1 /*
2  *  Copyright (c) 2016 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 #include <assert.h>
13 
14 #include "./vpx_config.h"
15 #include "./vpx_dsp_rtcd.h"
16 #include "vpx/vpx_integer.h"
17 #include "vpx_dsp/arm/transpose_neon.h"
18 #include "vpx_ports/mem.h"
19 
load_4x4(const int16_t * s,ptrdiff_t p,int16x4_t * s0,int16x4_t * s1,int16x4_t * s2,int16x4_t * s3)20 static INLINE void load_4x4(const int16_t *s, ptrdiff_t p, int16x4_t *s0,
21                             int16x4_t *s1, int16x4_t *s2, int16x4_t *s3) {
22   *s0 = vld1_s16(s);
23   s += p;
24   *s1 = vld1_s16(s);
25   s += p;
26   *s2 = vld1_s16(s);
27   s += p;
28   *s3 = vld1_s16(s);
29 }
30 
load_8x4(const uint16_t * s,ptrdiff_t p,uint16x8_t * s0,uint16x8_t * s1,uint16x8_t * s2,uint16x8_t * s3)31 static INLINE void load_8x4(const uint16_t *s, ptrdiff_t p, uint16x8_t *s0,
32                             uint16x8_t *s1, uint16x8_t *s2, uint16x8_t *s3) {
33   *s0 = vld1q_u16(s);
34   s += p;
35   *s1 = vld1q_u16(s);
36   s += p;
37   *s2 = vld1q_u16(s);
38   s += p;
39   *s3 = vld1q_u16(s);
40 }
41 
load_8x8(const int16_t * s,ptrdiff_t p,int16x8_t * s0,int16x8_t * s1,int16x8_t * s2,int16x8_t * s3,int16x8_t * s4,int16x8_t * s5,int16x8_t * s6,int16x8_t * s7)42 static INLINE void load_8x8(const int16_t *s, ptrdiff_t p, int16x8_t *s0,
43                             int16x8_t *s1, int16x8_t *s2, int16x8_t *s3,
44                             int16x8_t *s4, int16x8_t *s5, int16x8_t *s6,
45                             int16x8_t *s7) {
46   *s0 = vld1q_s16(s);
47   s += p;
48   *s1 = vld1q_s16(s);
49   s += p;
50   *s2 = vld1q_s16(s);
51   s += p;
52   *s3 = vld1q_s16(s);
53   s += p;
54   *s4 = vld1q_s16(s);
55   s += p;
56   *s5 = vld1q_s16(s);
57   s += p;
58   *s6 = vld1q_s16(s);
59   s += p;
60   *s7 = vld1q_s16(s);
61 }
62 
store_8x8(uint16_t * s,ptrdiff_t p,const uint16x8_t s0,const uint16x8_t s1,const uint16x8_t s2,const uint16x8_t s3,const uint16x8_t s4,const uint16x8_t s5,const uint16x8_t s6,const uint16x8_t s7)63 static INLINE void store_8x8(uint16_t *s, ptrdiff_t p, const uint16x8_t s0,
64                              const uint16x8_t s1, const uint16x8_t s2,
65                              const uint16x8_t s3, const uint16x8_t s4,
66                              const uint16x8_t s5, const uint16x8_t s6,
67                              const uint16x8_t s7) {
68   vst1q_u16(s, s0);
69   s += p;
70   vst1q_u16(s, s1);
71   s += p;
72   vst1q_u16(s, s2);
73   s += p;
74   vst1q_u16(s, s3);
75   s += p;
76   vst1q_u16(s, s4);
77   s += p;
78   vst1q_u16(s, s5);
79   s += p;
80   vst1q_u16(s, s6);
81   s += p;
82   vst1q_u16(s, s7);
83 }
84 
convolve8_4(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 filters)85 static INLINE int32x4_t convolve8_4(const int16x4_t s0, const int16x4_t s1,
86                                     const int16x4_t s2, const int16x4_t s3,
87                                     const int16x4_t s4, const int16x4_t s5,
88                                     const int16x4_t s6, const int16x4_t s7,
89                                     const int16x8_t filters) {
90   const int16x4_t filters_lo = vget_low_s16(filters);
91   const int16x4_t filters_hi = vget_high_s16(filters);
92   int32x4_t sum = vdupq_n_s32(0);
93 
94   sum = vmlal_lane_s16(sum, s0, filters_lo, 0);
95   sum = vmlal_lane_s16(sum, s1, filters_lo, 1);
96   sum = vmlal_lane_s16(sum, s2, filters_lo, 2);
97   sum = vmlal_lane_s16(sum, s3, filters_lo, 3);
98   sum = vmlal_lane_s16(sum, s4, filters_hi, 0);
99   sum = vmlal_lane_s16(sum, s5, filters_hi, 1);
100   sum = vmlal_lane_s16(sum, s6, filters_hi, 2);
101   sum = vmlal_lane_s16(sum, s7, filters_hi, 3);
102   return sum;
103 }
104 
convolve8_8(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 filters,const uint16x8_t max)105 static INLINE uint16x8_t convolve8_8(const int16x8_t s0, const int16x8_t s1,
106                                      const int16x8_t s2, const int16x8_t s3,
107                                      const int16x8_t s4, const int16x8_t s5,
108                                      const int16x8_t s6, const int16x8_t s7,
109                                      const int16x8_t filters,
110                                      const uint16x8_t max) {
111   const int16x4_t filters_lo = vget_low_s16(filters);
112   const int16x4_t filters_hi = vget_high_s16(filters);
113   int32x4_t sum0 = vdupq_n_s32(0);
114   int32x4_t sum1 = vdupq_n_s32(0);
115   uint16x8_t d;
116 
117   sum0 = vmlal_lane_s16(sum0, vget_low_s16(s0), filters_lo, 0);
118   sum0 = vmlal_lane_s16(sum0, vget_low_s16(s1), filters_lo, 1);
119   sum0 = vmlal_lane_s16(sum0, vget_low_s16(s2), filters_lo, 2);
120   sum0 = vmlal_lane_s16(sum0, vget_low_s16(s3), filters_lo, 3);
121   sum0 = vmlal_lane_s16(sum0, vget_low_s16(s4), filters_hi, 0);
122   sum0 = vmlal_lane_s16(sum0, vget_low_s16(s5), filters_hi, 1);
