1 /*
2 * Copyright (c) 2024, Alliance for Open Media. All rights reserved.
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include <arm_neon.h>
13 #include <assert.h>
14 #include <stdint.h>
15
16 #include "config/aom_config.h"
17 #include "config/aom_dsp_rtcd.h"
18
19 #include "aom_dsp/arm/aom_neon_sve_bridge.h"
20 #include "aom_dsp/arm/aom_filter.h"
21 #include "aom_dsp/arm/highbd_convolve8_neon.h"
22 #include "aom_dsp/arm/mem_neon.h"
23
highbd_convolve8_4_h(int16x8_t s[4],int16x8_t filter,uint16x4_t max)24 static inline uint16x4_t highbd_convolve8_4_h(int16x8_t s[4], int16x8_t filter,
25 uint16x4_t max) {
26 int64x2_t sum[4];
27
28 sum[0] = aom_sdotq_s16(vdupq_n_s64(0), s[0], filter);
29 sum[1] = aom_sdotq_s16(vdupq_n_s64(0), s[1], filter);
30 sum[2] = aom_sdotq_s16(vdupq_n_s64(0), s[2], filter);
31 sum[3] = aom_sdotq_s16(vdupq_n_s64(0), s[3], filter);
32
33 int64x2_t sum01 = vpaddq_s64(sum[0], sum[1]);
34 int64x2_t sum23 = vpaddq_s64(sum[2], sum[3]);
35
36 int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23));
37
38 uint16x4_t res = vqrshrun_n_s32(sum0123, FILTER_BITS);
39 return vmin_u16(res, max);
40 }
41
highbd_convolve8_8_h(int16x8_t s[8],int16x8_t filter,uint16x8_t max)42 static inline uint16x8_t highbd_convolve8_8_h(int16x8_t s[8], int16x8_t filter,
43 uint16x8_t max) {
44 int64x2_t sum[8];
45
46 sum[0] = aom_sdotq_s16(vdupq_n_s64(0), s[0], filter);
47 sum[1] = aom_sdotq_s16(vdupq_n_s64(0), s[1], filter);
48 sum[2] = aom_sdotq_s16(vdupq_n_s64(0), s[2], filter);
49 sum[3] = aom_sdotq_s16(vdupq_n_s64(0), s[3], filter);
50 sum[4] = aom_sdotq_s16(vdupq_n_s64(0), s[4], filter);
51 sum[5] = aom_sdotq_s16(vdupq_n_s64(0), s[5], filter);
52 sum[6] = aom_sdotq_s16(vdupq_n_s64(0), s[6], filter);
53 sum[7] = aom_sdotq_s16(vdupq_n_s64(0), s[7], filter);
54
55 int64x2_t sum01 = vpaddq_s64(sum[0], sum[1]);
56 int64x2_t sum23 = vpaddq_s64(sum[2], sum[3]);
57 int64x2_t sum45 = vpaddq_s64(sum[4], sum[5]);
58 int64x2_t sum67 = vpaddq_s64(sum[6], sum[7]);
59
60 int32x4_t sum0123 = vcombine_s32(vmovn_s64(sum01), vmovn_s64(sum23));
61 int32x4_t sum4567 = vcombine_s32(vmovn_s64(sum45), vmovn_s64(sum67));
62
63 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(sum0123, FILTER_BITS),
64 vqrshrun_n_s32(sum4567, FILTER_BITS));
65 return vminq_u16(res, max);
66 }
67
highbd_convolve8_horiz_8tap_sve(const uint16_t * src,ptrdiff_t src_stride,uint16_t * dst,ptrdiff_t dst_stride,const int16_t * filter_x,int width,int height,int bd)68 static inline void highbd_convolve8_horiz_8tap_sve(
69 const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst,
