Merge commit '6b110d3a739c31602b59887ad65c67025df3f49d'
[ffmpeg.git] / libavcodec / x86 / mpegaudiodsp.c
1 /*
2  * MMX optimized MP3 decoding functions
3  * Copyright (c) 2010 Vitor Sessak
4  *
5  * This file is part of FFmpeg.
6  *
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.
11  *
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.
16  *
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
20  */
21
22 #include "libavutil/attributes.h"
23 #include "libavutil/cpu.h"
24 #include "libavutil/internal.h"
25 #include "libavutil/x86/asm.h"
26 #include "libavutil/x86/cpu.h"
27 #include "libavcodec/mpegaudiodsp.h"
28
29 #define DECL(CPU)\
30 static void imdct36_blocks_ ## CPU(float *out, float *buf, float *in, int count, int switch_point, int block_type);\
31 void ff_imdct36_float_ ## CPU(float *out, float *buf, float *in, float *win);
32
33 DECL(sse)
34 DECL(sse2)
35 DECL(sse3)
36 DECL(ssse3)
37 DECL(avx)
38
39 void ff_four_imdct36_float_sse(float *out, float *buf, float *in, float *win,
40                                float *tmpbuf);
41 void ff_four_imdct36_float_avx(float *out, float *buf, float *in, float *win,
42                                float *tmpbuf);
43
44 DECLARE_ALIGNED(16, static float, mdct_win_sse)[2][4][4*40];
45
46 #if HAVE_SSE2_INLINE
47
48 #define MACS(rt, ra, rb) rt+=(ra)*(rb)
49 #define MLSS(rt, ra, rb) rt-=(ra)*(rb)
50
51 #define SUM8(op, sum, w, p)               \
52 {                                         \
53     op(sum, (w)[0 * 64], (p)[0 * 64]);    \
54     op(sum, (w)[1 * 64], (p)[1 * 64]);    \
55     op(sum, (w)[2 * 64], (p)[2 * 64]);    \
56     op(sum, (w)[3 * 64], (p)[3 * 64]);    \
57     op(sum, (w)[4 * 64], (p)[4 * 64]);    \
58     op(sum, (w)[5 * 64], (p)[5 * 64]);    \
59     op(sum, (w)[6 * 64], (p)[6 * 64]);    \
60     op(sum, (w)[7 * 64], (p)[7 * 64]);    \
61 }
62
63 static void apply_window(const float *buf, const float *win1,
64                          const float *win2, float *sum1, float *sum2, int len)
65 {
66     x86_reg count = - 4*len;
67     const float *win1a = win1+len;
68     const float *win2a = win2+len;
69     const float *bufa  = buf+len;
70     float *sum1a = sum1+len;
71     float *sum2a = sum2+len;
72
73
74 #define MULT(a, b)                                 \
75     "movaps " #a "(%1,%0), %%xmm1           \n\t"  \
76     "movaps " #a "(%3,%0), %%xmm2           \n\t"  \
77     "mulps         %%xmm2, %%xmm1           \n\t"  \
78     "subps         %%xmm1, %%xmm0           \n\t"  \
79     "mulps  " #b "(%2,%0), %%xmm2           \n\t"  \
80     "subps         %%xmm2, %%xmm4           \n\t"  \
81
82     __asm__ volatile(
83             "1:                                   \n\t"
84             "xorps       %%xmm0, %%xmm0           \n\t"
85             "xorps       %%xmm4, %%xmm4           \n\t"
86
87             MULT(   0,   0)
88             MULT( 256,  64)
89             MULT( 512, 128)
90             MULT( 768, 192)
91             MULT(1024, 256)
92             MULT(1280, 320)
93             MULT(1536, 384)
94             MULT(1792, 448)
95
96             "movaps      %%xmm0, (%4,%0)          \n\t"
97             "movaps      %%xmm4, (%5,%0)          \n\t"
98             "add            $16,  %0              \n\t"
99             "jl              1b                   \n\t"
100             :"+&r"(count)
101             :"r"(win1a), "r"(win2a), "r"(bufa), "r"(sum1a), "r"(sum2a)
102             );
103
104 #undef MULT
105 }
106
107 static void apply_window_mp3(float *in, float *win, int *unused, float *out,
108                              int incr)
109 {
