]> code.delx.au - pulseaudio/blob - src/pulsecore/svolume_sse.c
tests: modify mix-test to use new 'check' test framework
[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/rtclock.h>
28
29 #include <pulsecore/random.h>
30 #include <pulsecore/macro.h>
31 #include <pulsecore/endianmacros.h>
32
33 #include "cpu-x86.h"
34
35 #include "sample-util.h"
36
37 #if defined (__i386__) || defined (__amd64__)
38
39 #define VOLUME_32x16(s,v) /* .. | vh | vl | */ \
40 " pxor %%xmm4, %%xmm4 \n\t" /* .. | 0 | 0 | */ \
41 " punpcklwd %%xmm4, "#s" \n\t" /* .. | 0 | p0 | */ \
42 " pcmpgtw "#s", %%xmm4 \n\t" /* .. | 0 | s(p0) | */ \
43 " pand "#v", %%xmm4 \n\t" /* .. | 0 | (vl) | */ \
44 " movdqa "#s", %%xmm5 \n\t" \
45 " pmulhuw "#v", "#s" \n\t" /* .. | 0 | vl*p0 | */ \
46 " psubd %%xmm4, "#s" \n\t" /* .. | 0 | vl*p0 | + sign correct */ \
47 " psrld $16, "#v" \n\t" /* .. | 0 | vh | */ \
48 " pmaddwd %%xmm5, "#v" \n\t" /* .. | p0 * vh | */ \
49 " paddd "#s", "#v" \n\t" /* .. | p0 * v0 | */ \
50 " packssdw "#v", "#v" \n\t" /* .. | p1*v1 | p0*v0 | */
51
52 #define MOD_ADD(a,b) \
53 " add "#a", %3 \n\t" /* channel += inc */ \
54 " mov %3, %4 \n\t" \
55 " sub "#b", %4 \n\t" /* tmp = channel - channels */ \
56 " cmovae %4, %3 \n\t" /* if (tmp >= 0) channel = tmp */
57
58 /* swap 16 bits */
59 #define SWAP_16(s) \
60 " movdqa "#s", %%xmm4 \n\t" /* .. | h l | */ \
61 " psrlw $8, %%xmm4 \n\t" /* .. | 0 h | */ \
62 " psllw $8, "#s" \n\t" /* .. | l 0 | */ \
63 " por %%xmm4, "#s" \n\t" /* .. | l h | */
64
65 /* swap 2 registers 16 bits for better pairing */
66 #define SWAP_16_2(s1,s2) \
67 " movdqa "#s1", %%xmm4 \n\t" /* .. | h l | */ \
68 " movdqa "#s2", %%xmm5 \n\t" \
69 " psrlw $8, %%xmm4 \n\t" /* .. | 0 h | */ \
70 " psrlw $8, %%xmm5 \n\t" \
71 " psllw $8, "#s1" \n\t" /* .. | l 0 | */ \
72 " psllw $8, "#s2" \n\t" \
73 " por %%xmm4, "#s1" \n\t" /* .. | l h | */ \
74 " por %%xmm5, "#s2" \n\t"
75
76
77 static int channel_overread_table[8] = {8,8,8,12,8,10,12,14};
78
79 static void pa_volume_s16ne_sse2(int16_t *samples, const int32_t *volumes, unsigned channels, unsigned length) {
80 pa_reg_x86 channel, temp;
81
82 /* Channels must be at least 8 and always a multiple of the original number.
