/*
 * ARMv8 NEON optimizations for libjpeg-turbo
 *
 * Copyright (C) 2009-2011 Nokia Corporation and/or its subsidiary(-ies).
 * All rights reserved.
 * Author: Siarhei Siamashka <siarhei.siamashka@nokia.com>
 * Copyright (C) 2013-2014, Linaro Limited
 * Author: Ragesh Radhakrishnan <ragesh.r@linaro.org>
 * Copyright (C) 2014-2016, D. R. Commander.  All Rights Reserved.
 * Copyright (C) 2015-2016, Matthieu Darbois.  All Rights Reserved.
 * Copyright (C) 2016, Siarhei Siamashka.  All Rights Reserved.
 *
 * This software is provided 'as-is', without any express or implied
 * warranty.  In no event will the authors be held liable for any damages
 * arising from the use of this software.
 *
 * Permission is granted to anyone to use this software for any purpose,
 * including commercial applications, and to alter it and redistribute it
 * freely, subject to the following restrictions:
 *
 * 1. The origin of this software must not be misrepresented; you must not
 *    claim that you wrote the original software. If you use this software
 *    in a product, an acknowledgment in the product documentation would be
 *    appreciated but is not required.
 * 2. Altered source versions must be plainly marked as such, and must not be
 *    misrepresented as being the original software.
 * 3. This notice may not be removed or altered from any source distribution.
 */

#if defined(__linux__) && defined(__ELF__)
.section .note.GNU-stack, "", %progbits  /* mark stack as non-executable */
#endif

.text


#define RESPECT_STRICT_ALIGNMENT 1


/*****************************************************************************/

/* Supplementary macro for setting function attributes */
.macro asm_function fname
#ifdef __APPLE__
    .globl _\fname
_\fname:
#else
    .global \fname
#ifdef __ELF__
    .hidden \fname
    .type \fname, %function
#endif
\fname:
#endif
.endm

/* Transpose elements of single 128 bit registers */
.macro transpose_single x0, x1, xi, xilen, literal
    ins             \xi\xilen[0], \x0\xilen[0]
    ins             \x1\xilen[0], \x0\xilen[1]
    trn1            \x0\literal, \x0\literal, \x1\literal
    trn2            \x1\literal, \xi\literal, \x1\literal
.endm

/* Transpose elements of 2 differnet registers */
.macro transpose x0, x1, xi, xilen, literal
    mov             \xi\xilen, \x0\xilen
    trn1            \x0\literal, \x0\literal, \x1\literal
    trn2            \x1\literal, \xi\literal, \x1\literal
.endm

/* Transpose a block of 4x4 coefficients in four 64-bit registers */
.macro transpose_4x4_32 x0, x0len, x1, x1len, x2, x2len, x3, x3len, xi, xilen
    mov             \xi\xilen, \x0\xilen
    trn1            \x0\x0len, \x0\x0len, \x2\x2len
    trn2            \x2\x2len, \xi\x0len, \x2\x2len
    mov             \xi\xilen, \x1\xilen
    trn1            \x1\x1len, \x1\x1len, \x3\x3len
    trn2            \x3\x3len, \xi\x1len, \x3\x3len
.endm

.macro transpose_4x4_16 x0, x0len, x1, x1len, x2, x2len, x3, x3len, xi, xilen
    mov             \xi\xilen, \x0\xilen
    trn1            \x0\x0len, \x0\x0len, \x1\x1len
    trn2            \x1\x2len, \xi\x0len, \x1\x2len
    mov             \xi\xilen, \x2\xilen
    trn1            \x2\x2len, \x2\x2len, \x3\x3len
    trn2            \x3\x2len, \xi\x1len, \x3\x3len
.endm

.macro transpose_4x4 x0, x1, x2, x3, x5
    transpose_4x4_16 \x0, .4h, \x1, .4h, \x2, .4h, \x3, .4h, \x5, .16b
    transpose_4x4_32 \x0, .2s, \x1, .2s, \x2, .2s, \x3, .2s, \x5, .16b
.endm

.macro transpose_8x8 l0, l1, l2, l3, l4, l5, l6, l7, t0, t1, t2, t3
    trn1            \t0\().8h, \l0\().8h, \l1\().8h
    trn1            \t1\().8h, \l2\().8h, \l3\().8h
    trn1            \t2\().8h, \l4\().8h, \l5\().8h
    trn1            \t3\().8h, \l6\().8h, \l7\().8h
    trn2            \l1\().8h, \l0\().8h, \l1\().8h
    trn2            \l3\().8h, \l2\().8h, \l3\().8h
    trn2            \l5\().8h, \l4\().8h, \l5\().8h
    trn2            \l7\().8h, \l6\().8h, \l7\().8h

    trn1            \l4\().4s, \t2\().4s, \t3\().4s
    trn2            \t3\().4s, \t2\().4s, \t3\().4s
    trn1            \t2\().4s, \t0\().4s, \t1\().4s
    trn2            \l2\().4s, \t0\().4s, \t1\().4s
    trn1            \t0\().4s, \l1\().4s, \l3\().4s
    trn2            \l3\().4s, \l1\().4s, \l3\().4s
    trn2            \t1\().4s, \l5\().4s, \l7\().4s
    trn1            \l5\().4s, \l5\().4s, \l7\().4s

    trn2            \l6\().2d, \l2\().2d, \t3\().2d
    trn1            \l0\().2d, \t2\().2d, \l4\().2d
    trn1            \l1\().2d, \t0\().2d, \l5\().2d
    trn2            \l7\().2d, \l3\().2d, \t1\().2d
    trn1            \l2\().2d, \l2\().2d, \t3\().2d
    trn2            \l4\().2d, \t2\().2d, \l4\().2d
    trn1            \l3\().2d, \l3\().2d, \t1\().2d
    trn2            \l5\().2d, \t0\().2d, \l5\().2d
.endm


#define CENTERJSAMPLE 128

/*****************************************************************************/

/*
 * Perform dequantization and inverse DCT on one block of coefficients.
 *
 * GLOBAL(void)
 * jsimd_idct_islow_neon (void * dct_table, JCOEFPTR coef_block,
 *                        JSAMPARRAY output_buf, JDIMENSION output_col)
 */

#define CONST_BITS 13
#define PASS1_BITS 2

#define F_0_298  2446  /* FIX(0.298631336) */
#define F_0_390  3196  /* FIX(0.390180644) */
#define F_0_541  4433  /* FIX(0.541196100) */
#define F_0_765  6270  /* FIX(0.765366865) */
#define F_0_899  7373  /* FIX(0.899976223) */
#define F_1_175  9633  /* FIX(1.175875602) */
#define F_1_501 12299  /* FIX(1.501321110) */
#define F_1_847 15137  /* FIX(1.847759065) */
#define F_1_961 16069  /* FIX(1.961570560) */
#define F_2_053 16819  /* FIX(2.053119869) */
#define F_2_562 20995  /* FIX(2.562915447) */
#define F_3_072 25172  /* FIX(3.072711026) */

.balign 16
Ljsimd_idct_islow_neon_consts:
  .short F_0_298
  .short -F_0_390
  .short F_0_541
  .short F_0_765
  .short - F_0_899
  .short F_1_175
  .short F_1_501
  .short - F_1_847
  .short - F_1_961
  .short F_2_053
  .short - F_2_562
  .short F_3_072
  .short 0          /* padding */
  .short 0
  .short 0
  .short 0

#undef F_0_298
#undef F_0_390
#undef F_0_541
#undef F_0_765
#undef F_0_899
#undef F_1_175
#undef F_1_501
#undef F_1_847
#undef F_1_961
#undef F_2_053
#undef F_2_562
#undef F_3_072

#define XFIX_P_0_298 v0.h[0]
#define XFIX_N_0_390 v0.h[1]
#define XFIX_P_0_541 v0.h[2]
#define XFIX_P_0_765 v0.h[3]
#define XFIX_N_0_899 v0.h[4]
#define XFIX_P_1_175 v0.h[5]
#define XFIX_P_1_501 v0.h[6]
#define XFIX_N_1_847 v0.h[7]
#define XFIX_N_1_961 v1.h[0]
#define XFIX_P_2_053 v1.h[1]
#define XFIX_N_2_562 v1.h[2]
#define XFIX_P_3_072 v1.h[3]

asm_function jsimd_idct_islow_neon
    DCT_TABLE       .req x0
    COEF_BLOCK      .req x1
    OUTPUT_BUF      .req x2
    OUTPUT_COL      .req x3
    TMP1            .req x0
    TMP2            .req x1
    TMP3            .req x9
    TMP4            .req x10
    TMP5            .req x11
    TMP6            .req x12
    TMP7            .req x13
    TMP8            .req x14

    sub             sp, sp, #64
    adr             x15, Ljsimd_idct_islow_neon_consts
    st1             {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], #32
    st1             {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], #32
    ld1             {v0.8h, v1.8h}, [x15]
    ld1             {v2.8h, v3.8h, v4.8h, v5.8h}, [COEF_BLOCK], #64
    ld1             {v18.8h, v19.8h, v20.8h, v21.8h}, [DCT_TABLE], #64
    ld1             {v6.8h, v7.8h, v8.8h, v9.8h}, [COEF_BLOCK], #64
    ld1             {v22.8h, v23.8h, v24.8h, v25.8h}, [DCT_TABLE], #64

    cmeq            v16.8h, v3.8h, #0
    cmeq            v26.8h, v4.8h, #0
    cmeq            v27.8h, v5.8h, #0
    cmeq            v28.8h, v6.8h, #0
    cmeq            v29.8h, v7.8h, #0
    cmeq            v30.8h, v8.8h, #0
    cmeq            v31.8h, v9.8h, #0

    and             v10.16b, v16.16b, v26.16b
    and             v11.16b, v27.16b, v28.16b
    and             v12.16b, v29.16b, v30.16b
    and             v13.16b, v31.16b, v10.16b
    and             v14.16b, v11.16b, v12.16b
    mul             v2.8h, v2.8h, v18.8h
    and             v15.16b, v13.16b, v14.16b
    shl             v10.8h, v2.8h, #(PASS1_BITS)
    sqxtn           v16.8b, v15.8h
    mov             TMP1, v16.d[0]
    sub             sp, sp, #64
    mvn             TMP2, TMP1

    cbnz            TMP2, 2f
    /* case all AC coeffs are zeros */
    dup             v2.2d, v10.d[0]
    dup             v6.2d, v10.d[1]
    mov             v3.16b, v2.16b
    mov             v7.16b, v6.16b
    mov             v4.16b, v2.16b
    mov             v8.16b, v6.16b
    mov             v5.16b, v2.16b
    mov             v9.16b, v6.16b
1:
    /* for this transpose, we should organise data like this:
     * 00, 01, 02, 03, 40, 41, 42, 43
     * 10, 11, 12, 13, 50, 51, 52, 53
     * 20, 21, 22, 23, 60, 61, 62, 63
     * 30, 31, 32, 33, 70, 71, 72, 73
     * 04, 05, 06, 07, 44, 45, 46, 47
     * 14, 15, 16, 17, 54, 55, 56, 57
     * 24, 25, 26, 27, 64, 65, 66, 67
     * 34, 35, 36, 37, 74, 75, 76, 77
     */
    trn1            v28.8h, v2.8h, v3.8h
    trn1            v29.8h, v4.8h, v5.8h
    trn1            v30.8h, v6.8h, v7.8h
    trn1            v31.8h, v8.8h, v9.8h
    trn2            v16.8h, v2.8h, v3.8h
    trn2            v17.8h, v4.8h, v5.8h
    trn2            v18.8h, v6.8h, v7.8h
    trn2            v19.8h, v8.8h, v9.8h
    trn1            v2.4s, v28.4s, v29.4s
    trn1            v6.4s, v30.4s, v31.4s
    trn1            v3.4s, v16.4s, v17.4s
    trn1            v7.4s, v18.4s, v19.4s
    trn2            v4.4s, v28.4s, v29.4s
    trn2            v8.4s, v30.4s, v31.4s
    trn2            v5.4s, v16.4s, v17.4s
    trn2            v9.4s, v18.4s, v19.4s
    /* Even part: reverse the even part of the forward DCT. */
    add             v18.8h, v4.8h, v8.8h           /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]) + DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]) */
    add             v22.8h, v2.8h, v6.8h           /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) + DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */
    smull2          v19.4s, v18.8h, XFIX_P_0_541   /* z1h z1 = MULTIPLY(z2 + z3, FIX_0_541196100); */
    sub             v26.8h, v2.8h, v6.8h           /* z2 - z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) - DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */
    smull           v18.4s, v18.4h, XFIX_P_0_541   /* z1l z1 = MULTIPLY(z2 + z3, FIX_0_541196100); */
    sshll2          v23.4s, v22.8h, #(CONST_BITS)  /* tmp0h tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS); */
    mov             v21.16b, v19.16b               /* tmp3 = z1 */
    mov             v20.16b, v18.16b               /* tmp3 = z1 */
    smlal2          v19.4s, v8.8h, XFIX_N_1_847    /* tmp2h tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065); */
    smlal           v18.4s, v8.4h, XFIX_N_1_847    /* tmp2l tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065); */
    sshll2          v27.4s, v26.8h, #(CONST_BITS)  /* tmp1h tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS); */
    smlal2          v21.4s, v4.8h, XFIX_P_0_765    /* tmp3h tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865); */
    smlal           v20.4s, v4.4h, XFIX_P_0_765    /* tmp3l tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865); */
    sshll           v22.4s, v22.4h, #(CONST_BITS)  /* tmp0l tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS); */
    sshll           v26.4s, v26.4h, #(CONST_BITS)  /* tmp1l tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS); */
    add             v2.4s, v22.4s, v20.4s          /* tmp10l tmp10 = tmp0 + tmp3; */
    sub             v6.4s, v22.4s, v20.4s          /* tmp13l tmp13 = tmp0 - tmp3; */
    add             v8.4s, v26.4s, v18.4s          /* tmp11l tmp11 = tmp1 + tmp2; */
    sub             v4.4s, v26.4s, v18.4s          /* tmp12l tmp12 = tmp1 - tmp2; */
    add             v28.4s, v23.4s, v21.4s         /* tmp10h tmp10 = tmp0 + tmp3; */
    sub             v31.4s, v23.4s, v21.4s         /* tmp13h tmp13 = tmp0 - tmp3; */
    add             v29.4s, v27.4s, v19.4s         /* tmp11h tmp11 = tmp1 + tmp2; */
    sub             v30.4s, v27.4s, v19.4s         /* tmp12h tmp12 = tmp1 - tmp2; */

    /* Odd part per figure 8; the matrix is unitary and hence its
     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
     */

    add             v22.8h, v9.8h, v5.8h    /* z3 = tmp0 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */
    add             v24.8h, v7.8h, v3.8h    /* z4 = tmp1 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */
    add             v18.8h, v9.8h, v3.8h    /* z1 = tmp0 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */
    add             v20.8h, v7.8h, v5.8h    /* z2 = tmp1 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */
    add             v26.8h, v22.8h, v24.8h  /* z5 = z3 + z4 */

    smull2          v11.4s, v9.8h, XFIX_P_0_298   /* tmp0 = MULTIPLY(tmp0, FIX_0_298631336) */
    smull2          v13.4s, v7.8h, XFIX_P_2_053   /* tmp1 = MULTIPLY(tmp1, FIX_2_053119869) */
    smull2          v15.4s, v5.8h, XFIX_P_3_072   /* tmp2 = MULTIPLY(tmp2, FIX_3_072711026) */
    smull2          v17.4s, v3.8h, XFIX_P_1_501   /* tmp3 = MULTIPLY(tmp3, FIX_1_501321110) */
    smull2          v27.4s, v26.8h, XFIX_P_1_175  /* z5h z5 = MULTIPLY(z3 + z4, FIX_1_175875602) */
    smull2          v23.4s, v22.8h, XFIX_N_1_961  /* z3 = MULTIPLY(z3, - FIX_1_961570560) */
    smull2          v25.4s, v24.8h, XFIX_N_0_390  /* z4 = MULTIPLY(z4, - FIX_0_390180644) */
    smull2          v19.4s, v18.8h, XFIX_N_0_899  /* z1 = MULTIPLY(z1, - FIX_0_899976223) */
    smull2          v21.4s, v20.8h, XFIX_N_2_562  /* z2 = MULTIPLY(z2, - FIX_2_562915447) */

    smull           v10.4s, v9.4h, XFIX_P_0_298   /* tmp0 = MULTIPLY(tmp0, FIX_0_298631336) */
    smull           v12.4s, v7.4h, XFIX_P_2_053   /* tmp1 = MULTIPLY(tmp1, FIX_2_053119869) */
    smull           v14.4s, v5.4h, XFIX_P_3_072   /* tmp2 = MULTIPLY(tmp2, FIX_3_072711026) */
    smull           v16.4s, v3.4h, XFIX_P_1_501   /* tmp3 = MULTIPLY(tmp3, FIX_1_501321110) */
    smull           v26.4s, v26.4h, XFIX_P_1_175  /* z5l z5 = MULTIPLY(z3 + z4, FIX_1_175875602) */
    smull           v22.4s, v22.4h, XFIX_N_1_961  /* z3 = MULTIPLY(z3, - FIX_1_961570560) */
    smull           v24.4s, v24.4h, XFIX_N_0_390  /* z4 = MULTIPLY(z4, - FIX_0_390180644) */
    smull           v18.4s, v18.4h, XFIX_N_0_899  /* z1 = MULTIPLY(z1, - FIX_0_899976223) */
    smull           v20.4s, v20.4h, XFIX_N_2_562  /* z2 = MULTIPLY(z2, - FIX_2_562915447) */

    add             v23.4s, v23.4s, v27.4s  /* z3 += z5 */
    add             v22.4s, v22.4s, v26.4s  /* z3 += z5 */
    add             v25.4s, v25.4s, v27.4s  /* z4 += z5 */
    add             v24.4s, v24.4s, v26.4s  /* z4 += z5 */

    add             v11.4s, v11.4s, v19.4s  /* tmp0 += z1 */
    add             v10.4s, v10.4s, v18.4s  /* tmp0 += z1 */
    add             v13.4s, v13.4s, v21.4s  /* tmp1 += z2 */
    add             v12.4s, v12.4s, v20.4s  /* tmp1 += z2 */
    add             v15.4s, v15.4s, v21.4s  /* tmp2 += z2 */
    add             v14.4s, v14.4s, v20.4s  /* tmp2 += z2 */
    add             v17.4s, v17.4s, v19.4s  /* tmp3 += z1 */
    add             v16.4s, v16.4s, v18.4s  /* tmp3 += z1 */

    add             v11.4s, v11.4s, v23.4s  /* tmp0 += z3 */
    add             v10.4s, v10.4s, v22.4s  /* tmp0 += z3 */
    add             v13.4s, v13.4s, v25.4s  /* tmp1 += z4 */
    add             v12.4s, v12.4s, v24.4s  /* tmp1 += z4 */
    add             v17.4s, v17.4s, v25.4s  /* tmp3 += z4 */
    add             v16.4s, v16.4s, v24.4s  /* tmp3 += z4 */
    add             v15.4s, v15.4s, v23.4s  /* tmp2 += z3 */
    add             v14.4s, v14.4s, v22.4s  /* tmp2 += z3 */

