1 /*
2 * Copyright (c) 2014 The WebM project authors. All Rights Reserved.
3 *
4 * Use of this source code is governed by a BSD-style license
5 * that can be found in the LICENSE file in the root of the source
6 * tree. An additional intellectual property rights grant can be found
7 * in the file PATENTS. All contributing project authors may
8 * be found in the AUTHORS file in the root of the source tree.
9 */
10
11 #include <arm_neon.h>
12
13 #include "./vpx_config.h"
14 #include "./vpx_dsp_rtcd.h"
15 #include "vpx/vpx_integer.h"
16
17 //------------------------------------------------------------------------------
18 // DC 4x4
19
dc_sum_4(const uint8_t * ref)20 static INLINE uint16x4_t dc_sum_4(const uint8_t *ref) {
21 const uint8x8_t ref_u8 = vld1_u8(ref);
22 const uint16x4_t p0 = vpaddl_u8(ref_u8);
23 return vpadd_u16(p0, p0);
24 }
25
dc_store_4x4(uint8_t * dst,ptrdiff_t stride,const uint8x8_t dc)26 static INLINE void dc_store_4x4(uint8_t *dst, ptrdiff_t stride,
27 const uint8x8_t dc) {
28 const uint8x8_t dc_dup = vdup_lane_u8(dc, 0);
29 int i;
30 for (i = 0; i < 4; ++i, dst += stride) {
31 vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(dc_dup), 0);
32 }
33 }
34
vpx_dc_predictor_4x4_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)35 void vpx_dc_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
36 const uint8_t *above, const uint8_t *left) {
37 const uint8x8_t a = vld1_u8(above);
38 const uint8x8_t l = vld1_u8(left);
39 const uint16x8_t al = vaddl_u8(a, l);
40 uint16x4_t sum;
41 uint8x8_t dc;
42 sum = vpadd_u16(vget_low_u16(al), vget_low_u16(al));
43 sum = vpadd_u16(sum, sum);
44 dc = vreinterpret_u8_u16(vrshr_n_u16(sum, 3));
45 dc_store_4x4(dst, stride, dc);
46 }
47
vpx_dc_left_predictor_4x4_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)48 void vpx_dc_left_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
49 const uint8_t *above, const uint8_t *left) {
50 const uint16x4_t sum = dc_sum_4(left);
51 const uint8x8_t dc = vreinterpret_u8_u16(vrshr_n_u16(sum, 2));
52 (void)above;
53 dc_store_4x4(dst, stride, dc);
54 }
55
vpx_dc_top_predictor_4x4_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)56 void vpx_dc_top_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
57 const uint8_t *above, const uint8_t *left) {
58 const uint16x4_t sum = dc_sum_4(above);
59 const uint8x8_t dc = vreinterpret_u8_u16(vrshr_n_u16(sum, 2));
60 (void)left;
61 dc_store_4x4(dst, stride, dc);
62 }
63
vpx_dc_128_predictor_4x4_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)64 void vpx_dc_128_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
65 const uint8_t *above, const uint8_t *left) {
66 const uint8x8_t dc = vdup_n_u8(0x80);
67 (void)above;
68 (void)left;
69 dc_store_4x4(dst, stride, dc);
70 }
71
72 //------------------------------------------------------------------------------
73 // DC 8x8
74
dc_sum_8(const uint8_t * ref)75 static INLINE uint16x4_t dc_sum_8(const uint8_t *ref) {
76 const uint8x8_t ref_u8 = vld1_u8(ref);
77 uint16x4_t sum = vpaddl_u8(ref_u8);
78 sum = vpadd_u16(sum, sum);
79 return vpadd_u16(sum, sum);
80 }
81
dc_store_8x8(uint8_t * dst,ptrdiff_t stride,const uint8x8_t dc)82 static INLINE void dc_store_8x8(uint8_t *dst, ptrdiff_t stride,
83 const uint8x8_t dc) {
84 const uint8x8_t dc_dup = vdup_lane_u8(dc, 0);
85 int i;
86 for (i = 0; i < 8; ++i, dst += stride) {
87 vst1_u8(dst, dc_dup);
88 }
89 }
90
vpx_dc_predictor_8x8_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)91 void vpx_dc_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
92 const uint8_t *above, const uint8_t *left) {
93 const uint8x8_t above_u8 = vld1_u8(above);
94 const uint8x8_t left_u8 = vld1_u8(left);
95 const uint8x16_t above_and_left = vcombine_u8(above_u8, left_u8);
96 const uint16x8_t p0 = vpaddlq_u8(above_and_left);
97 uint16x4_t sum = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
98 uint8x8_t dc;
99 sum = vpadd_u16(sum, sum);
100 sum = vpadd_u16(sum, sum);
101 dc = vreinterpret_u8_u16(vrshr_n_u16(sum, 4));
102 dc_store_8x8(dst, stride, dc);
103 }
104
vpx_dc_left_predictor_8x8_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)105 void vpx_dc_left_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
106 const uint8_t *above, const uint8_t *left) {
107 const uint16x4_t sum = dc_sum_8(left);
108 const uint8x8_t dc = vreinterpret_u8_u16(vrshr_n_u16(sum, 3));
109 (void)above;
110 dc_store_8x8(dst, stride, dc);
111 }
112
vpx_dc_top_predictor_8x8_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)113 void vpx_dc_top_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
114 const uint8_t *above, const uint8_t *left) {
115 const uint16x4_t sum = dc_sum_8(above);
116 const uint8x8_t dc = vreinterpret_u8_u16(vrshr_n_u16(sum, 3));
117 (void)left;
118 dc_store_8x8(dst, stride, dc);
119 }
120
vpx_dc_128_predictor_8x8_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)121 void vpx_dc_128_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
122 const uint8_t *above, const uint8_t *left) {
123 const uint8x8_t dc = vdup_n_u8(0x80);
124 (void)above;
125 (void)left;
126 dc_store_8x8(dst, stride, dc);
127 }
128
129 //------------------------------------------------------------------------------
130 // DC 16x16
131
dc_sum_16(const uint8_t * ref)132 static INLINE uint16x4_t dc_sum_16(const uint8_t *ref) {
133 const uint8x16_t ref_u8 = vld1q_u8(ref);
134 const uint16x8_t p0 = vpaddlq_u8(ref_u8);
135 uint16x4_t sum = vadd_u16(vget_low_u16(p0), vget_high_u16(p0));
136 sum = vpadd_u16(sum, sum);
137 return vpadd_u16(sum, sum);
138 }
139
dc_store_16x16(uint8_t * dst,ptrdiff_t stride,const uint8x8_t dc)140 static INLINE void dc_store_16x16(uint8_t *dst, ptrdiff_t stride,
141 const uint8x8_t dc) {
142 const uint8x16_t dc_dup = vdupq_lane_u8(dc, 0);
143 int i;
144 for (i = 0; i < 16; ++i, dst += stride) {
145 vst1q_u8(dst, dc_dup);
146 }
147 }
148
vpx_dc_predictor_16x16_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)149 void vpx_dc_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
150 const uint8_t *above, const uint8_t *left) {