123   sum0 = vmlal_lane_s16(sum0, vget_low_s16(s6), filters_hi, 2);
124   sum0 = vmlal_lane_s16(sum0, vget_low_s16(s7), filters_hi, 3);
125   sum1 = vmlal_lane_s16(sum1, vget_high_s16(s0), filters_lo, 0);
126   sum1 = vmlal_lane_s16(sum1, vget_high_s16(s1), filters_lo, 1);
127   sum1 = vmlal_lane_s16(sum1, vget_high_s16(s2), filters_lo, 2);
128   sum1 = vmlal_lane_s16(sum1, vget_high_s16(s3), filters_lo, 3);
129   sum1 = vmlal_lane_s16(sum1, vget_high_s16(s4), filters_hi, 0);
130   sum1 = vmlal_lane_s16(sum1, vget_high_s16(s5), filters_hi, 1);
131   sum1 = vmlal_lane_s16(sum1, vget_high_s16(s6), filters_hi, 2);
132   sum1 = vmlal_lane_s16(sum1, vget_high_s16(s7), filters_hi, 3);
133   d = vcombine_u16(vqrshrun_n_s32(sum0, 7), vqrshrun_n_s32(sum1, 7));
134   d = vminq_u16(d, max);
135   return d;
136 }
137 
vpx_highbd_convolve8_horiz_neon(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 w,int h,int bd)138 void vpx_highbd_convolve8_horiz_neon(const uint8_t *src8, ptrdiff_t src_stride,
139                                      uint8_t *dst8, ptrdiff_t dst_stride,
140                                      const int16_t *filter_x, int x_step_q4,
141                                      const int16_t *filter_y,  // unused
142                                      int y_step_q4,            // unused
143                                      int w, int h, int bd) {
144   if (x_step_q4 != 16) {
145     vpx_highbd_convolve8_horiz_c(src8, src_stride, dst8, dst_stride, filter_x,
146                                  x_step_q4, filter_y, y_step_q4, w, h, bd);
147   } else {
148     const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
149     uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
150     const int16x8_t filters = vld1q_s16(filter_x);
151     const uint16x8_t max = vdupq_n_u16((1 << bd) - 1);
152     uint16x8_t t0, t1, t2, t3;
153 
154     assert(!((intptr_t)dst & 3));
155     assert(!(dst_stride & 3));
156 
157     src -= 3;
158 
159     if (h == 4) {
160       int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
161       int32x4_t d0, d1, d2, d3;
162       uint16x8_t d01, d23;
163 
164       __builtin_prefetch(src + 0 * src_stride);
165       __builtin_prefetch(src + 1 * src_stride);
166       __builtin_prefetch(src + 2 * src_stride);
167       __builtin_prefetch(src + 3 * src_stride);
168       load_8x4(src, src_stride, &t0, &t1, &t2, &t3);
169       transpose_u16_8x4(&t0, &t1, &t2, &t3);
170       s0 = vreinterpret_s16_u16(vget_low_u16(t0));
171       s1 = vreinterpret_s16_u16(vget_low_u16(t1));
172       s2 = vreinterpret_s16_u16(vget_low_u16(t2));
173       s3 = vreinterpret_s16_u16(vget_low_u16(t3));
174       s4 = vreinterpret_s16_u16(vget_high_u16(t0));
175       s5 = vreinterpret_s16_u16(vget_high_u16(t1));
176       s6 = vreinterpret_s16_u16(vget_high_u16(t2));
177       __builtin_prefetch(dst + 0 * dst_stride);
178       __builtin_prefetch(dst + 1 * dst_stride);
179       __builtin_prefetch(dst + 2 * dst_stride);
180       __builtin_prefetch(dst + 3 * dst_stride);
181       src += 7;
182 
183       do {
184         load_4x4((const int16_t *)src, src_stride, &s7, &s8, &s9, &s10);
185         transpose_s16_4x4d(&s7, &s8, &s9, &s10);
186 
187         d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
188         d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
189         d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
190         d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
191 
192         d01 = vcombine_u16(vqrshrun_n_s32(d0, 7), vqrshrun_n_s32(d1, 7));
193         d23 = vcombine_u16(vqrshrun_n_s32(d2, 7), vqrshrun_n_s32(d3, 7));
194         d01 = vminq_u16(d01, max);
195         d23 = vminq_u16(d23, max);
196         transpose_u16_4x4q(&d01, &d23);
197 
198         vst1_u16(dst + 0 * dst_stride, vget_low_u16(d01));
199         vst1_u16(dst + 1 * dst_stride, vget_low_u16(d23));
200         vst1_u16(dst + 2 * dst_stride, vget_high_u16(d01));
201         vst1_u16(dst + 3 * dst_stride, vget_high_u16(d23));
202 
203         s0 = s4;
204         s1 = s5;
205         s2 = s6;
206         s3 = s7;
207         s4 = s8;
208         s5 = s9;
209         s6 = s10;
210         src += 4;
211         dst += 4;
212         w -= 4;
213       } while (w > 0);
214     } else {
215       int16x8_t t4, t5, t6, t7;
216       int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
217       uint16x8_t d0, d1, d2, d3;
218 
219       if (w == 4) {
220         do {
221           load_8x8((const int16_t *)src, src_stride, &s0, &s1, &s2, &s3, &s4,
222                    &s5, &s6, &s7);
223           transpose_s16_8x8(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
224 
225           load_8x8((const int16_t *)(src + 7), src_stride, &s7, &s8, &s9, &s10,
226                    &t4, &t5, &t6, &t7);
227           src += 8 * src_stride;
228           __builtin_prefetch(dst + 0 * dst_stride);
229           __builtin_prefetch(dst + 1 * dst_stride);
