xref: /external/webp/src/dsp/enc_neon.c

// Copyright 2012 Google Inc. All Rights Reserved.
//
// Use of this source code is governed by a BSD-style license
// that can be found in the COPYING 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.
// -----------------------------------------------------------------------------
//
// ARM NEON version of speed-critical encoding functions.
//
// adapted from libvpx (http://www.webmproject.org/code/)

#include "./dsp.h"

#if defined(WEBP_USE_NEON)

#include <assert.h>

#include "./neon.h"
#include "../enc/vp8enci.h"

//------------------------------------------------------------------------------
// Transforms (Paragraph 14.4)

// Inverse transform.
// This code is pretty much the same as TransformOne in the dec_neon.c, except
// for subtraction to *ref. See the comments there for algorithmic explanations.

static const int16_t kC1 = 20091;
static const int16_t kC2 = 17734;  // half of kC2, actually. See comment above.

// This code works but is *slower* than the inlined-asm version below
// (with gcc-4.6). So we disable it for now. Later, it'll be conditional to
// USE_INTRINSICS define.
// With gcc-4.8, it's a little faster speed than inlined-assembly.
#if defined(USE_INTRINSICS)

// Treats 'v' as an uint8x8_t and zero extends to an int16x8_t.
static WEBP_INLINE int16x8_t ConvertU8ToS16(uint32x2_t v) {
  return vreinterpretq_s16_u16(vmovl_u8(vreinterpret_u8_u32(v)));
}

// Performs unsigned 8b saturation on 'dst01' and 'dst23' storing the result
// to the corresponding rows of 'dst'.
static WEBP_INLINE void SaturateAndStore4x4(uint8_t* const dst,
                                            const int16x8_t dst01,
                                            const int16x8_t dst23) {
  // Unsigned saturate to 8b.
  const uint8x8_t dst01_u8 = vqmovun_s16(dst01);
  const uint8x8_t dst23_u8 = vqmovun_s16(dst23);

  // Store the results.
  vst1_lane_u32((uint32_t*)(dst + 0 * BPS), vreinterpret_u32_u8(dst01_u8), 0);
  vst1_lane_u32((uint32_t*)(dst + 1 * BPS), vreinterpret_u32_u8(dst01_u8), 1);
  vst1_lane_u32((uint32_t*)(dst + 2 * BPS), vreinterpret_u32_u8(dst23_u8), 0);
  vst1_lane_u32((uint32_t*)(dst + 3 * BPS), vreinterpret_u32_u8(dst23_u8), 1);
}

static WEBP_INLINE void Add4x4(const int16x8_t row01, const int16x8_t row23,
                               const uint8_t* const ref, uint8_t* const dst) {
  uint32x2_t dst01 = vdup_n_u32(0);
  uint32x2_t dst23 = vdup_n_u32(0);

  // Load the source pixels.
  dst01 = vld1_lane_u32((uint32_t*)(ref + 0 * BPS), dst01, 0);
  dst23 = vld1_lane_u32((uint32_t*)(ref + 2 * BPS), dst23, 0);
  dst01 = vld1_lane_u32((uint32_t*)(ref + 1 * BPS), dst01, 1);
  dst23 = vld1_lane_u32((uint32_t*)(ref + 3 * BPS), dst23, 1);

  {
    // Convert to 16b.
    const int16x8_t dst01_s16 = ConvertU8ToS16(dst01);
    const int16x8_t dst23_s16 = ConvertU8ToS16(dst23);

    // Descale with rounding.
    const int16x8_t out01 = vrsraq_n_s16(dst01_s16, row01, 3);
    const int16x8_t out23 = vrsraq_n_s16(dst23_s16, row23, 3);
    // Add the inverse transform.
    SaturateAndStore4x4(dst, out01, out23);
  }
}

static WEBP_INLINE void Transpose8x2(const int16x8_t in0, const int16x8_t in1,
                                     int16x8x2_t* const out) {
  // a0 a1 a2 a3 | b0 b1 b2 b3   => a0 b0 c0 d0 | a1 b1 c1 d1
  // c0 c1 c2 c3 | d0 d1 d2 d3      a2 b2 c2 d2 | a3 b3 c3 d3
  const int16x8x2_t tmp0 = vzipq_s16(in0, in1);   // a0 c0 a1 c1 a2 c2 ...
                                                  // b0 d0 b1 d1 b2 d2 ...
  *out = vzipq_s16(tmp0.val[0], tmp0.val[1]);
}

static WEBP_INLINE void TransformPass(int16x8x2_t* const rows) {
  // {rows} = in0 | in4
  //          in8 | in12
  // B1 = in4 | in12
  const int16x8_t B1 =
      vcombine_s16(vget_high_s16(rows->val[0]), vget_high_s16(rows->val[1]));
  // C0 = kC1 * in4 | kC1 * in12
  // C1 = kC2 * in4 | kC2 * in12
  const int16x8_t C0 = vsraq_n_s16(B1, vqdmulhq_n_s16(B1, kC1), 1);
  const int16x8_t C1 = vqdmulhq_n_s16(B1, kC2);
  const int16x4_t a = vqadd_s16(vget_low_s16(rows->val[0]),
                                vget_low_s16(rows->val[1]));   // in0 + in8
  const int16x4_t b = vqsub_s16(vget_low_s16(rows->val[0]),
                                vget_low_s16(rows->val[1]));   // in0 - in8
  // c = kC2 * in4 - kC1 * in12
  // d = kC1 * in4 + kC2 * in12
  const int16x4_t c = vqsub_s16(vget_low_s16(C1), vget_high_s16(C0));
  const int16x4_t d = vqadd_s16(vget_low_s16(C0), vget_high_s16(C1));
  const int16x8_t D0 = vcombine_s16(a, b);      // D0 = a | b
  const int16x8_t D1 = vcombine_s16(d, c);      // D1 = d | c
  const int16x8_t E0 = vqaddq_s16(D0, D1);      // a+d | b+c
  const int16x8_t E_tmp = vqsubq_s16(D0, D1);   // a-d | b-c
  const int16x8_t E1 = vcombine_s16(vget_high_s16(E_tmp), vget_low_s16(E_tmp));
  Transpose8x2(E0, E1, rows);
}

static void ITransformOne(const uint8_t* ref,
                          const int16_t* in, uint8_t* dst) {
  int16x8x2_t rows;
  INIT_VECTOR2(rows, vld1q_s16(in + 0), vld1q_s16(in + 8));
  TransformPass(&rows);
  TransformPass(&rows);
  Add4x4(rows.val[0], rows.val[1], ref, dst);
}

#else

static void ITransformOne(const uint8_t* ref,
                          const int16_t* in, uint8_t* dst) {
  const int kBPS = BPS;
  const int16_t kC1C2[] = { kC1, kC2, 0, 0 };

  __asm__ volatile (
    "vld1.16         {q1, q2}, [%[in]]           \n"
    "vld1.16         {d0}, [%[kC1C2]]            \n"

