vp9_dct_intrin_sse2.c 78.4 KB
Newer Older
Christian Duvivier's avatar
Christian Duvivier committed
1 2 3 4 5 6 7 8 9 10
/*
 *  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.
 */

11
#include <assert.h>
Christian Duvivier's avatar
Christian Duvivier committed
12
#include <emmintrin.h>  // SSE2
13 14

#include "./vp9_rtcd.h"
15
#include "./vpx_dsp_rtcd.h"
Jingning Han's avatar
Jingning Han committed
16
#include "vpx_dsp/txfm_common.h"
17
#include "vpx_dsp/x86/fwd_txfm_sse2.h"
18
#include "vpx_dsp/x86/transpose_sse2.h"
Jingning Han's avatar
Jingning Han committed
19
#include "vpx_dsp/x86/txfm_common_sse2.h"
20
#include "vpx_ports/mem.h"
Christian Duvivier's avatar
Christian Duvivier committed
21

22 23
static INLINE void load_buffer_4x4(const int16_t *input, __m128i *in,
                                   int stride) {
24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42
  const __m128i k__nonzero_bias_a = _mm_setr_epi16(0, 1, 1, 1, 1, 1, 1, 1);
  const __m128i k__nonzero_bias_b = _mm_setr_epi16(1, 0, 0, 0, 0, 0, 0, 0);
  __m128i mask;

  in[0] = _mm_loadl_epi64((const __m128i *)(input + 0 * stride));
  in[1] = _mm_loadl_epi64((const __m128i *)(input + 1 * stride));
  in[2] = _mm_loadl_epi64((const __m128i *)(input + 2 * stride));
  in[3] = _mm_loadl_epi64((const __m128i *)(input + 3 * stride));

  in[0] = _mm_slli_epi16(in[0], 4);
  in[1] = _mm_slli_epi16(in[1], 4);
  in[2] = _mm_slli_epi16(in[2], 4);
  in[3] = _mm_slli_epi16(in[3], 4);

  mask = _mm_cmpeq_epi16(in[0], k__nonzero_bias_a);
  in[0] = _mm_add_epi16(in[0], mask);
  in[0] = _mm_add_epi16(in[0], k__nonzero_bias_b);
}

43
static INLINE void write_buffer_4x4(tran_low_t *output, __m128i *res) {
44 45 46 47 48 49 50
  const __m128i kOne = _mm_set1_epi16(1);
  __m128i in01 = _mm_unpacklo_epi64(res[0], res[1]);
  __m128i in23 = _mm_unpacklo_epi64(res[2], res[3]);
  __m128i out01 = _mm_add_epi16(in01, kOne);
  __m128i out23 = _mm_add_epi16(in23, kOne);
  out01 = _mm_srai_epi16(out01, 2);
  out23 = _mm_srai_epi16(out23, 2);
Deb Mukherjee's avatar
Deb Mukherjee committed
51 52
  store_output(&out01, (output + 0 * 8));
  store_output(&out23, (output + 1 * 8));
53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73
}

static INLINE void transpose_4x4(__m128i *res) {
  // Combine and transpose
  // 00 01 02 03 20 21 22 23
  // 10 11 12 13 30 31 32 33
  const __m128i tr0_0 = _mm_unpacklo_epi16(res[0], res[1]);
  const __m128i tr0_1 = _mm_unpackhi_epi16(res[0], res[1]);

  // 00 10 01 11 02 12 03 13
  // 20 30 21 31 22 32 23 33
  res[0] = _mm_unpacklo_epi32(tr0_0, tr0_1);
  res[2] = _mm_unpackhi_epi32(tr0_0, tr0_1);

  // 00 10 20 30 01 11 21 31
  // 02 12 22 32 03 13 23 33
  // only use the first 4 16-bit integers
  res[1] = _mm_unpackhi_epi64(res[0], res[0]);
  res[3] = _mm_unpackhi_epi64(res[2], res[2]);
}

74
static void fdct4_sse2(__m128i *in) {
75
  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
76
  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
77 78
  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
79 80 81
  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);

  __m128i u[4], v[4];
clang-format's avatar
clang-format committed
82 83
  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
  u[1] = _mm_unpacklo_epi16(in[3], in[2]);
84 85 86

  v[0] = _mm_add_epi16(u[0], u[1]);
  v[1] = _mm_sub_epi16(u[0], u[1]);
87 88 89

  u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16);  // 0
  u[1] = _mm_madd_epi16(v[0], k__cospi_p16_m16);  // 2
90 91
  u[2] = _mm_madd_epi16(v[1], k__cospi_p08_p24);  // 1
  u[3] = _mm_madd_epi16(v[1], k__cospi_p24_m08);  // 3
92 93 94 95 96 97 98 99 100 101 102 103 104 105 106

  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);

  in[0] = _mm_packs_epi32(u[0], u[1]);
  in[1] = _mm_packs_epi32(u[2], u[3]);
  transpose_4x4(in);
}

107
static void fadst4_sse2(__m128i *in) {
108 109 110 111
  const __m128i k__sinpi_p01_p02 = pair_set_epi16(sinpi_1_9, sinpi_2_9);
  const __m128i k__sinpi_p04_m01 = pair_set_epi16(sinpi_4_9, -sinpi_1_9);
  const __m128i k__sinpi_p03_p04 = pair_set_epi16(sinpi_3_9, sinpi_4_9);
  const __m128i k__sinpi_m03_p02 = pair_set_epi16(-sinpi_3_9, sinpi_2_9);
112
  const __m128i k__sinpi_p03_p03 = _mm_set1_epi16((int16_t)sinpi_3_9);
113 114 115 116 117 118 119 120 121
  const __m128i kZero = _mm_set1_epi16(0);
  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
  __m128i u[8], v[8];
  __m128i in7 = _mm_add_epi16(in[0], in[1]);

  u[0] = _mm_unpacklo_epi16(in[0], in[1]);
  u[1] = _mm_unpacklo_epi16(in[2], in[3]);
  u[2] = _mm_unpacklo_epi16(in7, kZero);
  u[3] = _mm_unpacklo_epi16(in[2], kZero);
122
  u[4] = _mm_unpacklo_epi16(in[3], kZero);
123 124 125 126 127 128 129

  v[0] = _mm_madd_epi16(u[0], k__sinpi_p01_p02);  // s0 + s2
  v[1] = _mm_madd_epi16(u[1], k__sinpi_p03_p04);  // s4 + s5
  v[2] = _mm_madd_epi16(u[2], k__sinpi_p03_p03);  // x1
  v[3] = _mm_madd_epi16(u[0], k__sinpi_p04_m01);  // s1 - s3
  v[4] = _mm_madd_epi16(u[1], k__sinpi_m03_p02);  // -s4 + s6
  v[5] = _mm_madd_epi16(u[3], k__sinpi_p03_p03);  // s4
130
  v[6] = _mm_madd_epi16(u[4], k__sinpi_p03_p03);
131 132

  u[0] = _mm_add_epi32(v[0], v[1]);
133
  u[1] = _mm_sub_epi32(v[2], v[6]);
134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154
  u[2] = _mm_add_epi32(v[3], v[4]);
  u[3] = _mm_sub_epi32(u[2], u[0]);
  u[4] = _mm_slli_epi32(v[5], 2);
  u[5] = _mm_sub_epi32(u[4], v[5]);
  u[6] = _mm_add_epi32(u[3], u[5]);

  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
  v[3] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);

  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);

  in[0] = _mm_packs_epi32(u[0], u[2]);
  in[1] = _mm_packs_epi32(u[1], u[3]);
  transpose_4x4(in);
}

clang-format's avatar
clang-format committed
155 156
void vp9_fht4x4_sse2(const int16_t *input, tran_low_t *output, int stride,
                     int tx_type) {
157
  __m128i in[4];
158

159
  switch (tx_type) {
clang-format's avatar
clang-format committed
160
    case DCT_DCT: vpx_fdct4x4_sse2(input, output, stride); break;
161 162
    case ADST_DCT:
      load_buffer_4x4(input, in, stride);
163 164
      fadst4_sse2(in);
      fdct4_sse2(in);
165
      write_buffer_4x4(output, in);
166
      break;
167 168
    case DCT_ADST:
      load_buffer_4x4(input, in, stride);
169 170
      fdct4_sse2(in);
      fadst4_sse2(in);
171
      write_buffer_4x4(output, in);
172
      break;
173 174
    case ADST_ADST:
      load_buffer_4x4(input, in, stride);
175 176
      fadst4_sse2(in);
      fadst4_sse2(in);
177
      write_buffer_4x4(output, in);
178
      break;
clang-format's avatar
clang-format committed
179
    default: assert(0); break;
180 181 182
  }
}

