Line data Source code
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 : #include <immintrin.h>
12 :
13 : #include "EbDefinitions.h"
14 : #include "aom_dsp_rtcd.h"
15 :
16 : #include "convolve.h"
17 : #include "convolve_avx2.h"
18 : #include "synonyms.h"
19 :
20 : // -----------------------------------------------------------------------------
21 : // Copy and average
22 :
23 0 : void eb_av1_highbd_convolve_y_sr_avx2(const uint16_t *src, int32_t src_stride,
24 : uint16_t *dst, int32_t dst_stride, int32_t w, int32_t h,
25 : const InterpFilterParams *filter_params_x,
26 : const InterpFilterParams *filter_params_y,
27 : const int32_t subpel_x_q4, const int32_t subpel_y_q4,
28 : ConvolveParams *conv_params, int32_t bd) {
29 : int32_t i, j;
30 0 : const int32_t fo_vert = filter_params_y->taps / 2 - 1;
31 0 : const uint16_t *const src_ptr = src - fo_vert * src_stride;
32 : (void)filter_params_x;
33 : (void)subpel_x_q4;
34 : (void)conv_params;
35 :
36 0 : assert(conv_params->round_0 <= FILTER_BITS);
37 0 : assert(((conv_params->round_0 + conv_params->round_1) <= (FILTER_BITS + 1)) ||
38 : ((conv_params->round_0 + conv_params->round_1) == (2 * FILTER_BITS)));
39 :
40 : __m256i s[8], coeffs_y[4];
41 :
42 0 : const int32_t bits = FILTER_BITS;
43 :
44 0 : const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
45 0 : const __m256i round_const_bits = _mm256_set1_epi32((1 << bits) >> 1);
46 : const __m256i clip_pixel =
47 0 : _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
48 0 : const __m256i zero = _mm256_setzero_si256();
49 :
50 0 : prepare_coeffs_8tap_avx2(filter_params_y, subpel_y_q4, coeffs_y);
51 :
52 0 : for (j = 0; j < w; j += 8) {
53 0 : const uint16_t *data = &src_ptr[j];
54 : /* Vertical filter */
55 : {
56 : __m256i src6;
57 0 : __m256i s01 = _mm256_permute2x128_si256(
58 : _mm256_castsi128_si256(
59 : _mm_loadu_si128((__m128i *)(data + 0 * src_stride))),
60 : _mm256_castsi128_si256(
61 : _mm_loadu_si128((__m128i *)(data + 1 * src_stride))),
62 : 0x20);
63 0 : __m256i s12 = _mm256_permute2x128_si256(
64 : _mm256_castsi128_si256(
65 : _mm_loadu_si128((__m128i *)(data + 1 * src_stride))),
66 : _mm256_castsi128_si256(
67 : _mm_loadu_si128((__m128i *)(data + 2 * src_stride))),
68 : 0x20);
69 0 : __m256i s23 = _mm256_permute2x128_si256(
70 : _mm256_castsi128_si256(
71 : _mm_loadu_si128((__m128i *)(data + 2 * src_stride))),
72 : _mm256_castsi128_si256(
73 : _mm_loadu_si128((__m128i *)(data + 3 * src_stride))),
74 : 0x20);
75 0 : __m256i s34 = _mm256_permute2x128_si256(
76 : _mm256_castsi128_si256(
77 : _mm_loadu_si128((__m128i *)(data + 3 * src_stride))),
78 : _mm256_castsi128_si256(
79 : _mm_loadu_si128((__m128i *)(data + 4 * src_stride))),
80 : 0x20);
81 0 : __m256i s45 = _mm256_permute2x128_si256(
82 : _mm256_castsi128_si256(
83 : _mm_loadu_si128((__m128i *)(data + 4 * src_stride))),
84 : _mm256_castsi128_si256(
85 : _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
86 : 0x20);
87 0 : src6 = _mm256_castsi128_si256(
88 0 : _mm_loadu_si128((__m128i *)(data + 6 * src_stride)));
89 0 : __m256i s56 = _mm256_permute2x128_si256(
90 : _mm256_castsi128_si256(
91 : _mm_loadu_si128((__m128i *)(data + 5 * src_stride))),
92 : src6, 0x20);
93 :
94 0 : s[0] = _mm256_unpacklo_epi16(s01, s12);
95 0 : s[1] = _mm256_unpacklo_epi16(s23, s34);
96 0 : s[2] = _mm256_unpacklo_epi16(s45, s56);
97 :