230           __builtin_prefetch(dst + 2 * dst_stride);
231           __builtin_prefetch(dst + 3 * dst_stride);
232           __builtin_prefetch(dst + 4 * dst_stride);
233           __builtin_prefetch(dst + 5 * dst_stride);
234           __builtin_prefetch(dst + 6 * dst_stride);
235           __builtin_prefetch(dst + 7 * dst_stride);
236           transpose_s16_8x8(&s7, &s8, &s9, &s10, &t4, &t5, &t6, &t7);
237 
238           __builtin_prefetch(src + 0 * src_stride);
239           __builtin_prefetch(src + 1 * src_stride);
240           __builtin_prefetch(src + 2 * src_stride);
241           __builtin_prefetch(src + 3 * src_stride);
242           __builtin_prefetch(src + 4 * src_stride);
243           __builtin_prefetch(src + 5 * src_stride);
244           __builtin_prefetch(src + 6 * src_stride);
245           __builtin_prefetch(src + 7 * src_stride);
246           d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
247           d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
248           d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
249           d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
250 
251           transpose_u16_8x4(&d0, &d1, &d2, &d3);
252           vst1_u16(dst, vget_low_u16(d0));
253           dst += dst_stride;
254           vst1_u16(dst, vget_low_u16(d1));
255           dst += dst_stride;
256           vst1_u16(dst, vget_low_u16(d2));
257           dst += dst_stride;
258           vst1_u16(dst, vget_low_u16(d3));
259           dst += dst_stride;
260           vst1_u16(dst, vget_high_u16(d0));
261           dst += dst_stride;
262           vst1_u16(dst, vget_high_u16(d1));
263           dst += dst_stride;
264           vst1_u16(dst, vget_high_u16(d2));
265           dst += dst_stride;
266           vst1_u16(dst, vget_high_u16(d3));
267           dst += dst_stride;
268           h -= 8;
269         } while (h > 0);
270       } else {
271         int width;
272         const uint16_t *s;
273         uint16_t *d;
274         int16x8_t s11, s12, s13, s14;
275         uint16x8_t d4, d5, d6, d7;
276 
277         do {
278           __builtin_prefetch(src + 0 * src_stride);
279           __builtin_prefetch(src + 1 * src_stride);
280           __builtin_prefetch(src + 2 * src_stride);
281           __builtin_prefetch(src + 3 * src_stride);
282           __builtin_prefetch(src + 4 * src_stride);
283           __builtin_prefetch(src + 5 * src_stride);
284           __builtin_prefetch(src + 6 * src_stride);
285           __builtin_prefetch(src + 7 * src_stride);
286           load_8x8((const int16_t *)src, src_stride, &s0, &s1, &s2, &s3, &s4,
287                    &s5, &s6, &s7);
288           transpose_s16_8x8(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
289 
290           width = w;
291           s = src + 7;
292           d = dst;
293           __builtin_prefetch(dst + 0 * dst_stride);
294           __builtin_prefetch(dst + 1 * dst_stride);
295           __builtin_prefetch(dst + 2 * dst_stride);
296           __builtin_prefetch(dst + 3 * dst_stride);
297           __builtin_prefetch(dst + 4 * dst_stride);
298           __builtin_prefetch(dst + 5 * dst_stride);
299           __builtin_prefetch(dst + 6 * dst_stride);
300           __builtin_prefetch(dst + 7 * dst_stride);
301 
302           do {
303             load_8x8((const int16_t *)s, src_stride, &s7, &s8, &s9, &s10, &s11,
304                      &s12, &s13, &s14);
305             transpose_s16_8x8(&s7, &s8, &s9, &s10, &s11, &s12, &s13, &s14);
306 
307             d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
308             d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
309             d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
310             d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
311             d4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filters, max);
312             d5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filters, max);
313             d6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filters, max);
314             d7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filters, max);
315 
316             transpose_u16_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
317             store_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7);
318 
319             s0 = s8;
320             s1 = s9;
321             s2 = s10;
322             s3 = s11;
323             s4 = s12;
324             s5 = s13;
325             s6 = s14;
326             s += 8;
327             d += 8;
328             width -= 8;
329           } while (width > 0);
330           src += 8 * src_stride;
331           dst += 8 * dst_stride;
332           h -= 8;
333         } while (h > 0);
334       }
335     }
336   }
337 }
338 
vpx_highbd_convolve8_avg_horiz_neon(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 w,int h,int bd)339 void vpx_highbd_convolve8_avg_horiz_neon(const uint8_t *src8,
