28
28
#ifndef X264_I386_DCT_H
29
29
#define X264_I386_DCT_H
31
#define x264_sub4x4_dct_mmx x264_template(sub4x4_dct_mmx)
31
32
void x264_sub4x4_dct_mmx ( dctcoef dct [16], pixel *pix1, pixel *pix2 );
33
#define x264_sub8x8_dct_mmx x264_template(sub8x8_dct_mmx)
32
34
void x264_sub8x8_dct_mmx ( dctcoef dct[ 4][16], pixel *pix1, pixel *pix2 );
35
#define x264_sub16x16_dct_mmx x264_template(sub16x16_dct_mmx)
33
36
void x264_sub16x16_dct_mmx ( dctcoef dct[16][16], pixel *pix1, pixel *pix2 );
37
#define x264_sub8x8_dct_sse2 x264_template(sub8x8_dct_sse2)
34
38
void x264_sub8x8_dct_sse2 ( int16_t dct[ 4][16], uint8_t *pix1, uint8_t *pix2 );
39
#define x264_sub16x16_dct_sse2 x264_template(sub16x16_dct_sse2)
35
40
void x264_sub16x16_dct_sse2 ( int16_t dct[16][16], uint8_t *pix1, uint8_t *pix2 );
41
#define x264_sub4x4_dct_ssse3 x264_template(sub4x4_dct_ssse3)
36
42
void x264_sub4x4_dct_ssse3 ( int16_t dct [16], uint8_t *pix1, uint8_t *pix2 );
43
#define x264_sub4x4_dct_avx512 x264_template(sub4x4_dct_avx512)
37
44
void x264_sub4x4_dct_avx512 ( int16_t dct [16], uint8_t *pix1, uint8_t *pix2 );
45
#define x264_sub8x8_dct_ssse3 x264_template(sub8x8_dct_ssse3)
38
46
void x264_sub8x8_dct_ssse3 ( int16_t dct[ 4][16], uint8_t *pix1, uint8_t *pix2 );
47
#define x264_sub16x16_dct_ssse3 x264_template(sub16x16_dct_ssse3)
39
48
void x264_sub16x16_dct_ssse3( int16_t dct[16][16], uint8_t *pix1, uint8_t *pix2 );
49
#define x264_sub8x8_dct_avx x264_template(sub8x8_dct_avx)
40
50
void x264_sub8x8_dct_avx ( int16_t dct[ 4][16], uint8_t *pix1, uint8_t *pix2 );
51
#define x264_sub16x16_dct_avx x264_template(sub16x16_dct_avx)
41
52
void x264_sub16x16_dct_avx ( int16_t dct[16][16], uint8_t *pix1, uint8_t *pix2 );
53
#define x264_sub8x8_dct_xop x264_template(sub8x8_dct_xop)
42
54
void x264_sub8x8_dct_xop ( int16_t dct[ 4][16], uint8_t *pix1, uint8_t *pix2 );
55
#define x264_sub16x16_dct_xop x264_template(sub16x16_dct_xop)
43
56
void x264_sub16x16_dct_xop ( int16_t dct[16][16], uint8_t *pix1, uint8_t *pix2 );
57
#define x264_sub8x8_dct_avx2 x264_template(sub8x8_dct_avx2)
44
58
void x264_sub8x8_dct_avx2 ( int16_t dct[ 4][16], uint8_t *pix1, uint8_t *pix2 );
59
#define x264_sub8x8_dct_avx512 x264_template(sub8x8_dct_avx512)
45
60
void x264_sub8x8_dct_avx512 ( int16_t dct[ 4][16], uint8_t *pix1, uint8_t *pix2 );
61
#define x264_sub16x16_dct_avx2 x264_template(sub16x16_dct_avx2)
46
62
void x264_sub16x16_dct_avx2 ( int16_t dct[16][16], uint8_t *pix1, uint8_t *pix2 );
63
#define x264_sub16x16_dct_avx512 x264_template(sub16x16_dct_avx512)
47
64
void x264_sub16x16_dct_avx512( int16_t dct[16][16], uint8_t *pix1, uint8_t *pix2 );
65
#define x264_sub8x8_dct_dc_mmx2 x264_template(sub8x8_dct_dc_mmx2)
48
66