98 0 : s[4] = _mm256_unpackhi_epi16(s01, s12);
99 0 : s[5] = _mm256_unpackhi_epi16(s23, s34);
100 0 : s[6] = _mm256_unpackhi_epi16(s45, s56);
101 :
102 0 : for (i = 0; i < h; i += 2) {
103 0 : data = &src_ptr[i * src_stride + j];
104 :
105 0 : const __m256i s67 = _mm256_permute2x128_si256(
106 : src6,
107 : _mm256_castsi128_si256(
108 : _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
109 : 0x20);
110 :
111 0 : src6 = _mm256_castsi128_si256(
112 0 : _mm_loadu_si128((__m128i *)(data + 8 * src_stride)));
113 :
114 0 : const __m256i s78 = _mm256_permute2x128_si256(
115 : _mm256_castsi128_si256(
116 : _mm_loadu_si128((__m128i *)(data + 7 * src_stride))),
117 : src6, 0x20);
118 :
119 0 : s[3] = _mm256_unpacklo_epi16(s67, s78);
120 0 : s[7] = _mm256_unpackhi_epi16(s67, s78);
121 :
122 0 : const __m256i res_a = convolve16_8tap_avx2(s, coeffs_y);
123 :
124 0 : __m256i res_a_round = _mm256_sra_epi32(
125 : _mm256_add_epi32(res_a, round_const_bits), round_shift_bits);
126 :
127 0 : if (w - j > 4) {
128 0 : const __m256i res_b = convolve16_8tap_avx2(s + 4, coeffs_y);
129 0 : __m256i res_b_round = _mm256_sra_epi32(
130 : _mm256_add_epi32(res_b, round_const_bits), round_shift_bits);
131 :
132 0 : __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);
133 0 : res_16bit = _mm256_min_epi16(res_16bit, clip_pixel);
134 0 : res_16bit = _mm256_max_epi16(res_16bit, zero);
135 :
136 0 : _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j],
137 : _mm256_castsi256_si128(res_16bit));
138 0 : _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
139 0 : _mm256_extracti128_si256(res_16bit, 1));
140 : }
141 0 : else if (w == 4) {
142 0 : res_a_round = _mm256_packs_epi32(res_a_round, res_a_round);
143 0 : res_a_round = _mm256_min_epi16(res_a_round, clip_pixel);
144 0 : res_a_round = _mm256_max_epi16(res_a_round, zero);
145 :
146 0 : _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j],
147 : _mm256_castsi256_si128(res_a_round));
148 0 : _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
149 0 : _mm256_extracti128_si256(res_a_round, 1));
150 : }
151 : else {
152 0 : res_a_round = _mm256_packs_epi32(res_a_round, res_a_round);
153 0 : res_a_round = _mm256_min_epi16(res_a_round, clip_pixel);
154 0 : res_a_round = _mm256_max_epi16(res_a_round, zero);
155 :
156 0 : xx_storel_32((__m128i *)&dst[i * dst_stride + j],
157 : _mm256_castsi256_si128(res_a_round));
158 0 : xx_storel_32((__m128i *)&dst[i * dst_stride + j + dst_stride],
159 0 : _mm256_extracti128_si256(res_a_round, 1));
160 : }
161 :
162 0 : s[0] = s[1];
163 0 : s[1] = s[2];
164 0 : s[2] = s[3];
165 :
166 0 : s[4] = s[5];
167 0 : s[5] = s[6];
168 0 : s[6] = s[7];
169 : }
170 : }
171 : }
172 0 : }
173 :
174 0 : void eb_av1_highbd_convolve_x_sr_avx2(const uint16_t *src, int32_t src_stride,
175 : uint16_t *dst, int32_t dst_stride, int32_t w, int32_t h,
176 : const InterpFilterParams *filter_params_x,
177 : const InterpFilterParams *filter_params_y,
178 : const int32_t subpel_x_q4, const int32_t subpel_y_q4,
179 : ConvolveParams *conv_params, int32_t bd) {
180 : int32_t i, j;
181 0 : const int32_t fo_horiz = filter_params_x->taps / 2 - 1;
182 0 : const uint16_t *const src_ptr = src - fo_horiz;
183 : (void)subpel_y_q4;
184 : (void)filter_params_y;
185 :
186 : // Check that, even with 12-bit input, the intermediate values will fit
187 : // into an unsigned 16-bit intermediate array.
