1 /*
2 *
3 * Copyright (c) 2020, Alliance for Open Media. All rights reserved
4 *
5 * This source code is subject to the terms of the BSD 2 Clause License and
6 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
7 * was not distributed with this source code in the LICENSE file, you can
8 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
9 * Media Patent License 1.0 was not distributed with this source code in the
10 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
11 */
12 #include <arm_neon.h>
13 #include <assert.h>
14
15 #include "aom_dsp/arm/mem_neon.h"
16 #include "aom_dsp/arm/transpose_neon.h"
17 #include "av1/common/resize.h"
18 #include "av1/common/arm/convolve_neon.h"
19 #include "config/av1_rtcd.h"
20 #include "config/aom_scale_rtcd.h"
21
scale_plane_2_to_1_phase_0(const uint8_t * src,const int src_stride,uint8_t * dst,const int dst_stride,const int w,const int h)22 static INLINE void scale_plane_2_to_1_phase_0(const uint8_t *src,
23 const int src_stride,
24 uint8_t *dst,
25 const int dst_stride, const int w,
26 const int h) {
27 const int max_width = (w + 15) & ~15;
28 int y = h;
29
30 assert(w && h);
31
32 do {
33 int x = max_width;
34 do {
35 const uint8x16x2_t s = vld2q_u8(src);
36 vst1q_u8(dst, s.val[0]);
37 src += 32;
38 dst += 16;
39 x -= 16;
40 } while (x);
41 src += 2 * (src_stride - max_width);
42 dst += dst_stride - max_width;
43 } while (--y);
44 }
45
scale_plane_4_to_1_phase_0(const uint8_t * src,const int src_stride,uint8_t * dst,const int dst_stride,const int w,const int h)46 static INLINE void scale_plane_4_to_1_phase_0(const uint8_t *src,
47 const int src_stride,
48 uint8_t *dst,
49 const int dst_stride, const int w,
50 const int h) {
51 const int max_width = (w + 15) & ~15;
52 int y = h;
53
54 assert(w && h);
55
56 do {
57 int x = max_width;
58 do {
59 const uint8x16x4_t s = vld4q_u8(src);
60 vst1q_u8(dst, s.val[0]);
61 src += 64;
62 dst += 16;
63 x -= 16;
64 } while (x);
65 src += 4 * (src_stride - max_width);
66 dst += dst_stride - max_width;
67 } while (--y);
68 }
69
scale_plane_bilinear_kernel(const uint8x16_t in0,const uint8x16_t in1,const uint8x16_t in2,const uint8x16_t in3,const uint8x8_t coef0,const uint8x8_t coef1,uint8_t * const dst)70 static INLINE void scale_plane_bilinear_kernel(
71 const uint8x16_t in0, const uint8x16_t in1, const uint8x16_t in2,
72 const uint8x16_t in3, const uint8x8_t coef0, const uint8x8_t coef1,
73 uint8_t *const dst) {
74 const uint16x8_t h0 = vmull_u8(vget_low_u8(in0), coef0);
75 const uint16x8_t h1 = vmull_u8(vget_high_u8(in0), coef0);
76 const uint16x8_t h2 = vmull_u8(vget_low_u8(in2), coef0);
77 const uint16x8_t h3 = vmull_u8(vget_high_u8(in2), coef0);
78 const uint16x8_t h4 = vmlal_u8(h0, vget_low_u8(in1), coef1);
79 const uint16x8_t h5 = vmlal_u8(h1, vget_high_u8(in1), coef1);
80 const uint16x8_t h6 = vmlal_u8(h2, vget_low_u8(in3), coef1);
81 const uint16x8_t h7 = vmlal_u8(h3, vget_high_u8(in3), coef1);
82
83 const uint8x8_t hor0 = vrshrn_n_u16(h4, 7); // temp: 00 01 02 03 04 05 06 07
84 const uint8x8_t hor1 = vrshrn_n_u16(h5, 7); // temp: 08 09 0A 0B 0C 0D 0E 0F
85 const uint8x8_t hor2 = vrshrn_n_u16(h6, 7); // temp: 10 11 12 13 14 15 16 17
86 const uint8x8_t hor3 = vrshrn_n_u16(h7, 7); // temp: 18 19 1A 1B 1C 1D 1E 1F
87 const uint16x8_t v0 = vmull_u8(hor0, coef0);
88 const uint16x8_t v1 = vmull_u8(hor1, coef0);
89 const uint16x8_t v2 = vmlal_u8(v0, hor2, coef1);
90 const uint16x8_t v3 = vmlal_u8(v1, hor3, coef1);
91 // dst: 0 1 2 3 4 5 6 7 8 9 A B C D E F
92 const uint8x16_t d = vcombine_u8(vrshrn_n_u16(v2, 7), vrshrn_n_u16(v3, 7));
93 vst1q_u8(dst, d);
94 }
95
scale_plane_2_to_1_bilinear(const uint8_t * const src,const int src_stride,uint8_t * dst,const int dst_stride,const int w,const int h,const int16_t c0,const int16_t c1)96 static INLINE void scale_plane_2_to_1_bilinear(