    // d2: in[0]
    // d3: in[8]
    // d4: in[4]
    // d5: in[12]
    "vswp            d3, d4                      \n"

    // q8 = {in[4], in[12]} * kC1 * 2 >> 16
    // q9 = {in[4], in[12]} * kC2 >> 16
    "vqdmulh.s16     q8, q2, d0[0]               \n"
    "vqdmulh.s16     q9, q2, d0[1]               \n"

    // d22 = a = in[0] + in[8]
    // d23 = b = in[0] - in[8]
    "vqadd.s16       d22, d2, d3                 \n"
    "vqsub.s16       d23, d2, d3                 \n"

    //  q8 = in[4]/[12] * kC1 >> 16
    "vshr.s16        q8, q8, #1                  \n"

    // Add {in[4], in[12]} back after the multiplication.
    "vqadd.s16       q8, q2, q8                  \n"

    // d20 = c = in[4]*kC2 - in[12]*kC1
    // d21 = d = in[4]*kC1 + in[12]*kC2
    "vqsub.s16       d20, d18, d17               \n"
    "vqadd.s16       d21, d19, d16               \n"

    // d2 = tmp[0] = a + d
    // d3 = tmp[1] = b + c
    // d4 = tmp[2] = b - c
    // d5 = tmp[3] = a - d
    "vqadd.s16       d2, d22, d21                \n"
    "vqadd.s16       d3, d23, d20                \n"
    "vqsub.s16       d4, d23, d20                \n"
    "vqsub.s16       d5, d22, d21                \n"

    "vzip.16         q1, q2                      \n"
    "vzip.16         q1, q2                      \n"

    "vswp            d3, d4                      \n"

    // q8 = {tmp[4], tmp[12]} * kC1 * 2 >> 16
    // q9 = {tmp[4], tmp[12]} * kC2 >> 16
    "vqdmulh.s16     q8, q2, d0[0]               \n"
    "vqdmulh.s16     q9, q2, d0[1]               \n"

    // d22 = a = tmp[0] + tmp[8]
    // d23 = b = tmp[0] - tmp[8]
    "vqadd.s16       d22, d2, d3                 \n"
    "vqsub.s16       d23, d2, d3                 \n"

    "vshr.s16        q8, q8, #1                  \n"
    "vqadd.s16       q8, q2, q8                  \n"

    // d20 = c = in[4]*kC2 - in[12]*kC1
    // d21 = d = in[4]*kC1 + in[12]*kC2
    "vqsub.s16       d20, d18, d17               \n"
    "vqadd.s16       d21, d19, d16               \n"

    // d2 = tmp[0] = a + d
    // d3 = tmp[1] = b + c
    // d4 = tmp[2] = b - c
    // d5 = tmp[3] = a - d
    "vqadd.s16       d2, d22, d21                \n"
    "vqadd.s16       d3, d23, d20                \n"
    "vqsub.s16       d4, d23, d20                \n"
    "vqsub.s16       d5, d22, d21                \n"

    "vld1.32         d6[0], [%[ref]], %[kBPS]    \n"
    "vld1.32         d6[1], [%[ref]], %[kBPS]    \n"
    "vld1.32         d7[0], [%[ref]], %[kBPS]    \n"
    "vld1.32         d7[1], [%[ref]], %[kBPS]    \n"

    "sub         %[ref], %[ref], %[kBPS], lsl #2 \n"

    // (val) + 4 >> 3
    "vrshr.s16       d2, d2, #3                  \n"
    "vrshr.s16       d3, d3, #3                  \n"
    "vrshr.s16       d4, d4, #3                  \n"
    "vrshr.s16       d5, d5, #3                  \n"

    "vzip.16         q1, q2                      \n"
    "vzip.16         q1, q2                      \n"

    // Must accumulate before saturating
    "vmovl.u8        q8, d6                      \n"
    "vmovl.u8        q9, d7                      \n"

    "vqadd.s16       q1, q1, q8                  \n"
    "vqadd.s16       q2, q2, q9                  \n"

    "vqmovun.s16     d0, q1                      \n"
    "vqmovun.s16     d1, q2                      \n"

    "vst1.32         d0[0], [%[dst]], %[kBPS]    \n"
    "vst1.32         d0[1], [%[dst]], %[kBPS]    \n"
    "vst1.32         d1[0], [%[dst]], %[kBPS]    \n"
    "vst1.32         d1[1], [%[dst]]             \n"

    : [in] "+r"(in), [dst] "+r"(dst)               // modified registers
    : [kBPS] "r"(kBPS), [kC1C2] "r"(kC1C2), [ref] "r"(ref)  // constants
    : "memory", "q0", "q1", "q2", "q8", "q9", "q10", "q11"  // clobbered
  );
}

#endif    // USE_INTRINSICS

static void ITransform(const uint8_t* ref,
                       const int16_t* in, uint8_t* dst, int do_two) {
  ITransformOne(ref, in, dst);
  if (do_two) {
    ITransformOne(ref + 4, in + 16, dst + 4);
  }
}

// Load all 4x4 pixels into a single uint8x16_t variable.
static uint8x16_t Load4x4(const uint8_t* src) {
  uint32x4_t out = vdupq_n_u32(0);
  out = vld1q_lane_u32((const uint32_t*)(src + 0 * BPS), out, 0);
  out = vld1q_lane_u32((const uint32_t*)(src + 1 * BPS), out, 1);
  out = vld1q_lane_u32((const uint32_t*)(src + 2 * BPS), out, 2);
  out = vld1q_lane_u32((const uint32_t*)(src + 3 * BPS), out, 3);
  return vreinterpretq_u8_u32(out);
}

// Forward transform.

#if defined(USE_INTRINSICS)

static WEBP_INLINE void Transpose4x4_S16(const int16x4_t A, const int16x4_t B,
                                         const int16x4_t C, const int16x4_t D,
                                         int16x8_t* const out01,
                                         int16x8_t* const out32) {
  const int16x4x2_t AB = vtrn_s16(A, B);
  const int16x4x2_t CD = vtrn_s16(C, D);
  const int32x2x2_t tmp02 = vtrn_s32(vreinterpret_s32_s16(AB.val[0]),
                                     vreinterpret_s32_s16(CD.val[0]));
  const int32x2x2_t tmp13 = vtrn_s32(vreinterpret_s32_s16(AB.val[1]),
                                     vreinterpret_s32_s16(CD.val[1]));
  *out01 = vreinterpretq_s16_s64(
      vcombine_s64(vreinterpret_s64_s32(tmp02.val[0]),
                   vreinterpret_s64_s32(tmp13.val[0])));
  *out32 = vreinterpretq_s16_s64(
      vcombine_s64(vreinterpret_s64_s32(tmp13.val[1]),
                   vreinterpret_s64_s32(tmp02.val[1])));
}

