1 /*
2 *
3 * Copyright (c) 2018, 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
13 #include <assert.h>
14 #include <arm_neon.h>
15
16 #include "config/aom_config.h"
17 #include "config/av1_rtcd.h"
18
19 #include "aom_dsp/aom_dsp_common.h"
20 #include "aom_dsp/arm/mem_neon.h"
21 #include "aom_dsp/arm/transpose_neon.h"
22 #include "aom_ports/mem.h"
23 #include "av1/common/convolve.h"
24 #include "av1/common/filter.h"
25 #include "av1/common/arm/convolve_neon.h"
26
convolve12_4_x(const int16x4_t s0,const int16x4_t s1,const int16x4_t s2,const int16x4_t s3,const int16x4_t s4,const int16x4_t s5,const int16x4_t s6,const int16x4_t s7,const int16x4_t s8,const int16x4_t s9,const int16x4_t s10,const int16x4_t s11,const int16x8_t x_filter_0_7,const int16x4_t x_filter_8_11,const int32x4_t horiz_const)27 static INLINE int16x4_t convolve12_4_x(const int16x4_t s0, const int16x4_t s1,
28 const int16x4_t s2, const int16x4_t s3,
29 const int16x4_t s4, const int16x4_t s5,
30 const int16x4_t s6, const int16x4_t s7,
31 const int16x4_t s8, const int16x4_t s9,
32 const int16x4_t s10, const int16x4_t s11,
33 const int16x8_t x_filter_0_7,
34 const int16x4_t x_filter_8_11,
35 const int32x4_t horiz_const) {
36 const int16x4_t x_filter_0_3 = vget_low_s16(x_filter_0_7);
37 const int16x4_t x_filter_4_7 = vget_high_s16(x_filter_0_7);
38
39 int32x4_t sum = horiz_const;
40 sum = vmlal_lane_s16(sum, s0, x_filter_0_3, 0);
41 sum = vmlal_lane_s16(sum, s1, x_filter_0_3, 1);
42 sum = vmlal_lane_s16(sum, s2, x_filter_0_3, 2);
43 sum = vmlal_lane_s16(sum, s3, x_filter_0_3, 3);
44 sum = vmlal_lane_s16(sum, s4, x_filter_4_7, 0);
45 sum = vmlal_lane_s16(sum, s5, x_filter_4_7, 1);
46 sum = vmlal_lane_s16(sum, s6, x_filter_4_7, 2);
47 sum = vmlal_lane_s16(sum, s7, x_filter_4_7, 3);
48 sum = vmlal_lane_s16(sum, s8, x_filter_8_11, 0);
49 sum = vmlal_lane_s16(sum, s9, x_filter_8_11, 1);
50 sum = vmlal_lane_s16(sum, s10, x_filter_8_11, 2);
51 sum = vmlal_lane_s16(sum, s11, x_filter_8_11, 3);
52
53 return vqrshrn_n_s32(sum, FILTER_BITS);
54 }
55
convolve_x_sr_12tap_neon(const uint8_t * src_ptr,int src_stride,uint8_t * dst_ptr,const int dst_stride,int w,int h,const int16_t * x_filter_ptr)56 static INLINE void convolve_x_sr_12tap_neon(const uint8_t *src_ptr,
57 int src_stride, uint8_t *dst_ptr,
58 const int dst_stride, int w, int h,
59 const int16_t *x_filter_ptr) {
60 const int16x8_t x_filter_0_7 = vld1q_s16(x_filter_ptr);
61 const int16x4_t x_filter_8_11 = vld1_s16(x_filter_ptr + 8);
62
63 // A shim of 1 << (ROUND0_BITS - 1) enables us to use a single rounding right
64 // shift by FILTER_BITS - instead of a first rounding right shift by
65 // ROUND0_BITS, followed by second rounding right shift by FILTER_BITS -
66 // ROUND0_BITS.
67 const int32x4_t horiz_const = vdupq_n_s32(1 << (ROUND0_BITS - 1));
68
69 #if AOM_ARCH_AARCH64
70 do {
71 const uint8_t *s = src_ptr;
72 uint8_t *d = dst_ptr;
73 int width = w;
74
75 uint8x8_t t0, t1, t2, t3;
76 load_u8_8x4(s, src_stride, &t0, &t1, &t2, &t3);
77 transpose_elems_inplace_u8_8x4(&t0, &t1, &t2, &t3);
78
79 int16x4_t s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
80 int16x4_t s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
81 int16x4_t s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
82 int16x4_t s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
83 int16x4_t s4 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
84 int16x4_t s5 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
85 int16x4_t s6 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
86 int16x4_t s7 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
87
88 load_u8_8x4(s + 8, src_stride, &t0, &t1, &t2, &t3);
89 transpose_elems_inplace_u8_8x4(&t0, &t1, &t2, &t3);
90
91 int16x4_t s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
92 int16x4_t s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
93 int16x4_t s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
94
95 s += 11;
96
97 do {
98 load_u8_8x4(s, src_stride, &t0, &t1, &t2, &t3);
99 transpose_elems_inplace_u8_8x4(&t0, &t1, &t2, &t3);
100
101 int16x4_t s11 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
102 int16x4_t s12 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
103 int16x4_t s13 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
104 int16x4_t s14 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
105
106 int16x4_t d0 =
107 convolve12_4_x(s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11,
108 x_filter_0_7, x_filter_8_11, horiz_const);
109 int16x4_t d1 =
110 convolve12_4_x(s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12,
111 x_filter_0_7, x_filter_8_11, horiz_const);
112 int16x4_t d2 =
113 convolve12_4_x(s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13,
114 x_filter_0_7, x_filter_8_11, horiz_const);
115 int16x4_t d3 =
116 convolve12_4_x(s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, s14,
117 x_filter_0_7, x_filter_8_11, horiz_const);
118
119 transpose_elems_inplace_s16_4x4(&d0, &d1, &d2, &d3);
120
121 uint8x8_t d01 = vqmovun_s16(vcombine_s16(d0, d1));
122 uint8x8_t d23 = vqmovun_s16(vcombine_s16(d2, d3));
123
124 store_u8x4_strided_x2(d, dst_stride, d01);
125 store_u8x4_strided_x2(d + 2 * dst_stride, dst_stride, d23);
126
127 s0 = s4;
128 s1 = s5;
129 s2 = s6;
130 s3 = s7;
131 s4 = s8;
132 s5 = s9;
133 s6 = s10;
134 s7 = s11;
135 s8 = s12;
136 s9 = s13;
137 s10 = s14;
138 s += 4;
139 d += 4;
140 width -= 4;
141 } while (width != 0);
142 src_ptr += 4 * src_stride;
143 dst_ptr += 4 * dst_stride;
144 h -= 4;
145 } while (h != 0);
146
147 #else // !AOM_ARCH_AARCH64
148 do {
149 const uint8_t *s = src_ptr;
150 uint8_t *d = dst_ptr;
151 int width = w;
152
153 do {
154 uint8x16_t t0 = vld1q_u8(s);
155 int16x8_t tt0 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(t0)));
156 int16x8_t tt8 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(t0)));
157
158 int16x4_t s0 = vget_low_s16(tt0);
159 int16x4_t s4 = vget_high_s16(tt0);
160 int16x4_t s8 = vget_low_s16(tt8);
161 int16x4_t s12 = vget_high_s16(tt8);
162
163 int16x4_t s1 = vext_s16(s0, s4, 1); // a1 a2 a3 a4
164 int16x4_t s2 = vext_s16(s0, s4, 2); // a2 a3 a4 a5
165 int16x4_t s3 = vext_s16(s0, s4, 3); // a3 a4 a5 a6
166 int16x4_t s5 = vext_s16(s4, s8, 1); // a5 a6 a7 a8
167 int16x4_t s6 = vext_s16(s4, s8, 2); // a6 a7 a8 a9
168 int16x4_t s7 = vext_s16(s4, s8, 3); // a7 a8 a9 a10
169 int16x4_t s9 = vext_s16(s8, s12, 1); // a9 a10 a11 a12
170 int16x4_t s10 = vext_s16(s8, s12, 2); // a10 a11 a12 a13
171 int16x4_t s11 = vext_s16(s8, s12, 3); // a11 a12 a13 a14
172
173 int16x4_t d0 =
174 convolve12_4_x(s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11,
175 x_filter_0_7, x_filter_8_11, horiz_const);
176
177 uint8x8_t dd0 = vqmovun_s16(vcombine_s16(d0, vdup_n_s16(0)));
178
179 store_u8_4x1(d, dd0);
180
181 s += 4;
182 d += 4;
183 width -= 4;
184 } while (width != 0);
185 src_ptr += src_stride;
186 dst_ptr += dst_stride;
187 } while (--h != 0);
188 #endif // AOM_ARCH_AARCH64
189 }
190
convolve4_4_x(const int16x4_t s0,const int16x4_t s1,const int16x4_t s2,const int16x4_t s3,const int16x4_t filter,const int16x4_t horiz_const)191 static INLINE uint8x8_t convolve4_4_x(const int16x4_t s0, const int16x4_t s1,
192 const int16x4_t s2, const int16x4_t s3,
193 const int16x4_t filter,
194 const int16x4_t horiz_const) {
195 int16x4_t sum = horiz_const;
196 sum = vmla_lane_s16(sum, s0, filter, 0);
197 sum = vmla_lane_s16(sum, s1, filter, 1);
198 sum = vmla_lane_s16(sum, s2, filter, 2);
199 sum = vmla_lane_s16(sum, s3, filter, 3);
200
201 // We halved the convolution filter values so - 1 from the right shift.
202 return vqrshrun_n_s16(vcombine_s16(sum, vdup_n_s16(0)), FILTER_BITS - 1);
203 }
204
convolve8_8_x(const int16x8_t s0,const int16x8_t s1,const int16x8_t s2,const int16x8_t s3,const int16x8_t s4,const int16x8_t s5,const int16x8_t s6,const int16x8_t s7,const int16x8_t filter,const int16x8_t horiz_const)205 static INLINE uint8x8_t convolve8_8_x(const int16x8_t s0, const int16x8_t s1,
206 const int16x8_t s2, const int16x8_t s3,
207 const int16x8_t s4, const int16x8_t s5,
208 const int16x8_t s6, const int16x8_t s7,
209 const int16x8_t filter,
210 const int16x8_t horiz_const) {
211 const int16x4_t filter_lo = vget_low_s16(filter);
212 const int16x4_t filter_hi = vget_high_s16(filter);
213
214 int16x8_t sum = horiz_const;
215 sum = vmlaq_lane_s16(sum, s0, filter_lo, 0);
216 sum = vmlaq_lane_s16(sum, s1, filter_lo, 1);
217 sum = vmlaq_lane_s16(sum, s2, filter_lo, 2);
218 sum = vmlaq_lane_s16(sum, s3, filter_lo, 3);
219 sum = vmlaq_lane_s16(sum, s4, filter_hi, 0);
220 sum = vmlaq_lane_s16(sum, s5, filter_hi, 1);
221 sum = vmlaq_lane_s16(sum, s6, filter_hi, 2);
222 sum = vmlaq_lane_s16(sum, s7, filter_hi, 3);
223
224 // We halved the convolution filter values so - 1 from the right shift.
