mirrors/pulseaudio
1
0
Fork 0
mirror of https://gitlab.freedesktop.org/pulseaudio/pulseaudio.git synced 2025-11-03 09:01:50 -05:00
Code Issues Projects Releases Packages Wiki Activity Actions

tests: Factor out Orc test code into cpu-test

Browse source 
Reorganises all the svolume core test code in cpu-test since it's the
same across MMX/SSE/etc.
This commit is contained in:
Arun Raghavan 2012-07-24 14:37:54 +05:30
parent c2f7987f09
commit b6f2ccf194
4 changed files with 93 additions and 219 deletions
Download patch file Download diff file Expand all files Collapse all files

8
src/pulsecore/cpu-orc.c
View file

@ -25,14 +25,18 @@
#include "cpu-orc.h" #include "cpu-orc.h"
void pa_cpu_init_orc(pa_cpu_info cpu_info) pa_bool_t pa_cpu_init_orc(pa_cpu_info cpu_info)
{ {
#ifndef DISABLE_ORC #ifndef DISABLE_ORC
/* Update these as we test on more architectures */ /* Update these as we test on more architectures */
pa_cpu_x86_flag_t x86_want_flags = PA_CPU_X86_MMX | PA_CPU_X86_SSE | PA_CPU_X86_SSE2 | PA_CPU_X86_SSE3 | PA_CPU_X86_SSSE3 | PA_CPU_X86_SSE4_1 | PA_CPU_X86_SSE4_2; pa_cpu_x86_flag_t x86_want_flags = PA_CPU_X86_MMX | PA_CPU_X86_SSE | PA_CPU_X86_SSE2 | PA_CPU_X86_SSE3 | PA_CPU_X86_SSSE3 | PA_CPU_X86_SSE4_1 | PA_CPU_X86_SSE4_2;
/* Enable Orc svolume optimizations */ /* Enable Orc svolume optimizations */
if ((cpu_info.cpu_type == PA_CPU_X86) && (cpu_info.flags.x86 & x86_want_flags)) if ((cpu_info.cpu_type == PA_CPU_X86) && (cpu_info.flags.x86 & x86_want_flags)) {
pa_volume_func_init_orc(); pa_volume_func_init_orc();
return TRUE;
}
#endif #endif
return FALSE;
} }

2
src/pulsecore/cpu-orc.h
View file

@ -26,7 +26,7 @@
/* Orc-optimised bits */ /* Orc-optimised bits */
void pa_cpu_init_orc(pa_cpu_info cpu_info); pa_bool_t pa_cpu_init_orc(pa_cpu_info cpu_info);
void pa_volume_func_init_orc(void); void pa_volume_func_init_orc(void);

