avg_neon.c revision 68e1c830ade592be74773e249bf94e2bbfb50de7 
1/*
2 *  Copyright (c) 2015 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_dsp_rtcd.h"
15#include "./vpx_config.h"
16
17#include "vpx/vpx_integer.h"
18
19static INLINE unsigned int horizontal_add_u16x8(const uint16x8_t v_16x8) {
20  const uint32x4_t a = vpaddlq_u16(v_16x8);
21  const uint64x2_t b = vpaddlq_u32(a);
22  const uint32x2_t c = vadd_u32(vreinterpret_u32_u64(vget_low_u64(b)),
23                                vreinterpret_u32_u64(vget_high_u64(b)));
24  return vget_lane_u32(c, 0);
25}
26
27unsigned int vpx_avg_4x4_neon(const uint8_t *s, int p) {
28  uint16x8_t v_sum;
29  uint32x2_t v_s0 = vdup_n_u32(0);
30  uint32x2_t v_s1 = vdup_n_u32(0);
31  v_s0 = vld1_lane_u32((const uint32_t *)s, v_s0, 0);
32  v_s0 = vld1_lane_u32((const uint32_t *)(s + p), v_s0, 1);
33  v_s1 = vld1_lane_u32((const uint32_t *)(s + 2 * p), v_s1, 0);
34  v_s1 = vld1_lane_u32((const uint32_t *)(s + 3 * p), v_s1, 1);
35  v_sum = vaddl_u8(vreinterpret_u8_u32(v_s0), vreinterpret_u8_u32(v_s1));
36  return (horizontal_add_u16x8(v_sum) + 8) >> 4;
37}
38
39unsigned int vpx_avg_8x8_neon(const uint8_t *s, int p) {
40  uint8x8_t v_s0 = vld1_u8(s);
41  const uint8x8_t v_s1 = vld1_u8(s + p);
42  uint16x8_t v_sum = vaddl_u8(v_s0, v_s1);
43
44  v_s0 = vld1_u8(s + 2 * p);
45  v_sum = vaddw_u8(v_sum, v_s0);
46
47  v_s0 = vld1_u8(s + 3 * p);
48  v_sum = vaddw_u8(v_sum, v_s0);
49
50  v_s0 = vld1_u8(s + 4 * p);
51  v_sum = vaddw_u8(v_sum, v_s0);
52
53  v_s0 = vld1_u8(s + 5 * p);
54  v_sum = vaddw_u8(v_sum, v_s0);
55
56  v_s0 = vld1_u8(s + 6 * p);
57  v_sum = vaddw_u8(v_sum, v_s0);
58
59  v_s0 = vld1_u8(s + 7 * p);
60  v_sum = vaddw_u8(v_sum, v_s0);
61
62  return (horizontal_add_u16x8(v_sum) + 32) >> 6;
63}
64
65// coeff: 16 bits, dynamic range [-32640, 32640].
66// length: value range {16, 64, 256, 1024}.
67int vpx_satd_neon(const int16_t *coeff, int length) {
68  const int16x4_t zero = vdup_n_s16(0);
69  int32x4_t accum = vdupq_n_s32(0);
70
71  do {
72    const int16x8_t src0 = vld1q_s16(coeff);
73    const int16x8_t src8 = vld1q_s16(coeff + 8);
74    accum = vabal_s16(accum, vget_low_s16(src0), zero);
75    accum = vabal_s16(accum, vget_high_s16(src0), zero);
76    accum = vabal_s16(accum, vget_low_s16(src8), zero);
77    accum = vabal_s16(accum, vget_high_s16(src8), zero);
78    length -= 16;
79    coeff += 16;
80  } while (length != 0);
81
82  {
83    // satd: 26 bits, dynamic range [-32640 * 1024, 32640 * 1024]
84    const int64x2_t s0 = vpaddlq_s32(accum);  // cascading summation of 'accum'.
