1 /*
2 * Copyright (c) 2023, Alliance for Open Media. All rights reserved.
3 *
4 * This source code is subject to the terms of the BSD 2 Clause License and
5 * the Alliance for Open Media Patent License 1.0. If the BSD 2 Clause License
6 * was not distributed with this source code in the LICENSE file, you can
7 * obtain it at www.aomedia.org/license/software. If the Alliance for Open
8 * Media Patent License 1.0 was not distributed with this source code in the
9 * PATENTS file, you can obtain it at www.aomedia.org/license/patent.
10 */
11
12 #include <arm_neon.h>
13 #include <assert.h>
14
15 #include "config/aom_config.h"
16 #include "config/av1_rtcd.h"
17
18 #include "aom_dsp/aom_dsp_common.h"
19 #include "aom_dsp/arm/aom_neon_sve_bridge.h"
20 #include "aom_dsp/arm/mem_neon.h"
21
av1_block_error_sve(const tran_low_t * coeff,const tran_low_t * dqcoeff,intptr_t block_size,int64_t * ssz)22 int64_t av1_block_error_sve(const tran_low_t *coeff, const tran_low_t *dqcoeff,
23 intptr_t block_size, int64_t *ssz) {
24 int64x2_t error[2] = { vdupq_n_s64(0), vdupq_n_s64(0) };
25 int64x2_t sqcoeff[2] = { vdupq_n_s64(0), vdupq_n_s64(0) };
26
27 assert(block_size >= 16);
28 assert((block_size % 16) == 0);
29
30 do {
31 const int16x8_t c0 = load_tran_low_to_s16q(coeff);
32 const int16x8_t c1 = load_tran_low_to_s16q(coeff + 8);
33 const int16x8_t d0 = load_tran_low_to_s16q(dqcoeff);
34 const int16x8_t d1 = load_tran_low_to_s16q(dqcoeff + 8);
35
36 const int16x8_t diff0 = vsubq_s16(c0, d0);
37 const int16x8_t diff1 = vsubq_s16(c1, d1);
38
39 error[0] = aom_sdotq_s16(error[0], diff0, diff0);
40 error[1] = aom_sdotq_s16(error[1], diff1, diff1);
41 sqcoeff[0] = aom_sdotq_s16(sqcoeff[0], c0, c0);
42 sqcoeff[1] = aom_sdotq_s16(sqcoeff[1], c1, c1);
43
44 coeff += 16;
45 dqcoeff += 16;
46 block_size -= 16;
47 } while (block_size != 0);
48
49 *ssz = vaddvq_s64(vaddq_s64(sqcoeff[0], sqcoeff[1]));
50 return vaddvq_s64(vaddq_s64(error[0], error[1]));
51 }
52
av1_block_error_lp_sve(const int16_t * coeff,const int16_t * dqcoeff,intptr_t block_size)53 int64_t av1_block_error_lp_sve(const int16_t *coeff, const int16_t *dqcoeff,
54 intptr_t block_size) {
55 if (block_size % 32 == 0) {
56 int64x2_t error[4] = { vdupq_n_s64(0), vdupq_n_s64(0), vdupq_n_s64(0),
57 vdupq_n_s64(0) };
58
59 do {
60 const int16x8_t c0 = vld1q_s16(coeff);
61 const int16x8_t c1 = vld1q_s16(coeff + 8);
62 const int16x8_t c2 = vld1q_s16(coeff + 16);
63 const int16x8_t c3 = vld1q_s16(coeff + 24);
64 const int16x8_t d0 = vld1q_s16(dqcoeff);
65 const int16x8_t d1 = vld1q_s16(dqcoeff + 8);
66 const int16x8_t d2 = vld1q_s16(dqcoeff + 16);
67 const int16x8_t d3 = vld1q_s16(dqcoeff + 24);
68
69 const int16x8_t diff0 = vsubq_s16(c0, d0);
70 const int16x8_t diff1 = vsubq_s16(c1, d1);
71 const int16x8_t diff2 = vsubq_s16(c2, d2);
72 const int16x8_t diff3 = vsubq_s16(c3, d3);
73
74 error[0] = aom_sdotq_s16(error[0], diff0, diff0);
75 error[1] = aom_sdotq_s16(error[1], diff1, diff1);
76 error[2] = aom_sdotq_s16(error[2], diff2, diff2);
77 error[3] = aom_sdotq_s16(error[3], diff3, diff3);
78
79 coeff += 32;
80 dqcoeff += 32;
81 block_size -= 32;
82 } while (block_size != 0);
83
84 error[0] = vaddq_s64(error[0], error[1]);
85 error[2] = vaddq_s64(error[2], error[3]);
86 error[0] = vaddq_s64(error[0], error[2]);
87 return vaddvq_s64(error[0]);
88 }
89 assert(block_size == 16);
90
91 int64x2_t error[2] = { vdupq_n_s64(0), vdupq_n_s64(0) };
92
93 do {
94 const int16x8_t c0 = vld1q_s16(coeff);
95 const int16x8_t c1 = vld1q_s16(coeff + 8);
96 const int16x8_t d0 = vld1q_s16(dqcoeff);
97 const int16x8_t d1 = vld1q_s16(dqcoeff + 8);
98
99 const int16x8_t diff0 = vsubq_s16(c0, d0);
100 const int16x8_t diff1 = vsubq_s16(c1, d1);
101
102 error[0] = aom_sdotq_s16(error[0], diff0, diff0);
103 error[1] = aom_sdotq_s16(error[1], diff1, diff1);
104
105 coeff += 16;
106 dqcoeff += 16;
107 block_size -= 16;
108 } while (block_size != 0);
109
110 return vaddvq_s64(vaddq_s64(error[0], error[1]));
111 }
112