1/*
2 *  Copyright (c) 2012 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 "vpx_config.h"
12#include "vp8_rtcd.h"
13#include "vpx_ports/x86.h"
14#include "vpx_mem/vpx_mem.h"
15#include "vp8/encoder/block.h"
16#include "vp8/common/entropy.h" /* vp8_default_inv_zig_zag */
17
18#include <mmintrin.h>  /* MMX */
19#include <xmmintrin.h> /* SSE */
20#include <emmintrin.h> /* SSE2 */
21
22#define SELECT_EOB(i, z)                    \
23  do {                                      \
24    short boost = *zbin_boost_ptr;          \
25    int cmp = (x[z] < boost) | (y[z] == 0); \
26    zbin_boost_ptr++;                       \
27    if (cmp) break;                         \
28    qcoeff_ptr[z] = y[z];                   \
29    eob = i;                                \
30    zbin_boost_ptr = b->zrun_zbin_boost;    \
31  } while (0)
32
33void vp8_regular_quantize_b_sse2(BLOCK *b, BLOCKD *d) {
34  char eob = 0;
35  short *zbin_boost_ptr;
36  short *qcoeff_ptr = d->qcoeff;
37  DECLARE_ALIGNED(16, short, x[16]);
38  DECLARE_ALIGNED(16, short, y[16]);
39
40  __m128i sz0, x0, sz1, x1, y0, y1, x_minus_zbin0, x_minus_zbin1;
41  __m128i quant_shift0 = _mm_load_si128((__m128i *)(b->quant_shift));
42  __m128i quant_shift1 = _mm_load_si128((__m128i *)(b->quant_shift + 8));
43  __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
44  __m128i z1 = _mm_load_si128((__m128i *)(b->coeff + 8));
45  __m128i zbin_extra = _mm_cvtsi32_si128(b->zbin_extra);
46  __m128i zbin0 = _mm_load_si128((__m128i *)(b->zbin));
47  __m128i zbin1 = _mm_load_si128((__m128i *)(b->zbin + 8));
48  __m128i round0 = _mm_load_si128((__m128i *)(b->round));
49  __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
50  __m128i quant0 = _mm_load_si128((__m128i *)(b->quant));
51  __m128i quant1 = _mm_load_si128((__m128i *)(b->quant + 8));
52  __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
53  __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
54
55  memset(qcoeff_ptr, 0, 32);
56
57  /* Duplicate to all lanes. */
58  zbin_extra = _mm_shufflelo_epi16(zbin_extra, 0);
59  zbin_extra = _mm_unpacklo_epi16(zbin_extra, zbin_extra);
60
61  /* Sign of z: z >> 15 */
62  sz0 = _mm_srai_epi16(z0, 15);
63  sz1 = _mm_srai_epi16(z1, 15);
64
65  /* x = abs(z): (z ^ sz) - sz */
66  x0 = _mm_xor_si128(z0, sz0);
67  x1 = _mm_xor_si128(z1, sz1);
68  x0 = _mm_sub_epi16(x0, sz0);
69  x1 = _mm_sub_epi16(x1, sz1);
70
71  /* zbin[] + zbin_extra */
72  zbin0 = _mm_add_epi16(zbin0, zbin_extra);
73  zbin1 = _mm_add_epi16(zbin1, zbin_extra);
74
75  /* In C x is compared to zbin where zbin = zbin[] + boost + extra. Rebalance
76   * the equation because boost is the only value which can change:
77   * x - (zbin[] + extra) >= boost */
78  x_minus_zbin0 = _mm_sub_epi16(x0, zbin0);
79  x_minus_zbin1 = _mm_sub_epi16(x1, zbin1);
80
81  _mm_store_si128((__m128i *)(x), x_minus_zbin0);
82  _mm_store_si128((__m128i *)(x + 8), x_minus_zbin1);
83
84  /* All the remaining calculations are valid whether they are done now with
85   * simd or later inside the loop one at a time. */
86  x0 = _mm_add_epi16(x0, round0);
87  x1 = _mm_add_epi16(x1, round1);
88
89  y0 = _mm_mulhi_epi16(x0, quant0);
90  y1 = _mm_mulhi_epi16(x1, quant1);
91
92  y0 = _mm_add_epi16(y0, x0);
93  y1 = _mm_add_epi16(y1, x1);
94
95  /* Instead of shifting each value independently we convert the scaling
96   * factor with 1 << (16 - shift) so we can use multiply/return high half. */
97  y0 = _mm_mulhi_epi16(y0, quant_shift0);
98  y1 = _mm_mulhi_epi16(y1, quant_shift1);
99
100  /* Return the sign: (y ^ sz) - sz */
101  y0 = _mm_xor_si128(y0, sz0);
102  y1 = _mm_xor_si128(y1, sz1);
103  y0 = _mm_sub_epi16(y0, sz0);
104  y1 = _mm_sub_epi16(y1, sz1);
105