183
void vp9_fdct8x8_quant_sse2(const int16_t *input, int stride,
clang-format's avatar
clang-format committed
184
                            int16_t *coeff_ptr, intptr_t n_coeffs,
Johann's avatar
Johann committed
185 186
                            int skip_block, const int16_t *round_ptr,
                            const int16_t *quant_ptr, int16_t *qcoeff_ptr,
clang-format's avatar
clang-format committed
187 188 189
                            int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
                            uint16_t *eob_ptr, const int16_t *scan_ptr,
                            const int16_t *iscan_ptr) {
190 191
  __m128i zero;
  int pass;
Johann's avatar
Johann committed
192

193 194 195 196
  // 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.
197
  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
198 199 200 201 202 203 204 205 206
  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_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__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
  // Load input
clang-format's avatar
clang-format committed
207 208 209 210 211 212 213 214
  __m128i in0 = _mm_load_si128((const __m128i *)(input + 0 * stride));
  __m128i in1 = _mm_load_si128((const __m128i *)(input + 1 * stride));
  __m128i in2 = _mm_load_si128((const __m128i *)(input + 2 * stride));
  __m128i in3 = _mm_load_si128((const __m128i *)(input + 3 * stride));
  __m128i in4 = _mm_load_si128((const __m128i *)(input + 4 * stride));
  __m128i in5 = _mm_load_si128((const __m128i *)(input + 5 * stride));
  __m128i in6 = _mm_load_si128((const __m128i *)(input + 6 * stride));
  __m128i in7 = _mm_load_si128((const __m128i *)(input + 7 * stride));
215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465
  __m128i *in[8];
  int index = 0;

  (void)scan_ptr;
  (void)coeff_ptr;

  // Pre-condition input (shift by two)
  in0 = _mm_slli_epi16(in0, 2);
  in1 = _mm_slli_epi16(in1, 2);
  in2 = _mm_slli_epi16(in2, 2);
  in3 = _mm_slli_epi16(in3, 2);
  in4 = _mm_slli_epi16(in4, 2);
  in5 = _mm_slli_epi16(in5, 2);
  in6 = _mm_slli_epi16(in6, 2);
  in7 = _mm_slli_epi16(in7, 2);

  in[0] = &in0;
  in[1] = &in1;
  in[2] = &in2;
  in[3] = &in3;
  in[4] = &in4;
  in[5] = &in5;
  in[6] = &in6;
  in[7] = &in7;

  // We do two passes, first the columns, then the rows. The results of the
  // first pass are transposed so that the same column code can be reused. The
  // results of the second pass are also transposed so that the rows (processed
  // as columns) are put back in row positions.
  for (pass = 0; pass < 2; pass++) {
    // To store results of each pass before the transpose.
    __m128i res0, res1, res2, res3, res4, res5, res6, res7;
    // Add/subtract
    const __m128i q0 = _mm_add_epi16(in0, in7);
    const __m128i q1 = _mm_add_epi16(in1, in6);
    const __m128i q2 = _mm_add_epi16(in2, in5);
    const __m128i q3 = _mm_add_epi16(in3, in4);
    const __m128i q4 = _mm_sub_epi16(in3, in4);
    const __m128i q5 = _mm_sub_epi16(in2, in5);
    const __m128i q6 = _mm_sub_epi16(in1, in6);
    const __m128i q7 = _mm_sub_epi16(in0, in7);
    // Work on first four results
    {
      // Add/subtract
      const __m128i r0 = _mm_add_epi16(q0, q3);
      const __m128i r1 = _mm_add_epi16(q1, q2);
      const __m128i r2 = _mm_sub_epi16(q1, q2);
      const __m128i r3 = _mm_sub_epi16(q0, q3);
      // Interleave to do the multiply by constants which gets us into 32bits
      const __m128i t0 = _mm_unpacklo_epi16(r0, r1);
      const __m128i t1 = _mm_unpackhi_epi16(r0, r1);
      const __m128i t2 = _mm_unpacklo_epi16(r2, r3);
      const __m128i t3 = _mm_unpackhi_epi16(r2, r3);
      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p16_p16);
      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p16_p16);
      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_p16_m16);
      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_p16_m16);
      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p24_p08);
      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p24_p08);
      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m08_p24);
      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m08_p24);
      // dct_const_round_shift
      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
      // Combine
      res0 = _mm_packs_epi32(w0, w1);
      res4 = _mm_packs_epi32(w2, w3);
      res2 = _mm_packs_epi32(w4, w5);
      res6 = _mm_packs_epi32(w6, w7);
    }
    // Work on next four results
    {
      // Interleave to do the multiply by constants which gets us into 32bits
      const __m128i d0 = _mm_unpacklo_epi16(q6, q5);
      const __m128i d1 = _mm_unpackhi_epi16(q6, q5);
      const __m128i e0 = _mm_madd_epi16(d0, k__cospi_p16_m16);
      const __m128i e1 = _mm_madd_epi16(d1, k__cospi_p16_m16);
      const __m128i e2 = _mm_madd_epi16(d0, k__cospi_p16_p16);
      const __m128i e3 = _mm_madd_epi16(d1, k__cospi_p16_p16);
      // dct_const_round_shift
      const __m128i f0 = _mm_add_epi32(e0, k__DCT_CONST_ROUNDING);
      const __m128i f1 = _mm_add_epi32(e1, k__DCT_CONST_ROUNDING);
      const __m128i f2 = _mm_add_epi32(e2, k__DCT_CONST_ROUNDING);
      const __m128i f3 = _mm_add_epi32(e3, k__DCT_CONST_ROUNDING);
      const __m128i s0 = _mm_srai_epi32(f0, DCT_CONST_BITS);
      const __m128i s1 = _mm_srai_epi32(f1, DCT_CONST_BITS);
      const __m128i s2 = _mm_srai_epi32(f2, DCT_CONST_BITS);
      const __m128i s3 = _mm_srai_epi32(f3, DCT_CONST_BITS);
      // Combine
      const __m128i r0 = _mm_packs_epi32(s0, s1);
      const __m128i r1 = _mm_packs_epi32(s2, s3);
      // Add/subtract
      const __m128i x0 = _mm_add_epi16(q4, r0);
      const __m128i x1 = _mm_sub_epi16(q4, r0);
      const __m128i x2 = _mm_sub_epi16(q7, r1);
      const __m128i x3 = _mm_add_epi16(q7, r1);
      // Interleave to do the multiply by constants which gets us into 32bits
      const __m128i t0 = _mm_unpacklo_epi16(x0, x3);
      const __m128i t1 = _mm_unpackhi_epi16(x0, x3);
      const __m128i t2 = _mm_unpacklo_epi16(x1, x2);
      const __m128i t3 = _mm_unpackhi_epi16(x1, x2);
      const __m128i u0 = _mm_madd_epi16(t0, k__cospi_p28_p04);
      const __m128i u1 = _mm_madd_epi16(t1, k__cospi_p28_p04);
      const __m128i u2 = _mm_madd_epi16(t0, k__cospi_m04_p28);
      const __m128i u3 = _mm_madd_epi16(t1, k__cospi_m04_p28);
      const __m128i u4 = _mm_madd_epi16(t2, k__cospi_p12_p20);
      const __m128i u5 = _mm_madd_epi16(t3, k__cospi_p12_p20);
      const __m128i u6 = _mm_madd_epi16(t2, k__cospi_m20_p12);
      const __m128i u7 = _mm_madd_epi16(t3, k__cospi_m20_p12);
      // dct_const_round_shift
      const __m128i v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
      const __m128i v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
      const __m128i v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
      const __m128i v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
      const __m128i v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
      const __m128i v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
      const __m128i v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
      const __m128i v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);
      const __m128i w0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
      const __m128i w1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
      const __m128i w2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
      const __m128i w3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
      const __m128i w4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
      const __m128i w5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
      const __m128i w6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
      const __m128i w7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
      // Combine
      res1 = _mm_packs_epi32(w0, w1);
      res7 = _mm_packs_epi32(w2, w3);
      res5 = _mm_packs_epi32(w4, w5);
      res3 = _mm_packs_epi32(w6, w7);
    }
    // Transpose the 8x8.
    {
      // 00 01 02 03 04 05 06 07
      // 10 11 12 13 14 15 16 17
      // 20 21 22 23 24 25 26 27
      // 30 31 32 33 34 35 36 37
      // 40 41 42 43 44 45 46 47
      // 50 51 52 53 54 55 56 57
      // 60 61 62 63 64 65 66 67
      // 70 71 72 73 74 75 76 77
      const __m128i tr0_0 = _mm_unpacklo_epi16(res0, res1);
      const __m128i tr0_1 = _mm_unpacklo_epi16(res2, res3);
      const __m128i tr0_2 = _mm_unpackhi_epi16(res0, res1);
      const __m128i tr0_3 = _mm_unpackhi_epi16(res2, res3);
      const __m128i tr0_4 = _mm_unpacklo_epi16(res4, res5);
      const __m128i tr0_5 = _mm_unpacklo_epi16(res6, res7);
      const __m128i tr0_6 = _mm_unpackhi_epi16(res4, res5);
      const __m128i tr0_7 = _mm_unpackhi_epi16(res6, res7);
      // 00 10 01 11 02 12 03 13
      // 20 30 21 31 22 32 23 33
      // 04 14 05 15 06 16 07 17
      // 24 34 25 35 26 36 27 37
      // 40 50 41 51 42 52 43 53
      // 60 70 61 71 62 72 63 73
      // 54 54 55 55 56 56 57 57
      // 64 74 65 75 66 76 67 77
      const __m128i tr1_0 = _mm_unpacklo_epi32(tr0_0, tr0_1);
      const __m128i tr1_1 = _mm_unpacklo_epi32(tr0_2, tr0_3);
      const __m128i tr1_2 = _mm_unpackhi_epi32(tr0_0, tr0_1);
      const __m128i tr1_3 = _mm_unpackhi_epi32(tr0_2, tr0_3);
      const __m128i tr1_4 = _mm_unpacklo_epi32(tr0_4, tr0_5);
      const __m128i tr1_5 = _mm_unpacklo_epi32(tr0_6, tr0_7);
      const __m128i tr1_6 = _mm_unpackhi_epi32(tr0_4, tr0_5);
      const __m128i tr1_7 = _mm_unpackhi_epi32(tr0_6, tr0_7);
      // 00 10 20 30 01 11 21 31
      // 40 50 60 70 41 51 61 71
      // 02 12 22 32 03 13 23 33
      // 42 52 62 72 43 53 63 73
      // 04 14 24 34 05 15 21 36
      // 44 54 64 74 45 55 61 76
      // 06 16 26 36 07 17 27 37
      // 46 56 66 76 47 57 67 77
      in0 = _mm_unpacklo_epi64(tr1_0, tr1_4);
      in1 = _mm_unpackhi_epi64(tr1_0, tr1_4);
      in2 = _mm_unpacklo_epi64(tr1_2, tr1_6);
      in3 = _mm_unpackhi_epi64(tr1_2, tr1_6);
      in4 = _mm_unpacklo_epi64(tr1_1, tr1_5);
      in5 = _mm_unpackhi_epi64(tr1_1, tr1_5);
      in6 = _mm_unpacklo_epi64(tr1_3, tr1_7);
      in7 = _mm_unpackhi_epi64(tr1_3, tr1_7);
      // 00 10 20 30 40 50 60 70
      // 01 11 21 31 41 51 61 71
      // 02 12 22 32 42 52 62 72
      // 03 13 23 33 43 53 63 73
      // 04 14 24 34 44 54 64 74
      // 05 15 25 35 45 55 65 75
      // 06 16 26 36 46 56 66 76
      // 07 17 27 37 47 57 67 77
    }
  }
  // Post-condition output and store it
  {
    // Post-condition (division by two)
    //    division of two 16 bits signed numbers using shifts
    //    n / 2 = (n - (n >> 15)) >> 1
    const __m128i sign_in0 = _mm_srai_epi16(in0, 15);
    const __m128i sign_in1 = _mm_srai_epi16(in1, 15);
    const __m128i sign_in2 = _mm_srai_epi16(in2, 15);
    const __m128i sign_in3 = _mm_srai_epi16(in3, 15);
    const __m128i sign_in4 = _mm_srai_epi16(in4, 15);
    const __m128i sign_in5 = _mm_srai_epi16(in5, 15);
    const __m128i sign_in6 = _mm_srai_epi16(in6, 15);
    const __m128i sign_in7 = _mm_srai_epi16(in7, 15);
    in0 = _mm_sub_epi16(in0, sign_in0);
    in1 = _mm_sub_epi16(in1, sign_in1);
    in2 = _mm_sub_epi16(in2, sign_in2);
    in3 = _mm_sub_epi16(in3, sign_in3);
    in4 = _mm_sub_epi16(in4, sign_in4);
    in5 = _mm_sub_epi16(in5, sign_in5);
    in6 = _mm_sub_epi16(in6, sign_in6);
    in7 = _mm_sub_epi16(in7, sign_in7);
    in0 = _mm_srai_epi16(in0, 1);
    in1 = _mm_srai_epi16(in1, 1);
    in2 = _mm_srai_epi16(in2, 1);
    in3 = _mm_srai_epi16(in3, 1);
    in4 = _mm_srai_epi16(in4, 1);
    in5 = _mm_srai_epi16(in5, 1);
    in6 = _mm_srai_epi16(in6, 1);
    in7 = _mm_srai_epi16(in7, 1);
  }