70 ptrdiff_t dst_stride, const int16_t *filter_x, int width, int height,
71 int bd) {
72 const int16x8_t filter = vld1q_s16(filter_x);
73
74 if (width == 4) {
75 const uint16x4_t max = vdup_n_u16((1 << bd) - 1);
76 const int16_t *s = (const int16_t *)src;
77 uint16_t *d = dst;
78
79 do {
80 int16x8_t s0[4], s1[4], s2[4], s3[4];
81 load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]);
82 load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]);
83 load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]);
84 load_s16_8x4(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3]);
85
86 uint16x4_t d0 = highbd_convolve8_4_h(s0, filter, max);
87 uint16x4_t d1 = highbd_convolve8_4_h(s1, filter, max);
88 uint16x4_t d2 = highbd_convolve8_4_h(s2, filter, max);
89 uint16x4_t d3 = highbd_convolve8_4_h(s3, filter, max);
90
91 store_u16_4x4(d, dst_stride, d0, d1, d2, d3);
92
93 s += 4 * src_stride;
94 d += 4 * dst_stride;
95 height -= 4;
96 } while (height > 0);
97 } else {
98 do {
99 const uint16x8_t max = vdupq_n_u16((1 << bd) - 1);
100 const int16_t *s = (const int16_t *)src;
101 uint16_t *d = dst;
102 int w = width;
103
104 do {
105 int16x8_t s0[8], s1[8], s2[8], s3[8];
106 load_s16_8x8(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3],
107 &s0[4], &s0[5], &s0[6], &s0[7]);
108 load_s16_8x8(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3],
109 &s1[4], &s1[5], &s1[6], &s1[7]);
110 load_s16_8x8(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3],
111 &s2[4], &s2[5], &s2[6], &s2[7]);
112 load_s16_8x8(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3],
113 &s3[4], &s3[5], &s3[6], &s3[7]);
114
115 uint16x8_t d0 = highbd_convolve8_8_h(s0, filter, max);
116 uint16x8_t d1 = highbd_convolve8_8_h(s1, filter, max);
117 uint16x8_t d2 = highbd_convolve8_8_h(s2, filter, max);
118 uint16x8_t d3 = highbd_convolve8_8_h(s3, filter, max);
119
120 store_u16_8x4(d, dst_stride, d0, d1, d2, d3);
121
122 s += 8;
123 d += 8;
124 w -= 8;
125 } while (w != 0);
126 src += 4 * src_stride;
127 dst += 4 * dst_stride;
128 height -= 4;
129 } while (height > 0);
130 }
131 }
132
133 // clang-format off
134 DECLARE_ALIGNED(16, static const uint16_t, kDotProdTbl[16]) = {
135 0, 1, 2, 3, 1, 2, 3, 4, 2, 3, 4, 5, 3, 4, 5, 6,
136 };
137
138 DECLARE_ALIGNED(16, static const uint16_t, kDeinterleaveTbl[8]) = {
139 0, 2, 4, 6, 1, 3, 5, 7,
140 };
141 // clang-format on
142
highbd_convolve4_4_h(int16x8_t s,int16x8_t filter,uint16x8x2_t permute_tbl,uint16x4_t max)143 static inline uint16x4_t highbd_convolve4_4_h(int16x8_t s, int16x8_t filter,
144 uint16x8x2_t permute_tbl,
145 uint16x4_t max) {
146 int16x8_t permuted_samples0 = aom_tbl_s16(s, permute_tbl.val[0]);
