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/timeval.h>
  28 #include <pulse/rtclock.h>
  29
  30 #include <pulsecore/random.h>
  31 #include <pulsecore/macro.h>
  32 #include <pulsecore/g711.h>
  33 #include <pulsecore/core-util.h>
  34
  35 #include "cpu-x86.h"
  36
  37 #include "sample-util.h"
  38 #include "endianmacros.h"
  39
  40 #if defined (__i386__) || defined (__amd64__)
  41
  42 #define VOLUME_32x16(s,v)                  /* .. |   vh  |   vl  | */                   \
  43       " pxor %%xmm4, %%xmm4          \n\t" /* .. |    0  |    0  | */                   \
  44       " punpcklwd %%xmm4, "#s"       \n\t" /* .. |    0  |   p0  | */                   \
  45       " pcmpgtw "#s", %%xmm4         \n\t" /* .. |    0  | s(p0) | */                   \
  46       " pand "#v", %%xmm4            \n\t" /* .. |    0  |  (vl) | */                   \
  47       " movdqa "#s", %%xmm5          \n\t"                                              \
  48       " pmulhuw "#v", "#s"           \n\t" /* .. |    0  | vl*p0 | */                   \
  49       " psubd %%xmm4, "#s"           \n\t" /* .. |    0  | vl*p0 | + sign correct */    \
  50       " psrld $16, "#v"              \n\t" /* .. |    0  |   vh  | */                   \
  51       " pmaddwd %%xmm5, "#v"         \n\t" /* .. |    p0 * vh    | */                   \
  52       " paddd "#s", "#v"             \n\t" /* .. |    p0 * v0    | */                   \
  53       " packssdw "#v", "#v"          \n\t" /* .. | p1*v1 | p0*v0 | */
  54
  55 #define MOD_ADD(a,b) \
  56       " add "#a", %3                 \n\t" /* channel += inc           */ \
  57       " mov %3, %4                   \n\t"                                \
  58       " sub "#b", %4                 \n\t" /* tmp = channel - channels */ \
  59       " cmovae %4, %3                \n\t" /* if (tmp >= 0) channel = tmp  */
  60
  61 /* swap 16 bits */
  62 #define SWAP_16(s) \
  63       " movdqa "#s", %%xmm4          \n\t" /* .. |  h  l |  */ \
  64       " psrlw $8, %%xmm4             \n\t" /* .. |  0  h |  */ \
  65       " psllw $8, "#s"               \n\t" /* .. |  l  0 |  */ \
  66       " por %%xmm4, "#s"             \n\t" /* .. |  l  h |  */
  67
  68 /* swap 2 registers 16 bits for better pairing */
  69 #define SWAP_16_2(s1,s2) \
  70       " movdqa "#s1", %%xmm4         \n\t" /* .. |  h  l |  */ \
  71       " movdqa "#s2", %%xmm5         \n\t"                     \
  72       " psrlw $8, %%xmm4             \n\t" /* .. |  0  h |  */ \
  73       " psrlw $8, %%xmm5             \n\t"                     \
  74       " psllw $8, "#s1"              \n\t" /* .. |  l  0 |  */ \
  75       " psllw $8, "#s2"              \n\t"                     \
  76       " por %%xmm4, "#s1"            \n\t" /* .. |  l  h |  */ \
  77       " por %%xmm5, "#s2"            \n\t"
  78
  79
  80 static int channel_overread_table[8] = {8,8,8,12,8,10,12,14};
  81
  82 static void pa_volume_s16ne_sse2(int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) {
  83     pa_reg_x86 channel, temp;
  84
  85     /* Channels must be at least 8 and always a multiple of the original number.