97 const uint8_t *const src, const int src_stride, uint8_t *dst,
98 const int dst_stride, const int w, const int h, const int16_t c0,
99 const int16_t c1) {
100 const int max_width = (w + 15) & ~15;
101 const uint8_t *src0 = src;
102 const uint8_t *src1 = src + src_stride;
103 const uint8x8_t coef0 = vdup_n_u8(c0);
104 const uint8x8_t coef1 = vdup_n_u8(c1);
105 int y = h;
106
107 assert(w && h);
108
109 do {
110 int x = max_width;
111 do {
112 // 000 002 004 006 008 00A 00C 00E 010 012 014 016 018 01A 01C 01E
113 // 001 003 005 007 009 00B 00D 00F 011 013 015 017 019 01B 01D 01F
114 const uint8x16x2_t s0 = vld2q_u8(src0);
115 // 100 102 104 106 108 10A 10C 10E 110 112 114 116 118 11A 11C 11E
116 // 101 103 105 107 109 10B 10D 10F 111 113 115 117 119 11B 11D 11F
117 const uint8x16x2_t s1 = vld2q_u8(src1);
118 scale_plane_bilinear_kernel(s0.val[0], s0.val[1], s1.val[0], s1.val[1],
119 coef0, coef1, dst);
120 src0 += 32;
121 src1 += 32;
122 dst += 16;
123 x -= 16;
124 } while (x);
125 src0 += 2 * (src_stride - max_width);
126 src1 += 2 * (src_stride - max_width);
127 dst += dst_stride - max_width;
128 } while (--y);
129 }
130
scale_plane_4_to_1_bilinear(const uint8_t * const src,const int src_stride,uint8_t * dst,const int dst_stride,const int w,const int h,const int16_t c0,const int16_t c1)131 static INLINE void scale_plane_4_to_1_bilinear(
132 const uint8_t *const src, const int src_stride, uint8_t *dst,
133 const int dst_stride, const int w, const int h, const int16_t c0,
134 const int16_t c1) {
135 const int max_width = (w + 15) & ~15;
136 const uint8_t *src0 = src;
137 const uint8_t *src1 = src + src_stride;
138 const uint8x8_t coef0 = vdup_n_u8(c0);
139 const uint8x8_t coef1 = vdup_n_u8(c1);
140 int y = h;
141
142 assert(w && h);
143
144 do {
145 int x = max_width;
146 do {
147 // (*) -- useless
148 // 000 004 008 00C 010 014 018 01C 020 024 028 02C 030 034 038 03C
149 // 001 005 009 00D 011 015 019 01D 021 025 029 02D 031 035 039 03D
150 // 002 006 00A 00E 012 016 01A 01E 022 026 02A 02E 032 036 03A 03E (*)
151 // 003 007 00B 00F 013 017 01B 01F 023 027 02B 02F 033 037 03B 03F (*)
152 const uint8x16x4_t s0 = vld4q_u8(src0);
153 // 100 104 108 10C 110 114 118 11C 120 124 128 12C 130 134 138 13C
154 // 101 105 109 10D 111 115 119 11D 121 125 129 12D 131 135 139 13D
155 // 102 106 10A 10E 112 116 11A 11E 122 126 12A 12E 132 136 13A 13E (*)
156 // 103 107 10B 10F 113 117 11B 11F 123 127 12B 12F 133 137 13B 13F (*)
157 const uint8x16x4_t s1 = vld4q_u8(src1);
158 scale_plane_bilinear_kernel(s0.val[0], s0.val[1], s1.val[0], s1.val[1],
159 coef0, coef1, dst);
160 src0 += 64;
161 src1 += 64;
162 dst += 16;
163 x -= 16;
164 } while (x);
165 src0 += 4 * (src_stride - max_width);
166 src1 += 4 * (src_stride - max_width);
167 dst += dst_stride - max_width;
168 } while (--y);
169 }
170
scale_plane_2_to_1_general(const uint8_t * src,const int src_stride,uint8_t * dst,const int dst_stride,const int w,const int h,const int16_t * const coef,uint8_t * const temp_buffer)171 static void scale_plane_2_to_1_general(const uint8_t *src, const int src_stride,
172 uint8_t *dst, const int dst_stride,
173 const int w, const int h,
174 const int16_t *const coef,
175 uint8_t *const temp_buffer) {
176 const int width_hor = (w + 3) & ~3;
177 const int width_ver = (w + 7) & ~7;
178 const int height_hor = (2 * h + SUBPEL_TAPS - 2 + 7) & ~7;
179 const int height_ver = (h + 3) & ~3;
180 const int16x8_t filters = vld1q_s16(coef);
181 int x, y = height_hor;
182 uint8_t *t = temp_buffer;
183 uint8x8_t s[14], d[4];
184
185 assert(w && h);
186
187 src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2 + 1;
188
189 // horizontal 4x8
190 // Note: processing 4x8 is about 20% faster than processing row by row using
191 // vld4_u8().
