tests: add cpu test

It tests only x86 architecture right now.

diff --git a/src/Makefile.am b/src/Makefile.am
index 8f54b1d..06bc5db 100644 (file)
--- a/src/Makefile.am
+++ b/src/Makefile.am
@@ -244,6 +244,7 @@ TESTS_default = \
                volume-test \
                mix-test \
                proplist-test \
+               cpu-test \
                lock-autospawn-test
 
 TESTS_norun = \
@@ -483,6 +484,11 @@ proplist_test_LDADD = $(AM_LDADD) libpulsecore-@PA_MAJORMINOR@.la libpulse.la li
 proplist_test_CFLAGS = $(AM_CFLAGS)
 proplist_test_LDFLAGS = $(AM_LDFLAGS) $(BINLDFLAGS)
 
+cpu_test_SOURCES = tests/cpu-test.c
+cpu_test_LDADD = $(AM_LDADD) libpulsecore-@PA_MAJORMINOR@.la libpulse.la libpulsecommon-@PA_MAJORMINOR@.la
+cpu_test_CFLAGS = $(AM_CFLAGS) $(LIBCHECK_CFLAGS)
+cpu_test_LDFLAGS = $(AM_LDFLAGS) $(BINLDFLAGS) $(LIBCHECK_LIBS)
+
 rtstutter_SOURCES = tests/rtstutter.c
 rtstutter_LDADD = $(AM_LDADD) libpulsecore-@PA_MAJORMINOR@.la libpulse.la libpulsecommon-@PA_MAJORMINOR@.la
 rtstutter_CFLAGS = $(AM_CFLAGS)

diff --git a/src/pulsecore/cpu-x86.c b/src/pulsecore/cpu-x86.c
index 05a4b2f..b2a558a 100644 (file)
--- a/src/pulsecore/cpu-x86.c
+++ b/src/pulsecore/cpu-x86.c
@@ -44,7 +44,7 @@ static void get_cpuid(uint32_t op, uint32_t *a, uint32_t *b, uint32_t *c, uint32
 }
 #endif
 
-pa_bool_t pa_cpu_init_x86(pa_cpu_x86_flag_t *flags) {
+void pa_cpu_get_x86_flags(pa_cpu_x86_flag_t *flags) {
 #if defined (__i386__) || defined (__amd64__)
     uint32_t eax, ebx, ecx, edx;
     uint32_t level;
@@ -111,6 +111,12 @@ pa_bool_t pa_cpu_init_x86(pa_cpu_x86_flag_t *flags) {
     (*flags & PA_CPU_X86_MMXEXT) ? "MMXEXT " : "",
     (*flags & PA_CPU_X86_3DNOW) ? "3DNOW " : "",
     (*flags & PA_CPU_X86_3DNOWEXT) ? "3DNOWEXT " : "");
+#endif /* defined (__i386__) || defined (__amd64__) */
+}
+
+pa_bool_t pa_cpu_init_x86(pa_cpu_x86_flag_t *flags) {
+#if defined (__i386__) || defined (__amd64__)
+    pa_cpu_get_x86_flags(flags);
 
     /* activate various optimisations */
     if (*flags & PA_CPU_X86_MMX) {

diff --git a/src/pulsecore/cpu-x86.h b/src/pulsecore/cpu-x86.h
index 0045ef6..9a9f3a1 100644 (file)
--- a/src/pulsecore/cpu-x86.h
+++ b/src/pulsecore/cpu-x86.h
@@ -40,6 +40,7 @@ typedef enum pa_cpu_x86_flag {
     PA_CPU_X86_CMOV      = (1 << 10)
 } pa_cpu_x86_flag_t;
 
+void pa_cpu_get_x86_flags(pa_cpu_x86_flag_t *flags);
 pa_bool_t pa_cpu_init_x86 (pa_cpu_x86_flag_t *flags);
 
 #if defined (__i386__)

diff --git a/src/pulsecore/sconv_sse.c b/src/pulsecore/sconv_sse.c
index 3c9a809..19de6fd 100644 (file)
--- a/src/pulsecore/sconv_sse.c
+++ b/src/pulsecore/sconv_sse.c
@@ -162,65 +162,11 @@ static void pa_sconv_s16le_from_f32ne_sse2(unsigned n, const float *a, int16_t *
     );
 }
 
