summaryrefslogtreecommitdiffstats
path: root/third_party/aom/av1/common/x86/highbd_wiener_convolve_avx2.c
blob: 0c8a8505b09fc02cc66c5977a5ba88c72f7da690 (plain)
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
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
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
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
161
162
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
191
192
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
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
/*
 * Copyright (c) 2018, Alliance for Open Media. All rights reserved
 *
 * This source code is subject to the terms of the BSD 2 Clause License and
 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
 * was not distributed with this source code in the LICENSE file, you can
 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
 * Media Patent License 1.0 was not distributed with this source code in the
 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
 */

#include <immintrin.h>
#include <assert.h>

#include "config/aom_dsp_rtcd.h"

#include "av1/common/convolve.h"
#include "aom_dsp/aom_dsp_common.h"
#include "aom_dsp/aom_filter.h"
#include "aom_dsp/x86/synonyms.h"
#include "aom_dsp/x86/synonyms_avx2.h"

// 128-bit xmmwords are written as [ ... ] with the MSB on the left.
// 256-bit ymmwords are written as two xmmwords, [ ... ][ ... ] with the MSB
// on the left.
// A row of, say, 16-bit pixels with values p0, p1, p2, ..., p14, p15 will be
// loaded and stored as [ p15 ... p9 p8 ][ p7 ... p1 p0 ].
void av1_highbd_wiener_convolve_add_src_avx2(
    const uint8_t *src8, ptrdiff_t src_stride, uint8_t *dst8,
    ptrdiff_t dst_stride, const int16_t *filter_x, int x_step_q4,
    const int16_t *filter_y, int y_step_q4, int w, int h,
    const ConvolveParams *conv_params, int bd) {
  assert(x_step_q4 == 16 && y_step_q4 == 16);
  assert(!(w & 7));
  assert(bd + FILTER_BITS - conv_params->round_0 + 2 <= 16);
  (void)x_step_q4;
  (void)y_step_q4;

  const uint16_t *const src = CONVERT_TO_SHORTPTR(src8);
  uint16_t *const dst = CONVERT_TO_SHORTPTR(dst8);

  DECLARE_ALIGNED(32, uint16_t,
                  temp[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]);
  int intermediate_height = h + SUBPEL_TAPS - 1;
  const int center_tap = ((SUBPEL_TAPS - 1) / 2);
  const uint16_t *const src_ptr = src - center_tap * src_stride - center_tap;

  const __m128i zero_128 = _mm_setzero_si128();
  const __m256i zero_256 = _mm256_setzero_si256();

  // Add an offset to account for the "add_src" part of the convolve function.
  const __m128i offset = _mm_insert_epi16(zero_128, 1 << FILTER_BITS, 3);

  const __m256i clamp_low = zero_256;

  /* Horizontal filter */
  {
    const __m256i clamp_high_ep =
        _mm256_set1_epi16(WIENER_CLAMP_LIMIT(conv_params->round_0, bd) - 1);

    // coeffs [ f7 f6 f5 f4 f3 f2 f1 f0 ]
    const __m128i coeffs_x = _mm_add_epi16(xx_loadu_128(filter_x), offset);

    // coeffs [ f3 f2 f3 f2 f1 f0 f1 f0 ]
    const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_x, coeffs_x);
    // coeffs [ f7 f6 f7 f6 f5 f4 f5 f4 ]
    const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_x, coeffs_x);

    // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ]
    const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123);
    // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ]
    const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123);
    // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ]
    const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567);
    // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ]
    const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567);

    // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ][ f1 f0 f1 f0 f1 f0 f1 f0 ]
    const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128);
    // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ][ f3 f2 f3 f2 f3 f2 f3 f2 ]
    const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128);
    // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ][ f5 f4 f5 f4 f5 f4 f5 f4 ]
    const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128);
    // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ][ f7 f6 f7 f6 f7 f6 f7 f6 ]
    const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128);

    const __m256i round_const = _mm256_set1_epi32(
        (1 << (conv_params->round_0 - 1)) + (1 << (bd + FILTER_BITS - 1)));

    for (int i = 0; i < intermediate_height; ++i) {
      for (int j = 0; j < w; j += 16) {
        const uint16_t *src_ij = src_ptr + i * src_stride + j;

