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
12#include <smmintrin.h> /* SSE4.1 */
13
14#include "./vp8_rtcd.h"
15#include "vp8/encoder/block.h"
16#include "vp8/common/entropy.h" /* vp8_default_inv_zig_zag */
17
18#define SELECT_EOB(i, z, x, y, q) \
19    do { \
20        short boost = *zbin_boost_ptr; \
21        short x_z = _mm_extract_epi16(x, z); \
22        short y_z = _mm_extract_epi16(y, z); \
23        int cmp = (x_z < boost) | (y_z == 0); \
24        zbin_boost_ptr++; \
25        if (cmp) \
26            break; \
27        q = _mm_insert_epi16(q, y_z, z); \
28        eob = i; \
29        zbin_boost_ptr = b->zrun_zbin_boost; \
30    } while (0)
31
32void vp8_regular_quantize_b_sse4_1(BLOCK *b, BLOCKD *d) {
33    char eob = 0;
34    short *zbin_boost_ptr  = b->zrun_zbin_boost;
35
36    __m128i sz0, x0, sz1, x1, y0, y1, x_minus_zbin0, x_minus_zbin1,
37            dqcoeff0, dqcoeff1;
38    __m128i quant_shift0 = _mm_load_si128((__m128i *)(b->quant_shift));
39    __m128i quant_shift1 = _mm_load_si128((__m128i *)(b->quant_shift + 8));
40    __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
41    __m128i z1 = _mm_load_si128((__m128i *)(b->coeff+8));
42    __m128i zbin_extra = _mm_cvtsi32_si128(b->zbin_extra);
43    __m128i zbin0 = _mm_load_si128((__m128i *)(b->zbin));
44    __m128i zbin1 = _mm_load_si128((__m128i *)(b->zbin + 8));
45    __m128i round0 = _mm_load_si128((__m128i *)(b->round));
46    __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
47    __m128i quant0 = _mm_load_si128((__m128i *)(b->quant));
48    __m128i quant1 = _mm_load_si128((__m128i *)(b->quant + 8));
49    __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
50    __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
51    __m128i qcoeff0 = _mm_setzero_si128();
52    __m128i qcoeff1 = _mm_setzero_si128();
53
54    /* Duplicate to all lanes. */
55    zbin_extra = _mm_shufflelo_epi16(zbin_extra, 0);
56    zbin_extra = _mm_unpacklo_epi16(zbin_extra, zbin_extra);
57
58    /* Sign of z: z >> 15 */
59    sz0 = _mm_srai_epi16(z0, 15);
60    sz1 = _mm_srai_epi16(z1, 15);
61
62    /* x = abs(z): (z ^ sz) - sz */
63    x0 = _mm_xor_si128(z0, sz0);
64    x1 = _mm_xor_si128(z1, sz1);
65    x0 = _mm_sub_epi16(x0, sz0);
66    x1 = _mm_sub_epi16(x1, sz1);
67
68    /* zbin[] + zbin_extra */
69    zbin0 = _mm_add_epi16(zbin0, zbin_extra);
70    zbin1 = _mm_add_epi16(zbin1, zbin_extra);
71
72    /* In C x is compared to zbin where zbin = zbin[] + boost + extra. Rebalance
73     * the equation because boost is the only value which can change:
74     * x - (zbin[] + extra) >= boost */
75    x_minus_zbin0 = _mm_sub_epi16(x0, zbin0);
76    x_minus_zbin1 = _mm_sub_epi16(x1, zbin1);
77
78    /* All the remaining calculations are valid whether they are done now with
79     * simd or later inside the loop one at a time. */
80    x0 = _mm_add_epi16(x0, round0);
81    x1 = _mm_add_epi16(x1, round1);
82
83    y0 = _mm_mulhi_epi16(x0, quant0);
84    y1 = _mm_mulhi_epi16(x1, quant1);
85
86    y0 = _mm_add_epi16(y0, x0);
87    y1 = _mm_add_epi16(y1, x1);
88
89    /* Instead of shifting each value independently we convert the scaling
90     * factor with 1 << (16 - shift) so we can use multiply/return high half. */
91    y0 = _mm_mulhi_epi16(y0, quant_shift0);
92    y1 = _mm_mulhi_epi16(y1, quant_shift1);
93
94    /* Return the sign: (y ^ sz) - sz */
95    y0 = _mm_xor_si128(y0, sz0);
96    y1 = _mm_xor_si128(y1, sz1);
97    y0 = _mm_sub_epi16(y0, sz0);
98    y1 = _mm_sub_epi16(y1, sz1);
99
100    /* The loop gets unrolled anyway. Avoid the vp8_default_zig_zag1d lookup. */
101    SELECT_EOB(1, 0, x_minus_zbin0, y0, qcoeff0);
102    SELECT_EOB(2, 1, x_minus_zbin0, y0, qcoeff0);
103    SELECT_EOB(3, 4, x_minus_zbin0, y0, qcoeff0);
104    SELECT_EOB(4, 0, x_minus_zbin1, y1, qcoeff1);
105    SELECT_EOB(5, 5, x_minus_zbin0, y0, qcoeff0);
106    SELECT_EOB(6, 2, x_minus_zbin0, y0, qcoeff0);
107    SELECT_EOB(7, 3, x_minus_zbin0, y0, qcoeff0);
108    SELECT_EOB(8, 6, x_minus_zbin0, y0, qcoeff0);
109    SELECT_EOB(9, 1, x_minus_zbin1, y1, qcoeff1);
110    SELECT_EOB(10, 4, x_minus_zbin1, y1, qcoeff1);
111    SELECT_EOB(11, 5, x_minus_zbin1, y1, qcoeff1);
112    SELECT_EOB(12, 2, x_minus_zbin1, y1, qcoeff1);
113    SELECT_EOB(13, 7, x_minus_zbin0, y0, qcoeff0);
114    SELECT_EOB(14, 3, x_minus_zbin1, y1, qcoeff1);
115    SELECT_EOB(15, 6, x_minus_zbin1, y1, qcoeff1);
116    SELECT_EOB(16, 7, x_minus_zbin1, y1, qcoeff1);
117
118    _mm_store_si128((__m128i *)(d->qcoeff), qcoeff0);
119    _mm_store_si128((__m128i *)(d->qcoeff + 8), qcoeff1);
120
121    dqcoeff0 = _mm_mullo_epi16(qcoeff0, dequant0);
122    dqcoeff1 = _mm_mullo_epi16(qcoeff1, dequant1);
123
124    _mm_store_si128((__m128i *)(d->dqcoeff), dqcoeff0);
125    _mm_store_si128((__m128i *)(d->dqcoeff + 8), dqcoeff1);
126
127    *d->eob = eob;
128}
129