5
+        add             x2,  sp,  #(\i*4)
+        mov             x3,  #\i
+        bl              \txfm2\()16_1d_4x16_pass2_neon
+.endr
+
+        add             sp,  sp,  #1024
+        ldp             d8,  d9,  [sp], 0x10
+.ifnc \txfm1\()_\txfm2,idct_idct
+        ldp             d10, d11, [sp], 0x10
+        ldp             d12, d13, [sp], 0x10
+        ldp             d14, d15, [sp], 0x10
+.endif
+        br              x15
+endfunc
+
+function ff_vp9_\txfm1\()_\txfm2\()_16x16_add_10_neon, export=1
+        mov             x13, #0x03ff
+        b               vp9_\txfm1\()_\txfm2\()_16x16_add_16_neon
+endfunc
+
+function ff_vp9_\txfm1\()_\txfm2\()_16x16_add_12_neon, export=1
+        mov             x13, #0x0fff
+        b               vp9_\txfm1\()_\txfm2\()_16x16_add_16_neon
+endfunc
+.endm
+
+itxfm_func16x16 idct,  idct
+itxfm_func16x16 iadst, idct
+itxfm_func16x16 idct,  iadst
+itxfm_func16x16 iadst, iadst
+
+
+function idct32x32_dc_add_neon
+        movrel          x4,  idct_coeffs
+        ld1             {v0.4h}, [x4]
+        sxtl            v0.4s,  v0.4h
+
+        movi            v1.4h,  #0
+
+        ld1             {v2.s}[0],  [x2]
+        smull           v2.2d,  v2.2s,  v0.s[0]
+        rshrn           v2.2s,  v2.2d,  #14
+        smull           v2.2d,  v2.2s,  v0.s[0]
+        rshrn           v2.2s,  v2.2d,  #14
+        st1             {v1.s}[0],  [x2]
+        dup             v2.4s,  v2.s[0]
+
+        srshr           v0.4s,  v2.4s,  #6
+
+        mov             x3,  x0
+        mov             x4,  #32
+        sub             x1,  x1,  #32
+        dup             v31.8h, w13
+1:
+        // Loop to add the constant v0 into all 32x32 outputs
+        subs            x4,  x4,  #1
+        ld1             {v1.8h,v2.8h},  [x0], #32
+        uaddw           v16.4s, v0.4s,  v1.4h
+        uaddw2          v17.4s, v0.4s,  v1.8h
+        ld1             {v3.8h,v4.8h},  [x0], x1
+        uaddw           v18.4s, v0.4s,  v2.4h
+        uaddw2          v19.4s, v0.4s,  v2.8h
+        uaddw           v20.4s, v0.4s,  v3.4h
+        uaddw2          v21.4s, v0.4s,  v3.8h
+        uaddw           v22.4s, v0.4s,  v4.4h
+        uaddw2          v23.4s, v0.4s,  v4.8h
+        sqxtun          v1.4h,  v16.4s
+        sqxtun2         v1.8h,  v17.4s
+        sqxtun          v2.4h,  v18.4s
+        sqxtun2         v2.8h,  v19.4s
+        sqxtun          v3.4h,  v20.4s
+        sqxtun2         v3.8h,  v21.4s
+        sqxtun          v4.4h,  v22.4s
+        sqxtun2         v4.8h,  v23.4s
+        umin            v1.8h,  v1.8h,  v31.8h
+        umin            v2.8h,  v2.8h,  v31.8h
+        st1             {v1.8h,v2.8h},  [x3], #32
+        umin            v3.8h,  v3.8h,  v31.8h
+        umin            v4.8h,  v4.8h,  v31.8h
+        st1             {v3.8h,v4.8h},  [x3], x1
+        b.ne            1b
+
+        ret
+endfunc
+
+.macro idct32_odd
+        dmbutterfly     v16, v31, v10.s[0], v10.s[1], v4, v5, v6, v7 // v16 = t16a, v31 = t31a
+        dmbutterfly     v24, v23, v10.s[2], v10.s[3], v4, v5, v6, v7 // v24 = t17a, v23 = t30a
+        dmbutterfly     v20, v27, v11.s[0], v11.s[1], v4, v5, v6, v7 // v20 = t18a, v27 = t29a
+        dmbutterfly     v28, v19, v11.s[2], v11.s[3], v4, v5, v6, v7 // v28 = t19a, v19 = t28a