192 do {
193 load_u8_8x8(src + 2, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
194 &s[6], &s[7]);
195 transpose_u8_8x8(&s[0], &s[1], &s[2], &s[3], &s[4], &s[5], &s[6], &s[7]);
196 x = width_hor;
197
198 do {
199 src += 8;
200 load_u8_8x8(src, src_stride, &s[6], &s[7], &s[8], &s[9], &s[10], &s[11],
201 &s[12], &s[13]);
202 transpose_u8_8x8(&s[6], &s[7], &s[8], &s[9], &s[10], &s[11], &s[12],
203 &s[13]);
204
205 d[0] = scale_filter_8(&s[0], filters); // 00 10 20 30 40 50 60 70
206 d[1] = scale_filter_8(&s[2], filters); // 01 11 21 31 41 51 61 71
207 d[2] = scale_filter_8(&s[4], filters); // 02 12 22 32 42 52 62 72
208 d[3] = scale_filter_8(&s[6], filters); // 03 13 23 33 43 53 63 73
209 // 00 01 02 03 40 41 42 43
210 // 10 11 12 13 50 51 52 53
211 // 20 21 22 23 60 61 62 63
212 // 30 31 32 33 70 71 72 73
213 transpose_u8_8x4(&d[0], &d[1], &d[2], &d[3]);
214 vst1_lane_u32((uint32_t *)(t + 0 * width_hor), vreinterpret_u32_u8(d[0]),
215 0);
216 vst1_lane_u32((uint32_t *)(t + 1 * width_hor), vreinterpret_u32_u8(d[1]),
217 0);
218 vst1_lane_u32((uint32_t *)(t + 2 * width_hor), vreinterpret_u32_u8(d[2]),
219 0);
220 vst1_lane_u32((uint32_t *)(t + 3 * width_hor), vreinterpret_u32_u8(d[3]),
221 0);
222 vst1_lane_u32((uint32_t *)(t + 4 * width_hor), vreinterpret_u32_u8(d[0]),
223 1);
224 vst1_lane_u32((uint32_t *)(t + 5 * width_hor), vreinterpret_u32_u8(d[1]),
225 1);
226 vst1_lane_u32((uint32_t *)(t + 6 * width_hor), vreinterpret_u32_u8(d[2]),
227 1);
228 vst1_lane_u32((uint32_t *)(t + 7 * width_hor), vreinterpret_u32_u8(d[3]),
229 1);
230
231 s[0] = s[8];
232 s[1] = s[9];
233 s[2] = s[10];
234 s[3] = s[11];
235 s[4] = s[12];
236 s[5] = s[13];
237
238 t += 4;
239 x -= 4;
240 } while (x);
241 src += 8 * src_stride - 2 * width_hor;
242 t += 7 * width_hor;
243 y -= 8;
244 } while (y);
245
246 // vertical 8x4
247 x = width_ver;
248 t = temp_buffer;
249 do {
250 load_u8_8x8(t, width_hor, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5], &s[6],
251 &s[7]);
252 t += 6 * width_hor;
253 y = height_ver;
254
255 do {
256 load_u8_8x8(t, width_hor, &s[6], &s[7], &s[8], &s[9], &s[10], &s[11],
257 &s[12], &s[13]);
258 t += 8 * width_hor;
259
260 d[0] = scale_filter_8(&s[0], filters); // 00 01 02 03 04 05 06 07
261 d[1] = scale_filter_8(&s[2], filters); // 10 11 12 13 14 15 16 17
262 d[2] = scale_filter_8(&s[4], filters); // 20 21 22 23 24 25 26 27
263 d[3] = scale_filter_8(&s[6], filters); // 30 31 32 33 34 35 36 37
264 vst1_u8(dst + 0 * dst_stride, d[0]);
265 vst1_u8(dst + 1 * dst_stride, d[1]);
266 vst1_u8(dst + 2 * dst_stride, d[2]);
267 vst1_u8(dst + 3 * dst_stride, d[3]);
268
269 s[0] = s[8];
270 s[1] = s[9];
271 s[2] = s[10];
272 s[3] = s[11];
273 s[4] = s[12];
274 s[5] = s[13];
275
276 dst += 4 * dst_stride;
277 y -= 4;
278 } while (y);
279 t -= width_hor * (2 * height_ver + 6);
280 t += 8;
281 dst -= height_ver * dst_stride;
282 dst += 8;
283 x -= 8;
284 } while (x);
285 }
286
scale_plane_4_to_1_general(const uint8_t * src,const int src_stride,uint8_t * dst,const int dst_stride,const int w,const int h,const int16_t * const coef,uint8_t * const temp_buffer)287 static void scale_plane_4_to_1_general(const uint8_t *src, const int src_stride,
288 uint8_t *dst, const int dst_stride,
289 const int w, const int h,
290 const int16_t *const coef,
291 uint8_t *const temp_buffer) {
292 const int width_hor = (w + 1) & ~1;
293 const int width_ver = (w + 7) & ~7;
294 const int height_hor = (4 * h + SUBPEL_TAPS - 2 + 7) & ~7;
295 const int height_ver = (h + 1) & ~1;
296 const int16x8_t filters = vld1q_s16(coef);
297 int x, y = height_hor;
298 uint8_t *t = temp_buffer;
299 uint8x8_t s[12], d[2];
300
301 assert(w && h);
302
303 src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2 + 3;
304
305 // horizontal 2x8
306 // Note: processing 2x8 is about 20% faster than processing row by row using
307 // vld4_u8().
