1 /*
2 * Copyright (c) 2016, 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 #ifndef AOM_AV1_COMMON_CDEF_BLOCK_SIMD_H_
13 #define AOM_AV1_COMMON_CDEF_BLOCK_SIMD_H_
14
15 #include "config/av1_rtcd.h"
16
17 #include "av1/common/cdef_block.h"
18
19 /* partial A is a 16-bit vector of the form:
20 [x8 x7 x6 x5 x4 x3 x2 x1] and partial B has the form:
21 [0 y1 y2 y3 y4 y5 y6 y7].
22 This function computes (x1^2+y1^2)*C1 + (x2^2+y2^2)*C2 + ...
23 (x7^2+y2^7)*C7 + (x8^2+0^2)*C8 where the C1..C8 constants are in const1
24 and const2. */
fold_mul_and_sum(v128 partiala,v128 partialb,v128 const1,v128 const2)25 static INLINE v128 fold_mul_and_sum(v128 partiala, v128 partialb, v128 const1,
26 v128 const2) {
27 v128 tmp;
28 /* Reverse partial B. */
29 partialb = v128_shuffle_8(
30 partialb, v128_from_32(0x0f0e0100, 0x03020504, 0x07060908, 0x0b0a0d0c));
31 /* Interleave the x and y values of identical indices and pair x8 with 0. */
32 tmp = partiala;
33 partiala = v128_ziplo_16(partialb, partiala);
34 partialb = v128_ziphi_16(partialb, tmp);
35 /* Square and add the corresponding x and y values. */
36 partiala = v128_madd_s16(partiala, partiala);
37 partialb = v128_madd_s16(partialb, partialb);
38 /* Multiply by constant. */
39 partiala = v128_mullo_s32(partiala, const1);
40 partialb = v128_mullo_s32(partialb, const2);
41 /* Sum all results. */
42 partiala = v128_add_32(partiala, partialb);
43 return partiala;
44 }
45
hsum4(v128 x0,v128 x1,v128 x2,v128 x3)46 static INLINE v128 hsum4(v128 x0, v128 x1, v128 x2, v128 x3) {
47 v128 t0, t1, t2, t3;
48 t0 = v128_ziplo_32(x1, x0);
49 t1 = v128_ziplo_32(x3, x2);
50 t2 = v128_ziphi_32(x1, x0);
51 t3 = v128_ziphi_32(x3, x2);
52 x0 = v128_ziplo_64(t1, t0);
53 x1 = v128_ziphi_64(t1, t0);
54 x2 = v128_ziplo_64(t3, t2);
55 x3 = v128_ziphi_64(t3, t2);
56 return v128_add_32(v128_add_32(x0, x1), v128_add_32(x2, x3));
57 }
58
59 /* Computes cost for directions 0, 5, 6 and 7. We can call this function again
60 to compute the remaining directions. */
compute_directions(v128 lines[8],int32_t tmp_cost1[4])61 static INLINE v128 compute_directions(v128 lines[8], int32_t tmp_cost1[4]) {
62 v128 partial4a, partial4b, partial5a, partial5b, partial7a, partial7b;
63 v128 partial6;
64 v128 tmp;
65 /* Partial sums for lines 0 and 1. */
66 partial4a = v128_shl_n_byte(lines[0], 14);
67 partial4b = v128_shr_n_byte(lines[0], 2);
68 partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[1], 12));
69 partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[1], 4));
70 tmp = v128_add_16(lines[0], lines[1]);
71 partial5a = v128_shl_n_byte(tmp, 10);
72 partial5b = v128_shr_n_byte(tmp, 6);
73 partial7a = v128_shl_n_byte(tmp, 4);
74 partial7b = v128_shr_n_byte(tmp, 12);
75 partial6 = tmp;
76
77 /* Partial sums for lines 2 and 3. */
78 partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[2], 10));
79 partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[2], 6));
80 partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[3], 8));
81 partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[3], 8));
82 tmp = v128_add_16(lines[2], lines[3]);
83 partial5a = v128_add_16(partial5a, v128_shl_n_byte(tmp, 8));
84 partial5b = v128_add_16(partial5b, v128_shr_n_byte(tmp, 8));
85 partial7a = v128_add_16(partial7a, v128_shl_n_byte(tmp, 6));
86 partial7b = v128_add_16(partial7b, v128_shr_n_byte(tmp, 10));
87 partial6 = v128_add_16(partial6, tmp);
88
89 /* Partial sums for lines 4 and 5. */
90 partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[4], 6));
91 partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[4], 10));
92 partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[5], 4));
93 partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[5], 12));
94 tmp = v128_add_16(lines[4], lines[5]);
95 partial5a = v128_add_16(partial5a, v128_shl_n_byte(tmp, 6));
96 partial5b = v128_add_16(partial5b, v128_shr_n_byte(tmp, 10));
97 partial7a = v128_add_16(partial7a, v128_shl_n_byte(tmp, 8));
98 partial7b = v128_add_16(partial7b, v128_shr_n_byte(tmp, 8));
99 partial6 = v128_add_16(partial6, tmp);
100
101 /* Partial sums for lines 6 and 7. */
102 partial4a = v128_add_16(partial4a, v128_shl_n_byte(lines[6], 2));
