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 <immintrin.h>
13
14 #include "config/av1_rtcd.h"
15
16 #include "aom_dsp/x86/convolve_avx2.h"
17 #include "aom_dsp/x86/convolve_common_intrin.h"
18 #include "aom_dsp/aom_dsp_common.h"
19 #include "aom_dsp/aom_filter.h"
20 #include "aom_dsp/x86/synonyms.h"
21 #include "av1/common/convolve.h"
22
av1_convolve_2d_sr_avx2(const uint8_t * src,int src_stride,uint8_t * dst,int dst_stride,int w,int h,const InterpFilterParams * filter_params_x,const InterpFilterParams * filter_params_y,const int subpel_x_q4,const int subpel_y_q4,ConvolveParams * conv_params)23 void av1_convolve_2d_sr_avx2(const uint8_t *src, int src_stride, uint8_t *dst,
24 int dst_stride, int w, int h,
25 const InterpFilterParams *filter_params_x,
26 const InterpFilterParams *filter_params_y,
27 const int subpel_x_q4, const int subpel_y_q4,
28 ConvolveParams *conv_params) {
29 const int bd = 8;
30 int im_stride = 8;
31 int i, is_horiz_4tap = 0, is_vert_4tap = 0;
32 DECLARE_ALIGNED(32, int16_t, im_block[(MAX_SB_SIZE + MAX_FILTER_TAP) * 8]);
33 const int bits =
34 FILTER_BITS * 2 - conv_params->round_0 - conv_params->round_1;
35 const int offset_bits = bd + 2 * FILTER_BITS - conv_params->round_0;
36
37 assert(conv_params->round_0 > 0);
38
39 const __m256i round_const_h = _mm256_set1_epi16(
40 ((1 << (conv_params->round_0 - 1)) >> 1) + (1 << (bd + FILTER_BITS - 2)));
41 const __m128i round_shift_h = _mm_cvtsi32_si128(conv_params->round_0 - 1);
42
43 const __m256i sum_round_v = _mm256_set1_epi32(
44 (1 << offset_bits) + ((1 << conv_params->round_1) >> 1));
45 const __m128i sum_shift_v = _mm_cvtsi32_si128(conv_params->round_1);
46
47 const __m256i round_const_v = _mm256_set1_epi32(
48 ((1 << bits) >> 1) - (1 << (offset_bits - conv_params->round_1)) -
49 ((1 << (offset_bits - conv_params->round_1)) >> 1));
50 const __m128i round_shift_v = _mm_cvtsi32_si128(bits);
51
52 __m256i filt[4], coeffs_h[4], coeffs_v[4];
53
54 filt[0] = _mm256_load_si256((__m256i const *)(filt_global_avx2));
55 filt[1] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32));
56
57 prepare_coeffs_lowbd(filter_params_x, subpel_x_q4, coeffs_h);
58 prepare_coeffs(filter_params_y, subpel_y_q4, coeffs_v);
59
60 // Condition for checking valid horz_filt taps
61 if (!(_mm256_extract_epi32(_mm256_or_si256(coeffs_h[0], coeffs_h[3]), 0)))
62 is_horiz_4tap = 1;
63
64 // Condition for checking valid vert_filt taps
65 if (!(_mm256_extract_epi32(_mm256_or_si256(coeffs_v[0], coeffs_v[3]), 0)))
66 is_vert_4tap = 1;
67
68 // horz_filt as 4 tap and vert_filt as 8 tap
69 if (is_horiz_4tap) {
70 int im_h = h + filter_params_y->taps - 1;
71 const int fo_vert = filter_params_y->taps / 2 - 1;
72 const int fo_horiz = 1;
73 const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
74
75 // horz-filter
76 for (int j = 0; j < w; j += 8) {
77 for (i = 0; i < (im_h - 2); i += 2) {
78 __m256i data = _mm256_castsi128_si256(
79 _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));
80
81 // Load the next line
82 data = _mm256_inserti128_si256(
83 data,
84 _mm_loadu_si128(
85 (__m128i *)&src_ptr[(i * src_stride) + j + src_stride]),
86 1);
87 __m256i res = convolve_lowbd_x_4tap(data, coeffs_h + 1, filt);
88
89 res = _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h),
90 round_shift_h);
91 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
92 }
93
94 __m256i data_1 = _mm256_castsi128_si256(
95 _mm_loadu_si128((__m128i *)&src_ptr[(i * src_stride) + j]));
96
97 __m256i res = convolve_lowbd_x_4tap(data_1, coeffs_h + 1, filt);
98 res =
99 _mm256_sra_epi16(_mm256_add_epi16(res, round_const_h), round_shift_h);
