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 <tmmintrin.h> /* SSSE3 */
12
13 #include "./vp8_rtcd.h"
14 #include "vp8/encoder/block.h"
15 #include "vpx_ports/bitops.h" /* get_msb */
16
vp8_fast_quantize_b_ssse3(BLOCK * b,BLOCKD * d)17 void vp8_fast_quantize_b_ssse3(BLOCK *b, BLOCKD *d) {
18 int eob, mask;
19
20 __m128i z0 = _mm_load_si128((__m128i *)(b->coeff));
21 __m128i z1 = _mm_load_si128((__m128i *)(b->coeff + 8));
22 __m128i round0 = _mm_load_si128((__m128i *)(b->round));
23 __m128i round1 = _mm_load_si128((__m128i *)(b->round + 8));
24 __m128i quant_fast0 = _mm_load_si128((__m128i *)(b->quant_fast));
25 __m128i quant_fast1 = _mm_load_si128((__m128i *)(b->quant_fast + 8));
26 __m128i dequant0 = _mm_load_si128((__m128i *)(d->dequant));
27 __m128i dequant1 = _mm_load_si128((__m128i *)(d->dequant + 8));
28
29 __m128i sz0, sz1, x, x0, x1, y0, y1, zeros, abs0, abs1;
30
31 DECLARE_ALIGNED(16, const uint8_t,
32 pshufb_zig_zag_mask[16]) = { 0, 1, 4, 8, 5, 2, 3, 6,
33 9, 12, 13, 10, 7, 11, 14, 15 };
34 __m128i zig_zag = _mm_load_si128((const __m128i *)pshufb_zig_zag_mask);
35
36 /* sign of z: z >> 15 */
37 sz0 = _mm_srai_epi16(z0, 15);
38 sz1 = _mm_srai_epi16(z1, 15);
39
40 /* x = abs(z) */
41 x0 = _mm_abs_epi16(z0);
42 x1 = _mm_abs_epi16(z1);
43
44 /* x += round */
45 x0 = _mm_add_epi16(x0, round0);
46 x1 = _mm_add_epi16(x1, round1);
47
48 /* y = (x * quant) >> 16 */
49 y0 = _mm_mulhi_epi16(x0, quant_fast0);
50 y1 = _mm_mulhi_epi16(x1, quant_fast1);
51
52 /* ASM saves Y for EOB */
53 /* I think we can ignore that because adding the sign doesn't change anything
54 * and multiplying 0 by dequant is OK as well */
55 abs0 = y0;
56 abs1 = y1;
57
58 /* Restore the sign bit. */
59 y0 = _mm_xor_si128(y0, sz0);
60 y1 = _mm_xor_si128(y1, sz1);
61 x0 = _mm_sub_epi16(y0, sz0);
62 x1 = _mm_sub_epi16(y1, sz1);
63
64 /* qcoeff = x */
65 _mm_store_si128((__m128i *)(d->qcoeff), x0);
66 _mm_store_si128((__m128i *)(d->qcoeff + 8), x1);
67
68 /* x * dequant */
69 x0 = _mm_mullo_epi16(x0, dequant0);
70 x1 = _mm_mullo_epi16(x1, dequant1);
71
72 /* dqcoeff = x * dequant */
73 _mm_store_si128((__m128i *)(d->dqcoeff), x0);
74 _mm_store_si128((__m128i *)(d->dqcoeff + 8), x1);
75
76 zeros = _mm_setzero_si128();
77
78 x0 = _mm_cmpgt_epi16(abs0, zeros);
79 x1 = _mm_cmpgt_epi16(abs1, zeros);
80
81 x = _mm_packs_epi16(x0, x1);
82
83 x = _mm_shuffle_epi8(x, zig_zag);
84
85 mask = _mm_movemask_epi8(x);
86
87 /* x2 is needed to increase the result from non-zero masks by 1,
88 * +1 is needed to mask undefined behavior for a null argument,
89 * the result of get_msb(1) is 0 */
90 eob = get_msb(mask * 2 + 1);
91
92 *d->eob = eob;
93 }
94