]>
code.delx.au - pulseaudio/blob - src/pulsecore/svolume_mmx.c
2 This file is part of PulseAudio.
4 Copyright 2004-2006 Lennart Poettering
5 Copyright 2009 Wim Taymans <wim.taymans@collabora.co.uk>
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.
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.
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
27 #include <pulse/timeval.h>
28 #include <pulse/rtclock.h>
30 #include <pulsecore/random.h>
31 #include <pulsecore/macro.h>
32 #include <pulsecore/g711.h>
33 #include <pulsecore/core-util.h>
34 #include <pulsecore/endianmacros.h>
38 #include "sample-util.h"
40 #if defined (__i386__) || defined (__amd64__)
41 /* in s: 2 int16_t samples
42 * in v: 2 int32_t volumes, fixed point 16:16
43 * out s: contains scaled and clamped int16_t samples.
45 * We calculate the high 32 bits of a 32x16 multiply which we then
46 * clamp to 16 bits. The calulcation is:
50 * s = ((s * vl) >> 16) + (s * vh);
52 * For the first multiply we have to do a sign correction as we need to
53 * multiply a signed int with an unsigned int. Hacker's delight 8-3 gives a
54 * simple formula to correct the sign of the high word after the signed
57 #define VOLUME_32x16(s,v) /* .. | vh | vl | */ \
58 " pxor %%mm4, %%mm4 \n\t" /* .. | 0 | 0 | */ \
59 " punpcklwd %%mm4, "#s" \n\t" /* .. | 0 | p0 | */ \
60 " pcmpgtw "#v", %%mm4 \n\t" /* .. | 0 | s(vl) | */ \
61 " pand "#s", %%mm4 \n\t" /* .. | 0 | (p0) | (vl >> 15) & p */ \
62 " movq "#s", %%mm5 \n\t" \
63 " pmulhw "#v", "#s" \n\t" /* .. | 0 | vl*p0 | */ \
64 " paddw %%mm4, "#s" \n\t" /* .. | 0 | vl*p0 | + sign correct */ \
65 " pslld $16, "#s" \n\t" /* .. | vl*p0 | 0 | */ \
66 " psrld $16, "#v" \n\t" /* .. | 0 | vh | */ \
67 " psrad $16, "#s" \n\t" /* .. | vl*p0 | sign extend */ \
68 " pmaddwd %%mm5, "#v" \n\t" /* .. | p0 * vh | */ \
69 " paddd "#s", "#v" \n\t" /* .. | p0 * v0 | */ \
70 " packssdw "#v", "#v" \n\t" /* .. | p1*v1 | p0*v0 | */
72 /* approximately advances %3 = (%3 + a) % b. This function requires that
74 #define MOD_ADD(a,b) \
75 " add "#a", %3 \n\t" \
77 " sub "#b", %4 \n\t" \
82 " movq "#s", %%mm4 \n\t" /* .. | h l | */ \
83 " psrlw $8, %%mm4 \n\t" /* .. | 0 h | */ \
84 " psllw $8, "#s" \n\t" /* .. | l 0 | */ \
85 " por %%mm4, "#s" \n\t" /* .. | l h | */
87 /* swap 2 registers 16 bits for better pairing */
88 #define SWAP_16_2(s1,s2) \
89 " movq "#s1", %%mm4 \n\t" /* .. | h l | */ \
90 " movq "#s2", %%mm5 \n\t" \
91 " psrlw $8, %%mm4 \n\t" /* .. | 0 h | */ \
92 " psrlw $8, %%mm5 \n\t" \
93 " psllw $8, "#s1" \n\t" /* .. | l 0 | */ \
94 " psllw $8, "#s2" \n\t" \
95 " por %%mm4, "#s1" \n\t" /* .. | l h | */ \
96 " por %%mm5, "#s2" \n\t"
98 static void pa_volume_s16ne_mmx(int16_t *samples
, int32_t *volumes
, unsigned channels
, unsigned length
) {
99 pa_reg_x86 channel
, temp
;
101 /* Channels must be at least 4, and always a multiple of the original number.