    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */

    add             v18.4s, v2.4s, v16.4s   /* tmp10 + tmp3 */
    add             v19.4s, v28.4s, v17.4s  /* tmp10 + tmp3 */
    sub             v20.4s, v2.4s, v16.4s   /* tmp10 - tmp3 */
    sub             v21.4s, v28.4s, v17.4s  /* tmp10 - tmp3 */
    add             v22.4s, v8.4s, v14.4s   /* tmp11 + tmp2 */
    add             v23.4s, v29.4s, v15.4s  /* tmp11 + tmp2 */
    sub             v24.4s, v8.4s, v14.4s   /* tmp11 - tmp2 */
    sub             v25.4s, v29.4s, v15.4s  /* tmp11 - tmp2 */
    add             v26.4s, v4.4s, v12.4s   /* tmp12 + tmp1 */
    add             v27.4s, v30.4s, v13.4s  /* tmp12 + tmp1 */
    sub             v28.4s, v4.4s, v12.4s   /* tmp12 - tmp1 */
    sub             v29.4s, v30.4s, v13.4s  /* tmp12 - tmp1 */
    add             v14.4s, v6.4s, v10.4s   /* tmp13 + tmp0 */
    add             v15.4s, v31.4s, v11.4s  /* tmp13 + tmp0 */
    sub             v16.4s, v6.4s, v10.4s   /* tmp13 - tmp0 */
    sub             v17.4s, v31.4s, v11.4s  /* tmp13 - tmp0 */

    shrn            v2.4h, v18.4s, #16  /* wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp3, CONST_BITS+PASS1_BITS+3) */
    shrn            v9.4h, v20.4s, #16  /* wsptr[DCTSIZE*7] = (int) DESCALE(tmp10 - tmp3, CONST_BITS+PASS1_BITS+3) */
    shrn            v3.4h, v22.4s, #16  /* wsptr[DCTSIZE*1] = (int) DESCALE(tmp11 + tmp2, CONST_BITS+PASS1_BITS+3) */
    shrn            v8.4h, v24.4s, #16  /* wsptr[DCTSIZE*6] = (int) DESCALE(tmp11 - tmp2, CONST_BITS+PASS1_BITS+3) */
    shrn            v4.4h, v26.4s, #16  /* wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 + tmp1, CONST_BITS+PASS1_BITS+3) */
    shrn            v7.4h, v28.4s, #16  /* wsptr[DCTSIZE*5] = (int) DESCALE(tmp12 - tmp1, CONST_BITS+PASS1_BITS+3) */
    shrn            v5.4h, v14.4s, #16  /* wsptr[DCTSIZE*3] = (int) DESCALE(tmp13 + tmp0, CONST_BITS+PASS1_BITS+3) */
    shrn            v6.4h, v16.4s, #16  /* wsptr[DCTSIZE*4] = (int) DESCALE(tmp13 - tmp0, CONST_BITS+PASS1_BITS+3) */
    shrn2           v2.8h, v19.4s, #16  /* wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp3, CONST_BITS+PASS1_BITS+3) */
    shrn2           v9.8h, v21.4s, #16  /* wsptr[DCTSIZE*7] = (int) DESCALE(tmp10 - tmp3, CONST_BITS+PASS1_BITS+3) */
    shrn2           v3.8h, v23.4s, #16  /* wsptr[DCTSIZE*1] = (int) DESCALE(tmp11 + tmp2, CONST_BITS+PASS1_BITS+3) */
    shrn2           v8.8h, v25.4s, #16  /* wsptr[DCTSIZE*6] = (int) DESCALE(tmp11 - tmp2, CONST_BITS+PASS1_BITS+3) */
    shrn2           v4.8h, v27.4s, #16  /* wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 + tmp1, CONST_BITS+PASS1_BITS+3) */
    shrn2           v7.8h, v29.4s, #16  /* wsptr[DCTSIZE*5] = (int) DESCALE(tmp12 - tmp1, CONST_BITS+PASS1_BITS+3) */
    shrn2           v5.8h, v15.4s, #16  /* wsptr[DCTSIZE*3] = (int) DESCALE(tmp13 + tmp0, CONST_BITS+PASS1_BITS+3) */
    shrn2           v6.8h, v17.4s, #16  /* wsptr[DCTSIZE*4] = (int) DESCALE(tmp13 - tmp0, CONST_BITS+PASS1_BITS+3) */
    movi            v0.16b, #(CENTERJSAMPLE)
    /* Prepare pointers (dual-issue with NEON instructions) */
      ldp             TMP1, TMP2, [OUTPUT_BUF], 16
    sqrshrn         v28.8b, v2.8h, #(CONST_BITS+PASS1_BITS+3-16)
      ldp             TMP3, TMP4, [OUTPUT_BUF], 16
    sqrshrn         v29.8b, v3.8h, #(CONST_BITS+PASS1_BITS+3-16)
      add             TMP1, TMP1, OUTPUT_COL
    sqrshrn         v30.8b, v4.8h, #(CONST_BITS+PASS1_BITS+3-16)
      add             TMP2, TMP2, OUTPUT_COL
    sqrshrn         v31.8b, v5.8h, #(CONST_BITS+PASS1_BITS+3-16)
      add             TMP3, TMP3, OUTPUT_COL
    sqrshrn2        v28.16b, v6.8h, #(CONST_BITS+PASS1_BITS+3-16)
      add             TMP4, TMP4, OUTPUT_COL
    sqrshrn2        v29.16b, v7.8h, #(CONST_BITS+PASS1_BITS+3-16)
      ldp             TMP5, TMP6, [OUTPUT_BUF], 16
    sqrshrn2        v30.16b, v8.8h, #(CONST_BITS+PASS1_BITS+3-16)
      ldp             TMP7, TMP8, [OUTPUT_BUF], 16
    sqrshrn2        v31.16b, v9.8h, #(CONST_BITS+PASS1_BITS+3-16)
      add             TMP5, TMP5, OUTPUT_COL
    add             v16.16b, v28.16b, v0.16b
      add             TMP6, TMP6, OUTPUT_COL
    add             v18.16b, v29.16b, v0.16b
      add             TMP7, TMP7, OUTPUT_COL
    add             v20.16b, v30.16b, v0.16b
      add             TMP8, TMP8, OUTPUT_COL
    add             v22.16b, v31.16b, v0.16b

    /* Transpose the final 8-bit samples */
    trn1            v28.16b, v16.16b, v18.16b
    trn1            v30.16b, v20.16b, v22.16b
    trn2            v29.16b, v16.16b, v18.16b
    trn2            v31.16b, v20.16b, v22.16b

    trn1            v16.8h, v28.8h, v30.8h
    trn2            v18.8h, v28.8h, v30.8h
    trn1            v20.8h, v29.8h, v31.8h
    trn2            v22.8h, v29.8h, v31.8h

    uzp1            v28.4s, v16.4s, v18.4s
    uzp2            v30.4s, v16.4s, v18.4s
    uzp1            v29.4s, v20.4s, v22.4s
    uzp2            v31.4s, v20.4s, v22.4s

    /* Store results to the output buffer */
    st1             {v28.d}[0], [TMP1]
    st1             {v29.d}[0], [TMP2]
    st1             {v28.d}[1], [TMP3]
    st1             {v29.d}[1], [TMP4]
    st1             {v30.d}[0], [TMP5]
    st1             {v31.d}[0], [TMP6]
    st1             {v30.d}[1], [TMP7]
    st1             {v31.d}[1], [TMP8]
    ld1             {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], #32
    ld1             {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], #32
    blr             x30

.balign 16
2:
    mul             v3.8h, v3.8h, v19.8h
    mul             v4.8h, v4.8h, v20.8h
    mul             v5.8h, v5.8h, v21.8h
    add             TMP4, xzr, TMP2, LSL #32
    mul             v6.8h, v6.8h, v22.8h
    mul             v7.8h, v7.8h, v23.8h
    adds            TMP3, xzr, TMP2, LSR #32
    mul             v8.8h, v8.8h, v24.8h
    mul             v9.8h, v9.8h, v25.8h
    b.ne            3f
    /* Right AC coef is zero */
    dup             v15.2d, v10.d[1]
    /* Even part: reverse the even part of the forward DCT. */
    add             v18.4h, v4.4h, v8.4h           /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]) + DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]) */
    add             v22.4h, v2.4h, v6.4h           /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) + DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */
    sub             v26.4h, v2.4h, v6.4h           /* z2 - z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) - DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */
    smull           v18.4s, v18.4h, XFIX_P_0_541   /* z1l z1 = MULTIPLY(z2 + z3, FIX_0_541196100); */
    sshll           v22.4s, v22.4h, #(CONST_BITS)  /* tmp0l tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS); */
    mov             v20.16b, v18.16b               /* tmp3 = z1 */
    sshll           v26.4s, v26.4h, #(CONST_BITS)  /* tmp1l tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS); */
    smlal           v18.4s, v8.4h, XFIX_N_1_847    /* tmp2l tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065); */
    smlal           v20.4s, v4.4h, XFIX_P_0_765    /* tmp3l tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865); */
    add             v2.4s, v22.4s, v20.4s          /* tmp10l tmp10 = tmp0 + tmp3; */
    sub             v6.4s, v22.4s, v20.4s          /* tmp13l tmp13 = tmp0 - tmp3; */
    add             v8.4s, v26.4s, v18.4s          /* tmp11l tmp11 = tmp1 + tmp2; */
    sub             v4.4s, v26.4s, v18.4s          /* tmp12l tmp12 = tmp1 - tmp2; */

    /* Odd part per figure 8; the matrix is unitary and hence its
     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
     */

    add             v22.4h, v9.4h, v5.4h    /* z3 = tmp0 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */
    add             v24.4h, v7.4h, v3.4h    /* z4 = tmp1 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */
    add             v18.4h, v9.4h, v3.4h    /* z1 = tmp0 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */
    add             v20.4h, v7.4h, v5.4h    /* z2 = tmp1 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */
    add             v26.4h, v22.4h, v24.4h  /* z5 = z3 + z4 */

    smull           v10.4s, v9.4h, XFIX_P_0_298   /* tmp0 = MULTIPLY(tmp0, FIX_0_298631336) */
    smull           v12.4s, v7.4h, XFIX_P_2_053   /* tmp1 = MULTIPLY(tmp1, FIX_2_053119869) */
    smull           v14.4s, v5.4h, XFIX_P_3_072   /* tmp2 = MULTIPLY(tmp2, FIX_3_072711026) */
    smull           v16.4s, v3.4h, XFIX_P_1_501   /* tmp3 = MULTIPLY(tmp3, FIX_1_501321110) */
    smull           v26.4s, v26.4h, XFIX_P_1_175  /* z5l z5 = MULTIPLY(z3 + z4, FIX_1_175875602) */
    smull           v22.4s, v22.4h, XFIX_N_1_961  /* z3 = MULTIPLY(z3, - FIX_1_961570560) */
    smull           v24.4s, v24.4h, XFIX_N_0_390  /* z4 = MULTIPLY(z4, - FIX_0_390180644) */
    smull           v18.4s, v18.4h, XFIX_N_0_899  /* z1 = MULTIPLY(z1, - FIX_0_899976223) */
    smull           v20.4s, v20.4h, XFIX_N_2_562  /* z2 = MULTIPLY(z2, - FIX_2_562915447) */

    add             v22.4s, v22.4s, v26.4s  /* z3 += z5 */
    add             v24.4s, v24.4s, v26.4s  /* z4 += z5 */

    add             v10.4s, v10.4s, v18.4s  /* tmp0 += z1 */
    add             v12.4s, v12.4s, v20.4s  /* tmp1 += z2 */
    add             v14.4s, v14.4s, v20.4s  /* tmp2 += z2 */
    add             v16.4s, v16.4s, v18.4s  /* tmp3 += z1 */

    add             v10.4s, v10.4s, v22.4s  /* tmp0 += z3 */
    add             v12.4s, v12.4s, v24.4s  /* tmp1 += z4 */
    add             v16.4s, v16.4s, v24.4s  /* tmp3 += z4 */
    add             v14.4s, v14.4s, v22.4s  /* tmp2 += z3 */

    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */

    add             v18.4s, v2.4s, v16.4s  /* tmp10 + tmp3 */
    sub             v20.4s, v2.4s, v16.4s  /* tmp10 - tmp3 */
    add             v22.4s, v8.4s, v14.4s  /* tmp11 + tmp2 */
    sub             v24.4s, v8.4s, v14.4s  /* tmp11 - tmp2 */
    add             v26.4s, v4.4s, v12.4s  /* tmp12 + tmp1 */
    sub             v28.4s, v4.4s, v12.4s  /* tmp12 - tmp1 */
    add             v14.4s, v6.4s, v10.4s  /* tmp13 + tmp0 */
    sub             v16.4s, v6.4s, v10.4s  /* tmp13 - tmp0 */

    rshrn           v2.4h, v18.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS) */
    rshrn           v3.4h, v22.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*1] = (int) DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS) */
    rshrn           v4.4h, v26.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS) */
    rshrn           v5.4h, v14.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*3] = (int) DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS) */
    rshrn2          v2.8h, v16.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*4] = (int) DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS) */
    rshrn2          v3.8h, v28.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*5] = (int) DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS) */
    rshrn2          v4.8h, v24.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*6] = (int) DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS) */
    rshrn2          v5.8h, v20.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*7] = (int) DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS) */
    mov             v6.16b, v15.16b
    mov             v7.16b, v15.16b
    mov             v8.16b, v15.16b
    mov             v9.16b, v15.16b
    b               1b

.balign 16
3:
    cbnz            TMP4, 4f
    /* Left AC coef is zero */
    dup             v14.2d, v10.d[0]
    /* Even part: reverse the even part of the forward DCT. */
    add             v18.8h, v4.8h, v8.8h           /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]) + DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]) */
    add             v22.8h, v2.8h, v6.8h           /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) + DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */
    smull2          v19.4s, v18.8h, XFIX_P_0_541   /* z1h z1 = MULTIPLY(z2 + z3, FIX_0_541196100); */
    sub             v26.8h, v2.8h, v6.8h           /* z2 - z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) - DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */
    sshll2          v23.4s, v22.8h, #(CONST_BITS)  /* tmp0h tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS); */
    mov             v21.16b, v19.16b               /* tmp3 = z1 */
    smlal2          v19.4s, v8.8h, XFIX_N_1_847    /* tmp2h tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065); */
    sshll2          v27.4s, v26.8h, #(CONST_BITS)  /* tmp1h tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS); */
    smlal2          v21.4s, v4.8h, XFIX_P_0_765    /* tmp3h tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865); */
    add             v28.4s, v23.4s, v21.4s         /* tmp10h tmp10 = tmp0 + tmp3; */
    sub             v31.4s, v23.4s, v21.4s         /* tmp13h tmp13 = tmp0 - tmp3; */
    add             v29.4s, v27.4s, v19.4s         /* tmp11h tmp11 = tmp1 + tmp2; */
    sub             v30.4s, v27.4s, v19.4s         /* tmp12h tmp12 = tmp1 - tmp2; */

    /* Odd part per figure 8; the matrix is unitary and hence its
     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
     */

    add             v22.8h, v9.8h, v5.8h    /* z3 = tmp0 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */
    add             v24.8h, v7.8h, v3.8h    /* z4 = tmp1 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */
    add             v18.8h, v9.8h, v3.8h    /* z1 = tmp0 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */
    add             v20.8h, v7.8h, v5.8h    /* z2 = tmp1 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */
    add             v26.8h, v22.8h, v24.8h  /* z5 = z3 + z4 */

    smull2          v11.4s, v9.8h, XFIX_P_0_298   /* tmp0 = MULTIPLY(tmp0, FIX_0_298631336) */
    smull2          v13.4s, v7.8h, XFIX_P_2_053   /* tmp1 = MULTIPLY(tmp1, FIX_2_053119869) */
    smull2          v15.4s, v5.8h, XFIX_P_3_072   /* tmp2 = MULTIPLY(tmp2, FIX_3_072711026) */
    smull2          v17.4s, v3.8h, XFIX_P_1_501   /* tmp3 = MULTIPLY(tmp3, FIX_1_501321110) */
    smull2          v27.4s, v26.8h, XFIX_P_1_175  /* z5h z5 = MULTIPLY(z3 + z4, FIX_1_175875602) */
    smull2          v23.4s, v22.8h, XFIX_N_1_961  /* z3 = MULTIPLY(z3, - FIX_1_961570560) */
    smull2          v25.4s, v24.8h, XFIX_N_0_390  /* z4 = MULTIPLY(z4, - FIX_0_390180644) */
    smull2          v19.4s, v18.8h, XFIX_N_0_899  /* z1 = MULTIPLY(z1, - FIX_0_899976223) */
    smull2          v21.4s, v20.8h, XFIX_N_2_562  /* z2 = MULTIPLY(z2, - FIX_2_562915447) */

    add             v23.4s, v23.4s, v27.4s  /* z3 += z5 */
    add             v22.4s, v22.4s, v26.4s  /* z3 += z5 */
    add             v25.4s, v25.4s, v27.4s  /* z4 += z5 */
    add             v24.4s, v24.4s, v26.4s  /* z4 += z5 */

    add             v11.4s, v11.4s, v19.4s  /* tmp0 += z1 */
    add             v13.4s, v13.4s, v21.4s  /* tmp1 += z2 */
    add             v15.4s, v15.4s, v21.4s  /* tmp2 += z2 */
    add             v17.4s, v17.4s, v19.4s  /* tmp3 += z1 */

    add             v11.4s, v11.4s, v23.4s  /* tmp0 += z3 */
    add             v13.4s, v13.4s, v25.4s  /* tmp1 += z4 */
    add             v17.4s, v17.4s, v25.4s  /* tmp3 += z4 */
    add             v15.4s, v15.4s, v23.4s  /* tmp2 += z3 */

    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */

    add             v19.4s, v28.4s, v17.4s  /* tmp10 + tmp3 */
    sub             v21.4s, v28.4s, v17.4s  /* tmp10 - tmp3 */
    add             v23.4s, v29.4s, v15.4s  /* tmp11 + tmp2 */
    sub             v25.4s, v29.4s, v15.4s  /* tmp11 - tmp2 */
    add             v27.4s, v30.4s, v13.4s  /* tmp12 + tmp1 */
    sub             v29.4s, v30.4s, v13.4s  /* tmp12 - tmp1 */
    add             v15.4s, v31.4s, v11.4s  /* tmp13 + tmp0 */
    sub             v17.4s, v31.4s, v11.4s  /* tmp13 - tmp0 */

    mov             v2.16b, v14.16b
    mov             v3.16b, v14.16b
    mov             v4.16b, v14.16b
    mov             v5.16b, v14.16b
    rshrn           v6.4h, v19.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS) */
    rshrn           v7.4h, v23.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*1] = (int) DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS) */
    rshrn           v8.4h, v27.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS) */
    rshrn           v9.4h, v15.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*3] = (int) DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS) */
    rshrn2          v6.8h, v17.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*4] = (int) DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS) */
    rshrn2          v7.8h, v29.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*5] = (int) DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS) */
    rshrn2          v8.8h, v25.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*6] = (int) DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS) */
    rshrn2          v9.8h, v21.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*7] = (int) DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS) */
    b               1b