225 return vqrshrun_n_s16(sum, FILTER_BITS - 1);
226 }
227
av1_convolve_x_sr_neon(const uint8_t * src,int src_stride,uint8_t * dst,int dst_stride,int w,int h,const InterpFilterParams * filter_params_x,const int subpel_x_qn,ConvolveParams * conv_params)228 void av1_convolve_x_sr_neon(const uint8_t *src, int src_stride, uint8_t *dst,
229 int dst_stride, int w, int h,
230 const InterpFilterParams *filter_params_x,
231 const int subpel_x_qn,
232 ConvolveParams *conv_params) {
233 if (w == 2 || h == 2) {
234 av1_convolve_x_sr_c(src, src_stride, dst, dst_stride, w, h, filter_params_x,
235 subpel_x_qn, conv_params);
236 return;
237 }
238
239 const uint8_t horiz_offset = filter_params_x->taps / 2 - 1;
240 src -= horiz_offset;
241
242 const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel(
243 filter_params_x, subpel_x_qn & SUBPEL_MASK);
244
245 if (filter_params_x->taps > 8) {
246 convolve_x_sr_12tap_neon(src, src_stride, dst, dst_stride, w, h,
247 x_filter_ptr);
248 return;
249 }
250
251 // This shim of 1 << ((ROUND0_BITS - 1) - 1) enables us to use a single
252 // rounding right shift by FILTER_BITS - instead of a first rounding right
253 // shift by ROUND0_BITS, followed by second rounding right shift by
254 // FILTER_BITS - ROUND0_BITS.
255 // The outermost -1 is needed because we will halve the filter values.
256 const int16x8_t horiz_const = vdupq_n_s16(1 << ((ROUND0_BITS - 1) - 1));
257
258 if (w <= 4) {
259 // 4-tap filters are used for blocks having width <= 4.
260 // Filter values are even, so halve to reduce intermediate precision reqs.
261 const int16x4_t x_filter = vshr_n_s16(vld1_s16(x_filter_ptr + 2), 1);
262
263 src += 2;
264
265 do {
266 uint8x8_t t0 = vld1_u8(src); // a0 a1 a2 a3 a4 a5 a6 a7
267 int16x4_t s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
268 int16x4_t s4 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
269
270 int16x4_t s1 = vext_s16(s0, s4, 1); // a1 a2 a3 a4
271 int16x4_t s2 = vext_s16(s0, s4, 2); // a2 a3 a4 a5
272 int16x4_t s3 = vext_s16(s0, s4, 3); // a3 a4 a5 a6
273
274 uint8x8_t d0 =
275 convolve4_4_x(s0, s1, s2, s3, x_filter, vget_low_s16(horiz_const));
276
277 store_u8_4x1(dst, d0);
278
279 src += src_stride;
280 dst += dst_stride;
281 } while (--h != 0);
282 } else {
283 // Filter values are even so halve to reduce precision requirements.
284 const int16x8_t x_filter = vshrq_n_s16(vld1q_s16(x_filter_ptr), 1);
285
286 #if AOM_ARCH_AARCH64
287 while (h >= 8) {
288 uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7;
289 load_u8_8x8(src, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
290
291 transpose_elems_inplace_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
292 int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
293 int16x8_t s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
294 int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
295 int16x8_t s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
296 int16x8_t s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
297 int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
298 int16x8_t s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
299
300 int width = w;
301 const uint8_t *s = src + 7;
302 uint8_t *d = dst;
303
304 __builtin_prefetch(d + 0 * dst_stride);
305 __builtin_prefetch(d + 1 * dst_stride);
306 __builtin_prefetch(d + 2 * dst_stride);
307 __builtin_prefetch(d + 3 * dst_stride);
308 __builtin_prefetch(d + 4 * dst_stride);
309 __builtin_prefetch(d + 5 * dst_stride);
310 __builtin_prefetch(d + 6 * dst_stride);
311 __builtin_prefetch(d + 7 * dst_stride);
312
313 do {
314 uint8x8_t t8, t9, t10, t11, t12, t13, t14;
315 load_u8_8x8(s, src_stride, &t7, &t8, &t9, &t10, &t11, &t12, &t13, &t14);
316
317 transpose_elems_inplace_u8_8x8(&t7, &t8, &t9, &t10, &t11, &t12, &t13,
318 &t14);
319 int16x8_t s7 = vreinterpretq_s16_u16(vmovl_u8(t7));
320 int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t8));
321 int16x8_t s9 = vreinterpretq_s16_u16(vmovl_u8(t9));
322 int16x8_t s10 = vreinterpretq_s16_u16(vmovl_u8(t10));
323 int16x8_t s11 = vreinterpretq_s16_u16(vmovl_u8(t11));
324 int16x8_t s12 = vreinterpretq_s16_u16(vmovl_u8(t12));
325 int16x8_t s13 = vreinterpretq_s16_u16(vmovl_u8(t13));
326 int16x8_t s14 = vreinterpretq_s16_u16(vmovl_u8(t14));
327
328 uint8x8_t d0 = convolve8_8_x(s0, s1, s2, s3, s4, s5, s6, s7, x_filter,
329 horiz_const);
330 uint8x8_t d1 = convolve8_8_x(s1, s2, s3, s4, s5, s6, s7, s8, x_filter,
331 horiz_const);
332 uint8x8_t d2 = convolve8_8_x(s2, s3, s4, s5, s6, s7, s8, s9, x_filter,
333 horiz_const);
334 uint8x8_t d3 = convolve8_8_x(s3, s4, s5, s6, s7, s8, s9, s10, x_filter,
335 horiz_const);
336 uint8x8_t d4 = convolve8_8_x(s4, s5, s6, s7, s8, s9, s10, s11, x_filter,
337 horiz_const);
338 uint8x8_t d5 = convolve8_8_x(s5, s6, s7, s8, s9, s10, s11, s12,
339 x_filter, horiz_const);
340 uint8x8_t d6 = convolve8_8_x(s6, s7, s8, s9, s10, s11, s12, s13,
341 x_filter, horiz_const);
342 uint8x8_t d7 = convolve8_8_x(s7, s8, s9, s10, s11, s12, s13, s14,
343 x_filter, horiz_const);
344
345 transpose_elems_inplace_u8_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
346
347 store_u8_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7);
348
349 s0 = s8;
350 s1 = s9;
351 s2 = s10;
352 s3 = s11;
353 s4 = s12;
354 s5 = s13;
355 s6 = s14;
356 s += 8;
357 d += 8;
358 width -= 8;
359 } while (width != 0);
360 src += 8 * src_stride;
361 dst += 8 * dst_stride;
362 h -= 8;
363 }
364 #endif // AOM_ARCH_AARCH64
365
366 while (h-- != 0) {
367 uint8x8_t t0 = vld1_u8(src); // a0 a1 a2 a3 a4 a5 a6 a7
368 int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
369
370 int width = w;
371 const uint8_t *s = src + 8;
372 uint8_t *d = dst;
373
374 __builtin_prefetch(d);
375
376 do {
377 uint8x8_t t8 = vld1_u8(s); // a8 a9 a10 a11 a12 a13 a14 a15
378 int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t8));
379
380 int16x8_t s1 = vextq_s16(s0, s8, 1); // a1 a2 a3 a4 a5 a6 a7 a8
381 int16x8_t s2 = vextq_s16(s0, s8, 2); // a2 a3 a4 a5 a6 a7 a8 a9
382 int16x8_t s3 = vextq_s16(s0, s8, 3); // a3 a4 a5 a6 a7 a8 a9 a10
383 int16x8_t s4 = vextq_s16(s0, s8, 4); // a4 a5 a6 a7 a8 a9 a10 a11
384 int16x8_t s5 = vextq_s16(s0, s8, 5); // a5 a6 a7 a8 a9 a10 a11 a12
385 int16x8_t s6 = vextq_s16(s0, s8, 6); // a6 a7 a8 a9 a10 a11 a12 a13
386 int16x8_t s7 = vextq_s16(s0, s8, 7); // a7 a8 a9 a10 a11 a12 a13 a14
387
388 uint8x8_t d0 = convolve8_8_x(s0, s1, s2, s3, s4, s5, s6, s7, x_filter,
389 horiz_const);
390
391 vst1_u8(d, d0);
392
393 s0 = s8;
394 s += 8;
395 d += 8;
396 width -= 8;
397 } while (width != 0);
398 src += src_stride;
399 dst += dst_stride;
400 }
401 }
402 }
403
convolve6_4_y(const int16x4_t s0,const int16x4_t s1,const int16x4_t s2,const int16x4_t s3,const int16x4_t s4,const int16x4_t s5,const int16x8_t y_filter_0_7)404 static INLINE int16x4_t convolve6_4_y(const int16x4_t s0, const int16x4_t s1,
405 const int16x4_t s2, const int16x4_t s3,
406 const int16x4_t s4, const int16x4_t s5,
407 const int16x8_t y_filter_0_7) {
408 const int16x4_t y_filter_0_3 = vget_low_s16(y_filter_0_7);
409 const int16x4_t y_filter_4_7 = vget_high_s16(y_filter_0_7);
410
411 // Filter values at indices 0 and 7 are 0.
412 int16x4_t sum = vmul_lane_s16(s0, y_filter_0_3, 1);
413 sum = vmla_lane_s16(sum, s1, y_filter_0_3, 2);
414 sum = vmla_lane_s16(sum, s2, y_filter_0_3, 3);
415 sum = vmla_lane_s16(sum, s3, y_filter_4_7, 0);
416 sum = vmla_lane_s16(sum, s4, y_filter_4_7, 1);
417 sum = vmla_lane_s16(sum, s5, y_filter_4_7, 2);
418
419 return sum;
420 }
421
convolve6_8_y(const int16x8_t s0,const int16x8_t s1,const int16x8_t s2,const int16x8_t s3,const int16x8_t s4,const int16x8_t s5,const int16x8_t y_filters)422 static INLINE uint8x8_t convolve6_8_y(const int16x8_t s0, const int16x8_t s1,
423 const int16x8_t s2, const int16x8_t s3,
424 const int16x8_t s4, const int16x8_t s5,
425 const int16x8_t y_filters) {
426 const int16x4_t y_filter_lo = vget_low_s16(y_filters);
427 const int16x4_t y_filter_hi = vget_high_s16(y_filters);
428
429 // Filter values at indices 0 and 7 are 0.