151 const uint8x16_t ref0 = vld1q_u8(above);
152 const uint8x16_t ref1 = vld1q_u8(left);
153 const uint16x8_t p0 = vpaddlq_u8(ref0);
154 const uint16x8_t p1 = vpaddlq_u8(ref1);
155 const uint16x8_t p2 = vaddq_u16(p0, p1);
156 uint16x4_t sum = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
157 uint8x8_t dc;
158 sum = vpadd_u16(sum, sum);
159 sum = vpadd_u16(sum, sum);
160 dc = vreinterpret_u8_u16(vrshr_n_u16(sum, 5));
161 dc_store_16x16(dst, stride, dc);
162 }
163
vpx_dc_left_predictor_16x16_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)164 void vpx_dc_left_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
165 const uint8_t *above,
166 const uint8_t *left) {
167 const uint16x4_t sum = dc_sum_16(left);
168 const uint8x8_t dc = vreinterpret_u8_u16(vrshr_n_u16(sum, 4));
169 (void)above;
170 dc_store_16x16(dst, stride, dc);
171 }
172
vpx_dc_top_predictor_16x16_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)173 void vpx_dc_top_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
174 const uint8_t *above,
175 const uint8_t *left) {
176 const uint16x4_t sum = dc_sum_16(above);
177 const uint8x8_t dc = vreinterpret_u8_u16(vrshr_n_u16(sum, 4));
178 (void)left;
179 dc_store_16x16(dst, stride, dc);
180 }
181
vpx_dc_128_predictor_16x16_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)182 void vpx_dc_128_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
183 const uint8_t *above,
184 const uint8_t *left) {
185 const uint8x8_t dc = vdup_n_u8(0x80);
186 (void)above;
187 (void)left;
188 dc_store_16x16(dst, stride, dc);
189 }
190
191 //------------------------------------------------------------------------------
192 // DC 32x32
193
dc_sum_32(const uint8_t * ref)194 static INLINE uint16x4_t dc_sum_32(const uint8_t *ref) {
195 const uint8x16x2_t r = vld2q_u8(ref);
196 const uint16x8_t p0 = vpaddlq_u8(r.val[0]);
197 const uint16x8_t p1 = vpaddlq_u8(r.val[1]);
198 const uint16x8_t p2 = vaddq_u16(p0, p1);
199 uint16x4_t sum = vadd_u16(vget_low_u16(p2), vget_high_u16(p2));
200 sum = vpadd_u16(sum, sum);
201 return vpadd_u16(sum, sum);
202 }
203
dc_store_32x32(uint8_t * dst,ptrdiff_t stride,const uint8x8_t dc)204 static INLINE void dc_store_32x32(uint8_t *dst, ptrdiff_t stride,
205 const uint8x8_t dc) {
206 uint8x16x2_t dc_dup;
207 int i;
208 dc_dup.val[0] = dc_dup.val[1] = vdupq_lane_u8(dc, 0);
209
210 for (i = 0; i < 32; ++i, dst += stride) {
211 vst2q_u8(dst, dc_dup);
212 }
213 }
214
vpx_dc_predictor_32x32_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)215 void vpx_dc_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
216 const uint8_t *above, const uint8_t *left) {
217 const uint8x16x2_t a = vld2q_u8(above);
218 const uint8x16x2_t l = vld2q_u8(left);
219 const uint16x8_t pa0 = vpaddlq_u8(a.val[0]);
220 const uint16x8_t pl0 = vpaddlq_u8(l.val[0]);
221 const uint16x8_t pa1 = vpaddlq_u8(a.val[1]);
222 const uint16x8_t pl1 = vpaddlq_u8(l.val[1]);
223 const uint16x8_t pa = vaddq_u16(pa0, pa1);
224 const uint16x8_t pl = vaddq_u16(pl0, pl1);
225 const uint16x8_t pal = vaddq_u16(pa, pl);
226 uint16x4_t sum = vadd_u16(vget_low_u16(pal), vget_high_u16(pal));
227 uint8x8_t dc;
228 sum = vpadd_u16(sum, sum);
229 sum = vpadd_u16(sum, sum);
230 dc = vreinterpret_u8_u16(vrshr_n_u16(sum, 6));
231 dc_store_32x32(dst, stride, dc);
232 }
233
vpx_dc_left_predictor_32x32_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)234 void vpx_dc_left_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
235 const uint8_t *above,
236 const uint8_t *left) {
237 const uint16x4_t sum = dc_sum_32(left);
238 const uint8x8_t dc = vreinterpret_u8_u16(vrshr_n_u16(sum, 5));
239 (void)above;
240 dc_store_32x32(dst, stride, dc);
241 }
242
vpx_dc_top_predictor_32x32_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)243 void vpx_dc_top_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
244 const uint8_t *above,
245 const uint8_t *left) {
246 const uint16x4_t sum = dc_sum_32(above);
247 const uint8x8_t dc = vreinterpret_u8_u16(vrshr_n_u16(sum, 5));
248 (void)left;
249 dc_store_32x32(dst, stride, dc);
250 }
251
vpx_dc_128_predictor_32x32_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)252 void vpx_dc_128_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
253 const uint8_t *above,
254 const uint8_t *left) {
255 const uint8x8_t dc = vdup_n_u8(0x80);
256 (void)above;
257 (void)left;
258 dc_store_32x32(dst, stride, dc);
259 }
260
261 // -----------------------------------------------------------------------------
262
vpx_d45_predictor_4x4_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)263 void vpx_d45_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
264 const uint8_t *above, const uint8_t *left) {
265 const uint8x8_t ABCDEFGH = vld1_u8(above);
266 const uint64x1_t A1 = vshr_n_u64(vreinterpret_u64_u8(ABCDEFGH), 8);
267 const uint64x1_t A2 = vshr_n_u64(vreinterpret_u64_u8(ABCDEFGH), 16);
268 const uint8x8_t BCDEFGH0 = vreinterpret_u8_u64(A1);
269 const uint8x8_t CDEFGH00 = vreinterpret_u8_u64(A2);
270 const uint8x8_t avg1 = vhadd_u8(ABCDEFGH, CDEFGH00);
271 const uint8x8_t avg2 = vrhadd_u8(avg1, BCDEFGH0);
272 const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2);
273 const uint32x2_t r0 = vreinterpret_u32_u8(avg2);
274 const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8));
275 const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16));
276 const uint32x2_t r3 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24));
277 (void)left;
278 vst1_lane_u32((uint32_t *)(dst + 0 * stride), r0, 0);
279 vst1_lane_u32((uint32_t *)(dst + 1 * stride), r1, 0);
280 vst1_lane_u32((uint32_t *)(dst + 2 * stride), r2, 0);
281 vst1_lane_u32((uint32_t *)(dst + 3 * stride), r3, 0);
282 vst1_lane_u8(dst + 3 * stride + 3, ABCDEFGH, 7);
283 }
284
d45_store_8(uint8_t ** dst,const ptrdiff_t stride,const uint8x8_t above_right,uint8x8_t * row)285 static INLINE void d45_store_8(uint8_t **dst, const ptrdiff_t stride,
286 const uint8x8_t above_right, uint8x8_t *row) {
287 *row = vext_u8(*row, above_right, 1);
288 vst1_u8(*dst, *row);
289 *dst += stride;