void x264_sub8x8_dct_dc_mmx2 ( int16_t dct [ 4], uint8_t *pix1, uint8_t *pix2 );
67
#define x264_sub8x8_dct_dc_sse2 x264_template(sub8x8_dct_dc_sse2)
49
68
void x264_sub8x8_dct_dc_sse2 ( dctcoef dct [ 4], pixel *pix1, pixel *pix2 );
69
#define x264_sub8x8_dct_dc_avx512 x264_template(sub8x8_dct_dc_avx512)
50
70
void x264_sub8x8_dct_dc_avx512 ( int16_t dct [ 4], uint8_t *pix1, uint8_t *pix2 );
71
#define x264_sub8x16_dct_dc_sse2 x264_template(sub8x16_dct_dc_sse2)
51
72
void x264_sub8x16_dct_dc_sse2 ( dctcoef dct [ 8], pixel *pix1, pixel *pix2 );
73
#define x264_sub8x16_dct_dc_ssse3 x264_template(sub8x16_dct_dc_ssse3)
52
74
void x264_sub8x16_dct_dc_ssse3 ( int16_t dct [ 8], uint8_t *pix1, uint8_t *pix2 );
75
#define x264_sub8x16_dct_dc_avx x264_template(sub8x16_dct_dc_avx)
53
76
void x264_sub8x16_dct_dc_avx ( dctcoef dct [ 8], pixel *pix1, pixel *pix2 );
77
#define x264_sub8x16_dct_dc_avx512 x264_template(sub8x16_dct_dc_avx512)
54
78
void x264_sub8x16_dct_dc_avx512( int16_t dct [ 8], uint8_t *pix1, uint8_t *pix2 );
80
#define x264_add4x4_idct_mmx x264_template(add4x4_idct_mmx)
56
81
void x264_add4x4_idct_mmx ( uint8_t *p_dst, int16_t dct [16] );
82
#define x264_add4x4_idct_sse2 x264_template(add4x4_idct_sse2)
57
83
void x264_add4x4_idct_sse2 ( uint16_t *p_dst, int32_t dct [16] );
84
#define x264_add4x4_idct_sse4 x264_template(add4x4_idct_sse4)
58
85
void x264_add4x4_idct_sse4 ( uint8_t *p_dst, int16_t dct [16] );
86
#define x264_add4x4_idct_avx x264_template(add4x4_idct_avx)
59
87
void x264_add4x4_idct_avx ( pixel *p_dst, dctcoef dct [16] );
88
#define x264_add8x8_idct_mmx x264_template(add8x8_idct_mmx)
60
89
void x264_add8x8_idct_mmx ( uint8_t *p_dst, int16_t dct[ 4][16] );
90
#define x264_add8x8_idct_dc_mmx2 x264_template(add8x8_idct_dc_mmx2)
61
91
void x264_add8x8_idct_dc_mmx2 ( uint8_t *p_dst, int16_t dct [ 4] );
92
#define x264_add16x16_idct_mmx x264_template(add16x16_idct_mmx)
62
93
void x264_add16x16_idct_mmx ( uint8_t *p_dst, int16_t dct[16][16] );
94
#define x264_add16x16_idct_dc_mmx2 x264_template(add16x16_idct_dc_mmx2)
63
95
void x264_add16x16_idct_dc_mmx2 ( uint8_t *p_dst, int16_t dct [16] );
96
#define x264_add8x8_idct_sse2 x264_template(add8x8_idct_sse2)
64
97
void x264_add8x8_idct_sse2 ( pixel *p_dst, dctcoef dct[ 4][16] );
98
#define x264_add8x8_idct_avx x264_template(add8x8_idct_avx)
65
99
void x264_add8x8_idct_avx ( pixel *p_dst, dctcoef dct[ 4][16] );
100
#define x264_add8x8_idct_avx2 x264_template(add8x8_idct_avx2)
66
101
void x264_add8x8_idct_avx2 ( pixel *p_dst, dctcoef dct[ 4][16] );
102
#define x264_add8x8_idct_avx512 x264_template(add8x8_idct_avx512)
67
103
void x264_add8x8_idct_avx512 ( uint8_t *p_dst, int16_t dct[ 4][16] );
104
#define x264_add16x16_idct_sse2 x264_template(add16x16_idct_sse2)
