1 /*
2 * Copyright (c) 2018, 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 <arm_neon.h>
13 #include <assert.h>
14
15 #include "config/aom_config.h"
16 #include "config/av1_rtcd.h"
17
18 #include "aom_dsp/aom_dsp_common.h"
19 #include "aom_dsp/txfm_common.h"
20 #include "aom_mem/aom_mem.h"
21 #include "aom_ports/mem.h"
22 #include "av1/common/common.h"
23 #include "av1/common/onyxc_int.h"
24 #include "av1/common/resize.h"
25 #include "av1/common/restoration.h"
26 #include "av1/common/arm/mem_neon.h"
27 #include "av1/common/arm/transpose_neon.h"
28
29 // Constants used for right shift in final_filter calculation.
30 #define NB_EVEN 5
31 #define NB_ODD 4
32
calc_ab_fast_internal_common(uint32x4_t s0,uint32x4_t s1,uint32x4_t s2,uint32x4_t s3,uint32x4_t s4,uint32x4_t s5,uint32x4_t s6,uint32x4_t s7,int32x4_t sr4,int32x4_t sr5,int32x4_t sr6,int32x4_t sr7,uint32x4_t const_n_val,uint32x4_t s_vec,uint32x4_t const_val,uint32x4_t one_by_n_minus_1_vec,uint16x4_t sgrproj_sgr,int32_t * src1,uint16_t * dst_A16,int32_t * src2,const int buf_stride)33 static INLINE void calc_ab_fast_internal_common(
34 uint32x4_t s0, uint32x4_t s1, uint32x4_t s2, uint32x4_t s3, uint32x4_t s4,
35 uint32x4_t s5, uint32x4_t s6, uint32x4_t s7, int32x4_t sr4, int32x4_t sr5,
36 int32x4_t sr6, int32x4_t sr7, uint32x4_t const_n_val, uint32x4_t s_vec,
37 uint32x4_t const_val, uint32x4_t one_by_n_minus_1_vec,
38 uint16x4_t sgrproj_sgr, int32_t *src1, uint16_t *dst_A16, int32_t *src2,
39 const int buf_stride) {
40 uint32x4_t q0, q1, q2, q3;
41 uint32x4_t p0, p1, p2, p3;
42 uint16x4_t d0, d1, d2, d3;
43
44 s0 = vmulq_u32(s0, const_n_val);
45 s1 = vmulq_u32(s1, const_n_val);
46 s2 = vmulq_u32(s2, const_n_val);
47 s3 = vmulq_u32(s3, const_n_val);
48
49 q0 = vmulq_u32(s4, s4);
50 q1 = vmulq_u32(s5, s5);
51 q2 = vmulq_u32(s6, s6);
52 q3 = vmulq_u32(s7, s7);
53
54 p0 = vcleq_u32(q0, s0);
55 p1 = vcleq_u32(q1, s1);
56 p2 = vcleq_u32(q2, s2);
57 p3 = vcleq_u32(q3, s3);
58
59 q0 = vsubq_u32(s0, q0);
60 q1 = vsubq_u32(s1, q1);
61 q2 = vsubq_u32(s2, q2);
62 q3 = vsubq_u32(s3, q3);
63
64 p0 = vandq_u32(p0, q0);
65 p1 = vandq_u32(p1, q1);
66 p2 = vandq_u32(p2, q2);
67 p3 = vandq_u32(p3, q3);
68
69 p0 = vmulq_u32(p0, s_vec);
70 p1 = vmulq_u32(p1, s_vec);
71 p2 = vmulq_u32(p2, s_vec);
72 p3 = vmulq_u32(p3, s_vec);
73
74 p0 = vrshrq_n_u32(p0, SGRPROJ_MTABLE_BITS);
75 p1 = vrshrq_n_u32(p1, SGRPROJ_MTABLE_BITS);
76 p2 = vrshrq_n_u32(p2, SGRPROJ_MTABLE_BITS);
77 p3 = vrshrq_n_u32(p3, SGRPROJ_MTABLE_BITS);
78
79 p0 = vminq_u32(p0, const_val);
80 p1 = vminq_u32(p1, const_val);
81 p2 = vminq_u32(p2, const_val);
82 p3 = vminq_u32(p3, const_val);
83
84 {
85 store_u32_4x4((uint32_t *)src1, buf_stride, p0, p1, p2, p3);
86
87 for (int x = 0; x < 4; x++) {
88 for (int y = 0; y < 4; y++) {
89 dst_A16[x * buf_stride + y] = x_by_xplus1[src1[x * buf_stride + y]];
90 }
91 }
92 load_u16_4x4(dst_A16, buf_stride, &d0, &d1, &d2, &d3);
93 }
94 p0 = vsubl_u16(sgrproj_sgr, d0);
95 p1 = vsubl_u16(sgrproj_sgr, d1);
96 p2 = vsubl_u16(sgrproj_sgr, d2);
97 p3 = vsubl_u16(sgrproj_sgr, d3);
98
99 s4 = vmulq_u32(vreinterpretq_u32_s32(sr4), one_by_n_minus_1_vec);
100 s5 = vmulq_u32(vreinterpretq_u32_s32(sr5), one_by_n_minus_1_vec);
101 s6 = vmulq_u32(vreinterpretq_u32_s32(sr6), one_by_n_minus_1_vec);
102 s7 = vmulq_u32(vreinterpretq_u32_s32(sr7), one_by_n_minus_1_vec);
103
104 s4 = vmulq_u32(s4, p0);
105 s5 = vmulq_u32(s5, p1);
106 s6 = vmulq_u32(s6, p2);
107 s7 = vmulq_u32(s7, p3);
108
109 p0 = vrshrq_n_u32(s4, SGRPROJ_RECIP_BITS);
110 p1 = vrshrq_n_u32(s5, SGRPROJ_RECIP_BITS);
111 p2 = vrshrq_n_u32(s6, SGRPROJ_RECIP_BITS);
112 p3 = vrshrq_n_u32(s7, SGRPROJ_RECIP_BITS);
113
114 store_s32_4x4(src2, buf_stride, vreinterpretq_s32_u32(p0),
115 vreinterpretq_s32_u32(p1), vreinterpretq_s32_u32(p2),
116 vreinterpretq_s32_u32(p3));
117 }
calc_ab_internal_common(uint32x4_t s0,uint32x4_t s1,uint32x4_t s2,uint32x4_t s3,uint32x4_t s4,uint32x4_t s5,uint32x4_t s6,uint32x4_t s7,uint16x8_t s16_0,uint16x8_t s16_1,uint16x8_t s16_2,uint16x8_t s16_3,uint16x8_t s16_4,uint16x8_t s16_5,uint16x8_t s16_6,uint16x8_t s16_7,uint32x4_t const_n_val,uint32x4_t s_vec,uint32x4_t const_val,uint16x4_t one_by_n_minus_1_vec,uint16x8_t sgrproj_sgr,int32_t * src1,uint16_t * dst_A16,int32_t * dst2,const int buf_stride)118 static INLINE void calc_ab_internal_common(
119 uint32x4_t s0, uint32x4_t s1, uint32x4_t s2, uint32x4_t s3, uint32x4_t s4,
120 uint32x4_t s5, uint32x4_t s6, uint32x4_t s7, uint16x8_t s16_0,
121 uint16x8_t s16_1, uint16x8_t s16_2, uint16x8_t s16_3, uint16x8_t s16_4,
122 uint16x8_t s16_5, uint16x8_t s16_6, uint16x8_t s16_7,
123 uint32x4_t const_n_val, uint32x4_t s_vec, uint32x4_t const_val,
124 uint16x4_t one_by_n_minus_1_vec, uint16x8_t sgrproj_sgr, int32_t *src1,
125 uint16_t *dst_A16, int32_t *dst2, const int buf_stride) {
126 uint16x4_t d0, d1, d2, d3, d4, d5, d6, d7;
127 uint32x4_t q0, q1, q2, q3, q4, q5, q6, q7;
128 uint32x4_t p0, p1, p2, p3, p4, p5, p6, p7;
129
130 s0 = vmulq_u32(s0, const_n_val);
131 s1 = vmulq_u32(s1, const_n_val);
132 s2 = vmulq_u32(s2, const_n_val);
133 s3 = vmulq_u32(s3, const_n_val);
134 s4 = vmulq_u32(s4, const_n_val);
135 s5 = vmulq_u32(s5, const_n_val);
136 s6 = vmulq_u32(s6, const_n_val);
137 s7 = vmulq_u32(s7, const_n_val);
138
139 d0 = vget_low_u16(s16_4);
140 d1 = vget_low_u16(s16_5);
141 d2 = vget_low_u16(s16_6);
142 d3 = vget_low_u16(s16_7);
143 d4 = vget_high_u16(s16_4);
144 d5 = vget_high_u16(s16_5);
145 d6 = vget_high_u16(s16_6);
146 d7 = vget_high_u16(s16_7);
147
148 q0 = vmull_u16(d0, d0);
149 q1 = vmull_u16(d1, d1);
150 q2 = vmull_u16(d2, d2);
151 q3 = vmull_u16(d3, d3);
152 q4 = vmull_u16(d4, d4);
153 q5 = vmull_u16(d5, d5);
154 q6 = vmull_u16(d6, d6);
155 q7 = vmull_u16(d7, d7);
156
157 p0 = vcleq_u32(q0, s0);
158 p1 = vcleq_u32(q1, s1);
159 p2 = vcleq_u32(q2, s2);
160 p3 = vcleq_u32(q3, s3);
161 p4 = vcleq_u32(q4, s4);
162 p5 = vcleq_u32(q5, s5);
163 p6 = vcleq_u32(q6, s6);
164 p7 = vcleq_u32(q7, s7);
165
166 q0 = vsubq_u32(s0, q0);
167 q1 = vsubq_u32(s1, q1);
168 q2 = vsubq_u32(s2, q2);
169 q3 = vsubq_u32(s3, q3);
170 q4 = vsubq_u32(s4, q4);
171 q5 = vsubq_u32(s5, q5);
172 q6 = vsubq_u32(s6, q6);
173 q7 = vsubq_u32(s7, q7);
174
175 p0 = vandq_u32(p0, q0);
176 p1 = vandq_u32(p1, q1);
177 p2 = vandq_u32(p2, q2);
178 p3 = vandq_u32(p3, q3);
179 p4 = vandq_u32(p4, q4);
180 p5 = vandq_u32(p5, q5);
181 p6 = vandq_u32(p6, q6);
182 p7 = vandq_u32(p7, q7);
183
184 p0 = vmulq_u32(p0, s_vec);
185 p1 = vmulq_u32(p1, s_vec);
186 p2 = vmulq_u32(p2, s_vec);
187 p3 = vmulq_u32(p3, s_vec);
188 p4 = vmulq_u32(p4, s_vec);
189 p5 = vmulq_u32(p5, s_vec);
190 p6 = vmulq_u32(p6, s_vec);
191 p7 = vmulq_u32(p7, s_vec);
192
193 p0 = vrshrq_n_u32(p0, SGRPROJ_MTABLE_BITS);
194 p1 = vrshrq_n_u32(p1, SGRPROJ_MTABLE_BITS);