83 * This is also the max amount we overread the volume array, which should
84 * have enough padding. */
85 if (channels < 8)
86 channels = channel_overread_table[channels];
87
88 __asm__ __volatile__ (
89 " xor %3, %3 \n\t"
90 " sar $1, %2 \n\t" /* length /= sizeof (int16_t) */
91
92 " test $1, %2 \n\t" /* check for odd samples */
93 " je 2f \n\t"
94
95 " movd (%q1, %3, 4), %%xmm0 \n\t" /* | v0h | v0l | */
96 " movw (%0), %w4 \n\t" /* .. | p0 | */
97 " movd %4, %%xmm1 \n\t"
98 VOLUME_32x16 (%%xmm1, %%xmm0)
99 " movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */
100 " movw %w4, (%0) \n\t"
101 " add $2, %0 \n\t"
102 MOD_ADD ($1, %5)
103
104 "2: \n\t"
105 " sar $1, %2 \n\t" /* prepare for processing 2 samples at a time */
106 " test $1, %2 \n\t"
107 " je 4f \n\t"
108
109 "3: \n\t" /* do samples in groups of 2 */
110 " movq (%q1, %3, 4), %%xmm0 \n\t" /* | v1h | v1l | v0h | v0l | */
111 " movd (%0), %%xmm1 \n\t" /* .. | p1 | p0 | */
112 VOLUME_32x16 (%%xmm1, %%xmm0)
113 " movd %%xmm0, (%0) \n\t" /* .. | p1*v1 | p0*v0 | */
114 " add $4, %0 \n\t"
115 MOD_ADD ($2, %5)
116
117 "4: \n\t"
118 " sar $1, %2 \n\t" /* prepare for processing 4 samples at a time */
119 " test $1, %2 \n\t"
120 " je 6f \n\t"
121
122 /* FIXME, we can do aligned access of the volume values if we can guarantee
123 * that the array is 16 bytes aligned, we probably have to do the odd values
124 * after this then. */
125 "5: \n\t" /* do samples in groups of 4 */
126 " movdqu (%q1, %3, 4), %%xmm0 \n\t" /* | v3h | v3l .. v0h | v0l | */
127 " movq (%0), %%xmm1 \n\t" /* .. | p3 .. p0 | */
128 VOLUME_32x16 (%%xmm1, %%xmm0)
129 " movq %%xmm0, (%0) \n\t" /* .. | p3*v3 .. p0*v0 | */
130 " add $8, %0 \n\t"
131 MOD_ADD ($4, %5)
132
133 "6: \n\t"
134 " sar $1, %2 \n\t" /* prepare for processing 8 samples at a time */
135 " cmp $0, %2 \n\t"
136 " je 8f \n\t"
137
138 "7: \n\t" /* do samples in groups of 8 */
139 " movdqu (%q1, %3, 4), %%xmm0 \n\t" /* | v3h | v3l .. v0h | v0l | */
140 " movdqu 16(%q1, %3, 4), %%xmm2 \n\t" /* | v7h | v7l .. v4h | v4l | */
141 " movq (%0), %%xmm1 \n\t" /* .. | p3 .. p0 | */
142 " movq 8(%0), %%xmm3 \n\t" /* .. | p7 .. p4 | */
143 VOLUME_32x16 (%%xmm1, %%xmm0)
144 VOLUME_32x16 (%%xmm3, %%xmm2)
145 " movq %%xmm0, (%0) \n\t" /* .. | p3*v3 .. p0*v0 | */
146 " movq %%xmm2, 8(%0) \n\t" /* .. | p7*v7 .. p4*v4 | */
147 " add $16, %0 \n\t"
148 MOD_ADD ($8, %5)
149 " dec %2 \n\t"
150 " jne 7b \n\t"
151 "8: \n\t"
152
153 : "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp)
154 #if defined (__i386__)
155 : "m" (channels)
156 #else
157 : "r" ((pa_reg_x86)channels)
158 #endif
159 : "cc"
160 );
161 }
162
163 static void pa_volume_s16re_sse2(int16_t *samples, const int32_t *volumes, unsigned channels, unsigned length) {
164 pa_reg_x86 channel, temp;
165
166 /* Channels must be at least 8 and always a multiple of the original number.