102 * This is also the max amount we overread the volume array, which should
103 * have enough padding. */
104 channels
= channels
== 3 ? 6 : PA_MAX (4U, channels
);
106 __asm__
__volatile__ (
108 " sar $1, %2 \n\t" /* length /= sizeof (int16_t) */
110 " test $1, %2 \n\t" /* check for odd samples */
113 " movd (%1, %3, 4), %%mm0 \n\t" /* | v0h | v0l | */
114 " movw (%0), %w4 \n\t" /* .. | p0 | */
115 " movd %4, %%mm1 \n\t"
116 VOLUME_32x16 (%%mm1
, %%mm0
)
117 " movd %%mm0, %4 \n\t" /* .. | p0*v0 | */
118 " movw %w4, (%0) \n\t"
123 " sar $1, %2 \n\t" /* prepare for processing 2 samples at a time */
124 " test $1, %2 \n\t" /* check for odd samples */
127 "3: \n\t" /* do samples in groups of 2 */
128 " movq (%1, %3, 4), %%mm0 \n\t" /* | v1h | v1l | v0h | v0l | */
129 " movd (%0), %%mm1 \n\t" /* .. | p1 | p0 | */
130 VOLUME_32x16 (%%mm1
, %%mm0
)
131 " movd %%mm0, (%0) \n\t" /* .. | p1*v1 | p0*v0 | */
136 " sar $1, %2 \n\t" /* prepare for processing 4 samples at a time */
140 "5: \n\t" /* do samples in groups of 4 */
141 " movq (%1, %3, 4), %%mm0 \n\t" /* | v1h | v1l | v0h | v0l | */
142 " movq 8(%1, %3, 4), %%mm2 \n\t" /* | v3h | v3l | v2h | v2l | */
143 " movd (%0), %%mm1 \n\t" /* .. | p1 | p0 | */
144 " movd 4(%0), %%mm3 \n\t" /* .. | p3 | p2 | */
145 VOLUME_32x16 (%%mm1
, %%mm0
)
146 VOLUME_32x16 (%%mm3
, %%mm2
)
147 " movd %%mm0, (%0) \n\t" /* .. | p1*v1 | p0*v0 | */
148 " movd %%mm2, 4(%0) \n\t" /* .. | p3*v3 | p2*v2 | */
157 : "+r" (samples
), "+r" (volumes
), "+r" (length
), "=D" (channel
), "=&r" (temp
)
158 #if defined (__i386__)
159 : "m" ((pa_reg_x86
)channels
)
161 : "r" ((pa_reg_x86
)channels
)
167 static void pa_volume_s16re_mmx(int16_t *samples
, int32_t *volumes
, unsigned channels
, unsigned length
) {
168 pa_reg_x86 channel
, temp
;
170 /* Channels must be at least 4, and always a multiple of the original number.
171 * This is also the max amount we overread the volume array, which should
172 * have enough padding. */
173 channels
= channels
== 3 ? 6 : PA_MAX (4U, channels
);
175 __asm__
__volatile__ (
177 " sar $1, %2 \n\t" /* length /= sizeof (int16_t) */
178 " pcmpeqw %%mm6, %%mm6 \n\t" /* .. | ffff | ffff | */
179 " pcmpeqw %%mm7, %%mm7 \n\t" /* .. | ffff | ffff | */
180 " pslld $16, %%mm6 \n\t" /* .. | ffff | 0 | */
181 " psrld $31, %%mm7 \n\t" /* .. | 0 | 1 | */
183 " test $1, %2 \n\t" /* check for odd samples */
186 " movd (%1, %3, 4), %%mm0 \n\t" /* | v0h | v0l | */
187 " movw (%0), %w4 \n\t" /* .. | p0 | */
189 " movd %4, %%mm1 \n\t"
190 VOLUME_32x16 (%%mm1
, %%mm0
)
191 " movd %%mm0, %4 \n\t" /* .. | p0*v0 | */
193 " movw %w4, (%0) \n\t"
198 " sar $1, %2 \n\t" /* prepare for processing 2 samples at a time */
199 " test $1, %2 \n\t" /* check for odd samples */
202 "3: \n\t" /* do samples in groups of 2 */
203 " movq (%1, %3, 4), %%mm0 \n\t" /* | v1h | v1l | v0h | v0l | */
204 " movd (%0), %%mm1 \n\t" /* .. | p1 | p0 | */
206 VOLUME_32x16 (%%mm1
, %%mm0
)
208 " movd %%mm0, (%0) \n\t" /* .. | p1*v1 | p0*v0 | */
213 " sar $1, %2 \n\t" /* prepare for processing 4 samples at a time */
217 "5: \n\t" /* do samples in groups of 4 */
218 " movq (%1, %3, 4), %%mm0 \n\t" /* | v1h | v1l | v0h | v0l | */
219 " movq 8(%1, %3, 4), %%mm2 \n\t" /* | v3h | v3l | v2h | v2l | */
220 " movd (%0), %%mm1 \n\t" /* .. | p1 | p0 | */
221 " movd 4(%0), %%mm3 \n\t" /* .. | p3 | p2 | */
222 SWAP_16_2 (%%mm1
, %%mm3
