1 /*****************************************************************************
2 * dct.c: h264 encoder library
3 *****************************************************************************
4 * Copyright (C) 2003 Laurent Aimar
7 * Authors: Eric Petit <titer@m0k.org>
9 * This program is free software; you can redistribute it and/or modify
10 * it under the terms of the GNU General Public License as published by
11 * the Free Software Foundation; either version 2 of the License, or
12 * (at your option) any later version.
14 * This program is distributed in the hope that it will be useful,
15 * but WITHOUT ANY WARRANTY; without even the implied warranty of
16 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
17 * GNU General Public License for more details.
19 * You should have received a copy of the GNU General Public License
20 * along with this program; if not, write to the Free Software
21 * Foundation, Inc., 59 Temple Place - Suite 330, Boston, MA 02111, USA.
22 *****************************************************************************/
28 #include "common/common.h"
29 #include "ppccommon.h"
31 #define VEC_DCT(a0,a1,a2,a3,b0,b1,b2,b3) \
32 b1 = vec_add( a0, a3 ); \
33 b3 = vec_add( a1, a2 ); \
34 b0 = vec_add( b1, b3 ); \
35 b2 = vec_sub( b1, b3 ); \
36 a0 = vec_sub( a0, a3 ); \
37 a1 = vec_sub( a1, a2 ); \
38 b1 = vec_add( a0, a0 ); \
39 b1 = vec_add( b1, a1 ); \
40 b3 = vec_sub( a0, a1 ); \
41 b3 = vec_sub( b3, a1 )
43 void x264_sub4x4_dct_altivec( int16_t dct[4][4],
44 uint8_t *pix1, int i_pix1, uint8_t *pix2, int i_pix2 )
48 vec_s16_t dct0v, dct1v, dct2v, dct3v;
49 vec_s16_t tmp0v, tmp1v, tmp2v, tmp3v;
51 VEC_DIFF_H( pix1, i_pix1, pix2, i_pix2, 4, dct0v );
52 VEC_DIFF_H( pix1, i_pix1, pix2, i_pix2, 4, dct1v );
53 VEC_DIFF_H( pix1, i_pix1, pix2, i_pix2, 4, dct2v );
54 VEC_DIFF_H( pix1, i_pix1, pix2, i_pix2, 4, dct3v );
55 VEC_DCT( dct0v, dct1v, dct2v, dct3v, tmp0v, tmp1v, tmp2v, tmp3v );
56 VEC_TRANSPOSE_4( tmp0v, tmp1v, tmp2v, tmp3v,
57 dct0v, dct1v, dct2v, dct3v );
58 VEC_DCT( dct0v, dct1v, dct2v, dct3v, tmp0v, tmp1v, tmp2v, tmp3v );
59 VEC_TRANSPOSE_4( tmp0v, tmp1v, tmp2v, tmp3v,
60 dct0v, dct1v, dct2v, dct3v );
61 VEC_STORE8( dct0v, dct[0] );
62 VEC_STORE8( dct1v, dct[1] );
63 VEC_STORE8( dct2v, dct[2] );
64 VEC_STORE8( dct3v, dct[3] );
67 void x264_sub8x8_dct_altivec( int16_t dct[4][4][4],
68 uint8_t *pix1, int i_pix1, uint8_t *pix2, int i_pix2 )
72 vec_s16_t dct0v, dct1v, dct2v, dct3v, dct4v, dct5v, dct6v, dct7v;
73 vec_s16_t tmp0v, tmp1v, tmp2v, tmp3v, tmp4v, tmp5v, tmp6v, tmp7v;
75 VEC_DIFF_H( pix1, i_pix1, pix2, i_pix2, 8, dct0v );
76 VEC_DIFF_H( pix1, i_pix1, pix2, i_pix2, 8, dct1v );
77 VEC_DIFF_H( pix1, i_pix1, pix2, i_pix2, 8, dct2v );