  iscan_ptr += n_coeffs;
  qcoeff_ptr += n_coeffs;
  dqcoeff_ptr += n_coeffs;
  n_coeffs = -n_coeffs;
  zero = _mm_setzero_si128();

  if (!skip_block) {
    __m128i eob;
    __m128i round, quant, dequant;
    {
      __m128i coeff0, coeff1;

      // Setup global values
      {
clang-format's avatar
clang-format committed
466 467 468
        round = _mm_load_si128((const __m128i *)round_ptr);
        quant = _mm_load_si128((const __m128i *)quant_ptr);
        dequant = _mm_load_si128((const __m128i *)dequant_ptr);
469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499
      }

      {
        __m128i coeff0_sign, coeff1_sign;
        __m128i qcoeff0, qcoeff1;
        __m128i qtmp0, qtmp1;
        // Do DC and first 15 AC
        coeff0 = *in[0];
        coeff1 = *in[1];

        // Poor man's sign extract
        coeff0_sign = _mm_srai_epi16(coeff0, 15);
        coeff1_sign = _mm_srai_epi16(coeff1, 15);
        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);

        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
        round = _mm_unpackhi_epi64(round, round);
        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
        quant = _mm_unpackhi_epi64(quant, quant);
        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);

        // Reinsert signs
        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);

clang-format's avatar
clang-format committed
500 501
        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
502 503 504 505 506

        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
        dequant = _mm_unpackhi_epi64(dequant, dequant);
        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);

clang-format's avatar
clang-format committed
507 508
        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
509 510 511 512 513 514 515 516 517 518 519 520
      }

      {
        // Scan for eob
        __m128i zero_coeff0, zero_coeff1;
        __m128i nzero_coeff0, nzero_coeff1;
        __m128i iscan0, iscan1;
        __m128i eob1;
        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
clang-format's avatar
clang-format committed
521 522
        iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
        iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541
        // Add one to convert from indices to counts
        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
        eob = _mm_and_si128(iscan0, nzero_coeff0);
        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
        eob = _mm_max_epi16(eob, eob1);
      }
      n_coeffs += 8 * 2;
    }

    // AC only loop
    index = 2;
    while (n_coeffs < 0) {
      __m128i coeff0, coeff1;
      {
        __m128i coeff0_sign, coeff1_sign;
        __m128i qcoeff0, qcoeff1;
        __m128i qtmp0, qtmp1;

542
        assert(index < (int)(sizeof(in) / sizeof(in[0])) - 1);
543 544 545 546 547 548 549 550 551 552 553 554 555 556 557 558 559 560 561 562 563 564
        coeff0 = *in[index];
        coeff1 = *in[index + 1];

        // Poor man's sign extract
        coeff0_sign = _mm_srai_epi16(coeff0, 15);
        coeff1_sign = _mm_srai_epi16(coeff1, 15);
        qcoeff0 = _mm_xor_si128(coeff0, coeff0_sign);
        qcoeff1 = _mm_xor_si128(coeff1, coeff1_sign);
        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);

        qcoeff0 = _mm_adds_epi16(qcoeff0, round);
        qcoeff1 = _mm_adds_epi16(qcoeff1, round);
        qtmp0 = _mm_mulhi_epi16(qcoeff0, quant);
        qtmp1 = _mm_mulhi_epi16(qcoeff1, quant);

        // Reinsert signs
        qcoeff0 = _mm_xor_si128(qtmp0, coeff0_sign);
        qcoeff1 = _mm_xor_si128(qtmp1, coeff1_sign);
        qcoeff0 = _mm_sub_epi16(qcoeff0, coeff0_sign);
        qcoeff1 = _mm_sub_epi16(qcoeff1, coeff1_sign);

clang-format's avatar
clang-format committed
565 566
        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), qcoeff0);
        _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, qcoeff1);
567 568 569 570

        coeff0 = _mm_mullo_epi16(qcoeff0, dequant);
        coeff1 = _mm_mullo_epi16(qcoeff1, dequant);

clang-format's avatar
clang-format committed
571 572
        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), coeff0);
        _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, coeff1);
573 574 575 576 577 578 579 580 581 582 583 584
      }

