annotate ps2/dsputil_mmi.c @ 1878:838c18d1e7fc libavcodec

be smarter about the variable size of the extra SVQ3 data
author melanson
date Sat, 13 Mar 2004 21:26:54 +0000
parents 72ac356803ea
children ef2149182f1c
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
1 /*
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
2 * MMI optimized DSP utils
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
3 * Copyright (c) 2000, 2001 Fabrice Bellard.
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
4 *
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
5 * This library is free software; you can redistribute it and/or
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
6 * modify it under the terms of the GNU Lesser General Public
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
7 * License as published by the Free Software Foundation; either
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
8 * version 2 of the License, or (at your option) any later version.
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
9 *
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
10 * This library is distributed in the hope that it will be useful,
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
11 * but WITHOUT ANY WARRANTY; without even the implied warranty of
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
12 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
13 * Lesser General Public License for more details.
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
14 *
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
15 * You should have received a copy of the GNU Lesser General Public
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
16 * License along with this library; if not, write to the Free Software
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
17 * Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307 USA
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
18 *
1876
72ac356803ea credit update
melanson
parents: 1324
diff changeset
19 * MMI optimization by Leon van Stuivenberg
1143
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
20 * clear_blocks_mmi() by BroadQ
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
21 */
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
22
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
23 #include "../dsputil.h"
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
24 #include "mmi.h"
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
25
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
26 void ff_mmi_idct_put(uint8_t *dest, int line_size, DCTELEM *block);
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
27 void ff_mmi_idct_add(uint8_t *dest, int line_size, DCTELEM *block);
1324
7d328fd9d8a5 the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents: 1146
diff changeset
28 void ff_mmi_idct(DCTELEM *block);
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
29
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
30 static void clear_blocks_mmi(DCTELEM * blocks)
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
31 {
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
32 asm volatile(
1143
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
33 ".set noreorder \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
34 "addiu $9, %0, 768 \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
35 "nop \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
36 "1: \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
37 "sq $0, 0(%0) \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
38 "move $8, %0 \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
39 "addi %0, %0, 64 \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
40 "sq $0, 16($8) \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
41 "slt $10, %0, $9 \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
42 "sq $0, 32($8) \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
43 "bnez $10, 1b \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
44 "sq $0, 48($8) \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
45 ".set reorder \n"
a4facfd78935 clear_blocks_mmi() optimization by BroadQ (and patch by Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1092
diff changeset
46 : "+r" (blocks) :: "$8", "$9", "memory" );
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
47 }
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
48
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
49
1064
b32afefe7d33 * UINTX -> uintx_t INTX -> intx_t
kabi
parents: 856
diff changeset
50 static void get_pixels_mmi(DCTELEM *block, const uint8_t *pixels, int line_size)
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
51 {
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
52 asm volatile(
772
311ac50375e4 dsputil mmi cleanup patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 721
diff changeset
53 ".set push \n\t"
311ac50375e4 dsputil mmi cleanup patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 721
diff changeset
54 ".set mips3 \n\t"
1146
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
55 "ld $8, 0(%0) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
56 "add %0, %0, %2 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
57 "ld $9, 0(%0) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
58 "add %0, %0, %2 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
59 "ld $10, 0(%0) \n\t"
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
60 "pextlb $8, $0, $8 \n\t"
1146
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
61 "sq $8, 0(%1) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
62 "add %0, %0, %2 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
63 "ld $8, 0(%0) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
64 "pextlb $9, $0, $9 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
65 "sq $9, 16(%1) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
66 "add %0, %0, %2 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
67 "ld $9, 0(%0) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
68 "pextlb $10, $0, $10 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
69 "sq $10, 32(%1) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
70 "add %0, %0, %2 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
71 "ld $10, 0(%0) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
72 "pextlb $8, $0, $8 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
73 "sq $8, 48(%1) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
74 "add %0, %0, %2 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
75 "ld $8, 0(%0) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
76 "pextlb $9, $0, $9 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
77 "sq $9, 64(%1) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
78 "add %0, %0, %2 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
79 "ld $9, 0(%0) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
80 "pextlb $10, $0, $10 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
81 "sq $10, 80(%1) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
82 "pextlb $8, $0, $8 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
83 "sq $8, 96(%1) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
