Quelle  mc_dotprod.S   Sprache: Sparc

/*
 * Copyright © 2024, VideoLAN and dav1d authors
 * Copyright © 2024, Janne Grunau
 * Copyright © 2024, Martin Storsjo
 * Copyright © 2024, Arm Limited
 * All rights reserved.
 *
 * Redistribution and use in source and binary forms, with or without
 * modification, are permitted provided that the following conditions are met:
 *
 * 1. Redistributions of source code must retain the above copyright notice, this
 *    list of conditions and the following disclaimer.
 *
 * 2. Redistributions in binary form must reproduce the above copyright notice,
 *    this list of conditions and the following disclaimer in the documentation
 *    and/or other materials provided with the distribution.
 *
 * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS" AND
 * ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE IMPLIED
 * WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
 * DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR
 * ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES
 * (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES;
 * LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED AND
 * ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY, OR TORT
 * (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE OF THIS
 * SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
 */

#include "src/arm/asm.S"
#include "util.S"


#if HAVE_DOTPROD
ENABLE_DOTPROD

// No spaces in these expressions, due to gas-preprocessor. It is translated by
// -1 to save the negative offset at getting the address of `mc_subpel_filters`.
#define REGULAR1        (((0*15-1)<<7)|(3*15-1))
#define SMOOTH1         (((1*15-1)<<7)|(4*15-1))
#define SHARP1          (((2*15-1)<<7)|(3*15-1))

#define FUNC_ALIGN      2
#define JUMP_ALIGN      2
#define LOOP_ALIGN      2


const h_tbl_neon_dotprod, align=4
        // Shuffle indices to permute horizontal samples in preparation for
        // input to SDOT instructions. The 8-tap horizontal convolution uses
        // sample indices in the interval of [-3, 4] relative to the current
        // sample position.
        .byte  0,  1,  2,  3,   1,  2,  3,  4,   2,  3,  4,  5,   3,  4,  5,  6
        .byte  4,  5,  6,  7,   5,  6,  7,  8,   6,  7,  8,  9,   7,  8,  9, 10
        .byte  8,  9, 10, 11,   9, 10, 11, 12,  10, 11, 12, 13,  11, 12, 13, 14

        // Shuffle indices to permute horizontal samples in preparation for
        // input to USMMLA instructions.
#define OFFSET_USMMLA 48
        .byte  0,  1,  2,  3,   4,  5,  6,  7,   2,  3,  4,  5,   6,  7,  8,  9
        .byte  4,  5,  6,  7,   8,  9, 10, 11,   6,  7,  8,  9,  10, 11, 12, 13

        // Lookup table used to help conversion of shifted 32-bit values to 8-bit.
#define OFFSET_CVT_32_8 80
        .byte  1,  2,  5,  6,   9, 10, 13, 14,  17, 18, 21, 22,  25, 26, 29, 30
endconst

const v_tbl_neon_dotprod, align=4
        // Vertical convolutions are also using SDOT instructions, where a
        // 128-bit register contains a transposed 4x4 matrix of values.
        // Subsequent iterations of the vertical convolution can reuse the
        // 3x4 sub-matrix from the previous loop iteration. These shuffle
        // indices shift and merge this 4x4 matrix with the values of a new
        // line.
        .byte  1,  2,  3, 16,   5,  6,  7, 20,   9, 10, 11, 24,  13, 14, 15, 28
        .byte  1,  2,  3, 16,   5,  6,  7, 17,   9, 10, 11, 18,  13, 14, 15, 19
        .byte  1,  2,  3, 20,   5,  6,  7, 21,   9, 10, 11, 22,  13, 14, 15, 23
        .byte  1,  2,  3, 24,   5,  6,  7, 25,   9, 10, 11, 26,  13, 14, 15, 27
        .byte  1,  2,  3, 28,   5,  6,  7, 29,   9, 10, 11, 30,  13, 14, 15, 31
endconst


.macro make_8tap_fn op, type, type_h, type_v, isa, jump=1
function \op\()_8tap_\type\()_8bpc_\isa, export=1, align=FUNC_ALIGN
        mov             x9,  \type_h
        mov             x10, \type_v
    .if \jump
        b               \op\()_8tap_\isa
    .endif
endfunc
.endm

.macro filter_8tap_fn type, dot, isa, dst, d_strd, src, s_strd, w, h, mx, my, xmx, xmy, ldst, lsrc, wd_strd
make_8tap_fn \type, sharp,          SHARP1,   SHARP1,   \isa
make_8tap_fn \type, sharp_smooth,   SHARP1,   SMOOTH1,  \isa
make_8tap_fn \type, sharp_regular,  SHARP1,   REGULAR1, \isa
make_8tap_fn \type, smooth_sharp,   SMOOTH1,  SHARP1,   \isa
make_8tap_fn \type, smooth,         SMOOTH1,  SMOOTH1,  \isa
make_8tap_fn \type, smooth_regular, SMOOTH1,  REGULAR1, \isa
make_8tap_fn \type, regular_sharp,  REGULAR1, SHARP1,   \isa
make_8tap_fn \type, regular_smooth, REGULAR1, SMOOTH1,  \isa
make_8tap_fn \type, regular,        REGULAR1, REGULAR1, \isa, jump=0

function \type\()_8tap_\isa, align=FUNC_ALIGN
        clz             w8, \w
        mov             w11,  #0x4081   // (1 << 14) | (1 << 7) | (1 << 0)
        sub             w8, w8, #24     // for jump tables
        movrel          x12, X(mc_subpel_filters)
        cbnz            \mx, L(\type\()_8tap_h_hv_\isa)
        cbnz            \my, L(\type\()_8tap_v_\isa)
.ifc \type, prep
        add             \wd_strd, \w, \w    // prep_neon needs w * 2 as stride
.endif
        b               X(\type\()_neon)

        .align JUMP_ALIGN
L(\type\()_8tap_v_\isa):
        madd            \my, \my, w11, w10
        movrel          x13, v_tbl_neon_dotprod
        sub             \src, \src, \s_strd
.ifc \isa, neon_dotprod
    .ifc \type, prep
        mov             w8, #0x2002         // FILTER_WEIGHT * 128 + rounding
        dup             v4.4s, w8
    .else
        movi            v4.4s, #32, lsl #8  // FILTER_WEIGHT * 128, bias for SDOT
    .endif
.endif
        ubfx            w11, \my, #7, #7
        and             \my, \my, #0x7F
        ldp             q6, q28, [x13]
        cmp             \h, #4
        csel            \my, \my, w11, le
        sub             \src, \src, \s_strd, lsl #1     // src - s_strd * 3
        add             \xmy, x12, \xmy, lsl #3         // subpel V filter address
        ldr             q29, [x13, #32]
.ifc \isa, neon_dotprod
        movi            v5.16b, #128
.endif
        ldr             d7, [\xmy]
        cmp             \w, #8
        b.eq            80f
        b.lt            40f

        // .align JUMP_ALIGN    // fallthrough
160:    // V - 16xN+
        ldp             q30, q31, [x13, #48]
.ifc \type, prep
        add             \wd_strd, \w, \w
.endif
        .align LOOP_ALIGN
161:
        mov             \lsrc, \src
        mov             \ldst, \dst
        sub             w8, \h, #1

        ldr             q16, [\lsrc]
        ldr             q17, [\lsrc, \s_strd]
        add             \lsrc, \lsrc, \s_strd, lsl #1
        ldr             q18, [\lsrc]
        ldr             q19, [\lsrc, \s_strd]
        add             \lsrc, \lsrc, \s_strd, lsl #1

        zip1            v0.16b, v16.16b, v17.16b
        zip2            v1.16b, v16.16b, v17.16b
        zip1            v2.16b, v18.16b, v19.16b
        zip2            v3.16b, v18.16b, v19.16b

        ldr             q20, [\lsrc]
        ldr             q21, [\lsrc, \s_strd]
        add             \lsrc, \lsrc, \s_strd, lsl #1
        ldr             q22, [\lsrc]
        ldr             q23, [\lsrc, \s_strd]
        add             \lsrc, \lsrc, \s_strd, lsl #1

        zip1            v18.16b, v20.16b, v21.16b
        zip2            v21.16b, v20.16b, v21.16b
        zip1            v24.16b, v22.16b, v23.16b
        zip2            v27.16b, v22.16b, v23.16b

        zip1            v16.8h, v0.8h, v2.8h
        zip2            v19.8h, v0.8h, v2.8h
        zip1            v22.8h, v1.8h, v3.8h
        zip2            v25.8h, v1.8h, v3.8h