290 }
291
vpx_d45_predictor_8x8_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)292 void vpx_d45_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
293 const uint8_t *above, const uint8_t *left) {
294 const uint8x8_t A0 = vld1_u8(above);
295 const uint8x8_t above_right = vdup_lane_u8(A0, 7);
296 const uint8x8_t A1 = vext_u8(A0, above_right, 1);
297 const uint8x8_t A2 = vext_u8(A0, above_right, 2);
298 const uint8x8_t avg1 = vhadd_u8(A0, A2);
299 uint8x8_t row = vrhadd_u8(avg1, A1);
300 (void)left;
301
302 vst1_u8(dst, row);
303 dst += stride;
304 d45_store_8(&dst, stride, above_right, &row);
305 d45_store_8(&dst, stride, above_right, &row);
306 d45_store_8(&dst, stride, above_right, &row);
307 d45_store_8(&dst, stride, above_right, &row);
308 d45_store_8(&dst, stride, above_right, &row);
309 d45_store_8(&dst, stride, above_right, &row);
310 vst1_u8(dst, above_right);
311 }
312
d45_store_16(uint8_t ** dst,const ptrdiff_t stride,const uint8x16_t above_right,uint8x16_t * row)313 static INLINE void d45_store_16(uint8_t **dst, const ptrdiff_t stride,
314 const uint8x16_t above_right, uint8x16_t *row) {
315 *row = vextq_u8(*row, above_right, 1);
316 vst1q_u8(*dst, *row);
317 *dst += stride;
318 }
319
vpx_d45_predictor_16x16_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)320 void vpx_d45_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
321 const uint8_t *above, const uint8_t *left) {
322 const uint8x16_t A0 = vld1q_u8(above);
323 const uint8x16_t above_right = vdupq_lane_u8(vget_high_u8(A0), 7);
324 const uint8x16_t A1 = vextq_u8(A0, above_right, 1);
325 const uint8x16_t A2 = vextq_u8(A0, above_right, 2);
326 const uint8x16_t avg1 = vhaddq_u8(A0, A2);
327 uint8x16_t row = vrhaddq_u8(avg1, A1);
328 (void)left;
329
330 vst1q_u8(dst, row);
331 dst += stride;
332 d45_store_16(&dst, stride, above_right, &row);
333 d45_store_16(&dst, stride, above_right, &row);
334 d45_store_16(&dst, stride, above_right, &row);
335 d45_store_16(&dst, stride, above_right, &row);
336 d45_store_16(&dst, stride, above_right, &row);
337 d45_store_16(&dst, stride, above_right, &row);
338 d45_store_16(&dst, stride, above_right, &row);
339 d45_store_16(&dst, stride, above_right, &row);
340 d45_store_16(&dst, stride, above_right, &row);
341 d45_store_16(&dst, stride, above_right, &row);
342 d45_store_16(&dst, stride, above_right, &row);
343 d45_store_16(&dst, stride, above_right, &row);
344 d45_store_16(&dst, stride, above_right, &row);
345 d45_store_16(&dst, stride, above_right, &row);
346 vst1q_u8(dst, above_right);
347 }
348
vpx_d45_predictor_32x32_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)349 void vpx_d45_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
350 const uint8_t *above, const uint8_t *left) {
351 const uint8x16_t A0_0 = vld1q_u8(above);
352 const uint8x16_t A0_1 = vld1q_u8(above + 16);
353 const uint8x16_t above_right = vdupq_lane_u8(vget_high_u8(A0_1), 7);
354 const uint8x16_t A1_0 = vld1q_u8(above + 1);
355 const uint8x16_t A1_1 = vld1q_u8(above + 17);
356 const uint8x16_t A2_0 = vld1q_u8(above + 2);
357 const uint8x16_t A2_1 = vld1q_u8(above + 18);
358 const uint8x16_t avg_0 = vhaddq_u8(A0_0, A2_0);
359 const uint8x16_t avg_1 = vhaddq_u8(A0_1, A2_1);
360 uint8x16_t row_0 = vrhaddq_u8(avg_0, A1_0);
361 uint8x16_t row_1 = vrhaddq_u8(avg_1, A1_1);
362 int i;
363 (void)left;
364
365 vst1q_u8(dst, row_0);
366 dst += 16;
367 vst1q_u8(dst, row_1);
368 dst += stride - 16;
369
370 for (i = 0; i < 30; ++i) {
371 row_0 = vextq_u8(row_0, row_1, 1);
372 row_1 = vextq_u8(row_1, above_right, 1);
373 vst1q_u8(dst, row_0);
374 dst += 16;
375 vst1q_u8(dst, row_1);
376 dst += stride - 16;
377 }
378
379 vst1q_u8(dst, above_right);
380 dst += 16;
381 vst1q_u8(dst, row_1);
382 }
383
384 // -----------------------------------------------------------------------------
385
vpx_d135_predictor_4x4_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)386 void vpx_d135_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
387 const uint8_t *above, const uint8_t *left) {
388 const uint8x8_t XA0123 = vld1_u8(above - 1);
389 const uint8x8_t L0123 = vld1_u8(left);
390 const uint8x8_t L3210 = vrev64_u8(L0123);
391 const uint8x8_t L3210XA012 = vext_u8(L3210, XA0123, 4);
392 const uint8x8_t L210XA0123 = vext_u8(L3210, XA0123, 5);
393 const uint8x8_t L10XA0123_ =
394 vreinterpret_u8_u64(vshr_n_u64(vreinterpret_u64_u8(L210XA0123), 8));
395 const uint8x8_t avg1 = vhadd_u8(L10XA0123_, L3210XA012);
396 const uint8x8_t avg2 = vrhadd_u8(avg1, L210XA0123);
397 const uint64x1_t avg2_u64 = vreinterpret_u64_u8(avg2);
398 const uint32x2_t r3 = vreinterpret_u32_u8(avg2);
399 const uint32x2_t r2 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 8));
400 const uint32x2_t r1 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 16));
401 const uint32x2_t r0 = vreinterpret_u32_u64(vshr_n_u64(avg2_u64, 24));
402 vst1_lane_u32((uint32_t *)dst, r0, 0);
403 dst += stride;
404 vst1_lane_u32((uint32_t *)dst, r1, 0);
405 dst += stride;
406 vst1_lane_u32((uint32_t *)dst, r2, 0);
407 dst += stride;
408 vst1_lane_u32((uint32_t *)dst, r3, 0);
409 }
410
vpx_d135_predictor_8x8_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)411 void vpx_d135_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
412 const uint8_t *above, const uint8_t *left) {
413 const uint8x8_t XA0123456 = vld1_u8(above - 1);
414 const uint8x8_t A01234567 = vld1_u8(above);
415 const uint8x8_t A1234567_ = vld1_u8(above + 1);
416 const uint8x8_t L01234567 = vld1_u8(left);
417 const uint8x8_t L76543210 = vrev64_u8(L01234567);
418 const uint8x8_t L6543210X = vext_u8(L76543210, XA0123456, 1);
419 const uint8x8_t L543210XA0 = vext_u8(L76543210, XA0123456, 2);
420 const uint8x16_t L76543210XA0123456 = vcombine_u8(L76543210, XA0123456);
421 const uint8x16_t L6543210XA01234567 = vcombine_u8(L6543210X, A01234567);
422 const uint8x16_t L543210XA01234567_ = vcombine_u8(L543210XA0, A1234567_);
423 const uint8x16_t avg = vhaddq_u8(L76543210XA0123456, L543210XA01234567_);
424 const uint8x16_t row = vrhaddq_u8(avg, L6543210XA01234567);
