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>
37 #include "sample-util.h"
38 #include "endianmacros.h"
40 #if defined (__i386__) || defined (__amd64__)
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 | */
55 #define MOD_ADD(a,b) \
56 " add "#a", %3 \n\t" /* channel += inc */ \
58 " sub "#b", %4 \n\t" /* tmp = channel - channels */ \
59 " cmovae %4, %3 \n\t" /* if (tmp >= 0) channel = tmp */
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 | */
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"
80 static int channel_overread_table[8] = {8,8,8,12,8,10,12,14};
83 pa_volume_s16ne_sse2 (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
85 pa_reg_x86 channel, temp;
87 /* Channels must be at least 8 and always a multiple of the original number.
88 * This is also the max amount we overread the volume array, which should
89 * have enough padding. */
91 channels = channel_overread_table[channels];
93 __asm__ __volatile__ (
95 " sar $1, %2 \n\t" /* length /= sizeof (int16_t) */
97 " test $1, %2 \n\t" /* check for odd samples */
100 " movd (%1, %3, 4), %%xmm0 \n\t" /* | v0h | v0l | */
101 " movw (%0), %w4 \n\t" /* .. | p0 | */
102 " movd %4, %%xmm1 \n\t"
103 VOLUME_32x16 (%%xmm1, %%xmm0)
104 " movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */
105 " movw %w4, (%0) \n\t"
110 " sar $1, %2 \n\t" /* prepare for processing 2 samples at a time */
114 "3: \n\t" /* do samples in groups of 2 */
115 " movq (%1, %3, 4), %%xmm0 \n\t" /* | v1h | v1l | v0h | v0l | */
116 " movd (%0), %%xmm1 \n\t" /* .. | p1 | p0 | */
117 VOLUME_32x16 (%%xmm1, %%xmm0)
118 " movd %%xmm0, (%0) \n\t" /* .. | p1*v1 | p0*v0 | */
123 " sar $1, %2 \n\t" /* prepare for processing 4 samples at a time */
127 /* FIXME, we can do aligned access of the volume values if we can guarantee
128 * that the array is 16 bytes aligned, we probably have to do the odd values
129 * after this then. */
130 "5: \n\t" /* do samples in groups of 4 */
131 " movdqu (%1, %3, 4), %%xmm0 \n\t" /* | v3h | v3l .. v0h | v0l | */
132 " movq (%0), %%xmm1 \n\t" /* .. | p3 .. p0 | */
133 VOLUME_32x16 (%%xmm1, %%xmm0)
134 " movq %%xmm0, (%0) \n\t" /* .. | p3*v3 .. p0*v0 | */
139 " sar $1, %2 \n\t" /* prepare for processing 8 samples at a time */
143 "7: \n\t" /* do samples in groups of 8 */
144 " movdqu (%1, %3, 4), %%xmm0 \n\t" /* | v3h | v3l .. v0h | v0l | */
145 " movdqu 16(%1, %3, 4), %%xmm2 \n\t" /* | v7h | v7l .. v4h | v4l | */
146 " movq (%0), %%xmm1 \n\t" /* .. | p3 .. p0 | */
147 " movq 8(%0), %%xmm3 \n\t" /* .. | p7 .. p4 | */
148 VOLUME_32x16 (%%xmm1, %%xmm0)
149 VOLUME_32x16 (%%xmm3, %%xmm2)
150 " movq %%xmm0, (%0) \n\t" /* .. | p3*v3 .. p0*v0 | */
151 " movq %%xmm2, 8(%0) \n\t" /* .. | p7*v7 .. p4*v4 | */
158 : "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp)
159 : "rm" ((pa_reg_x86)channels)
165 pa_volume_s16re_sse2 (int16_t *samples, int32_t *volumes, unsigned channels, unsigned length)
167 pa_reg_x86 channel, temp;
169 /* Channels must be at least 8 and always a multiple of the original number.