85    const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
86                                  vreinterpret_s32_s64(vget_high_s64(s0)));
87    const int satd = vget_lane_s32(s1, 0);
88    return satd;
89  }
90}
91
92void vpx_int_pro_row_neon(int16_t hbuf[16], uint8_t const *ref,
93                          const int ref_stride, const int height) {
94  int i;
95  uint16x8_t vec_sum_lo = vdupq_n_u16(0);
96  uint16x8_t vec_sum_hi = vdupq_n_u16(0);
97  const int shift_factor = ((height >> 5) + 3) * -1;
98  const int16x8_t vec_shift = vdupq_n_s16(shift_factor);
99
100  for (i = 0; i < height; i += 8) {
101    const uint8x16_t vec_row1 = vld1q_u8(ref);
102    const uint8x16_t vec_row2 = vld1q_u8(ref + ref_stride);
103    const uint8x16_t vec_row3 = vld1q_u8(ref + ref_stride * 2);
104    const uint8x16_t vec_row4 = vld1q_u8(ref + ref_stride * 3);
105    const uint8x16_t vec_row5 = vld1q_u8(ref + ref_stride * 4);
106    const uint8x16_t vec_row6 = vld1q_u8(ref + ref_stride * 5);
107    const uint8x16_t vec_row7 = vld1q_u8(ref + ref_stride * 6);
108    const uint8x16_t vec_row8 = vld1q_u8(ref + ref_stride * 7);
109
110    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row1));
111    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row1));
112
113    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row2));
114    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row2));
115
116    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row3));
117    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row3));
118
119    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row4));
120    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row4));
121
122    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row5));
123    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row5));
124
125    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row6));
126    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row6));
127
128    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row7));
129    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row7));
130
131    vec_sum_lo = vaddw_u8(vec_sum_lo, vget_low_u8(vec_row8));
132    vec_sum_hi = vaddw_u8(vec_sum_hi, vget_high_u8(vec_row8));
133
134    ref += ref_stride * 8;
135  }
136
137  vec_sum_lo = vshlq_u16(vec_sum_lo, vec_shift);
138  vec_sum_hi = vshlq_u16(vec_sum_hi, vec_shift);
139
140  vst1q_s16(hbuf, vreinterpretq_s16_u16(vec_sum_lo));
141  hbuf += 8;
142  vst1q_s16(hbuf, vreinterpretq_s16_u16(vec_sum_hi));
143}
144
145int16_t vpx_int_pro_col_neon(uint8_t const *ref, const int width) {
146  int i;
147  uint16x8_t vec_sum = vdupq_n_u16(0);
148
149  for (i = 0; i < width; i += 16) {
150    const uint8x16_t vec_row = vld1q_u8(ref);
151    vec_sum = vaddw_u8(vec_sum, vget_low_u8(vec_row));
152    vec_sum = vaddw_u8(vec_sum, vget_high_u8(vec_row));
153    ref += 16;
154  }
155
156  return horizontal_add_u16x8(vec_sum);
157}
158
159// ref, src = [0, 510] - max diff = 16-bits
160// bwl = {2, 3, 4}, width = {16, 32, 64}
161int vpx_vector_var_neon(int16_t const *ref, int16_t const *src, const int bwl) {
162  int width = 4 << bwl;
163  int32x4_t sse = vdupq_n_s32(0);
164  int16x8_t total = vdupq_n_s16(0);
165
166  assert(width >= 8);
167  assert((width % 8) == 0);
168
169  do {
170    const int16x8_t r = vld1q_s16(ref);
171    const int16x8_t s = vld1q_s16(src);
172    const int16x8_t diff = vsubq_s16(r, s);  // [-510, 510], 10 bits.
173    const int16x4_t diff_lo = vget_low_s16(diff);
174    const int16x4_t diff_hi = vget_high_s16(diff);
175    sse = vmlal_s16(sse, diff_lo, diff_lo);  // dynamic range 26 bits.
176    sse = vmlal_s16(sse, diff_hi, diff_hi);
177    total = vaddq_s16(total, diff);  // dynamic range 16 bits.
178
179    ref += 8;
180    src += 8;
181    width -= 8;
182  } while (width != 0);
183
184  {
185    // Note: 'total''s pairwise addition could be implemented similarly to
186    // horizontal_add_u16x8(), but one less vpaddl with 'total' when paired
187    // with the summation of 'sse' performed better on a Cortex-A15.
