1 /*
2 * Copyright (c) 2022, 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
14 #include "aom_dsp/arm/mem_neon.h"
15
16 #include "av1/common/quant_common.h"
17 #include "av1/encoder/av1_quantize.h"
18
quantize_4(const tran_low_t * coeff_ptr,tran_low_t * qcoeff_ptr,tran_low_t * dqcoeff_ptr,int32x4_t v_quant_s32,int32x4_t v_dequant_s32,int32x4_t v_round_s32,int log_scale)19 static INLINE uint16x4_t quantize_4(const tran_low_t *coeff_ptr,
20 tran_low_t *qcoeff_ptr,
21 tran_low_t *dqcoeff_ptr,
22 int32x4_t v_quant_s32,
23 int32x4_t v_dequant_s32,
24 int32x4_t v_round_s32, int log_scale) {
25 const int32x4_t v_coeff = vld1q_s32(coeff_ptr);
26 const int32x4_t v_coeff_sign =
27 vreinterpretq_s32_u32(vcltq_s32(v_coeff, vdupq_n_s32(0)));
28 const int32x4_t v_log_scale = vdupq_n_s32(log_scale);
29 const int32x4_t v_abs_coeff = vabsq_s32(v_coeff);
30 // ((abs_coeff << (1 + log_scale)) >= dequant_ptr[rc01])
31 const int32x4_t v_abs_coeff_scaled =
32 vshlq_s32(v_abs_coeff, vdupq_n_s32(1 + log_scale));
33 const uint32x4_t v_mask = vcgeq_s32(v_abs_coeff_scaled, v_dequant_s32);
34 // const int64_t tmp = vmask ? (int64_t)abs_coeff + log_scaled_round : 0
35 const int32x4_t v_tmp = vandq_s32(vaddq_s32(v_abs_coeff, v_round_s32),
36 vreinterpretq_s32_u32(v_mask));
37 // const int abs_qcoeff = (int)((tmp * quant) >> (16 - log_scale));
38 const int32x4_t v_abs_qcoeff =
39 vqdmulhq_s32(vshlq_s32(v_tmp, v_log_scale), v_quant_s32);
40 // qcoeff_ptr[rc] = (tran_low_t)((abs_qcoeff ^ coeff_sign) - coeff_sign);
41 const int32x4_t v_qcoeff =
42 vsubq_s32(veorq_s32(v_abs_qcoeff, v_coeff_sign), v_coeff_sign);
43 // vshlq_s32 will shift right if shift value is negative.
44 const int32x4_t v_abs_dqcoeff =
45 vshlq_s32(vmulq_s32(v_abs_qcoeff, v_dequant_s32), vnegq_s32(v_log_scale));
46 // dqcoeff_ptr[rc] = (tran_low_t)((abs_dqcoeff ^ coeff_sign) - coeff_sign);
47 const int32x4_t v_dqcoeff =
48 vsubq_s32(veorq_s32(v_abs_dqcoeff, v_coeff_sign), v_coeff_sign);
49
50 vst1q_s32(qcoeff_ptr, v_qcoeff);
51 vst1q_s32(dqcoeff_ptr, v_dqcoeff);
52
53 // Used to find eob.