195 p2 = vrshrq_n_u32(p2, SGRPROJ_MTABLE_BITS);
196 p3 = vrshrq_n_u32(p3, SGRPROJ_MTABLE_BITS);
197 p4 = vrshrq_n_u32(p4, SGRPROJ_MTABLE_BITS);
198 p5 = vrshrq_n_u32(p5, SGRPROJ_MTABLE_BITS);
199 p6 = vrshrq_n_u32(p6, SGRPROJ_MTABLE_BITS);
200 p7 = vrshrq_n_u32(p7, SGRPROJ_MTABLE_BITS);
201
202 p0 = vminq_u32(p0, const_val);
203 p1 = vminq_u32(p1, const_val);
204 p2 = vminq_u32(p2, const_val);
205 p3 = vminq_u32(p3, const_val);
206 p4 = vminq_u32(p4, const_val);
207 p5 = vminq_u32(p5, const_val);
208 p6 = vminq_u32(p6, const_val);
209 p7 = vminq_u32(p7, const_val);
210
211 {
212 store_u32_4x4((uint32_t *)src1, buf_stride, p0, p1, p2, p3);
213 store_u32_4x4((uint32_t *)src1 + 4, buf_stride, p4, p5, p6, p7);
214
215 for (int x = 0; x < 4; x++) {
216 for (int y = 0; y < 8; y++) {
217 dst_A16[x * buf_stride + y] = x_by_xplus1[src1[x * buf_stride + y]];
218 }
219 }
220 load_u16_8x4(dst_A16, buf_stride, &s16_4, &s16_5, &s16_6, &s16_7);
221 }
222
223 s16_4 = vsubq_u16(sgrproj_sgr, s16_4);
224 s16_5 = vsubq_u16(sgrproj_sgr, s16_5);
225 s16_6 = vsubq_u16(sgrproj_sgr, s16_6);
226 s16_7 = vsubq_u16(sgrproj_sgr, s16_7);
227
228 s0 = vmull_u16(vget_low_u16(s16_0), one_by_n_minus_1_vec);
229 s1 = vmull_u16(vget_low_u16(s16_1), one_by_n_minus_1_vec);
230 s2 = vmull_u16(vget_low_u16(s16_2), one_by_n_minus_1_vec);
231 s3 = vmull_u16(vget_low_u16(s16_3), one_by_n_minus_1_vec);
232 s4 = vmull_u16(vget_high_u16(s16_0), one_by_n_minus_1_vec);
233 s5 = vmull_u16(vget_high_u16(s16_1), one_by_n_minus_1_vec);
234 s6 = vmull_u16(vget_high_u16(s16_2), one_by_n_minus_1_vec);
235 s7 = vmull_u16(vget_high_u16(s16_3), one_by_n_minus_1_vec);
236
237 s0 = vmulq_u32(s0, vmovl_u16(vget_low_u16(s16_4)));
238 s1 = vmulq_u32(s1, vmovl_u16(vget_low_u16(s16_5)));
239 s2 = vmulq_u32(s2, vmovl_u16(vget_low_u16(s16_6)));
240 s3 = vmulq_u32(s3, vmovl_u16(vget_low_u16(s16_7)));
241 s4 = vmulq_u32(s4, vmovl_u16(vget_high_u16(s16_4)));
242 s5 = vmulq_u32(s5, vmovl_u16(vget_high_u16(s16_5)));
243 s6 = vmulq_u32(s6, vmovl_u16(vget_high_u16(s16_6)));
244 s7 = vmulq_u32(s7, vmovl_u16(vget_high_u16(s16_7)));
245
246 p0 = vrshrq_n_u32(s0, SGRPROJ_RECIP_BITS);
247 p1 = vrshrq_n_u32(s1, SGRPROJ_RECIP_BITS);
248 p2 = vrshrq_n_u32(s2, SGRPROJ_RECIP_BITS);
249 p3 = vrshrq_n_u32(s3, SGRPROJ_RECIP_BITS);
250 p4 = vrshrq_n_u32(s4, SGRPROJ_RECIP_BITS);
251 p5 = vrshrq_n_u32(s5, SGRPROJ_RECIP_BITS);
252 p6 = vrshrq_n_u32(s6, SGRPROJ_RECIP_BITS);
253 p7 = vrshrq_n_u32(s7, SGRPROJ_RECIP_BITS);
254
255 store_s32_4x4(dst2, buf_stride, vreinterpretq_s32_u32(p0),
256 vreinterpretq_s32_u32(p1), vreinterpretq_s32_u32(p2),
257 vreinterpretq_s32_u32(p3));
258 store_s32_4x4(dst2 + 4, buf_stride, vreinterpretq_s32_u32(p4),
259 vreinterpretq_s32_u32(p5), vreinterpretq_s32_u32(p6),
260 vreinterpretq_s32_u32(p7));
261 }
262
boxsum2_square_sum_calc(int16x4_t t1,int16x4_t t2,int16x4_t t3,int16x4_t t4,int16x4_t t5,int16x4_t t6,int16x4_t t7,int16x4_t t8,int16x4_t t9,int16x4_t t10,int16x4_t t11,int32x4_t * r0,int32x4_t * r1,int32x4_t * r2,int32x4_t * r3)263 static INLINE void boxsum2_square_sum_calc(
264 int16x4_t t1, int16x4_t t2, int16x4_t t3, int16x4_t t4, int16x4_t t5,
265 int16x4_t t6, int16x4_t t7, int16x4_t t8, int16x4_t t9, int16x4_t t10,
266 int16x4_t t11, int32x4_t *r0, int32x4_t *r1, int32x4_t *r2, int32x4_t *r3) {
267 int32x4_t d1, d2, d3, d4, d5, d6, d7, d8, d9, d10, d11;
268 int32x4_t r12, r34, r67, r89, r1011;
269 int32x4_t r345, r6789, r789;
270
271 d1 = vmull_s16(t1, t1);
272 d2 = vmull_s16(t2, t2);
273 d3 = vmull_s16(t3, t3);
274 d4 = vmull_s16(t4, t4);
275 d5 = vmull_s16(t5, t5);
276 d6 = vmull_s16(t6, t6);
277 d7 = vmull_s16(t7, t7);
278 d8 = vmull_s16(t8, t8);
279 d9 = vmull_s16(t9, t9);
280 d10 = vmull_s16(t10, t10);
281 d11 = vmull_s16(t11, t11);
282
283 r12 = vaddq_s32(d1, d2);
284 r34 = vaddq_s32(d3, d4);
285 r67 = vaddq_s32(d6, d7);
286 r89 = vaddq_s32(d8, d9);
287 r1011 = vaddq_s32(d10, d11);
288 r345 = vaddq_s32(r34, d5);
289 r6789 = vaddq_s32(r67, r89);
290 r789 = vsubq_s32(r6789, d6);
291 *r0 = vaddq_s32(r12, r345);
292 *r1 = vaddq_s32(r67, r345);
293 *r2 = vaddq_s32(d5, r6789);
294 *r3 = vaddq_s32(r789, r1011);
295 }
296
boxsum2(int16_t * src,const int src_stride,int16_t * dst16,int32_t * dst32,int32_t * dst2,const int dst_stride,const int width,const int height)297 static INLINE void boxsum2(int16_t *src, const int src_stride, int16_t *dst16,
298 int32_t *dst32, int32_t *dst2, const int dst_stride,
299 const int width, const int height) {
300 assert(width > 2 * SGRPROJ_BORDER_HORZ);
301 assert(height > 2 * SGRPROJ_BORDER_VERT);
302
303 int16_t *dst1_16_ptr, *src_ptr;
304 int32_t *dst2_ptr;
305 int h, w, count = 0;
306 const int dst_stride_2 = (dst_stride << 1);
307 const int dst_stride_8 = (dst_stride << 3);
308
309 dst1_16_ptr = dst16;
310 dst2_ptr = dst2;
311 src_ptr = src;
312 w = width;
313 {
314 int16x8_t t1, t2, t3, t4, t5, t6, t7;
315 int16x8_t t8, t9, t10, t11, t12;
316
317 int16x8_t q12345, q56789, q34567, q7891011;
318 int16x8_t q12, q34, q67, q89, q1011;
319 int16x8_t q345, q6789, q789;
320
321 int32x4_t r12345, r56789, r34567, r7891011;
322
323 do {
324 h = height;
325 dst1_16_ptr = dst16 + (count << 3);
326 dst2_ptr = dst2 + (count << 3);
327 src_ptr = src + (count << 3);
328
329 dst1_16_ptr += dst_stride_2;
330 dst2_ptr += dst_stride_2;
331 do {
332 load_s16_8x4(src_ptr, src_stride, &t1, &t2, &t3, &t4);
333 src_ptr += 4 * src_stride;
334 load_s16_8x4(src_ptr, src_stride, &t5, &t6, &t7, &t8);
335 src_ptr += 4 * src_stride;
336 load_s16_8x4(src_ptr, src_stride, &t9, &t10, &t11, &t12);
337
338 q12 = vaddq_s16(t1, t2);
339 q34 = vaddq_s16(t3, t4);
340 q67 = vaddq_s16(t6, t7);
341 q89 = vaddq_s16(t8, t9);
342 q1011 = vaddq_s16(t10, t11);
343 q345 = vaddq_s16(q34, t5);
344 q6789 = vaddq_s16(q67, q89);
345 q789 = vaddq_s16(q89, t7);
346 q12345 = vaddq_s16(q12, q345);
347 q34567 = vaddq_s16(q67, q345);
348 q56789 = vaddq_s16(t5, q6789);
349 q7891011 = vaddq_s16(q789, q1011);
350
351 store_s16_8x4(dst1_16_ptr, dst_stride_2, q12345, q34567, q56789,
352 q7891011);
353 dst1_16_ptr += dst_stride_8;
354
355 boxsum2_square_sum_calc(
356 vget_low_s16(t1), vget_low_s16(t2), vget_low_s16(t3),
357 vget_low_s16(t4), vget_low_s16(t5), vget_low_s16(t6),
358 vget_low_s16(t7), vget_low_s16(t8), vget_low_s16(t9),
359 vget_low_s16(t10), vget_low_s16(t11), &r12345, &r34567, &r56789,
360 &r7891011);
361
362 store_s32_4x4(dst2_ptr, dst_stride_2, r12345, r34567, r56789, r7891011);
363
364 boxsum2_square_sum_calc(
365 vget_high_s16(t1), vget_high_s16(t2), vget_high_s16(t3),
366 vget_high_s16(t4), vget_high_s16(t5), vget_high_s16(t6),
367 vget_high_s16(t7), vget_high_s16(t8), vget_high_s16(t9),
368 vget_high_s16(t10), vget_high_s16(t11), &r12345, &r34567, &r56789,
369 &r7891011);
370