425 const uint8x8_t row_0 = vget_low_u8(row);
426 const uint8x8_t row_1 = vget_high_u8(row);
427 const uint8x8_t r0 = vext_u8(row_0, row_1, 7);
428 const uint8x8_t r1 = vext_u8(row_0, row_1, 6);
429 const uint8x8_t r2 = vext_u8(row_0, row_1, 5);
430 const uint8x8_t r3 = vext_u8(row_0, row_1, 4);
431 const uint8x8_t r4 = vext_u8(row_0, row_1, 3);
432 const uint8x8_t r5 = vext_u8(row_0, row_1, 2);
433 const uint8x8_t r6 = vext_u8(row_0, row_1, 1);
434
435 vst1_u8(dst, r0);
436 dst += stride;
437 vst1_u8(dst, r1);
438 dst += stride;
439 vst1_u8(dst, r2);
440 dst += stride;
441 vst1_u8(dst, r3);
442 dst += stride;
443 vst1_u8(dst, r4);
444 dst += stride;
445 vst1_u8(dst, r5);
446 dst += stride;
447 vst1_u8(dst, r6);
448 dst += stride;
449 vst1_u8(dst, row_0);
450 }
451
d135_store_16x8(uint8_t ** dst,const ptrdiff_t stride,const uint8x16_t row_0,const uint8x16_t row_1,const uint8x16_t row_2,const uint8x16_t row_3,const uint8x16_t row_4,const uint8x16_t row_5,const uint8x16_t row_6,const uint8x16_t row_7)452 static INLINE void d135_store_16x8(
453 uint8_t **dst, const ptrdiff_t stride, const uint8x16_t row_0,
454 const uint8x16_t row_1, const uint8x16_t row_2, const uint8x16_t row_3,
455 const uint8x16_t row_4, const uint8x16_t row_5, const uint8x16_t row_6,
456 const uint8x16_t row_7) {
457 vst1q_u8(*dst, row_0);
458 *dst += stride;
459 vst1q_u8(*dst, row_1);
460 *dst += stride;
461 vst1q_u8(*dst, row_2);
462 *dst += stride;
463 vst1q_u8(*dst, row_3);
464 *dst += stride;
465 vst1q_u8(*dst, row_4);
466 *dst += stride;
467 vst1q_u8(*dst, row_5);
468 *dst += stride;
469 vst1q_u8(*dst, row_6);
470 *dst += stride;
471 vst1q_u8(*dst, row_7);
472 *dst += stride;
473 }
474
vpx_d135_predictor_16x16_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)475 void vpx_d135_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
476 const uint8_t *above, const uint8_t *left) {
477 const uint8x16_t XA0123456789abcde = vld1q_u8(above - 1);
478 const uint8x16_t A0123456789abcdef = vld1q_u8(above);
479 const uint8x16_t A123456789abcdef_ = vld1q_u8(above + 1);
480 const uint8x16_t L0123456789abcdef = vld1q_u8(left);
481 const uint8x8_t L76543210 = vrev64_u8(vget_low_u8(L0123456789abcdef));
482 const uint8x8_t Lfedcba98 = vrev64_u8(vget_high_u8(L0123456789abcdef));
483 const uint8x16_t Lfedcba9876543210 = vcombine_u8(Lfedcba98, L76543210);
484 const uint8x16_t Ledcba9876543210X =
485 vextq_u8(Lfedcba9876543210, XA0123456789abcde, 1);
486 const uint8x16_t Ldcba9876543210XA0 =
487 vextq_u8(Lfedcba9876543210, XA0123456789abcde, 2);
488 const uint8x16_t avg_0 = vhaddq_u8(Lfedcba9876543210, Ldcba9876543210XA0);
489 const uint8x16_t avg_1 = vhaddq_u8(XA0123456789abcde, A123456789abcdef_);
490 const uint8x16_t row_0 = vrhaddq_u8(avg_0, Ledcba9876543210X);
491 const uint8x16_t row_1 = vrhaddq_u8(avg_1, A0123456789abcdef);
492 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 15);
493 const uint8x16_t r_1 = vextq_u8(row_0, row_1, 14);
494 const uint8x16_t r_2 = vextq_u8(row_0, row_1, 13);
495 const uint8x16_t r_3 = vextq_u8(row_0, row_1, 12);
496 const uint8x16_t r_4 = vextq_u8(row_0, row_1, 11);
497 const uint8x16_t r_5 = vextq_u8(row_0, row_1, 10);
498 const uint8x16_t r_6 = vextq_u8(row_0, row_1, 9);
499 const uint8x16_t r_7 = vcombine_u8(vget_high_u8(row_0), vget_low_u8(row_1));
500 const uint8x16_t r_8 = vextq_u8(row_0, row_1, 7);
501 const uint8x16_t r_9 = vextq_u8(row_0, row_1, 6);
502 const uint8x16_t r_a = vextq_u8(row_0, row_1, 5);
503 const uint8x16_t r_b = vextq_u8(row_0, row_1, 4);
504 const uint8x16_t r_c = vextq_u8(row_0, row_1, 3);
505 const uint8x16_t r_d = vextq_u8(row_0, row_1, 2);
506 const uint8x16_t r_e = vextq_u8(row_0, row_1, 1);
507
508 d135_store_16x8(&dst, stride, r_0, r_1, r_2, r_3, r_4, r_5, r_6, r_7);
509 d135_store_16x8(&dst, stride, r_8, r_9, r_a, r_b, r_c, r_d, r_e, row_0);
510 }
511
d135_store_32x2(uint8_t ** dst,const ptrdiff_t stride,const uint8x16_t row_0,const uint8x16_t row_1,const uint8x16_t row_2)512 static INLINE void d135_store_32x2(uint8_t **dst, const ptrdiff_t stride,
513 const uint8x16_t row_0,
514 const uint8x16_t row_1,
515 const uint8x16_t row_2) {
516 uint8_t *dst2 = *dst;
517 vst1q_u8(dst2, row_1);
518 dst2 += 16;
519 vst1q_u8(dst2, row_2);
520 dst2 += 16 * stride - 16;
521 vst1q_u8(dst2, row_0);
522 dst2 += 16;
523 vst1q_u8(dst2, row_1);
524 *dst += stride;
525 }
526
vpx_d135_predictor_32x32_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)527 void vpx_d135_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
528 const uint8_t *above, const uint8_t *left) {
529 const uint8x16_t LL0123456789abcdef = vld1q_u8(left + 16);
530 const uint8x16_t LU0123456789abcdef = vld1q_u8(left);
531 const uint8x8_t LL76543210 = vrev64_u8(vget_low_u8(LL0123456789abcdef));
532 const uint8x8_t LU76543210 = vrev64_u8(vget_low_u8(LU0123456789abcdef));
533 const uint8x8_t LLfedcba98 = vrev64_u8(vget_high_u8(LL0123456789abcdef));
534 const uint8x8_t LUfedcba98 = vrev64_u8(vget_high_u8(LU0123456789abcdef));
535 const uint8x16_t LLfedcba9876543210 = vcombine_u8(LLfedcba98, LL76543210);
536 const uint8x16_t LUfedcba9876543210 = vcombine_u8(LUfedcba98, LU76543210);
537 const uint8x16_t LLedcba9876543210Uf =
538 vextq_u8(LLfedcba9876543210, LUfedcba9876543210, 1);
539 const uint8x16_t LLdcba9876543210Ufe =
540 vextq_u8(LLfedcba9876543210, LUfedcba9876543210, 2);
541 const uint8x16_t avg_0 = vhaddq_u8(LLfedcba9876543210, LLdcba9876543210Ufe);
542 const uint8x16_t row_0 = vrhaddq_u8(avg_0, LLedcba9876543210Uf);
543
544 const uint8x16_t XAL0123456789abcde = vld1q_u8(above - 1);
545 const uint8x16_t LUedcba9876543210X =
546 vextq_u8(LUfedcba9876543210, XAL0123456789abcde, 1);
547 const uint8x16_t LUdcba9876543210XA0 =
548 vextq_u8(LUfedcba9876543210, XAL0123456789abcde, 2);
549 const uint8x16_t avg_1 = vhaddq_u8(LUfedcba9876543210, LUdcba9876543210XA0);
550 const uint8x16_t row_1 = vrhaddq_u8(avg_1, LUedcba9876543210X);
551
552 const uint8x16_t AL0123456789abcdef = vld1q_u8(above);
553 const uint8x16_t AL123456789abcdefg = vld1q_u8(above + 1);
