tor-browser

The Tor Browser
git clone https://git.dasho.dev/tor-browser.git
Log | Files | Refs | README | LICENSE

dec_msa.c (41156B)


      1 // Copyright 2016 Google Inc. All Rights Reserved.
      2 //
      3 // Use of this source code is governed by a BSD-style license
      4 // that can be found in the COPYING file in the root of the source
      5 // tree. An additional intellectual property rights grant can be found
      6 // in the file PATENTS. All contributing project authors may
      7 // be found in the AUTHORS file in the root of the source tree.
      8 // -----------------------------------------------------------------------------
      9 //
     10 // MSA version of dsp functions
     11 //
     12 // Author(s):  Prashant Patil   (prashant.patil@imgtec.com)
     13 
     14 
     15 #include "src/dsp/dsp.h"
     16 
     17 #if defined(WEBP_USE_MSA)
     18 
     19 #include "src/dsp/msa_macro.h"
     20 
     21 //------------------------------------------------------------------------------
     22 // Transforms
     23 
     24 #define IDCT_1D_W(in0, in1, in2, in3, out0, out1, out2, out3) {  \
     25  v4i32 a1_m, b1_m, c1_m, d1_m;                                  \
     26  v4i32 c_tmp1_m, c_tmp2_m, d_tmp1_m, d_tmp2_m;                  \
     27  const v4i32 cospi8sqrt2minus1 = __msa_fill_w(20091);           \
     28  const v4i32 sinpi8sqrt2 = __msa_fill_w(35468);                 \
     29                                                                 \
     30  a1_m = in0 + in2;                                              \
     31  b1_m = in0 - in2;                                              \
     32  c_tmp1_m = (in1 * sinpi8sqrt2) >> 16;                          \
     33  c_tmp2_m = in3 + ((in3 * cospi8sqrt2minus1) >> 16);            \
     34  c1_m = c_tmp1_m - c_tmp2_m;                                    \
     35  d_tmp1_m = in1 + ((in1 * cospi8sqrt2minus1) >> 16);            \
     36  d_tmp2_m = (in3 * sinpi8sqrt2) >> 16;                          \
     37  d1_m = d_tmp1_m + d_tmp2_m;                                    \
     38  BUTTERFLY_4(a1_m, b1_m, c1_m, d1_m, out0, out1, out2, out3);   \
     39 }
     40 
     41 static void TransformOne(const int16_t* WEBP_RESTRICT in,
     42                         uint8_t* WEBP_RESTRICT dst) {
     43  v8i16 input0, input1;
     44  v4i32 in0, in1, in2, in3, hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3;
     45  v4i32 res0, res1, res2, res3;
     46  const v16i8 zero = { 0 };
     47  v16i8 dest0, dest1, dest2, dest3;
     48 
     49  LD_SH2(in, 8, input0, input1);
     50  UNPCK_SH_SW(input0, in0, in1);
     51  UNPCK_SH_SW(input1, in2, in3);
     52  IDCT_1D_W(in0, in1, in2, in3, hz0, hz1, hz2, hz3);
     53  TRANSPOSE4x4_SW_SW(hz0, hz1, hz2, hz3, hz0, hz1, hz2, hz3);
     54  IDCT_1D_W(hz0, hz1, hz2, hz3, vt0, vt1, vt2, vt3);
     55  SRARI_W4_SW(vt0, vt1, vt2, vt3, 3);
     56  TRANSPOSE4x4_SW_SW(vt0, vt1, vt2, vt3, vt0, vt1, vt2, vt3);
     57  LD_SB4(dst, BPS, dest0, dest1, dest2, dest3);
     58  ILVR_B4_SW(zero, dest0, zero, dest1, zero, dest2, zero, dest3,
     59             res0, res1, res2, res3);
     60  ILVR_H4_SW(zero, res0, zero, res1, zero, res2, zero, res3,
     61             res0, res1, res2, res3);
     62  ADD4(res0, vt0, res1, vt1, res2, vt2, res3, vt3, res0, res1, res2, res3);
     63  CLIP_SW4_0_255(res0, res1, res2, res3);
     64  PCKEV_B2_SW(res0, res1, res2, res3, vt0, vt1);
     65  res0 = (v4i32)__msa_pckev_b((v16i8)vt0, (v16i8)vt1);
     66  ST4x4_UB(res0, res0, 3, 2, 1, 0, dst, BPS);
     67 }
     68 
     69 static void TransformTwo(const int16_t* WEBP_RESTRICT in,
     70                         uint8_t* WEBP_RESTRICT dst, int do_two) {
     71  TransformOne(in, dst);
     72  if (do_two) {
     73    TransformOne(in + 16, dst + 4);
     74  }
     75 }
     76 
     77 static void TransformWHT(const int16_t* WEBP_RESTRICT in,
     78                         int16_t* WEBP_RESTRICT out) {
     79  v8i16 input0, input1;
     80  const v8i16 mask0 = { 0, 1, 2, 3, 8, 9, 10, 11 };
     81  const v8i16 mask1 = { 4, 5, 6, 7, 12, 13, 14, 15 };
     82  const v8i16 mask2 = { 0, 4, 8, 12, 1, 5, 9, 13 };
     83  const v8i16 mask3 = { 3, 7, 11, 15, 2, 6, 10, 14 };
     84  v8i16 tmp0, tmp1, tmp2, tmp3;
     85  v8i16 out0, out1;
     86 
     87  LD_SH2(in, 8, input0, input1);
     88  input1 = SLDI_SH(input1, input1, 8);
     89  tmp0 = input0 + input1;
     90  tmp1 = input0 - input1;
     91  VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3);
     92  out0 = tmp2 + tmp3;
     93  out1 = tmp2 - tmp3;
     94  VSHF_H2_SH(out0, out1, out0, out1, mask2, mask3, input0, input1);
     95  tmp0 = input0 + input1;
     96  tmp1 = input0 - input1;
     97  VSHF_H2_SH(tmp0, tmp1, tmp0, tmp1, mask0, mask1, tmp2, tmp3);
     98  tmp0 = tmp2 + tmp3;
     99  tmp1 = tmp2 - tmp3;
    100  ADDVI_H2_SH(tmp0, 3, tmp1, 3, out0, out1);
    101  SRAI_H2_SH(out0, out1, 3);
    102  out[0] = __msa_copy_s_h(out0, 0);
    103  out[16] = __msa_copy_s_h(out0, 4);
    104  out[32] = __msa_copy_s_h(out1, 0);
    105  out[48] = __msa_copy_s_h(out1, 4);
    106  out[64] = __msa_copy_s_h(out0, 1);
    107  out[80] = __msa_copy_s_h(out0, 5);
    108  out[96] = __msa_copy_s_h(out1, 1);
    109  out[112] = __msa_copy_s_h(out1, 5);
    110  out[128] = __msa_copy_s_h(out0, 2);
    111  out[144] = __msa_copy_s_h(out0, 6);
    112  out[160] = __msa_copy_s_h(out1, 2);
    113  out[176] = __msa_copy_s_h(out1, 6);
    114  out[192] = __msa_copy_s_h(out0, 3);
    115  out[208] = __msa_copy_s_h(out0, 7);
    116  out[224] = __msa_copy_s_h(out1, 3);
    117  out[240] = __msa_copy_s_h(out1, 7);
    118 }
    119 
    120 static void TransformDC(const int16_t* WEBP_RESTRICT in,
    121                        uint8_t* WEBP_RESTRICT dst) {
    122  const int DC = (in[0] + 4) >> 3;
    123  const v8i16 tmp0 = __msa_fill_h(DC);
    124  ADDBLK_ST4x4_UB(tmp0, tmp0, tmp0, tmp0, dst, BPS);
    125 }
    126 
    127 static void TransformAC3(const int16_t* WEBP_RESTRICT in,
    128                         uint8_t* WEBP_RESTRICT dst) {