371 store_s32_4x4(dst2_ptr + 4, dst_stride_2, r12345, r34567, r56789,
372 r7891011);
373 dst2_ptr += (dst_stride_8);
374 h -= 8;
375 } while (h > 0);
376 w -= 8;
377 count++;
378 } while (w > 0);
379 }
380
381 {
382 int16x4_t s1, s2, s3, s4, s5, s6, s7, s8;
383 int32x4_t d1, d2, d3, d4, d5, d6, d7, d8;
384 int32x4_t q12345, q34567, q23456, q45678;
385 int32x4_t q23, q45, q67;
386 int32x4_t q2345, q4567;
387
388 int32x4_t r12345, r34567, r23456, r45678;
389 int32x4_t r23, r45, r67;
390 int32x4_t r2345, r4567;
391
392 int32_t *src2_ptr, *dst1_32_ptr;
393 int16_t *src1_ptr;
394 count = 0;
395 h = height;
396 do {
397 dst1_32_ptr = dst32 + count * dst_stride_8 + (dst_stride_2);
398 dst2_ptr = dst2 + count * dst_stride_8 + (dst_stride_2);
399 src1_ptr = dst16 + count * dst_stride_8 + (dst_stride_2);
400 src2_ptr = dst2 + count * dst_stride_8 + (dst_stride_2);
401 w = width;
402
403 dst1_32_ptr += 2;
404 dst2_ptr += 2;
405 load_s16_4x4(src1_ptr, dst_stride_2, &s1, &s2, &s3, &s4);
406 transpose_s16_4x4d(&s1, &s2, &s3, &s4);
407 load_s32_4x4(src2_ptr, dst_stride_2, &d1, &d2, &d3, &d4);
408 transpose_s32_4x4(&d1, &d2, &d3, &d4);
409 do {
410 src1_ptr += 4;
411 src2_ptr += 4;
412 load_s16_4x4(src1_ptr, dst_stride_2, &s5, &s6, &s7, &s8);
413 transpose_s16_4x4d(&s5, &s6, &s7, &s8);
414 load_s32_4x4(src2_ptr, dst_stride_2, &d5, &d6, &d7, &d8);
415 transpose_s32_4x4(&d5, &d6, &d7, &d8);
416 q23 = vaddl_s16(s2, s3);
417 q45 = vaddl_s16(s4, s5);
418 q67 = vaddl_s16(s6, s7);
419 q2345 = vaddq_s32(q23, q45);
420 q4567 = vaddq_s32(q45, q67);
421 q12345 = vaddq_s32(vmovl_s16(s1), q2345);
422 q23456 = vaddq_s32(q2345, vmovl_s16(s6));
423 q34567 = vaddq_s32(q4567, vmovl_s16(s3));
424 q45678 = vaddq_s32(q4567, vmovl_s16(s8));
425
426 transpose_s32_4x4(&q12345, &q23456, &q34567, &q45678);
427 store_s32_4x4(dst1_32_ptr, dst_stride_2, q12345, q23456, q34567,
428 q45678);
429 dst1_32_ptr += 4;
430 s1 = s5;
431 s2 = s6;
432 s3 = s7;
433 s4 = s8;
434
435 r23 = vaddq_s32(d2, d3);
436 r45 = vaddq_s32(d4, d5);
437 r67 = vaddq_s32(d6, d7);
438 r2345 = vaddq_s32(r23, r45);
439 r4567 = vaddq_s32(r45, r67);
440 r12345 = vaddq_s32(d1, r2345);
441 r23456 = vaddq_s32(r2345, d6);
442 r34567 = vaddq_s32(r4567, d3);
443 r45678 = vaddq_s32(r4567, d8);
444
445 transpose_s32_4x4(&r12345, &r23456, &r34567, &r45678);
446 store_s32_4x4(dst2_ptr, dst_stride_2, r12345, r23456, r34567, r45678);
447 dst2_ptr += 4;
448 d1 = d5;
449 d2 = d6;
450 d3 = d7;
451 d4 = d8;
452 w -= 4;
453 } while (w > 0);
454 h -= 8;
455 count++;
456 } while (h > 0);
457 }
458 }
459
calc_ab_internal_lbd(int32_t * A,uint16_t * A16,uint16_t * B16,int32_t * B,const int buf_stride,const int width,const int height,const int r,const int s,const int ht_inc)460 static INLINE void calc_ab_internal_lbd(int32_t *A, uint16_t *A16,
461 uint16_t *B16, int32_t *B,
462 const int buf_stride, const int width,
463 const int height, const int r,
464 const int s, const int ht_inc) {
465 int32_t *src1, *dst2, count = 0;
466 uint16_t *dst_A16, *src2;
467 const uint32_t n = (2 * r + 1) * (2 * r + 1);
468 const uint32x4_t const_n_val = vdupq_n_u32(n);
469 const uint16x8_t sgrproj_sgr = vdupq_n_u16(SGRPROJ_SGR);
470 const uint16x4_t one_by_n_minus_1_vec = vdup_n_u16(one_by_x[n - 1]);
471 const uint32x4_t const_val = vdupq_n_u32(255);
472
473 uint16x8_t s16_0, s16_1, s16_2, s16_3, s16_4, s16_5, s16_6, s16_7;
474
475 uint32x4_t s0, s1, s2, s3, s4, s5, s6, s7;
476
477 const uint32x4_t s_vec = vdupq_n_u32(s);
478 int w, h = height;
479
480 do {
481 dst_A16 = A16 + (count << 2) * buf_stride;
482 src1 = A + (count << 2) * buf_stride;
483 src2 = B16 + (count << 2) * buf_stride;
484 dst2 = B + (count << 2) * buf_stride;
485 w = width;
486 do {
487 load_u32_4x4((uint32_t *)src1, buf_stride, &s0, &s1, &s2, &s3);
488 load_u32_4x4((uint32_t *)src1 + 4, buf_stride, &s4, &s5, &s6, &s7);
489 load_u16_8x4(src2, buf_stride, &s16_0, &s16_1, &s16_2, &s16_3);
490
491 s16_4 = s16_0;
492 s16_5 = s16_1;
493 s16_6 = s16_2;
494 s16_7 = s16_3;
495
496 calc_ab_internal_common(
497 s0, s1, s2, s3, s4, s5, s6, s7, s16_0, s16_1, s16_2, s16_3, s16_4,
498 s16_5, s16_6, s16_7, const_n_val, s_vec, const_val,
499 one_by_n_minus_1_vec, sgrproj_sgr, src1, dst_A16, dst2, buf_stride);
500
501 w -= 8;
502 dst2 += 8;
503 src1 += 8;
504 src2 += 8;
505 dst_A16 += 8;
506 } while (w > 0);
507 count++;
508 h -= (ht_inc * 4);
509 } while (h > 0);
510 }
511
calc_ab_internal_hbd(int32_t * A,uint16_t * A16,uint16_t * B16,int32_t * B,const int buf_stride,const int width,const int height,const int bit_depth,const int r,const int s,const int ht_inc)512 static INLINE void calc_ab_internal_hbd(int32_t *A, uint16_t *A16,
513 uint16_t *B16, int32_t *B,
514 const int buf_stride, const int width,
515 const int height, const int bit_depth,
516 const int r, const int s,
517 const int ht_inc) {
518 int32_t *src1, *dst2, count = 0;
519 uint16_t *dst_A16, *src2;
520 const uint32_t n = (2 * r + 1) * (2 * r + 1);
521 const int16x8_t bd_min_2_vec = vdupq_n_s16(-(bit_depth - 8));
522 const int32x4_t bd_min_1_vec = vdupq_n_s32(-((bit_depth - 8) << 1));
523 const uint32x4_t const_n_val = vdupq_n_u32(n);
524 const uint16x8_t sgrproj_sgr = vdupq_n_u16(SGRPROJ_SGR);
525 const uint16x4_t one_by_n_minus_1_vec = vdup_n_u16(one_by_x[n - 1]);
526 const uint32x4_t const_val = vdupq_n_u32(255);
527
528 int32x4_t sr0, sr1, sr2, sr3, sr4, sr5, sr6, sr7;
529 uint16x8_t s16_0, s16_1, s16_2, s16_3;
530 uint16x8_t s16_4, s16_5, s16_6, s16_7;
531 uint32x4_t s0, s1, s2, s3, s4, s5, s6, s7;
532
533 const uint32x4_t s_vec = vdupq_n_u32(s);
534 int w, h = height;
535
536 do {
537 src1 = A + (count << 2) * buf_stride;
538 src2 = B16 + (count << 2) * buf_stride;
539 dst2 = B + (count << 2) * buf_stride;
540 dst_A16 = A16 + (count << 2) * buf_stride;
541 w = width;
542 do {
543 load_s32_4x4(src1, buf_stride, &sr0, &sr1, &sr2, &sr3);
544 load_s32_4x4(src1 + 4, buf_stride, &sr4, &sr5, &sr6, &sr7);
545 load_u16_8x4(src2, buf_stride, &s16_0, &s16_1, &s16_2, &s16_3);
546
547 s0 = vrshlq_u32(vreinterpretq_u32_s32(sr0), bd_min_1_vec);
548 s1 = vrshlq_u32(vreinterpretq_u32_s32(sr1), bd_min_1_vec);
549 s2 = vrshlq_u32(vreinterpretq_u32_s32(sr2), bd_min_1_vec);
550 s3 = vrshlq_u32(vreinterpretq_u32_s32(sr3), bd_min_1_vec);
551 s4 = vrshlq_u32(vreinterpretq_u32_s32(sr4), bd_min_1_vec);
552 s5 = vrshlq_u32(vreinterpretq_u32_s32(sr5), bd_min_1_vec);
553 s6 = vrshlq_u32(vreinterpretq_u32_s32(sr6), bd_min_1_vec);
554 s7 = vrshlq_u32(vreinterpretq_u32_s32(sr7), bd_min_1_vec);
555
556 s16_4 = vrshlq_u16(s16_0, bd_min_2_vec);
557 s16_5 = vrshlq_u16(s16_1, bd_min_2_vec);
558 s16_6 = vrshlq_u16(s16_2, bd_min_2_vec);
559 s16_7 = vrshlq_u16(s16_3, bd_min_2_vec);
560
561 calc_ab_internal_common(
562 s0, s1, s2, s3, s4, s5, s6, s7, s16_0, s16_1, s16_2, s16_3, s16_4,