340                                          ptrdiff_t src_stride, uint8_t *dst8,
341                                          ptrdiff_t dst_stride,
342                                          const int16_t *filter_x, int x_step_q4,
343                                          const int16_t *filter_y,  // unused
344                                          int y_step_q4,            // unused
345                                          int w, int h, int bd) {
346   if (x_step_q4 != 16) {
347     vpx_highbd_convolve8_avg_horiz_c(src8, src_stride, dst8, dst_stride,
348                                      filter_x, x_step_q4, filter_y, y_step_q4,
349                                      w, h, bd);
350   } else {
351     const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
352     uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
353     const int16x8_t filters = vld1q_s16(filter_x);
354     const uint16x8_t max = vdupq_n_u16((1 << bd) - 1);
355     uint16x8_t t0, t1, t2, t3;
356 
357     assert(!((intptr_t)dst & 3));
358     assert(!(dst_stride & 3));
359 
360     src -= 3;
361 
362     if (h == 4) {
363       int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
364       int32x4_t d0, d1, d2, d3;
365       uint16x8_t d01, d23, t01, t23;
366 
367       __builtin_prefetch(src + 0 * src_stride);
368       __builtin_prefetch(src + 1 * src_stride);
369       __builtin_prefetch(src + 2 * src_stride);
370       __builtin_prefetch(src + 3 * src_stride);
371       load_8x4(src, src_stride, &t0, &t1, &t2, &t3);
372       transpose_u16_8x4(&t0, &t1, &t2, &t3);
373       s0 = vreinterpret_s16_u16(vget_low_u16(t0));
374       s1 = vreinterpret_s16_u16(vget_low_u16(t1));
375       s2 = vreinterpret_s16_u16(vget_low_u16(t2));
376       s3 = vreinterpret_s16_u16(vget_low_u16(t3));
377       s4 = vreinterpret_s16_u16(vget_high_u16(t0));
378       s5 = vreinterpret_s16_u16(vget_high_u16(t1));
379       s6 = vreinterpret_s16_u16(vget_high_u16(t2));
380       __builtin_prefetch(dst + 0 * dst_stride);
381       __builtin_prefetch(dst + 1 * dst_stride);
382       __builtin_prefetch(dst + 2 * dst_stride);
383       __builtin_prefetch(dst + 3 * dst_stride);
384       src += 7;
385 
386       do {
387         load_4x4((const int16_t *)src, src_stride, &s7, &s8, &s9, &s10);
388         transpose_s16_4x4d(&s7, &s8, &s9, &s10);
389 
390         d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
391         d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
392         d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
393         d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
394 
395         t01 = vcombine_u16(vqrshrun_n_s32(d0, 7), vqrshrun_n_s32(d1, 7));
396         t23 = vcombine_u16(vqrshrun_n_s32(d2, 7), vqrshrun_n_s32(d3, 7));
397         t01 = vminq_u16(t01, max);
398         t23 = vminq_u16(t23, max);
399         transpose_u16_4x4q(&t01, &t23);
400 
401         d01 = vcombine_u16(vld1_u16(dst + 0 * dst_stride),
402                            vld1_u16(dst + 2 * dst_stride));
403         d23 = vcombine_u16(vld1_u16(dst + 1 * dst_stride),
404                            vld1_u16(dst + 3 * dst_stride));
405         d01 = vrhaddq_u16(d01, t01);
406         d23 = vrhaddq_u16(d23, t23);
407 
408         vst1_u16(dst + 0 * dst_stride, vget_low_u16(d01));
409         vst1_u16(dst + 1 * dst_stride, vget_low_u16(d23));
410         vst1_u16(dst + 2 * dst_stride, vget_high_u16(d01));
411         vst1_u16(dst + 3 * dst_stride, vget_high_u16(d23));
412 
413         s0 = s4;
414         s1 = s5;
415         s2 = s6;
416         s3 = s7;
417         s4 = s8;
418         s5 = s9;
419         s6 = s10;
420         src += 4;
421         dst += 4;
422         w -= 4;
423       } while (w > 0);
424     } else {
425       int16x8_t t4, t5, t6, t7;
426       int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
427       uint16x8_t d0, d1, d2, d3, t0, t1, t2, t3;
428 
429       if (w == 4) {
430         do {
431           load_8x8((const int16_t *)src, src_stride, &s0, &s1, &s2, &s3, &s4,
432                    &s5, &s6, &s7);
433           transpose_s16_8x8(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
434 
435           load_8x8((const int16_t *)(src + 7), src_stride, &s7, &s8, &s9, &s10,
436                    &t4, &t5, &t6, &t7);
437           src += 8 * src_stride;
438           __builtin_prefetch(dst + 0 * dst_stride);
439           __builtin_prefetch(dst + 1 * dst_stride);
440           __builtin_prefetch(dst + 2 * dst_stride);
441           __builtin_prefetch(dst + 3 * dst_stride);
442           __builtin_prefetch(dst + 4 * dst_stride);
443           __builtin_prefetch(dst + 5 * dst_stride);
444           __builtin_prefetch(dst + 6 * dst_stride);
445           __builtin_prefetch(dst + 7 * dst_stride);
446           transpose_s16_8x8(&s7, &s8, &s9, &s10, &t4, &t5, &t6, &t7);