    129  const int a = in[0] + 4;
    130  const int c4 = WEBP_TRANSFORM_AC3_MUL2(in[4]);
    131  const int d4 = WEBP_TRANSFORM_AC3_MUL1(in[4]);
    132  const int in2 = WEBP_TRANSFORM_AC3_MUL2(in[1]);
    133  const int in3 = WEBP_TRANSFORM_AC3_MUL1(in[1]);
    134  v4i32 tmp0 = { 0 };
    135  v4i32 out0 = __msa_fill_w(a + d4);
    136  v4i32 out1 = __msa_fill_w(a + c4);
    137  v4i32 out2 = __msa_fill_w(a - c4);
    138  v4i32 out3 = __msa_fill_w(a - d4);
    139  v4i32 res0, res1, res2, res3;
    140  const v4i32 zero = { 0 };
    141  v16u8 dest0, dest1, dest2, dest3;
    142 
    143  INSERT_W4_SW(in3, in2, -in2, -in3, tmp0);
    144  ADD4(out0, tmp0, out1, tmp0, out2, tmp0, out3, tmp0,
    145       out0, out1, out2, out3);
    146  SRAI_W4_SW(out0, out1, out2, out3, 3);
    147  LD_UB4(dst, BPS, dest0, dest1, dest2, dest3);
    148  ILVR_B4_SW(zero, dest0, zero, dest1, zero, dest2, zero, dest3,
    149             res0, res1, res2, res3);
    150  ILVR_H4_SW(zero, res0, zero, res1, zero, res2, zero, res3,
    151             res0, res1, res2, res3);
    152  ADD4(res0, out0, res1, out1, res2, out2, res3, out3, res0, res1, res2, res3);
    153  CLIP_SW4_0_255(res0, res1, res2, res3);
    154  PCKEV_B2_SW(res0, res1, res2, res3, out0, out1);
    155  res0 = (v4i32)__msa_pckev_b((v16i8)out0, (v16i8)out1);
    156  ST4x4_UB(res0, res0, 3, 2, 1, 0, dst, BPS);
    157 }
    158 
    159 //------------------------------------------------------------------------------
    160 // Edge filtering functions
    161 
    162 #define FLIP_SIGN2(in0, in1, out0, out1) {  \
    163  out0 = (v16i8)__msa_xori_b(in0, 0x80);    \
    164  out1 = (v16i8)__msa_xori_b(in1, 0x80);    \
    165 }
    166 
    167 #define FLIP_SIGN4(in0, in1, in2, in3, out0, out1, out2, out3) {  \
    168  FLIP_SIGN2(in0, in1, out0, out1);                               \
    169  FLIP_SIGN2(in2, in3, out2, out3);                               \
    170 }
    171 
    172 #define FILT_VAL(q0_m, p0_m, mask, filt) do {  \
    173  v16i8 q0_sub_p0;                             \
    174  q0_sub_p0 = __msa_subs_s_b(q0_m, p0_m);      \
    175  filt = __msa_adds_s_b(filt, q0_sub_p0);      \
    176  filt = __msa_adds_s_b(filt, q0_sub_p0);      \
    177  filt = __msa_adds_s_b(filt, q0_sub_p0);      \
    178  filt = filt & mask;                          \
    179 } while (0)
    180 
    181 #define FILT2(q_m, p_m, q, p) do {            \
    182  u_r = SRAI_H(temp1, 7);                     \
    183  u_r = __msa_sat_s_h(u_r, 7);                \
    184  u_l = SRAI_H(temp3, 7);                     \
    185  u_l = __msa_sat_s_h(u_l, 7);                \
    186  u = __msa_pckev_b((v16i8)u_l, (v16i8)u_r);  \
    187  q_m = __msa_subs_s_b(q_m, u);               \
    188  p_m = __msa_adds_s_b(p_m, u);               \
    189  q = __msa_xori_b((v16u8)q_m, 0x80);         \
    190  p = __msa_xori_b((v16u8)p_m, 0x80);         \
    191 } while (0)
    192 
    193 #define LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev) do {  \
    194  v16i8 p1_m, p0_m, q0_m, q1_m;                         \
    195  v16i8 filt, t1, t2;                                   \
    196  const v16i8 cnst4b = __msa_ldi_b(4);                  \
    197  const v16i8 cnst3b = __msa_ldi_b(3);                  \
    198                                                        \
    199  FLIP_SIGN4(p1, p0, q0, q1, p1_m, p0_m, q0_m, q1_m);   \
    200  filt = __msa_subs_s_b(p1_m, q1_m);                    \
    201  filt = filt & hev;                                    \
    202  FILT_VAL(q0_m, p0_m, mask, filt);                     \
    203  t1 = __msa_adds_s_b(filt, cnst4b);                    \
    204  t1 = SRAI_B(t1, 3);                                   \
    205  t2 = __msa_adds_s_b(filt, cnst3b);                    \
    206  t2 = SRAI_B(t2, 3);                                   \
    207  q0_m = __msa_subs_s_b(q0_m, t1);                      \
    208  q0 = __msa_xori_b((v16u8)q0_m, 0x80);                 \
    209  p0_m = __msa_adds_s_b(p0_m, t2);                      \
    210  p0 = __msa_xori_b((v16u8)p0_m, 0x80);                 \
    211  filt = __msa_srari_b(t1, 1);                          \
    212  hev = __msa_xori_b(hev, 0xff);                        \
    213  filt = filt & hev;                                    \
    214  q1_m = __msa_subs_s_b(q1_m, filt);                    \
    215  q1 = __msa_xori_b((v16u8)q1_m, 0x80);                 \
    216  p1_m = __msa_adds_s_b(p1_m, filt);                    \
    217  p1 = __msa_xori_b((v16u8)p1_m, 0x80);                 \
    218 } while (0)
    219 
    220 #define LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev) do {  \
    221  v16i8 p2_m, p1_m, p0_m, q2_m, q1_m, q0_m;                   \
    222  v16i8 u, filt, t1, t2, filt_sign;                           \
    223  v8i16 filt_r, filt_l, u_r, u_l;                             \
    224  v8i16 temp0, temp1, temp2, temp3;                           \
    225  const v16i8 cnst4b = __msa_ldi_b(4);                        \
    226  const v16i8 cnst3b = __msa_ldi_b(3);                        \
    227  const v8i16 cnst9h = __msa_ldi_h(9);                        \
    228  const v8i16 cnst63h = __msa_ldi_h(63);                      \
    229                                                              \
    230  FLIP_SIGN4(p1, p0, q0, q1, p1_m, p0_m, q0_m, q1_m);         \
    231  filt = __msa_subs_s_b(p1_m, q1_m);                          \
    232  FILT_VAL(q0_m, p0_m, mask, filt);                           \
    233  FLIP_SIGN2(p2, q2, p2_m, q2_m);                             \
    234  t2 = filt & hev;                                            \
    235  /* filt_val &= ~hev */                                      \
    236  hev = __msa_xori_b(hev, 0xff);                              \
    237  filt = filt & hev;                                          \
    238  t1 = __msa_adds_s_b(t2, cnst4b);                            \
    239  t1 = SRAI_B(t1, 3);                                         \
