ac3dsp: Reverse matrix in/out order in downmix()
[ffmpeg.git] / libavcodec / x86 / ac3dsp_init.c
1 /*
2  * x86-optimized AC-3 DSP functions
3  * Copyright (c) 2011 Justin Ruggles
4  *
5  * This file is part of Libav.
6  *
7  * Libav 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  * Libav 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 Libav; 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/mem.h"
24 #include "libavutil/x86/asm.h"
25 #include "libavutil/x86/cpu.h"
26 #include "libavcodec/ac3.h"
27 #include "libavcodec/ac3dsp.h"
28
29 void ff_ac3_exponent_min_mmx   (uint8_t *exp, int num_reuse_blocks, int nb_coefs);
30 void ff_ac3_exponent_min_mmxext(uint8_t *exp, int num_reuse_blocks, int nb_coefs);
31 void ff_ac3_exponent_min_sse2  (uint8_t *exp, int num_reuse_blocks, int nb_coefs);
32
33 int ff_ac3_max_msb_abs_int16_mmx  (const int16_t *src, int len);
34 int ff_ac3_max_msb_abs_int16_mmxext(const int16_t *src, int len);
35 int ff_ac3_max_msb_abs_int16_sse2 (const int16_t *src, int len);
36 int ff_ac3_max_msb_abs_int16_ssse3(const int16_t *src, int len);
37
38 void ff_ac3_lshift_int16_mmx (int16_t *src, unsigned int len, unsigned int shift);
39 void ff_ac3_lshift_int16_sse2(int16_t *src, unsigned int len, unsigned int shift);
40
41 void ff_ac3_rshift_int32_mmx (int32_t *src, unsigned int len, unsigned int shift);
42 void ff_ac3_rshift_int32_sse2(int32_t *src, unsigned int len, unsigned int shift);
43
44 void ff_float_to_fixed24_3dnow(int32_t *dst, const float *src, unsigned int len);
45 void ff_float_to_fixed24_sse  (int32_t *dst, const float *src, unsigned int len);
46 void ff_float_to_fixed24_sse2 (int32_t *dst, const float *src, unsigned int len);
47
48 int ff_ac3_compute_mantissa_size_sse2(uint16_t mant_cnt[6][16]);
49
50 void ff_ac3_extract_exponents_sse2 (uint8_t *exp, int32_t *coef, int nb_coefs);
51 void ff_ac3_extract_exponents_ssse3(uint8_t *exp, int32_t *coef, int nb_coefs);
52
53 void ff_apply_window_int16_round_mmxext(int16_t *output, const int16_t *input,
54                                         const int16_t *window, unsigned int len);
55 void ff_apply_window_int16_round_sse2(int16_t *output, const int16_t *input,
56                                       const int16_t *window, unsigned int len);
57 void ff_apply_window_int16_mmxext(int16_t *output, const int16_t *input,
58                                   const int16_t *window, unsigned int len);
59 void ff_apply_window_int16_sse2(int16_t *output, const int16_t *input,
60                                 const int16_t *window, unsigned int len);
61 void ff_apply_window_int16_ssse3(int16_t *output, const int16_t *input,
62                                  const int16_t *window, unsigned int len);
63 void ff_apply_window_int16_ssse3_atom(int16_t *output, const int16_t *input,
64                                       const int16_t *window, unsigned int len);
65
66 #if HAVE_SSE_INLINE && HAVE_7REGS
67
68 #define IF1(x) x
69 #define IF0(x)
70
71 #define MIX5(mono, stereo)                                      \
72     __asm__ volatile (                                          \
73         "movss           0(%1), %%xmm5          \n"             \
74         "movss           4(%1), %%xmm6          \n"             \
75         "movss          12(%1), %%xmm7          \n"             \
76         "shufps     $0, %%xmm5, %%xmm5          \n"             \
77         "shufps     $0, %%xmm6, %%xmm6          \n"             \
78         "shufps     $0, %%xmm7, %%xmm7          \n"             \
79         "1:                                     \n"             \
80         "movaps       (%0, %2), %%xmm0          \n"             \
81         "movaps       (%0, %3), %%xmm1          \n"             \
82         "movaps       (%0, %4), %%xmm2          \n"             \
83         "movaps       (%0, %5), %%xmm3          \n"             \
84         "movaps       (%0, %6), %%xmm4          \n"             \
85         "mulps          %%xmm5, %%xmm0          \n"             \
86         "mulps          %%xmm6, %%xmm1          \n"             \
