Coverage Report

Created: 2025-06-13 07:07

/src/aom/av1/common/x86/convolve_sse2.c
Line
Count
Source (jump to first uncovered line)
1
/*
2
 * Copyright (c) 2017, Alliance for Open Media. All rights reserved.
3
 *
4
 * This source code is subject to the terms of the BSD 2 Clause License and
5
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6
 * was not distributed with this source code in the LICENSE file, you can
7
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8
 * Media Patent License 1.0 was not distributed with this source code in the
9
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10
 */
11
12
#include <emmintrin.h>
13
14
#include "config/av1_rtcd.h"
15
16
#include "aom_dsp/aom_dsp_common.h"
17
#include "aom_dsp/aom_filter.h"
18
#include "aom_dsp/x86/convolve_common_intrin.h"
19
#include "aom_dsp/x86/synonyms.h"
20
#include "av1/common/convolve.h"
21
22
static inline void prepare_coeffs(const InterpFilterParams *const filter_params,
23
                                  const int subpel_q4,
24
0
                                  __m128i *const coeffs /* [4] */) {
25
0
  const int16_t *const y_filter = av1_get_interp_filter_subpel_kernel(
26
0
      filter_params, subpel_q4 & SUBPEL_MASK);
27
0
  const __m128i coeffs_y = _mm_loadu_si128((__m128i *)y_filter);
28
  // coeffs 0 1 0 1 2 3 2 3
29
0
  const __m128i tmp_0 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
30
  // coeffs 4 5 4 5 6 7 6 7
31
0
  const __m128i tmp_1 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);
32
33
0
  coeffs[0] = _mm_unpacklo_epi64(tmp_0, tmp_0);  // coeffs 0 1 0 1 0 1 0 1
34
0
  coeffs[1] = _mm_unpackhi_epi64(tmp_0, tmp_0);  // coeffs 2 3 2 3 2 3 2 3
35
0
  coeffs[2] = _mm_unpacklo_epi64(tmp_1, tmp_1);  // coeffs 4 5 4 5 4 5 4 5
36
0
  coeffs[3] = _mm_unpackhi_epi64(tmp_1, tmp_1);  // coeffs 6 7 6 7 6 7 6 7
37
0
}
38
39
static inline __m128i convolve(const __m128i *const s,
40
0
                               const __m128i *const coeffs) {
41
0
  const __m128i d0 = _mm_madd_epi16(s[0], coeffs[0]);
42
0
  const __m128i d1 = _mm_madd_epi16(s[1], coeffs[1]);
43
0
  const __m128i d2 = _mm_madd_epi16(s[2], coeffs[2]);
44
0
  const __m128i d3 = _mm_madd_epi16(s[3], coeffs[3]);
45
0
  const __m128i d = _mm_add_epi32(_mm_add_epi32(d0, d1), _mm_add_epi32(d2, d3));
46
0
  return d;
47
0
}
48
49
static inline __m128i convolve_lo_x(const __m128i *const s,
50
0
                                    const __m128i *const coeffs) {
51
0
  __m128i ss[4];
52
0
  ss[0] = _mm_unpacklo_epi8(s[0], _mm_setzero_si128());
53
0
  ss[1] = _mm_unpacklo_epi8(s[1], _mm_setzero_si128());
54
0
  ss[2] = _mm_unpacklo_epi8(s[2], _mm_setzero_si128());
55
0
  ss[3] = _mm_unpacklo_epi8(s[3], _mm_setzero_si128());
56
0
  return convolve(ss, coeffs);
57
0
}
58
59
static inline __m128i convolve_lo_y(const __m128i *const s,
60
0
                                    const __m128i *const coeffs) {
61
0
  __m128i ss[4];
62
0
  ss[0] = _mm_unpacklo_epi8(s[0], _mm_setzero_si128());
63
0
  ss[1] = _mm_unpacklo_epi8(s[2], _mm_setzero_si128());
64
0
  ss[2] = _mm_unpacklo_epi8(s[4], _mm_setzero_si128());
65
0
  ss[3] = _mm_unpacklo_epi8(s[6], _mm_setzero_si128());
66
0
  return convolve(ss, coeffs);
67
0
}
68
69
static inline __m128i convolve_hi_y(const __m128i *const s,
70
0
                                    const __m128i *const coeffs) {
71
0
  __m128i ss[4];
72
0
  ss[0] = _mm_unpackhi_epi8(s[0], _mm_setzero_si128());
73
0
  ss[1] = _mm_unpackhi_epi8(s[2], _mm_setzero_si128());
74
0
  ss[2] = _mm_unpackhi_epi8(s[4], _mm_setzero_si128());
75
0
  ss[3] = _mm_unpackhi_epi8(s[6], _mm_setzero_si128());
76
0
  return convolve(ss, coeffs);
77
0
}
78
79
static void convolve_y_sr_12tap_sse2(const uint8_t *src, int src_stride,