    240  t2 = __msa_adds_s_b(t2, cnst3b);                            \
    241  t2 = SRAI_B(t2, 3);                                         \
    242  q0_m = __msa_subs_s_b(q0_m, t1);                            \
    243  p0_m = __msa_adds_s_b(p0_m, t2);                            \
    244  filt_sign = __msa_clti_s_b(filt, 0);                        \
    245  ILVRL_B2_SH(filt_sign, filt, filt_r, filt_l);               \
    246  /* update q2/p2 */                                          \
    247  temp0 = filt_r * cnst9h;                                    \
    248  temp1 = temp0 + cnst63h;                                    \
    249  temp2 = filt_l * cnst9h;                                    \
    250  temp3 = temp2 + cnst63h;                                    \
    251  FILT2(q2_m, p2_m, q2, p2);                                  \
    252  /* update q1/p1 */                                          \
    253  temp1 = temp1 + temp0;                                      \
    254  temp3 = temp3 + temp2;                                      \
    255  FILT2(q1_m, p1_m, q1, p1);                                  \
    256  /* update q0/p0 */                                          \
    257  temp1 = temp1 + temp0;                                      \
    258  temp3 = temp3 + temp2;                                      \
    259  FILT2(q0_m, p0_m, q0, p0);                                  \
    260 } while (0)
    261 
    262 #define LPF_MASK_HEV(p3_in, p2_in, p1_in, p0_in,                 \
    263                     q0_in, q1_in, q2_in, q3_in,                 \
    264                     limit_in, b_limit_in, thresh_in,            \
    265                     hev_out, mask_out) do {                     \
    266  v16u8 p3_asub_p2_m, p2_asub_p1_m, p1_asub_p0_m, q1_asub_q0_m;  \
    267  v16u8 p1_asub_q1_m, p0_asub_q0_m, q3_asub_q2_m, q2_asub_q1_m;  \
    268  v16u8 flat_out;                                                \
    269                                                                 \
    270  /* absolute subtraction of pixel values */                     \
    271  p3_asub_p2_m = __msa_asub_u_b(p3_in, p2_in);                   \
    272  p2_asub_p1_m = __msa_asub_u_b(p2_in, p1_in);                   \
    273  p1_asub_p0_m = __msa_asub_u_b(p1_in, p0_in);                   \
    274  q1_asub_q0_m = __msa_asub_u_b(q1_in, q0_in);                   \
    275  q2_asub_q1_m = __msa_asub_u_b(q2_in, q1_in);                   \
    276  q3_asub_q2_m = __msa_asub_u_b(q3_in, q2_in);                   \
    277  p0_asub_q0_m = __msa_asub_u_b(p0_in, q0_in);                   \
    278  p1_asub_q1_m = __msa_asub_u_b(p1_in, q1_in);                   \
    279  /* calculation of hev */                                       \
    280  flat_out = __msa_max_u_b(p1_asub_p0_m, q1_asub_q0_m);          \
    281  hev_out = (thresh_in < flat_out);                              \
    282  /* calculation of mask */                                      \
    283  p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p0_asub_q0_m);     \
    284  p1_asub_q1_m = SRAI_B(p1_asub_q1_m, 1);                        \
    285  p0_asub_q0_m = __msa_adds_u_b(p0_asub_q0_m, p1_asub_q1_m);     \
    286  mask_out = (b_limit_in < p0_asub_q0_m);                        \
    287  mask_out = __msa_max_u_b(flat_out, mask_out);                  \
    288  p3_asub_p2_m = __msa_max_u_b(p3_asub_p2_m, p2_asub_p1_m);      \
    289  mask_out = __msa_max_u_b(p3_asub_p2_m, mask_out);              \
    290  q2_asub_q1_m = __msa_max_u_b(q2_asub_q1_m, q3_asub_q2_m);      \
    291  mask_out = __msa_max_u_b(q2_asub_q1_m, mask_out);              \
    292  mask_out = (limit_in < mask_out);                              \
    293  mask_out = __msa_xori_b(mask_out, 0xff);                       \
    294 } while (0)
    295 
    296 #define ST6x1_UB(in0, in0_idx, in1, in1_idx, pdst, stride) do { \
    297  const uint16_t tmp0_h = __msa_copy_s_h((v8i16)in1, in1_idx);  \
    298  const uint32_t tmp0_w = __msa_copy_s_w((v4i32)in0, in0_idx);  \
    299  SW(tmp0_w, pdst);                                             \
    300  SH(tmp0_h, pdst + stride);                                    \
    301 } while (0)
    302 
    303 #define ST6x4_UB(in0, start_in0_idx, in1, start_in1_idx, pdst, stride) do { \
    304  uint8_t* ptmp1 = (uint8_t*)pdst;                                          \
    305  ST6x1_UB(in0, start_in0_idx, in1, start_in1_idx, ptmp1, 4);               \
    306  ptmp1 += stride;                                                          \
    307  ST6x1_UB(in0, start_in0_idx + 1, in1, start_in1_idx + 1, ptmp1, 4);       \
    308  ptmp1 += stride;                                                          \
    309  ST6x1_UB(in0, start_in0_idx + 2, in1, start_in1_idx + 2, ptmp1, 4);       \
    310  ptmp1 += stride;                                                          \
    311  ST6x1_UB(in0, start_in0_idx + 3, in1, start_in1_idx + 3, ptmp1, 4);       \
    312 } while (0)
    313 
    314 #define LPF_SIMPLE_FILT(p1_in, p0_in, q0_in, q1_in, mask) do {       \
    315    v16i8 p1_m, p0_m, q0_m, q1_m, filt, filt1, filt2;                \
    316    const v16i8 cnst4b = __msa_ldi_b(4);                             \
    317    const v16i8 cnst3b =  __msa_ldi_b(3);                            \
    318                                                                     \
    319    FLIP_SIGN4(p1_in, p0_in, q0_in, q1_in, p1_m, p0_m, q0_m, q1_m);  \
    320    filt = __msa_subs_s_b(p1_m, q1_m);                               \
    321    FILT_VAL(q0_m, p0_m, mask, filt);                                \
    322    filt1 = __msa_adds_s_b(filt, cnst4b);                            \
    323    filt1 = SRAI_B(filt1, 3);                                        \
    324    filt2 = __msa_adds_s_b(filt, cnst3b);                            \
    325    filt2 = SRAI_B(filt2, 3);                                        \
    326    q0_m = __msa_subs_s_b(q0_m, filt1);                              \
    327    p0_m = __msa_adds_s_b(p0_m, filt2);                              \
