Quelle  row_sve.h   Sprache: C

/*
 *  Copyright 2024 The LibYuv Project Authors. All rights reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS. All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#ifndef INCLUDE_LIBYUV_ROW_SVE_H_
#define INCLUDE_LIBYUV_ROW_SVE_H_

#include "libyuv/row.h"

#ifdef __cplusplus
namespace libyuv {
extern "C" {
#endif

#if !defined(LIBYUV_DISABLE_SVE) && defined(__aarch64__)

#if !defined(LIBYUV_DISABLE_SME) && defined(CLANG_HAS_SME) && \
    defined(__aarch64__)
#define STREAMING_COMPATIBLE __arm_streaming_compatible
#else  // defined(LIBYUV_DISABLE_SME) || !defined(CLANG_HAS_SME) ||
       // !defined(__aarch64__)
#define STREAMING_COMPATIBLE
#endif  // !defined(LIBYUV_DISABLE_SME) && defined(CLANG_HAS_SME) &&
        // defined(__aarch64__)

#define YUVTORGB_SVE_SETUP                          \
  "ld1rb {z28.b}, p0/z, [%[kUVCoeff], #0] \n" \
  "ld1rb {z29.b}, p0/z, [%[kUVCoeff], #1] \n" \
  "ld1rb {z30.b}, p0/z, [%[kUVCoeff], #2] \n" \
  "ld1rb {z31.b}, p0/z, [%[kUVCoeff], #3] \n" \
  "ld1rh {z24.h}, p0/z, [%[kRGBCoeffBias], #0] \n" \
  "ld1rh {z25.h}, p0/z, [%[kRGBCoeffBias], #2] \n" \
  "ld1rh {z26.h}, p0/z, [%[kRGBCoeffBias], #4] \n" \
  "ld1rh {z27.h}, p0/z, [%[kRGBCoeffBias], #6] \n"

#define READYUV444_SVE                           \
  "ld1b {z0.h}, p1/z, [%[src_y]] \n" \
  "ld1b {z1.h}, p1/z, [%[src_u]] \n" \
  "ld1b {z2.h}, p1/z, [%[src_v]] \n" \
  "add %[src_y], %[src_y], %[vl] \n" \
  "add %[src_u], %[src_u], %[vl] \n" \
  "add %[src_v], %[src_v], %[vl] \n" \
  "prfm pldl1keep, [%[src_y], 448] \n" \
  "prfm pldl1keep, [%[src_u], 448] \n" \
  "trn1 z0.b, z0.b, z0.b \n" \
  "prfm pldl1keep, [%[src_v], 448] \n"

#define READYUV400_SVE                           \
  "ld1b {z0.h}, p1/z, [%[src_y]] \n" \
  "inch %[src_y] \n" \
  "prfm pldl1keep, [%[src_y], 448] \n" \
  "trn1 z0.b, z0.b, z0.b \n"

#define READYUV422_SVE                           \
  "ld1b {z0.h}, p1/z, [%[src_y]] \n" \
  "ld1b {z1.s}, p1/z, [%[src_u]] \n" \
  "ld1b {z2.s}, p1/z, [%[src_v]] \n" \
  "inch %[src_y] \n" \
  "incw %[src_u] \n" \
  "incw %[src_v] \n" \
  "prfm pldl1keep, [%[src_y], 448] \n" \
  "prfm pldl1keep, [%[src_u], 128] \n" \
  "prfm pldl1keep, [%[src_v], 128] \n" \
  "trn1 z0.b, z0.b, z0.b \n" \
  "trn1 z1.h, z1.h, z1.h \n" \
  "trn1 z2.h, z2.h, z2.h \n"

// Read twice as much data from YUV, putting the even elements from the Y data
// in z0.h and odd elements in z1.h. U/V data is not duplicated, stored in
// z2.h/z3.h.
#define READYUV422_SVE_2X                        \
  "ld1b {z0.b}, p1/z, [%[src_y]] \n" \
  "ld1b {z2.h}, p1/z, [%[src_u]] \n" \
  "ld1b {z3.h}, p1/z, [%[src_v]] \n" \
  "incb %[src_y] \n" \
  "inch %[src_u] \n" \
  "inch %[src_v] \n" \
  "prfm pldl1keep, [%[src_y], 448] \n" \
  "prfm pldl1keep, [%[src_u], 128] \n" \
  "prfm pldl1keep, [%[src_v], 128] \n" \
  "trn2 z1.b, z0.b, z0.b \n" \
  "trn1 z0.b, z0.b, z0.b \n"

#define READI210_SVE                             \
  "ld1h {z3.h}, p1/z, [%[src_y]] \n" \
  "ld1h {z1.s}, p1/z, [%[src_u]] \n" \
  "ld1h {z2.s}, p1/z, [%[src_v]] \n" \
  "incb %[src_y] \n" \
  "inch %[src_u] \n" \
  "inch %[src_v] \n" \
  "lsl z0.h, z3.h, #6 \n" \
  "trn1 z1.h, z1.h, z1.h \n" \
  "trn1 z2.h, z2.h, z2.h \n" \
  "prfm pldl1keep, [%[src_y], 448] \n" \
  "prfm pldl1keep, [%[src_u], 128] \n" \
  "prfm pldl1keep, [%[src_v], 128] \n" \
  "usra z0.h, z3.h, #4 \n" \
  "uqshrnb z1.b, z1.h, #2 \n" \
  "uqshrnb z2.b, z2.h, #2 \n"

#define READP210_SVE                             \
  "ld1h {z0.h}, p1/z, [%[src_y]] \n" \
  "ld1h {z1.h}, p2/z, [%[src_uv]] \n" \
  "incb %[src_y] \n" \
  "incb %[src_uv] \n" \
  "prfm pldl1keep, [%[src_y], 448] \n" \
  "prfm pldl1keep, [%[src_uv], 256] \n" \
  "tbl z1.b, {z1.b}, z22.b \n"

#define READI410_SVE                             \
  "ld1h {z3.h}, p1/z, [%[src_y]] \n" \
  "lsl z0.h, z3.h, #6 \n" \
  "usra z0.h, z3.h, #4 \n" \
  "ld1h {z1.h}, p1/z, [%[src_u]] \n" \
  "ld1h {z2.h}, p1/z, [%[src_v]] \n" \
  "incb %[src_y] \n" \
  "incb %[src_u] \n" \
  "incb %[src_v] \n" \
  "prfm pldl1keep, [%[src_y], 448] \n" \
  "prfm pldl1keep, [%[src_u], 128] \n" \
  "prfm pldl1keep, [%[src_v], 128] \n" \
  "uqshrnb z1.b, z1.h, #2 \n" \
  "uqshrnb z2.b, z2.h, #2 \n"

