fwd_dct32x32_impl_sse2.h 163 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11
/*
 *  Copyright (c) 2012 The WebM 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.
 */

#include <emmintrin.h>  // SSE2
12

13
#include "vpx_dsp/fwd_txfm.h"
Jingning Han's avatar
Jingning Han committed
14 15
#include "vpx_dsp/txfm_common.h"
#include "vpx_dsp/x86/txfm_common_sse2.h"
16

17 18 19
// TODO(jingning) The high bit-depth version needs re-work for performance.
// The current SSE2 implementation also causes cross reference to the static
// functions in the C implementation file.
20 21 22
#if DCT_HIGH_BIT_DEPTH
#define ADD_EPI16 _mm_adds_epi16
#define SUB_EPI16 _mm_subs_epi16
23
#if FDCT32x32_HIGH_PRECISION
24
void vpx_fdct32x32_rows_c(const int16_t *intermediate, tran_low_t *out) {
25 26 27 28 29
    int i, j;
    for (i = 0; i < 32; ++i) {
      tran_high_t temp_in[32], temp_out[32];
      for (j = 0; j < 32; ++j)
        temp_in[j] = intermediate[j * 32 + i];
30
      vpx_fdct32(temp_in, temp_out, 0);
31
      for (j = 0; j < 32; ++j)
Yaowu Xu's avatar
Yaowu Xu committed
32 33
        out[j + i * 32] =
            (tran_low_t)((temp_out[j] + 1 + (temp_out[j] < 0)) >> 2);
34
    }
35
}
36 37
  #define HIGH_FDCT32x32_2D_C vpx_highbd_fdct32x32_c
  #define HIGH_FDCT32x32_2D_ROWS_C vpx_fdct32x32_rows_c
38
#else
39
void vpx_fdct32x32_rd_rows_c(const int16_t *intermediate, tran_low_t *out) {
40 41 42 43 44
    int i, j;
    for (i = 0; i < 32; ++i) {
      tran_high_t temp_in[32], temp_out[32];
      for (j = 0; j < 32; ++j)
        temp_in[j] = intermediate[j * 32 + i];
45
      vpx_fdct32(temp_in, temp_out, 1);
46
      for (j = 0; j < 32; ++j)
Yaowu Xu's avatar
Yaowu Xu committed
47
        out[j + i * 32] = (tran_low_t)temp_out[j];
48
    }
49
}
50 51
  #define HIGH_FDCT32x32_2D_C vpx_highbd_fdct32x32_rd_c
  #define HIGH_FDCT32x32_2D_ROWS_C vpx_fdct32x32_rd_rows_c
52 53 54 55 56 57
#endif  // FDCT32x32_HIGH_PRECISION
#else
#define ADD_EPI16 _mm_add_epi16
#define SUB_EPI16 _mm_sub_epi16
#endif  // DCT_HIGH_BIT_DEPTH