-#undef RUN_TEST
-
-#ifdef RUN_TEST
-#define SAMPLES 1019
-#define TIMES 1000
-
-static void run_test(void) {
-    int16_t samples[SAMPLES];
-    int16_t samples_ref[SAMPLES];
-    float floats[SAMPLES];
-    int i;
-    pa_usec_t start, stop;
-    pa_convert_func_t func;
-
-    printf("checking SSE %zd\n", sizeof(samples));
-
-    memset(samples_ref, 0, sizeof(samples_ref));
-    memset(samples, 0, sizeof(samples));
-
-    for (i = 0; i < SAMPLES; i++) {
-        floats[i] = 2.1f * (rand()/(float) RAND_MAX - 0.5f);
-    }
-
-    func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE);
-    func(SAMPLES, floats, samples_ref);
-    pa_sconv_s16le_from_f32ne_sse2(SAMPLES, floats, samples);
-
-    for (i = 0; i < SAMPLES; i++) {
-        if (samples[i] != samples_ref[i]) {
-            printf ("%d: %04x != %04x (%f)\n", i, samples[i], samples_ref[i],
-                      floats[i]);
-        }
-    }
-
-    start = pa_rtclock_now();
-    for (i = 0; i < TIMES; i++) {
-        pa_sconv_s16le_from_f32ne_sse2(SAMPLES, floats, samples);
-    }
-    stop = pa_rtclock_now();
-    pa_log_info("SSE: %llu usec.", (long long unsigned int)(stop - start));
-
-    start = pa_rtclock_now();
-    for (i = 0; i < TIMES; i++) {
-        func(SAMPLES, floats, samples_ref);
-    }
-    stop = pa_rtclock_now();
-    pa_log_info("ref: %llu usec.", (long long unsigned int)(stop - start));
-}
-#endif
 #endif /* defined (__i386__) || defined (__amd64__) */
 
-
 void pa_convert_func_init_sse(pa_cpu_x86_flag_t flags) {
 #if !defined(__APPLE__) && defined (__i386__) || defined (__amd64__)
 
-#ifdef RUN_TEST
-    run_test();
-#endif
-
     if (flags & PA_CPU_X86_SSE2) {
         pa_log_info("Initialising SSE2 optimized conversions.");
         pa_set_convert_from_float32ne_function(PA_SAMPLE_S16LE, (pa_convert_func_t) pa_sconv_s16le_from_f32ne_sse2);

diff --git a/src/pulsecore/svolume_mmx.c b/src/pulsecore/svolume_mmx.c
index c470ffb..d68a105 100644 (file)
--- a/src/pulsecore/svolume_mmx.c
+++ b/src/pulsecore/svolume_mmx.c
@@ -240,100 +240,10 @@ static void pa_volume_s16re_mmx(int16_t *samples, const int32_t *volumes, unsign
     );
 }
 
-#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 MMX %zd\n", sizeof(samples));
-
-    pa_random(samples, sizeof(samples));
-    /* for (i = 0; i < SAMPLES; i++)
-       samples[i] = -1; */
-    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);
-        /* volumes[i] = 0x0000ffff; */
-    for (padding = 0; padding < PADDING; padding++, i++)
-        volumes[i] = volumes[padding];
-
-    func(samples_ref, volumes, CHANNELS, sizeof(samples));
-    pa_volume_s16ne_mmx(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]);
-        }
-    }
-
-    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_mmx(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));
-            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
-
 #endif /* defined (__i386__) || defined (__amd64__) */
 
-
 void pa_volume_func_init_mmx(pa_cpu_x86_flag_t flags) {
 #if defined (__i386__) || defined (__amd64__)
-
-#ifdef RUN_TEST
-    run_test();
-#endif
-
     if ((flags & PA_CPU_X86_MMX) && (flags & PA_CPU_X86_CMOV)) {
         pa_log_info("Initialising MMX optimized volume functions.");
 

diff --git a/src/pulsecore/svolume_sse.c b/src/pulsecore/svolume_sse.c
index 771bf53..daa87ca 100644 (file)
--- a/src/pulsecore/svolume_sse.c
+++ b/src/pulsecore/svolume_sse.c
@@ -252,95 +252,10 @@ static void pa_volume_s16re_sse2(int16_t *samples, const int32_t *volumes, unsig
     );
 }
 
