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
23 * Some useful links for those who may be interested in optimizing code for ARM.
24 * ARM Architecture Reference Manual: http://www.arm.com/community/academy/resources.html
25 * Instructions timings and optimization guide for ARM9E: http://www.arm.com/pdfs/DDI0222B_9EJS_r1p2.pdf
29 #include "mpegvideo.h"
33 #ifdef ENABLE_ARM_TESTS
35 * h263 dequantizer supplementary function, it is performance critical and needs to
36 * have optimized implementations for each architecture. Is also used as a reference
37 * implementation in regression tests
39 static inline void dct_unquantize_h263_helper_c(DCTELEM *block, int qmul, int qadd, int count)
42 for (i = 0; i < count; i++) {
46 level = level * qmul - qadd;
48 level = level * qmul + qadd;
56 /* GCC 3.1 or higher is required to support symbolic names in assembly code */
57 #if (__GNUC__ > 3) || ((__GNUC__ == 3) && (__GNUC_MINOR__ >= 1))
60 * Special optimized version of dct_unquantize_h263_helper_c, it requires the block
61 * to be at least 8 bytes aligned, and may process more elements than requested.
62 * But it is guaranteed to never process more than 64 elements provided that
63 * xxcount argument is <= 64, so it is safe. This macro is optimized for a common
64 * distribution of values for nCoeffs (they are mostly multiple of 8 plus one or
65 * two extra elements). So this macro processes data as 8 elements per loop iteration
66 * and contains optional 2 elements processing in the end.
68 * Inner loop should take 6 cycles per element on arm926ej-s (Nokia 770)
70 #define dct_unquantize_h263_special_helper_armv5te(xxblock, xxqmul, xxqadd, xxcount) \
71 ({ DCTELEM *xblock = xxblock; \
72 int xqmul = xxqmul, xqadd = xxqadd, xcount = xxcount, xtmp; \
74 __asm__ __volatile__( \
75 "subs %[count], %[count], #2 \n\t" \
77 "ldrd r4, [%[block], #0] \n\t" \
79 "ldrd r6, [%[block], #8] \n\t" \
81 "rsbs %[data1], %[zero], r4, asr #16 \n\t" \
82 "addgt %[data1], %[qadd], #0 \n\t" \
83 "rsblt %[data1], %[qadd], #0 \n\t" \
84 "smlatbne %[data1], r4, %[qmul], %[data1] \n\t" \
86 "rsbs %[data2], %[zero], r5, asr #16 \n\t" \
87 "addgt %[data2], %[qadd], #0 \n\t" \
88 "rsblt %[data2], %[qadd], #0 \n\t" \
89 "smlatbne %[data2], r5, %[qmul], %[data2] \n\t" \
91 "rsbs %[tmp], %[zero], r4, asl #16 \n\t" \
92 "addgt %[tmp], %[qadd], #0 \n\t" \
93 "rsblt %[tmp], %[qadd], #0 \n\t" \
94 "smlabbne r4, r4, %[qmul], %[tmp] \n\t" \
96 "rsbs %[tmp], %[zero], r5, asl #16 \n\t" \
97 "addgt %[tmp], %[qadd], #0 \n\t" \
98 "rsblt %[tmp], %[qadd], #0 \n\t" \
99 "smlabbne r5, r5, %[qmul], %[tmp] \n\t" \
101 "strh r4, [%[block]], #2 \n\t" \
102 "strh %[data1], [%[block]], #2 \n\t" \
103 "strh r5, [%[block]], #2 \n\t" \
104 "strh %[data2], [%[block]], #2 \n\t" \
106 "rsbs %[data1], %[zero], r6, asr #16 \n\t" \