147 int16x8_t permuted_samples1 = aom_tbl_s16(s, permute_tbl.val[1]);
148
149 int64x2_t sum0 =
150 aom_svdot_lane_s16(vdupq_n_s64(0), permuted_samples0, filter, 0);
151 int64x2_t sum1 =
152 aom_svdot_lane_s16(vdupq_n_s64(0), permuted_samples1, filter, 0);
153
154 int32x4_t res_s32 = vcombine_s32(vmovn_s64(sum0), vmovn_s64(sum1));
155 uint16x4_t res = vqrshrun_n_s32(res_s32, FILTER_BITS);
156
157 return vmin_u16(res, max);
158 }
159
highbd_convolve4_8_h(int16x8_t s[4],int16x8_t filter,uint16x8_t idx,uint16x8_t max)160 static inline uint16x8_t highbd_convolve4_8_h(int16x8_t s[4], int16x8_t filter,
161 uint16x8_t idx, uint16x8_t max) {
162 int64x2_t sum04 = aom_svdot_lane_s16(vdupq_n_s64(0), s[0], filter, 0);
163 int64x2_t sum15 = aom_svdot_lane_s16(vdupq_n_s64(0), s[1], filter, 0);
164 int64x2_t sum26 = aom_svdot_lane_s16(vdupq_n_s64(0), s[2], filter, 0);
165 int64x2_t sum37 = aom_svdot_lane_s16(vdupq_n_s64(0), s[3], filter, 0);
166
167 int32x4_t res0 = vcombine_s32(vmovn_s64(sum04), vmovn_s64(sum15));
168 int32x4_t res1 = vcombine_s32(vmovn_s64(sum26), vmovn_s64(sum37));
169
170 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(res0, FILTER_BITS),
171 vqrshrun_n_s32(res1, FILTER_BITS));
172
173 res = aom_tbl_u16(res, idx);
174
175 return vminq_u16(res, max);
176 }
177
highbd_convolve8_horiz_4tap_sve(const uint16_t * src,ptrdiff_t src_stride,uint16_t * dst,ptrdiff_t dst_stride,const int16_t * filter_x,int width,int height,int bd)178 static inline void highbd_convolve8_horiz_4tap_sve(
179 const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst,
180 ptrdiff_t dst_stride, const int16_t *filter_x, int width, int height,
181 int bd) {
182 const int16x8_t filter = vcombine_s16(vld1_s16(filter_x + 2), vdup_n_s16(0));
183
184 if (width == 4) {
185 const uint16x4_t max = vdup_n_u16((1 << bd) - 1);
186 uint16x8x2_t permute_tbl = vld1q_u16_x2(kDotProdTbl);
187
188 const int16_t *s = (const int16_t *)src;
189 uint16_t *d = dst;
190
191 do {
192 int16x8_t s0, s1, s2, s3;
193 load_s16_8x4(s, src_stride, &s0, &s1, &s2, &s3);
194
195 uint16x4_t d0 = highbd_convolve4_4_h(s0, filter, permute_tbl, max);
196 uint16x4_t d1 = highbd_convolve4_4_h(s1, filter, permute_tbl, max);
197 uint16x4_t d2 = highbd_convolve4_4_h(s2, filter, permute_tbl, max);
198 uint16x4_t d3 = highbd_convolve4_4_h(s3, filter, permute_tbl, max);
199
200 store_u16_4x4(d, dst_stride, d0, d1, d2, d3);
201
202 s += 4 * src_stride;
203 d += 4 * dst_stride;
204 height -= 4;
205 } while (height > 0);
206 } else {
207 const uint16x8_t max = vdupq_n_u16((1 << bd) - 1);