84 "pextlb $9, $0, $9 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
85 "sq $9, 112(%1) \n\t"
772
311ac50375e4 dsputil mmi cleanup patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 721
diff changeset
86 ".set pop \n\t"
1146
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
87 : "+r" (pixels) : "r" (block), "r" (line_size) : "$8", "$9", "$10", "memory" );
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
88 }
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
89
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
90
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
91 static void put_pixels8_mmi(uint8_t *block, const uint8_t *pixels, int line_size, int h)
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
92 {
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
93 asm volatile(
772
311ac50375e4 dsputil mmi cleanup patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 721
diff changeset
94 ".set push \n\t"
311ac50375e4 dsputil mmi cleanup patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 721
diff changeset
95 ".set mips3 \n\t"
1146
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
96 "1: \n\t"
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
97 "ldr $8, 0(%1) \n\t"
1146
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
98 "addiu %2, %2, -1 \n\t"
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
99 "ldl $8, 7(%1) \n\t"
1146
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
100 "add %1, %1, %3 \n\t"
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
101 "sd $8, 0(%0) \n\t"
1146
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
102 "add %0, %0, %3 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
103 "bgtz %2, 1b \n\t"
772
311ac50375e4 dsputil mmi cleanup patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 721
diff changeset
104 ".set pop \n\t"
1146
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
105 : "+r" (block), "+r" (pixels), "+r" (h) : "r" (line_size)
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
106 : "$8", "memory" );
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
107 }
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
108
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
109
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
110 static void put_pixels16_mmi(uint8_t *block, const uint8_t *pixels, int line_size, int h)
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
111 {
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
112 asm volatile (
772
311ac50375e4 dsputil mmi cleanup patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 721
diff changeset
113 ".set push \n\t"
311ac50375e4 dsputil mmi cleanup patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 721
diff changeset
114 ".set mips3 \n\t"
1146
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
115 "1: \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
116 "ldr $8, 0(%1) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
117 "add $11, %1, %3 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
118 "ldl $8, 7(%1) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
119 "add $10, %0, %3 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
120 "ldr $9, 8(%1) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
121 "ldl $9, 15(%1) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
122 "ldr $12, 0($11) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
123 "add %1, $11, %3 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
124 "ldl $12, 7($11) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
125 "pcpyld $8, $9, $8 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
126 "sq $8, 0(%0) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
127 "ldr $13, 8($11) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
128 "addiu %2, %2, -2 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
129 "ldl $13, 15($11) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
130 "add %0, $10, %3 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
131 "pcpyld $12, $13, $12 \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
132 "sq $12, 0($10) \n\t"
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
133 "bgtz %2, 1b \n\t"
772
311ac50375e4 dsputil mmi cleanup patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 721
diff changeset
134 ".set pop \n\t"
1146
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
135 : "+r" (block), "+r" (pixels), "+r" (h) : "r" (line_size)
868650fc345a new versions; put_pix's about 20% faster, get_pix 3.5x faster. patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 1143
diff changeset
136 : "$8", "$9", "$10", "$11", "$12", "$13", "memory" );
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
137 }
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
138
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
139
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
140 void dsputil_init_mmi(DSPContext* c, AVCodecContext *avctx)
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
141 {
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
142 const int idct_algo= avctx->idct_algo;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
143
856
3c6df37177dd * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 772
diff changeset
144 c->clear_blocks = clear_blocks_mmi;
3c6df37177dd * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 772
diff changeset
145
3c6df37177dd * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 772
diff changeset
146 c->put_pixels_tab[1][0] = put_pixels8_mmi;
3c6df37177dd * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 772
diff changeset
147 c->put_no_rnd_pixels_tab[1][0] = put_pixels8_mmi;
3c6df37177dd * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 772
diff changeset
148
3c6df37177dd * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 772
diff changeset
149 c->put_pixels_tab[0][0] = put_pixels16_mmi;
3c6df37177dd * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 772
diff changeset
150 c->put_no_rnd_pixels_tab[0][0] = put_pixels16_mmi;
3c6df37177dd * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 772
diff changeset
151
3c6df37177dd * using DSPContext - so each codec could use its local (sub)set of CPU extension
kabi
parents: 772
diff changeset
152 c->get_pixels = get_pixels_mmi;
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
153
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
154 if(idct_algo==FF_IDCT_AUTO || idct_algo==FF_IDCT_PS2){
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
155 c->idct_put= ff_mmi_idct_put;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
156 c->idct_add= ff_mmi_idct_add;
1324
7d328fd9d8a5 the return of the idct with 16bit output by ("Ivan Kalvachev" <ivan at cacad dot com>)
michaelni
parents: 1146
diff changeset
157 c->idct = ff_mmi_idct;
1092
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
158 c->idct_permutation_type= FF_LIBMPEG2_IDCT_PERM;
f59c3f66363b MpegEncContext.(i)dct_* -> DspContext.(i)dct_*
michaelni
parents: 1064
diff changeset
159 }
689
efcbfbd18864 ps2 idct patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents:
diff changeset
160 }
721
71f669e9f633 ps2 optimizations update patch by (Leon van Stuivenberg <leonvs at iae dot nl>)
michaelni
parents: 689
diff changeset
161