80
                                     uint8_t *dst, int dst_stride, int w, int h,
81
                                     const InterpFilterParams *filter_params_y,
82
0
                                     int subpel_y_qn) {
83
0
  const int fo_vert = filter_params_y->taps / 2 - 1;
84
0
  const uint8_t *src_ptr = src - fo_vert * src_stride;
85
0
  const __m128i round_const = _mm_set1_epi32((1 << FILTER_BITS) >> 1);
86
0
  const __m128i round_shift = _mm_cvtsi32_si128(FILTER_BITS);
87
0
  __m128i coeffs[6];
88
89
0
  prepare_coeffs_12tap(filter_params_y, subpel_y_qn, coeffs);
90
91
0
  int j = 0;
92
0
  do {
93
0
    __m128i s[12], src10, res_lo, res_hi;
94
0
    __m128i res_lo_round, res_hi_round, res16, res;
95
0
    const uint8_t *data = &src_ptr[j];
96
97
0
    src10 = _mm_loadl_epi64((__m128i *)(data + 10 * src_stride));
98
0
    s[0] =
99
0
        _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(data + 0 * src_stride)),
100
0
                          _mm_loadl_epi64((__m128i *)(data + 1 * src_stride)));
101
0
    s[1] =
102
0
        _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(data + 1 * src_stride)),
103
0
                          _mm_loadl_epi64((__m128i *)(data + 2 * src_stride)));
104
0
    s[2] =
105
0
        _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(data + 2 * src_stride)),
106
0
                          _mm_loadl_epi64((__m128i *)(data + 3 * src_stride)));
107
0
    s[3] =
108
0
        _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(data + 3 * src_stride)),
109
0
                          _mm_loadl_epi64((__m128i *)(data + 4 * src_stride)));
110
0
    s[4] =
111
0
        _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(data + 4 * src_stride)),
112
0
                          _mm_loadl_epi64((__m128i *)(data + 5 * src_stride)));
113
0
    s[5] =
114
0
        _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(data + 5 * src_stride)),
115
0
                          _mm_loadl_epi64((__m128i *)(data + 6 * src_stride)));
116
0
    s[6] =
117
0
        _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(data + 6 * src_stride)),
118
0
                          _mm_loadl_epi64((__m128i *)(data + 7 * src_stride)));
119
0
    s[7] =
120
0
        _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(data + 7 * src_stride)),
121
0
                          _mm_loadl_epi64((__m128i *)(data + 8 * src_stride)));
122
0
    s[8] =
123
0
        _mm_unpacklo_epi8(_mm_loadl_epi64((__m128i *)(data + 8 * src_stride)),
124
0
                          _mm_loadl_epi64((__m128i *)(data + 9 * src_stride)));
125
0
    s[9] = _mm_unpacklo_epi8(
126
0
        _mm_loadl_epi64((__m128i *)(data + 9 * src_stride)), src10);
127
128
0
    int i = 0;
129
0
    do {
130
0
      data = &src_ptr[i * src_stride + j];
131
0
      s[10] = _mm_unpacklo_epi8(
132
0
          src10, _mm_loadl_epi64((__m128i *)(data + 11 * src_stride)));
133
0
      src10 = _mm_loadl_epi64((__m128i *)(data + 12 * src_stride));
134
0
      s[11] = _mm_unpacklo_epi8(
135
0
          _mm_loadl_epi64((__m128i *)(data + 11 * src_stride)), src10);
136
137
0
      res_lo = convolve_lo_y_12tap(s, coeffs);  // Filter low index pixels
138
0
      res_hi = convolve_hi_y_12tap(s, coeffs);  // Filter high index pixels
139
140
0
      res_lo_round =
141
0
          _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
142
0
      res_hi_round =
143
0
          _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift);
144
145
0
      res16 = _mm_packs_epi32(res_lo_round, res_hi_round);
146
0
      res = _mm_packus_epi16(res16, res16);
147
148
0
      _mm_storel_epi64((__m128i *)(dst + i * dst_stride + j), res);
149
0
      i++;
150
151
0
      res_lo = convolve_lo_y_12tap(s + 1, coeffs);  // Filter low index pixels