208 uint16x8_t idx = vld1q_u16(kDeinterleaveTbl);
209
210 do {
211 const int16_t *s = (const int16_t *)src;
212 uint16_t *d = dst;
213 int w = width;
214
215 do {
216 int16x8_t s0[4], s1[4], s2[4], s3[4];
217 load_s16_8x4(s + 0 * src_stride, 1, &s0[0], &s0[1], &s0[2], &s0[3]);
218 load_s16_8x4(s + 1 * src_stride, 1, &s1[0], &s1[1], &s1[2], &s1[3]);
219 load_s16_8x4(s + 2 * src_stride, 1, &s2[0], &s2[1], &s2[2], &s2[3]);
220 load_s16_8x4(s + 3 * src_stride, 1, &s3[0], &s3[1], &s3[2], &s3[3]);
221
222 uint16x8_t d0 = highbd_convolve4_8_h(s0, filter, idx, max);
223 uint16x8_t d1 = highbd_convolve4_8_h(s1, filter, idx, max);
224 uint16x8_t d2 = highbd_convolve4_8_h(s2, filter, idx, max);
225 uint16x8_t d3 = highbd_convolve4_8_h(s3, filter, idx, max);
226
227 store_u16_8x4(d, dst_stride, d0, d1, d2, d3);
228
229 s += 8;
230 d += 8;
231 w -= 8;
232 } while (w != 0);
233 src += 4 * src_stride;
234 dst += 4 * dst_stride;
235 height -= 4;
236 } while (height > 0);
237 }
238 }
239
aom_highbd_convolve8_horiz_sve(const uint8_t * src8,ptrdiff_t src_stride,uint8_t * dst8,ptrdiff_t dst_stride,const int16_t * filter_x,int x_step_q4,const int16_t * filter_y,int y_step_q4,int width,int height,int bd)240 void aom_highbd_convolve8_horiz_sve(const uint8_t *src8, ptrdiff_t src_stride,
241 uint8_t *dst8, ptrdiff_t dst_stride,
242 const int16_t *filter_x, int x_step_q4,
243 const int16_t *filter_y, int y_step_q4,
244 int width, int height, int bd) {
245 assert(x_step_q4 == 16);
246 assert(width >= 4 && height >= 4);
247 (void)filter_y;
248 (void)x_step_q4;
249 (void)y_step_q4;
250
251 const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
252 uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
253
254 src -= SUBPEL_TAPS / 2 - 1;
255
256 const int filter_taps = get_filter_taps_convolve8(filter_x);
257
258 if (filter_taps == 2) {
259 highbd_convolve8_horiz_2tap_neon(src + 3, src_stride, dst, dst_stride,
260 filter_x, width, height, bd);
261 } else if (filter_taps == 4) {
262 highbd_convolve8_horiz_4tap_sve(src + 2, src_stride, dst, dst_stride,
263 filter_x, width, height, bd);
264 } else {
265 highbd_convolve8_horiz_8tap_sve(src, src_stride, dst, dst_stride, filter_x,
266 width, height, bd);
267 }
268 }
269
270 DECLARE_ALIGNED(16, static const uint8_t, kDotProdMergeBlockTbl[48]) = {
271 // Shift left and insert new last column in transposed 4x4 block.
272 2, 3, 4, 5, 6, 7, 16, 17, 10, 11, 12, 13, 14, 15, 24, 25,
273 // Shift left and insert two new columns in transposed 4x4 block.
274 4, 5, 6, 7, 16, 17, 18, 19, 12, 13, 14, 15, 24, 25, 26, 27,
275 // Shift left and insert three new columns in transposed 4x4 block.
