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_dsp_rtcd.h"
14#include "./vpx_config.h"
15
16#include "vpx/vpx_integer.h"
17#include "vpx_ports/mem.h"
18
19static INLINE int horizontal_add_s16x8(const int16x8_t v_16x8) {
20  const int32x4_t a = vpaddlq_s16(v_16x8);
21  const int64x2_t b = vpaddlq_s32(a);
22  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
23                               vreinterpret_s32_s64(vget_high_s64(b)));
24  return vget_lane_s32(c, 0);
25}
26
27static INLINE int horizontal_add_s32x4(const int32x4_t v_32x4) {
28  const int64x2_t b = vpaddlq_s32(v_32x4);
29  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
30                               vreinterpret_s32_s64(vget_high_s64(b)));
31  return vget_lane_s32(c, 0);
32}
33
34// w * h must be less than 2048 or local variable v_sum may overflow.
35static void variance_neon_w8(const uint8_t *a, int a_stride,
36                             const uint8_t *b, int b_stride,
37                             int w, int h, uint32_t *sse, int *sum) {
38  int i, j;
39  int16x8_t v_sum = vdupq_n_s16(0);
40  int32x4_t v_sse_lo = vdupq_n_s32(0);
41  int32x4_t v_sse_hi = vdupq_n_s32(0);
42
43  for (i = 0; i < h; ++i) {
44    for (j = 0; j < w; j += 8) {
45      const uint8x8_t v_a = vld1_u8(&a[j]);
46      const uint8x8_t v_b = vld1_u8(&b[j]);
47      const uint16x8_t v_diff = vsubl_u8(v_a, v_b);
48      const int16x8_t sv_diff = vreinterpretq_s16_u16(v_diff);
49      v_sum = vaddq_s16(v_sum, sv_diff);
50      v_sse_lo = vmlal_s16(v_sse_lo,
51                           vget_low_s16(sv_diff),
52                           vget_low_s16(sv_diff));
53      v_sse_hi = vmlal_s16(v_sse_hi,
54                           vget_high_s16(sv_diff),
55                           vget_high_s16(sv_diff));
56    }
57    a += a_stride;
58    b += b_stride;
59  }
60
61  *sum = horizontal_add_s16x8(v_sum);
62  *sse = (unsigned int)horizontal_add_s32x4(vaddq_s32(v_sse_lo, v_sse_hi));
63}
64
65void vpx_get8x8var_neon(const uint8_t *a, int a_stride,
66                        const uint8_t *b, int b_stride,
67                        unsigned int *sse, int *sum) {
68  variance_neon_w8(a, a_stride, b, b_stride, 8, 8, sse, sum);
69}
70
71void vpx_get16x16var_neon(const uint8_t *a, int a_stride,
72                          const uint8_t *b, int b_stride,
73                          unsigned int *sse, int *sum) {
74  variance_neon_w8(a, a_stride, b, b_stride, 16, 16, sse, sum);
75}
76
77unsigned int vpx_variance8x8_neon(const uint8_t *a, int a_stride,
78                                  const uint8_t *b, int b_stride,
79                                  unsigned int *sse) {
80  int sum;
81  variance_neon_w8(a, a_stride, b, b_stride, 8, 8, sse, &sum);
82  return *sse - (((int64_t)sum * sum) >> 6);  //  >> 6 = / 8 * 8
83}
84
85unsigned int vpx_variance16x16_neon(const uint8_t *a, int a_stride,
86                                    const uint8_t *b, int b_stride,
87                                    unsigned int *sse) {
88  int sum;
89  variance_neon_w8(a, a_stride, b, b_stride, 16, 16, sse, &sum);
90  return *sse - (((int64_t)sum * sum) >> 8);  //  >> 8 = / 16 * 16
91}
92
93unsigned int vpx_variance32x32_neon(const uint8_t *a, int a_stride,
94                                    const uint8_t *b, int b_stride,
95                                    unsigned int *sse) {
96  int sum;
97  variance_neon_w8(a, a_stride, b, b_stride, 32, 32, sse, &sum);
98  return *sse - (((int64_t)sum * sum) >> 10);  // >> 10 = / 32 * 32
99}
100
101unsigned int vpx_variance32x64_neon(const uint8_t *a, int a_stride,
102                                    const uint8_t *b, int b_stride,
103                                    unsigned int *sse) {
104  int sum1, sum2;
105  uint32_t sse1, sse2;
106  variance_neon_w8(a, a_stride, b, b_stride, 32, 32, &sse1, &sum1);
107  variance_neon_w8(a + (32 * a_stride), a_stride,
108                   b + (32 * b_stride), b_stride, 32, 32,