152
0
      res_hi = convolve_hi_y_12tap(s + 1, coeffs);  // Filter high index pixels
153
154
0
      res_lo_round =
155
0
          _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
156
0
      res_hi_round =
157
0
          _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift);
158
159
0
      res16 = _mm_packs_epi32(res_lo_round, res_hi_round);
160
0
      res = _mm_packus_epi16(res16, res16);
161
162
0
      _mm_storel_epi64((__m128i *)(dst + i * dst_stride + j), res);
163
0
      i++;
164
165
0
      s[0] = s[2];
166
0
      s[1] = s[3];
167
0
      s[2] = s[4];
168
0
      s[3] = s[5];
169
0
      s[4] = s[6];
170
0
      s[5] = s[7];
171
0
      s[6] = s[8];
172
0
      s[7] = s[9];
173
0
      s[8] = s[10];
174
0
      s[9] = s[11];
175
0
    } while (i < h);
176
0
    j += 8;
177
0
  } while (j < w);
178
0
}
179
180
void av1_convolve_y_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst,
181
                            int dst_stride, int w, int h,
182
                            const InterpFilterParams *filter_params_y,
183
0
                            const int subpel_y_qn) {
184
0
  if (filter_params_y->taps > 8) {
185
0
    if (w < 8) {
186
0
      av1_convolve_y_sr_c(src, src_stride, dst, dst_stride, w, h,
187
0
                          filter_params_y, subpel_y_qn);
188
0
    } else {
189
0
      convolve_y_sr_12tap_sse2(src, src_stride, dst, dst_stride, w, h,
190
0
                               filter_params_y, subpel_y_qn);
191
0
    }
192
0
  } else {
193
0
    const int fo_vert = filter_params_y->taps / 2 - 1;
194
0
    const uint8_t *src_ptr = src - fo_vert * src_stride;
195
0
    const __m128i round_const = _mm_set1_epi32((1 << FILTER_BITS) >> 1);
196
0
    const __m128i round_shift = _mm_cvtsi32_si128(FILTER_BITS);
197
0
    __m128i coeffs[4];
198
199
0
    prepare_coeffs(filter_params_y, subpel_y_qn, coeffs);
200
201
0
    if (w <= 4) {
202
0
      __m128i s[8], src6, res, res_round, res16;
203
0
      int res_int;
204
0
      s[0] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 0 * src_stride),
205
0
                               xx_loadl_32(src_ptr + 1 * src_stride));
206
0
      s[1] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 1 * src_stride),
207
0
                               xx_loadl_32(src_ptr + 2 * src_stride));
208
0
      s[2] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 2 * src_stride),
209
0
                               xx_loadl_32(src_ptr + 3 * src_stride));
210
0
      s[3] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 3 * src_stride),
211
0
                               xx_loadl_32(src_ptr + 4 * src_stride));
212
0
      s[4] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 4 * src_stride),
213
0
                               xx_loadl_32(src_ptr + 5 * src_stride));
214
0
      src6 = xx_loadl_32(src_ptr + 6 * src_stride);
215
0
      s[5] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 5 * src_stride), src6);
216
217
0
      do {
218
0
        s[6] = _mm_unpacklo_epi8(src6, xx_loadl_32(src_ptr + 7 * src_stride));
219
0
        src6 = xx_loadl_32(src_ptr + 8 * src_stride);
220
0
        s[7] = _mm_unpacklo_epi8(xx_loadl_32(src_ptr + 7 * src_stride), src6);
221
222
0
        res = convolve_lo_y(s + 0, coeffs);
223
0
        res_round = _mm_sra_epi32(_mm_add_epi32(res, round_const), round_shift);
224
0
        res16 = _mm_packs_epi32(res_round, res_round);
225
0
        res_int = _mm_cvtsi128_si32(_mm_packus_epi16(res16, res16));
226
227
0
        if (w == 2)
228
0
          *(uint16_t *)dst = (uint16_t)res_int;
229
0
        else
230
0
          *(int *)dst = res_int;
231
232
0
        src_ptr += src_stride;
233
0
        dst += dst_stride;
234
235
