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 "./vp8_rtcd.h"
14 #include "vp8/encoder/block.h"
15 
16 static const uint16_t inv_zig_zag[16] = { 1, 2, 6,  7,  3,  5,  8,  13,
17                                           4, 9, 12, 14, 10, 11, 15, 16 };
18 
vp8_fast_quantize_b_neon(BLOCK * b,BLOCKD * d)19 void vp8_fast_quantize_b_neon(BLOCK *b, BLOCKD *d) {
20   const int16x8_t one_q = vdupq_n_s16(-1), z0 = vld1q_s16(b->coeff),
21                   z1 = vld1q_s16(b->coeff + 8), round0 = vld1q_s16(b->round),
22                   round1 = vld1q_s16(b->round + 8),
23                   quant0 = vld1q_s16(b->quant_fast),
24                   quant1 = vld1q_s16(b->quant_fast + 8),
25                   dequant0 = vld1q_s16(d->dequant),
26                   dequant1 = vld1q_s16(d->dequant + 8);
27   const uint16x8_t zig_zag0 = vld1q_u16(inv_zig_zag),
28                    zig_zag1 = vld1q_u16(inv_zig_zag + 8);
29   int16x8_t x0, x1, sz0, sz1, y0, y1;
30   uint16x8_t eob0, eob1;
31 #ifndef __aarch64__
32   uint16x4_t eob_d16;
33   uint32x2_t eob_d32;
34   uint32x4_t eob_q32;
35 #endif  // __arch64__
36 
37   /* sign of z: z >> 15 */
38   sz0 = vshrq_n_s16(z0, 15);
39   sz1 = vshrq_n_s16(z1, 15);
40 
41   /* x = abs(z) */
42   x0 = vabsq_s16(z0);
43   x1 = vabsq_s16(z1);
44 
45   /* x += round */
46   x0 = vaddq_s16(x0, round0);
47   x1 = vaddq_s16(x1, round1);
48 
49   /* y = 2 * (x * quant) >> 16 */
50   y0 = vqdmulhq_s16(x0, quant0);
51   y1 = vqdmulhq_s16(x1, quant1);
52 
53   /* Compensate for doubling in vqdmulhq */
54   y0 = vshrq_n_s16(y0, 1);
55   y1 = vshrq_n_s16(y1, 1);
56 
57   /* Restore sign bit */
58   y0 = veorq_s16(y0, sz0);
59   y1 = veorq_s16(y1, sz1);
60   x0 = vsubq_s16(y0, sz0);
61   x1 = vsubq_s16(y1, sz1);
62 
63   /* find non-zero elements */
64   eob0 = vtstq_s16(x0, one_q);
65   eob1 = vtstq_s16(x1, one_q);
66 
67   /* mask zig zag */
68   eob0 = vandq_u16(eob0, zig_zag0);
69   eob1 = vandq_u16(eob1, zig_zag1);
70 
71   /* select the largest value */
72   eob0 = vmaxq_u16(eob0, eob1);
73 #ifdef __aarch64__
74   *d->eob = (int8_t)vmaxvq_u16(eob0);
75 #else
76   eob_d16 = vmax_u16(vget_low_u16(eob0), vget_high_u16(eob0));
77   eob_q32 = vmovl_u16(eob_d16);
78   eob_d32 = vmax_u32(vget_low_u32(eob_q32), vget_high_u32(eob_q32));
79   eob_d32 = vpmax_u32(eob_d32, eob_d32);
80 
81   vst1_lane_s8((int8_t *)d->eob, vreinterpret_s8_u32(eob_d32), 0);
82 #endif  // __aarch64__
83 
84   /* qcoeff = x */
85   vst1q_s16(d->qcoeff, x0);
86   vst1q_s16(d->qcoeff + 8, x1);
87 
88   /* dqcoeff = x * dequant */
89   vst1q_s16(d->dqcoeff, vmulq_s16(dequant0, x0));
90   vst1q_s16(d->dqcoeff + 8, vmulq_s16(dequant1, x1));
91 }
92