ffmpeg / libavcodec / ppc / dsputil_ppc.c @ 8dbe5856
History | View | Annotate | Download (6.36 KB)
1 |
/*
|
---|---|
2 |
* Copyright (c) 2002 Brian Foley
|
3 |
* Copyright (c) 2002 Dieter Shirley
|
4 |
* Copyright (c) 2003-2004 Romain Dolbeau <romain@dolbeau.org>
|
5 |
*
|
6 |
* This file is part of FFmpeg.
|
7 |
*
|
8 |
* FFmpeg is free software; you can redistribute it and/or
|
9 |
* modify it under the terms of the GNU Lesser General Public
|
10 |
* License as published by the Free Software Foundation; either
|
11 |
* version 2.1 of the License, or (at your option) any later version.
|
12 |
*
|
13 |
* FFmpeg is distributed in the hope that it will be useful,
|
14 |
* but WITHOUT ANY WARRANTY; without even the implied warranty of
|
15 |
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
|
16 |
* Lesser General Public License for more details.
|
17 |
*
|
18 |
* You should have received a copy of the GNU Lesser General Public
|
19 |
* License along with FFmpeg; if not, write to the Free Software
|
20 |
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
|
21 |
*/
|
22 |
|
23 |
#include "libavutil/cpu.h" |
24 |
#include "libavcodec/dsputil.h" |
25 |
#include "dsputil_altivec.h" |
26 |
|
27 |
/* ***** WARNING ***** WARNING ***** WARNING ***** */
|
28 |
/*
|
29 |
clear_blocks_dcbz32_ppc will not work properly on PowerPC processors with a
|
30 |
cache line size not equal to 32 bytes.
|
31 |
Fortunately all processor used by Apple up to at least the 7450 (aka second
|
32 |
generation G4) use 32 bytes cache line.
|
33 |
This is due to the use of the 'dcbz' instruction. It simply clear to zero a
|
34 |
single cache line, so you need to know the cache line size to use it !
|
35 |
It's absurd, but it's fast...
|
36 |
|
37 |
update 24/06/2003 : Apple released yesterday the G5, with a PPC970. cache line
|
38 |
size: 128 bytes. Oups.
|
39 |
The semantic of dcbz was changed, it always clear 32 bytes. so the function
|
40 |
below will work, but will be slow. So I fixed check_dcbz_effect to use dcbzl,
|
41 |
which is defined to clear a cache line (as dcbz before). So we still can
|
42 |
distinguish, and use dcbz (32 bytes) or dcbzl (one cache line) as required.
|
43 |
|
44 |
see <http://developer.apple.com/technotes/tn/tn2087.html>
|
45 |
and <http://developer.apple.com/technotes/tn/tn2086.html>
|
46 |
*/
|
47 |
static void clear_blocks_dcbz32_ppc(DCTELEM *blocks) |
48 |
{ |
49 |
register int misal = ((unsigned long)blocks & 0x00000010); |
50 |
register int i = 0; |
51 |
#if 1 |
52 |
if (misal) {
|
53 |
((unsigned long*)blocks)[0] = 0L; |
54 |
((unsigned long*)blocks)[1] = 0L; |
55 |
((unsigned long*)blocks)[2] = 0L; |
56 |
((unsigned long*)blocks)[3] = 0L; |
57 |
i += 16;
|
58 |
} |
59 |
for ( ; i < sizeof(DCTELEM)*6*64-31 ; i += 32) { |
60 |
__asm__ volatile("dcbz %0,%1" : : "b" (blocks), "r" (i) : "memory"); |
61 |
} |
62 |
if (misal) {
|
63 |
((unsigned long*)blocks)[188] = 0L; |
64 |
((unsigned long*)blocks)[189] = 0L; |
65 |
((unsigned long*)blocks)[190] = 0L; |
66 |
((unsigned long*)blocks)[191] = 0L; |
67 |
i += 16;
|
68 |
} |
69 |
#else
|
70 |
memset(blocks, 0, sizeof(DCTELEM)*6*64); |
71 |
#endif
|
72 |
} |
73 |
|
74 |
/* same as above, when dcbzl clear a whole 128B cache line
|
75 |
i.e. the PPC970 aka G5 */
|
76 |
#if HAVE_DCBZL
|
77 |
static void clear_blocks_dcbz128_ppc(DCTELEM *blocks) |
78 |
{ |
79 |
register int misal = ((unsigned long)blocks & 0x0000007f); |
80 |
register int i = 0; |
81 |
#if 1 |
82 |
if (misal) {
|
83 |
// we could probably also optimize this case,
|
84 |
// but there's not much point as the machines
|
85 |
// aren't available yet (2003-06-26)
|
86 |
memset(blocks, 0, sizeof(DCTELEM)*6*64); |