188 0 : assert(bd + FILTER_BITS + 2 - conv_params->round_0 <= 16);
189 :
190 : __m256i s[4], coeffs_x[4];
191 :
192 : const __m256i round_const_x =
193 0 : _mm256_set1_epi32(((1 << conv_params->round_0) >> 1));
194 0 : const __m128i round_shift_x = _mm_cvtsi32_si128(conv_params->round_0);
195 :
196 0 : const int32_t bits = FILTER_BITS - conv_params->round_0;
197 0 : const __m128i round_shift_bits = _mm_cvtsi32_si128(bits);
198 0 : const __m256i round_const_bits = _mm256_set1_epi32((1 << bits) >> 1);
199 : const __m256i clip_pixel =
200 0 : _mm256_set1_epi16(bd == 10 ? 1023 : (bd == 12 ? 4095 : 255));
201 0 : const __m256i zero = _mm256_setzero_si256();
202 :
203 0 : assert(bits >= 0);
204 0 : assert((FILTER_BITS - conv_params->round_1) >= 0 ||
205 : ((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS));
206 :
207 0 : prepare_coeffs_8tap_avx2(filter_params_x, subpel_x_q4, coeffs_x);
208 :
209 0 : for (j = 0; j < w; j += 8) {
210 : /* Horizontal filter */
211 0 : for (i = 0; i < h; i += 2) {
212 : const __m256i row0 =
213 0 : _mm256_loadu_si256((__m256i *)&src_ptr[i * src_stride + j]);
214 : __m256i row1 =
215 0 : _mm256_loadu_si256((__m256i *)&src_ptr[(i + 1) * src_stride + j]);
216 :
217 0 : const __m256i r0 = _mm256_permute2x128_si256(row0, row1, 0x20);
218 0 : const __m256i r1 = _mm256_permute2x128_si256(row0, row1, 0x31);
219 :
220 : // even pixels
221 0 : s[0] = _mm256_alignr_epi8(r1, r0, 0);
222 0 : s[1] = _mm256_alignr_epi8(r1, r0, 4);
223 0 : s[2] = _mm256_alignr_epi8(r1, r0, 8);
224 0 : s[3] = _mm256_alignr_epi8(r1, r0, 12);
225 :
226 0 : __m256i res_even = convolve16_8tap_avx2(s, coeffs_x);
227 0 : res_even = _mm256_sra_epi32(_mm256_add_epi32(res_even, round_const_x),
228 : round_shift_x);
229 :
230 : // odd pixels
231 0 : s[0] = _mm256_alignr_epi8(r1, r0, 2);
232 0 : s[1] = _mm256_alignr_epi8(r1, r0, 6);
233 0 : s[2] = _mm256_alignr_epi8(r1, r0, 10);
234 0 : s[3] = _mm256_alignr_epi8(r1, r0, 14);
235 :
236 0 : __m256i res_odd = convolve16_8tap_avx2(s, coeffs_x);
237 0 : res_odd = _mm256_sra_epi32(_mm256_add_epi32(res_odd, round_const_x),
238 : round_shift_x);
239 :
240 0 : res_even = _mm256_sra_epi32(_mm256_add_epi32(res_even, round_const_bits),
241 : round_shift_bits);
242 0 : res_odd = _mm256_sra_epi32(_mm256_add_epi32(res_odd, round_const_bits),
243 : round_shift_bits);
244 :
245 0 : __m256i res_even1 = _mm256_packs_epi32(res_even, res_even);
246 0 : __m256i res_odd1 = _mm256_packs_epi32(res_odd, res_odd);
247 :
248 0 : __m256i res = _mm256_unpacklo_epi16(res_even1, res_odd1);
249 0 : res = _mm256_min_epi16(res, clip_pixel);
250 0 : res = _mm256_max_epi16(res, zero);
251 :
252 0 : if (w - j > 4) {
253 0 : _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j],
254 : _mm256_castsi256_si128(res));
255 0 : _mm_storeu_si128((__m128i *)&dst[i * dst_stride + j + dst_stride],
256 0 : _mm256_extracti128_si256(res, 1));
257 : }
258 0 : else if (w == 4) {
259 0 : _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j],
260 : _mm256_castsi256_si128(res));
261 0 : _mm_storel_epi64((__m128i *)&dst[i * dst_stride + j + dst_stride],
262 0 : _mm256_extracti128_si256(res, 1));
263 : }
264 : else {
265 0 : xx_storel_32((__m128i *)&dst[i * dst_stride + j],
266 : _mm256_castsi256_si128(res));
267 0 : xx_storel_32((__m128i *)&dst[i * dst_stride + j + dst_stride],
268 0 : _mm256_extracti128_si256(res, 1));
269 : }
270 : }
271 : }
272 0 : }
273 :
274 : // -----------------------------------------------------------------------------
275 : // Horizontal Filtering
276 :
277 : //HIGH_FUN_CONV_1D(horiz, x_step_q4, filter_x, h, src, , avx2);
278 : //HIGH_FUN_CONV_1D(vert, y_step_q4, filter_y, v, src - src_stride * 3, , avx2);
279 :
280 : #undef HIGHBD_FUNC
|