103 partial4b = v128_add_16(partial4b, v128_shr_n_byte(lines[6], 14));
104 partial4a = v128_add_16(partial4a, lines[7]);
105 tmp = v128_add_16(lines[6], lines[7]);
106 partial5a = v128_add_16(partial5a, v128_shl_n_byte(tmp, 4));
107 partial5b = v128_add_16(partial5b, v128_shr_n_byte(tmp, 12));
108 partial7a = v128_add_16(partial7a, v128_shl_n_byte(tmp, 10));
109 partial7b = v128_add_16(partial7b, v128_shr_n_byte(tmp, 6));
110 partial6 = v128_add_16(partial6, tmp);
111
112 /* Compute costs in terms of partial sums. */
113 partial4a =
114 fold_mul_and_sum(partial4a, partial4b, v128_from_32(210, 280, 420, 840),
115 v128_from_32(105, 120, 140, 168));
116 partial7a =
117 fold_mul_and_sum(partial7a, partial7b, v128_from_32(210, 420, 0, 0),
118 v128_from_32(105, 105, 105, 140));
119 partial5a =
120 fold_mul_and_sum(partial5a, partial5b, v128_from_32(210, 420, 0, 0),
121 v128_from_32(105, 105, 105, 140));
122 partial6 = v128_madd_s16(partial6, partial6);
123 partial6 = v128_mullo_s32(partial6, v128_dup_32(105));
124
125 partial4a = hsum4(partial4a, partial5a, partial6, partial7a);
126 v128_store_unaligned(tmp_cost1, partial4a);
127 return partial4a;
128 }
129
130 /* transpose and reverse the order of the lines -- equivalent to a 90-degree
131 counter-clockwise rotation of the pixels. */
array_reverse_transpose_8x8(v128 * in,v128 * res)132 static INLINE void array_reverse_transpose_8x8(v128 *in, v128 *res) {
133 const v128 tr0_0 = v128_ziplo_16(in[1], in[0]);
134 const v128 tr0_1 = v128_ziplo_16(in[3], in[2]);
135 const v128 tr0_2 = v128_ziphi_16(in[1], in[0]);
136 const v128 tr0_3 = v128_ziphi_16(in[3], in[2]);
137 const v128 tr0_4 = v128_ziplo_16(in[5], in[4]);
138 const v128 tr0_5 = v128_ziplo_16(in[7], in[6]);
139 const v128 tr0_6 = v128_ziphi_16(in[5], in[4]);
140 const v128 tr0_7 = v128_ziphi_16(in[7], in[6]);
141
142 const v128 tr1_0 = v128_ziplo_32(tr0_1, tr0_0);
143 const v128 tr1_1 = v128_ziplo_32(tr0_5, tr0_4);
144 const v128 tr1_2 = v128_ziphi_32(tr0_1, tr0_0);
145 const v128 tr1_3 = v128_ziphi_32(tr0_5, tr0_4);
146 const v128 tr1_4 = v128_ziplo_32(tr0_3, tr0_2);
147 const v128 tr1_5 = v128_ziplo_32(tr0_7, tr0_6);
148 const v128 tr1_6 = v128_ziphi_32(tr0_3, tr0_2);
149 const v128 tr1_7 = v128_ziphi_32(tr0_7, tr0_6);
150
151 res[7] = v128_ziplo_64(tr1_1, tr1_0);
152 res[6] = v128_ziphi_64(tr1_1, tr1_0);
153 res[5] = v128_ziplo_64(tr1_3, tr1_2);
154 res[4] = v128_ziphi_64(tr1_3, tr1_2);
155 res[3] = v128_ziplo_64(tr1_5, tr1_4);
156 res[2] = v128_ziphi_64(tr1_5, tr1_4);
157 res[1] = v128_ziplo_64(tr1_7, tr1_6);
158 res[0] = v128_ziphi_64(tr1_7, tr1_6);
159 }
160
SIMD_FUNC(cdef_find_dir)161 int SIMD_FUNC(cdef_find_dir)(const uint16_t *img, int stride, int32_t *var,
162 int coeff_shift) {
163 int i;
164 int32_t cost[8];
165 int32_t best_cost = 0;
166 int best_dir = 0;
167 v128 lines[8];
168 for (i = 0; i < 8; i++) {
169 lines[i] = v128_load_unaligned(&img[i * stride]);
170 lines[i] =
171 v128_sub_16(v128_shr_s16(lines[i], coeff_shift), v128_dup_16(128));
172 }
173
174 /* Compute "mostly vertical" directions. */
175 v128 dir47 = compute_directions(lines, cost + 4);
176
177 array_reverse_transpose_8x8(lines, lines);
178
179 /* Compute "mostly horizontal" directions. */
180 v128 dir03 = compute_directions(lines, cost);
181
182 v128 max = v128_max_s32(dir03, dir47);
183 max = v128_max_s32(max, v128_align(max, max, 8));
184 max = v128_max_s32(max, v128_align(max, max, 4));
185 best_cost = v128_low_u32(max);
186 v128 t =
187 v128_pack_s32_s16(v128_cmpeq_32(max, dir47), v128_cmpeq_32(max, dir03));
188 best_dir = v128_movemask_8(v128_pack_s16_s8(t, t));
189 best_dir = get_msb(best_dir ^ (best_dir - 1)); // Count trailing zeros
190
191 /* Difference between the optimal variance and the variance along the
192 orthogonal direction. Again, the sum(x^2) terms cancel out. */
193 *var = best_cost - cost[(best_dir + 4) & 7];
194 /* We'd normally divide by 840, but dividing by 1024 is close enough
195 for what we're going to do with this. */
196 *var >>= 10;
197 return best_dir;
198 }
199
200 // Work around compiler out of memory issues with Win32 builds. This issue has
201 // been observed with Visual Studio 2017, 2019, and 2022 (version 17.4).