static WEBP_INLINE int16x8_t DiffU8ToS16(const uint8x8_t a,
                                         const uint8x8_t b) {
  return vreinterpretq_s16_u16(vsubl_u8(a, b));
}

static void FTransform(const uint8_t* src, const uint8_t* ref,
                       int16_t* out) {
  int16x8_t d0d1, d3d2;   // working 4x4 int16 variables
  {
    const uint8x16_t S0 = Load4x4(src);
    const uint8x16_t R0 = Load4x4(ref);
    const int16x8_t D0D1 = DiffU8ToS16(vget_low_u8(S0), vget_low_u8(R0));
    const int16x8_t D2D3 = DiffU8ToS16(vget_high_u8(S0), vget_high_u8(R0));
    const int16x4_t D0 = vget_low_s16(D0D1);
    const int16x4_t D1 = vget_high_s16(D0D1);
    const int16x4_t D2 = vget_low_s16(D2D3);
    const int16x4_t D3 = vget_high_s16(D2D3);
    Transpose4x4_S16(D0, D1, D2, D3, &d0d1, &d3d2);
  }
  {    // 1rst pass
    const int32x4_t kCst937 = vdupq_n_s32(937);
    const int32x4_t kCst1812 = vdupq_n_s32(1812);
    const int16x8_t a0a1 = vaddq_s16(d0d1, d3d2);   // d0+d3 | d1+d2   (=a0|a1)
    const int16x8_t a3a2 = vsubq_s16(d0d1, d3d2);   // d0-d3 | d1-d2   (=a3|a2)
    const int16x8_t a0a1_2 = vshlq_n_s16(a0a1, 3);
    const int16x4_t tmp0 = vadd_s16(vget_low_s16(a0a1_2),
                                    vget_high_s16(a0a1_2));
    const int16x4_t tmp2 = vsub_s16(vget_low_s16(a0a1_2),
                                    vget_high_s16(a0a1_2));
    const int32x4_t a3_2217 = vmull_n_s16(vget_low_s16(a3a2), 2217);
    const int32x4_t a2_2217 = vmull_n_s16(vget_high_s16(a3a2), 2217);
    const int32x4_t a2_p_a3 = vmlal_n_s16(a2_2217, vget_low_s16(a3a2), 5352);
    const int32x4_t a3_m_a2 = vmlsl_n_s16(a3_2217, vget_high_s16(a3a2), 5352);
    const int16x4_t tmp1 = vshrn_n_s32(vaddq_s32(a2_p_a3, kCst1812), 9);
    const int16x4_t tmp3 = vshrn_n_s32(vaddq_s32(a3_m_a2, kCst937), 9);
    Transpose4x4_S16(tmp0, tmp1, tmp2, tmp3, &d0d1, &d3d2);
  }
  {    // 2nd pass
    // the (1<<16) addition is for the replacement: a3!=0  <-> 1-(a3==0)
    const int32x4_t kCst12000 = vdupq_n_s32(12000 + (1 << 16));
    const int32x4_t kCst51000 = vdupq_n_s32(51000);
    const int16x8_t a0a1 = vaddq_s16(d0d1, d3d2);   // d0+d3 | d1+d2   (=a0|a1)
    const int16x8_t a3a2 = vsubq_s16(d0d1, d3d2);   // d0-d3 | d1-d2   (=a3|a2)
    const int16x4_t a0_k7 = vadd_s16(vget_low_s16(a0a1), vdup_n_s16(7));
    const int16x4_t out0 = vshr_n_s16(vadd_s16(a0_k7, vget_high_s16(a0a1)), 4);
    const int16x4_t out2 = vshr_n_s16(vsub_s16(a0_k7, vget_high_s16(a0a1)), 4);
    const int32x4_t a3_2217 = vmull_n_s16(vget_low_s16(a3a2), 2217);
    const int32x4_t a2_2217 = vmull_n_s16(vget_high_s16(a3a2), 2217);
    const int32x4_t a2_p_a3 = vmlal_n_s16(a2_2217, vget_low_s16(a3a2), 5352);
    const int32x4_t a3_m_a2 = vmlsl_n_s16(a3_2217, vget_high_s16(a3a2), 5352);
    const int16x4_t tmp1 = vaddhn_s32(a2_p_a3, kCst12000);
    const int16x4_t out3 = vaddhn_s32(a3_m_a2, kCst51000);
    const int16x4_t a3_eq_0 =
        vreinterpret_s16_u16(vceq_s16(vget_low_s16(a3a2), vdup_n_s16(0)));
    const int16x4_t out1 = vadd_s16(tmp1, a3_eq_0);
    vst1_s16(out +  0, out0);
    vst1_s16(out +  4, out1);
    vst1_s16(out +  8, out2);
    vst1_s16(out + 12, out3);
  }
}

#else

// adapted from vp8/encoder/arm/neon/shortfdct_neon.asm
static const int16_t kCoeff16[] = {
  5352,  5352,  5352, 5352, 2217,  2217,  2217, 2217
};
static const int32_t kCoeff32[] = {
   1812,  1812,  1812,  1812,
    937,   937,   937,   937,
  12000, 12000, 12000, 12000,
  51000, 51000, 51000, 51000
};

static void FTransform(const uint8_t* src, const uint8_t* ref,
                       int16_t* out) {
  const int kBPS = BPS;
  const uint8_t* src_ptr = src;
  const uint8_t* ref_ptr = ref;
  const int16_t* coeff16 = kCoeff16;
  const int32_t* coeff32 = kCoeff32;

  __asm__ volatile (
    // load src into q4, q5 in high half
    "vld1.8 {d8},  [%[src_ptr]], %[kBPS]      \n"
    "vld1.8 {d10}, [%[src_ptr]], %[kBPS]      \n"
    "vld1.8 {d9},  [%[src_ptr]], %[kBPS]      \n"
    "vld1.8 {d11}, [%[src_ptr]]               \n"

    // load ref into q6, q7 in high half
    "vld1.8 {d12}, [%[ref_ptr]], %[kBPS]      \n"
    "vld1.8 {d14}, [%[ref_ptr]], %[kBPS]      \n"
    "vld1.8 {d13}, [%[ref_ptr]], %[kBPS]      \n"
    "vld1.8 {d15}, [%[ref_ptr]]               \n"

    // Pack the high values in to q4 and q6
    "vtrn.32     q4, q5                       \n"
    "vtrn.32     q6, q7                       \n"

    // d[0-3] = src - ref
    "vsubl.u8    q0, d8, d12                  \n"
    "vsubl.u8    q1, d9, d13                  \n"

    // load coeff16 into q8(d16=5352, d17=2217)
    "vld1.16     {q8}, [%[coeff16]]           \n"

    // load coeff32 high half into q9 = 1812, q10 = 937
    "vld1.32     {q9, q10}, [%[coeff32]]!     \n"

    // load coeff32 low half into q11=12000, q12=51000
    "vld1.32     {q11,q12}, [%[coeff32]]      \n"