563 s16_5, s16_6, s16_7, const_n_val, s_vec, const_val,
564 one_by_n_minus_1_vec, sgrproj_sgr, src1, dst_A16, dst2, buf_stride);
565
566 w -= 8;
567 dst2 += 8;
568 src1 += 8;
569 src2 += 8;
570 dst_A16 += 8;
571 } while (w > 0);
572 count++;
573 h -= (ht_inc * 4);
574 } while (h > 0);
575 }
576
calc_ab_fast_internal_lbd(int32_t * A,uint16_t * A16,int32_t * B,const int buf_stride,const int width,const int height,const int r,const int s,const int ht_inc)577 static INLINE void calc_ab_fast_internal_lbd(int32_t *A, uint16_t *A16,
578 int32_t *B, const int buf_stride,
579 const int width, const int height,
580 const int r, const int s,
581 const int ht_inc) {
582 int32_t *src1, *src2, count = 0;
583 uint16_t *dst_A16;
584 const uint32_t n = (2 * r + 1) * (2 * r + 1);
585 const uint32x4_t const_n_val = vdupq_n_u32(n);
586 const uint16x4_t sgrproj_sgr = vdup_n_u16(SGRPROJ_SGR);
587 const uint32x4_t one_by_n_minus_1_vec = vdupq_n_u32(one_by_x[n - 1]);
588 const uint32x4_t const_val = vdupq_n_u32(255);
589
590 int32x4_t sr0, sr1, sr2, sr3, sr4, sr5, sr6, sr7;
591 uint32x4_t s0, s1, s2, s3, s4, s5, s6, s7;
592
593 const uint32x4_t s_vec = vdupq_n_u32(s);
594 int w, h = height;
595
596 do {
597 src1 = A + (count << 2) * buf_stride;
598 src2 = B + (count << 2) * buf_stride;
599 dst_A16 = A16 + (count << 2) * buf_stride;
600 w = width;
601 do {
602 load_s32_4x4(src1, buf_stride, &sr0, &sr1, &sr2, &sr3);
603 load_s32_4x4(src2, buf_stride, &sr4, &sr5, &sr6, &sr7);
604
605 s0 = vreinterpretq_u32_s32(sr0);
606 s1 = vreinterpretq_u32_s32(sr1);
607 s2 = vreinterpretq_u32_s32(sr2);
608 s3 = vreinterpretq_u32_s32(sr3);
609 s4 = vreinterpretq_u32_s32(sr4);
610 s5 = vreinterpretq_u32_s32(sr5);
611 s6 = vreinterpretq_u32_s32(sr6);
612 s7 = vreinterpretq_u32_s32(sr7);
613
614 calc_ab_fast_internal_common(s0, s1, s2, s3, s4, s5, s6, s7, sr4, sr5,
615 sr6, sr7, const_n_val, s_vec, const_val,
616 one_by_n_minus_1_vec, sgrproj_sgr, src1,
617 dst_A16, src2, buf_stride);
618
619 w -= 4;
620 src1 += 4;
621 src2 += 4;
622 dst_A16 += 4;
623 } while (w > 0);
624 count++;
625 h -= (ht_inc * 4);
626 } while (h > 0);
627 }
628
calc_ab_fast_internal_hbd(int32_t * A,uint16_t * A16,int32_t * B,const int buf_stride,const int width,const int height,const int bit_depth,const int r,const int s,const int ht_inc)629 static INLINE void calc_ab_fast_internal_hbd(int32_t *A, uint16_t *A16,
630 int32_t *B, const int buf_stride,
631 const int width, const int height,
632 const int bit_depth, const int r,
633 const int s, const int ht_inc) {
634 int32_t *src1, *src2, count = 0;
635 uint16_t *dst_A16;
636 const uint32_t n = (2 * r + 1) * (2 * r + 1);
637 const int32x4_t bd_min_2_vec = vdupq_n_s32(-(bit_depth - 8));
638 const int32x4_t bd_min_1_vec = vdupq_n_s32(-((bit_depth - 8) << 1));
639 const uint32x4_t const_n_val = vdupq_n_u32(n);
640 const uint16x4_t sgrproj_sgr = vdup_n_u16(SGRPROJ_SGR);
641 const uint32x4_t one_by_n_minus_1_vec = vdupq_n_u32(one_by_x[n - 1]);
642 const uint32x4_t const_val = vdupq_n_u32(255);
643
644 int32x4_t sr0, sr1, sr2, sr3, sr4, sr5, sr6, sr7;
645 uint32x4_t s0, s1, s2, s3, s4, s5, s6, s7;
646
647 const uint32x4_t s_vec = vdupq_n_u32(s);
648 int w, h = height;
649
650 do {
651 src1 = A + (count << 2) * buf_stride;
652 src2 = B + (count << 2) * buf_stride;
653 dst_A16 = A16 + (count << 2) * buf_stride;
654 w = width;
655 do {
656 load_s32_4x4(src1, buf_stride, &sr0, &sr1, &sr2, &sr3);
657 load_s32_4x4(src2, buf_stride, &sr4, &sr5, &sr6, &sr7);
658
659 s0 = vrshlq_u32(vreinterpretq_u32_s32(sr0), bd_min_1_vec);
660 s1 = vrshlq_u32(vreinterpretq_u32_s32(sr1), bd_min_1_vec);
661 s2 = vrshlq_u32(vreinterpretq_u32_s32(sr2), bd_min_1_vec);
662 s3 = vrshlq_u32(vreinterpretq_u32_s32(sr3), bd_min_1_vec);
663 s4 = vrshlq_u32(vreinterpretq_u32_s32(sr4), bd_min_2_vec);
664 s5 = vrshlq_u32(vreinterpretq_u32_s32(sr5), bd_min_2_vec);
665 s6 = vrshlq_u32(vreinterpretq_u32_s32(sr6), bd_min_2_vec);
666 s7 = vrshlq_u32(vreinterpretq_u32_s32(sr7), bd_min_2_vec);
667
668 calc_ab_fast_internal_common(s0, s1, s2, s3, s4, s5, s6, s7, sr4, sr5,
669 sr6, sr7, const_n_val, s_vec, const_val,
670 one_by_n_minus_1_vec, sgrproj_sgr, src1,
671 dst_A16, src2, buf_stride);
672
673 w -= 4;
674 src1 += 4;
675 src2 += 4;
676 dst_A16 += 4;
677 } while (w > 0);
678 count++;
679 h -= (ht_inc * 4);
680 } while (h > 0);
681 }
682
boxsum1(int16_t * src,const int src_stride,uint16_t * dst1,int32_t * dst2,const int dst_stride,const int width,const int height)683 static INLINE void boxsum1(int16_t *src, const int src_stride, uint16_t *dst1,
684 int32_t *dst2, const int dst_stride, const int width,
685 const int height) {
686 assert(width > 2 * SGRPROJ_BORDER_HORZ);
687 assert(height > 2 * SGRPROJ_BORDER_VERT);
688
689 int16_t *src_ptr;
690 int32_t *dst2_ptr;
691 uint16_t *dst1_ptr;
692 int h, w, count = 0;
693
694 w = width;
695 {
696 int16x8_t s1, s2, s3, s4, s5, s6, s7, s8;
697 int16x8_t q23, q34, q56, q234, q345, q456, q567;
698 int32x4_t r23, r56, r345, r456, r567, r78, r678;
699 int32x4_t r4_low, r4_high, r34_low, r34_high, r234_low, r234_high;
700 int32x4_t r2, r3, r5, r6, r7, r8;
701 int16x8_t q678, q78;
702
703 do {
704 dst1_ptr = dst1 + (count << 3);
705 dst2_ptr = dst2 + (count << 3);
706 src_ptr = src + (count << 3);
707 h = height;
708
709 load_s16_8x4(src_ptr, src_stride, &s1, &s2, &s3, &s4);
710 src_ptr += 4 * src_stride;
711
712 q23 = vaddq_s16(s2, s3);
713 q234 = vaddq_s16(q23, s4);
714 q34 = vaddq_s16(s3, s4);
715 dst1_ptr += (dst_stride << 1);
716
717 r2 = vmull_s16(vget_low_s16(s2), vget_low_s16(s2));
718 r3 = vmull_s16(vget_low_s16(s3), vget_low_s16(s3));
719 r4_low = vmull_s16(vget_low_s16(s4), vget_low_s16(s4));
720 r23 = vaddq_s32(r2, r3);
721 r234_low = vaddq_s32(r23, r4_low);
722 r34_low = vaddq_s32(r3, r4_low);
723
724 r2 = vmull_s16(vget_high_s16(s2), vget_high_s16(s2));
725 r3 = vmull_s16(vget_high_s16(s3), vget_high_s16(s3));
726 r4_high = vmull_s16(vget_high_s16(s4), vget_high_s16(s4));
727 r23 = vaddq_s32(r2, r3);
728 r234_high = vaddq_s32(r23, r4_high);
729 r34_high = vaddq_s32(r3, r4_high);
730
731 dst2_ptr += (dst_stride << 1);
732
733 do {
734 load_s16_8x4(src_ptr, src_stride, &s5, &s6, &s7, &s8);
735 src_ptr += 4 * src_stride;
736
737 q345 = vaddq_s16(s5, q34);
738 q56 = vaddq_s16(s5, s6);
739 q456 = vaddq_s16(s4, q56);
740 q567 = vaddq_s16(s7, q56);
741 q78 = vaddq_s16(s7, s8);
742 q678 = vaddq_s16(s6, q78);
743
744 store_s16_8x4((int16_t *)dst1_ptr, dst_stride, q234, q345, q456, q567);
745 dst1_ptr += (dst_stride << 2);
746
747 s4 = s8;
748 q34 = q78;
749 q234 = q678;
750
751 r5 = vmull_s16(vget_low_s16(s5), vget_low_s16(s5));
752 r6 = vmull_s16(vget_low_s16(s6), vget_low_s16(s6));
753 r7 = vmull_s16(vget_low_s16(s7), vget_low_s16(s7));
