/src/libvpx/vpx_dsp/x86/convolve_ssse3.h
Line | Count | Source (jump to first uncovered line) |
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_VPX_DSP_X86_CONVOLVE_SSSE3_H_ |
12 | | #define VPX_VPX_DSP_X86_CONVOLVE_SSSE3_H_ |
13 | | |
14 | | #include <assert.h> |
15 | | #include <tmmintrin.h> // SSSE3 |
16 | | |
17 | | #include "./vpx_config.h" |
18 | | |
19 | | static INLINE void shuffle_filter_ssse3(const int16_t *const filter, |
20 | 0 | __m128i *const f) { |
21 | 0 | const __m128i f_values = _mm_load_si128((const __m128i *)filter); |
22 | | // pack and duplicate the filter values |
23 | 0 | f[0] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0200u)); |
24 | 0 | f[1] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0604u)); |
25 | 0 | f[2] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0a08u)); |
26 | 0 | f[3] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0e0cu)); |
27 | 0 | } Unexecuted instantiation: vp9_frame_scale_ssse3.c:shuffle_filter_ssse3 Unexecuted instantiation: vpx_subpixel_8t_intrin_avx2.c:shuffle_filter_ssse3 Unexecuted instantiation: vpx_subpixel_8t_intrin_ssse3.c:shuffle_filter_ssse3 |
28 | | |
29 | | static INLINE void shuffle_filter_odd_ssse3(const int16_t *const filter, |
30 | 0 | __m128i *const f) { |
31 | 0 | 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 | 0 | assert(filter[3] >= 0 && filter[3] < 256); |
36 | 0 | f[0] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0007u)); |
37 | 0 | f[1] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0402u)); |
38 | 0 | f[2] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0806u)); |
39 | 0 | f[3] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x0c0au)); |
40 | 0 | f[4] = _mm_shuffle_epi8(f_values, _mm_set1_epi16(0x070eu)); |
41 | 0 | } Unexecuted instantiation: vp9_frame_scale_ssse3.c:shuffle_filter_odd_ssse3 Unexecuted instantiation: vpx_subpixel_8t_intrin_avx2.c:shuffle_filter_odd_ssse3 Unexecuted instantiation: vpx_subpixel_8t_intrin_ssse3.c:shuffle_filter_odd_ssse3 |
42 | | |
43 | | static INLINE __m128i convolve8_8_ssse3(const __m128i *const s, |
44 | 7.85M | const __m128i *const f) { |
45 | | // multiply 2 adjacent elements with the filter and add the result |
46 | 7.85M | const __m128i k_64 = _mm_set1_epi16(1 << 6); |
47 | 7.85M | const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]); |
48 | 7.85M | const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]); |
49 | 7.85M | const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]); |
50 | 7.85M | const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]); |
51 | 7.85M | __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 | 7.85M | sum1 = _mm_add_epi16(x0, x2); |
57 | 7.85M | sum2 = _mm_add_epi16(x1, x3); |
58 | | // add the rounding offset early to avoid another saturated add |
59 | 7.85M | sum1 = _mm_add_epi16(sum1, k_64); |
60 | 7.85M | sum1 = _mm_adds_epi16(sum1, sum2); |
61 | | // shift by 7 bit each 16 bit |
62 | 7.85M | sum1 = _mm_srai_epi16(sum1, 7); |
63 | 7.85M | return sum1; |
