code.delx.au - pulseaudio/blob - src/pulsecore/svolume_sse.c

   1 /***
   2   This file is part of PulseAudio.
   3
   4   Copyright 2004-2006 Lennart Poettering
   5   Copyright 2009 Wim Taymans <wim.taymans@collabora.co.uk>
   6
   7   PulseAudio is free software; you can redistribute it and/or modify
   8   it under the terms of the GNU Lesser General Public License as published
   9   by the Free Software Foundation; either version 2.1 of the License,
  10   or (at your option) any later version.
  11
  12   PulseAudio is distributed in the hope that it will be useful, but
  13   WITHOUT ANY WARRANTY; without even the implied warranty of
  14   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU
  15   General Public License for more details.
  16
  17   You should have received a copy of the GNU Lesser General Public License
  18   along with PulseAudio; if not, write to the Free Software
  19   Foundation, Inc., 59 Temple Place, Suite 330, Boston, MA 02111-1307
  20   USA.
  21 ***/
  22
  23 #ifdef HAVE_CONFIG_H
  24 #include <config.h>
  25 #endif
  26
  27 #include <pulse/rtclock.h>
  28
  29 #include <pulsecore/random.h>
  30 #include <pulsecore/macro.h>
  31 #include <pulsecore/endianmacros.h>
  32
  33 #include "cpu-x86.h"
  34
  35 #include "sample-util.h"
  36
  37 #if (!defined(__FreeBSD__) && defined (__i386__)) || defined (__amd64__)
  38
  39 #define VOLUME_32x16(s,v)                  /* .. |   vh  |   vl  | */                   \
  40       " pxor %%xmm4, %%xmm4          \n\t" /* .. |    0  |    0  | */                   \
  41       " punpcklwd %%xmm4, "#s"       \n\t" /* .. |    0  |   p0  | */                   \
  42       " pcmpgtw "#s", %%xmm4         \n\t" /* .. |    0  | s(p0) | */                   \
  43       " pand "#v", %%xmm4            \n\t" /* .. |    0  |  (vl) | */                   \
  44       " movdqa "#s", %%xmm5          \n\t"                                              \
  45       " pmulhuw "#v", "#s"           \n\t" /* .. |    0  | vl*p0 | */                   \
  46       " psubd %%xmm4, "#s"           \n\t" /* .. |    0  | vl*p0 | + sign correct */    \
  47       " psrld $16, "#v"              \n\t" /* .. |    0  |   vh  | */                   \
  48       " pmaddwd %%xmm5, "#v"         \n\t" /* .. |    p0 * vh    | */                   \
  49       " paddd "#s", "#v"             \n\t" /* .. |    p0 * v0    | */                   \
  50       " packssdw "#v", "#v"          \n\t" /* .. | p1*v1 | p0*v0 | */
  51
  52 #define MOD_ADD(a,b) \
  53       " add "#a", %3                 \n\t" /* channel += inc           */ \
  54       " mov %3, %4                   \n\t"                                \
  55       " sub "#b", %4                 \n\t" /* tmp = channel - channels */ \
  56       " cmovae %4, %3                \n\t" /* if (tmp >= 0) channel = tmp  */
  57
  58 /* swap 16 bits */
  59 #define SWAP_16(s) \
  60       " movdqa "#s", %%xmm4          \n\t" /* .. |  h  l |  */ \
  61       " psrlw $8, %%xmm4             \n\t" /* .. |  0  h |  */ \
  62       " psllw $8, "#s"               \n\t" /* .. |  l  0 |  */ \
  63       " por %%xmm4, "#s"             \n\t" /* .. |  l  h |  */
  64
  65 /* swap 2 registers 16 bits for better pairing */
  66 #define SWAP_16_2(s1,s2) \
  67       " movdqa "#s1", %%xmm4         \n\t" /* .. |  h  l |  */ \
  68       " movdqa "#s2", %%xmm5         \n\t"                     \
  69       " psrlw $8, %%xmm4             \n\t" /* .. |  0  h |  */ \
  70       " psrlw $8, %%xmm5             \n\t"                     \
  71       " psllw $8, "#s1"              \n\t" /* .. |  l  0 |  */ \
  72       " psllw $8, "#s2"              \n\t"                     \
  73       " por %%xmm4, "#s1"            \n\t" /* .. |  l  h |  */ \
  74       " por %%xmm5, "#s2"            \n\t"
  75
  76 static int channel_overread_table[8] = {8,8,8,12,8,10,12,14};
  77
  78 static void pa_volume_s16ne_sse2(int16_t *samples, const int32_t *volumes, unsigned channels, unsigned length) {
  79     pa_reg_x86 channel, temp;
  80
  81     /* Channels must be at least 8 and always a multiple of the original number.