85
src/pulsecore/svolume_orc.c
View file

@ -45,94 +45,9 @@ pa_volume_s16ne_orc(int16_t *samples, const int32_t *volumes, unsigned channels,
fallback(samples, volumes, channels, length); fallback(samples, volumes, channels, length);
} }
#undef RUN_TEST
#ifdef RUN_TEST
#define CHANNELS 2
#define SAMPLES 1022
#define TIMES 1000
#define TIMES2 100
#define PADDING 16
static void run_test(void) {
int16_t samples[SAMPLES];
int16_t samples_ref[SAMPLES];
int16_t samples_orig[SAMPLES];
int32_t volumes[CHANNELS + PADDING];
int i, j, padding;
pa_do_volume_func_t func;
pa_usec_t start, stop;
int k;
pa_usec_t min = INT_MAX, max = 0;
double s1 = 0, s2 = 0;
func = pa_get_volume_func(PA_SAMPLE_S16NE);
printf("checking ORC %zd\n", sizeof(samples));
pa_random(samples, sizeof(samples));
memcpy(samples_ref, samples, sizeof(samples));
memcpy(samples_orig, samples, sizeof(samples));
for (i = 0; i < CHANNELS; i++)
volumes[i] = PA_CLAMP_VOLUME(rand() >> 15);
for (padding = 0; padding < PADDING; padding++, i++)
volumes[i] = volumes[padding];
func(samples_ref, volumes, CHANNELS, sizeof(samples));
pa_volume_s16ne_orc(samples, volumes, CHANNELS, sizeof(samples));
for (i = 0; i < SAMPLES; i++) {
if (samples[i] != samples_ref[i]) {
printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i],
samples_orig[i], volumes[i % CHANNELS]);
}
}
for (k = 0; k < TIMES2; k++) {
start = pa_rtclock_now();
for (j = 0; j < TIMES; j++) {
memcpy(samples, samples_orig, sizeof(samples));
pa_volume_s16ne_orc(samples, volumes, CHANNELS, sizeof(samples));
}
stop = pa_rtclock_now();
if (min > (stop - start)) min = stop - start;
if (max < (stop - start)) max = stop - start;
s1 += stop - start;
s2 += (stop - start) * (stop - start);
}
pa_log_info("ORC: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1,
(long long unsigned int)min, (long long unsigned int)max, sqrt(TIMES2 * s2 - s1 * s1) / TIMES2);
min = INT_MAX; max = 0;
s1 = s2 = 0;
for (k = 0; k < TIMES2; k++) {
start = pa_rtclock_now();
for (j = 0; j < TIMES; j++) {
memcpy(samples_ref, samples_orig, sizeof(samples));
func(samples_ref, volumes, CHANNELS, sizeof(samples));
}
stop = pa_rtclock_now();
if (min > (stop - start)) min = stop - start;
if (max < (stop - start)) max = stop - start;
s1 += stop - start;
s2 += (stop - start) * (stop - start);
}
pa_log_info("ref: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1,
(long long unsigned int)min, (long long unsigned int)max, sqrt(TIMES2 * s2 - s1 * s1) / TIMES2);
pa_assert_se(memcmp(samples_ref, samples, sizeof(samples)) == 0);
}
#endif
void pa_volume_func_init_orc(void) { void pa_volume_func_init_orc(void) {
pa_log_info("Initialising ORC optimized volume functions."); pa_log_info("Initialising ORC optimized volume functions.");
#ifdef RUN_TEST
run_test();
#endif
fallback = pa_get_volume_func(PA_SAMPLE_S16NE); fallback = pa_get_volume_func(PA_SAMPLE_S16NE);
pa_set_volume_func(PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_orc); pa_set_volume_func(PA_SAMPLE_S16NE, (pa_do_volume_func_t) pa_volume_s16ne_orc);
} }