68
105
void x264_add16x16_idct_sse2 ( pixel *p_dst, dctcoef dct[16][16] );
106
#define x264_add16x16_idct_avx x264_template(add16x16_idct_avx)
69
107
void x264_add16x16_idct_avx ( pixel *p_dst, dctcoef dct[16][16] );
108
#define x264_add16x16_idct_avx2 x264_template(add16x16_idct_avx2)
70
109
void x264_add16x16_idct_avx2 ( pixel *p_dst, dctcoef dct[16][16] );
110
#define x264_add8x8_idct_dc_sse2 x264_template(add8x8_idct_dc_sse2)
71
111
void x264_add8x8_idct_dc_sse2 ( pixel *p_dst, dctcoef dct [ 4] );
112
#define x264_add16x16_idct_dc_sse2 x264_template(add16x16_idct_dc_sse2)
72
113
void x264_add16x16_idct_dc_sse2 ( pixel *p_dst, dctcoef dct [16] );
114
#define x264_add8x8_idct_dc_ssse3 x264_template(add8x8_idct_dc_ssse3)
73
115
void x264_add8x8_idct_dc_ssse3 ( uint8_t *p_dst, int16_t dct [ 4] );
116
#define x264_add16x16_idct_dc_ssse3 x264_template(add16x16_idct_dc_ssse3)
74
117
void x264_add16x16_idct_dc_ssse3( uint8_t *p_dst, int16_t dct [16] );
118
#define x264_add8x8_idct_dc_avx x264_template(add8x8_idct_dc_avx)
75
119
void x264_add8x8_idct_dc_avx ( pixel *p_dst, dctcoef dct [ 4] );
120
#define x264_add16x16_idct_dc_avx x264_template(add16x16_idct_dc_avx)
76
121
void x264_add16x16_idct_dc_avx ( pixel *p_dst, dctcoef dct [16] );
122
#define x264_add16x16_idct_dc_avx2 x264_template(add16x16_idct_dc_avx2)
77
123
void x264_add16x16_idct_dc_avx2 ( uint8_t *p_dst, int16_t dct [16] );
125
#define x264_dct4x4dc_mmx2 x264_template(dct4x4dc_mmx2)
79
126
void x264_dct4x4dc_mmx2 ( int16_t d[16] );
127
#define x264_dct4x4dc_sse2 x264_template(dct4x4dc_sse2)
80
128
void x264_dct4x4dc_sse2 ( int32_t d[16] );
129
#define x264_dct4x4dc_avx x264_template(dct4x4dc_avx)
81
130
void x264_dct4x4dc_avx ( int32_t d[16] );
131
#define x264_idct4x4dc_mmx x264_template(idct4x4dc_mmx)
82
132
void x264_idct4x4dc_mmx ( int16_t d[16] );
133
#define x264_idct4x4dc_sse2 x264_template(idct4x4dc_sse2)
83
134
void x264_idct4x4dc_sse2 ( int32_t d[16] );
135
#define x264_idct4x4dc_avx x264_template(idct4x4dc_avx)
84
136
void x264_idct4x4dc_avx ( int32_t d[16] );
138
#define x264_dct2x4dc_mmx2 x264_template(dct2x4dc_mmx2)
86
139
void x264_dct2x4dc_mmx2( dctcoef dct[8], dctcoef dct4x4[8][16] );
140
#define x264_dct2x4dc_sse2 x264_template(dct2x4dc_sse2)
87
141
void x264_dct2x4dc_sse2( dctcoef dct[8], dctcoef dct4x4[8][16] );
142
#define x264_dct2x4dc_avx x264_template(dct2x4dc_avx)
88
143
void x264_dct2x4dc_avx ( dctcoef dct[8], dctcoef dct4x4[8][16] );
145
#define x264_sub8x8_dct8_mmx x264_template(sub8x8_dct8_mmx)
90
146
void x264_sub8x8_dct8_mmx ( int16_t dct [64], uint8_t *pix1, uint8_t *pix2 );
147
#define x264_sub16x16_dct8_mmx x264_template(sub16x16_dct8_mmx)
91
148
void x264_sub16x16_dct8_mmx ( int16_t dct[4][64], uint8_t *pix1, uint8_t *pix2 );