430 int16x8_t sum = vmulq_lane_s16(s0, y_filter_lo, 1);
431 sum = vmlaq_lane_s16(sum, s1, y_filter_lo, 2);
432 sum = vmlaq_lane_s16(sum, s2, y_filter_lo, 3);
433 sum = vmlaq_lane_s16(sum, s3, y_filter_hi, 0);
434 sum = vmlaq_lane_s16(sum, s4, y_filter_hi, 1);
435 sum = vmlaq_lane_s16(sum, s5, y_filter_hi, 2);
436 // We halved the convolution filter values so -1 from the right shift.
437 return vqrshrun_n_s16(sum, FILTER_BITS - 1);
438 }
439
convolve_y_sr_6tap_neon(const uint8_t * src_ptr,int src_stride,uint8_t * dst_ptr,const int dst_stride,int w,int h,const int16x8_t y_filter)440 static INLINE void convolve_y_sr_6tap_neon(const uint8_t *src_ptr,
441 int src_stride, uint8_t *dst_ptr,
442 const int dst_stride, int w, int h,
443 const int16x8_t y_filter) {
444 if (w <= 4) {
445 uint8x8_t t0 = load_unaligned_u8_4x1(src_ptr + 0 * src_stride);
446 uint8x8_t t1 = load_unaligned_u8_4x1(src_ptr + 1 * src_stride);
447 uint8x8_t t2 = load_unaligned_u8_4x1(src_ptr + 2 * src_stride);
448 uint8x8_t t3 = load_unaligned_u8_4x1(src_ptr + 3 * src_stride);
449 uint8x8_t t4 = load_unaligned_u8_4x1(src_ptr + 4 * src_stride);
450
451 int16x4_t s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
452 int16x4_t s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
453 int16x4_t s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
454 int16x4_t s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
455 int16x4_t s4 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t4)));
456
457 src_ptr += 5 * src_stride;
458
459 do {
460 #if AOM_ARCH_AARCH64
461 uint8x8_t t5 = load_unaligned_u8_4x1(src_ptr + 0 * src_stride);
462 uint8x8_t t6 = load_unaligned_u8_4x1(src_ptr + 1 * src_stride);
463 uint8x8_t t7 = load_unaligned_u8_4x1(src_ptr + 2 * src_stride);
464 uint8x8_t t8 = load_unaligned_u8_4x1(src_ptr + 3 * src_stride);
465
466 int16x4_t s5 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t5)));
467 int16x4_t s6 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t6)));
468 int16x4_t s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t7)));
469 int16x4_t s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t8)));
470
471 int16x4_t d0 = convolve6_4_y(s0, s1, s2, s3, s4, s5, y_filter);
472 int16x4_t d1 = convolve6_4_y(s1, s2, s3, s4, s5, s6, y_filter);
473 int16x4_t d2 = convolve6_4_y(s2, s3, s4, s5, s6, s7, y_filter);
474 int16x4_t d3 = convolve6_4_y(s3, s4, s5, s6, s7, s8, y_filter);
475
476 // We halved the convolution filter values so -1 from the right shift.
477 uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS - 1);
478 uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS - 1);
479
480 store_u8x4_strided_x2(dst_ptr, dst_stride, d01);
481 store_u8x4_strided_x2(dst_ptr + 2 * dst_stride, dst_stride, d23);
482
483 s0 = s4;
484 s1 = s5;
485 s2 = s6;
486 s3 = s7;
487 s4 = s8;
488 src_ptr += 4 * src_stride;
489 dst_ptr += 4 * dst_stride;
490 h -= 4;
491 #else // !AOM_ARCH_AARCH64
492 uint8x8_t t5 = load_unaligned_u8_4x1(src_ptr);
493 int16x4_t s5 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t5)));
494
495 int16x4_t d0 = convolve6_4_y(s0, s1, s2, s3, s4, s5, y_filter);
496 // We halved the convolution filter values so -1 from the right shift.
497 uint8x8_t d01 =
498 vqrshrun_n_s16(vcombine_s16(d0, vdup_n_s16(0)), FILTER_BITS - 1);
499
500 store_u8_4x1(dst_ptr, d01);
501
502 s0 = s1;
503 s1 = s2;
504 s2 = s3;
505 s3 = s4;
506 s4 = s5;
507 src_ptr += src_stride;
508 dst_ptr += dst_stride;
509 h--;
510 #endif // AOM_ARCH_AARCH64
511 } while (h != 0);
512
513 } else {
514 do {
515 const uint8_t *s = src_ptr;
516 uint8_t *d = dst_ptr;
517 int height = h;
518
519 uint8x8_t t0, t1, t2, t3, t4;
520 load_u8_8x5(s, src_stride, &t0, &t1, &t2, &t3, &t4);
521
522 int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
523 int16x8_t s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
524 int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
525 int16x8_t s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
526 int16x8_t s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
527
528 s += 5 * src_stride;
529
530 do {
531 #if AOM_ARCH_AARCH64
532 uint8x8_t t5, t6, t7, t8;
533 load_u8_8x4(s, src_stride, &t5, &t6, &t7, &t8);
534
535 int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
536 int16x8_t s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
537 int16x8_t s7 = vreinterpretq_s16_u16(vmovl_u8(t7));
538 int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t8));
539
540 uint8x8_t d0 = convolve6_8_y(s0, s1, s2, s3, s4, s5, y_filter);
541 uint8x8_t d1 = convolve6_8_y(s1, s2, s3, s4, s5, s6, y_filter);
542 uint8x8_t d2 = convolve6_8_y(s2, s3, s4, s5, s6, s7, y_filter);
543 uint8x8_t d3 = convolve6_8_y(s3, s4, s5, s6, s7, s8, y_filter);
544
545 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
546
547 s0 = s4;
548 s1 = s5;
549 s2 = s6;
550 s3 = s7;
551 s4 = s8;
552 s += 4 * src_stride;
553 d += 4 * dst_stride;
554 height -= 4;
555 #else // !AOM_ARCH_AARCH64
556 int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s)));
557
558 uint8x8_t d0 = convolve6_8_y(s0, s1, s2, s3, s4, s5, y_filter);
559
560 vst1_u8(d, d0);
561
562 s0 = s1;
563 s1 = s2;
564 s2 = s3;
565 s3 = s4;
566 s4 = s5;
567 s += src_stride;
568 d += dst_stride;
569 height--;
570 #endif // AOM_ARCH_AARCH64
571 } while (height != 0);
572 src_ptr += 8;
573 dst_ptr += 8;
574 w -= 8;
575 } while (w != 0);
576 }
577 }
578
convolve8_4_y(const int16x4_t s0,const int16x4_t s1,const int16x4_t s2,const int16x4_t s3,const int16x4_t s4,const int16x4_t s5,const int16x4_t s6,const int16x4_t s7,const int16x8_t filter)579 static INLINE int16x4_t convolve8_4_y(const int16x4_t s0, const int16x4_t s1,
580 const int16x4_t s2, const int16x4_t s3,
581 const int16x4_t s4, const int16x4_t s5,
582 const int16x4_t s6, const int16x4_t s7,
583 const int16x8_t filter) {
584 const int16x4_t filter_lo = vget_low_s16(filter);
585 const int16x4_t filter_hi = vget_high_s16(filter);
586
587 int16x4_t sum = vmul_lane_s16(s0, filter_lo, 0);
588 sum = vmla_lane_s16(sum, s1, filter_lo, 1);
589 sum = vmla_lane_s16(sum, s2, filter_lo, 2);
590 sum = vmla_lane_s16(sum, s3, filter_lo, 3);
591 sum = vmla_lane_s16(sum, s4, filter_hi, 0);
592 sum = vmla_lane_s16(sum, s5, filter_hi, 1);
593 sum = vmla_lane_s16(sum, s6, filter_hi, 2);
594 sum = vmla_lane_s16(sum, s7, filter_hi, 3);
595
596 return sum;
597 }
598
convolve8_8_y(const int16x8_t s0,const int16x8_t s1,const int16x8_t s2,const int16x8_t s3,const int16x8_t s4,const int16x8_t s5,const int16x8_t s6,const int16x8_t s7,const int16x8_t filter)599 static INLINE uint8x8_t convolve8_8_y(const int16x8_t s0, const int16x8_t s1,
600 const int16x8_t s2, const int16x8_t s3,
601 const int16x8_t s4, const int16x8_t s5,
602 const int16x8_t s6, const int16x8_t s7,
603 const int16x8_t filter) {
604 const int16x4_t filter_lo = vget_low_s16(filter);
605 const int16x4_t filter_hi = vget_high_s16(filter);
606
607 int16x8_t sum = vmulq_lane_s16(s0, filter_lo, 0);
608 sum = vmlaq_lane_s16(sum, s1, filter_lo, 1);
609 sum = vmlaq_lane_s16(sum, s2, filter_lo, 2);
610 sum = vmlaq_lane_s16(sum, s3, filter_lo, 3);
611 sum = vmlaq_lane_s16(sum, s4, filter_hi, 0);
612 sum = vmlaq_lane_s16(sum, s5, filter_hi, 1);
613 sum = vmlaq_lane_s16(sum, s6, filter_hi, 2);
614 sum = vmlaq_lane_s16(sum, s7, filter_hi, 3);
615
616 // We halved the convolution filter values so -1 from the right shift.