754 r8 = vmull_s16(vget_low_s16(s8), vget_low_s16(s8));
755
756 r345 = vaddq_s32(r5, r34_low);
757 r56 = vaddq_s32(r5, r6);
758 r456 = vaddq_s32(r4_low, r56);
759 r567 = vaddq_s32(r7, r56);
760 r78 = vaddq_s32(r7, r8);
761 r678 = vaddq_s32(r6, r78);
762 store_s32_4x4(dst2_ptr, dst_stride, r234_low, r345, r456, r567);
763
764 r4_low = r8;
765 r34_low = r78;
766 r234_low = r678;
767
768 r5 = vmull_s16(vget_high_s16(s5), vget_high_s16(s5));
769 r6 = vmull_s16(vget_high_s16(s6), vget_high_s16(s6));
770 r7 = vmull_s16(vget_high_s16(s7), vget_high_s16(s7));
771 r8 = vmull_s16(vget_high_s16(s8), vget_high_s16(s8));
772
773 r345 = vaddq_s32(r5, r34_high);
774 r56 = vaddq_s32(r5, r6);
775 r456 = vaddq_s32(r4_high, r56);
776 r567 = vaddq_s32(r7, r56);
777 r78 = vaddq_s32(r7, r8);
778 r678 = vaddq_s32(r6, r78);
779 store_s32_4x4((dst2_ptr + 4), dst_stride, r234_high, r345, r456, r567);
780 dst2_ptr += (dst_stride << 2);
781
782 r4_high = r8;
783 r34_high = r78;
784 r234_high = r678;
785
786 h -= 4;
787 } while (h > 0);
788 w -= 8;
789 count++;
790 } while (w > 0);
791 }
792
793 {
794 int16x4_t d1, d2, d3, d4, d5, d6, d7, d8;
795 int16x4_t q23, q34, q56, q234, q345, q456, q567;
796 int32x4_t r23, r56, r234, r345, r456, r567, r34, r78, r678;
797 int32x4_t r1, r2, r3, r4, r5, r6, r7, r8;
798 int16x4_t q678, q78;
799
800 int32_t *src2_ptr;
801 uint16_t *src1_ptr;
802 count = 0;
803 h = height;
804 w = width;
805 do {
806 dst1_ptr = dst1 + (count << 2) * dst_stride;
807 dst2_ptr = dst2 + (count << 2) * dst_stride;
808 src1_ptr = dst1 + (count << 2) * dst_stride;
809 src2_ptr = dst2 + (count << 2) * dst_stride;
810 w = width;
811
812 load_s16_4x4((int16_t *)src1_ptr, dst_stride, &d1, &d2, &d3, &d4);
813 transpose_s16_4x4d(&d1, &d2, &d3, &d4);
814 load_s32_4x4(src2_ptr, dst_stride, &r1, &r2, &r3, &r4);
815 transpose_s32_4x4(&r1, &r2, &r3, &r4);
816 src1_ptr += 4;
817 src2_ptr += 4;
818
819 q23 = vadd_s16(d2, d3);
820 q234 = vadd_s16(q23, d4);
821 q34 = vadd_s16(d3, d4);
822 dst1_ptr += 2;
823 r23 = vaddq_s32(r2, r3);
824 r234 = vaddq_s32(r23, r4);
825 r34 = vaddq_s32(r3, r4);
826 dst2_ptr += 2;
827
828 do {
829 load_s16_4x4((int16_t *)src1_ptr, dst_stride, &d5, &d6, &d7, &d8);
830 transpose_s16_4x4d(&d5, &d6, &d7, &d8);
831 load_s32_4x4(src2_ptr, dst_stride, &r5, &r6, &r7, &r8);
832 transpose_s32_4x4(&r5, &r6, &r7, &r8);
833 src1_ptr += 4;
834 src2_ptr += 4;
835
836 q345 = vadd_s16(d5, q34);
837 q56 = vadd_s16(d5, d6);
838 q456 = vadd_s16(d4, q56);
839 q567 = vadd_s16(d7, q56);
840 q78 = vadd_s16(d7, d8);
841 q678 = vadd_s16(d6, q78);
842 transpose_s16_4x4d(&q234, &q345, &q456, &q567);
843 store_s16_4x4((int16_t *)dst1_ptr, dst_stride, q234, q345, q456, q567);
844 dst1_ptr += 4;
845
846 d4 = d8;
847 q34 = q78;
848 q234 = q678;
849
850 r345 = vaddq_s32(r5, r34);
851 r56 = vaddq_s32(r5, r6);
852 r456 = vaddq_s32(r4, r56);
853 r567 = vaddq_s32(r7, r56);
854 r78 = vaddq_s32(r7, r8);
855 r678 = vaddq_s32(r6, r78);
856 transpose_s32_4x4(&r234, &r345, &r456, &r567);
857 store_s32_4x4(dst2_ptr, dst_stride, r234, r345, r456, r567);
858 dst2_ptr += 4;
859
860 r4 = r8;
861 r34 = r78;
862 r234 = r678;
863 w -= 4;
864 } while (w > 0);
865 h -= 4;
866 count++;
867 } while (h > 0);
868 }
869 }
870
cross_sum_inp_s32(int32_t * buf,int buf_stride)871 static INLINE int32x4_t cross_sum_inp_s32(int32_t *buf, int buf_stride) {
872 int32x4_t xtr, xt, xtl, xl, x, xr, xbr, xb, xbl;
873 int32x4_t fours, threes, res;
874
875 xtl = vld1q_s32(buf - buf_stride - 1);
876 xt = vld1q_s32(buf - buf_stride);
877 xtr = vld1q_s32(buf - buf_stride + 1);
878 xl = vld1q_s32(buf - 1);
879 x = vld1q_s32(buf);
880 xr = vld1q_s32(buf + 1);
881 xbl = vld1q_s32(buf + buf_stride - 1);
882 xb = vld1q_s32(buf + buf_stride);
883 xbr = vld1q_s32(buf + buf_stride + 1);
884
885 fours = vaddq_s32(xl, vaddq_s32(xt, vaddq_s32(xr, vaddq_s32(xb, x))));
886 threes = vaddq_s32(xtl, vaddq_s32(xtr, vaddq_s32(xbr, xbl)));
887 res = vsubq_s32(vshlq_n_s32(vaddq_s32(fours, threes), 2), threes);
888 return res;
889 }
890
cross_sum_inp_u16(uint16_t * buf,int buf_stride,int32x4_t * a0,int32x4_t * a1)891 static INLINE void cross_sum_inp_u16(uint16_t *buf, int buf_stride,
892 int32x4_t *a0, int32x4_t *a1) {
893 uint16x8_t xtr, xt, xtl, xl, x, xr, xbr, xb, xbl;
894 uint16x8_t r0, r1;
895
896 xtl = vld1q_u16(buf - buf_stride - 1);
897 xt = vld1q_u16(buf - buf_stride);
898 xtr = vld1q_u16(buf - buf_stride + 1);
899 xl = vld1q_u16(buf - 1);
900 x = vld1q_u16(buf);
901 xr = vld1q_u16(buf + 1);
902 xbl = vld1q_u16(buf + buf_stride - 1);
903 xb = vld1q_u16(buf + buf_stride);
904 xbr = vld1q_u16(buf + buf_stride + 1);
905
906 xb = vaddq_u16(xb, x);
907 xt = vaddq_u16(xt, xr);
908 xl = vaddq_u16(xl, xb);
909 xl = vaddq_u16(xl, xt);
910
911 r0 = vshlq_n_u16(xl, 2);
912
913 xbl = vaddq_u16(xbl, xbr);
914 xtl = vaddq_u16(xtl, xtr);
915 xtl = vaddq_u16(xtl, xbl);
916
917 r1 = vshlq_n_u16(xtl, 2);
918 r1 = vsubq_u16(r1, xtl);
919
920 *a0 = vreinterpretq_s32_u32(
921 vaddq_u32(vmovl_u16(vget_low_u16(r0)), vmovl_u16(vget_low_u16(r1))));
922 *a1 = vreinterpretq_s32_u32(
923 vaddq_u32(vmovl_u16(vget_high_u16(r0)), vmovl_u16(vget_high_u16(r1))));
924 }
925
cross_sum_fast_even_row(int32_t * buf,int buf_stride)926 static INLINE int32x4_t cross_sum_fast_even_row(int32_t *buf, int buf_stride) {
927 int32x4_t xtr, xt, xtl, xbr, xb, xbl;
928 int32x4_t fives, sixes, fives_plus_sixes;
929
930 xtl = vld1q_s32(buf - buf_stride - 1);
931 xt = vld1q_s32(buf - buf_stride);
932 xtr = vld1q_s32(buf - buf_stride + 1);
933 xbl = vld1q_s32(buf + buf_stride - 1);
934 xb = vld1q_s32(buf + buf_stride);
935 xbr = vld1q_s32(buf + buf_stride + 1);
936
937 fives = vaddq_s32(xtl, vaddq_s32(xtr, vaddq_s32(xbr, xbl)));
938 sixes = vaddq_s32(xt, xb);
939 fives_plus_sixes = vaddq_s32(fives, sixes);
940
941 return vaddq_s32(
942 vaddq_s32(vshlq_n_s32(fives_plus_sixes, 2), fives_plus_sixes), sixes);
943 }
944
cross_sum_fast_even_row_inp16(uint16_t * buf,int buf_stride,int32x4_t * a0,int32x4_t * a1)945 static INLINE void cross_sum_fast_even_row_inp16(uint16_t *buf, int buf_stride,
946 int32x4_t *a0, int32x4_t *a1) {
947 uint16x8_t xtr, xt, xtl, xbr, xb, xbl, xb0;
948
949 xtl = vld1q_u16(buf - buf_stride - 1);
950 xt = vld1q_u16(buf - buf_stride);
951 xtr = vld1q_u16(buf - buf_stride + 1);
952 xbl = vld1q_u16(buf + buf_stride - 1);
953 xb = vld1q_u16(buf + buf_stride);
954 xbr = vld1q_u16(buf + buf_stride + 1);
955
956 xbr = vaddq_u16(xbr, xbl);
957 xtr = vaddq_u16(xtr, xtl);
958 xbr = vaddq_u16(xbr, xtr);
959 xtl = vshlq_n_u16(xbr, 2);
960 xbr = vaddq_u16(xtl, xbr);
961
962 xb = vaddq_u16(xb, xt);
963 xb0 = vshlq_n_u16(xb, 1);
964 xb = vshlq_n_u16(xb, 2);
965 xb = vaddq_u16(xb, xb0);
966
967 *a0 = vreinterpretq_s32_u32(
968 vaddq_u32(vmovl_u16(vget_low_u16(xbr)), vmovl_u16(vget_low_u16(xb))));