554 const uint8x16_t ALfR0123456789abcde = vld1q_u8(above + 15);
555 const uint8x16_t AR0123456789abcdef = vld1q_u8(above + 16);
556 const uint8x16_t AR123456789abcdef_ = vld1q_u8(above + 17);
557 const uint8x16_t avg_2 = vhaddq_u8(XAL0123456789abcde, AL123456789abcdefg);
558 const uint8x16_t row_2 = vrhaddq_u8(avg_2, AL0123456789abcdef);
559 const uint8x16_t avg_3 = vhaddq_u8(ALfR0123456789abcde, AR123456789abcdef_);
560 const uint8x16_t row_3 = vrhaddq_u8(avg_3, AR0123456789abcdef);
561
562 {
563 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 15);
564 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 15);
565 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 15);
566 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
567 }
568
569 {
570 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 14);
571 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 14);
572 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 14);
573 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
574 }
575
576 {
577 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 13);
578 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 13);
579 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 13);
580 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
581 }
582
583 {
584 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 12);
585 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 12);
586 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 12);
587 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
588 }
589
590 {
591 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 11);
592 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 11);
593 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 11);
594 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
595 }
596
597 {
598 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 10);
599 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 10);
600 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 10);
601 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
602 }
603
604 {
605 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 9);
606 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 9);
607 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 9);
608 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
609 }
610
611 {
612 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 8);
613 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 8);
614 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 8);
615 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
616 }
617
618 {
619 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 7);
620 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 7);
621 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 7);
622 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
623 }
624
625 {
626 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 6);
627 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 6);
628 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 6);
629 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
630 }
631
632 {
633 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 5);
634 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 5);
635 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 5);
636 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
637 }
638
639 {
640 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 4);
641 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 4);
642 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 4);
643 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
644 }
645
646 {
647 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 3);
648 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 3);
649 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 3);
650 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
651 }
652
653 {
654 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 2);
655 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 2);
656 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 2);
657 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
658 }
659
660 {
661 const uint8x16_t r_0 = vextq_u8(row_0, row_1, 1);
662 const uint8x16_t r_1 = vextq_u8(row_1, row_2, 1);
663 const uint8x16_t r_2 = vextq_u8(row_2, row_3, 1);
664 d135_store_32x2(&dst, stride, r_0, r_1, r_2);
665 }
666
667 d135_store_32x2(&dst, stride, row_0, row_1, row_2);
668 }
669
670 // -----------------------------------------------------------------------------
671
672 #if !HAVE_NEON_ASM
673
vpx_v_predictor_4x4_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)674 void vpx_v_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
675 const uint8_t *above, const uint8_t *left) {
676 const uint32_t d = *(const uint32_t *)above;
677 int i;
678 (void)left;
679
680 for (i = 0; i < 4; i++, dst += stride) {
681 *(uint32_t *)dst = d;
682 }
683 }
684
vpx_v_predictor_8x8_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)685 void vpx_v_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
686 const uint8_t *above, const uint8_t *left) {
687 const uint8x8_t d = vld1_u8(above);
688 int i;
689 (void)left;
690
691 for (i = 0; i < 8; i++, dst += stride) {
692 vst1_u8(dst, d);
693 }
694 }
695
vpx_v_predictor_16x16_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)696 void vpx_v_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
697 const uint8_t *above, const uint8_t *left) {
698 const uint8x16_t d = vld1q_u8(above);
699 int i;
700 (void)left;
701
702 for (i = 0; i < 16; i++, dst += stride) {
703 vst1q_u8(dst, d);
704 }
705 }
706
vpx_v_predictor_32x32_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)707 void vpx_v_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
708 const uint8_t *above, const uint8_t *left) {
709 const uint8x16_t d0 = vld1q_u8(above);
710 const uint8x16_t d1 = vld1q_u8(above + 16);
711 int i;
712 (void)left;
713
714 for (i = 0; i < 32; i++) {
715 // Note: performance was worse using vst2q_u8 under gcc-4.9 & clang-3.8.