78 VEC_DIFF_H( pix1, i_pix1, pix2, i_pix2, 8, dct3v );
79 VEC_DIFF_H( pix1, i_pix1, pix2, i_pix2, 8, dct4v );
80 VEC_DIFF_H( pix1, i_pix1, pix2, i_pix2, 8, dct5v );
81 VEC_DIFF_H( pix1, i_pix1, pix2, i_pix2, 8, dct6v );
82 VEC_DIFF_H( pix1, i_pix1, pix2, i_pix2, 8, dct7v );
83 VEC_DCT( dct0v, dct1v, dct2v, dct3v, tmp0v, tmp1v, tmp2v, tmp3v );
84 VEC_DCT( dct4v, dct5v, dct6v, dct7v, tmp4v, tmp5v, tmp6v, tmp7v );
85 VEC_TRANSPOSE_8( tmp0v, tmp1v, tmp2v, tmp3v,
86 tmp4v, tmp5v, tmp6v, tmp7v,
87 dct0v, dct1v, dct2v, dct3v,
88 dct4v, dct5v, dct6v, dct7v );
89 VEC_DCT( dct0v, dct1v, dct2v, dct3v, tmp0v, tmp1v, tmp2v, tmp3v );
90 VEC_DCT( dct4v, dct5v, dct6v, dct7v, tmp4v, tmp5v, tmp6v, tmp7v );
91 VEC_TRANSPOSE_8( tmp0v, tmp1v, tmp2v, tmp3v,
92 tmp4v, tmp5v, tmp6v, tmp7v,
93 dct0v, dct1v, dct2v, dct3v,
94 dct4v, dct5v, dct6v, dct7v );
95 VEC_STORE8_H( dct0v, dct[0][0] );
96 VEC_STORE8_L( dct0v, dct[1][0] );
97 VEC_STORE8_H( dct1v, dct[0][1] );
98 VEC_STORE8_L( dct1v, dct[1][1] );
99 VEC_STORE8_H( dct2v, dct[0][2] );
100 VEC_STORE8_L( dct2v, dct[1][2] );
101 VEC_STORE8_H( dct3v, dct[0][3] );
102 VEC_STORE8_L( dct3v, dct[1][3] );
103 VEC_STORE8_H( dct4v, dct[2][0] );
104 VEC_STORE8_L( dct4v, dct[3][0] );
105 VEC_STORE8_H( dct5v, dct[2][1] );
106 VEC_STORE8_L( dct5v, dct[3][1] );
107 VEC_STORE8_H( dct6v, dct[2][2] );
108 VEC_STORE8_L( dct6v, dct[3][2] );
109 VEC_STORE8_H( dct7v, dct[2][3] );
110 VEC_STORE8_L( dct7v, dct[3][3] );
113 void x264_sub16x16_dct_altivec( int16_t dct[16][4][4],
114 uint8_t *pix1, int i_pix1, uint8_t *pix2, int i_pix2 )
118 vec_s16_t dcth0v, dcth1v, dcth2v, dcth3v,
119 dcth4v, dcth5v, dcth6v, dcth7v,
120 dctl0v, dctl1v, dctl2v, dctl3v,
121 dctl4v, dctl5v, dctl6v, dctl7v;
122 vec_s16_t temp0v, temp1v, temp2v, temp3v,
123 temp4v, temp5v, temp6v, temp7v;
125 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth0v, dctl0v );
126 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth1v, dctl1v );
127 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth2v, dctl2v );
128 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth3v, dctl3v );
129 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth4v, dctl4v );
130 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth5v, dctl5v );
131 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth6v, dctl6v );
132 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth7v, dctl7v );
134 VEC_DCT( dcth0v, dcth1v, dcth2v, dcth3v,
135 temp0v, temp1v, temp2v, temp3v );
136 VEC_DCT( dcth4v, dcth5v, dcth6v, dcth7v,
137 temp4v, temp5v, temp6v, temp7v );
138 VEC_TRANSPOSE_8( temp0v, temp1v, temp2v, temp3v,
139 temp4v, temp5v, temp6v, temp7v,
140 dcth0v, dcth1v, dcth2v, dcth3v,
141 dcth4v, dcth5v, dcth6v, dcth7v );