217
src/tests/cpu-test.c
View file

@ -8,30 +8,90 @@
#include <pulse/rtclock.h> #include <pulse/rtclock.h>
#include <pulsecore/cpu-x86.h> #include <pulsecore/cpu-x86.h>
#include <pulsecore/cpu-orc.h>
#include <pulsecore/random.h> #include <pulsecore/random.h>
#include <pulsecore/macro.h> #include <pulsecore/macro.h>
#include <pulsecore/endianmacros.h> #include <pulsecore/endianmacros.h>
#include <pulsecore/sconv.h> #include <pulsecore/sconv.h>
#include <pulsecore/sample-util.h> #include <pulsecore/sample-util.h>
START_TEST (svolume_mmx_test) { /* Common defines for svolume tests */
#define CHANNELS 2 #define CHANNELS 2
#define SAMPLES 1022 #define SAMPLES 1022
#define TIMES 1000 #define TIMES 1000
#define TIMES2 100 #define TIMES2 100
#define PADDING 16 #define PADDING 16
static void run_volume_test(pa_do_volume_func_t func, pa_do_volume_func_t orig_func) {
int16_t samples[SAMPLES]; int16_t samples[SAMPLES];
int16_t samples_ref[SAMPLES]; int16_t samples_ref[SAMPLES];
int16_t samples_orig[SAMPLES]; int16_t samples_orig[SAMPLES];
int32_t volumes[CHANNELS + PADDING]; int32_t volumes[CHANNELS + PADDING];
int i, j, padding; int i, j, padding;
pa_do_volume_func_t orig_func, mmx_func;
pa_usec_t start, stop; pa_usec_t start, stop;
int k; int k;
pa_usec_t min = INT_MAX, max = 0; pa_usec_t min = INT_MAX, max = 0;
double s1 = 0, s2 = 0; double s1 = 0, s2 = 0;
pa_random(samples, sizeof(samples));
memcpy(samples_ref, samples, sizeof(samples));
memcpy(samples_orig, samples, sizeof(samples));
for (i = 0; i < CHANNELS; i++)
volumes[i] = PA_CLAMP_VOLUME((pa_volume_t)(rand() >> 15));
for (padding = 0; padding < PADDING; padding++, i++)
volumes[i] = volumes[padding];
orig_func(samples_ref, volumes, CHANNELS, sizeof(samples));
func(samples, volumes, CHANNELS, sizeof(samples));
for (i = 0; i < SAMPLES; i++) {
if (samples[i] != samples_ref[i]) {
printf("%d: %04x != %04x (%04x * %08x)\n", i, samples[i], samples_ref[i],
samples_orig[i], volumes[i % CHANNELS]);
fail();
}
}
for (k = 0; k < TIMES2; k++) {
start = pa_rtclock_now();
for (j = 0; j < TIMES; j++) {
memcpy(samples, samples_orig, sizeof(samples));
func(samples, volumes, CHANNELS, sizeof(samples));
}
stop = pa_rtclock_now();
if (min > (stop - start)) min = stop - start;
if (max < (stop - start)) max = stop - start;
s1 += stop - start;
s2 += (stop - start) * (stop - start);
}
pa_log_info("func: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1,
(long long unsigned int)min, (long long unsigned int)max, sqrt(times2 * s2 - s1 * s1) / times2);
min = INT_MAX; max = 0;
s1 = s2 = 0;
for (k = 0; k < TIMES2; k++) {
start = pa_rtclock_now();
for (j = 0; j < TIMES; j++) {
memcpy(samples_ref, samples_orig, sizeof(samples));
orig_func(samples_ref, volumes, CHANNELS, sizeof(samples));
}
stop = pa_rtclock_now();
if (min > (stop - start)) min = stop - start;
if (max < (stop - start)) max = stop - start;
s1 += stop - start;
s2 += (stop - start) * (stop - start);
}
pa_log_info("orig: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1,
(long long unsigned int)min, (long long unsigned int)max, sqrt(times2 * s2 - s1 * s1) / times2);
fail_unless(memcmp(samples_ref, samples, sizeof(samples)) == 0);
}
START_TEST (svolume_mmx_test) {
