xref: /aosp_15_r20/external/libvpx/vp9/encoder/arm/neon/vp9_error_neon.c (revision fb1b10ab9aebc7c7068eedab379b749d7e3900be)
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 "./vp9_rtcd.h"
15 #include "vpx_dsp/arm/mem_neon.h"
16 #include "vpx_dsp/arm/sum_neon.h"
17 
vp9_block_error_neon(const tran_low_t * coeff,const tran_low_t * dqcoeff,intptr_t block_size,int64_t * ssz)18 int64_t vp9_block_error_neon(const tran_low_t *coeff, const tran_low_t *dqcoeff,
19                              intptr_t block_size, int64_t *ssz) {
20   uint64x2_t err_u64 = vdupq_n_u64(0);
21   int64x2_t ssz_s64 = vdupq_n_s64(0);
22 
23   assert(block_size >= 16);
24   assert((block_size % 16) == 0);
25 
26   do {
27     uint32x4_t err;
28     int32x4_t ssz0, ssz1;
29 
30     const int16x8_t c0 = load_tran_low_to_s16q(coeff);
31     const int16x8_t c1 = load_tran_low_to_s16q(coeff + 8);
32     const int16x8_t d0 = load_tran_low_to_s16q(dqcoeff);
33     const int16x8_t d1 = load_tran_low_to_s16q(dqcoeff + 8);
34 
35     const uint16x8_t diff0 = vreinterpretq_u16_s16(vabdq_s16(c0, d0));
36     const uint16x8_t diff1 = vreinterpretq_u16_s16(vabdq_s16(c1, d1));
37 
38     // diff is 15-bits, the squares 30, so we can store 4 in 32-bits before
39     // accumulating them in 64-bits.
40     err = vmull_u16(vget_low_u16(diff0), vget_low_u16(diff0));
41     err = vmlal_u16(err, vget_high_u16(diff0), vget_high_u16(diff0));
42     err = vmlal_u16(err, vget_low_u16(diff1), vget_low_u16(diff1));
43     err = vmlal_u16(err, vget_high_u16(diff1), vget_high_u16(diff1));
44     err_u64 = vpadalq_u32(err_u64, err);
45 
46     // We can't do the same here as we're operating on signed integers, so we
47     // can store 2 15-bit diff before accumulating into 64-bits.
48     ssz0 = vmull_s16(vget_low_s16(c0), vget_low_s16(c0));
49     ssz0 = vmlal_s16(ssz0, vget_high_s16(c0), vget_high_s16(c0));
50     ssz_s64 = vpadalq_s32(ssz_s64, ssz0);
51 
52     ssz1 = vmull_s16(vget_low_s16(c1), vget_low_s16(c1));
53     ssz1 = vmlal_s16(ssz1, vget_high_s16(c1), vget_high_s16(c1));
54     ssz_s64 = vpadalq_s32(ssz_s64, ssz1);
55 
56     coeff += 16;
57     dqcoeff += 16;
58     block_size -= 16;
59   } while (block_size != 0);
60 
61   *ssz = horizontal_add_int64x2(ssz_s64);
62   return (int64_t)horizontal_add_uint64x2(err_u64);
63 }
64 
vp9_block_error_fp_neon(const tran_low_t * coeff,const tran_low_t * dqcoeff,int block_size)65 int64_t vp9_block_error_fp_neon(const tran_low_t *coeff,
66                                 const tran_low_t *dqcoeff, int block_size) {
67   uint64x2_t err_u64[2] = { vdupq_n_u64(0), vdupq_n_u64(0) };
68 
69   assert(block_size >= 16);
70   assert((block_size % 16) == 0);
71 
72   do {
73     uint32x4_t err0, err1;
74 
75     const int16x8_t c0 = load_tran_low_to_s16q(coeff);
76     const int16x8_t c1 = load_tran_low_to_s16q(coeff + 8);
77     const int16x8_t d0 = load_tran_low_to_s16q(dqcoeff);
78     const int16x8_t d1 = load_tran_low_to_s16q(dqcoeff + 8);
79 
80     const uint16x8_t diff0 = vreinterpretq_u16_s16(vabdq_s16(c0, d0));
81     const uint16x8_t diff1 = vreinterpretq_u16_s16(vabdq_s16(c1, d1));
82 
83     // diff is 15-bits, the squares 30, so in theory we can store 4 in 32-bits
84     // before accumulating them in 64-bits. However splitting into 2 mull, mlal
85     // pairs is beneficial since it allows us to use both Neon
86     // multiply-accumulate pipes - on CPUs that have them - rather than having
87     // a single chain of 4 instructions executing serially.
88     err0 = vmull_u16(vget_low_u16(diff0), vget_low_u16(diff0));
89     err0 = vmlal_u16(err0, vget_high_u16(diff0), vget_high_u16(diff0));
90     err_u64[0] = vpadalq_u32(err_u64[0], err0);
91 
92     err1 = vmull_u16(vget_low_u16(diff1), vget_low_u16(diff1));
93     err1 = vmlal_u16(err1, vget_high_u16(diff1), vget_high_u16(diff1));
94     err_u64[1] = vpadalq_u32(err_u64[1], err1);
95 
96     coeff += 16;
97     dqcoeff += 16;
98     block_size -= 16;
99   } while (block_size != 0);
100 
101   return horizontal_add_uint64x2(vaddq_u64(err_u64[0], err_u64[1]));
102 }
103