      {
        // Scan for eob
        __m128i zero_coeff0, zero_coeff1;
        __m128i nzero_coeff0, nzero_coeff1;
        __m128i iscan0, iscan1;
        __m128i eob0, eob1;
        zero_coeff0 = _mm_cmpeq_epi16(coeff0, zero);
        zero_coeff1 = _mm_cmpeq_epi16(coeff1, zero);
        nzero_coeff0 = _mm_cmpeq_epi16(zero_coeff0, zero);
        nzero_coeff1 = _mm_cmpeq_epi16(zero_coeff1, zero);
clang-format's avatar
clang-format committed
585 586
        iscan0 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs));
        iscan1 = _mm_load_si128((const __m128i *)(iscan_ptr + n_coeffs) + 1);
587 588 589 590 591 592 593 594 595 596 597 598 599 600 601 602 603 604 605 606 607 608 609 610 611
        // Add one to convert from indices to counts
        iscan0 = _mm_sub_epi16(iscan0, nzero_coeff0);
        iscan1 = _mm_sub_epi16(iscan1, nzero_coeff1);
        eob0 = _mm_and_si128(iscan0, nzero_coeff0);
        eob1 = _mm_and_si128(iscan1, nzero_coeff1);
        eob0 = _mm_max_epi16(eob0, eob1);
        eob = _mm_max_epi16(eob, eob0);
      }
      n_coeffs += 8 * 2;
      index += 2;
    }

    // Accumulate EOB
    {
      __m128i eob_shuffled;
      eob_shuffled = _mm_shuffle_epi32(eob, 0xe);
      eob = _mm_max_epi16(eob, eob_shuffled);
      eob_shuffled = _mm_shufflelo_epi16(eob, 0xe);
      eob = _mm_max_epi16(eob, eob_shuffled);
      eob_shuffled = _mm_shufflelo_epi16(eob, 0x1);
      eob = _mm_max_epi16(eob, eob_shuffled);
      *eob_ptr = _mm_extract_epi16(eob, 1);
    }
  } else {
    do {
clang-format's avatar
clang-format committed
612 613 614 615
      _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs), zero);
      _mm_store_si128((__m128i *)(dqcoeff_ptr + n_coeffs) + 1, zero);
      _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs), zero);
      _mm_store_si128((__m128i *)(qcoeff_ptr + n_coeffs) + 1, zero);
616 617 618 619 620 621
      n_coeffs += 8 * 2;
    } while (n_coeffs < 0);
    *eob_ptr = 0;
  }
}

622
// load 8x8 array
623 624
static INLINE void load_buffer_8x8(const int16_t *input, __m128i *in,
                                   int stride) {
clang-format's avatar
clang-format committed
625 626 627 628 629 630 631 632
  in[0] = _mm_load_si128((const __m128i *)(input + 0 * stride));
  in[1] = _mm_load_si128((const __m128i *)(input + 1 * stride));
  in[2] = _mm_load_si128((const __m128i *)(input + 2 * stride));
  in[3] = _mm_load_si128((const __m128i *)(input + 3 * stride));
  in[4] = _mm_load_si128((const __m128i *)(input + 4 * stride));
  in[5] = _mm_load_si128((const __m128i *)(input + 5 * stride));
  in[6] = _mm_load_si128((const __m128i *)(input + 6 * stride));
  in[7] = _mm_load_si128((const __m128i *)(input + 7 * stride));
633 634 635 636 637 638 639 640 641 642 643

  in[0] = _mm_slli_epi16(in[0], 2);
  in[1] = _mm_slli_epi16(in[1], 2);
  in[2] = _mm_slli_epi16(in[2], 2);
  in[3] = _mm_slli_epi16(in[3], 2);
  in[4] = _mm_slli_epi16(in[4], 2);
  in[5] = _mm_slli_epi16(in[5], 2);
  in[6] = _mm_slli_epi16(in[6], 2);
  in[7] = _mm_slli_epi16(in[7], 2);
}

644
// right shift and rounding
645
static INLINE void right_shift_8x8(__m128i *res, const int bit) {
646 647 648 649 650 651 652 653 654
  __m128i sign0 = _mm_srai_epi16(res[0], 15);
  __m128i sign1 = _mm_srai_epi16(res[1], 15);
  __m128i sign2 = _mm_srai_epi16(res[2], 15);
  __m128i sign3 = _mm_srai_epi16(res[3], 15);
  __m128i sign4 = _mm_srai_epi16(res[4], 15);
  __m128i sign5 = _mm_srai_epi16(res[5], 15);
  __m128i sign6 = _mm_srai_epi16(res[6], 15);
  __m128i sign7 = _mm_srai_epi16(res[7], 15);

655 656 657 658 659 660 661 662 663 664
  if (bit == 2) {
    const __m128i const_rounding = _mm_set1_epi16(1);
    res[0] = _mm_add_epi16(res[0], const_rounding);
    res[1] = _mm_add_epi16(res[1], const_rounding);
    res[2] = _mm_add_epi16(res[2], const_rounding);
    res[3] = _mm_add_epi16(res[3], const_rounding);
    res[4] = _mm_add_epi16(res[4], const_rounding);
    res[5] = _mm_add_epi16(res[5], const_rounding);
    res[6] = _mm_add_epi16(res[6], const_rounding);
    res[7] = _mm_add_epi16(res[7], const_rounding);
665 666
  }

667 668 669 670 671 672 673 674 675
  res[0] = _mm_sub_epi16(res[0], sign0);
  res[1] = _mm_sub_epi16(res[1], sign1);
  res[2] = _mm_sub_epi16(res[2], sign2);
  res[3] = _mm_sub_epi16(res[3], sign3);
  res[4] = _mm_sub_epi16(res[4], sign4);
  res[5] = _mm_sub_epi16(res[5], sign5);
  res[6] = _mm_sub_epi16(res[6], sign6);
  res[7] = _mm_sub_epi16(res[7], sign7);

676 677 678 679 680 681 682 683 684 685 686 687 688 689 690 691 692 693 694
  if (bit == 1) {
    res[0] = _mm_srai_epi16(res[0], 1);
    res[1] = _mm_srai_epi16(res[1], 1);
    res[2] = _mm_srai_epi16(res[2], 1);
    res[3] = _mm_srai_epi16(res[3], 1);
    res[4] = _mm_srai_epi16(res[4], 1);
    res[5] = _mm_srai_epi16(res[5], 1);
    res[6] = _mm_srai_epi16(res[6], 1);
    res[7] = _mm_srai_epi16(res[7], 1);
  } else {
    res[0] = _mm_srai_epi16(res[0], 2);
    res[1] = _mm_srai_epi16(res[1], 2);
    res[2] = _mm_srai_epi16(res[2], 2);
    res[3] = _mm_srai_epi16(res[3], 2);
    res[4] = _mm_srai_epi16(res[4], 2);
    res[5] = _mm_srai_epi16(res[5], 2);
    res[6] = _mm_srai_epi16(res[6], 2);
    res[7] = _mm_srai_epi16(res[7], 2);
  }
695 696 697
}

// write 8x8 array
698 699
static INLINE void write_buffer_8x8(tran_low_t *output, __m128i *res,
                                    int stride) {
Deb Mukherjee's avatar
Deb Mukherjee committed
700 701 702 703 704 705 706 707
  store_output(&res[0], (output + 0 * stride));
  store_output(&res[1], (output + 1 * stride));
  store_output(&res[2], (output + 2 * stride));
  store_output(&res[3], (output + 3 * stride));
  store_output(&res[4], (output + 4 * stride));
  store_output(&res[5], (output + 5 * stride));
  store_output(&res[6], (output + 6 * stride));
  store_output(&res[7], (output + 7 * stride));
708 709
}

710
static void fdct8_sse2(__m128i *in) {
711
  // constants
712
  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
713 714 715 716 717 718 719 720 721 722 723 724 725 726 727 728 729 730 731 732 733 734 735 736 737 738 739 740 741 742 743 744 745 746 747 748 749 750 751 752 753 754 755 756 757 758 759 760 761 762 763 764 765 766 767 768 769 770 771 772 773 774 775 776 777 778 779 780 781 782 783 784 785 786 787 788 789 790 791 792 793 794 795 796 797 798 799 800 801 802 803 804 805 806 807 808 809 810 811 812 813 814 815 816 817 818 819 820 821 822 823 824 825 826 827 828 829 830 831 832 833 834 835 836 837 838 839 840 841 842 843 844 845 846
  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_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__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);
  __m128i u0, u1, u2, u3, u4, u5, u6, u7;
  __m128i v0, v1, v2, v3, v4, v5, v6, v7;
  __m128i s0, s1, s2, s3, s4, s5, s6, s7;