308 do {
309 load_u8_8x8(src + 4, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
310 &s[6], &s[7]);
311 transpose_u8_4x8(&s[0], &s[1], &s[2], &s[3], s[4], s[5], s[6], s[7]);
312 x = width_hor;
313
314 do {
315 uint8x8x2_t dd;
316 src += 8;
317 load_u8_8x8(src, src_stride, &s[4], &s[5], &s[6], &s[7], &s[8], &s[9],
318 &s[10], &s[11]);
319 transpose_u8_8x8(&s[4], &s[5], &s[6], &s[7], &s[8], &s[9], &s[10],
320 &s[11]);
321
322 d[0] = scale_filter_8(&s[0], filters); // 00 10 20 30 40 50 60 70
323 d[1] = scale_filter_8(&s[4], filters); // 01 11 21 31 41 51 61 71
324 // dd.val[0]: 00 01 20 21 40 41 60 61
325 // dd.val[1]: 10 11 30 31 50 51 70 71
326 dd = vtrn_u8(d[0], d[1]);
327 vst1_lane_u16((uint16_t *)(t + 0 * width_hor),
328 vreinterpret_u16_u8(dd.val[0]), 0);
329 vst1_lane_u16((uint16_t *)(t + 1 * width_hor),
330 vreinterpret_u16_u8(dd.val[1]), 0);
331 vst1_lane_u16((uint16_t *)(t + 2 * width_hor),
332 vreinterpret_u16_u8(dd.val[0]), 1);
333 vst1_lane_u16((uint16_t *)(t + 3 * width_hor),
334 vreinterpret_u16_u8(dd.val[1]), 1);
335 vst1_lane_u16((uint16_t *)(t + 4 * width_hor),
336 vreinterpret_u16_u8(dd.val[0]), 2);
337 vst1_lane_u16((uint16_t *)(t + 5 * width_hor),
338 vreinterpret_u16_u8(dd.val[1]), 2);
339 vst1_lane_u16((uint16_t *)(t + 6 * width_hor),
340 vreinterpret_u16_u8(dd.val[0]), 3);
341 vst1_lane_u16((uint16_t *)(t + 7 * width_hor),
342 vreinterpret_u16_u8(dd.val[1]), 3);
343
344 s[0] = s[8];
345 s[1] = s[9];
346 s[2] = s[10];
347 s[3] = s[11];
348
349 t += 2;
350 x -= 2;
351 } while (x);
352 src += 8 * src_stride - 4 * width_hor;
353 t += 7 * width_hor;
354 y -= 8;
355 } while (y);
356
357 // vertical 8x2
358 x = width_ver;
359 t = temp_buffer;
360 do {
361 load_u8_8x4(t, width_hor, &s[0], &s[1], &s[2], &s[3]);
362 t += 4 * width_hor;
363 y = height_ver;
364
365 do {
366 load_u8_8x8(t, width_hor, &s[4], &s[5], &s[6], &s[7], &s[8], &s[9],
367 &s[10], &s[11]);
368 t += 8 * width_hor;
369
370 d[0] = scale_filter_8(&s[0], filters); // 00 01 02 03 04 05 06 07
371 d[1] = scale_filter_8(&s[4], filters); // 10 11 12 13 14 15 16 17
372 vst1_u8(dst + 0 * dst_stride, d[0]);
373 vst1_u8(dst + 1 * dst_stride, d[1]);
374
375 s[0] = s[8];
376 s[1] = s[9];
377 s[2] = s[10];
378 s[3] = s[11];
379
380 dst += 2 * dst_stride;
381 y -= 2;
382 } while (y);
383 t -= width_hor * (4 * height_ver + 4);
384 t += 8;
385 dst -= height_ver * dst_stride;
386 dst += 8;
387 x -= 8;
388 } while (x);
389 }
390
scale_filter_bilinear(const uint8x8_t * const s,const uint8x8_t * const coef)391 static INLINE uint8x8_t scale_filter_bilinear(const uint8x8_t *const s,
392 const uint8x8_t *const coef) {
393 const uint16x8_t h0 = vmull_u8(s[0], coef[0]);
394 const uint16x8_t h1 = vmlal_u8(h0, s[1], coef[1]);
395
396 return vrshrn_n_u16(h1, 7);
397 }
398
399 // Notes for 4 to 3 scaling:
400 //
401 // 1. 6 rows are calculated in each horizontal inner loop, so width_hor must be
402 // multiple of 6, and no less than w.
403 //
404 // 2. 8 rows are calculated in each vertical inner loop, so width_ver must be
405 // multiple of 8, and no less than w.
406 //
407 // 3. 8 columns are calculated in each horizontal inner loop for further
408 // vertical scaling, so height_hor must be multiple of 8, and no less than
409 // 4 * h / 3.