+        dmbutterfly     v18, v29, v12.s[0], v12.s[1], v4, v5, v6, v7 // v18 = t20a, v29 = t27a
+        dmbutterfly     v26, v21, v12.s[2], v12.s[3], v4, v5, v6, v7 // v26 = t21a, v21 = t26a
+        dmbutterfly     v22, v25, v13.s[0], v13.s[1], v4, v5, v6, v7 // v22 = t22a, v25 = t25a
+        dmbutterfly     v30, v17, v13.s[2], v13.s[3], v4, v5, v6, v7 // v30 = t23a, v17 = t24a
+
+        butterfly_4s    v4,  v24, v16, v24 // v4  = t16, v24 = t17
+        butterfly_4s    v5,  v20, v28, v20 // v5  = t19, v20 = t18
+        butterfly_4s    v6,  v26, v18, v26 // v6  = t20, v26 = t21
+        butterfly_4s    v7,  v22, v30, v22 // v7  = t23, v22 = t22
+        butterfly_4s    v28, v25, v17, v25 // v28 = t24, v25 = t25
+        butterfly_4s    v30, v21, v29, v21 // v30 = t27, v21 = t26
+        butterfly_4s    v29, v23, v31, v23 // v29 = t31, v23 = t30
+        butterfly_4s    v31, v27, v19, v27 // v31 = t28, v27 = t29
+
+        dmbutterfly     v23, v24, v1.s[0], v1.s[1], v16, v17, v18, v19        // v23 = t17a, v24 = t30a
+        dmbutterfly     v27, v20, v1.s[0], v1.s[1], v16, v17, v18, v19, neg=1 // v27 = t29a, v20 = t18a
+        dmbutterfly     v21, v26, v1.s[2], v1.s[3], v16, v17, v18, v19        // v21 = t21a, v26 = t26a
+        dmbutterfly     v25, v22, v1.s[2], v1.s[3], v16, v17, v18, v19, neg=1 // v25 = t25a, v22 = t22a
+
+        butterfly_4s    v16, v5,  v4,  v5  // v16 = t16a, v5  = t19a
+        butterfly_4s    v17, v20, v23, v20 // v17 = t17,  v20 = t18
+        butterfly_4s    v18, v6,  v7,  v6  // v18 = t23a, v6  = t20a
+        butterfly_4s    v19, v21, v22, v21 // v19 = t22,  v21 = t21
+        butterfly_4s    v4,  v28, v28, v30 // v4  = t24a, v28 = t27a
+        butterfly_4s    v23, v26, v25, v26 // v23 = t25,  v26 = t26
+        butterfly_4s    v7,  v8,  v29, v31 // v7  = t31a, v3  = t28a
+        butterfly_4s    v22, v27, v24, v27 // v22 = t30,  v27 = t29
+
+        dmbutterfly     v27, v20, v0.s[2], v0.s[3], v24, v25, v30, v31        // v27 = t18a, v20 = t29a
+        dmbutterfly     v8,  v5,  v0.s[2], v0.s[3], v24, v25, v30, v31        // v3  = t19,  v5  = t28
+        dmbutterfly     v28, v6,  v0.s[2], v0.s[3], v24, v25, v30, v31, neg=1 // v28 = t27,  v6  = t20
+        dmbutterfly     v26, v21, v0.s[2], v0.s[3], v24, v25, v30, v31, neg=1 // v26 = t26a, v21 = t21a
+
+        butterfly_4s    v31, v24, v7,  v4  // v31 = t31,  v24 = t24
+        butterfly_4s    v30, v25, v22, v23 // v30 = t30a, v25 = t25a
+        butterfly_4s_r  v23, v16, v16, v18 // v23 = t23,  v16 = t16
+        butterfly_4s_r  v22, v17, v17, v19 // v22 = t22a, v17 = t17a
+        butterfly_4s    v18, v21, v27, v21 // v18 = t18,  v21 = t21
+        butterfly_4s_r  v27, v28, v5,  v28 // v27 = t27a, v28 = t28a
+        butterfly_4s    v29, v26, v20, v26 // v29 = t29,  v26 = t26
+        butterfly_4s    v19, v20, v8,  v6  // v19 = t19a, v20 = t20
+
+        dmbutterfly0    v27, v20, v27, v20, v4, v5, v6, v7, v8, v9 // v27 = t27,  v20 = t20
+        dmbutterfly0    v26, v21, v26, v21, v4, v5, v6, v7, v8, v9 // v26 = t26a, v21 = t21a
+        dmbutterfly0    v25, v22, v25, v22, v4, v5, v6, v7, v8, v9 // v25 = t25,  v22 = t22
+        dmbutterfly0    v24, v23, v24, v23, v4, v5, v6, v7, v8, v9 // v24 = t24a, v23 = t23a
+.endm
+
+// Do an 32-point IDCT of a 4x32 slice out of a 32x32 matrix.