58

59
void FDCT32x32_2D(const int16_t *input,
60
                  tran_low_t *output_org, int stride) {
61
  // Calculate pre-multiplied strides
62 63 64
  const int str1 = stride;
  const int str2 = 2 * stride;
  const int str3 = 2 * stride + str1;
65 66 67 68 69 70
  // We need an intermediate buffer between passes.
  DECLARE_ALIGNED(16, int16_t, intermediate[32 * 32]);
  // Constants
  //    When we use them, in one case, they are all the same. In all others
  //    it's a pair of them that we need to repeat four times. This is done
  //    by constructing the 32 bit constant corresponding to that pair.
71
  const __m128i k__cospi_p16_p16 = _mm_set1_epi16((int16_t)cospi_16_64);
72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110
  const __m128i k__cospi_p16_m16 = pair_set_epi16(+cospi_16_64, -cospi_16_64);
  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64,   cospi_24_64);
  const __m128i k__cospi_m24_m08 = pair_set_epi16(-cospi_24_64, -cospi_8_64);
  const __m128i k__cospi_p24_p08 = pair_set_epi16(+cospi_24_64,  cospi_8_64);
  const __m128i k__cospi_p12_p20 = pair_set_epi16(+cospi_12_64,  cospi_20_64);
  const __m128i k__cospi_m20_p12 = pair_set_epi16(-cospi_20_64,  cospi_12_64);
  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64,   cospi_28_64);
  const __m128i k__cospi_p28_p04 = pair_set_epi16(+cospi_28_64,  cospi_4_64);
  const __m128i k__cospi_m28_m04 = pair_set_epi16(-cospi_28_64, -cospi_4_64);
  const __m128i k__cospi_m12_m20 = pair_set_epi16(-cospi_12_64, -cospi_20_64);
  const __m128i k__cospi_p30_p02 = pair_set_epi16(+cospi_30_64,  cospi_2_64);
  const __m128i k__cospi_p14_p18 = pair_set_epi16(+cospi_14_64,  cospi_18_64);
  const __m128i k__cospi_p22_p10 = pair_set_epi16(+cospi_22_64,  cospi_10_64);
  const __m128i k__cospi_p06_p26 = pair_set_epi16(+cospi_6_64,   cospi_26_64);
  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64,  cospi_6_64);
  const __m128i k__cospi_m10_p22 = pair_set_epi16(-cospi_10_64,  cospi_22_64);
  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64,  cospi_14_64);
  const __m128i k__cospi_m02_p30 = pair_set_epi16(-cospi_2_64,   cospi_30_64);
  const __m128i k__cospi_p31_p01 = pair_set_epi16(+cospi_31_64,  cospi_1_64);
  const __m128i k__cospi_p15_p17 = pair_set_epi16(+cospi_15_64,  cospi_17_64);
  const __m128i k__cospi_p23_p09 = pair_set_epi16(+cospi_23_64,  cospi_9_64);
  const __m128i k__cospi_p07_p25 = pair_set_epi16(+cospi_7_64,   cospi_25_64);
  const __m128i k__cospi_m25_p07 = pair_set_epi16(-cospi_25_64,  cospi_7_64);
  const __m128i k__cospi_m09_p23 = pair_set_epi16(-cospi_9_64,   cospi_23_64);
  const __m128i k__cospi_m17_p15 = pair_set_epi16(-cospi_17_64,  cospi_15_64);
  const __m128i k__cospi_m01_p31 = pair_set_epi16(-cospi_1_64,   cospi_31_64);
  const __m128i k__cospi_p27_p05 = pair_set_epi16(+cospi_27_64,  cospi_5_64);
  const __m128i k__cospi_p11_p21 = pair_set_epi16(+cospi_11_64,  cospi_21_64);
  const __m128i k__cospi_p19_p13 = pair_set_epi16(+cospi_19_64,  cospi_13_64);
  const __m128i k__cospi_p03_p29 = pair_set_epi16(+cospi_3_64,   cospi_29_64);
  const __m128i k__cospi_m29_p03 = pair_set_epi16(-cospi_29_64,  cospi_3_64);
  const __m128i k__cospi_m13_p19 = pair_set_epi16(-cospi_13_64,  cospi_19_64);
  const __m128i k__cospi_m21_p11 = pair_set_epi16(-cospi_21_64,  cospi_11_64);
  const __m128i k__cospi_m05_p27 = pair_set_epi16(-cospi_5_64,   cospi_27_64);
  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
  const __m128i kZero = _mm_set1_epi16(0);
  const __m128i kOne  = _mm_set1_epi16(1);
  // Do the two transform/transpose passes
  int pass;
111 112 113
#if DCT_HIGH_BIT_DEPTH
  int overflow;
#endif
114 115 116 117 118 119 120 121 122 123 124 125
  for (pass = 0; pass < 2; ++pass) {
    // We process eight columns (transposed rows in second pass) at a time.
    int column_start;
    for (column_start = 0; column_start < 32; column_start += 8) {
      __m128i step1[32];
      __m128i step2[32];
      __m128i step3[32];
      __m128i out[32];
      // Stage 1
      // Note: even though all the loads below are aligned, using the aligned
      //       intrinsic make the code slightly slower.
      if (0 == pass) {
126
        const int16_t *in  = &input[column_start];
127 128 129 130
        // step1[i] =  (in[ 0 * stride] + in[(32 -  1) * stride]) << 2;
        // Note: the next four blocks could be in a loop. That would help the
        //       instruction cache but is actually slower.
        {
131 132
          const int16_t *ina =  in +  0 * str1;
          const int16_t *inb =  in + 31 * str1;
133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160
          __m128i *step1a = &step1[ 0];
          __m128i *step1b = &step1[31];
          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
          step1a[ 0] = _mm_add_epi16(ina0, inb0);
          step1a[ 1] = _mm_add_epi16(ina1, inb1);
          step1a[ 2] = _mm_add_epi16(ina2, inb2);
          step1a[ 3] = _mm_add_epi16(ina3, inb3);
          step1b[-3] = _mm_sub_epi16(ina3, inb3);
          step1b[-2] = _mm_sub_epi16(ina2, inb2);
          step1b[-1] = _mm_sub_epi16(ina1, inb1);
          step1b[-0] = _mm_sub_epi16(ina0, inb0);
          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
        }
        {
161 162
          const int16_t *ina =  in +  4 * str1;
          const int16_t *inb =  in + 27 * str1;
163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190
          __m128i *step1a = &step1[ 4];
          __m128i *step1b = &step1[27];
          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
          step1a[ 0] = _mm_add_epi16(ina0, inb0);
          step1a[ 1] = _mm_add_epi16(ina1, inb1);
          step1a[ 2] = _mm_add_epi16(ina2, inb2);
          step1a[ 3] = _mm_add_epi16(ina3, inb3);
          step1b[-3] = _mm_sub_epi16(ina3, inb3);
          step1b[-2] = _mm_sub_epi16(ina2, inb2);
          step1b[-1] = _mm_sub_epi16(ina1, inb1);
          step1b[-0] = _mm_sub_epi16(ina0, inb0);
          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
        }
        {
191 192
          const int16_t *ina =  in +  8 * str1;
          const int16_t *inb =  in + 23 * str1;
193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220
          __m128i *step1a = &step1[ 8];
          __m128i *step1b = &step1[23];
          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
          step1a[ 0] = _mm_add_epi16(ina0, inb0);
          step1a[ 1] = _mm_add_epi16(ina1, inb1);
          step1a[ 2] = _mm_add_epi16(ina2, inb2);
          step1a[ 3] = _mm_add_epi16(ina3, inb3);
          step1b[-3] = _mm_sub_epi16(ina3, inb3);
          step1b[-2] = _mm_sub_epi16(ina2, inb2);
          step1b[-1] = _mm_sub_epi16(ina1, inb1);
          step1b[-0] = _mm_sub_epi16(ina0, inb0);
          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
        }
        {
221 222
          const int16_t *ina =  in + 12 * str1;
          const int16_t *inb =  in + 19 * str1;
223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266
          __m128i *step1a = &step1[12];
          __m128i *step1b = &step1[19];
          const __m128i ina0  = _mm_loadu_si128((const __m128i *)(ina));
          const __m128i ina1  = _mm_loadu_si128((const __m128i *)(ina + str1));
          const __m128i ina2  = _mm_loadu_si128((const __m128i *)(ina + str2));
          const __m128i ina3  = _mm_loadu_si128((const __m128i *)(ina + str3));
          const __m128i inb3  = _mm_loadu_si128((const __m128i *)(inb - str3));
          const __m128i inb2  = _mm_loadu_si128((const __m128i *)(inb - str2));
          const __m128i inb1  = _mm_loadu_si128((const __m128i *)(inb - str1));
          const __m128i inb0  = _mm_loadu_si128((const __m128i *)(inb));
          step1a[ 0] = _mm_add_epi16(ina0, inb0);
          step1a[ 1] = _mm_add_epi16(ina1, inb1);
          step1a[ 2] = _mm_add_epi16(ina2, inb2);
          step1a[ 3] = _mm_add_epi16(ina3, inb3);
          step1b[-3] = _mm_sub_epi16(ina3, inb3);
          step1b[-2] = _mm_sub_epi16(ina2, inb2);
          step1b[-1] = _mm_sub_epi16(ina1, inb1);
          step1b[-0] = _mm_sub_epi16(ina0, inb0);
          step1a[ 0] = _mm_slli_epi16(step1a[ 0], 2);
          step1a[ 1] = _mm_slli_epi16(step1a[ 1], 2);
          step1a[ 2] = _mm_slli_epi16(step1a[ 2], 2);
          step1a[ 3] = _mm_slli_epi16(step1a[ 3], 2);
          step1b[-3] = _mm_slli_epi16(step1b[-3], 2);
          step1b[-2] = _mm_slli_epi16(step1b[-2], 2);
          step1b[-1] = _mm_slli_epi16(step1b[-1], 2);
          step1b[-0] = _mm_slli_epi16(step1b[-0], 2);
        }
      } else {
        int16_t *in = &intermediate[column_start];
        // step1[i] =  in[ 0 * 32] + in[(32 -  1) * 32];
        // Note: using the same approach as above to have common offset is
        //       counter-productive as all offsets can be calculated at compile
        //       time.
        // Note: the next four blocks could be in a loop. That would help the
        //       instruction cache but is actually slower.
        {
          __m128i in00  = _mm_loadu_si128((const __m128i *)(in +  0 * 32));
          __m128i in01  = _mm_loadu_si128((const __m128i *)(in +  1 * 32));
          __m128i in02  = _mm_loadu_si128((const __m128i *)(in +  2 * 32));
          __m128i in03  = _mm_loadu_si128((const __m128i *)(in +  3 * 32));
          __m128i in28  = _mm_loadu_si128((const __m128i *)(in + 28 * 32));
          __m128i in29  = _mm_loadu_si128((const __m128i *)(in + 29 * 32));
          __m128i in30  = _mm_loadu_si128((const __m128i *)(in + 30 * 32));
          __m128i in31  = _mm_loadu_si128((const __m128i *)(in + 31 * 32));
267 268 269 270 271 272 273 274 275
          step1[0] = ADD_EPI16(in00, in31);
          step1[1] = ADD_EPI16(in01, in30);
          step1[2] = ADD_EPI16(in02, in29);
          step1[3] = ADD_EPI16(in03, in28);
          step1[28] = SUB_EPI16(in03, in28);
          step1[29] = SUB_EPI16(in02, in29);
          step1[30] = SUB_EPI16(in01, in30);
          step1[31] = SUB_EPI16(in00, in31);
#if DCT_HIGH_BIT_DEPTH
276 277 278
          overflow = check_epi16_overflow_x8(&step1[0], &step1[1], &step1[2],
                                             &step1[3], &step1[28], &step1[29],
                                             &step1[30], &step1[31]);
279 280 281 282 283
          if (overflow) {
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
            return;
          }
#endif  // DCT_HIGH_BIT_DEPTH
284 285 286 287 288 289 290 291 292 293
        }
        {
          __m128i in04  = _mm_loadu_si128((const __m128i *)(in +  4 * 32));
          __m128i in05  = _mm_loadu_si128((const __m128i *)(in +  5 * 32));
          __m128i in06  = _mm_loadu_si128((const __m128i *)(in +  6 * 32));
          __m128i in07  = _mm_loadu_si128((const __m128i *)(in +  7 * 32));
          __m128i in24  = _mm_loadu_si128((const __m128i *)(in + 24 * 32));
          __m128i in25  = _mm_loadu_si128((const __m128i *)(in + 25 * 32));
          __m128i in26  = _mm_loadu_si128((const __m128i *)(in + 26 * 32));
          __m128i in27  = _mm_loadu_si128((const __m128i *)(in + 27 * 32));
294 295 296 297 298 299 300 301 302
          step1[4] = ADD_EPI16(in04, in27);
          step1[5] = ADD_EPI16(in05, in26);
          step1[6] = ADD_EPI16(in06, in25);
          step1[7] = ADD_EPI16(in07, in24);
          step1[24] = SUB_EPI16(in07, in24);
          step1[25] = SUB_EPI16(in06, in25);
          step1[26] = SUB_EPI16(in05, in26);
          step1[27] = SUB_EPI16(in04, in27);
#if DCT_HIGH_BIT_DEPTH
303 304 305
          overflow = check_epi16_overflow_x8(&step1[4], &step1[5], &step1[6],
                                             &step1[7], &step1[24], &step1[25],
                                             &step1[26], &step1[27]);
306 307 308 309 310
          if (overflow) {
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
            return;
          }
#endif  // DCT_HIGH_BIT_DEPTH
311 312 313 314 315 316 317 318 319 320
        }
        {
          __m128i in08  = _mm_loadu_si128((const __m128i *)(in +  8 * 32));
          __m128i in09  = _mm_loadu_si128((const __m128i *)(in +  9 * 32));
          __m128i in10  = _mm_loadu_si128((const __m128i *)(in + 10 * 32));
          __m128i in11  = _mm_loadu_si128((const __m128i *)(in + 11 * 32));
          __m128i in20  = _mm_loadu_si128((const __m128i *)(in + 20 * 32));
          __m128i in21  = _mm_loadu_si128((const __m128i *)(in + 21 * 32));
          __m128i in22  = _mm_loadu_si128((const __m128i *)(in + 22 * 32));
          __m128i in23  = _mm_loadu_si128((const __m128i *)(in + 23 * 32));
321 322 323 324 325 326 327 328 329
          step1[8] = ADD_EPI16(in08, in23);
          step1[9] = ADD_EPI16(in09, in22);
          step1[10] = ADD_EPI16(in10, in21);
          step1[11] = ADD_EPI16(in11, in20);
          step1[20] = SUB_EPI16(in11, in20);
          step1[21] = SUB_EPI16(in10, in21);
          step1[22] = SUB_EPI16(in09, in22);
          step1[23] = SUB_EPI16(in08, in23);
#if DCT_HIGH_BIT_DEPTH
330 331 332
          overflow = check_epi16_overflow_x8(&step1[8], &step1[9], &step1[10],
                                             &step1[11], &step1[20], &step1[21],
                                             &step1[22], &step1[23]);
333 334 335 336 337
          if (overflow) {
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
            return;
          }
#endif  // DCT_HIGH_BIT_DEPTH
338 339 340 341 342 343 344 345 346 347
        }
        {
          __m128i in12  = _mm_loadu_si128((const __m128i *)(in + 12 * 32));
          __m128i in13  = _mm_loadu_si128((const __m128i *)(in + 13 * 32));
          __m128i in14  = _mm_loadu_si128((const __m128i *)(in + 14 * 32));
          __m128i in15  = _mm_loadu_si128((const __m128i *)(in + 15 * 32));
          __m128i in16  = _mm_loadu_si128((const __m128i *)(in + 16 * 32));
          __m128i in17  = _mm_loadu_si128((const __m128i *)(in + 17 * 32));
          __m128i in18  = _mm_loadu_si128((const __m128i *)(in + 18 * 32));
          __m128i in19  = _mm_loadu_si128((const __m128i *)(in + 19 * 32));
348 349 350 351 352 353 354 355 356
          step1[12] = ADD_EPI16(in12, in19);
          step1[13] = ADD_EPI16(in13, in18);
          step1[14] = ADD_EPI16(in14, in17);
          step1[15] = ADD_EPI16(in15, in16);
          step1[16] = SUB_EPI16(in15, in16);
          step1[17] = SUB_EPI16(in14, in17);
          step1[18] = SUB_EPI16(in13, in18);
          step1[19] = SUB_EPI16(in12, in19);
#if DCT_HIGH_BIT_DEPTH
357 358 359
          overflow = check_epi16_overflow_x8(&step1[12], &step1[13], &step1[14],
                                             &step1[15], &step1[16], &step1[17],
                                             &step1[18], &step1[19]);
360 361 362 363 364
          if (overflow) {
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
            return;
          }
#endif  // DCT_HIGH_BIT_DEPTH
365 366 367 368
        }
      }
      // Stage 2
      {
369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386
        step2[0] = ADD_EPI16(step1[0], step1[15]);
        step2[1] = ADD_EPI16(step1[1], step1[14]);
        step2[2] = ADD_EPI16(step1[2], step1[13]);
        step2[3] = ADD_EPI16(step1[3], step1[12]);
        step2[4] = ADD_EPI16(step1[4], step1[11]);
        step2[5] = ADD_EPI16(step1[5], step1[10]);
        step2[6] = ADD_EPI16(step1[6], step1[ 9]);
        step2[7] = ADD_EPI16(step1[7], step1[ 8]);
        step2[8] = SUB_EPI16(step1[7], step1[ 8]);
        step2[9] = SUB_EPI16(step1[6], step1[ 9]);
        step2[10] = SUB_EPI16(step1[5], step1[10]);
        step2[11] = SUB_EPI16(step1[4], step1[11]);
        step2[12] = SUB_EPI16(step1[3], step1[12]);
        step2[13] = SUB_EPI16(step1[2], step1[13]);
        step2[14] = SUB_EPI16(step1[1], step1[14]);
        step2[15] = SUB_EPI16(step1[0], step1[15]);
#if DCT_HIGH_BIT_DEPTH
        overflow = check_epi16_overflow_x16(
387 388 389 390
            &step2[0], &step2[1], &step2[2], &step2[3],
            &step2[4], &step2[5], &step2[6], &step2[7],
            &step2[8], &step2[9], &step2[10], &step2[11],
            &step2[12], &step2[13], &step2[14], &step2[15]);
391 392 393 394 395 396 397 398
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466
      }
      {
        const __m128i s2_20_0 = _mm_unpacklo_epi16(step1[27], step1[20]);
        const __m128i s2_20_1 = _mm_unpackhi_epi16(step1[27], step1[20]);
        const __m128i s2_21_0 = _mm_unpacklo_epi16(step1[26], step1[21]);
        const __m128i s2_21_1 = _mm_unpackhi_epi16(step1[26], step1[21]);
        const __m128i s2_22_0 = _mm_unpacklo_epi16(step1[25], step1[22]);
        const __m128i s2_22_1 = _mm_unpackhi_epi16(step1[25], step1[22]);
        const __m128i s2_23_0 = _mm_unpacklo_epi16(step1[24], step1[23]);
        const __m128i s2_23_1 = _mm_unpackhi_epi16(step1[24], step1[23]);
        const __m128i s2_20_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_m16);
        const __m128i s2_20_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_m16);
        const __m128i s2_21_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_m16);
        const __m128i s2_21_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_m16);
        const __m128i s2_22_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_m16);
        const __m128i s2_22_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_m16);
        const __m128i s2_23_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_m16);
        const __m128i s2_23_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_m16);
        const __m128i s2_24_2 = _mm_madd_epi16(s2_23_0, k__cospi_p16_p16);
        const __m128i s2_24_3 = _mm_madd_epi16(s2_23_1, k__cospi_p16_p16);
        const __m128i s2_25_2 = _mm_madd_epi16(s2_22_0, k__cospi_p16_p16);
        const __m128i s2_25_3 = _mm_madd_epi16(s2_22_1, k__cospi_p16_p16);
        const __m128i s2_26_2 = _mm_madd_epi16(s2_21_0, k__cospi_p16_p16);
        const __m128i s2_26_3 = _mm_madd_epi16(s2_21_1, k__cospi_p16_p16);
        const __m128i s2_27_2 = _mm_madd_epi16(s2_20_0, k__cospi_p16_p16);
        const __m128i s2_27_3 = _mm_madd_epi16(s2_20_1, k__cospi_p16_p16);
        // dct_const_round_shift
        const __m128i s2_20_4 = _mm_add_epi32(s2_20_2, k__DCT_CONST_ROUNDING);
        const __m128i s2_20_5 = _mm_add_epi32(s2_20_3, k__DCT_CONST_ROUNDING);
        const __m128i s2_21_4 = _mm_add_epi32(s2_21_2, k__DCT_CONST_ROUNDING);
        const __m128i s2_21_5 = _mm_add_epi32(s2_21_3, k__DCT_CONST_ROUNDING);
        const __m128i s2_22_4 = _mm_add_epi32(s2_22_2, k__DCT_CONST_ROUNDING);
        const __m128i s2_22_5 = _mm_add_epi32(s2_22_3, k__DCT_CONST_ROUNDING);
        const __m128i s2_23_4 = _mm_add_epi32(s2_23_2, k__DCT_CONST_ROUNDING);
        const __m128i s2_23_5 = _mm_add_epi32(s2_23_3, k__DCT_CONST_ROUNDING);
        const __m128i s2_24_4 = _mm_add_epi32(s2_24_2, k__DCT_CONST_ROUNDING);
        const __m128i s2_24_5 = _mm_add_epi32(s2_24_3, k__DCT_CONST_ROUNDING);
        const __m128i s2_25_4 = _mm_add_epi32(s2_25_2, k__DCT_CONST_ROUNDING);
        const __m128i s2_25_5 = _mm_add_epi32(s2_25_3, k__DCT_CONST_ROUNDING);
        const __m128i s2_26_4 = _mm_add_epi32(s2_26_2, k__DCT_CONST_ROUNDING);
        const __m128i s2_26_5 = _mm_add_epi32(s2_26_3, k__DCT_CONST_ROUNDING);
        const __m128i s2_27_4 = _mm_add_epi32(s2_27_2, k__DCT_CONST_ROUNDING);
        const __m128i s2_27_5 = _mm_add_epi32(s2_27_3, k__DCT_CONST_ROUNDING);
        const __m128i s2_20_6 = _mm_srai_epi32(s2_20_4, DCT_CONST_BITS);
        const __m128i s2_20_7 = _mm_srai_epi32(s2_20_5, DCT_CONST_BITS);
        const __m128i s2_21_6 = _mm_srai_epi32(s2_21_4, DCT_CONST_BITS);
        const __m128i s2_21_7 = _mm_srai_epi32(s2_21_5, DCT_CONST_BITS);
        const __m128i s2_22_6 = _mm_srai_epi32(s2_22_4, DCT_CONST_BITS);
        const __m128i s2_22_7 = _mm_srai_epi32(s2_22_5, DCT_CONST_BITS);
        const __m128i s2_23_6 = _mm_srai_epi32(s2_23_4, DCT_CONST_BITS);
        const __m128i s2_23_7 = _mm_srai_epi32(s2_23_5, DCT_CONST_BITS);
        const __m128i s2_24_6 = _mm_srai_epi32(s2_24_4, DCT_CONST_BITS);
        const __m128i s2_24_7 = _mm_srai_epi32(s2_24_5, DCT_CONST_BITS);
        const __m128i s2_25_6 = _mm_srai_epi32(s2_25_4, DCT_CONST_BITS);
        const __m128i s2_25_7 = _mm_srai_epi32(s2_25_5, DCT_CONST_BITS);
        const __m128i s2_26_6 = _mm_srai_epi32(s2_26_4, DCT_CONST_BITS);
        const __m128i s2_26_7 = _mm_srai_epi32(s2_26_5, DCT_CONST_BITS);
        const __m128i s2_27_6 = _mm_srai_epi32(s2_27_4, DCT_CONST_BITS);
        const __m128i s2_27_7 = _mm_srai_epi32(s2_27_5, DCT_CONST_BITS);
        // Combine
        step2[20] = _mm_packs_epi32(s2_20_6, s2_20_7);
        step2[21] = _mm_packs_epi32(s2_21_6, s2_21_7);
        step2[22] = _mm_packs_epi32(s2_22_6, s2_22_7);
        step2[23] = _mm_packs_epi32(s2_23_6, s2_23_7);
        step2[24] = _mm_packs_epi32(s2_24_6, s2_24_7);
        step2[25] = _mm_packs_epi32(s2_25_6, s2_25_7);
        step2[26] = _mm_packs_epi32(s2_26_6, s2_26_7);
        step2[27] = _mm_packs_epi32(s2_27_6, s2_27_7);
467
#if DCT_HIGH_BIT_DEPTH
468 469 470
        overflow = check_epi16_overflow_x8(&step2[20], &step2[21], &step2[22],
                                           &step2[23], &step2[24], &step2[25],
                                           &step2[26], &step2[27]);
471 472 473 474 475 476 477 478
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
479
      }
480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517

#if !FDCT32x32_HIGH_PRECISION
      // dump the magnitude by half, hence the intermediate values are within
      // the range of 16 bits.
      if (1 == pass) {
        __m128i s3_00_0 = _mm_cmplt_epi16(step2[ 0], kZero);
        __m128i s3_01_0 = _mm_cmplt_epi16(step2[ 1], kZero);
        __m128i s3_02_0 = _mm_cmplt_epi16(step2[ 2], kZero);
        __m128i s3_03_0 = _mm_cmplt_epi16(step2[ 3], kZero);
        __m128i s3_04_0 = _mm_cmplt_epi16(step2[ 4], kZero);
        __m128i s3_05_0 = _mm_cmplt_epi16(step2[ 5], kZero);
        __m128i s3_06_0 = _mm_cmplt_epi16(step2[ 6], kZero);
        __m128i s3_07_0 = _mm_cmplt_epi16(step2[ 7], kZero);
        __m128i s2_08_0 = _mm_cmplt_epi16(step2[ 8], kZero);
        __m128i s2_09_0 = _mm_cmplt_epi16(step2[ 9], kZero);
        __m128i s3_10_0 = _mm_cmplt_epi16(step2[10], kZero);
        __m128i s3_11_0 = _mm_cmplt_epi16(step2[11], kZero);
        __m128i s3_12_0 = _mm_cmplt_epi16(step2[12], kZero);
        __m128i s3_13_0 = _mm_cmplt_epi16(step2[13], kZero);
        __m128i s2_14_0 = _mm_cmplt_epi16(step2[14], kZero);
        __m128i s2_15_0 = _mm_cmplt_epi16(step2[15], kZero);
        __m128i s3_16_0 = _mm_cmplt_epi16(step1[16], kZero);
        __m128i s3_17_0 = _mm_cmplt_epi16(step1[17], kZero);
        __m128i s3_18_0 = _mm_cmplt_epi16(step1[18], kZero);
        __m128i s3_19_0 = _mm_cmplt_epi16(step1[19], kZero);
        __m128i s3_20_0 = _mm_cmplt_epi16(step2[20], kZero);
        __m128i s3_21_0 = _mm_cmplt_epi16(step2[21], kZero);
        __m128i s3_22_0 = _mm_cmplt_epi16(step2[22], kZero);
        __m128i s3_23_0 = _mm_cmplt_epi16(step2[23], kZero);
        __m128i s3_24_0 = _mm_cmplt_epi16(step2[24], kZero);
        __m128i s3_25_0 = _mm_cmplt_epi16(step2[25], kZero);
        __m128i s3_26_0 = _mm_cmplt_epi16(step2[26], kZero);
        __m128i s3_27_0 = _mm_cmplt_epi16(step2[27], kZero);
        __m128i s3_28_0 = _mm_cmplt_epi16(step1[28], kZero);
        __m128i s3_29_0 = _mm_cmplt_epi16(step1[29], kZero);
        __m128i s3_30_0 = _mm_cmplt_epi16(step1[30], kZero);
        __m128i s3_31_0 = _mm_cmplt_epi16(step1[31], kZero);