  // stage 1
  s0 = _mm_add_epi16(in[0], in[7]);
  s1 = _mm_add_epi16(in[1], in[6]);
  s2 = _mm_add_epi16(in[2], in[5]);
  s3 = _mm_add_epi16(in[3], in[4]);
  s4 = _mm_sub_epi16(in[3], in[4]);
  s5 = _mm_sub_epi16(in[2], in[5]);
  s6 = _mm_sub_epi16(in[1], in[6]);
  s7 = _mm_sub_epi16(in[0], in[7]);

  u0 = _mm_add_epi16(s0, s3);
  u1 = _mm_add_epi16(s1, s2);
  u2 = _mm_sub_epi16(s1, s2);
  u3 = _mm_sub_epi16(s0, s3);
  // interleave and perform butterfly multiplication/addition
  v0 = _mm_unpacklo_epi16(u0, u1);
  v1 = _mm_unpackhi_epi16(u0, u1);
  v2 = _mm_unpacklo_epi16(u2, u3);
  v3 = _mm_unpackhi_epi16(u2, u3);

  u0 = _mm_madd_epi16(v0, k__cospi_p16_p16);
  u1 = _mm_madd_epi16(v1, k__cospi_p16_p16);
  u2 = _mm_madd_epi16(v0, k__cospi_p16_m16);
  u3 = _mm_madd_epi16(v1, k__cospi_p16_m16);
  u4 = _mm_madd_epi16(v2, k__cospi_p24_p08);
  u5 = _mm_madd_epi16(v3, k__cospi_p24_p08);
  u6 = _mm_madd_epi16(v2, k__cospi_m08_p24);
  u7 = _mm_madd_epi16(v3, k__cospi_m08_p24);

  // shift and rounding
  v0 = _mm_add_epi32(u0, k__DCT_CONST_ROUNDING);
  v1 = _mm_add_epi32(u1, k__DCT_CONST_ROUNDING);
  v2 = _mm_add_epi32(u2, k__DCT_CONST_ROUNDING);
  v3 = _mm_add_epi32(u3, k__DCT_CONST_ROUNDING);
  v4 = _mm_add_epi32(u4, k__DCT_CONST_ROUNDING);
  v5 = _mm_add_epi32(u5, k__DCT_CONST_ROUNDING);
  v6 = _mm_add_epi32(u6, k__DCT_CONST_ROUNDING);
  v7 = _mm_add_epi32(u7, k__DCT_CONST_ROUNDING);

  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);

  in[0] = _mm_packs_epi32(u0, u1);
  in[2] = _mm_packs_epi32(u4, u5);
  in[4] = _mm_packs_epi32(u2, u3);
  in[6] = _mm_packs_epi32(u6, u7);

  // stage 2
  // interleave and perform butterfly multiplication/addition
  u0 = _mm_unpacklo_epi16(s6, s5);
  u1 = _mm_unpackhi_epi16(s6, s5);
  v0 = _mm_madd_epi16(u0, k__cospi_p16_m16);
  v1 = _mm_madd_epi16(u1, k__cospi_p16_m16);
  v2 = _mm_madd_epi16(u0, k__cospi_p16_p16);
  v3 = _mm_madd_epi16(u1, k__cospi_p16_p16);

  // shift and rounding
  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);

  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);

  u0 = _mm_packs_epi32(v0, v1);
  u1 = _mm_packs_epi32(v2, v3);

  // stage 3
  s0 = _mm_add_epi16(s4, u0);
  s1 = _mm_sub_epi16(s4, u0);
  s2 = _mm_sub_epi16(s7, u1);
  s3 = _mm_add_epi16(s7, u1);

  // stage 4
  u0 = _mm_unpacklo_epi16(s0, s3);
  u1 = _mm_unpackhi_epi16(s0, s3);
  u2 = _mm_unpacklo_epi16(s1, s2);
  u3 = _mm_unpackhi_epi16(s1, s2);

  v0 = _mm_madd_epi16(u0, k__cospi_p28_p04);
  v1 = _mm_madd_epi16(u1, k__cospi_p28_p04);
  v2 = _mm_madd_epi16(u2, k__cospi_p12_p20);
  v3 = _mm_madd_epi16(u3, k__cospi_p12_p20);
  v4 = _mm_madd_epi16(u2, k__cospi_m20_p12);
  v5 = _mm_madd_epi16(u3, k__cospi_m20_p12);
  v6 = _mm_madd_epi16(u0, k__cospi_m04_p28);
  v7 = _mm_madd_epi16(u1, k__cospi_m04_p28);

  // shift and rounding
  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);

  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);

  in[1] = _mm_packs_epi32(v0, v1);
  in[3] = _mm_packs_epi32(v4, v5);
  in[5] = _mm_packs_epi32(v2, v3);
  in[7] = _mm_packs_epi32(v6, v7);

  // transpose
847
  transpose_16bit_8x8(in, in);
848 849
}

850
static void fadst8_sse2(__m128i *in) {
851 852 853 854 855 856 857 858 859 860 861 862 863
  // Constants
  const __m128i k__cospi_p02_p30 = pair_set_epi16(cospi_2_64, cospi_30_64);
  const __m128i k__cospi_p30_m02 = pair_set_epi16(cospi_30_64, -cospi_2_64);
  const __m128i k__cospi_p10_p22 = pair_set_epi16(cospi_10_64, cospi_22_64);
  const __m128i k__cospi_p22_m10 = pair_set_epi16(cospi_22_64, -cospi_10_64);
  const __m128i k__cospi_p18_p14 = pair_set_epi16(cospi_18_64, cospi_14_64);
  const __m128i k__cospi_p14_m18 = pair_set_epi16(cospi_14_64, -cospi_18_64);
  const __m128i k__cospi_p26_p06 = pair_set_epi16(cospi_26_64, cospi_6_64);
  const __m128i k__cospi_p06_m26 = pair_set_epi16(cospi_6_64, -cospi_26_64);
  const __m128i k__cospi_p08_p24 = pair_set_epi16(cospi_8_64, cospi_24_64);
  const __m128i k__cospi_p24_m08 = pair_set_epi16(cospi_24_64, -cospi_8_64);
  const __m128i k__cospi_m24_p08 = pair_set_epi16(-cospi_24_64, cospi_8_64);
  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
864
  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
865 866 867 868 869 870 871 872 873 874
  const __m128i k__const_0 = _mm_set1_epi16(0);
  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);

  __m128i u0, u1, u2, u3, u4, u5, u6, u7, u8, u9, u10, u11, u12, u13, u14, u15;
  __m128i v0, v1, v2, v3, v4, v5, v6, v7, v8, v9, v10, v11, v12, v13, v14, v15;
  __m128i w0, w1, w2, w3, w4, w5, w6, w7, w8, w9, w10, w11, w12, w13, w14, w15;
  __m128i s0, s1, s2, s3, s4, s5, s6, s7;
  __m128i in0, in1, in2, in3, in4, in5, in6, in7;

  // properly aligned for butterfly input
clang-format's avatar
clang-format committed
875 876 877 878 879 880 881 882
  in0 = in[7];
  in1 = in[0];
  in2 = in[5];
  in3 = in[2];
  in4 = in[3];
  in5 = in[4];
  in6 = in[1];
  in7 = in[6];
883 884 885 886 887 888 889 890 891 892 893 894 895 896 897 898 899 900 901 902 903 904 905 906 907 908 909 910 911 912 913 914 915 916 917 918 919 920 921 922 923 924 925 926 927 928 929 930 931 932 933 934 935 936 937 938 939 940 941 942 943 944 945 946 947 948 949 950 951 952 953 954 955 956 957 958 959 960 961 962 963 964 965 966 967 968 969 970 971 972 973 974 975 976 977 978 979 980 981 982 983 984 985 986 987 988 989 990 991 992 993 994 995 996 997 998 999 1000 1001 1002 1003 1004 1005 1006 1007 1008 1009 1010 1011 1012 1013 1014 1015 1016 1017 1018 1019 1020 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 1058 1059 1060 1061 1062 1063 1064 1065 1066 1067 1068 1069 1070 1071 1072 1073 1074 1075 1076