+// The 32-point IDCT can be decomposed into two 16-point IDCTs;
+// a normal IDCT16 with every other input component (the even ones, with
+// each output written twice), followed by a separate 16-point IDCT
+// of the odd inputs, added/subtracted onto the outputs of the first idct16.
+// x0 = dst (temp buffer)
+// x1 = unused
+// x2 = src
+// x9 = double input stride
+function idct32_1d_4x32_pass1_neon
+        movi            v4.4s,  #0
+
+        // v16 = IN(0), v17 = IN(2) ... v31 = IN(30)
+.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
+        ld1             {v\i\().4s}, [x2]
+        st1             {v4.4s},  [x2], x9
+.endr
+
+        idct16
+
+        // Do four 4x4 transposes. Originally, v16-v31 contain the
+        // 16 rows. Afterwards, v16-v19, v20-v23, v24-v27 and v28-v31
+        // contain the four transposed 4x4 blocks.
+        transpose_4x4s  v16, v17, v18, v19, v4, v5, v6, v7
+        transpose_4x4s  v20, v21, v22, v23, v4, v5, v6, v7
+        transpose_4x4s  v24, v25, v26, v27, v4, v5, v6, v7
+        transpose_4x4s  v28, v29, v30, v31, v4, v5, v6, v7
+
+        // Store the registers a, b, c, d horizontally, followed by the
+        // same registers d, c, b, a mirrored.
+.macro store_rev a, b, c, d
+        // There's no rev128 instruction, but we reverse each 64 bit
+        // half, and then flip them using an ext with 8 bytes offset.
+        rev64           v7.4s, v\d\().4s
+        st1             {v\a\().4s},  [x0], #16
+        ext             v7.16b, v7.16b, v7.16b, #8
+        st1             {v\b\().4s},  [x0], #16
+        rev64           v6.4s, v\c\().4s
+        st1             {v\c\().4s},  [x0], #16
+        ext             v6.16b, v6.16b, v6.16b, #8
+        st1             {v\d\().4s},  [x0], #16
+        rev64           v5.4s, v\b\().4s
+        st1             {v7.4s},  [x0], #16
+        ext             v5.16b, v5.16b, v5.16b, #8
+        st1             {v6.4s},  [x0], #16
+        rev64           v4.4s, v\a\().4s
+        st1             {v5.4s},  [x0], #16
+        ext             v4.16b, v4.16b, v4.16b, #8
+        st1             {v4.4s},  [x0], #16
+.endm
+        store_rev       16, 20, 24, 28
+        store_rev       17, 21, 25, 29
+        store_rev       18, 22, 26, 30
+        store_rev       19, 23, 27, 31
+        sub             x0,  x0,  #512
+.purgem store_rev
+
+        // Move x2 back to the start of the input, and move
+        // to the first odd row
+        sub             x2,  x2,  x9, lsl #4
+        add             x2,  x2,  #128
+
+        movi            v4.4s,  #0
+        // v16 = IN(1), v17 = IN(3) ... v31 = IN(31)
+.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
+        ld1             {v\i\().4s}, [x2]
+        st1             {v4.4s},  [x2], x9
+.endr
+
+        idct32_odd
+
+        transpose_4x4s  v31, v30, v29, v28, v4, v5, v6, v7
+        transpose_4x4s  v27, v26, v25, v24, v4, v5, v6, v7
+        transpose_4x4s  v23, v22, v21, v20, v4, v5, v6, v7
+        transpose_4x4s  v19, v18, v17, v16, v4, v5, v6, v7
+
+        // Store the registers a, b, c, d horizontally,
+        // adding into the output first, and the mirrored,
+        // subtracted from the output.