518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551
        step2[0] = SUB_EPI16(step2[ 0], s3_00_0);
        step2[1] = SUB_EPI16(step2[ 1], s3_01_0);
        step2[2] = SUB_EPI16(step2[ 2], s3_02_0);
        step2[3] = SUB_EPI16(step2[ 3], s3_03_0);
        step2[4] = SUB_EPI16(step2[ 4], s3_04_0);
        step2[5] = SUB_EPI16(step2[ 5], s3_05_0);
        step2[6] = SUB_EPI16(step2[ 6], s3_06_0);
        step2[7] = SUB_EPI16(step2[ 7], s3_07_0);
        step2[8] = SUB_EPI16(step2[ 8], s2_08_0);
        step2[9] = SUB_EPI16(step2[ 9], s2_09_0);
        step2[10] = SUB_EPI16(step2[10], s3_10_0);
        step2[11] = SUB_EPI16(step2[11], s3_11_0);
        step2[12] = SUB_EPI16(step2[12], s3_12_0);
        step2[13] = SUB_EPI16(step2[13], s3_13_0);
        step2[14] = SUB_EPI16(step2[14], s2_14_0);
        step2[15] = SUB_EPI16(step2[15], s2_15_0);
        step1[16] = SUB_EPI16(step1[16], s3_16_0);
        step1[17] = SUB_EPI16(step1[17], s3_17_0);
        step1[18] = SUB_EPI16(step1[18], s3_18_0);
        step1[19] = SUB_EPI16(step1[19], s3_19_0);
        step2[20] = SUB_EPI16(step2[20], s3_20_0);
        step2[21] = SUB_EPI16(step2[21], s3_21_0);
        step2[22] = SUB_EPI16(step2[22], s3_22_0);
        step2[23] = SUB_EPI16(step2[23], s3_23_0);
        step2[24] = SUB_EPI16(step2[24], s3_24_0);
        step2[25] = SUB_EPI16(step2[25], s3_25_0);
        step2[26] = SUB_EPI16(step2[26], s3_26_0);
        step2[27] = SUB_EPI16(step2[27], s3_27_0);
        step1[28] = SUB_EPI16(step1[28], s3_28_0);
        step1[29] = SUB_EPI16(step1[29], s3_29_0);
        step1[30] = SUB_EPI16(step1[30], s3_30_0);
        step1[31] = SUB_EPI16(step1[31], s3_31_0);
#if DCT_HIGH_BIT_DEPTH
        overflow = check_epi16_overflow_x32(
552 553 554 555 556 557 558 559
            &step2[0], &step2[1], &step2[2], &step2[3],
            &step2[4], &step2[5], &step2[6], &step2[7],
            &step2[8], &step2[9], &step2[10], &step2[11],
            &step2[12], &step2[13], &step2[14], &step2[15],
            &step1[16], &step1[17], &step1[18], &step1[19],
            &step2[20], &step2[21], &step2[22], &step2[23],
            &step2[24], &step2[25], &step2[26], &step2[27],
            &step1[28], &step1[29], &step1[30], &step1[31]);
560 561 562 563 564 565 566 567 568 569 570 571 572 573 574
        if (overflow) {
          HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
        step2[0] = _mm_add_epi16(step2[ 0], kOne);
        step2[1] = _mm_add_epi16(step2[ 1], kOne);
        step2[2] = _mm_add_epi16(step2[ 2], kOne);
        step2[3] = _mm_add_epi16(step2[ 3], kOne);
        step2[4] = _mm_add_epi16(step2[ 4], kOne);
        step2[5] = _mm_add_epi16(step2[ 5], kOne);
        step2[6] = _mm_add_epi16(step2[ 6], kOne);
        step2[7] = _mm_add_epi16(step2[ 7], kOne);
        step2[8] = _mm_add_epi16(step2[ 8], kOne);
        step2[9] = _mm_add_epi16(step2[ 9], kOne);
575 576 577 578 579 580 581 582 583 584 585 586 587 588 589 590 591 592 593 594 595 596 597
        step2[10] = _mm_add_epi16(step2[10], kOne);
        step2[11] = _mm_add_epi16(step2[11], kOne);
        step2[12] = _mm_add_epi16(step2[12], kOne);
        step2[13] = _mm_add_epi16(step2[13], kOne);
        step2[14] = _mm_add_epi16(step2[14], kOne);
        step2[15] = _mm_add_epi16(step2[15], kOne);
        step1[16] = _mm_add_epi16(step1[16], kOne);
        step1[17] = _mm_add_epi16(step1[17], kOne);
        step1[18] = _mm_add_epi16(step1[18], kOne);
        step1[19] = _mm_add_epi16(step1[19], kOne);
        step2[20] = _mm_add_epi16(step2[20], kOne);
        step2[21] = _mm_add_epi16(step2[21], kOne);
        step2[22] = _mm_add_epi16(step2[22], kOne);
        step2[23] = _mm_add_epi16(step2[23], kOne);
        step2[24] = _mm_add_epi16(step2[24], kOne);
        step2[25] = _mm_add_epi16(step2[25], kOne);
        step2[26] = _mm_add_epi16(step2[26], kOne);
        step2[27] = _mm_add_epi16(step2[27], kOne);
        step1[28] = _mm_add_epi16(step1[28], kOne);
        step1[29] = _mm_add_epi16(step1[29], kOne);
        step1[30] = _mm_add_epi16(step1[30], kOne);
        step1[31] = _mm_add_epi16(step1[31], kOne);

598 599 600 601 602 603 604 605 606 607
        step2[0] = _mm_srai_epi16(step2[ 0], 2);
        step2[1] = _mm_srai_epi16(step2[ 1], 2);
        step2[2] = _mm_srai_epi16(step2[ 2], 2);
        step2[3] = _mm_srai_epi16(step2[ 3], 2);
        step2[4] = _mm_srai_epi16(step2[ 4], 2);
        step2[5] = _mm_srai_epi16(step2[ 5], 2);
        step2[6] = _mm_srai_epi16(step2[ 6], 2);
        step2[7] = _mm_srai_epi16(step2[ 7], 2);
        step2[8] = _mm_srai_epi16(step2[ 8], 2);
        step2[9] = _mm_srai_epi16(step2[ 9], 2);
608 609 610 611 612 613 614 615 616 617 618 619 620 621 622 623 624 625 626 627 628 629 630
        step2[10] = _mm_srai_epi16(step2[10], 2);
        step2[11] = _mm_srai_epi16(step2[11], 2);
        step2[12] = _mm_srai_epi16(step2[12], 2);
        step2[13] = _mm_srai_epi16(step2[13], 2);
        step2[14] = _mm_srai_epi16(step2[14], 2);
        step2[15] = _mm_srai_epi16(step2[15], 2);
        step1[16] = _mm_srai_epi16(step1[16], 2);
        step1[17] = _mm_srai_epi16(step1[17], 2);
        step1[18] = _mm_srai_epi16(step1[18], 2);
        step1[19] = _mm_srai_epi16(step1[19], 2);
        step2[20] = _mm_srai_epi16(step2[20], 2);
        step2[21] = _mm_srai_epi16(step2[21], 2);
        step2[22] = _mm_srai_epi16(step2[22], 2);
        step2[23] = _mm_srai_epi16(step2[23], 2);
        step2[24] = _mm_srai_epi16(step2[24], 2);
        step2[25] = _mm_srai_epi16(step2[25], 2);
        step2[26] = _mm_srai_epi16(step2[26], 2);
        step2[27] = _mm_srai_epi16(step2[27], 2);
        step1[28] = _mm_srai_epi16(step1[28], 2);
        step1[29] = _mm_srai_epi16(step1[29], 2);
        step1[30] = _mm_srai_epi16(step1[30], 2);
        step1[31] = _mm_srai_epi16(step1[31], 2);
      }
631
#endif  // !FDCT32x32_HIGH_PRECISION
632 633 634 635

#if FDCT32x32_HIGH_PRECISION
      if (pass == 0) {
#endif
636 637
      // Stage 3
      {
638 639 640 641 642 643 644 645 646
        step3[0] = ADD_EPI16(step2[(8 - 1)], step2[0]);
        step3[1] = ADD_EPI16(step2[(8 - 2)], step2[1]);
        step3[2] = ADD_EPI16(step2[(8 - 3)], step2[2]);
        step3[3] = ADD_EPI16(step2[(8 - 4)], step2[3]);
        step3[4] = SUB_EPI16(step2[(8 - 5)], step2[4]);
        step3[5] = SUB_EPI16(step2[(8 - 6)], step2[5]);
        step3[6] = SUB_EPI16(step2[(8 - 7)], step2[6]);
        step3[7] = SUB_EPI16(step2[(8 - 8)], step2[7]);
#if DCT_HIGH_BIT_DEPTH
647 648 649
        overflow = check_epi16_overflow_x8(&step3[0], &step3[1], &step3[2],
                                           &step3[3], &step3[4], &step3[5],
                                           &step3[6], &step3[7]);
650 651 652 653 654 655 656 657
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
658 659 660 661 662 663 664 665 666 667 668 669 670 671 672 673 674 675 676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693
      }
      {
        const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
        const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
        const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
        const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
        const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
        const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
        const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
        const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
        const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
        const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
        const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
        const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
        // dct_const_round_shift
        const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
        const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
        const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
        const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
        const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
        const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
        const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
        const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
        const __m128i s3_10_6 = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
        const __m128i s3_10_7 = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
        const __m128i s3_11_6 = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
        const __m128i s3_11_7 = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
        const __m128i s3_12_6 = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
        const __m128i s3_12_7 = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
        const __m128i s3_13_6 = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
        const __m128i s3_13_7 = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
        // Combine
        step3[10] = _mm_packs_epi32(s3_10_6, s3_10_7);
        step3[11] = _mm_packs_epi32(s3_11_6, s3_11_7);
        step3[12] = _mm_packs_epi32(s3_12_6, s3_12_7);
        step3[13] = _mm_packs_epi32(s3_13_6, s3_13_7);
694
#if DCT_HIGH_BIT_DEPTH
695 696
        overflow = check_epi16_overflow_x4(&step3[10], &step3[11],
                                           &step3[12], &step3[13]);
697 698 699 700 701 702 703 704
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
705 706
      }
      {
707 708 709 710 711 712 713 714 715 716 717 718 719 720 721 722 723 724
        step3[16] = ADD_EPI16(step2[23], step1[16]);
        step3[17] = ADD_EPI16(step2[22], step1[17]);
        step3[18] = ADD_EPI16(step2[21], step1[18]);
        step3[19] = ADD_EPI16(step2[20], step1[19]);
        step3[20] = SUB_EPI16(step1[19], step2[20]);
        step3[21] = SUB_EPI16(step1[18], step2[21]);
        step3[22] = SUB_EPI16(step1[17], step2[22]);
        step3[23] = SUB_EPI16(step1[16], step2[23]);
        step3[24] = SUB_EPI16(step1[31], step2[24]);
        step3[25] = SUB_EPI16(step1[30], step2[25]);
        step3[26] = SUB_EPI16(step1[29], step2[26]);
        step3[27] = SUB_EPI16(step1[28], step2[27]);
        step3[28] = ADD_EPI16(step2[27], step1[28]);
        step3[29] = ADD_EPI16(step2[26], step1[29]);
        step3[30] = ADD_EPI16(step2[25], step1[30]);
        step3[31] = ADD_EPI16(step2[24], step1[31]);
#if DCT_HIGH_BIT_DEPTH
        overflow = check_epi16_overflow_x16(
725 726 727 728
            &step3[16], &step3[17], &step3[18], &step3[19],
            &step3[20], &step3[21], &step3[22], &step3[23],
            &step3[24], &step3[25], &step3[26], &step3[27],
            &step3[28], &step3[29], &step3[30], &step3[31]);
729 730 731 732 733 734 735 736
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
737
      }
738