447 
448           __builtin_prefetch(src + 0 * src_stride);
449           __builtin_prefetch(src + 1 * src_stride);
450           __builtin_prefetch(src + 2 * src_stride);
451           __builtin_prefetch(src + 3 * src_stride);
452           __builtin_prefetch(src + 4 * src_stride);
453           __builtin_prefetch(src + 5 * src_stride);
454           __builtin_prefetch(src + 6 * src_stride);
455           __builtin_prefetch(src + 7 * src_stride);
456           t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
457           t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
458           t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
459           t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
460           transpose_u16_8x4(&t0, &t1, &t2, &t3);
461 
462           d0 = vcombine_u16(vld1_u16(dst + 0 * dst_stride),
463                             vld1_u16(dst + 4 * dst_stride));
464           d1 = vcombine_u16(vld1_u16(dst + 1 * dst_stride),
465                             vld1_u16(dst + 5 * dst_stride));
466           d2 = vcombine_u16(vld1_u16(dst + 2 * dst_stride),
467                             vld1_u16(dst + 6 * dst_stride));
468           d3 = vcombine_u16(vld1_u16(dst + 3 * dst_stride),
469                             vld1_u16(dst + 7 * dst_stride));
470           d0 = vrhaddq_u16(d0, t0);
471           d1 = vrhaddq_u16(d1, t1);
472           d2 = vrhaddq_u16(d2, t2);
473           d3 = vrhaddq_u16(d3, t3);
474 
475           vst1_u16(dst, vget_low_u16(d0));
476           dst += dst_stride;
477           vst1_u16(dst, vget_low_u16(d1));
478           dst += dst_stride;
479           vst1_u16(dst, vget_low_u16(d2));
480           dst += dst_stride;
481           vst1_u16(dst, vget_low_u16(d3));
482           dst += dst_stride;
483           vst1_u16(dst, vget_high_u16(d0));
484           dst += dst_stride;
485           vst1_u16(dst, vget_high_u16(d1));
486           dst += dst_stride;
487           vst1_u16(dst, vget_high_u16(d2));
488           dst += dst_stride;
489           vst1_u16(dst, vget_high_u16(d3));
490           dst += dst_stride;
491           h -= 8;
492         } while (h > 0);
493       } else {
494         int width;
495         const uint16_t *s;
496         uint16_t *d;
497         int16x8_t s11, s12, s13, s14;
498         uint16x8_t d4, d5, d6, d7;
499 
500         do {
501           __builtin_prefetch(src + 0 * src_stride);
502           __builtin_prefetch(src + 1 * src_stride);
503           __builtin_prefetch(src + 2 * src_stride);
504           __builtin_prefetch(src + 3 * src_stride);
505           __builtin_prefetch(src + 4 * src_stride);
506           __builtin_prefetch(src + 5 * src_stride);
507           __builtin_prefetch(src + 6 * src_stride);
508           __builtin_prefetch(src + 7 * src_stride);
509           load_8x8((const int16_t *)src, src_stride, &s0, &s1, &s2, &s3, &s4,
510                    &s5, &s6, &s7);
511           transpose_s16_8x8(&s0, &s1, &s2, &s3, &s4, &s5, &s6, &s7);
512 
513           width = w;
514           s = src + 7;
515           d = dst;
516           __builtin_prefetch(dst + 0 * dst_stride);
517           __builtin_prefetch(dst + 1 * dst_stride);
518           __builtin_prefetch(dst + 2 * dst_stride);
519           __builtin_prefetch(dst + 3 * dst_stride);
520           __builtin_prefetch(dst + 4 * dst_stride);
521           __builtin_prefetch(dst + 5 * dst_stride);
522           __builtin_prefetch(dst + 6 * dst_stride);
523           __builtin_prefetch(dst + 7 * dst_stride);
524 
525           do {
526             load_8x8((const int16_t *)s, src_stride, &s7, &s8, &s9, &s10, &s11,
527                      &s12, &s13, &s14);
528             transpose_s16_8x8(&s7, &s8, &s9, &s10, &s11, &s12, &s13, &s14);
529 
530             d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
531             d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
532             d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
533             d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
534             d4 = convolve8_8(s4, s5, s6, s7, s8, s9, s10, s11, filters, max);
535             d5 = convolve8_8(s5, s6, s7, s8, s9, s10, s11, s12, filters, max);
536             d6 = convolve8_8(s6, s7, s8, s9, s10, s11, s12, s13, filters, max);
537             d7 = convolve8_8(s7, s8, s9, s10, s11, s12, s13, s14, filters, max);
538 
539             transpose_u16_8x8(&d0, &d1, &d2, &d3, &d4, &d5, &d6, &d7);
540 
541             d0 = vrhaddq_u16(d0, vld1q_u16(d + 0 * dst_stride));
542             d1 = vrhaddq_u16(d1, vld1q_u16(d + 1 * dst_stride));
543             d2 = vrhaddq_u16(d2, vld1q_u16(d + 2 * dst_stride));
544             d3 = vrhaddq_u16(d3, vld1q_u16(d + 3 * dst_stride));
545             d4 = vrhaddq_u16(d4, vld1q_u16(d + 4 * dst_stride));
546             d5 = vrhaddq_u16(d5, vld1q_u16(d + 5 * dst_stride));
