1/*
2 *  Copyright (c) 2014 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
13#include <math.h>
14
15#include "vpx_mem/vpx_mem.h"
16
17#include "vp9/common/vp9_quant_common.h"
18#include "vp9/common/vp9_seg_common.h"
19
20#include "vp9/encoder/vp9_encoder.h"
21#include "vp9/encoder/vp9_quantize.h"
22#include "vp9/encoder/vp9_rd.h"
23
24void vp9_quantize_fp_neon(const int16_t *coeff_ptr, intptr_t count,
25                          int skip_block, const int16_t *zbin_ptr,
26                          const int16_t *round_ptr, const int16_t *quant_ptr,
27                          const int16_t *quant_shift_ptr, int16_t *qcoeff_ptr,
28                          int16_t *dqcoeff_ptr, const int16_t *dequant_ptr,
29                          uint16_t *eob_ptr,
30                          const int16_t *scan, const int16_t *iscan) {
31  // TODO(jingning) Decide the need of these arguments after the
32  // quantization process is completed.
33  (void)zbin_ptr;
34  (void)quant_shift_ptr;
35  (void)scan;
36
37  if (!skip_block) {
38    // Quantization pass: All coefficients with index >= zero_flag are
39    // skippable. Note: zero_flag can be zero.
40    int i;
41    const int16x8_t v_zero = vdupq_n_s16(0);
42    const int16x8_t v_one = vdupq_n_s16(1);
43    int16x8_t v_eobmax_76543210 = vdupq_n_s16(-1);
44    int16x8_t v_round = vmovq_n_s16(round_ptr[1]);
45    int16x8_t v_quant = vmovq_n_s16(quant_ptr[1]);
46    int16x8_t v_dequant = vmovq_n_s16(dequant_ptr[1]);
47    // adjust for dc
48    v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
49    v_quant = vsetq_lane_s16(quant_ptr[0], v_quant, 0);
50    v_dequant = vsetq_lane_s16(dequant_ptr[0], v_dequant, 0);
51    // process dc and the first seven ac coeffs
52    {
53      const int16x8_t v_iscan = vld1q_s16(&iscan[0]);
54      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[0]);
55      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
56      const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
57      const int32x4_t v_tmp_lo = vmull_s16(vget_low_s16(v_tmp),
58                                           vget_low_s16(v_quant));
59      const int32x4_t v_tmp_hi = vmull_s16(vget_high_s16(v_tmp),
60                                           vget_high_s16(v_quant));
61      const int16x8_t v_tmp2 = vcombine_s16(vshrn_n_s32(v_tmp_lo, 16),
62                                            vshrn_n_s32(v_tmp_hi, 16));
63      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
64      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
65      const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
66      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
67      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
68      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
69      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
70      vst1q_s16(&qcoeff_ptr[0], v_qcoeff);
71      vst1q_s16(&dqcoeff_ptr[0], v_dqcoeff);
72      v_round = vmovq_n_s16(round_ptr[1]);
73      v_quant = vmovq_n_s16(quant_ptr[1]);
74      v_dequant = vmovq_n_s16(dequant_ptr[1]);
75    }
76    // now process the rest of the ac coeffs
77    for (i = 8; i < count; i += 8) {
78      const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
79      const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[i]);
80      const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
81      const int16x8_t v_tmp = vabaq_s16(v_round, v_coeff, v_zero);
82      const int32x4_t v_tmp_lo = vmull_s16(vget_low_s16(v_tmp),
83                                           vget_low_s16(v_quant));
84      const int32x4_t v_tmp_hi = vmull_s16(vget_high_s16(v_tmp),
85                                           vget_high_s16(v_quant));
86      const int16x8_t v_tmp2 = vcombine_s16(vshrn_n_s32(v_tmp_lo, 16),
87                                            vshrn_n_s32(v_tmp_hi, 16));
88      const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
89      const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
90      const int16x8_t v_nz_iscan = vbslq_s16(v_nz_mask, v_zero, v_iscan_plus1);
91      const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
92      const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
93      const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
94      v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
95      vst1q_s16(&qcoeff_ptr[i], v_qcoeff);
96      vst1q_s16(&dqcoeff_ptr[i], v_dqcoeff);
97    }
98    {
99      const int16x4_t v_eobmax_3210 =
100          vmax_s16(vget_low_s16(v_eobmax_76543210),
101                   vget_high_s16(v_eobmax_76543210));
102      const int64x1_t v_eobmax_xx32 =
103          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
104      const int16x4_t v_eobmax_tmp =
105          vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
106      const int64x1_t v_eobmax_xxx3 =
107          vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
108      const int16x4_t v_eobmax_final =
109          vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
110
111      *eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
112    }
113  } else {
114    memset(qcoeff_ptr, 0, count * sizeof(int16_t));
115    memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
116    *eob_ptr = 0;
117  }
118}
119