        zip1            v17.8h, v18.8h, v24.8h
        zip2            v20.8h, v18.8h, v24.8h
        zip1            v23.8h, v21.8h, v27.8h
        zip2            v26.8h, v21.8h, v27.8h
.ifc \isa, neon_dotprod
        sub             v16.16b, v16.16b, v5.16b
        sub             v19.16b, v19.16b, v5.16b
        sub             v22.16b, v22.16b, v5.16b
        sub             v25.16b, v25.16b, v5.16b

        sub             v17.16b, v17.16b, v5.16b
        sub             v20.16b, v20.16b, v5.16b
        sub             v23.16b, v23.16b, v5.16b
        sub             v26.16b, v26.16b, v5.16b
.endif
        .align LOOP_ALIGN
16:
.ifc \isa, neon_i8mm
        ld1             {v18.16b}, [\lsrc], \s_strd
        movi            v0.4s, #0
        movi            v1.4s, #0
        movi            v2.4s, #0
        movi            v3.4s, #0
        mov             v21.16b, v18.16b
        mov             v24.16b, v18.16b
        mov             v27.16b, v18.16b
.else   // neon_dotprod
        ld1             {v27.16b}, [\lsrc], \s_strd
        mov             v0.16b, v4.16b
        mov             v1.16b, v4.16b
        mov             v2.16b, v4.16b
        mov             v3.16b, v4.16b
        sub             v18.16b, v27.16b, v5.16b
        sub             v21.16b, v27.16b, v5.16b
        sub             v24.16b, v27.16b, v5.16b
        sub             v27.16b, v27.16b, v5.16b
.endif
        \dot            v0.4s, v16.16b, v7.4b[0]
        \dot            v1.4s, v19.16b, v7.4b[0]
        \dot            v2.4s, v22.16b, v7.4b[0]
        \dot            v3.4s, v25.16b, v7.4b[0]

        tbl             v16.16b, {v16.16b, v17.16b}, v6.16b
        tbl             v19.16b, {v19.16b, v20.16b}, v6.16b
        tbl             v22.16b, {v22.16b, v23.16b}, v6.16b
        tbl             v25.16b, {v25.16b, v26.16b}, v6.16b

        \dot            v0.4s, v17.16b, v7.4b[1]
        \dot            v1.4s, v20.16b, v7.4b[1]
        \dot            v2.4s, v23.16b, v7.4b[1]
        \dot            v3.4s, v26.16b, v7.4b[1]

        tbl             v17.16b, {v17.16b, v18.16b}, v28.16b
        tbl             v20.16b, {v20.16b, v21.16b}, v29.16b
        tbl             v23.16b, {v23.16b, v24.16b}, v30.16b
        tbl             v26.16b, {v26.16b, v27.16b}, v31.16b

        subs            w8, w8, #1
        uzp1            v0.8h, v0.8h, v1.8h
        uzp1            v2.8h, v2.8h, v3.8h
.ifc \type, prep
    .ifc \isa, neon_i8mm
        srshr           v0.8h, v0.8h, #2
        srshr           v1.8h, v2.8h, #2
    .else
        sshr            v0.8h, v0.8h, #2
        sshr            v1.8h, v2.8h, #2
    .endif
        st1             {v0.8h, v1.8h}, [\ldst], \d_strd
.else   // put
        sqrshrun        v0.8b, v0.8h, #6
        sqrshrun2       v0.16b, v2.8h, #6
        st1             {v0.16b}, [\ldst], \d_strd
.endif
        b.gt            16b

.ifc \isa, neon_i8mm
        movi            v0.4s, #0
        movi            v1.4s, #0
        movi            v2.4s, #0
        movi            v3.4s, #0
.else   // neon_dotprod
        mov             v0.16b, v4.16b
        mov             v1.16b, v4.16b
        mov             v2.16b, v4.16b
        mov             v3.16b, v4.16b
.endif
        \dot            v0.4s, v16.16b, v7.4b[0]
        \dot            v1.4s, v19.16b, v7.4b[0]
        \dot            v2.4s, v22.16b, v7.4b[0]
        \dot            v3.4s, v25.16b, v7.4b[0]

        \dot            v0.4s, v17.16b, v7.4b[1]
        \dot            v1.4s, v20.16b, v7.4b[1]
        \dot            v2.4s, v23.16b, v7.4b[1]
        \dot            v3.4s, v26.16b, v7.4b[1]

        subs            \w, \w, #16
        uzp1            v0.8h, v0.8h, v1.8h
        uzp1            v2.8h, v2.8h, v3.8h
.ifc \type, prep
    .ifc \isa, neon_i8mm
        srshr           v0.8h, v0.8h, #2
        srshr           v1.8h, v2.8h, #2
    .else
        sshr            v0.8h, v0.8h, #2
        sshr            v1.8h, v2.8h, #2
    .endif
        stp             q0, q1, [\ldst]
        add             \dst, \dst, #32
.else   // put
        sqrshrun        v0.8b, v0.8h, #6
        sqrshrun2       v0.16b, v2.8h, #6
        str             q0, [\ldst]
        add             \dst, \dst, #16
.endif
        add             \src, \src, #16
        b.gt            161b
        ret

        .align JUMP_ALIGN
80:     // V - 8xN
        ldr             d16, [\src]
        ldr             d17, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
        ldr             d18, [\src]
        ldr             d19, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1

        ldr             d20, [\src]
        ldr             d21, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
        ldr             d22, [\src]
        ldr             d23, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
        subs            \h, \h, #2  // for prep: sub is enough

        zip1            v0.16b, v16.16b, v17.16b
        zip1            v2.16b, v18.16b, v19.16b
        zip1            v18.16b, v20.16b, v21.16b
        zip1            v24.16b, v22.16b, v23.16b

        zip1            v16.8h,  v0.8h,  v2.8h
        zip2            v19.8h,  v0.8h,  v2.8h
        zip1            v17.8h, v18.8h, v24.8h
        zip2            v20.8h, v18.8h, v24.8h
.ifc \isa, neon_dotprod
        sub             v16.16b, v16.16b, v5.16b
        sub             v19.16b, v19.16b, v5.16b
        sub             v17.16b, v17.16b, v5.16b
        sub             v20.16b, v20.16b, v5.16b
.endif
.ifc \type, put
        b.eq            82f
.endif
        .align LOOP_ALIGN
8:
.ifc \isa, neon_i8mm
        ldr             d18, [\src]
        movi            v0.4s, #0
        movi            v1.4s, #0
        ldr             d24, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
        movi            v2.4s, #0
        movi            v3.4s, #0
        mov             v21.8b, v18.8b
        mov             v27.8b, v24.8b
.else   // neon_dotprod
        ldr             d21, [\src]
        ldr             d27, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
        mov             v0.16b, v4.16b
        mov             v1.16b, v4.16b
        mov             v2.16b, v4.16b
        mov             v3.16b, v4.16b
        sub             v18.16b, v21.16b, v5.16b
        sub             v21.16b, v21.16b, v5.16b
        sub             v24.16b, v27.16b, v5.16b
        sub             v27.16b, v27.16b, v5.16b
.endif
        tbl             v22.16b, {v16.16b, v17.16b}, v6.16b
        tbl             v25.16b, {v19.16b, v20.16b}, v6.16b
        tbl             v23.16b, {v17.16b, v18.16b}, v28.16b
        tbl             v26.16b, {v20.16b, v21.16b}, v29.16b

        \dot            v0.4s, v16.16b, v7.4b[0]
        \dot            v0.4s, v17.16b, v7.4b[1]
        \dot            v1.4s, v19.16b, v7.4b[0]
        \dot            v1.4s, v20.16b, v7.4b[1]

        tbl             v16.16b, {v22.16b, v23.16b}, v6.16b
        tbl             v19.16b, {v25.16b, v26.16b}, v6.16b
        tbl             v17.16b, {v23.16b, v24.16b}, v28.16b
        tbl             v20.16b, {v26.16b, v27.16b}, v29.16b

        \dot            v2.4s, v22.16b, v7.4b[0]
        \dot            v2.4s, v23.16b, v7.4b[1]
        \dot            v3.4s, v25.16b, v7.4b[0]
        \dot            v3.4s, v26.16b, v7.4b[1]

        subs            \h, \h, #2
        uzp1            v0.8h, v0.8h, v1.8h
        uzp1            v2.8h, v2.8h, v3.8h
.ifc \type, prep
    .ifc \isa, neon_i8mm
        srshr           v0.8h, v0.8h, #2
        srshr           v1.8h, v2.8h, #2
    .else
        sshr            v0.8h, v0.8h, #2
        sshr            v1.8h, v2.8h, #2
    .endif
        stp             q0, q1, [\dst], #32
.else   // put
        sqrshrun        v0.8b, v0.8h, #6
        sqrshrun        v1.8b, v2.8h, #6
        str             d0, [\dst]
        str             d1, [\dst, \d_strd]
        add             \dst, \dst, \d_strd, lsl #1
.endif
        b.gt            8b