109                   &sse2, &sum2);
110  *sse = sse1 + sse2;
111  sum1 += sum2;
112  return *sse - (((int64_t)sum1 * sum1) >> 11);  // >> 11 = / 32 * 64
113}
114
115unsigned int vpx_variance64x32_neon(const uint8_t *a, int a_stride,
116                                    const uint8_t *b, int b_stride,
117                                    unsigned int *sse) {
118  int sum1, sum2;
119  uint32_t sse1, sse2;
120  variance_neon_w8(a, a_stride, b, b_stride, 64, 16, &sse1, &sum1);
121  variance_neon_w8(a + (16 * a_stride), a_stride,
122                   b + (16 * b_stride), b_stride, 64, 16,
123                   &sse2, &sum2);
124  *sse = sse1 + sse2;
125  sum1 += sum2;
126  return *sse - (((int64_t)sum1 * sum1) >> 11);  // >> 11 = / 32 * 64
127}
128
129unsigned int vpx_variance64x64_neon(const uint8_t *a, int a_stride,
130                                    const uint8_t *b, int b_stride,
131                                    unsigned int *sse) {
132  int sum1, sum2;
133  uint32_t sse1, sse2;
134
135  variance_neon_w8(a, a_stride, b, b_stride, 64, 16, &sse1, &sum1);
136  variance_neon_w8(a + (16 * a_stride), a_stride,
137                   b + (16 * b_stride), b_stride, 64, 16,
138                   &sse2, &sum2);
139  sse1 += sse2;
140  sum1 += sum2;
141
142  variance_neon_w8(a + (16 * 2 * a_stride), a_stride,
143                   b + (16 * 2 * b_stride), b_stride,
144                   64, 16, &sse2, &sum2);
145  sse1 += sse2;
146  sum1 += sum2;
147
148  variance_neon_w8(a + (16 * 3 * a_stride), a_stride,
149                   b + (16 * 3 * b_stride), b_stride,
150                   64, 16, &sse2, &sum2);
151  *sse = sse1 + sse2;
152  sum1 += sum2;
153  return *sse - (((int64_t)sum1 * sum1) >> 12);  // >> 12 = / 64 * 64
154}
155
156unsigned int vpx_variance16x8_neon(
157        const unsigned char *src_ptr,
158        int source_stride,
159        const unsigned char *ref_ptr,
160        int recon_stride,
161        unsigned int *sse) {
162    int i;
163    int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
164    uint32x2_t d0u32, d10u32;
165    int64x1_t d0s64, d1s64;
166    uint8x16_t q0u8, q1u8, q2u8, q3u8;
167    uint16x8_t q11u16, q12u16, q13u16, q14u16;
168    int32x4_t q8s32, q9s32, q10s32;
169    int64x2_t q0s64, q1s64, q5s64;
170
171    q8s32 = vdupq_n_s32(0);
172    q9s32 = vdupq_n_s32(0);
173    q10s32 = vdupq_n_s32(0);
174
175    for (i = 0; i < 4; i++) {
176        q0u8 = vld1q_u8(src_ptr);
177        src_ptr += source_stride;
178        q1u8 = vld1q_u8(src_ptr);
179        src_ptr += source_stride;
180        __builtin_prefetch(src_ptr);
181
182        q2u8 = vld1q_u8(ref_ptr);
183        ref_ptr += recon_stride;
184        q3u8 = vld1q_u8(ref_ptr);
185        ref_ptr += recon_stride;
186        __builtin_prefetch(ref_ptr);
187
188        q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
189        q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
190        q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
191        q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
192
193        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
194        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
195        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
196        q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
197        q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
198
199        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
200        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
201        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
202        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
203        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
204
205        d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
206        d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
207        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