276 6, 7, 16, 17, 18, 19, 20, 21, 14, 15, 24, 25, 26, 27, 28, 29
277 };
278
transpose_concat_4x4(int16x4_t s0,int16x4_t s1,int16x4_t s2,int16x4_t s3,int16x8_t res[2])279 static inline void transpose_concat_4x4(int16x4_t s0, int16x4_t s1,
280 int16x4_t s2, int16x4_t s3,
281 int16x8_t res[2]) {
282 // Transpose 16-bit elements and concatenate result rows as follows:
283 // s0: 00, 01, 02, 03
284 // s1: 10, 11, 12, 13
285 // s2: 20, 21, 22, 23
286 // s3: 30, 31, 32, 33
287 //
288 // res[0]: 00 10 20 30 01 11 21 31
289 // res[1]: 02 12 22 32 03 13 23 33
290
291 int16x8_t s0q = vcombine_s16(s0, vdup_n_s16(0));
292 int16x8_t s1q = vcombine_s16(s1, vdup_n_s16(0));
293 int16x8_t s2q = vcombine_s16(s2, vdup_n_s16(0));
294 int16x8_t s3q = vcombine_s16(s3, vdup_n_s16(0));
295
296 int32x4_t s01 = vreinterpretq_s32_s16(vzip1q_s16(s0q, s1q));
297 int32x4_t s23 = vreinterpretq_s32_s16(vzip1q_s16(s2q, s3q));
298
299 int32x4x2_t s0123 = vzipq_s32(s01, s23);
300
301 res[0] = vreinterpretq_s16_s32(s0123.val[0]);
302 res[1] = vreinterpretq_s16_s32(s0123.val[1]);
303 }
304
transpose_concat_8x4(int16x8_t s0,int16x8_t s1,int16x8_t s2,int16x8_t s3,int16x8_t res[4])305 static inline void transpose_concat_8x4(int16x8_t s0, int16x8_t s1,
306 int16x8_t s2, int16x8_t s3,
307 int16x8_t res[4]) {
308 // Transpose 16-bit elements and concatenate result rows as follows:
309 // s0: 00, 01, 02, 03, 04, 05, 06, 07
310 // s1: 10, 11, 12, 13, 14, 15, 16, 17
311 // s2: 20, 21, 22, 23, 24, 25, 26, 27
312 // s3: 30, 31, 32, 33, 34, 35, 36, 37
313 //
314 // res_lo[0]: 00 10 20 30 01 11 21 31
315 // res_lo[1]: 02 12 22 32 03 13 23 33
316 // res_hi[0]: 04 14 24 34 05 15 25 35
317 // res_hi[1]: 06 16 26 36 07 17 27 37
318
319 int16x8x2_t tr01_16 = vzipq_s16(s0, s1);
320 int16x8x2_t tr23_16 = vzipq_s16(s2, s3);
321
322 int32x4x2_t tr01_32 = vzipq_s32(vreinterpretq_s32_s16(tr01_16.val[0]),
323 vreinterpretq_s32_s16(tr23_16.val[0]));
324 int32x4x2_t tr23_32 = vzipq_s32(vreinterpretq_s32_s16(tr01_16.val[1]),
325 vreinterpretq_s32_s16(tr23_16.val[1]));
326
327 res[0] = vreinterpretq_s16_s32(tr01_32.val[0]);
328 res[1] = vreinterpretq_s16_s32(tr01_32.val[1]);
329 res[2] = vreinterpretq_s16_s32(tr23_32.val[0]);
330 res[3] = vreinterpretq_s16_s32(tr23_32.val[1]);
331 }
332
aom_tbl2x4_s16(int16x8_t t0[4],int16x8_t t1[4],uint8x16_t tbl,int16x8_t res[4])333 static inline void aom_tbl2x4_s16(int16x8_t t0[4], int16x8_t t1[4],
334 uint8x16_t tbl, int16x8_t res[4]) {
335 int8x16x2_t samples0 = { vreinterpretq_s8_s16(t0[0]),
336 vreinterpretq_s8_s16(t1[0]) };
337 int8x16x2_t samples1 = { vreinterpretq_s8_s16(t0[1]),
338 vreinterpretq_s8_s16(t1[1]) };
339 int8x16x2_t samples2 = { vreinterpretq_s8_s16(t0[2]),
340 vreinterpretq_s8_s16(t1[2]) };
341 int8x16x2_t samples3 = { vreinterpretq_s8_s16(t0[3]),
342 vreinterpretq_s8_s16(t1[3]) };
343
344 res[0] = vreinterpretq_s16_s8(vqtbl2q_s8(samples0, tbl));
345 res[1] = vreinterpretq_s16_s8(vqtbl2q_s8(samples1, tbl));
346 res[2] = vreinterpretq_s16_s8(vqtbl2q_s8(samples2, tbl));
347 res[3] = vreinterpretq_s16_s8(vqtbl2q_s8(samples3, tbl));
348 }
349
aom_tbl2x2_s16(int16x8_t t0[2],int16x8_t t1[2],uint8x16_t tbl,int16x8_t res[2])350 static inline void aom_tbl2x2_s16(int16x8_t t0[2], int16x8_t t1[2],
351 uint8x16_t tbl, int16x8_t res[2]) {