188    const int32x4_t t0 = vpaddlq_s16(total);  // cascading summation of 'total'
189    const int32x2_t t1 = vadd_s32(vget_low_s32(t0), vget_high_s32(t0));
190    const int32x2_t t2 = vpadd_s32(t1, t1);
191    const int t = vget_lane_s32(t2, 0);
192    const int64x2_t s0 = vpaddlq_s32(sse);  // cascading summation of 'sse'.
193    const int32x2_t s1 = vadd_s32(vreinterpret_s32_s64(vget_low_s64(s0)),
194                                  vreinterpret_s32_s64(vget_high_s64(s0)));
195    const int s = vget_lane_s32(s1, 0);
196    const int shift_factor = bwl + 2;
197    return s - ((t * t) >> shift_factor);
198  }
199}
200
201void vpx_minmax_8x8_neon(const uint8_t *a, int a_stride,
202                         const uint8_t *b, int b_stride,
203                         int *min, int *max) {
204  // Load and concatenate.
205  const uint8x16_t a01 = vcombine_u8(vld1_u8(a),
206                                     vld1_u8(a + a_stride));
207  const uint8x16_t a23 = vcombine_u8(vld1_u8(a + 2 * a_stride),
208                                     vld1_u8(a + 3 * a_stride));
209  const uint8x16_t a45 = vcombine_u8(vld1_u8(a + 4 * a_stride),
210                                     vld1_u8(a + 5 * a_stride));
211  const uint8x16_t a67 = vcombine_u8(vld1_u8(a + 6 * a_stride),
212                                     vld1_u8(a + 7 * a_stride));
213
214  const uint8x16_t b01 = vcombine_u8(vld1_u8(b),
215                                     vld1_u8(b + b_stride));
216  const uint8x16_t b23 = vcombine_u8(vld1_u8(b + 2 * b_stride),
217                                     vld1_u8(b + 3 * b_stride));
218  const uint8x16_t b45 = vcombine_u8(vld1_u8(b + 4 * b_stride),
219                                     vld1_u8(b + 5 * b_stride));
220  const uint8x16_t b67 = vcombine_u8(vld1_u8(b + 6 * b_stride),
221                                     vld1_u8(b + 7 * b_stride));
222
223  // Absolute difference.
224  const uint8x16_t ab01_diff = vabdq_u8(a01, b01);
225  const uint8x16_t ab23_diff = vabdq_u8(a23, b23);
226  const uint8x16_t ab45_diff = vabdq_u8(a45, b45);
227  const uint8x16_t ab67_diff = vabdq_u8(a67, b67);
228
229  // Max values between the Q vectors.
230  const uint8x16_t ab0123_max = vmaxq_u8(ab01_diff, ab23_diff);
231  const uint8x16_t ab4567_max = vmaxq_u8(ab45_diff, ab67_diff);
232  const uint8x16_t ab0123_min = vminq_u8(ab01_diff, ab23_diff);
233  const uint8x16_t ab4567_min = vminq_u8(ab45_diff, ab67_diff);
234
235  const uint8x16_t ab07_max = vmaxq_u8(ab0123_max, ab4567_max);
236  const uint8x16_t ab07_min = vminq_u8(ab0123_min, ab4567_min);
237
238  // Split to D and start doing pairwise.
239  uint8x8_t ab_max = vmax_u8(vget_high_u8(ab07_max), vget_low_u8(ab07_max));
240  uint8x8_t ab_min = vmin_u8(vget_high_u8(ab07_min), vget_low_u8(ab07_min));
241
242  // Enough runs of vpmax/min propogate the max/min values to every position.
243  ab_max = vpmax_u8(ab_max, ab_max);
244  ab_min = vpmin_u8(ab_min, ab_min);
245
246  ab_max = vpmax_u8(ab_max, ab_max);
247  ab_min = vpmin_u8(ab_min, ab_min);
248
249  ab_max = vpmax_u8(ab_max, ab_max);
250  ab_min = vpmin_u8(ab_min, ab_min);
251
252  *min = *max = 0;  // Clear high bits
253  // Store directly to avoid costly neon->gpr transfer.
254  vst1_lane_u8((uint8_t *)max, ab_max, 0);
255  vst1_lane_u8((uint8_t *)min, ab_min, 0);
256}
257