617 return vqrshrun_n_s16(sum, FILTER_BITS - 1);
618 }
619
convolve_y_sr_8tap_neon(const uint8_t * src_ptr,int src_stride,uint8_t * dst_ptr,const int dst_stride,int w,int h,const int16x8_t y_filter)620 static INLINE void convolve_y_sr_8tap_neon(const uint8_t *src_ptr,
621 int src_stride, uint8_t *dst_ptr,
622 const int dst_stride, int w, int h,
623 const int16x8_t y_filter) {
624 if (w <= 4) {
625 uint8x8_t t0 = load_unaligned_u8_4x1(src_ptr + 0 * src_stride);
626 uint8x8_t t1 = load_unaligned_u8_4x1(src_ptr + 1 * src_stride);
627 uint8x8_t t2 = load_unaligned_u8_4x1(src_ptr + 2 * src_stride);
628 uint8x8_t t3 = load_unaligned_u8_4x1(src_ptr + 3 * src_stride);
629 uint8x8_t t4 = load_unaligned_u8_4x1(src_ptr + 4 * src_stride);
630 uint8x8_t t5 = load_unaligned_u8_4x1(src_ptr + 5 * src_stride);
631 uint8x8_t t6 = load_unaligned_u8_4x1(src_ptr + 6 * src_stride);
632
633 int16x4_t s0 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t0)));
634 int16x4_t s1 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t1)));
635 int16x4_t s2 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t2)));
636 int16x4_t s3 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t3)));
637 int16x4_t s4 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t4)));
638 int16x4_t s5 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t5)));
639 int16x4_t s6 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t6)));
640
641 src_ptr += 7 * src_stride;
642
643 do {
644 #if AOM_ARCH_AARCH64
645 uint8x8_t t7 = load_unaligned_u8_4x1(src_ptr + 0 * src_stride);
646 uint8x8_t t8 = load_unaligned_u8_4x1(src_ptr + 1 * src_stride);
647 uint8x8_t t9 = load_unaligned_u8_4x1(src_ptr + 2 * src_stride);
648 uint8x8_t t10 = load_unaligned_u8_4x1(src_ptr + 3 * src_stride);
649
650 int16x4_t s7 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t7)));
651 int16x4_t s8 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t8)));
652 int16x4_t s9 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t9)));
653 int16x4_t s10 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t10)));
654
655 int16x4_t d0 = convolve8_4_y(s0, s1, s2, s3, s4, s5, s6, s7, y_filter);
656 int16x4_t d1 = convolve8_4_y(s1, s2, s3, s4, s5, s6, s7, s8, y_filter);
657 int16x4_t d2 = convolve8_4_y(s2, s3, s4, s5, s6, s7, s8, s9, y_filter);
658 int16x4_t d3 = convolve8_4_y(s3, s4, s5, s6, s7, s8, s9, s10, y_filter);
659
660 // We halved the convolution filter values so -1 from the right shift.
661 uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS - 1);
662 uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS - 1);
663
664 store_u8x4_strided_x2(dst_ptr, dst_stride, d01);
665 store_u8x4_strided_x2(dst_ptr + 2 * dst_stride, dst_stride, d23);
666
667 s0 = s4;
668 s1 = s5;
669 s2 = s6;
670 s3 = s7;
671 s4 = s8;
672 s5 = s9;
673 s6 = s10;
674 src_ptr += 4 * src_stride;
675 dst_ptr += 4 * dst_stride;
676 h -= 4;
677 #else // !AOM_ARCH_AARCH64
678 uint8x8_t t7 = load_unaligned_u8_4x1(src_ptr);
679 int16x4_t s7 = vreinterpret_s16_u16(vget_low_u16(vmovl_u8(t7)));
680
681 int16x4_t d0 = convolve8_4_y(s0, s1, s2, s3, s4, s5, s6, s7, y_filter);
682 // We halved the convolution filter values so -1 from the right shift.
683 uint8x8_t d01 =
684 vqrshrun_n_s16(vcombine_s16(d0, vdup_n_s16(0)), FILTER_BITS - 1);
685
686 store_u8_4x1(dst_ptr, d01);
687
688 s0 = s1;
689 s1 = s2;
690 s2 = s3;
691 s3 = s4;
692 s4 = s5;
693 s5 = s6;
694 s6 = s7;
695 src_ptr += src_stride;
696 dst_ptr += dst_stride;
697 h--;
698 #endif // AOM_ARCH_AARCH64
699 } while (h != 0);
700 } else {
701 do {
702 const uint8_t *s = src_ptr;
703 uint8_t *d = dst_ptr;
704 int height = h;
705
706 uint8x8_t t0, t1, t2, t3, t4, t5, t6;
707 load_u8_8x7(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6);
708
709 int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
710 int16x8_t s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
711 int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
712 int16x8_t s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
713 int16x8_t s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
714 int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
715 int16x8_t s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
716
717 s += 7 * src_stride;
718
719 do {
720 #if AOM_ARCH_AARCH64
721 uint8x8_t t7, t8, t9, t10;
722 load_u8_8x4(s, src_stride, &t7, &t8, &t9, &t10);
723
724 int16x8_t s7 = vreinterpretq_s16_u16(vmovl_u8(t7));
725 int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t8));
726 int16x8_t s9 = vreinterpretq_s16_u16(vmovl_u8(t9));
727 int16x8_t s10 = vreinterpretq_s16_u16(vmovl_u8(t10));
728
729 uint8x8_t d0 = convolve8_8_y(s0, s1, s2, s3, s4, s5, s6, s7, y_filter);
730 uint8x8_t d1 = convolve8_8_y(s1, s2, s3, s4, s5, s6, s7, s8, y_filter);
731 uint8x8_t d2 = convolve8_8_y(s2, s3, s4, s5, s6, s7, s8, s9, y_filter);
732 uint8x8_t d3 = convolve8_8_y(s3, s4, s5, s6, s7, s8, s9, s10, y_filter);
733
734 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
735
736 s0 = s4;
737 s1 = s5;
738 s2 = s6;
739 s3 = s7;
740 s4 = s8;
741 s5 = s9;
742 s6 = s10;
743 s += 4 * src_stride;
744 d += 4 * dst_stride;
745 height -= 4;
746 #else // !AOM_ARCH_AARCH64
747 int16x8_t s7 = vreinterpretq_s16_u16(vmovl_u8(vld1_u8(s)));
748
749 uint8x8_t d0 = convolve8_8_y(s0, s1, s2, s3, s4, s5, s6, s7, y_filter);
750
751 vst1_u8(d, d0);
752
753 s0 = s1;
754 s1 = s2;
755 s2 = s3;
756 s3 = s4;
757 s4 = s5;
758 s5 = s6;
759 s6 = s7;
760 s += src_stride;
761 d += dst_stride;
762 height--;
763 #endif // AOM_ARCH_AARCH64
764 } while (height != 0);
765 src_ptr += 8;
766 dst_ptr += 8;
767 w -= 8;
768 } while (w != 0);
769 }
770 }
771
convolve12_4_y(const int16x4_t s0,const int16x4_t s1,const int16x4_t s2,const int16x4_t s3,const int16x4_t s4,const int16x4_t s5,const int16x4_t s6,const int16x4_t s7,const int16x4_t s8,const int16x4_t s9,const int16x4_t s10,const int16x4_t s11,const int16x8_t y_filter_0_7,const int16x4_t y_filter_8_11)772 static INLINE int16x4_t convolve12_4_y(const int16x4_t s0, const int16x4_t s1,
773 const int16x4_t s2, const int16x4_t s3,
774 const int16x4_t s4, const int16x4_t s5,
775 const int16x4_t s6, const int16x4_t s7,
776 const int16x4_t s8, const int16x4_t s9,
777 const int16x4_t s10, const int16x4_t s11,
778 const int16x8_t y_filter_0_7,
779 const int16x4_t y_filter_8_11) {
780 const int16x4_t y_filter_0_3 = vget_low_s16(y_filter_0_7);
781 const int16x4_t y_filter_4_7 = vget_high_s16(y_filter_0_7);
782 int16x4_t sum;
783
784 sum = vmul_lane_s16(s0, y_filter_0_3, 0);
785 sum = vmla_lane_s16(sum, s1, y_filter_0_3, 1);
786 sum = vmla_lane_s16(sum, s2, y_filter_0_3, 2);
787 sum = vmla_lane_s16(sum, s3, y_filter_0_3, 3);
788 sum = vmla_lane_s16(sum, s4, y_filter_4_7, 0);
789
790 sum = vmla_lane_s16(sum, s7, y_filter_4_7, 3);
791 sum = vmla_lane_s16(sum, s8, y_filter_8_11, 0);
792 sum = vmla_lane_s16(sum, s9, y_filter_8_11, 1);
793 sum = vmla_lane_s16(sum, s10, y_filter_8_11, 2);
794 sum = vmla_lane_s16(sum, s11, y_filter_8_11, 3);
795
796 // Saturating addition is required for the largest filter taps to avoid
797 // overflow (while staying in 16-bit elements.)
798 sum = vqadd_s16(sum, vmul_lane_s16(s5, y_filter_4_7, 1));
799 sum = vqadd_s16(sum, vmul_lane_s16(s6, y_filter_4_7, 2));
800
801 return sum;
802 }
803
convolve12_8_y(const int16x8_t s0,const int16x8_t s1,const int16x8_t s2,const int16x8_t s3,const int16x8_t s4,const int16x8_t s5,const int16x8_t s6,const int16x8_t s7,const int16x8_t s8,const int16x8_t s9,const int16x8_t s10,const int16x8_t s11,const int16x8_t y_filter_0_7,const int16x4_t y_filter_8_11)804 static INLINE uint8x8_t convolve12_8_y(const int16x8_t s0, const int16x8_t s1,
805 const int16x8_t s2, const int16x8_t s3,
806 const int16x8_t s4, const int16x8_t s5,
807 const int16x8_t s6, const int16x8_t s7,
808 const int16x8_t s8, const int16x8_t s9,
809 const int16x8_t s10, const int16x8_t s11,
810 const int16x8_t y_filter_0_7,
811 const int16x4_t y_filter_8_11) {
812 const int16x4_t y_filter_0_3 = vget_low_s16(y_filter_0_7);
813 const int16x4_t y_filter_4_7 = vget_high_s16(y_filter_0_7);
814 int16x8_t sum;
815
816 sum = vmulq_lane_s16(s0, y_filter_0_3, 0);
817 sum = vmlaq_lane_s16(sum, s1, y_filter_0_3, 1);
818 sum = vmlaq_lane_s16(sum, s2, y_filter_0_3, 2);
819 sum = vmlaq_lane_s16(sum, s3, y_filter_0_3, 3);
820 sum = vmlaq_lane_s16(sum, s4, y_filter_4_7, 0);
821
822 sum = vmlaq_lane_s16(sum, s7, y_filter_4_7, 3);
823 sum = vmlaq_lane_s16(sum, s8, y_filter_8_11, 0);
824 sum = vmlaq_lane_s16(sum, s9, y_filter_8_11, 1);
825 sum = vmlaq_lane_s16(sum, s10, y_filter_8_11, 2);
826 sum = vmlaq_lane_s16(sum, s11, y_filter_8_11, 3);
827
828 // Saturating addition is required for the largest filter taps to avoid
829 // overflow (while staying in 16-bit elements.)