0
        res = convolve_lo_y(s + 1, coeffs);
236
0
        res_round = _mm_sra_epi32(_mm_add_epi32(res, round_const), round_shift);
237
0
        res16 = _mm_packs_epi32(res_round, res_round);
238
0
        res_int = _mm_cvtsi128_si32(_mm_packus_epi16(res16, res16));
239
240
0
        if (w == 2)
241
0
          *(uint16_t *)dst = (uint16_t)res_int;
242
0
        else
243
0
          *(int *)dst = res_int;
244
245
0
        src_ptr += src_stride;
246
0
        dst += dst_stride;
247
248
0
        s[0] = s[2];
249
0
        s[1] = s[3];
250
0
        s[2] = s[4];
251
0
        s[3] = s[5];
252
0
        s[4] = s[6];
253
0
        s[5] = s[7];
254
0
        h -= 2;
255
0
      } while (h);
256
0
    } else {
257
0
      assert(!(w % 8));
258
0
      int j = 0;
259
0
      do {
260
0
        __m128i s[8], src6, res_lo, res_hi;
261
0
        __m128i res_lo_round, res_hi_round, res16, res;
262
0
        const uint8_t *data = &src_ptr[j];
263
264
0
        src6 = _mm_loadl_epi64((__m128i *)(data + 6 * src_stride));
265
0
        s[0] = _mm_unpacklo_epi8(
266
0
            _mm_loadl_epi64((__m128i *)(data + 0 * src_stride)),
267
0
            _mm_loadl_epi64((__m128i *)(data + 1 * src_stride)));
268
0
        s[1] = _mm_unpacklo_epi8(
269
0
            _mm_loadl_epi64((__m128i *)(data + 1 * src_stride)),
270
0
            _mm_loadl_epi64((__m128i *)(data + 2 * src_stride)));
271
0
        s[2] = _mm_unpacklo_epi8(
272
0
            _mm_loadl_epi64((__m128i *)(data + 2 * src_stride)),
273
0
            _mm_loadl_epi64((__m128i *)(data + 3 * src_stride)));
274
0
        s[3] = _mm_unpacklo_epi8(
275
0
            _mm_loadl_epi64((__m128i *)(data + 3 * src_stride)),
276
0
            _mm_loadl_epi64((__m128i *)(data + 4 * src_stride)));
277
0
        s[4] = _mm_unpacklo_epi8(
278
0
            _mm_loadl_epi64((__m128i *)(data + 4 * src_stride)),
279
0
            _mm_loadl_epi64((__m128i *)(data + 5 * src_stride)));
280
0
        s[5] = _mm_unpacklo_epi8(
281
0
            _mm_loadl_epi64((__m128i *)(data + 5 * src_stride)), src6);
282
283
0
        int i = 0;
284
0
        do {
285
0
          data = &src_ptr[i * src_stride + j];
286
0
          s[6] = _mm_unpacklo_epi8(
287
0
              src6, _mm_loadl_epi64((__m128i *)(data + 7 * src_stride)));
288
0
          src6 = _mm_loadl_epi64((__m128i *)(data + 8 * src_stride));
289
0
          s[7] = _mm_unpacklo_epi8(
290
0
              _mm_loadl_epi64((__m128i *)(data + 7 * src_stride)), src6);
291
292
0
          res_lo = convolve_lo_y(s, coeffs);  // Filter low index pixels
293
0
          res_hi = convolve_hi_y(s, coeffs);  // Filter high index pixels
294
295
0
          res_lo_round =
296
0
              _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
297
0
          res_hi_round =
298
0
              _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift);
299
300
0
          res16 = _mm_packs_epi32(res_lo_round, res_hi_round);
301
0
          res = _mm_packus_epi16(res16, res16);
302
303
0
          _mm_storel_epi64((__m128i *)(dst + i * dst_stride + j), res);
304
0
          i++;
305
306
0
          res_lo = convolve_lo_y(s + 1, coeffs);  // Filter low index pixels
307
0
          res_hi = convolve_hi_y(s + 1, coeffs);  // Filter high index pixels
308
309
0
          res_lo_round =
310
0
              _mm_sra_epi32(_mm_add_epi32(res_lo, round_const), round_shift);
311
0
          res_hi_round =
312
0
              _mm_sra_epi32(_mm_add_epi32(res_hi, round_const), round_shift);
313
314
0
          res16 = _mm_packs_epi32(res_lo_round, res_hi_round);
315
0
          res = _mm_packus_epi16(res16, res16);
316
317
0
          _mm_storel_epi64((__m128i *)(dst + i * dst_stride + j), res);
318
0
          i++;
319
320
0
          s[0] = s[2];