54 const uint32x4_t nz_qcoeff_mask = vcgtq_s32(v_abs_qcoeff, vdupq_n_s32(0));
55 return vmovn_u32(nz_qcoeff_mask);
56 }
57
get_max_lane_eob(const int16_t * iscan,int16x8_t v_eobmax,uint16x8_t v_mask)58 static INLINE int16x8_t get_max_lane_eob(const int16_t *iscan,
59 int16x8_t v_eobmax,
60 uint16x8_t v_mask) {
61 const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
62 const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, vdupq_n_s16(1));
63 const int16x8_t v_nz_iscan = vbslq_s16(v_mask, v_iscan_plus1, vdupq_n_s16(0));
64 return vmaxq_s16(v_eobmax, v_nz_iscan);
65 }
66
get_max_eob(int16x8_t v_eobmax)67 static INLINE uint16_t get_max_eob(int16x8_t v_eobmax) {
68 #ifdef __aarch64__
69 return (uint16_t)vmaxvq_s16(v_eobmax);
70 #else
71 const int16x4_t v_eobmax_3210 =
72 vmax_s16(vget_low_s16(v_eobmax), vget_high_s16(v_eobmax));
73 const int64x1_t v_eobmax_xx32 =
74 vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
75 const int16x4_t v_eobmax_tmp =
76 vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
77 const int64x1_t v_eobmax_xxx3 =
78 vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
79 const int16x4_t v_eobmax_final =
80 vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
81 return (uint16_t)vget_lane_s16(v_eobmax_final, 0);
82 #endif
83 }
84
av1_highbd_quantize_fp_neon(const tran_low_t * coeff_ptr,intptr_t count,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)85 void av1_highbd_quantize_fp_neon(
86 const tran_low_t *coeff_ptr, intptr_t count, const int16_t *zbin_ptr,
87 const int16_t *round_ptr, const int16_t *quant_ptr,
88 const int16_t *quant_shift_ptr, tran_low_t *qcoeff_ptr,
89 tran_low_t *dqcoeff_ptr, const int16_t *dequant_ptr, uint16_t *eob_ptr,
90 const int16_t *scan, const int16_t *iscan, int log_scale) {
91 (void)scan;
92 (void)zbin_ptr;
93 (void)quant_shift_ptr;
94
95 const int16x4_t v_quant = vld1_s16(quant_ptr);
96 const int16x4_t v_dequant = vld1_s16(dequant_ptr);
97 const int16x4_t v_zero = vdup_n_s16(0);
98 const uint16x4_t v_round_select = vcgt_s16(vdup_n_s16(log_scale), v_zero);
99 const int16x4_t v_round_no_scale = vld1_s16(round_ptr);
100 const int16x4_t v_round_log_scale =
101 vqrdmulh_n_s16(v_round_no_scale, (int16_t)(1 << (15 - log_scale)));
102 const int16x4_t v_round =
103 vbsl_s16(v_round_select, v_round_log_scale, v_round_no_scale);
104 int32x4_t v_round_s32 = vaddl_s16(v_round, v_zero);
105 int32x4_t v_quant_s32 = vshlq_n_s32(vaddl_s16(v_quant, v_zero), 15);
106 int32x4_t v_dequant_s32 = vaddl_s16(v_dequant, v_zero);
107 uint16x4_t v_mask_lo, v_mask_hi;
108 int16x8_t v_eobmax = vdupq_n_s16(-1);
109
110 // DC and first 3 AC
111 v_mask_lo = quantize_4(coeff_ptr, qcoeff_ptr, dqcoeff_ptr, v_quant_s32,
112 v_dequant_s32, v_round_s32, log_scale);
113
114 // overwrite the DC constants with AC constants
115 v_round_s32 = vdupq_lane_s32(vget_low_s32(v_round_s32), 1);
116 v_quant_s32 = vdupq_lane_s32(vget_low_s32(v_quant_s32), 1);
117 v_dequant_s32 = vdupq_lane_s32(vget_low_s32(v_dequant_s32), 1);
118
119 // 4 more AC
120 v_mask_hi = quantize_4(coeff_ptr + 4, qcoeff_ptr + 4, dqcoeff_ptr + 4,
121 v_quant_s32, v_dequant_s32, v_round_s32, log_scale);
122
123 // Find the max lane eob for the first 8 coeffs.
124 v_eobmax =
125 get_max_lane_eob(iscan, v_eobmax, vcombine_u16(v_mask_lo, v_mask_hi));
126
127 count -= 8;
128 do {
129 coeff_ptr += 8;
130 qcoeff_ptr += 8;
131 dqcoeff_ptr += 8;
132 iscan += 8;
133 v_mask_lo = quantize_4(coeff_ptr, qcoeff_ptr, dqcoeff_ptr, v_quant_s32,
134 v_dequant_s32, v_round_s32, log_scale);
135 v_mask_hi = quantize_4(coeff_ptr + 4, qcoeff_ptr + 4, dqcoeff_ptr + 4,
136 v_quant_s32, v_dequant_s32, v_round_s32, log_scale);
137 // Find the max lane eob for 8 coeffs.
138 v_eobmax =
139 get_max_lane_eob(iscan, v_eobmax, vcombine_u16(v_mask_lo, v_mask_hi));
140 count -= 8;
141 } while (count);
142
143 *eob_ptr = get_max_eob(v_eobmax);
144 }
145