410 //
411 // 4. 6 columns are calculated in each vertical inner loop, so height_ver must
412 // be multiple of 6, and no less than h.
413 //
414 // 5. The physical location of the last row of the 4 to 3 scaled frame is
415 // decided by phase_scaler, and are always less than 1 pixel below the last row
416 // of the original image.
scale_plane_4_to_3_bilinear(const uint8_t * src,const int src_stride,uint8_t * dst,const int dst_stride,const int w,const int h,const int phase_scaler,uint8_t * const temp_buffer)417 static void scale_plane_4_to_3_bilinear(const uint8_t *src,
418 const int src_stride, uint8_t *dst,
419 const int dst_stride, const int w,
420 const int h, const int phase_scaler,
421 uint8_t *const temp_buffer) {
422 static const int step_q4 = 16 * 4 / 3;
423 const int width_hor = (w + 5) - ((w + 5) % 6);
424 const int stride_hor = width_hor + 2; // store 2 extra pixels
425 const int width_ver = (w + 7) & ~7;
426 // We only need 1 extra row below because there are only 2 bilinear
427 // coefficients.
428 const int height_hor = (4 * h / 3 + 1 + 7) & ~7;
429 const int height_ver = (h + 5) - ((h + 5) % 6);
430 int x, y = height_hor;
431 uint8_t *t = temp_buffer;
432 uint8x8_t s[9], d[8], c[6];
433 const InterpKernel *interp_kernel =
434 (const InterpKernel *)av1_interp_filter_params_list[BILINEAR].filter_ptr;
435 assert(w && h);
436
437 c[0] = vdup_n_u8((uint8_t)interp_kernel[phase_scaler][3]);
438 c[1] = vdup_n_u8((uint8_t)interp_kernel[phase_scaler][4]);
439 c[2] = vdup_n_u8(
440 (uint8_t)interp_kernel[(phase_scaler + 1 * step_q4) & SUBPEL_MASK][3]);
441 c[3] = vdup_n_u8(
442 (uint8_t)interp_kernel[(phase_scaler + 1 * step_q4) & SUBPEL_MASK][4]);
443 c[4] = vdup_n_u8(
444 (uint8_t)interp_kernel[(phase_scaler + 2 * step_q4) & SUBPEL_MASK][3]);
445 c[5] = vdup_n_u8(
446 (uint8_t)interp_kernel[(phase_scaler + 2 * step_q4) & SUBPEL_MASK][4]);
447
448 d[6] = vdup_n_u8(0);
449 d[7] = vdup_n_u8(0);
450
451 // horizontal 6x8
452 do {
453 load_u8_8x8(src, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
454 &s[6], &s[7]);
455 src += 1;
456 transpose_u8_8x8(&s[0], &s[1], &s[2], &s[3], &s[4], &s[5], &s[6], &s[7]);
457 x = width_hor;
458
459 do {
460 load_u8_8x8(src, src_stride, &s[1], &s[2], &s[3], &s[4], &s[5], &s[6],
461 &s[7], &s[8]);
462 src += 8;
463 transpose_u8_8x8(&s[1], &s[2], &s[3], &s[4], &s[5], &s[6], &s[7], &s[8]);
464
465 // 00 10 20 30 40 50 60 70
466 // 01 11 21 31 41 51 61 71
467 // 02 12 22 32 42 52 62 72
468 // 03 13 23 33 43 53 63 73
469 // 04 14 24 34 44 54 64 74
470 // 05 15 25 35 45 55 65 75
471 d[0] = scale_filter_bilinear(&s[0], &c[0]);
472 d[1] =
473 scale_filter_bilinear(&s[(phase_scaler + 1 * step_q4) >> 4], &c[2]);
474 d[2] =
475 scale_filter_bilinear(&s[(phase_scaler + 2 * step_q4) >> 4], &c[4]);
476 d[3] = scale_filter_bilinear(&s[4], &c[0]);
477 d[4] = scale_filter_bilinear(&s[4 + ((phase_scaler + 1 * step_q4) >> 4)],
478 &c[2]);
479 d[5] = scale_filter_bilinear(&s[4 + ((phase_scaler + 2 * step_q4) >> 4)],
480 &c[4]);
481
482 // 00 01 02 03 04 05 xx xx
483 // 10 11 12 13 14 15 xx xx
484 // 20 21 22 23 24 25 xx xx
485 // 30 31 32 33 34 35 xx xx
486 // 40 41 42 43 44 45 xx xx
487 // 50 51 52 53 54 55 xx xx
488 // 60 61 62 63 64 65 xx xx
489 // 70 71 72 73 74 75 xx xx
490 transpose_u8_8x8(&d[0], &d[1], &d[2], &d[3], &d[4], &d[5], &d[6], &d[7]);
491 // store 2 extra pixels
492 vst1_u8(t + 0 * stride_hor, d[0]);
493 vst1_u8(t + 1 * stride_hor, d[1]);
494 vst1_u8(t + 2 * stride_hor, d[2]);