202 #if defined(_MSC_VER) && defined(_M_IX86) && _MSC_VER < 1940
203 #define CDEF_INLINE static INLINE
204 #else
205 #define CDEF_INLINE SIMD_INLINE
206 #endif
207
208 // sign(a-b) * min(abs(a-b), max(0, threshold - (abs(a-b) >> adjdamp)))
constrain16(v256 a,v256 b,unsigned int threshold,unsigned int adjdamp)209 CDEF_INLINE v256 constrain16(v256 a, v256 b, unsigned int threshold,
210 unsigned int adjdamp) {
211 v256 diff = v256_sub_16(a, b);
212 const v256 sign = v256_shr_n_s16(diff, 15);
213 diff = v256_abs_s16(diff);
214 const v256 s =
215 v256_ssub_u16(v256_dup_16(threshold), v256_shr_u16(diff, adjdamp));
216 return v256_xor(v256_add_16(sign, v256_min_s16(diff, s)), sign);
217 }
218
get_max_primary(const int is_lowbd,v256 * tap,v256 max,v256 cdef_large_value_mask)219 SIMD_INLINE v256 get_max_primary(const int is_lowbd, v256 *tap, v256 max,
220 v256 cdef_large_value_mask) {
221 if (is_lowbd) {
222 v256 max_u8;
223 max_u8 = tap[0];
224 max_u8 = v256_max_u8(max_u8, tap[1]);
225 max_u8 = v256_max_u8(max_u8, tap[2]);
226 max_u8 = v256_max_u8(max_u8, tap[3]);
227 /* The source is 16 bits, however, we only really care about the lower
228 8 bits. The upper 8 bits contain the "large" flag. After the final
229 primary max has been calculated, zero out the upper 8 bits. Use this
230 to find the "16 bit" max. */
231 max = v256_max_s16(max, v256_and(max_u8, cdef_large_value_mask));
232 } else {
233 /* Convert CDEF_VERY_LARGE to 0 before calculating max. */
234 max = v256_max_s16(max, v256_and(tap[0], cdef_large_value_mask));
235 max = v256_max_s16(max, v256_and(tap[1], cdef_large_value_mask));
236 max = v256_max_s16(max, v256_and(tap[2], cdef_large_value_mask));
237 max = v256_max_s16(max, v256_and(tap[3], cdef_large_value_mask));
238 }
239 return max;
240 }
241
get_max_secondary(const int is_lowbd,v256 * tap,v256 max,v256 cdef_large_value_mask)242 SIMD_INLINE v256 get_max_secondary(const int is_lowbd, v256 *tap, v256 max,
243 v256 cdef_large_value_mask) {
244 if (is_lowbd) {
245 v256 max_u8;
246 max_u8 = tap[0];
247 max_u8 = v256_max_u8(max_u8, tap[1]);
248 max_u8 = v256_max_u8(max_u8, tap[2]);
249 max_u8 = v256_max_u8(max_u8, tap[3]);
250 max_u8 = v256_max_u8(max_u8, tap[4]);
251 max_u8 = v256_max_u8(max_u8, tap[5]);
252 max_u8 = v256_max_u8(max_u8, tap[6]);
253 max_u8 = v256_max_u8(max_u8, tap[7]);
254 /* The source is 16 bits, however, we only really care about the lower
255 8 bits. The upper 8 bits contain the "large" flag. After the final
256 primary max has been calculated, zero out the upper 8 bits. Use this
257 to find the "16 bit" max. */
258 max = v256_max_s16(max, v256_and(max_u8, cdef_large_value_mask));
259 } else {
260 /* Convert CDEF_VERY_LARGE to 0 before calculating max. */
261 max = v256_max_s16(max, v256_and(tap[0], cdef_large_value_mask));
262 max = v256_max_s16(max, v256_and(tap[1], cdef_large_value_mask));
263 max = v256_max_s16(max, v256_and(tap[2], cdef_large_value_mask));
264 max = v256_max_s16(max, v256_and(tap[3], cdef_large_value_mask));
265 max = v256_max_s16(max, v256_and(tap[4], cdef_large_value_mask));
266 max = v256_max_s16(max, v256_and(tap[5], cdef_large_value_mask));
267 max = v256_max_s16(max, v256_and(tap[6], cdef_large_value_mask));
268 max = v256_max_s16(max, v256_and(tap[7], cdef_large_value_mask));
269 }
270 return max;
271 }
272
273 // MSVC takes far too much time optimizing these.