87         "mulps          %%xmm5, %%xmm2          \n"             \
88         "mulps          %%xmm7, %%xmm3          \n"             \
89         "mulps          %%xmm7, %%xmm4          \n"             \
90  stereo("addps          %%xmm1, %%xmm0          \n")            \
91         "addps          %%xmm1, %%xmm2          \n"             \
92         "addps          %%xmm3, %%xmm0          \n"             \
93         "addps          %%xmm4, %%xmm2          \n"             \
94    mono("addps          %%xmm2, %%xmm0          \n")            \
95         "movaps         %%xmm0, (%0, %2)        \n"             \
96  stereo("movaps         %%xmm2, (%0, %3)        \n")            \
97         "add               $16, %0              \n"             \
98         "jl                 1b                  \n"             \
99         : "+&r"(i)                                              \
100         : "r"(matrix[0]),                                          \
101           "r"(samples[0] + len),                                \
102           "r"(samples[1] + len),                                \
103           "r"(samples[2] + len),                                \
104           "r"(samples[3] + len),                                \
105           "r"(samples[4] + len)                                 \
106         : XMM_CLOBBERS("%xmm0", "%xmm1", "%xmm2", "%xmm3",      \
107                       "%xmm4", "%xmm5", "%xmm6", "%xmm7",)      \
108          "memory"                                               \
109     );
110
111 #define MIX_MISC(stereo)                                        \
112     __asm__ volatile (                                          \
113         "mov              %5, %2            \n"                 \
114         "1:                                 \n"                 \
115         "mov -%c7(%6, %2, %c8), %3          \n"                 \
116         "movaps     (%3, %0), %%xmm0        \n"                 \
117  stereo("movaps       %%xmm0, %%xmm1        \n")                \
118         "mulps        %%xmm4, %%xmm0        \n"                 \
119  stereo("mulps        %%xmm5, %%xmm1        \n")                \
120         "2:                                 \n"                 \
121         "mov   (%6, %2, %c8), %1            \n"                 \
122         "movaps     (%1, %0), %%xmm2        \n"                 \
123  stereo("movaps       %%xmm2, %%xmm3        \n")                \
124         "mulps   (%4, %2, 8), %%xmm2        \n"                 \
125  stereo("mulps 16(%4, %2, 8), %%xmm3        \n")                \
126         "addps        %%xmm2, %%xmm0        \n"                 \
127  stereo("addps        %%xmm3, %%xmm1        \n")                \
128         "add              $4, %2            \n"                 \
129         "jl               2b                \n"                 \
130         "mov              %5, %2            \n"                 \
131  stereo("mov   (%6, %2, %c8), %1            \n")                \
132         "movaps       %%xmm0, (%3, %0)      \n"                 \
133  stereo("movaps       %%xmm1, (%1, %0)      \n")                \
134         "add             $16, %0            \n"                 \
135         "jl               1b                \n"                 \
136         : "+&r"(i), "=&r"(j), "=&r"(k), "=&r"(m)                \
137         : "r"(matrix_simd + in_ch),                             \
138           "g"((intptr_t) - 4 * (in_ch - 1)),                    \
139           "r"(samp + in_ch),                                    \
140           "i"(sizeof(float *)), "i"(sizeof(float *)/4)          \
141         : "memory"                                              \
142     );
143
144 static void ac3_downmix_sse(float **samples, float **matrix,
145                             int out_ch, int in_ch, int len)
146 {
147     int **matrix_cmp = (int **)matrix;
148     intptr_t i, j, k, m;
149
150     i = -len * sizeof(float);
151     if (in_ch == 5 && out_ch == 2 &&
152         !(matrix_cmp[1][0] | matrix_cmp[0][2]   |