352 int8x16x2_t samples0 = { vreinterpretq_s8_s16(t0[0]),
353 vreinterpretq_s8_s16(t1[0]) };
354 int8x16x2_t samples1 = { vreinterpretq_s8_s16(t0[1]),
355 vreinterpretq_s8_s16(t1[1]) };
356
357 res[0] = vreinterpretq_s16_s8(vqtbl2q_s8(samples0, tbl));
358 res[1] = vreinterpretq_s16_s8(vqtbl2q_s8(samples1, tbl));
359 }
360
highbd_convolve8_4_v(int16x8_t samples_lo[2],int16x8_t samples_hi[2],int16x8_t filter,uint16x4_t max)361 static inline uint16x4_t highbd_convolve8_4_v(int16x8_t samples_lo[2],
362 int16x8_t samples_hi[2],
363 int16x8_t filter,
364 uint16x4_t max) {
365 int64x2_t sum[2];
366
367 sum[0] = aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[0], filter, 0);
368 sum[0] = aom_svdot_lane_s16(sum[0], samples_hi[0], filter, 1);
369
370 sum[1] = aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[1], filter, 0);
371 sum[1] = aom_svdot_lane_s16(sum[1], samples_hi[1], filter, 1);
372
373 int32x4_t res_s32 = vcombine_s32(vmovn_s64(sum[0]), vmovn_s64(sum[1]));
374
375 uint16x4_t res = vqrshrun_n_s32(res_s32, FILTER_BITS);
376
377 return vmin_u16(res, max);
378 }
379
highbd_convolve8_8_v(int16x8_t samples_lo[4],int16x8_t samples_hi[4],int16x8_t filter,uint16x8_t max)380 static inline uint16x8_t highbd_convolve8_8_v(int16x8_t samples_lo[4],
381 int16x8_t samples_hi[4],
382 int16x8_t filter,
383 uint16x8_t max) {
384 int64x2_t sum[4];
385
386 sum[0] = aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[0], filter, 0);
387 sum[0] = aom_svdot_lane_s16(sum[0], samples_hi[0], filter, 1);
388
389 sum[1] = aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[1], filter, 0);
390 sum[1] = aom_svdot_lane_s16(sum[1], samples_hi[1], filter, 1);
391
392 sum[2] = aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[2], filter, 0);
393 sum[2] = aom_svdot_lane_s16(sum[2], samples_hi[2], filter, 1);
394
395 sum[3] = aom_svdot_lane_s16(vdupq_n_s64(0), samples_lo[3], filter, 0);
396 sum[3] = aom_svdot_lane_s16(sum[3], samples_hi[3], filter, 1);
397
398 int32x4_t res0 = vcombine_s32(vmovn_s64(sum[0]), vmovn_s64(sum[1]));
399 int32x4_t res1 = vcombine_s32(vmovn_s64(sum[2]), vmovn_s64(sum[3]));
400
401 uint16x8_t res = vcombine_u16(vqrshrun_n_s32(res0, FILTER_BITS),
402 vqrshrun_n_s32(res1, FILTER_BITS));
403
404 return vminq_u16(res, max);
405 }
406
highbd_convolve8_vert_8tap_sve(const uint16_t * src,ptrdiff_t src_stride,uint16_t * dst,ptrdiff_t dst_stride,const int16_t * filter_y,int width,int height,int bd)407 static inline void highbd_convolve8_vert_8tap_sve(
408 const uint16_t *src, ptrdiff_t src_stride, uint16_t *dst,
409 ptrdiff_t dst_stride, const int16_t *filter_y, int width, int height,
410 int bd) {
411 const int16x8_t y_filter = vld1q_s16(filter_y);
412
413 uint8x16_t merge_block_tbl[3];
414 merge_block_tbl[0] = vld1q_u8(kDotProdMergeBlockTbl);
415 merge_block_tbl[1] = vld1q_u8(kDotProdMergeBlockTbl + 16);
416 merge_block_tbl[2] = vld1q_u8(kDotProdMergeBlockTbl + 32);
417
418 if (width == 4) {
419 const uint16x4_t max = vdup_n_u16((1 << bd) - 1);
420 int16_t *s = (int16_t *)src;
421
422 int16x4_t s0, s1, s2, s3, s4, s5, s6;
423 load_s16_4x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
424 s += 7 * src_stride;
425
426 // This operation combines a conventional transpose and the sample permute
427 // required before computing the dot product.