    328    q0_in = __msa_xori_b((v16u8)q0_m, 0x80);                         \
    329    p0_in = __msa_xori_b((v16u8)p0_m, 0x80);                         \
    330 } while (0)
    331 
    332 #define LPF_SIMPLE_MASK(p1, p0, q0, q1, b_limit, mask) do {    \
    333    v16u8 p1_a_sub_q1, p0_a_sub_q0;                            \
    334                                                               \
    335    p0_a_sub_q0 = __msa_asub_u_b(p0, q0);                      \
    336    p1_a_sub_q1 = __msa_asub_u_b(p1, q1);                      \
    337    p1_a_sub_q1 = (v16u8)__msa_srli_b((v16i8)p1_a_sub_q1, 1);  \
    338    p0_a_sub_q0 = __msa_adds_u_b(p0_a_sub_q0, p0_a_sub_q0);    \
    339    mask = __msa_adds_u_b(p0_a_sub_q0, p1_a_sub_q1);           \
    340    mask = (mask <= b_limit);                                  \
    341 } while (0)
    342 
    343 static void VFilter16(uint8_t* src, int stride,
    344                      int b_limit_in, int limit_in, int thresh_in) {
    345  uint8_t* ptemp = src - 4 * stride;
    346  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
    347  v16u8 mask, hev;
    348  const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in);
    349  const v16u8 limit = (v16u8)__msa_fill_b(limit_in);
    350  const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
    351 
    352  LD_UB8(ptemp, stride, p3, p2, p1, p0, q0, q1, q2, q3);
    353  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
    354               hev, mask);
    355  LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev);
    356  ptemp = src - 3 * stride;
    357  ST_UB4(p2, p1, p0, q0, ptemp, stride);
    358  ptemp += (4 * stride);
    359  ST_UB2(q1, q2, ptemp, stride);
    360 }
    361 
    362 static void HFilter16(uint8_t* src, int stride,
    363                      int b_limit_in, int limit_in, int thresh_in) {
    364  uint8_t* ptmp  = src - 4;
    365  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
    366  v16u8 mask, hev;
    367  v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8;
    368  v16u8 row9, row10, row11, row12, row13, row14, row15;
    369  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
    370  const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
    371  const v16u8 limit = (v16u8)__msa_fill_b(limit_in);
    372  const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in);
    373 
    374  LD_UB8(ptmp, stride, row0, row1, row2, row3, row4, row5, row6, row7);
    375  ptmp += (8 * stride);
    376  LD_UB8(ptmp, stride, row8, row9, row10, row11, row12, row13, row14, row15);
    377  TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
    378                      row8, row9, row10, row11, row12, row13, row14, row15,
    379                      p3, p2, p1, p0, q0, q1, q2, q3);
    380  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
    381               hev, mask);
    382  LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev);
    383  ILVR_B2_SH(p1, p2, q0, p0, tmp0, tmp1);
    384  ILVRL_H2_SH(tmp1, tmp0, tmp3, tmp4);
    385  ILVL_B2_SH(p1, p2, q0, p0, tmp0, tmp1);
    386  ILVRL_H2_SH(tmp1, tmp0, tmp6, tmp7);
    387  ILVRL_B2_SH(q2, q1, tmp2, tmp5);
    388  ptmp = src - 3;
    389  ST6x1_UB(tmp3, 0, tmp2, 0, ptmp, 4);
    390  ptmp += stride;
    391  ST6x1_UB(tmp3, 1, tmp2, 1, ptmp, 4);
    392  ptmp += stride;
    393  ST6x1_UB(tmp3, 2, tmp2, 2, ptmp, 4);
    394  ptmp += stride;
    395  ST6x1_UB(tmp3, 3, tmp2, 3, ptmp, 4);
    396  ptmp += stride;
    397  ST6x1_UB(tmp4, 0, tmp2, 4, ptmp, 4);
    398  ptmp += stride;
    399  ST6x1_UB(tmp4, 1, tmp2, 5, ptmp, 4);
    400  ptmp += stride;
    401  ST6x1_UB(tmp4, 2, tmp2, 6, ptmp, 4);
    402  ptmp += stride;
    403  ST6x1_UB(tmp4, 3, tmp2, 7, ptmp, 4);
    404  ptmp += stride;
    405  ST6x1_UB(tmp6, 0, tmp5, 0, ptmp, 4);
    406  ptmp += stride;
    407  ST6x1_UB(tmp6, 1, tmp5, 1, ptmp, 4);
    408  ptmp += stride;
    409  ST6x1_UB(tmp6, 2, tmp5, 2, ptmp, 4);
    410  ptmp += stride;
    411  ST6x1_UB(tmp6, 3, tmp5, 3, ptmp, 4);
    412  ptmp += stride;
    413  ST6x1_UB(tmp7, 0, tmp5, 4, ptmp, 4);
    414  ptmp += stride;
    415  ST6x1_UB(tmp7, 1, tmp5, 5, ptmp, 4);
    416  ptmp += stride;
    417  ST6x1_UB(tmp7, 2, tmp5, 6, ptmp, 4);
    418  ptmp += stride;
    419  ST6x1_UB(tmp7, 3, tmp5, 7, ptmp, 4);
    420 }
    421 
    422 // on three inner edges
    423 static void VFilterHorEdge16i(uint8_t* src, int stride,
    424                              int b_limit, int limit, int thresh) {
    425  v16u8 mask, hev;
    426  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
    427  const v16u8 thresh0 = (v16u8)__msa_fill_b(thresh);
    428  const v16u8 b_limit0 = (v16u8)__msa_fill_b(b_limit);
    429  const v16u8 limit0 = (v16u8)__msa_fill_b(limit);
    430 
    431  LD_UB8((src - 4 * stride), stride, p3, p2, p1, p0, q0, q1, q2, q3);
    432  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
    433               hev, mask);
    434  LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev);
    435  ST_UB4(p1, p0, q0, q1, (src - 2 * stride), stride);
    436 }
    437 
    438 static void VFilter16i(uint8_t* src_y, int stride,
    439                       int b_limit, int limit, int thresh) {
    440  VFilterHorEdge16i(src_y +  4 * stride, stride, b_limit, limit, thresh);
    441  VFilterHorEdge16i(src_y +  8 * stride, stride, b_limit, limit, thresh);
    442  VFilterHorEdge16i(src_y + 12 * stride, stride, b_limit, limit, thresh);
    443 }
    444 
    445 static void HFilterVertEdge16i(uint8_t* src, int stride,
    446                               int b_limit, int limit, int thresh) {
    447  v16u8 mask, hev;
    448  v16u8 p3, p2, p1, p0, q3, q2, q1, q0;
    449  v16u8 row0, row1, row2, row3, row4, row5, row6, row7;
    450  v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
    451  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
    452  const v16u8 thresh0 = (v16u8)__msa_fill_b(thresh);
    453  const v16u8 b_limit0 = (v16u8)__msa_fill_b(b_limit);
    454  const v16u8 limit0 = (v16u8)__msa_fill_b(limit);
