variance_neon.c 14.7 KB
Newer Older
Johann's avatar
Johann committed
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
/*
 *  Copyright (c) 2014 The WebM project authors. All Rights Reserved.
 *
 *  Use of this source code is governed by a BSD-style license
 *  that can be found in the LICENSE file in the root of the source
 *  tree. An additional intellectual property rights grant can be found
 *  in the file PATENTS.  All contributing project authors may
 *  be found in the AUTHORS file in the root of the source tree.
 */

#include <arm_neon.h>

#include "./vpx_dsp_rtcd.h"
#include "./vpx_config.h"

#include "vpx/vpx_integer.h"
17
#include "vpx_ports/mem.h"
Johann's avatar
Johann committed
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418

static INLINE int horizontal_add_s16x8(const int16x8_t v_16x8) {
  const int32x4_t a = vpaddlq_s16(v_16x8);
  const int64x2_t b = vpaddlq_s32(a);
  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
                               vreinterpret_s32_s64(vget_high_s64(b)));
  return vget_lane_s32(c, 0);
}

static INLINE int horizontal_add_s32x4(const int32x4_t v_32x4) {
  const int64x2_t b = vpaddlq_s32(v_32x4);
  const int32x2_t c = vadd_s32(vreinterpret_s32_s64(vget_low_s64(b)),
                               vreinterpret_s32_s64(vget_high_s64(b)));
  return vget_lane_s32(c, 0);
}

// w * h must be less than 2048 or local variable v_sum may overflow.
static void variance_neon_w8(const uint8_t *a, int a_stride,
                             const uint8_t *b, int b_stride,
                             int w, int h, uint32_t *sse, int *sum) {
  int i, j;
  int16x8_t v_sum = vdupq_n_s16(0);
  int32x4_t v_sse_lo = vdupq_n_s32(0);
  int32x4_t v_sse_hi = vdupq_n_s32(0);

  for (i = 0; i < h; ++i) {
    for (j = 0; j < w; j += 8) {
      const uint8x8_t v_a = vld1_u8(&a[j]);
      const uint8x8_t v_b = vld1_u8(&b[j]);
      const uint16x8_t v_diff = vsubl_u8(v_a, v_b);
      const int16x8_t sv_diff = vreinterpretq_s16_u16(v_diff);
      v_sum = vaddq_s16(v_sum, sv_diff);
      v_sse_lo = vmlal_s16(v_sse_lo,
                           vget_low_s16(sv_diff),
                           vget_low_s16(sv_diff));
      v_sse_hi = vmlal_s16(v_sse_hi,
                           vget_high_s16(sv_diff),
                           vget_high_s16(sv_diff));
    }
    a += a_stride;
    b += b_stride;
  }

  *sum = horizontal_add_s16x8(v_sum);
  *sse = (unsigned int)horizontal_add_s32x4(vaddq_s32(v_sse_lo, v_sse_hi));
}

void vpx_get8x8var_neon(const uint8_t *a, int a_stride,
                        const uint8_t *b, int b_stride,
                        unsigned int *sse, int *sum) {
  variance_neon_w8(a, a_stride, b, b_stride, 8, 8, sse, sum);
}

void vpx_get16x16var_neon(const uint8_t *a, int a_stride,
                          const uint8_t *b, int b_stride,
                          unsigned int *sse, int *sum) {
  variance_neon_w8(a, a_stride, b, b_stride, 16, 16, sse, sum);
}

unsigned int vpx_variance8x8_neon(const uint8_t *a, int a_stride,
                                  const uint8_t *b, int b_stride,
                                  unsigned int *sse) {
  int sum;
  variance_neon_w8(a, a_stride, b, b_stride, 8, 8, sse, &sum);
  return *sse - (((int64_t)sum * sum) >> 6);  //  >> 6 = / 8 * 8
}

unsigned int vpx_variance16x16_neon(const uint8_t *a, int a_stride,
                                    const uint8_t *b, int b_stride,
                                    unsigned int *sse) {
  int sum;
  variance_neon_w8(a, a_stride, b, b_stride, 16, 16, sse, &sum);
  return *sse - (((int64_t)sum * sum) >> 8);  //  >> 8 = / 16 * 16
}