// We need different predicates for the UV components since we are reading
// 32-bit (pairs of UV) elements rather than 16-bit Y elements.
#define READP410_SVE                                    \
  "ld1h {z0.h}, p1/z, [%[src_y]] \n" \
  "ld1w {z1.s}, p2/z, [%[src_uv]] \n" \
  "ld1w {z2.s}, p3/z, [%[src_uv], #1, mul vl] \n" \
  "incb %[src_y] \n" \
  "incb %[src_uv], all, mul #2 \n" \
  "prfm pldl1keep, [%[src_y], 448] \n" \
  "prfm pldl1keep, [%[src_uv], 256] \n" \
  "uzp2 z1.b, z1.b, z2.b \n"

#define READI212_SVE                             \
  "ld1h {z3.h}, p1/z, [%[src_y]] \n" \
  "ld1h {z1.s}, p1/z, [%[src_u]] \n" \
  "ld1h {z2.s}, p1/z, [%[src_v]] \n" \
  "incb %[src_y] \n" \
  "inch %[src_u] \n" \
  "inch %[src_v] \n" \
  "lsl z0.h, z3.h, #4 \n" \
  "trn1 z1.h, z1.h, z1.h \n" \
  "trn1 z2.h, z2.h, z2.h \n" \
  "prfm pldl1keep, [%[src_y], 448] \n" \
  "prfm pldl1keep, [%[src_u], 128] \n" \
  "prfm pldl1keep, [%[src_v], 128] \n" \
  "usra z0.h, z3.h, #8 \n" \
  "uqshrnb z1.b, z1.h, #4 \n" \
  "uqshrnb z2.b, z2.h, #4 \n"

#define I400TORGB_SVE                                    \
  "umulh z18.h, z24.h, z0.h \n" /* Y */ \
  "movprfx z16, z18 \n"         \
  "usqadd z16.h, p0/m, z16.h, z4.h \n" /* B */ \
  "movprfx z17, z18 \n"         \
  "usqadd z17.h, p0/m, z17.h, z6.h \n" /* G */ \
  "usqadd z18.h, p0/m, z18.h, z5.h \n" /* R */

// We need a different predicate for the UV component to handle the tail.
// If there is a single element remaining then we want to load one Y element
// but two UV elements.
#define READNV_SVE_2X                                               \
  "ld1b {z0.b}, p1/z, [%[src_y]] \n" /* Y0Y0 */         \
  "ld1b {z2.b}, p2/z, [%[src_uv]] \n" /* U0V0 or V0U0 */ \
  "incb %[src_y] \n"                    \
  "incb %[src_uv] \n"                    \
  "prfm pldl1keep, [%[src_y], 448] \n"                    \
  "prfm pldl1keep, [%[src_uv], 256] \n"                    \
  "trn2 z1.b, z0.b, z0.b \n" /* YYYY */         \
  "trn1 z0.b, z0.b, z0.b \n" /* YYYY */

// Like NVTORGB_SVE but U/V components are stored in widened .h elements of
// z1/z2 rather than even/odd .b lanes of z1.
#define I4XXTORGB_SVE                                     \
  "umulh z0.h, z24.h, z0.h \n" /* Y */  \
  "umullb z6.h, z30.b, z1.b \n"          \
  "umullb z4.h, z28.b, z1.b \n" /* DB */ \
  "umullb z5.h, z29.b, z2.b \n" /* DR */ \
  "umlalb z6.h, z31.b, z2.b \n" /* DG */ \
  "add z17.h, z0.h, z26.h \n" /* G */  \
  "add z16.h, z0.h, z4.h \n" /* B */  \
  "add z18.h, z0.h, z5.h \n" /* R */  \
  "uqsub z17.h, z17.h, z6.h \n" /* G */  \
  "uqsub z16.h, z16.h, z25.h \n" /* B */  \
  "uqsub z18.h, z18.h, z27.h \n" /* R */

// Like I4XXTORGB_SVE but U/V components are stored in even/odd .b lanes of z1
// rather than widened .h elements of z1/z2.
#define NVTORGB_SVE                                       \
  "umulh z0.h, z24.h, z0.h \n" /* Y */  \
  "umullb z6.h, z30.b, z1.b \n"          \
  "umullb z4.h, z28.b, z1.b \n" /* DB */ \
  "umullt z5.h, z29.b, z1.b \n" /* DR */ \
  "umlalt z6.h, z31.b, z1.b \n" /* DG */ \
  "add z17.h, z0.h, z26.h \n" /* G */  \
  "add z16.h, z0.h, z4.h \n" /* B */  \
  "add z18.h, z0.h, z5.h \n" /* R */  \
  "uqsub z17.h, z17.h, z6.h \n" /* G */  \
  "uqsub z16.h, z16.h, z25.h \n" /* B */  \
  "uqsub z18.h, z18.h, z27.h \n" /* R */

// The U/V component multiplies do not need to be duplicated in I422, we just
// need to combine them with Y0/Y1 correctly.
#define I422TORGB_SVE_2X                                  \
  "umulh z0.h, z24.h, z0.h \n" /* Y0 */ \
  "umulh z1.h, z24.h, z1.h \n" /* Y1 */ \
  "umullb z6.h, z30.b, z2.b \n"          \
  "umullb z4.h, z28.b, z2.b \n" /* DB */ \
  "umullb z5.h, z29.b, z3.b \n" /* DR */ \
  "umlalb z6.h, z31.b, z3.b \n" /* DG */ \
                                                          \
  "add z17.h, z0.h, z26.h \n" /* G0 */ \
  "add z21.h, z1.h, z26.h \n" /* G1 */ \
  "add z16.h, z0.h, z4.h \n" /* B0 */ \
  "add z20.h, z1.h, z4.h \n" /* B1 */ \
  "add z18.h, z0.h, z5.h \n" /* R0 */ \
  "add z22.h, z1.h, z5.h \n" /* R1 */ \
  "uqsub z17.h, z17.h, z6.h \n" /* G0 */ \
  "uqsub z21.h, z21.h, z6.h \n" /* G1 */ \
  "uqsub z16.h, z16.h, z25.h \n" /* B0 */ \
  "uqsub z20.h, z20.h, z25.h \n" /* B1 */ \
  "uqsub z18.h, z18.h, z27.h \n" /* R0 */ \
  "uqsub z22.h, z22.h, z27.h \n" /* R1 */