830 sum = vqaddq_s16(sum, vmulq_lane_s16(s5, y_filter_4_7, 1));
831 sum = vqaddq_s16(sum, vmulq_lane_s16(s6, y_filter_4_7, 2));
832
833 return vqrshrun_n_s16(sum, FILTER_BITS);
834 }
835
convolve_y_sr_12tap_neon(const uint8_t * src_ptr,int src_stride,uint8_t * dst_ptr,int dst_stride,int w,int h,const int16_t * y_filter_ptr)836 static INLINE void convolve_y_sr_12tap_neon(const uint8_t *src_ptr,
837 int src_stride, uint8_t *dst_ptr,
838 int dst_stride, int w, int h,
839 const int16_t *y_filter_ptr) {
840 const int16x8_t y_filter_0_7 = vld1q_s16(y_filter_ptr);
841 const int16x4_t y_filter_8_11 = vld1_s16(y_filter_ptr + 8);
842
843 if (w <= 4) {
844 uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
845 load_u8_8x11(src_ptr, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7,
846 &t8, &t9, &t10);
847 int16x4_t s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
848 int16x4_t s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
849 int16x4_t s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
850 int16x4_t s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
851 int16x4_t s4 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t4)));
852 int16x4_t s5 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t5)));
853 int16x4_t s6 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t6)));
854 int16x4_t s7 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t7)));
855 int16x4_t s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t8)));
856 int16x4_t s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t9)));
857 int16x4_t s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t10)));
858
859 src_ptr += 11 * src_stride;
860
861 do {
862 uint8x8_t t11, t12, t13, t14;
863 load_u8_8x4(src_ptr, src_stride, &t11, &t12, &t13, &t14);
864
865 int16x4_t s11 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t11)));
866 int16x4_t s12 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t12)));
867 int16x4_t s13 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t13)));
868 int16x4_t s14 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t14)));
869
870 int16x4_t d0 = convolve12_4_y(s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10,
871 s11, y_filter_0_7, y_filter_8_11);
872 int16x4_t d1 = convolve12_4_y(s1, s2, s3, s4, s5, s6, s7, s8, s9, s10,
873 s11, s12, y_filter_0_7, y_filter_8_11);
874 int16x4_t d2 = convolve12_4_y(s2, s3, s4, s5, s6, s7, s8, s9, s10, s11,
875 s12, s13, y_filter_0_7, y_filter_8_11);
876 int16x4_t d3 = convolve12_4_y(s3, s4, s5, s6, s7, s8, s9, s10, s11, s12,
877 s13, s14, y_filter_0_7, y_filter_8_11);
878
879 uint8x8_t d01 = vqrshrun_n_s16(vcombine_s16(d0, d1), FILTER_BITS);
880 uint8x8_t d23 = vqrshrun_n_s16(vcombine_s16(d2, d3), FILTER_BITS);
881
882 store_u8x4_strided_x2(dst_ptr, dst_stride, d01);
883 store_u8x4_strided_x2(dst_ptr + 2 * dst_stride, dst_stride, d23);
884
885 s0 = s4;
886 s1 = s5;
887 s2 = s6;
888 s3 = s7;
889 s4 = s8;
890 s5 = s9;
891 s6 = s10;
892 s7 = s11;
893 s8 = s12;
894 s9 = s13;
895 s10 = s14;
896 src_ptr += 4 * src_stride;
897 dst_ptr += 4 * dst_stride;
898 h -= 4;
899 } while (h != 0);
900
901 } else {
902 do {
903 const uint8_t *s = src_ptr;
904 uint8_t *d = dst_ptr;
905 int height = h;
906
907 uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7, t8, t9, t10;
908 load_u8_8x11(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7, &t8,
909 &t9, &t10);
910 int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
911 int16x8_t s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
912 int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
913 int16x8_t s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
914 int16x8_t s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
915 int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
916 int16x8_t s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
917 int16x8_t s7 = vreinterpretq_s16_u16(vmovl_u8(t7));
918 int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t8));
919 int16x8_t s9 = vreinterpretq_s16_u16(vmovl_u8(t9));
920 int16x8_t s10 = vreinterpretq_s16_u16(vmovl_u8(t10));
921
922 s += 11 * src_stride;
923
924 do {
925 uint8x8_t t11, t12, t13, t14;
926 load_u8_8x4(s, src_stride, &t11, &t12, &t13, &t14);
927
928 int16x8_t s11 = vreinterpretq_s16_u16(vmovl_u8(t11));
929 int16x8_t s12 = vreinterpretq_s16_u16(vmovl_u8(t12));
930 int16x8_t s13 = vreinterpretq_s16_u16(vmovl_u8(t13));
931 int16x8_t s14 = vreinterpretq_s16_u16(vmovl_u8(t14));
932
933 uint8x8_t d0 = convolve12_8_y(s0, s1, s2, s3, s4, s5, s6, s7, s8, s9,
934 s10, s11, y_filter_0_7, y_filter_8_11);
935 uint8x8_t d1 = convolve12_8_y(s1, s2, s3, s4, s5, s6, s7, s8, s9, s10,
936 s11, s12, y_filter_0_7, y_filter_8_11);
937 uint8x8_t d2 = convolve12_8_y(s2, s3, s4, s5, s6, s7, s8, s9, s10, s11,
938 s12, s13, y_filter_0_7, y_filter_8_11);
939 uint8x8_t d3 = convolve12_8_y(s3, s4, s5, s6, s7, s8, s9, s10, s11, s12,
940 s13, s14, y_filter_0_7, y_filter_8_11);
941
942 store_u8_8x4(d, dst_stride, d0, d1, d2, d3);
943
944 s0 = s4;
945 s1 = s5;
946 s2 = s6;
947 s3 = s7;
948 s4 = s8;
949 s5 = s9;
950 s6 = s10;
951 s7 = s11;
952 s8 = s12;
953 s9 = s13;
954 s10 = s14;
955 s += 4 * src_stride;
956 d += 4 * dst_stride;
957 height -= 4;
958 } while (height != 0);
959 src_ptr += 8;
960 dst_ptr += 8;
961 w -= 8;
962 } while (w != 0);
963 }
964 }
965
av1_convolve_y_sr_neon(const uint8_t * src,int src_stride,uint8_t * dst,int dst_stride,int w,int h,const InterpFilterParams * filter_params_y,const int subpel_y_qn)966 void av1_convolve_y_sr_neon(const uint8_t *src, int src_stride, uint8_t *dst,
967 int dst_stride, int w, int h,
968 const InterpFilterParams *filter_params_y,
969 const int subpel_y_qn) {
970 if (w == 2 || h == 2) {
971 av1_convolve_y_sr_c(src, src_stride, dst, dst_stride, w, h, filter_params_y,
972 subpel_y_qn);
973 return;
974 }
975
976 const int y_filter_taps = get_filter_tap(filter_params_y, subpel_y_qn);
977 const int clamped_y_taps = y_filter_taps < 6 ? 6 : y_filter_taps;
978 const int vert_offset = clamped_y_taps / 2 - 1;
979
980 src -= vert_offset * src_stride;
981
982 const int16_t *y_filter_ptr = av1_get_interp_filter_subpel_kernel(
983 filter_params_y, subpel_y_qn & SUBPEL_MASK);
984
985 if (y_filter_taps > 8) {
986 convolve_y_sr_12tap_neon(src, src_stride, dst, dst_stride, w, h,
987 y_filter_ptr);
988 return;
989 }
990
991 // Filter values are even so halve to reduce precision requirements.
992 const int16x8_t y_filter = vshrq_n_s16(vld1q_s16(y_filter_ptr), 1);
993
994 if (y_filter_taps < 8) {
995 convolve_y_sr_6tap_neon(src, src_stride, dst, dst_stride, w, h, y_filter);
996 } else {
997 convolve_y_sr_8tap_neon(src, src_stride, dst, dst_stride, w, h, y_filter);
998 }
999 }
1000
1001 static INLINE int16x4_t
convolve12_4_2d_h(const int16x4_t s0,const int16x4_t s1,const int16x4_t s2,const int16x4_t s3,const int16x4_t s4,const int16x4_t s5,const int16x4_t s6,const int16x4_t s7,const int16x4_t s8,const int16x4_t s9,const int16x4_t s10,const int16x4_t s11,const int16x8_t x_filter_0_7,const int16x4_t x_filter_8_11,const int32x4_t horiz_const)1002 convolve12_4_2d_h(const int16x4_t s0, const int16x4_t s1, const int16x4_t s2,
1003 const int16x4_t s3, const int16x4_t s4, const int16x4_t s5,
1004 const int16x4_t s6, const int16x4_t s7, const int16x4_t s8,
1005 const int16x4_t s9, const int16x4_t s10, const int16x4_t s11,
1006 const int16x8_t x_filter_0_7, const int16x4_t x_filter_8_11,
1007 const int32x4_t horiz_const) {
1008 const int16x4_t x_filter_0_3 = vget_low_s16(x_filter_0_7);
1009 const int16x4_t x_filter_4_7 = vget_high_s16(x_filter_0_7);
1010
1011 int32x4_t sum = horiz_const;
1012 sum = vmlal_lane_s16(sum, s0, x_filter_0_3, 0);
1013 sum = vmlal_lane_s16(sum, s1, x_filter_0_3, 1);
1014 sum = vmlal_lane_s16(sum, s2, x_filter_0_3, 2);
1015 sum = vmlal_lane_s16(sum, s3, x_filter_0_3, 3);
1016 sum = vmlal_lane_s16(sum, s4, x_filter_4_7, 0);
1017 sum = vmlal_lane_s16(sum, s5, x_filter_4_7, 1);
1018 sum = vmlal_lane_s16(sum, s6, x_filter_4_7, 2);
1019 sum = vmlal_lane_s16(sum, s7, x_filter_4_7, 3);
1020 sum = vmlal_lane_s16(sum, s8, x_filter_8_11, 0);
1021 sum = vmlal_lane_s16(sum, s9, x_filter_8_11, 1);
1022 sum = vmlal_lane_s16(sum, s10, x_filter_8_11, 2);
1023 sum = vmlal_lane_s16(sum, s11, x_filter_8_11, 3);
1024
1025 return vshrn_n_s32(sum, ROUND0_BITS);
1026 }
1027
convolve_2d_sr_horiz_12tap_neon(const uint8_t * src_ptr,int src_stride,int16_t * dst_ptr,const int dst_stride,int w,int h,const int16x8_t x_filter_0_7,const int16x4_t x_filter_8_11)1028 static INLINE void convolve_2d_sr_horiz_12tap_neon(
1029 const uint8_t *src_ptr, int src_stride, int16_t *dst_ptr,
1030 const int dst_stride, int w, int h, const int16x8_t x_filter_0_7,
1031 const int16x4_t x_filter_8_11) {
1032 const int bd = 8;
1033 // A shim of 1 << (ROUND0_BITS - 1) enables us to use non-rounding shifts -
1034 // which are generally faster than rounding shifts on modern CPUs.