.ifc \type, put
        .align JUMP_ALIGN
82:
.endif
.ifc \isa, neon_i8mm
        ldr             d18, [\src]
        movi            v0.4s, #0
        movi            v1.4s, #0
        movi            v2.4s, #0
        movi            v3.4s, #0
        mov             v21.8b, v18.8b
.else   // neon_dotprod
        ldr             d21, [\src]
        mov             v0.16b, v4.16b
        mov             v1.16b, v4.16b
        mov             v2.16b, v4.16b
        mov             v3.16b, v4.16b
        sub             v18.16b, v21.16b, v5.16b
        sub             v21.16b, v21.16b, v5.16b
.endif
        tbl             v22.16b, {v16.16b, v17.16b}, v6.16b
        tbl             v25.16b, {v19.16b, v20.16b}, v6.16b
        tbl             v23.16b, {v17.16b, v18.16b}, v28.16b
        tbl             v26.16b, {v20.16b, v21.16b}, v29.16b

        \dot            v0.4s, v16.16b, v7.4b[0]
        \dot            v0.4s, v17.16b, v7.4b[1]
        \dot            v1.4s, v19.16b, v7.4b[0]
        \dot            v1.4s, v20.16b, v7.4b[1]

        \dot            v2.4s, v22.16b, v7.4b[0]
        \dot            v2.4s, v23.16b, v7.4b[1]
        \dot            v3.4s, v25.16b, v7.4b[0]
        \dot            v3.4s, v26.16b, v7.4b[1]

        uzp1            v0.8h, v0.8h, v1.8h
        uzp1            v2.8h, v2.8h, v3.8h
.ifc \type, prep
    .ifc \isa, neon_i8mm
        srshr           v0.8h, v0.8h, #2
        srshr           v1.8h, v2.8h, #2
    .else
        sshr            v0.8h, v0.8h, #2
        sshr            v1.8h, v2.8h, #2
    .endif
        stp             q0, q1, [\dst]
.else   // put
        sqrshrun        v0.8b, v0.8h, #6
        sqrshrun        v1.8b, v2.8h, #6
        str             d0, [\dst]
        str             d1, [\dst, \d_strd]
.endif
        ret

        .align JUMP_ALIGN
40:     // V - 4xN or 2xN (put only)
.ifc \type, put
        cmp             \w, #2
        b.eq            20f
.endif
        ldr             s16, [\src]
        ldr             s17, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
        ldr             s18, [\src]
        ldr             s19, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1

        ldr             s20, [\src]
        ldr             s21, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
        ldr             s22, [\src]
        ldr             s23, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
        subs            \h, \h, #2  // for prep: sub is enough

        zip1            v0.8b, v16.8b, v17.8b
        zip1            v2.8b, v18.8b, v19.8b
        zip1            v18.8b, v20.8b, v21.8b
        zip1            v24.8b, v22.8b, v23.8b

        zip1            v16.8h, v0.8h, v2.8h
        zip1            v17.8h, v18.8h, v24.8h
.ifc \isa, neon_dotprod
        sub             v16.16b, v16.16b, v5.16b
        sub             v17.16b, v17.16b, v5.16b
.endif
.ifc \type, put
        b.eq            42f
.endif
        .align LOOP_ALIGN
4:
        ldr             s18, [\src]
        ldr             s21, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
.ifc \isa, neon_i8mm
        movi            v0.4s, #0
        movi            v1.4s, #0
.else   // neon_dotprod
        mov             v0.16b, v4.16b
        mov             v1.16b, v4.16b
        sub             v18.16b, v18.16b, v5.16b
        sub             v21.16b, v21.16b, v5.16b
.endif
        tbl             v19.16b, {v16.16b, v17.16b}, v6.16b
        tbl             v20.16b, {v17.16b, v18.16b}, v28.16b

        \dot            v0.4s, v16.16b, v7.4b[0]
        \dot            v0.4s, v17.16b, v7.4b[1]

        tbl             v16.16b, {v19.16b, v20.16b}, v6.16b
        tbl             v17.16b, {v20.16b, v21.16b}, v28.16b

        \dot            v1.4s, v19.16b, v7.4b[0]
        \dot            v1.4s, v20.16b, v7.4b[1]
.ifc \type, prep
        subs            \h, \h, #2
    .ifc \isa, neon_i8mm
        rshrn           v0.4h, v0.4s, #2
        rshrn2          v0.8h, v1.4s, #2
    .else
        shrn            v0.4h, v0.4s, #2
        shrn2           v0.8h, v1.4s, #2
    .endif
        str             q0, [\dst], #16
.else
        uzp1            v0.8h, v0.8h, v1.8h
        sqrshrun        v0.8b, v0.8h, #6
        subs            \h, \h, #2
        fmov            x8, d0
        lsr             x9, x8, #32
        str             w8, [\dst]
        str             w9, [\dst, \d_strd]
        add             \dst, \dst, \d_strd, lsl #1
.endif
        b.gt            4b

.ifc \type, put
        .align JUMP_ALIGN
42:
.endif
        ldr             s18, [\src]
.ifc \isa, neon_i8mm
        movi            v0.4s, #0
        movi            v1.4s, #0
.else   // neon_dotprod
        mov             v0.16b, v4.16b
        mov             v1.16b, v4.16b
        sub             v18.16b, v18.16b, v5.16b
.endif
        tbl             v19.16b, {v16.16b, v17.16b}, v6.16b
        tbl             v20.16b, {v17.16b, v18.16b}, v28.16b

        \dot            v0.4s, v16.16b, v7.4b[0]
        \dot            v0.4s, v17.16b, v7.4b[1]

        \dot            v1.4s, v19.16b, v7.4b[0]
        \dot            v1.4s, v20.16b, v7.4b[1]
.ifc \type, prep
    .ifc \isa, neon_i8mm
        rshrn           v0.4h, v0.4s, #2
        rshrn2          v0.8h, v1.4s, #2
    .else
        shrn            v0.4h, v0.4s, #2
        shrn2           v0.8h, v1.4s, #2
    .endif
        str             q0, [\dst]
.else
        uzp1            v0.8h, v0.8h, v1.8h
        sqrshrun        v0.8b, v0.8h, #6
        fmov            x8, d0
        lsr             x9, x8, #32
        str             w8, [\dst]
        str             w9, [\dst, \d_strd]
.endif
        ret

.ifc \type, put
        .align JUMP_ALIGN
20:     // V - 2xN
        ldr             h16, [\src]
        ldr             h17, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
        ldr             h18, [\src]
        ldr             h19, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1

        ldr             h20, [\src]
        ldr             h21, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
        ldr             h22, [\src]
        ldr             h23, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
        subs            \h, \h, #2

        zip1            v0.8b, v16.8b, v17.8b
        zip1            v2.8b, v18.8b, v19.8b
        zip1            v18.8b, v20.8b, v21.8b
        zip1            v24.8b, v22.8b, v23.8b

        zip1            v16.4h, v0.4h, v2.4h
        zip1            v17.4h, v18.4h, v24.4h
    .ifc \isa, neon_dotprod
        sub             v16.8b, v16.8b, v5.8b
        sub             v17.8b, v17.8b, v5.8b
    .endif
        b.eq            22f

        .align LOOP_ALIGN
2:
        ldr             h18, [\src]
        ldr             h21, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
    .ifc \isa, neon_i8mm
        movi            v0.4s, #0
        movi            v1.4s, #0
    .else   // put
        mov             v0.16b, v4.16b
        mov             v1.16b, v4.16b
        sub             v18.8b, v18.8b, v5.8b
        sub             v21.8b, v21.8b, v5.8b
    .endif
        tbl             v19.16b, {v16.16b, v17.16b}, v6.16b
        tbl             v20.16b, {v17.16b, v18.16b}, v28.16b

        \dot            v0.4s, v16.16b, v7.4b[0]
        \dot            v0.4s, v17.16b, v7.4b[1]

        tbl             v16.16b, {v19.16b, v20.16b}, v6.16b
        tbl             v17.16b, {v20.16b, v21.16b}, v28.16b

        \dot            v1.4s, v19.16b, v7.4b[0]
        \dot            v1.4s, v20.16b, v7.4b[1]

        uzp1            v0.8h, v0.8h, v1.8h
        sqrshrun        v0.8b, v0.8h, #6

        subs            \h, \h, #2
        fmov            x8, d0
        lsr             x9, x8, #32
        strh            w8, [\dst]
        strh            w9, [\dst, \d_strd]
        add             \dst, \dst, \d_strd, lsl #1
        b.gt            2b

        .align JUMP_ALIGN
22:
        ldr             h18, [\src]
    .ifc \isa, neon_i8mm
        movi            v0.4s, #0
        movi            v1.4s, #0
    .else   // put
        mov             v0.16b, v4.16b
        mov             v1.16b, v4.16b
        sub             v18.8b, v18.8b, v5.8b
    .endif
        tbl             v19.16b, {v16.16b, v17.16b}, v6.16b
        tbl             v20.16b, {v17.16b, v18.16b}, v28.16b

