annotate x86/mpegaudiodec_mmx.c @ 12206:fe243bb5ef61 libavcodec

Move ff_pw_* from vc1dsp_mmx.c to dsputil_mmx.c Should fix compilation with icc and should help prevent any future duplicates
author conrad
date Wed, 21 Jul 2010 10:02:03 +0000
parents ce4b71c171aa
children 3fc4c625b6f3
Ignore whitespace changes - Everywhere: Within whitespace: At end of lines:
rev   line source
11939
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
1 /*
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
2 * MMX optimized MP3 decoding functions
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
3 * Copyright (c) 2010 Vitor Sessak
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
4 *
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
5 * This file is part of FFmpeg.
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
6 *
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
7 * FFmpeg is free software; you can redistribute it and/or
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
8 * modify it under the terms of the GNU Lesser General Public
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
9 * License as published by the Free Software Foundation; either
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
10 * version 2.1 of the License, or (at your option) any later version.
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
11 *
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
12 * FFmpeg is distributed in the hope that it will be useful,
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
13 * but WITHOUT ANY WARRANTY; without even the implied warranty of
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
14 * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
15 * Lesser General Public License for more details.
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
16 *
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
17 * You should have received a copy of the GNU Lesser General Public
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
18 * License along with FFmpeg; if not, write to the Free Software
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
19 * Foundation, Inc., 51 Franklin Street, Fifth Floor, Boston, MA 02110-1301 USA
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
20 */
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
21
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
22 #include "libavutil/x86_cpu.h"
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
23
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
24 #define CONFIG_FLOAT 1
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
25 #include "libavcodec/mpegaudio.h"
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
26
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
27 #define MACS(rt, ra, rb) rt+=(ra)*(rb)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
28 #define MLSS(rt, ra, rb) rt-=(ra)*(rb)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
29
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
30 #define SUM8(op, sum, w, p) \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
31 { \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
32 op(sum, (w)[0 * 64], (p)[0 * 64]); \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
33 op(sum, (w)[1 * 64], (p)[1 * 64]); \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
34 op(sum, (w)[2 * 64], (p)[2 * 64]); \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
35 op(sum, (w)[3 * 64], (p)[3 * 64]); \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
36 op(sum, (w)[4 * 64], (p)[4 * 64]); \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
37 op(sum, (w)[5 * 64], (p)[5 * 64]); \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
38 op(sum, (w)[6 * 64], (p)[6 * 64]); \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
39 op(sum, (w)[7 * 64], (p)[7 * 64]); \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
40 }
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
41
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
42 static void apply_window(const float *buf, const float *win1,
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
43 const float *win2, float *sum1, float *sum2, int len)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
44 {
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
45 x86_reg count = - 4*len;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
46 const float *win1a = win1+len;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
47 const float *win2a = win2+len;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
48 const float *bufa = buf+len;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
49 float *sum1a = sum1+len;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
50 float *sum2a = sum2+len;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
51
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
52
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
53 #define MULT(a, b) \
11941
9e12316c508a Fix asm constraints in apply_window()
vitor
parents: 11939
diff changeset
54 "movaps " #a "(%1,%0), %%xmm1 \n\t" \
9e12316c508a Fix asm constraints in apply_window()
vitor
parents: 11939
diff changeset
55 "movaps " #a "(%3,%0), %%xmm2 \n\t" \
11939
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
56 "mulps %%xmm2, %%xmm1 \n\t" \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
57 "subps %%xmm1, %%xmm0 \n\t" \
11941
9e12316c508a Fix asm constraints in apply_window()
vitor
parents: 11939
diff changeset
58 "mulps " #b "(%2,%0), %%xmm2 \n\t" \
11939
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
59 "subps %%xmm2, %%xmm4 \n\t" \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
60
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
61 __asm__ volatile(
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
62 "1: \n\t"
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
63 "xorps %%xmm0, %%xmm0 \n\t"
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
64 "xorps %%xmm4, %%xmm4 \n\t"
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
65
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
66 MULT( 0, 0)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
67 MULT( 256, 64)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
68 MULT( 512, 128)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
69 MULT( 768, 192)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
70 MULT(1024, 256)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
71 MULT(1280, 320)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
72 MULT(1536, 384)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
73 MULT(1792, 448)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
74
11941
9e12316c508a Fix asm constraints in apply_window()
vitor
parents: 11939
diff changeset
75 "movaps %%xmm0, (%4,%0) \n\t"
9e12316c508a Fix asm constraints in apply_window()
vitor
parents: 11939
diff changeset
76 "movaps %%xmm4, (%5,%0) \n\t"
11942
ce4b71c171aa Fix compilation on x64.