.balign 16
4:
    /* "No" AC coef is zero */
    /* Even part: reverse the even part of the forward DCT. */
    add             v18.8h, v4.8h, v8.8h           /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*2], quantptr[DCTSIZE*2]) + DEQUANTIZE(inptr[DCTSIZE*6], quantptr[DCTSIZE*6]) */
    add             v22.8h, v2.8h, v6.8h           /* z2 + z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) + DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */
    smull2          v19.4s, v18.8h, XFIX_P_0_541   /* z1h z1 = MULTIPLY(z2 + z3, FIX_0_541196100); */
    sub             v26.8h, v2.8h, v6.8h           /* z2 - z3 = DEQUANTIZE(inptr[DCTSIZE*0], quantptr[DCTSIZE*0]) - DEQUANTIZE(inptr[DCTSIZE*4], quantptr[DCTSIZE*4]) */
    smull           v18.4s, v18.4h, XFIX_P_0_541   /* z1l z1 = MULTIPLY(z2 + z3, FIX_0_541196100); */
    sshll2          v23.4s, v22.8h, #(CONST_BITS)  /* tmp0h tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS); */
    mov             v21.16b, v19.16b               /* tmp3 = z1 */
    mov             v20.16b, v18.16b               /* tmp3 = z1 */
    smlal2          v19.4s, v8.8h, XFIX_N_1_847    /* tmp2h tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065); */
    smlal           v18.4s, v8.4h, XFIX_N_1_847    /* tmp2l tmp2 = z1 + MULTIPLY(z3, - FIX_1_847759065); */
    sshll2          v27.4s, v26.8h, #(CONST_BITS)  /* tmp1h tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS); */
    smlal2          v21.4s, v4.8h, XFIX_P_0_765    /* tmp3h tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865); */
    smlal           v20.4s, v4.4h, XFIX_P_0_765    /* tmp3l tmp3 = z1 + MULTIPLY(z2, FIX_0_765366865); */
    sshll           v22.4s, v22.4h, #(CONST_BITS)  /* tmp0l tmp0 = LEFT_SHIFT(z2 + z3, CONST_BITS); */
    sshll           v26.4s, v26.4h, #(CONST_BITS)  /* tmp1l tmp1 = LEFT_SHIFT(z2 - z3, CONST_BITS); */
    add             v2.4s, v22.4s, v20.4s          /* tmp10l tmp10 = tmp0 + tmp3; */
    sub             v6.4s, v22.4s, v20.4s          /* tmp13l tmp13 = tmp0 - tmp3; */
    add             v8.4s, v26.4s, v18.4s          /* tmp11l tmp11 = tmp1 + tmp2; */
    sub             v4.4s, v26.4s, v18.4s          /* tmp12l tmp12 = tmp1 - tmp2; */
    add             v28.4s, v23.4s, v21.4s         /* tmp10h tmp10 = tmp0 + tmp3; */
    sub             v31.4s, v23.4s, v21.4s         /* tmp13h tmp13 = tmp0 - tmp3; */
    add             v29.4s, v27.4s, v19.4s         /* tmp11h tmp11 = tmp1 + tmp2; */
    sub             v30.4s, v27.4s, v19.4s         /* tmp12h tmp12 = tmp1 - tmp2; */

    /* Odd part per figure 8; the matrix is unitary and hence its
     * transpose is its inverse.  i0..i3 are y7,y5,y3,y1 respectively.
     */

    add             v22.8h, v9.8h, v5.8h    /* z3 = tmp0 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */
    add             v24.8h, v7.8h, v3.8h    /* z4 = tmp1 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */
    add             v18.8h, v9.8h, v3.8h    /* z1 = tmp0 + tmp3 = DEQUANTIZE(inptr[DCTSIZE*7], quantptr[DCTSIZE*7]) + DEQUANTIZE(inptr[DCTSIZE*1], quantptr[DCTSIZE*1]) */
    add             v20.8h, v7.8h, v5.8h    /* z2 = tmp1 + tmp2 = DEQUANTIZE(inptr[DCTSIZE*5], quantptr[DCTSIZE*5]) + DEQUANTIZE(inptr[DCTSIZE*3], quantptr[DCTSIZE*3]) */
    add             v26.8h, v22.8h, v24.8h  /* z5 = z3 + z4 */

    smull2          v11.4s, v9.8h, XFIX_P_0_298   /* tmp0 = MULTIPLY(tmp0, FIX_0_298631336) */
    smull2          v13.4s, v7.8h, XFIX_P_2_053   /* tmp1 = MULTIPLY(tmp1, FIX_2_053119869) */
    smull2          v15.4s, v5.8h, XFIX_P_3_072   /* tmp2 = MULTIPLY(tmp2, FIX_3_072711026) */
    smull2          v17.4s, v3.8h, XFIX_P_1_501   /* tmp3 = MULTIPLY(tmp3, FIX_1_501321110) */
    smull2          v27.4s, v26.8h, XFIX_P_1_175  /* z5h z5 = MULTIPLY(z3 + z4, FIX_1_175875602) */
    smull2          v23.4s, v22.8h, XFIX_N_1_961  /* z3 = MULTIPLY(z3, - FIX_1_961570560) */
    smull2          v25.4s, v24.8h, XFIX_N_0_390  /* z4 = MULTIPLY(z4, - FIX_0_390180644) */
    smull2          v19.4s, v18.8h, XFIX_N_0_899  /* z1 = MULTIPLY(z1, - FIX_0_899976223) */
    smull2          v21.4s, v20.8h, XFIX_N_2_562  /* z2 = MULTIPLY(z2, - FIX_2_562915447) */

    smull           v10.4s, v9.4h, XFIX_P_0_298   /* tmp0 = MULTIPLY(tmp0, FIX_0_298631336) */
    smull           v12.4s, v7.4h, XFIX_P_2_053   /* tmp1 = MULTIPLY(tmp1, FIX_2_053119869) */
    smull           v14.4s, v5.4h, XFIX_P_3_072   /* tmp2 = MULTIPLY(tmp2, FIX_3_072711026) */
    smull           v16.4s, v3.4h, XFIX_P_1_501   /* tmp3 = MULTIPLY(tmp3, FIX_1_501321110) */
    smull           v26.4s, v26.4h, XFIX_P_1_175  /* z5l z5 = MULTIPLY(z3 + z4, FIX_1_175875602) */
    smull           v22.4s, v22.4h, XFIX_N_1_961  /* z3 = MULTIPLY(z3, - FIX_1_961570560) */
    smull           v24.4s, v24.4h, XFIX_N_0_390  /* z4 = MULTIPLY(z4, - FIX_0_390180644) */
    smull           v18.4s, v18.4h, XFIX_N_0_899  /* z1 = MULTIPLY(z1, - FIX_0_899976223) */
    smull           v20.4s, v20.4h, XFIX_N_2_562  /* z2 = MULTIPLY(z2, - FIX_2_562915447) */

    add             v23.4s, v23.4s, v27.4s  /* z3 += z5 */
    add             v22.4s, v22.4s, v26.4s  /* z3 += z5 */
    add             v25.4s, v25.4s, v27.4s  /* z4 += z5 */
    add             v24.4s, v24.4s, v26.4s  /* z4 += z5 */

    add             v11.4s, v11.4s, v19.4s  /* tmp0 += z1 */
    add             v10.4s, v10.4s, v18.4s  /* tmp0 += z1 */
    add             v13.4s, v13.4s, v21.4s  /* tmp1 += z2 */
    add             v12.4s, v12.4s, v20.4s  /* tmp1 += z2 */
    add             v15.4s, v15.4s, v21.4s  /* tmp2 += z2 */
    add             v14.4s, v14.4s, v20.4s  /* tmp2 += z2 */
    add             v17.4s, v17.4s, v19.4s  /* tmp3 += z1 */
    add             v16.4s, v16.4s, v18.4s  /* tmp3 += z1 */

    add             v11.4s, v11.4s, v23.4s  /* tmp0 += z3 */
    add             v10.4s, v10.4s, v22.4s  /* tmp0 += z3 */
    add             v13.4s, v13.4s, v25.4s  /* tmp1 += z4 */
    add             v12.4s, v12.4s, v24.4s  /* tmp1 += z4 */
    add             v17.4s, v17.4s, v25.4s  /* tmp3 += z4 */
    add             v16.4s, v16.4s, v24.4s  /* tmp3 += z4 */
    add             v15.4s, v15.4s, v23.4s  /* tmp2 += z3 */
    add             v14.4s, v14.4s, v22.4s  /* tmp2 += z3 */

    /* Final output stage: inputs are tmp10..tmp13, tmp0..tmp3 */

    add             v18.4s, v2.4s, v16.4s   /* tmp10 + tmp3 */
    add             v19.4s, v28.4s, v17.4s  /* tmp10 + tmp3 */
    sub             v20.4s, v2.4s, v16.4s   /* tmp10 - tmp3 */
    sub             v21.4s, v28.4s, v17.4s  /* tmp10 - tmp3 */
    add             v22.4s, v8.4s, v14.4s   /* tmp11 + tmp2 */
    add             v23.4s, v29.4s, v15.4s  /* tmp11 + tmp2 */
    sub             v24.4s, v8.4s, v14.4s   /* tmp11 - tmp2 */
    sub             v25.4s, v29.4s, v15.4s  /* tmp11 - tmp2 */
    add             v26.4s, v4.4s, v12.4s   /* tmp12 + tmp1 */
    add             v27.4s, v30.4s, v13.4s  /* tmp12 + tmp1 */
    sub             v28.4s, v4.4s, v12.4s   /* tmp12 - tmp1 */
    sub             v29.4s, v30.4s, v13.4s  /* tmp12 - tmp1 */
    add             v14.4s, v6.4s, v10.4s   /* tmp13 + tmp0 */
    add             v15.4s, v31.4s, v11.4s  /* tmp13 + tmp0 */
    sub             v16.4s, v6.4s, v10.4s   /* tmp13 - tmp0 */
    sub             v17.4s, v31.4s, v11.4s  /* tmp13 - tmp0 */

    rshrn           v2.4h, v18.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS) */
    rshrn           v3.4h, v22.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*1] = (int) DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS) */
    rshrn           v4.4h, v26.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS) */
    rshrn           v5.4h, v14.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*3] = (int) DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS) */
    rshrn           v6.4h, v19.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*0] = (int) DESCALE(tmp10 + tmp3, CONST_BITS-PASS1_BITS) */
    rshrn           v7.4h, v23.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*1] = (int) DESCALE(tmp11 + tmp2, CONST_BITS-PASS1_BITS) */
    rshrn           v8.4h, v27.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*2] = (int) DESCALE(tmp12 + tmp1, CONST_BITS-PASS1_BITS) */
    rshrn           v9.4h, v15.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*3] = (int) DESCALE(tmp13 + tmp0, CONST_BITS-PASS1_BITS) */
    rshrn2          v2.8h, v16.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*4] = (int) DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS) */
    rshrn2          v3.8h, v28.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*5] = (int) DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS) */
    rshrn2          v4.8h, v24.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*6] = (int) DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS) */
    rshrn2          v5.8h, v20.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*7] = (int) DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS) */
    rshrn2          v6.8h, v17.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*4] = (int) DESCALE(tmp13 - tmp0, CONST_BITS-PASS1_BITS) */
    rshrn2          v7.8h, v29.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*5] = (int) DESCALE(tmp12 - tmp1, CONST_BITS-PASS1_BITS) */
    rshrn2          v8.8h, v25.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*6] = (int) DESCALE(tmp11 - tmp2, CONST_BITS-PASS1_BITS) */
    rshrn2          v9.8h, v21.4s, #(CONST_BITS-PASS1_BITS)  /* wsptr[DCTSIZE*7] = (int) DESCALE(tmp10 - tmp3, CONST_BITS-PASS1_BITS) */
    b               1b

    .unreq          DCT_TABLE
    .unreq          COEF_BLOCK
    .unreq          OUTPUT_BUF
    .unreq          OUTPUT_COL
    .unreq          TMP1
    .unreq          TMP2
    .unreq          TMP3
    .unreq          TMP4
    .unreq          TMP5
    .unreq          TMP6
    .unreq          TMP7
    .unreq          TMP8

#undef CENTERJSAMPLE
#undef CONST_BITS
#undef PASS1_BITS
#undef XFIX_P_0_298
#undef XFIX_N_0_390
#undef XFIX_P_0_541
#undef XFIX_P_0_765
#undef XFIX_N_0_899
#undef XFIX_P_1_175
#undef XFIX_P_1_501
#undef XFIX_N_1_847
#undef XFIX_N_1_961
#undef XFIX_P_2_053
#undef XFIX_N_2_562
#undef XFIX_P_3_072


/*****************************************************************************/

/*
 * jsimd_idct_ifast_neon
 *
 * This function contains a fast, not so accurate integer implementation of
 * the inverse DCT (Discrete Cosine Transform). It uses the same calculations
 * and produces exactly the same output as IJG's original 'jpeg_idct_ifast'
 * function from jidctfst.c
 *
 * Normally 1-D AAN DCT needs 5 multiplications and 29 additions.
 * But in ARM NEON case some extra additions are required because VQDMULH
 * instruction can't handle the constants larger than 1. So the expressions
 * like "x * 1.082392200" have to be converted to "x * 0.082392200 + x",
 * which introduces an extra addition. Overall, there are 6 extra additions
 * per 1-D IDCT pass, totalling to 5 VQDMULH and 35 VADD/VSUB instructions.
 */

#define XFIX_1_082392200 v0.h[0]
#define XFIX_1_414213562 v0.h[1]
#define XFIX_1_847759065 v0.h[2]
#define XFIX_2_613125930 v0.h[3]

.balign 16
Ljsimd_idct_ifast_neon_consts:
  .short (277 * 128 - 256 * 128)  /* XFIX_1_082392200 */
  .short (362 * 128 - 256 * 128)  /* XFIX_1_414213562 */
  .short (473 * 128 - 256 * 128)  /* XFIX_1_847759065 */
  .short (669 * 128 - 512 * 128)  /* XFIX_2_613125930 */

asm_function jsimd_idct_ifast_neon

    DCT_TABLE       .req x0
    COEF_BLOCK      .req x1
    OUTPUT_BUF      .req x2
    OUTPUT_COL      .req x3
    TMP1            .req x0
    TMP2            .req x1
    TMP3            .req x9
    TMP4            .req x10
    TMP5            .req x11
    TMP6            .req x12
    TMP7            .req x13
    TMP8            .req x14

    /* Load and dequantize coefficients into NEON registers
     * with the following allocation:
     *       0 1 2 3 | 4 5 6 7
     *      ---------+--------
     *   0 | d16     | d17     ( v16.8h )
     *   1 | d18     | d19     ( v17.8h )
     *   2 | d20     | d21     ( v18.8h )
     *   3 | d22     | d23     ( v19.8h )
     *   4 | d24     | d25     ( v20.8h )
     *   5 | d26     | d27     ( v21.8h )
     *   6 | d28     | d29     ( v22.8h )
     *   7 | d30     | d31     ( v23.8h )
     */
    /* Save NEON registers used in fast IDCT */
    adr             TMP5, Ljsimd_idct_ifast_neon_consts
    ld1             {v16.8h, v17.8h}, [COEF_BLOCK], 32
    ld1             {v0.8h, v1.8h}, [DCT_TABLE], 32
    ld1             {v18.8h, v19.8h}, [COEF_BLOCK], 32
    mul             v16.8h, v16.8h, v0.8h
    ld1             {v2.8h, v3.8h}, [DCT_TABLE], 32
    mul             v17.8h, v17.8h, v1.8h
    ld1             {v20.8h, v21.8h}, [COEF_BLOCK], 32
    mul             v18.8h, v18.8h, v2.8h
    ld1             {v0.8h, v1.8h}, [DCT_TABLE], 32
    mul             v19.8h, v19.8h, v3.8h
    ld1             {v22.8h, v23.8h}, [COEF_BLOCK], 32
    mul             v20.8h, v20.8h, v0.8h
    ld1             {v2.8h, v3.8h}, [DCT_TABLE], 32
    mul             v22.8h, v22.8h, v2.8h
    mul             v21.8h, v21.8h, v1.8h
    ld1             {v0.4h}, [TMP5]        /* load constants */
    mul             v23.8h, v23.8h, v3.8h