  86      * This is also the max amount we overread the volume array, which should
  87      * have enough padding. */
  88     if (channels < 8)
  89         channels = channel_overread_table[channels];
  90
  91     __asm__ __volatile__ (
  92         " xor %3, %3                    \n\t"
  93         " sar $1, %2                    \n\t" /* length /= sizeof (int16_t) */
  94
  95         " test $1, %2                   \n\t" /* check for odd samples */
  96         " je 2f                         \n\t"
  97
  98         " movd (%1, %3, 4), %%xmm0      \n\t" /* |  v0h  |  v0l  | */
  99         " movw (%0), %w4                \n\t" /*     ..  |   p0  | */
 100         " movd %4, %%xmm1               \n\t"
 101         VOLUME_32x16 (%%xmm1, %%xmm0)
 102         " movd %%xmm0, %4               \n\t" /*     ..  | p0*v0 | */
 103         " movw %w4, (%0)                \n\t"
 104         " add $2, %0                    \n\t"
 105         MOD_ADD ($1, %5)
 106
 107         "2:                             \n\t"
 108         " sar $1, %2                    \n\t" /* prepare for processing 2 samples at a time */
 109         " test $1, %2                   \n\t"
 110         " je 4f                         \n\t"
 111
 112         "3:                             \n\t" /* do samples in groups of 2 */
 113         " movq (%1, %3, 4), %%xmm0      \n\t" /* |  v1h  |  v1l  |  v0h  |  v0l  | */
 114         " movd (%0), %%xmm1             \n\t" /*              .. |   p1  |  p0   | */
 115         VOLUME_32x16 (%%xmm1, %%xmm0)
 116         " movd %%xmm0, (%0)             \n\t" /*              .. | p1*v1 | p0*v0 | */
 117         " add $4, %0                    \n\t"
 118         MOD_ADD ($2, %5)
 119
 120         "4:                             \n\t"
 121         " sar $1, %2                    \n\t" /* prepare for processing 4 samples at a time */
 122         " test $1, %2                   \n\t"
 123         " je 6f                         \n\t"
 124
 125         /* FIXME, we can do aligned access of the volume values if we can guarantee
 126          * that the array is 16 bytes aligned, we probably have to do the odd values
 127          * after this then. */
 128         "5:                             \n\t" /* do samples in groups of 4 */
 129         " movdqu (%1, %3, 4), %%xmm0    \n\t" /* |  v3h  |  v3l  ..  v0h  |  v0l  | */
 130         " movq (%0), %%xmm1             \n\t" /*              .. |   p3  ..  p0   | */
 131         VOLUME_32x16 (%%xmm1, %%xmm0)
 132         " movq %%xmm0, (%0)             \n\t" /*              .. | p3*v3 .. p0*v0 | */
 133         " add $8, %0                    \n\t"
 134         MOD_ADD ($4, %5)
 135
 136         "6:                             \n\t"
 137         " sar $1, %2                    \n\t" /* prepare for processing 8 samples at a time */
 138         " cmp $0, %2                    \n\t"
 139         " je 8f                         \n\t"
 140
 141         "7:                             \n\t" /* do samples in groups of 8 */
 142         " movdqu (%1, %3, 4), %%xmm0    \n\t" /* |  v3h  |  v3l  ..  v0h  |  v0l  | */
 143         " movdqu 16(%1, %3, 4), %%xmm2  \n\t" /* |  v7h  |  v7l  ..  v4h  |  v4l  | */
 144         " movq (%0), %%xmm1             \n\t" /*              .. |   p3  ..  p0   | */
 145         " movq 8(%0), %%xmm3            \n\t" /*              .. |   p7  ..  p4   | */
 146         VOLUME_32x16 (%%xmm1, %%xmm0)
 147         VOLUME_32x16 (%%xmm3, %%xmm2)
 148         " movq %%xmm0, (%0)             \n\t" /*              .. | p3*v3 .. p0*v0 | */
 149         " movq %%xmm2, 8(%0)            \n\t" /*              .. | p7*v7 .. p4*v4 | */