547             d6 = vrhaddq_u16(d6, vld1q_u16(d + 6 * dst_stride));
548             d7 = vrhaddq_u16(d7, vld1q_u16(d + 7 * dst_stride));
549 
550             store_8x8(d, dst_stride, d0, d1, d2, d3, d4, d5, d6, d7);
551 
552             s0 = s8;
553             s1 = s9;
554             s2 = s10;
555             s3 = s11;
556             s4 = s12;
557             s5 = s13;
558             s6 = s14;
559             s += 8;
560             d += 8;
561             width -= 8;
562           } while (width > 0);
563           src += 8 * src_stride;
564           dst += 8 * dst_stride;
565           h -= 8;
566         } while (h > 0);
567       }
568     }
569   }
570 }
571 
vpx_highbd_convolve8_vert_neon(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 w,int h,int bd)572 void vpx_highbd_convolve8_vert_neon(const uint8_t *src8, ptrdiff_t src_stride,
573                                     uint8_t *dst8, ptrdiff_t dst_stride,
574                                     const int16_t *filter_x,  // unused
575                                     int x_step_q4,            // unused
576                                     const int16_t *filter_y, int y_step_q4,
577                                     int w, int h, int bd) {
578   if (y_step_q4 != 16) {
579     vpx_highbd_convolve8_vert_c(src8, src_stride, dst8, dst_stride, filter_x,
580                                 x_step_q4, filter_y, y_step_q4, w, h, bd);
581   } else {
582     const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
583     uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
584     const int16x8_t filters = vld1q_s16(filter_y);
585     const uint16x8_t max = vdupq_n_u16((1 << bd) - 1);
586 
587     assert(!((intptr_t)dst & 3));
588     assert(!(dst_stride & 3));
589 
590     src -= 3 * src_stride;
591 
592     if (w == 4) {
593       int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
594       int32x4_t d0, d1, d2, d3;
595       uint16x8_t d01, d23;
596 
597       s0 = vreinterpret_s16_u16(vld1_u16(src));
598       src += src_stride;
599       s1 = vreinterpret_s16_u16(vld1_u16(src));
600       src += src_stride;
601       s2 = vreinterpret_s16_u16(vld1_u16(src));
602       src += src_stride;
603       s3 = vreinterpret_s16_u16(vld1_u16(src));
604       src += src_stride;
605       s4 = vreinterpret_s16_u16(vld1_u16(src));
606       src += src_stride;
607       s5 = vreinterpret_s16_u16(vld1_u16(src));
608       src += src_stride;
609       s6 = vreinterpret_s16_u16(vld1_u16(src));
610       src += src_stride;
611 
612       do {
613         s7 = vreinterpret_s16_u16(vld1_u16(src));
614         src += src_stride;
615         s8 = vreinterpret_s16_u16(vld1_u16(src));
616         src += src_stride;
617         s9 = vreinterpret_s16_u16(vld1_u16(src));
618         src += src_stride;
619         s10 = vreinterpret_s16_u16(vld1_u16(src));
620         src += src_stride;
621 
622         __builtin_prefetch(dst + 0 * dst_stride);
623         __builtin_prefetch(dst + 1 * dst_stride);
624         __builtin_prefetch(dst + 2 * dst_stride);
625         __builtin_prefetch(dst + 3 * dst_stride);
626         __builtin_prefetch(src + 0 * src_stride);
627         __builtin_prefetch(src + 1 * src_stride);
628         __builtin_prefetch(src + 2 * src_stride);
629         __builtin_prefetch(src + 3 * src_stride);
630         d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
631         d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
632         d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
633         d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
634 
635         d01 = vcombine_u16(vqrshrun_n_s32(d0, 7), vqrshrun_n_s32(d1, 7));
636         d23 = vcombine_u16(vqrshrun_n_s32(d2, 7), vqrshrun_n_s32(d3, 7));
637         d01 = vminq_u16(d01, max);
638         d23 = vminq_u16(d23, max);
639         vst1_u16(dst, vget_low_u16(d01));
640         dst += dst_stride;
641         vst1_u16(dst, vget_high_u16(d01));
642         dst += dst_stride;
643         vst1_u16(dst, vget_low_u16(d23));
644         dst += dst_stride;
645         vst1_u16(dst, vget_high_u16(d23));
646         dst += dst_stride;
647 
648         s0 = s4;
649         s1 = s5;
650         s2 = s6;
651         s3 = s7;
652         s4 = s8;
653         s5 = s9;
654         s6 = s10;
655         h -= 4;
656       } while (h > 0);
657     } else {
658       int height;
659       const uint16_t *s;
660       uint16_t *d;
661       int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
662       uint16x8_t d0, d1, d2, d3;
663 
664       do {
665         __builtin_prefetch(src + 0 * src_stride);
666         __builtin_prefetch(src + 1 * src_stride);
667         __builtin_prefetch(src + 2 * src_stride);
668         __builtin_prefetch(src + 3 * src_stride);
669         __builtin_prefetch(src + 4 * src_stride);
670         __builtin_prefetch(src + 5 * src_stride);
671         __builtin_prefetch(src + 6 * src_stride);