142 VEC_DCT( dcth0v, dcth1v, dcth2v, dcth3v,
143 temp0v, temp1v, temp2v, temp3v );
144 VEC_DCT( dcth4v, dcth5v, dcth6v, dcth7v,
145 temp4v, temp5v, temp6v, temp7v );
146 VEC_TRANSPOSE_8( temp0v, temp1v, temp2v, temp3v,
147 temp4v, temp5v, temp6v, temp7v,
148 dcth0v, dcth1v, dcth2v, dcth3v,
149 dcth4v, dcth5v, dcth6v, dcth7v );
150 VEC_STORE8_H( dcth0v, dct[0][0] );
151 VEC_STORE8_L( dcth0v, dct[1][0] );
152 VEC_STORE8_H( dcth1v, dct[0][1] );
153 VEC_STORE8_L( dcth1v, dct[1][1] );
154 VEC_STORE8_H( dcth2v, dct[0][2] );
155 VEC_STORE8_L( dcth2v, dct[1][2] );
156 VEC_STORE8_H( dcth3v, dct[0][3] );
157 VEC_STORE8_L( dcth3v, dct[1][3] );
158 VEC_STORE8_H( dcth4v, dct[2][0] );
159 VEC_STORE8_L( dcth4v, dct[3][0] );
160 VEC_STORE8_H( dcth5v, dct[2][1] );
161 VEC_STORE8_L( dcth5v, dct[3][1] );
162 VEC_STORE8_H( dcth6v, dct[2][2] );
163 VEC_STORE8_L( dcth6v, dct[3][2] );
164 VEC_STORE8_H( dcth7v, dct[2][3] );
165 VEC_STORE8_L( dcth7v, dct[3][3] );
167 VEC_DCT( dctl0v, dctl1v, dctl2v, dctl3v,
168 temp0v, temp1v, temp2v, temp3v );
169 VEC_DCT( dctl4v, dctl5v, dctl6v, dctl7v,
170 temp4v, temp5v, temp6v, temp7v );
171 VEC_TRANSPOSE_8( temp0v, temp1v, temp2v, temp3v,
172 temp4v, temp5v, temp6v, temp7v,
173 dctl0v, dctl1v, dctl2v, dctl3v,
174 dctl4v, dctl5v, dctl6v, dctl7v );
175 VEC_DCT( dctl0v, dctl1v, dctl2v, dctl3v,
176 temp0v, temp1v, temp2v, temp3v );
177 VEC_DCT( dctl4v, dctl5v, dctl6v, dctl7v,
178 temp4v, temp5v, temp6v, temp7v );
179 VEC_TRANSPOSE_8( temp0v, temp1v, temp2v, temp3v,
180 temp4v, temp5v, temp6v, temp7v,
181 dctl0v, dctl1v, dctl2v, dctl3v,
182 dctl4v, dctl5v, dctl6v, dctl7v );
183 VEC_STORE8_H( dctl0v, dct[4][0] );
184 VEC_STORE8_L( dctl0v, dct[5][0] );
185 VEC_STORE8_H( dctl1v, dct[4][1] );
186 VEC_STORE8_L( dctl1v, dct[5][1] );
187 VEC_STORE8_H( dctl2v, dct[4][2] );
188 VEC_STORE8_L( dctl2v, dct[5][2] );
189 VEC_STORE8_H( dctl3v, dct[4][3] );
190 VEC_STORE8_L( dctl3v, dct[5][3] );
191 VEC_STORE8_H( dctl4v, dct[6][0] );
192 VEC_STORE8_L( dctl4v, dct[7][0] );
193 VEC_STORE8_H( dctl5v, dct[6][1] );
194 VEC_STORE8_L( dctl5v, dct[7][1] );
195 VEC_STORE8_H( dctl6v, dct[6][2] );
196 VEC_STORE8_L( dctl6v, dct[7][2] );
197 VEC_STORE8_H( dctl7v, dct[6][3] );
198 VEC_STORE8_L( dctl7v, dct[7][3] );
200 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth0v, dctl0v );
201 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth1v, dctl1v );
202 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth2v, dctl2v );
203 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth3v, dctl3v );
204 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth4v, dctl4v );
205 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth5v, dctl5v );