    /* 1-D IDCT, pass 1 */
    sub             v2.8h, v18.8h, v22.8h
    add             v22.8h, v18.8h, v22.8h
    sub             v1.8h, v19.8h, v21.8h
    add             v21.8h, v19.8h, v21.8h
    sub             v5.8h, v17.8h, v23.8h
    add             v23.8h, v17.8h, v23.8h
    sqdmulh         v4.8h, v2.8h, XFIX_1_414213562
    sqdmulh         v6.8h, v1.8h, XFIX_2_613125930
    add             v3.8h, v1.8h, v1.8h
    sub             v1.8h, v5.8h, v1.8h
    add             v18.8h, v2.8h, v4.8h
    sqdmulh         v4.8h, v1.8h, XFIX_1_847759065
    sub             v2.8h, v23.8h, v21.8h
    add             v3.8h, v3.8h, v6.8h
    sqdmulh         v6.8h, v2.8h, XFIX_1_414213562
    add             v1.8h, v1.8h, v4.8h
    sqdmulh         v4.8h, v5.8h, XFIX_1_082392200
    sub             v18.8h, v18.8h, v22.8h
    add             v2.8h, v2.8h, v6.8h
    sub             v6.8h, v16.8h, v20.8h
    add             v20.8h, v16.8h, v20.8h
    add             v17.8h, v5.8h, v4.8h
    add             v5.8h, v6.8h, v18.8h
    sub             v18.8h, v6.8h, v18.8h
    add             v6.8h, v23.8h, v21.8h
    add             v16.8h, v20.8h, v22.8h
    sub             v3.8h, v6.8h, v3.8h
    sub             v20.8h, v20.8h, v22.8h
    sub             v3.8h, v3.8h, v1.8h
    sub             v1.8h, v17.8h, v1.8h
    add             v2.8h, v3.8h, v2.8h
    sub             v23.8h, v16.8h, v6.8h
    add             v1.8h, v1.8h, v2.8h
    add             v16.8h, v16.8h, v6.8h
    add             v22.8h, v5.8h, v3.8h
    sub             v17.8h, v5.8h, v3.8h
    sub             v21.8h, v18.8h, v2.8h
    add             v18.8h, v18.8h, v2.8h
    sub             v19.8h, v20.8h, v1.8h
    add             v20.8h, v20.8h, v1.8h
    transpose_8x8   v16, v17, v18, v19, v20, v21, v22, v23, v28, v29, v30, v31
    /* 1-D IDCT, pass 2 */
    sub             v2.8h, v18.8h, v22.8h
    add             v22.8h, v18.8h, v22.8h
    sub             v1.8h, v19.8h, v21.8h
    add             v21.8h, v19.8h, v21.8h
    sub             v5.8h, v17.8h, v23.8h
    add             v23.8h, v17.8h, v23.8h
    sqdmulh         v4.8h, v2.8h, XFIX_1_414213562
    sqdmulh         v6.8h, v1.8h, XFIX_2_613125930
    add             v3.8h, v1.8h, v1.8h
    sub             v1.8h, v5.8h, v1.8h
    add             v18.8h, v2.8h, v4.8h
    sqdmulh         v4.8h, v1.8h, XFIX_1_847759065
    sub             v2.8h, v23.8h, v21.8h
    add             v3.8h, v3.8h, v6.8h
    sqdmulh         v6.8h, v2.8h, XFIX_1_414213562
    add             v1.8h, v1.8h, v4.8h
    sqdmulh         v4.8h, v5.8h, XFIX_1_082392200
    sub             v18.8h, v18.8h, v22.8h
    add             v2.8h, v2.8h, v6.8h
    sub             v6.8h, v16.8h, v20.8h
    add             v20.8h, v16.8h, v20.8h
    add             v17.8h, v5.8h, v4.8h
    add             v5.8h, v6.8h, v18.8h
    sub             v18.8h, v6.8h, v18.8h
    add             v6.8h, v23.8h, v21.8h
    add             v16.8h, v20.8h, v22.8h
    sub             v3.8h, v6.8h, v3.8h
    sub             v20.8h, v20.8h, v22.8h
    sub             v3.8h, v3.8h, v1.8h
    sub             v1.8h, v17.8h, v1.8h
    add             v2.8h, v3.8h, v2.8h
    sub             v23.8h, v16.8h, v6.8h
    add             v1.8h, v1.8h, v2.8h
    add             v16.8h, v16.8h, v6.8h
    add             v22.8h, v5.8h, v3.8h
    sub             v17.8h, v5.8h, v3.8h
    sub             v21.8h, v18.8h, v2.8h
    add             v18.8h, v18.8h, v2.8h
    sub             v19.8h, v20.8h, v1.8h
    add             v20.8h, v20.8h, v1.8h
    /* Descale to 8-bit and range limit */
    movi            v0.16b, #0x80
      /* Prepare pointers (dual-issue with NEON instructions) */
      ldp             TMP1, TMP2, [OUTPUT_BUF], 16
    sqshrn          v28.8b, v16.8h, #5
      ldp             TMP3, TMP4, [OUTPUT_BUF], 16
    sqshrn          v29.8b, v17.8h, #5
      add             TMP1, TMP1, OUTPUT_COL
    sqshrn          v30.8b, v18.8h, #5
      add             TMP2, TMP2, OUTPUT_COL
    sqshrn          v31.8b, v19.8h, #5
      add             TMP3, TMP3, OUTPUT_COL
    sqshrn2         v28.16b, v20.8h, #5
      add             TMP4, TMP4, OUTPUT_COL
    sqshrn2         v29.16b, v21.8h, #5
      ldp             TMP5, TMP6, [OUTPUT_BUF], 16
    sqshrn2         v30.16b, v22.8h, #5
      ldp             TMP7, TMP8, [OUTPUT_BUF], 16
    sqshrn2         v31.16b, v23.8h, #5
      add             TMP5, TMP5, OUTPUT_COL
    add             v16.16b, v28.16b, v0.16b
      add             TMP6, TMP6, OUTPUT_COL
    add             v18.16b, v29.16b, v0.16b
      add             TMP7, TMP7, OUTPUT_COL
    add             v20.16b, v30.16b, v0.16b
      add             TMP8, TMP8, OUTPUT_COL
    add             v22.16b, v31.16b, v0.16b

    /* Transpose the final 8-bit samples */
    trn1            v28.16b, v16.16b, v18.16b
    trn1            v30.16b, v20.16b, v22.16b
    trn2            v29.16b, v16.16b, v18.16b
    trn2            v31.16b, v20.16b, v22.16b

    trn1            v16.8h, v28.8h, v30.8h
    trn2            v18.8h, v28.8h, v30.8h
    trn1            v20.8h, v29.8h, v31.8h
    trn2            v22.8h, v29.8h, v31.8h

    uzp1            v28.4s, v16.4s, v18.4s
    uzp2            v30.4s, v16.4s, v18.4s
    uzp1            v29.4s, v20.4s, v22.4s
    uzp2            v31.4s, v20.4s, v22.4s

    /* Store results to the output buffer */
    st1             {v28.d}[0], [TMP1]
    st1             {v29.d}[0], [TMP2]
    st1             {v28.d}[1], [TMP3]
    st1             {v29.d}[1], [TMP4]
    st1             {v30.d}[0], [TMP5]
    st1             {v31.d}[0], [TMP6]
    st1             {v30.d}[1], [TMP7]
    st1             {v31.d}[1], [TMP8]
    blr             x30

    .unreq          DCT_TABLE
    .unreq          COEF_BLOCK
    .unreq          OUTPUT_BUF
    .unreq          OUTPUT_COL
    .unreq          TMP1
    .unreq          TMP2
    .unreq          TMP3
    .unreq          TMP4
    .unreq          TMP5
    .unreq          TMP6
    .unreq          TMP7
    .unreq          TMP8


/*****************************************************************************/

/*
 * jsimd_idct_4x4_neon
 *
 * This function contains inverse-DCT code for getting reduced-size
 * 4x4 pixels output from an 8x8 DCT block. It uses the same  calculations
 * and produces exactly the same output as IJG's original 'jpeg_idct_4x4'
 * function from jpeg-6b (jidctred.c).
 *
 * NOTE: jpeg-8 has an improved implementation of 4x4 inverse-DCT, which
 *       requires much less arithmetic operations and hence should be faster.
 *       The primary purpose of this particular NEON optimized function is
 *       bit exact compatibility with jpeg-6b.
 *
 * TODO: a bit better instructions scheduling can be achieved by expanding
 *       idct_helper/transpose_4x4 macros and reordering instructions,
 *       but readability will suffer somewhat.
 */

#define CONST_BITS  13

#define FIX_0_211164243 (1730)   /* FIX(0.211164243) */
#define FIX_0_509795579 (4176)   /* FIX(0.509795579) */
#define FIX_0_601344887 (4926)   /* FIX(0.601344887) */
#define FIX_0_720959822 (5906)   /* FIX(0.720959822) */
#define FIX_0_765366865 (6270)   /* FIX(0.765366865) */
#define FIX_0_850430095 (6967)   /* FIX(0.850430095) */
#define FIX_0_899976223 (7373)   /* FIX(0.899976223) */
#define FIX_1_061594337 (8697)   /* FIX(1.061594337) */
#define FIX_1_272758580 (10426)  /* FIX(1.272758580) */
#define FIX_1_451774981 (11893)  /* FIX(1.451774981) */
#define FIX_1_847759065 (15137)  /* FIX(1.847759065) */
#define FIX_2_172734803 (17799)  /* FIX(2.172734803) */
#define FIX_2_562915447 (20995)  /* FIX(2.562915447) */
#define FIX_3_624509785 (29692)  /* FIX(3.624509785) */

.balign 16
Ljsimd_idct_4x4_neon_consts:
  .short FIX_1_847759065      /* v0.h[0] */
  .short -FIX_0_765366865     /* v0.h[1] */
  .short -FIX_0_211164243     /* v0.h[2] */
  .short FIX_1_451774981      /* v0.h[3] */
  .short -FIX_2_172734803     /* d1[0] */
  .short FIX_1_061594337      /* d1[1] */
  .short -FIX_0_509795579     /* d1[2] */
  .short -FIX_0_601344887     /* d1[3] */
  .short FIX_0_899976223      /* v2.h[0] */
  .short FIX_2_562915447      /* v2.h[1] */
  .short 1 << (CONST_BITS+1)  /* v2.h[2] */
  .short 0                    /* v2.h[3] */

.macro idct_helper x4, x6, x8, x10, x12, x14, x16, shift, y26, y27, y28, y29
    smull           v28.4s, \x4, v2.h[2]
    smlal           v28.4s, \x8, v0.h[0]
    smlal           v28.4s, \x14, v0.h[1]

    smull           v26.4s, \x16, v1.h[2]
    smlal           v26.4s, \x12, v1.h[3]
    smlal           v26.4s, \x10, v2.h[0]
    smlal           v26.4s, \x6, v2.h[1]

    smull           v30.4s, \x4, v2.h[2]
    smlsl           v30.4s, \x8, v0.h[0]
    smlsl           v30.4s, \x14, v0.h[1]

    smull           v24.4s, \x16, v0.h[2]
    smlal           v24.4s, \x12, v0.h[3]
    smlal           v24.4s, \x10, v1.h[0]
    smlal           v24.4s, \x6, v1.h[1]

    add             v20.4s, v28.4s, v26.4s
    sub             v28.4s, v28.4s, v26.4s

  .if \shift > 16
    srshr           v20.4s, v20.4s, #\shift
    srshr           v28.4s, v28.4s, #\shift
    xtn             \y26, v20.4s
    xtn             \y29, v28.4s
  .else
    rshrn           \y26, v20.4s, #\shift
    rshrn           \y29, v28.4s, #\shift
  .endif

    add             v20.4s, v30.4s, v24.4s
    sub             v30.4s, v30.4s, v24.4s

  .if \shift > 16
    srshr           v20.4s, v20.4s, #\shift
    srshr           v30.4s, v30.4s, #\shift
    xtn             \y27, v20.4s
    xtn             \y28, v30.4s
  .else
    rshrn           \y27, v20.4s, #\shift
    rshrn           \y28, v30.4s, #\shift
  .endif
.endm

asm_function jsimd_idct_4x4_neon

    DCT_TABLE       .req x0
    COEF_BLOCK      .req x1
    OUTPUT_BUF      .req x2
    OUTPUT_COL      .req x3
    TMP1            .req x0
    TMP2            .req x1
    TMP3            .req x2
    TMP4            .req x15

    /* Save all used NEON registers */
    sub             sp, sp, 272
    str             x15, [sp], 16
    /* Load constants (v3.4h is just used for padding) */
    adr             TMP4, Ljsimd_idct_4x4_neon_consts
    st1             {v0.8b, v1.8b, v2.8b, v3.8b}, [sp], 32
    st1             {v4.8b, v5.8b, v6.8b, v7.8b}, [sp], 32
    st1             {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], 32
    st1             {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], 32
    st1             {v16.8b, v17.8b, v18.8b, v19.8b}, [sp], 32
    st1             {v20.8b, v21.8b, v22.8b, v23.8b}, [sp], 32
    st1             {v24.8b, v25.8b, v26.8b, v27.8b}, [sp], 32
    st1             {v28.8b, v29.8b, v30.8b, v31.8b}, [sp], 32
    ld1             {v0.4h, v1.4h, v2.4h, v3.4h}, [TMP4]

    /* Load all COEF_BLOCK into NEON registers with the following allocation:
     *       0 1 2 3 | 4 5 6 7
     *      ---------+--------
     *   0 | v4.4h   | v5.4h
     *   1 | v6.4h   | v7.4h
     *   2 | v8.4h   | v9.4h
     *   3 | v10.4h  | v11.4h
     *   4 | -       | -
     *   5 | v12.4h  | v13.4h
     *   6 | v14.4h  | v15.4h
     *   7 | v16.4h  | v17.4h
     */
    ld1             {v4.4h, v5.4h, v6.4h, v7.4h}, [COEF_BLOCK], 32
    ld1             {v8.4h, v9.4h, v10.4h, v11.4h}, [COEF_BLOCK], 32
    add             COEF_BLOCK, COEF_BLOCK, #16
    ld1             {v12.4h, v13.4h, v14.4h, v15.4h}, [COEF_BLOCK], 32
    ld1             {v16.4h, v17.4h}, [COEF_BLOCK], 16
    /* dequantize */
    ld1             {v18.4h, v19.4h, v20.4h, v21.4h}, [DCT_TABLE], 32
    mul             v4.4h, v4.4h, v18.4h
    mul             v5.4h, v5.4h, v19.4h
    ins             v4.d[1], v5.d[0]              /* 128 bit q4 */
    ld1             {v22.4h, v23.4h, v24.4h, v25.4h}, [DCT_TABLE], 32
    mul             v6.4h, v6.4h, v20.4h
    mul             v7.4h, v7.4h, v21.4h
    ins             v6.d[1], v7.d[0]              /* 128 bit q6 */
    mul             v8.4h, v8.4h, v22.4h
    mul             v9.4h, v9.4h, v23.4h
    ins             v8.d[1], v9.d[0]              /* 128 bit q8 */
    add             DCT_TABLE, DCT_TABLE, #16
    ld1             {v26.4h, v27.4h, v28.4h, v29.4h}, [DCT_TABLE], 32
    mul             v10.4h, v10.4h, v24.4h
    mul             v11.4h, v11.4h, v25.4h
    ins             v10.d[1], v11.d[0]            /* 128 bit q10 */
    mul             v12.4h, v12.4h, v26.4h
    mul             v13.4h, v13.4h, v27.4h
    ins             v12.d[1], v13.d[0]            /* 128 bit q12 */
    ld1             {v30.4h, v31.4h}, [DCT_TABLE], 16
    mul             v14.4h, v14.4h, v28.4h
    mul             v15.4h, v15.4h, v29.4h
    ins             v14.d[1], v15.d[0]            /* 128 bit q14 */
    mul             v16.4h, v16.4h, v30.4h
    mul             v17.4h, v17.4h, v31.4h
    ins             v16.d[1], v17.d[0]            /* 128 bit q16 */

    /* Pass 1 */
    idct_helper     v4.4h, v6.4h, v8.4h, v10.4h, v12.4h, v14.4h, v16.4h, 12, \
                    v4.4h, v6.4h, v8.4h, v10.4h
    transpose_4x4   v4, v6, v8, v10, v3
    ins             v10.d[1], v11.d[0]
    idct_helper     v5.4h, v7.4h, v9.4h, v11.4h, v13.4h, v15.4h, v17.4h, 12, \
                    v5.4h, v7.4h, v9.4h, v11.4h
    transpose_4x4   v5, v7, v9, v11, v3
    ins             v10.d[1], v11.d[0]

    /* Pass 2 */
    idct_helper     v4.4h, v6.4h, v8.4h, v10.4h, v7.4h, v9.4h, v11.4h, 19, \
                    v26.4h, v27.4h, v28.4h, v29.4h
    transpose_4x4   v26, v27, v28, v29, v3

    /* Range limit */
    movi            v30.8h, #0x80
    ins             v26.d[1], v27.d[0]
    ins             v28.d[1], v29.d[0]
    add             v26.8h, v26.8h, v30.8h
    add             v28.8h, v28.8h, v30.8h
    sqxtun          v26.8b, v26.8h
    sqxtun          v27.8b, v28.8h

    /* Store results to the output buffer */
    ldp             TMP1, TMP2, [OUTPUT_BUF], 16
    ldp             TMP3, TMP4, [OUTPUT_BUF]
    add             TMP1, TMP1, OUTPUT_COL
    add             TMP2, TMP2, OUTPUT_COL
    add             TMP3, TMP3, OUTPUT_COL
    add             TMP4, TMP4, OUTPUT_COL

#if defined(__ARMEL__) && !RESPECT_STRICT_ALIGNMENT
    /* We can use much less instructions on little endian systems if the
     * OS kernel is not configured to trap unaligned memory accesses
     */
    st1             {v26.s}[0], [TMP1], 4
    st1             {v27.s}[0], [TMP3], 4
    st1             {v26.s}[1], [TMP2], 4
    st1             {v27.s}[1], [TMP4], 4
#else
    st1             {v26.b}[0], [TMP1], 1
    st1             {v27.b}[0], [TMP3], 1
    st1             {v26.b}[1], [TMP1], 1
    st1             {v27.b}[1], [TMP3], 1
    st1             {v26.b}[2], [TMP1], 1
    st1             {v27.b}[2], [TMP3], 1
    st1             {v26.b}[3], [TMP1], 1
    st1             {v27.b}[3], [TMP3], 1

    st1             {v26.b}[4], [TMP2], 1
    st1             {v27.b}[4], [TMP4], 1
    st1             {v26.b}[5], [TMP2], 1
    st1             {v27.b}[5], [TMP4], 1
    st1             {v26.b}[6], [TMP2], 1
    st1             {v27.b}[6], [TMP4], 1
    st1             {v26.b}[7], [TMP2], 1
    st1             {v27.b}[7], [TMP4], 1
#endif

    /* vpop            {v8.4h - v15.4h}    ;not available */
    sub             sp, sp, #272
    ldr             x15, [sp], 16
    ld1             {v0.8b, v1.8b, v2.8b, v3.8b}, [sp], 32
    ld1             {v4.8b, v5.8b, v6.8b, v7.8b}, [sp], 32
    ld1             {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], 32
    ld1             {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], 32
    ld1             {v16.8b, v17.8b, v18.8b, v19.8b}, [sp], 32
    ld1             {v20.8b, v21.8b, v22.8b, v23.8b}, [sp], 32
    ld1             {v24.8b, v25.8b, v26.8b, v27.8b}, [sp], 32
    ld1             {v28.8b, v29.8b, v30.8b, v31.8b}, [sp], 32
    blr             x30

    .unreq          DCT_TABLE
    .unreq          COEF_BLOCK
    .unreq          OUTPUT_BUF
    .unreq          OUTPUT_COL
    .unreq          TMP1
    .unreq          TMP2
    .unreq          TMP3
    .unreq          TMP4

.purgem idct_helper


/*****************************************************************************/

/*
 * jsimd_idct_2x2_neon
 *
 * This function contains inverse-DCT code for getting reduced-size
 * 2x2 pixels output from an 8x8 DCT block. It uses the same  calculations
 * and produces exactly the same output as IJG's original 'jpeg_idct_2x2'
 * function from jpeg-6b (jidctred.c).
 *
 * NOTE: jpeg-8 has an improved implementation of 2x2 inverse-DCT, which
 *       requires much less arithmetic operations and hence should be faster.
 *       The primary purpose of this particular NEON optimized function is
 *       bit exact compatibility with jpeg-6b.
 */

.balign 8
Ljsimd_idct_2x2_neon_consts:
  .short -FIX_0_720959822  /* v14[0] */
  .short FIX_0_850430095   /* v14[1] */
  .short -FIX_1_272758580  /* v14[2] */
  .short FIX_3_624509785   /* v14[3] */

.macro idct_helper x4, x6, x10, x12, x16, shift, y26, y27
    sshll           v15.4s, \x4, #15
    smull           v26.4s, \x6, v14.h[3]
    smlal           v26.4s, \x10, v14.h[2]
    smlal           v26.4s, \x12, v14.h[1]
    smlal           v26.4s, \x16, v14.h[0]

    add             v20.4s, v15.4s, v26.4s
    sub             v15.4s, v15.4s, v26.4s

  .if \shift > 16
    srshr           v20.4s, v20.4s, #\shift
    srshr           v15.4s, v15.4s, #\shift
    xtn             \y26, v20.4s
    xtn             \y27, v15.4s
  .else
    rshrn           \y26, v20.4s, #\shift
    rshrn           \y27, v15.4s, #\shift
  .endif
.endm

asm_function jsimd_idct_2x2_neon

    DCT_TABLE       .req x0
    COEF_BLOCK      .req x1
    OUTPUT_BUF      .req x2
    OUTPUT_COL      .req x3
    TMP1            .req x0
    TMP2            .req x15

    /* vpush           {v8.4h - v15.4h}            ; not available */
    sub             sp, sp, 208
    str             x15, [sp], 16