        \dot            v0.4s, v16.16b, v7.4b[0]
        \dot            v0.4s, v17.16b, v7.4b[1]

        \dot            v1.4s, v19.16b, v7.4b[0]
        \dot            v1.4s, v20.16b, v7.4b[1]

        uzp1            v0.8h, v0.8h, v1.8h
        sqrshrun        v0.8b, v0.8h, #6

        fmov            x8, d0
        lsr             x9, x8, #32
        strh            w8, [\dst]
        strh            w9, [\dst, \d_strd]
        ret
.endif

        .align JUMP_ALIGN
L(\type\()_8tap_h_hv_\isa):
        madd            \mx, \mx, w11, w9
        madd            w14, \my, w11, w10      // for HV
.ifc \isa, neon_dotprod
        mov             w13, #0x2002            // FILTER_WEIGHT * 128 + rounding
        dup             v27.4s, w13             // put H overrides this
.endif
        movrel          x13, h_tbl_neon_dotprod
        sub             \src, \src, #3          // src - 3
        ldr             q28, [x13]              // for 4-tap & 8-tap H filters
        ubfx            w15, \mx, #7, #7
        and             \mx, \mx, #0x7F
        ubfx            w11, w14, #7, #7        // for HV
        and             w14, w14, #0x7F         // for HV
        cmp             \w, #4
        csel            \mx, \mx, w15, le
        add             \xmx, x12, \xmx, lsl #3 // subpel H filter address
.ifc \isa, neon_dotprod
        movi            v24.16b, #128
.endif
        cbz             \my, L(\type\()_8tap_h_\isa)

        // HV cases
        cmp             \h, #4
        csel            w14, w14, w11, le
        sub             \src, \src, \s_strd, lsl #1 // src - s_strd * 2 - 3
        add             \xmy, x12, x14, lsl #3      // subpel V filter address
        mov             x15, x30
        ldr             d7, [\xmy]
.ifc \type, put
        ldr             q25, [x13, #(OFFSET_CVT_32_8)] // LUT to help conversion
.endif                                                 // of 32b values to 8b
        sxtl            v7.8h, v7.8b
        cmp             w10, #SHARP1
        b.ne            L(\type\()_6tap_hv_\isa)    // vertical != SHARP1

        // HV 8-tap cases
        sub             \src, \src, \s_strd         // src - s_strd * 3 - 3
        cmp             \w, #4
        b.eq            40f
.ifc \type, put
        b.lt            20f
.endif

        // .align JUMP_ALIGN    // fallthrough
80:     // HV8 - 8xN+
        ldp             q29, q30, [x13, #16]
        ldr             d26, [\xmx]
.ifc \type, prep
        add             \wd_strd, \w, \w
.endif
        .align LOOP_ALIGN
81:
        mov             \lsrc, \src
        mov             \ldst, \dst
        mov             w8, \h
.ifc \isa, neon_i8mm
        bl              L(\type\()_hv_filter8_\isa)
        srshr           v16.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        srshr           v17.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        srshr           v18.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        srshr           v19.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        srshr           v20.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        srshr           v21.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        srshr           v22.8h, v22.8h, #2
.else
        bl              L(\type\()_hv_filter8_\isa)
        sshr            v16.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        sshr            v17.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        sshr            v18.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        sshr            v19.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        sshr            v20.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        sshr            v21.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        sshr            v22.8h, v22.8h, #2
.endif
        .align LOOP_ALIGN
8:
        ldr             q23, [\lsrc]
        add             \lsrc, \lsrc, \s_strd

        smull           v0.4s, v16.4h, v7.h[0]
        smull2          v1.4s, v16.8h, v7.h[0]
        mov             v16.16b, v17.16b
.ifc \isa, neon_i8mm
        movi            v5.4s, #0
        movi            v6.4s, #0
        tbl             v2.16b, {v23.16b}, v28.16b
        tbl             v3.16b, {v23.16b}, v29.16b
.else   // neon_dotprod
        sub             v23.16b, v23.16b, v24.16b
        mov             v5.16b, v27.16b
        mov             v6.16b, v27.16b
.endif
        smlal           v0.4s, v17.4h, v7.h[1]
        smlal2          v1.4s, v17.8h, v7.h[1]
.ifc \isa, neon_i8mm
        tbl             v4.16b, {v23.16b}, v30.16b
        mov             v17.16b, v18.16b
.else   // neon_dotprod
        mov             v17.16b, v18.16b
        tbl             v2.16b, {v23.16b}, v28.16b
        tbl             v3.16b, {v23.16b}, v29.16b
        tbl             v4.16b, {v23.16b}, v30.16b
.endif
        smlal           v0.4s, v18.4h, v7.h[2]
        smlal2          v1.4s, v18.8h, v7.h[2]
        mov             v18.16b, v19.16b

        \dot            v5.4s, v2.16b, v26.4b[0]
        \dot            v6.4s, v3.16b, v26.4b[0]

        smlal           v0.4s, v19.4h, v7.h[3]
        smlal2          v1.4s, v19.8h, v7.h[3]
        mov             v19.16b, v20.16b

        \dot            v5.4s, v3.16b, v26.4b[1]
        \dot            v6.4s, v4.16b, v26.4b[1]

        smlal           v0.4s, v20.4h, v7.h[4]
        smlal2          v1.4s, v20.8h, v7.h[4]
        mov             v20.16b, v21.16b

        smlal           v0.4s, v21.4h, v7.h[5]
        smlal2          v1.4s, v21.8h, v7.h[5]
.ifc \type, prep
        uzp1            v23.8h, v5.8h, v6.8h
.endif
        mov             v21.16b, v22.16b
        smlal           v0.4s, v22.4h, v7.h[6]
        smlal2          v1.4s, v22.8h, v7.h[6]
.ifc \isa, neon_i8mm
        subs            w8, w8, #1
.endif
.ifc \type, prep
    .ifc \isa, neon_i8mm
        srshr           v22.8h, v23.8h, #2
    .else
        sshr            v22.8h, v23.8h, #2
    .endif
        smlal           v0.4s, v22.4h, v7.h[7]
        smlal2          v1.4s, v22.8h, v7.h[7]
        rshrn           v0.4h, v0.4s, #6
        rshrn2          v0.8h, v1.4s, #6
.else   // put
    .ifc \isa, neon_i8mm
        rshrn           v22.4h, v5.4s, #2
        rshrn2          v22.8h, v6.4s, #2
    .else
        shrn            v22.4h, v5.4s, #2
        shrn2           v22.8h, v6.4s, #2
    .endif
        smlal           v0.4s, v22.4h, v7.h[7]
        smlal2          v1.4s, v22.8h, v7.h[7]
        tbl             v0.16b, {v0.16b, v1.16b}, v25.16b
        sqrshrun        v0.8b, v0.8h, #2
.endif
.ifc \isa, neon_dotprod
        subs            w8, w8, #1
.endif
.ifc \type, prep
        st1             {v0.8h}, [\ldst], \d_strd
        b.gt            8b
        add             \dst, \dst, #16
.else
        st1             {v0.8b}, [\ldst], \d_strd
        b.gt            8b
        add             \dst, \dst, #8
.endif
        add             \src, \src, #8
        subs            \w, \w, #8
        b.gt            81b
        ret             x15

        .align JUMP_ALIGN
40:     // HV8 - 4xN
        ldur            s26, [\xmx, #2]
        add             \src, \src, #2

        bl              L(\type\()_hv_filter4_\isa)
        shrn            v16.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v17.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v18.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v19.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v20.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v21.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v22.4h, v22.4s, #2

        .align LOOP_ALIGN
4:
        ld1             {v4.8b}, [\src], \s_strd

        smull           v0.4s, v16.4h, v7.h[0]
        smlal           v0.4s, v17.4h, v7.h[1]
        mov             v16.16b, v17.16b
        mov             v17.16b, v18.16b
.ifc \isa, neon_dotprod
        sub             v4.16b, v4.16b, v24.16b
.endif
        smlal           v0.4s, v18.4h, v7.h[2]
        smlal           v0.4s, v19.4h, v7.h[3]
        tbl             v2.16b, {v4.16b}, v28.16b
.ifc \isa, neon_i8mm
        movi            v5.4s, #0
.else
        mov             v5.16b, v27.16b
.endif
        mov             v18.16b, v19.16b
        mov             v19.16b, v20.16b

        smlal           v0.4s, v20.4h, v7.h[4]
        smlal           v0.4s, v21.4h, v7.h[5]