110     LOCAL_ALIGNED_16(float, suma, [17]);
111     LOCAL_ALIGNED_16(float, sumb, [17]);
112     LOCAL_ALIGNED_16(float, sumc, [17]);
113     LOCAL_ALIGNED_16(float, sumd, [17]);
114
115     float sum;
116
117     /* copy to avoid wrap */
118     __asm__ volatile(
119             "movaps    0(%0), %%xmm0   \n\t" \
120             "movaps   16(%0), %%xmm1   \n\t" \
121             "movaps   32(%0), %%xmm2   \n\t" \
122             "movaps   48(%0), %%xmm3   \n\t" \
123             "movaps   %%xmm0,   0(%1) \n\t" \
124             "movaps   %%xmm1,  16(%1) \n\t" \
125             "movaps   %%xmm2,  32(%1) \n\t" \
126             "movaps   %%xmm3,  48(%1) \n\t" \
127             "movaps   64(%0), %%xmm0   \n\t" \
128             "movaps   80(%0), %%xmm1   \n\t" \
129             "movaps   96(%0), %%xmm2   \n\t" \
130             "movaps  112(%0), %%xmm3   \n\t" \
131             "movaps   %%xmm0,  64(%1) \n\t" \
132             "movaps   %%xmm1,  80(%1) \n\t" \
133             "movaps   %%xmm2,  96(%1) \n\t" \
134             "movaps   %%xmm3, 112(%1) \n\t"
135             ::"r"(in), "r"(in+512)
136             :"memory"
137             );
138
139     apply_window(in + 16, win     , win + 512, suma, sumc, 16);
140     apply_window(in + 32, win + 48, win + 640, sumb, sumd, 16);
141
142     SUM8(MACS, suma[0], win + 32, in + 48);
143
144     sumc[ 0] = 0;
145     sumb[16] = 0;
146     sumd[16] = 0;
147
148 #define SUMS(suma, sumb, sumc, sumd, out1, out2)               \
149             "movups " #sumd "(%4),       %%xmm0          \n\t" \
150             "shufps         $0x1b,       %%xmm0, %%xmm0  \n\t" \
151             "subps  " #suma "(%1),       %%xmm0          \n\t" \
152             "movaps        %%xmm0," #out1 "(%0)          \n\t" \
153 \
154             "movups " #sumc "(%3),       %%xmm0          \n\t" \
155             "shufps         $0x1b,       %%xmm0, %%xmm0  \n\t" \
156             "addps  " #sumb "(%2),       %%xmm0          \n\t" \
157             "movaps        %%xmm0," #out2 "(%0)          \n\t"
158
159     if (incr == 1) {
160         __asm__ volatile(
161             SUMS( 0, 48,  4, 52,  0, 112)
162             SUMS(16, 32, 20, 36, 16,  96)
163             SUMS(32, 16, 36, 20, 32,  80)
164             SUMS(48,  0, 52,  4, 48,  64)
165
166             :"+&r"(out)
167             :"r"(&suma[0]), "r"(&sumb[0]), "r"(&sumc[0]), "r"(&sumd[0])
168             :"memory"
169             );
170         out += 16*incr;
171     } else {
172         int j;
173         float *out2 = out + 32 * incr;
174         out[0  ]  = -suma[   0];
175         out += incr;
176         out2 -= incr;
177         for(j=1;j<16;j++) {
178             *out  = -suma[   j] + sumd[16-j];
179             *out2 =  sumb[16-j] + sumc[   j];
180             out  += incr;
181             out2 -= incr;
182         }
183     }
184
185     sum = 0;
186     SUM8(MLSS, sum, win + 16 + 32, in + 32);
187     *out = sum;
188 }
189
190 #endif /* HAVE_SSE2_INLINE */
191
192 #if HAVE_YASM
193 #define DECL_IMDCT_BLOCKS(CPU1, CPU2)                                       \
194 static void imdct36_blocks_ ## CPU1(float *out, float *buf, float *in,      \
195                                int count, int switch_point, int block_type) \
196 {                                                                           \
197     int align_end = count - (count & 3);                                \
198     int j;                                                              \
199     for (j = 0; j < align_end; j+= 4) {                                 \
