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 <pulsecore/random.h>
  29 #include <pulsecore/macro.h>
  30 #include <pulsecore/g711.h>
  31 #include <pulsecore/core-util.h>
  32
  33 #include "cpu-x86.h"
  34
  35 #include "sample-util.h"
  36 #include "endianmacros.h"
  37
  38 #define VOLUME_32x16(s,v)                  /* .. |   vh  |   vl  | */                   \
  39       " pxor %%xmm4, %%xmm4          \n\t" /* .. |    0  |    0  | */                   \
  40       " punpcklwd %%xmm4, "#s"       \n\t" /* .. |    0  |   p0  | */                   \
  41       " pcmpgtw "#s", %%xmm4         \n\t" /* .. |    0  | s(p0) | */                   \
  42       " pand "#v", %%xmm4            \n\t" /* .. |    0  |  (vl) | */                   \
  43       " movdqa "#s", %%xmm5          \n\t"                                              \
  44       " pmulhuw "#v", "#s"           \n\t" /* .. |    0  | vl*p0 | */                   \
  45       " psubd %%xmm4, "#s"           \n\t" /* .. |    0  | vl*p0 | + sign correct */    \
  46       " psrld $16, "#v"              \n\t" /* .. |   p0  |    0  | */                   \
  47       " pmaddwd %%xmm5, "#v"         \n\t" /* .. |    p0 * vh    | */                   \
  48       " paddd "#s", "#v"             \n\t" /* .. |    p0 * v0    | */                   \
  49       " packssdw "#v", "#v"          \n\t" /* .. | p1*v1 | p0*v0 | */
  50
  51 #define MOD_ADD(a,b) \
  52       " add "#a", %3                 \n\t" /* channel += inc           */ \
  53       " mov %3, %4                   \n\t"                                \
  54       " sub "#b", %4                 \n\t" /* tmp = channel - channels */ \
  55       " cmovae %4, %3                \n\t" /* if (tmp >= 0) channel = tmp  */
  56
  57 /* swap 16 bits */
  58 #define SWAP_16(s) \
  59       " movdqa "#s", %%xmm4          \n\t" /* .. |  h  l |  */ \
  60       " psrlw $8, %%xmm4             \n\t" /* .. |  0  h |  */ \
  61       " psllw $8, "#s"               \n\t" /* .. |  l  0 |  */ \
  62       " por %%xmm4, "#s"             \n\t" /* .. |  l  h |  */
  63
  64 /* swap 2 registers 16 bits for better pairing */
  65 #define SWAP_16_2(s1,s2) \
  66       " movdqa "#s1", %%xmm4         \n\t" /* .. |  h  l |  */ \
  67       " movdqa "#s2", %%xmm5         \n\t"                     \
  68       " psrlw $8, %%xmm4             \n\t" /* .. |  0  h |  */ \
  69       " psrlw $8, %%xmm5             \n\t"                     \
  70       " psllw $8, "#s1"              \n\t" /* .. |  l  0 |  */ \
  71       " psllw $8, "#s2"              \n\t"                     \
  72       " por %%xmm4, "#s1"            \n\t" /* .. |  l  h |  */ \
  73       " por %%xmm5, "#s2"            \n\t"
  74
  75 static void
  76 pa_volume_s16ne_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
  77 {
  78   pa_reg_x86 channel, temp;
  79
  80   /* the max number of samples we process at a time, this is also the max amount
  81    * we overread the volume array, which should have enough padding. */
  82   channels = MAX (8, channels);
  83
  84   __asm__ __volatile__ (
  85     " xor %3, %3                    \n\t"
  86     " sar $1, %2                    \n\t" /* length /= sizeof (int16_t) */
  87