// clang-format off
#define NVTORGB_SVE_2X(bt_u, bt_v)                        \
  "umulh z0.h, z24.h, z0.h \n" /* Y0 */ \
  "umulh z1.h, z24.h, z1.h \n" /* Y1 */ \
  "umull" #bt_u " z6.h, z30.b, z2.b \n"          \
  "umull" #bt_u " z4.h, z28.b, z2.b \n" /* DB */ \
  "umull" #bt_v " z5.h, z29.b, z2.b \n" /* DR */ \
  "umlal" #bt_v " z6.h, z31.b, z2.b \n" /* DG */ \
                                                          \
  "add z17.h, z0.h, z26.h \n" /* G0 */ \
  "add z21.h, z1.h, z26.h \n" /* G1 */ \
  "add z16.h, z0.h, z4.h \n" /* B0 */ \
  "add z20.h, z1.h, z4.h \n" /* B1 */ \
  "add z18.h, z0.h, z5.h \n" /* R0 */ \
  "add z22.h, z1.h, z5.h \n" /* R1 */ \
  "uqsub z17.h, z17.h, z6.h \n" /* G0 */ \
  "uqsub z21.h, z21.h, z6.h \n" /* G1 */ \
  "uqsub z16.h, z16.h, z25.h \n" /* B0 */ \
  "uqsub z20.h, z20.h, z25.h \n" /* B1 */ \
  "uqsub z18.h, z18.h, z27.h \n" /* R0 */ \
  "uqsub z22.h, z22.h, z27.h \n" /* R1 */
// clang-format on

#define RGBTOARGB8_SVE_TOP_2X                        \
  /* Inputs: B: z16.h,  G: z17.h,  R: z18.h */       \
  "uqshl z16.h, p0/m, z16.h, #2 \n" /* B0 */ \
  "uqshl z17.h, p0/m, z17.h, #2 \n" /* G0 */ \
  "uqshl z18.h, p0/m, z18.h, #2 \n" /* R0 */ \
  "uqshl z20.h, p0/m, z20.h, #2 \n" /* B1 */ \
  "uqshl z21.h, p0/m, z21.h, #2 \n" /* G1 */ \
  "uqshl z22.h, p0/m, z22.h, #2 \n" /* R1 */

// Convert from 2.14 fixed point RGB to 8 bit ARGB, interleaving as BG and RA
// pairs to allow us to use ST2 for storing rather than ST4.
#define RGBTOARGB8_SVE                                    \
  /* Inputs: B: z16.h,  G: z17.h,  R: z18.h,  A: z19.b */ \
  "uqshrnb z16.b, z16.h, #6 \n" /* B0 */          \
  "uqshrnb z18.b, z18.h, #6 \n" /* R0 */          \
  "uqshrnt z16.b, z17.h, #6 \n" /* BG */          \
  "trn1 z17.b, z18.b, z19.b \n" /* RA */

// Convert from 2.14 fixed point RGBA to 8 bit ARGB, interleaving as BG and RA
// pairs to allow us to use ST2 for storing rather than ST4.
#define RGBATOARGB8_SVE                                   \
  /* Inputs: B: z16.h,  G: z17.h,  R: z18.h,  A: z19.h */ \
  "uqshrnb z16.b, z16.h, #6 \n" /* B0 */          \
  "uqshrnt z16.b, z17.h, #6 \n" /* BG */          \
  "uqshrnb z17.b, z18.h, #6 \n" /* R0 */          \
  "uqshrnt z17.b, z19.h, #2 \n" /* RA */

// Convert from 2.14 fixed point RGB to 8 bit RGBA, interleaving as AB and GR
// pairs to allow us to use ST2 for storing rather than ST4.
#define RGBTORGBA8_SVE                                    \
  /* Inputs: B: z16.h,  G: z17.h,  R: z18.h,  A: z19.b */ \
  "uqshrnt z19.b, z16.h, #6 \n" /* AB */          \
  "uqshrnb z20.b, z17.h, #6 \n" /* G0 */          \
  "uqshrnt z20.b, z18.h, #6 \n" /* GR */

#define RGBTOARGB8_SVE_2X                                 \
  /* Inputs: B: z16.h,  G: z17.h,  R: z18.h,  A: z19.b */ \
  "uqshrnb z16.b, z16.h, #6 \n" /* B0 */          \
  "uqshrnb z17.b, z17.h, #6 \n" /* G0 */          \
  "uqshrnb z18.b, z18.h, #6 \n" /* R0 */          \
  "uqshrnt z16.b, z20.h, #6 \n" /* B1 */          \
  "uqshrnt z17.b, z21.h, #6 \n" /* G1 */          \
  "uqshrnt z18.b, z22.h, #6 \n" /* R1 */

// Store AR30 elements. Inputs are 2.14 fixed point RGB. We expect z23 to be
// populated with 0x3ff0 (0x3fff would also work) to saturate the R input
// rather than needing a pair of shifts to saturate and then insert into the
// correct position in the lane.
#define STOREAR30_SVE                                                    \
  "uqshl z16.h, p0/m, z16.h, #2 \n" /* bbbbbbbbbbxxxxxx */ \
  "uqshl z17.h, p0/m, z17.h, #2 \n" /* ggggggggggxxxxxx */ \
  "umin z18.h, p0/m, z18.h, z23.h \n" /* 00rrrrrrrrrrxxxx */ \
  "orr z18.h, z18.h, #0xc000 \n" /* 11rrrrrrrrrrxxxx */ \
  "sri z18.h, z17.h, #12 \n" /* 11rrrrrrrrrrgggg */ \
  "lsl z17.h, z17.h, #4 \n" /* ggggggxxxxxx0000 */ \
  "sri z17.h, z16.h, #6 \n" /* ggggggbbbbbbbbbb */ \
  "st2h {z17.h, z18.h}, p1, [%[dst_ar30]] \n"                        \
  "incb %[dst_ar30], all, mul #2 \n"

