ffmpeg / libavcodec / ppc / dsputil_ppc.c @ ccf22d3e
History | View | Annotate | Download (6.27 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 "libavcodec/dsputil.h" |
24 |
#include "dsputil_altivec.h" |
25 |
|
26 |
/* ***** WARNING ***** WARNING ***** WARNING ***** */
|
27 |
/*
|
28 |
clear_blocks_dcbz32_ppc will not work properly on PowerPC processors with a
|
29 |
cache line size not equal to 32 bytes.
|
30 |
Fortunately all processor used by Apple up to at least the 7450 (aka second
|
31 |
generation G4) use 32 bytes cache line.
|
32 |
This is due to the use of the 'dcbz' instruction. It simply clear to zero a
|
33 |
single cache line, so you need to know the cache line size to use it !
|
34 |
It's absurd, but it's fast...
|
35 |
|
36 |
update 24/06/2003 : Apple released yesterday the G5, with a PPC970. cache line
|
37 |
size: 128 bytes. Oups.
|
38 |
The semantic of dcbz was changed, it always clear 32 bytes. so the function
|
39 |
below will work, but will be slow. So I fixed check_dcbz_effect to use dcbzl,
|
40 |
which is defined to clear a cache line (as dcbz before). So we still can
|
41 |
distinguish, and use dcbz (32 bytes) or dcbzl (one cache line) as required.
|
42 |
|
43 |
see <http://developer.apple.com/technotes/tn/tn2087.html>
|
44 |
and <http://developer.apple.com/technotes/tn/tn2086.html>
|
45 |
*/
|
46 |
static void clear_blocks_dcbz32_ppc(DCTELEM *blocks) |
47 |
{ |
48 |
register int misal = ((unsigned long)blocks & 0x00000010); |
49 |
register int i = 0; |
50 |
#if 1 |
51 |
if (misal) {
|
52 |
((unsigned long*)blocks)[0] = 0L; |
53 |
((unsigned long*)blocks)[1] = 0L; |
54 |
((unsigned long*)blocks)[2] = 0L; |
55 |
((unsigned long*)blocks)[3] = 0L; |
56 |
i += 16;
|
57 |
} |
58 |
for ( ; i < sizeof(DCTELEM)*6*64-31 ; i += 32) { |
59 |
__asm__ volatile("dcbz %0,%1" : : "b" (blocks), "r" (i) : "memory"); |
60 |
} |
61 |
if (misal) {
|
62 |
((unsigned long*)blocks)[188] = 0L; |
63 |
((unsigned long*)blocks)[189] = 0L; |
64 |
((unsigned long*)blocks)[190] = 0L; |
65 |
((unsigned long*)blocks)[191] = 0L; |
66 |
i += 16;
|
67 |
} |
68 |
#else
|
69 |
memset(blocks, 0, sizeof(DCTELEM)*6*64); |
70 |
#endif
|
71 |
} |
72 |
|
73 |
/* same as above, when dcbzl clear a whole 128B cache line
|
74 |
i.e. the PPC970 aka G5 */
|
75 |
#if HAVE_DCBZL
|
76 |
static void clear_blocks_dcbz128_ppc(DCTELEM *blocks) |
77 |
{ |
78 |
register int misal = ((unsigned long)blocks & 0x0000007f); |
79 |
register int i = 0; |
80 |
#if 1 |
81 |
if (misal) {
|
82 |
// we could probably also optimize this case,
|
83 |
// but there's not much point as the machines
|
84 |
// aren't available yet (2003-06-26)
|
85 |
memset(blocks, 0, sizeof(DCTELEM)*6*64); |
86 |
} |
87 |
else
|
88 |
for ( ; i < sizeof(DCTELEM)*6*64 ; i += 128) { |
89 |
__asm__ volatile("dcbzl %0,%1" : : "b" (blocks), "r" (i) : "memory"); |
90 |
} |
91 |
#else
|
92 |
memset(blocks, 0, sizeof(DCTELEM)*6*64); |
93 |
#endif
|
94 |
} |
95 |
#else
|
96 |
static void clear_blocks_dcbz128_ppc(DCTELEM *blocks) |
97 |
{ |
98 |
memset(blocks, 0, sizeof(DCTELEM)*6*64); |
99 |
} |
100 |
#endif
|
101 |
|
102 |
#if HAVE_DCBZL
|