  82      * This is also the max amount we overread the volume array, which should
  83      * have enough padding. */
  84     if (channels < 8)
  85         channels = channel_overread_table[channels];
  86
  87     __asm__ __volatile__ (
  88         " xor %3, %3                    \n\t"
  89         " sar $1, %2                    \n\t" /* length /= sizeof (int16_t) */
  90
  91         " test $1, %2                   \n\t" /* check for odd samples */
  92         " je 2f                         \n\t"
  93
  94         " movd (%q1, %3, 4), %%xmm0     \n\t" /* |  v0h  |  v0l  | */
  95         " movw (%0), %w4                \n\t" /*     ..  |   p0  | */
  96         " movd %4, %%xmm1               \n\t"
  97         VOLUME_32x16 (%%xmm1, %%xmm0)
  98         " movd %%xmm0, %4               \n\t" /*     ..  | p0*v0 | */
  99         " movw %w4, (%0)                \n\t"
 100         " add $2, %0                    \n\t"
 101         MOD_ADD ($1, %5)
 102
 103         "2:                             \n\t"
 104         " sar $1, %2                    \n\t" /* prepare for processing 2 samples at a time */
 105         " test $1, %2                   \n\t"
 106         " je 4f                         \n\t"
 107
 108         "3:                             \n\t" /* do samples in groups of 2 */
 109         " movq (%q1, %3, 4), %%xmm0     \n\t" /* |  v1h  |  v1l  |  v0h  |  v0l  | */
 110         " movd (%0), %%xmm1             \n\t" /*              .. |   p1  |  p0   | */
 111         VOLUME_32x16 (%%xmm1, %%xmm0)
 112         " movd %%xmm0, (%0)             \n\t" /*              .. | p1*v1 | p0*v0 | */
 113         " add $4, %0                    \n\t"
 114         MOD_ADD ($2, %5)
 115
 116         "4:                             \n\t"
 117         " sar $1, %2                    \n\t" /* prepare for processing 4 samples at a time */
 118         " test $1, %2                   \n\t"
 119         " je 6f                         \n\t"
 120
 121         /* FIXME, we can do aligned access of the volume values if we can guarantee
 122          * that the array is 16 bytes aligned, we probably have to do the odd values
 123          * after this then. */
 124         "5:                             \n\t" /* do samples in groups of 4 */
 125         " movdqu (%q1, %3, 4), %%xmm0   \n\t" /* |  v3h  |  v3l  ..  v0h  |  v0l  | */
 126         " movq (%0), %%xmm1             \n\t" /*              .. |   p3  ..  p0   | */
 127         VOLUME_32x16 (%%xmm1, %%xmm0)
 128         " movq %%xmm0, (%0)             \n\t" /*              .. | p3*v3 .. p0*v0 | */
 129         " add $8, %0                    \n\t"
 130         MOD_ADD ($4, %5)
 131
 132         "6:                             \n\t"
 133         " sar $1, %2                    \n\t" /* prepare for processing 8 samples at a time */
 134         " cmp $0, %2                    \n\t"
 135         " je 8f                         \n\t"
 136
 137         "7:                             \n\t" /* do samples in groups of 8 */
 138         " movdqu (%q1, %3, 4), %%xmm0   \n\t" /* |  v3h  |  v3l  ..  v0h  |  v0l  | */
 139         " movdqu 16(%q1, %3, 4), %%xmm2 \n\t" /* |  v7h  |  v7l  ..  v4h  |  v4l  | */
 140         " movq (%0), %%xmm1             \n\t" /*              .. |   p3  ..  p0   | */
 141         " movq 8(%0), %%xmm3            \n\t" /*              .. |   p7  ..  p4   | */
 142         VOLUME_32x16 (%%xmm1, %%xmm0)
 143         VOLUME_32x16 (%%xmm3, %%xmm2)
 144         " movq %%xmm0, (%0)             \n\t" /*              .. | p3*v3 .. p0*v0 | */
 145         " movq %%xmm2, 8(%0)            \n\t" /*              .. | p7*v7 .. p4*v4 | */
 146         " add $16, %0                   \n\t"
 147         MOD_ADD ($8, %5)
 148         " dec %2                        \n\t"
 149         " jne 7b                        \n\t"
 150         "8:                             \n\t"
 151
 152         : "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp)
 153 #if defined (__i386__)
 154         : "m" (channels)
 155 #else
 156         : "r" ((pa_reg_x86)channels)
 157 #endif
 158         : "cc"
 159     );
 160 }
 161
 162 static void pa_volume_s16re_sse2(int16_t *samples, const int32_t *volumes, unsigned channels, unsigned length) {
 163     pa_reg_x86 channel, temp;
 164
 165     /* Channels must be at least 8 and always a multiple of the original number.