739 740
      // Stage 4
      {
741 742 743 744 745 746 747 748 749 750 751 752 753 754
        step1[0] = ADD_EPI16(step3[ 3], step3[ 0]);
        step1[1] = ADD_EPI16(step3[ 2], step3[ 1]);
        step1[2] = SUB_EPI16(step3[ 1], step3[ 2]);
        step1[3] = SUB_EPI16(step3[ 0], step3[ 3]);
        step1[8] = ADD_EPI16(step3[11], step2[ 8]);
        step1[9] = ADD_EPI16(step3[10], step2[ 9]);
        step1[10] = SUB_EPI16(step2[ 9], step3[10]);
        step1[11] = SUB_EPI16(step2[ 8], step3[11]);
        step1[12] = SUB_EPI16(step2[15], step3[12]);
        step1[13] = SUB_EPI16(step2[14], step3[13]);
        step1[14] = ADD_EPI16(step3[13], step2[14]);
        step1[15] = ADD_EPI16(step3[12], step2[15]);
#if DCT_HIGH_BIT_DEPTH
        overflow = check_epi16_overflow_x16(
755 756 757 758
            &step1[0], &step1[1], &step1[2], &step1[3],
            &step1[4], &step1[5], &step1[6], &step1[7],
            &step1[8], &step1[9], &step1[10], &step1[11],
            &step1[12], &step1[13], &step1[14], &step1[15]);
759 760 761 762 763 764 765 766
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786
      }
      {
        const __m128i s1_05_0 = _mm_unpacklo_epi16(step3[6], step3[5]);
        const __m128i s1_05_1 = _mm_unpackhi_epi16(step3[6], step3[5]);
        const __m128i s1_05_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_m16);
        const __m128i s1_05_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_m16);
        const __m128i s1_06_2 = _mm_madd_epi16(s1_05_0, k__cospi_p16_p16);
        const __m128i s1_06_3 = _mm_madd_epi16(s1_05_1, k__cospi_p16_p16);
        // dct_const_round_shift
        const __m128i s1_05_4 = _mm_add_epi32(s1_05_2, k__DCT_CONST_ROUNDING);
        const __m128i s1_05_5 = _mm_add_epi32(s1_05_3, k__DCT_CONST_ROUNDING);
        const __m128i s1_06_4 = _mm_add_epi32(s1_06_2, k__DCT_CONST_ROUNDING);
        const __m128i s1_06_5 = _mm_add_epi32(s1_06_3, k__DCT_CONST_ROUNDING);
        const __m128i s1_05_6 = _mm_srai_epi32(s1_05_4, DCT_CONST_BITS);
        const __m128i s1_05_7 = _mm_srai_epi32(s1_05_5, DCT_CONST_BITS);
        const __m128i s1_06_6 = _mm_srai_epi32(s1_06_4, DCT_CONST_BITS);
        const __m128i s1_06_7 = _mm_srai_epi32(s1_06_5, DCT_CONST_BITS);
        // Combine
        step1[5] = _mm_packs_epi32(s1_05_6, s1_05_7);
        step1[6] = _mm_packs_epi32(s1_06_6, s1_06_7);
787
#if DCT_HIGH_BIT_DEPTH
788
        overflow = check_epi16_overflow_x2(&step1[5], &step1[6]);
789 790 791 792 793 794 795 796
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846 847 848 849 850 851 852 853 854 855 856 857 858 859 860 861 862 863 864
      }
      {
        const __m128i s1_18_0 = _mm_unpacklo_epi16(step3[18], step3[29]);
        const __m128i s1_18_1 = _mm_unpackhi_epi16(step3[18], step3[29]);
        const __m128i s1_19_0 = _mm_unpacklo_epi16(step3[19], step3[28]);
        const __m128i s1_19_1 = _mm_unpackhi_epi16(step3[19], step3[28]);
        const __m128i s1_20_0 = _mm_unpacklo_epi16(step3[20], step3[27]);
        const __m128i s1_20_1 = _mm_unpackhi_epi16(step3[20], step3[27]);
        const __m128i s1_21_0 = _mm_unpacklo_epi16(step3[21], step3[26]);
        const __m128i s1_21_1 = _mm_unpackhi_epi16(step3[21], step3[26]);
        const __m128i s1_18_2 = _mm_madd_epi16(s1_18_0, k__cospi_m08_p24);
        const __m128i s1_18_3 = _mm_madd_epi16(s1_18_1, k__cospi_m08_p24);
        const __m128i s1_19_2 = _mm_madd_epi16(s1_19_0, k__cospi_m08_p24);
        const __m128i s1_19_3 = _mm_madd_epi16(s1_19_1, k__cospi_m08_p24);
        const __m128i s1_20_2 = _mm_madd_epi16(s1_20_0, k__cospi_m24_m08);
        const __m128i s1_20_3 = _mm_madd_epi16(s1_20_1, k__cospi_m24_m08);
        const __m128i s1_21_2 = _mm_madd_epi16(s1_21_0, k__cospi_m24_m08);
        const __m128i s1_21_3 = _mm_madd_epi16(s1_21_1, k__cospi_m24_m08);
        const __m128i s1_26_2 = _mm_madd_epi16(s1_21_0, k__cospi_m08_p24);
        const __m128i s1_26_3 = _mm_madd_epi16(s1_21_1, k__cospi_m08_p24);
        const __m128i s1_27_2 = _mm_madd_epi16(s1_20_0, k__cospi_m08_p24);
        const __m128i s1_27_3 = _mm_madd_epi16(s1_20_1, k__cospi_m08_p24);
        const __m128i s1_28_2 = _mm_madd_epi16(s1_19_0, k__cospi_p24_p08);
        const __m128i s1_28_3 = _mm_madd_epi16(s1_19_1, k__cospi_p24_p08);
        const __m128i s1_29_2 = _mm_madd_epi16(s1_18_0, k__cospi_p24_p08);
        const __m128i s1_29_3 = _mm_madd_epi16(s1_18_1, k__cospi_p24_p08);
        // dct_const_round_shift
        const __m128i s1_18_4 = _mm_add_epi32(s1_18_2, k__DCT_CONST_ROUNDING);
        const __m128i s1_18_5 = _mm_add_epi32(s1_18_3, k__DCT_CONST_ROUNDING);
        const __m128i s1_19_4 = _mm_add_epi32(s1_19_2, k__DCT_CONST_ROUNDING);
        const __m128i s1_19_5 = _mm_add_epi32(s1_19_3, k__DCT_CONST_ROUNDING);
        const __m128i s1_20_4 = _mm_add_epi32(s1_20_2, k__DCT_CONST_ROUNDING);
        const __m128i s1_20_5 = _mm_add_epi32(s1_20_3, k__DCT_CONST_ROUNDING);
        const __m128i s1_21_4 = _mm_add_epi32(s1_21_2, k__DCT_CONST_ROUNDING);
        const __m128i s1_21_5 = _mm_add_epi32(s1_21_3, k__DCT_CONST_ROUNDING);
        const __m128i s1_26_4 = _mm_add_epi32(s1_26_2, k__DCT_CONST_ROUNDING);
        const __m128i s1_26_5 = _mm_add_epi32(s1_26_3, k__DCT_CONST_ROUNDING);
        const __m128i s1_27_4 = _mm_add_epi32(s1_27_2, k__DCT_CONST_ROUNDING);
        const __m128i s1_27_5 = _mm_add_epi32(s1_27_3, k__DCT_CONST_ROUNDING);
        const __m128i s1_28_4 = _mm_add_epi32(s1_28_2, k__DCT_CONST_ROUNDING);
        const __m128i s1_28_5 = _mm_add_epi32(s1_28_3, k__DCT_CONST_ROUNDING);
        const __m128i s1_29_4 = _mm_add_epi32(s1_29_2, k__DCT_CONST_ROUNDING);
        const __m128i s1_29_5 = _mm_add_epi32(s1_29_3, k__DCT_CONST_ROUNDING);
        const __m128i s1_18_6 = _mm_srai_epi32(s1_18_4, DCT_CONST_BITS);
        const __m128i s1_18_7 = _mm_srai_epi32(s1_18_5, DCT_CONST_BITS);
        const __m128i s1_19_6 = _mm_srai_epi32(s1_19_4, DCT_CONST_BITS);
        const __m128i s1_19_7 = _mm_srai_epi32(s1_19_5, DCT_CONST_BITS);
        const __m128i s1_20_6 = _mm_srai_epi32(s1_20_4, DCT_CONST_BITS);
        const __m128i s1_20_7 = _mm_srai_epi32(s1_20_5, DCT_CONST_BITS);
        const __m128i s1_21_6 = _mm_srai_epi32(s1_21_4, DCT_CONST_BITS);
        const __m128i s1_21_7 = _mm_srai_epi32(s1_21_5, DCT_CONST_BITS);
        const __m128i s1_26_6 = _mm_srai_epi32(s1_26_4, DCT_CONST_BITS);
        const __m128i s1_26_7 = _mm_srai_epi32(s1_26_5, DCT_CONST_BITS);
        const __m128i s1_27_6 = _mm_srai_epi32(s1_27_4, DCT_CONST_BITS);
        const __m128i s1_27_7 = _mm_srai_epi32(s1_27_5, DCT_CONST_BITS);
        const __m128i s1_28_6 = _mm_srai_epi32(s1_28_4, DCT_CONST_BITS);
        const __m128i s1_28_7 = _mm_srai_epi32(s1_28_5, DCT_CONST_BITS);
        const __m128i s1_29_6 = _mm_srai_epi32(s1_29_4, DCT_CONST_BITS);
        const __m128i s1_29_7 = _mm_srai_epi32(s1_29_5, DCT_CONST_BITS);
        // Combine
        step1[18] = _mm_packs_epi32(s1_18_6, s1_18_7);
        step1[19] = _mm_packs_epi32(s1_19_6, s1_19_7);
        step1[20] = _mm_packs_epi32(s1_20_6, s1_20_7);
        step1[21] = _mm_packs_epi32(s1_21_6, s1_21_7);
        step1[26] = _mm_packs_epi32(s1_26_6, s1_26_7);
        step1[27] = _mm_packs_epi32(s1_27_6, s1_27_7);
        step1[28] = _mm_packs_epi32(s1_28_6, s1_28_7);
        step1[29] = _mm_packs_epi32(s1_29_6, s1_29_7);
865
#if DCT_HIGH_BIT_DEPTH
866 867 868
        overflow = check_epi16_overflow_x8(&step1[18], &step1[19], &step1[20],
                                           &step1[21], &step1[26], &step1[27],
                                           &step1[28], &step1[29]);
869 870 871 872 873 874 875 876
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
877 878 879
      }
      // Stage 5
      {
880 881 882 883 884
        step2[4] = ADD_EPI16(step1[5], step3[4]);
        step2[5] = SUB_EPI16(step3[4], step1[5]);
        step2[6] = SUB_EPI16(step3[7], step1[6]);
        step2[7] = ADD_EPI16(step1[6], step3[7]);
#if DCT_HIGH_BIT_DEPTH
885 886
        overflow = check_epi16_overflow_x4(&step2[4], &step2[5],
                                           &step2[6], &step2[7]);
887 888 889 890 891 892 893 894
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930
      }
      {
        const __m128i out_00_0 = _mm_unpacklo_epi16(step1[0], step1[1]);
        const __m128i out_00_1 = _mm_unpackhi_epi16(step1[0], step1[1]);
        const __m128i out_08_0 = _mm_unpacklo_epi16(step1[2], step1[3]);
        const __m128i out_08_1 = _mm_unpackhi_epi16(step1[2], step1[3]);
        const __m128i out_00_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_p16);
        const __m128i out_00_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_p16);
        const __m128i out_16_2 = _mm_madd_epi16(out_00_0, k__cospi_p16_m16);
        const __m128i out_16_3 = _mm_madd_epi16(out_00_1, k__cospi_p16_m16);
        const __m128i out_08_2 = _mm_madd_epi16(out_08_0, k__cospi_p24_p08);
        const __m128i out_08_3 = _mm_madd_epi16(out_08_1, k__cospi_p24_p08);
        const __m128i out_24_2 = _mm_madd_epi16(out_08_0, k__cospi_m08_p24);
        const __m128i out_24_3 = _mm_madd_epi16(out_08_1, k__cospi_m08_p24);
        // dct_const_round_shift
        const __m128i out_00_4 = _mm_add_epi32(out_00_2, k__DCT_CONST_ROUNDING);
        const __m128i out_00_5 = _mm_add_epi32(out_00_3, k__DCT_CONST_ROUNDING);
        const __m128i out_16_4 = _mm_add_epi32(out_16_2, k__DCT_CONST_ROUNDING);
        const __m128i out_16_5 = _mm_add_epi32(out_16_3, k__DCT_CONST_ROUNDING);
        const __m128i out_08_4 = _mm_add_epi32(out_08_2, k__DCT_CONST_ROUNDING);
        const __m128i out_08_5 = _mm_add_epi32(out_08_3, k__DCT_CONST_ROUNDING);
        const __m128i out_24_4 = _mm_add_epi32(out_24_2, k__DCT_CONST_ROUNDING);
        const __m128i out_24_5 = _mm_add_epi32(out_24_3, k__DCT_CONST_ROUNDING);
        const __m128i out_00_6 = _mm_srai_epi32(out_00_4, DCT_CONST_BITS);
        const __m128i out_00_7 = _mm_srai_epi32(out_00_5, DCT_CONST_BITS);
        const __m128i out_16_6 = _mm_srai_epi32(out_16_4, DCT_CONST_BITS);
        const __m128i out_16_7 = _mm_srai_epi32(out_16_5, DCT_CONST_BITS);
        const __m128i out_08_6 = _mm_srai_epi32(out_08_4, DCT_CONST_BITS);
        const __m128i out_08_7 = _mm_srai_epi32(out_08_5, DCT_CONST_BITS);
        const __m128i out_24_6 = _mm_srai_epi32(out_24_4, DCT_CONST_BITS);
        const __m128i out_24_7 = _mm_srai_epi32(out_24_5, DCT_CONST_BITS);
        // Combine
        out[ 0] = _mm_packs_epi32(out_00_6, out_00_7);
        out[16] = _mm_packs_epi32(out_16_6, out_16_7);
        out[ 8] = _mm_packs_epi32(out_08_6, out_08_7);
        out[24] = _mm_packs_epi32(out_24_6, out_24_7);
931
#if DCT_HIGH_BIT_DEPTH
932 933
        overflow = check_epi16_overflow_x4(&out[0], &out[16],
                                           &out[8], &out[24]);
934 935 936 937 938 939 940 941
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977
      }
      {
        const __m128i s2_09_0 = _mm_unpacklo_epi16(step1[ 9], step1[14]);
        const __m128i s2_09_1 = _mm_unpackhi_epi16(step1[ 9], step1[14]);
        const __m128i s2_10_0 = _mm_unpacklo_epi16(step1[10], step1[13]);
        const __m128i s2_10_1 = _mm_unpackhi_epi16(step1[10], step1[13]);
        const __m128i s2_09_2 = _mm_madd_epi16(s2_09_0, k__cospi_m08_p24);
        const __m128i s2_09_3 = _mm_madd_epi16(s2_09_1, k__cospi_m08_p24);
        const __m128i s2_10_2 = _mm_madd_epi16(s2_10_0, k__cospi_m24_m08);
        const __m128i s2_10_3 = _mm_madd_epi16(s2_10_1, k__cospi_m24_m08);
        const __m128i s2_13_2 = _mm_madd_epi16(s2_10_0, k__cospi_m08_p24);
        const __m128i s2_13_3 = _mm_madd_epi16(s2_10_1, k__cospi_m08_p24);
        const __m128i s2_14_2 = _mm_madd_epi16(s2_09_0, k__cospi_p24_p08);
        const __m128i s2_14_3 = _mm_madd_epi16(s2_09_1, k__cospi_p24_p08);
        // dct_const_round_shift
        const __m128i s2_09_4 = _mm_add_epi32(s2_09_2, k__DCT_CONST_ROUNDING);
        const __m128i s2_09_5 = _mm_add_epi32(s2_09_3, k__DCT_CONST_ROUNDING);
        const __m128i s2_10_4 = _mm_add_epi32(s2_10_2, k__DCT_CONST_ROUNDING);
        const __m128i s2_10_5 = _mm_add_epi32(s2_10_3, k__DCT_CONST_ROUNDING);
        const __m128i s2_13_4 = _mm_add_epi32(s2_13_2, k__DCT_CONST_ROUNDING);
        const __m128i s2_13_5 = _mm_add_epi32(s2_13_3, k__DCT_CONST_ROUNDING);
        const __m128i s2_14_4 = _mm_add_epi32(s2_14_2, k__DCT_CONST_ROUNDING);
        const __m128i s2_14_5 = _mm_add_epi32(s2_14_3, k__DCT_CONST_ROUNDING);
        const __m128i s2_09_6 = _mm_srai_epi32(s2_09_4, DCT_CONST_BITS);
        const __m128i s2_09_7 = _mm_srai_epi32(s2_09_5, DCT_CONST_BITS);
        const __m128i s2_10_6 = _mm_srai_epi32(s2_10_4, DCT_CONST_BITS);
        const __m128i s2_10_7 = _mm_srai_epi32(s2_10_5, DCT_CONST_BITS);
        const __m128i s2_13_6 = _mm_srai_epi32(s2_13_4, DCT_CONST_BITS);
        const __m128i s2_13_7 = _mm_srai_epi32(s2_13_5, DCT_CONST_BITS);
        const __m128i s2_14_6 = _mm_srai_epi32(s2_14_4, DCT_CONST_BITS);
        const __m128i s2_14_7 = _mm_srai_epi32(s2_14_5, DCT_CONST_BITS);
        // Combine
        step2[ 9] = _mm_packs_epi32(s2_09_6, s2_09_7);
        step2[10] = _mm_packs_epi32(s2_10_6, s2_10_7);
        step2[13] = _mm_packs_epi32(s2_13_6, s2_13_7);
        step2[14] = _mm_packs_epi32(s2_14_6, s2_14_7);
978
#if DCT_HIGH_BIT_DEPTH
979 980
        overflow = check_epi16_overflow_x4(&step2[9], &step2[10],
                                           &step2[13], &step2[14]);
981 982 983 984 985 986 987 988
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
989 990
      }
      {
991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008
        step2[16] = ADD_EPI16(step1[19], step3[16]);
        step2[17] = ADD_EPI16(step1[18], step3[17]);
        step2[18] = SUB_EPI16(step3[17], step1[18]);
        step2[19] = SUB_EPI16(step3[16], step1[19]);
        step2[20] = SUB_EPI16(step3[23], step1[20]);
        step2[21] = SUB_EPI16(step3[22], step1[21]);
        step2[22] = ADD_EPI16(step1[21], step3[22]);
        step2[23] = ADD_EPI16(step1[20], step3[23]);
        step2[24] = ADD_EPI16(step1[27], step3[24]);
        step2[25] = ADD_EPI16(step1[26], step3[25]);
        step2[26] = SUB_EPI16(step3[25], step1[26]);
        step2[27] = SUB_EPI16(step3[24], step1[27]);
        step2[28] = SUB_EPI16(step3[31], step1[28]);
        step2[29] = SUB_EPI16(step3[30], step1[29]);
        step2[30] = ADD_EPI16(step1[29], step3[30]);
        step2[31] = ADD_EPI16(step1[28], step3[31]);
#if DCT_HIGH_BIT_DEPTH
        overflow = check_epi16_overflow_x16(
1009 1010 1011 1012
            &step2[16], &step2[17], &step2[18], &step2[19],
            &step2[20], &step2[21], &step2[22], &step2[23],
            &step2[24], &step2[25], &step2[26], &step2[27],
            &step2[28], &step2[29], &step2[30], &step2[31]);
1013 1014 1015 1016 1017 1018 1019 1020
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
1021 1022 1023 1024 1025 1026 1027 1028 1029 1030 1031 1032 1033 1034 1035 1036 1037 1038 1039 1040 1041 1042 1043 1044 1045 1046 1047 1048 1049 1050 1051 1052 1053 1054 1055 1056 1057
      }
      // Stage 6
      {
        const __m128i out_04_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
        const __m128i out_04_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
        const __m128i out_20_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
        const __m128i out_20_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
        const __m128i out_12_0 = _mm_unpacklo_epi16(step2[5], step2[6]);
        const __m128i out_12_1 = _mm_unpackhi_epi16(step2[5], step2[6]);
        const __m128i out_28_0 = _mm_unpacklo_epi16(step2[4], step2[7]);
        const __m128i out_28_1 = _mm_unpackhi_epi16(step2[4], step2[7]);
        const __m128i out_04_2 = _mm_madd_epi16(out_04_0, k__cospi_p28_p04);
        const __m128i out_04_3 = _mm_madd_epi16(out_04_1, k__cospi_p28_p04);
        const __m128i out_20_2 = _mm_madd_epi16(out_20_0, k__cospi_p12_p20);
        const __m128i out_20_3 = _mm_madd_epi16(out_20_1, k__cospi_p12_p20);
        const __m128i out_12_2 = _mm_madd_epi16(out_12_0, k__cospi_m20_p12);
        const __m128i out_12_3 = _mm_madd_epi16(out_12_1, k__cospi_m20_p12);
        const __m128i out_28_2 = _mm_madd_epi16(out_28_0, k__cospi_m04_p28);
        const __m128i out_28_3 = _mm_madd_epi16(out_28_1, k__cospi_m04_p28);
        // dct_const_round_shift
        const __m128i out_04_4 = _mm_add_epi32(out_04_2, k__DCT_CONST_ROUNDING);
        const __m128i out_04_5 = _mm_add_epi32(out_04_3, k__DCT_CONST_ROUNDING);
        const __m128i out_20_4 = _mm_add_epi32(out_20_2, k__DCT_CONST_ROUNDING);
        const __m128i out_20_5 = _mm_add_epi32(out_20_3, k__DCT_CONST_ROUNDING);
        const __m128i out_12_4 = _mm_add_epi32(out_12_2, k__DCT_CONST_ROUNDING);
        const __m128i out_12_5 = _mm_add_epi32(out_12_3, k__DCT_CONST_ROUNDING);
        const __m128i out_28_4 = _mm_add_epi32(out_28_2, k__DCT_CONST_ROUNDING);
        const __m128i out_28_5 = _mm_add_epi32(out_28_3, k__DCT_CONST_ROUNDING);
        const __m128i out_04_6 = _mm_srai_epi32(out_04_4, DCT_CONST_BITS);
        const __m128i out_04_7 = _mm_srai_epi32(out_04_5, DCT_CONST_BITS);
        const __m128i out_20_6 = _mm_srai_epi32(out_20_4, DCT_CONST_BITS);
        const __m128i out_20_7 = _mm_srai_epi32(out_20_5, DCT_CONST_BITS);
        const __m128i out_12_6 = _mm_srai_epi32(out_12_4, DCT_CONST_BITS);
        const __m128i out_12_7 = _mm_srai_epi32(out_12_5, DCT_CONST_BITS);
        const __m128i out_28_6 = _mm_srai_epi32(out_28_4, DCT_CONST_BITS);
        const __m128i out_28_7 = _mm_srai_epi32(out_28_5, DCT_CONST_BITS);
        // Combine
1058
        out[4] = _mm_packs_epi32(out_04_6, out_04_7);
1059 1060 1061
        out[20] = _mm_packs_epi32(out_20_6, out_20_7);
        out[12] = _mm_packs_epi32(out_12_6, out_12_7);
        out[28] = _mm_packs_epi32(out_28_6, out_28_7);
1062
#if DCT_HIGH_BIT_DEPTH
1063 1064
        overflow = check_epi16_overflow_x4(&out[4], &out[20],
                                           &out[12], &out[28]);
1065 1066 1067 1068 1069 1070 1071 1072
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
1073 1074
      }
      {
1075 1076 1077 1078 1079 1080 1081 1082 1083
        step3[8] = ADD_EPI16(step2[ 9], step1[ 8]);
        step3[9] = SUB_EPI16(step1[ 8], step2[ 9]);
        step3[10] = SUB_EPI16(step1[11], step2[10]);
        step3[11] = ADD_EPI16(step2[10], step1[11]);
        step3[12] = ADD_EPI16(step2[13], step1[12]);
        step3[13] = SUB_EPI16(step1[12], step2[13]);
        step3[14] = SUB_EPI16(step1[15], step2[14]);
        step3[15] = ADD_EPI16(step2[14], step1[15]);
#if DCT_HIGH_BIT_DEPTH
1084 1085 1086
        overflow = check_epi16_overflow_x8(&step3[8], &step3[9], &step3[10],
                                           &step3[11], &step3[12], &step3[13],
                                           &step3[14], &step3[15]);
1087 1088 1089 1090 1091 1092 1093 1094
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
1095 1096 1097 1098 1099 1100 1101 1102 1103 1104 1105 1106 1107 1108 1109 1110 1111 1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122 1123 1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141 1142 1143 1144 1145 1146 1147 1148 1149 1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163
      }
      {
        const __m128i s3_17_0 = _mm_unpacklo_epi16(step2[17], step2[30]);
        const __m128i s3_17_1 = _mm_unpackhi_epi16(step2[17], step2[30]);
        const __m128i s3_18_0 = _mm_unpacklo_epi16(step2[18], step2[29]);
        const __m128i s3_18_1 = _mm_unpackhi_epi16(step2[18], step2[29]);
        const __m128i s3_21_0 = _mm_unpacklo_epi16(step2[21], step2[26]);
        const __m128i s3_21_1 = _mm_unpackhi_epi16(step2[21], step2[26]);
        const __m128i s3_22_0 = _mm_unpacklo_epi16(step2[22], step2[25]);
        const __m128i s3_22_1 = _mm_unpackhi_epi16(step2[22], step2[25]);
        const __m128i s3_17_2 = _mm_madd_epi16(s3_17_0, k__cospi_m04_p28);
        const __m128i s3_17_3 = _mm_madd_epi16(s3_17_1, k__cospi_m04_p28);
        const __m128i s3_18_2 = _mm_madd_epi16(s3_18_0, k__cospi_m28_m04);
        const __m128i s3_18_3 = _mm_madd_epi16(s3_18_1, k__cospi_m28_m04);
        const __m128i s3_21_2 = _mm_madd_epi16(s3_21_0, k__cospi_m20_p12);
        const __m128i s3_21_3 = _mm_madd_epi16(s3_21_1, k__cospi_m20_p12);
        const __m128i s3_22_2 = _mm_madd_epi16(s3_22_0, k__cospi_m12_m20);
        const __m128i s3_22_3 = _mm_madd_epi16(s3_22_1, k__cospi_m12_m20);
        const __m128i s3_25_2 = _mm_madd_epi16(s3_22_0, k__cospi_m20_p12);
        const __m128i s3_25_3 = _mm_madd_epi16(s3_22_1, k__cospi_m20_p12);
        const __m128i s3_26_2 = _mm_madd_epi16(s3_21_0, k__cospi_p12_p20);
        const __m128i s3_26_3 = _mm_madd_epi16(s3_21_1, k__cospi_p12_p20);
        const __m128i s3_29_2 = _mm_madd_epi16(s3_18_0, k__cospi_m04_p28);
        const __m128i s3_29_3 = _mm_madd_epi16(s3_18_1, k__cospi_m04_p28);
        const __m128i s3_30_2 = _mm_madd_epi16(s3_17_0, k__cospi_p28_p04);
        const __m128i s3_30_3 = _mm_madd_epi16(s3_17_1, k__cospi_p28_p04);
        // dct_const_round_shift
        const __m128i s3_17_4 = _mm_add_epi32(s3_17_2, k__DCT_CONST_ROUNDING);
        const __m128i s3_17_5 = _mm_add_epi32(s3_17_3, k__DCT_CONST_ROUNDING);
        const __m128i s3_18_4 = _mm_add_epi32(s3_18_2, k__DCT_CONST_ROUNDING);
        const __m128i s3_18_5 = _mm_add_epi32(s3_18_3, k__DCT_CONST_ROUNDING);
        const __m128i s3_21_4 = _mm_add_epi32(s3_21_2, k__DCT_CONST_ROUNDING);
        const __m128i s3_21_5 = _mm_add_epi32(s3_21_3, k__DCT_CONST_ROUNDING);
        const __m128i s3_22_4 = _mm_add_epi32(s3_22_2, k__DCT_CONST_ROUNDING);
        const __m128i s3_22_5 = _mm_add_epi32(s3_22_3, k__DCT_CONST_ROUNDING);
        const __m128i s3_17_6 = _mm_srai_epi32(s3_17_4, DCT_CONST_BITS);
        const __m128i s3_17_7 = _mm_srai_epi32(s3_17_5, DCT_CONST_BITS);
        const __m128i s3_18_6 = _mm_srai_epi32(s3_18_4, DCT_CONST_BITS);
        const __m128i s3_18_7 = _mm_srai_epi32(s3_18_5, DCT_CONST_BITS);
        const __m128i s3_21_6 = _mm_srai_epi32(s3_21_4, DCT_CONST_BITS);
        const __m128i s3_21_7 = _mm_srai_epi32(s3_21_5, DCT_CONST_BITS);
        const __m128i s3_22_6 = _mm_srai_epi32(s3_22_4, DCT_CONST_BITS);
        const __m128i s3_22_7 = _mm_srai_epi32(s3_22_5, DCT_CONST_BITS);
        const __m128i s3_25_4 = _mm_add_epi32(s3_25_2, k__DCT_CONST_ROUNDING);
        const __m128i s3_25_5 = _mm_add_epi32(s3_25_3, k__DCT_CONST_ROUNDING);
        const __m128i s3_26_4 = _mm_add_epi32(s3_26_2, k__DCT_CONST_ROUNDING);
        const __m128i s3_26_5 = _mm_add_epi32(s3_26_3, k__DCT_CONST_ROUNDING);