    /* Load constants */
    adr             TMP2, Ljsimd_idct_2x2_neon_consts
    st1             {v4.8b, v5.8b, v6.8b, v7.8b}, [sp], 32
    st1             {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], 32
    st1             {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], 32
    st1             {v16.8b, v17.8b, v18.8b, v19.8b}, [sp], 32
    st1             {v21.8b, v22.8b}, [sp], 16
    st1             {v24.8b, v25.8b, v26.8b, v27.8b}, [sp], 32
    st1             {v30.8b, v31.8b}, [sp], 16
    ld1             {v14.4h}, [TMP2]

    /* Load all COEF_BLOCK into NEON registers with the following allocation:
     *       0 1 2 3 | 4 5 6 7
     *      ---------+--------
     *   0 | v4.4h   | v5.4h
     *   1 | v6.4h   | v7.4h
     *   2 | -       | -
     *   3 | v10.4h  | v11.4h
     *   4 | -       | -
     *   5 | v12.4h  | v13.4h
     *   6 | -       | -
     *   7 | v16.4h  | v17.4h
     */
    ld1             {v4.4h, v5.4h, v6.4h, v7.4h}, [COEF_BLOCK], 32
    add             COEF_BLOCK, COEF_BLOCK, #16
    ld1             {v10.4h, v11.4h}, [COEF_BLOCK], 16
    add             COEF_BLOCK, COEF_BLOCK, #16
    ld1             {v12.4h, v13.4h}, [COEF_BLOCK], 16
    add             COEF_BLOCK, COEF_BLOCK, #16
    ld1             {v16.4h, v17.4h}, [COEF_BLOCK], 16
    /* Dequantize */
    ld1             {v18.4h, v19.4h, v20.4h, v21.4h}, [DCT_TABLE], 32
    mul             v4.4h, v4.4h, v18.4h
    mul             v5.4h, v5.4h, v19.4h
    ins             v4.d[1], v5.d[0]
    mul             v6.4h, v6.4h, v20.4h
    mul             v7.4h, v7.4h, v21.4h
    ins             v6.d[1], v7.d[0]
    add             DCT_TABLE, DCT_TABLE, #16
    ld1             {v24.4h, v25.4h}, [DCT_TABLE], 16
    mul             v10.4h, v10.4h, v24.4h
    mul             v11.4h, v11.4h, v25.4h
    ins             v10.d[1], v11.d[0]
    add             DCT_TABLE, DCT_TABLE, #16
    ld1             {v26.4h, v27.4h}, [DCT_TABLE], 16
    mul             v12.4h, v12.4h, v26.4h
    mul             v13.4h, v13.4h, v27.4h
    ins             v12.d[1], v13.d[0]
    add             DCT_TABLE, DCT_TABLE, #16
    ld1             {v30.4h, v31.4h}, [DCT_TABLE], 16
    mul             v16.4h, v16.4h, v30.4h
    mul             v17.4h, v17.4h, v31.4h
    ins             v16.d[1], v17.d[0]

    /* Pass 1 */
#if 0
    idct_helper     v4.4h, v6.4h, v10.4h, v12.4h, v16.4h, 13, v4.4h, v6.4h
    transpose_4x4   v4.4h, v6.4h, v8.4h, v10.4h
    idct_helper     v5.4h, v7.4h, v11.4h, v13.4h, v17.4h, 13, v5.4h, v7.4h
    transpose_4x4   v5.4h, v7.4h, v9.4h, v11.4h
#else
    smull           v26.4s, v6.4h, v14.h[3]
    smlal           v26.4s, v10.4h, v14.h[2]
    smlal           v26.4s, v12.4h, v14.h[1]
    smlal           v26.4s, v16.4h, v14.h[0]
    smull           v24.4s, v7.4h, v14.h[3]
    smlal           v24.4s, v11.4h, v14.h[2]
    smlal           v24.4s, v13.4h, v14.h[1]
    smlal           v24.4s, v17.4h, v14.h[0]
    sshll           v15.4s, v4.4h, #15
    sshll           v30.4s, v5.4h, #15
    add             v20.4s, v15.4s, v26.4s
    sub             v15.4s, v15.4s, v26.4s
    rshrn           v4.4h, v20.4s, #13
    rshrn           v6.4h, v15.4s, #13
    add             v20.4s, v30.4s, v24.4s
    sub             v15.4s, v30.4s, v24.4s
    rshrn           v5.4h, v20.4s, #13
    rshrn           v7.4h, v15.4s, #13
    ins             v4.d[1], v5.d[0]
    ins             v6.d[1], v7.d[0]
    transpose       v4, v6, v3, .16b, .8h
    transpose       v6, v10, v3, .16b, .4s
    ins             v11.d[0], v10.d[1]
    ins             v7.d[0], v6.d[1]
#endif

    /* Pass 2 */
    idct_helper     v4.4h, v6.4h, v10.4h, v7.4h, v11.4h, 20, v26.4h, v27.4h

    /* Range limit */
    movi            v30.8h, #0x80
    ins             v26.d[1], v27.d[0]
    add             v26.8h, v26.8h, v30.8h
    sqxtun          v30.8b, v26.8h
    ins             v26.d[0], v30.d[0]
    sqxtun          v27.8b, v26.8h

    /* Store results to the output buffer */
    ldp             TMP1, TMP2, [OUTPUT_BUF]
    add             TMP1, TMP1, OUTPUT_COL
    add             TMP2, TMP2, OUTPUT_COL

    st1             {v26.b}[0], [TMP1], 1
    st1             {v27.b}[4], [TMP1], 1
    st1             {v26.b}[1], [TMP2], 1
    st1             {v27.b}[5], [TMP2], 1

    sub             sp, sp, #208
    ldr             x15, [sp], 16
    ld1             {v4.8b, v5.8b, v6.8b, v7.8b}, [sp], 32
    ld1             {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], 32
    ld1             {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], 32
    ld1             {v16.8b, v17.8b, v18.8b, v19.8b}, [sp], 32
    ld1             {v21.8b, v22.8b}, [sp], 16
    ld1             {v24.8b, v25.8b, v26.8b, v27.8b}, [sp], 32
    ld1             {v30.8b, v31.8b}, [sp], 16
    blr             x30

    .unreq          DCT_TABLE
    .unreq          COEF_BLOCK
    .unreq          OUTPUT_BUF
    .unreq          OUTPUT_COL
    .unreq          TMP1
    .unreq          TMP2

.purgem idct_helper


/*****************************************************************************/

/*
 * jsimd_ycc_extrgb_convert_neon
 * jsimd_ycc_extbgr_convert_neon
 * jsimd_ycc_extrgbx_convert_neon
 * jsimd_ycc_extbgrx_convert_neon
 * jsimd_ycc_extxbgr_convert_neon
 * jsimd_ycc_extxrgb_convert_neon
 *
 * Colorspace conversion YCbCr -> RGB
 */

#if defined(__APPLE__) || defined(__ANDROID__)
/* TODO: expand this to include other devices that are known not to have a slow
 *       ld3 implementation. */
#define ST3_IS_FAST
#endif

.macro do_load size
  .if \size == 8
    ld1             {v4.8b}, [U], 8
    ld1             {v5.8b}, [V], 8
    ld1             {v0.8b}, [Y], 8
    prfm            pldl1keep, [U, #64]
    prfm            pldl1keep, [V, #64]
    prfm            pldl1keep, [Y, #64]
  .elseif \size == 4
    ld1             {v4.b}[0], [U], 1
    ld1             {v4.b}[1], [U], 1
    ld1             {v4.b}[2], [U], 1
    ld1             {v4.b}[3], [U], 1
    ld1             {v5.b}[0], [V], 1
    ld1             {v5.b}[1], [V], 1
    ld1             {v5.b}[2], [V], 1
    ld1             {v5.b}[3], [V], 1
    ld1             {v0.b}[0], [Y], 1
    ld1             {v0.b}[1], [Y], 1
    ld1             {v0.b}[2], [Y], 1
    ld1             {v0.b}[3], [Y], 1
  .elseif \size == 2
    ld1             {v4.b}[4], [U], 1
    ld1             {v4.b}[5], [U], 1
    ld1             {v5.b}[4], [V], 1
    ld1             {v5.b}[5], [V], 1
    ld1             {v0.b}[4], [Y], 1
    ld1             {v0.b}[5], [Y], 1
  .elseif \size == 1
    ld1             {v4.b}[6], [U], 1
    ld1             {v5.b}[6], [V], 1
    ld1             {v0.b}[6], [Y], 1
  .else
    .error unsupported macroblock size
  .endif
.endm

.macro do_store bpp, size
  .if \bpp == 24
    .if \size == 8
#ifdef ST3_IS_FAST
      st3           {v10.8b, v11.8b, v12.8b}, [RGB], 24
#else
      st1           {v10.b}[0], [RGB], #1
      st1           {v11.b}[0], [RGB], #1
      st1           {v12.b}[0], [RGB], #1

      st1           {v10.b}[1], [RGB], #1
      st1           {v11.b}[1], [RGB], #1
      st1           {v12.b}[1], [RGB], #1

      st1           {v10.b}[2], [RGB], #1
      st1           {v11.b}[2], [RGB], #1
      st1           {v12.b}[2], [RGB], #1

      st1           {v10.b}[3], [RGB], #1
      st1           {v11.b}[3], [RGB], #1
      st1           {v12.b}[3], [RGB], #1

      st1           {v10.b}[4], [RGB], #1
      st1           {v11.b}[4], [RGB], #1
      st1           {v12.b}[4], [RGB], #1

      st1           {v10.b}[5], [RGB], #1
      st1           {v11.b}[5], [RGB], #1
      st1           {v12.b}[5], [RGB], #1

      st1           {v10.b}[6], [RGB], #1
      st1           {v11.b}[6], [RGB], #1
      st1           {v12.b}[6], [RGB], #1

      st1           {v10.b}[7], [RGB], #1
      st1           {v11.b}[7], [RGB], #1
      st1           {v12.b}[7], [RGB], #1
#endif
    .elseif \size == 4
      st3           {v10.b, v11.b, v12.b}[0], [RGB], 3
      st3           {v10.b, v11.b, v12.b}[1], [RGB], 3
      st3           {v10.b, v11.b, v12.b}[2], [RGB], 3
      st3           {v10.b, v11.b, v12.b}[3], [RGB], 3
    .elseif \size == 2
      st3           {v10.b, v11.b, v12.b}[4], [RGB], 3
      st3           {v10.b, v11.b, v12.b}[5], [RGB], 3
    .elseif \size == 1
      st3           {v10.b, v11.b, v12.b}[6], [RGB], 3
    .else
     .error unsupported macroblock size
    .endif
  .elseif \bpp == 32
    .if \size == 8
      st4           {v10.8b, v11.8b, v12.8b, v13.8b}, [RGB], 32
    .elseif \size == 4
      st4           {v10.b, v11.b, v12.b, v13.b}[0], [RGB], 4
      st4           {v10.b, v11.b, v12.b, v13.b}[1], [RGB], 4
      st4           {v10.b, v11.b, v12.b, v13.b}[2], [RGB], 4
      st4           {v10.b, v11.b, v12.b, v13.b}[3], [RGB], 4
    .elseif \size == 2
      st4           {v10.b, v11.b, v12.b, v13.b}[4], [RGB], 4
      st4           {v10.b, v11.b, v12.b, v13.b}[5], [RGB], 4
    .elseif \size == 1
      st4           {v10.b, v11.b, v12.b, v13.b}[6], [RGB], 4
    .else
      .error unsupported macroblock size
    .endif
  .elseif \bpp==16
    .if \size == 8
      st1           {v25.8h}, [RGB], 16
    .elseif \size == 4
      st1           {v25.4h}, [RGB], 8
    .elseif \size == 2
      st1           {v25.h}[4], [RGB], 2
      st1           {v25.h}[5], [RGB], 2
    .elseif \size == 1
      st1           {v25.h}[6], [RGB], 2
    .else
      .error unsupported macroblock size
    .endif
  .else
    .error unsupported bpp
  .endif
.endm

.macro generate_jsimd_ycc_rgb_convert_neon colorid, bpp, r_offs, rsize, g_offs, gsize, b_offs, bsize, defsize

/*
 * 2-stage pipelined YCbCr->RGB conversion
 */

.macro do_yuv_to_rgb_stage1
    uaddw           v6.8h, v2.8h, v4.8b     /* q3 = u - 128 */
    uaddw           v8.8h, v2.8h, v5.8b     /* q2 = v - 128 */
    smull           v20.4s, v6.4h, v1.h[1]  /* multiply by -11277 */
    smlal           v20.4s, v8.4h, v1.h[2]  /* multiply by -23401 */
    smull2          v22.4s, v6.8h, v1.h[1]  /* multiply by -11277 */
    smlal2          v22.4s, v8.8h, v1.h[2]  /* multiply by -23401 */
    smull           v24.4s, v8.4h, v1.h[0]  /* multiply by 22971 */
    smull2          v26.4s, v8.8h, v1.h[0]  /* multiply by 22971 */
    smull           v28.4s, v6.4h, v1.h[3]  /* multiply by 29033 */
    smull2          v30.4s, v6.8h, v1.h[3]  /* multiply by 29033 */
.endm

.macro do_yuv_to_rgb_stage2
    rshrn           v20.4h, v20.4s, #15
    rshrn2          v20.8h, v22.4s, #15
    rshrn           v24.4h, v24.4s, #14
    rshrn2          v24.8h, v26.4s, #14
    rshrn           v28.4h, v28.4s, #14
    rshrn2          v28.8h, v30.4s, #14
    uaddw           v20.8h, v20.8h, v0.8b
    uaddw           v24.8h, v24.8h, v0.8b
    uaddw           v28.8h, v28.8h, v0.8b
  .if \bpp != 16
    sqxtun          v1\g_offs\defsize, v20.8h
    sqxtun          v1\r_offs\defsize, v24.8h
    sqxtun          v1\b_offs\defsize, v28.8h
  .else
    sqshlu          v21.8h, v20.8h, #8
    sqshlu          v25.8h, v24.8h, #8
    sqshlu          v29.8h, v28.8h, #8
    sri             v25.8h, v21.8h, #5
    sri             v25.8h, v29.8h, #11
  .endif
.endm

.macro do_yuv_to_rgb_stage2_store_load_stage1
    rshrn           v20.4h, v20.4s, #15
    rshrn           v24.4h, v24.4s, #14
    rshrn           v28.4h, v28.4s, #14
    ld1             {v4.8b}, [U], 8
    rshrn2          v20.8h, v22.4s, #15
    rshrn2          v24.8h, v26.4s, #14
    rshrn2          v28.8h, v30.4s, #14
    ld1             {v5.8b}, [V], 8
    uaddw           v20.8h, v20.8h, v0.8b
    uaddw           v24.8h, v24.8h, v0.8b
    uaddw           v28.8h, v28.8h, v0.8b
  .if \bpp != 16  /**************** rgb24/rgb32 ******************************/
    sqxtun          v1\g_offs\defsize, v20.8h
    ld1             {v0.8b}, [Y], 8
    sqxtun          v1\r_offs\defsize, v24.8h
    prfm            pldl1keep, [U, #64]
    prfm            pldl1keep, [V, #64]
    prfm            pldl1keep, [Y, #64]
    sqxtun          v1\b_offs\defsize, v28.8h
    uaddw           v6.8h, v2.8h, v4.8b     /* v6.16b = u - 128 */
    uaddw           v8.8h, v2.8h, v5.8b     /* q2 = v - 128 */
    smull           v20.4s, v6.4h, v1.h[1]  /* multiply by -11277 */
    smlal           v20.4s, v8.4h, v1.h[2]  /* multiply by -23401 */
    smull2          v22.4s, v6.8h, v1.h[1]  /* multiply by -11277 */
    smlal2          v22.4s, v8.8h, v1.h[2]  /* multiply by -23401 */
    smull           v24.4s, v8.4h, v1.h[0]  /* multiply by 22971 */
    smull2          v26.4s, v8.8h, v1.h[0]  /* multiply by 22971 */
  .else  /**************************** rgb565 ********************************/
    sqshlu          v21.8h, v20.8h, #8
    sqshlu          v25.8h, v24.8h, #8
    sqshlu          v29.8h, v28.8h, #8
    uaddw           v6.8h, v2.8h, v4.8b     /* v6.16b = u - 128 */
    uaddw           v8.8h, v2.8h, v5.8b     /* q2 = v - 128 */
    ld1             {v0.8b}, [Y], 8
    smull           v20.4s, v6.4h, v1.h[1]  /* multiply by -11277 */
    smlal           v20.4s, v8.4h, v1.h[2]  /* multiply by -23401 */
    smull2          v22.4s, v6.8h, v1.h[1]  /* multiply by -11277 */
    smlal2          v22.4s, v8.8h, v1.h[2]  /* multiply by -23401 */
    sri             v25.8h, v21.8h, #5
    smull           v24.4s, v8.4h, v1.h[0]  /* multiply by 22971 */
    smull2          v26.4s, v8.8h, v1.h[0]  /* multiply by 22971 */
    prfm            pldl1keep, [U, #64]
    prfm            pldl1keep, [V, #64]
    prfm            pldl1keep, [Y, #64]
    sri             v25.8h, v29.8h, #11
  .endif
    do_store        \bpp, 8
    smull           v28.4s, v6.4h, v1.h[3]  /* multiply by 29033 */
    smull2          v30.4s, v6.8h, v1.h[3]  /* multiply by 29033 */
.endm

.macro do_yuv_to_rgb
    do_yuv_to_rgb_stage1
    do_yuv_to_rgb_stage2
.endm

/* Apple gas crashes on adrl, work around that by using adr.
 * But this requires a copy of these constants for each function.
 */

.balign 16
Ljsimd_ycc_\colorid\()_neon_consts:
  .short 0,      0,     0,      0
  .short 22971, -11277, -23401, 29033
  .short -128,  -128,   -128,   -128
  .short -128,  -128,   -128,   -128

asm_function jsimd_ycc_\colorid\()_convert_neon
    OUTPUT_WIDTH    .req x0
    INPUT_BUF       .req x1
    INPUT_ROW       .req x2
    OUTPUT_BUF      .req x3
    NUM_ROWS        .req x4

    INPUT_BUF0      .req x5
    INPUT_BUF1      .req x6
    INPUT_BUF2      .req x1

    RGB             .req x7
    Y               .req x8
    U               .req x9
    V               .req x10
    N               .req x15

    sub             sp, sp, 336
    str             x15, [sp], 16

    /* Load constants to d1, d2, d3 (v0.4h is just used for padding) */
    adr             x15, Ljsimd_ycc_\colorid\()_neon_consts

    /* Save NEON registers */
    st1             {v0.8b, v1.8b, v2.8b, v3.8b}, [sp], 32
    st1             {v4.8b, v5.8b, v6.8b, v7.8b}, [sp], 32
    st1             {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], 32
    st1             {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], 32
    st1             {v16.8b, v17.8b, v18.8b, v19.8b}, [sp], 32
    st1             {v20.8b, v21.8b, v22.8b, v23.8b}, [sp], 32
    st1             {v24.8b, v25.8b, v26.8b, v27.8b}, [sp], 32
    st1             {v28.8b, v29.8b, v30.8b, v31.8b}, [sp], 32
    ld1             {v0.4h, v1.4h}, [x15], 16
    ld1             {v2.8h}, [x15]

