1 /*
2 * Copyright (c) 2017 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 #ifndef VPX_DSP_X86_CONVOLVE_SSSE3_H_
12 #define VPX_DSP_X86_CONVOLVE_SSSE3_H_
13
14 #include <assert.h>
15 #include <tmmintrin.h> // SSSE3
16
17 #include "./vpx_config.h"
18
shuffle_filter_ssse3(const int16_t * const filter,__m128i * const f)19 static INLINE void shuffle_filter_ssse3(const int16_t *const filter,
20 __m128i *const f) {
21 const __m128i f_values = _mm_load_si128((const __m128i *)filter);
22 // pack and duplicate the filter values
23 f[0] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u));
24 f[1] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u));
25 f[2] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u));
26 f[3] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu));
27 }
28
shuffle_filter_odd_ssse3(const int16_t * const filter,__m128i * const f)29 static INLINE void shuffle_filter_odd_ssse3(const int16_t *const filter,
30 __m128i *const f) {
31 const __m128i f_values = _mm_load_si128((const __m128i *)filter);
32 // pack and duplicate the filter values
33 // It utilizes the fact that the high byte of filter[3] is always 0 to clean
34 // half of f[0] and f[4].
35 assert(filter[3] >= 0 && filter[3] < 256);
36 f[0] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0007u));
37 f[1] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0402u));
38 f[2] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0806u));
39 f[3] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0c0au));
40 f[4] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x070eu));
41 }
42
convolve8_8_ssse3(const __m128i * const s,const __m128i * const f)43 static INLINE __m128i convolve8_8_ssse3(const __m128i *const s,
44 const __m128i *const f) {
45 // multiply 2 adjacent elements with the filter and add the result
46 const __m128i k_64 = _mm_set1_epi16(1 << 6);
47 const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]);
48 const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]);
49 const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]);
50 const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]);
51 __m128i sum1, sum2;
52
53 // sum the results together, saturating only on the final step
54 // adding x0 with x2 and x1 with x3 is the only order that prevents
55 // outranges for all filters
56 sum1 = _mm_add_epi16(x0, x2);
57 sum2 = _mm_add_epi16(x1, x3);
58 // add the rounding offset early to avoid another saturated add
59 sum1 = _mm_add_epi16(sum1, k_64);
60 sum1 = _mm_adds_epi16(sum1, sum2);
61 // shift by 7 bit each 16 bit
62 sum1 = _mm_srai_epi16(sum1, 7);
63 return sum1;
64 }
65
convolve8_8_even_offset_ssse3(const __m128i * const s,const __m128i * const f)66 static INLINE __m128i convolve8_8_even_offset_ssse3(const __m128i *const s,
67 const __m128i *const f) {
68 // multiply 2 adjacent elements with the filter and add the result
69 const __m128i k_64 = _mm_set1_epi16(1 << 6);
70 const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]);
71 const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]);
72 const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]);
73 const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]);
74 // compensate the subtracted 64 in f[1]. x4 is always non negative.
75 const __m128i x4 = _mm_maddubs_epi16(s[1], _mm_set1_epi8(64));
76 // add and saturate the results together
77 __m128i temp = _mm_adds_epi16(x0, x3);
78 temp = _mm_adds_epi16(temp, x1);
79 temp = _mm_adds_epi16(temp, x2);
80 temp = _mm_adds_epi16(temp, x4);
81 // round and shift by 7 bit each 16 bit
82 temp = _mm_adds_epi16(temp, k_64);
83 temp = _mm_srai_epi16(temp, 7);
84 return temp;
85 }
86
convolve8_8_odd_offset_ssse3(const __m128i * const s,const __m128i * const f)87 static INLINE __m128i convolve8_8_odd_offset_ssse3(const __m128i *const s,
88 const __m128i *const f) {
89 // multiply 2 adjacent elements with the filter and add the result
90 const __m128i k_64 = _mm_set1_epi16(1 << 6);
91 const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]);
92 const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]);
93 const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]);
94 const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]);
95 const __m128i x4 = _mm_maddubs_epi16(s[4], f[4]);
96 // compensate the subtracted 64 in f[2]. x5 is always non negative.
97 const __m128i x5 = _mm_maddubs_epi16(s[2], _mm_set1_epi8(64));
98 __m128i temp;
99
100 // add and saturate the results together
101 temp = _mm_adds_epi16(x0, x1);
102 temp = _mm_adds_epi16(temp, x2);
103 temp = _mm_adds_epi16(temp, x3);
104 temp = _mm_adds_epi16(temp, x4);
105 temp = _mm_adds_epi16(temp, x5);
106 // round and shift by 7 bit each 16 bit
107 temp = _mm_adds_epi16(temp, k_64);
108 temp = _mm_srai_epi16(temp, 7);
109 return temp;
110 }
111
112 #endif // VPX_DSP_X86_CONVOLVE_SSSE3_H_
113