#define YUVTORGB_SVE_REGS                                                     \
  "z0", "z1", "z2", "z3", "z4", "z5", "z6", "z7", "z16", "z17", "z18", "z19", \
      "z20", "z22", "z23", "z24", "z25", "z26", "z27", "z28", "z29", "z30",   \
      "z31", "p0", "p1", "p2", "p3"

static inline void I400ToARGBRow_SVE_SC(const uint8_t* src_y,
                                        uint8_t* dst_argb,
                                        const struct YuvConstants* yuvconstants,
                                        int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm volatile(
      "cnth %[vl] \n"
      "ptrue p0.b \n"
      "dup z19.b, #255 \n"  // Alpha
      YUVTORGB_SVE_SETUP
      "cmp %w[width], %w[vl] \n"
      "mov z1.h, #128 \n"  // U/V
      "umullb z6.h, z30.b, z1.b \n"
      "umullb z4.h, z28.b, z1.b \n"  // DB
      "umullb z5.h, z29.b, z1.b \n"  // DR
      "mla z6.h, p0/m, z31.h, z1.h \n"  // DG
      "sub z4.h, z4.h, z25.h \n"
      "sub z5.h, z5.h, z27.h \n"
      "sub z6.h, z26.h, z6.h \n"
      "b.le 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.h \n"
      "sub %w[width], %w[width], %w[vl] \n"
      "1: \n"  //
      READYUV400_SVE I400TORGB_SVE RGBTOARGB8_SVE
      "subs %w[width], %w[width], %w[vl] \n"
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"
      "add %[dst_argb], %[dst_argb], %[vl], lsl #2 \n"
      "b.gt 1b \n"
      "add %w[width], %w[width], %w[vl] \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "2: \n"
      "whilelt p1.h, wzr, %w[width] \n"  //
      READYUV400_SVE I400TORGB_SVE RGBTOARGB8_SVE
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"
      : [src_y] "+r"(src_y),                               // %[src_y]
        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
        [width] "+r"(width),                               // %[width]
        [vl] "=&r"(vl)                                     // %[vl]
      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
      : "cc", "memory", YUVTORGB_SVE_REGS);
}

static inline void I422ToARGBRow_SVE_SC(const uint8_t* src_y,
                                        const uint8_t* src_u,
                                        const uint8_t* src_v,
                                        uint8_t* dst_argb,
                                        const struct YuvConstants* yuvconstants,
                                        int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm volatile(
      "cntb %[vl] \n"
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "dup z19.b, #255 \n"  // Alpha
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.b \n"
      "1: \n"  //
      READYUV422_SVE_2X I422TORGB_SVE_2X RGBTOARGB8_SVE_2X
      "subs %w[width], %w[width], %w[vl] \n"
      "st4b {z16.b, z17.b, z18.b, z19.b}, p1, [%[dst_argb]] \n"
      "incb %[dst_argb], all, mul #4 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "cnth %[vl] \n"
      "whilelt p1.b, wzr, %w[width] \n"  //
      READYUV422_SVE_2X I422TORGB_SVE_2X RGBTOARGB8_SVE_2X
      "st4b {z16.b, z17.b, z18.b, z19.b}, p1, [%[dst_argb]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                               // %[src_y]
        [src_u] "+r"(src_u),                               // %[src_u]
        [src_v] "+r"(src_v),                               // %[src_v]
        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
        [width] "+r"(width),                               // %[width]
        [vl] "=&r"(vl)                                     // %[vl]
      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
      : "cc", "memory", YUVTORGB_SVE_REGS);
}

static inline void I422ToRGB24Row_SVE_SC(
    const uint8_t* src_y,
    const uint8_t* src_u,
    const uint8_t* src_v,
    uint8_t* dst_argb,
    const struct YuvConstants* yuvconstants,
    int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm volatile(
      "cntb %[vl] \n"
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.b \n"
      "1: \n"  //
      READYUV422_SVE_2X I422TORGB_SVE_2X RGBTOARGB8_SVE_2X
      "subs %w[width], %w[width], %w[vl] \n"
      "st3b {z16.b, z17.b, z18.b}, p1, [%[dst_argb]] \n"
      "incb %[dst_argb], all, mul #3 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "cnth %[vl] \n"
      "whilelt p1.b, wzr, %w[width] \n"  //
      READYUV422_SVE_2X I422TORGB_SVE_2X RGBTOARGB8_SVE_2X
      "st3b {z16.b, z17.b, z18.b}, p1, [%[dst_argb]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                               // %[src_y]
        [src_u] "+r"(src_u),                               // %[src_u]
        [src_v] "+r"(src_v),                               // %[src_v]
        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
        [width] "+r"(width),                               // %[width]
        [vl] "=&r"(vl)                                     // %[vl]
      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
      : "cc", "memory", YUVTORGB_SVE_REGS);
}

#define RGB8TORGB565_SVE_FROM_TOP_2X                        \
  "sri z18.h, z17.h, #5 \n" /* rrrrrgggggg00000 */ \
  "sri z22.h, z21.h, #5 \n" /* rrrrrgggggg00000 */ \
  "sri z18.h, z16.h, #11 \n" /* rrrrrggggggbbbbb */ \
  "sri z22.h, z20.h, #11 \n" /* rrrrrggggggbbbbb */ \
  "mov z19.d, z22.d \n"

static inline void I422ToRGB565Row_SVE_SC(
    const uint8_t* src_y,
    const uint8_t* src_u,
    const uint8_t* src_v,
    uint8_t* dst_rgb565,
    const struct YuvConstants* yuvconstants,
    int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm volatile(
      "cntb %[vl] \n"
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.b \n"
      "1: \n"  //
      READYUV422_SVE_2X I422TORGB_SVE_2X RGBTOARGB8_SVE_TOP_2X
      "subs %w[width], %w[width], %w[vl] \n"  //
      RGB8TORGB565_SVE_FROM_TOP_2X
      "st2h {z18.h, z19.h}, p1, [%[dst]] \n"
      "incb %[dst], all, mul #2 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "cnth %[vl] \n"
      "whilelt p1.b, wzr, %w[width] \n"  //
      READYUV422_SVE_2X I422TORGB_SVE_2X RGBTOARGB8_SVE_TOP_2X
          RGB8TORGB565_SVE_FROM_TOP_2X
      "st2h {z18.h, z19.h}, p1, [%[dst]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                               // %[src_y]
        [src_u] "+r"(src_u),                               // %[src_u]
        [src_v] "+r"(src_v),                               // %[src_v]
        [dst] "+r"(dst_rgb565),                            // %[dst]
        [width] "+r"(width),                               // %[width]
        [vl] "=&r"(vl)                                     // %[vl]
      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
      : "cc", "memory", YUVTORGB_SVE_REGS);
}