321
0
          s[1] = s[3];
322
0
          s[2] = s[4];
323
0
          s[3] = s[5];
324
0
          s[4] = s[6];
325
0
          s[5] = s[7];
326
0
        } while (i < h);
327
0
        j += 8;
328
0
      } while (j < w);
329
0
    }
330
0
  }
331
0
}
332
333
static void convolve_x_sr_12tap_sse2(const uint8_t *src, int src_stride,
334
                                     uint8_t *dst, int dst_stride, int w, int h,
335
                                     const InterpFilterParams *filter_params_x,
336
                                     int subpel_x_qn,
337
0
                                     ConvolveParams *conv_params) {
338
0
  const int fo_horiz = filter_params_x->taps / 2 - 1;
339
0
  const uint8_t *src_ptr = src - fo_horiz;
340
0
  const int bits = FILTER_BITS - conv_params->round_0;
341
0
  const __m128i round_0_const =
342
0
      _mm_set1_epi32((1 << conv_params->round_0) >> 1);
343
0
  const __m128i round_const = _mm_set1_epi32((1 << bits) >> 1);
344
0
  const __m128i round_0_shift = _mm_cvtsi32_si128(conv_params->round_0);
345
0
  const __m128i round_shift = _mm_cvtsi32_si128(bits);
346
0
  const __m128i zero = _mm_setzero_si128();
347
0
  __m128i coeffs[6];
348
349
0
  assert(bits >= 0);
350
0
  assert((FILTER_BITS - conv_params->round_1) >= 0 ||
351
0
         ((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS));
352
353
0
  prepare_coeffs_12tap(filter_params_x, subpel_x_qn, coeffs);
354
355
0
  int i = 0;
356
0
  do {
357
0
    int j = 0;
358
0
    do {
359
0
      const __m128i data =
360
0
          _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
361
0
      __m128i s[4];
362
363
0
      s[0] = _mm_unpacklo_epi16(data, _mm_srli_si128(data, 1));
364
0
      s[1] =
365
0
          _mm_unpacklo_epi16(_mm_srli_si128(data, 2), _mm_srli_si128(data, 3));
366
0
      s[2] =
367
0
          _mm_unpacklo_epi16(_mm_srli_si128(data, 4), _mm_srli_si128(data, 5));
368
0
      s[3] =
369
0
          _mm_unpacklo_epi16(_mm_srli_si128(data, 6), _mm_srli_si128(data, 7));
370
371
0
      const __m128i res32 = convolve_lo_x_12tap(s, coeffs, zero);
372
373
0
      __m128i res32_round =
374
0
          _mm_sra_epi32(_mm_add_epi32(res32, round_0_const), round_0_shift);
375
0
      res32_round =
376
0
          _mm_sra_epi32(_mm_add_epi32(res32_round, round_const), round_shift);
377
378
0
      const __m128i res16 = _mm_packs_epi32(res32_round, zero);
379
0
      const __m128i res = _mm_packus_epi16(res16, zero);
380
381
0
      const int val = _mm_cvtsi128_si32(res);
382
0
      memcpy((dst + i * dst_stride + j), &val, sizeof(val));
383
0
      j += 4;
384
0
    } while (j < w);
385
0
  } while (++i < h);
386
0
}
387
388
void av1_convolve_x_sr_sse2(const uint8_t *src, int src_stride, uint8_t *dst,
389
                            int dst_stride, int w, int h,
390
                            const InterpFilterParams *filter_params_x,
391
                            const int subpel_x_qn,
392
0
                            ConvolveParams *conv_params) {
393
0
  if (filter_params_x->taps > 8) {
394
0
    if (w < 4) {
395
0
      av1_convolve_x_sr_c(src, src_stride, dst, dst_stride, w, h,
396
0
                          filter_params_x, subpel_x_qn, conv_params);
397
0
    } else {
398
0
      convolve_x_sr_12tap_sse2(src, src_stride, dst, dst_stride, w, h,
399
0
                               filter_params_x, subpel_x_qn, conv_params);
400
0
    }
401
0
  } else {
402
0
    const int fo_horiz = filter_params_x->taps / 2 - 1;
403
0
    const uint8_t *src_ptr = src - fo_horiz;
404
0
    const int bits = FILTER_BITS - conv_params->round_0;
405
0
    const __m128i round_0_const =
406
0
        _mm_set1_epi32((1 << conv_params->round_0) >> 1);
407
0
