2
* Optimization of some functions from mpegvideo.c for armv5te
3
* Copyright (c) 2007 Siarhei Siamashka <ssvb@users.sourceforge.net>
5
* This file is part of FFmpeg.
7
* FFmpeg is free software; you can redistribute it and/or
8
* modify it under the terms of the GNU Lesser General Public
9
* License as published by the Free Software Foundation; either
10
* version 2.1 of the License, or (at your option) any later version.
12
* FFmpeg is distributed in the hope that it will be useful,
13
* but WITHOUT ANY WARRANTY; without even the implied warranty of
14
* MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
15
* Lesser General Public License for more details.
17
* You should have received a copy of the GNU Lesser General Public
18
* License along with FFmpeg; if not, write to the Free Software
19
* Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
22
#include "libavcodec/avcodec.h"
23
#include "libavcodec/dsputil.h"
24
#include "libavcodec/mpegvideo.h"
27
#ifdef ENABLE_ARM_TESTS
29
* h263 dequantizer supplementary function, it is performance critical and needs to
30
* have optimized implementations for each architecture. Is also used as a reference
31
* implementation in regression tests
33
static inline void dct_unquantize_h263_helper_c(DCTELEM *block, int qmul, int qadd, int count)
36
for (i = 0; i < count; i++) {
40
level = level * qmul - qadd;
42
level = level * qmul + qadd;
50
/* GCC 3.1 or higher is required to support symbolic names in assembly code */
51
#if (__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 1))
54
* Special optimized version of dct_unquantize_h263_helper_c, it requires the block
55
* to be at least 8 bytes aligned, and may process more elements than requested.
56
* But it is guaranteed to never process more than 64 elements provided that
57
* xxcount argument is <= 64, so it is safe. This macro is optimized for a common
58
* distribution of values for nCoeffs (they are mostly multiple of 8 plus one or
59
* two extra elements). So this macro processes data as 8 elements per loop iteration
60
* and contains optional 2 elements processing in the end.
62
* Inner loop should take 6 cycles per element on arm926ej-s (Nokia 770)
64
#define dct_unquantize_h263_special_helper_armv5te(xxblock, xxqmul, xxqadd, xxcount) \
65
({ DCTELEM *xblock = xxblock; \
66
int xqmul = xxqmul, xqadd = xxqadd, xcount = xxcount, xtmp; \
69
"subs %[count], %[count], #2 \n\t" \
71
"ldrd r4, [%[block], #0] \n\t" \
73
"ldrd r6, [%[block], #8] \n\t" \
75
"rsbs %[data1], %[zero], r4, asr #16 \n\t" \
76
"addgt %[data1], %[qadd], #0 \n\t" \
77
"rsblt %[data1], %[qadd], #0 \n\t" \
78
"smlatbne %[data1], r4, %[qmul], %[data1] \n\t" \
80
"rsbs %[data2], %[zero], r5, asr #16 \n\t" \
81
"addgt %[data2], %[qadd], #0 \n\t" \
82
"rsblt %[data2], %[qadd], #0 \n\t" \
83
"smlatbne %[data2], r5, %[qmul], %[data2] \n\t" \
85
"rsbs %[tmp], %[zero], r4, asl #16 \n\t" \
86
"addgt %[tmp], %[qadd], #0 \n\t" \
87
"rsblt %[tmp], %[qadd], #0 \n\t" \
88
"smlabbne r4, r4, %[qmul], %[tmp] \n\t" \
90
"rsbs %[tmp], %[zero], r5, asl #16 \n\t" \
91