#define RGB8TOARGB1555_SVE_FROM_TOP_2X                      \
  "dup z0.h, #0x8000 \n" /* 1000000000000000 */ \
  "dup z1.h, #0x8000 \n" /* 1000000000000000 */ \
  "sri z0.h, z18.h, #1 \n" /* 1rrrrrxxxxxxxxxx */ \
  "sri z1.h, z22.h, #1 \n" /* 1rrrrrxxxxxxxxxx */ \
  "sri z0.h, z17.h, #6 \n" /* 1rrrrrgggggxxxxx */ \
  "sri z1.h, z21.h, #6 \n" /* 1rrrrrgggggxxxxx */ \
  "sri z0.h, z16.h, #11 \n" /* 1rrrrrgggggbbbbb */ \
  "sri z1.h, z20.h, #11 \n" /* 1rrrrrgggggbbbbb */

static inline void I422ToARGB1555Row_SVE_SC(
    const uint8_t* src_y,
    const uint8_t* src_u,
    const uint8_t* src_v,
    uint8_t* dst_argb1555,
    const struct YuvConstants* yuvconstants,
    int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm volatile(
      "cntb %[vl] \n"
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.b \n"
      "1: \n"  //
      READYUV422_SVE_2X I422TORGB_SVE_2X RGBTOARGB8_SVE_TOP_2X
      "subs %w[width], %w[width], %w[vl] \n"  //
      RGB8TOARGB1555_SVE_FROM_TOP_2X
      "st2h {z0.h, z1.h}, p1, [%[dst]] \n"
      "incb %[dst], all, mul #2 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "cnth %[vl] \n"
      "whilelt p1.b, wzr, %w[width] \n"  //
      READYUV422_SVE_2X I422TORGB_SVE_2X RGBTOARGB8_SVE_TOP_2X
          RGB8TOARGB1555_SVE_FROM_TOP_2X
      "st2h {z0.h, z1.h}, p1, [%[dst]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                               // %[src_y]
        [src_u] "+r"(src_u),                               // %[src_u]
        [src_v] "+r"(src_v),                               // %[src_v]
        [dst] "+r"(dst_argb1555),                          // %[dst]
        [width] "+r"(width),                               // %[width]
        [vl] "=&r"(vl)                                     // %[vl]
      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
      : "cc", "memory", YUVTORGB_SVE_REGS);
}

#define RGB8TOARGB4444_SVE_FROM_TOP_2X                      \
  "dup z0.h, #0xf000 \n" /* 1111000000000000 */ \
  "dup z1.h, #0xf000 \n" /* 1111000000000000 */ \
  "sri z0.h, z18.h, #4 \n" /* 1111rrrrxxxxxxxx */ \
  "sri z1.h, z22.h, #4 \n" /* 1111rrrrxxxxxxxx */ \
  "sri z0.h, z17.h, #8 \n" /* 1111rrrrggggxxxx */ \
  "sri z1.h, z21.h, #8 \n" /* 1111rrrrggggxxxx */ \
  "sri z0.h, z16.h, #12 \n" /* 1111rrrrggggbbbb */ \
  "sri z1.h, z20.h, #12 \n" /* 1111rrrrggggbbbb */

static inline void I422ToARGB4444Row_SVE_SC(
    const uint8_t* src_y,
    const uint8_t* src_u,
    const uint8_t* src_v,
    uint8_t* dst_argb4444,
    const struct YuvConstants* yuvconstants,
    int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm volatile(
      "cntb %[vl] \n"
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.b \n"
      "1: \n"  //
      READYUV422_SVE_2X I422TORGB_SVE_2X RGBTOARGB8_SVE_TOP_2X
      "subs %w[width], %w[width], %w[vl] \n"  //
      RGB8TOARGB4444_SVE_FROM_TOP_2X
      "st2h {z0.h, z1.h}, p1, [%[dst]] \n"
      "incb %[dst], all, mul #2 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "cnth %[vl] \n"
      "whilelt p1.b, wzr, %w[width] \n"  //
      READYUV422_SVE_2X I422TORGB_SVE_2X RGBTOARGB8_SVE_TOP_2X
          RGB8TOARGB4444_SVE_FROM_TOP_2X
      "st2h {z0.h, z1.h}, p1, [%[dst]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                               // %[src_y]
        [src_u] "+r"(src_u),                               // %[src_u]
        [src_v] "+r"(src_v),                               // %[src_v]
        [dst] "+r"(dst_argb4444),                          // %[dst]
        [width] "+r"(width),                               // %[width]
        [vl] "=&r"(vl)                                     // %[vl]
      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
      : "cc", "memory", YUVTORGB_SVE_REGS);
}

static inline void I422ToRGBARow_SVE_SC(const uint8_t* src_y,
                                        const uint8_t* src_u,
                                        const uint8_t* src_v,
                                        uint8_t* dst_argb,
                                        const struct YuvConstants* yuvconstants,
                                        int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm volatile(
      "cnth %[vl] \n"
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "dup z19.b, #255 \n"  // Alpha
      "subs %w[width], %w[width], %w[vl] \n"
      "b.le 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.h \n"
      "1: \n"  //
      READYUV422_SVE I4XXTORGB_SVE RGBTORGBA8_SVE
      "subs %w[width], %w[width], %w[vl] \n"
      "st2h {z19.h, z20.h}, p1, [%[dst_argb]] \n"
      "add %[dst_argb], %[dst_argb], %[vl], lsl #2 \n"
      "b.gt 1b \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      "whilelt p1.h, wzr, %w[width] \n"  //
      READYUV422_SVE I4XXTORGB_SVE RGBTORGBA8_SVE
      "st2h {z19.h, z20.h}, p1, [%[dst_argb]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                               // %[src_y]
        [src_u] "+r"(src_u),                               // %[src_u]
        [src_v] "+r"(src_v),                               // %[src_v]
        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
        [width] "+r"(width),                               // %[width]
        [vl] "=&r"(vl)                                     // %[vl]
      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
      : "cc", "memory", YUVTORGB_SVE_REGS);
}

static inline void I422AlphaToARGBRow_SVE_SC(
    const uint8_t* src_y,
    const uint8_t* src_u,
    const uint8_t* src_v,
    const uint8_t* src_a,
    uint8_t* dst_argb,
    const struct YuvConstants* yuvconstants,
    int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm volatile(
      "cntb %[vl] \n"
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.b \n"
      "1: \n"  //
      READYUV422_SVE_2X
      "ld1b {z19.b}, p1/z, [%[src_a]] \n"
      "add %[src_a], %[src_a], %[vl] \n"  // Alpha
      I422TORGB_SVE_2X RGBTOARGB8_SVE_2X
      "subs %w[width], %w[width], %w[vl] \n"
      "st4b {z16.b, z17.b, z18.b, z19.b}, p1, [%[dst_argb]] \n"
      "incb %[dst_argb], all, mul #4 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "cnth %[vl] \n"
      "whilelt p1.b, wzr, %w[width] \n"  //
      READYUV422_SVE_2X
      "ld1b {z19.b}, p1/z, [%[src_a]] \n"  // Alpha
      I422TORGB_SVE_2X RGBTOARGB8_SVE_2X
      "st4b {z16.b, z17.b, z18.b, z19.b}, p1, [%[dst_argb]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                               // %[src_y]
        [src_u] "+r"(src_u),                               // %[src_u]
        [src_v] "+r"(src_v),                               // %[src_v]
        [src_a] "+r"(src_a),                               // %[src_a]
        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
        [width] "+r"(width),                               // %[width]
        [vl] "=&r"(vl)                                     // %[vl]
      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
      : "cc", "memory", YUVTORGB_SVE_REGS);
}