unsigned int vpx_variance32x32_neon(const uint8_t *a, int a_stride,
                                    const uint8_t *b, int b_stride,
                                    unsigned int *sse) {
  int sum;
  variance_neon_w8(a, a_stride, b, b_stride, 32, 32, sse, &sum);
  return *sse - (((int64_t)sum * sum) >> 10);  // >> 10 = / 32 * 32
}

unsigned int vpx_variance32x64_neon(const uint8_t *a, int a_stride,
                                    const uint8_t *b, int b_stride,
                                    unsigned int *sse) {
  int sum1, sum2;
  uint32_t sse1, sse2;
  variance_neon_w8(a, a_stride, b, b_stride, 32, 32, &sse1, &sum1);
  variance_neon_w8(a + (32 * a_stride), a_stride,
                   b + (32 * b_stride), b_stride, 32, 32,
                   &sse2, &sum2);
  *sse = sse1 + sse2;
  sum1 += sum2;
  return *sse - (((int64_t)sum1 * sum1) >> 11);  // >> 11 = / 32 * 64
}

unsigned int vpx_variance64x32_neon(const uint8_t *a, int a_stride,
                                    const uint8_t *b, int b_stride,
                                    unsigned int *sse) {
  int sum1, sum2;
  uint32_t sse1, sse2;
  variance_neon_w8(a, a_stride, b, b_stride, 64, 16, &sse1, &sum1);
  variance_neon_w8(a + (16 * a_stride), a_stride,
                   b + (16 * b_stride), b_stride, 64, 16,
                   &sse2, &sum2);
  *sse = sse1 + sse2;
  sum1 += sum2;
  return *sse - (((int64_t)sum1 * sum1) >> 11);  // >> 11 = / 32 * 64
}

unsigned int vpx_variance64x64_neon(const uint8_t *a, int a_stride,
                                    const uint8_t *b, int b_stride,
                                    unsigned int *sse) {
  int sum1, sum2;
  uint32_t sse1, sse2;

  variance_neon_w8(a, a_stride, b, b_stride, 64, 16, &sse1, &sum1);
  variance_neon_w8(a + (16 * a_stride), a_stride,
                   b + (16 * b_stride), b_stride, 64, 16,
                   &sse2, &sum2);
  sse1 += sse2;
  sum1 += sum2;

  variance_neon_w8(a + (16 * 2 * a_stride), a_stride,
                   b + (16 * 2 * b_stride), b_stride,
                   64, 16, &sse2, &sum2);
  sse1 += sse2;
  sum1 += sum2;

  variance_neon_w8(a + (16 * 3 * a_stride), a_stride,
                   b + (16 * 3 * b_stride), b_stride,
                   64, 16, &sse2, &sum2);
  *sse = sse1 + sse2;
  sum1 += sum2;
  return *sse - (((int64_t)sum1 * sum1) >> 12);  // >> 12 = / 64 * 64
}

unsigned int vpx_variance16x8_neon(
        const unsigned char *src_ptr,
        int source_stride,
        const unsigned char *ref_ptr,
        int recon_stride,
        unsigned int *sse) {
    int i;
    int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
    uint32x2_t d0u32, d10u32;
    int64x1_t d0s64, d1s64;
    uint8x16_t q0u8, q1u8, q2u8, q3u8;
    uint16x8_t q11u16, q12u16, q13u16, q14u16;
    int32x4_t q8s32, q9s32, q10s32;
    int64x2_t q0s64, q1s64, q5s64;

    q8s32 = vdupq_n_s32(0);
    q9s32 = vdupq_n_s32(0);
    q10s32 = vdupq_n_s32(0);

    for (i = 0; i < 4; i++) {
        q0u8 = vld1q_u8(src_ptr);
        src_ptr += source_stride;
        q1u8 = vld1q_u8(src_ptr);
        src_ptr += source_stride;
        __builtin_prefetch(src_ptr);

        q2u8 = vld1q_u8(ref_ptr);
        ref_ptr += recon_stride;
        q3u8 = vld1q_u8(ref_ptr);
        ref_ptr += recon_stride;
        __builtin_prefetch(ref_ptr);

        q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
        q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
        q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
        q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));

        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
        q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
        q10s32 = vmlal_s16(q10s32, d23s16, d23s16);