716 // clang-3.8 unrolled the loop fully with no filler so the cause is likely
717 // the latency of the instruction.
718 vst1q_u8(dst, d0);
719 dst += 16;
720 vst1q_u8(dst, d1);
721 dst += stride - 16;
722 }
723 }
724
725 // -----------------------------------------------------------------------------
726
vpx_h_predictor_4x4_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)727 void vpx_h_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
728 const uint8_t *above, const uint8_t *left) {
729 const uint32x2_t zero = vdup_n_u32(0);
730 const uint8x8_t left_u8 =
731 vreinterpret_u8_u32(vld1_lane_u32((const uint32_t *)left, zero, 0));
732 uint8x8_t d;
733 (void)above;
734
735 d = vdup_lane_u8(left_u8, 0);
736 vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d), 0);
737 dst += stride;
738 d = vdup_lane_u8(left_u8, 1);
739 vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d), 0);
740 dst += stride;
741 d = vdup_lane_u8(left_u8, 2);
742 vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d), 0);
743 dst += stride;
744 d = vdup_lane_u8(left_u8, 3);
745 vst1_lane_u32((uint32_t *)dst, vreinterpret_u32_u8(d), 0);
746 }
747
vpx_h_predictor_8x8_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)748 void vpx_h_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
749 const uint8_t *above, const uint8_t *left) {
750 const uint8x8_t left_u8 = vld1_u8(left);
751 uint8x8_t d;
752 (void)above;
753
754 d = vdup_lane_u8(left_u8, 0);
755 vst1_u8(dst, d);
756 dst += stride;
757 d = vdup_lane_u8(left_u8, 1);
758 vst1_u8(dst, d);
759 dst += stride;
760 d = vdup_lane_u8(left_u8, 2);
761 vst1_u8(dst, d);
762 dst += stride;
763 d = vdup_lane_u8(left_u8, 3);
764 vst1_u8(dst, d);
765 dst += stride;
766 d = vdup_lane_u8(left_u8, 4);
767 vst1_u8(dst, d);
768 dst += stride;
769 d = vdup_lane_u8(left_u8, 5);
770 vst1_u8(dst, d);
771 dst += stride;
772 d = vdup_lane_u8(left_u8, 6);
773 vst1_u8(dst, d);
774 dst += stride;
775 d = vdup_lane_u8(left_u8, 7);
776 vst1_u8(dst, d);
777 }
778
h_store_16x8(uint8_t ** dst,const ptrdiff_t stride,const uint8x8_t left)779 static INLINE void h_store_16x8(uint8_t **dst, const ptrdiff_t stride,
780 const uint8x8_t left) {
781 const uint8x16_t row_0 = vdupq_lane_u8(left, 0);
782 const uint8x16_t row_1 = vdupq_lane_u8(left, 1);
783 const uint8x16_t row_2 = vdupq_lane_u8(left, 2);
784 const uint8x16_t row_3 = vdupq_lane_u8(left, 3);
785 const uint8x16_t row_4 = vdupq_lane_u8(left, 4);
786 const uint8x16_t row_5 = vdupq_lane_u8(left, 5);
787 const uint8x16_t row_6 = vdupq_lane_u8(left, 6);
788 const uint8x16_t row_7 = vdupq_lane_u8(left, 7);
789
790 vst1q_u8(*dst, row_0);
791 *dst += stride;
792 vst1q_u8(*dst, row_1);
793 *dst += stride;
794 vst1q_u8(*dst, row_2);
795 *dst += stride;
796 vst1q_u8(*dst, row_3);
797 *dst += stride;
798 vst1q_u8(*dst, row_4);
799 *dst += stride;
800 vst1q_u8(*dst, row_5);
801 *dst += stride;
802 vst1q_u8(*dst, row_6);
803 *dst += stride;
804 vst1q_u8(*dst, row_7);
805 *dst += stride;
806 }
807
vpx_h_predictor_16x16_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)808 void vpx_h_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
809 const uint8_t *above, const uint8_t *left) {
810 const uint8x16_t left_u8q = vld1q_u8(left);
811 (void)above;
812
813 h_store_16x8(&dst, stride, vget_low_u8(left_u8q));
814 h_store_16x8(&dst, stride, vget_high_u8(left_u8q));
815 }
816
h_store_32x8(uint8_t ** dst,const ptrdiff_t stride,const uint8x8_t left)817 static INLINE void h_store_32x8(uint8_t **dst, const ptrdiff_t stride,
818 const uint8x8_t left) {
819 const uint8x16_t row_0 = vdupq_lane_u8(left, 0);
820 const uint8x16_t row_1 = vdupq_lane_u8(left, 1);
821 const uint8x16_t row_2 = vdupq_lane_u8(left, 2);
822 const uint8x16_t row_3 = vdupq_lane_u8(left, 3);
823 const uint8x16_t row_4 = vdupq_lane_u8(left, 4);
824 const uint8x16_t row_5 = vdupq_lane_u8(left, 5);
825 const uint8x16_t row_6 = vdupq_lane_u8(left, 6);
826 const uint8x16_t row_7 = vdupq_lane_u8(left, 7);
827
828 vst1q_u8(*dst, row_0); // Note clang-3.8 produced poor code w/vst2q_u8
829 *dst += 16;
830 vst1q_u8(*dst, row_0);
831 *dst += stride - 16;
832 vst1q_u8(*dst, row_1);
833 *dst += 16;
834 vst1q_u8(*dst, row_1);
835 *dst += stride - 16;
836 vst1q_u8(*dst, row_2);
837 *dst += 16;
838 vst1q_u8(*dst, row_2);