    455 
    456  LD_UB8(src - 4, stride, row0, row1, row2, row3, row4, row5, row6, row7);
    457  LD_UB8(src - 4 + (8 * stride), stride,
    458         row8, row9, row10, row11, row12, row13, row14, row15);
    459  TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
    460                      row8, row9, row10, row11, row12, row13, row14, row15,
    461                      p3, p2, p1, p0, q0, q1, q2, q3);
    462  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit0, b_limit0, thresh0,
    463               hev, mask);
    464  LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev);
    465  ILVR_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
    466  ILVRL_H2_SH(tmp1, tmp0, tmp2, tmp3);
    467  ILVL_B2_SH(p0, p1, q1, q0, tmp0, tmp1);
    468  ILVRL_H2_SH(tmp1, tmp0, tmp4, tmp5);
    469  src -= 2;
    470  ST4x8_UB(tmp2, tmp3, src, stride);
    471  src += (8 * stride);
    472  ST4x8_UB(tmp4, tmp5, src, stride);
    473 }
    474 
    475 static void HFilter16i(uint8_t* src_y, int stride,
    476                       int b_limit, int limit, int thresh) {
    477  HFilterVertEdge16i(src_y +  4, stride, b_limit, limit, thresh);
    478  HFilterVertEdge16i(src_y +  8, stride, b_limit, limit, thresh);
    479  HFilterVertEdge16i(src_y + 12, stride, b_limit, limit, thresh);
    480 }
    481 
    482 // 8-pixels wide variants, for chroma filtering
    483 static void VFilter8(uint8_t* WEBP_RESTRICT src_u, uint8_t* WEBP_RESTRICT src_v,
    484                     int stride, int b_limit_in, int limit_in, int thresh_in) {
    485  uint8_t* ptmp_src_u = src_u - 4 * stride;
    486  uint8_t* ptmp_src_v = src_v - 4 * stride;
    487  uint64_t p2_d, p1_d, p0_d, q0_d, q1_d, q2_d;
    488  v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev;
    489  v16u8 p3_u, p2_u, p1_u, p0_u, q3_u, q2_u, q1_u, q0_u;
    490  v16u8 p3_v, p2_v, p1_v, p0_v, q3_v, q2_v, q1_v, q0_v;
    491  const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
    492  const v16u8 limit = (v16u8)__msa_fill_b(limit_in);
    493  const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in);
    494 
    495  LD_UB8(ptmp_src_u, stride, p3_u, p2_u, p1_u, p0_u, q0_u, q1_u, q2_u, q3_u);
    496  LD_UB8(ptmp_src_v, stride, p3_v, p2_v, p1_v, p0_v, q0_v, q1_v, q2_v, q3_v);
    497  ILVR_D4_UB(p3_v, p3_u, p2_v, p2_u, p1_v, p1_u, p0_v, p0_u, p3, p2, p1, p0);
    498  ILVR_D4_UB(q0_v, q0_u, q1_v, q1_u, q2_v, q2_u, q3_v, q3_u, q0, q1, q2, q3);
    499  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
    500               hev, mask);
    501  LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev);
    502  p2_d = __msa_copy_s_d((v2i64)p2, 0);
    503  p1_d = __msa_copy_s_d((v2i64)p1, 0);
    504  p0_d = __msa_copy_s_d((v2i64)p0, 0);
    505  q0_d = __msa_copy_s_d((v2i64)q0, 0);
    506  q1_d = __msa_copy_s_d((v2i64)q1, 0);
    507  q2_d = __msa_copy_s_d((v2i64)q2, 0);
    508  ptmp_src_u += stride;
    509  SD4(p2_d, p1_d, p0_d, q0_d, ptmp_src_u, stride);
    510  ptmp_src_u += (4 * stride);
    511  SD(q1_d, ptmp_src_u);
    512  ptmp_src_u += stride;
    513  SD(q2_d, ptmp_src_u);
    514  p2_d = __msa_copy_s_d((v2i64)p2, 1);
    515  p1_d = __msa_copy_s_d((v2i64)p1, 1);
    516  p0_d = __msa_copy_s_d((v2i64)p0, 1);
    517  q0_d = __msa_copy_s_d((v2i64)q0, 1);
    518  q1_d = __msa_copy_s_d((v2i64)q1, 1);
    519  q2_d = __msa_copy_s_d((v2i64)q2, 1);
    520  ptmp_src_v += stride;
    521  SD4(p2_d, p1_d, p0_d, q0_d, ptmp_src_v, stride);
    522  ptmp_src_v += (4 * stride);
    523  SD(q1_d, ptmp_src_v);
    524  ptmp_src_v += stride;
    525  SD(q2_d, ptmp_src_v);
    526 }
    527 
    528 static void HFilter8(uint8_t* WEBP_RESTRICT src_u, uint8_t* WEBP_RESTRICT src_v,
    529                     int stride, int b_limit_in, int limit_in, int thresh_in) {
    530  uint8_t* ptmp_src_u = src_u - 4;
    531  uint8_t* ptmp_src_v = src_v - 4;
    532  v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev;
    533  v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8;
    534  v16u8 row9, row10, row11, row12, row13, row14, row15;
    535  v8i16 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5, tmp6, tmp7;
    536  const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
    537  const v16u8 limit = (v16u8)__msa_fill_b(limit_in);
    538  const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in);
    539 
    540  LD_UB8(ptmp_src_u, stride, row0, row1, row2, row3, row4, row5, row6, row7);
    541  LD_UB8(ptmp_src_v, stride,
    542         row8, row9, row10, row11, row12, row13, row14, row15);
    543  TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
    544                      row8, row9, row10, row11, row12, row13, row14, row15,
    545                      p3, p2, p1, p0, q0, q1, q2, q3);
    546  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
    547               hev, mask);
    548  LPF_MBFILTER(p2, p1, p0, q0, q1, q2, mask, hev);
    549  ILVR_B2_SH(p1, p2, q0, p0, tmp0, tmp1);
    550  ILVRL_H2_SH(tmp1, tmp0, tmp3, tmp4);
    551  ILVL_B2_SH(p1, p2, q0, p0, tmp0, tmp1);
    552  ILVRL_H2_SH(tmp1, tmp0, tmp6, tmp7);
    553  ILVRL_B2_SH(q2, q1, tmp2, tmp5);
    554  ptmp_src_u += 1;
    555  ST6x4_UB(tmp3, 0, tmp2, 0, ptmp_src_u, stride);
    556  ptmp_src_u += 4 * stride;
    557  ST6x4_UB(tmp4, 0, tmp2, 4, ptmp_src_u, stride);
    558  ptmp_src_v += 1;
    559  ST6x4_UB(tmp6, 0, tmp5, 0, ptmp_src_v, stride);
    560  ptmp_src_v += 4 * stride;
    561  ST6x4_UB(tmp7, 0, tmp5, 4, ptmp_src_v, stride);
    562 }
    563 
    564 static void VFilter8i(uint8_t* WEBP_RESTRICT src_u,
    565                      uint8_t* WEBP_RESTRICT src_v, int stride,
    566                      int b_limit_in, int limit_in, int thresh_in) {
    567  uint64_t p1_d, p0_d, q0_d, q1_d;
    568  v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev;
    569  v16u8 p3_u, p2_u, p1_u, p0_u, q3_u, q2_u, q1_u, q0_u;
    570  v16u8 p3_v, p2_v, p1_v, p0_v, q3_v, q2_v, q1_v, q0_v;
    571  const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in);
    572  const v16u8 limit = (v16u8)__msa_fill_b(limit_in);