    // part 1
    // Transpose. Register dN is the same as dN in C
    "vtrn.32         d0, d2                   \n"
    "vtrn.32         d1, d3                   \n"
    "vtrn.16         d0, d1                   \n"
    "vtrn.16         d2, d3                   \n"

    "vadd.s16        d4, d0, d3               \n" // a0 = d0 + d3
    "vadd.s16        d5, d1, d2               \n" // a1 = d1 + d2
    "vsub.s16        d6, d1, d2               \n" // a2 = d1 - d2
    "vsub.s16        d7, d0, d3               \n" // a3 = d0 - d3

    "vadd.s16        d0, d4, d5               \n" // a0 + a1
    "vshl.s16        d0, d0, #3               \n" // temp[0+i*4] = (a0+a1) << 3
    "vsub.s16        d2, d4, d5               \n" // a0 - a1
    "vshl.s16        d2, d2, #3               \n" // (temp[2+i*4] = (a0-a1) << 3

    "vmlal.s16       q9, d7, d16              \n" // a3*5352 + 1812
    "vmlal.s16       q10, d7, d17             \n" // a3*2217 + 937
    "vmlal.s16       q9, d6, d17              \n" // a2*2217 + a3*5352 + 1812
    "vmlsl.s16       q10, d6, d16             \n" // a3*2217 + 937 - a2*5352

    // temp[1+i*4] = (d2*2217 + d3*5352 + 1812) >> 9
    // temp[3+i*4] = (d3*2217 + 937 - d2*5352) >> 9
    "vshrn.s32       d1, q9, #9               \n"
    "vshrn.s32       d3, q10, #9              \n"

    // part 2
    // transpose d0=ip[0], d1=ip[4], d2=ip[8], d3=ip[12]
    "vtrn.32         d0, d2                   \n"
    "vtrn.32         d1, d3                   \n"
    "vtrn.16         d0, d1                   \n"
    "vtrn.16         d2, d3                   \n"

    "vmov.s16        d26, #7                  \n"

    "vadd.s16        d4, d0, d3               \n" // a1 = ip[0] + ip[12]
    "vadd.s16        d5, d1, d2               \n" // b1 = ip[4] + ip[8]
    "vsub.s16        d6, d1, d2               \n" // c1 = ip[4] - ip[8]
    "vadd.s16        d4, d4, d26              \n" // a1 + 7
    "vsub.s16        d7, d0, d3               \n" // d1 = ip[0] - ip[12]

    "vadd.s16        d0, d4, d5               \n" // op[0] = a1 + b1 + 7
    "vsub.s16        d2, d4, d5               \n" // op[8] = a1 - b1 + 7

    "vmlal.s16       q11, d7, d16             \n" // d1*5352 + 12000
    "vmlal.s16       q12, d7, d17             \n" // d1*2217 + 51000

    "vceq.s16        d4, d7, #0               \n"

    "vshr.s16        d0, d0, #4               \n"
    "vshr.s16        d2, d2, #4               \n"

    "vmlal.s16       q11, d6, d17             \n" // c1*2217 + d1*5352 + 12000
    "vmlsl.s16       q12, d6, d16             \n" // d1*2217 - c1*5352 + 51000

    "vmvn            d4, d4                   \n" // !(d1 == 0)
    // op[4] = (c1*2217 + d1*5352 + 12000)>>16
    "vshrn.s32       d1, q11, #16             \n"
    // op[4] += (d1!=0)
    "vsub.s16        d1, d1, d4               \n"
    // op[12]= (d1*2217 - c1*5352 + 51000)>>16
    "vshrn.s32       d3, q12, #16             \n"

    // set result to out array
    "vst1.16         {q0, q1}, [%[out]]   \n"
    : [src_ptr] "+r"(src_ptr), [ref_ptr] "+r"(ref_ptr),
      [coeff32] "+r"(coeff32)          // modified registers
    : [kBPS] "r"(kBPS), [coeff16] "r"(coeff16),
      [out] "r"(out)                   // constants
    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
      "q10", "q11", "q12", "q13"       // clobbered
  );
}

#endif

#define LOAD_LANE_16b(VALUE, LANE) do {             \
  (VALUE) = vld1_lane_s16(src, (VALUE), (LANE));    \
  src += stride;                                    \
} while (0)

static void FTransformWHT(const int16_t* src, int16_t* out) {
  const int stride = 16;
  const int16x4_t zero = vdup_n_s16(0);
  int32x4x4_t tmp0;
  int16x4x4_t in;
  INIT_VECTOR4(in, zero, zero, zero, zero);
  LOAD_LANE_16b(in.val[0], 0);
  LOAD_LANE_16b(in.val[1], 0);
  LOAD_LANE_16b(in.val[2], 0);
  LOAD_LANE_16b(in.val[3], 0);
  LOAD_LANE_16b(in.val[0], 1);
  LOAD_LANE_16b(in.val[1], 1);
  LOAD_LANE_16b(in.val[2], 1);
  LOAD_LANE_16b(in.val[3], 1);
  LOAD_LANE_16b(in.val[0], 2);
  LOAD_LANE_16b(in.val[1], 2);
  LOAD_LANE_16b(in.val[2], 2);
  LOAD_LANE_16b(in.val[3], 2);
  LOAD_LANE_16b(in.val[0], 3);
  LOAD_LANE_16b(in.val[1], 3);
  LOAD_LANE_16b(in.val[2], 3);
  LOAD_LANE_16b(in.val[3], 3);

  {
    // a0 = in[0 * 16] + in[2 * 16]
    // a1 = in[1 * 16] + in[3 * 16]
    // a2 = in[1 * 16] - in[3 * 16]
    // a3 = in[0 * 16] - in[2 * 16]
    const int32x4_t a0 = vaddl_s16(in.val[0], in.val[2]);
    const int32x4_t a1 = vaddl_s16(in.val[1], in.val[3]);
    const int32x4_t a2 = vsubl_s16(in.val[1], in.val[3]);
    const int32x4_t a3 = vsubl_s16(in.val[0], in.val[2]);
    tmp0.val[0] = vaddq_s32(a0, a1);
    tmp0.val[1] = vaddq_s32(a3, a2);
    tmp0.val[2] = vsubq_s32(a3, a2);
    tmp0.val[3] = vsubq_s32(a0, a1);
  }
  {
    const int32x4x4_t tmp1 = Transpose4x4(tmp0);
    // a0 = tmp[0 + i] + tmp[ 8 + i]
    // a1 = tmp[4 + i] + tmp[12 + i]
    // a2 = tmp[4 + i] - tmp[12 + i]
    // a3 = tmp[0 + i] - tmp[ 8 + i]
    const int32x4_t a0 = vaddq_s32(tmp1.val[0], tmp1.val[2]);
    const int32x4_t a1 = vaddq_s32(tmp1.val[1], tmp1.val[3]);
    const int32x4_t a2 = vsubq_s32(tmp1.val[1], tmp1.val[3]);
    const int32x4_t a3 = vsubq_s32(tmp1.val[0], tmp1.val[2]);
    const int32x4_t b0 = vhaddq_s32(a0, a1);  // (a0 + a1) >> 1
    const int32x4_t b1 = vhaddq_s32(a3, a2);  // (a3 + a2) >> 1
    const int32x4_t b2 = vhsubq_s32(a3, a2);  // (a3 - a2) >> 1
    const int32x4_t b3 = vhsubq_s32(a0, a1);  // (a0 - a1) >> 1
    const int16x4_t out0 = vmovn_s32(b0);
    const int16x4_t out1 = vmovn_s32(b1);
    const int16x4_t out2 = vmovn_s32(b2);
    const int16x4_t out3 = vmovn_s32(b3);

    vst1_s16(out +  0, out0);
    vst1_s16(out +  4, out1);
    vst1_s16(out +  8, out2);
    vst1_s16(out + 12, out3);
  }
}
#undef LOAD_LANE_16b