839 *dst += stride - 16;
840 vst1q_u8(*dst, row_3);
841 *dst += 16;
842 vst1q_u8(*dst, row_3);
843 *dst += stride - 16;
844 vst1q_u8(*dst, row_4);
845 *dst += 16;
846 vst1q_u8(*dst, row_4);
847 *dst += stride - 16;
848 vst1q_u8(*dst, row_5);
849 *dst += 16;
850 vst1q_u8(*dst, row_5);
851 *dst += stride - 16;
852 vst1q_u8(*dst, row_6);
853 *dst += 16;
854 vst1q_u8(*dst, row_6);
855 *dst += stride - 16;
856 vst1q_u8(*dst, row_7);
857 *dst += 16;
858 vst1q_u8(*dst, row_7);
859 *dst += stride - 16;
860 }
861
vpx_h_predictor_32x32_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)862 void vpx_h_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
863 const uint8_t *above, const uint8_t *left) {
864 int i;
865 (void)above;
866
867 for (i = 0; i < 2; i++, left += 16) {
868 const uint8x16_t left_u8 = vld1q_u8(left);
869 h_store_32x8(&dst, stride, vget_low_u8(left_u8));
870 h_store_32x8(&dst, stride, vget_high_u8(left_u8));
871 }
872 }
873
874 // -----------------------------------------------------------------------------
875
convert_u8_to_s16(uint8x8_t v)876 static INLINE int16x8_t convert_u8_to_s16(uint8x8_t v) {
877 return vreinterpretq_s16_u16(vmovl_u8(v));
878 }
879
vpx_tm_predictor_4x4_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)880 void vpx_tm_predictor_4x4_neon(uint8_t *dst, ptrdiff_t stride,
881 const uint8_t *above, const uint8_t *left) {
882 const uint8x8_t top_left = vld1_dup_u8(above - 1);
883 const uint8x8_t left_u8 = vld1_u8(left);
884 const uint8x8_t above_u8 = vld1_u8(above);
885 const int16x4_t left_s16 = vget_low_s16(convert_u8_to_s16(left_u8));
886 int16x8_t sub, sum;
887 uint32x2_t d;
888
889 sub = vreinterpretq_s16_u16(vsubl_u8(above_u8, top_left));
890 // Avoid vcombine_s16() which generates lots of redundant code with clang-3.8.
891 sub = vreinterpretq_s16_s64(
892 vdupq_lane_s64(vreinterpret_s64_s16(vget_low_s16(sub)), 0));
893
894 sum = vcombine_s16(vdup_lane_s16(left_s16, 0), vdup_lane_s16(left_s16, 1));
895 sum = vaddq_s16(sum, sub);
896 d = vreinterpret_u32_u8(vqmovun_s16(sum));
897 vst1_lane_u32((uint32_t *)dst, d, 0);
898 dst += stride;
899 vst1_lane_u32((uint32_t *)dst, d, 1);
900 dst += stride;
901
902 sum = vcombine_s16(vdup_lane_s16(left_s16, 2), vdup_lane_s16(left_s16, 3));
903 sum = vaddq_s16(sum, sub);
904 d = vreinterpret_u32_u8(vqmovun_s16(sum));
905 vst1_lane_u32((uint32_t *)dst, d, 0);
906 dst += stride;
907 vst1_lane_u32((uint32_t *)dst, d, 1);
908 }
909
tm_8_kernel(uint8_t ** dst,const ptrdiff_t stride,const int16x8_t left_dup,const int16x8_t sub)910 static INLINE void tm_8_kernel(uint8_t **dst, const ptrdiff_t stride,
911 const int16x8_t left_dup, const int16x8_t sub) {
912 const int16x8_t sum = vaddq_s16(left_dup, sub);
913 const uint8x8_t d = vqmovun_s16(sum);
914 vst1_u8(*dst, d);
915 *dst += stride;
916 }
917
vpx_tm_predictor_8x8_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)918 void vpx_tm_predictor_8x8_neon(uint8_t *dst, ptrdiff_t stride,
919 const uint8_t *above, const uint8_t *left) {
920 const uint8x8_t top_left = vld1_dup_u8(above - 1);
921 const uint8x8_t above_u8 = vld1_u8(above);
922 const uint8x8_t left_u8 = vld1_u8(left);
923 const int16x8_t left_s16q = convert_u8_to_s16(left_u8);
924 const int16x8_t sub = vreinterpretq_s16_u16(vsubl_u8(above_u8, top_left));
925 int16x4_t left_s16d = vget_low_s16(left_s16q);
926 int i;
927
928 for (i = 0; i < 2; i++, left_s16d = vget_high_s16(left_s16q)) {
929 int16x8_t left_dup;
930
931 left_dup = vdupq_lane_s16(left_s16d, 0);
932 tm_8_kernel(&dst, stride, left_dup, sub);
933 left_dup = vdupq_lane_s16(left_s16d, 1);
934 tm_8_kernel(&dst, stride, left_dup, sub);
935 left_dup = vdupq_lane_s16(left_s16d, 2);
936 tm_8_kernel(&dst, stride, left_dup, sub);
937 left_dup = vdupq_lane_s16(left_s16d, 3);
938 tm_8_kernel(&dst, stride, left_dup, sub);
939 }
940 }
941
tm_16_kernel(uint8_t ** dst,const ptrdiff_t stride,const int16x8_t left_dup,const int16x8_t sub0,const int16x8_t sub1)942 static INLINE void tm_16_kernel(uint8_t **dst, const ptrdiff_t stride,
943 const int16x8_t left_dup, const int16x8_t sub0,
944 const int16x8_t sub1) {
945 const int16x8_t sum0 = vaddq_s16(left_dup, sub0);
946 const int16x8_t sum1 = vaddq_s16(left_dup, sub1);
947 const uint8x8_t d0 = vqmovun_s16(sum0);
948 const uint8x8_t d1 = vqmovun_s16(sum1);