100 _mm256_store_si256((__m256i *)&im_block[i * im_stride], res);
101
102 // vert filter
103 CONVOLVE_SR_VERTICAL_FILTER_8TAP;
104 }
105 } else if (is_vert_4tap) {
106 int im_h = h + 3;
107 const int fo_vert = 1;
108 const int fo_horiz = filter_params_x->taps / 2 - 1;
109 const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
110
111 filt[2] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
112 filt[3] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
113
114 for (int j = 0; j < w; j += 8) {
115 // horz_filter
116 CONVOLVE_SR_HORIZONTAL_FILTER_8TAP;
117 // vert_filter
118 __m256i s[6];
119 __m256i src_0 = _mm256_loadu_si256((__m256i *)(im_block + 0 * im_stride));
120 __m256i src_1 = _mm256_loadu_si256((__m256i *)(im_block + 1 * im_stride));
121 __m256i src_2 = _mm256_loadu_si256((__m256i *)(im_block + 2 * im_stride));
122 __m256i src_3 = _mm256_loadu_si256((__m256i *)(im_block + 3 * im_stride));
123
124 s[0] = _mm256_unpacklo_epi16(src_0, src_1);
125 s[1] = _mm256_unpacklo_epi16(src_2, src_3);
126 s[3] = _mm256_unpackhi_epi16(src_0, src_1);
127 s[4] = _mm256_unpackhi_epi16(src_2, src_3);
128
129 for (i = 0; i < h; i += 2) {
130 const int16_t *data = &im_block[i * im_stride];
131
132 const __m256i s4 =
133 _mm256_loadu_si256((__m256i *)(data + 4 * im_stride));
134 const __m256i s5 =
135 _mm256_loadu_si256((__m256i *)(data + 5 * im_stride));
136
137 s[2] = _mm256_unpacklo_epi16(s4, s5);
138 s[5] = _mm256_unpackhi_epi16(s4, s5);
139
140 __m256i res_a = convolve_4tap(s, coeffs_v + 1);
141 __m256i res_b = convolve_4tap(s + 3, coeffs_v + 1);
142
143 // Combine V round and 2F-H-V round into a single rounding
144 res_a =
145 _mm256_sra_epi32(_mm256_add_epi32(res_a, sum_round_v), sum_shift_v);
146 res_b =
147 _mm256_sra_epi32(_mm256_add_epi32(res_b, sum_round_v), sum_shift_v);
148
149 const __m256i res_a_round = _mm256_sra_epi32(
150 _mm256_add_epi32(res_a, round_const_v), round_shift_v);
151 const __m256i res_b_round = _mm256_sra_epi32(
152 _mm256_add_epi32(res_b, round_const_v), round_shift_v);
153
154 /* rounding code */
155 // 16 bit conversion
156 const __m256i res_16bit = _mm256_packs_epi32(res_a_round, res_b_round);
157 // 8 bit conversion and saturation to uint8
158 const __m256i res_8b = _mm256_packus_epi16(res_16bit, res_16bit);
159
160 const __m128i res_0 = _mm256_castsi256_si128(res_8b);
161 const __m128i res_1 = _mm256_extracti128_si256(res_8b, 1);
162
163 // Store values into the destination buffer
164 __m128i *const p_0 = (__m128i *)&dst[i * dst_stride + j];
165 __m128i *const p_1 = (__m128i *)&dst[i * dst_stride + j + dst_stride];
166 if (w - j > 4) {
167 _mm_storel_epi64(p_0, res_0);
168 _mm_storel_epi64(p_1, res_1);
169 } else if (w == 4) {
170 xx_storel_32(p_0, res_0);
171 xx_storel_32(p_1, res_1);
172 } else {
173 *(uint16_t *)p_0 = _mm_cvtsi128_si32(res_0);
174 *(uint16_t *)p_1 = _mm_cvtsi128_si32(res_1);
175 }
176
177 s[0] = s[1];
178 s[1] = s[2];
179 s[3] = s[4];
180 s[4] = s[5];
181 }
182 }
183 } else {
184 int j;
185 int im_h = h + filter_params_y->taps - 1;
186 const int fo_vert = filter_params_y->taps / 2 - 1;
187 const int fo_horiz = filter_params_x->taps / 2 - 1;
188 const uint8_t *const src_ptr = src - fo_vert * src_stride - fo_horiz;
189
190 filt[2] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 2));
191 filt[3] = _mm256_load_si256((__m256i const *)(filt_global_avx2 + 32 * 3));
192
193 for (j = 0; j < w; j += 8) {
194 CONVOLVE_SR_HORIZONTAL_FILTER_8TAP;
195
196 CONVOLVE_SR_VERTICAL_FILTER_8TAP;
197 }
198 }
199 }
200
copy_128(const uint8_t * src,uint8_t * dst)201 static INLINE void copy_128(const uint8_t *src, uint8_t *dst) {
202 __m256i s[4];