    /* Save ARM registers and handle input arguments */
    /* push            {x4, x5, x6, x7, x8, x9, x10, x30} */
    stp             x4, x5, [sp], 16
    stp             x6, x7, [sp], 16
    stp             x8, x9, [sp], 16
    stp             x10, x30, [sp], 16
    ldr             INPUT_BUF0, [INPUT_BUF]
    ldr             INPUT_BUF1, [INPUT_BUF, #8]
    ldr             INPUT_BUF2, [INPUT_BUF, #16]
    .unreq          INPUT_BUF

    /* Initially set v10, v11.4h, v12.8b, d13 to 0xFF */
    movi            v10.16b, #255
    movi            v13.16b, #255

    /* Outer loop over scanlines */
    cmp             NUM_ROWS, #1
    b.lt            9f
0:
    lsl             x16, INPUT_ROW, #3
    ldr             Y, [INPUT_BUF0, x16]
    ldr             U, [INPUT_BUF1, x16]
    mov             N, OUTPUT_WIDTH
    ldr             V, [INPUT_BUF2, x16]
    add             INPUT_ROW, INPUT_ROW, #1
    ldr             RGB, [OUTPUT_BUF], #8

    /* Inner loop over pixels */
    subs            N, N, #8
    b.lt            3f
    do_load         8
    do_yuv_to_rgb_stage1
    subs            N, N, #8
    b.lt            2f
1:
    do_yuv_to_rgb_stage2_store_load_stage1
    subs            N, N, #8
    b.ge            1b
2:
    do_yuv_to_rgb_stage2
    do_store        \bpp, 8
    tst             N, #7
    b.eq            8f
3:
    tst             N, #4
    b.eq            3f
    do_load         4
3:
    tst             N, #2
    b.eq            4f
    do_load         2
4:
    tst             N, #1
    b.eq            5f
    do_load         1
5:
    do_yuv_to_rgb
    tst             N, #4
    b.eq            6f
    do_store        \bpp, 4
6:
    tst             N, #2
    b.eq            7f
    do_store        \bpp, 2
7:
    tst             N, #1
    b.eq            8f
    do_store        \bpp, 1
8:
    subs            NUM_ROWS, NUM_ROWS, #1
    b.gt            0b
9:
    /* Restore all registers and return */
    sub             sp, sp, #336
    ldr             x15, [sp], 16
    ld1             {v0.8b, v1.8b, v2.8b, v3.8b}, [sp], 32
    ld1             {v4.8b, v5.8b, v6.8b, v7.8b}, [sp], 32
    ld1             {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], 32
    ld1             {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], 32
    ld1             {v16.8b, v17.8b, v18.8b, v19.8b}, [sp], 32
    ld1             {v20.8b, v21.8b, v22.8b, v23.8b}, [sp], 32
    ld1             {v24.8b, v25.8b, v26.8b, v27.8b}, [sp], 32
    ld1             {v28.8b, v29.8b, v30.8b, v31.8b}, [sp], 32
    /* pop             {r4, r5, r6, r7, r8, r9, r10, pc} */
    ldp             x4, x5, [sp], 16
    ldp             x6, x7, [sp], 16
    ldp             x8, x9, [sp], 16
    ldp             x10, x30, [sp], 16
    br              x30
    .unreq          OUTPUT_WIDTH
    .unreq          INPUT_ROW
    .unreq          OUTPUT_BUF
    .unreq          NUM_ROWS
    .unreq          INPUT_BUF0
    .unreq          INPUT_BUF1
    .unreq          INPUT_BUF2
    .unreq          RGB
    .unreq          Y
    .unreq          U
    .unreq          V
    .unreq          N

.purgem do_yuv_to_rgb
.purgem do_yuv_to_rgb_stage1
.purgem do_yuv_to_rgb_stage2
.purgem do_yuv_to_rgb_stage2_store_load_stage1

.endm

/*--------------------------------- id ----- bpp R  rsize G  gsize B  bsize defsize */
generate_jsimd_ycc_rgb_convert_neon extrgb,  24, 0, .4h,  1, .4h,  2, .4h,  .8b
generate_jsimd_ycc_rgb_convert_neon extbgr,  24, 2, .4h,  1, .4h,  0, .4h,  .8b
generate_jsimd_ycc_rgb_convert_neon extrgbx, 32, 0, .4h,  1, .4h,  2, .4h,  .8b
generate_jsimd_ycc_rgb_convert_neon extbgrx, 32, 2, .4h,  1, .4h,  0, .4h,  .8b
generate_jsimd_ycc_rgb_convert_neon extxbgr, 32, 3, .4h,  2, .4h,  1, .4h,  .8b
generate_jsimd_ycc_rgb_convert_neon extxrgb, 32, 1, .4h,  2, .4h,  3, .4h,  .8b
generate_jsimd_ycc_rgb_convert_neon rgb565,  16, 0, .4h,  0, .4h,  0, .4h,  .8b

.purgem do_load
.purgem do_store


/*****************************************************************************/

/*
 * jsimd_extrgb_ycc_convert_neon
 * jsimd_extbgr_ycc_convert_neon
 * jsimd_extrgbx_ycc_convert_neon
 * jsimd_extbgrx_ycc_convert_neon
 * jsimd_extxbgr_ycc_convert_neon
 * jsimd_extxrgb_ycc_convert_neon
 *
 * Colorspace conversion RGB -> YCbCr
 */

.macro do_store size
  .if \size == 8
    st1             {v20.8b}, [Y], #8
    st1             {v21.8b}, [U], #8
    st1             {v22.8b}, [V], #8
  .elseif \size == 4
    st1             {v20.b}[0], [Y], #1
    st1             {v20.b}[1], [Y], #1
    st1             {v20.b}[2], [Y], #1
    st1             {v20.b}[3], [Y], #1
    st1             {v21.b}[0], [U], #1
    st1             {v21.b}[1], [U], #1
    st1             {v21.b}[2], [U], #1
    st1             {v21.b}[3], [U], #1
    st1             {v22.b}[0], [V], #1
    st1             {v22.b}[1], [V], #1
    st1             {v22.b}[2], [V], #1
    st1             {v22.b}[3], [V], #1
  .elseif \size == 2
    st1             {v20.b}[4], [Y], #1
    st1             {v20.b}[5], [Y], #1
    st1             {v21.b}[4], [U], #1
    st1             {v21.b}[5], [U], #1
    st1             {v22.b}[4], [V], #1
    st1             {v22.b}[5], [V], #1
  .elseif \size == 1
    st1             {v20.b}[6], [Y], #1
    st1             {v21.b}[6], [U], #1
    st1             {v22.b}[6], [V], #1
  .else
    .error unsupported macroblock size
  .endif
.endm

#if defined(__APPLE__) || defined(__ANDROID__)
/* TODO: expand this to include other devices that are known not to have a slow
 *       ld3 implementation. */
#define LD3_IS_FAST
#endif

.macro do_load bpp, size
  .if \bpp == 24
    .if \size == 8
#ifdef LD3_IS_FAST
      ld3           {v10.8b, v11.8b, v12.8b}, [RGB], #24
#else
      ld1           {v10.b}[0], [RGB], #1
      ld1           {v11.b}[0], [RGB], #1
      ld1           {v12.b}[0], [RGB], #1

      ld1           {v10.b}[1], [RGB], #1
      ld1           {v11.b}[1], [RGB], #1
      ld1           {v12.b}[1], [RGB], #1

      ld1           {v10.b}[2], [RGB], #1
      ld1           {v11.b}[2], [RGB], #1
      ld1           {v12.b}[2], [RGB], #1

      ld1           {v10.b}[3], [RGB], #1
      ld1           {v11.b}[3], [RGB], #1
      ld1           {v12.b}[3], [RGB], #1

      ld1           {v10.b}[4], [RGB], #1
      ld1           {v11.b}[4], [RGB], #1
      ld1           {v12.b}[4], [RGB], #1

      ld1           {v10.b}[5], [RGB], #1
      ld1           {v11.b}[5], [RGB], #1
      ld1           {v12.b}[5], [RGB], #1

      ld1           {v10.b}[6], [RGB], #1
      ld1           {v11.b}[6], [RGB], #1
      ld1           {v12.b}[6], [RGB], #1

      ld1           {v10.b}[7], [RGB], #1
      ld1           {v11.b}[7], [RGB], #1
      ld1           {v12.b}[7], [RGB], #1
#endif
      prfm          pldl1keep, [RGB, #128]
    .elseif \size == 4
      ld3           {v10.b, v11.b, v12.b}[0], [RGB], #3
      ld3           {v10.b, v11.b, v12.b}[1], [RGB], #3
      ld3           {v10.b, v11.b, v12.b}[2], [RGB], #3
      ld3           {v10.b, v11.b, v12.b}[3], [RGB], #3
    .elseif \size == 2
      ld3           {v10.b, v11.b, v12.b}[4], [RGB], #3
      ld3           {v10.b, v11.b, v12.b}[5], [RGB], #3
    .elseif \size == 1
      ld3           {v10.b, v11.b, v12.b}[6], [RGB], #3
    .else
      .error unsupported macroblock size
    .endif
  .elseif \bpp == 32
    .if \size == 8
      ld4           {v10.8b, v11.8b, v12.8b, v13.8b}, [RGB], #32
      prfm          pldl1keep, [RGB, #128]
    .elseif \size == 4
      ld4           {v10.b, v11.b, v12.b, v13.b}[0], [RGB], #4
      ld4           {v10.b, v11.b, v12.b, v13.b}[1], [RGB], #4
      ld4           {v10.b, v11.b, v12.b, v13.b}[2], [RGB], #4
      ld4           {v10.b, v11.b, v12.b, v13.b}[3], [RGB], #4
    .elseif \size == 2
      ld4           {v10.b, v11.b, v12.b, v13.b}[4], [RGB], #4
      ld4           {v10.b, v11.b, v12.b, v13.b}[5], [RGB], #4
    .elseif \size == 1
      ld4           {v10.b, v11.b, v12.b, v13.b}[6], [RGB], #4
    .else
      .error unsupported macroblock size
    .endif
  .else
    .error unsupported bpp
  .endif
.endm

.macro generate_jsimd_rgb_ycc_convert_neon colorid, bpp, r_offs, g_offs, b_offs

/*
 * 2-stage pipelined RGB->YCbCr conversion
 */

.macro do_rgb_to_yuv_stage1
    ushll           v4.8h, v1\r_offs\().8b, #0  /* r = v4 */
    ushll           v6.8h, v1\g_offs\().8b, #0  /* g = v6 */
    ushll           v8.8h, v1\b_offs\().8b, #0  /* b = v8 */
    rev64           v18.4s, v1.4s
    rev64           v26.4s, v1.4s
    rev64           v28.4s, v1.4s
    rev64           v30.4s, v1.4s
    umull           v14.4s, v4.4h, v0.h[0]
    umull2          v16.4s, v4.8h, v0.h[0]
    umlsl           v18.4s, v4.4h, v0.h[3]
    umlsl2          v26.4s, v4.8h, v0.h[3]
    umlal           v28.4s, v4.4h, v0.h[5]
    umlal2          v30.4s, v4.8h, v0.h[5]
    umlal           v14.4s, v6.4h, v0.h[1]
    umlal2          v16.4s, v6.8h, v0.h[1]
    umlsl           v18.4s, v6.4h, v0.h[4]
    umlsl2          v26.4s, v6.8h, v0.h[4]
    umlsl           v28.4s, v6.4h, v0.h[6]
    umlsl2          v30.4s, v6.8h, v0.h[6]
    umlal           v14.4s, v8.4h, v0.h[2]
    umlal2          v16.4s, v8.8h, v0.h[2]
    umlal           v18.4s, v8.4h, v0.h[5]
    umlal2          v26.4s, v8.8h, v0.h[5]
    umlsl           v28.4s, v8.4h, v0.h[7]
    umlsl2          v30.4s, v8.8h, v0.h[7]
.endm

.macro do_rgb_to_yuv_stage2
    rshrn           v20.4h, v14.4s, #16
    shrn            v22.4h, v18.4s, #16
    shrn            v24.4h, v28.4s, #16
    rshrn2          v20.8h, v16.4s, #16
    shrn2           v22.8h, v26.4s, #16
    shrn2           v24.8h, v30.4s, #16
    xtn             v20.8b, v20.8h       /* v20 = y */
    xtn             v21.8b, v22.8h       /* v21 = u */
    xtn             v22.8b, v24.8h       /* v22 = v */
.endm

.macro do_rgb_to_yuv
    do_rgb_to_yuv_stage1
    do_rgb_to_yuv_stage2
.endm

/* TODO: expand macros and interleave instructions if some in-order
 *       ARM64 processor actually can dual-issue LOAD/STORE with ALU */
.macro do_rgb_to_yuv_stage2_store_load_stage1
    do_rgb_to_yuv_stage2
    do_load         \bpp, 8
    st1             {v20.8b}, [Y], #8
    st1             {v21.8b}, [U], #8
    st1             {v22.8b}, [V], #8
    do_rgb_to_yuv_stage1
.endm

.balign 16
Ljsimd_\colorid\()_ycc_neon_consts:
  .short 19595, 38470, 7471, 11059
  .short 21709, 32768, 27439, 5329
  .short 32767, 128, 32767, 128
  .short 32767, 128, 32767, 128

asm_function jsimd_\colorid\()_ycc_convert_neon
    OUTPUT_WIDTH    .req w0
    INPUT_BUF       .req x1
    OUTPUT_BUF      .req x2
    OUTPUT_ROW      .req x3
    NUM_ROWS        .req x4

    OUTPUT_BUF0     .req x5
    OUTPUT_BUF1     .req x6
    OUTPUT_BUF2     .req x2  /* OUTPUT_BUF */

    RGB             .req x7
    Y               .req x9
    U               .req x10
    V               .req x11
    N               .req w12

    /* Load constants to d0, d1, d2, d3 */
    adr             x13, Ljsimd_\colorid\()_ycc_neon_consts
    ld1             {v0.8h, v1.8h}, [x13]

    ldr             OUTPUT_BUF0, [OUTPUT_BUF]
    ldr             OUTPUT_BUF1, [OUTPUT_BUF, #8]
    ldr             OUTPUT_BUF2, [OUTPUT_BUF, #16]
    .unreq          OUTPUT_BUF

    /* Save NEON registers */
    sub             sp, sp, #64
    st1             {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], 32
    st1             {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], 32

    /* Outer loop over scanlines */
    cmp             NUM_ROWS, #1
    b.lt            9f
0:
    ldr             Y, [OUTPUT_BUF0, OUTPUT_ROW, lsl #3]
    ldr             U, [OUTPUT_BUF1, OUTPUT_ROW, lsl #3]
    mov             N, OUTPUT_WIDTH
    ldr             V, [OUTPUT_BUF2, OUTPUT_ROW, lsl #3]
    add             OUTPUT_ROW, OUTPUT_ROW, #1
    ldr             RGB, [INPUT_BUF], #8

    /* Inner loop over pixels */
    subs            N, N, #8
    b.lt            3f
    do_load         \bpp, 8
    do_rgb_to_yuv_stage1
    subs            N, N, #8
    b.lt            2f
1:
    do_rgb_to_yuv_stage2_store_load_stage1
    subs            N, N, #8
    b.ge            1b
2:
    do_rgb_to_yuv_stage2
    do_store        8
    tst             N, #7
    b.eq            8f
3:
    tbz             N, #2, 3f
    do_load         \bpp, 4
3:
    tbz             N, #1, 4f
    do_load         \bpp, 2
4:
    tbz             N, #0, 5f
    do_load         \bpp, 1
5:
    do_rgb_to_yuv
    tbz             N, #2, 6f
    do_store        4
6:
    tbz             N, #1, 7f
    do_store        2
7:
    tbz             N, #0, 8f
    do_store        1
8:
    subs            NUM_ROWS, NUM_ROWS, #1
    b.gt            0b
9:
    /* Restore all registers and return */
    sub             sp, sp, #64
    ld1             {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], 32
    ld1             {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], 32
    br              x30

    .unreq          OUTPUT_WIDTH
    .unreq          OUTPUT_ROW
    .unreq          INPUT_BUF
    .unreq          NUM_ROWS
    .unreq          OUTPUT_BUF0
    .unreq          OUTPUT_BUF1
    .unreq          OUTPUT_BUF2
    .unreq          RGB
    .unreq          Y
    .unreq          U
    .unreq          V
    .unreq          N

.purgem do_rgb_to_yuv
.purgem do_rgb_to_yuv_stage1
.purgem do_rgb_to_yuv_stage2
.purgem do_rgb_to_yuv_stage2_store_load_stage1

.endm

/*--------------------------------- id ----- bpp R  G  B */
generate_jsimd_rgb_ycc_convert_neon extrgb,  24, 0, 1, 2
generate_jsimd_rgb_ycc_convert_neon extbgr,  24, 2, 1, 0
generate_jsimd_rgb_ycc_convert_neon extrgbx, 32, 0, 1, 2
generate_jsimd_rgb_ycc_convert_neon extbgrx, 32, 2, 1, 0
generate_jsimd_rgb_ycc_convert_neon extxbgr, 32, 3, 2, 1
generate_jsimd_rgb_ycc_convert_neon extxrgb, 32, 1, 2, 3

.purgem do_load
.purgem do_store


/*****************************************************************************/

/*
 * Load data into workspace, applying unsigned->signed conversion
 *
 * TODO: can be combined with 'jsimd_fdct_ifast_neon' to get
 *       rid of VST1.16 instructions
 */

asm_function jsimd_convsamp_neon
    SAMPLE_DATA     .req x0
    START_COL       .req x1
    WORKSPACE       .req x2
    TMP1            .req x9
    TMP2            .req x10
    TMP3            .req x11
    TMP4            .req x12
    TMP5            .req x13
    TMP6            .req x14
    TMP7            .req x15
    TMP8            .req x4
    TMPDUP          .req w3

    mov             TMPDUP, #128
    ldp             TMP1, TMP2, [SAMPLE_DATA], 16
    ldp             TMP3, TMP4, [SAMPLE_DATA], 16
    dup             v0.8b, TMPDUP
    add             TMP1, TMP1, START_COL
    add             TMP2, TMP2, START_COL
    ldp             TMP5, TMP6, [SAMPLE_DATA], 16
    add             TMP3, TMP3, START_COL
    add             TMP4, TMP4, START_COL
    ldp             TMP7, TMP8, [SAMPLE_DATA], 16
    add             TMP5, TMP5, START_COL
    add             TMP6, TMP6, START_COL
    ld1             {v16.8b}, [TMP1]
    add             TMP7, TMP7, START_COL
    add             TMP8, TMP8, START_COL
    ld1             {v17.8b}, [TMP2]
    usubl           v16.8h, v16.8b, v0.8b
    ld1             {v18.8b}, [TMP3]
    usubl           v17.8h, v17.8b, v0.8b
    ld1             {v19.8b}, [TMP4]
    usubl           v18.8h, v18.8b, v0.8b
    ld1             {v20.8b}, [TMP5]
    usubl           v19.8h, v19.8b, v0.8b
    ld1             {v21.8b}, [TMP6]
    st1             {v16.8h, v17.8h, v18.8h, v19.8h}, [WORKSPACE], 64
    usubl           v20.8h, v20.8b, v0.8b
    ld1             {v22.8b}, [TMP7]
    usubl           v21.8h, v21.8b, v0.8b
    ld1             {v23.8b}, [TMP8]
    usubl           v22.8h, v22.8b, v0.8b
    usubl           v23.8h, v23.8b, v0.8b
    st1             {v20.8h, v21.8h, v22.8h, v23.8h}, [WORKSPACE], 64

    br              x30

    .unreq          SAMPLE_DATA
    .unreq          START_COL
    .unreq          WORKSPACE
    .unreq          TMP1
    .unreq          TMP2
    .unreq          TMP3
    .unreq          TMP4
    .unreq          TMP5
    .unreq          TMP6
    .unreq          TMP7
    .unreq          TMP8
    .unreq          TMPDUP