        const __m128i s3_29_4 = _mm_add_epi32(s3_29_2, k__DCT_CONST_ROUNDING);
        const __m128i s3_29_5 = _mm_add_epi32(s3_29_3, k__DCT_CONST_ROUNDING);
        const __m128i s3_30_4 = _mm_add_epi32(s3_30_2, k__DCT_CONST_ROUNDING);
        const __m128i s3_30_5 = _mm_add_epi32(s3_30_3, k__DCT_CONST_ROUNDING);
        const __m128i s3_25_6 = _mm_srai_epi32(s3_25_4, DCT_CONST_BITS);
        const __m128i s3_25_7 = _mm_srai_epi32(s3_25_5, DCT_CONST_BITS);
        const __m128i s3_26_6 = _mm_srai_epi32(s3_26_4, DCT_CONST_BITS);
        const __m128i s3_26_7 = _mm_srai_epi32(s3_26_5, DCT_CONST_BITS);
        const __m128i s3_29_6 = _mm_srai_epi32(s3_29_4, DCT_CONST_BITS);
        const __m128i s3_29_7 = _mm_srai_epi32(s3_29_5, DCT_CONST_BITS);
        const __m128i s3_30_6 = _mm_srai_epi32(s3_30_4, DCT_CONST_BITS);
        const __m128i s3_30_7 = _mm_srai_epi32(s3_30_5, DCT_CONST_BITS);
        // Combine
        step3[17] = _mm_packs_epi32(s3_17_6, s3_17_7);
        step3[18] = _mm_packs_epi32(s3_18_6, s3_18_7);
        step3[21] = _mm_packs_epi32(s3_21_6, s3_21_7);
        step3[22] = _mm_packs_epi32(s3_22_6, s3_22_7);
        // Combine
        step3[25] = _mm_packs_epi32(s3_25_6, s3_25_7);
        step3[26] = _mm_packs_epi32(s3_26_6, s3_26_7);
        step3[29] = _mm_packs_epi32(s3_29_6, s3_29_7);
        step3[30] = _mm_packs_epi32(s3_30_6, s3_30_7);
1164
#if DCT_HIGH_BIT_DEPTH
1165 1166 1167
        overflow = check_epi16_overflow_x8(&step3[17], &step3[18], &step3[21],
                                           &step3[22], &step3[25], &step3[26],
                                           &step3[29], &step3[30]);
1168 1169 1170 1171 1172 1173 1174 1175
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
1176 1177 1178 1179 1180 1181 1182 1183 1184 1185 1186 1187 1188 1189 1190 1191 1192 1193 1194 1195 1196 1197 1198 1199 1200 1201 1202 1203 1204 1205 1206 1207 1208 1209 1210 1211 1212 1213 1214 1215 1216 1217 1218 1219 1220 1221 1222 1223 1224 1225 1226 1227 1228 1229 1230 1231 1232 1233 1234 1235 1236 1237 1238 1239 1240 1241 1242 1243 1244
      }
      // Stage 7
      {
        const __m128i out_02_0 = _mm_unpacklo_epi16(step3[ 8], step3[15]);
        const __m128i out_02_1 = _mm_unpackhi_epi16(step3[ 8], step3[15]);
        const __m128i out_18_0 = _mm_unpacklo_epi16(step3[ 9], step3[14]);
        const __m128i out_18_1 = _mm_unpackhi_epi16(step3[ 9], step3[14]);
        const __m128i out_10_0 = _mm_unpacklo_epi16(step3[10], step3[13]);
        const __m128i out_10_1 = _mm_unpackhi_epi16(step3[10], step3[13]);
        const __m128i out_26_0 = _mm_unpacklo_epi16(step3[11], step3[12]);
        const __m128i out_26_1 = _mm_unpackhi_epi16(step3[11], step3[12]);
        const __m128i out_02_2 = _mm_madd_epi16(out_02_0, k__cospi_p30_p02);
        const __m128i out_02_3 = _mm_madd_epi16(out_02_1, k__cospi_p30_p02);
        const __m128i out_18_2 = _mm_madd_epi16(out_18_0, k__cospi_p14_p18);
        const __m128i out_18_3 = _mm_madd_epi16(out_18_1, k__cospi_p14_p18);
        const __m128i out_10_2 = _mm_madd_epi16(out_10_0, k__cospi_p22_p10);
        const __m128i out_10_3 = _mm_madd_epi16(out_10_1, k__cospi_p22_p10);
        const __m128i out_26_2 = _mm_madd_epi16(out_26_0, k__cospi_p06_p26);
        const __m128i out_26_3 = _mm_madd_epi16(out_26_1, k__cospi_p06_p26);
        const __m128i out_06_2 = _mm_madd_epi16(out_26_0, k__cospi_m26_p06);
        const __m128i out_06_3 = _mm_madd_epi16(out_26_1, k__cospi_m26_p06);
        const __m128i out_22_2 = _mm_madd_epi16(out_10_0, k__cospi_m10_p22);
        const __m128i out_22_3 = _mm_madd_epi16(out_10_1, k__cospi_m10_p22);
        const __m128i out_14_2 = _mm_madd_epi16(out_18_0, k__cospi_m18_p14);
        const __m128i out_14_3 = _mm_madd_epi16(out_18_1, k__cospi_m18_p14);
        const __m128i out_30_2 = _mm_madd_epi16(out_02_0, k__cospi_m02_p30);
        const __m128i out_30_3 = _mm_madd_epi16(out_02_1, k__cospi_m02_p30);
        // dct_const_round_shift
        const __m128i out_02_4 = _mm_add_epi32(out_02_2, k__DCT_CONST_ROUNDING);
        const __m128i out_02_5 = _mm_add_epi32(out_02_3, k__DCT_CONST_ROUNDING);
        const __m128i out_18_4 = _mm_add_epi32(out_18_2, k__DCT_CONST_ROUNDING);
        const __m128i out_18_5 = _mm_add_epi32(out_18_3, k__DCT_CONST_ROUNDING);
        const __m128i out_10_4 = _mm_add_epi32(out_10_2, k__DCT_CONST_ROUNDING);
        const __m128i out_10_5 = _mm_add_epi32(out_10_3, k__DCT_CONST_ROUNDING);
        const __m128i out_26_4 = _mm_add_epi32(out_26_2, k__DCT_CONST_ROUNDING);
        const __m128i out_26_5 = _mm_add_epi32(out_26_3, k__DCT_CONST_ROUNDING);
        const __m128i out_06_4 = _mm_add_epi32(out_06_2, k__DCT_CONST_ROUNDING);
        const __m128i out_06_5 = _mm_add_epi32(out_06_3, k__DCT_CONST_ROUNDING);
        const __m128i out_22_4 = _mm_add_epi32(out_22_2, k__DCT_CONST_ROUNDING);
        const __m128i out_22_5 = _mm_add_epi32(out_22_3, k__DCT_CONST_ROUNDING);
        const __m128i out_14_4 = _mm_add_epi32(out_14_2, k__DCT_CONST_ROUNDING);
        const __m128i out_14_5 = _mm_add_epi32(out_14_3, k__DCT_CONST_ROUNDING);
        const __m128i out_30_4 = _mm_add_epi32(out_30_2, k__DCT_CONST_ROUNDING);
        const __m128i out_30_5 = _mm_add_epi32(out_30_3, k__DCT_CONST_ROUNDING);
        const __m128i out_02_6 = _mm_srai_epi32(out_02_4, DCT_CONST_BITS);
        const __m128i out_02_7 = _mm_srai_epi32(out_02_5, DCT_CONST_BITS);
        const __m128i out_18_6 = _mm_srai_epi32(out_18_4, DCT_CONST_BITS);
        const __m128i out_18_7 = _mm_srai_epi32(out_18_5, DCT_CONST_BITS);
        const __m128i out_10_6 = _mm_srai_epi32(out_10_4, DCT_CONST_BITS);
        const __m128i out_10_7 = _mm_srai_epi32(out_10_5, DCT_CONST_BITS);
        const __m128i out_26_6 = _mm_srai_epi32(out_26_4, DCT_CONST_BITS);
        const __m128i out_26_7 = _mm_srai_epi32(out_26_5, DCT_CONST_BITS);
        const __m128i out_06_6 = _mm_srai_epi32(out_06_4, DCT_CONST_BITS);
        const __m128i out_06_7 = _mm_srai_epi32(out_06_5, DCT_CONST_BITS);
        const __m128i out_22_6 = _mm_srai_epi32(out_22_4, DCT_CONST_BITS);
        const __m128i out_22_7 = _mm_srai_epi32(out_22_5, DCT_CONST_BITS);
        const __m128i out_14_6 = _mm_srai_epi32(out_14_4, DCT_CONST_BITS);
        const __m128i out_14_7 = _mm_srai_epi32(out_14_5, DCT_CONST_BITS);
        const __m128i out_30_6 = _mm_srai_epi32(out_30_4, DCT_CONST_BITS);
        const __m128i out_30_7 = _mm_srai_epi32(out_30_5, DCT_CONST_BITS);
        // Combine
        out[ 2] = _mm_packs_epi32(out_02_6, out_02_7);
        out[18] = _mm_packs_epi32(out_18_6, out_18_7);
        out[10] = _mm_packs_epi32(out_10_6, out_10_7);
        out[26] = _mm_packs_epi32(out_26_6, out_26_7);
        out[ 6] = _mm_packs_epi32(out_06_6, out_06_7);
        out[22] = _mm_packs_epi32(out_22_6, out_22_7);
        out[14] = _mm_packs_epi32(out_14_6, out_14_7);
        out[30] = _mm_packs_epi32(out_30_6, out_30_7);
1245
#if DCT_HIGH_BIT_DEPTH
1246 1247 1248
        overflow = check_epi16_overflow_x8(&out[2], &out[18], &out[10],
                                           &out[26], &out[6], &out[22],
                                           &out[14], &out[30]);
1249 1250 1251 1252 1253 1254 1255 1256
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
1257 1258
      }
      {
1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276
        step1[16] = ADD_EPI16(step3[17], step2[16]);
        step1[17] = SUB_EPI16(step2[16], step3[17]);
        step1[18] = SUB_EPI16(step2[19], step3[18]);
        step1[19] = ADD_EPI16(step3[18], step2[19]);
        step1[20] = ADD_EPI16(step3[21], step2[20]);
        step1[21] = SUB_EPI16(step2[20], step3[21]);
        step1[22] = SUB_EPI16(step2[23], step3[22]);
        step1[23] = ADD_EPI16(step3[22], step2[23]);
        step1[24] = ADD_EPI16(step3[25], step2[24]);
        step1[25] = SUB_EPI16(step2[24], step3[25]);
        step1[26] = SUB_EPI16(step2[27], step3[26]);
        step1[27] = ADD_EPI16(step3[26], step2[27]);
        step1[28] = ADD_EPI16(step3[29], step2[28]);
        step1[29] = SUB_EPI16(step2[28], step3[29]);
        step1[30] = SUB_EPI16(step2[31], step3[30]);
        step1[31] = ADD_EPI16(step3[30], step2[31]);
#if DCT_HIGH_BIT_DEPTH
        overflow = check_epi16_overflow_x16(
1277 1278 1279 1280
            &step1[16], &step1[17], &step1[18], &step1[19],
            &step1[20], &step1[21], &step1[22], &step1[23],
            &step1[24], &step1[25], &step1[26], &step1[27],
            &step1[28], &step1[29], &step1[30], &step1[31]);
1281 1282 1283 1284 1285 1286 1287 1288
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
             HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
1289 1290 1291 1292 1293 1294 1295 1296 1297 1298 1299 1300 1301 1302 1303 1304 1305 1306 1307 1308 1309 1310 1311 1312 1313 1314 1315 1316 1317 1318 1319 1320 1321 1322 1323 1324 1325 1326 1327 1328 1329 1330 1331 1332 1333 1334 1335 1336 1337 1338 1339 1340 1341 1342 1343 1344 1345 1346 1347 1348 1349 1350 1351 1352 1353 1354 1355 1356 1357
      }
      // Final stage --- outputs indices are bit-reversed.
      {
        const __m128i out_01_0 = _mm_unpacklo_epi16(step1[16], step1[31]);
        const __m128i out_01_1 = _mm_unpackhi_epi16(step1[16], step1[31]);
        const __m128i out_17_0 = _mm_unpacklo_epi16(step1[17], step1[30]);
        const __m128i out_17_1 = _mm_unpackhi_epi16(step1[17], step1[30]);
        const __m128i out_09_0 = _mm_unpacklo_epi16(step1[18], step1[29]);
        const __m128i out_09_1 = _mm_unpackhi_epi16(step1[18], step1[29]);
        const __m128i out_25_0 = _mm_unpacklo_epi16(step1[19], step1[28]);
        const __m128i out_25_1 = _mm_unpackhi_epi16(step1[19], step1[28]);
        const __m128i out_01_2 = _mm_madd_epi16(out_01_0, k__cospi_p31_p01);
        const __m128i out_01_3 = _mm_madd_epi16(out_01_1, k__cospi_p31_p01);
        const __m128i out_17_2 = _mm_madd_epi16(out_17_0, k__cospi_p15_p17);
        const __m128i out_17_3 = _mm_madd_epi16(out_17_1, k__cospi_p15_p17);
        const __m128i out_09_2 = _mm_madd_epi16(out_09_0, k__cospi_p23_p09);
        const __m128i out_09_3 = _mm_madd_epi16(out_09_1, k__cospi_p23_p09);
        const __m128i out_25_2 = _mm_madd_epi16(out_25_0, k__cospi_p07_p25);
        const __m128i out_25_3 = _mm_madd_epi16(out_25_1, k__cospi_p07_p25);
        const __m128i out_07_2 = _mm_madd_epi16(out_25_0, k__cospi_m25_p07);
        const __m128i out_07_3 = _mm_madd_epi16(out_25_1, k__cospi_m25_p07);
        const __m128i out_23_2 = _mm_madd_epi16(out_09_0, k__cospi_m09_p23);
        const __m128i out_23_3 = _mm_madd_epi16(out_09_1, k__cospi_m09_p23);
        const __m128i out_15_2 = _mm_madd_epi16(out_17_0, k__cospi_m17_p15);
        const __m128i out_15_3 = _mm_madd_epi16(out_17_1, k__cospi_m17_p15);
        const __m128i out_31_2 = _mm_madd_epi16(out_01_0, k__cospi_m01_p31);
        const __m128i out_31_3 = _mm_madd_epi16(out_01_1, k__cospi_m01_p31);
        // dct_const_round_shift
        const __m128i out_01_4 = _mm_add_epi32(out_01_2, k__DCT_CONST_ROUNDING);
        const __m128i out_01_5 = _mm_add_epi32(out_01_3, k__DCT_CONST_ROUNDING);
        const __m128i out_17_4 = _mm_add_epi32(out_17_2, k__DCT_CONST_ROUNDING);
        const __m128i out_17_5 = _mm_add_epi32(out_17_3, k__DCT_CONST_ROUNDING);
        const __m128i out_09_4 = _mm_add_epi32(out_09_2, k__DCT_CONST_ROUNDING);
        const __m128i out_09_5 = _mm_add_epi32(out_09_3, k__DCT_CONST_ROUNDING);
        const __m128i out_25_4 = _mm_add_epi32(out_25_2, k__DCT_CONST_ROUNDING);
        const __m128i out_25_5 = _mm_add_epi32(out_25_3, k__DCT_CONST_ROUNDING);
        const __m128i out_07_4 = _mm_add_epi32(out_07_2, k__DCT_CONST_ROUNDING);
        const __m128i out_07_5 = _mm_add_epi32(out_07_3, k__DCT_CONST_ROUNDING);
        const __m128i out_23_4 = _mm_add_epi32(out_23_2, k__DCT_CONST_ROUNDING);
        const __m128i out_23_5 = _mm_add_epi32(out_23_3, k__DCT_CONST_ROUNDING);
        const __m128i out_15_4 = _mm_add_epi32(out_15_2, k__DCT_CONST_ROUNDING);
        const __m128i out_15_5 = _mm_add_epi32(out_15_3, k__DCT_CONST_ROUNDING);
        const __m128i out_31_4 = _mm_add_epi32(out_31_2, k__DCT_CONST_ROUNDING);
        const __m128i out_31_5 = _mm_add_epi32(out_31_3, k__DCT_CONST_ROUNDING);
        const __m128i out_01_6 = _mm_srai_epi32(out_01_4, DCT_CONST_BITS);
        const __m128i out_01_7 = _mm_srai_epi32(out_01_5, DCT_CONST_BITS);
        const __m128i out_17_6 = _mm_srai_epi32(out_17_4, DCT_CONST_BITS);
        const __m128i out_17_7 = _mm_srai_epi32(out_17_5, DCT_CONST_BITS);
        const __m128i out_09_6 = _mm_srai_epi32(out_09_4, DCT_CONST_BITS);
        const __m128i out_09_7 = _mm_srai_epi32(out_09_5, DCT_CONST_BITS);
        const __m128i out_25_6 = _mm_srai_epi32(out_25_4, DCT_CONST_BITS);
        const __m128i out_25_7 = _mm_srai_epi32(out_25_5, DCT_CONST_BITS);
        const __m128i out_07_6 = _mm_srai_epi32(out_07_4, DCT_CONST_BITS);
        const __m128i out_07_7 = _mm_srai_epi32(out_07_5, DCT_CONST_BITS);
        const __m128i out_23_6 = _mm_srai_epi32(out_23_4, DCT_CONST_BITS);
        const __m128i out_23_7 = _mm_srai_epi32(out_23_5, DCT_CONST_BITS);
        const __m128i out_15_6 = _mm_srai_epi32(out_15_4, DCT_CONST_BITS);
        const __m128i out_15_7 = _mm_srai_epi32(out_15_5, DCT_CONST_BITS);
        const __m128i out_31_6 = _mm_srai_epi32(out_31_4, DCT_CONST_BITS);
        const __m128i out_31_7 = _mm_srai_epi32(out_31_5, DCT_CONST_BITS);
        // Combine
        out[ 1] = _mm_packs_epi32(out_01_6, out_01_7);
        out[17] = _mm_packs_epi32(out_17_6, out_17_7);
        out[ 9] = _mm_packs_epi32(out_09_6, out_09_7);
        out[25] = _mm_packs_epi32(out_25_6, out_25_7);
        out[ 7] = _mm_packs_epi32(out_07_6, out_07_7);
        out[23] = _mm_packs_epi32(out_23_6, out_23_7);
        out[15] = _mm_packs_epi32(out_15_6, out_15_7);
        out[31] = _mm_packs_epi32(out_31_6, out_31_7);
1358
#if DCT_HIGH_BIT_DEPTH
1359 1360 1361
        overflow = check_epi16_overflow_x8(&out[1], &out[17], &out[9],
                                           &out[25], &out[7], &out[23],
                                           &out[15], &out[31]);
1362 1363 1364 1365 1366 1367 1368 1369
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
1370 1371 1372 1373 1374 1375 1376 1377 1378 1379 1380 1381 1382 1383 1384 1385 1386 1387 1388 1389 1390 1391 1392 1393 1394 1395 1396 1397 1398 1399 1400 1401 1402 1403 1404 1405 1406 1407 1408 1409 1410 1411 1412 1413 1414 1415 1416 1417 1418 1419 1420 1421 1422 1423 1424 1425 1426 1427 1428 1429 1430 1431 1432 1433 1434 1435 1436 1437
      }
      {
        const __m128i out_05_0 = _mm_unpacklo_epi16(step1[20], step1[27]);
        const __m128i out_05_1 = _mm_unpackhi_epi16(step1[20], step1[27]);
        const __m128i out_21_0 = _mm_unpacklo_epi16(step1[21], step1[26]);
        const __m128i out_21_1 = _mm_unpackhi_epi16(step1[21], step1[26]);
        const __m128i out_13_0 = _mm_unpacklo_epi16(step1[22], step1[25]);
        const __m128i out_13_1 = _mm_unpackhi_epi16(step1[22], step1[25]);
        const __m128i out_29_0 = _mm_unpacklo_epi16(step1[23], step1[24]);
        const __m128i out_29_1 = _mm_unpackhi_epi16(step1[23], step1[24]);
        const __m128i out_05_2 = _mm_madd_epi16(out_05_0, k__cospi_p27_p05);
        const __m128i out_05_3 = _mm_madd_epi16(out_05_1, k__cospi_p27_p05);
        const __m128i out_21_2 = _mm_madd_epi16(out_21_0, k__cospi_p11_p21);
        const __m128i out_21_3 = _mm_madd_epi16(out_21_1, k__cospi_p11_p21);
        const __m128i out_13_2 = _mm_madd_epi16(out_13_0, k__cospi_p19_p13);
        const __m128i out_13_3 = _mm_madd_epi16(out_13_1, k__cospi_p19_p13);
        const __m128i out_29_2 = _mm_madd_epi16(out_29_0, k__cospi_p03_p29);
        const __m128i out_29_3 = _mm_madd_epi16(out_29_1, k__cospi_p03_p29);
        const __m128i out_03_2 = _mm_madd_epi16(out_29_0, k__cospi_m29_p03);
        const __m128i out_03_3 = _mm_madd_epi16(out_29_1, k__cospi_m29_p03);
        const __m128i out_19_2 = _mm_madd_epi16(out_13_0, k__cospi_m13_p19);
        const __m128i out_19_3 = _mm_madd_epi16(out_13_1, k__cospi_m13_p19);
        const __m128i out_11_2 = _mm_madd_epi16(out_21_0, k__cospi_m21_p11);
        const __m128i out_11_3 = _mm_madd_epi16(out_21_1, k__cospi_m21_p11);
        const __m128i out_27_2 = _mm_madd_epi16(out_05_0, k__cospi_m05_p27);
        const __m128i out_27_3 = _mm_madd_epi16(out_05_1, k__cospi_m05_p27);
        // dct_const_round_shift
        const __m128i out_05_4 = _mm_add_epi32(out_05_2, k__DCT_CONST_ROUNDING);
        const __m128i out_05_5 = _mm_add_epi32(out_05_3, k__DCT_CONST_ROUNDING);
        const __m128i out_21_4 = _mm_add_epi32(out_21_2, k__DCT_CONST_ROUNDING);
        const __m128i out_21_5 = _mm_add_epi32(out_21_3, k__DCT_CONST_ROUNDING);
        const __m128i out_13_4 = _mm_add_epi32(out_13_2, k__DCT_CONST_ROUNDING);
        const __m128i out_13_5 = _mm_add_epi32(out_13_3, k__DCT_CONST_ROUNDING);
        const __m128i out_29_4 = _mm_add_epi32(out_29_2, k__DCT_CONST_ROUNDING);
        const __m128i out_29_5 = _mm_add_epi32(out_29_3, k__DCT_CONST_ROUNDING);
        const __m128i out_03_4 = _mm_add_epi32(out_03_2, k__DCT_CONST_ROUNDING);
        const __m128i out_03_5 = _mm_add_epi32(out_03_3, k__DCT_CONST_ROUNDING);
        const __m128i out_19_4 = _mm_add_epi32(out_19_2, k__DCT_CONST_ROUNDING);
        const __m128i out_19_5 = _mm_add_epi32(out_19_3, k__DCT_CONST_ROUNDING);
        const __m128i out_11_4 = _mm_add_epi32(out_11_2, k__DCT_CONST_ROUNDING);
        const __m128i out_11_5 = _mm_add_epi32(out_11_3, k__DCT_CONST_ROUNDING);
        const __m128i out_27_4 = _mm_add_epi32(out_27_2, k__DCT_CONST_ROUNDING);
        const __m128i out_27_5 = _mm_add_epi32(out_27_3, k__DCT_CONST_ROUNDING);
        const __m128i out_05_6 = _mm_srai_epi32(out_05_4, DCT_CONST_BITS);
        const __m128i out_05_7 = _mm_srai_epi32(out_05_5, DCT_CONST_BITS);
        const __m128i out_21_6 = _mm_srai_epi32(out_21_4, DCT_CONST_BITS);
        const __m128i out_21_7 = _mm_srai_epi32(out_21_5, DCT_CONST_BITS);
        const __m128i out_13_6 = _mm_srai_epi32(out_13_4, DCT_CONST_BITS);
        const __m128i out_13_7 = _mm_srai_epi32(out_13_5, DCT_CONST_BITS);
        const __m128i out_29_6 = _mm_srai_epi32(out_29_4, DCT_CONST_BITS);
        const __m128i out_29_7 = _mm_srai_epi32(out_29_5, DCT_CONST_BITS);
        const __m128i out_03_6 = _mm_srai_epi32(out_03_4, DCT_CONST_BITS);
        const __m128i out_03_7 = _mm_srai_epi32(out_03_5, DCT_CONST_BITS);
        const __m128i out_19_6 = _mm_srai_epi32(out_19_4, DCT_CONST_BITS);
        const __m128i out_19_7 = _mm_srai_epi32(out_19_5, DCT_CONST_BITS);
        const __m128i out_11_6 = _mm_srai_epi32(out_11_4, DCT_CONST_BITS);
        const __m128i out_11_7 = _mm_srai_epi32(out_11_5, DCT_CONST_BITS);
        const __m128i out_27_6 = _mm_srai_epi32(out_27_4, DCT_CONST_BITS);
        const __m128i out_27_7 = _mm_srai_epi32(out_27_5, DCT_CONST_BITS);
        // Combine
        out[ 5] = _mm_packs_epi32(out_05_6, out_05_7);
        out[21] = _mm_packs_epi32(out_21_6, out_21_7);
        out[13] = _mm_packs_epi32(out_13_6, out_13_7);
        out[29] = _mm_packs_epi32(out_29_6, out_29_7);
        out[ 3] = _mm_packs_epi32(out_03_6, out_03_7);
        out[19] = _mm_packs_epi32(out_19_6, out_19_7);
        out[11] = _mm_packs_epi32(out_11_6, out_11_7);
        out[27] = _mm_packs_epi32(out_27_6, out_27_7);
1438
#if DCT_HIGH_BIT_DEPTH
1439 1440 1441
        overflow = check_epi16_overflow_x8(&out[5], &out[21], &out[13],
                                           &out[29], &out[3], &out[19],
                                           &out[11], &out[27]);
1442 1443 1444 1445 1446 1447 1448 1449
        if (overflow) {
          if (pass == 0)
            HIGH_FDCT32x32_2D_C(input, output_org, stride);
          else
            HIGH_FDCT32x32_2D_ROWS_C(intermediate, output_org);
          return;
        }
#endif  // DCT_HIGH_BIT_DEPTH
1450
      }
1451 1452 1453 1454 1455 1456
#if FDCT32x32_HIGH_PRECISION
      } else {
        __m128i lstep1[64], lstep2[64], lstep3[64];
        __m128i u[32], v[32], sign[16];
        const __m128i K32One = _mm_set_epi32(1, 1, 1, 1);
        // start using 32-bit operations
1457 1458 1459
        // stage 3
        {
          // expanding to 32-bit length priori to addition operations
1460 1461 1462 1463 1464 1465 1466 1467 1468 1469 1470 1471 1472 1473 1474 1475 1476 1477 1478 1479 1480 1481 1482 1483 1484 1485 1486 1487 1488 1489 1490 1491
          lstep2[ 0] = _mm_unpacklo_epi16(step2[ 0], kZero);
          lstep2[ 1] = _mm_unpackhi_epi16(step2[ 0], kZero);
          lstep2[ 2] = _mm_unpacklo_epi16(step2[ 1], kZero);
          lstep2[ 3] = _mm_unpackhi_epi16(step2[ 1], kZero);
          lstep2[ 4] = _mm_unpacklo_epi16(step2[ 2], kZero);
          lstep2[ 5] = _mm_unpackhi_epi16(step2[ 2], kZero);
          lstep2[ 6] = _mm_unpacklo_epi16(step2[ 3], kZero);
          lstep2[ 7] = _mm_unpackhi_epi16(step2[ 3], kZero);
          lstep2[ 8] = _mm_unpacklo_epi16(step2[ 4], kZero);
          lstep2[ 9] = _mm_unpackhi_epi16(step2[ 4], kZero);
          lstep2[10] = _mm_unpacklo_epi16(step2[ 5], kZero);
          lstep2[11] = _mm_unpackhi_epi16(step2[ 5], kZero);
          lstep2[12] = _mm_unpacklo_epi16(step2[ 6], kZero);
          lstep2[13] = _mm_unpackhi_epi16(step2[ 6], kZero);
          lstep2[14] = _mm_unpacklo_epi16(step2[ 7], kZero);
          lstep2[15] = _mm_unpackhi_epi16(step2[ 7], kZero);
          lstep2[ 0] = _mm_madd_epi16(lstep2[ 0], kOne);
          lstep2[ 1] = _mm_madd_epi16(lstep2[ 1], kOne);
          lstep2[ 2] = _mm_madd_epi16(lstep2[ 2], kOne);
          lstep2[ 3] = _mm_madd_epi16(lstep2[ 3], kOne);
          lstep2[ 4] = _mm_madd_epi16(lstep2[ 4], kOne);
          lstep2[ 5] = _mm_madd_epi16(lstep2[ 5], kOne);
          lstep2[ 6] = _mm_madd_epi16(lstep2[ 6], kOne);
          lstep2[ 7] = _mm_madd_epi16(lstep2[ 7], kOne);
          lstep2[ 8] = _mm_madd_epi16(lstep2[ 8], kOne);
          lstep2[ 9] = _mm_madd_epi16(lstep2[ 9], kOne);
          lstep2[10] = _mm_madd_epi16(lstep2[10], kOne);
          lstep2[11] = _mm_madd_epi16(lstep2[11], kOne);
          lstep2[12] = _mm_madd_epi16(lstep2[12], kOne);
          lstep2[13] = _mm_madd_epi16(lstep2[13], kOne);
          lstep2[14] = _mm_madd_epi16(lstep2[14], kOne);
          lstep2[15] = _mm_madd_epi16(lstep2[15], kOne);
1492 1493 1494 1495 1496 1497 1498 1499 1500 1501 1502 1503 1504 1505 1506 1507 1508 1509 1510 1511 1512 1513 1514 1515 1516 1517 1518 1519 1520 1521 1522 1523 1524 1525 1526 1527 1528 1529 1530 1531 1532 1533 1534 1535 1536 1537 1538 1539 1540 1541