107 "addgt %[data1], %[qadd], #0 \n\t" \
108 "rsblt %[data1], %[qadd], #0 \n\t" \
109 "smlatbne %[data1], r6, %[qmul], %[data1] \n\t" \
111 "rsbs %[data2], %[zero], r7, asr #16 \n\t" \
112 "addgt %[data2], %[qadd], #0 \n\t" \
113 "rsblt %[data2], %[qadd], #0 \n\t" \
114 "smlatbne %[data2], r7, %[qmul], %[data2] \n\t" \
116 "rsbs %[tmp], %[zero], r6, asl #16 \n\t" \
117 "addgt %[tmp], %[qadd], #0 \n\t" \
118 "rsblt %[tmp], %[qadd], #0 \n\t" \
119 "smlabbne r6, r6, %[qmul], %[tmp] \n\t" \
121 "rsbs %[tmp], %[zero], r7, asl #16 \n\t" \
122 "addgt %[tmp], %[qadd], #0 \n\t" \
123 "rsblt %[tmp], %[qadd], #0 \n\t" \
124 "smlabbne r7, r7, %[qmul], %[tmp] \n\t" \
126 "strh r6, [%[block]], #2 \n\t" \
127 "strh %[data1], [%[block]], #2 \n\t" \
128 "strh r7, [%[block]], #2 \n\t" \
129 "strh %[data2], [%[block]], #2 \n\t" \
131 "subs %[count], %[count], #8 \n\t" \
132 "ldrgtd r4, [%[block], #0] \n\t" /* load data early to avoid load/use pipeline stall */ \
135 "adds %[count], %[count], #2 \n\t" \
138 "ldrsh %[data1], [%[block], #0] \n\t" \
139 "ldrsh %[data2], [%[block], #2] \n\t" \
140 "mov %[tmp], %[qadd] \n\t" \
141 "cmp %[data1], #0 \n\t" \
142 "rsblt %[tmp], %[qadd], #0 \n\t" \
143 "smlabbne %[data1], %[data1], %[qmul], %[tmp] \n\t" \
144 "mov %[tmp], %[qadd] \n\t" \
145 "cmp %[data2], #0 \n\t" \
146 "rsblt %[tmp], %[qadd], #0 \n\t" \
147 "smlabbne %[data2], %[data2], %[qmul], %[tmp] \n\t" \
148 "strh %[data1], [%[block]], #2 \n\t" \
149 "strh %[data2], [%[block]], #2 \n\t" \
151 : [block] "+&r" (xblock), [count] "+&r" (xcount), [tmp] "=&r" (xtmp), \
152 [data1] "=&r" (xdata1), [data2] "=&r" (xdata2) \
153 : [qmul] "r" (xqmul), [qadd] "r" (xqadd), [zero] "r" (0) \
154 : "r4", "r5", "r6", "r7", "cc", "memory" \
158 static void dct_unquantize_h263_intra_armv5te(MpegEncContext *s,
159 DCTELEM *block, int n, int qscale)
161 int level, qmul, qadd;
164 assert(s->block_last_index[n]>=0);
170 level = block[0] * s->y_dc_scale;
172 level = block[0] * s->c_dc_scale;
173 qadd = (qscale - 1) | 1;
181 nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
183 dct_unquantize_h263_special_helper_armv5te(block, qmul, qadd, nCoeffs + 1);
187 static void dct_unquantize_h263_inter_armv5te(MpegEncContext *s,
188 DCTELEM *block, int n, int qscale)
193 assert(s->block_last_index[n]>=0);
195 qadd = (qscale - 1) | 1;
198 nCoeffs= s->inter_scantable.raster_end[ s->block_last_index[n] ];
200 dct_unquantize_h263_special_helper_armv5te(block, qmul, qadd, nCoeffs + 1);
203 #define HAVE_DCT_UNQUANTIZE_H263_ARMV5TE_OPTIMIZED
207 void MPV_common_init_armv5te(MpegEncContext *s)
209 #ifdef HAVE_DCT_UNQUANTIZE_H263_ARMV5TE_OPTIMIZED
210 s->dct_unquantize_h263_intra = dct_unquantize_h263_intra_armv5te;
211 s->dct_unquantize_h263_inter = dct_unquantize_h263_inter_armv5te;