pa_do_volume_func_t orig_func, mmx_func;
pa_cpu_x86_flag_t flags = 0; pa_cpu_x86_flag_t flags = 0;
pa_cpu_get_x86_flags(&flags); pa_cpu_get_x86_flags(&flags);
@ -45,88 +105,13 @@ START_TEST (svolume_mmx_test) {
pa_volume_func_init_mmx(flags); pa_volume_func_init_mmx(flags);
mmx_func = pa_get_volume_func(PA_SAMPLE_S16NE); mmx_func = pa_get_volume_func(PA_SAMPLE_S16NE);
pa_log_debug("Checking MMX svolume (%zd)\n", sizeof(samples)); pa_log_debug("Checking MMX svolume");
run_volume_test(mmx_func, orig_func);
pa_random(samples, sizeof(samples));
memcpy(samples_ref, samples, sizeof(samples));
memcpy(samples_orig, samples, sizeof(samples));
for (i = 0; i < CHANNELS; i++)
volumes[i] = PA_CLAMP_VOLUME((pa_volume_t)(rand() >> 15));
for (padding = 0; padding < PADDING; padding++, i++)
volumes[i] = volumes[padding];
orig_func(samples_ref, volumes, CHANNELS, sizeof(samples));
mmx_func(samples, volumes, CHANNELS, sizeof(samples));
for (i = 0; i < SAMPLES; i++) {
if (samples[i] != samples_ref[i]) {
printf("%d: %04x != %04x (%04x * %08x)\n", i, samples[i], samples_ref[i],
samples_orig[i], volumes[i % CHANNELS]);
fail();
}
}
for (k = 0; k < TIMES2; k++) {
start = pa_rtclock_now();
for (j = 0; j < TIMES; j++) {
memcpy(samples, samples_orig, sizeof(samples));
mmx_func(samples, volumes, CHANNELS, sizeof(samples));
}
stop = pa_rtclock_now();
if (min > (stop - start)) min = stop - start;
if (max < (stop - start)) max = stop - start;
s1 += stop - start;
s2 += (stop - start) * (stop - start);
}
pa_log_info("MMX: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1,
(long long unsigned int)min, (long long unsigned int)max, sqrt(TIMES2 * s2 - s1 * s1) / TIMES2);
min = INT_MAX; max = 0;
s1 = s2 = 0;
for (k = 0; k < TIMES2; k++) {
start = pa_rtclock_now();
for (j = 0; j < TIMES; j++) {
memcpy(samples_ref, samples_orig, sizeof(samples));
orig_func(samples_ref, volumes, CHANNELS, sizeof(samples));
}
stop = pa_rtclock_now();
if (min > (stop - start)) min = stop - start;
if (max < (stop - start)) max = stop - start;
s1 += stop - start;
s2 += (stop - start) * (stop - start);
}
pa_log_info("ref: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1,
(long long unsigned int)min, (long long unsigned int)max, sqrt(TIMES2 * s2 - s1 * s1) / TIMES2);
fail_unless(memcmp(samples_ref, samples, sizeof(samples)) == 0);
#undef CHANNELS
#undef SAMPLES
#undef TIMES
#undef TIMES2
#undef PADDING
} }
END_TEST END_TEST
START_TEST (svolume_sse_test) { START_TEST (svolume_sse_test) {
#define CHANNELS 2
#define SAMPLES 1022
#define TIMES 1000
#define TIMES2 100
#define PADDING 16
int16_t samples[SAMPLES];
int16_t samples_ref[SAMPLES];
int16_t samples_orig[SAMPLES];
int32_t volumes[CHANNELS + PADDING];
int i, j, padding;
pa_do_volume_func_t orig_func, sse_func; pa_do_volume_func_t orig_func, sse_func;
pa_usec_t start, stop;
int k;
pa_usec_t min = INT_MAX, max = 0;
double s1 = 0, s2 = 0;
pa_cpu_x86_flag_t flags = 0; pa_cpu_x86_flag_t flags = 0;
pa_cpu_get_x86_flags(&flags); pa_cpu_get_x86_flags(&flags);
@ -140,70 +125,39 @@ START_TEST (svolume_sse_test) {
pa_volume_func_init_sse(flags); pa_volume_func_init_sse(flags);
sse_func = pa_get_volume_func(PA_SAMPLE_S16NE); sse_func = pa_get_volume_func(PA_SAMPLE_S16NE);