428 int16x8_t s0123[2], s1234[2], s2345[2], s3456[2];
429 transpose_concat_4x4(s0, s1, s2, s3, s0123);
430 transpose_concat_4x4(s1, s2, s3, s4, s1234);
431 transpose_concat_4x4(s2, s3, s4, s5, s2345);
432 transpose_concat_4x4(s3, s4, s5, s6, s3456);
433
434 do {
435 int16x4_t s7, s8, s9, s10;
436 load_s16_4x4(s, src_stride, &s7, &s8, &s9, &s10);
437
438 int16x8_t s4567[2], s5678[2], s6789[2], s78910[2];
439
440 // Transpose and shuffle the 4 lines that were loaded.
441 transpose_concat_4x4(s7, s8, s9, s10, s78910);
442
443 // Merge new data into block from previous iteration.
444 aom_tbl2x2_s16(s3456, s78910, merge_block_tbl[0], s4567);
445 aom_tbl2x2_s16(s3456, s78910, merge_block_tbl[1], s5678);
446 aom_tbl2x2_s16(s3456, s78910, merge_block_tbl[2], s6789);
447
448 uint16x4_t d0 = highbd_convolve8_4_v(s0123, s4567, y_filter, max);
449 uint16x4_t d1 = highbd_convolve8_4_v(s1234, s5678, y_filter, max);
450 uint16x4_t d2 = highbd_convolve8_4_v(s2345, s6789, y_filter, max);
451 uint16x4_t d3 = highbd_convolve8_4_v(s3456, s78910, y_filter, max);
452
453 store_u16_4x4(dst, dst_stride, d0, d1, d2, d3);
454
455 // Prepare block for next iteration - re-using as much as possible.
456 // Shuffle everything up four rows.
457 s0123[0] = s4567[0];
458 s0123[1] = s4567[1];
459 s1234[0] = s5678[0];
460 s1234[1] = s5678[1];
461 s2345[0] = s6789[0];
462 s2345[1] = s6789[1];
463 s3456[0] = s78910[0];
464 s3456[1] = s78910[1];
465
466 s += 4 * src_stride;
467 dst += 4 * dst_stride;
468 height -= 4;
469 } while (height != 0);
470 } else {
471 const uint16x8_t max = vdupq_n_u16((1 << bd) - 1);
472 do {
473 int h = height;
474 int16_t *s = (int16_t *)src;
475 uint16_t *d = dst;
476
477 int16x8_t s0, s1, s2, s3, s4, s5, s6;
478 load_s16_8x7(s, src_stride, &s0, &s1, &s2, &s3, &s4, &s5, &s6);
479 s += 7 * src_stride;
480
481 // This operation combines a conventional transpose and the sample permute
482 // required before computing the dot product.