          lstep3[ 0] = _mm_add_epi32(lstep2[14], lstep2[ 0]);
          lstep3[ 1] = _mm_add_epi32(lstep2[15], lstep2[ 1]);
          lstep3[ 2] = _mm_add_epi32(lstep2[12], lstep2[ 2]);
          lstep3[ 3] = _mm_add_epi32(lstep2[13], lstep2[ 3]);
          lstep3[ 4] = _mm_add_epi32(lstep2[10], lstep2[ 4]);
          lstep3[ 5] = _mm_add_epi32(lstep2[11], lstep2[ 5]);
          lstep3[ 6] = _mm_add_epi32(lstep2[ 8], lstep2[ 6]);
          lstep3[ 7] = _mm_add_epi32(lstep2[ 9], lstep2[ 7]);
          lstep3[ 8] = _mm_sub_epi32(lstep2[ 6], lstep2[ 8]);
          lstep3[ 9] = _mm_sub_epi32(lstep2[ 7], lstep2[ 9]);
          lstep3[10] = _mm_sub_epi32(lstep2[ 4], lstep2[10]);
          lstep3[11] = _mm_sub_epi32(lstep2[ 5], lstep2[11]);
          lstep3[12] = _mm_sub_epi32(lstep2[ 2], lstep2[12]);
          lstep3[13] = _mm_sub_epi32(lstep2[ 3], lstep2[13]);
          lstep3[14] = _mm_sub_epi32(lstep2[ 0], lstep2[14]);
          lstep3[15] = _mm_sub_epi32(lstep2[ 1], lstep2[15]);
        }
        {
          const __m128i s3_10_0 = _mm_unpacklo_epi16(step2[13], step2[10]);
          const __m128i s3_10_1 = _mm_unpackhi_epi16(step2[13], step2[10]);
          const __m128i s3_11_0 = _mm_unpacklo_epi16(step2[12], step2[11]);
          const __m128i s3_11_1 = _mm_unpackhi_epi16(step2[12], step2[11]);
          const __m128i s3_10_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_m16);
          const __m128i s3_10_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_m16);
          const __m128i s3_11_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_m16);
          const __m128i s3_11_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_m16);
          const __m128i s3_12_2 = _mm_madd_epi16(s3_11_0, k__cospi_p16_p16);
          const __m128i s3_12_3 = _mm_madd_epi16(s3_11_1, k__cospi_p16_p16);
          const __m128i s3_13_2 = _mm_madd_epi16(s3_10_0, k__cospi_p16_p16);
          const __m128i s3_13_3 = _mm_madd_epi16(s3_10_1, k__cospi_p16_p16);
          // dct_const_round_shift
          const __m128i s3_10_4 = _mm_add_epi32(s3_10_2, k__DCT_CONST_ROUNDING);
          const __m128i s3_10_5 = _mm_add_epi32(s3_10_3, k__DCT_CONST_ROUNDING);
          const __m128i s3_11_4 = _mm_add_epi32(s3_11_2, k__DCT_CONST_ROUNDING);
          const __m128i s3_11_5 = _mm_add_epi32(s3_11_3, k__DCT_CONST_ROUNDING);
          const __m128i s3_12_4 = _mm_add_epi32(s3_12_2, k__DCT_CONST_ROUNDING);
          const __m128i s3_12_5 = _mm_add_epi32(s3_12_3, k__DCT_CONST_ROUNDING);
          const __m128i s3_13_4 = _mm_add_epi32(s3_13_2, k__DCT_CONST_ROUNDING);
          const __m128i s3_13_5 = _mm_add_epi32(s3_13_3, k__DCT_CONST_ROUNDING);
          lstep3[20] = _mm_srai_epi32(s3_10_4, DCT_CONST_BITS);
          lstep3[21] = _mm_srai_epi32(s3_10_5, DCT_CONST_BITS);
          lstep3[22] = _mm_srai_epi32(s3_11_4, DCT_CONST_BITS);
          lstep3[23] = _mm_srai_epi32(s3_11_5, DCT_CONST_BITS);
          lstep3[24] = _mm_srai_epi32(s3_12_4, DCT_CONST_BITS);
          lstep3[25] = _mm_srai_epi32(s3_12_5, DCT_CONST_BITS);
          lstep3[26] = _mm_srai_epi32(s3_13_4, DCT_CONST_BITS);
          lstep3[27] = _mm_srai_epi32(s3_13_5, DCT_CONST_BITS);
        }
        {
1542 1543 1544 1545 1546 1547 1548 1549 1550 1551 1552 1553 1554 1555 1556 1557 1558 1559 1560 1561 1562 1563 1564 1565 1566 1567 1568 1569 1570 1571 1572 1573
          lstep2[40] = _mm_unpacklo_epi16(step2[20], kZero);
          lstep2[41] = _mm_unpackhi_epi16(step2[20], kZero);
          lstep2[42] = _mm_unpacklo_epi16(step2[21], kZero);
          lstep2[43] = _mm_unpackhi_epi16(step2[21], kZero);
          lstep2[44] = _mm_unpacklo_epi16(step2[22], kZero);
          lstep2[45] = _mm_unpackhi_epi16(step2[22], kZero);
          lstep2[46] = _mm_unpacklo_epi16(step2[23], kZero);
          lstep2[47] = _mm_unpackhi_epi16(step2[23], kZero);
          lstep2[48] = _mm_unpacklo_epi16(step2[24], kZero);
          lstep2[49] = _mm_unpackhi_epi16(step2[24], kZero);
          lstep2[50] = _mm_unpacklo_epi16(step2[25], kZero);
          lstep2[51] = _mm_unpackhi_epi16(step2[25], kZero);
          lstep2[52] = _mm_unpacklo_epi16(step2[26], kZero);
          lstep2[53] = _mm_unpackhi_epi16(step2[26], kZero);
          lstep2[54] = _mm_unpacklo_epi16(step2[27], kZero);
          lstep2[55] = _mm_unpackhi_epi16(step2[27], kZero);
          lstep2[40] = _mm_madd_epi16(lstep2[40], kOne);
          lstep2[41] = _mm_madd_epi16(lstep2[41], kOne);
          lstep2[42] = _mm_madd_epi16(lstep2[42], kOne);
          lstep2[43] = _mm_madd_epi16(lstep2[43], kOne);
          lstep2[44] = _mm_madd_epi16(lstep2[44], kOne);
          lstep2[45] = _mm_madd_epi16(lstep2[45], kOne);
          lstep2[46] = _mm_madd_epi16(lstep2[46], kOne);
          lstep2[47] = _mm_madd_epi16(lstep2[47], kOne);
          lstep2[48] = _mm_madd_epi16(lstep2[48], kOne);
          lstep2[49] = _mm_madd_epi16(lstep2[49], kOne);
          lstep2[50] = _mm_madd_epi16(lstep2[50], kOne);
          lstep2[51] = _mm_madd_epi16(lstep2[51], kOne);
          lstep2[52] = _mm_madd_epi16(lstep2[52], kOne);
          lstep2[53] = _mm_madd_epi16(lstep2[53], kOne);
          lstep2[54] = _mm_madd_epi16(lstep2[54], kOne);
          lstep2[55] = _mm_madd_epi16(lstep2[55], kOne);
1574