        \dot            v5.4s, v2.16b, v26.4b[0]
        mov             v20.16b, v21.16b
        mov             v21.16b, v22.16b
        smlal           v0.4s, v22.4h, v7.h[6]
.ifc \isa, neon_i8mm
        rshrn           v22.4h, v5.4s, #2
.else
        shrn            v22.4h, v5.4s, #2
.endif
        smlal           v0.4s, v22.4h, v7.h[7]
.ifc \type, prep
        rshrn           v0.4h, v0.4s, #6
        str             d0, [\dst], #8
        subs            \h, \h, #1
.else
        subs            \h, \h, #1
        tbl             v0.8b, {v0.16b}, v25.8b
        sqrshrun        v0.8b, v0.8h, #2
        str             s0, [\dst]
        add             \dst, \dst, \d_strd
.endif
        b.gt            4b
        ret             x15

.ifc \type, put
        .align JUMP_ALIGN
20:     // HV8 - 2xN
        ldur            s26, [\xmx, #2]
        add             \src, \src, #2

        bl              L(\type\()_hv_filter4_\isa)
        shrn            v16.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v17.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v18.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v19.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v20.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v21.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v22.4h, v22.4s, #2

        .align LOOP_ALIGN
2:
        ld1             {v4.8b}, [\src], \s_strd

        smull           v0.4s, v16.4h, v7.h[0]
        smlal           v0.4s, v17.4h, v7.h[1]
        mov             v16.16b, v17.16b
        mov             v17.16b, v18.16b
    .ifc \isa, neon_dotprod
        sub             v4.16b, v4.16b, v24.16b
    .endif
        smlal           v0.4s, v18.4h, v7.h[2]
        smlal           v0.4s, v19.4h, v7.h[3]
        tbl             v2.16b, {v4.16b}, v28.16b
    .ifc \isa, neon_i8mm
        movi            v5.4s, #0
    .else
        mov             v5.16b, v27.16b
    .endif
        mov             v18.16b, v19.16b
        mov             v19.16b, v20.16b

        smlal           v0.4s, v20.4h, v7.h[4]
        smlal           v0.4s, v21.4h, v7.h[5]

        \dot            v5.4s, v2.16b, v26.4b[0]
        mov             v20.16b, v21.16b
        mov             v21.16b, v22.16b

        smlal           v0.4s, v22.4h, v7.h[6]
    .ifc \isa, neon_i8mm
        rshrn           v22.4h, v5.4s, #2
    .else
        shrn            v22.4h, v5.4s, #2
    .endif
        smlal           v0.4s, v22.4h, v7.h[7]
        subs            \h, \h, #1

        tbl             v0.8b, {v0.16b}, v25.8b
        sqrshrun        v0.8b, v0.8h, #2

        str             h0, [\dst]
        add             \dst, \dst, \d_strd
        b.gt            2b
        ret             x15
.endif

        .align JUMP_ALIGN
L(\type\()_6tap_hv_\isa):
        cmp             \w, #4
        b.eq            40f
.ifc \type, put
        b.lt            20f
.endif

        // .align JUMP_ALIGN    // fallthrough
80:     // HV6 - 8xN+
        ldr             d26, [\xmx]
.ifc \type, prep
        add             \wd_strd, \w, \w
.endif
.ifc \isa, neon_i8mm
        cmp             w9, #SHARP1
        b.eq            88f             // horizontal == SHARP1

        ldp             q29, q30, [x13, #(OFFSET_USMMLA)]
        ext             v0.8b, v26.8b, v26.8b, #7
        ins             v26.d[1], v0.d[0]

        .align LOOP_ALIGN
81:
        mov             \lsrc, \src
        mov             \ldst, \dst
        mov             w8, \h

        bl              L(\type\()_hv_filter6_neon_i8mm)
        srshr           v16.8h, v22.8h, #2
        bl              L(\type\()_hv_filter6_neon_i8mm)
        srshr           v17.8h, v22.8h, #2
        bl              L(\type\()_hv_filter6_neon_i8mm)
        srshr           v18.8h, v22.8h, #2
        bl              L(\type\()_hv_filter6_neon_i8mm)
        srshr           v19.8h, v22.8h, #2
        bl              L(\type\()_hv_filter6_neon_i8mm)
        srshr           v20.8h, v22.8h, #2

        .align LOOP_ALIGN
8:
        ld1             {v23.16b}, [\lsrc], \s_strd

        smull           v0.4s, v16.4h, v7.h[1]
        smull2          v1.4s, v16.8h, v7.h[1]
        mov             v16.16b, v17.16b
        movi            v5.4s, #0
        movi            v6.4s, #0
        tbl             v2.16b, {v23.16b}, v29.16b
        tbl             v3.16b, {v23.16b}, v30.16b

        smlal           v0.4s, v17.4h, v7.h[2]
        smlal2          v1.4s, v17.8h, v7.h[2]
        mov             v17.16b, v18.16b

        usmmla          v5.4s, v2.16b, v26.16b
        usmmla          v6.4s, v3.16b, v26.16b

        smlal           v0.4s, v18.4h, v7.h[3]
        smlal2          v1.4s, v18.8h, v7.h[3]
        mov             v18.16b, v19.16b
        subs            w8, w8, #1

        smlal           v0.4s, v19.4h, v7.h[4]
        smlal2          v1.4s, v19.8h, v7.h[4]
        uzp1            v23.8h, v5.8h, v6.8h
        mov             v19.16b, v20.16b

        smlal           v0.4s, v20.4h, v7.h[5]
        smlal2          v1.4s, v20.8h, v7.h[5]
        srshr           v20.8h, v23.8h, #2
        smlal           v0.4s, v20.4h, v7.h[6]
        smlal2          v1.4s, v20.8h, v7.h[6]
    .ifc \type, prep
        rshrn           v0.4h, v0.4s, #6
        rshrn2          v0.8h, v1.4s, #6
        st1             {v0.8h}, [\ldst], \d_strd
        b.gt            8b
        add             \dst, \dst, #16
    .else
        tbl             v0.16b, {v0.16b, v1.16b}, v25.16b
        sqrshrun        v0.8b, v0.8h, #2
        st1             {v0.8b}, [\ldst], \d_strd
        b.gt            8b
        add             \dst, \dst, #8
    .endif
        add             \src, \src, #8
        subs            \w, \w, #8
        b.gt            81b
        ret             x15

        .align JUMP_ALIGN
88:
.endif  // neon_i8mm
        ldp             q29, q30, [x13, #16]

        .align LOOP_ALIGN
81:
        mov             \lsrc, \src
        mov             \ldst, \dst
        mov             w8, \h
.ifc \isa, neon_i8mm
        bl              L(\type\()_hv_filter8_\isa)
        srshr           v16.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        srshr           v17.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        srshr           v18.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        srshr           v19.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        srshr           v20.8h, v22.8h, #2
.else
        bl              L(\type\()_hv_filter8_\isa)
        sshr            v16.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        sshr            v17.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        sshr            v18.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        sshr            v19.8h, v22.8h, #2
        bl              L(\type\()_hv_filter8_\isa)
        sshr            v20.8h, v22.8h, #2
.endif
        .align LOOP_ALIGN
8:
        ldr             q23, [\lsrc]
        add             \lsrc, \lsrc, \s_strd

        smull           v0.4s, v16.4h, v7.h[1]
        smull2          v1.4s, v16.8h, v7.h[1]
.ifc \isa, neon_dotprod
        sub             v23.16b, v23.16b, v24.16b
.endif
        mov             v16.16b, v17.16b
.ifc \isa, neon_i8mm
        movi            v5.4s, #0
        movi            v6.4s, #0
.else
        mov             v5.16b, v27.16b
        mov             v6.16b, v27.16b
.endif
        tbl             v2.16b, {v23.16b}, v28.16b
        tbl             v3.16b, {v23.16b}, v29.16b

        smlal           v0.4s, v17.4h, v7.h[2]
        smlal2          v1.4s, v17.8h, v7.h[2]
        tbl             v4.16b, {v23.16b}, v30.16b
        mov             v17.16b, v18.16b

        \dot            v5.4s, v2.16b, v26.4b[0]
        \dot            v6.4s, v3.16b, v26.4b[0]

        smlal           v0.4s, v18.4h, v7.h[3]
        smlal2          v1.4s, v18.8h, v7.h[3]
        mov             v18.16b, v19.16b