274 // https://bugs.chromium.org/p/aomedia/issues/detail?id=3395
275 #if defined(_MSC_VER) && !defined(__clang__)
276 #pragma optimize("", off)
277 #endif
278
filter_block_4x4(const int is_lowbd,void * dest,int dstride,const uint16_t * in,int pri_strength,int sec_strength,int dir,int pri_damping,int sec_damping,int coeff_shift,int height,int enable_primary,int enable_secondary)279 CDEF_INLINE void filter_block_4x4(const int is_lowbd, void *dest, int dstride,
280 const uint16_t *in, int pri_strength,
281 int sec_strength, int dir, int pri_damping,
282 int sec_damping, int coeff_shift, int height,
283 int enable_primary, int enable_secondary) {
284 uint8_t *dst8 = (uint8_t *)dest;
285 uint16_t *dst16 = (uint16_t *)dest;
286 const int clipping_required = enable_primary && enable_secondary;
287 v256 p0, p1, p2, p3;
288 v256 sum, row, res;
289 v256 max, min;
290 const v256 cdef_large_value_mask = v256_dup_16((uint16_t)~CDEF_VERY_LARGE);
291 const int po1 = cdef_directions[dir][0];
292 const int po2 = cdef_directions[dir][1];
293 const int s1o1 = cdef_directions[dir + 2][0];
294 const int s1o2 = cdef_directions[dir + 2][1];
295 const int s2o1 = cdef_directions[dir - 2][0];
296 const int s2o2 = cdef_directions[dir - 2][1];
297 const int *pri_taps = cdef_pri_taps[(pri_strength >> coeff_shift) & 1];
298 const int *sec_taps = cdef_sec_taps;
299 int i;
300
301 if (enable_primary && pri_strength)
302 pri_damping = AOMMAX(0, pri_damping - get_msb(pri_strength));
303 if (enable_secondary && sec_strength)
304 sec_damping = AOMMAX(0, sec_damping - get_msb(sec_strength));
305
306 for (i = 0; i < height; i += 4) {
307 sum = v256_zero();
308 row = v256_from_v64(v64_load_aligned(&in[(i + 0) * CDEF_BSTRIDE]),
309 v64_load_aligned(&in[(i + 1) * CDEF_BSTRIDE]),
310 v64_load_aligned(&in[(i + 2) * CDEF_BSTRIDE]),
311 v64_load_aligned(&in[(i + 3) * CDEF_BSTRIDE]));
312 max = min = row;
313
314 if (enable_primary) {
315 v256 tap[4];
316 // Primary near taps
317 tap[0] =
318 v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE + po1]),
319 v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + po1]),
320 v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE + po1]),
321 v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE + po1]));
322 p0 = constrain16(tap[0], row, pri_strength, pri_damping);
323 tap[1] =
324 v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE - po1]),
325 v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - po1]),
326 v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE - po1]),
327 v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE - po1]));
328 p1 = constrain16(tap[1], row, pri_strength, pri_damping);
329
330 // sum += pri_taps[0] * (p0 + p1)
331 sum = v256_add_16(
332 sum, v256_mullo_s16(v256_dup_16(pri_taps[0]), v256_add_16(p0, p1)));
333
334 // Primary far taps
335 tap[2] =
336 v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE + po2]),
337 v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + po2]),
338 v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE + po2]),
339 v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE + po2]));
340 p0 = constrain16(tap[2], row, pri_strength, pri_damping);
341 tap[3] =
342 v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE - po2]),
343 v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - po2]),
344 v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE - po2]),
345 v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE - po2]));
346 p1 = constrain16(tap[3], row, pri_strength, pri_damping);
347
348 // sum += pri_taps[1] * (p0 + p1)
349 sum = v256_add_16(
350 sum, v256_mullo_s16(v256_dup_16(pri_taps[1]), v256_add_16(p0, p1)));
351 if (clipping_required) {
352 max = get_max_primary(is_lowbd, tap, max, cdef_large_value_mask);
353
354 min = v256_min_s16(min, tap[0]);
355 min = v256_min_s16(min, tap[1]);
356 min = v256_min_s16(min, tap[2]);
357 min = v256_min_s16(min, tap[3]);
358 }
359 }
360
361 if (enable_secondary) {
362 v256 tap[8];
363 // Secondary near taps
364 tap[0] =
365 v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE + s1o1]),
366 v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s1o1]),
367 v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE + s1o1]),
368 v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE + s1o1]));
369 p0 = constrain16(tap[0], row, sec_strength, sec_damping);
370 tap[1] =
371 v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE - s1o1]),
372 v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s1o1]),
373 v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE - s1o1]),
374 v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE - s1o1]));
375 p1 = constrain16(tap[1], row, sec_strength, sec_damping);
376 tap[2] =