1575 1576 1577 1578 1579 1580 1581 1582 1583 1584 1585 1586 1587 1588 1589 1590 1591 1592 1593 1594 1595 1596 1597 1598 1599 1600 1601 1602 1603 1604 1605 1606
          lstep1[32] = _mm_unpacklo_epi16(step1[16], kZero);
          lstep1[33] = _mm_unpackhi_epi16(step1[16], kZero);
          lstep1[34] = _mm_unpacklo_epi16(step1[17], kZero);
          lstep1[35] = _mm_unpackhi_epi16(step1[17], kZero);
          lstep1[36] = _mm_unpacklo_epi16(step1[18], kZero);
          lstep1[37] = _mm_unpackhi_epi16(step1[18], kZero);
          lstep1[38] = _mm_unpacklo_epi16(step1[19], kZero);
          lstep1[39] = _mm_unpackhi_epi16(step1[19], kZero);
          lstep1[56] = _mm_unpacklo_epi16(step1[28], kZero);
          lstep1[57] = _mm_unpackhi_epi16(step1[28], kZero);
          lstep1[58] = _mm_unpacklo_epi16(step1[29], kZero);
          lstep1[59] = _mm_unpackhi_epi16(step1[29], kZero);
          lstep1[60] = _mm_unpacklo_epi16(step1[30], kZero);
          lstep1[61] = _mm_unpackhi_epi16(step1[30], kZero);
          lstep1[62] = _mm_unpacklo_epi16(step1[31], kZero);
          lstep1[63] = _mm_unpackhi_epi16(step1[31], kZero);
          lstep1[32] = _mm_madd_epi16(lstep1[32], kOne);
          lstep1[33] = _mm_madd_epi16(lstep1[33], kOne);
          lstep1[34] = _mm_madd_epi16(lstep1[34], kOne);
          lstep1[35] = _mm_madd_epi16(lstep1[35], kOne);
          lstep1[36] = _mm_madd_epi16(lstep1[36], kOne);
          lstep1[37] = _mm_madd_epi16(lstep1[37], kOne);
          lstep1[38] = _mm_madd_epi16(lstep1[38], kOne);
          lstep1[39] = _mm_madd_epi16(lstep1[39], kOne);
          lstep1[56] = _mm_madd_epi16(lstep1[56], kOne);
          lstep1[57] = _mm_madd_epi16(lstep1[57], kOne);
          lstep1[58] = _mm_madd_epi16(lstep1[58], kOne);
          lstep1[59] = _mm_madd_epi16(lstep1[59], kOne);
          lstep1[60] = _mm_madd_epi16(lstep1[60], kOne);
          lstep1[61] = _mm_madd_epi16(lstep1[61], kOne);
          lstep1[62] = _mm_madd_epi16(lstep1[62], kOne);
          lstep1[63] = _mm_madd_epi16(lstep1[63], kOne);
1607 1608 1609