    573  const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
    574 
    575  LD_UB8(src_u, stride, p3_u, p2_u, p1_u, p0_u, q0_u, q1_u, q2_u, q3_u);
    576  src_u += (5 * stride);
    577  LD_UB8(src_v, stride, p3_v, p2_v, p1_v, p0_v, q0_v, q1_v, q2_v, q3_v);
    578  src_v += (5 * stride);
    579  ILVR_D4_UB(p3_v, p3_u, p2_v, p2_u, p1_v, p1_u, p0_v, p0_u, p3, p2, p1, p0);
    580  ILVR_D4_UB(q0_v, q0_u, q1_v, q1_u, q2_v, q2_u, q3_v, q3_u, q0, q1, q2, q3);
    581  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
    582               hev, mask);
    583  LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev);
    584  p1_d = __msa_copy_s_d((v2i64)p1, 0);
    585  p0_d = __msa_copy_s_d((v2i64)p0, 0);
    586  q0_d = __msa_copy_s_d((v2i64)q0, 0);
    587  q1_d = __msa_copy_s_d((v2i64)q1, 0);
    588  SD4(q1_d, q0_d, p0_d, p1_d, src_u, -stride);
    589  p1_d = __msa_copy_s_d((v2i64)p1, 1);
    590  p0_d = __msa_copy_s_d((v2i64)p0, 1);
    591  q0_d = __msa_copy_s_d((v2i64)q0, 1);
    592  q1_d = __msa_copy_s_d((v2i64)q1, 1);
    593  SD4(q1_d, q0_d, p0_d, p1_d, src_v, -stride);
    594 }
    595 
    596 static void HFilter8i(uint8_t* WEBP_RESTRICT src_u,
    597                      uint8_t* WEBP_RESTRICT src_v, int stride,
    598                      int b_limit_in, int limit_in, int thresh_in) {
    599  v16u8 p3, p2, p1, p0, q3, q2, q1, q0, mask, hev;
    600  v16u8 row0, row1, row2, row3, row4, row5, row6, row7, row8;
    601  v16u8 row9, row10, row11, row12, row13, row14, row15;
    602  v4i32 tmp0, tmp1, tmp2, tmp3, tmp4, tmp5;
    603  const v16u8 thresh = (v16u8)__msa_fill_b(thresh_in);
    604  const v16u8 limit = (v16u8)__msa_fill_b(limit_in);
    605  const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
    606 
    607  LD_UB8(src_u, stride, row0, row1, row2, row3, row4, row5, row6, row7);
    608  LD_UB8(src_v, stride,
    609         row8, row9, row10, row11, row12, row13, row14, row15);
    610  TRANSPOSE16x8_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
    611                      row8, row9, row10, row11, row12, row13, row14, row15,
    612                      p3, p2, p1, p0, q0, q1, q2, q3);
    613  LPF_MASK_HEV(p3, p2, p1, p0, q0, q1, q2, q3, limit, b_limit, thresh,
    614               hev, mask);
    615  LPF_FILTER4_4W(p1, p0, q0, q1, mask, hev);
    616  ILVR_B2_SW(p0, p1, q1, q0, tmp0, tmp1);
    617  ILVRL_H2_SW(tmp1, tmp0, tmp2, tmp3);
    618  ILVL_B2_SW(p0, p1, q1, q0, tmp0, tmp1);
    619  ILVRL_H2_SW(tmp1, tmp0, tmp4, tmp5);
    620  src_u += 2;
    621  ST4x4_UB(tmp2, tmp2, 0, 1, 2, 3, src_u, stride);
    622  src_u += 4 * stride;
    623  ST4x4_UB(tmp3, tmp3, 0, 1, 2, 3, src_u, stride);
    624  src_v += 2;
    625  ST4x4_UB(tmp4, tmp4, 0, 1, 2, 3, src_v, stride);
    626  src_v += 4 * stride;
    627  ST4x4_UB(tmp5, tmp5, 0, 1, 2, 3, src_v, stride);
    628 }
    629 
    630 static void SimpleVFilter16(uint8_t* src, int stride, int b_limit_in) {
    631  v16u8 p1, p0, q1, q0, mask;
    632  const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
    633 
    634  LD_UB4(src - 2 * stride, stride, p1, p0, q0, q1);
    635  LPF_SIMPLE_MASK(p1, p0, q0, q1, b_limit, mask);
    636  LPF_SIMPLE_FILT(p1, p0, q0, q1, mask);
    637  ST_UB2(p0, q0, src - stride, stride);
    638 }
    639 
    640 static void SimpleHFilter16(uint8_t* src, int stride, int b_limit_in) {
    641  v16u8 p1, p0, q1, q0, mask, row0, row1, row2, row3, row4, row5, row6, row7;
    642  v16u8 row8, row9, row10, row11, row12, row13, row14, row15;
    643  v8i16 tmp0, tmp1;
    644  const v16u8 b_limit = (v16u8)__msa_fill_b(b_limit_in);
    645  uint8_t* ptemp_src = src - 2;
    646 
    647  LD_UB8(ptemp_src, stride, row0, row1, row2, row3, row4, row5, row6, row7);
    648  LD_UB8(ptemp_src + 8 * stride, stride,
    649         row8, row9, row10, row11, row12, row13, row14, row15);
    650  TRANSPOSE16x4_UB_UB(row0, row1, row2, row3, row4, row5, row6, row7,
    651                      row8, row9, row10, row11, row12, row13, row14, row15,
    652                      p1, p0, q0, q1);
    653  LPF_SIMPLE_MASK(p1, p0, q0, q1, b_limit, mask);
    654  LPF_SIMPLE_FILT(p1, p0, q0, q1, mask);
    655  ILVRL_B2_SH(q0, p0, tmp1, tmp0);
    656  ptemp_src += 1;
    657  ST2x4_UB(tmp1, 0, ptemp_src, stride);
    658  ptemp_src += 4 * stride;
    659  ST2x4_UB(tmp1, 4, ptemp_src, stride);
    660  ptemp_src += 4 * stride;
    661  ST2x4_UB(tmp0, 0, ptemp_src, stride);
    662  ptemp_src += 4 * stride;
    663  ST2x4_UB(tmp0, 4, ptemp_src, stride);
    664  ptemp_src += 4 * stride;
    665 }
    666 
    667 static void SimpleVFilter16i(uint8_t* src_y, int stride, int b_limit_in) {
    668  SimpleVFilter16(src_y +  4 * stride, stride, b_limit_in);
    669  SimpleVFilter16(src_y +  8 * stride, stride, b_limit_in);
    670  SimpleVFilter16(src_y + 12 * stride, stride, b_limit_in);
    671 }
    672 
    673 static void SimpleHFilter16i(uint8_t* src_y, int stride, int b_limit_in) {
    674  SimpleHFilter16(src_y +  4, stride, b_limit_in);
    675  SimpleHFilter16(src_y +  8, stride, b_limit_in);
    676  SimpleHFilter16(src_y + 12, stride, b_limit_in);
    677 }
    678 
    679 //------------------------------------------------------------------------------
    680 // Intra predictions
    681 //------------------------------------------------------------------------------
    682 
    683 // 4x4
    684 
    685 static void DC4(uint8_t* dst) {   // DC
    686  uint32_t dc = 4;
    687  int i;
    688  for (i = 0; i < 4; ++i) dc += dst[i - BPS] + dst[-1 + i * BPS];
    689  dc >>= 3;
    690  dc = dc | (dc << 8) | (dc << 16) | (dc << 24);
    691  SW4(dc, dc, dc, dc, dst, BPS);
    692 }
    693 
    694 static void TM4(uint8_t* dst) {
    695  const uint8_t* const ptemp = dst - BPS - 1;
    696  v8i16 T, d, r0, r1, r2, r3;
    697  const v16i8 zero = { 0 };
    698  const v8i16 TL = (v8i16)__msa_fill_h(ptemp[0 * BPS]);
    699  const v8i16 L0 = (v8i16)__msa_fill_h(ptemp[1 * BPS]);
    700  const v8i16 L1 = (v8i16)__msa_fill_h(ptemp[2 * BPS]);