  88     " test $1, %2                   \n\t" /* check for odd samples */
  89     " je 2f                         \n\t"
  90
  91     " movd (%1, %3, 4), %%xmm0      \n\t" /* |  v0h  |  v0l  | */
  92     " movw (%0), %4                 \n\t" /*     ..  |   p0  | */
  93     " movd %4, %%xmm1               \n\t"
  94     VOLUME_32x16 (%%xmm1, %%xmm0)
  95     " movd %%xmm0, %4               \n\t" /*     ..  | p0*v0 | */
  96     " movw %4, (%0)                 \n\t"
  97     " add $2, %0                    \n\t"
  98     MOD_ADD ($1, %5)
  99
 100     "2:                             \n\t"
 101     " sar $1, %2                    \n\t" /* prepare for processing 2 samples at a time */
 102     " test $1, %2                   \n\t"
 103     " je 4f                         \n\t"
 104
 105     "3:                             \n\t" /* do samples in groups of 2 */
 106     " movq (%1, %3, 4), %%xmm0      \n\t" /* |  v1h  |  v1l  |  v0h  |  v0l  | */
 107     " movd (%0), %%xmm1             \n\t" /*              .. |   p1  |  p0   | */
 108     VOLUME_32x16 (%%xmm1, %%xmm0)
 109     " movd %%xmm0, (%0)             \n\t" /*              .. | p1*v1 | p0*v0 | */
 110     " add $4, %0                    \n\t"
 111     MOD_ADD ($2, %5)
 112
 113     "4:                             \n\t"
 114     " sar $1, %2                    \n\t" /* prepare for processing 4 samples at a time */
 115     " test $1, %2                   \n\t"
 116     " je 6f                         \n\t"
 117
 118     "5:                             \n\t" /* do samples in groups of 4 */
 119     " movdqu (%1, %3, 4), %%xmm0    \n\t" /* |  v3h  |  v3l  ..  v0h  |  v0l  | */
 120     " movq (%0), %%xmm1             \n\t" /*              .. |   p3  ..  p0   | */
 121     VOLUME_32x16 (%%xmm1, %%xmm0)
 122     " movq %%xmm0, (%0)             \n\t" /*              .. | p3*v3 .. p0*v0 | */
 123     " add $8, %0                    \n\t"
 124     MOD_ADD ($4, %5)
 125
 126     "6:                             \n\t"
 127     " sar $1, %2                    \n\t" /* prepare for processing 8 samples at a time */
 128     " cmp $0, %2                    \n\t"
 129     " je 8f                         \n\t"
 130
 131     "7:                             \n\t" /* do samples in groups of 8 */
 132     " movdqu (%1, %3, 4), %%xmm0    \n\t" /* |  v3h  |  v3l  ..  v0h  |  v0l  | */
 133     " movdqu 16(%1, %3, 4), %%xmm2  \n\t" /* |  v7h  |  v7l  ..  v4h  |  v4l  | */
 134     " movq (%0), %%xmm1             \n\t" /*              .. |   p3  ..  p0   | */
 135     " movq 8(%0), %%xmm3            \n\t" /*              .. |   p7  ..  p4   | */
 136     VOLUME_32x16 (%%xmm1, %%xmm0)
 137     VOLUME_32x16 (%%xmm3, %%xmm2)
 138     " movq %%xmm0, (%0)             \n\t" /*              .. | p3*v3 .. p0*v0 | */
 139     " movq %%xmm2, 8(%0)            \n\t" /*              .. | p7*v7 .. p4*v4 | */
 140     " add $16, %0                   \n\t"
 141     MOD_ADD ($8, %5)
 142     " dec %2                        \n\t"
 143     " jne 7b                        \n\t"
 144     "8:                             \n\t"
 145
 146     : "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp)
 147     : "r" ((pa_reg_x86)channels)
 148     : "cc"
 149   );
 150 }
 151
 152 static void