//------------------------------------------------------------------------------
// Texture distortion
//
// We try to match the spectral content (weighted) between source and
// reconstructed samples.

// This code works but is *slower* than the inlined-asm version below
// (with gcc-4.6). So we disable it for now. Later, it'll be conditional to
// USE_INTRINSICS define.
// With gcc-4.8, it's only slightly slower than the inlined.
#if defined(USE_INTRINSICS)

// Zero extend an uint16x4_t 'v' to an int32x4_t.
static WEBP_INLINE int32x4_t ConvertU16ToS32(uint16x4_t v) {
  return vreinterpretq_s32_u32(vmovl_u16(v));
}

// Does a regular 4x4 transpose followed by an adjustment of the upper columns
// in the inner rows to restore the source order of differences,
// i.e., a0 - a1 | a3 - a2.
static WEBP_INLINE int32x4x4_t DistoTranspose4x4(const int32x4x4_t rows) {
  int32x4x4_t out = Transpose4x4(rows);
  // restore source order in the columns containing differences.
  const int32x2_t r1h = vget_high_s32(out.val[1]);
  const int32x2_t r2h = vget_high_s32(out.val[2]);
  out.val[1] = vcombine_s32(vget_low_s32(out.val[1]), r2h);
  out.val[2] = vcombine_s32(vget_low_s32(out.val[2]), r1h);
  return out;
}

static WEBP_INLINE int32x4x4_t DistoHorizontalPass(const uint8x8_t r0r1,
                                                   const uint8x8_t r2r3) {
  // a0 = in[0] + in[2] | a1 = in[1] + in[3]
  const uint16x8_t a0a1 = vaddl_u8(r0r1, r2r3);
  // a3 = in[0] - in[2] | a2 = in[1] - in[3]
  const uint16x8_t a3a2 = vsubl_u8(r0r1, r2r3);
  const int32x4_t tmp0 = vpaddlq_s16(vreinterpretq_s16_u16(a0a1));  // a0 + a1
  const int32x4_t tmp1 = vpaddlq_s16(vreinterpretq_s16_u16(a3a2));  // a3 + a2
  // no pairwise subtraction; reorder to perform tmp[2]/tmp[3] calculations.
  // a0a0 a3a3 a0a0 a3a3 a0a0 a3a3 a0a0 a3a3
  // a1a1 a2a2 a1a1 a2a2 a1a1 a2a2 a1a1 a2a2
  const int16x8x2_t transpose =
      vtrnq_s16(vreinterpretq_s16_u16(a0a1), vreinterpretq_s16_u16(a3a2));
  // tmp[3] = a0 - a1 | tmp[2] = a3 - a2
  const int32x4_t tmp32_1 = vsubl_s16(vget_low_s16(transpose.val[0]),
                                      vget_low_s16(transpose.val[1]));
  const int32x4_t tmp32_2 = vsubl_s16(vget_high_s16(transpose.val[0]),
                                      vget_high_s16(transpose.val[1]));
  // [0]: tmp[3] [1]: tmp[2]
  const int32x4x2_t split = vtrnq_s32(tmp32_1, tmp32_2);
  const int32x4x4_t res = { { tmp0, tmp1, split.val[1], split.val[0] } };
  return res;
}

static WEBP_INLINE int32x4x4_t DistoVerticalPass(const int32x4x4_t rows) {
  // a0 = tmp[0 + i] + tmp[8 + i];
  const int32x4_t a0 = vaddq_s32(rows.val[0], rows.val[1]);
  // a1 = tmp[4 + i] + tmp[12+ i];
  const int32x4_t a1 = vaddq_s32(rows.val[2], rows.val[3]);
  // a2 = tmp[4 + i] - tmp[12+ i];
  const int32x4_t a2 = vsubq_s32(rows.val[2], rows.val[3]);
  // a3 = tmp[0 + i] - tmp[8 + i];
  const int32x4_t a3 = vsubq_s32(rows.val[0], rows.val[1]);
  const int32x4_t b0 = vqabsq_s32(vaddq_s32(a0, a1));  // abs(a0 + a1)
  const int32x4_t b1 = vqabsq_s32(vaddq_s32(a3, a2));  // abs(a3 + a2)
  const int32x4_t b2 = vabdq_s32(a3, a2);              // abs(a3 - a2)
  const int32x4_t b3 = vabdq_s32(a0, a1);              // abs(a0 - a1)
  const int32x4x4_t res = { { b0, b1, b2, b3 } };
  return res;
}

// Calculate the weighted sum of the rows in 'b'.
static WEBP_INLINE int64x1_t DistoSum(const int32x4x4_t b,
                                      const int32x4_t w0, const int32x4_t w1,
                                      const int32x4_t w2, const int32x4_t w3) {
  const int32x4_t s0 = vmulq_s32(w0, b.val[0]);
  const int32x4_t s1 = vmlaq_s32(s0, w1, b.val[1]);
  const int32x4_t s2 = vmlaq_s32(s1, w2, b.val[2]);
  const int32x4_t s3 = vmlaq_s32(s2, w3, b.val[3]);
  const int64x2_t sum1 = vpaddlq_s32(s3);
  const int64x1_t sum2 = vadd_s64(vget_low_s64(sum1), vget_high_s64(sum1));
  return sum2;
}

#define LOAD_LANE_32b(src, VALUE, LANE) \
    (VALUE) = vld1q_lane_u32((const uint32_t*)(src), (VALUE), (LANE))