    701  const v8i16 L2 = (v8i16)__msa_fill_h(ptemp[3 * BPS]);
    702  const v8i16 L3 = (v8i16)__msa_fill_h(ptemp[4 * BPS]);
    703  const v16u8 T1 = LD_UB(ptemp + 1);
    704 
    705  T  = (v8i16)__msa_ilvr_b(zero, (v16i8)T1);
    706  d = T - TL;
    707  ADD4(d, L0, d, L1, d, L2, d, L3, r0, r1, r2, r3);
    708  CLIP_SH4_0_255(r0, r1, r2, r3);
    709  PCKEV_ST4x4_UB(r0, r1, r2, r3, dst, BPS);
    710 }
    711 
    712 static void VE4(uint8_t* dst) {    // vertical
    713  const uint8_t* const ptop = dst - BPS - 1;
    714  const uint32_t val0 = LW(ptop + 0);
    715  const uint32_t val1 = LW(ptop + 4);
    716  uint32_t out;
    717  v16u8 A = { 0 }, B, C, AC, B2, R;
    718 
    719  INSERT_W2_UB(val0, val1, A);
    720  B = SLDI_UB(A, A, 1);
    721  C = SLDI_UB(A, A, 2);
    722  AC = __msa_ave_u_b(A, C);
    723  B2 = __msa_ave_u_b(B, B);
    724  R = __msa_aver_u_b(AC, B2);
    725  out = __msa_copy_s_w((v4i32)R, 0);
    726  SW4(out, out, out, out, dst, BPS);
    727 }
    728 
    729 static void RD4(uint8_t* dst) {   // Down-right
    730  const uint8_t* const ptop = dst - 1 - BPS;
    731  uint32_t val0 = LW(ptop + 0);
    732  uint32_t val1 = LW(ptop + 4);
    733  uint32_t val2, val3;
    734  v16u8 A, B, C, AC, B2, R, A1 = { 0 };
    735 
    736  INSERT_W2_UB(val0, val1, A1);
    737  A = SLDI_UB(A1, A1, 12);
    738  A = (v16u8)__msa_insert_b((v16i8)A, 3, ptop[1 * BPS]);
    739  A = (v16u8)__msa_insert_b((v16i8)A, 2, ptop[2 * BPS]);
    740  A = (v16u8)__msa_insert_b((v16i8)A, 1, ptop[3 * BPS]);
    741  A = (v16u8)__msa_insert_b((v16i8)A, 0, ptop[4 * BPS]);
    742  B = SLDI_UB(A, A, 1);
    743  C = SLDI_UB(A, A, 2);
    744  AC = __msa_ave_u_b(A, C);
    745  B2 = __msa_ave_u_b(B, B);
    746  R = __msa_aver_u_b(AC, B2);
    747  val3 = __msa_copy_s_w((v4i32)R, 0);
    748  R = SLDI_UB(R, R, 1);
    749  val2 = __msa_copy_s_w((v4i32)R, 0);
    750  R = SLDI_UB(R, R, 1);
    751  val1 = __msa_copy_s_w((v4i32)R, 0);
    752  R = SLDI_UB(R, R, 1);
    753  val0 = __msa_copy_s_w((v4i32)R, 0);
    754  SW4(val0, val1, val2, val3, dst, BPS);
    755 }
    756 
    757 static void LD4(uint8_t* dst) {   // Down-Left
    758  const uint8_t* const ptop = dst - BPS;
    759  uint32_t val0 = LW(ptop + 0);
    760  uint32_t val1 = LW(ptop + 4);
    761  uint32_t val2, val3;
    762  v16u8 A = { 0 }, B, C, AC, B2, R;
    763 
    764  INSERT_W2_UB(val0, val1, A);
    765  B = SLDI_UB(A, A, 1);
    766  C = SLDI_UB(A, A, 2);
    767  C = (v16u8)__msa_insert_b((v16i8)C, 6, ptop[7]);
    768  AC = __msa_ave_u_b(A, C);
    769  B2 = __msa_ave_u_b(B, B);
    770  R = __msa_aver_u_b(AC, B2);
    771  val0 = __msa_copy_s_w((v4i32)R, 0);
    772  R = SLDI_UB(R, R, 1);
    773  val1 = __msa_copy_s_w((v4i32)R, 0);
    774  R = SLDI_UB(R, R, 1);
    775  val2 = __msa_copy_s_w((v4i32)R, 0);
    776  R = SLDI_UB(R, R, 1);
    777  val3 = __msa_copy_s_w((v4i32)R, 0);
    778  SW4(val0, val1, val2, val3, dst, BPS);
    779 }
    780 
    781 // 16x16
    782 
    783 static void DC16(uint8_t* dst) {   // DC
    784  uint32_t dc = 16;
    785  int i;
    786  const v16u8 rtop = LD_UB(dst - BPS);
    787  const v8u16 dctop = __msa_hadd_u_h(rtop, rtop);
    788  v16u8 out;
    789 
    790  for (i = 0; i < 16; ++i) {
    791    dc += dst[-1 + i * BPS];
    792  }
    793  dc += HADD_UH_U32(dctop);
    794  out = (v16u8)__msa_fill_b(dc >> 5);
    795  ST_UB8(out, out, out, out, out, out, out, out, dst, BPS);
    796  ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS);
    797 }
    798 
    799 static void TM16(uint8_t* dst) {
    800  int j;
    801  v8i16 d1, d2;
    802  const v16i8 zero = { 0 };
    803  const v8i16 TL = (v8i16)__msa_fill_h(dst[-1 - BPS]);
    804  const v16i8 T = LD_SB(dst - BPS);
    805 
    806  ILVRL_B2_SH(zero, T, d1, d2);
    807  SUB2(d1, TL, d2, TL, d1, d2);
    808  for (j = 0; j < 16; j += 4) {
    809    v16i8 t0, t1, t2, t3;
    810    v8i16 r0, r1, r2, r3, r4, r5, r6, r7;
    811    const v8i16 L0 = (v8i16)__msa_fill_h(dst[-1 + 0 * BPS]);
    812    const v8i16 L1 = (v8i16)__msa_fill_h(dst[-1 + 1 * BPS]);
    813    const v8i16 L2 = (v8i16)__msa_fill_h(dst[-1 + 2 * BPS]);
    814    const v8i16 L3 = (v8i16)__msa_fill_h(dst[-1 + 3 * BPS]);
    815    ADD4(d1, L0, d1, L1, d1, L2, d1, L3, r0, r1, r2, r3);
    816    ADD4(d2, L0, d2, L1, d2, L2, d2, L3, r4, r5, r6, r7);
    817    CLIP_SH4_0_255(r0, r1, r2, r3);
    818    CLIP_SH4_0_255(r4, r5, r6, r7);
    819    PCKEV_B4_SB(r4, r0, r5, r1, r6, r2, r7, r3, t0, t1, t2, t3);
    820    ST_SB4(t0, t1, t2, t3, dst, BPS);
    821    dst += 4 * BPS;
    822  }
    823 }
    824 
    825 static void VE16(uint8_t* dst) {   // vertical
    826  const v16u8 rtop = LD_UB(dst - BPS);
    827  ST_UB8(rtop, rtop, rtop, rtop, rtop, rtop, rtop, rtop, dst, BPS);
    828  ST_UB8(rtop, rtop, rtop, rtop, rtop, rtop, rtop, rtop, dst + 8 * BPS, BPS);
    829 }
    830 
    831 static void HE16(uint8_t* dst) {   // horizontal
    832  int j;
    833  for (j = 16; j > 0; j -= 4) {
    834    const v16u8 L0 = (v16u8)__msa_fill_b(dst[-1 + 0 * BPS]);
    835    const v16u8 L1 = (v16u8)__msa_fill_b(dst[-1 + 1 * BPS]);
    836    const v16u8 L2 = (v16u8)__msa_fill_b(dst[-1 + 2 * BPS]);
    837    const v16u8 L3 = (v16u8)__msa_fill_b(dst[-1 + 3 * BPS]);
    838    ST_UB4(L0, L1, L2, L3, dst, BPS);
    839    dst += 4 * BPS;
    840  }
    841 }
    842 
    843 static void DC16NoTop(uint8_t* dst) {   // DC with top samples not available
    844  int j;
    845  uint32_t dc = 8;
    846  v16u8 out;
    847 
    848  for (j = 0; j < 16; ++j) {
    849    dc += dst[-1 + j * BPS];
    850  }
    851  out = (v16u8)__msa_fill_b(dc >> 4);
    852  ST_UB8(out, out, out, out, out, out, out, out, dst, BPS);
    853  ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS);
    854 }
    855 
    856 static void DC16NoLeft(uint8_t* dst) {   // DC with left samples not available
    857  uint32_t dc = 8;
    858  const v16u8 rtop = LD_UB(dst - BPS);
    859  const v8u16 dctop = __msa_hadd_u_h(rtop, rtop);
    860  v16u8 out;
    861 
    862  dc += HADD_UH_U32(dctop);