static inline void I444AlphaToARGBRow_SVE_SC(
    const uint8_t* src_y,
    const uint8_t* src_u,
    const uint8_t* src_v,
    const uint8_t* src_a,
    uint8_t* dst_argb,
    const struct YuvConstants* yuvconstants,
    int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm volatile(
      "cnth %[vl] \n"
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.h \n"
      "1: \n"  //
      READYUV444_SVE
      "ld1b {z19.h}, p1/z, [%[src_a]] \n"
      "add %[src_a], %[src_a], %[vl] \n"  // Alpha
      I4XXTORGB_SVE RGBTOARGB8_SVE
      "subs %w[width], %w[width], %w[vl] \n"
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"
      "add %[dst_argb], %[dst_argb], %[vl], lsl #2 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "whilelt p1.h, wzr, %w[width] \n"  //
      READYUV444_SVE
      "ld1b {z19.h}, p1/z, [%[src_a]] \n"  // Alpha
      I4XXTORGB_SVE RGBTOARGB8_SVE
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                               // %[src_y]
        [src_u] "+r"(src_u),                               // %[src_u]
        [src_v] "+r"(src_v),                               // %[src_v]
        [src_a] "+r"(src_a),                               // %[src_a]
        [dst_argb] "+r"(dst_argb),                         // %[dst_argb]
        [width] "+r"(width),                               // %[width]
        [vl] "=&r"(vl)                                     // %[vl]
      : [kUVCoeff] "r"(&yuvconstants->kUVCoeff),           // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias)  // %[kRGBCoeffBias]
      : "cc", "memory", YUVTORGB_SVE_REGS);
}

static inline void NV12ToARGBRow_SVE_SC(const uint8_t* src_y,
                                        const uint8_t* src_uv,
                                        uint8_t* dst_argb,
                                        const struct YuvConstants* yuvconstants,
                                        int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm("cntb %0" : "=r"(vl));
  int width_last_y = width & (vl - 1);
  int width_last_uv = width_last_y + (width_last_y & 1);
  asm volatile(
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "dup z19.b, #255 \n"  // Alpha
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.b \n"
      "ptrue p2.b \n"
      "1: \n"  //
      READNV_SVE_2X NVTORGB_SVE_2X(b, t) RGBTOARGB8_SVE_2X
      "subs %w[width], %w[width], %w[vl] \n"
      "st4b {z16.b, z17.b, z18.b, z19.b}, p1, [%[dst_argb]] \n"
      "add %[dst_argb], %[dst_argb], %[vl], lsl #2 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "whilelt p1.b, wzr, %w[width_last_y] \n"
      "whilelt p2.b, wzr, %w[width_last_uv] \n"  //
      READNV_SVE_2X NVTORGB_SVE_2X(b, t) RGBTOARGB8_SVE_2X
      "st4b {z16.b, z17.b, z18.b, z19.b}, p1, [%[dst_argb]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                                // %[src_y]
        [src_uv] "+r"(src_uv),                              // %[src_uv]
        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
        [width] "+r"(width)                                 // %[width]
      : [vl] "r"(vl),                                       // %[vl]
        [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
        [width_last_y] "r"(width_last_y),                   // %[width_last_y]
        [width_last_uv] "r"(width_last_uv)                  // %[width_last_uv]
      : "cc", "memory", YUVTORGB_SVE_REGS, "p2");
}

static inline void NV21ToARGBRow_SVE_SC(const uint8_t* src_y,
                                        const uint8_t* src_vu,
                                        uint8_t* dst_argb,
                                        const struct YuvConstants* yuvconstants,
                                        int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm("cntb %0" : "=r"(vl));
  int width_last_y = width & (vl - 1);
  int width_last_uv = width_last_y + (width_last_y & 1);
  asm volatile(
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "dup z19.b, #255 \n"  // Alpha
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.b \n"
      "ptrue p2.b \n"
      "1: \n"  //
      READNV_SVE_2X NVTORGB_SVE_2X(t, b) RGBTOARGB8_SVE_2X
      "subs %w[width], %w[width], %w[vl] \n"
      "st4b {z16.b, z17.b, z18.b, z19.b}, p1, [%[dst_argb]] \n"
      "add %[dst_argb], %[dst_argb], %[vl], lsl #2 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "whilelt p1.b, wzr, %w[width_last_y] \n"
      "whilelt p2.b, wzr, %w[width_last_uv] \n"  //
      READNV_SVE_2X NVTORGB_SVE_2X(t, b) RGBTOARGB8_SVE_2X
      "st4b {z16.b, z17.b, z18.b, z19.b}, p1, [%[dst_argb]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                                // %[src_y]
        [src_uv] "+r"(src_vu),                              // %[src_vu]
        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
        [width] "+r"(width)                                 // %[width]
      : [vl] "r"(vl),                                       // %[vl]
        [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
        [width_last_y] "r"(width_last_y),                   // %[width_last_y]
        [width_last_uv] "r"(width_last_uv)                  // %[width_last_uv]
      : "cc", "memory", YUVTORGB_SVE_REGS, "p2");
}

static inline void NV12ToRGB24Row_SVE_SC(
    const uint8_t* src_y,
    const uint8_t* src_uv,
    uint8_t* dst_rgb24,
    const struct YuvConstants* yuvconstants,
    int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm("cntb %0" : "=r"(vl));
  int width_last_y = width & (vl - 1);
  int width_last_uv = width_last_y + (width_last_y & 1);
  asm volatile(
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "dup z19.b, #255 \n"  // Alpha
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.b \n"
      "ptrue p2.b \n"
      "1: \n"  //
      READNV_SVE_2X NVTORGB_SVE_2X(b, t) RGBTOARGB8_SVE_2X
      "subs %w[width], %w[width], %w[vl] \n"
      "st3b {z16.b, z17.b, z18.b}, p1, [%[dst_rgb24]] \n"
      "incb %[dst_rgb24], all, mul #3 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "whilelt p1.b, wzr, %w[width_last_y] \n"
      "whilelt p2.b, wzr, %w[width_last_uv] \n"  //
      READNV_SVE_2X NVTORGB_SVE_2X(b, t) RGBTOARGB8_SVE_2X
      "st3b {z16.b, z17.b, z18.b}, p1, [%[dst_rgb24]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                                // %[src_y]
        [src_uv] "+r"(src_uv),                              // %[src_uv]
        [dst_rgb24] "+r"(dst_rgb24),                        // %[dst_rgb24]
        [width] "+r"(width)                                 // %[width]
      : [vl] "r"(vl),                                       // %[vl]
        [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
        [width_last_y] "r"(width_last_y),                   // %[width_last_y]
        [width_last_uv] "r"(width_last_uv)                  // %[width_last_uv]
      : "cc", "memory", YUVTORGB_SVE_REGS, "p2");
}