200         LOCAL_ALIGNED_16(float, tmpbuf, [1024]);                        \
201         float *win = mdct_win_sse[switch_point && j < 4][block_type];   \
202         /* apply window & overlap with previous buffer */               \
203                                                                         \
204         /* select window */                                             \
205         ff_four_imdct36_float_ ## CPU2(out, buf, in, win, tmpbuf);      \
206         in      += 4*18;                                                \
207         buf     += 4*18;                                                \
208         out     += 4;                                                   \
209     }                                                                   \
210     for (; j < count; j++) {                                            \
211         /* apply window & overlap with previous buffer */               \
212                                                                         \
213         /* select window */                                             \
214         int win_idx = (switch_point && j < 2) ? 0 : block_type;         \
215         float *win = ff_mdct_win_float[win_idx + (4 & -(j & 1))];       \
216                                                                         \
217         ff_imdct36_float_ ## CPU1(out, buf, in, win);                   \
218                                                                         \
219         in  += 18;                                                      \
220         buf++;                                                          \
221         out++;                                                          \
222     }                                                                   \
223 }
224
225 #if HAVE_SSE
226 DECL_IMDCT_BLOCKS(sse,sse)
227 DECL_IMDCT_BLOCKS(sse2,sse)
228 DECL_IMDCT_BLOCKS(sse3,sse)
229 DECL_IMDCT_BLOCKS(ssse3,sse)
230 #endif
231 #if HAVE_AVX_EXTERNAL
232 DECL_IMDCT_BLOCKS(avx,avx)
233 #endif
234 #endif /* HAVE_YASM */
235
236 av_cold void ff_mpadsp_init_x86(MPADSPContext *s)
237 {
238     int mm_flags = av_get_cpu_flags();
239
240     int i, j;
241     for (j = 0; j < 4; j++) {
242         for (i = 0; i < 40; i ++) {
243             mdct_win_sse[0][j][4*i    ] = ff_mdct_win_float[j    ][i];
244             mdct_win_sse[0][j][4*i + 1] = ff_mdct_win_float[j + 4][i];
245             mdct_win_sse[0][j][4*i + 2] = ff_mdct_win_float[j    ][i];
246             mdct_win_sse[0][j][4*i + 3] = ff_mdct_win_float[j + 4][i];
247             mdct_win_sse[1][j][4*i    ] = ff_mdct_win_float[0    ][i];
248             mdct_win_sse[1][j][4*i + 1] = ff_mdct_win_float[4    ][i];
249             mdct_win_sse[1][j][4*i + 2] = ff_mdct_win_float[j    ][i];
250             mdct_win_sse[1][j][4*i + 3] = ff_mdct_win_float[j + 4][i];
251         }
252     }
253
254 #if HAVE_SSE2_INLINE
255     if (mm_flags & AV_CPU_FLAG_SSE2) {
256         s->apply_window_float = apply_window_mp3;
257     }
258 #endif /* HAVE_SSE2_INLINE */
259
260 #if HAVE_YASM
261     if (EXTERNAL_AVX(mm_flags)) {
262         s->imdct36_blocks_float = imdct36_blocks_avx;
263     } else if (EXTERNAL_SSSE3(mm_flags)) {
264         s->imdct36_blocks_float = imdct36_blocks_ssse3;
265     } else if (EXTERNAL_SSE3(mm_flags)) {
266         s->imdct36_blocks_float = imdct36_blocks_sse3;
267     } else if (EXTERNAL_SSE2(mm_flags)) {
268         s->imdct36_blocks_float = imdct36_blocks_sse2;
269     } else if (EXTERNAL_SSE(mm_flags)) {
270         s->imdct36_blocks_float = imdct36_blocks_sse;
271     }
272 #endif /* HAVE_YASM */
273 }