969 *a1 = vreinterpretq_s32_u32(
970 vaddq_u32(vmovl_u16(vget_high_u16(xbr)), vmovl_u16(vget_high_u16(xb))));
971 }
972
cross_sum_fast_odd_row(int32_t * buf)973 static INLINE int32x4_t cross_sum_fast_odd_row(int32_t *buf) {
974 int32x4_t xl, x, xr;
975 int32x4_t fives, sixes, fives_plus_sixes;
976
977 xl = vld1q_s32(buf - 1);
978 x = vld1q_s32(buf);
979 xr = vld1q_s32(buf + 1);
980 fives = vaddq_s32(xl, xr);
981 sixes = x;
982 fives_plus_sixes = vaddq_s32(fives, sixes);
983
984 return vaddq_s32(
985 vaddq_s32(vshlq_n_s32(fives_plus_sixes, 2), fives_plus_sixes), sixes);
986 }
987
cross_sum_fast_odd_row_inp16(uint16_t * buf,int32x4_t * a0,int32x4_t * a1)988 static INLINE void cross_sum_fast_odd_row_inp16(uint16_t *buf, int32x4_t *a0,
989 int32x4_t *a1) {
990 uint16x8_t xl, x, xr;
991 uint16x8_t x0;
992
993 xl = vld1q_u16(buf - 1);
994 x = vld1q_u16(buf);
995 xr = vld1q_u16(buf + 1);
996 xl = vaddq_u16(xl, xr);
997 x0 = vshlq_n_u16(xl, 2);
998 xl = vaddq_u16(xl, x0);
999
1000 x0 = vshlq_n_u16(x, 1);
1001 x = vshlq_n_u16(x, 2);
1002 x = vaddq_u16(x, x0);
1003
1004 *a0 = vreinterpretq_s32_u32(
1005 vaddq_u32(vmovl_u16(vget_low_u16(xl)), vmovl_u16(vget_low_u16(x))));
1006 *a1 = vreinterpretq_s32_u32(
1007 vaddq_u32(vmovl_u16(vget_high_u16(xl)), vmovl_u16(vget_high_u16(x))));
1008 }
1009
final_filter_fast_internal(uint16_t * A,int32_t * B,const int buf_stride,int16_t * src,const int src_stride,int32_t * dst,const int dst_stride,const int width,const int height)1010 static void final_filter_fast_internal(uint16_t *A, int32_t *B,
1011 const int buf_stride, int16_t *src,
1012 const int src_stride, int32_t *dst,
1013 const int dst_stride, const int width,
1014 const int height) {
1015 int16x8_t s0;
1016 int32_t *B_tmp, *dst_ptr;
1017 uint16_t *A_tmp;
1018 int16_t *src_ptr;
1019 int32x4_t a_res0, a_res1, b_res0, b_res1;
1020 int w, h, count = 0;
1021 assert(SGRPROJ_SGR_BITS == 8);
1022 assert(SGRPROJ_RST_BITS == 4);
1023
1024 A_tmp = A;
1025 B_tmp = B;
1026 src_ptr = src;
1027 dst_ptr = dst;
1028 h = height;
1029 do {
1030 A_tmp = (A + count * buf_stride);
1031 B_tmp = (B + count * buf_stride);
1032 src_ptr = (src + count * src_stride);
1033 dst_ptr = (dst + count * dst_stride);
1034 w = width;
1035 if (!(count & 1)) {
1036 do {
1037 s0 = vld1q_s16(src_ptr);
1038 cross_sum_fast_even_row_inp16(A_tmp, buf_stride, &a_res0, &a_res1);
1039 a_res0 = vmulq_s32(vmovl_s16(vget_low_s16(s0)), a_res0);
1040 a_res1 = vmulq_s32(vmovl_s16(vget_high_s16(s0)), a_res1);
1041
1042 b_res0 = cross_sum_fast_even_row(B_tmp, buf_stride);
1043 b_res1 = cross_sum_fast_even_row(B_tmp + 4, buf_stride);
1044 a_res0 = vaddq_s32(a_res0, b_res0);
1045 a_res1 = vaddq_s32(a_res1, b_res1);
1046
1047 a_res0 =
1048 vrshrq_n_s32(a_res0, SGRPROJ_SGR_BITS + NB_EVEN - SGRPROJ_RST_BITS);
1049 a_res1 =
1050 vrshrq_n_s32(a_res1, SGRPROJ_SGR_BITS + NB_EVEN - SGRPROJ_RST_BITS);
1051
1052 vst1q_s32(dst_ptr, a_res0);
1053 vst1q_s32(dst_ptr + 4, a_res1);
1054
1055 A_tmp += 8;
1056 B_tmp += 8;
1057 src_ptr += 8;
1058 dst_ptr += 8;
1059 w -= 8;
1060 } while (w > 0);
1061 } else {
1062 do {
1063 s0 = vld1q_s16(src_ptr);
1064 cross_sum_fast_odd_row_inp16(A_tmp, &a_res0, &a_res1);
1065 a_res0 = vmulq_s32(vmovl_s16(vget_low_s16(s0)), a_res0);
1066 a_res1 = vmulq_s32(vmovl_s16(vget_high_s16(s0)), a_res1);
1067
1068 b_res0 = cross_sum_fast_odd_row(B_tmp);
1069 b_res1 = cross_sum_fast_odd_row(B_tmp + 4);
1070 a_res0 = vaddq_s32(a_res0, b_res0);
1071 a_res1 = vaddq_s32(a_res1, b_res1);
1072
1073 a_res0 =
1074 vrshrq_n_s32(a_res0, SGRPROJ_SGR_BITS + NB_ODD - SGRPROJ_RST_BITS);
1075 a_res1 =
1076 vrshrq_n_s32(a_res1, SGRPROJ_SGR_BITS + NB_ODD - SGRPROJ_RST_BITS);
1077
1078 vst1q_s32(dst_ptr, a_res0);
1079 vst1q_s32(dst_ptr + 4, a_res1);
1080
1081 A_tmp += 8;
1082 B_tmp += 8;
1083 src_ptr += 8;
1084 dst_ptr += 8;
1085 w -= 8;
1086 } while (w > 0);
1087 }
1088 count++;
1089 h -= 1;
1090 } while (h > 0);
1091 }
1092
final_filter_internal(uint16_t * A,int32_t * B,const int buf_stride,int16_t * src,const int src_stride,int32_t * dst,const int dst_stride,const int width,const int height)1093 void final_filter_internal(uint16_t *A, int32_t *B, const int buf_stride,
1094 int16_t *src, const int src_stride, int32_t *dst,
1095 const int dst_stride, const int width,
1096 const int height) {
1097 int16x8_t s0;
1098 int32_t *B_tmp, *dst_ptr;
1099 uint16_t *A_tmp;
1100 int16_t *src_ptr;
1101 int32x4_t a_res0, a_res1, b_res0, b_res1;
1102 int w, h, count = 0;
1103
1104 assert(SGRPROJ_SGR_BITS == 8);
1105 assert(SGRPROJ_RST_BITS == 4);
1106 h = height;
1107
1108 do {
1109 A_tmp = (A + count * buf_stride);
1110 B_tmp = (B + count * buf_stride);
1111 src_ptr = (src + count * src_stride);
1112 dst_ptr = (dst + count * dst_stride);
1113 w = width;
1114 do {
1115 s0 = vld1q_s16(src_ptr);
1116 cross_sum_inp_u16(A_tmp, buf_stride, &a_res0, &a_res1);
1117 a_res0 = vmulq_s32(vmovl_s16(vget_low_s16(s0)), a_res0);
1118 a_res1 = vmulq_s32(vmovl_s16(vget_high_s16(s0)), a_res1);
1119
1120 b_res0 = cross_sum_inp_s32(B_tmp, buf_stride);
1121 b_res1 = cross_sum_inp_s32(B_tmp + 4, buf_stride);
1122 a_res0 = vaddq_s32(a_res0, b_res0);
1123 a_res1 = vaddq_s32(a_res1, b_res1);
1124
1125 a_res0 =
1126 vrshrq_n_s32(a_res0, SGRPROJ_SGR_BITS + NB_EVEN - SGRPROJ_RST_BITS);
1127 a_res1 =
1128 vrshrq_n_s32(a_res1, SGRPROJ_SGR_BITS + NB_EVEN - SGRPROJ_RST_BITS);
1129 vst1q_s32(dst_ptr, a_res0);
1130 vst1q_s32(dst_ptr + 4, a_res1);
1131
1132 A_tmp += 8;
1133 B_tmp += 8;
1134 src_ptr += 8;
1135 dst_ptr += 8;
1136 w -= 8;
1137 } while (w > 0);
1138 count++;
1139 h -= 1;
1140 } while (h > 0);
1141 }
1142
restoration_fast_internal(uint16_t * dgd16,int width,int height,int dgd_stride,int32_t * dst,int dst_stride,int bit_depth,int sgr_params_idx,int radius_idx)1143 static INLINE void restoration_fast_internal(uint16_t *dgd16, int width,
1144 int height, int dgd_stride,
1145 int32_t *dst, int dst_stride,
1146 int bit_depth, int sgr_params_idx,
1147 int radius_idx) {
1148 const sgr_params_type *const params = &sgr_params[sgr_params_idx];
1149 const int r = params->r[radius_idx];
1150 const int width_ext = width + 2 * SGRPROJ_BORDER_HORZ;
1151 const int height_ext = height + 2 * SGRPROJ_BORDER_VERT;
1152
1153 const int buf_stride = ((width_ext + 3) & ~3) + 16;
1154 int32_t A_[RESTORATION_PROC_UNIT_PELS];
1155 uint16_t A16_[RESTORATION_PROC_UNIT_PELS];
1156 int32_t B_[RESTORATION_PROC_UNIT_PELS];
1157 int32_t *square_sum_buf = A_;
1158 int32_t *sum_buf = B_;
1159 uint16_t *tmp16_buf = A16_;
1160
1161 assert(r <= MAX_RADIUS && "Need MAX_RADIUS >= r");
1162 assert(r <= SGRPROJ_BORDER_VERT - 1 && r <= SGRPROJ_BORDER_HORZ - 1 &&
1163 "Need SGRPROJ_BORDER_* >= r+1");
1164
1165 assert(radius_idx == 0);
1166 assert(r == 2);
1167
1168 // input(dgd16) is 16bit.