483 int16x8_t s0123[4], s1234[4], s2345[4], s3456[4];
484 transpose_concat_8x4(s0, s1, s2, s3, s0123);
485 transpose_concat_8x4(s1, s2, s3, s4, s1234);
486 transpose_concat_8x4(s2, s3, s4, s5, s2345);
487 transpose_concat_8x4(s3, s4, s5, s6, s3456);
488
489 do {
490 int16x8_t s7, s8, s9, s10;
491 load_s16_8x4(s, src_stride, &s7, &s8, &s9, &s10);
492
493 int16x8_t s4567[4], s5678[4], s6789[4], s78910[4];
494
495 // Transpose and shuffle the 4 lines that were loaded.
496 transpose_concat_8x4(s7, s8, s9, s10, s78910);
497
498 // Merge new data into block from previous iteration.
499 aom_tbl2x4_s16(s3456, s78910, merge_block_tbl[0], s4567);
500 aom_tbl2x4_s16(s3456, s78910, merge_block_tbl[1], s5678);
501 aom_tbl2x4_s16(s3456, s78910, merge_block_tbl[2], s6789);
502
503 uint16x8_t d0 = highbd_convolve8_8_v(s0123, s4567, y_filter, max);
504 uint16x8_t d1 = highbd_convolve8_8_v(s1234, s5678, y_filter, max);
505 uint16x8_t d2 = highbd_convolve8_8_v(s2345, s6789, y_filter, max);
506 uint16x8_t d3 = highbd_convolve8_8_v(s3456, s78910, y_filter, max);
507
508 store_u16_8x4(d, dst_stride, d0, d1, d2, d3);
509
510 // Prepare block for next iteration - re-using as much as possible.
511 // Shuffle everything up four rows.
512 s0123[0] = s4567[0];
513 s0123[1] = s4567[1];
514 s0123[2] = s4567[2];
515 s0123[3] = s4567[3];
516
517 s1234[0] = s5678[0];
518 s1234[1] = s5678[1];
519 s1234[2] = s5678[2];
520 s1234[3] = s5678[3];
521
522 s2345[0] = s6789[0];
523 s2345[1] = s6789[1];
524 s2345[2] = s6789[2];
525 s2345[3] = s6789[3];
526
527 s3456[0] = s78910[0];
528 s3456[1] = s78910[1];
529 s3456[2] = s78910[2];
530 s3456[3] = s78910[3];
531
532 s += 4 * src_stride;
533 d += 4 * dst_stride;
534 h -= 4;
535 } while (h != 0);
536 src += 8;
537 dst += 8;
538 width -= 8;
539 } while (width != 0);
540 }
541 }
542
aom_highbd_convolve8_vert_sve(const uint8_t * src8,ptrdiff_t src_stride,uint8_t * dst8,ptrdiff_t dst_stride,const int16_t * filter_x,int x_step_q4,const int16_t * filter_y,int y_step_q4,int width,int height,int bd)543 void aom_highbd_convolve8_vert_sve(const uint8_t *src8, ptrdiff_t src_stride,
544 uint8_t *dst8, ptrdiff_t dst_stride,
545 const int16_t *filter_x, int x_step_q4,
546 const int16_t *filter_y, int y_step_q4,
547 int width, int height, int bd) {
548 assert(y_step_q4 == 16);
549 assert(width >= 4 && height >= 4);
550 (void)filter_x;
551 (void)y_step_q4;
552 (void)x_step_q4;
553
554 const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
555 uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
556
557 src -= (SUBPEL_TAPS / 2 - 1) * src_stride;
558
559 const int filter_taps = get_filter_taps_convolve8(filter_y);
560
561 if (filter_taps == 2) {
562 highbd_convolve8_vert_2tap_neon(src + 3 * src_stride, src_stride, dst,
563 dst_stride, filter_y, width, height, bd);
564 } else if (filter_taps == 4) {
565 highbd_convolve8_vert_4tap_neon(src + 2 * src_stride, src_stride, dst,
566 dst_stride, filter_y, width, height, bd);
567 } else {
568 highbd_convolve8_vert_8tap_sve(src, src_stride, dst, dst_stride, filter_y,
569 width, height, bd);
570 }
571 }
572