Altivec Patch (Mark III) by (Dieter Shirley <dieters at schemasoft dot com>)
[ffmpeg.git] / libavcodec / ppc / idct_altivec.c
1 /*
2  * Copyright (c) 2001 Michel Lespinasse
3  *
4  * This library is free software; you can redistribute it and/or
5  * modify it under the terms of the GNU Lesser General Public
6  * License as published by the Free Software Foundation; either
7  * version 2 of the License, or (at your option) any later version.
8  *
9  * This library is distributed in the hope that it will be useful,
10  * but WITHOUT ANY WARRANTY; without even the implied warranty of
11  * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
12  * Lesser General Public License for more details.
13  *
14  * You should have received a copy of the GNU Lesser General Public
15  * License along with this library; if not, write to the Free Software
16  * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA  02111-1307  USA
17  *
18  */
19
20 /*
21  * NOTE: This code is based on GPL code from the libmpeg2 project.  The
22  * author, Michel Lespinasses, has given explicit permission to release
23  * under LGPL as part of ffmpeg.
24  *
25  */
26
27 /*
28  * FFMpeg integration by Dieter Shirley
29  *
30  * This file is a direct copy of the altivec idct module from the libmpeg2
31  * project.  I've deleted all of the libmpeg2 specific code, renamed the functions and
32  * re-ordered the function parameters.  The only change to the IDCT function
33  * itself was to factor out the partial transposition, and to perform a full
34  * transpose at the end of the function.
35  */
36
37
38 #include <stdlib.h>                                      /* malloc(), free() */
39 #include <string.h>
40 #include "../dsputil.h"
41
42 #define vector_s16_t vector signed short
43 #define vector_u16_t vector unsigned short
44 #define vector_s8_t vector signed char
45 #define vector_u8_t vector unsigned char
46 #define vector_s32_t vector signed int
47 #define vector_u32_t vector unsigned int
48
49 #define IDCT_HALF                                       \
50     /* 1st stage */                                     \
51     t1 = vec_mradds (a1, vx7, vx1 );                    \
52     t8 = vec_mradds (a1, vx1, vec_subs (zero, vx7));    \
53     t7 = vec_mradds (a2, vx5, vx3);                     \
54     t3 = vec_mradds (ma2, vx3, vx5);                    \
55                                                         \
56     /* 2nd stage */                                     \
57     t5 = vec_adds (vx0, vx4);                           \
58     t0 = vec_subs (vx0, vx4);                           \
59     t2 = vec_mradds (a0, vx6, vx2);                     \
60     t4 = vec_mradds (a0, vx2, vec_subs (zero, vx6));    \
61     t6 = vec_adds (t8, t3);                             \
62     t3 = vec_subs (t8, t3);                             \
63     t8 = vec_subs (t1, t7);                             \
64     t1 = vec_adds (t1, t7);                             \
65                                                         \
66     /* 3rd stage */                                     \
67     t7 = vec_adds (t5, t2);                             \
68     t2 = vec_subs (t5, t2);                             \
69     t5 = vec_adds (t0, t4);                             \
70     t0 = vec_subs (t0, t4);                             \
71     t4 = vec_subs (t8, t3);                             \
72     t3 = vec_adds (t8, t3);                             \
73                                                         \
74     /* 4th stage */                                     \
75     vy0 = vec_adds (t7, t1);                            \
76     vy7 = vec_subs (t7, t1);                            \
77     vy1 = vec_mradds (c4, t3, t5);                      \
78     vy6 = vec_mradds (mc4, t3, t5);                     \
79     vy2 = vec_mradds (c4, t4, t0);                      \
80     vy5 = vec_mradds (mc4, t4, t0);                     \
81     vy3 = vec_adds (t2, t6);                            \
82     vy4 = vec_subs (t2, t6);
83
84         
85 #define IDCT                                                            \
86     vector_s16_t vx0, vx1, vx2, vx3, vx4, vx5, vx6, vx7;                \
87     vector_s16_t vy0, vy1, vy2, vy3, vy4, vy5, vy6, vy7;                \
88     vector_s16_t a0, a1, a2, ma2, c4, mc4, zero, bias;                  \
89     vector_s16_t t0, t1, t2, t3, t4, t5, t6, t7, t8;                    \
90     vector_u16_t shift;                                                 \
91                                                                         \
92     c4 = vec_splat (constants[0], 0);                                   \
93     a0 = vec_splat (constants[0], 1);                                   \
94     a1 = vec_splat (constants[0], 2);                                   \
95     a2 = vec_splat (constants[0], 3);                                   \
96     mc4 = vec_splat (constants[0], 4);                                  \
97     ma2 = vec_splat (constants[0], 5);                                  \
98     bias = (vector_s16_t)vec_splat ((vector_s32_t)constants[0], 3);     \
99                                                                         \
100     zero = vec_splat_s16 (0);                                           \
101     shift = vec_splat_u16 (4);                                          \
102                                                                         \
103     vx0 = vec_mradds (vec_sl (block[0], shift), constants[1], zero);    \
104     vx1 = vec_mradds (vec_sl (block[1], shift), constants[2], zero);    \
105     vx2 = vec_mradds (vec_sl (block[2], shift), constants[3], zero);    \
106     vx3 = vec_mradds (vec_sl (block[3], shift), constants[4], zero);    \
107     vx4 = vec_mradds (vec_sl (block[4], shift), constants[1], zero);    \
108     vx5 = vec_mradds (vec_sl (block[5], shift), constants[4], zero);    \