672         s = src;
673         s0 = vreinterpretq_s16_u16(vld1q_u16(s));
674         s += src_stride;
675         s1 = vreinterpretq_s16_u16(vld1q_u16(s));
676         s += src_stride;
677         s2 = vreinterpretq_s16_u16(vld1q_u16(s));
678         s += src_stride;
679         s3 = vreinterpretq_s16_u16(vld1q_u16(s));
680         s += src_stride;
681         s4 = vreinterpretq_s16_u16(vld1q_u16(s));
682         s += src_stride;
683         s5 = vreinterpretq_s16_u16(vld1q_u16(s));
684         s += src_stride;
685         s6 = vreinterpretq_s16_u16(vld1q_u16(s));
686         s += src_stride;
687         d = dst;
688         height = h;
689 
690         do {
691           s7 = vreinterpretq_s16_u16(vld1q_u16(s));
692           s += src_stride;
693           s8 = vreinterpretq_s16_u16(vld1q_u16(s));
694           s += src_stride;
695           s9 = vreinterpretq_s16_u16(vld1q_u16(s));
696           s += src_stride;
697           s10 = vreinterpretq_s16_u16(vld1q_u16(s));
698           s += src_stride;
699 
700           __builtin_prefetch(d + 0 * dst_stride);
701           __builtin_prefetch(d + 1 * dst_stride);
702           __builtin_prefetch(d + 2 * dst_stride);
703           __builtin_prefetch(d + 3 * dst_stride);
704           __builtin_prefetch(s + 0 * src_stride);
705           __builtin_prefetch(s + 1 * src_stride);
706           __builtin_prefetch(s + 2 * src_stride);
707           __builtin_prefetch(s + 3 * src_stride);
708           d0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
709           d1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
710           d2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
711           d3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
712 
713           vst1q_u16(d, d0);
714           d += dst_stride;
715           vst1q_u16(d, d1);
716           d += dst_stride;
717           vst1q_u16(d, d2);
718           d += dst_stride;
719           vst1q_u16(d, d3);
720           d += dst_stride;
721 
722           s0 = s4;
723           s1 = s5;
724           s2 = s6;
725           s3 = s7;
726           s4 = s8;
727           s5 = s9;
728           s6 = s10;
729           height -= 4;
730         } while (height > 0);
731         src += 8;
732         dst += 8;
733         w -= 8;
734       } while (w > 0);
735     }
736   }
737 }
738 
vpx_highbd_convolve8_avg_vert_neon(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 w,int h,int bd)739 void vpx_highbd_convolve8_avg_vert_neon(const uint8_t *src8,
740                                         ptrdiff_t src_stride, uint8_t *dst8,
741                                         ptrdiff_t dst_stride,
742                                         const int16_t *filter_x,  // unused
743                                         int x_step_q4,            // unused
744                                         const int16_t *filter_y, int y_step_q4,
745                                         int w, int h, int bd) {
746   if (y_step_q4 != 16) {
747     vpx_highbd_convolve8_avg_vert_c(src8, src_stride, dst8, dst_stride,
748                                     filter_x, x_step_q4, filter_y, y_step_q4, w,
749                                     h, bd);
750   } else {
751     const uint16_t *src = CONVERT_TO_SHORTPTR(src8);
752     uint16_t *dst = CONVERT_TO_SHORTPTR(dst8);
753     const int16x8_t filters = vld1q_s16(filter_y);
754     const uint16x8_t max = vdupq_n_u16((1 << bd) - 1);
755 
756     assert(!((intptr_t)dst & 3));
757     assert(!(dst_stride & 3));
758 
759     src -= 3 * src_stride;
760 
761     if (w == 4) {
762       int16x4_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
763       int32x4_t d0, d1, d2, d3;
764       uint16x8_t d01, d23, t01, t23;
765 
766       s0 = vreinterpret_s16_u16(vld1_u16(src));
767       src += src_stride;
768       s1 = vreinterpret_s16_u16(vld1_u16(src));
769       src += src_stride;
770       s2 = vreinterpret_s16_u16(vld1_u16(src));
771       src += src_stride;
772       s3 = vreinterpret_s16_u16(vld1_u16(src));
773       src += src_stride;
774       s4 = vreinterpret_s16_u16(vld1_u16(src));
775       src += src_stride;
776       s5 = vreinterpret_s16_u16(vld1_u16(src));
777       src += src_stride;
778       s6 = vreinterpret_s16_u16(vld1_u16(src));
779       src += src_stride;
780 
781       do {
782         s7 = vreinterpret_s16_u16(vld1_u16(src));
783         src += src_stride;
784         s8 = vreinterpret_s16_u16(vld1_u16(src));
785         src += src_stride;
786         s9 = vreinterpret_s16_u16(vld1_u16(src));
787         src += src_stride;
788         s10 = vreinterpret_s16_u16(vld1_u16(src));
789         src += src_stride;
790 
791         __builtin_prefetch(dst + 0 * dst_stride);
792         __builtin_prefetch(dst + 1 * dst_stride);
793         __builtin_prefetch(dst + 2 * dst_stride);
794         __builtin_prefetch(dst + 3 * dst_stride);
