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 
vp9_quantize_fp_neon(const int16_t * coeff_ptr,intptr_t count,int skip_block,const int16_t * zbin_ptr,const int16_t * round_ptr,const int16_t * quant_ptr,const int16_t * quant_shift_ptr,int16_t * qcoeff_ptr,int16_t * dqcoeff_ptr,const int16_t * dequant_ptr,int zbin_oq_value,uint16_t * eob_ptr,const int16_t * scan,const int16_t * iscan)24 void 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                           int zbin_oq_value, uint16_t *eob_ptr,
30                           const int16_t *scan, const int16_t *iscan) {
31   int i;
32   // TODO(jingning) Decide the need of these arguments after the
33   // quantization process is completed.
34   (void)zbin_ptr;
35   (void)quant_shift_ptr;
36   (void)zbin_oq_value;
37   (void)scan;
38 
39   if (!skip_block) {
40     // Quantization pass: All coefficients with index >= zero_flag are
41     // skippable. Note: zero_flag can be zero.
42 
43     const int16x8_t v_zero = vdupq_n_s16(0);
44     const int16x8_t v_one = vdupq_n_s16(1);
45     int16x8_t v_eobmax_76543210 = vdupq_n_s16(-1);
46     int16x8_t v_round = vmovq_n_s16(round_ptr[1]);
47     int16x8_t v_quant = vmovq_n_s16(quant_ptr[1]);
48     int16x8_t v_dequant = vmovq_n_s16(dequant_ptr[1]);
49     // adjust for dc
50     v_round = vsetq_lane_s16(round_ptr[0], v_round, 0);
51     v_quant = vsetq_lane_s16(quant_ptr[0], v_quant, 0);
52     v_dequant = vsetq_lane_s16(dequant_ptr[0], v_dequant, 0);
53 
54     for (i = 0; i < count; i += 8) {
55       const int16x8_t v_iscan = vld1q_s16(&iscan[i]);
56       const int16x8_t v_coeff = vld1q_s16(&coeff_ptr[i]);
57       const int16x8_t v_coeff_sign = vshrq_n_s16(v_coeff, 15);
58       const int16x8_t v_abs_coeff = vabsq_s16(v_coeff);
59       const int16x8_t v_tmp = vqaddq_s16(v_abs_coeff, v_round);
60       const int32x4_t v_tmp_lo = vmull_s16(vget_low_s16(v_tmp),
61                                            vget_low_s16(v_quant));
62       const int32x4_t v_tmp_hi = vmull_s16(vget_high_s16(v_tmp),
63                                            vget_high_s16(v_quant));
64       const int16x8_t v_tmp2 = vcombine_s16(vshrn_n_s32(v_tmp_lo, 16),
65                                             vshrn_n_s32(v_tmp_hi, 16));
66       const uint16x8_t v_nz_mask = vceqq_s16(v_tmp2, v_zero);
67       const int16x8_t v_iscan_plus1 = vaddq_s16(v_iscan, v_one);
68       const int16x8_t v_nz_iscan =
69           vandq_s16(vmvnq_s16(vreinterpretq_s16_u16(v_nz_mask)), v_iscan_plus1);
70       const int16x8_t v_qcoeff_a = veorq_s16(v_tmp2, v_coeff_sign);
71       const int16x8_t v_qcoeff = vsubq_s16(v_qcoeff_a, v_coeff_sign);
72       const int16x8_t v_dqcoeff = vmulq_s16(v_qcoeff, v_dequant);
73 
74       v_eobmax_76543210 = vmaxq_s16(v_eobmax_76543210, v_nz_iscan);
75 
76       vst1q_s16(&qcoeff_ptr[i], v_qcoeff);
77       vst1q_s16(&dqcoeff_ptr[i], v_dqcoeff);
78       v_round = vmovq_n_s16(round_ptr[1]);
79       v_quant = vmovq_n_s16(quant_ptr[1]);
80       v_dequant = vmovq_n_s16(dequant_ptr[1]);
81     }
82     {
83       const int16x4_t v_eobmax_3210 =
84           vmax_s16(vget_low_s16(v_eobmax_76543210),
85                    vget_high_s16(v_eobmax_76543210));
86       const int64x1_t v_eobmax_xx32 =
87           vshr_n_s64(vreinterpret_s64_s16(v_eobmax_3210), 32);
88       const int16x4_t v_eobmax_tmp =
89           vmax_s16(v_eobmax_3210, vreinterpret_s16_s64(v_eobmax_xx32));
90       const int64x1_t v_eobmax_xxx3 =
91           vshr_n_s64(vreinterpret_s64_s16(v_eobmax_tmp), 16);
92       const int16x4_t v_eobmax_final =
93           vmax_s16(v_eobmax_tmp, vreinterpret_s16_s64(v_eobmax_xxx3));
94 
95       *eob_ptr = (uint16_t)vget_lane_s16(v_eobmax_final, 0);
96     }
97   } else {
98     vpx_memset(qcoeff_ptr, 0, count * sizeof(int16_t));
99     vpx_memset(dqcoeff_ptr, 0, count * sizeof(int16_t));
100     *eob_ptr = 0;
101   }
102 }
103