109     vx6 = vec_mradds (vec_sl (block[6], shift), constants[3], zero);    \
110     vx7 = vec_mradds (vec_sl (block[7], shift), constants[2], zero);    \
111                                                                         \
112     IDCT_HALF                                                           \
113                                                                         \
114     vx0 = vec_mergeh (vy0, vy4);                                        \
115     vx1 = vec_mergel (vy0, vy4);                                        \
116     vx2 = vec_mergeh (vy1, vy5);                                        \
117     vx3 = vec_mergel (vy1, vy5);                                        \
118     vx4 = vec_mergeh (vy2, vy6);                                        \
119     vx5 = vec_mergel (vy2, vy6);                                        \
120     vx6 = vec_mergeh (vy3, vy7);                                        \
121     vx7 = vec_mergel (vy3, vy7);                                        \
122                                                                         \
123     vy0 = vec_mergeh (vx0, vx4);                                        \
124     vy1 = vec_mergel (vx0, vx4);                                        \
125     vy2 = vec_mergeh (vx1, vx5);                                        \
126     vy3 = vec_mergel (vx1, vx5);                                        \
127     vy4 = vec_mergeh (vx2, vx6);                                        \
128     vy5 = vec_mergel (vx2, vx6);                                        \
129     vy6 = vec_mergeh (vx3, vx7);                                        \
130     vy7 = vec_mergel (vx3, vx7);                                        \
131                                                                         \
132     vx0 = vec_adds (vec_mergeh (vy0, vy4), bias);                       \
133     vx1 = vec_mergel (vy0, vy4);                                        \
134     vx2 = vec_mergeh (vy1, vy5);                                        \
135     vx3 = vec_mergel (vy1, vy5);                                        \
136     vx4 = vec_mergeh (vy2, vy6);                                        \
137     vx5 = vec_mergel (vy2, vy6);                                        \
138     vx6 = vec_mergeh (vy3, vy7);                                        \
139     vx7 = vec_mergel (vy3, vy7);                                        \
140                                                                         \
141     IDCT_HALF                                                           \
142                                                                         \
143     shift = vec_splat_u16 (6);                                          \
144     vx0 = vec_sra (vy0, shift);                                         \
145     vx1 = vec_sra (vy1, shift);                                         \
146     vx2 = vec_sra (vy2, shift);                                         \
147     vx3 = vec_sra (vy3, shift);                                         \
148     vx4 = vec_sra (vy4, shift);                                         \
149     vx5 = vec_sra (vy5, shift);                                         \
150     vx6 = vec_sra (vy6, shift);                                         \
151     vx7 = vec_sra (vy7, shift);
152
153 static const vector_s16_t constants[5] = {
154     (vector_s16_t)(23170, 13573, 6518, 21895, -23170, -21895, 32, 31),
155     (vector_s16_t)(16384, 22725, 21407, 19266, 16384, 19266, 21407, 22725),
156     (vector_s16_t)(22725, 31521, 29692, 26722, 22725, 26722, 29692, 31521),
157     (vector_s16_t)(21407, 29692, 27969, 25172, 21407, 25172, 27969, 29692),
158     (vector_s16_t)(19266, 26722, 25172, 22654, 19266, 22654, 25172, 26722)
159 };
160
161 void idct_put_altivec(uint8_t* dest, int stride, vector_s16_t* block)
162 {
163     vector_u8_t tmp;
164
165     IDCT
166
167 #define COPY(dest,src)                                          \
168     tmp = vec_packsu (src, src);                                \
169     vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest);       \
170     vec_ste ((vector_u32_t)tmp, 4, (unsigned int *)dest);
171
172     COPY (dest, vx0)    dest += stride;
173     COPY (dest, vx1)    dest += stride;
174     COPY (dest, vx2)    dest += stride;
175     COPY (dest, vx3)    dest += stride;
176     COPY (dest, vx4)    dest += stride;
177     COPY (dest, vx5)    dest += stride;
178     COPY (dest, vx6)    dest += stride;
179     COPY (dest, vx7)
180 }
181
182 void idct_add_altivec(uint8_t* dest, int stride, vector_s16_t* block)
183 {
184     vector_u8_t tmp;
185     vector_s16_t tmp2, tmp3;
186     vector_u8_t perm0;
187     vector_u8_t perm1;
188     vector_u8_t p0, p1, p;
189
190     IDCT
191
192     p0 = vec_lvsl (0, dest);
193     p1 = vec_lvsl (stride, dest);
194     p = vec_splat_u8 (-1);
195     perm0 = vec_mergeh (p, p0);
196     perm1 = vec_mergeh (p, p1);
197
198 #define ADD(dest,src,perm)                                              \
199     /* *(uint64_t *)&tmp = *(uint64_t *)dest; */                        \
200     tmp = vec_ld (0, dest);                                             \
201     tmp2 = (vector_s16_t)vec_perm (tmp, (vector_u8_t)zero, perm);       \
202     tmp3 = vec_adds (tmp2, src);                                        \
203     tmp = vec_packsu (tmp3, tmp3);                                      \
204     vec_ste ((vector_u32_t)tmp, 0, (unsigned int *)dest);               \
205     vec_ste ((vector_u32_t)tmp, 4, (unsigned int *)dest);
206
207     ADD (dest, vx0, perm0)      dest += stride;
208     ADD (dest, vx1, perm1)      dest += stride;
209     ADD (dest, vx2, perm0)      dest += stride;
210     ADD (dest, vx3, perm1)      dest += stride;
211     ADD (dest, vx4, perm0)      dest += stride;
212     ADD (dest, vx5, perm1)      dest += stride;
213     ADD (dest, vx6, perm0)      dest += stride;
214     ADD (dest, vx7, perm1)
215 }
216