 153 pa_volume_s16re_sse (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
 154 {
 155   pa_reg_x86 channel, temp;
 156
 157   /* the max number of samples we process at a time, this is also the max amount
 158    * we overread the volume array, which should have enough padding. */
 159   channels = MAX (8, channels);
 160
 161   __asm__ __volatile__ (
 162     " xor %3, %3                    \n\t"
 163     " sar $1, %2                    \n\t" /* length /= sizeof (int16_t) */
 164
 165     " test $1, %2                   \n\t" /* check for odd samples */
 166     " je 2f                         \n\t"
 167
 168     " movd (%1, %3, 4), %%xmm0      \n\t" /* |  v0h  |  v0l  | */
 169     " movw (%0), %4                 \n\t" /*     ..  |   p0  | */
 170     " rorw $8, %4                   \n\t"
 171     " movd %4, %%xmm1               \n\t"
 172     VOLUME_32x16 (%%xmm1, %%xmm0)
 173     " movd %%xmm0, %4               \n\t" /*     ..  | p0*v0 | */
 174     " rorw $8, %4                   \n\t"
 175     " movw %4, (%0)                 \n\t"
 176     " add $2, %0                    \n\t"
 177     MOD_ADD ($1, %5)
 178
 179     "2:                             \n\t"
 180     " sar $1, %2                    \n\t" /* prepare for processing 2 samples at a time */
 181     " test $1, %2                   \n\t"
 182     " je 4f                         \n\t"
 183
 184     "3:                             \n\t" /* do samples in groups of 2 */
 185     " movq (%1, %3, 4), %%xmm0      \n\t" /* |  v1h  |  v1l  |  v0h  |  v0l  | */
 186     " movd (%0), %%xmm1             \n\t" /*              .. |   p1  |  p0   | */
 187     SWAP_16 (%%xmm1)
 188     VOLUME_32x16 (%%xmm1, %%xmm0)
 189     SWAP_16 (%%xmm0)
 190     " movd %%xmm0, (%0)             \n\t" /*              .. | p1*v1 | p0*v0 | */
 191     " add $4, %0                    \n\t"
 192     MOD_ADD ($2, %5)
 193
 194     "4:                             \n\t"
 195     " sar $1, %2                    \n\t" /* prepare for processing 4 samples at a time */
 196     " test $1, %2                   \n\t"
 197     " je 6f                         \n\t"
 198
 199     /* FIXME, we can do aligned access of the volume values if we can guarantee
 200      * that the array is 16 bytes aligned, we probably have to do the odd values
 201      * after this then. */
 202     "5:                             \n\t" /* do samples in groups of 4 */
 203     " movdqu (%1, %3, 4), %%xmm0    \n\t" /* |  v3h  |  v3l  ..  v0h  |  v0l  | */
 204     " movq (%0), %%xmm1             \n\t" /*              .. |   p3  ..  p0   | */
 205     SWAP_16 (%%xmm1)
 206     VOLUME_32x16 (%%xmm1, %%xmm0)
 207     SWAP_16 (%%xmm0)
 208     " movq %%xmm0, (%0)             \n\t" /*              .. | p3*v3 .. p0*v0 | */
 209     " add $8, %0                    \n\t"
 210     MOD_ADD ($4, %5)
 211
 212     "6:                             \n\t"
 213     " sar $1, %2                    \n\t" /* prepare for processing 8 samples at a time */
 214     " cmp $0, %2                    \n\t"
 215     " je 8f                         \n\t"
 216
 217     "7:                             \n\t" /* do samples in groups of 8 */
 218     " movdqu (%1, %3, 4), %%xmm0    \n\t" /* |  v3h  |  v3l  ..  v0h  |  v0l  | */
 219     " movdqu 16(%1, %3, 4), %%xmm2  \n\t" /* |  v7h  |  v7l  ..  v4h  |  v4l  | */
