1 /*
2 * Copyright (c) 2017, 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 <immintrin.h>
13
14 #include "config/av1_rtcd.h"
15
16 #include "aom/aom_integer.h"
17 #include "aom_dsp/aom_dsp_common.h"
18
init_one_qp(const __m128i * p,__m256i * qp)19 static inline void init_one_qp(const __m128i *p, __m256i *qp) {
20 const __m128i zero = _mm_setzero_si128();
21 const __m128i dc = _mm_unpacklo_epi16(*p, zero);
22 const __m128i ac = _mm_unpackhi_epi16(*p, zero);
23 *qp = _mm256_insertf128_si256(_mm256_castsi128_si256(dc), ac, 1);
24 }
25
update_qp(__m256i * qp)26 static inline void update_qp(__m256i *qp) {
27 qp[0] = _mm256_permute2x128_si256(qp[0], qp[0], 0x11);
28 qp[1] = _mm256_permute2x128_si256(qp[1], qp[1], 0x11);
29 qp[2] = _mm256_permute2x128_si256(qp[2], qp[2], 0x11);
30 }
31
init_qp(const int16_t * round_ptr,const int16_t * quant_ptr,const int16_t * dequant_ptr,int log_scale,__m256i * qp)32 static inline void init_qp(const int16_t *round_ptr, const int16_t *quant_ptr,
33 const int16_t *dequant_ptr, int log_scale,
34 __m256i *qp) {
35 __m128i round = _mm_loadu_si128((const __m128i *)round_ptr);
36 if (log_scale) {
37 const __m128i round_scale = _mm_set1_epi16(1 << (15 - log_scale));
38 round = _mm_mulhrs_epi16(round, round_scale);
39 }
40 const __m128i quant = _mm_loadu_si128((const __m128i *)quant_ptr);
41 const __m128i dequant = _mm_loadu_si128((const __m128i *)dequant_ptr);
42
43 init_one_qp(&round, &qp[0]);
44 init_one_qp(&quant, &qp[1]);
45 init_one_qp(&dequant, &qp[2]);
46 }
47
quantize(const __m256i * qp,__m256i * c,const int16_t * iscan_ptr,int log_scale,tran_low_t * qcoeff,tran_low_t * dqcoeff,__m256i * eob)48 static inline void quantize(const __m256i *qp, __m256i *c,
49 const int16_t *iscan_ptr, int log_scale,
50 tran_low_t *qcoeff, tran_low_t *dqcoeff,
51 __m256i *eob) {
52 const __m256i abs_coeff = _mm256_abs_epi32(*c);
53 __m256i q = _mm256_add_epi32(abs_coeff, qp[0]);
54
55 __m256i q_lo = _mm256_mul_epi32(q, qp[1]);
56 __m256i q_hi = _mm256_srli_epi64(q, 32);
57 const __m256i qp_hi = _mm256_srli_epi64(qp[1], 32);
58 q_hi = _mm256_mul_epi32(q_hi, qp_hi);
59 q_lo = _mm256_srli_epi64(q_lo, 16 - log_scale);
60 q_hi = _mm256_srli_epi64(q_hi, 16 - log_scale);
61 q_hi = _mm256_slli_epi64(q_hi, 32);
62 q = _mm256_or_si256(q_lo, q_hi);
63 const __m256i abs_s = _mm256_slli_epi32(abs_coeff, 1 + log_scale);
64 const __m256i mask = _mm256_cmpgt_epi32(qp[2], abs_s);
65 q = _mm256_andnot_si256(mask, q);
66
67 __m256i dq = _mm256_mullo_epi32(q, qp[2]);
68 dq = _mm256_srai_epi32(dq, log_scale);
69 q = _mm256_sign_epi32(q, *c);
70 dq = _mm256_sign_epi32(dq, *c);
71
72 _mm256_storeu_si256((__m256i *)qcoeff, q);
73 _mm256_storeu_si256((__m256i *)dqcoeff, dq);
74
75 const __m128i isc = _mm_loadu_si128((const __m128i *)iscan_ptr);
76 const __m128i zr = _mm_setzero_si128();
77 const __m128i lo = _mm_unpacklo_epi16(isc, zr);
78 const __m128i hi = _mm_unpackhi_epi16(isc, zr);
79 const __m256i iscan =
80 _mm256_insertf128_si256(_mm256_castsi128_si256(lo), hi, 1);
81
82 const __m256i zero = _mm256_setzero_si256();
83 const __m256i zc = _mm256_cmpeq_epi32(dq, zero);
84 const __m256i nz = _mm256_cmpeq_epi32(zc, zero);
85 __m256i cur_eob = _mm256_sub_epi32(iscan, nz);
86 cur_eob = _mm256_and_si256(cur_eob, nz);
87 *eob = _mm256_max_epi32(cur_eob, *eob);
88 }
89
av1_highbd_quantize_fp_avx2(const tran_low_t * coeff_ptr,intptr_t n_coeffs,const int16_t * zbin_ptr,const int16_t * round_ptr,const int16_t * quant_ptr,const int16_t * quant_shift_ptr,tran_low_t * qcoeff_ptr,tran_low_t * dqcoeff_ptr,const int16_t * dequant_ptr,uint16_t * eob_ptr,const int16_t * scan,const int16_t * iscan,int log_scale)90 void av1_highbd_quantize_fp_avx2(
91 const tran_low_t *coeff_ptr, intptr_t n_coeffs, const int16_t *zbin_ptr,
92 const int16_t *round_ptr, const int16_t *quant_ptr,
93 const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
94 tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
95 const int16_t *scan, const int16_t *iscan, int log_scale) {
96 (void)scan;
97 (void)zbin_ptr;
98 (void)quant_shift_ptr;
99 const unsigned int step = 8;
100 __m256i qp[3], coeff;
101
102 init_qp(round_ptr, quant_ptr, dequant_ptr, log_scale, qp);
103 coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr);
104
105 __m256i eob = _mm256_setzero_si256();
106 quantize(qp, &coeff, iscan, log_scale, qcoeff_ptr, dqcoeff_ptr, &eob);
107
108 coeff_ptr += step;
109 qcoeff_ptr += step;
110 dqcoeff_ptr += step;
111 iscan += step;
112 n_coeffs -= step;
113
114 update_qp(qp);
115 while (n_coeffs > 0) {
116 coeff = _mm256_loadu_si256((const __m256i *)coeff_ptr);
117 quantize(qp, &coeff, iscan, log_scale, qcoeff_ptr, dqcoeff_ptr, &eob);
118
119 coeff_ptr += step;
120 qcoeff_ptr += step;
121 dqcoeff_ptr += step;
122 iscan += step;
123 n_coeffs -= step;
124 }
125 {
126 __m256i eob_s;
127 eob_s = _mm256_shuffle_epi32(eob, 0xe);
128 eob = _mm256_max_epi16(eob, eob_s);
129 eob_s = _mm256_shufflelo_epi16(eob, 0xe);
130 eob = _mm256_max_epi16(eob, eob_s);
131 eob_s = _mm256_shufflelo_epi16(eob, 1);
132 eob = _mm256_max_epi16(eob, eob_s);
133 const __m128i final_eob = _mm_max_epi16(_mm256_castsi256_si128(eob),
134 _mm256_extractf128_si256(eob, 1));
135 *eob_ptr = _mm_extract_epi16(final_eob, 0);
136 }
137 }
138