208        q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
209        q10s32 = vmlal_s16(q10s32, d27s16, d27s16);
210
211        d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
212        d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
213        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
214        q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
215        q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
216    }
217
218    q10s32 = vaddq_s32(q10s32, q9s32);
219    q0s64 = vpaddlq_s32(q8s32);
220    q1s64 = vpaddlq_s32(q10s32);
221
222    d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
223    d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
224
225    q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
226                      vreinterpret_s32_s64(d0s64));
227    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
228
229    d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
230    d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
231
232    return vget_lane_u32(d0u32, 0);
233}
234
235unsigned int vpx_variance8x16_neon(
236        const unsigned char *src_ptr,
237        int source_stride,
238        const unsigned char *ref_ptr,
239        int recon_stride,
240        unsigned int *sse) {
241    int i;
242    uint8x8_t d0u8, d2u8, d4u8, d6u8;
243    int16x4_t d22s16, d23s16, d24s16, d25s16;
244    uint32x2_t d0u32, d10u32;
245    int64x1_t d0s64, d1s64;
246    uint16x8_t q11u16, q12u16;
247    int32x4_t q8s32, q9s32, q10s32;
248    int64x2_t q0s64, q1s64, q5s64;
249
250    q8s32 = vdupq_n_s32(0);
251    q9s32 = vdupq_n_s32(0);
252    q10s32 = vdupq_n_s32(0);
253
254    for (i = 0; i < 8; i++) {
255        d0u8 = vld1_u8(src_ptr);
256        src_ptr += source_stride;
257        d2u8 = vld1_u8(src_ptr);
258        src_ptr += source_stride;
259        __builtin_prefetch(src_ptr);
260
261        d4u8 = vld1_u8(ref_ptr);
262        ref_ptr += recon_stride;
263        d6u8 = vld1_u8(ref_ptr);
264        ref_ptr += recon_stride;
265        __builtin_prefetch(ref_ptr);
266
267        q11u16 = vsubl_u8(d0u8, d4u8);
268        q12u16 = vsubl_u8(d2u8, d6u8);
269
270        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
271        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
272        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
273        q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
274        q10s32 = vmlal_s16(q10s32, d23s16, d23s16);
275
276        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
277        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
278        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
279        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
280        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
281    }
282
283    q10s32 = vaddq_s32(q10s32, q9s32);
284    q0s64 = vpaddlq_s32(q8s32);
285    q1s64 = vpaddlq_s32(q10s32);
286
287    d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
288    d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
289
290    q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
291                      vreinterpret_s32_s64(d0s64));
292    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);
293
294    d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
295    d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);
296
297    return vget_lane_u32(d0u32, 0);
298}
299
300unsigned int vpx_mse16x16_neon(
301        const unsigned char *src_ptr,
302        int source_stride,
303        const unsigned char *ref_ptr,
304        int recon_stride,
305        unsigned int *sse) {
306    int i;
307    int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
308    int64x1_t d0s64;
309    uint8x16_t q0u8, q1u8, q2u8, q3u8;
310    int32x4_t q7s32, q8s32, q9s32, q10s32;
311    uint16x8_t q11u16, q12u16, q13u16, q14u16;
312    int64x2_t q1s64;
313
314    q7s32 = vdupq_n_s32(0);
315    q8s32 = vdupq_n_s32(0);