949 vst1_u8(*dst, d0);
950 *dst += 8;
951 vst1_u8(*dst, d1);
952 *dst += stride - 8;
953 }
954
vpx_tm_predictor_16x16_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)955 void vpx_tm_predictor_16x16_neon(uint8_t *dst, ptrdiff_t stride,
956 const uint8_t *above, const uint8_t *left) {
957 const uint8x16_t top_left = vld1q_dup_u8(above - 1);
958 const uint8x16_t above_u8 = vld1q_u8(above);
959 const int16x8_t sub0 = vreinterpretq_s16_u16(
960 vsubl_u8(vget_low_u8(above_u8), vget_low_u8(top_left)));
961 const int16x8_t sub1 = vreinterpretq_s16_u16(
962 vsubl_u8(vget_high_u8(above_u8), vget_high_u8(top_left)));
963 int16x8_t left_dup;
964 int i;
965
966 for (i = 0; i < 2; i++, left += 8) {
967 const uint8x8_t left_u8 = vld1_u8(left);
968 const int16x8_t left_s16q = convert_u8_to_s16(left_u8);
969 const int16x4_t left_low = vget_low_s16(left_s16q);
970 const int16x4_t left_high = vget_high_s16(left_s16q);
971
972 left_dup = vdupq_lane_s16(left_low, 0);
973 tm_16_kernel(&dst, stride, left_dup, sub0, sub1);
974 left_dup = vdupq_lane_s16(left_low, 1);
975 tm_16_kernel(&dst, stride, left_dup, sub0, sub1);
976 left_dup = vdupq_lane_s16(left_low, 2);
977 tm_16_kernel(&dst, stride, left_dup, sub0, sub1);
978 left_dup = vdupq_lane_s16(left_low, 3);
979 tm_16_kernel(&dst, stride, left_dup, sub0, sub1);
980
981 left_dup = vdupq_lane_s16(left_high, 0);
982 tm_16_kernel(&dst, stride, left_dup, sub0, sub1);
983 left_dup = vdupq_lane_s16(left_high, 1);
984 tm_16_kernel(&dst, stride, left_dup, sub0, sub1);
985 left_dup = vdupq_lane_s16(left_high, 2);
986 tm_16_kernel(&dst, stride, left_dup, sub0, sub1);
987 left_dup = vdupq_lane_s16(left_high, 3);
988 tm_16_kernel(&dst, stride, left_dup, sub0, sub1);
989 }
990 }
991
tm_32_kernel(uint8_t ** dst,const ptrdiff_t stride,const int16x8_t left_dup,const int16x8_t sub0,const int16x8_t sub1,const int16x8_t sub2,const int16x8_t sub3)992 static INLINE void tm_32_kernel(uint8_t **dst, const ptrdiff_t stride,
993 const int16x8_t left_dup, const int16x8_t sub0,
994 const int16x8_t sub1, const int16x8_t sub2,
995 const int16x8_t sub3) {
996 const int16x8_t sum0 = vaddq_s16(left_dup, sub0);
997 const int16x8_t sum1 = vaddq_s16(left_dup, sub1);
998 const int16x8_t sum2 = vaddq_s16(left_dup, sub2);
999 const int16x8_t sum3 = vaddq_s16(left_dup, sub3);
1000 const uint8x8_t d0 = vqmovun_s16(sum0);
1001 const uint8x8_t d1 = vqmovun_s16(sum1);
1002 const uint8x8_t d2 = vqmovun_s16(sum2);
1003 const uint8x8_t d3 = vqmovun_s16(sum3);
1004
1005 vst1q_u8(*dst, vcombine_u8(d0, d1));
1006 *dst += 16;
1007 vst1q_u8(*dst, vcombine_u8(d2, d3));
1008 *dst += stride - 16;
1009 }
1010
vpx_tm_predictor_32x32_neon(uint8_t * dst,ptrdiff_t stride,const uint8_t * above,const uint8_t * left)1011 void vpx_tm_predictor_32x32_neon(uint8_t *dst, ptrdiff_t stride,
1012 const uint8_t *above, const uint8_t *left) {
1013 const uint8x16_t top_left = vld1q_dup_u8(above - 1);
1014 const uint8x16_t above_low = vld1q_u8(above);
1015 const uint8x16_t above_high = vld1q_u8(above + 16);
1016 const int16x8_t sub0 = vreinterpretq_s16_u16(
1017 vsubl_u8(vget_low_u8(above_low), vget_low_u8(top_left)));
1018 const int16x8_t sub1 = vreinterpretq_s16_u16(
1019 vsubl_u8(vget_high_u8(above_low), vget_high_u8(top_left)));
1020 const int16x8_t sub2 = vreinterpretq_s16_u16(
1021 vsubl_u8(vget_low_u8(above_high), vget_low_u8(top_left)));
1022 const int16x8_t sub3 = vreinterpretq_s16_u16(
1023 vsubl_u8(vget_high_u8(above_high), vget_high_u8(top_left)));
1024 int16x8_t left_dup;
1025 int i, j;
1026
1027 for (j = 0; j < 4; j++, left += 8) {
1028 const uint8x8_t left_u8 = vld1_u8(left);
1029 const int16x8_t left_s16q = convert_u8_to_s16(left_u8);
1030 int16x4_t left_s16d = vget_low_s16(left_s16q);
1031 for (i = 0; i < 2; i++, left_s16d = vget_high_s16(left_s16q)) {
1032 left_dup = vdupq_lane_s16(left_s16d, 0);
1033 tm_32_kernel(&dst, stride, left_dup, sub0, sub1, sub2, sub3);
1034 left_dup = vdupq_lane_s16(left_s16d, 1);
1035 tm_32_kernel(&dst, stride, left_dup, sub0, sub1, sub2, sub3);
1036 left_dup = vdupq_lane_s16(left_s16d, 2);
1037 tm_32_kernel(&dst, stride, left_dup, sub0, sub1, sub2, sub3);
1038 left_dup = vdupq_lane_s16(left_s16d, 3);
1039 tm_32_kernel(&dst, stride, left_dup, sub0, sub1, sub2, sub3);
1040 }
1041 }
1042 }
1043 #endif // !HAVE_NEON_ASM
1044