495 vst1_u8(t + 3 * stride_hor, d[3]);
496 vst1_u8(t + 4 * stride_hor, d[4]);
497 vst1_u8(t + 5 * stride_hor, d[5]);
498 vst1_u8(t + 6 * stride_hor, d[6]);
499 vst1_u8(t + 7 * stride_hor, d[7]);
500
501 s[0] = s[8];
502
503 t += 6;
504 x -= 6;
505 } while (x);
506 src += 8 * src_stride - 4 * width_hor / 3 - 1;
507 t += 7 * stride_hor + 2;
508 y -= 8;
509 } while (y);
510
511 // vertical 8x6
512 x = width_ver;
513 t = temp_buffer;
514 do {
515 load_u8_8x8(t, stride_hor, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5], &s[6],
516 &s[7]);
517 t += stride_hor;
518 y = height_ver;
519
520 do {
521 load_u8_8x8(t, stride_hor, &s[1], &s[2], &s[3], &s[4], &s[5], &s[6],
522 &s[7], &s[8]);
523 t += 8 * stride_hor;
524
525 d[0] = scale_filter_bilinear(&s[0], &c[0]);
526 d[1] =
527 scale_filter_bilinear(&s[(phase_scaler + 1 * step_q4) >> 4], &c[2]);
528 d[2] =
529 scale_filter_bilinear(&s[(phase_scaler + 2 * step_q4) >> 4], &c[4]);
530 d[3] = scale_filter_bilinear(&s[4], &c[0]);
531 d[4] = scale_filter_bilinear(&s[4 + ((phase_scaler + 1 * step_q4) >> 4)],
532 &c[2]);
533 d[5] = scale_filter_bilinear(&s[4 + ((phase_scaler + 2 * step_q4) >> 4)],
534 &c[4]);
535 vst1_u8(dst + 0 * dst_stride, d[0]);
536 vst1_u8(dst + 1 * dst_stride, d[1]);
537 vst1_u8(dst + 2 * dst_stride, d[2]);
538 vst1_u8(dst + 3 * dst_stride, d[3]);
539 vst1_u8(dst + 4 * dst_stride, d[4]);
540 vst1_u8(dst + 5 * dst_stride, d[5]);
541
542 s[0] = s[8];
543
544 dst += 6 * dst_stride;
545 y -= 6;
546 } while (y);
547 t -= stride_hor * (4 * height_ver / 3 + 1);
548 t += 8;
549 dst -= height_ver * dst_stride;
550 dst += 8;
551 x -= 8;
552 } while (x);
553 }
554
scale_plane_4_to_3_general(const uint8_t * src,const int src_stride,uint8_t * dst,const int dst_stride,const int w,const int h,const InterpKernel * const coef,const int phase_scaler,uint8_t * const temp_buffer)555 static void scale_plane_4_to_3_general(const uint8_t *src, const int src_stride,
556 uint8_t *dst, const int dst_stride,
557 const int w, const int h,
558 const InterpKernel *const coef,
559 const int phase_scaler,
560 uint8_t *const temp_buffer) {
561 static const int step_q4 = 16 * 4 / 3;
562 const int width_hor = (w + 5) - ((w + 5) % 6);
563 const int stride_hor = width_hor + 2; // store 2 extra pixels
564 const int width_ver = (w + 7) & ~7;
565 // We need (SUBPEL_TAPS - 1) extra rows: (SUBPEL_TAPS / 2 - 1) extra rows
566 // above and (SUBPEL_TAPS / 2) extra rows below.
567 const int height_hor = (4 * h / 3 + SUBPEL_TAPS - 1 + 7) & ~7;
568 const int height_ver = (h + 5) - ((h + 5) % 6);
569 const int16x8_t filters0 = vld1q_s16(
570 (const int16_t *)&coef[(phase_scaler + 0 * step_q4) & SUBPEL_MASK]);
571 const int16x8_t filters1 = vld1q_s16(
572 (const int16_t *)&coef[(phase_scaler + 1 * step_q4) & SUBPEL_MASK]);
573 const int16x8_t filters2 = vld1q_s16(
574 (const int16_t *)&coef[(phase_scaler + 2 * step_q4) & SUBPEL_MASK]);
575 int x, y = height_hor;
576 uint8_t *t = temp_buffer;
577 uint8x8_t s[15], d[8];
578
579 assert(w && h);
580
581 src -= (SUBPEL_TAPS / 2 - 1) * src_stride + SUBPEL_TAPS / 2;
582 d[6] = vdup_n_u8(0);
583 d[7] = vdup_n_u8(0);
584
585 // horizontal 6x8
586 do {
587 load_u8_8x8(src + 1, src_stride, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5],
588 &s[6], &s[7]);
589 transpose_u8_8x8(&s[0], &s[1], &s[2], &s[3], &s[4], &s[5], &s[6], &s[7]);
590 x = width_hor;
591
592 do {
593 src += 8;
594 load_u8_8x8(src, src_stride, &s[7], &s[8], &s[9], &s[10], &s[11], &s[12],