87 |
} |
88 |
else
|
89 |
for ( ; i < sizeof(DCTELEM)*6*64 ; i += 128) { |
90 |
__asm__ volatile("dcbzl %0,%1" : : "b" (blocks), "r" (i) : "memory"); |
91 |
} |
92 |
#else
|
93 |
memset(blocks, 0, sizeof(DCTELEM)*6*64); |
94 |
#endif
|
95 |
} |
96 |
#else
|
97 |
static void clear_blocks_dcbz128_ppc(DCTELEM *blocks) |
98 |
{ |
99 |
memset(blocks, 0, sizeof(DCTELEM)*6*64); |
100 |
} |
101 |
#endif
|
102 |
|
103 |
#if HAVE_DCBZL
|
104 |
/* check dcbz report how many bytes are set to 0 by dcbz */
|
105 |
/* update 24/06/2003 : replace dcbz by dcbzl to get
|
106 |
the intended effect (Apple "fixed" dcbz)
|
107 |
unfortunately this cannot be used unless the assembler
|
108 |
knows about dcbzl ... */
|
109 |
static long check_dcbzl_effect(void) |
110 |
{ |
111 |
register char *fakedata = av_malloc(1024); |
112 |
register char *fakedata_middle; |
113 |
register long zero = 0; |
114 |
register long i = 0; |
115 |
long count = 0; |
116 |
|
117 |
if (!fakedata) {
|
118 |
return 0L; |
119 |
} |
120 |
|
121 |
fakedata_middle = (fakedata + 512);
|
122 |
|
123 |
memset(fakedata, 0xFF, 1024); |
124 |
|
125 |
/* below the constraint "b" seems to mean "Address base register"
|
126 |
in gcc-3.3 / RS/6000 speaks. seems to avoid using r0, so.... */
|
127 |
__asm__ volatile("dcbzl %0, %1" : : "b" (fakedata_middle), "r" (zero)); |
128 |
|
129 |
for (i = 0; i < 1024 ; i ++) { |
130 |
if (fakedata[i] == (char)0) |
131 |
count++; |
132 |
} |
133 |
|
134 |
av_free(fakedata); |
135 |
|
136 |
return count;
|
137 |
} |
138 |
#else
|
139 |
static long check_dcbzl_effect(void) |
140 |
{ |
141 |
return 0; |
142 |
} |
143 |
#endif
|
144 |
|
145 |
static void prefetch_ppc(void *mem, int stride, int h) |
146 |
{ |
147 |
register const uint8_t *p = mem; |
148 |
do {
|
149 |
__asm__ volatile ("dcbt 0,%0" : : "r" (p)); |
150 |
p+= stride; |
151 |
} while(--h);
|
152 |
} |
153 |
|
154 |
void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx)
|
155 |
{ |
156 |
const int h264_high_depth = avctx->codec_id == CODEC_ID_H264 && avctx->bits_per_raw_sample > 8; |
157 |
|
158 |
// Common optimizations whether AltiVec is available or not
|
159 |
c->prefetch = prefetch_ppc; |
160 |
if (!h264_high_depth) {
|
161 |
switch (check_dcbzl_effect()) {
|
162 |
case 32: |
163 |
c->clear_blocks = clear_blocks_dcbz32_ppc; |
164 |
break;
|
165 |
case 128: |
166 |
c->clear_blocks = clear_blocks_dcbz128_ppc; |
167 |
break;
|
168 |
default:
|
169 |
break;
|
170 |
} |
171 |
} |
172 |
|
173 |
#if HAVE_ALTIVEC
|
174 |
if(CONFIG_H264_DECODER) dsputil_h264_init_ppc(c, avctx);
|
175 |
|
176 |
if (av_get_cpu_flags() & AV_CPU_FLAG_ALTIVEC) {
|
177 |
dsputil_init_altivec(c, avctx); |
178 |
float_init_altivec(c, avctx); |
179 |
int_init_altivec(c, avctx); |
180 |
c->gmc1 = gmc1_altivec; |
181 |
|
182 |
#if CONFIG_ENCODERS
|
183 |
if (avctx->dct_algo == FF_DCT_AUTO ||
|
184 |
avctx->dct_algo == FF_DCT_ALTIVEC) { |
185 |
c->fdct = fdct_altivec; |
186 |
} |
187 |
#endif //CONFIG_ENCODERS |
188 |
|
189 |
if (avctx->lowres==0) { |
190 |
if ((avctx->idct_algo == FF_IDCT_AUTO) ||
|
191 |
(avctx->idct_algo == FF_IDCT_ALTIVEC)) { |
192 |
c->idct_put = idct_put_altivec; |
193 |
c->idct_add = idct_add_altivec; |
194 |
c->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM; |
195 |
}else if((CONFIG_VP3_DECODER || CONFIG_VP5_DECODER || CONFIG_VP6_DECODER) && |
196 |
avctx->idct_algo==FF_IDCT_VP3){ |
197 |
c->idct_put = ff_vp3_idct_put_altivec; |
198 |
c->idct_add = ff_vp3_idct_add_altivec; |
199 |
c->idct = ff_vp3_idct_altivec; |
200 |
c->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM; |
201 |
} |
202 |
} |
203 |
|
204 |
} |
205 |
#endif /* HAVE_ALTIVEC */ |
206 |
} |