203 s[0] = _mm256_loadu_si256((__m256i *)(src + 0 * 32));
204 s[1] = _mm256_loadu_si256((__m256i *)(src + 1 * 32));
205 s[2] = _mm256_loadu_si256((__m256i *)(src + 2 * 32));
206 s[3] = _mm256_loadu_si256((__m256i *)(src + 3 * 32));
207 _mm256_storeu_si256((__m256i *)(dst + 0 * 32), s[0]);
208 _mm256_storeu_si256((__m256i *)(dst + 1 * 32), s[1]);
209 _mm256_storeu_si256((__m256i *)(dst + 2 * 32), s[2]);
210 _mm256_storeu_si256((__m256i *)(dst + 3 * 32), s[3]);
211 }
212
av1_convolve_2d_copy_sr_avx2(const uint8_t * src,int src_stride,uint8_t * dst,int dst_stride,int w,int h,const InterpFilterParams * filter_params_x,const InterpFilterParams * filter_params_y,const int subpel_x_q4,const int subpel_y_q4,ConvolveParams * conv_params)213 void av1_convolve_2d_copy_sr_avx2(const uint8_t *src, int src_stride,
214 uint8_t *dst, int dst_stride, int w, int h,
215 const InterpFilterParams *filter_params_x,
216 const InterpFilterParams *filter_params_y,
217 const int subpel_x_q4, const int subpel_y_q4,
218 ConvolveParams *conv_params) {
219 (void)filter_params_x;
220 (void)filter_params_y;
221 (void)subpel_x_q4;
222 (void)subpel_y_q4;
223 (void)conv_params;
224
225 if (w >= 16) {
226 assert(!((intptr_t)dst % 16));
227 assert(!(dst_stride % 16));
228 }
229
230 if (w == 2) {
231 do {
232 memmove(dst, src, 2 * sizeof(*src));
233 src += src_stride;
234 dst += dst_stride;
235 memmove(dst, src, 2 * sizeof(*src));
236 src += src_stride;
237 dst += dst_stride;
238 h -= 2;
239 } while (h);
240 } else if (w == 4) {
241 do {
242 memmove(dst, src, 4 * sizeof(*src));
243 src += src_stride;
244 dst += dst_stride;
245 memmove(dst, src, 4 * sizeof(*src));
246 src += src_stride;
247 dst += dst_stride;
248 h -= 2;
249 } while (h);
250 } else if (w == 8) {
251 do {
252 __m128i s[2];
253 s[0] = _mm_loadl_epi64((__m128i *)src);
254 src += src_stride;
255 s[1] = _mm_loadl_epi64((__m128i *)src);
256 src += src_stride;
257 _mm_storel_epi64((__m128i *)dst, s[0]);
258 dst += dst_stride;
259 _mm_storel_epi64((__m128i *)dst, s[1]);
260 dst += dst_stride;
261 h -= 2;
262 } while (h);
263 } else if (w == 16) {
264 do {
265 __m128i s[2];
266 s[0] = _mm_loadu_si128((__m128i *)src);
267 src += src_stride;
268 s[1] = _mm_loadu_si128((__m128i *)src);
269 src += src_stride;
270 _mm_store_si128((__m128i *)dst, s[0]);
271 dst += dst_stride;
272 _mm_store_si128((__m128i *)dst, s[1]);
273 dst += dst_stride;
274 h -= 2;
275 } while (h);
276 } else if (w == 32) {
277 do {
278 __m256i s[2];
279 s[0] = _mm256_loadu_si256((__m256i *)src);
280 src += src_stride;
281 s[1] = _mm256_loadu_si256((__m256i *)src);
282 src += src_stride;
283 _mm256_storeu_si256((__m256i *)dst, s[0]);
284 dst += dst_stride;
285 _mm256_storeu_si256((__m256i *)dst, s[1]);
286 dst += dst_stride;
287 h -= 2;
288 } while (h);
289 } else if (w == 64) {
290 do {
291 __m256i s[4];
292 s[0] = _mm256_loadu_si256((__m256i *)(src + 0 * 32));
293 s[1] = _mm256_loadu_si256((__m256i *)(src + 1 * 32));
294 src += src_stride;
295 s[2] = _mm256_loadu_si256((__m256i *)(src + 0 * 32));
296 s[3] = _mm256_loadu_si256((__m256i *)(src + 1 * 32));
297 src += src_stride;
298 _mm256_storeu_si256((__m256i *)(dst + 0 * 32), s[0]);
299 _mm256_storeu_si256((__m256i *)(dst + 1 * 32), s[1]);
300 dst += dst_stride;
301 _mm256_storeu_si256((__m256i *)(dst + 0 * 32), s[2]);
302 _mm256_storeu_si256((__m256i *)(dst + 1 * 32), s[3]);
303 dst += dst_stride;
304 h -= 2;
305 } while (h);
306 } else {
307 do {
308 copy_128(src, dst);
309 src += src_stride;
310 dst += dst_stride;
311 copy_128(src, dst);
312 src += src_stride;
313 dst += dst_stride;
314 h -= 2;
315 } while (h);
316 }
317 }
318