1035 const int32x4_t horiz_const =
1036 vdupq_n_s32((1 << (bd + FILTER_BITS - 1)) + (1 << (ROUND0_BITS - 1)));
1037
1038 #if AOM_ARCH_AARCH64
1039 do {
1040 const uint8_t *s = src_ptr;
1041 int16_t *d = dst_ptr;
1042 int width = w;
1043
1044 uint8x8_t t0, t1, t2, t3;
1045 load_u8_8x4(s, src_stride, &t0, &t1, &t2, &t3);
1046 transpose_elems_inplace_u8_8x4(&t0, &t1, &t2, &t3);
1047
1048 int16x4_t s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
1049 int16x4_t s1 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
1050 int16x4_t s2 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
1051 int16x4_t s3 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
1052 int16x4_t s4 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
1053 int16x4_t s5 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
1054 int16x4_t s6 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
1055 int16x4_t s7 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
1056
1057 load_u8_8x4(s + 8, src_stride, &t0, &t1, &t2, &t3);
1058 transpose_elems_inplace_u8_8x4(&t0, &t1, &t2, &t3);
1059
1060 int16x4_t s8 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
1061 int16x4_t s9 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
1062 int16x4_t s10 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
1063
1064 s += 11;
1065
1066 do {
1067 load_u8_8x4(s, src_stride, &t0, &t1, &t2, &t3);
1068 transpose_elems_inplace_u8_8x4(&t0, &t1, &t2, &t3);
1069
1070 int16x4_t s11 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
1071 int16x4_t s12 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t1)));
1072 int16x4_t s13 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t2)));
1073 int16x4_t s14 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t3)));
1074
1075 int16x4_t d0 =
1076 convolve12_4_2d_h(s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11,
1077 x_filter_0_7, x_filter_8_11, horiz_const);
1078 int16x4_t d1 =
1079 convolve12_4_2d_h(s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12,
1080 x_filter_0_7, x_filter_8_11, horiz_const);
1081 int16x4_t d2 =
1082 convolve12_4_2d_h(s2, s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13,
1083 x_filter_0_7, x_filter_8_11, horiz_const);
1084 int16x4_t d3 =
1085 convolve12_4_2d_h(s3, s4, s5, s6, s7, s8, s9, s10, s11, s12, s13, s14,
1086 x_filter_0_7, x_filter_8_11, horiz_const);
1087
1088 transpose_elems_inplace_s16_4x4(&d0, &d1, &d2, &d3);
1089 store_s16_4x4(d, dst_stride, d0, d1, d2, d3);
1090
1091 s0 = s4;
1092 s1 = s5;
1093 s2 = s6;
1094 s3 = s7;
1095 s4 = s8;
1096 s5 = s9;
1097 s6 = s10;
1098 s7 = s11;
1099 s8 = s12;
1100 s9 = s13;
1101 s10 = s14;
1102 s += 4;
1103 d += 4;
1104 width -= 4;
1105 } while (width != 0);
1106 src_ptr += 4 * src_stride;
1107 dst_ptr += 4 * dst_stride;
1108 h -= 4;
1109 } while (h > 4);
1110 #endif // AOM_ARCH_AARCH64
1111
1112 do {
1113 const uint8_t *s = src_ptr;
1114 int16_t *d = dst_ptr;
1115 int width = w;
1116
1117 do {
1118 uint8x16_t t0 = vld1q_u8(s);
1119 int16x8_t tt0 = vreinterpretq_s16_u16(vmovl_u8(vget_low_u8(t0)));
1120 int16x8_t tt1 = vreinterpretq_s16_u16(vmovl_u8(vget_high_u8(t0)));
1121
1122 int16x4_t s0 = vget_low_s16(tt0);
1123 int16x4_t s4 = vget_high_s16(tt0);
1124 int16x4_t s8 = vget_low_s16(tt1);
1125 int16x4_t s12 = vget_high_s16(tt1);
1126
1127 int16x4_t s1 = vext_s16(s0, s4, 1); // a1 a2 a3 a4
1128 int16x4_t s2 = vext_s16(s0, s4, 2); // a2 a3 a4 a5
1129 int16x4_t s3 = vext_s16(s0, s4, 3); // a3 a4 a5 a6
1130 int16x4_t s5 = vext_s16(s4, s8, 1); // a5 a6 a7 a8
1131 int16x4_t s6 = vext_s16(s4, s8, 2); // a6 a7 a8 a9
1132 int16x4_t s7 = vext_s16(s4, s8, 3); // a7 a8 a9 a10
1133 int16x4_t s9 = vext_s16(s8, s12, 1); // a9 a10 a11 a12
1134 int16x4_t s10 = vext_s16(s8, s12, 2); // a10 a11 a12 a13
1135 int16x4_t s11 = vext_s16(s8, s12, 3); // a11 a12 a13 a14
1136
1137 int16x4_t d0 =
1138 convolve12_4_2d_h(s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10, s11,
1139 x_filter_0_7, x_filter_8_11, horiz_const);
1140 vst1_s16(d, d0);
1141
1142 s += 4;
1143 d += 4;
1144 width -= 4;
1145 } while (width != 0);
1146 src_ptr += src_stride;
1147 dst_ptr += dst_stride;
1148 } while (--h != 0);
1149 }
1150
convolve4_4_2d_h(const int16x4_t s0,const int16x4_t s1,const int16x4_t s2,const int16x4_t s3,const int16x4_t filter,const int16x4_t horiz_const)1151 static INLINE int16x4_t convolve4_4_2d_h(const int16x4_t s0, const int16x4_t s1,
1152 const int16x4_t s2, const int16x4_t s3,
1153 const int16x4_t filter,
1154 const int16x4_t horiz_const) {
1155 int16x4_t sum = horiz_const;
1156 sum = vmla_lane_s16(sum, s0, filter, 0);
1157 sum = vmla_lane_s16(sum, s1, filter, 1);
1158 sum = vmla_lane_s16(sum, s2, filter, 2);
1159 sum = vmla_lane_s16(sum, s3, filter, 3);
1160
1161 // We halved the convolution filter values so -1 from the right shift.
1162 return vshr_n_s16(sum, ROUND0_BITS - 1);
1163 }
1164
convolve8_8_2d_h(const int16x8_t s0,const int16x8_t s1,const int16x8_t s2,const int16x8_t s3,const int16x8_t s4,const int16x8_t s5,const int16x8_t s6,const int16x8_t s7,const int16x8_t filter,const int16x8_t horiz_const)1165 static INLINE int16x8_t convolve8_8_2d_h(const int16x8_t s0, const int16x8_t s1,
1166 const int16x8_t s2, const int16x8_t s3,
1167 const int16x8_t s4, const int16x8_t s5,
1168 const int16x8_t s6, const int16x8_t s7,
1169 const int16x8_t filter,
1170 const int16x8_t horiz_const) {
1171 const int16x4_t filter_lo = vget_low_s16(filter);
1172 const int16x4_t filter_hi = vget_high_s16(filter);
1173
1174 int16x8_t sum = horiz_const;
1175 sum = vmlaq_lane_s16(sum, s0, filter_lo, 0);
1176 sum = vmlaq_lane_s16(sum, s1, filter_lo, 1);
1177 sum = vmlaq_lane_s16(sum, s2, filter_lo, 2);
1178 sum = vmlaq_lane_s16(sum, s3, filter_lo, 3);
1179 sum = vmlaq_lane_s16(sum, s4, filter_hi, 0);
1180 sum = vmlaq_lane_s16(sum, s5, filter_hi, 1);
1181 sum = vmlaq_lane_s16(sum, s6, filter_hi, 2);
1182 sum = vmlaq_lane_s16(sum, s7, filter_hi, 3);
1183
1184 // We halved the convolution filter values so -1 from the right shift.
1185 return vshrq_n_s16(sum, ROUND0_BITS - 1);
1186 }
1187
convolve_2d_sr_horiz_neon(const uint8_t * src,int src_stride,int16_t * im_block,int im_stride,int w,int im_h,const int16_t * x_filter_ptr)1188 static INLINE void convolve_2d_sr_horiz_neon(const uint8_t *src, int src_stride,
1189 int16_t *im_block, int im_stride,
1190 int w, int im_h,
1191 const int16_t *x_filter_ptr) {
1192 const int bd = 8;
1193
1194 const uint8_t *src_ptr = src;
1195 int16_t *dst_ptr = im_block;
1196 int dst_stride = im_stride;
1197 int height = im_h;
1198
1199 if (w <= 4) {
1200 // A shim of 1 << ((ROUND0_BITS - 1) - 1) enables us to use non-rounding
1201 // shifts - which are generally faster than rounding shifts on modern CPUs.
1202 // (The extra -1 is needed because we halved the filter values.)
1203 const int16x4_t horiz_const = vdup_n_s16((1 << (bd + FILTER_BITS - 2)) +
1204 (1 << ((ROUND0_BITS - 1) - 1)));
1205 // 4-tap filters are used for blocks having width <= 4.
1206 // Filter values are even, so halve to reduce intermediate precision reqs.
1207 const int16x4_t x_filter = vshr_n_s16(vld1_s16(x_filter_ptr + 2), 1);
1208
1209 src_ptr += 2;
1210
1211 do {
1212 uint8x8_t t0 = vld1_u8(src_ptr); // a0 a1 a2 a3 a4 a5 a6 a7
1213 int16x4_t s0 = vget_low_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
1214 int16x4_t s4 = vget_high_s16(vreinterpretq_s16_u16(vmovl_u8(t0)));
1215
1216 int16x4_t s1 = vext_s16(s0, s4, 1); // a1 a2 a3 a4
1217 int16x4_t s2 = vext_s16(s0, s4, 2); // a2 a3 a4 a5
1218 int16x4_t s3 = vext_s16(s0, s4, 3); // a3 a4 a5 a6
1219
1220 int16x4_t d0 = convolve4_4_2d_h(s0, s1, s2, s3, x_filter, horiz_const);
1221
1222 vst1_s16(dst_ptr, d0);
1223
1224 src_ptr += src_stride;
1225 dst_ptr += dst_stride;
1226 } while (--height != 0);
1227 } else {
1228 // A shim of 1 << ((ROUND0_BITS - 1) - 1) enables us to use non-rounding
1229 // shifts - which are generally faster than rounding shifts on modern CPUs.