static inline void NV21ToRGB24Row_SVE_SC(
    const uint8_t* src_y,
    const uint8_t* src_vu,
    uint8_t* dst_rgb24,
    const struct YuvConstants* yuvconstants,
    int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm("cntb %0" : "=r"(vl));
  int width_last_y = width & (vl - 1);
  int width_last_uv = width_last_y + (width_last_y & 1);
  asm volatile(
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "dup z19.b, #255 \n"  // Alpha
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.b \n"
      "ptrue p2.b \n"
      "1: \n"  //
      READNV_SVE_2X NVTORGB_SVE_2X(t, b) RGBTOARGB8_SVE_2X
      "subs %w[width], %w[width], %w[vl] \n"
      "st3b {z16.b, z17.b, z18.b}, p1, [%[dst_rgb24]] \n"
      "incb %[dst_rgb24], all, mul #3 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "whilelt p1.b, wzr, %w[width_last_y] \n"
      "whilelt p2.b, wzr, %w[width_last_uv] \n"  //
      READNV_SVE_2X NVTORGB_SVE_2X(t, b) RGBTOARGB8_SVE_2X
      "st3b {z16.b, z17.b, z18.b}, p1, [%[dst_rgb24]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                                // %[src_y]
        [src_uv] "+r"(src_vu),                              // %[src_vu]
        [dst_rgb24] "+r"(dst_rgb24),                        // %[dst_rgb24]
        [width] "+r"(width)                                 // %[width]
      : [vl] "r"(vl),                                       // %[vl]
        [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
        [width_last_y] "r"(width_last_y),                   // %[width_last_y]
        [width_last_uv] "r"(width_last_uv)                  // %[width_last_uv]
      : "cc", "memory", YUVTORGB_SVE_REGS, "p2");
}

#define READYUY2_SVE                                        \
  "ld1w {z0.s}, p2/z, [%[src_yuy2]] \n" /* YUYV */ \
  "incb %[src_yuy2] \n"            \
  "prfm pldl1keep, [%[src_yuy2], 448] \n"            \
  "tbl z1.b, {z0.b}, z22.b \n" /* UVUV */ \
  "trn1 z0.b, z0.b, z0.b \n" /* YYYY */

static inline void YUY2ToARGBRow_SVE_SC(const uint8_t* src_yuy2,
                                        uint8_t* dst_argb,
                                        const struct YuvConstants* yuvconstants,
                                        int width) STREAMING_COMPATIBLE {
  uint32_t nv_uv_start = 0x03010301U;
  uint32_t nv_uv_step = 0x04040404U;
  uint64_t vl;
  asm("cnth %0" : "=r"(vl));
  int width_last_y = width & (vl - 1);
  int width_last_uv = width_last_y + (width_last_y & 1);
  asm volatile(
      "ptrue p0.b \n"
      "index z22.s, %w[nv_uv_start], %w[nv_uv_step] \n"
      "dup z19.b, #255 \n"  // Alpha
      YUVTORGB_SVE_SETUP
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.h \n"
      "ptrue p2.h \n"
      "1: \n"  //
      READYUY2_SVE NVTORGB_SVE RGBTOARGB8_SVE
      "subs %w[width], %w[width], %w[vl] \n"
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"
      "add %[dst_argb], %[dst_argb], %[vl], lsl #2 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "whilelt p1.h, wzr, %w[width_last_y] \n"
      "whilelt p2.h, wzr, %w[width_last_uv] \n"  //
      READYUY2_SVE NVTORGB_SVE RGBTOARGB8_SVE
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"

      "99: \n"
      : [src_yuy2] "+r"(src_yuy2),                          // %[src_yuy2]
        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
        [width] "+r"(width)                                 // %[width]
      : [vl] "r"(vl),                                       // %[vl]
        [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
        [nv_uv_start] "r"(nv_uv_start),                     // %[nv_uv_start]
        [nv_uv_step] "r"(nv_uv_step),                       // %[nv_uv_step]
        [width_last_y] "r"(width_last_y),                   // %[width_last_y]
        [width_last_uv] "r"(width_last_uv)                  // %[width_last_uv]
      : "cc", "memory", YUVTORGB_SVE_REGS, "p2");
}

#define READUYVY_SVE                                        \
  "ld1w {z0.s}, p2/z, [%[src_uyvy]] \n" /* UYVY */ \
  "incb %[src_uyvy] \n"            \
  "prfm pldl1keep, [%[src_uyvy], 448] \n"            \
  "tbl z1.b, {z0.b}, z22.b \n" /* UVUV */ \
  "trn2 z0.b, z0.b, z0.b \n" /* YYYY */

static inline void UYVYToARGBRow_SVE_SC(const uint8_t* src_uyvy,
                                        uint8_t* dst_argb,
                                        const struct YuvConstants* yuvconstants,
                                        int width) STREAMING_COMPATIBLE {
  uint32_t nv_uv_start = 0x02000200U;
  uint32_t nv_uv_step = 0x04040404U;
  uint64_t vl;
  asm("cnth %0" : "=r"(vl));
  int width_last_y = width & (vl - 1);
  int width_last_uv = width_last_y + (width_last_y & 1);
  asm volatile(
      "ptrue p0.b \n"
      "index z22.s, %w[nv_uv_start], %w[nv_uv_step] \n"
      "dup z19.b, #255 \n"  // Alpha
      YUVTORGB_SVE_SETUP
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.h \n"
      "ptrue p2.h \n"
      "1: \n"  //
      READUYVY_SVE NVTORGB_SVE RGBTOARGB8_SVE
      "subs %w[width], %w[width], %w[vl] \n"
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"
      "add %[dst_argb], %[dst_argb], %[vl], lsl #2 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "2: \n"
      "whilelt p1.h, wzr, %w[width_last_y] \n"
      "whilelt p2.h, wzr, %w[width_last_uv] \n"  //
      READUYVY_SVE NVTORGB_SVE RGBTOARGB8_SVE
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"