 166      * This is also the max amount we overread the volume array, which should
 167      * have enough padding. */
 168     if (channels < 8)
 169         channels = channel_overread_table[channels];
 170
 171     __asm__ __volatile__ (
 172         " xor %3, %3                    \n\t"
 173         " sar $1, %2                    \n\t" /* length /= sizeof (int16_t) */
 174
 175         " test $1, %2                   \n\t" /* check for odd samples */
 176         " je 2f                         \n\t"
 177
 178         " movd (%q1, %3, 4), %%xmm0     \n\t" /* |  v0h  |  v0l  | */
 179         " movw (%0), %w4                \n\t" /*     ..  |   p0  | */
 180         " rorw $8, %w4                  \n\t"
 181         " movd %4, %%xmm1               \n\t"
 182         VOLUME_32x16 (%%xmm1, %%xmm0)
 183         " movd %%xmm0, %4               \n\t" /*     ..  | p0*v0 | */
 184         " rorw $8, %w4                  \n\t"
 185         " movw %w4, (%0)                \n\t"
 186         " add $2, %0                    \n\t"
 187         MOD_ADD ($1, %5)
 188
 189         "2:                             \n\t"
 190         " sar $1, %2                    \n\t" /* prepare for processing 2 samples at a time */
 191         " test $1, %2                   \n\t"
 192         " je 4f                         \n\t"
 193
 194         "3:                             \n\t" /* do samples in groups of 2 */
 195         " movq (%q1, %3, 4), %%xmm0     \n\t" /* |  v1h  |  v1l  |  v0h  |  v0l  | */
 196         " movd (%0), %%xmm1             \n\t" /*              .. |   p1  |  p0   | */
 197         SWAP_16 (%%xmm1)
 198         VOLUME_32x16 (%%xmm1, %%xmm0)
 199         SWAP_16 (%%xmm0)
 200         " movd %%xmm0, (%0)             \n\t" /*              .. | p1*v1 | p0*v0 | */
 201         " add $4, %0                    \n\t"
 202         MOD_ADD ($2, %5)
 203
 204         "4:                             \n\t"
 205         " sar $1, %2                    \n\t" /* prepare for processing 4 samples at a time */
 206         " test $1, %2                   \n\t"
 207         " je 6f                         \n\t"
 208
 209         /* FIXME, we can do aligned access of the volume values if we can guarantee
 210          * that the array is 16 bytes aligned, we probably have to do the odd values
 211          * after this then. */
 212         "5:                             \n\t" /* do samples in groups of 4 */
 213         " movdqu (%q1, %3, 4), %%xmm0   \n\t" /* |  v3h  |  v3l  ..  v0h  |  v0l  | */
 214         " movq (%0), %%xmm1             \n\t" /*              .. |   p3  ..  p0   | */
 215         SWAP_16 (%%xmm1)
 216         VOLUME_32x16 (%%xmm1, %%xmm0)
 217         SWAP_16 (%%xmm0)
 218         " movq %%xmm0, (%0)             \n\t" /*              .. | p3*v3 .. p0*v0 | */
 219         " add $8, %0                    \n\t"
 220         MOD_ADD ($4, %5)
 221
 222         "6:                             \n\t"
 223         " sar $1, %2                    \n\t" /* prepare for processing 8 samples at a time */
 224         " cmp $0, %2                    \n\t"
 225         " je 8f                         \n\t"
 226
 227         "7:                             \n\t" /* do samples in groups of 8 */
 228         " movdqu (%q1, %3, 4), %%xmm0   \n\t" /* |  v3h  |  v3l  ..  v0h  |  v0l  | */
 229         " movdqu 16(%q1, %3, 4), %%xmm2 \n\t" /* |  v7h  |  v7l  ..  v4h  |  v4l  | */
 230         " movq (%0), %%xmm1             \n\t" /*              .. |   p3  ..  p0   | */
 231         " movq 8(%0), %%xmm3            \n\t" /*              .. |   p7  ..  p4   | */
 232         SWAP_16_2 (%%xmm1, %%xmm3)
 233         VOLUME_32x16 (%%xmm1, %%xmm0)
 234         VOLUME_32x16 (%%xmm3, %%xmm2)
 235         SWAP_16_2 (%%xmm0, %%xmm2)
 236         " movq %%xmm0, (%0)             \n\t" /*              .. | p3*v3 .. p0*v0 | */
 237         " movq %%xmm2, 8(%0)            \n\t" /*              .. | p7*v7 .. p4*v4 | */
 238         " add $16, %0                   \n\t"
 239         MOD_ADD ($8, %5)
 240         " dec %2                        \n\t"
 241         " jne 7b                        \n\t"
 242         "8:                             \n\t"
 243
 244         : "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp)
 245 #if defined (__i386__)
 246         : "m" (channels)
 247 #else
 248         : "r" ((pa_reg_x86)channels)
 249 #endif
 250         : "cc"
 251     );
 252 }
 253
 254 #endif /* (!defined(__FreeBSD__) && defined (__i386__)) || defined (__amd64__) */
 255
 256 void pa_volume_func_init_sse(pa_cpu_x86_flag_t flags) {
 257 #if (!defined(__FreeBSD__) && defined (__i386__)) || defined (__amd64__)
 258     if (flags & PA_CPU_X86_SSE2) {
 259         pa_log_info("Initialising SSE2 optimized volume functions.");
 260
 261         pa_set_volume_func(PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_sse2);
 262         pa_set_volume_func(PA_SAMPLE_S16RE, (pa_do_volume_func_t) pa_volume_s16re_sse2);
 263     }
 264 #endif /* (!defined(__FreeBSD__) && defined (__i386__)) || defined (__amd64__) */
 265 }