153           matrix_cmp[1][3] | matrix_cmp[0][4]   |
154           (matrix_cmp[0][1] ^ matrix_cmp[1][1]) |
155           (matrix_cmp[0][0] ^ matrix_cmp[1][2]))) {
156         MIX5(IF0, IF1);
157     } else if (in_ch == 5 && out_ch == 1 &&
158                matrix_cmp[0][0] == matrix_cmp[0][2] &&
159                matrix_cmp[0][3] == matrix_cmp[0][4]) {
160         MIX5(IF1, IF0);
161     } else {
162         DECLARE_ALIGNED(16, float, matrix_simd)[AC3_MAX_CHANNELS][2][4];
163         float *samp[AC3_MAX_CHANNELS];
164
165         for (j = 0; j < in_ch; j++)
166             samp[j] = samples[j] + len;
167
168         j = 2 * in_ch * sizeof(float);
169         k =     in_ch * sizeof(float);
170         __asm__ volatile (
171             "1:                                 \n"
172             "sub             $4, %1             \n"
173             "sub             $8, %0             \n"
174             "movss     (%3, %1), %%xmm4         \n"
175             "movss     (%4, %1), %%xmm5         \n"
176             "shufps          $0, %%xmm4, %%xmm4 \n"
177             "shufps          $0, %%xmm5, %%xmm5 \n"
178             "movaps      %%xmm4,   (%2, %0, 4)  \n"
179             "movaps      %%xmm5, 16(%2, %0, 4)  \n"
180             "jg              1b                 \n"
181             : "+&r"(j), "+&r"(k)
182             : "r"(matrix_simd), "r"(matrix[0]), "r"(matrix[1])
183             : "memory"
184         );
185         if (out_ch == 2) {
186             MIX_MISC(IF1);
187         } else {
188             MIX_MISC(IF0);
189         }
190     }
191 }
192
193 #endif /* HAVE_SSE_INLINE && HAVE_7REGS */
194
195 av_cold void ff_ac3dsp_init_x86(AC3DSPContext *c, int bit_exact)
196 {
197     int cpu_flags = av_get_cpu_flags();
198
199     if (EXTERNAL_MMX(cpu_flags)) {
200         c->ac3_exponent_min = ff_ac3_exponent_min_mmx;
201         c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_mmx;
202         c->ac3_lshift_int16 = ff_ac3_lshift_int16_mmx;
203         c->ac3_rshift_int32 = ff_ac3_rshift_int32_mmx;
204     }
205     if (EXTERNAL_AMD3DNOW(cpu_flags)) {
206         if (!bit_exact) {
207             c->float_to_fixed24 = ff_float_to_fixed24_3dnow;
208         }
209     }
210     if (EXTERNAL_MMXEXT(cpu_flags)) {
211         c->ac3_exponent_min = ff_ac3_exponent_min_mmxext;
212         c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_mmxext;
213         if (bit_exact) {
214             c->apply_window_int16 = ff_apply_window_int16_mmxext;
215         } else {
216             c->apply_window_int16 = ff_apply_window_int16_round_mmxext;
217         }
218     }
219     if (EXTERNAL_SSE(cpu_flags)) {
220         c->float_to_fixed24 = ff_float_to_fixed24_sse;
221     }
222     if (EXTERNAL_SSE2(cpu_flags)) {
223         c->ac3_exponent_min = ff_ac3_exponent_min_sse2;
224         c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_sse2;
225         c->float_to_fixed24 = ff_float_to_fixed24_sse2;
226         c->compute_mantissa_size = ff_ac3_compute_mantissa_size_sse2;
227         c->extract_exponents = ff_ac3_extract_exponents_sse2;
228         if (bit_exact) {
229             c->apply_window_int16 = ff_apply_window_int16_sse2;
230         }
231     }
232
233     if (EXTERNAL_SSE2_FAST(cpu_flags)) {
234         c->ac3_lshift_int16 = ff_ac3_lshift_int16_sse2;
235         c->ac3_rshift_int32 = ff_ac3_rshift_int32_sse2;
236         if (!bit_exact) {
237             c->apply_window_int16 = ff_apply_window_int16_round_sse2;
238         }
239     }
240
241     if (EXTERNAL_SSSE3(cpu_flags)) {
242         c->ac3_max_msb_abs_int16 = ff_ac3_max_msb_abs_int16_ssse3;
243         if (cpu_flags & AV_CPU_FLAG_ATOM) {
244             c->apply_window_int16 = ff_apply_window_int16_ssse3_atom;
245         } else {
246             c->extract_exponents = ff_ac3_extract_exponents_ssse3;
247             c->apply_window_int16 = ff_apply_window_int16_ssse3;
248         }
249     }
250
251 #if HAVE_SSE_INLINE && HAVE_7REGS
252     if (INLINE_SSE(cpu_flags)) {
253         c->downmix = ac3_downmix_sse;
254     }
255 #endif
256 }