149
#define x264_sub8x8_dct8_sse2 x264_template(sub8x8_dct8_sse2)
92
150
void x264_sub8x8_dct8_sse2 ( dctcoef dct [64], pixel *pix1, pixel *pix2 );
151
#define x264_sub16x16_dct8_sse2 x264_template(sub16x16_dct8_sse2)
93
152
void x264_sub16x16_dct8_sse2 ( dctcoef dct[4][64], pixel *pix1, pixel *pix2 );
153
#define x264_sub8x8_dct8_ssse3 x264_template(sub8x8_dct8_ssse3)
94
154
void x264_sub8x8_dct8_ssse3 ( int16_t dct [64], uint8_t *pix1, uint8_t *pix2 );
155
#define x264_sub16x16_dct8_ssse3 x264_template(sub16x16_dct8_ssse3)
95
156
void x264_sub16x16_dct8_ssse3( int16_t dct[4][64], uint8_t *pix1, uint8_t *pix2 );
157
#define x264_sub8x8_dct8_sse4 x264_template(sub8x8_dct8_sse4)
96
158
void x264_sub8x8_dct8_sse4 ( int32_t dct [64], uint16_t *pix1, uint16_t *pix2 );
159
#define x264_sub16x16_dct8_sse4 x264_template(sub16x16_dct8_sse4)
97
160
void x264_sub16x16_dct8_sse4 ( int32_t dct[4][64], uint16_t *pix1, uint16_t *pix2 );
161
#define x264_sub8x8_dct8_avx x264_template(sub8x8_dct8_avx)
98
162
void x264_sub8x8_dct8_avx ( dctcoef dct [64], pixel *pix1, pixel *pix2 );
163
#define x264_sub16x16_dct8_avx x264_template(sub16x16_dct8_avx)
99
164
void x264_sub16x16_dct8_avx ( dctcoef dct[4][64], pixel *pix1, pixel *pix2 );
165
#define x264_sub16x16_dct8_avx2 x264_template(sub16x16_dct8_avx2)
100
166
void x264_sub16x16_dct8_avx2 ( dctcoef dct[4][64], pixel *pix1, pixel *pix2 );
169
#define x264_add8x8_idct8_mmx x264_template(add8x8_idct8_mmx)
103
170
void x264_add8x8_idct8_mmx ( uint8_t *dst, int16_t dct [64] );
171
#define x264_add16x16_idct8_mmx x264_template(add16x16_idct8_mmx)
104
172
void x264_add16x16_idct8_mmx ( uint8_t *dst, int16_t dct[4][64] );
173
#define x264_add8x8_idct8_sse2 x264_template(add8x8_idct8_sse2)
105
174
void x264_add8x8_idct8_sse2 ( pixel *dst, dctcoef dct [64] );
175
#define x264_add16x16_idct8_sse2 x264_template(add16x16_idct8_sse2)
106
176
void x264_add16x16_idct8_sse2( pixel *dst, dctcoef dct[4][64] );
177
#define x264_add8x8_idct8_avx x264_template(add8x8_idct8_avx)
107
178
void x264_add8x8_idct8_avx ( pixel *dst, dctcoef dct [64] );
179
#define x264_add16x16_idct8_avx x264_template(add16x16_idct8_avx)
108
180
void x264_add16x16_idct8_avx ( pixel *dst, dctcoef dct[4][64] );
182
#define x264_zigzag_scan_8x8_frame_mmx2 x264_template(zigzag_scan_8x8_frame_mmx2)
110
183
void x264_zigzag_scan_8x8_frame_mmx2 ( int16_t level[64], int16_t dct[64] );
184
#define x264_zigzag_scan_8x8_frame_sse2 x264_template(zigzag_scan_8x8_frame_sse2)
111
185
void x264_zigzag_scan_8x8_frame_sse2 ( dctcoef level[64], dctcoef dct[64] );
186
#define x264_zigzag_scan_8x8_frame_ssse3 x264_template(zigzag_scan_8x8_frame_ssse3)
112
187