        \dot            v5.4s, v3.16b, v26.4b[1]
        \dot            v6.4s, v4.16b, v26.4b[1]

        smlal           v0.4s, v19.4h, v7.h[4]
        smlal2          v1.4s, v19.8h, v7.h[4]
        mov             v19.16b, v20.16b
        uzp1            v23.8h, v5.8h, v6.8h

        smlal           v0.4s, v20.4h, v7.h[5]
        smlal2          v1.4s, v20.8h, v7.h[5]
.ifc \isa, neon_i8mm
        srshr           v20.8h, v23.8h, #2
.else
        sshr            v20.8h, v23.8h, #2
.endif
        subs            w8, w8, #1
        smlal           v0.4s, v20.4h, v7.h[6]
        smlal2          v1.4s, v20.8h, v7.h[6]
.ifc \type, prep
        rshrn           v0.4h, v0.4s, #6
        rshrn2          v0.8h, v1.4s, #6
        st1             {v0.8h}, [\ldst], \d_strd
        b.gt            8b
        add             \dst, \dst, #16
.else
        tbl             v0.16b, {v0.16b, v1.16b}, v25.16b
        sqrshrun        v0.8b, v0.8h, #2
        st1             {v0.8b}, [\ldst], \d_strd
        b.gt            8b
        add             \dst, \dst, #8
.endif
        add             \src, \src, #8
        subs            \w, \w, #8
        b.gt            81b
        ret             x15

        .align FUNC_ALIGN
L(\type\()_hv_filter8_\isa):
        ld1             {v4.16b}, [\lsrc], \s_strd
.ifc \isa, neon_i8mm
        movi            v22.4s, #0
        movi            v23.4s, #0
.else   // neon_dotprod
        sub             v4.16b, v4.16b, v24.16b
        mov             v22.16b, v27.16b
        mov             v23.16b, v27.16b
.endif
        tbl             v2.16b, {v4.16b}, v28.16b
        tbl             v3.16b, {v4.16b}, v29.16b
        tbl             v4.16b, {v4.16b}, v30.16b
        \dot            v22.4s, v2.16b, v26.4b[0]
        \dot            v23.4s, v3.16b, v26.4b[0]
        \dot            v22.4s, v3.16b, v26.4b[1]
        \dot            v23.4s, v4.16b, v26.4b[1]
        uzp1            v22.8h, v22.8h, v23.8h
        ret

.ifc \isa, neon_i8mm
        .align FUNC_ALIGN
L(\type\()_hv_filter6_neon_i8mm):
        ld1             {v4.16b}, [\lsrc], \s_strd
        movi            v22.4s, #0
        movi            v23.4s, #0
        tbl             v2.16b, {v4.16b}, v29.16b
        tbl             v3.16b, {v4.16b}, v30.16b
        usmmla          v22.4s, v2.16b, v26.16b
        usmmla          v23.4s, v3.16b, v26.16b
        uzp1            v22.8h, v22.8h, v23.8h
        ret
.endif

        .align FUNC_ALIGN
L(\type\()_hv_filter4_\isa):
        ld1             {v4.8b}, [\src], \s_strd
.ifc \isa, neon_i8mm
        movi            v22.4s, #2
.else
        mov             v22.16b, v27.16b
        sub             v4.16b, v4.16b, v24.16b
.endif
        tbl             v2.16b, {v4.16b}, v28.16b
        \dot            v22.4s, v2.16b, v26.4b[0]
        ret

        .align JUMP_ALIGN
40:     // HV6 - 4xN
        ldur            s26, [\xmx, #2]
        add             \src, \src, #2

        bl              L(\type\()_hv_filter4_\isa)
        shrn            v16.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v17.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v18.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v19.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v20.4h, v22.4s, #2

        .align LOOP_ALIGN
4:
        ld1             {v4.8b}, [\src], \s_strd

        smull           v0.4s, v16.4h, v7.h[1]
        smlal           v0.4s, v17.4h, v7.h[2]
.ifc \isa, neon_dotprod
        sub             v4.16b, v4.16b, v24.16b
.endif
        mov             v16.16b, v17.16b
        mov             v17.16b, v18.16b

        smlal           v0.4s, v18.4h, v7.h[3]
        smlal           v0.4s, v19.4h, v7.h[4]
        tbl             v2.16b, {v4.16b}, v28.16b
.ifc \isa, neon_i8mm
        movi            v5.4s, #0
.else
        mov             v5.16b, v27.16b
.endif
        mov             v18.16b, v19.16b
        mov             v19.16b, v20.16b
        \dot            v5.4s, v2.16b, v26.4b[0]

        smlal           v0.4s, v20.4h, v7.h[5]
.ifc \isa, neon_i8mm
        rshrn           v20.4h, v5.4s, #2
.else
        shrn            v20.4h, v5.4s, #2
.endif
        subs            \h, \h, #1
        smlal           v0.4s, v20.4h, v7.h[6]
.ifc \type, prep
        rshrn           v0.4h, v0.4s, #6
        str             d0, [\dst], #8
.else
        tbl             v0.8b, {v0.16b}, v25.8b
        sqrshrun        v0.8b, v0.8h, #2
        str             s0, [\dst]
        add             \dst, \dst, \d_strd
.endif
        b.gt            4b
        ret             x15

.ifc \type, put
        .align JUMP_ALIGN
20:     // HV6 - 2xN
        ldur            s26, [\xmx, #2]
        add             \src, \src, #2

        bl              L(\type\()_hv_filter4_\isa)
        shrn            v16.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v17.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v18.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v19.4h, v22.4s, #2
        bl              L(\type\()_hv_filter4_\isa)
        shrn            v20.4h, v22.4s, #2

        .align LOOP_ALIGN
2:
        ld1             {v4.8b}, [\src], \s_strd

        smull           v0.4s, v16.4h, v7.h[1]
        smlal           v0.4s, v17.4h, v7.h[2]
    .ifc \isa, neon_dotprod
        sub             v4.16b, v4.16b, v24.16b
    .endif
        mov             v16.16b, v17.16b
        mov             v17.16b, v18.16b

        smlal           v0.4s, v18.4h, v7.h[3]
        smlal           v0.4s, v19.4h, v7.h[4]
        tbl             v2.16b, {v4.16b}, v28.16b
    .ifc \isa, neon_i8mm
        movi            v5.4s, #0
    .else
        mov             v5.16b, v27.16b
    .endif

        mov             v18.16b, v19.16b
        mov             v19.16b, v20.16b
        \dot            v5.4s, v2.16b, v26.4b[0]

        smlal           v0.4s, v20.4h, v7.h[5]
    .ifc \isa, neon_i8mm
        rshrn           v20.4h, v5.4s, #2
    .else
        shrn            v20.4h, v5.4s, #2
    .endif

        subs            \h, \h, #1
        smlal           v0.4s, v20.4h, v7.h[6]

        tbl             v0.8b, {v0.16b}, v25.8b
        sqrshrun        v0.8b, v0.8h, #2

        str             h0, [\dst]
        add             \dst, \dst, \d_strd
        b.gt            2b
        ret             x15
.endif

        .align JUMP_ALIGN
L(\type\()_8tap_h_\isa):
        movrel          x11, \type\()_8tap_h_\isa\()_tbl
        ldrsw           x8, [x11, x8, lsl #2]
.ifc \type, put
    .ifc \isa, neon_i8mm
        movi            v27.4s, #34     // special rounding
    .else
        mov             w10, #0x2022    // 64 * 128 + 34, bias and rounding for SDOT
        dup             v27.4s, w10
    .endif
.endif
        add             x11, x11, x8
        br              x11

.ifc \type, put
        .align JUMP_ALIGN
20:     // H - 2xN
        AARCH64_VALID_JUMP_TARGET
        add             \src, \src, #2
        ldur            s26, [\xmx, #2]

        .align LOOP_ALIGN
2:
        ldr             d0, [\src]
        ldr             d1, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
    .ifc \isa, neon_dotprod
        sub             v0.8b, v0.8b, v24.8b
        sub             v1.8b, v1.8b, v24.8b
    .endif
        mov             v4.16b, v27.16b
        mov             v5.16b, v27.16b

        tbl             v2.16b, {v0.16b}, v28.16b
        tbl             v3.16b, {v1.16b}, v28.16b

        \dot            v4.4s, v2.16b, v26.4b[0]
        \dot            v5.4s, v3.16b, v26.4b[0]

        uzp1            v4.8h, v4.8h, v5.8h
        sqshrun         v4.8b, v4.8h, #6

        subs            \h, \h, #2
        fmov            x8, d4
        lsr             x9, x8, #32
        strh            w8, [\dst]
        strh            w9, [\dst, \d_strd]
        add             \dst, \dst, \d_strd, lsl #1
        b.gt            2b
        ret
.endif

        .align JUMP_ALIGN
40:     // H - 4xN
        AARCH64_VALID_JUMP_TARGET
        add             \src, \src, #2
        ldur            s26, [\xmx, #2]

        .align LOOP_ALIGN
4:
        ldr             d0, [\src]
        ldr             d1, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
.ifc \type\()_\isa, prep_neon_i8mm
        movi            v4.4s, #0
        movi            v5.4s, #0
.else
    .ifc \isa, neon_dotprod
        sub             v0.8b, v0.8b, v24.8b
        sub             v1.8b, v1.8b, v24.8b
    .endif
        mov             v4.16b, v27.16b
        mov             v5.16b, v27.16b
.endif
        tbl             v2.16b, {v0.16b}, v28.16b
        tbl             v3.16b, {v1.16b}, v28.16b