// Hadamard transform
// Returns the weighted sum of the absolute value of transformed coefficients.
static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
                    const uint16_t* const w) {
  uint32x4_t d0d1 = { 0, 0, 0, 0 };
  uint32x4_t d2d3 = { 0, 0, 0, 0 };
  LOAD_LANE_32b(a + 0 * BPS, d0d1, 0);  // a00 a01 a02 a03
  LOAD_LANE_32b(a + 1 * BPS, d0d1, 1);  // a10 a11 a12 a13
  LOAD_LANE_32b(b + 0 * BPS, d0d1, 2);  // b00 b01 b02 b03
  LOAD_LANE_32b(b + 1 * BPS, d0d1, 3);  // b10 b11 b12 b13
  LOAD_LANE_32b(a + 2 * BPS, d2d3, 0);  // a20 a21 a22 a23
  LOAD_LANE_32b(a + 3 * BPS, d2d3, 1);  // a30 a31 a32 a33
  LOAD_LANE_32b(b + 2 * BPS, d2d3, 2);  // b20 b21 b22 b23
  LOAD_LANE_32b(b + 3 * BPS, d2d3, 3);  // b30 b31 b32 b33

  {
    // a00 a01 a20 a21 a10 a11 a30 a31 b00 b01 b20 b21 b10 b11 b30 b31
    // a02 a03 a22 a23 a12 a13 a32 a33 b02 b03 b22 b23 b12 b13 b32 b33
    const uint16x8x2_t tmp =
        vtrnq_u16(vreinterpretq_u16_u32(d0d1), vreinterpretq_u16_u32(d2d3));
    const uint8x16_t d0d1u8 = vreinterpretq_u8_u16(tmp.val[0]);
    const uint8x16_t d2d3u8 = vreinterpretq_u8_u16(tmp.val[1]);
    const int32x4x4_t hpass_a = DistoHorizontalPass(vget_low_u8(d0d1u8),
                                                    vget_low_u8(d2d3u8));
    const int32x4x4_t hpass_b = DistoHorizontalPass(vget_high_u8(d0d1u8),
                                                    vget_high_u8(d2d3u8));
    const int32x4x4_t tmp_a = DistoTranspose4x4(hpass_a);
    const int32x4x4_t tmp_b = DistoTranspose4x4(hpass_b);
    const int32x4x4_t vpass_a = DistoVerticalPass(tmp_a);
    const int32x4x4_t vpass_b = DistoVerticalPass(tmp_b);
    const int32x4_t w0 = ConvertU16ToS32(vld1_u16(w + 0));
    const int32x4_t w1 = ConvertU16ToS32(vld1_u16(w + 4));
    const int32x4_t w2 = ConvertU16ToS32(vld1_u16(w + 8));
    const int32x4_t w3 = ConvertU16ToS32(vld1_u16(w + 12));
    const int64x1_t sum1 = DistoSum(vpass_a, w0, w1, w2, w3);
    const int64x1_t sum2 = DistoSum(vpass_b, w0, w1, w2, w3);
    const int32x2_t diff = vabd_s32(vreinterpret_s32_s64(sum1),
                                    vreinterpret_s32_s64(sum2));
    const int32x2_t res = vshr_n_s32(diff, 5);
    return vget_lane_s32(res, 0);
  }
}

#undef LOAD_LANE_32b

#else

// Hadamard transform
// Returns the weighted sum of the absolute value of transformed coefficients.
static int Disto4x4(const uint8_t* const a, const uint8_t* const b,
                    const uint16_t* const w) {
  const int kBPS = BPS;
  const uint8_t* A = a;
  const uint8_t* B = b;
  const uint16_t* W = w;
  int sum;
  __asm__ volatile (
    "vld1.32         d0[0], [%[a]], %[kBPS]   \n"
    "vld1.32         d0[1], [%[a]], %[kBPS]   \n"
    "vld1.32         d2[0], [%[a]], %[kBPS]   \n"
    "vld1.32         d2[1], [%[a]]            \n"

    "vld1.32         d1[0], [%[b]], %[kBPS]   \n"
    "vld1.32         d1[1], [%[b]], %[kBPS]   \n"
    "vld1.32         d3[0], [%[b]], %[kBPS]   \n"
    "vld1.32         d3[1], [%[b]]            \n"

    // a d0/d2, b d1/d3
    // d0/d1: 01 01 01 01
    // d2/d3: 23 23 23 23
    // But: it goes 01 45 23 67
    // Notice the middle values are transposed
    "vtrn.16         q0, q1                   \n"

    // {a0, a1} = {in[0] + in[2], in[1] + in[3]}
    "vaddl.u8        q2, d0, d2               \n"
    "vaddl.u8        q10, d1, d3              \n"
    // {a3, a2} = {in[0] - in[2], in[1] - in[3]}
    "vsubl.u8        q3, d0, d2               \n"
    "vsubl.u8        q11, d1, d3              \n"

    // tmp[0] = a0 + a1
    "vpaddl.s16      q0, q2                   \n"
    "vpaddl.s16      q8, q10                  \n"

    // tmp[1] = a3 + a2
    "vpaddl.s16      q1, q3                   \n"
    "vpaddl.s16      q9, q11                  \n"

    // No pair subtract
    // q2 = {a0, a3}
    // q3 = {a1, a2}
    "vtrn.16         q2, q3                   \n"
    "vtrn.16         q10, q11                 \n"

    // {tmp[3], tmp[2]} = {a0 - a1, a3 - a2}
    "vsubl.s16       q12, d4, d6              \n"
    "vsubl.s16       q13, d5, d7              \n"
    "vsubl.s16       q14, d20, d22            \n"
    "vsubl.s16       q15, d21, d23            \n"

    // separate tmp[3] and tmp[2]
    // q12 = tmp[3]
    // q13 = tmp[2]
    "vtrn.32         q12, q13                 \n"
    "vtrn.32         q14, q15                 \n"

    // Transpose tmp for a
    "vswp            d1, d26                  \n" // vtrn.64
    "vswp            d3, d24                  \n" // vtrn.64
    "vtrn.32         q0, q1                   \n"
    "vtrn.32         q13, q12                 \n"

    // Transpose tmp for b
    "vswp            d17, d30                 \n" // vtrn.64
    "vswp            d19, d28                 \n" // vtrn.64
    "vtrn.32         q8, q9                   \n"
    "vtrn.32         q15, q14                 \n"

    // The first Q register is a, the second b.
    // q0/8 tmp[0-3]
    // q13/15 tmp[4-7]
    // q1/9 tmp[8-11]
    // q12/14 tmp[12-15]

    // These are still in 01 45 23 67 order. We fix it easily in the addition
    // case but the subtraction propagates them.
    "vswp            d3, d27                  \n"
    "vswp            d19, d31                 \n"

    // a0 = tmp[0] + tmp[8]
    "vadd.s32        q2, q0, q1               \n"
    "vadd.s32        q3, q8, q9               \n"

    // a1 = tmp[4] + tmp[12]
    "vadd.s32        q10, q13, q12            \n"
    "vadd.s32        q11, q15, q14            \n"

    // a2 = tmp[4] - tmp[12]
    "vsub.s32        q13, q13, q12            \n"
    "vsub.s32        q15, q15, q14            \n"

    // a3 = tmp[0] - tmp[8]
    "vsub.s32        q0, q0, q1               \n"
    "vsub.s32        q8, q8, q9               \n"

    // b0 = a0 + a1
    "vadd.s32        q1, q2, q10              \n"
    "vadd.s32        q9, q3, q11              \n"

