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