167 * This is also the max amount we overread the volume array, which should
168 * have enough padding. */
169 if (channels < 8)
170 channels = channel_overread_table[channels];
171
172 __asm__ __volatile__ (
173 " xor %3, %3 \n\t"
174 " sar $1, %2 \n\t" /* length /= sizeof (int16_t) */
175
176 " test $1, %2 \n\t" /* check for odd samples */
177 " je 2f \n\t"
178
179 " movd (%q1, %3, 4), %%xmm0 \n\t" /* | v0h | v0l | */
180 " movw (%0), %w4 \n\t" /* .. | p0 | */
181 " rorw $8, %w4 \n\t"
182 " movd %4, %%xmm1 \n\t"
183 VOLUME_32x16 (%%xmm1, %%xmm0)
184 " movd %%xmm0, %4 \n\t" /* .. | p0*v0 | */
185 " rorw $8, %w4 \n\t"
186 " movw %w4, (%0) \n\t"
187 " add $2, %0 \n\t"
188 MOD_ADD ($1, %5)
189
190 "2: \n\t"
191 " sar $1, %2 \n\t" /* prepare for processing 2 samples at a time */
192 " test $1, %2 \n\t"
193 " je 4f \n\t"
194
195 "3: \n\t" /* do samples in groups of 2 */
196 " movq (%q1, %3, 4), %%xmm0 \n\t" /* | v1h | v1l | v0h | v0l | */
197 " movd (%0), %%xmm1 \n\t" /* .. | p1 | p0 | */
198 SWAP_16 (%%xmm1)
199 VOLUME_32x16 (%%xmm1, %%xmm0)
200 SWAP_16 (%%xmm0)
201 " movd %%xmm0, (%0) \n\t" /* .. | p1*v1 | p0*v0 | */
202 " add $4, %0 \n\t"
203 MOD_ADD ($2, %5)
204
205 "4: \n\t"
206 " sar $1, %2 \n\t" /* prepare for processing 4 samples at a time */
207 " test $1, %2 \n\t"
208 " je 6f \n\t"
209
210 /* FIXME, we can do aligned access of the volume values if we can guarantee
211 * that the array is 16 bytes aligned, we probably have to do the odd values
212 * after this then. */
213 "5: \n\t" /* do samples in groups of 4 */
214 " movdqu (%q1, %3, 4), %%xmm0 \n\t" /* | v3h | v3l .. v0h | v0l | */
215 " movq (%0), %%xmm1 \n\t" /* .. | p3 .. p0 | */
216 SWAP_16 (%%xmm1)
217 VOLUME_32x16 (%%xmm1, %%xmm0)
218 SWAP_16 (%%xmm0)
219 " movq %%xmm0, (%0) \n\t" /* .. | p3*v3 .. p0*v0 | */
220 " add $8, %0 \n\t"
221 MOD_ADD ($4, %5)
222
223 "6: \n\t"
224 " sar $1, %2 \n\t" /* prepare for processing 8 samples at a time */
225 " cmp $0, %2 \n\t"
226 " je 8f \n\t"
227
228 "7: \n\t" /* do samples in groups of 8 */
229 " movdqu (%q1, %3, 4), %%xmm0 \n\t" /* | v3h | v3l .. v0h | v0l | */
230 " movdqu 16(%q1, %3, 4), %%xmm2 \n\t" /* | v7h | v7l .. v4h | v4l | */
231 " movq (%0), %%xmm1 \n\t" /* .. | p3 .. p0 | */
232 " movq 8(%0), %%xmm3 \n\t" /* .. | p7 .. p4 | */
233 SWAP_16_2 (%%xmm1, %%xmm3)
234 VOLUME_32x16 (%%xmm1, %%xmm0)
235 VOLUME_32x16 (%%xmm3, %%xmm2)
236 SWAP_16_2 (%%xmm0, %%xmm2)
237 " movq %%xmm0, (%0) \n\t" /* .. | p3*v3 .. p0*v0 | */
238 " movq %%xmm2, 8(%0) \n\t" /* .. | p7*v7 .. p4*v4 | */
239 " add $16, %0 \n\t"
240 MOD_ADD ($8, %5)
241 " dec %2 \n\t"
242 " jne 7b \n\t"
243 "8: \n\t"
244
245 : "+r" (samples), "+r" (volumes), "+r" (length), "=D" (channel), "=&r" (temp)
246 #if defined (__i386__)
247 : "m" (channels)
248 #else
249 : "r" ((pa_reg_x86)channels)
250 #endif
251 : "cc"
252 );
253 }
254
255 #undef RUN_TEST
256
257 #ifdef RUN_TEST
258 #define CHANNELS 2
259 #define SAMPLES 1022
260 #define TIMES 1000
261 #define TIMES2 100
262 #define PADDING 16
263
