]> code.delx.au - pulseaudio/blob - src/pulsecore/svolume_sse.c
SSE/MMX: Fix problem with highpitched noise on i386
[pulseaudio] / 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 #include <pulsecore/endianmacros.h>
35
36 #include "cpu-x86.h"
37
38 #include "sample-util.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 #if defined (__i386__)
158 : "m" ((pa_reg_x86)channels)
159 #else
160 : "r" ((pa_reg_x86)channels)
161 #endif
162 : "cc"
163 );
164 }
165
166 static void pa_volume_s16re_sse2(int16_t *samples, int32_t *volumes, unsigned channels, unsigned length) {
167 pa_reg_x86 channel, temp;
168
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. */
172 if (channels < 8)
173 channels = channel_overread_table[channels];
174
175 __asm__ __volatile__ (
176 " xor %3, %3 \n\t"
177 " sar $1, %2 \n\t" /* length /= sizeof (int16_t) */
178
179 " test $1, %2 \n\t" /* check for odd samples */
180 " je 2f \n\t"
181
182 " movd (%1, %3, 4), %%xmm0 \n\t" /* | v0h | v0l | */
183 " movw (%0), %w4 \n\t" /* .. | p0 | */
184 " rorw $8, %w4 \n\t"
185 " movd %4, %%xmm1 \n\t"
186 VOLUME_32x16 (%%xmm1, %%xmm0)
187 " movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */
188 " rorw $8, %w4 \n\t"
189 " movw %w4, (%0) \n\t"
190 " add $2, %0 \n\t"
191 MOD_ADD ($1, %5)
192
193 "2: \n\t"
194 " sar $1, %2 \n\t" /* prepare for processing 2 samples at a time */
195 " test $1, %2 \n\t"
196 " je 4f \n\t"
197
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 | */
201 SWAP_16 (%%xmm1)
202 VOLUME_32x16 (%%xmm1, %%xmm0)
203 SWAP_16 (%%xmm0)
204 " movd %%xmm0, (%0) \n\t" /* .. | p1*v1 | p0*v0 | */
205 " add $4, %0 \n\t"
206 MOD_ADD ($2, %5)
207
208 "4: \n\t"
209 " sar $1, %2 \n\t" /* prepare for processing 4 samples at a time */
210 " test $1, %2 \n\t"
211 " je 6f \n\t"
212
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 | */
219 SWAP_16 (%%xmm1)
220 VOLUME_32x16 (%%xmm1, %%xmm0)
221 SWAP_16 (%%xmm0)
222 " movq %%xmm0, (%0) \n\t" /* .. | p3*v3 .. p0*v0 | */
223 " add $8, %0 \n\t"
224 MOD_ADD ($4, %5)
225
226 "6: \n\t"
227 " sar $1, %2 \n\t" /* prepare for processing 8 samples at a time */
228 " cmp $0, %2 \n\t"
229 " je 8f \n\t"
230
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 | */
242 " add $16, %0 \n\t"
243 MOD_ADD ($8, %5)
244 " dec %2 \n\t"
245 " jne 7b \n\t"
246 "8: \n\t"
247
248 : "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp)
249 #if defined (__i386__)
250 : "m" ((pa_reg_x86)channels)
251 #else
252 : "r" ((pa_reg_x86)channels)
253 #endif
254 : "cc"
255 );
256 }
257
258 #undef RUN_TEST
259
260 #ifdef RUN_TEST
261 #define CHANNELS 2
262 #define SAMPLES 1022
263 #define TIMES 1000
264 #define TIMES2 100
265 #define PADDING 16
266
267 static void run_test(void) {
268 int16_t samples[SAMPLES];
269 int16_t samples_ref[SAMPLES];
270 int16_t samples_orig[SAMPLES];
271 int32_t volumes[CHANNELS + PADDING];
272 int i, j, padding;
273 pa_do_volume_func_t func;
274 pa_usec_t start, stop;
275 int k;
276 pa_usec_t min = INT_MAX, max = 0;
277 double s1 = 0, s2 = 0;
278
279 func = pa_get_volume_func(PA_SAMPLE_S16NE);
280
281 printf("checking SSE2 %zd\n", sizeof(samples));
282
283 pa_random(samples, sizeof(samples));
284 memcpy(samples_ref, samples, sizeof(samples));
285 memcpy(samples_orig, samples, sizeof(samples));
286
287 for (i = 0; i < CHANNELS; i++)
288 volumes[i] = PA_CLAMP_VOLUME(rand() >> 1);
289 for (padding = 0; padding < PADDING; padding++, i++)
290 volumes[i] = volumes[padding];
291
292 func(samples_ref, volumes, CHANNELS, sizeof(samples));
293 pa_volume_s16ne_sse2(samples, volumes, CHANNELS, sizeof(samples));
294 for (i = 0; i < SAMPLES; i++) {
295 if (samples[i] != samples_ref[i]) {
296 printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i],
297 samples_orig[i], volumes[i % CHANNELS]);
298 }
299 }
300
301 for (k = 0; k < TIMES2; k++) {
302 start = pa_rtclock_now();
303 for (j = 0; j < TIMES; j++) {
304 memcpy(samples, samples_orig, sizeof(samples));
305 pa_volume_s16ne_sse2(samples, volumes, CHANNELS, sizeof(samples));
306 }
307 stop = pa_rtclock_now();
308
309 if (min > (stop - start)) min = stop - start;
310 if (max < (stop - start)) max = stop - start;
311 s1 += stop - start;
312 s2 += (stop - start) * (stop - start);
313 }
314 pa_log_info("SSE: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1,
315 (long long unsigned int)min, (long long unsigned int)max, sqrt(TIMES2 * s2 - s1 * s1) / TIMES2);
316
317 min = INT_MAX; max = 0;
318 s1 = s2 = 0;
319 for (k = 0; k < TIMES2; k++) {
320 start = pa_rtclock_now();
321 for (j = 0; j < TIMES; j++) {
322 memcpy(samples_ref, samples_orig, sizeof(samples));
323 func(samples_ref, volumes, CHANNELS, sizeof(samples));
324 }
325 stop = pa_rtclock_now();
326
327 if (min > (stop - start)) min = stop - start;
328 if (max < (stop - start)) max = stop - start;
329 s1 += stop - start;
330 s2 += (stop - start) * (stop - start);
331 }
332 pa_log_info("ref: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1,
333 (long long unsigned int)min, (long long unsigned int)max, sqrt(TIMES2 * s2 - s1 * s1) / TIMES2);
334
335 pa_assert_se(memcmp(samples_ref, samples, sizeof(samples)) == 0);
336 }
337 #endif
338 #endif /* defined (__i386__) || defined (__amd64__) */
339
340 void pa_volume_func_init_sse(pa_cpu_x86_flag_t flags) {
341 #if defined (__i386__) || defined (__amd64__)
342
343 #ifdef RUN_TEST
344 run_test();
345 #endif
346
347 if (flags & PA_CPU_X86_SSE2) {
348 pa_log_info("Initialising SSE2 optimized functions.");
349
350 pa_set_volume_func(PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_sse2);
351 pa_set_volume_func(PA_SAMPLE_S16RE, (pa_do_volume_func_t) pa_volume_s16re_sse2);
352 }
353 #endif /* defined (__i386__) || defined (__amd64__) */
354 }