1230 // (The extra -1 is needed because we halved the filter values.)
1231 const int16x8_t horiz_const = vdupq_n_s16((1 << (bd + FILTER_BITS - 2)) +
1232 (1 << ((ROUND0_BITS - 1) - 1)));
1233 // Filter values are even, so halve to reduce intermediate precision reqs.
1234 const int16x8_t x_filter = vshrq_n_s16(vld1q_s16(x_filter_ptr), 1);
1235
1236 #if AOM_ARCH_AARCH64
1237 while (height > 8) {
1238 const uint8_t *s = src_ptr;
1239 int16_t *d = dst_ptr;
1240 int width = w;
1241
1242 uint8x8_t t0, t1, t2, t3, t4, t5, t6, t7;
1243 load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
1244 transpose_elems_inplace_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
1245
1246 int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
1247 int16x8_t s1 = vreinterpretq_s16_u16(vmovl_u8(t1));
1248 int16x8_t s2 = vreinterpretq_s16_u16(vmovl_u8(t2));
1249 int16x8_t s3 = vreinterpretq_s16_u16(vmovl_u8(t3));
1250 int16x8_t s4 = vreinterpretq_s16_u16(vmovl_u8(t4));
1251 int16x8_t s5 = vreinterpretq_s16_u16(vmovl_u8(t5));
1252 int16x8_t s6 = vreinterpretq_s16_u16(vmovl_u8(t6));
1253
1254 s += 7;
1255
1256 do {
1257 load_u8_8x8(s, src_stride, &t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
1258
1259 transpose_elems_inplace_u8_8x8(&t0, &t1, &t2, &t3, &t4, &t5, &t6, &t7);
1260
1261 int16x8_t s7 = vreinterpretq_s16_u16(vmovl_u8(t0));
1262 int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
1263 int16x8_t s9 = vreinterpretq_s16_u16(vmovl_u8(t2));
1264 int16x8_t s10 = vreinterpretq_s16_u16(vmovl_u8(t3));
1265 int16x8_t s11 = vreinterpretq_s16_u16(vmovl_u8(t4));
1266 int16x8_t s12 = vreinterpretq_s16_u16(vmovl_u8(t5));
1267 int16x8_t s13 = vreinterpretq_s16_u16(vmovl_u8(t6));
1268 int16x8_t s14 = vreinterpretq_s16_u16(vmovl_u8(t7));
1269
1270 int16x8_t d0 = convolve8_8_2d_h(s0, s1, s2, s3, s4, s5, s6, s7,
1271 x_filter, horiz_const);
1272 int16x8_t d1 = convolve8_8_2d_h(s1, s2, s3, s4, s5, s6, s7, s8,
1273 x_filter, horiz_const);
1274 int16x8_t d2 = convolve8_8_2d_h(s2, s3, s4, s5, s6, s7, s8, s9,
1275 x_filter, horiz_const);
1276 int16x8_t d3 = convolve8_8_2d_h(s3, s4, s5, s6, s7, s8, s9, s10,
1277 x_filter, horiz_const);
1278 int16x8_t d4 = convolve8_8_2d_h(s4, s5, s6, s7, s8, s9, s10, s11,
1279 x_filter, horiz_const);
1280 int16x8_t d5 = convolve8_8_2d_h(s5, s6, s7, s8, s9, s10, s11, s12,
1281 x_filter, horiz_const);
1282 int16x8_t d6 = convolve8_8_2d_h(s6, s7, s8, s9, s10, s11, s12, s13,
1283 x_filter, horiz_const);
1284 int16x8_t d7 = convolve8_8_2d_h(s7, s8, s9, s10, s11, s12, s13, s14,
1285 x_filter, horiz_const);
1286
1287 transpose_elems_inplace_s16_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
1288
1289 store_s16_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7);
1290
1291 s0 = s8;
1292 s1 = s9;
1293 s2 = s10;
1294 s3 = s11;
1295 s4 = s12;
1296 s5 = s13;
1297 s6 = s14;
1298 s += 8;
1299 d += 8;
1300 width -= 8;
1301 } while (width != 0);
1302 src_ptr += 8 * src_stride;
1303 dst_ptr += 8 * dst_stride;
1304 height -= 8;
1305 }
1306 #endif // AOM_ARCH_AARCH64
1307
1308 do {
1309 const uint8_t *s = src_ptr;
1310 int16_t *d = dst_ptr;
1311 int width = w;
1312
1313 uint8x8_t t0 = vld1_u8(s); // a0 a1 a2 a3 a4 a5 a6 a7
1314 int16x8_t s0 = vreinterpretq_s16_u16(vmovl_u8(t0));
1315
1316 do {
1317 uint8x8_t t1 = vld1_u8(s + 8); // a8 a9 a10 a11 a12 a13 a14 a15
1318 int16x8_t s8 = vreinterpretq_s16_u16(vmovl_u8(t1));
1319
1320 int16x8_t s1 = vextq_s16(s0, s8, 1); // a1 a2 a3 a4 a5 a6 a7 a8
1321 int16x8_t s2 = vextq_s16(s0, s8, 2); // a2 a3 a4 a5 a6 a7 a8 a9
1322 int16x8_t s3 = vextq_s16(s0, s8, 3); // a3 a4 a5 a6 a7 a8 a9 a10
1323 int16x8_t s4 = vextq_s16(s0, s8, 4); // a4 a5 a6 a7 a8 a9 a10 a11
1324 int16x8_t s5 = vextq_s16(s0, s8, 5); // a5 a6 a7 a8 a9 a10 a11 a12
1325 int16x8_t s6 = vextq_s16(s0, s8, 6); // a6 a7 a8 a9 a10 a11 a12 a13
1326 int16x8_t s7 = vextq_s16(s0, s8, 7); // a7 a8 a9 a10 a11 a12 a13 a14
1327
1328 int16x8_t d0 = convolve8_8_2d_h(s0, s1, s2, s3, s4, s5, s6, s7,
1329 x_filter, horiz_const);
1330
1331 vst1q_s16(d, d0);
1332
1333 s0 = s8;
1334 s += 8;
1335 d += 8;
1336 width -= 8;
1337 } while (width != 0);
1338 src_ptr += src_stride;
1339 dst_ptr += dst_stride;
1340 } while (--height != 0);
1341 }
1342 }
1343
av1_convolve_2d_sr_neon(const uint8_t * src,int src_stride,uint8_t * dst,int dst_stride,int w,int h,const InterpFilterParams * filter_params_x,const InterpFilterParams * filter_params_y,const int subpel_x_qn,const int subpel_y_qn,ConvolveParams * conv_params)1344 void av1_convolve_2d_sr_neon(const uint8_t *src, int src_stride, uint8_t *dst,
1345 int dst_stride, int w, int h,
1346 const InterpFilterParams *filter_params_x,
1347 const InterpFilterParams *filter_params_y,
1348 const int subpel_x_qn, const int subpel_y_qn,
1349 ConvolveParams *conv_params) {
1350 if (w == 2 || h == 2) {
1351 av1_convolve_2d_sr_c(src, src_stride, dst, dst_stride, w, h,
1352 filter_params_x, filter_params_y, subpel_x_qn,
1353 subpel_y_qn, conv_params);
1354 return;
1355 }
1356
1357 const int y_filter_taps = get_filter_tap(filter_params_y, subpel_y_qn);
1358 const int clamped_y_taps = y_filter_taps < 6 ? 6 : y_filter_taps;
1359 const int im_h = h + clamped_y_taps - 1;
1360 const int im_stride = MAX_SB_SIZE;
1361 const int vert_offset = clamped_y_taps / 2 - 1;
1362 const int horiz_offset = filter_params_x->taps / 2 - 1;
1363 const uint8_t *src_ptr = src - vert_offset * src_stride - horiz_offset;
1364
1365 const int16_t *x_filter_ptr = av1_get_interp_filter_subpel_kernel(
1366 filter_params_x, subpel_x_qn & SUBPEL_MASK);
1367 const int16_t *y_filter_ptr = av1_get_interp_filter_subpel_kernel(
1368 filter_params_y, subpel_y_qn & SUBPEL_MASK);
1369
1370 if (filter_params_x->taps > 8) {
1371 DECLARE_ALIGNED(16, int16_t,
1372 im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE]);
1373
1374 const int16x8_t x_filter_0_7 = vld1q_s16(x_filter_ptr);
1375 const int16x4_t x_filter_8_11 = vld1_s16(x_filter_ptr + 8);
1376 const int16x8_t y_filter_0_7 = vld1q_s16(y_filter_ptr);
1377 const int16x4_t y_filter_8_11 = vld1_s16(y_filter_ptr + 8);
1378
1379 convolve_2d_sr_horiz_12tap_neon(src_ptr, src_stride, im_block, im_stride, w,
1380 im_h, x_filter_0_7, x_filter_8_11);
1381
1382 convolve_2d_sr_vert_12tap_neon(im_block, im_stride, dst, dst_stride, w, h,
1383 y_filter_0_7, y_filter_8_11);
1384 } else {
1385 DECLARE_ALIGNED(16, int16_t,
1386 im_block[(MAX_SB_SIZE + SUBPEL_TAPS - 1) * MAX_SB_SIZE]);
1387
1388 convolve_2d_sr_horiz_neon(src_ptr, src_stride, im_block, im_stride, w, im_h,
1389 x_filter_ptr);
1390
1391 const int16x8_t y_filter = vld1q_s16(y_filter_ptr);
1392
1393 if (clamped_y_taps <= 6) {
1394 convolve_2d_sr_vert_6tap_neon(im_block, im_stride, dst, dst_stride, w, h,
1395 y_filter);
1396 } else {
1397 convolve_2d_sr_vert_8tap_neon(im_block, im_stride, dst, dst_stride, w, h,
1398 y_filter);
1399 }
1400 }
1401 }
1402
av1_convolve_x_sr_intrabc_neon(const uint8_t * src,int src_stride,uint8_t * dst,int dst_stride,int w,int h,const InterpFilterParams * filter_params_x,const int subpel_x_qn,ConvolveParams * conv_params)1403 void av1_convolve_x_sr_intrabc_neon(const uint8_t *src, int src_stride,