264 static void run_test(void) {
265 int16_t samples[SAMPLES];
266 int16_t samples_ref[SAMPLES];
267 int16_t samples_orig[SAMPLES];
268 int32_t volumes[CHANNELS + PADDING];
269 int i, j, padding;
270 pa_do_volume_func_t func;
271 pa_usec_t start, stop;
272 int k;
273 pa_usec_t min = INT_MAX, max = 0;
274 double s1 = 0, s2 = 0;
275
276 func = pa_get_volume_func(PA_SAMPLE_S16NE);
277
278 printf("checking SSE2 %zd\n", sizeof(samples));
279
280 pa_random(samples, sizeof(samples));
281 memcpy(samples_ref, samples, sizeof(samples));
282 memcpy(samples_orig, samples, sizeof(samples));
283
284 for (i = 0; i < CHANNELS; i++)
285 volumes[i] = PA_CLAMP_VOLUME(rand() >> 15);
286 for (padding = 0; padding < PADDING; padding++, i++)
287 volumes[i] = volumes[padding];
288
289 func(samples_ref, volumes, CHANNELS, sizeof(samples));
290 pa_volume_s16ne_sse2(samples, volumes, CHANNELS, sizeof(samples));
291 for (i = 0; i < SAMPLES; i++) {
292 if (samples[i] != samples_ref[i]) {
293 printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i],
294 samples_orig[i], volumes[i % CHANNELS]);
295 }
296 }
297
298 for (k = 0; k < TIMES2; k++) {
299 start = pa_rtclock_now();
300 for (j = 0; j < TIMES; j++) {
301 memcpy(samples, samples_orig, sizeof(samples));
302 pa_volume_s16ne_sse2(samples, volumes, CHANNELS, sizeof(samples));
303 }
304 stop = pa_rtclock_now();
305
306 if (min > (stop - start)) min = stop - start;
307 if (max < (stop - start)) max = stop - start;
308 s1 += stop - start;
309 s2 += (stop - start) * (stop - start);
310 }
311 pa_log_info("SSE: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1,
312 (long long unsigned int)min, (long long unsigned int)max, sqrt(TIMES2 * s2 - s1 * s1) / TIMES2);
313
314 min = INT_MAX; max = 0;
315 s1 = s2 = 0;
316 for (k = 0; k < TIMES2; k++) {
317 start = pa_rtclock_now();
318 for (j = 0; j < TIMES; j++) {
319 memcpy(samples_ref, samples_orig, sizeof(samples));
320 func(samples_ref, volumes, CHANNELS, sizeof(samples));
321 }
322 stop = pa_rtclock_now();
323
324 if (min > (stop - start)) min = stop - start;
325 if (max < (stop - start)) max = stop - start;
326 s1 += stop - start;
327 s2 += (stop - start) * (stop - start);
328 }
329 pa_log_info("ref: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1,
330 (long long unsigned int)min, (long long unsigned int)max, sqrt(TIMES2 * s2 - s1 * s1) / TIMES2);
331
332 pa_assert_se(memcmp(samples_ref, samples, sizeof(samples)) == 0);
333 }
334 #endif
335 #endif /* defined (__i386__) || defined (__amd64__) */
336
337 void pa_volume_func_init_sse(pa_cpu_x86_flag_t flags) {
338 #if defined (__i386__) || defined (__amd64__)
339
340 #ifdef RUN_TEST
341 run_test();
342 #endif
343
344 if (flags & PA_CPU_X86_SSE2) {
345 pa_log_info("Initialising SSE2 optimized volume functions.");
346
347 pa_set_volume_func(PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_sse2);
348 pa_set_volume_func(PA_SAMPLE_S16RE, (pa_do_volume_func_t) pa_volume_s16re_sse2);
349 }
350 #endif /* defined (__i386__) || defined (__amd64__) */
351 }