+.macro store_rev a, b, c, d
+        ld1             {v4.4s},  [x0]
+        rev64           v9.4s, v\d\().4s
+        add             v4.4s, v4.4s, v\a\().4s
+        st1             {v4.4s},  [x0], #16
+        rev64           v8.4s, v\c\().4s
+        ld1             {v4.4s},  [x0]
+        ext             v9.16b, v9.16b, v9.16b, #8
+        add             v4.4s, v4.4s, v\b\().4s
+        st1             {v4.4s},  [x0], #16
+        ext             v8.16b, v8.16b, v8.16b, #8
+        ld1             {v4.4s},  [x0]
+        rev64           v\b\().4s, v\b\().4s
+        add             v4.4s, v4.4s, v\c\().4s
+        st1             {v4.4s},  [x0], #16
+        rev64           v\a\().4s, v\a\().4s
+        ld1             {v4.4s},  [x0]
+        ext             v\b\().16b, v\b\().16b, v\b\().16b, #8
+        add             v4.4s, v4.4s, v\d\().4s
+        st1             {v4.4s},  [x0], #16
+        ext             v\a\().16b, v\a\().16b, v\a\().16b, #8
+        ld1             {v4.4s},  [x0]
+        sub             v4.4s, v4.4s, v9.4s
+        st1             {v4.4s},  [x0], #16
+        ld1             {v4.4s},  [x0]
+        sub             v4.4s, v4.4s, v8.4s
+        st1             {v4.4s},  [x0], #16
+        ld1             {v4.4s},  [x0]
+        sub             v4.4s, v4.4s, v\b\().4s
+        st1             {v4.4s},  [x0], #16
+        ld1             {v4.4s},  [x0]
+        sub             v4.4s, v4.4s, v\a\().4s
+        st1             {v4.4s},  [x0], #16
+.endm
+
+        store_rev       31, 27, 23, 19
+        store_rev       30, 26, 22, 18
+        store_rev       29, 25, 21, 17
+        store_rev       28, 24, 20, 16
+.purgem store_rev
+        ret
+endfunc
+
+// This is mostly the same as 4x32_pass1, but without the transpose,
+// and use the source as temp buffer between the two idct passes, and
+// add into the destination.
+// x0 = dst
+// x1 = dst stride
+// x2 = src (temp buffer)
+// x7 = negative double temp buffer stride
+// x9 = double temp buffer stride
+function idct32_1d_4x32_pass2_neon
+        // v16 = IN(0), v17 = IN(2) ... v31 = IN(30)
+.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
+        ld1             {v\i\().4s}, [x2], x9
+.endr
+        sub             x2,  x2,  x9, lsl #4
+
+        idct16
+
+.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
+        st1             {v\i\().4s}, [x2], x9
+.endr
+
+        sub             x2,  x2,  x9, lsl #4
+        add             x2,  x2,  #128
+
+        // v16 = IN(1), v17 = IN(3) ... v31 = IN(31)
+.irp i, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 26, 27, 28, 29, 30, 31
+        ld1             {v\i\().4s}, [x2], x9
+.endr
+        sub             x2,  x2,  x9, lsl #4
+        sub             x2,  x2,  #128
+
+        idct32_odd
+
+.macro load_acc_store a, b, c, d, neg=0
+.if \neg == 0
+        ld1             {v4.4s},  [x2], x9
+        ld1             {v5.4s},  [x2], x9
+        add             v4.4s, v4.4s, v\a\().4s
+        ld1             {v6.4s},  [x2], x9
+        add             v5.4s, v5.4s, v\b\().4s
+        ld1             {v7.4s},  [x2], x9
+        add             v6.4s, v6.4s, v\c\().4s
+        add             v7.4s, v7.4s, v\d\().4s
+.else
+        ld1             {v4.4s},  [x2], x7
+        ld1             {v5.4s},  [x2], x7
+        sub             v4.4s, v4.4s, v\a\().4s
+        ld1             {v6.4s},  [x2], x7
+        sub             v5.4s, v5.4s, v\b\().4s
+        ld1             {v7.4s},  [x2], x7
+        sub             v6.4s, v6.4s, v\c\().4s
+        sub             v7.4s, v7.4s, v\d\().4s
+.endif
+        ld1             {v8.4h},   [x0], x1
+        ld1             {v8.d}[1], [x0], x1
+        srshr           v4.4s, v4.4s, #6
+        ld1             {v9.4h},   [x0], x1
+        srshr           v5.4s, v5.4s, #6
+        uaddw           v4.4s, v4.4s, v8.4h
+        ld1             {v9.d}[1], [x0], x1
+        srshr           v6.4s, v6.4s, #6