64 | 7.85M | } Unexecuted instantiation: vp9_frame_scale_ssse3.c:convolve8_8_ssse3 vpx_subpixel_8t_intrin_avx2.c:convolve8_8_ssse3 Line | Count | Source | 44 | 7.85M | const __m128i *const f) { | 45 | | // multiply 2 adjacent elements with the filter and add the result | 46 | 7.85M | const __m128i k_64 = _mm_set1_epi16(1 << 6); | 47 | 7.85M | const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]); | 48 | 7.85M | const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]); | 49 | 7.85M | const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]); | 50 | 7.85M | const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]); | 51 | 7.85M | __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 | 7.85M | sum1 = _mm_add_epi16(x0, x2); | 57 | 7.85M | sum2 = _mm_add_epi16(x1, x3); | 58 | | // add the rounding offset early to avoid another saturated add | 59 | 7.85M | sum1 = _mm_add_epi16(sum1, k_64); | 60 | 7.85M | sum1 = _mm_adds_epi16(sum1, sum2); | 61 | | // shift by 7 bit each 16 bit | 62 | 7.85M | sum1 = _mm_srai_epi16(sum1, 7); | 63 | 7.85M | return sum1; | 64 | 7.85M | } |
Unexecuted instantiation: vpx_subpixel_8t_intrin_ssse3.c:convolve8_8_ssse3 |
65 | | |
66 | | static INLINE __m128i convolve8_8_even_offset_ssse3(const __m128i *const s, |
67 | 0 | const __m128i *const f) { |
68 | | // multiply 2 adjacent elements with the filter and add the result |
69 | 0 | const __m128i k_64 = _mm_set1_epi16(1 << 6); |
70 | 0 | const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]); |
71 | 0 | const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]); |
72 | 0 | const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]); |
73 | 0 | const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]); |
74 | | // compensate the subtracted 64 in f[1]. x4 is always non negative. |
75 | 0 | const __m128i x4 = _mm_maddubs_epi16(s[1], _mm_set1_epi8(64)); |
76 | | // add and saturate the results together |
77 | 0 | __m128i temp = _mm_adds_epi16(x0, x3); |
78 | 0 | temp = _mm_adds_epi16(temp, x1); |
79 | 0 | temp = _mm_adds_epi16(temp, x2); |
80 | 0 | temp = _mm_adds_epi16(temp, x4); |
81 | | // round and shift by 7 bit each 16 bit |
82 | 0 | temp = _mm_adds_epi16(temp, k_64); |
83 | 0 | temp = _mm_srai_epi16(temp, 7); |
84 | 0 | return temp; |
85 | 0 | } Unexecuted instantiation: vp9_frame_scale_ssse3.c:convolve8_8_even_offset_ssse3 Unexecuted instantiation: vpx_subpixel_8t_intrin_avx2.c:convolve8_8_even_offset_ssse3 Unexecuted instantiation: vpx_subpixel_8t_intrin_ssse3.c:convolve8_8_even_offset_ssse3 |
86 | | |
87 | | static INLINE __m128i convolve8_8_odd_offset_ssse3(const __m128i *const s, |
88 | 0 | const __m128i *const f) { |
89 | | // multiply 2 adjacent elements with the filter and add the result |
90 | 0 | const __m128i k_64 = _mm_set1_epi16(1 << 6); |
91 | 0 | const __m128i x0 = _mm_maddubs_epi16(s[0], f[0]); |
92 | 0 | const __m128i x1 = _mm_maddubs_epi16(s[1], f[1]); |
93 | 0 | const __m128i x2 = _mm_maddubs_epi16(s[2], f[2]); |
94 | 0 | const __m128i x3 = _mm_maddubs_epi16(s[3], f[3]); |
95 | 0 | const __m128i x4 = _mm_maddubs_epi16(s[4], f[4]); |
96 | | // compensate the subtracted 64 in f[2]. x5 is always non negative. |
97 | 0 | const __m128i x5 = _mm_maddubs_epi16(s[2], _mm_set1_epi8(64)); |
98 | 0 | __m128i temp; |
99 | | |
100 | | // add and saturate the results together |
101 | 0 | temp = _mm_adds_epi16(x0, x1); |
102 | 0 | temp = _mm_adds_epi16(temp, x2); |
103 | 0 | temp = _mm_adds_epi16(temp, x3); |
104 | 0 | temp = _mm_adds_epi16(temp, x4); |
105 | 0 | temp = _mm_adds_epi16(temp, x5); |
106 | | // round and shift by 7 bit each 16 bit |
107 | 0 | temp = _mm_adds_epi16(temp, k_64); |
108 | 0 | temp = _mm_srai_epi16(temp, 7); |
109 | 0 | return temp; |
110 | 0 | } Unexecuted instantiation: vp9_frame_scale_ssse3.c:convolve8_8_odd_offset_ssse3 Unexecuted instantiation: vpx_subpixel_8t_intrin_avx2.c:convolve8_8_odd_offset_ssse3 Unexecuted instantiation: vpx_subpixel_8t_intrin_ssse3.c:convolve8_8_odd_offset_ssse3 |
111 | | |
112 | | #endif // VPX_VPX_DSP_X86_CONVOLVE_SSSE3_H_ |