  // column transformation
  // stage 1
  // interleave and multiply/add into 32-bit integer
  s0 = _mm_unpacklo_epi16(in0, in1);
  s1 = _mm_unpackhi_epi16(in0, in1);
  s2 = _mm_unpacklo_epi16(in2, in3);
  s3 = _mm_unpackhi_epi16(in2, in3);
  s4 = _mm_unpacklo_epi16(in4, in5);
  s5 = _mm_unpackhi_epi16(in4, in5);
  s6 = _mm_unpacklo_epi16(in6, in7);
  s7 = _mm_unpackhi_epi16(in6, in7);

  u0 = _mm_madd_epi16(s0, k__cospi_p02_p30);
  u1 = _mm_madd_epi16(s1, k__cospi_p02_p30);
  u2 = _mm_madd_epi16(s0, k__cospi_p30_m02);
  u3 = _mm_madd_epi16(s1, k__cospi_p30_m02);
  u4 = _mm_madd_epi16(s2, k__cospi_p10_p22);
  u5 = _mm_madd_epi16(s3, k__cospi_p10_p22);
  u6 = _mm_madd_epi16(s2, k__cospi_p22_m10);
  u7 = _mm_madd_epi16(s3, k__cospi_p22_m10);
  u8 = _mm_madd_epi16(s4, k__cospi_p18_p14);
  u9 = _mm_madd_epi16(s5, k__cospi_p18_p14);
  u10 = _mm_madd_epi16(s4, k__cospi_p14_m18);
  u11 = _mm_madd_epi16(s5, k__cospi_p14_m18);
  u12 = _mm_madd_epi16(s6, k__cospi_p26_p06);
  u13 = _mm_madd_epi16(s7, k__cospi_p26_p06);
  u14 = _mm_madd_epi16(s6, k__cospi_p06_m26);
  u15 = _mm_madd_epi16(s7, k__cospi_p06_m26);

  // addition
  w0 = _mm_add_epi32(u0, u8);
  w1 = _mm_add_epi32(u1, u9);
  w2 = _mm_add_epi32(u2, u10);
  w3 = _mm_add_epi32(u3, u11);
  w4 = _mm_add_epi32(u4, u12);
  w5 = _mm_add_epi32(u5, u13);
  w6 = _mm_add_epi32(u6, u14);
  w7 = _mm_add_epi32(u7, u15);
  w8 = _mm_sub_epi32(u0, u8);
  w9 = _mm_sub_epi32(u1, u9);
  w10 = _mm_sub_epi32(u2, u10);
  w11 = _mm_sub_epi32(u3, u11);
  w12 = _mm_sub_epi32(u4, u12);
  w13 = _mm_sub_epi32(u5, u13);
  w14 = _mm_sub_epi32(u6, u14);
  w15 = _mm_sub_epi32(u7, u15);

  // shift and rounding
  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);
  v8 = _mm_add_epi32(w8, k__DCT_CONST_ROUNDING);
  v9 = _mm_add_epi32(w9, k__DCT_CONST_ROUNDING);
  v10 = _mm_add_epi32(w10, k__DCT_CONST_ROUNDING);
  v11 = _mm_add_epi32(w11, k__DCT_CONST_ROUNDING);
  v12 = _mm_add_epi32(w12, k__DCT_CONST_ROUNDING);
  v13 = _mm_add_epi32(w13, k__DCT_CONST_ROUNDING);
  v14 = _mm_add_epi32(w14, k__DCT_CONST_ROUNDING);
  v15 = _mm_add_epi32(w15, k__DCT_CONST_ROUNDING);

  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);
  u8 = _mm_srai_epi32(v8, DCT_CONST_BITS);
  u9 = _mm_srai_epi32(v9, DCT_CONST_BITS);
  u10 = _mm_srai_epi32(v10, DCT_CONST_BITS);
  u11 = _mm_srai_epi32(v11, DCT_CONST_BITS);
  u12 = _mm_srai_epi32(v12, DCT_CONST_BITS);
  u13 = _mm_srai_epi32(v13, DCT_CONST_BITS);
  u14 = _mm_srai_epi32(v14, DCT_CONST_BITS);
  u15 = _mm_srai_epi32(v15, DCT_CONST_BITS);

  // back to 16-bit and pack 8 integers into __m128i
  in[0] = _mm_packs_epi32(u0, u1);
  in[1] = _mm_packs_epi32(u2, u3);
  in[2] = _mm_packs_epi32(u4, u5);
  in[3] = _mm_packs_epi32(u6, u7);
  in[4] = _mm_packs_epi32(u8, u9);
  in[5] = _mm_packs_epi32(u10, u11);
  in[6] = _mm_packs_epi32(u12, u13);
  in[7] = _mm_packs_epi32(u14, u15);

  // stage 2
  s0 = _mm_add_epi16(in[0], in[2]);
  s1 = _mm_add_epi16(in[1], in[3]);
  s2 = _mm_sub_epi16(in[0], in[2]);
  s3 = _mm_sub_epi16(in[1], in[3]);
  u0 = _mm_unpacklo_epi16(in[4], in[5]);
  u1 = _mm_unpackhi_epi16(in[4], in[5]);
  u2 = _mm_unpacklo_epi16(in[6], in[7]);
  u3 = _mm_unpackhi_epi16(in[6], in[7]);

  v0 = _mm_madd_epi16(u0, k__cospi_p08_p24);
  v1 = _mm_madd_epi16(u1, k__cospi_p08_p24);
  v2 = _mm_madd_epi16(u0, k__cospi_p24_m08);
  v3 = _mm_madd_epi16(u1, k__cospi_p24_m08);
  v4 = _mm_madd_epi16(u2, k__cospi_m24_p08);
  v5 = _mm_madd_epi16(u3, k__cospi_m24_p08);
  v6 = _mm_madd_epi16(u2, k__cospi_p08_p24);
  v7 = _mm_madd_epi16(u3, k__cospi_p08_p24);

  w0 = _mm_add_epi32(v0, v4);
  w1 = _mm_add_epi32(v1, v5);
  w2 = _mm_add_epi32(v2, v6);
  w3 = _mm_add_epi32(v3, v7);
  w4 = _mm_sub_epi32(v0, v4);
  w5 = _mm_sub_epi32(v1, v5);
  w6 = _mm_sub_epi32(v2, v6);
  w7 = _mm_sub_epi32(v3, v7);

  v0 = _mm_add_epi32(w0, k__DCT_CONST_ROUNDING);
  v1 = _mm_add_epi32(w1, k__DCT_CONST_ROUNDING);
  v2 = _mm_add_epi32(w2, k__DCT_CONST_ROUNDING);
  v3 = _mm_add_epi32(w3, k__DCT_CONST_ROUNDING);
  v4 = _mm_add_epi32(w4, k__DCT_CONST_ROUNDING);
  v5 = _mm_add_epi32(w5, k__DCT_CONST_ROUNDING);
  v6 = _mm_add_epi32(w6, k__DCT_CONST_ROUNDING);
  v7 = _mm_add_epi32(w7, k__DCT_CONST_ROUNDING);

  u0 = _mm_srai_epi32(v0, DCT_CONST_BITS);
  u1 = _mm_srai_epi32(v1, DCT_CONST_BITS);
  u2 = _mm_srai_epi32(v2, DCT_CONST_BITS);
  u3 = _mm_srai_epi32(v3, DCT_CONST_BITS);
  u4 = _mm_srai_epi32(v4, DCT_CONST_BITS);
  u5 = _mm_srai_epi32(v5, DCT_CONST_BITS);
  u6 = _mm_srai_epi32(v6, DCT_CONST_BITS);
  u7 = _mm_srai_epi32(v7, DCT_CONST_BITS);

  // back to 16-bit intergers
  s4 = _mm_packs_epi32(u0, u1);
  s5 = _mm_packs_epi32(u2, u3);
  s6 = _mm_packs_epi32(u4, u5);
  s7 = _mm_packs_epi32(u6, u7);