    // b1 = a3 + a2
    "vadd.s32        q12, q0, q13             \n"
    "vadd.s32        q14, q8, q15             \n"

    // b2 = a3 - a2
    "vsub.s32        q0, q0, q13              \n"
    "vsub.s32        q8, q8, q15              \n"

    // b3 = a0 - a1
    "vsub.s32        q2, q2, q10              \n"
    "vsub.s32        q3, q3, q11              \n"

    "vld1.64         {q10, q11}, [%[w]]       \n"

    // abs(b0)
    "vabs.s32        q1, q1                   \n"
    "vabs.s32        q9, q9                   \n"
    // abs(b1)
    "vabs.s32        q12, q12                 \n"
    "vabs.s32        q14, q14                 \n"
    // abs(b2)
    "vabs.s32        q0, q0                   \n"
    "vabs.s32        q8, q8                   \n"
    // abs(b3)
    "vabs.s32        q2, q2                   \n"
    "vabs.s32        q3, q3                   \n"

    // expand w before using.
    "vmovl.u16       q13, d20                 \n"
    "vmovl.u16       q15, d21                 \n"

    // w[0] * abs(b0)
    "vmul.u32        q1, q1, q13              \n"
    "vmul.u32        q9, q9, q13              \n"

    // w[4] * abs(b1)
    "vmla.u32        q1, q12, q15             \n"
    "vmla.u32        q9, q14, q15             \n"

    // expand w before using.
    "vmovl.u16       q13, d22                 \n"
    "vmovl.u16       q15, d23                 \n"

    // w[8] * abs(b1)
    "vmla.u32        q1, q0, q13              \n"
    "vmla.u32        q9, q8, q13              \n"

    // w[12] * abs(b1)
    "vmla.u32        q1, q2, q15              \n"
    "vmla.u32        q9, q3, q15              \n"

    // Sum the arrays
    "vpaddl.u32      q1, q1                   \n"
    "vpaddl.u32      q9, q9                   \n"
    "vadd.u64        d2, d3                   \n"
    "vadd.u64        d18, d19                 \n"

    // Hadamard transform needs 4 bits of extra precision (2 bits in each
    // direction) for dynamic raw. Weights w[] are 16bits at max, so the maximum
    // precision for coeff is 8bit of input + 4bits of Hadamard transform +
    // 16bits for w[] + 2 bits of abs() summation.
    //
    // This uses a maximum of 31 bits (signed). Discarding the top 32 bits is
    // A-OK.

    // sum2 - sum1
    "vsub.u32        d0, d2, d18              \n"
    // abs(sum2 - sum1)
    "vabs.s32        d0, d0                   \n"
    // abs(sum2 - sum1) >> 5
    "vshr.u32        d0, #5                   \n"

    // It would be better to move the value straight into r0 but I'm not
    // entirely sure how this works with inline assembly.
    "vmov.32         %[sum], d0[0]            \n"

    : [sum] "=r"(sum), [a] "+r"(A), [b] "+r"(B), [w] "+r"(W)
    : [kBPS] "r"(kBPS)
    : "memory", "q0", "q1", "q2", "q3", "q4", "q5", "q6", "q7", "q8", "q9",
      "q10", "q11", "q12", "q13", "q14", "q15"  // clobbered
  ) ;

  return sum;
}

#endif  // USE_INTRINSICS

static int Disto16x16(const uint8_t* const a, const uint8_t* const b,
                      const uint16_t* const w) {
  int D = 0;
  int x, y;
  for (y = 0; y < 16 * BPS; y += 4 * BPS) {
    for (x = 0; x < 16; x += 4) {
      D += Disto4x4(a + x + y, b + x + y, w);
    }
  }
  return D;
}

//------------------------------------------------------------------------------

static void CollectHistogram(const uint8_t* ref, const uint8_t* pred,
                             int start_block, int end_block,
                             VP8Histogram* const histo) {
  const uint16x8_t max_coeff_thresh = vdupq_n_u16(MAX_COEFF_THRESH);
  int j;
  for (j = start_block; j < end_block; ++j) {
    int16_t out[16];
    FTransform(ref + VP8DspScan[j], pred + VP8DspScan[j], out);
    {
      int k;
      const int16x8_t a0 = vld1q_s16(out + 0);
      const int16x8_t b0 = vld1q_s16(out + 8);
      const uint16x8_t a1 = vreinterpretq_u16_s16(vabsq_s16(a0));
      const uint16x8_t b1 = vreinterpretq_u16_s16(vabsq_s16(b0));
      const uint16x8_t a2 = vshrq_n_u16(a1, 3);
      const uint16x8_t b2 = vshrq_n_u16(b1, 3);
      const uint16x8_t a3 = vminq_u16(a2, max_coeff_thresh);
      const uint16x8_t b3 = vminq_u16(b2, max_coeff_thresh);
      vst1q_s16(out + 0, vreinterpretq_s16_u16(a3));
      vst1q_s16(out + 8, vreinterpretq_s16_u16(b3));
      // Convert coefficients to bin.
      for (k = 0; k < 16; ++k) {
        histo->distribution[out[k]]++;
      }
    }
  }
}

//------------------------------------------------------------------------------

static WEBP_INLINE void AccumulateSSE16(const uint8_t* const a,
                                        const uint8_t* const b,
                                        uint32x4_t* const sum) {
  const uint8x16_t a0 = vld1q_u8(a);
  const uint8x16_t b0 = vld1q_u8(b);
  const uint8x16_t abs_diff = vabdq_u8(a0, b0);
  uint16x8_t prod = vmull_u8(vget_low_u8(abs_diff), vget_low_u8(abs_diff));
  prod = vmlal_u8(prod, vget_high_u8(abs_diff), vget_high_u8(abs_diff));
  *sum = vpadalq_u16(*sum, prod);      // pair-wise add and accumulate
}

// Horizontal sum of all four uint32_t values in 'sum'.
static int SumToInt(uint32x4_t sum) {
  const uint64x2_t sum2 = vpaddlq_u32(sum);
  const uint64_t sum3 = vgetq_lane_u64(sum2, 0) + vgetq_lane_u64(sum2, 1);
  return (int)sum3;
}

static int SSE16x16(const uint8_t* a, const uint8_t* b) {
  uint32x4_t sum = vdupq_n_u32(0);
  int y;
  for (y = 0; y < 16; ++y) {
    AccumulateSSE16(a + y * BPS, b + y * BPS, &sum);
  }
  return SumToInt(sum);
}

static int SSE16x8(const uint8_t* a, const uint8_t* b) {
  uint32x4_t sum = vdupq_n_u32(0);
  int y;
  for (y = 0; y < 8; ++y) {
    AccumulateSSE16(a + y * BPS, b + y * BPS, &sum);
  }
  return SumToInt(sum);
}

static int SSE8x8(const uint8_t* a, const uint8_t* b) {
  uint32x4_t sum = vdupq_n_u32(0);
  int y;
  for (y = 0; y < 8; ++y) {
    const uint8x8_t a0 = vld1_u8(a + y * BPS);
    const uint8x8_t b0 = vld1_u8(b + y * BPS);
    const uint8x8_t abs_diff = vabd_u8(a0, b0);
    const uint16x8_t prod = vmull_u8(abs_diff, abs_diff);
    sum = vpadalq_u16(sum, prod);
  }
  return SumToInt(sum);
}

static int SSE4x4(const uint8_t* a, const uint8_t* b) {
  const uint8x16_t a0 = Load4x4(a);
  const uint8x16_t b0 = Load4x4(b);
  const uint8x16_t abs_diff = vabdq_u8(a0, b0);
  uint16x8_t prod = vmull_u8(vget_low_u8(abs_diff), vget_low_u8(abs_diff));
  prod = vmlal_u8(prod, vget_high_u8(abs_diff), vget_high_u8(abs_diff));
  return SumToInt(vpaddlq_u16(prod));
}