        \dot            v4.4s, v2.16b, v26.4b[0]
        \dot            v5.4s, v3.16b, v26.4b[0]
.ifc \type, prep
        subs            \h, \h, #2
    .ifc \isa, neon_i8mm
        uzp1            v4.8h, v4.8h, v5.8h
        srshr           v4.8h, v4.8h, #2
    .else
        shrn            v4.4h, v4.4s, #2
        shrn2           v4.8h, v5.4s, #2
    .endif
        str             q4, [\dst], #16
.else   // put
        uzp1            v4.8h, v4.8h, v5.8h
        sqshrun         v4.8b, v4.8h, #6
        subs            \h, \h, #2
        fmov            x8, d4
        lsr             x9, x8, #32
        str             w8, [\dst]
        str             w9, [\dst, \d_strd]
        add             \dst, \dst, \d_strd, lsl #1
.endif
        b.gt            4b
        ret

        .align JUMP_ALIGN
80:     // H - 8xN
        AARCH64_VALID_JUMP_TARGET
        ldr             d26, [\xmx]
.ifc \isa, neon_i8mm
        cmp             w9, #SHARP1
        b.eq            88f             // horizontal == SHARP1

        ldp             q29, q30, [x13, #(OFFSET_USMMLA)]
        ext             v0.8b, v26.8b, v26.8b, #7
        ins             v26.d[1], v0.d[0]

        .align LOOP_ALIGN
8:
        ldr             q0, [\src]
        ldr             q16, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
    .ifc \type, prep
        movi            v4.4s, #0
        movi            v5.4s, #0
        movi            v20.4s, #0
        movi            v21.4s, #0
    .else
        mov             v4.16b, v27.16b
        mov             v5.16b, v27.16b
        mov             v20.16b, v27.16b
        mov             v21.16b, v27.16b
    .endif
        tbl             v1.16b, {v0.16b}, v29.16b
        tbl             v2.16b, {v0.16b}, v30.16b
        tbl             v17.16b, {v16.16b}, v29.16b
        tbl             v18.16b, {v16.16b}, v30.16b

        usmmla          v4.4s, v1.16b, v26.16b
        usmmla          v5.4s, v2.16b, v26.16b
        usmmla          v20.4s, v17.16b, v26.16b
        usmmla          v21.4s, v18.16b, v26.16b

        uzp1            v4.8h, v4.8h, v5.8h
        uzp1            v20.8h, v20.8h, v21.8h
    .ifc \type, prep
        srshr           v4.8h, v4.8h, #2
        srshr           v20.8h, v20.8h, #2
        subs            \h, \h, #2
        stp             q4, q20, [\dst], #32
    .else   // put
        sqshrun         v4.8b, v4.8h, #6
        sqshrun         v20.8b, v20.8h, #6
        subs            \h, \h, #2
        str             d4, [\dst]
        str             d20, [\dst, \d_strd]
        add             \dst, \dst, \d_strd, lsl #1
    .endif
        b.gt            8b
        ret

        .align JUMP_ALIGN
88:
.endif  // neon_i8mm
        ldp             q29, q30, [x13, #16]

        .align LOOP_ALIGN
8:
        ldr             q0, [\src]
        ldr             q16, [\src, \s_strd]
        add             \src, \src, \s_strd, lsl #1
.ifc \type\()_\isa, prep_neon_i8mm
        movi            v4.4s, #0
        movi            v5.4s, #0
        movi            v20.4s, #0
        movi            v21.4s, #0
.else
    .ifc \isa, neon_dotprod
        sub             v0.16b, v0.16b, v24.16b
        sub             v16.16b, v16.16b, v24.16b
    .endif
        mov             v4.16b, v27.16b
        mov             v5.16b, v27.16b
        mov             v20.16b, v27.16b
        mov             v21.16b, v27.16b
.endif
        tbl             v1.16b, {v0.16b}, v28.16b
        tbl             v2.16b, {v0.16b}, v29.16b
        tbl             v3.16b, {v0.16b}, v30.16b
        tbl             v17.16b, {v16.16b}, v28.16b
        tbl             v18.16b, {v16.16b}, v29.16b
        tbl             v19.16b, {v16.16b}, v30.16b

        \dot            v4.4s, v1.16b, v26.4b[0]
        \dot            v5.4s, v2.16b, v26.4b[0]
        \dot            v20.4s, v17.16b, v26.4b[0]
        \dot            v21.4s, v18.16b, v26.4b[0]
        \dot            v4.4s, v2.16b, v26.4b[1]
        \dot            v5.4s, v3.16b, v26.4b[1]
        \dot            v20.4s, v18.16b, v26.4b[1]
        \dot            v21.4s, v19.16b, v26.4b[1]

        uzp1            v4.8h, v4.8h, v5.8h
        uzp1            v20.8h, v20.8h, v21.8h
.ifc \type, prep
    .ifc \isa, neon_i8mm
        srshr           v4.8h, v4.8h, #2
        srshr           v20.8h, v20.8h, #2
    .else
        sshr            v4.8h, v4.8h, #2
        sshr            v20.8h, v20.8h, #2
    .endif
        subs            \h, \h, #2
        stp             q4, q20, [\dst], #32
.else   // put
        sqshrun         v4.8b, v4.8h, #6
        sqshrun         v20.8b, v20.8h, #6
        subs            \h, \h, #2
        str             d4, [\dst]
        str             d20, [\dst, \d_strd]
        add             \dst, \dst, \d_strd, lsl #1
.endif
        b.gt            8b
        ret

        .align JUMP_ALIGN
160:    // H - 16xN
        AARCH64_VALID_JUMP_TARGET
        ldr             d26, [\xmx]
.ifc \isa, neon_i8mm
        cmp             w9, #SHARP1
        b.eq            168f            // horizontal == SHARP1

        ldp             q29, q30, [x13, #(OFFSET_USMMLA)]
        ext             v0.8b, v26.8b, v26.8b, #7
        ins             v26.d[1], v0.d[0]

        .align LOOP_ALIGN
16:
        ldr             q16, [\src]
        ldur            q17, [\src, #8] // avoid 2 register TBL for small cores
        add             \src, \src, \s_strd
    .ifc \type, prep
        movi            v6.4s, #0
        movi            v7.4s, #0
        movi            v22.4s, #0
        movi            v23.4s, #0
    .else
        mov             v6.16b, v27.16b
        mov             v7.16b, v27.16b
        mov             v22.16b, v27.16b
        mov             v23.16b, v27.16b
    .endif
        tbl             v0.16b, {v16.16b}, v29.16b
        tbl             v1.16b, {v16.16b}, v30.16b
        tbl             v2.16b, {v17.16b}, v29.16b
        tbl             v3.16b, {v17.16b}, v30.16b

        usmmla          v6.4s, v0.16b, v26.16b
        usmmla          v7.4s, v1.16b, v26.16b
        usmmla          v22.4s, v2.16b, v26.16b
        usmmla          v23.4s, v3.16b, v26.16b

        uzp1            v6.8h, v6.8h, v7.8h
        uzp1            v22.8h, v22.8h, v23.8h
    .ifc \type, prep
        srshr           v6.8h, v6.8h, #2
        srshr           v22.8h, v22.8h, #2
        subs            \h, \h, #1
        stp             q6, q22, [\dst], #32
    .else   // put
        sqshrun         v6.8b, v6.8h, #6
        sqshrun2        v6.16b, v22.8h, #6
        subs            \h, \h, #1
        st1             {v6.16b}, [\dst], \d_strd
    .endif
        b.gt            16b
        ret

        .align JUMP_ALIGN
168:
.endif  // neon_i8mm
        ldp             q29, q30, [x13, #16]

        .align LOOP_ALIGN
16:
        ldr             q16, [\src]
        ldur            q17, [\src, #12]  // avoid 2 register TBL for small cores
        add             \src, \src, \s_strd
.ifc \type\()_\isa, prep_neon_i8mm
        movi            v6.4s, #0
        movi            v7.4s, #0
        movi            v22.4s, #0
        movi            v23.4s, #0
.else
    .ifc \isa, neon_dotprod
        sub             v16.16b, v16.16b, v24.16b
        sub             v17.16b, v17.16b, v24.16b
    .endif
        mov             v6.16b, v27.16b
        mov             v7.16b, v27.16b
        mov             v22.16b, v27.16b
        mov             v23.16b, v27.16b
.endif
        tbl             v0.16b, {v16.16b}, v28.16b
        tbl             v1.16b, {v16.16b}, v29.16b
        tbl             v2.16b, {v16.16b}, v30.16b
        tbl             v3.16b, {v17.16b}, v28.16b
        tbl             v4.16b, {v17.16b}, v29.16b