      "99: \n"
      : [src_uyvy] "+r"(src_uyvy),                          // %[src_yuy2]
        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
        [width] "+r"(width)                                 // %[width]
      : [vl] "r"(vl),                                       // %[vl]
        [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
        [nv_uv_start] "r"(nv_uv_start),                     // %[nv_uv_start]
        [nv_uv_step] "r"(nv_uv_step),                       // %[nv_uv_step]
        [width_last_y] "r"(width_last_y),                   // %[width_last_y]
        [width_last_uv] "r"(width_last_uv)                  // %[width_last_uv]
      : "cc", "memory", YUVTORGB_SVE_REGS, "p2");
}

static inline void I210ToARGBRow_SVE_SC(const uint16_t* src_y,
                                        const uint16_t* src_u,
                                        const uint16_t* src_v,
                                        uint8_t* dst_argb,
                                        const struct YuvConstants* yuvconstants,
                                        int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm("cnth %0" : "=r"(vl));
  int width_last_y = width & (vl - 1);
  asm volatile(
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "dup z19.b, #255 \n"  // Alpha
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.h \n"
      "1: \n"  //
      READI210_SVE I4XXTORGB_SVE RGBTOARGB8_SVE
      "subs %w[width], %w[width], %w[vl] \n"
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"
      "add %[dst_argb], %[dst_argb], %[vl], lsl #2 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "whilelt p1.h, wzr, %w[width_last_y] \n"  //
      READI210_SVE I4XXTORGB_SVE RGBTOARGB8_SVE
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                                // %[src_y]
        [src_u] "+r"(src_u),                                // %[src_u]
        [src_v] "+r"(src_v),                                // %[src_v]
        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
        [width] "+r"(width)                                 // %[width]
      : [vl] "r"(vl),                                       // %[vl]
        [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
        [width_last_y] "r"(width_last_y)                    // %[width_last_y]
      : "cc", "memory", YUVTORGB_SVE_REGS);
}

static inline void I210AlphaToARGBRow_SVE_SC(
    const uint16_t* src_y,
    const uint16_t* src_u,
    const uint16_t* src_v,
    const uint16_t* src_a,
    uint8_t* dst_argb,
    const struct YuvConstants* yuvconstants,
    int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm("cnth %0" : "=r"(vl));
  int width_last_y = width & (vl - 1);
  asm volatile(
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.h \n"
      "1: \n"  //
      READI210_SVE
      "ld1h {z19.h}, p1/z, [%[src_a]] \n"  //
      I4XXTORGB_SVE
      "incb %[src_a] \n"  //
      RGBATOARGB8_SVE
      "subs %w[width], %w[width], %w[vl] \n"
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"
      "add %[dst_argb], %[dst_argb], %[vl], lsl #2 \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "whilelt p1.h, wzr, %w[width_last_y] \n"  //
      READI210_SVE
      "ld1h {z19.h}, p1/z, [%[src_a]] \n"  //
      I4XXTORGB_SVE RGBATOARGB8_SVE
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"

      "99: \n"
      : [src_y] "+r"(src_y),                                // %[src_y]
        [src_u] "+r"(src_u),                                // %[src_u]
        [src_v] "+r"(src_v),                                // %[src_v]
        [src_a] "+r"(src_a),                                // %[src_a]
        [dst_argb] "+r"(dst_argb),                          // %[dst_argb]
        [width] "+r"(width)                                 // %[width]
      : [vl] "r"(vl),                                       // %[vl]
        [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
        [width_last_y] "r"(width_last_y)                    // %[width_last_y]
      : "cc", "memory", YUVTORGB_SVE_REGS);
}

static inline void I210ToAR30Row_SVE_SC(const uint16_t* src_y,
                                        const uint16_t* src_u,
                                        const uint16_t* src_v,
                                        uint8_t* dst_ar30,
                                        const struct YuvConstants* yuvconstants,
                                        int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm("cnth %0" : "=r"(vl));
  int width_last_y = width & (vl - 1);
  // The limit is used for saturating the 2.14 red channel in STOREAR30_SVE.
  uint16_t limit = 0x3ff0;
  asm volatile(
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "dup z23.h, %w[limit] \n"
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.h \n"
      "1: \n"  //
      READI210_SVE I4XXTORGB_SVE STOREAR30_SVE
      "subs %w[width], %w[width], %w[vl] \n"
      "b.ge 1b \n"

      "2: \n"
      "adds %w[width], %w[width], %w[vl] \n"
      "b.eq 99f \n"

      // Calculate a predicate for the final iteration to deal with the tail.
      "whilelt p1.h, wzr, %w[width_last_y] \n"  //
      READI210_SVE I4XXTORGB_SVE STOREAR30_SVE

      "99: \n"
      : [src_y] "+r"(src_y),                                // %[src_y]
        [src_u] "+r"(src_u),                                // %[src_u]
        [src_v] "+r"(src_v),                                // %[src_v]
        [dst_ar30] "+r"(dst_ar30),                          // %[dst_ar30]
        [width] "+r"(width)                                 // %[width]
      : [vl] "r"(vl),                                       // %[vl]
        [kUVCoeff] "r"(&yuvconstants->kUVCoeff),            // %[kUVCoeff]
        [kRGBCoeffBias] "r"(&yuvconstants->kRGBCoeffBias),  // %[kRGBCoeffBias]
        [width_last_y] "r"(width_last_y),                   // %[width_last_y]
        [limit] "r"(limit)                                  // %[limit]
      : "cc", "memory", YUVTORGB_SVE_REGS);
}

// P210 has 10 bits in msb of 16 bit NV12 style layout.
static inline void P210ToARGBRow_SVE_SC(const uint16_t* src_y,
                                        const uint16_t* src_uv,
                                        uint8_t* dst_argb,
                                        const struct YuvConstants* yuvconstants,
                                        int width) STREAMING_COMPATIBLE {
  uint64_t vl;
  asm("cnth %0" : "=r"(vl));
  int width_last_y = width & (vl - 1);
  int width_last_uv = width_last_y + (width_last_y & 1);
  uint32_t nv_uv_start = 0x03010301U;
  uint32_t nv_uv_step = 0x04040404U;
  asm volatile(
      "ptrue p0.b \n"  //
      YUVTORGB_SVE_SETUP
      "index z22.s, %w[nv_uv_start], %w[nv_uv_step] \n"
      "dup z19.b, #255 \n"  // Alpha
      "subs %w[width], %w[width], %w[vl] \n"
      "b.lt 2f \n"

      // Run bulk of computation with an all-true predicate to avoid predicate
      // generation overhead.
      "ptrue p1.h \n"
      "ptrue p2.h \n"
      "1: \n"  //
      READP210_SVE NVTORGB_SVE RGBTOARGB8_SVE
      "subs %w[width], %w[width], %w[vl] \n"
      "st2h {z16.h, z17.h}, p1, [%[dst_argb]] \n"
      "add %[dst_argb], %[dst_argb], %[vl], lsl #2 \n"
--> --------------------

--> maximum size reached

--> --------------------