  // stage 3
  u0 = _mm_unpacklo_epi16(s2, s3);
  u1 = _mm_unpackhi_epi16(s2, s3);
  u2 = _mm_unpacklo_epi16(s6, s7);
  u3 = _mm_unpackhi_epi16(s6, s7);

  v0 = _mm_madd_epi16(u0, k__cospi_p16_p16);
  v1 = _mm_madd_epi16(u1, k__cospi_p16_p16);
  v2 = _mm_madd_epi16(u0, k__cospi_p16_m16);
  v3 = _mm_madd_epi16(u1, k__cospi_p16_m16);
  v4 = _mm_madd_epi16(u2, k__cospi_p16_p16);
  v5 = _mm_madd_epi16(u3, k__cospi_p16_p16);
  v6 = _mm_madd_epi16(u2, k__cospi_p16_m16);
  v7 = _mm_madd_epi16(u3, k__cospi_p16_m16);

  u0 = _mm_add_epi32(v0, k__DCT_CONST_ROUNDING);
  u1 = _mm_add_epi32(v1, k__DCT_CONST_ROUNDING);
  u2 = _mm_add_epi32(v2, k__DCT_CONST_ROUNDING);
  u3 = _mm_add_epi32(v3, k__DCT_CONST_ROUNDING);
  u4 = _mm_add_epi32(v4, k__DCT_CONST_ROUNDING);
  u5 = _mm_add_epi32(v5, k__DCT_CONST_ROUNDING);
  u6 = _mm_add_epi32(v6, k__DCT_CONST_ROUNDING);
  u7 = _mm_add_epi32(v7, k__DCT_CONST_ROUNDING);

  v0 = _mm_srai_epi32(u0, DCT_CONST_BITS);
  v1 = _mm_srai_epi32(u1, DCT_CONST_BITS);
  v2 = _mm_srai_epi32(u2, DCT_CONST_BITS);
  v3 = _mm_srai_epi32(u3, DCT_CONST_BITS);
  v4 = _mm_srai_epi32(u4, DCT_CONST_BITS);
  v5 = _mm_srai_epi32(u5, DCT_CONST_BITS);
  v6 = _mm_srai_epi32(u6, DCT_CONST_BITS);
  v7 = _mm_srai_epi32(u7, DCT_CONST_BITS);

  s2 = _mm_packs_epi32(v0, v1);
  s3 = _mm_packs_epi32(v2, v3);
  s6 = _mm_packs_epi32(v4, v5);
  s7 = _mm_packs_epi32(v6, v7);

  // FIXME(jingning): do subtract using bit inversion?
  in[0] = s0;
  in[1] = _mm_sub_epi16(k__const_0, s4);
  in[2] = s6;
  in[3] = _mm_sub_epi16(k__const_0, s2);
  in[4] = s3;
  in[5] = _mm_sub_epi16(k__const_0, s7);
  in[6] = s5;
  in[7] = _mm_sub_epi16(k__const_0, s1);

  // transpose
1077
  transpose_16bit_8x8(in, in);
1078 1079
}

clang-format's avatar
clang-format committed
1080 1081
void vp9_fht8x8_sse2(const int16_t *input, tran_low_t *output, int stride,
                     int tx_type) {
1082
  __m128i in[8];
1083

1084
  switch (tx_type) {
clang-format's avatar
clang-format committed
1085
    case DCT_DCT: vpx_fdct8x8_sse2(input, output, stride); break;
1086 1087
    case ADST_DCT:
      load_buffer_8x8(input, in, stride);
1088 1089
      fadst8_sse2(in);
      fdct8_sse2(in);
1090 1091
      right_shift_8x8(in, 1);
      write_buffer_8x8(output, in, 8);
1092
      break;
1093 1094
    case DCT_ADST:
      load_buffer_8x8(input, in, stride);
1095 1096
      fdct8_sse2(in);
      fadst8_sse2(in);
1097 1098
      right_shift_8x8(in, 1);
      write_buffer_8x8(output, in, 8);
1099
      break;
1100 1101
    case ADST_ADST:
      load_buffer_8x8(input, in, stride);
1102 1103
      fadst8_sse2(in);
      fadst8_sse2(in);
1104 1105
      right_shift_8x8(in, 1);
      write_buffer_8x8(output, in, 8);
1106
      break;
clang-format's avatar
clang-format committed
1107
    default: assert(0); break;
1108 1109 1110
  }
}

clang-format's avatar
clang-format committed
1111
static INLINE void load_buffer_16x16(const int16_t *input, __m128i *in0,
1112 1113 1114 1115 1116 1117 1118 1119 1120 1121 1122
                                     __m128i *in1, int stride) {
  // load first 8 columns
  load_buffer_8x8(input, in0, stride);
  load_buffer_8x8(input + 8 * stride, in0 + 8, stride);

  input += 8;
  // load second 8 columns
  load_buffer_8x8(input, in1, stride);
  load_buffer_8x8(input + 8 * stride, in1 + 8, stride);
}

1123
static INLINE void write_buffer_16x16(tran_low_t *output, __m128i *in0,
1124 1125 1126 1127 1128 1129 1130 1131 1132 1133 1134 1135 1136 1137 1138 1139 1140 1141
                                      __m128i *in1, int stride) {
  // write first 8 columns
  write_buffer_8x8(output, in0, stride);
  write_buffer_8x8(output + 8 * stride, in0 + 8, stride);
  // write second 8 columns
  output += 8;
  write_buffer_8x8(output, in1, stride);
  write_buffer_8x8(output + 8 * stride, in1 + 8, stride);
}

static INLINE void right_shift_16x16(__m128i *res0, __m128i *res1) {
  // perform rounding operations
  right_shift_8x8(res0, 2);
  right_shift_8x8(res0 + 8, 2);
  right_shift_8x8(res1, 2);
  right_shift_8x8(res1 + 8, 2);
}

1142
static void fdct16_8col(__m128i *in) {
1143 1144
  // perform 16x16 1-D DCT for 8 columns
  __m128i i[8], s[8], p[8], t[8], u[16], v[16];
1145
  const __m128i k__cospi_p16_p16 = _mm_set1_epi16(cospi_16_64);
1146 1147 1148
  const __m128i k__cospi_p16_m16 = pair_set_epi16(cospi_16_64, -cospi_16_64);
  const __m128i k__cospi_m16_p16 = pair_set_epi16(-cospi_16_64, cospi_16_64);
  const __m128i k__cospi_p24_p08 = pair_set_epi16(cospi_24_64, cospi_8_64);
1149
  const __m128i k__cospi_p08_m24 = pair_set_epi16(cospi_8_64, -cospi_24_64);
1150 1151 1152 1153 1154 1155 1156 1157 1158 1159 1160 1161 1162 1163 1164 1165 1166 1167 1168 1169 1170 1171 1172 1173 1174 1175 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 1245 1246 1247 1248 1249 1250 1251 1252 1253 1254 1255 1256 1257 1258 1259 1260 1261 1262 1263 1264 1265 1266 1267 1268 1269 1270 1271 1272 1273 1274 1275 1276 1277 1278 1279 1280 1281 1282 1283 1284 1285 1286 1287 1288 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
  const __m128i k__cospi_m08_p24 = pair_set_epi16(-cospi_8_64, cospi_24_64);
  const __m128i k__cospi_p28_p04 = pair_set_epi16(cospi_28_64, cospi_4_64);
  const __m128i k__cospi_m04_p28 = pair_set_epi16(-cospi_4_64, cospi_28_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_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_m02_p30 = pair_set_epi16(-cospi_2_64, cospi_30_64);
  const __m128i k__cospi_m18_p14 = pair_set_epi16(-cospi_18_64, cospi_14_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_m10_p22 = pair_set_epi16(-cospi_10_64, cospi_22_64);
  const __m128i k__cospi_m26_p06 = pair_set_epi16(-cospi_26_64, cospi_6_64);
  const __m128i k__DCT_CONST_ROUNDING = _mm_set1_epi32(DCT_CONST_ROUNDING);

  // stage 1
  i[0] = _mm_add_epi16(in[0], in[15]);
  i[1] = _mm_add_epi16(in[1], in[14]);
  i[2] = _mm_add_epi16(in[2], in[13]);
  i[3] = _mm_add_epi16(in[3], in[12]);
  i[4] = _mm_add_epi16(in[4], in[11]);
  i[5] = _mm_add_epi16(in[5], in[10]);
  i[6] = _mm_add_epi16(in[6], in[9]);
  i[7] = _mm_add_epi16(in[7], in[8]);

  s[0] = _mm_sub_epi16(in[7], in[8]);
  s[1] = _mm_sub_epi16(in[6], in[9]);
  s[2] = _mm_sub_epi16(in[5], in[10]);
  s[3] = _mm_sub_epi16(in[4], in[11]);
  s[4] = _mm_sub_epi16(in[3], in[12]);
  s[5] = _mm_sub_epi16(in[2], in[13]);
  s[6] = _mm_sub_epi16(in[1], in[14]);
  s[7] = _mm_sub_epi16(in[0], in[15]);

  p[0] = _mm_add_epi16(i[0], i[7]);
  p[1] = _mm_add_epi16(i[1], i[6]);
  p[2] = _mm_add_epi16(i[2], i[5]);
  p[3] = _mm_add_epi16(i[3], i[4]);
  p[4] = _mm_sub_epi16(i[3], i[4]);
  p[5] = _mm_sub_epi16(i[2], i[5]);
  p[6] = _mm_sub_epi16(i[1], i[6]);
  p[7] = _mm_sub_epi16(i[0], i[7]);

  u[0] = _mm_add_epi16(p[0], p[3]);
  u[1] = _mm_add_epi16(p[1], p[2]);
  u[2] = _mm_sub_epi16(p[1], p[2]);
  u[3] = _mm_sub_epi16(p[0], p[3]);