 220     " movq (%0), %%xmm1             \n\t" /*              .. |   p3  ..  p0   | */
 221     " movq 8(%0), %%xmm3            \n\t" /*              .. |   p7  ..  p4   | */
 222     SWAP_16_2 (%%xmm1, %%xmm3)
 223     VOLUME_32x16 (%%xmm1, %%xmm0)
 224     VOLUME_32x16 (%%xmm3, %%xmm2)
 225     SWAP_16_2 (%%xmm0, %%xmm2)
 226     " movq %%xmm0, (%0)             \n\t" /*              .. | p3*v3 .. p0*v0 | */
 227     " movq %%xmm2, 8(%0)            \n\t" /*              .. | p7*v7 .. p4*v4 | */
 228     " add $16, %0                   \n\t"
 229     MOD_ADD ($8, %5)
 230     " dec %2                        \n\t"
 231     " jne 7b                        \n\t"
 232     "8:                             \n\t"
 233
 234     : "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp)
 235     : "r" ((pa_reg_x86)channels)
 236     : "cc"
 237   );
 238 }
 239
 240 #undef RUN_TEST
 241
 242 #ifdef RUN_TEST
 243 #define CHANNELS 2
 244 #define SAMPLES 1021
 245 #define TIMES 1000
 246 #define PADDING 16
 247
 248 static void run_test (void) {
 249   int16_t samples[SAMPLES];
 250   int16_t samples_ref[SAMPLES];
 251   int16_t samples_orig[SAMPLES];
 252   int32_t volumes[CHANNELS + PADDING];
 253   int i, j, padding;
 254   pa_do_volume_func_t func;
 255   struct timeval start, stop;
 256
 257   func = pa_get_volume_func (PA_SAMPLE_S16NE);
 258
 259   printf ("checking SSE %zd\n", sizeof (samples));
 260
 261   pa_random (samples, sizeof (samples));
 262   memcpy (samples_ref, samples, sizeof (samples));
 263   memcpy (samples_orig, samples, sizeof (samples));
 264
 265   for (i = 0; i < CHANNELS; i++)
 266     volumes[i] = rand() >> 1;
 267   for (padding = 0; padding < PADDING; padding++, i++)
 268     volumes[i] = volumes[padding];
 269
 270   func (samples_ref, volumes, CHANNELS, sizeof (samples));
 271   pa_volume_s16ne_sse (samples, volumes, CHANNELS, sizeof (samples));
 272   for (i = 0; i < SAMPLES; i++) {
 273     if (samples[i] != samples_ref[i]) {
 274       printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i],
 275               samples_orig[i], volumes[i % CHANNELS]);
 276     }
 277   }
 278
 279   pa_gettimeofday(&start);
 280   for (j = 0; j < TIMES; j++) {
 281     memcpy (samples, samples_orig, sizeof (samples));
 282     pa_volume_s16ne_sse (samples, volumes, CHANNELS, sizeof (samples));
 283   }
 284   pa_gettimeofday(&stop);
 285   pa_log_info("SSE: %llu usec.", (long long unsigned int)pa_timeval_diff (&stop, &start));
 286
 287   pa_gettimeofday(&start);
 288   for (j = 0; j < TIMES; j++) {
 289     memcpy (samples_ref, samples_orig, sizeof (samples));
 290     func (samples_ref, volumes, CHANNELS, sizeof (samples));
 291   }
 292   pa_gettimeofday(&stop);
 293   pa_log_info("ref: %llu usec.", (long long unsigned int)pa_timeval_diff (&stop, &start));
 294 }
 295 #endif
 296
 297 void pa_volume_func_init_sse (pa_cpu_x86_flag_t flags) {
 298   pa_log_info("Initialising SSE optimized functions.");
 299
 300 #ifdef RUN_TEST
 301   run_test ();
 302 #endif
 303
 304   pa_set_volume_func (PA_SAMPLE_S16NE,     (pa_do_volume_func_t) pa_volume_s16ne_sse);
 305   pa_set_volume_func (PA_SAMPLE_S16RE,     (pa_do_volume_func_t) pa_volume_s16re_sse);
 306 }