        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);

        d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
        d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q13u16));
        q9s32 = vmlal_s16(q9s32, d26s16, d26s16);
        q10s32 = vmlal_s16(q10s32, d27s16, d27s16);

        d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
        d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q14u16));
        q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
        q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
    }

    q10s32 = vaddq_s32(q10s32, q9s32);
    q0s64 = vpaddlq_s32(q8s32);
    q1s64 = vpaddlq_s32(q10s32);

    d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
    d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));

    q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
                      vreinterpret_s32_s64(d0s64));
    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);

    d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
    d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);

    return vget_lane_u32(d0u32, 0);
}

unsigned int vpx_variance8x16_neon(
        const unsigned char *src_ptr,
        int source_stride,
        const unsigned char *ref_ptr,
        int recon_stride,
        unsigned int *sse) {
    int i;
    uint8x8_t d0u8, d2u8, d4u8, d6u8;
    int16x4_t d22s16, d23s16, d24s16, d25s16;
    uint32x2_t d0u32, d10u32;
    int64x1_t d0s64, d1s64;
    uint16x8_t q11u16, q12u16;
    int32x4_t q8s32, q9s32, q10s32;
    int64x2_t q0s64, q1s64, q5s64;

    q8s32 = vdupq_n_s32(0);
    q9s32 = vdupq_n_s32(0);
    q10s32 = vdupq_n_s32(0);

    for (i = 0; i < 8; i++) {
        d0u8 = vld1_u8(src_ptr);
        src_ptr += source_stride;
        d2u8 = vld1_u8(src_ptr);
        src_ptr += source_stride;
        __builtin_prefetch(src_ptr);

        d4u8 = vld1_u8(ref_ptr);
        ref_ptr += recon_stride;
        d6u8 = vld1_u8(ref_ptr);
        ref_ptr += recon_stride;
        __builtin_prefetch(ref_ptr);

        q11u16 = vsubl_u8(d0u8, d4u8);
        q12u16 = vsubl_u8(d2u8, d6u8);

        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q11u16));
        q9s32 = vmlal_s16(q9s32, d22s16, d22s16);
        q10s32 = vmlal_s16(q10s32, d23s16, d23s16);

        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
        q8s32 = vpadalq_s16(q8s32, vreinterpretq_s16_u16(q12u16));
        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);
    }

    q10s32 = vaddq_s32(q10s32, q9s32);
    q0s64 = vpaddlq_s32(q8s32);
    q1s64 = vpaddlq_s32(q10s32);

    d0s64 = vadd_s64(vget_low_s64(q0s64), vget_high_s64(q0s64));
    d1s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));

    q5s64 = vmull_s32(vreinterpret_s32_s64(d0s64),
                      vreinterpret_s32_s64(d0s64));
    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d1s64), 0);

    d10u32 = vshr_n_u32(vreinterpret_u32_s64(vget_low_s64(q5s64)), 7);
    d0u32 = vsub_u32(vreinterpret_u32_s64(d1s64), d10u32);

    return vget_lane_u32(d0u32, 0);
}

unsigned int vpx_mse16x16_neon(
        const unsigned char *src_ptr,
        int source_stride,
        const unsigned char *ref_ptr,
        int recon_stride,
        unsigned int *sse) {
    int i;
    int16x4_t d22s16, d23s16, d24s16, d25s16, d26s16, d27s16, d28s16, d29s16;
    int64x1_t d0s64;
    uint8x16_t q0u8, q1u8, q2u8, q3u8;
    int32x4_t q7s32, q8s32, q9s32, q10s32;
    uint16x8_t q11u16, q12u16, q13u16, q14u16;
    int64x2_t q1s64;

    q7s32 = vdupq_n_s32(0);
    q8s32 = vdupq_n_s32(0);
    q9s32 = vdupq_n_s32(0);
    q10s32 = vdupq_n_s32(0);

    for (i = 0; i < 8; i++) {  // mse16x16_neon_loop
        q0u8 = vld1q_u8(src_ptr);
        src_ptr += source_stride;
        q1u8 = vld1q_u8(src_ptr);
        src_ptr += source_stride;
        q2u8 = vld1q_u8(ref_ptr);
        ref_ptr += recon_stride;
        q3u8 = vld1q_u8(ref_ptr);
        ref_ptr += recon_stride;

        q11u16 = vsubl_u8(vget_low_u8(q0u8), vget_low_u8(q2u8));
        q12u16 = vsubl_u8(vget_high_u8(q0u8), vget_high_u8(q2u8));
        q13u16 = vsubl_u8(vget_low_u8(q1u8), vget_low_u8(q3u8));
        q14u16 = vsubl_u8(vget_high_u8(q1u8), vget_high_u8(q3u8));