1169 // sum of pixels 1st stage output will be in 16bit(tmp16_buf). End output is
1170 // kept in 32bit [sum_buf]. sum of squares output is kept in 32bit
1171 // buffer(square_sum_buf).
1172 boxsum2((int16_t *)(dgd16 - dgd_stride * SGRPROJ_BORDER_VERT -
1173 SGRPROJ_BORDER_HORZ),
1174 dgd_stride, (int16_t *)tmp16_buf, sum_buf, square_sum_buf, buf_stride,
1175 width_ext, height_ext);
1176
1177 square_sum_buf += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
1178 sum_buf += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
1179 tmp16_buf += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
1180
1181 // Calculation of a, b. a output is in 16bit tmp_buf which is in range of
1182 // [1, 256] for all bit depths. b output is kept in 32bit buffer.
1183
1184 if (8 == bit_depth) {
1185 calc_ab_fast_internal_lbd(
1186 (square_sum_buf - buf_stride - 1), (tmp16_buf - buf_stride - 1),
1187 (sum_buf - buf_stride - 1), buf_stride * 2, width + 2, height + 2, r,
1188 params->s[radius_idx], 2);
1189 } else {
1190 calc_ab_fast_internal_hbd(
1191 (square_sum_buf - buf_stride - 1), (tmp16_buf - buf_stride - 1),
1192 (sum_buf - buf_stride - 1), buf_stride * 2, width + 2, height + 2,
1193 bit_depth, r, params->s[radius_idx], 2);
1194 }
1195 final_filter_fast_internal(tmp16_buf, sum_buf, buf_stride, (int16_t *)dgd16,
1196 dgd_stride, dst, dst_stride, width, height);
1197 }
1198
restoration_internal(uint16_t * dgd16,int width,int height,int dgd_stride,int32_t * dst,int dst_stride,int bit_depth,int sgr_params_idx,int radius_idx)1199 static INLINE void restoration_internal(uint16_t *dgd16, int width, int height,
1200 int dgd_stride, int32_t *dst,
1201 int dst_stride, int bit_depth,
1202 int sgr_params_idx, int radius_idx) {
1203 const sgr_params_type *const params = &sgr_params[sgr_params_idx];
1204 const int r = params->r[radius_idx];
1205 const int width_ext = width + 2 * SGRPROJ_BORDER_HORZ;
1206 const int height_ext = height + 2 * SGRPROJ_BORDER_VERT;
1207
1208 int buf_stride = ((width_ext + 3) & ~3) + 16;
1209 int32_t A_[RESTORATION_PROC_UNIT_PELS];
1210 uint16_t A16_[RESTORATION_PROC_UNIT_PELS];
1211 uint16_t B16_[RESTORATION_PROC_UNIT_PELS];
1212 int32_t B_[RESTORATION_PROC_UNIT_PELS];
1213 int32_t *square_sum_buf = A_;
1214 uint16_t *sum_buf = B16_;
1215 uint16_t *A16 = A16_;
1216 int32_t *B = B_;
1217
1218 assert(r <= MAX_RADIUS && "Need MAX_RADIUS >= r");
1219 assert(r <= SGRPROJ_BORDER_VERT - 1 && r <= SGRPROJ_BORDER_HORZ - 1 &&
1220 "Need SGRPROJ_BORDER_* >= r+1");
1221
1222 assert(radius_idx == 1);
1223 assert(r == 1);
1224
1225 // input(dgd16) is 16bit.
1226 // sum of pixels output will be in 16bit(sum_buf).
1227 // sum of squares output is kept in 32bit buffer(square_sum_buf).
1228 boxsum1((int16_t *)(dgd16 - dgd_stride * SGRPROJ_BORDER_VERT -
1229 SGRPROJ_BORDER_HORZ),
1230 dgd_stride, sum_buf, square_sum_buf, buf_stride, width_ext,
1231 height_ext);
1232
1233 square_sum_buf += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
1234 B += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
1235 A16 += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
1236 sum_buf += SGRPROJ_BORDER_VERT * buf_stride + SGRPROJ_BORDER_HORZ;
1237
1238 // Calculation of a, b. a output is in 16bit tmp_buf which is in range of
1239 // [1, 256] for all bit depths. b output is kept in 32bit buffer.
1240 if (8 == bit_depth) {
1241 calc_ab_internal_lbd((square_sum_buf - buf_stride - 1),
1242 (A16 - buf_stride - 1), (sum_buf - buf_stride - 1),
1243 (B - buf_stride - 1), buf_stride, width + 2,
1244 height + 2, r, params->s[radius_idx], 1);
1245 } else {
1246 calc_ab_internal_hbd((square_sum_buf - buf_stride - 1),
1247 (A16 - buf_stride - 1), (sum_buf - buf_stride - 1),
1248 (B - buf_stride - 1), buf_stride, width + 2,
1249 height + 2, bit_depth, r, params->s[radius_idx], 1);
1250 }
1251 final_filter_internal(A16, B, buf_stride, (int16_t *)dgd16, dgd_stride, dst,
1252 dst_stride, width, height);
1253 }
1254
src_convert_u8_to_u16(const uint8_t * src,const int src_stride,uint16_t * dst,const int dst_stride,const int width,const int height)1255 static INLINE void src_convert_u8_to_u16(const uint8_t *src,
1256 const int src_stride, uint16_t *dst,
1257 const int dst_stride, const int width,
1258 const int height) {
1259 const uint8_t *src_ptr;
1260 uint16_t *dst_ptr;
1261 int h, w, count = 0;
1262
1263 uint8x8_t t1, t2, t3, t4;
1264 uint16x8_t s1, s2, s3, s4;
1265 h = height;
1266 do {
1267 src_ptr = src + (count << 2) * src_stride;
1268 dst_ptr = dst + (count << 2) * dst_stride;
1269 w = width;
1270 if (w >= 7) {
1271 do {
1272 load_u8_8x4(src_ptr, src_stride, &t1, &t2, &t3, &t4);
1273 s1 = vmovl_u8(t1);
1274 s2 = vmovl_u8(t2);
1275 s3 = vmovl_u8(t3);
1276 s4 = vmovl_u8(t4);
1277 store_u16_8x4(dst_ptr, dst_stride, s1, s2, s3, s4);
1278
1279 src_ptr += 8;
1280 dst_ptr += 8;
1281 w -= 8;
1282 } while (w > 7);
1283 }
1284
1285 for (int y = 0; y < w; y++) {
1286 dst_ptr[y] = src_ptr[y];
1287 dst_ptr[y + 1 * dst_stride] = src_ptr[y + 1 * src_stride];
1288 dst_ptr[y + 2 * dst_stride] = src_ptr[y + 2 * src_stride];
1289 dst_ptr[y + 3 * dst_stride] = src_ptr[y + 3 * src_stride];
1290 }
1291 count++;
1292 h -= 4;
1293 } while (h > 3);
1294
1295 src_ptr = src + (count << 2) * src_stride;
1296 dst_ptr = dst + (count << 2) * dst_stride;
1297 for (int x = 0; x < h; x++) {
1298 for (int y = 0; y < width; y++) {
1299 dst_ptr[y + x * dst_stride] = src_ptr[y + x * src_stride];
1300 }
1301 }
1302 }
1303
src_convert_hbd_copy(const uint16_t * src,int src_stride,uint16_t * dst,const int dst_stride,int width,int height)1304 static INLINE void src_convert_hbd_copy(const uint16_t *src, int src_stride,
1305 uint16_t *dst, const int dst_stride,
1306 int width, int height) {
1307 const uint16_t *src_ptr;
1308 uint16_t *dst_ptr;
1309 int h, w, count = 0;
1310 uint16x8_t s1, s2, s3, s4;
1311
1312 h = height;
1313 do {
1314 src_ptr = src + (count << 2) * src_stride;
1315 dst_ptr = dst + (count << 2) * dst_stride;
1316 w = width;
1317 do {
1318 load_u16_8x4(src_ptr, src_stride, &s1, &s2, &s3, &s4);
1319 store_u16_8x4(dst_ptr, dst_stride, s1, s2, s3, s4);
1320 src_ptr += 8;
1321 dst_ptr += 8;
1322 w -= 8;
1323 } while (w > 7);
1324
1325 for (int y = 0; y < w; y++) {
1326 dst_ptr[y] = src_ptr[y];
1327 dst_ptr[y + 1 * dst_stride] = src_ptr[y + 1 * src_stride];