377 v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE + s2o1]),
378 v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s2o1]),
379 v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE + s2o1]),
380 v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE + s2o1]));
381 p2 = constrain16(tap[2], row, sec_strength, sec_damping);
382 tap[3] =
383 v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE - s2o1]),
384 v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s2o1]),
385 v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE - s2o1]),
386 v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE - s2o1]));
387 p3 = constrain16(tap[3], row, sec_strength, sec_damping);
388
389 // sum += sec_taps[0] * (p0 + p1 + p2 + p3)
390 sum = v256_add_16(sum, v256_mullo_s16(v256_dup_16(sec_taps[0]),
391 v256_add_16(v256_add_16(p0, p1),
392 v256_add_16(p2, p3))));
393
394 // Secondary far taps
395 tap[4] =
396 v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE + s1o2]),
397 v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s1o2]),
398 v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE + s1o2]),
399 v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE + s1o2]));
400 p0 = constrain16(tap[4], row, sec_strength, sec_damping);
401 tap[5] =
402 v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE - s1o2]),
403 v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s1o2]),
404 v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE - s1o2]),
405 v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE - s1o2]));
406 p1 = constrain16(tap[5], row, sec_strength, sec_damping);
407 tap[6] =
408 v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE + s2o2]),
409 v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s2o2]),
410 v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE + s2o2]),
411 v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE + s2o2]));
412 p2 = constrain16(tap[6], row, sec_strength, sec_damping);
413 tap[7] =
414 v256_from_v64(v64_load_unaligned(&in[(i + 0) * CDEF_BSTRIDE - s2o2]),
415 v64_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s2o2]),
416 v64_load_unaligned(&in[(i + 2) * CDEF_BSTRIDE - s2o2]),
417 v64_load_unaligned(&in[(i + 3) * CDEF_BSTRIDE - s2o2]));
418 p3 = constrain16(tap[7], row, sec_strength, sec_damping);
419
420 // sum += sec_taps[1] * (p0 + p1 + p2 + p3)
421 sum = v256_add_16(sum, v256_mullo_s16(v256_dup_16(sec_taps[1]),
422 v256_add_16(v256_add_16(p0, p1),
423 v256_add_16(p2, p3))));
424
425 if (clipping_required) {
426 max = get_max_secondary(is_lowbd, tap, max, cdef_large_value_mask);
427
428 min = v256_min_s16(min, tap[0]);
429 min = v256_min_s16(min, tap[1]);
430 min = v256_min_s16(min, tap[2]);
431 min = v256_min_s16(min, tap[3]);
432 min = v256_min_s16(min, tap[4]);
433 min = v256_min_s16(min, tap[5]);
434 min = v256_min_s16(min, tap[6]);
435 min = v256_min_s16(min, tap[7]);
436 }
437 }
438
439 // res = row + ((sum - (sum < 0) + 8) >> 4)
440 sum = v256_add_16(sum, v256_cmplt_s16(sum, v256_zero()));
441 res = v256_add_16(sum, v256_dup_16(8));
442 res = v256_shr_n_s16(res, 4);
443 res = v256_add_16(row, res);
444 if (clipping_required) {
445 res = v256_min_s16(v256_max_s16(res, min), max);
446 }
447
448 if (is_lowbd) {
449 const v128 res_128 = v256_low_v128(v256_pack_s16_u8(res, res));
450 u32_store_aligned(&dst8[(i + 0) * dstride],
451 v64_high_u32(v128_high_v64(res_128)));
452 u32_store_aligned(&dst8[(i + 1) * dstride],
453 v64_low_u32(v128_high_v64(res_128)));
454 u32_store_aligned(&dst8[(i + 2) * dstride],
455 v64_high_u32(v128_low_v64(res_128)));
456 u32_store_aligned(&dst8[(i + 3) * dstride],
457 v64_low_u32(v128_low_v64(res_128)));
458 } else {
459 v64_store_aligned(&dst16[(i + 0) * dstride],
460 v128_high_v64(v256_high_v128(res)));
461 v64_store_aligned(&dst16[(i + 1) * dstride],
462 v128_low_v64(v256_high_v128(res)));
463 v64_store_aligned(&dst16[(i + 2) * dstride],
464 v128_high_v64(v256_low_v128(res)));
465 v64_store_aligned(&dst16[(i + 3) * dstride],
466 v128_low_v64(v256_low_v128(res)));
467 }
468 }
469 }
470
filter_block_8x8(const int is_lowbd,void * dest,int dstride,const uint16_t * in,int pri_strength,int sec_strength,int dir,int pri_damping,int sec_damping,int coeff_shift,int height,int enable_primary,int enable_secondary)471 CDEF_INLINE void filter_block_8x8(const int is_lowbd, void *dest, int dstride,
472 const uint16_t *in, int pri_strength,
473 int sec_strength, int dir, int pri_damping,
474 int sec_damping, int coeff_shift, int height,
475 int enable_primary, int enable_secondary) {
476 uint8_t *dst8 = (uint8_t *)dest;
477 uint16_t *dst16 = (uint16_t *)dest;
478 const int clipping_required = enable_primary && enable_secondary;
479 int i;
480 v256 sum, p0, p1, p2, p3, row, res;