        // Load 16-bit src data
        const __m256i src_0 = yy_loadu_256(src_ij + 0);
        const __m256i src_1 = yy_loadu_256(src_ij + 1);
        const __m256i src_2 = yy_loadu_256(src_ij + 2);
        const __m256i src_3 = yy_loadu_256(src_ij + 3);
        const __m256i src_4 = yy_loadu_256(src_ij + 4);
        const __m256i src_5 = yy_loadu_256(src_ij + 5);
        const __m256i src_6 = yy_loadu_256(src_ij + 6);
        const __m256i src_7 = yy_loadu_256(src_ij + 7);

        // Multiply src data by filter coeffs and sum pairs
        const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01);
        const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01);
        const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23);
        const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23);
        const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45);
        const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45);
        const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67);
        const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67);

        // Calculate scalar product for even- and odd-indices separately,
        // increasing to 32-bit precision
        const __m256i res_even_sum = _mm256_add_epi32(
            _mm256_add_epi32(res_0, res_4), _mm256_add_epi32(res_2, res_6));
        const __m256i res_even = _mm256_srai_epi32(
            _mm256_add_epi32(res_even_sum, round_const), conv_params->round_0);

        const __m256i res_odd_sum = _mm256_add_epi32(
            _mm256_add_epi32(res_1, res_5), _mm256_add_epi32(res_3, res_7));
        const __m256i res_odd = _mm256_srai_epi32(
            _mm256_add_epi32(res_odd_sum, round_const), conv_params->round_0);

        // Reduce to 16-bit precision and pack even- and odd-index results
        // back into one register. The _mm256_packs_epi32 intrinsic returns
        // a register with the pixels ordered as follows:
        // [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ]
        const __m256i res = _mm256_packs_epi32(res_even, res_odd);
        const __m256i res_clamped =
            _mm256_min_epi16(_mm256_max_epi16(res, clamp_low), clamp_high_ep);

        // Store in a temporary array
        yy_storeu_256(temp + i * MAX_SB_SIZE + j, res_clamped);
      }
    }
  }

  /* Vertical filter */
  {
    const __m256i clamp_high = _mm256_set1_epi16((1 << bd) - 1);

    // coeffs [ f7 f6 f5 f4 f3 f2 f1 f0 ]
    const __m128i coeffs_y = _mm_add_epi16(xx_loadu_128(filter_y), offset);

    // coeffs [ f3 f2 f3 f2 f1 f0 f1 f0 ]
    const __m128i coeffs_0123 = _mm_unpacklo_epi32(coeffs_y, coeffs_y);
    // coeffs [ f7 f6 f7 f6 f5 f4 f5 f4 ]
    const __m128i coeffs_4567 = _mm_unpackhi_epi32(coeffs_y, coeffs_y);

    // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ]
    const __m128i coeffs_01_128 = _mm_unpacklo_epi64(coeffs_0123, coeffs_0123);
    // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ]
    const __m128i coeffs_23_128 = _mm_unpackhi_epi64(coeffs_0123, coeffs_0123);
    // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ]
    const __m128i coeffs_45_128 = _mm_unpacklo_epi64(coeffs_4567, coeffs_4567);
    // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ]
    const __m128i coeffs_67_128 = _mm_unpackhi_epi64(coeffs_4567, coeffs_4567);

    // coeffs [ f1 f0 f1 f0 f1 f0 f1 f0 ][ f1 f0 f1 f0 f1 f0 f1 f0 ]
    const __m256i coeffs_01 = yy_set_m128i(coeffs_01_128, coeffs_01_128);
    // coeffs [ f3 f2 f3 f2 f3 f2 f3 f2 ][ f3 f2 f3 f2 f3 f2 f3 f2 ]
    const __m256i coeffs_23 = yy_set_m128i(coeffs_23_128, coeffs_23_128);
    // coeffs [ f5 f4 f5 f4 f5 f4 f5 f4 ][ f5 f4 f5 f4 f5 f4 f5 f4 ]
    const __m256i coeffs_45 = yy_set_m128i(coeffs_45_128, coeffs_45_128);
    // coeffs [ f7 f6 f7 f6 f7 f6 f7 f6 ][ f7 f6 f7 f6 f7 f6 f7 f6 ]
    const __m256i coeffs_67 = yy_set_m128i(coeffs_67_128, coeffs_67_128);

    const __m256i round_const =
        _mm256_set1_epi32((1 << (conv_params->round_1 - 1)) -
                          (1 << (bd + conv_params->round_1 - 1)));

    for (int i = 0; i < h; ++i) {
      for (int j = 0; j < w; j += 16) {
        const uint16_t *temp_ij = temp + i * MAX_SB_SIZE + j;