1404 uint8_t *dst, int dst_stride, int w, int h,
1405 const InterpFilterParams *filter_params_x,
1406 const int subpel_x_qn,
1407 ConvolveParams *conv_params) {
1408 assert(subpel_x_qn == 8);
1409 assert(filter_params_x->taps == 2);
1410 assert((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS);
1411 (void)filter_params_x;
1412 (void)subpel_x_qn;
1413 (void)conv_params;
1414
1415 if (w <= 4) {
1416 do {
1417 uint8x8_t s0_0 = vld1_u8(src);
1418 uint8x8_t s0_1 = vld1_u8(src + 1);
1419 uint8x8_t s1_0 = vld1_u8(src + src_stride);
1420 uint8x8_t s1_1 = vld1_u8(src + src_stride + 1);
1421
1422 uint8x8_t d0 = vrhadd_u8(s0_0, s0_1);
1423 uint8x8_t d1 = vrhadd_u8(s1_0, s1_1);
1424
1425 if (w == 2) {
1426 store_u8_2x1(dst + 0 * dst_stride, d0);
1427 store_u8_2x1(dst + 1 * dst_stride, d1);
1428 } else {
1429 store_u8_4x1(dst + 0 * dst_stride, d0);
1430 store_u8_4x1(dst + 1 * dst_stride, d1);
1431 }
1432
1433 src += 2 * src_stride;
1434 dst += 2 * dst_stride;
1435 h -= 2;
1436 } while (h != 0);
1437 } else if (w == 8) {
1438 do {
1439 uint8x8_t s0_0 = vld1_u8(src);
1440 uint8x8_t s0_1 = vld1_u8(src + 1);
1441 uint8x8_t s1_0 = vld1_u8(src + src_stride);
1442 uint8x8_t s1_1 = vld1_u8(src + src_stride + 1);
1443
1444 uint8x8_t d0 = vrhadd_u8(s0_0, s0_1);
1445 uint8x8_t d1 = vrhadd_u8(s1_0, s1_1);
1446
1447 vst1_u8(dst, d0);
1448 vst1_u8(dst + dst_stride, d1);
1449
1450 src += 2 * src_stride;
1451 dst += 2 * dst_stride;
1452 h -= 2;
1453 } while (h != 0);
1454 } else {
1455 do {
1456 const uint8_t *src_ptr = src;
1457 uint8_t *dst_ptr = dst;
1458 int width = w;
1459
1460 do {
1461 uint8x16_t s0 = vld1q_u8(src_ptr);
1462 uint8x16_t s1 = vld1q_u8(src_ptr + 1);
1463
1464 uint8x16_t d0 = vrhaddq_u8(s0, s1);
1465
1466 vst1q_u8(dst_ptr, d0);
1467
1468 src_ptr += 16;
1469 dst_ptr += 16;
1470 width -= 16;
1471 } while (width != 0);
1472 src += src_stride;
1473 dst += dst_stride;
1474 } while (--h != 0);
1475 }
1476 }
1477
av1_convolve_y_sr_intrabc_neon(const uint8_t * src,int src_stride,uint8_t * dst,int dst_stride,int w,int h,const InterpFilterParams * filter_params_y,const int subpel_y_qn)1478 void av1_convolve_y_sr_intrabc_neon(const uint8_t *src, int src_stride,
1479 uint8_t *dst, int dst_stride, int w, int h,
1480 const InterpFilterParams *filter_params_y,
1481 const int subpel_y_qn) {
1482 assert(subpel_y_qn == 8);
1483 assert(filter_params_y->taps == 2);
1484 (void)filter_params_y;
1485 (void)subpel_y_qn;
1486
1487 if (w <= 4) {
1488 do {
1489 uint8x8_t s0 = load_unaligned_u8_4x1(src);
1490 uint8x8_t s1 = load_unaligned_u8_4x1(src + src_stride);
1491 uint8x8_t s2 = load_unaligned_u8_4x1(src + 2 * src_stride);
1492
1493 uint8x8_t d0 = vrhadd_u8(s0, s1);
1494 uint8x8_t d1 = vrhadd_u8(s1, s2);
1495
1496 if (w == 2) {
1497 store_u8_2x1(dst + 0 * dst_stride, d0);
1498 store_u8_2x1(dst + 1 * dst_stride, d1);
1499 } else {
1500 store_u8_4x1(dst + 0 * dst_stride, d0);
1501 store_u8_4x1(dst + 1 * dst_stride, d1);
1502 }
1503
1504 src += 2 * src_stride;
1505 dst += 2 * dst_stride;
1506 h -= 2;
1507 } while (h != 0);
1508 } else if (w == 8) {
1509 do {
1510 uint8x8_t s0 = vld1_u8(src);
1511 uint8x8_t s1 = vld1_u8(src + src_stride);
1512 uint8x8_t s2 = vld1_u8(src + 2 * src_stride);
1513
1514 uint8x8_t d0 = vrhadd_u8(s0, s1);
1515 uint8x8_t d1 = vrhadd_u8(s1, s2);
1516
1517 vst1_u8(dst, d0);
1518 vst1_u8(dst + dst_stride, d1);
1519
1520 src += 2 * src_stride;
1521 dst += 2 * dst_stride;
1522 h -= 2;
1523 } while (h != 0);
1524 } else {
1525 do {
1526 const uint8_t *src_ptr = src;
1527 uint8_t *dst_ptr = dst;
1528 int height = h;
1529
1530 do {
1531 uint8x16_t s0 = vld1q_u8(src_ptr);
1532 uint8x16_t s1 = vld1q_u8(src_ptr + src_stride);
1533
1534 uint8x16_t d0 = vrhaddq_u8(s0, s1);
1535
1536 vst1q_u8(dst_ptr, d0);
1537
1538 src_ptr += src_stride;
1539 dst_ptr += dst_stride;
1540 } while (--height != 0);
1541 src += 16;
1542 dst += 16;
1543 w -= 16;
1544 } while (w != 0);
1545 }
1546 }
1547
av1_convolve_2d_sr_intrabc_neon(const uint8_t * src,int src_stride,uint8_t * dst,int dst_stride,int w,int h,const InterpFilterParams * filter_params_x,const InterpFilterParams * filter_params_y,const int subpel_x_qn,const int subpel_y_qn,ConvolveParams * conv_params)1548 void av1_convolve_2d_sr_intrabc_neon(const uint8_t *src, int src_stride,
1549 uint8_t *dst, int dst_stride, int w, int h,
1550 const InterpFilterParams *filter_params_x,
1551 const InterpFilterParams *filter_params_y,
1552 const int subpel_x_qn,
1553 const int subpel_y_qn,
1554 ConvolveParams *conv_params) {
1555 assert(subpel_x_qn == 8);
1556 assert(subpel_y_qn == 8);
1557 assert(filter_params_x->taps == 2 && filter_params_y->taps == 2);
1558 assert((conv_params->round_0 + conv_params->round_1) == 2 * FILTER_BITS);
1559 (void)filter_params_x;
1560 (void)subpel_x_qn;
1561 (void)filter_params_y;
1562 (void)subpel_y_qn;
1563 (void)conv_params;
1564
1565 uint16_t im_block[(MAX_SB_SIZE + MAX_FILTER_TAP - 1) * MAX_SB_SIZE];
1566 int im_h = h + 1;
1567 int im_stride = w;
1568 assert(w <= MAX_SB_SIZE && h <= MAX_SB_SIZE);
1569
1570 uint16_t *im = im_block;
1571
1572 // Horizontal filter.
1573 if (w <= 4) {
1574 do {
1575 uint8x8_t s0 = vld1_u8(src);
1576 uint8x8_t s1 = vld1_u8(src + 1);
1577
1578 uint16x4_t sum = vget_low_u16(vaddl_u8(s0, s1));
1579
1580 // Safe to store the whole vector, the im buffer is big enough.
1581 vst1_u16(im, sum);
1582
1583 src += src_stride;
1584 im += im_stride;
1585 } while (--im_h != 0);
1586 } else {
1587 do {
1588 const uint8_t *src_ptr = src;
1589 uint16_t *im_ptr = im;
1590 int width = w;
1591
1592 do {
1593 uint8x8_t s0 = vld1_u8(src_ptr);
1594 uint8x8_t s1 = vld1_u8(src_ptr + 1);
1595
1596 uint16x8_t sum = vaddl_u8(s0, s1);
1597
1598 vst1q_u16(im_ptr, sum);
1599
1600 src_ptr += 8;
1601 im_ptr += 8;
1602 width -= 8;
1603 } while (width != 0);
1604 src += src_stride;
1605 im += im_stride;
1606 } while (--im_h != 0);
1607 }
1608
1609 im = im_block;
1610
1611 // Vertical filter.
1612 if (w <= 4) {
1613 do {
1614 uint16x4_t s0 = vld1_u16(im);
1615 uint16x4_t s1 = vld1_u16(im + im_stride);
1616 uint16x4_t s2 = vld1_u16(im + 2 * im_stride);
1617
1618 uint16x4_t sum0 = vadd_u16(s0, s1);
1619 uint16x4_t sum1 = vadd_u16(s1, s2);
1620
1621 uint8x8_t d0 = vqrshrn_n_u16(vcombine_u16(sum0, vdup_n_u16(0)), 2);
1622 uint8x8_t d1 = vqrshrn_n_u16(vcombine_u16(sum1, vdup_n_u16(0)), 2);
1623
1624 if (w == 2) {
1625 store_u8_2x1(dst + 0 * dst_stride, d0);
1626 store_u8_2x1(dst + 1 * dst_stride, d1);
1627 } else {
1628 store_u8_4x1(dst + 0 * dst_stride, d0);
1629 store_u8_4x1(dst + 1 * dst_stride, d1);
1630 }
1631
1632 im += 2 * im_stride;
1633 dst += 2 * dst_stride;
1634 h -= 2;
1635 } while (h != 0);
1636 } else {
1637 do {
1638 uint16_t *im_ptr = im;
1639 uint8_t *dst_ptr = dst;
1640 int height = h;
1641
1642 do {
1643 uint16x8_t s0 = vld1q_u16(im_ptr);
1644 uint16x8_t s1 = vld1q_u16(im_ptr + im_stride);
1645
1646 uint16x8_t sum = vaddq_u16(s0, s1);
1647 uint8x8_t d0 = vqrshrn_n_u16(sum, 2);
1648
1649 vst1_u8(dst_ptr, d0);
1650
1651 im_ptr += im_stride;
1652 dst_ptr += dst_stride;
1653 } while (--height != 0);
1654 im += 8;
1655 dst += 8;
1656 w -= 8;
1657 } while (w != 0);
1658 }
1659 }
1660