481 const v256 cdef_large_value_mask = v256_dup_16((uint16_t)~CDEF_VERY_LARGE);
482 v256 max, min;
483 const int po1 = cdef_directions[dir][0];
484 const int po2 = cdef_directions[dir][1];
485 const int s1o1 = cdef_directions[dir + 2][0];
486 const int s1o2 = cdef_directions[dir + 2][1];
487 const int s2o1 = cdef_directions[dir - 2][0];
488 const int s2o2 = cdef_directions[dir - 2][1];
489 const int *pri_taps = cdef_pri_taps[(pri_strength >> coeff_shift) & 1];
490 const int *sec_taps = cdef_sec_taps;
491
492 if (enable_primary && pri_strength)
493 pri_damping = AOMMAX(0, pri_damping - get_msb(pri_strength));
494 if (enable_secondary && sec_strength)
495 sec_damping = AOMMAX(0, sec_damping - get_msb(sec_strength));
496
497 for (i = 0; i < height; i += 2) {
498 v256 tap[8];
499 sum = v256_zero();
500 row = v256_from_v128(v128_load_aligned(&in[i * CDEF_BSTRIDE]),
501 v128_load_aligned(&in[(i + 1) * CDEF_BSTRIDE]));
502
503 min = max = row;
504 if (enable_primary) {
505 // Primary near taps
506 tap[0] = v256_from_v128(
507 v128_load_unaligned(&in[i * CDEF_BSTRIDE + po1]),
508 v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + po1]));
509 tap[1] = v256_from_v128(
510 v128_load_unaligned(&in[i * CDEF_BSTRIDE - po1]),
511 v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - po1]));
512 p0 = constrain16(tap[0], row, pri_strength, pri_damping);
513 p1 = constrain16(tap[1], row, pri_strength, pri_damping);
514
515 // sum += pri_taps[0] * (p0 + p1)
516 sum = v256_add_16(
517 sum, v256_mullo_s16(v256_dup_16(pri_taps[0]), v256_add_16(p0, p1)));
518
519 // Primary far taps
520 tap[2] = v256_from_v128(
521 v128_load_unaligned(&in[i * CDEF_BSTRIDE + po2]),
522 v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + po2]));
523 tap[3] = v256_from_v128(
524 v128_load_unaligned(&in[i * CDEF_BSTRIDE - po2]),
525 v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - po2]));
526 p0 = constrain16(tap[2], row, pri_strength, pri_damping);
527 p1 = constrain16(tap[3], row, pri_strength, pri_damping);
528
529 // sum += pri_taps[1] * (p0 + p1)
530 sum = v256_add_16(
531 sum, v256_mullo_s16(v256_dup_16(pri_taps[1]), v256_add_16(p0, p1)));
532
533 if (clipping_required) {
534 max = get_max_primary(is_lowbd, tap, max, cdef_large_value_mask);
535
536 min = v256_min_s16(min, tap[0]);
537 min = v256_min_s16(min, tap[1]);
538 min = v256_min_s16(min, tap[2]);
539 min = v256_min_s16(min, tap[3]);
540 }
541 // End primary
542 }
543
544 if (enable_secondary) {
545 // Secondary near taps
546 tap[0] = v256_from_v128(
547 v128_load_unaligned(&in[i * CDEF_BSTRIDE + s1o1]),
548 v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s1o1]));
549 tap[1] = v256_from_v128(
550 v128_load_unaligned(&in[i * CDEF_BSTRIDE - s1o1]),
551 v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s1o1]));
552 tap[2] = v256_from_v128(
553 v128_load_unaligned(&in[i * CDEF_BSTRIDE + s2o1]),
554 v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s2o1]));
555 tap[3] = v256_from_v128(
556 v128_load_unaligned(&in[i * CDEF_BSTRIDE - s2o1]),
557 v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s2o1]));
558 p0 = constrain16(tap[0], row, sec_strength, sec_damping);
559 p1 = constrain16(tap[1], row, sec_strength, sec_damping);
560 p2 = constrain16(tap[2], row, sec_strength, sec_damping);
561 p3 = constrain16(tap[3], row, sec_strength, sec_damping);
562
563 // sum += sec_taps[0] * (p0 + p1 + p2 + p3)
564 sum = v256_add_16(sum, v256_mullo_s16(v256_dup_16(sec_taps[0]),
565 v256_add_16(v256_add_16(p0, p1),
566 v256_add_16(p2, p3))));
567
568 // Secondary far taps
569 tap[4] = v256_from_v128(
570 v128_load_unaligned(&in[i * CDEF_BSTRIDE + s1o2]),
571 v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s1o2]));
572 tap[5] = v256_from_v128(
573 v128_load_unaligned(&in[i * CDEF_BSTRIDE - s1o2]),
574 v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s1o2]));
575 tap[6] = v256_from_v128(
576 v128_load_unaligned(&in[i * CDEF_BSTRIDE + s2o2]),
577 v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE + s2o2]));
578 tap[7] = v256_from_v128(
579 v128_load_unaligned(&in[i * CDEF_BSTRIDE - s2o2]),
580 v128_load_unaligned(&in[(i + 1) * CDEF_BSTRIDE - s2o2]));
581 p0 = constrain16(tap[4], row, sec_strength, sec_damping);
582 p1 = constrain16(tap[5], row, sec_strength, sec_damping);
583 p2 = constrain16(tap[6], row, sec_strength, sec_damping);
584 p3 = constrain16(tap[7], row, sec_strength, sec_damping);
585
586 // sum += sec_taps[1] * (p0 + p1 + p2 + p3)
587 sum = v256_add_16(sum, v256_mullo_s16(v256_dup_16(sec_taps[1]),
588 v256_add_16(v256_add_16(p0, p1),
589 v256_add_16(p2, p3))));
590
591 if (clipping_required) {
592 max = get_max_secondary(is_lowbd, tap, max, cdef_large_value_mask);
593
594 min = v256_min_s16(min, tap[0]);
595 min = v256_min_s16(min, tap[1]);