    const __m128i round_const = _mm_set1_epi32((1 << bits) >> 1);
408
0
    const __m128i round_0_shift = _mm_cvtsi32_si128(conv_params->round_0);
409
0
    const __m128i round_shift = _mm_cvtsi32_si128(bits);
410
0
    __m128i coeffs[4];
411
412
0
    assert(bits >= 0);
413
0
    assert((FILTER_BITS - conv_params->round_1) >= 0 ||
414
0
           ((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS));
415
416
0
    prepare_coeffs(filter_params_x, subpel_x_qn, coeffs);
417
418
0
    if (w <= 4) {
419
0
      do {
420
0
        const __m128i data = _mm_loadu_si128((__m128i *)src_ptr);
421
0
        __m128i s[4];
422
423
0
        s[0] = _mm_unpacklo_epi8(data, _mm_srli_si128(data, 1));
424
0
        s[1] =
425
0
            _mm_unpacklo_epi8(_mm_srli_si128(data, 2), _mm_srli_si128(data, 3));
426
0
        s[2] =
427
0
            _mm_unpacklo_epi8(_mm_srli_si128(data, 4), _mm_srli_si128(data, 5));
428
0
        s[3] =
429
0
            _mm_unpacklo_epi8(_mm_srli_si128(data, 6), _mm_srli_si128(data, 7));
430
0
        const __m128i res_lo = convolve_lo_x(s, coeffs);
431
0
        __m128i res_lo_round =
432
0
            _mm_sra_epi32(_mm_add_epi32(res_lo, round_0_const), round_0_shift);
433
0
        res_lo_round = _mm_sra_epi32(_mm_add_epi32(res_lo_round, round_const),
434
0
                                     round_shift);
435
436
0
        const __m128i res16 = _mm_packs_epi32(res_lo_round, res_lo_round);
437
0
        const __m128i res = _mm_packus_epi16(res16, res16);
438
439
0
        int r = _mm_cvtsi128_si32(res);
440
0
        if (w == 2)
441
0
          *(uint16_t *)dst = (uint16_t)r;
442
0
        else
443
0
          *(int *)dst = r;
444
445
0
        src_ptr += src_stride;
446
0
        dst += dst_stride;
447
0
      } while (--h);
448
0
    } else {
449
0
      assert(!(w % 8));
450
0
      int i = 0;
451
0
      do {
452
0
        int j = 0;
453
0
        do {
454
0
          const __m128i data =
455
0
              _mm_loadu_si128((__m128i *)&src_ptr[i * src_stride + j]);
456
0
          __m128i s[4];
457
458
          // Filter even-index pixels
459
0
          s[0] = data;
460
0
          s[1] = _mm_srli_si128(data, 2);
461
0
          s[2] = _mm_srli_si128(data, 4);
462
0
          s[3] = _mm_srli_si128(data, 6);
463
0
          const __m128i res_even = convolve_lo_x(s, coeffs);
464
465
          // Filter odd-index pixels
466
0
          s[0] = _mm_srli_si128(data, 1);
467
0
          s[1] = _mm_srli_si128(data, 3);
468
0
          s[2] = _mm_srli_si128(data, 5);
469
0
          s[3] = _mm_srli_si128(data, 7);
470
0
          const __m128i res_odd = convolve_lo_x(s, coeffs);
471
472
          // Rearrange pixels back into the order 0 ... 7
473
0
          const __m128i res_lo = _mm_unpacklo_epi32(res_even, res_odd);
474
0
          const __m128i res_hi = _mm_unpackhi_epi32(res_even, res_odd);
475
0
          __m128i res_lo_round = _mm_sra_epi32(
476
0
              _mm_add_epi32(res_lo, round_0_const), round_0_shift);
477
0
          res_lo_round = _mm_sra_epi32(_mm_add_epi32(res_lo_round, round_const),
478
0
                                       round_shift);
479
0
          __m128i res_hi_round = _mm_sra_epi32(
480
0
              _mm_add_epi32(res_hi, round_0_const), round_0_shift);
481
0
          res_hi_round = _mm_sra_epi32(_mm_add_epi32(res_hi_round, round_const),
482
0
                                       round_shift);
483
484
0
          const __m128i res16 = _mm_packs_epi32(res_lo_round, res_hi_round);
485
0
          const __m128i res = _mm_packus_epi16(res16, res16);
486
487
0
          _mm_storel_epi64((__m128i *)(dst + i * dst_stride + j), res);
488
0
          j += 8;
489
0
        } while (j < w);
490
0
      } while (++i < h);
491
0
    }
492
0
  }
493
0
}