595 &s[13], &s[14]);
596 transpose_u8_8x8(&s[7], &s[8], &s[9], &s[10], &s[11], &s[12], &s[13],
597 &s[14]);
598
599 // 00 10 20 30 40 50 60 70
600 // 01 11 21 31 41 51 61 71
601 // 02 12 22 32 42 52 62 72
602 // 03 13 23 33 43 53 63 73
603 // 04 14 24 34 44 54 64 74
604 // 05 15 25 35 45 55 65 75
605 d[0] = scale_filter_8(&s[0], filters0);
606 d[1] = scale_filter_8(&s[(phase_scaler + 1 * step_q4) >> 4], filters1);
607 d[2] = scale_filter_8(&s[(phase_scaler + 2 * step_q4) >> 4], filters2);
608 d[3] = scale_filter_8(&s[4], filters0);
609 d[4] =
610 scale_filter_8(&s[4 + ((phase_scaler + 1 * step_q4) >> 4)], filters1);
611 d[5] =
612 scale_filter_8(&s[4 + ((phase_scaler + 2 * step_q4) >> 4)], filters2);
613
614 // 00 01 02 03 04 05 xx xx
615 // 10 11 12 13 14 15 xx xx
616 // 20 21 22 23 24 25 xx xx
617 // 30 31 32 33 34 35 xx xx
618 // 40 41 42 43 44 45 xx xx
619 // 50 51 52 53 54 55 xx xx
620 // 60 61 62 63 64 65 xx xx
621 // 70 71 72 73 74 75 xx xx
622 transpose_u8_8x8(&d[0], &d[1], &d[2], &d[3], &d[4], &d[5], &d[6], &d[7]);
623 // store 2 extra pixels
624 vst1_u8(t + 0 * stride_hor, d[0]);
625 vst1_u8(t + 1 * stride_hor, d[1]);
626 vst1_u8(t + 2 * stride_hor, d[2]);
627 vst1_u8(t + 3 * stride_hor, d[3]);
628 vst1_u8(t + 4 * stride_hor, d[4]);
629 vst1_u8(t + 5 * stride_hor, d[5]);
630 vst1_u8(t + 6 * stride_hor, d[6]);
631 vst1_u8(t + 7 * stride_hor, d[7]);
632
633 s[0] = s[8];
634 s[1] = s[9];
635 s[2] = s[10];
636 s[3] = s[11];
637 s[4] = s[12];
638 s[5] = s[13];
639 s[6] = s[14];
640
641 t += 6;
642 x -= 6;
643 } while (x);
644 src += 8 * src_stride - 4 * width_hor / 3;
645 t += 7 * stride_hor + 2;
646 y -= 8;
647 } while (y);
648
649 // vertical 8x6
650 x = width_ver;
651 t = temp_buffer;
652 do {
653 load_u8_8x8(t, stride_hor, &s[0], &s[1], &s[2], &s[3], &s[4], &s[5], &s[6],
654 &s[7]);
655 t += 7 * stride_hor;
656 y = height_ver;
657
658 do {
659 load_u8_8x8(t, stride_hor, &s[7], &s[8], &s[9], &s[10], &s[11], &s[12],
660 &s[13], &s[14]);
661 t += 8 * stride_hor;
662
663 d[0] = scale_filter_8(&s[0], filters0);
664 d[1] = scale_filter_8(&s[(phase_scaler + 1 * step_q4) >> 4], filters1);
665 d[2] = scale_filter_8(&s[(phase_scaler + 2 * step_q4) >> 4], filters2);
666 d[3] = scale_filter_8(&s[4], filters0);
667 d[4] =
668 scale_filter_8(&s[4 + ((phase_scaler + 1 * step_q4) >> 4)], filters1);
669 d[5] =
670 scale_filter_8(&s[4 + ((phase_scaler + 2 * step_q4) >> 4)], filters2);
671 vst1_u8(dst + 0 * dst_stride, d[0]);
672 vst1_u8(dst + 1 * dst_stride, d[1]);
673 vst1_u8(dst + 2 * dst_stride, d[2]);
674 vst1_u8(dst + 3 * dst_stride, d[3]);
675 vst1_u8(dst + 4 * dst_stride, d[4]);
676 vst1_u8(dst + 5 * dst_stride, d[5]);
677
678 s[0] = s[8];
679 s[1] = s[9];
680 s[2] = s[10];
681 s[3] = s[11];
682 s[4] = s[12];
683 s[5] = s[13];
684 s[6] = s[14];
685
686 dst += 6 * dst_stride;
687 y -= 6;
688 } while (y);
689 t -= stride_hor * (4 * height_ver / 3 + 7);
690 t += 8;
691 dst -= height_ver * dst_stride;
692 dst += 8;
693 x -= 8;
694 } while (x);
695 }
696
av1_resize_and_extend_frame_neon(const YV12_BUFFER_CONFIG * src,YV12_BUFFER_CONFIG * dst,const InterpFilter filter,const int phase,const int num_planes)697 void av1_resize_and_extend_frame_neon(const YV12_BUFFER_CONFIG *src,
698 YV12_BUFFER_CONFIG *dst,
699 const InterpFilter filter,
700 const int phase, const int num_planes) {
701 // We use AOMMIN(num_planes, MAX_MB_PLANE) instead of num_planes to quiet
702 // the static analysis warnings.