206 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth6v, dctl6v );
207 VEC_DIFF_HL( pix1, i_pix1, pix2, i_pix2, dcth7v, dctl7v );
209 VEC_DCT( dcth0v, dcth1v, dcth2v, dcth3v,
210 temp0v, temp1v, temp2v, temp3v );
211 VEC_DCT( dcth4v, dcth5v, dcth6v, dcth7v,
212 temp4v, temp5v, temp6v, temp7v );
213 VEC_TRANSPOSE_8( temp0v, temp1v, temp2v, temp3v,
214 temp4v, temp5v, temp6v, temp7v,
215 dcth0v, dcth1v, dcth2v, dcth3v,
216 dcth4v, dcth5v, dcth6v, dcth7v );
217 VEC_DCT( dcth0v, dcth1v, dcth2v, dcth3v,
218 temp0v, temp1v, temp2v, temp3v );
219 VEC_DCT( dcth4v, dcth5v, dcth6v, dcth7v,
220 temp4v, temp5v, temp6v, temp7v );
221 VEC_TRANSPOSE_8( temp0v, temp1v, temp2v, temp3v,
222 temp4v, temp5v, temp6v, temp7v,
223 dcth0v, dcth1v, dcth2v, dcth3v,
224 dcth4v, dcth5v, dcth6v, dcth7v );
225 VEC_STORE8_H( dcth0v, dct[8][0] );
226 VEC_STORE8_L( dcth0v, dct[9][0] );
227 VEC_STORE8_H( dcth1v, dct[8][1] );
228 VEC_STORE8_L( dcth1v, dct[9][1] );
229 VEC_STORE8_H( dcth2v, dct[8][2] );
230 VEC_STORE8_L( dcth2v, dct[9][2] );
231 VEC_STORE8_H( dcth3v, dct[8][3] );
232 VEC_STORE8_L( dcth3v, dct[9][3] );
233 VEC_STORE8_H( dcth4v, dct[10][0] );
234 VEC_STORE8_L( dcth4v, dct[11][0] );
235 VEC_STORE8_H( dcth5v, dct[10][1] );
236 VEC_STORE8_L( dcth5v, dct[11][1] );
237 VEC_STORE8_H( dcth6v, dct[10][2] );
238 VEC_STORE8_L( dcth6v, dct[11][2] );
239 VEC_STORE8_H( dcth7v, dct[10][3] );
240 VEC_STORE8_L( dcth7v, dct[11][3] );
242 VEC_DCT( dctl0v, dctl1v, dctl2v, dctl3v,
243 temp0v, temp1v, temp2v, temp3v );
244 VEC_DCT( dctl4v, dctl5v, dctl6v, dctl7v,
245 temp4v, temp5v, temp6v, temp7v );
246 VEC_TRANSPOSE_8( temp0v, temp1v, temp2v, temp3v,
247 temp4v, temp5v, temp6v, temp7v,
248 dctl0v, dctl1v, dctl2v, dctl3v,
249 dctl4v, dctl5v, dctl6v, dctl7v );
250 VEC_DCT( dctl0v, dctl1v, dctl2v, dctl3v,
251 temp0v, temp1v, temp2v, temp3v );
252 VEC_DCT( dctl4v, dctl5v, dctl6v, dctl7v,
253 temp4v, temp5v, temp6v, temp7v );
254 VEC_TRANSPOSE_8( temp0v, temp1v, temp2v, temp3v,
255 temp4v, temp5v, temp6v, temp7v,
256 dctl0v, dctl1v, dctl2v, dctl3v,
257 dctl4v, dctl5v, dctl6v, dctl7v );
258 VEC_STORE8_H( dctl0v, dct[12][0] );
259 VEC_STORE8_L( dctl0v, dct[13][0] );
260 VEC_STORE8_H( dctl1v, dct[12][1] );
261 VEC_STORE8_L( dctl1v, dct[13][1] );
262 VEC_STORE8_H( dctl2v, dct[12][2] );
263 VEC_STORE8_L( dctl2v, dct[13][2] );
264 VEC_STORE8_H( dctl3v, dct[12][3] );
265 VEC_STORE8_L( dctl3v, dct[13][3] );
266 VEC_STORE8_H( dctl4v, dct[14][0] );
267 VEC_STORE8_L( dctl4v, dct[15][0] );
268 VEC_STORE8_H( dctl5v, dct[14][1] );
269 VEC_STORE8_L( dctl5v, dct[15][1] );
270 VEC_STORE8_H( dctl6v, dct[14][2] );
271 VEC_STORE8_L( dctl6v, dct[15][2] );
272 VEC_STORE8_H( dctl7v, dct[14][3] );
273 VEC_STORE8_L( dctl7v, dct[15][3] );