    863  out = (v16u8)__msa_fill_b(dc >> 4);
    864  ST_UB8(out, out, out, out, out, out, out, out, dst, BPS);
    865  ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS);
    866 }
    867 
    868 static void DC16NoTopLeft(uint8_t* dst) {   // DC with nothing
    869  const v16u8 out = (v16u8)__msa_fill_b(0x80);
    870  ST_UB8(out, out, out, out, out, out, out, out, dst, BPS);
    871  ST_UB8(out, out, out, out, out, out, out, out, dst + 8 * BPS, BPS);
    872 }
    873 
    874 // Chroma
    875 
    876 #define STORE8x8(out, dst) do {                 \
    877  SD4(out, out, out, out, dst + 0 * BPS, BPS);  \
    878  SD4(out, out, out, out, dst + 4 * BPS, BPS);  \
    879 } while (0)
    880 
    881 static void DC8uv(uint8_t* dst) {   // DC
    882  uint32_t dc = 8;
    883  int i;
    884  uint64_t out;
    885  const v16u8 rtop = LD_UB(dst - BPS);
    886  const v8u16 temp0 = __msa_hadd_u_h(rtop, rtop);
    887  const v4u32 temp1 = __msa_hadd_u_w(temp0, temp0);
    888  const v2u64 temp2 = __msa_hadd_u_d(temp1, temp1);
    889  v16u8 dctemp;
    890 
    891  for (i = 0; i < 8; ++i) {
    892    dc += dst[-1 + i * BPS];
    893  }
    894  dc += __msa_copy_s_w((v4i32)temp2, 0);
    895  dctemp = (v16u8)__msa_fill_b(dc >> 4);
    896  out = __msa_copy_s_d((v2i64)dctemp, 0);
    897  STORE8x8(out, dst);
    898 }
    899 
    900 static void TM8uv(uint8_t* dst) {
    901  int j;
    902  const v16i8 T1 = LD_SB(dst - BPS);
    903  const v16i8 zero = { 0 };
    904  const v8i16 T  = (v8i16)__msa_ilvr_b(zero, T1);
    905  const v8i16 TL = (v8i16)__msa_fill_h(dst[-1 - BPS]);
    906  const v8i16 d = T - TL;
    907 
    908  for (j = 0; j < 8; j += 4) {
    909    v16i8 t0, t1;
    910    v8i16 r0 = (v8i16)__msa_fill_h(dst[-1 + 0 * BPS]);
    911    v8i16 r1 = (v8i16)__msa_fill_h(dst[-1 + 1 * BPS]);
    912    v8i16 r2 = (v8i16)__msa_fill_h(dst[-1 + 2 * BPS]);
    913    v8i16 r3 = (v8i16)__msa_fill_h(dst[-1 + 3 * BPS]);
    914    ADD4(d, r0, d, r1, d, r2, d, r3, r0, r1, r2, r3);
    915    CLIP_SH4_0_255(r0, r1, r2, r3);
    916    PCKEV_B2_SB(r1, r0, r3, r2, t0, t1);
    917    ST4x4_UB(t0, t1, 0, 2, 0, 2, dst, BPS);
    918    ST4x4_UB(t0, t1, 1, 3, 1, 3, dst + 4, BPS);
    919    dst += 4 * BPS;
    920  }
    921 }
    922 
    923 static void VE8uv(uint8_t* dst) {   // vertical
    924  const v16u8 rtop = LD_UB(dst - BPS);
    925  const uint64_t out = __msa_copy_s_d((v2i64)rtop, 0);
    926  STORE8x8(out, dst);
    927 }
    928 
    929 static void HE8uv(uint8_t* dst) {   // horizontal
    930  int j;
    931  for (j = 0; j < 8; j += 4) {
    932    const v16u8 L0 = (v16u8)__msa_fill_b(dst[-1 + 0 * BPS]);
    933    const v16u8 L1 = (v16u8)__msa_fill_b(dst[-1 + 1 * BPS]);
    934    const v16u8 L2 = (v16u8)__msa_fill_b(dst[-1 + 2 * BPS]);
    935    const v16u8 L3 = (v16u8)__msa_fill_b(dst[-1 + 3 * BPS]);
    936    const uint64_t out0 = __msa_copy_s_d((v2i64)L0, 0);
    937    const uint64_t out1 = __msa_copy_s_d((v2i64)L1, 0);
    938    const uint64_t out2 = __msa_copy_s_d((v2i64)L2, 0);
    939    const uint64_t out3 = __msa_copy_s_d((v2i64)L3, 0);
    940    SD4(out0, out1, out2, out3, dst, BPS);
    941    dst += 4 * BPS;
    942  }
    943 }
    944 
    945 static void DC8uvNoLeft(uint8_t* dst) {   // DC with no left samples
    946  const uint32_t dc = 4;
    947  const v16u8 rtop = LD_UB(dst - BPS);
    948  const v8u16 temp0 = __msa_hadd_u_h(rtop, rtop);
    949  const v4u32 temp1 = __msa_hadd_u_w(temp0, temp0);
    950  const v2u64 temp2 = __msa_hadd_u_d(temp1, temp1);
    951  const uint32_t sum_m = __msa_copy_s_w((v4i32)temp2, 0);
    952  const v16u8 dcval = (v16u8)__msa_fill_b((dc + sum_m) >> 3);
    953  const uint64_t out = __msa_copy_s_d((v2i64)dcval, 0);
    954  STORE8x8(out, dst);
    955 }
    956 
    957 static void DC8uvNoTop(uint8_t* dst) {   // DC with no top samples
    958  uint32_t dc = 4;
    959  int i;
    960  uint64_t out;
    961  v16u8 dctemp;
    962 
    963  for (i = 0; i < 8; ++i) {
    964    dc += dst[-1 + i * BPS];
    965  }
    966  dctemp = (v16u8)__msa_fill_b(dc >> 3);
    967  out = __msa_copy_s_d((v2i64)dctemp, 0);
    968  STORE8x8(out, dst);
    969 }
    970 
    971 static void DC8uvNoTopLeft(uint8_t* dst) {   // DC with nothing
    972  const uint64_t out = 0x8080808080808080ULL;
    973  STORE8x8(out, dst);
    974 }
    975 
    976 //------------------------------------------------------------------------------
    977 // Entry point
    978 
    979 extern void VP8DspInitMSA(void);
    980 
    981 WEBP_TSAN_IGNORE_FUNCTION void VP8DspInitMSA(void) {
    982  VP8TransformWHT = TransformWHT;
    983  VP8Transform = TransformTwo;
    984  VP8TransformDC = TransformDC;
    985  VP8TransformAC3 = TransformAC3;
    986 
    987  VP8VFilter16  = VFilter16;
    988  VP8HFilter16  = HFilter16;
    989  VP8VFilter16i = VFilter16i;
    990  VP8HFilter16i = HFilter16i;
    991  VP8VFilter8  = VFilter8;
    992  VP8HFilter8  = HFilter8;
    993  VP8VFilter8i = VFilter8i;
    994  VP8HFilter8i = HFilter8i;
    995  VP8SimpleVFilter16  = SimpleVFilter16;
    996  VP8SimpleHFilter16  = SimpleHFilter16;
    997  VP8SimpleVFilter16i = SimpleVFilter16i;
    998  VP8SimpleHFilter16i = SimpleHFilter16i;
    999 
   1000  VP8PredLuma4[0] = DC4;
   1001  VP8PredLuma4[1] = TM4;
   1002  VP8PredLuma4[2] = VE4;
   1003  VP8PredLuma4[4] = RD4;
   1004  VP8PredLuma4[6] = LD4;
   1005  VP8PredLuma16[0] = DC16;
   1006  VP8PredLuma16[1] = TM16;
   1007  VP8PredLuma16[2] = VE16;
   1008  VP8PredLuma16[3] = HE16;
   1009  VP8PredLuma16[4] = DC16NoTop;
   1010  VP8PredLuma16[5] = DC16NoLeft;
   1011  VP8PredLuma16[6] = DC16NoTopLeft;
   1012  VP8PredChroma8[0] = DC8uv;
   1013  VP8PredChroma8[1] = TM8uv;
   1014  VP8PredChroma8[2] = VE8uv;
   1015  VP8PredChroma8[3] = HE8uv;
   1016  VP8PredChroma8[4] = DC8uvNoTop;
   1017  VP8PredChroma8[5] = DC8uvNoLeft;
   1018  VP8PredChroma8[6] = DC8uvNoTopLeft;
   1019 }
   1020 
   1021 #else  // !WEBP_USE_MSA
   1022 
   1023 WEBP_DSP_INIT_STUB(VP8DspInitMSA)
   1024 
   1025 #endif  // WEBP_USE_MSA