        \dot            v6.4s, v0.16b, v26.4b[0]
        \dot            v7.4s, v1.16b, v26.4b[0]
        \dot            v22.4s, v2.16b, v26.4b[0]
        \dot            v23.4s, v3.16b, v26.4b[0]
        \dot            v6.4s, v1.16b, v26.4b[1]
        \dot            v7.4s, v2.16b, v26.4b[1]
        \dot            v22.4s, v3.16b, v26.4b[1]
        \dot            v23.4s, v4.16b, v26.4b[1]

        uzp1            v6.8h, v6.8h, v7.8h
        uzp1            v22.8h, v22.8h, v23.8h
.ifc \type, prep
    .ifc \isa, neon_i8mm
        srshr           v6.8h, v6.8h, #2
        srshr           v22.8h, v22.8h, #2
    .else
        sshr            v6.8h, v6.8h, #2
        sshr            v22.8h, v22.8h, #2
    .endif
        subs            \h, \h, #1
        stp             q6, q22, [\dst], #32
.else   // put
        sqshrun         v6.8b, v6.8h, #6
        sqshrun2        v6.16b, v22.8h, #6
        subs            \h, \h, #1
        st1             {v6.16b}, [\dst], \d_strd
.endif
        b.gt            16b
        ret

        .align JUMP_ALIGN
320:    // H - 32xN+
640:
1280:
        AARCH64_VALID_JUMP_TARGET
        ldr             d26, [\xmx]
.ifc \type, put
        sub             \d_strd, \d_strd, \w, uxtw
.endif
        sub             \s_strd, \s_strd, \w, uxtw
        mov             w8, \w

.ifc \isa, neon_i8mm
        cmp             w9, #SHARP1
        b.eq            328f            // horizontal == SHARP1

        ldp             q29, q30, [x13, #(OFFSET_USMMLA)]
        ext             v0.8b, v26.8b, v26.8b, #7
        ins             v26.d[1], v0.d[0]

        .align LOOP_ALIGN
32:
        ldr             q16, [\src]
        ldur            q17, [\src, #8] // avoid 2 register TBL for small cores
        add             \src, \src, #16
    .ifc \type, prep
        movi            v6.4s, #0
        movi            v7.4s, #0
        movi            v22.4s, #0
        movi            v23.4s, #0
    .else
        mov             v6.16b, v27.16b
        mov             v7.16b, v27.16b
        mov             v22.16b, v27.16b
        mov             v23.16b, v27.16b
    .endif
        tbl             v0.16b, {v16.16b}, v29.16b
        tbl             v1.16b, {v16.16b}, v30.16b
        tbl             v2.16b, {v17.16b}, v29.16b
        tbl             v3.16b, {v17.16b}, v30.16b

        usmmla          v6.4s, v0.16b, v26.16b
        usmmla          v7.4s, v1.16b, v26.16b
        usmmla          v22.4s, v2.16b, v26.16b
        usmmla          v23.4s, v3.16b, v26.16b

        uzp1            v6.8h, v6.8h, v7.8h
        uzp1            v22.8h, v22.8h, v23.8h
    .ifc \type, prep
        srshr           v6.8h, v6.8h, #2
        srshr           v22.8h, v22.8h, #2
        subs            w8, w8, #16
        stp             q6, q22, [\dst], #32
    .else   // put
        sqshrun         v6.8b, v6.8h, #6
        sqshrun2        v6.16b, v22.8h, #6
        subs            w8, w8, #16
        str             q6, [\dst], #16
    .endif
        b.gt            32b

        add             \src, \src, \s_strd
    .ifc \type, put
        add             \dst, \dst, \d_strd
    .endif
        mov             w8, \w
        subs            \h, \h, #1
        b.gt            32b
        ret

        .align JUMP_ALIGN
328:
.endif  // neon_i8mm
        ldp             q29, q30, [x13, #16]

        .align LOOP_ALIGN
32:
        ldr             q16, [\src]
        ldur            q17, [\src, #12]  // avoid 2 register TBL for small cores
        add             \src, \src, #16
.ifc \type\()_\isa, prep_neon_i8mm
        movi            v6.4s, #0
        movi            v7.4s, #0
        movi            v22.4s, #0
        movi            v23.4s, #0
.else
    .ifc \isa, neon_dotprod
        sub             v16.16b, v16.16b, v24.16b
        sub             v17.16b, v17.16b, v24.16b
    .endif
        mov             v6.16b, v27.16b
        mov             v7.16b, v27.16b
        mov             v22.16b, v27.16b
        mov             v23.16b, v27.16b
.endif
        tbl             v0.16b, {v16.16b}, v28.16b
        tbl             v1.16b, {v16.16b}, v29.16b
        tbl             v2.16b, {v16.16b}, v30.16b
        tbl             v3.16b, {v17.16b}, v28.16b
        tbl             v4.16b, {v17.16b}, v29.16b

        \dot            v6.4s, v0.16b, v26.4b[0]
        \dot            v7.4s, v1.16b, v26.4b[0]
        \dot            v22.4s, v2.16b, v26.4b[0]
        \dot            v23.4s, v3.16b, v26.4b[0]
        \dot            v6.4s, v1.16b, v26.4b[1]
        \dot            v7.4s, v2.16b, v26.4b[1]
        \dot            v22.4s, v3.16b, v26.4b[1]
        \dot            v23.4s, v4.16b, v26.4b[1]

        uzp1            v6.8h, v6.8h, v7.8h
        uzp1            v22.8h, v22.8h, v23.8h
.ifc \type, prep
    .ifc \isa, neon_i8mm
        srshr           v6.8h, v6.8h, #2
        srshr           v22.8h, v22.8h, #2
    .else
        sshr            v6.8h, v6.8h, #2
        sshr            v22.8h, v22.8h, #2
    .endif
        subs            w8, w8, #16
        stp             q6, q22, [\dst], #32
.else   // put
        sqshrun         v6.8b, v6.8h, #6
        sqshrun2        v6.16b, v22.8h, #6
        subs            w8, w8, #16
        str             q6, [\dst], #16
.endif
        b.gt            32b

        add             \src, \src, \s_strd
.ifc \type, put
        add             \dst, \dst, \d_strd
.endif
        mov             w8, \w
        subs            \h, \h, #1
        b.gt            32b
        ret
endfunc

jumptable \type\()_8tap_h_\isa\()_tbl
        .word 1280b - \type\()_8tap_h_\isa\()_tbl
        .word 640b  - \type\()_8tap_h_\isa\()_tbl
        .word 320b  - \type\()_8tap_h_\isa\()_tbl
        .word 160b  - \type\()_8tap_h_\isa\()_tbl
        .word 80b   - \type\()_8tap_h_\isa\()_tbl
        .word 40b   - \type\()_8tap_h_\isa\()_tbl
.ifc \type, put
        .word 20b   - \type\()_8tap_h_\isa\()_tbl
.endif
endjumptable
.endm

// dst(x0), d_strd(x7), src(x1), s_strd(x2), w(w3), h(w4), mx(w5), my(w6)
// xmx(x5), xmy(x6), ldst(x5), lsrc(x6), wd_strd(w7)
filter_8tap_fn prep, sdot, neon_dotprod, x0, x7, x1, x2, w3, w4, w5, w6, x5, x6, x5, x6, w7

// dst(x0) d_strd(x1) src(x2) s_strd(x3) w(w4) h(w5) mx(w6) my(w7)
// xmx(x6), xmy(x7), ldst(x6), lsrc(x7), wd_strd(w1)
filter_8tap_fn  put, sdot, neon_dotprod, x0, x1, x2, x3, w4, w5, w6, w7, x6, x7, x6, x7, w1

#if HAVE_I8MM
ENABLE_I8MM

// dst(x0), d_strd(x7), src(x1), s_strd(x2), w(w3), h(w4), mx(w5), my(w6)
// xmx(x5), xmy(x6), ldst(x5), lsrc(x6), wd_strd(w7)
filter_8tap_fn prep, usdot, neon_i8mm, x0, x7, x1, x2, w3, w4, w5, w6, x5, x6, x5, x6, w7

// dst(x0) d_strd(x1) src(x2) s_strd(x3) w(w4) h(w5) mx(w6) my(w7)
// xmx(x6), xmy(x7), ldst(x6), lsrc(x7), wd_strd(w1)
filter_8tap_fn  put, usdot, neon_i8mm, x0, x1, x2, x3, w4, w5, w6, w7, x6, x7, x6, x7, w1

DISABLE_I8MM
#endif  // HAVE_I8MM

DISABLE_DOTPROD
#endif  // HAVE_DOTPROD