        d22s16 = vreinterpret_s16_u16(vget_low_u16(q11u16));
        d23s16 = vreinterpret_s16_u16(vget_high_u16(q11u16));
        q7s32 = vmlal_s16(q7s32, d22s16, d22s16);
        q8s32 = vmlal_s16(q8s32, d23s16, d23s16);

        d24s16 = vreinterpret_s16_u16(vget_low_u16(q12u16));
        d25s16 = vreinterpret_s16_u16(vget_high_u16(q12u16));
        q9s32 = vmlal_s16(q9s32, d24s16, d24s16);
        q10s32 = vmlal_s16(q10s32, d25s16, d25s16);

        d26s16 = vreinterpret_s16_u16(vget_low_u16(q13u16));
        d27s16 = vreinterpret_s16_u16(vget_high_u16(q13u16));
        q7s32 = vmlal_s16(q7s32, d26s16, d26s16);
        q8s32 = vmlal_s16(q8s32, d27s16, d27s16);

        d28s16 = vreinterpret_s16_u16(vget_low_u16(q14u16));
        d29s16 = vreinterpret_s16_u16(vget_high_u16(q14u16));
        q9s32 = vmlal_s16(q9s32, d28s16, d28s16);
        q10s32 = vmlal_s16(q10s32, d29s16, d29s16);
    }

    q7s32 = vaddq_s32(q7s32, q8s32);
    q9s32 = vaddq_s32(q9s32, q10s32);
    q10s32 = vaddq_s32(q7s32, q9s32);

    q1s64 = vpaddlq_s32(q10s32);
    d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));

    vst1_lane_u32((uint32_t *)sse, vreinterpret_u32_s64(d0s64), 0);
    return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
}

unsigned int vpx_get4x4sse_cs_neon(
        const unsigned char *src_ptr,
        int source_stride,
        const unsigned char *ref_ptr,
        int recon_stride) {
    int16x4_t d22s16, d24s16, d26s16, d28s16;
    int64x1_t d0s64;
    uint8x8_t d0u8, d1u8, d2u8, d3u8, d4u8, d5u8, d6u8, d7u8;
    int32x4_t q7s32, q8s32, q9s32, q10s32;
    uint16x8_t q11u16, q12u16, q13u16, q14u16;
    int64x2_t q1s64;

    d0u8 = vld1_u8(src_ptr);
    src_ptr += source_stride;
    d4u8 = vld1_u8(ref_ptr);
    ref_ptr += recon_stride;
    d1u8 = vld1_u8(src_ptr);
    src_ptr += source_stride;
    d5u8 = vld1_u8(ref_ptr);
    ref_ptr += recon_stride;
    d2u8 = vld1_u8(src_ptr);
    src_ptr += source_stride;
    d6u8 = vld1_u8(ref_ptr);
    ref_ptr += recon_stride;
    d3u8 = vld1_u8(src_ptr);
    src_ptr += source_stride;
    d7u8 = vld1_u8(ref_ptr);
    ref_ptr += recon_stride;

    q11u16 = vsubl_u8(d0u8, d4u8);
    q12u16 = vsubl_u8(d1u8, d5u8);
    q13u16 = vsubl_u8(d2u8, d6u8);
    q14u16 = vsubl_u8(d3u8, d7u8);

    d22s16 = vget_low_s16(vreinterpretq_s16_u16(q11u16));
    d24s16 = vget_low_s16(vreinterpretq_s16_u16(q12u16));
    d26s16 = vget_low_s16(vreinterpretq_s16_u16(q13u16));
    d28s16 = vget_low_s16(vreinterpretq_s16_u16(q14u16));

    q7s32 = vmull_s16(d22s16, d22s16);
    q8s32 = vmull_s16(d24s16, d24s16);
    q9s32 = vmull_s16(d26s16, d26s16);
    q10s32 = vmull_s16(d28s16, d28s16);

    q7s32 = vaddq_s32(q7s32, q8s32);
    q9s32 = vaddq_s32(q9s32, q10s32);
    q9s32 = vaddq_s32(q7s32, q9s32);

    q1s64 = vpaddlq_s32(q9s32);
    d0s64 = vadd_s64(vget_low_s64(q1s64), vget_high_s64(q1s64));

    return vget_lane_u32(vreinterpret_u32_s64(d0s64), 0);
}