vitor
parents: 11941
diff changeset
77 "add $16, %0 \n\t"
11939
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
78 "jl 1b \n\t"
11941
9e12316c508a Fix asm constraints in apply_window()
vitor
parents: 11939
diff changeset
79 :"+&r"(count)
9e12316c508a Fix asm constraints in apply_window()
vitor
parents: 11939
diff changeset
80 :"r"(win1a), "r"(win2a), "r"(bufa), "r"(sum1a), "r"(sum2a)
11939
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
81 );
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
82
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
83 #undef MULT
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
84 }
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
85
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
86 static void apply_window_mp3(float *in, float *win, int *unused, float *out,
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
87 int incr)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
88 {
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
89 LOCAL_ALIGNED_16(float, suma, [17]);
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
90 LOCAL_ALIGNED_16(float, sumb, [17]);
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
91 LOCAL_ALIGNED_16(float, sumc, [17]);
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
92 LOCAL_ALIGNED_16(float, sumd, [17]);
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
93
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
94 float sum;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
95
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
96 /* copy to avoid wrap */
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
97 memcpy(in + 512, in, 32 * sizeof(*in));
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
98
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
99 apply_window(in + 16, win , win + 512, suma, sumc, 16);
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
100 apply_window(in + 32, win + 48, win + 640, sumb, sumd, 16);
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
101
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
102 SUM8(MACS, suma[0], win + 32, in + 48);
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
103
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
104 sumc[ 0] = 0;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
105 sumb[16] = 0;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
106 sumd[16] = 0;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
107
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
108 #define SUMS(suma, sumb, sumc, sumd, out1, out2) \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
109 "movups " #sumd "(%4), %%xmm0 \n\t" \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
110 "shufps $0x1b, %%xmm0, %%xmm0 \n\t" \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
111 "subps " #suma "(%1), %%xmm0 \n\t" \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
112 "movaps %%xmm0," #out1 "(%0) \n\t" \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
113 \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
114 "movups " #sumc "(%3), %%xmm0 \n\t" \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
115 "shufps $0x1b, %%xmm0, %%xmm0 \n\t" \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
116 "addps " #sumb "(%2), %%xmm0 \n\t" \
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
117 "movaps %%xmm0," #out2 "(%0) \n\t"
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
118
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
119 if (incr == 1) {
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
120 __asm__ volatile(
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
121 SUMS( 0, 48, 4, 52, 0, 112)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
122 SUMS(16, 32, 20, 36, 16, 96)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
123 SUMS(32, 16, 36, 20, 32, 80)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
124 SUMS(48, 0, 52, 4, 48, 64)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
125
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
126 :"+&r"(out)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
127 :"r"(&suma[0]), "r"(&sumb[0]), "r"(&sumc[0]), "r"(&sumd[0])
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
128 :"memory"
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
129 );
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
130 out += 16*incr;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
131 } else {
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
132 int j;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
133 float *out2 = out + 32 * incr;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
134 out[0 ] = -suma[ 0];
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
135 out += incr;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
136 out2 -= incr;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
137 for(j=1;j<16;j++) {
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
138 *out = -suma[ j] + sumd[16-j];
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
139 *out2 = sumb[16-j] + sumc[ j];
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
140 out += incr;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
141 out2 -= incr;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
142 }
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
143 }
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
144
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
145 sum = 0;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
146 SUM8(MLSS, sum, win + 16 + 32, in + 32);
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
147 *out = sum;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
148 }
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
149
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
150 void ff_mpegaudiodec_init_mmx(MPADecodeContext *s)
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
151 {
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
152 mm_flags = mm_support();
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
153
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
154 if (mm_flags & FF_MM_SSE2) {
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
155 s->apply_window_mp3 = apply_window_mp3;
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
156 }
ef338bd70180 SSE-optimized MP3 floating point windowing functions
vitor
parents:
diff changeset
157 }