        // Load 16-bit data from the output of the horizontal filter in
        // which the pixels are ordered as follows:
        // [ 15 13 11 9 14 12 10 8 ] [ 7 5 3 1 6 4 2 0 ]
        const __m256i data_0 = yy_loadu_256(temp_ij + 0 * MAX_SB_SIZE);
        const __m256i data_1 = yy_loadu_256(temp_ij + 1 * MAX_SB_SIZE);
        const __m256i data_2 = yy_loadu_256(temp_ij + 2 * MAX_SB_SIZE);
        const __m256i data_3 = yy_loadu_256(temp_ij + 3 * MAX_SB_SIZE);
        const __m256i data_4 = yy_loadu_256(temp_ij + 4 * MAX_SB_SIZE);
        const __m256i data_5 = yy_loadu_256(temp_ij + 5 * MAX_SB_SIZE);
        const __m256i data_6 = yy_loadu_256(temp_ij + 6 * MAX_SB_SIZE);
        const __m256i data_7 = yy_loadu_256(temp_ij + 7 * MAX_SB_SIZE);

        // Filter the even-indices, increasing to 32-bit precision
        const __m256i src_0 = _mm256_unpacklo_epi16(data_0, data_1);
        const __m256i src_2 = _mm256_unpacklo_epi16(data_2, data_3);
        const __m256i src_4 = _mm256_unpacklo_epi16(data_4, data_5);
        const __m256i src_6 = _mm256_unpacklo_epi16(data_6, data_7);

        const __m256i res_0 = _mm256_madd_epi16(src_0, coeffs_01);
        const __m256i res_2 = _mm256_madd_epi16(src_2, coeffs_23);
        const __m256i res_4 = _mm256_madd_epi16(src_4, coeffs_45);
        const __m256i res_6 = _mm256_madd_epi16(src_6, coeffs_67);

        const __m256i res_even = _mm256_add_epi32(
            _mm256_add_epi32(res_0, res_2), _mm256_add_epi32(res_4, res_6));

        // Filter the odd-indices, increasing to 32-bit precision
        const __m256i src_1 = _mm256_unpackhi_epi16(data_0, data_1);
        const __m256i src_3 = _mm256_unpackhi_epi16(data_2, data_3);
        const __m256i src_5 = _mm256_unpackhi_epi16(data_4, data_5);
        const __m256i src_7 = _mm256_unpackhi_epi16(data_6, data_7);

        const __m256i res_1 = _mm256_madd_epi16(src_1, coeffs_01);
        const __m256i res_3 = _mm256_madd_epi16(src_3, coeffs_23);
        const __m256i res_5 = _mm256_madd_epi16(src_5, coeffs_45);
        const __m256i res_7 = _mm256_madd_epi16(src_7, coeffs_67);

        const __m256i res_odd = _mm256_add_epi32(
            _mm256_add_epi32(res_1, res_3), _mm256_add_epi32(res_5, res_7));

        // Pixels are currently in the following order:
        // res_even order: [ 14 12 10 8 ] [ 6 4 2 0 ]
        // res_odd order:  [ 15 13 11 9 ] [ 7 5 3 1 ]
        //
        // Rearrange the pixels into the following order:
        // res_lo order: [ 11 10  9  8 ] [ 3 2 1 0 ]
        // res_hi order: [ 15 14 13 12 ] [ 7 6 5 4 ]
        const __m256i res_lo = _mm256_unpacklo_epi32(res_even, res_odd);
        const __m256i res_hi = _mm256_unpackhi_epi32(res_even, res_odd);

        const __m256i res_lo_round = _mm256_srai_epi32(
            _mm256_add_epi32(res_lo, round_const), conv_params->round_1);
        const __m256i res_hi_round = _mm256_srai_epi32(
            _mm256_add_epi32(res_hi, round_const), conv_params->round_1);

        // Reduce to 16-bit precision and pack into the correct order:
        // [ 15 14 13 12 11 10 9 8 ][ 7 6 5 4 3 2 1 0 ]
        const __m256i res_16bit =
            _mm256_packs_epi32(res_lo_round, res_hi_round);
        const __m256i res_16bit_clamped = _mm256_min_epi16(
            _mm256_max_epi16(res_16bit, clamp_low), clamp_high);

        // Store in the dst array
        yy_storeu_256(dst + i * dst_stride + j, res_16bit_clamped);
      }
    }
  }
}