1328 dst_ptr[y + 2 * dst_stride] = src_ptr[y + 2 * src_stride];
1329 dst_ptr[y + 3 * dst_stride] = src_ptr[y + 3 * src_stride];
1330 }
1331 count++;
1332 h -= 4;
1333 } while (h > 3);
1334
1335 src_ptr = src + (count << 2) * src_stride;
1336 dst_ptr = dst + (count << 2) * dst_stride;
1337
1338 for (int x = 0; x < h; x++) {
1339 memcpy((dst_ptr + x * dst_stride), (src_ptr + x * src_stride),
1340 sizeof(uint16_t) * width);
1341 }
1342 }
1343
av1_selfguided_restoration_neon(const uint8_t * dat8,int width,int height,int stride,int32_t * flt0,int32_t * flt1,int flt_stride,int sgr_params_idx,int bit_depth,int highbd)1344 int av1_selfguided_restoration_neon(const uint8_t *dat8, int width, int height,
1345 int stride, int32_t *flt0, int32_t *flt1,
1346 int flt_stride, int sgr_params_idx,
1347 int bit_depth, int highbd) {
1348 const sgr_params_type *const params = &sgr_params[sgr_params_idx];
1349 assert(!(params->r[0] == 0 && params->r[1] == 0));
1350
1351 uint16_t dgd16_[RESTORATION_PROC_UNIT_PELS];
1352 const int dgd16_stride = width + 2 * SGRPROJ_BORDER_HORZ;
1353 uint16_t *dgd16 =
1354 dgd16_ + dgd16_stride * SGRPROJ_BORDER_VERT + SGRPROJ_BORDER_HORZ;
1355 const int width_ext = width + 2 * SGRPROJ_BORDER_HORZ;
1356 const int height_ext = height + 2 * SGRPROJ_BORDER_VERT;
1357 const int dgd_stride = stride;
1358
1359 if (highbd) {
1360 const uint16_t *dgd16_tmp = CONVERT_TO_SHORTPTR(dat8);
1361 src_convert_hbd_copy(
1362 dgd16_tmp - SGRPROJ_BORDER_VERT * dgd_stride - SGRPROJ_BORDER_HORZ,
1363 dgd_stride,
1364 dgd16 - SGRPROJ_BORDER_VERT * dgd16_stride - SGRPROJ_BORDER_HORZ,
1365 dgd16_stride, width_ext, height_ext);
1366 } else {
1367 src_convert_u8_to_u16(
1368 dat8 - SGRPROJ_BORDER_VERT * dgd_stride - SGRPROJ_BORDER_HORZ,
1369 dgd_stride,
1370 dgd16 - SGRPROJ_BORDER_VERT * dgd16_stride - SGRPROJ_BORDER_HORZ,
1371 dgd16_stride, width_ext, height_ext);
1372 }
1373
1374 if (params->r[0] > 0)
1375 restoration_fast_internal(dgd16, width, height, dgd16_stride, flt0,
1376 flt_stride, bit_depth, sgr_params_idx, 0);
1377 if (params->r[1] > 0)
1378 restoration_internal(dgd16, width, height, dgd16_stride, flt1, flt_stride,
1379 bit_depth, sgr_params_idx, 1);
1380 return 0;
1381 }
1382
apply_selfguided_restoration_neon(const uint8_t * dat8,int width,int height,int stride,int eps,const int * xqd,uint8_t * dst8,int dst_stride,int32_t * tmpbuf,int bit_depth,int highbd)1383 void apply_selfguided_restoration_neon(const uint8_t *dat8, int width,
1384 int height, int stride, int eps,
1385 const int *xqd, uint8_t *dst8,
1386 int dst_stride, int32_t *tmpbuf,
1387 int bit_depth, int highbd) {
1388 int32_t *flt0 = tmpbuf;
1389 int32_t *flt1 = flt0 + RESTORATION_UNITPELS_MAX;
1390 assert(width * height <= RESTORATION_UNITPELS_MAX);
1391 uint16_t dgd16_[RESTORATION_PROC_UNIT_PELS];
1392 const int dgd16_stride = width + 2 * SGRPROJ_BORDER_HORZ;
1393 uint16_t *dgd16 =
1394 dgd16_ + dgd16_stride * SGRPROJ_BORDER_VERT + SGRPROJ_BORDER_HORZ;
1395 const int width_ext = width + 2 * SGRPROJ_BORDER_HORZ;
1396 const int height_ext = height + 2 * SGRPROJ_BORDER_VERT;
1397 const int dgd_stride = stride;
1398 const sgr_params_type *const params = &sgr_params[eps];
1399 int xq[2];
1400
1401 assert(!(params->r[0] == 0 && params->r[1] == 0));
1402
1403 if (highbd) {
1404 const uint16_t *dgd16_tmp = CONVERT_TO_SHORTPTR(dat8);
1405 src_convert_hbd_copy(
1406 dgd16_tmp - SGRPROJ_BORDER_VERT * dgd_stride - SGRPROJ_BORDER_HORZ,
1407 dgd_stride,
1408 dgd16 - SGRPROJ_BORDER_VERT * dgd16_stride - SGRPROJ_BORDER_HORZ,
1409 dgd16_stride, width_ext, height_ext);
1410 } else {
1411 src_convert_u8_to_u16(
1412 dat8 - SGRPROJ_BORDER_VERT * dgd_stride - SGRPROJ_BORDER_HORZ,
1413 dgd_stride,
1414 dgd16 - SGRPROJ_BORDER_VERT * dgd16_stride - SGRPROJ_BORDER_HORZ,
1415 dgd16_stride, width_ext, height_ext);
1416 }
1417
1418 if (params->r[0] > 0)
1419 restoration_fast_internal(dgd16, width, height, dgd16_stride, flt0, width,
1420 bit_depth, eps, 0);
1421 if (params->r[1] > 0)
1422 restoration_internal(dgd16, width, height, dgd16_stride, flt1, width,
1423 bit_depth, eps, 1);
1424
1425 decode_xq(xqd, xq, params);
1426
1427 {
1428 int16_t *src_ptr;
1429 uint8_t *dst_ptr;
1430 uint16_t *dst16_ptr;
1431 int16x4_t d0, d4;
1432 int16x8_t r0, s0;
1433 uint16x8_t r4;
1434 int32x4_t u0, u4, v0, v4, f00, f10;
1435 uint8x8_t t0;
1436 int count = 0, w = width, h = height, rc = 0;
1437
1438 const int32x4_t xq0_vec = vdupq_n_s32(xq[0]);
1439 const int32x4_t xq1_vec = vdupq_n_s32(xq[1]);
1440 const int16x8_t zero = vdupq_n_s16(0);
1441 const uint16x8_t max = vdupq_n_u16((1 << bit_depth) - 1);
1442 uint16_t *dst16 = CONVERT_TO_SHORTPTR(dst8);
1443 dst_ptr = dst8;
1444 src_ptr = (int16_t *)dgd16;
1445 do {
1446 w = width;
1447 count = 0;
1448 dst_ptr = dst8 + rc * dst_stride;
1449 dst16_ptr = dst16 + rc * dst_stride;
1450 do {
1451 s0 = vld1q_s16(src_ptr + count);
1452
1453 u0 = vshll_n_s16(vget_low_s16(s0), SGRPROJ_RST_BITS);
1454 u4 = vshll_n_s16(vget_high_s16(s0), SGRPROJ_RST_BITS);
1455
1456 v0 = vshlq_n_s32(u0, SGRPROJ_PRJ_BITS);
1457 v4 = vshlq_n_s32(u4, SGRPROJ_PRJ_BITS);
1458
1459 if (params->r[0] > 0) {
1460 f00 = vld1q_s32(flt0 + count);
1461 f10 = vld1q_s32(flt0 + count + 4);
1462
1463 f00 = vsubq_s32(f00, u0);
1464 f10 = vsubq_s32(f10, u4);
1465
1466 v0 = vmlaq_s32(v0, xq0_vec, f00);
1467 v4 = vmlaq_s32(v4, xq0_vec, f10);
1468 }
1469
1470 if (params->r[1] > 0) {
1471 f00 = vld1q_s32(flt1 + count);
1472 f10 = vld1q_s32(flt1 + count + 4);
1473
1474 f00 = vsubq_s32(f00, u0);
1475 f10 = vsubq_s32(f10, u4);
1476
1477 v0 = vmlaq_s32(v0, xq1_vec, f00);
1478 v4 = vmlaq_s32(v4, xq1_vec, f10);
1479 }
1480
1481 d0 = vqrshrn_n_s32(v0, SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
1482 d4 = vqrshrn_n_s32(v4, SGRPROJ_PRJ_BITS + SGRPROJ_RST_BITS);
1483
1484 r0 = vcombine_s16(d0, d4);
1485
1486 r4 = vreinterpretq_u16_s16(vmaxq_s16(r0, zero));
1487
1488 if (highbd) {
1489 r4 = vminq_u16(r4, max);
1490 vst1q_u16(dst16_ptr, r4);
1491 } else {
1492 t0 = vqmovn_u16(r4);
1493 vst1_u8(dst_ptr, t0);
1494 }
1495 w -= 8;
1496 count += 8;
1497 dst_ptr += 8;
1498 dst16_ptr += 8;
1499 } while (w > 0);
1500
1501 src_ptr += dgd16_stride;
1502 flt1 += width;
1503 flt0 += width;
1504 rc++;
1505 h--;
1506 } while (h > 0);
1507 }
1508 }
1509