170 * This is also the max amount we overread the volume array, which should
171 * have enough padding. */
173 channels = channel_overread_table[channels];
175 __asm__ __volatile__ (
177 " sar $1, %2 \n\t" /* length /= sizeof (int16_t) */
179 " test $1, %2 \n\t" /* check for odd samples */
182 " movd (%1, %3, 4), %%xmm0 \n\t" /* | v0h | v0l | */
183 " movw (%0), %w4 \n\t" /* .. | p0 | */
185 " movd %4, %%xmm1 \n\t"
186 VOLUME_32x16 (%%xmm1, %%xmm0)
187 " movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */
189 " movw %w4, (%0) \n\t"
194 " sar $1, %2 \n\t" /* prepare for processing 2 samples at a time */
198 "3: \n\t" /* do samples in groups of 2 */
199 " movq (%1, %3, 4), %%xmm0 \n\t" /* | v1h | v1l | v0h | v0l | */
200 " movd (%0), %%xmm1 \n\t" /* .. | p1 | p0 | */
202 VOLUME_32x16 (%%xmm1, %%xmm0)
204 " movd %%xmm0, (%0) \n\t" /* .. | p1*v1 | p0*v0 | */
209 " sar $1, %2 \n\t" /* prepare for processing 4 samples at a time */
213 /* FIXME, we can do aligned access of the volume values if we can guarantee
214 * that the array is 16 bytes aligned, we probably have to do the odd values
215 * after this then. */
216 "5: \n\t" /* do samples in groups of 4 */
217 " movdqu (%1, %3, 4), %%xmm0 \n\t" /* | v3h | v3l .. v0h | v0l | */
218 " movq (%0), %%xmm1 \n\t" /* .. | p3 .. p0 | */
220 VOLUME_32x16 (%%xmm1, %%xmm0)
222 " movq %%xmm0, (%0) \n\t" /* .. | p3*v3 .. p0*v0 | */
227 " sar $1, %2 \n\t" /* prepare for processing 8 samples at a time */
231 "7: \n\t" /* do samples in groups of 8 */
232 " movdqu (%1, %3, 4), %%xmm0 \n\t" /* | v3h | v3l .. v0h | v0l | */
233 " movdqu 16(%1, %3, 4), %%xmm2 \n\t" /* | v7h | v7l .. v4h | v4l | */
234 " movq (%0), %%xmm1 \n\t" /* .. | p3 .. p0 | */
235 " movq 8(%0), %%xmm3 \n\t" /* .. | p7 .. p4 | */
236 SWAP_16_2 (%%xmm1, %%xmm3)
237 VOLUME_32x16 (%%xmm1, %%xmm0)
238 VOLUME_32x16 (%%xmm3, %%xmm2)
239 SWAP_16_2 (%%xmm0, %%xmm2)
240 " movq %%xmm0, (%0) \n\t" /* .. | p3*v3 .. p0*v0 | */
241 " movq %%xmm2, 8(%0) \n\t" /* .. | p7*v7 .. p4*v4 | */
248 : "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp)
249 : "rm" ((pa_reg_x86)channels)
262 static void run_test (void) {
263 int16_t samples[SAMPLES];
264 int16_t samples_ref[SAMPLES];
265 int16_t samples_orig[SAMPLES];
266 int32_t volumes[CHANNELS + PADDING];
268 pa_do_volume_func_t func;
269 pa_usec_t start, stop;
271 func = pa_get_volume_func (PA_SAMPLE_S16NE);
273 printf ("checking SSE2 %zd\n", sizeof (samples));
275 pa_random (samples, sizeof (samples));
276 memcpy (samples_ref, samples, sizeof (samples));
277 memcpy (samples_orig, samples, sizeof (samples));
279 for (i = 0; i < CHANNELS; i++)
280 volumes[i] = rand() >> 1;
281 for (padding = 0; padding < PADDING; padding++, i++)
282 volumes[i] = volumes[padding];
284 func (samples_ref, volumes, CHANNELS, sizeof (samples));
285 pa_volume_s16ne_sse2 (samples, volumes, CHANNELS, sizeof (samples));
286 for (i = 0; i < SAMPLES; i++) {
287 if (samples[i] != samples_ref[i]) {
288 printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i],
289 samples_orig[i], volumes[i % CHANNELS]);
293 start = pa_rtclock_now();
294 for (j = 0; j < TIMES; j++) {
295 memcpy (samples, samples_orig, sizeof (samples));
296 pa_volume_s16ne_sse2 (samples, volumes, CHANNELS, sizeof (samples));
298 stop = pa_rtclock_now();
299 pa_log_info("SSE: %llu usec.", (long long unsigned int)(stop - start));
301 start = pa_rtclock_now();
302 for (j = 0; j < TIMES; j++) {
303 memcpy (samples_ref, samples_orig, sizeof (samples));
304 func (samples_ref, volumes, CHANNELS, sizeof (samples));
306 stop = pa_rtclock_now();
307 pa_log_info("ref: %llu usec.", (long long unsigned int)(stop - start));
309 pa_assert_se(memcmp(samples_ref, samples, sizeof(samples)) == 0);
312 #endif /* defined (__i386__) || defined (__amd64__) */
314 void pa_volume_func_init_sse (pa_cpu_x86_flag_t flags) {
315 #if defined (__i386__) || defined (__amd64__)
321 if (flags & PA_CPU_X86_SSE2) {
322 pa_log_info("Initialising SSE2 optimized functions.");
324 pa_set_volume_func (PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_sse2);
325 pa_set_volume_func (PA_SAMPLE_S16RE, (pa_do_volume_func_t) pa_volume_s16re_sse2);
327 #endif /* defined (__i386__) || defined (__amd64__) */