596 min = v256_min_s16(min, tap[2]);
597 min = v256_min_s16(min, tap[3]);
598 min = v256_min_s16(min, tap[4]);
599 min = v256_min_s16(min, tap[5]);
600 min = v256_min_s16(min, tap[6]);
601 min = v256_min_s16(min, tap[7]);
602 }
603 // End secondary
604 }
605
606 // res = row + ((sum - (sum < 0) + 8) >> 4)
607 sum = v256_add_16(sum, v256_cmplt_s16(sum, v256_zero()));
608 res = v256_add_16(sum, v256_dup_16(8));
609 res = v256_shr_n_s16(res, 4);
610 res = v256_add_16(row, res);
611 if (clipping_required) {
612 res = v256_min_s16(v256_max_s16(res, min), max);
613 }
614
615 if (is_lowbd) {
616 const v128 res_128 = v256_low_v128(v256_pack_s16_u8(res, res));
617 v64_store_aligned(&dst8[i * dstride], v128_high_v64(res_128));
618 v64_store_aligned(&dst8[(i + 1) * dstride], v128_low_v64(res_128));
619 } else {
620 v128_store_unaligned(&dst16[i * dstride], v256_high_v128(res));
621 v128_store_unaligned(&dst16[(i + 1) * dstride], v256_low_v128(res));
622 }
623 }
624 }
625
626 #if defined(_MSC_VER) && !defined(__clang__)
627 #pragma optimize("", on)
628 #endif
629
copy_block_4xh(const int is_lowbd,void * dest,int dstride,const uint16_t * in,int height)630 SIMD_INLINE void copy_block_4xh(const int is_lowbd, void *dest, int dstride,
631 const uint16_t *in, int height) {
632 uint8_t *dst8 = (uint8_t *)dest;
633 uint16_t *dst16 = (uint16_t *)dest;
634 int i;
635 for (i = 0; i < height; i += 4) {
636 const v128 row0 =
637 v128_from_v64(v64_load_aligned(&in[(i + 0) * CDEF_BSTRIDE]),
638 v64_load_aligned(&in[(i + 1) * CDEF_BSTRIDE]));
639 const v128 row1 =
640 v128_from_v64(v64_load_aligned(&in[(i + 2) * CDEF_BSTRIDE]),
641 v64_load_aligned(&in[(i + 3) * CDEF_BSTRIDE]));
642 if (is_lowbd) {
643 /* Note: v128_pack_s16_u8(). The parameter order is swapped internally */
644 const v128 res_128 = v128_pack_s16_u8(row1, row0);
645 u32_store_aligned(&dst8[(i + 0) * dstride],
646 v64_high_u32(v128_low_v64(res_128)));
647 u32_store_aligned(&dst8[(i + 1) * dstride],
648 v64_low_u32(v128_low_v64(res_128)));
649 u32_store_aligned(&dst8[(i + 2) * dstride],
650 v64_high_u32(v128_high_v64(res_128)));
651 u32_store_aligned(&dst8[(i + 3) * dstride],
652 v64_low_u32(v128_high_v64(res_128)));
653 } else {
654 v64_store_aligned(&dst16[(i + 0) * dstride], v128_high_v64(row0));
655 v64_store_aligned(&dst16[(i + 1) * dstride], v128_low_v64(row0));
656 v64_store_aligned(&dst16[(i + 2) * dstride], v128_high_v64(row1));
657 v64_store_aligned(&dst16[(i + 3) * dstride], v128_low_v64(row1));
658 }
659 }
660 }
661
copy_block_8xh(const int is_lowbd,void * dest,int dstride,const uint16_t * in,int height)662 SIMD_INLINE void copy_block_8xh(const int is_lowbd, void *dest, int dstride,
663 const uint16_t *in, int height) {
664 uint8_t *dst8 = (uint8_t *)dest;
665 uint16_t *dst16 = (uint16_t *)dest;
666 int i;
667 for (i = 0; i < height; i += 2) {
668 const v128 row0 = v128_load_aligned(&in[i * CDEF_BSTRIDE]);
669 const v128 row1 = v128_load_aligned(&in[(i + 1) * CDEF_BSTRIDE]);
670 if (is_lowbd) {
671 /* Note: v128_pack_s16_u8(). The parameter order is swapped internally */
672 const v128 res_128 = v128_pack_s16_u8(row1, row0);
673 v64_store_aligned(&dst8[i * dstride], v128_low_v64(res_128));
674 v64_store_aligned(&dst8[(i + 1) * dstride], v128_high_v64(res_128));
675 } else {
676 v128_store_unaligned(&dst16[i * dstride], row0);
677 v128_store_unaligned(&dst16[(i + 1) * dstride], row1);
678 }
679 }
680 }
681
SIMD_FUNC(cdef_filter_8_0)682 void SIMD_FUNC(cdef_filter_8_0)(void *dest, int dstride, const uint16_t *in,
683 int pri_strength, int sec_strength, int dir,
684 int pri_damping, int sec_damping,
685 int coeff_shift, int block_width,
686 int block_height) {
687 if (block_width == 8) {
688 filter_block_8x8(/*is_lowbd=*/1, dest, dstride, in, pri_strength,
689 sec_strength, dir, pri_damping, sec_damping, coeff_shift,
690 block_height, /*enable_primary=*/1,
691 /*enable_secondary=*/1);
692 } else {
693 filter_block_4x4(/*is_lowbd=*/1, dest, dstride, in, pri_strength,
694 sec_strength, dir, pri_damping, sec_damping, coeff_shift,
695 block_height, /*enable_primary=*/1,
696 /*enable_secondary=*/1);
697 }
698 }
699
SIMD_FUNC(cdef_filter_8_1)700 void SIMD_FUNC(cdef_filter_8_1)(void *dest, int dstride, const uint16_t *in,
701 int pri_strength, int sec_strength, int dir,
702 int pri_damping, int sec_damping,
703 int coeff_shift, int block_width,
704 int block_height) {
705 if (block_width == 8) {
706 filter_block_8x8(/*is_lowbd=*/1, dest, dstride, in, pri_strength,
707 sec_strength, dir, pri_damping, sec_damping, coeff_shift,
708 block_height, /*enable_primary=*/1,
709 /*enable_secondary=*/0);
710 } else {
711 filter_block_4x4(/*is_lowbd=*/1, dest, dstride, in, pri_strength,
712 sec_strength, dir, pri_damping, sec_damping, coeff_shift,