void x264_zigzag_scan_8x8_frame_ssse3 ( int16_t level[64], int16_t dct[64] );
188
#define x264_zigzag_scan_8x8_frame_avx x264_template(zigzag_scan_8x8_frame_avx)
113
189
void x264_zigzag_scan_8x8_frame_avx ( dctcoef level[64], dctcoef dct[64] );
190
#define x264_zigzag_scan_8x8_frame_xop x264_template(zigzag_scan_8x8_frame_xop)
114
191
void x264_zigzag_scan_8x8_frame_xop ( int16_t level[64], int16_t dct[64] );
192
#define x264_zigzag_scan_8x8_frame_avx512 x264_template(zigzag_scan_8x8_frame_avx512)
115
193
void x264_zigzag_scan_8x8_frame_avx512( dctcoef level[64], dctcoef dct[64] );
194
#define x264_zigzag_scan_4x4_frame_mmx x264_template(zigzag_scan_4x4_frame_mmx)
116
195
void x264_zigzag_scan_4x4_frame_mmx ( int16_t level[16], int16_t dct[16] );
196
#define x264_zigzag_scan_4x4_frame_sse2 x264_template(zigzag_scan_4x4_frame_sse2)
117
197
void x264_zigzag_scan_4x4_frame_sse2 ( int32_t level[16], int32_t dct[16] );
198
#define x264_zigzag_scan_4x4_frame_ssse3 x264_template(zigzag_scan_4x4_frame_ssse3)
118
199
void x264_zigzag_scan_4x4_frame_ssse3 ( int16_t level[16], int16_t dct[16] );
200
#define x264_zigzag_scan_4x4_frame_avx x264_template(zigzag_scan_4x4_frame_avx)
119
201
void x264_zigzag_scan_4x4_frame_avx ( dctcoef level[16], dctcoef dct[16] );
202
#define x264_zigzag_scan_4x4_frame_xop x264_template(zigzag_scan_4x4_frame_xop)
120
203
void x264_zigzag_scan_4x4_frame_xop ( dctcoef level[16], dctcoef dct[16] );
204
#define x264_zigzag_scan_4x4_frame_avx512 x264_template(zigzag_scan_4x4_frame_avx512)
121
205
void x264_zigzag_scan_4x4_frame_avx512( dctcoef level[16], dctcoef dct[16] );
206
#define x264_zigzag_scan_4x4_field_sse x264_template(zigzag_scan_4x4_field_sse)
122
207
void x264_zigzag_scan_4x4_field_sse ( int16_t level[16], int16_t dct[16] );
208
#define x264_zigzag_scan_4x4_field_sse2 x264_template(zigzag_scan_4x4_field_sse2)
123
209
void x264_zigzag_scan_4x4_field_sse2 ( int32_t level[16], int32_t dct[16] );
210
#define x264_zigzag_scan_4x4_field_avx512 x264_template(zigzag_scan_4x4_field_avx512)
124
211
void x264_zigzag_scan_4x4_field_avx512( dctcoef level[16], dctcoef dct[16] );
212
#define x264_zigzag_scan_8x8_field_mmx2 x264_template(zigzag_scan_8x8_field_mmx2)
125
213
void x264_zigzag_scan_8x8_field_mmx2 ( int16_t level[64], int16_t dct[64] );
214
#define x264_zigzag_scan_8x8_field_sse4 x264_template(zigzag_scan_8x8_field_sse4)
126
215
void x264_zigzag_scan_8x8_field_sse4 ( int32_t level[64], int32_t dct[64] );
216
#define x264_zigzag_scan_8x8_field_avx x264_template(zigzag_scan_8x8_field_avx)
127
217
void x264_zigzag_scan_8x8_field_avx ( int32_t level[64], int32_t dct[64] );
218
#define x264_zigzag_scan_8x8_field_xop x264_template(zigzag_scan_8x8_field_xop)
128
219