 150         " add $16, %0                   \n\t"
 151         MOD_ADD ($8, %5)
 152         " dec %2                        \n\t"
 153         " jne 7b                        \n\t"
 154         "8:                             \n\t"
 155
 156         : "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp)
 157         : "rm" ((pa_reg_x86)channels)
 158         : "cc"
 159     );
 160 }
 161
 162 static void pa_volume_s16re_sse2(int16_t *samples, 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 (%1, %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 (%1, %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 (%1, %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 (%1, %3, 4), %%xmm0    \n\t" /* |  v3h  |  v3l  ..  v0h  |  v0l  | */
 229         " movdqu 16(%1, %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         : "rm" ((pa_reg_x86)channels)
 246         : "cc"
 247     );
 248 }
 249
 250 #undef RUN_TEST
 251
 252 #ifdef RUN_TEST
 253 #define CHANNELS 2
 254 #define SAMPLES 1022
 255 #define TIMES 1000
 256 #define PADDING 16
 257
 258 static void run_test(void) {
 259     int16_t samples[SAMPLES];
 260     int16_t samples_ref[SAMPLES];
 261     int16_t samples_orig[SAMPLES];
 262     int32_t volumes[CHANNELS + PADDING];
 263     int i, j, padding;
 264     pa_do_volume_func_t func;
 265     pa_usec_t start, stop;
 266
 267     func = pa_get_volume_func(PA_SAMPLE_S16NE);
 268
 269     printf("checking SSE2 %zd\n", sizeof(samples));
 270
 271     pa_random(samples, sizeof(samples));
 272     memcpy(samples_ref, samples, sizeof(samples));
 273     memcpy(samples_orig, samples, sizeof(samples));
 274
 275     for (i = 0; i < CHANNELS; i++)
 276         volumes[i] = PA_CLAMP_VOLUME(rand() >> 1);
 277     for (padding = 0; padding < PADDING; padding++, i++)
 278         volumes[i] = volumes[padding];
 279
 280     func(samples_ref, volumes, CHANNELS, sizeof(samples));
 281     pa_volume_s16ne_sse2(samples, volumes, CHANNELS, sizeof(samples));
 282     for (i = 0; i < SAMPLES; i++) {
 283         if (samples[i] != samples_ref[i]) {
 284             printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i],
 285                       samples_orig[i], volumes[i % CHANNELS]);
 286         }
 287     }
 288
 289     start = pa_rtclock_now();
 290     for (j = 0; j < TIMES; j++) {
 291         memcpy(samples, samples_orig, sizeof(samples));
 292         pa_volume_s16ne_sse2(samples, volumes, CHANNELS, sizeof(samples));
 293     }
 294     stop = pa_rtclock_now();
 295     pa_log_info("SSE: %llu usec.", (long long unsigned int)(stop - start));
 296
 297     start = pa_rtclock_now();
 298     for (j = 0; j < TIMES; j++) {
 299         memcpy(samples_ref, samples_orig, sizeof(samples));
 300         func(samples_ref, volumes, CHANNELS, sizeof (samples));
 301     }
 302     stop = pa_rtclock_now();
 303     pa_log_info("ref: %llu usec.", (long long unsigned int)(stop - start));
 304
 305     pa_assert_se(memcmp(samples_ref, samples, sizeof(samples)) == 0);
 306 }
 307 #endif
 308 #endif /* defined (__i386__) || defined (__amd64__) */
 309
 310 void pa_volume_func_init_sse(pa_cpu_x86_flag_t flags) {
 311 #if defined (__i386__) || defined (__amd64__)
 312
 313 #ifdef RUN_TEST
 314     run_test();
 315 #endif
 316
 317     if (flags & PA_CPU_X86_SSE2) {
 318         pa_log_info("Initialising SSE2 optimized functions.");
 319
 320         pa_set_volume_func(PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_sse2);
 321         pa_set_volume_func(PA_SAMPLE_S16RE, (pa_do_volume_func_t) pa_volume_s16re_sse2);
 322     }
 323 #endif /* defined (__i386__) || defined (__amd64__) */
 324 }