/*****************************************************************************/

/*
 * jsimd_fdct_islow_neon
 *
 * This file contains a slow-but-accurate integer implementation of the
 * forward DCT (Discrete Cosine Transform). The following code is based
 * directly on the IJG''s original jfdctint.c; see the jfdctint.c for
 * more details.
 *
 * TODO: can be combined with 'jsimd_convsamp_neon' to get
 *       rid of a bunch of VLD1.16 instructions
 */

#define CONST_BITS 13
#define PASS1_BITS 2

#define DESCALE_P1 (CONST_BITS-PASS1_BITS)
#define DESCALE_P2 (CONST_BITS+PASS1_BITS)

#define F_0_298  2446  /* FIX(0.298631336) */
#define F_0_390  3196  /* FIX(0.390180644) */
#define F_0_541  4433  /* FIX(0.541196100) */
#define F_0_765  6270  /* FIX(0.765366865) */
#define F_0_899  7373  /* FIX(0.899976223) */
#define F_1_175  9633  /* FIX(1.175875602) */
#define F_1_501 12299  /* FIX(1.501321110) */
#define F_1_847 15137  /* FIX(1.847759065) */
#define F_1_961 16069  /* FIX(1.961570560) */
#define F_2_053 16819  /* FIX(2.053119869) */
#define F_2_562 20995  /* FIX(2.562915447) */
#define F_3_072 25172  /* FIX(3.072711026) */

.balign 16
Ljsimd_fdct_islow_neon_consts:
  .short F_0_298
  .short -F_0_390
  .short F_0_541
  .short F_0_765
  .short - F_0_899
  .short F_1_175
  .short F_1_501
  .short - F_1_847
  .short - F_1_961
  .short F_2_053
  .short - F_2_562
  .short F_3_072
  .short 0          /* padding */
  .short 0
  .short 0
  .short 0

#undef F_0_298
#undef F_0_390
#undef F_0_541
#undef F_0_765
#undef F_0_899
#undef F_1_175
#undef F_1_501
#undef F_1_847
#undef F_1_961
#undef F_2_053
#undef F_2_562
#undef F_3_072
#define XFIX_P_0_298 v0.h[0]
#define XFIX_N_0_390 v0.h[1]
#define XFIX_P_0_541 v0.h[2]
#define XFIX_P_0_765 v0.h[3]
#define XFIX_N_0_899 v0.h[4]
#define XFIX_P_1_175 v0.h[5]
#define XFIX_P_1_501 v0.h[6]
#define XFIX_N_1_847 v0.h[7]
#define XFIX_N_1_961 v1.h[0]
#define XFIX_P_2_053 v1.h[1]
#define XFIX_N_2_562 v1.h[2]
#define XFIX_P_3_072 v1.h[3]

asm_function jsimd_fdct_islow_neon

    DATA            .req x0
    TMP             .req x9

    /* Load constants */
    adr             TMP, Ljsimd_fdct_islow_neon_consts
    ld1             {v0.8h, v1.8h}, [TMP]

    /* Save NEON registers */
    sub             sp, sp, #64
    st1             {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], 32
    st1             {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], 32

    /* Load all DATA into NEON registers with the following allocation:
     *       0 1 2 3 | 4 5 6 7
     *      ---------+--------
     *   0 | d16     | d17    | v16.8h
     *   1 | d18     | d19    | v17.8h
     *   2 | d20     | d21    | v18.8h
     *   3 | d22     | d23    | v19.8h
     *   4 | d24     | d25    | v20.8h
     *   5 | d26     | d27    | v21.8h
     *   6 | d28     | d29    | v22.8h
     *   7 | d30     | d31    | v23.8h
     */

    ld1             {v16.8h, v17.8h, v18.8h, v19.8h}, [DATA], 64
    ld1             {v20.8h, v21.8h, v22.8h, v23.8h}, [DATA]
    sub             DATA, DATA, #64

    /* Transpose */
    transpose_8x8   v16, v17, v18, v19, v20, v21, v22, v23, v31, v2, v3, v4
    /* 1-D FDCT */
    add             v24.8h, v16.8h, v23.8h  /* tmp0 = dataptr[0] + dataptr[7]; */
    sub             v31.8h, v16.8h, v23.8h  /* tmp7 = dataptr[0] - dataptr[7]; */
    add             v25.8h, v17.8h, v22.8h  /* tmp1 = dataptr[1] + dataptr[6]; */
    sub             v30.8h, v17.8h, v22.8h  /* tmp6 = dataptr[1] - dataptr[6]; */
    add             v26.8h, v18.8h, v21.8h  /* tmp2 = dataptr[2] + dataptr[5]; */
    sub             v29.8h, v18.8h, v21.8h  /* tmp5 = dataptr[2] - dataptr[5]; */
    add             v27.8h, v19.8h, v20.8h  /* tmp3 = dataptr[3] + dataptr[4]; */
    sub             v28.8h, v19.8h, v20.8h  /* tmp4 = dataptr[3] - dataptr[4]; */

    /* even part */

    add             v8.8h, v24.8h, v27.8h   /* tmp10 = tmp0 + tmp3; */
    sub             v9.8h, v24.8h, v27.8h   /* tmp13 = tmp0 - tmp3; */
    add             v10.8h, v25.8h, v26.8h  /* tmp11 = tmp1 + tmp2; */
    sub             v11.8h, v25.8h, v26.8h  /* tmp12 = tmp1 - tmp2; */

    add             v16.8h, v8.8h, v10.8h  /* tmp10 + tmp11 */
    sub             v20.8h, v8.8h, v10.8h  /* tmp10 - tmp11 */

    add             v18.8h, v11.8h, v9.8h  /* tmp12 + tmp13 */

    shl             v16.8h, v16.8h, #PASS1_BITS  /* dataptr[0] = (DCTELEM) LEFT_SHIFT(tmp10 + tmp11, PASS1_BITS); */
    shl             v20.8h, v20.8h, #PASS1_BITS  /* dataptr[4] = (DCTELEM) LEFT_SHIFT(tmp10 - tmp11, PASS1_BITS); */

    smull2          v24.4s, v18.8h, XFIX_P_0_541  /* z1 hi = MULTIPLY(tmp12 + tmp13, XFIX_P_0_541); */
    smull           v18.4s, v18.4h, XFIX_P_0_541  /* z1 lo = MULTIPLY(tmp12 + tmp13, XFIX_P_0_541); */
    mov             v22.16b, v18.16b
    mov             v25.16b, v24.16b

    smlal           v18.4s, v9.4h, XFIX_P_0_765   /* lo z1 + MULTIPLY(tmp13, XFIX_P_0_765) */
    smlal2          v24.4s, v9.8h, XFIX_P_0_765   /* hi z1 + MULTIPLY(tmp13, XFIX_P_0_765) */
    smlal           v22.4s, v11.4h, XFIX_N_1_847  /* lo z1 + MULTIPLY(tmp12, XFIX_N_1_847) */
    smlal2          v25.4s, v11.8h, XFIX_N_1_847  /* hi z1 + MULTIPLY(tmp12, XFIX_N_1_847) */

    rshrn           v18.4h, v18.4s, #DESCALE_P1
    rshrn           v22.4h, v22.4s, #DESCALE_P1
    rshrn2          v18.8h, v24.4s, #DESCALE_P1  /* dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, XFIX_P_0_765), CONST_BITS-PASS1_BITS); */
    rshrn2          v22.8h, v25.4s, #DESCALE_P1  /* dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, XFIX_N_1_847), CONST_BITS-PASS1_BITS); */

    /* Odd part */

    add             v8.8h, v28.8h, v31.8h        /* z1 = tmp4 + tmp7; */
    add             v9.8h, v29.8h, v30.8h        /* z2 = tmp5 + tmp6; */
    add             v10.8h, v28.8h, v30.8h       /* z3 = tmp4 + tmp6; */
    add             v11.8h, v29.8h, v31.8h       /* z4 = tmp5 + tmp7; */
    smull           v4.4s, v10.4h, XFIX_P_1_175  /* z5 lo = z3 lo * XFIX_P_1_175 */
    smull2          v5.4s, v10.8h, XFIX_P_1_175
    smlal           v4.4s, v11.4h, XFIX_P_1_175  /* z5 = MULTIPLY(z3 + z4, FIX_1_175875602); */
    smlal2          v5.4s, v11.8h, XFIX_P_1_175

    smull2          v24.4s, v28.8h, XFIX_P_0_298
    smull2          v25.4s, v29.8h, XFIX_P_2_053
    smull2          v26.4s, v30.8h, XFIX_P_3_072
    smull2          v27.4s, v31.8h, XFIX_P_1_501
    smull           v28.4s, v28.4h, XFIX_P_0_298  /* tmp4 = MULTIPLY(tmp4, FIX_0_298631336); */
    smull           v29.4s, v29.4h, XFIX_P_2_053  /* tmp5 = MULTIPLY(tmp5, FIX_2_053119869); */
    smull           v30.4s, v30.4h, XFIX_P_3_072  /* tmp6 = MULTIPLY(tmp6, FIX_3_072711026); */
    smull           v31.4s, v31.4h, XFIX_P_1_501  /* tmp7 = MULTIPLY(tmp7, FIX_1_501321110); */

    smull2          v12.4s, v8.8h, XFIX_N_0_899
    smull2          v13.4s, v9.8h, XFIX_N_2_562
    smull2          v14.4s, v10.8h, XFIX_N_1_961
    smull2          v15.4s, v11.8h, XFIX_N_0_390
    smull           v8.4s, v8.4h, XFIX_N_0_899    /* z1 = MULTIPLY(z1, - FIX_0_899976223); */
    smull           v9.4s, v9.4h, XFIX_N_2_562    /* z2 = MULTIPLY(z2, - FIX_2_562915447); */
    smull           v10.4s, v10.4h, XFIX_N_1_961  /* z3 = MULTIPLY(z3, - FIX_1_961570560); */
    smull           v11.4s, v11.4h, XFIX_N_0_390  /* z4 = MULTIPLY(z4, - FIX_0_390180644); */

    add             v10.4s, v10.4s, v4.4s  /* z3 += z5 */
    add             v14.4s, v14.4s, v5.4s
    add             v11.4s, v11.4s, v4.4s  /* z4 += z5 */
    add             v15.4s, v15.4s, v5.4s

    add             v28.4s, v28.4s, v8.4s   /* tmp4 += z1 */
    add             v24.4s, v24.4s, v12.4s
    add             v29.4s, v29.4s, v9.4s   /* tmp5 += z2 */
    add             v25.4s, v25.4s, v13.4s
    add             v30.4s, v30.4s, v10.4s  /* tmp6 += z3 */
    add             v26.4s, v26.4s, v14.4s
    add             v31.4s, v31.4s, v11.4s  /* tmp7 += z4 */
    add             v27.4s, v27.4s, v15.4s

    add             v28.4s, v28.4s, v10.4s  /* tmp4 += z3 */
    add             v24.4s, v24.4s, v14.4s
    add             v29.4s, v29.4s, v11.4s  /* tmp5 += z4 */
    add             v25.4s, v25.4s, v15.4s
    add             v30.4s, v30.4s, v9.4s   /* tmp6 += z2 */
    add             v26.4s, v26.4s, v13.4s
    add             v31.4s, v31.4s, v8.4s   /* tmp7 += z1 */
    add             v27.4s, v27.4s, v12.4s

    rshrn           v23.4h, v28.4s, #DESCALE_P1
    rshrn           v21.4h, v29.4s, #DESCALE_P1
    rshrn           v19.4h, v30.4s, #DESCALE_P1
    rshrn           v17.4h, v31.4s, #DESCALE_P1
    rshrn2          v23.8h, v24.4s, #DESCALE_P1  /* dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS); */
    rshrn2          v21.8h, v25.4s, #DESCALE_P1  /* dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS); */
    rshrn2          v19.8h, v26.4s, #DESCALE_P1  /* dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS); */
    rshrn2          v17.8h, v27.4s, #DESCALE_P1  /* dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS); */

    /* Transpose */
    transpose_8x8   v16, v17, v18, v19, v20, v21, v22, v23, v31, v2, v3, v4

    /* 1-D FDCT */
    add             v24.8h, v16.8h, v23.8h  /* tmp0 = dataptr[0] + dataptr[7]; */
    sub             v31.8h, v16.8h, v23.8h  /* tmp7 = dataptr[0] - dataptr[7]; */
    add             v25.8h, v17.8h, v22.8h  /* tmp1 = dataptr[1] + dataptr[6]; */
    sub             v30.8h, v17.8h, v22.8h  /* tmp6 = dataptr[1] - dataptr[6]; */
    add             v26.8h, v18.8h, v21.8h  /* tmp2 = dataptr[2] + dataptr[5]; */
    sub             v29.8h, v18.8h, v21.8h  /* tmp5 = dataptr[2] - dataptr[5]; */
    add             v27.8h, v19.8h, v20.8h  /* tmp3 = dataptr[3] + dataptr[4]; */
    sub             v28.8h, v19.8h, v20.8h  /* tmp4 = dataptr[3] - dataptr[4]; */

    /* even part */
    add             v8.8h, v24.8h, v27.8h   /* tmp10 = tmp0 + tmp3; */
    sub             v9.8h, v24.8h, v27.8h   /* tmp13 = tmp0 - tmp3; */
    add             v10.8h, v25.8h, v26.8h  /* tmp11 = tmp1 + tmp2; */
    sub             v11.8h, v25.8h, v26.8h  /* tmp12 = tmp1 - tmp2; */

    add             v16.8h, v8.8h, v10.8h  /* tmp10 + tmp11 */
    sub             v20.8h, v8.8h, v10.8h  /* tmp10 - tmp11 */

    add             v18.8h, v11.8h, v9.8h  /* tmp12 + tmp13 */

    srshr           v16.8h, v16.8h, #PASS1_BITS  /* dataptr[0] = (DCTELEM) DESCALE(tmp10 + tmp11, PASS1_BITS); */
    srshr           v20.8h, v20.8h, #PASS1_BITS  /* dataptr[4] = (DCTELEM) DESCALE(tmp10 - tmp11, PASS1_BITS); */

    smull2          v24.4s, v18.8h, XFIX_P_0_541  /* z1 hi = MULTIPLY(tmp12 + tmp13, XFIX_P_0_541); */
    smull           v18.4s, v18.4h, XFIX_P_0_541  /* z1 lo = MULTIPLY(tmp12 + tmp13, XFIX_P_0_541); */
    mov             v22.16b, v18.16b
    mov             v25.16b, v24.16b

    smlal           v18.4s, v9.4h, XFIX_P_0_765   /* lo z1 + MULTIPLY(tmp13, XFIX_P_0_765) */
    smlal2          v24.4s, v9.8h, XFIX_P_0_765   /* hi z1 + MULTIPLY(tmp13, XFIX_P_0_765) */
    smlal           v22.4s, v11.4h, XFIX_N_1_847  /* lo z1 + MULTIPLY(tmp12, XFIX_N_1_847) */
    smlal2          v25.4s, v11.8h, XFIX_N_1_847  /* hi z1 + MULTIPLY(tmp12, XFIX_N_1_847) */

    rshrn           v18.4h, v18.4s, #DESCALE_P2
    rshrn           v22.4h, v22.4s, #DESCALE_P2
    rshrn2          v18.8h, v24.4s, #DESCALE_P2  /* dataptr[2] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp13, XFIX_P_0_765), CONST_BITS-PASS1_BITS); */
    rshrn2          v22.8h, v25.4s, #DESCALE_P2  /* dataptr[6] = (DCTELEM) DESCALE(z1 + MULTIPLY(tmp12, XFIX_N_1_847), CONST_BITS-PASS1_BITS); */

    /* Odd part */
    add             v8.8h, v28.8h, v31.8h   /* z1 = tmp4 + tmp7; */
    add             v9.8h, v29.8h, v30.8h   /* z2 = tmp5 + tmp6; */
    add             v10.8h, v28.8h, v30.8h  /* z3 = tmp4 + tmp6; */
    add             v11.8h, v29.8h, v31.8h  /* z4 = tmp5 + tmp7; */

    smull           v4.4s, v10.4h, XFIX_P_1_175  /* z5 lo = z3 lo * XFIX_P_1_175 */
    smull2          v5.4s, v10.8h, XFIX_P_1_175
    smlal           v4.4s, v11.4h, XFIX_P_1_175  /* z5 = MULTIPLY(z3 + z4, FIX_1_175875602); */
    smlal2          v5.4s, v11.8h, XFIX_P_1_175

    smull2          v24.4s, v28.8h, XFIX_P_0_298
    smull2          v25.4s, v29.8h, XFIX_P_2_053
    smull2          v26.4s, v30.8h, XFIX_P_3_072
    smull2          v27.4s, v31.8h, XFIX_P_1_501
    smull           v28.4s, v28.4h, XFIX_P_0_298  /* tmp4 = MULTIPLY(tmp4, FIX_0_298631336); */
    smull           v29.4s, v29.4h, XFIX_P_2_053  /* tmp5 = MULTIPLY(tmp5, FIX_2_053119869); */
    smull           v30.4s, v30.4h, XFIX_P_3_072  /* tmp6 = MULTIPLY(tmp6, FIX_3_072711026); */
    smull           v31.4s, v31.4h, XFIX_P_1_501  /* tmp7 = MULTIPLY(tmp7, FIX_1_501321110); */

    smull2          v12.4s, v8.8h, XFIX_N_0_899
    smull2          v13.4s, v9.8h, XFIX_N_2_562
    smull2          v14.4s, v10.8h, XFIX_N_1_961
    smull2          v15.4s, v11.8h, XFIX_N_0_390
    smull           v8.4s, v8.4h, XFIX_N_0_899    /* z1 = MULTIPLY(z1, - FIX_0_899976223); */
    smull           v9.4s, v9.4h, XFIX_N_2_562    /* z2 = MULTIPLY(z2, - FIX_2_562915447); */
    smull           v10.4s, v10.4h, XFIX_N_1_961  /* z3 = MULTIPLY(z3, - FIX_1_961570560); */
    smull           v11.4s, v11.4h, XFIX_N_0_390  /* z4 = MULTIPLY(z4, - FIX_0_390180644); */

    add             v10.4s, v10.4s, v4.4s
    add             v14.4s, v14.4s, v5.4s
    add             v11.4s, v11.4s, v4.4s
    add             v15.4s, v15.4s, v5.4s

    add             v28.4s, v28.4s, v8.4s   /* tmp4 += z1 */
    add             v24.4s, v24.4s, v12.4s
    add             v29.4s, v29.4s, v9.4s   /* tmp5 += z2 */
    add             v25.4s, v25.4s, v13.4s
    add             v30.4s, v30.4s, v10.4s  /* tmp6 += z3 */
    add             v26.4s, v26.4s, v14.4s
    add             v31.4s, v31.4s, v11.4s  /* tmp7 += z4 */
    add             v27.4s, v27.4s, v15.4s

    add             v28.4s, v28.4s, v10.4s  /* tmp4 += z3 */
    add             v24.4s, v24.4s, v14.4s
    add             v29.4s, v29.4s, v11.4s  /* tmp5 += z4 */
    add             v25.4s, v25.4s, v15.4s
    add             v30.4s, v30.4s, v9.4s   /* tmp6 += z2 */
    add             v26.4s, v26.4s, v13.4s
    add             v31.4s, v31.4s, v8.4s   /* tmp7 += z1 */
    add             v27.4s, v27.4s, v12.4s

    rshrn           v23.4h, v28.4s, #DESCALE_P2
    rshrn           v21.4h, v29.4s, #DESCALE_P2
    rshrn           v19.4h, v30.4s, #DESCALE_P2
    rshrn           v17.4h, v31.4s, #DESCALE_P2
    rshrn2          v23.8h, v24.4s, #DESCALE_P2  /* dataptr[7] = (DCTELEM) DESCALE(tmp4 + z1 + z3, CONST_BITS-PASS1_BITS); */
    rshrn2          v21.8h, v25.4s, #DESCALE_P2  /* dataptr[5] = (DCTELEM) DESCALE(tmp5 + z2 + z4, CONST_BITS-PASS1_BITS); */
    rshrn2          v19.8h, v26.4s, #DESCALE_P2  /* dataptr[3] = (DCTELEM) DESCALE(tmp6 + z2 + z3, CONST_BITS-PASS1_BITS); */
    rshrn2          v17.8h, v27.4s, #DESCALE_P2  /* dataptr[1] = (DCTELEM) DESCALE(tmp7 + z1 + z4, CONST_BITS-PASS1_BITS); */