+        uaddw2          v5.4s, v5.4s, v8.8h
+        srshr           v7.4s, v7.4s, #6
+        sub             x0,  x0,  x1, lsl #2
+        uaddw           v6.4s, v6.4s, v9.4h
+        sqxtun          v4.4h, v4.4s
+        uaddw2          v7.4s, v7.4s, v9.8h
+        sqxtun2         v4.8h, v5.4s
+        umin            v4.8h, v4.8h, v15.8h
+        st1             {v4.4h},   [x0], x1
+        sqxtun          v5.4h, v6.4s
+        st1             {v4.d}[1], [x0], x1
+        sqxtun2         v5.8h, v7.4s
+        umin            v5.8h, v5.8h, v15.8h
+        st1             {v5.4h},   [x0], x1
+        st1             {v5.d}[1], [x0], x1
+.endm
+        load_acc_store  31, 30, 29, 28
+        load_acc_store  27, 26, 25, 24
+        load_acc_store  23, 22, 21, 20
+        load_acc_store  19, 18, 17, 16
+        sub             x2,  x2,  x9
+        load_acc_store  16, 17, 18, 19, 1
+        load_acc_store  20, 21, 22, 23, 1
+        load_acc_store  24, 25, 26, 27, 1
+        load_acc_store  28, 29, 30, 31, 1
+.purgem load_acc_store
+        ret
+endfunc
+
+const min_eob_idct_idct_32, align=4
+        .short  0, 9, 34, 70, 135, 240, 336, 448
+endconst
+
+function vp9_idct_idct_32x32_add_16_neon
+        cmp             w3,  #1
+        b.eq            idct32x32_dc_add_neon
+
+        movrel          x10, idct_coeffs
+        movrel          x12, min_eob_idct_idct_32, 2
+
+        mov             x15, x30
+        stp             d8,  d9,  [sp, #-0x10]!
+        stp             d10, d11, [sp, #-0x10]!
+        stp             d12, d13, [sp, #-0x10]!
+        stp             d14, d15, [sp, #-0x10]!
+
+        sub             sp,  sp,  #4096
+
+        mov             x4,  x0
+        mov             x5,  x1
+        mov             x6,  x2
+
+        // Double stride of the input, since we only read every other line
+        mov             x9,  #256
+        neg             x7,  x9
+
+        ld1             {v0.8h,v1.8h},   [x10], #32
+        sxtl            v2.4s,  v1.4h
+        sxtl2           v3.4s,  v1.8h
+        sxtl2           v1.4s,  v0.8h
+        sxtl            v0.4s,  v0.4h
+        ld1             {v10.8h,v11.8h}, [x10]
+        sxtl            v12.4s, v11.4h
+        sxtl2           v13.4s, v11.8h
+        sxtl2           v11.4s, v10.8h
+        sxtl            v10.4s, v10.4h
+
+        dup             v15.8h, w13
+
+.irp i, 0, 4, 8, 12, 16, 20, 24, 28
+        add             x0,  sp,  #(\i*128)
+.if \i > 0
+        ldrh            w1,  [x12], #2
+        cmp             w3,  w1
+        mov             x1,  #(32 - \i)/4
+        b.le            1f
+.endif
+        add             x2,  x6,  #(\i*4)
+        bl              idct32_1d_4x32_pass1_neon
+.endr
+        b               3f
+
+1:
+        // Write zeros to the temp buffer for pass 2
+        movi            v16.4s,  #0
+        movi            v17.4s,  #0
+        movi            v18.4s,  #0
+        movi            v19.4s,  #0
+2:
+        subs            x1,  x1,  #1
+.rept 4
+        st1             {v16.4s-v19.4s},  [x0], #64
+        st1             {v16.4s-v19.4s},  [x0], #64
+.endr
+        b.ne            2b
+3:
+.irp i, 0, 4, 8, 12, 16, 20, 24, 28
+        add             x0,  x4,  #(\i*2)
+        mov             x1,  x5
+        add             x2,  sp,  #(\i*4)
+        bl              idct32_1d_4x32_pass2_neon
+.endr
+
+        add             sp,  sp,  #4096
+        ldp             d14, d15, [sp], 0x10
+        ldp             d12, d13, [sp], 0x10
+        ldp             d10, d11, [sp], 0x10
+        ldp             d8,  d9,  [sp], 0x10
+
+        br              x15
+endfunc
+
+function ff_vp9_idct_idct_32x32_add_10_neon, export=1
+        mov             x13, #0x03ff
+        b               vp9_idct_idct_32x32_add_16_neon
+endfunc
+
+function ff_vp9_idct_idct_32x32_add_12_neon, export=1
+        mov             x13, #0x0fff
+        b               vp9_idct_idct_32x32_add_16_neon
+endfunc