795         __builtin_prefetch(src + 0 * src_stride);
796         __builtin_prefetch(src + 1 * src_stride);
797         __builtin_prefetch(src + 2 * src_stride);
798         __builtin_prefetch(src + 3 * src_stride);
799         d0 = convolve8_4(s0, s1, s2, s3, s4, s5, s6, s7, filters);
800         d1 = convolve8_4(s1, s2, s3, s4, s5, s6, s7, s8, filters);
801         d2 = convolve8_4(s2, s3, s4, s5, s6, s7, s8, s9, filters);
802         d3 = convolve8_4(s3, s4, s5, s6, s7, s8, s9, s10, filters);
803 
804         t01 = vcombine_u16(vqrshrun_n_s32(d0, 7), vqrshrun_n_s32(d1, 7));
805         t23 = vcombine_u16(vqrshrun_n_s32(d2, 7), vqrshrun_n_s32(d3, 7));
806         t01 = vminq_u16(t01, max);
807         t23 = vminq_u16(t23, max);
808 
809         d01 = vcombine_u16(vld1_u16(dst + 0 * dst_stride),
810                            vld1_u16(dst + 1 * dst_stride));
811         d23 = vcombine_u16(vld1_u16(dst + 2 * dst_stride),
812                            vld1_u16(dst + 3 * dst_stride));
813         d01 = vrhaddq_u16(d01, t01);
814         d23 = vrhaddq_u16(d23, t23);
815 
816         vst1_u16(dst, vget_low_u16(d01));
817         dst += dst_stride;
818         vst1_u16(dst, vget_high_u16(d01));
819         dst += dst_stride;
820         vst1_u16(dst, vget_low_u16(d23));
821         dst += dst_stride;
822         vst1_u16(dst, vget_high_u16(d23));
823         dst += dst_stride;
824 
825         s0 = s4;
826         s1 = s5;
827         s2 = s6;
828         s3 = s7;
829         s4 = s8;
830         s5 = s9;
831         s6 = s10;
832         h -= 4;
833       } while (h > 0);
834     } else {
835       int height;
836       const uint16_t *s;
837       uint16_t *d;
838       int16x8_t s0, s1, s2, s3, s4, s5, s6, s7, s8, s9, s10;
839       uint16x8_t d0, d1, d2, d3, t0, t1, t2, t3;
840 
841       do {
842         __builtin_prefetch(src + 0 * src_stride);
843         __builtin_prefetch(src + 1 * src_stride);
844         __builtin_prefetch(src + 2 * src_stride);
845         __builtin_prefetch(src + 3 * src_stride);
846         __builtin_prefetch(src + 4 * src_stride);
847         __builtin_prefetch(src + 5 * src_stride);
848         __builtin_prefetch(src + 6 * src_stride);
849         s = src;
850         s0 = vreinterpretq_s16_u16(vld1q_u16(s));
851         s += src_stride;
852         s1 = vreinterpretq_s16_u16(vld1q_u16(s));
853         s += src_stride;
854         s2 = vreinterpretq_s16_u16(vld1q_u16(s));
855         s += src_stride;
856         s3 = vreinterpretq_s16_u16(vld1q_u16(s));
857         s += src_stride;
858         s4 = vreinterpretq_s16_u16(vld1q_u16(s));
859         s += src_stride;
860         s5 = vreinterpretq_s16_u16(vld1q_u16(s));
861         s += src_stride;
862         s6 = vreinterpretq_s16_u16(vld1q_u16(s));
863         s += src_stride;
864         d = dst;
865         height = h;
866 
867         do {
868           s7 = vreinterpretq_s16_u16(vld1q_u16(s));
869           s += src_stride;
870           s8 = vreinterpretq_s16_u16(vld1q_u16(s));
871           s += src_stride;
872           s9 = vreinterpretq_s16_u16(vld1q_u16(s));
873           s += src_stride;
874           s10 = vreinterpretq_s16_u16(vld1q_u16(s));
875           s += src_stride;
876 
877           __builtin_prefetch(d + 0 * dst_stride);
878           __builtin_prefetch(d + 1 * dst_stride);
879           __builtin_prefetch(d + 2 * dst_stride);
880           __builtin_prefetch(d + 3 * dst_stride);
881           __builtin_prefetch(s + 0 * src_stride);
882           __builtin_prefetch(s + 1 * src_stride);
883           __builtin_prefetch(s + 2 * src_stride);
884           __builtin_prefetch(s + 3 * src_stride);
885           t0 = convolve8_8(s0, s1, s2, s3, s4, s5, s6, s7, filters, max);
886           t1 = convolve8_8(s1, s2, s3, s4, s5, s6, s7, s8, filters, max);
887           t2 = convolve8_8(s2, s3, s4, s5, s6, s7, s8, s9, filters, max);
888           t3 = convolve8_8(s3, s4, s5, s6, s7, s8, s9, s10, filters, max);
889 
890           d0 = vld1q_u16(d + 0 * dst_stride);
891           d1 = vld1q_u16(d + 1 * dst_stride);
892           d2 = vld1q_u16(d + 2 * dst_stride);
893           d3 = vld1q_u16(d + 3 * dst_stride);
894           d0 = vrhaddq_u16(d0, t0);
895           d1 = vrhaddq_u16(d1, t1);
896           d2 = vrhaddq_u16(d2, t2);
897           d3 = vrhaddq_u16(d3, t3);
898 
899           vst1q_u16(d, d0);
900           d += dst_stride;
901           vst1q_u16(d, d1);
902           d += dst_stride;
903           vst1q_u16(d, d2);
904           d += dst_stride;
905           vst1q_u16(d, d3);
906           d += dst_stride;
907 
908           s0 = s4;
909           s1 = s5;
910           s2 = s6;
911           s3 = s7;
912           s4 = s8;
913           s5 = s9;
914           s6 = s10;
915           height -= 4;
916         } while (height > 0);
917         src += 8;
918         dst += 8;
919         w -= 8;
920       } while (w > 0);
921     }
922   }
923 }
924