703 int scaled = 0;
704 for (int i = 0; i < AOMMIN(num_planes, MAX_MB_PLANE); ++i) {
705 const int is_uv = i > 0;
706 const int src_w = src->crop_widths[is_uv];
707 const int src_h = src->crop_heights[is_uv];
708 const int dst_w = dst->crop_widths[is_uv];
709 const int dst_h = dst->crop_heights[is_uv];
710 const int dst_y_w = (dst->crop_widths[0] + 1) & ~1;
711 const int dst_y_h = (dst->crop_heights[0] + 1) & ~1;
712
713 if (2 * dst_w == src_w && 2 * dst_h == src_h) {
714 scaled = 1;
715 if (phase == 0) {
716 scale_plane_2_to_1_phase_0(src->buffers[i], src->strides[is_uv],
717 dst->buffers[i], dst->strides[is_uv], dst_w,
718 dst_h);
719 } else if (filter == BILINEAR) {
720 const int16_t c0 = av1_bilinear_filters[phase][3];
721 const int16_t c1 = av1_bilinear_filters[phase][4];
722 scale_plane_2_to_1_bilinear(src->buffers[i], src->strides[is_uv],
723 dst->buffers[i], dst->strides[is_uv], dst_w,
724 dst_h, c0, c1);
725 } else {
726 const int buffer_stride = (dst_y_w + 3) & ~3;
727 const int buffer_height = (2 * dst_y_h + SUBPEL_TAPS - 2 + 7) & ~7;
728 uint8_t *const temp_buffer =
729 (uint8_t *)malloc(buffer_stride * buffer_height);
730 if (temp_buffer) {
731 const InterpKernel *interp_kernel =
732 (const InterpKernel *)av1_interp_filter_params_list[filter]
733 .filter_ptr;
734 scale_plane_2_to_1_general(src->buffers[i], src->strides[is_uv],
735 dst->buffers[i], dst->strides[is_uv],
736 dst_w, dst_h, interp_kernel[phase],
737 temp_buffer);
738 free(temp_buffer);
739 } else {
740 scaled = 0;
741 }
742 }
743 } else if (4 * dst_w == src_w && 4 * dst_h == src_h) {
744 scaled = 1;
745 if (phase == 0) {
746 scale_plane_4_to_1_phase_0(src->buffers[i], src->strides[is_uv],
747 dst->buffers[i], dst->strides[is_uv], dst_w,
748 dst_h);
749 } else if (filter == BILINEAR) {
750 const int16_t c0 = av1_bilinear_filters[phase][3];
751 const int16_t c1 = av1_bilinear_filters[phase][4];
752 scale_plane_4_to_1_bilinear(src->buffers[i], src->strides[is_uv],
753 dst->buffers[i], dst->strides[is_uv], dst_w,
754 dst_h, c0, c1);
755 } else {
756 const int buffer_stride = (dst_y_w + 1) & ~1;
757 const int buffer_height = (4 * dst_y_h + SUBPEL_TAPS - 2 + 7) & ~7;
758 uint8_t *const temp_buffer =
759 (uint8_t *)malloc(buffer_stride * buffer_height);
760 if (temp_buffer) {
761 const InterpKernel *interp_kernel =
762 (const InterpKernel *)av1_interp_filter_params_list[filter]
763 .filter_ptr;
764 scale_plane_4_to_1_general(src->buffers[i], src->strides[is_uv],
765 dst->buffers[i], dst->strides[is_uv],
766 dst_w, dst_h, interp_kernel[phase],
767 temp_buffer);
768 free(temp_buffer);
769 } else {
770 scaled = 0;
771 }
772 }
773 } else if (4 * dst_w == 3 * src_w && 4 * dst_h == 3 * src_h) {
774 // 4 to 3
775 const int buffer_stride = (dst_y_w + 5) - ((dst_y_w + 5) % 6) + 2;
776 const int buffer_height = (4 * dst_y_h / 3 + SUBPEL_TAPS - 1 + 7) & ~7;
777 uint8_t *const temp_buffer =
778 (uint8_t *)malloc(buffer_stride * buffer_height);
779 if (temp_buffer) {
780 scaled = 1;
781 if (filter == BILINEAR) {
782 scale_plane_4_to_3_bilinear(src->buffers[i], src->strides[is_uv],
783 dst->buffers[i], dst->strides[is_uv],
784 dst_w, dst_h, phase, temp_buffer);
785 } else {
786 const InterpKernel *interp_kernel =
787 (const InterpKernel *)av1_interp_filter_params_list[filter]
788 .filter_ptr;
789 scale_plane_4_to_3_general(src->buffers[i], src->strides[is_uv],
790 dst->buffers[i], dst->strides[is_uv],
791 dst_w, dst_h, interp_kernel, phase,
792 temp_buffer);
793 }
794 free(temp_buffer);
795 } else {
796 scaled = 0;
797 }
798 }
799 }
800 if (!scaled) {
801 av1_resize_and_extend_frame_c(src, dst, filter, phase, num_planes);
802 } else {
803 aom_extend_frame_borders(dst, num_planes);
804 }
805 }
806