          lstep3[32] = _mm_add_epi32(lstep2[46], lstep1[32]);
          lstep3[33] = _mm_add_epi32(lstep2[47], lstep1[33]);
1610

1611 1612 1613 1614 1615 1616 1617 1618 1619 1620 1621 1622 1623 1624 1625 1626 1627 1628 1629 1630 1631 1632 1633 1634 1635 1636 1637 1638 1639 1640 1641 1642
          lstep3[34] = _mm_add_epi32(lstep2[44], lstep1[34]);
          lstep3[35] = _mm_add_epi32(lstep2[45], lstep1[35]);
          lstep3[36] = _mm_add_epi32(lstep2[42], lstep1[36]);
          lstep3[37] = _mm_add_epi32(lstep2[43], lstep1[37]);
          lstep3[38] = _mm_add_epi32(lstep2[40], lstep1[38]);
          lstep3[39] = _mm_add_epi32(lstep2[41], lstep1[39]);
          lstep3[40] = _mm_sub_epi32(lstep1[38], lstep2[40]);
          lstep3[41] = _mm_sub_epi32(lstep1[39], lstep2[41]);
          lstep3[42] = _mm_sub_epi32(lstep1[36], lstep2[42]);
          lstep3[43] = _mm_sub_epi32(lstep1[37], lstep2[43]);
          lstep3[44] = _mm_sub_epi32(lstep1[34], lstep2[44]);
          lstep3[45] = _mm_sub_epi32(lstep1[35], lstep2[45]);
          lstep3[46] = _mm_sub_epi32(lstep1[32], lstep2[46]);
          lstep3[47] = _mm_sub_epi32(lstep1[33], lstep2[47]);
          lstep3[48] = _mm_sub_epi32(lstep1[62], lstep2[48]);
          lstep3[49] = _mm_sub_epi32(lstep1[63], lstep2[49]);
          lstep3[50] = _mm_sub_epi32(lstep1[60], lstep2[50]);
          lstep3[51] = _mm_sub_epi32(lstep1[61], lstep2[51]);
          lstep3[52] = _mm_sub_epi32(lstep1[58], lstep2[52]);
          lstep3[53] = _mm_sub_epi32(lstep1[59], lstep2[53]);
          lstep3[54] = _mm_sub_epi32(lstep1[56], lstep2[54]);
          lstep3[55] = _mm_sub_epi32(lstep1[57], lstep2[55]);
          lstep3[56] = _mm_add_epi32(lstep2[54], lstep1[56]);
          lstep3[57] = _mm_add_epi32(lstep2[55], lstep1[57]);
          lstep3[58] = _mm_add_epi32(lstep2[52], lstep1[58]);
          lstep3[59] = _mm_add_epi32(lstep2[53], lstep1[59]);
          lstep3[60] = _mm_add_epi32(lstep2[50], lstep1[60]);
          lstep3[61] = _mm_add_epi32(lstep2[51], lstep1[61]);
          lstep3[62] = _mm_add_epi32(lstep2[48], lstep1[62]);
          lstep3[63] = _mm_add_epi32(lstep2[49], lstep1[63]);
        }

1643 1644 1645
        // stage 4
        {
          // expanding to 32-bit length priori to addition operations
1646 1647 1648 1649 1650 1651 1652 1653 1654 1655 1656 1657 1658 1659 1660 1661
          lstep2[16] = _mm_unpacklo_epi16(step2[ 8], kZero);
          lstep2[17] = _mm_unpackhi_epi16(step2[ 8], kZero);
          lstep2[18] = _mm_unpacklo_epi16(step2[ 9], kZero);
          lstep2[19] = _mm_unpackhi_epi16(step2[ 9], kZero);
          lstep2[28] = _mm_unpacklo_epi16(step2[14], kZero);
          lstep2[29] = _mm_unpackhi_epi16(step2[14], kZero);
          lstep2[30] = _mm_unpacklo_epi16(step2[15], kZero);
          lstep2[31] = _mm_unpackhi_epi16(step2[15], kZero);
          lstep2[16] = _mm_madd_epi16(lstep2[16], kOne);
          lstep2[17] = _mm_madd_epi16(lstep2[17], kOne);
          lstep2[18] = _mm_madd_epi16(lstep2[18], kOne);
          lstep2[19] = _mm_madd_epi16(lstep2[19], kOne);
          lstep2[28] = _mm_madd_epi16(lstep2[28], kOne);
          lstep2[29] = _mm_madd_epi16(lstep2[29], kOne);
          lstep2[30] = _mm_madd_epi16(lstep2[30], kOne);
          lstep2[31] = _mm_madd_epi16(lstep2[31], kOne);
1662 1663 1664 1665 1666 1667 1668 1669 1670 1671 1672 1673 1674 1675 1676 1677 1678 1679 1680 1681 1682 1683 1684 1685 1686 1687 1688

          lstep1[ 0] = _mm_add_epi32(lstep3[ 6], lstep3[ 0]);
          lstep1[ 1] = _mm_add_epi32(lstep3[ 7], lstep3[ 1]);
          lstep1[ 2] = _mm_add_epi32(lstep3[ 4], lstep3[ 2]);
          lstep1[ 3] = _mm_add_epi32(lstep3[ 5], lstep3[ 3]);
          lstep1[ 4] = _mm_sub_epi32(lstep3[ 2], lstep3[ 4]);
          lstep1[ 5] = _mm_sub_epi32(lstep3[ 3], lstep3[ 5]);
          lstep1[ 6] = _mm_sub_epi32(lstep3[ 0], lstep3[ 6]);
          lstep1[ 7] = _mm_sub_epi32(lstep3[ 1], lstep3[ 7]);
          lstep1[16] = _mm_add_epi32(lstep3[22], lstep2[16]);
          lstep1[17] = _mm_add_epi32(lstep3[23], lstep2[17]);
          lstep1[18] = _mm_add_epi32(lstep3[20], lstep2[18]);
          lstep1[19] = _mm_add_epi32(lstep3[21], lstep2[19]);
          lstep1[20] = _mm_sub_epi32(lstep2[18], lstep3[20]);
          lstep1[21] = _mm_sub_epi32(lstep2[19], lstep3[21]);
          lstep1[22] = _mm_sub_epi32(lstep2[16], lstep3[22]);
          lstep1[23] = _mm_sub_epi32(lstep2[17], lstep3[23]);
          lstep1[24] = _mm_sub_epi32(lstep2[30], lstep3[24]);
          lstep1[25] = _mm_sub_epi32(lstep2[31], lstep3[25]);
          lstep1[26] = _mm_sub_epi32(lstep2[28], lstep3[26]);
          lstep1[27] = _mm_sub_epi32(lstep2[29], lstep3[27]);
          lstep1[28] = _mm_add_epi32(lstep3[26], lstep2[28]);
          lstep1[29] = _mm_add_epi32(lstep3[27], lstep2[29]);
          lstep1[30] = _mm_add_epi32(lstep3[24], lstep2[30]);
          lstep1[31] = _mm_add_epi32(lstep3[25], lstep2[31]);
        }
        {
1689 1690 1691 1692
        // to be continued...
        //
        const __m128i k32_p16_p16 = pair_set_epi32(cospi_16_64, cospi_16_64);
        const __m128i k32_p16_m16 = pair_set_epi32(cospi_16_64, -cospi_16_64);
1693

1694 1695 1696 1697
        u[0] = _mm_unpacklo_epi32(lstep3[12], lstep3[10]);
        u[1] = _mm_unpackhi_epi32(lstep3[12], lstep3[10]);
        u[2] = _mm_unpacklo_epi32(lstep3[13], lstep3[11]);
        u[3] = _mm_unpackhi_epi32(lstep3[13], lstep3[11]);
1698

1699 1700
        // TODO(jingning): manually inline k_madd_epi32_ to further hide
        // instruction latency.
1701 1702 1703 1704 1705 1706 1707 1708 1709
        v[0] = k_madd_epi32(u[0], k32_p16_m16);