106  _mm_store_si128((__m128i *)(y), y0);
107  _mm_store_si128((__m128i *)(y + 8), y1);
108
109  zbin_boost_ptr = b->zrun_zbin_boost;
110
111  /* The loop gets unrolled anyway. Avoid the vp8_default_zig_zag1d lookup. */
112  SELECT_EOB(1, 0);
113  SELECT_EOB(2, 1);
114  SELECT_EOB(3, 4);
115  SELECT_EOB(4, 8);
116  SELECT_EOB(5, 5);
117  SELECT_EOB(6, 2);
118  SELECT_EOB(7, 3);
119  SELECT_EOB(8, 6);
120  SELECT_EOB(9, 9);
121  SELECT_EOB(10, 12);
122  SELECT_EOB(11, 13);
123  SELECT_EOB(12, 10);
124  SELECT_EOB(13, 7);
125  SELECT_EOB(14, 11);
126  SELECT_EOB(15, 14);
127  SELECT_EOB(16, 15);
128
129  y0 = _mm_load_si128((__m128i *)(d->qcoeff));
130  y1 = _mm_load_si128((__m128i *)(d->qcoeff + 8));
131
132  /* dqcoeff = qcoeff * dequant */
133  y0 = _mm_mullo_epi16(y0, dequant0);
134  y1 = _mm_mullo_epi16(y1, dequant1);
135
136  _mm_store_si128((__m128i *)(d->dqcoeff), y0);
137  _mm_store_si128((__m128i *)(d->dqcoeff + 8), y1);
138
139  *d->eob = eob;
140}
141
142void vp8_fast_quantize_b_sse2(BLOCK *b, BLOCKD *d) {
143  __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
144  __m128i z1 = _mm_load_si128((__m128i *)(b->coeff + 8));
145  __m128i round0 = _mm_load_si128((__m128i *)(b->round));
146  __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
147  __m128i quant_fast0 = _mm_load_si128((__m128i *)(b->quant_fast));
148  __m128i quant_fast1 = _mm_load_si128((__m128i *)(b->quant_fast + 8));
149  __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
150  __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
151  __m128i inv_zig_zag0 =
152      _mm_load_si128((const __m128i *)(vp8_default_inv_zig_zag));
153  __m128i inv_zig_zag1 =
154      _mm_load_si128((const __m128i *)(vp8_default_inv_zig_zag + 8));
155
156  __m128i sz0, sz1, x0, x1, y0, y1, xdq0, xdq1, zeros, ones;
157
158  /* sign of z: z >> 15 */
159  sz0 = _mm_srai_epi16(z0, 15);
160  sz1 = _mm_srai_epi16(z1, 15);
161
162  /* x = abs(z): (z ^ sz) - sz */
163  x0 = _mm_xor_si128(z0, sz0);
164  x1 = _mm_xor_si128(z1, sz1);
165  x0 = _mm_sub_epi16(x0, sz0);
166  x1 = _mm_sub_epi16(x1, sz1);
167
168  /* x += round */
169  x0 = _mm_add_epi16(x0, round0);
170  x1 = _mm_add_epi16(x1, round1);
171
172  /* y = (x * quant) >> 16 */
173  y0 = _mm_mulhi_epi16(x0, quant_fast0);
174  y1 = _mm_mulhi_epi16(x1, quant_fast1);
175
176  /* x = abs(y) = (y ^ sz) - sz */
177  y0 = _mm_xor_si128(y0, sz0);
178  y1 = _mm_xor_si128(y1, sz1);
179  x0 = _mm_sub_epi16(y0, sz0);
180  x1 = _mm_sub_epi16(y1, sz1);
181
182  /* qcoeff = x */
183  _mm_store_si128((__m128i *)(d->qcoeff), x0);
184  _mm_store_si128((__m128i *)(d->qcoeff + 8), x1);
185
186  /* x * dequant */
187  xdq0 = _mm_mullo_epi16(x0, dequant0);
188  xdq1 = _mm_mullo_epi16(x1, dequant1);
189
190  /* dqcoeff = x * dequant */
191  _mm_store_si128((__m128i *)(d->dqcoeff), xdq0);
192  _mm_store_si128((__m128i *)(d->dqcoeff + 8), xdq1);
193
194  /* build a mask for the zig zag */
195  zeros = _mm_setzero_si128();
196
197  x0 = _mm_cmpeq_epi16(x0, zeros);
198  x1 = _mm_cmpeq_epi16(x1, zeros);
199
200  ones = _mm_cmpeq_epi16(zeros, zeros);
201
202  x0 = _mm_xor_si128(x0, ones);
203  x1 = _mm_xor_si128(x1, ones);
204
205  x0 = _mm_and_si128(x0, inv_zig_zag0);
206  x1 = _mm_and_si128(x1, inv_zig_zag1);
207
208  x0 = _mm_max_epi16(x0, x1);
209
210  /* now down to 8 */
211  x1 = _mm_shuffle_epi32(x0, 0xE);  // 0b00001110
212
213  x0 = _mm_max_epi16(x0, x1);
214
215  /* only 4 left */
216  x1 = _mm_shufflelo_epi16(x0, 0xE);  // 0b00001110
217
218  x0 = _mm_max_epi16(x0, x1);
219
220  /* okay, just 2! */
221  x1 = _mm_shufflelo_epi16(x0, 0x1);  // 0b00000001
222
223  x0 = _mm_max_epi16(x0, x1);
224
225  *d->eob = 0xFF & _mm_cvtsi128_si32(x0);
226}
227