"addgt %[tmp], %[qadd], #0 \n\t" \
92
"rsblt %[tmp], %[qadd], #0 \n\t" \
93
"smlabbne r5, r5, %[qmul], %[tmp] \n\t" \
95
"strh r4, [%[block]], #2 \n\t" \
96
"strh %[data1], [%[block]], #2 \n\t" \
97
"strh r5, [%[block]], #2 \n\t" \
98
"strh %[data2], [%[block]], #2 \n\t" \
100
"rsbs %[data1], %[zero], r6, asr #16 \n\t" \
101
"addgt %[data1], %[qadd], #0 \n\t" \
102
"rsblt %[data1], %[qadd], #0 \n\t" \
103
"smlatbne %[data1], r6, %[qmul], %[data1] \n\t" \
105
"rsbs %[data2], %[zero], r7, asr #16 \n\t" \
106
"addgt %[data2], %[qadd], #0 \n\t" \
107
"rsblt %[data2], %[qadd], #0 \n\t" \
108
"smlatbne %[data2], r7, %[qmul], %[data2] \n\t" \
110
"rsbs %[tmp], %[zero], r6, asl #16 \n\t" \
111
"addgt %[tmp], %[qadd], #0 \n\t" \
112
"rsblt %[tmp], %[qadd], #0 \n\t" \
113
"smlabbne r6, r6, %[qmul], %[tmp] \n\t" \
115
"rsbs %[tmp], %[zero], r7, asl #16 \n\t" \
116
"addgt %[tmp], %[qadd], #0 \n\t" \
117
"rsblt %[tmp], %[qadd], #0 \n\t" \
118
"smlabbne r7, r7, %[qmul], %[tmp] \n\t" \
120
"strh r6, [%[block]], #2 \n\t" \
121
"strh %[data1], [%[block]], #2 \n\t" \
122
"strh r7, [%[block]], #2 \n\t" \
123
"strh %[data2], [%[block]], #2 \n\t" \
125
"subs %[count], %[count], #8 \n\t" \
126
"ldrgtd r4, [%[block], #0] \n\t" /* load data early to avoid load/use pipeline stall */ \
129
"adds %[count], %[count], #2 \n\t" \
132
"ldrsh %[data1], [%[block], #0] \n\t" \
133
"ldrsh %[data2], [%[block], #2] \n\t" \
134
"mov %[tmp], %[qadd] \n\t" \
135
"cmp %[data1], #0 \n\t" \
136
"rsblt %[tmp], %[qadd], #0 \n\t" \
137
"smlabbne %[data1], %[data1], %[qmul], %[tmp] \n\t" \
138
"mov %[tmp], %[qadd] \n\t" \
139
"cmp %[data2], #0 \n\t" \
140
"rsblt %[tmp], %[qadd], #0 \n\t" \
141
"smlabbne %[data2], %[data2], %[qmul], %[tmp] \n\t" \
142
"strh %[data1], [%[block]], #2 \n\t" \
143
"strh %[data2], [%[block]], #2 \n\t" \
145
: [block] "+&r" (xblock), [count] "+&r" (xcount), [tmp] "=&r" (xtmp), \
146
[data1] "=&r" (xdata1), [data2] "=&r" (xdata2) \
147
: [qmul] "r" (xqmul), [qadd] "r" (xqadd), [zero] "r" (0) \
148
: "r4", "r5", "r6", "r7", "cc", "memory" \
152
static void dct_unquantize_h263_intra_armv5te(MpegEncContext *s,
153
DCTELEM *block, int n, int qscale)
155
int level, qmul, qadd;
158
assert(s->block_last_index[n]>=0);
164
level = block[0] * s->y_dc_scale;
166
level = block[0] * s->c_dc_scale;
167
qadd = (qscale - 1) | 1;
175
nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
177
dct_unquantize_h263_special_helper_armv5te(block, qmul, qadd, nCoeffs + 1);
181
static void dct_unquantize_h263_inter_armv5te(MpegEncContext *s,
182
DCTELEM *block, int n, int qscale)
187
assert(s->block_last_index[n]>=0);
189
qadd = (qscale - 1) | 1;
192
nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
194
dct_unquantize_h263_special_helper_armv5te(block, qmul, qadd, nCoeffs + 1);
197
#define HAVE_DCT_UNQUANTIZE_H263_ARMV5TE_OPTIMIZED
201
void MPV_common_init_armv5te(MpegEncContext *s)
203
#ifdef HAVE_DCT_UNQUANTIZE_H263_ARMV5TE_OPTIMIZED
204
s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_armv5te;
205
s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_armv5te;