void x264_zigzag_scan_8x8_field_xop ( int16_t level[64], int16_t dct[64] );
220
#define x264_zigzag_scan_8x8_field_avx512 x264_template(zigzag_scan_8x8_field_avx512)
129
221
void x264_zigzag_scan_8x8_field_avx512( dctcoef level[64], dctcoef dct[64] );
222
#define x264_zigzag_sub_4x4_frame_avx x264_template(zigzag_sub_4x4_frame_avx)
130
223
int x264_zigzag_sub_4x4_frame_avx ( int16_t level[16], const uint8_t *src, uint8_t *dst );
224
#define x264_zigzag_sub_4x4_frame_ssse3 x264_template(zigzag_sub_4x4_frame_ssse3)
131
225
int x264_zigzag_sub_4x4_frame_ssse3 ( int16_t level[16], const uint8_t *src, uint8_t *dst );
226
#define x264_zigzag_sub_4x4ac_frame_avx x264_template(zigzag_sub_4x4ac_frame_avx)
132
227
int x264_zigzag_sub_4x4ac_frame_avx ( int16_t level[16], const uint8_t *src, uint8_t *dst, int16_t *dc );
228
#define x264_zigzag_sub_4x4ac_frame_ssse3 x264_template(zigzag_sub_4x4ac_frame_ssse3)
133
229
int x264_zigzag_sub_4x4ac_frame_ssse3( int16_t level[16], const uint8_t *src, uint8_t *dst, int16_t *dc );
230
#define x264_zigzag_sub_4x4_field_avx x264_template(zigzag_sub_4x4_field_avx)
134
231
int x264_zigzag_sub_4x4_field_avx ( int16_t level[16], const uint8_t *src, uint8_t *dst );
232
#define x264_zigzag_sub_4x4_field_ssse3 x264_template(zigzag_sub_4x4_field_ssse3)
135
233
int x264_zigzag_sub_4x4_field_ssse3 ( int16_t level[16], const uint8_t *src, uint8_t *dst );
234
#define x264_zigzag_sub_4x4ac_field_avx x264_template(zigzag_sub_4x4ac_field_avx)
136
235
int x264_zigzag_sub_4x4ac_field_avx ( int16_t level[16], const uint8_t *src, uint8_t *dst, int16_t *dc );
236
#define x264_zigzag_sub_4x4ac_field_ssse3 x264_template(zigzag_sub_4x4ac_field_ssse3)
137
237
int x264_zigzag_sub_4x4ac_field_ssse3( int16_t level[16], const uint8_t *src, uint8_t *dst, int16_t *dc );
238
#define x264_zigzag_interleave_8x8_cavlc_mmx x264_template(zigzag_interleave_8x8_cavlc_mmx)
138
239
void x264_zigzag_interleave_8x8_cavlc_mmx ( int16_t *dst, int16_t *src, uint8_t *nnz );
240
#define x264_zigzag_interleave_8x8_cavlc_sse2 x264_template(zigzag_interleave_8x8_cavlc_sse2)
139
241
void x264_zigzag_interleave_8x8_cavlc_sse2 ( dctcoef *dst, dctcoef *src, uint8_t *nnz );
242
#define x264_zigzag_interleave_8x8_cavlc_avx x264_template(zigzag_interleave_8x8_cavlc_avx)
140
243
void x264_zigzag_interleave_8x8_cavlc_avx ( dctcoef *dst, dctcoef *src, uint8_t *nnz );
244
#define x264_zigzag_interleave_8x8_cavlc_avx2 x264_template(zigzag_interleave_8x8_cavlc_avx2)
141
245
void x264_zigzag_interleave_8x8_cavlc_avx2 ( int16_t *dst, int16_t *src, uint8_t *nnz );
246
#define x264_zigzag_interleave_8x8_cavlc_avx512 x264_template(zigzag_interleave_8x8_cavlc_avx512)
142
247
void x264_zigzag_interleave_8x8_cavlc_avx512( dctcoef *dst, dctcoef *src, uint8_t *nnz );