316    q9s32 = vdupq_n_s32(0);
317    q10s32 = vdupq_n_s32(0);
318
319    for (i = 0; i < 8; i++) {  // mse16x16_neon_loop
320        q0u8 = vld1q_u8(src_ptr);
321        src_ptr += source_stride;
322        q1u8 = vld1q_u8(src_ptr);
323        src_ptr += source_stride;
324        q2u8 = vld1q_u8(ref_ptr);
325        ref_ptr += recon_stride;
326        q3u8 = vld1q_u8(ref_ptr);
327        ref_ptr += recon_stride;
328
329        q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
330        q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
331        q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
332        q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));
333
334        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
335        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
336        q7s32 = vmlal_s16(q7s32, d22s16, d22s16);
337        q8s32 = vmlal_s16(q8s32, d23s16, d23s16);
338
339        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
340        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
341        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
342        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
343
344        d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
345        d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
346        q7s32 = vmlal_s16(q7s32, d26s16, d26s16);
347        q8s32 = vmlal_s16(q8s32, d27s16, d27s16);
348
349        d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
350        d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
351        q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
352        q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
353    }
354
355    q7s32 = vaddq_s32(q7s32, q8s32);
356    q9s32 = vaddq_s32(q9s32, q10s32);
357    q10s32 = vaddq_s32(q7s32, q9s32);
358
359    q1s64 = vpaddlq_s32(q10s32);
360    d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
361
362    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d0s64), 0);
363    return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
364}
365
366unsigned int vpx_get4x4sse_cs_neon(
367        const unsigned char *src_ptr,
368        int source_stride,
369        const unsigned char *ref_ptr,
370        int recon_stride) {
371    int16x4_t d22s16, d24s16, d26s16, d28s16;
372    int64x1_t d0s64;
373    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
374    int32x4_t q7s32, q8s32, q9s32, q10s32;
375    uint16x8_t q11u16, q12u16, q13u16, q14u16;
376    int64x2_t q1s64;
377
378    d0u8 = vld1_u8(src_ptr);
379    src_ptr += source_stride;
380    d4u8 = vld1_u8(ref_ptr);
381    ref_ptr += recon_stride;
382    d1u8 = vld1_u8(src_ptr);
383    src_ptr += source_stride;
384    d5u8 = vld1_u8(ref_ptr);
385    ref_ptr += recon_stride;
386    d2u8 = vld1_u8(src_ptr);
387    src_ptr += source_stride;
388    d6u8 = vld1_u8(ref_ptr);
389    ref_ptr += recon_stride;
390    d3u8 = vld1_u8(src_ptr);
391    src_ptr += source_stride;
392    d7u8 = vld1_u8(ref_ptr);
393    ref_ptr += recon_stride;
394
395    q11u16 = vsubl_u8(d0u8, d4u8);
396    q12u16 = vsubl_u8(d1u8, d5u8);
397    q13u16 = vsubl_u8(d2u8, d6u8);
398    q14u16 = vsubl_u8(d3u8, d7u8);
399
400    d22s16 = vget_low_s16(vreinterpretq_s16_u16(q11u16));
401    d24s16 = vget_low_s16(vreinterpretq_s16_u16(q12u16));
402    d26s16 = vget_low_s16(vreinterpretq_s16_u16(q13u16));
403    d28s16 = vget_low_s16(vreinterpretq_s16_u16(q14u16));
404
405    q7s32 = vmull_s16(d22s16, d22s16);
406    q8s32 = vmull_s16(d24s16, d24s16);
407    q9s32 = vmull_s16(d26s16, d26s16);
408    q10s32 = vmull_s16(d28s16, d28s16);
409
410    q7s32 = vaddq_s32(q7s32, q8s32);
411    q9s32 = vaddq_s32(q9s32, q10s32);
412    q9s32 = vaddq_s32(q7s32, q9s32);
413
414    q1s64 = vpaddlq_s32(q9s32);
415    d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));
416
417    return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
418}
419