pa_log_debug("Checking SSE2 svolume (%zd)\n", sizeof(samples)); pa_log_debug("Checking SSE2 svolume");
run_volume_test(sse_func, orig_func);
}
END_TEST
pa_random(samples, sizeof(samples)); START_TEST (svolume_orc_test) {
memcpy(samples_ref, samples, sizeof(samples)); pa_do_volume_func_t orig_func, orc_func;
memcpy(samples_orig, samples, sizeof(samples)); pa_cpu_info cpu_info;
for (i = 0; i < CHANNELS; i++) #if defined (__i386__) || defined (__amd64__)
volumes[i] = PA_CLAMP_VOLUME((pa_volume_t)(rand() >> 15)); pa_zero(cpu_info);
for (padding = 0; padding < PADDING; padding++, i++) cpu_info.cpu_type = PA_CPU_X86;
volumes[i] = volumes[padding]; pa_cpu_get_x86_flags(&cpu_info.flags.x86);
#endif
orig_func(samples_ref, volumes, CHANNELS, sizeof(samples)); orig_func = pa_get_volume_func(PA_SAMPLE_S16NE);
sse_func(samples, volumes, CHANNELS, sizeof(samples));
for (i = 0; i < SAMPLES; i++) { if (!pa_cpu_init_orc(cpu_info)) {
if (samples[i] != samples_ref[i]) { pa_log_info("Orc not supported. Skipping");
printf ("%d: %04x != %04x (%04x * %04x)\n", i, samples[i], samples_ref[i], return;
samples_orig[i], volumes[i % CHANNELS]);
fail();
}
} }
for (k = 0; k < TIMES2; k++) { orc_func = pa_get_volume_func(PA_SAMPLE_S16NE);
start = pa_rtclock_now();
for (j = 0; j < TIMES; j++) {
memcpy(samples, samples_orig, sizeof(samples));
sse_func(samples, volumes, CHANNELS, sizeof(samples));
}
stop = pa_rtclock_now();
if (min > (stop - start)) min = stop - start; pa_log_debug("Checking SSE2 svolume");
if (max < (stop - start)) max = stop - start; run_volume_test(orc_func, orig_func, CHANNELS, SAMPLES, TIMES, TIMES2, PADDING);
s1 += stop - start;
s2 += (stop - start) * (stop - start);
}
pa_log_info("SSE: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1,
(long long unsigned int)min, (long long unsigned int)max, sqrt(TIMES2 * s2 - s1 * s1) / TIMES2);
min = INT_MAX; max = 0;
s1 = s2 = 0;
for (k = 0; k < TIMES2; k++) {
start = pa_rtclock_now();
for (j = 0; j < TIMES; j++) {
memcpy(samples_ref, samples_orig, sizeof(samples));
orig_func(samples_ref, volumes, CHANNELS, sizeof(samples));
}
stop = pa_rtclock_now();
if (min > (stop - start)) min = stop - start;
if (max < (stop - start)) max = stop - start;
s1 += stop - start;
s2 += (stop - start) * (stop - start);
}
pa_log_info("ref: %llu usec (min = %llu, max = %llu, stddev = %g).", (long long unsigned int)s1,
(long long unsigned int)min, (long long unsigned int)max, sqrt(TIMES2 * s2 - s1 * s1) / TIMES2);
fail_unless(memcmp(samples_ref, samples, sizeof(samples)) == 0);
#undef CHANNELS #undef CHANNELS
#undef SAMPLES #undef SAMPLES
#undef TIMES #undef TIMES
#undef TIMES2 #undef TIMES2
#undef PADDING #undef PADDING
} /* End svolume tests */
END_TEST
START_TEST (sconv_sse_test) { START_TEST (sconv_sse_test) {
#define SAMPLES 1019 #define SAMPLES 1019
@ -280,6 +234,7 @@ int main(int argc, char *argv[]) {
tc = tcase_create("x86"); tc = tcase_create("x86");
tcase_add_test(tc, svolume_mmx_test); tcase_add_test(tc, svolume_mmx_test);
tcase_add_test(tc, svolume_sse_test); tcase_add_test(tc, svolume_sse_test);
tcase_add_test(tc, svolume_orc_test);
tcase_add_test(tc, sconv_sse_test); tcase_add_test(tc, sconv_sse_test);
suite_add_tcase(s, tc); suite_add_tcase(s, tc);