103 |
/* check dcbz report how many bytes are set to 0 by dcbz */
|
104 |
/* update 24/06/2003 : replace dcbz by dcbzl to get
|
105 |
the intended effect (Apple "fixed" dcbz)
|
106 |
unfortunately this cannot be used unless the assembler
|
107 |
knows about dcbzl ... */
|
108 |
static long check_dcbzl_effect(void) |
109 |
{ |
110 |
register char *fakedata = av_malloc(1024); |
111 |
register char *fakedata_middle; |
112 |
register long zero = 0; |
113 |
register long i = 0; |
114 |
long count = 0; |
115 |
|
116 |
if (!fakedata) {
|
117 |
return 0L; |
118 |
} |
119 |
|
120 |
fakedata_middle = (fakedata + 512);
|
121 |
|
122 |
memset(fakedata, 0xFF, 1024); |
123 |
|
124 |
/* below the constraint "b" seems to mean "Address base register"
|
125 |
in gcc-3.3 / RS/6000 speaks. seems to avoid using r0, so.... */
|
126 |
__asm__ volatile("dcbzl %0, %1" : : "b" (fakedata_middle), "r" (zero)); |
127 |
|
128 |
for (i = 0; i < 1024 ; i ++) { |
129 |
if (fakedata[i] == (char)0) |
130 |
count++; |
131 |
} |
132 |
|
133 |
av_free(fakedata); |
134 |
|
135 |
return count;
|
136 |
} |
137 |
#else
|
138 |
static long check_dcbzl_effect(void) |
139 |
{ |
140 |
return 0; |
141 |
} |
142 |
#endif
|
143 |
|
144 |
static void prefetch_ppc(void *mem, int stride, int h) |
145 |
{ |
146 |
register const uint8_t *p = mem; |
147 |
do {
|
148 |
__asm__ volatile ("dcbt 0,%0" : : "r" (p)); |
149 |
p+= stride; |
150 |
} while(--h);
|
151 |
} |
152 |
|
153 |
void dsputil_init_ppc(DSPContext* c, AVCodecContext *avctx)
|
154 |
{ |
155 |
// Common optimizations whether AltiVec is available or not
|
156 |
c->prefetch = prefetch_ppc; |
157 |
switch (check_dcbzl_effect()) {
|
158 |
case 32: |
159 |
c->clear_blocks = clear_blocks_dcbz32_ppc; |
160 |
break;
|
161 |
case 128: |
162 |
c->clear_blocks = clear_blocks_dcbz128_ppc; |
163 |
break;
|
164 |
default:
|
165 |
break;
|
166 |
} |
167 |
|
168 |
#if HAVE_ALTIVEC
|
169 |
if(CONFIG_H264_DECODER) dsputil_h264_init_ppc(c, avctx);
|
170 |
|
171 |
if (mm_support() & AV_CPU_FLAG_ALTIVEC) {
|
172 |
dsputil_init_altivec(c, avctx); |
173 |
if(CONFIG_VC1_DECODER)
|
174 |
vc1dsp_init_altivec(c, avctx); |
175 |
float_init_altivec(c, avctx); |
176 |
int_init_altivec(c, avctx); |
177 |
c->gmc1 = gmc1_altivec; |
178 |
|
179 |
#if CONFIG_ENCODERS
|
180 |
if (avctx->dct_algo == FF_DCT_AUTO ||
|
181 |
avctx->dct_algo == FF_DCT_ALTIVEC) { |
182 |
c->fdct = fdct_altivec; |
183 |
} |
184 |
#endif //CONFIG_ENCODERS |
185 |
|
186 |
if (avctx->lowres==0) { |
187 |
if ((avctx->idct_algo == FF_IDCT_AUTO) ||
|
188 |
(avctx->idct_algo == FF_IDCT_ALTIVEC)) { |
189 |
c->idct_put = idct_put_altivec; |
190 |
c->idct_add = idct_add_altivec; |
191 |
c->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM; |
192 |
}else if((CONFIG_VP3_DECODER || CONFIG_VP5_DECODER || CONFIG_VP6_DECODER) && |
193 |
avctx->idct_algo==FF_IDCT_VP3){ |
194 |
c->idct_put = ff_vp3_idct_put_altivec; |
195 |
c->idct_add = ff_vp3_idct_add_altivec; |
196 |
c->idct = ff_vp3_idct_altivec; |
197 |
c->idct_permutation_type = FF_TRANSPOSE_IDCT_PERM; |
198 |
} |
199 |
} |
200 |
|
201 |
} |
202 |
#endif /* HAVE_ALTIVEC */ |
203 |
} |