)
223 VOLUME_32x16 (%%mm1
, %%mm0
)
224 VOLUME_32x16 (%%mm3
, %%mm2
)
225 SWAP_16_2 (%%mm0
, %%mm2
)
226 " movd %%mm0, (%0) \n\t" /* .. | p1*v1 | p0*v0 | */
227 " movd %%mm2, 4(%0) \n\t" /* .. | p3*v3 | p2*v2 | */
236 : "+r" (samples
), "+r" (volumes
), "+r" (length
), "=D" (channel
), "=&r" (temp
)
237 #if defined (__i386__)
238 : "m" ((pa_reg_x86
)channels
)
240 : "r" ((pa_reg_x86
)channels
)
255 static void run_test(void) {
256 int16_t samples
[SAMPLES
];
257 int16_t samples_ref
[SAMPLES
];
258 int16_t samples_orig
[SAMPLES
];
259 int32_t volumes
[CHANNELS
+ PADDING
];
261 pa_do_volume_func_t func
;
262 pa_usec_t start
, stop
;
264 pa_usec_t min
= INT_MAX
, max
= 0;
265 double s1
= 0, s2
= 0;
267 func
= pa_get_volume_func(PA_SAMPLE_S16NE
);
269 printf("checking MMX %zd\n", sizeof(samples
));
271 pa_random(samples
, sizeof(samples
));
272 /* for (i = 0; i < SAMPLES; i++)
274 memcpy(samples_ref
, samples
, sizeof(samples
));
275 memcpy(samples_orig
, samples
, sizeof(samples
));
277 for (i
= 0; i
< CHANNELS
; i
++)
278 volumes
[i
] = PA_CLAMP_VOLUME(rand() >> 1);
279 /* volumes[i] = 0x0000ffff; */
280 for (padding
= 0; padding
< PADDING
; padding
++, i
++)
281 volumes
[i
] = volumes
[padding
];
283 func(samples_ref
, volumes
, CHANNELS
, sizeof(samples
));
284 pa_volume_s16ne_mmx(samples
, volumes
, CHANNELS
, sizeof(samples
));
285 for (i
= 0; i
< SAMPLES
; i
++) {
286 if (samples
[i
] != samples_ref
[i
]) {
287 printf("%d: %04x != %04x (%04x * %08x)\n", i
, samples
[i
], samples_ref
[i
],
288 samples_orig
[i
], volumes
[i
% CHANNELS
]);
292 for (k
= 0; k
< TIMES2
; k
++) {
293 start
= pa_rtclock_now();
294 for (j
= 0; j
< TIMES
; j
++) {
295 memcpy(samples
, samples_orig
, sizeof(samples
));
296 pa_volume_s16ne_mmx(samples
, volumes
, CHANNELS
, sizeof(samples
));
298 stop
= pa_rtclock_now();
300 if (min
> (stop
- start
)) min
= stop
- start
;
301 if (max
< (stop
- start
)) max
= stop
- start
;
303 s2
+= (stop
- start
) * (stop
- start
);
305 pa_log_info("MMX: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1
,
306 (long long unsigned int)min
, (long long unsigned int)max
, sqrt(TIMES2
* s2
- s1
* s1
) / TIMES2
);
308 min
= INT_MAX
; max
= 0;
310 for (k
= 0; k
< TIMES2
; k
++) {
311 start
= pa_rtclock_now();
312 for (j
= 0; j
< TIMES
; j
++) {
313 memcpy(samples_ref
, samples_orig
, sizeof(samples
));
314 func(samples_ref
, volumes
, CHANNELS
, sizeof(samples
));
316 stop
= pa_rtclock_now();
318 if (min
> (stop
- start
)) min
= stop
- start
;
319 if (max
< (stop
- start
)) max
= stop
- start
;
321 s2
+= (stop
- start
) * (stop
- start
);
323 pa_log_info("ref: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1
,
324 (long long unsigned int)min
, (long long unsigned int)max
, sqrt(TIMES2
* s2
- s1
* s1
) / TIMES2
);
326 pa_assert_se(memcmp(samples_ref
, samples
, sizeof(samples
)) == 0);
330 #endif /* defined (__i386__) || defined (__amd64__) */
333 void pa_volume_func_init_mmx(pa_cpu_x86_flag_t flags
) {
334 #if defined (__i386__) || defined (__amd64__)
340 if ((flags
& PA_CPU_X86_MMX
) && (flags
& PA_CPU_X86_CMOV
)) {
341 pa_log_info("Initialising MMX optimized functions.");
343 pa_set_volume_func(PA_SAMPLE_S16NE
, (pa_do_volume_func_t
) pa_volume_s16ne_mmx
);
344 pa_set_volume_func(PA_SAMPLE_S16RE
, (pa_do_volume_func_t
) pa_volume_s16re_mmx
);
346 #endif /* defined (__i386__) || defined (__amd64__) */