713 block_height, /*enable_primary=*/1,
714 /*enable_secondary=*/0);
715 }
716 }
SIMD_FUNC(cdef_filter_8_2)717 void SIMD_FUNC(cdef_filter_8_2)(void *dest, int dstride, const uint16_t *in,
718 int pri_strength, int sec_strength, int dir,
719 int pri_damping, int sec_damping,
720 int coeff_shift, int block_width,
721 int block_height) {
722 if (block_width == 8) {
723 filter_block_8x8(/*is_lowbd=*/1, dest, dstride, in, pri_strength,
724 sec_strength, dir, pri_damping, sec_damping, coeff_shift,
725 block_height, /*enable_primary=*/0,
726 /*enable_secondary=*/1);
727 } else {
728 filter_block_4x4(/*is_lowbd=*/1, dest, dstride, in, pri_strength,
729 sec_strength, dir, pri_damping, sec_damping, coeff_shift,
730 block_height, /*enable_primary=*/0,
731 /*enable_secondary=*/1);
732 }
733 }
734
SIMD_FUNC(cdef_filter_8_3)735 void SIMD_FUNC(cdef_filter_8_3)(void *dest, int dstride, const uint16_t *in,
736 int pri_strength, int sec_strength, int dir,
737 int pri_damping, int sec_damping,
738 int coeff_shift, int block_width,
739 int block_height) {
740 (void)pri_strength;
741 (void)sec_strength;
742 (void)dir;
743 (void)pri_damping;
744 (void)sec_damping;
745 (void)coeff_shift;
746 (void)block_width;
747
748 if (block_width == 8) {
749 copy_block_8xh(/*is_lowbd=*/1, dest, dstride, in, block_height);
750 } else {
751 copy_block_4xh(/*is_lowbd=*/1, dest, dstride, in, block_height);
752 }
753 }
754
SIMD_FUNC(cdef_filter_16_0)755 void SIMD_FUNC(cdef_filter_16_0)(void *dest, int dstride, const uint16_t *in,
756 int pri_strength, int sec_strength, int dir,
757 int pri_damping, int sec_damping,
758 int coeff_shift, int block_width,
759 int block_height) {
760 if (block_width == 8) {
761 filter_block_8x8(/*is_lowbd=*/0, dest, dstride, in, pri_strength,
762 sec_strength, dir, pri_damping, sec_damping, coeff_shift,
763 block_height, /*enable_primary=*/1,
764 /*enable_secondary=*/1);
765 } else {
766 filter_block_4x4(/*is_lowbd=*/0, dest, dstride, in, pri_strength,
767 sec_strength, dir, pri_damping, sec_damping, coeff_shift,
768 block_height, /*enable_primary=*/1,
769 /*enable_secondary=*/1);
770 }
771 }
772
SIMD_FUNC(cdef_filter_16_1)773 void SIMD_FUNC(cdef_filter_16_1)(void *dest, int dstride, const uint16_t *in,
774 int pri_strength, int sec_strength, int dir,
775 int pri_damping, int sec_damping,
776 int coeff_shift, int block_width,
777 int block_height) {
778 if (block_width == 8) {
779 filter_block_8x8(/*is_lowbd=*/0, dest, dstride, in, pri_strength,
780 sec_strength, dir, pri_damping, sec_damping, coeff_shift,
781 block_height, /*enable_primary=*/1,
782 /*enable_secondary=*/0);
783 } else {
784 filter_block_4x4(/*is_lowbd=*/0, dest, dstride, in, pri_strength,
785 sec_strength, dir, pri_damping, sec_damping, coeff_shift,
786 block_height, /*enable_primary=*/1,
787 /*enable_secondary=*/0);
788 }
789 }
SIMD_FUNC(cdef_filter_16_2)790 void SIMD_FUNC(cdef_filter_16_2)(void *dest, int dstride, const uint16_t *in,
791 int pri_strength, int sec_strength, int dir,
792 int pri_damping, int sec_damping,
793 int coeff_shift, int block_width,
794 int block_height) {
795 if (block_width == 8) {
796 filter_block_8x8(/*is_lowbd=*/0, dest, dstride, in, pri_strength,
797 sec_strength, dir, pri_damping, sec_damping, coeff_shift,
798 block_height, /*enable_primary=*/0,
799 /*enable_secondary=*/1);
800 } else {
801 filter_block_4x4(/*is_lowbd=*/0, dest, dstride, in, pri_strength,
802 sec_strength, dir, pri_damping, sec_damping, coeff_shift,
803 block_height, /*enable_primary=*/0,
804 /*enable_secondary=*/1);
805 }
806 }
807
SIMD_FUNC(cdef_filter_16_3)808 void SIMD_FUNC(cdef_filter_16_3)(void *dest, int dstride, const uint16_t *in,
809 int pri_strength, int sec_strength, int dir,
810 int pri_damping, int sec_damping,
811 int coeff_shift, int block_width,
812 int block_height) {
813 (void)pri_strength;
814 (void)sec_strength;
815 (void)dir;
816 (void)pri_damping;
817 (void)sec_damping;
818 (void)coeff_shift;
819 (void)block_width;
820 if (block_width == 8) {
821 copy_block_8xh(/*is_lowbd=*/0, dest, dstride, in, block_height);
822 } else {
823 copy_block_4xh(/*is_lowbd=*/0, dest, dstride, in, block_height);
824 }
825 }
826
SIMD_FUNC(cdef_copy_rect8_16bit_to_16bit)827 void SIMD_FUNC(cdef_copy_rect8_16bit_to_16bit)(uint16_t *dst, int dstride,
828 const uint16_t *src, int sstride,
829 int width, int height) {
830 int i, j;
831 for (i = 0; i < height; i++) {
832 for (j = 0; j < (width & ~0x7); j += 8) {
833 v128 row = v128_load_unaligned(&src[i * sstride + j]);
834 v128_store_unaligned(&dst[i * dstride + j], row);
835 }
836 for (; j < width; j++) {
837 dst[i * dstride + j] = src[i * sstride + j];
838 }
839 }
840 }
841
842 #undef CDEF_INLINE
843
844 #endif // AOM_AV1_COMMON_CDEF_BLOCK_SIMD_H_
845