    /* store results */
    st1             {v16.8h, v17.8h, v18.8h, v19.8h}, [DATA], 64
    st1             {v20.8h, v21.8h, v22.8h, v23.8h}, [DATA]

    /* Restore NEON registers */
    sub             sp, sp, #64
    ld1             {v8.8b, v9.8b, v10.8b, v11.8b}, [sp], 32
    ld1             {v12.8b, v13.8b, v14.8b, v15.8b}, [sp], 32

    br              x30

    .unreq          DATA
    .unreq          TMP

#undef XFIX_P_0_298
#undef XFIX_N_0_390
#undef XFIX_P_0_541
#undef XFIX_P_0_765
#undef XFIX_N_0_899
#undef XFIX_P_1_175
#undef XFIX_P_1_501
#undef XFIX_N_1_847
#undef XFIX_N_1_961
#undef XFIX_P_2_053
#undef XFIX_N_2_562
#undef XFIX_P_3_072


/*****************************************************************************/

/*
 * jsimd_fdct_ifast_neon
 *
 * This function contains a fast, not so accurate integer implementation of
 * the forward DCT (Discrete Cosine Transform). It uses the same calculations
 * and produces exactly the same output as IJG's original 'jpeg_fdct_ifast'
 * function from jfdctfst.c
 *
 * TODO: can be combined with 'jsimd_convsamp_neon' to get
 *       rid of a bunch of VLD1.16 instructions
 */

#undef XFIX_0_541196100
#define XFIX_0_382683433 v0.h[0]
#define XFIX_0_541196100 v0.h[1]
#define XFIX_0_707106781 v0.h[2]
#define XFIX_1_306562965 v0.h[3]

.balign 16
Ljsimd_fdct_ifast_neon_consts:
  .short (98 * 128)               /* XFIX_0_382683433 */
  .short (139 * 128)              /* XFIX_0_541196100 */
  .short (181 * 128)              /* XFIX_0_707106781 */
  .short (334 * 128 - 256 * 128)  /* XFIX_1_306562965 */

asm_function jsimd_fdct_ifast_neon

    DATA            .req x0
    TMP             .req x9

    /* Load constants */
    adr             TMP, Ljsimd_fdct_ifast_neon_consts
    ld1             {v0.4h}, [TMP]

    /* Load all DATA into NEON registers with the following allocation:
     *       0 1 2 3 | 4 5 6 7
     *      ---------+--------
     *   0 | d16     | d17    | v0.8h
     *   1 | d18     | d19    | q9
     *   2 | d20     | d21    | q10
     *   3 | d22     | d23    | q11
     *   4 | d24     | d25    | q12
     *   5 | d26     | d27    | q13
     *   6 | d28     | d29    | q14
     *   7 | d30     | d31    | q15
     */

    ld1             {v16.8h, v17.8h, v18.8h, v19.8h}, [DATA], 64
    ld1             {v20.8h, v21.8h, v22.8h, v23.8h}, [DATA]
    mov             TMP, #2
    sub             DATA, DATA, #64
1:
    /* Transpose */
    transpose_8x8   v16, v17, v18, v19, v20, v21, v22, v23, v1, v2, v3, v4
    subs            TMP, TMP, #1
    /* 1-D FDCT */
    add             v4.8h, v19.8h, v20.8h
    sub             v20.8h, v19.8h, v20.8h
    sub             v28.8h, v18.8h, v21.8h
    add             v18.8h, v18.8h, v21.8h
    sub             v29.8h, v17.8h, v22.8h
    add             v17.8h, v17.8h, v22.8h
    sub             v21.8h, v16.8h, v23.8h
    add             v16.8h, v16.8h, v23.8h
    sub             v6.8h, v17.8h, v18.8h
    sub             v7.8h, v16.8h, v4.8h
    add             v5.8h, v17.8h, v18.8h
    add             v6.8h, v6.8h, v7.8h
    add             v4.8h, v16.8h, v4.8h
    sqdmulh         v6.8h, v6.8h, XFIX_0_707106781
    add             v19.8h, v20.8h, v28.8h
    add             v16.8h, v4.8h, v5.8h
    sub             v20.8h, v4.8h, v5.8h
    add             v5.8h, v28.8h, v29.8h
    add             v29.8h, v29.8h, v21.8h
    sqdmulh         v5.8h, v5.8h, XFIX_0_707106781
    sub             v28.8h, v19.8h, v29.8h
    add             v18.8h, v7.8h, v6.8h
    sqdmulh         v28.8h, v28.8h, XFIX_0_382683433
    sub             v22.8h, v7.8h, v6.8h
    sqdmulh         v19.8h, v19.8h, XFIX_0_541196100
    sqdmulh         v7.8h, v29.8h, XFIX_1_306562965
    add             v6.8h, v21.8h, v5.8h
    sub             v5.8h, v21.8h, v5.8h
    add             v29.8h, v29.8h, v28.8h
    add             v19.8h, v19.8h, v28.8h
    add             v29.8h, v29.8h, v7.8h
    add             v21.8h, v5.8h, v19.8h
    sub             v19.8h, v5.8h, v19.8h
    add             v17.8h, v6.8h, v29.8h
    sub             v23.8h, v6.8h, v29.8h

    b.ne            1b

    /* store results */
    st1             {v16.8h, v17.8h, v18.8h, v19.8h}, [DATA], 64
    st1             {v20.8h, v21.8h, v22.8h, v23.8h}, [DATA]

    br              x30

    .unreq          DATA
    .unreq          TMP
#undef XFIX_0_382683433
#undef XFIX_0_541196100
#undef XFIX_0_707106781
#undef XFIX_1_306562965


/*****************************************************************************/

/*
 * GLOBAL(void)
 * jsimd_quantize_neon (JCOEFPTR coef_block, DCTELEM * divisors,
 *                      DCTELEM * workspace);
 *
 */
asm_function jsimd_quantize_neon

    COEF_BLOCK      .req x0
    DIVISORS        .req x1
    WORKSPACE       .req x2

    RECIPROCAL      .req DIVISORS
    CORRECTION      .req x9
    SHIFT           .req x10
    LOOP_COUNT      .req x11

    mov             LOOP_COUNT, #2
    add             CORRECTION, DIVISORS, #(64 * 2)
    add             SHIFT, DIVISORS, #(64 * 6)
1:
    subs            LOOP_COUNT, LOOP_COUNT, #1
    ld1             {v0.8h, v1.8h, v2.8h, v3.8h}, [WORKSPACE], 64
    ld1             {v4.8h, v5.8h, v6.8h, v7.8h}, [CORRECTION], 64
    abs             v20.8h, v0.8h
    abs             v21.8h, v1.8h
    abs             v22.8h, v2.8h
    abs             v23.8h, v3.8h
    ld1             {v28.8h, v29.8h, v30.8h, v31.8h}, [RECIPROCAL], 64
    add             v20.8h, v20.8h, v4.8h  /* add correction */
    add             v21.8h, v21.8h, v5.8h
    add             v22.8h, v22.8h, v6.8h
    add             v23.8h, v23.8h, v7.8h
    umull           v4.4s, v20.4h, v28.4h  /* multiply by reciprocal */
    umull2          v16.4s, v20.8h, v28.8h
    umull           v5.4s, v21.4h, v29.4h
    umull2          v17.4s, v21.8h, v29.8h
    umull           v6.4s, v22.4h, v30.4h  /* multiply by reciprocal */
    umull2          v18.4s, v22.8h, v30.8h
    umull           v7.4s, v23.4h, v31.4h
    umull2          v19.4s, v23.8h, v31.8h
    ld1             {v24.8h, v25.8h, v26.8h, v27.8h}, [SHIFT], 64
    shrn            v4.4h, v4.4s, #16
    shrn            v5.4h, v5.4s, #16
    shrn            v6.4h, v6.4s, #16
    shrn            v7.4h, v7.4s, #16
    shrn2           v4.8h, v16.4s, #16
    shrn2           v5.8h, v17.4s, #16
    shrn2           v6.8h, v18.4s, #16
    shrn2           v7.8h, v19.4s, #16
    neg             v24.8h, v24.8h
    neg             v25.8h, v25.8h
    neg             v26.8h, v26.8h
    neg             v27.8h, v27.8h
    sshr            v0.8h, v0.8h, #15  /* extract sign */
    sshr            v1.8h, v1.8h, #15
    sshr            v2.8h, v2.8h, #15
    sshr            v3.8h, v3.8h, #15
    ushl            v4.8h, v4.8h, v24.8h  /* shift */
    ushl            v5.8h, v5.8h, v25.8h
    ushl            v6.8h, v6.8h, v26.8h
    ushl            v7.8h, v7.8h, v27.8h

    eor             v4.16b, v4.16b, v0.16b  /* restore sign */
    eor             v5.16b, v5.16b, v1.16b
    eor             v6.16b, v6.16b, v2.16b
    eor             v7.16b, v7.16b, v3.16b
    sub             v4.8h, v4.8h, v0.8h
    sub             v5.8h, v5.8h, v1.8h
    sub             v6.8h, v6.8h, v2.8h
    sub             v7.8h, v7.8h, v3.8h
    st1             {v4.8h, v5.8h, v6.8h, v7.8h}, [COEF_BLOCK], 64

    b.ne            1b

    br              x30  /* return */

    .unreq          COEF_BLOCK
    .unreq          DIVISORS
    .unreq          WORKSPACE
    .unreq          RECIPROCAL
    .unreq          CORRECTION
    .unreq          SHIFT
    .unreq          LOOP_COUNT


/*****************************************************************************/

/*
 * Downsample pixel values of a single component.
 * This version handles the common case of 2:1 horizontal and 1:1 vertical,
 * without smoothing.
 *
 * GLOBAL(void)
 * jsimd_h2v1_downsample_neon (JDIMENSION image_width, int max_v_samp_factor,
 *                             JDIMENSION v_samp_factor,
 *                             JDIMENSION width_blocks, JSAMPARRAY input_data,
 *                             JSAMPARRAY output_data);
 */

.balign 16
Ljsimd_h2_downsample_neon_consts:
  .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, \
        0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0F  /* diff 0 */
  .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, \
        0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0E, 0x0E  /* diff 1 */
  .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, \
        0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0D, 0x0D, 0x0D  /* diff 2 */
  .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, \
        0x08, 0x09, 0x0A, 0x0B, 0x0C, 0x0C, 0x0C, 0x0C  /* diff 3 */
  .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, \
        0x08, 0x09, 0x0A, 0x0B, 0x0B, 0x0B, 0x0B, 0x0B  /* diff 4 */
  .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, \
        0x08, 0x09, 0x0A, 0x0A, 0x0A, 0x0A, 0x0A, 0x0A  /* diff 5 */
  .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, \
        0x08, 0x09, 0x09, 0x09, 0x09, 0x09, 0x09, 0x09  /* diff 6 */
  .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, \
        0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08, 0x08  /* diff 7 */
  .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x07, \
        0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07, 0x07  /* diff 8 */
  .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x06, 0x06, \
        0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06, 0x06  /* diff 9 */
  .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x05, 0x05, 0x05, \
        0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05, 0x05  /* diff 10 */
  .byte 0x00, 0x01, 0x02, 0x03, 0x04, 0x04, 0x04, 0x04, \
        0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04, 0x04  /* diff 11 */
  .byte 0x00, 0x01, 0x02, 0x03, 0x03, 0x03, 0x03, 0x03, \
        0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03, 0x03  /* diff 12 */
  .byte 0x00, 0x01, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, \
        0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02, 0x02  /* diff 13 */
  .byte 0x00, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, \
        0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01, 0x01  /* diff 14 */
  .byte 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, \
        0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00  /* diff 15 */

asm_function jsimd_h2v1_downsample_neon
    IMAGE_WIDTH     .req x0
    MAX_V_SAMP      .req x1
    V_SAMP          .req x2
    BLOCK_WIDTH     .req x3
    INPUT_DATA      .req x4
    OUTPUT_DATA     .req x5
    OUTPTR          .req x9
    INPTR           .req x10
    TMP1            .req x11
    TMP2            .req x12
    TMP3            .req x13
    TMPDUP          .req w15

    mov             TMPDUP, #0x10000
    lsl             TMP2, BLOCK_WIDTH, #4
    sub             TMP2, TMP2, IMAGE_WIDTH
    adr             TMP3, Ljsimd_h2_downsample_neon_consts
    add             TMP3, TMP3, TMP2, lsl #4
    dup             v16.4s, TMPDUP
    ld1             {v18.16b}, [TMP3]

1:  /* row loop */
    ldr             INPTR, [INPUT_DATA], #8
    ldr             OUTPTR, [OUTPUT_DATA], #8
    subs            TMP1, BLOCK_WIDTH, #1
    b.eq            3f
2:  /* columns */
    ld1             {v0.16b}, [INPTR], #16
    mov             v4.16b, v16.16b
    subs            TMP1, TMP1, #1
    uadalp          v4.8h, v0.16b
    shrn            v6.8b, v4.8h, #1
    st1             {v6.8b}, [OUTPTR], #8
    b.ne            2b
3:  /* last columns */
    ld1             {v0.16b}, [INPTR]
    mov             v4.16b, v16.16b
    subs            V_SAMP, V_SAMP, #1
    /* expand right */
    tbl             v2.16b, {v0.16b}, v18.16b
    uadalp          v4.8h, v2.16b
    shrn            v6.8b, v4.8h, #1
    st1             {v6.8b}, [OUTPTR], #8
    b.ne            1b

    br              x30

    .unreq          IMAGE_WIDTH
    .unreq          MAX_V_SAMP
    .unreq          V_SAMP
    .unreq          BLOCK_WIDTH
    .unreq          INPUT_DATA
    .unreq          OUTPUT_DATA
    .unreq          OUTPTR
    .unreq          INPTR
    .unreq          TMP1
    .unreq          TMP2
    .unreq          TMP3
    .unreq          TMPDUP


/*****************************************************************************/

/*
 * Downsample pixel values of a single component.
 * This version handles the common case of 2:1 horizontal and 2:1 vertical,
 * without smoothing.
 *
 * GLOBAL(void)
 * jsimd_h2v2_downsample_neon (JDIMENSION image_width, int max_v_samp_factor,
 *                             JDIMENSION v_samp_factor, JDIMENSION width_blocks,
 *                             JSAMPARRAY input_data, JSAMPARRAY output_data);
 */

.balign 16
asm_function jsimd_h2v2_downsample_neon
    IMAGE_WIDTH     .req x0
    MAX_V_SAMP      .req x1
    V_SAMP          .req x2
    BLOCK_WIDTH     .req x3
    INPUT_DATA      .req x4
    OUTPUT_DATA     .req x5
    OUTPTR          .req x9
    INPTR0          .req x10
    INPTR1          .req x14
    TMP1            .req x11
    TMP2            .req x12
    TMP3            .req x13
    TMPDUP          .req w15

    mov             TMPDUP, #1
    lsl             TMP2, BLOCK_WIDTH, #4
    lsl             TMPDUP, TMPDUP, #17
    sub             TMP2, TMP2, IMAGE_WIDTH
    adr             TMP3, Ljsimd_h2_downsample_neon_consts
    orr             TMPDUP, TMPDUP, #1
    add             TMP3, TMP3, TMP2, lsl #4
    dup             v16.4s, TMPDUP
    ld1             {v18.16b}, [TMP3]

1:  /* row loop */
    ldr             INPTR0, [INPUT_DATA], #8
    ldr             OUTPTR, [OUTPUT_DATA], #8
    ldr             INPTR1, [INPUT_DATA], #8
    subs            TMP1, BLOCK_WIDTH, #1
    b.eq            3f
2:  /* columns */
    ld1             {v0.16b}, [INPTR0], #16
    ld1             {v1.16b}, [INPTR1], #16
    mov             v4.16b, v16.16b
    subs            TMP1, TMP1, #1
    uadalp          v4.8h, v0.16b
    uadalp          v4.8h, v1.16b
    shrn            v6.8b, v4.8h, #2
    st1             {v6.8b}, [OUTPTR], #8
    b.ne            2b
3:  /* last columns */
    ld1             {v0.16b}, [INPTR0], #16
    ld1             {v1.16b}, [INPTR1], #16
    mov             v4.16b, v16.16b
    subs            V_SAMP, V_SAMP, #1
    /* expand right */
    tbl             v2.16b, {v0.16b}, v18.16b
    tbl             v3.16b, {v1.16b}, v18.16b
    uadalp          v4.8h, v2.16b
    uadalp          v4.8h, v3.16b
    shrn            v6.8b, v4.8h, #2
    st1             {v6.8b}, [OUTPTR], #8
    b.ne            1b

    br              x30

    .unreq          IMAGE_WIDTH
    .unreq          MAX_V_SAMP
    .unreq          V_SAMP
    .unreq          BLOCK_WIDTH
    .unreq          INPUT_DATA
    .unreq          OUTPUT_DATA
    .unreq          OUTPTR
    .unreq          INPTR0
    .unreq          INPTR1
    .unreq          TMP1
    .unreq          TMP2
    .unreq          TMP3
    .unreq          TMPDUP