  v[0] = _mm_unpacklo_epi16(u[0], u[1]);
  v[1] = _mm_unpackhi_epi16(u[0], u[1]);
  v[2] = _mm_unpacklo_epi16(u[2], u[3]);
  v[3] = _mm_unpackhi_epi16(u[2], u[3]);

  u[0] = _mm_madd_epi16(v[0], k__cospi_p16_p16);
  u[1] = _mm_madd_epi16(v[1], k__cospi_p16_p16);
  u[2] = _mm_madd_epi16(v[0], k__cospi_p16_m16);
  u[3] = _mm_madd_epi16(v[1], k__cospi_p16_m16);
  u[4] = _mm_madd_epi16(v[2], k__cospi_p24_p08);
  u[5] = _mm_madd_epi16(v[3], k__cospi_p24_p08);
  u[6] = _mm_madd_epi16(v[2], k__cospi_m08_p24);
  u[7] = _mm_madd_epi16(v[3], k__cospi_m08_p24);

  v[0] = _mm_add_epi32(u[0], k__DCT_CONST_ROUNDING);
  v[1] = _mm_add_epi32(u[1], k__DCT_CONST_ROUNDING);
  v[2] = _mm_add_epi32(u[2], k__DCT_CONST_ROUNDING);
  v[3] = _mm_add_epi32(u[3], k__DCT_CONST_ROUNDING);
  v[4] = _mm_add_epi32(u[4], k__DCT_CONST_ROUNDING);
  v[5] = _mm_add_epi32(u[5], k__DCT_CONST_ROUNDING);
  v[6] = _mm_add_epi32(u[6], k__DCT_CONST_ROUNDING);
  v[7] = _mm_add_epi32(u[7], k__DCT_CONST_ROUNDING);

  u[0] = _mm_srai_epi32(v[0], DCT_CONST_BITS);
  u[1] = _mm_srai_epi32(v[1], DCT_CONST_BITS);
  u[2] = _mm_srai_epi32(v[2], DCT_CONST_BITS);
  u[3] = _mm_srai_epi32(v[3], DCT_CONST_BITS);
  u[4] = _mm_srai_epi32(v[4], DCT_CONST_BITS);
  u[5] = _mm_srai_epi32(v[5], DCT_CONST_BITS);
  u[6] = _mm_srai_epi32(v[6], DCT_CONST_BITS);
  u[7] = _mm_srai_epi32(v[7], DCT_CONST_BITS);

  in[0] = _mm_packs_epi32(u[0], u[1]);
  in[4] = _mm_packs_epi32(u[4], u[5]);
  in[8] = _mm_packs_epi32(u[2], u[3]);
  in[12] = _mm_packs_epi32(u[6], u[7]);

  u[0] = _mm_unpacklo_epi16(p[5], p[6]);
  u[1] = _mm_unpackhi_epi16(p[5], p[6]);
  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
  v[2] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
  v[3] = _mm_madd_epi16(u[1], k__cospi_p16_p16);

  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);

  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);

  u[0] = _mm_packs_epi32(v[0], v[1]);
  u[1] = _mm_packs_epi32(v[2], v[3]);

  t[0] = _mm_add_epi16(p[4], u[0]);
  t[1] = _mm_sub_epi16(p[4], u[0]);
  t[2] = _mm_sub_epi16(p[7], u[1]);
  t[3] = _mm_add_epi16(p[7], u[1]);

  u[0] = _mm_unpacklo_epi16(t[0], t[3]);
  u[1] = _mm_unpackhi_epi16(t[0], t[3]);
  u[2] = _mm_unpacklo_epi16(t[1], t[2]);
  u[3] = _mm_unpackhi_epi16(t[1], t[2]);

  v[0] = _mm_madd_epi16(u[0], k__cospi_p28_p04);
  v[1] = _mm_madd_epi16(u[1], k__cospi_p28_p04);
  v[2] = _mm_madd_epi16(u[2], k__cospi_p12_p20);
  v[3] = _mm_madd_epi16(u[3], k__cospi_p12_p20);
  v[4] = _mm_madd_epi16(u[2], k__cospi_m20_p12);
  v[5] = _mm_madd_epi16(u[3], k__cospi_m20_p12);
  v[6] = _mm_madd_epi16(u[0], k__cospi_m04_p28);
  v[7] = _mm_madd_epi16(u[1], k__cospi_m04_p28);

  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);

  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);

  in[2] = _mm_packs_epi32(v[0], v[1]);
  in[6] = _mm_packs_epi32(v[4], v[5]);
  in[10] = _mm_packs_epi32(v[2], v[3]);
  in[14] = _mm_packs_epi32(v[6], v[7]);

  // stage 2
  u[0] = _mm_unpacklo_epi16(s[2], s[5]);
  u[1] = _mm_unpackhi_epi16(s[2], s[5]);
  u[2] = _mm_unpacklo_epi16(s[3], s[4]);
  u[3] = _mm_unpackhi_epi16(s[3], s[4]);

  v[0] = _mm_madd_epi16(u[0], k__cospi_m16_p16);
  v[1] = _mm_madd_epi16(u[1], k__cospi_m16_p16);
  v[2] = _mm_madd_epi16(u[2], k__cospi_m16_p16);
  v[3] = _mm_madd_epi16(u[3], k__cospi_m16_p16);
  v[4] = _mm_madd_epi16(u[2], k__cospi_p16_p16);
  v[5] = _mm_madd_epi16(u[3], k__cospi_p16_p16);
  v[6] = _mm_madd_epi16(u[0], k__cospi_p16_p16);
  v[7] = _mm_madd_epi16(u[1], k__cospi_p16_p16);

  u[0] = _mm_add_epi32(v[0], k__DCT_CONST_ROUNDING);
  u[1] = _mm_add_epi32(v[1], k__DCT_CONST_ROUNDING);
  u[2] = _mm_add_epi32(v[2], k__DCT_CONST_ROUNDING);
  u[3] = _mm_add_epi32(v[3], k__DCT_CONST_ROUNDING);
  u[4] = _mm_add_epi32(v[4], k__DCT_CONST_ROUNDING);
  u[5] = _mm_add_epi32(v[5], k__DCT_CONST_ROUNDING);
  u[6] = _mm_add_epi32(v[6], k__DCT_CONST_ROUNDING);
  u[7] = _mm_add_epi32(v[7], k__DCT_CONST_ROUNDING);

  v[0] = _mm_srai_epi32(u[0], DCT_CONST_BITS);
  v[1] = _mm_srai_epi32(u[1], DCT_CONST_BITS);
  v[2] = _mm_srai_epi32(u[2], DCT_CONST_BITS);
  v[3] = _mm_srai_epi32(u[3], DCT_CONST_BITS);
  v[4] = _mm_srai_epi32(u[4], DCT_CONST_BITS);
  v[5] = _mm_srai_epi32(u[5], DCT_CONST_BITS);
  v[6] = _mm_srai_epi32(u[6], DCT_CONST_BITS);
  v[7] = _mm_srai_epi32(u[7], DCT_CONST_BITS);

  t[2] = _mm_packs_epi32(v[0], v[1]);
  t[3] = _mm_packs_epi32(v[2], v[3]);
  t[4] = _mm_packs_epi32(v[4], v[5]);
  t[5] = _mm_packs_epi32(v[6], v[7]);

  // stage 3
  p[0] = _mm_add_epi16(s[0], t[3]);
  p[1] = _mm_add_epi16(s[1], t[2]);
  p[2] = _mm_sub_epi16(s[1], t[2]);
  p[3] = _mm_sub_epi16(s[0], t[3]);
  p[4] = _mm_sub_epi16(s[7], t[4]);
  p[5] = _mm_sub_epi16(s[6], t[5]);
  p[6] = _mm_add_epi16(s[6], t[5]);
  p[7] = _mm_add_epi16(s[7], t[4]);

  // stage 4
  u[0] = _mm_unpacklo_epi16(p[1], p[6]);
  u[1] = _mm_unpackhi_epi16(p[1], p[6]);
  u[2] = _mm_unpacklo_epi16(p[2], p[5]);
  u[3] = _mm_unpackhi_epi16(p[2], p[5]);

  v[0] = _mm_madd_epi16(u[0], k__cospi_m08_p24);
  v[1] = _mm_madd_epi16(u[1], k__cospi_m08_p24);
1353 1354 1355 1356
  v[2] = _mm_madd_epi16(u[2], k__cospi_p24_p08);
  v[3] = _mm_madd_epi16(u[3], k__cospi_p24_p08);
  v[4] = _mm_madd_epi16(u[2], k__cospi_p08_m24);
  v[5] = _mm_madd_epi16(u[3], k__cospi_p08_m24);
1357 1358 1359 1360 1361 1362 1363 1364 1365 1366 1367 1368 1369 1370 1371 1372 1373 1374 1375 1376