//------------------------------------------------------------------------------

// Compilation with gcc-4.6.x is problematic for now.
#if !defined(WORK_AROUND_GCC)

static int16x8_t Quantize(int16_t* const in,
                          const VP8Matrix* const mtx, int offset) {
  const uint16x8_t sharp = vld1q_u16(&mtx->sharpen_[offset]);
  const uint16x8_t q = vld1q_u16(&mtx->q_[offset]);
  const uint16x8_t iq = vld1q_u16(&mtx->iq_[offset]);
  const uint32x4_t bias0 = vld1q_u32(&mtx->bias_[offset + 0]);
  const uint32x4_t bias1 = vld1q_u32(&mtx->bias_[offset + 4]);

  const int16x8_t a = vld1q_s16(in + offset);                // in
  const uint16x8_t b = vreinterpretq_u16_s16(vabsq_s16(a));  // coeff = abs(in)
  const int16x8_t sign = vshrq_n_s16(a, 15);                 // sign
  const uint16x8_t c = vaddq_u16(b, sharp);                  // + sharpen
  const uint32x4_t m0 = vmull_u16(vget_low_u16(c), vget_low_u16(iq));
  const uint32x4_t m1 = vmull_u16(vget_high_u16(c), vget_high_u16(iq));
  const uint32x4_t m2 = vhaddq_u32(m0, bias0);
  const uint32x4_t m3 = vhaddq_u32(m1, bias1);     // (coeff * iQ + bias) >> 1
  const uint16x8_t c0 = vcombine_u16(vshrn_n_u32(m2, 16),
                                     vshrn_n_u32(m3, 16));   // QFIX=17 = 16+1
  const uint16x8_t c1 = vminq_u16(c0, vdupq_n_u16(MAX_LEVEL));
  const int16x8_t c2 = veorq_s16(vreinterpretq_s16_u16(c1), sign);
  const int16x8_t c3 = vsubq_s16(c2, sign);                  // restore sign
  const int16x8_t c4 = vmulq_s16(c3, vreinterpretq_s16_u16(q));
  vst1q_s16(in + offset, c4);
  assert(QFIX == 17);  // this function can't work as is if QFIX != 16+1
  return c3;
}

static const uint8_t kShuffles[4][8] = {
  { 0,   1,  2,  3,  8,  9, 16, 17 },
  { 10, 11,  4,  5,  6,  7, 12, 13 },
  { 18, 19, 24, 25, 26, 27, 20, 21 },
  { 14, 15, 22, 23, 28, 29, 30, 31 }
};

static int QuantizeBlock(int16_t in[16], int16_t out[16],
                         const VP8Matrix* const mtx) {
  const int16x8_t out0 = Quantize(in, mtx, 0);
  const int16x8_t out1 = Quantize(in, mtx, 8);
  uint8x8x4_t shuffles;
  // vtbl?_u8 are marked unavailable for iOS arm64 with Xcode < 6.3, use
  // non-standard versions there.
#if defined(__APPLE__) && defined(__aarch64__) && \
    defined(__apple_build_version__) && (__apple_build_version__< 6020037)
  uint8x16x2_t all_out;
  INIT_VECTOR2(all_out, vreinterpretq_u8_s16(out0), vreinterpretq_u8_s16(out1));
  INIT_VECTOR4(shuffles,
               vtbl2q_u8(all_out, vld1_u8(kShuffles[0])),
               vtbl2q_u8(all_out, vld1_u8(kShuffles[1])),
               vtbl2q_u8(all_out, vld1_u8(kShuffles[2])),
               vtbl2q_u8(all_out, vld1_u8(kShuffles[3])));
#else
  uint8x8x4_t all_out;
  INIT_VECTOR4(all_out,
               vreinterpret_u8_s16(vget_low_s16(out0)),
               vreinterpret_u8_s16(vget_high_s16(out0)),
               vreinterpret_u8_s16(vget_low_s16(out1)),
               vreinterpret_u8_s16(vget_high_s16(out1)));
  INIT_VECTOR4(shuffles,
               vtbl4_u8(all_out, vld1_u8(kShuffles[0])),
               vtbl4_u8(all_out, vld1_u8(kShuffles[1])),
               vtbl4_u8(all_out, vld1_u8(kShuffles[2])),
               vtbl4_u8(all_out, vld1_u8(kShuffles[3])));
#endif
  // Zigzag reordering
  vst1_u8((uint8_t*)(out +  0), shuffles.val[0]);
  vst1_u8((uint8_t*)(out +  4), shuffles.val[1]);
  vst1_u8((uint8_t*)(out +  8), shuffles.val[2]);
  vst1_u8((uint8_t*)(out + 12), shuffles.val[3]);
  // test zeros
  if (*(uint64_t*)(out +  0) != 0) return 1;
  if (*(uint64_t*)(out +  4) != 0) return 1;
  if (*(uint64_t*)(out +  8) != 0) return 1;
  if (*(uint64_t*)(out + 12) != 0) return 1;
  return 0;
}

#endif   // !WORK_AROUND_GCC

#endif   // WEBP_USE_NEON

//------------------------------------------------------------------------------
// Entry point

extern void VP8EncDspInitNEON(void);

void VP8EncDspInitNEON(void) {
#if defined(WEBP_USE_NEON)
  VP8ITransform = ITransform;
  VP8FTransform = FTransform;

  VP8FTransformWHT = FTransformWHT;

  VP8TDisto4x4 = Disto4x4;
  VP8TDisto16x16 = Disto16x16;
  VP8CollectHistogram = CollectHistogram;
  VP8SSE16x16 = SSE16x16;
  VP8SSE16x8 = SSE16x8;
  VP8SSE8x8 = SSE8x8;
  VP8SSE4x4 = SSE4x4;
#if !defined(WORK_AROUND_GCC)
  VP8EncQuantizeBlock = QuantizeBlock;
#endif
#endif   // WEBP_USE_NEON
}