-#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 SSE2 %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_sse2(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_sse2(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("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));
-            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
 #endif /* defined (__i386__) || defined (__amd64__) */
 
 void pa_volume_func_init_sse(pa_cpu_x86_flag_t flags) {
 #if defined (__i386__) || defined (__amd64__)
-
-#ifdef RUN_TEST
-    run_test();
-#endif
-
     if (flags & PA_CPU_X86_SSE2) {
         pa_log_info("Initialising SSE2 optimized volume functions.");
 

diff --git a/src/tests/cpu-test.c b/src/tests/cpu-test.c
new file mode 100644 (file)
index 0000000..de25bc6
--- /dev/null
+++ b/src/tests/cpu-test.c
@@ -0,0 +1,292 @@
+#ifdef HAVE_CONFIG_H
+#include <config.h>
+#endif
+
+#include <check.h>
+#include <unistd.h>
+#include <math.h>
+
+#include <pulse/rtclock.h>
+#include <pulsecore/cpu-x86.h>
+#include <pulsecore/random.h>
+#include <pulsecore/macro.h>
+#include <pulsecore/endianmacros.h>
+#include <pulsecore/sconv.h>
+#include <pulsecore/sample-util.h>
+
+START_TEST (svolume_mmx_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, mmx_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_get_x86_flags(&flags);
+
+    if (!((flags & PA_CPU_X86_MMX) && (flags & PA_CPU_X86_CMOV))) {
+        pa_log_info("MMX/CMOV not supported. Skipping");
+        return;
+    }
+
+    orig_func = pa_get_volume_func(PA_SAMPLE_S16NE);
+    pa_volume_func_init_mmx(flags);
+    mmx_func = pa_get_volume_func(PA_SAMPLE_S16NE);
+
+    pa_log_debug("Checking MMX svolume (%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((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
+
+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_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_get_x86_flags(&flags);
+
+    if (!(flags & PA_CPU_X86_SSE2)) {
+        pa_log_info("SSE2 not supported. Skipping");
+        return;
+    }
+
+    orig_func = pa_get_volume_func(PA_SAMPLE_S16NE);
+    pa_volume_func_init_sse(flags);
+    sse_func = pa_get_volume_func(PA_SAMPLE_S16NE);
+
+    pa_log_debug("Checking SSE2 svolume (%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((pa_volume_t)(rand() >> 15));
+    for (padding = 0; padding < PADDING; padding++, i++)
+        volumes[i] = volumes[padding];
+
+    orig_func(samples_ref, volumes, CHANNELS, sizeof(samples));
+    sse_func(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]);
+            fail();
+        }
+    }
+
+    for (k = 0; k < TIMES2; k++) {
+        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;
+        if (max < (stop - start)) max = stop - start;
+        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 SAMPLES
+#undef TIMES
+#undef TIMES2
+#undef PADDING
+}
+END_TEST
+
+START_TEST (sconv_sse_test) {
+#define SAMPLES 1019
+#define TIMES 1000
+
+    int16_t samples[SAMPLES];
+    int16_t samples_ref[SAMPLES];
+    float floats[SAMPLES];
+    int i;
+    pa_usec_t start, stop;
+    pa_convert_func_t orig_func, sse_func;
+    pa_cpu_x86_flag_t flags = 0;
+
+    pa_cpu_get_x86_flags(&flags);
+
+    if (!(flags & PA_CPU_X86_SSE2)) {
+        pa_log_info("SSE2 not supported. Skipping");
+        return;
+    }
+
+    pa_log_debug("Checking SSE sconv (%zd)\n", sizeof(samples));
+
+    memset(samples_ref, 0, sizeof(samples_ref));
+    memset(samples, 0, sizeof(samples));
+
+    for (i = 0; i < SAMPLES; i++) {
+        floats[i] = 2.1f * (rand()/(float) RAND_MAX - 0.5f);
+    }
+
+    orig_func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE);
+    pa_convert_func_init_sse(flags);
+    sse_func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE);
+
+    orig_func(SAMPLES, floats, samples_ref);
+    sse_func(SAMPLES, floats, samples);
+
+    for (i = 0; i < SAMPLES; i++) {
+        if (samples[i] != samples_ref[i]) {
+            printf ("%d: %04x != %04x (%f)\n", i, samples[i], samples_ref[i],
+                      floats[i]);
+            fail();
+        }
+    }
+
+    start = pa_rtclock_now();
+    for (i = 0; i < TIMES; i++) {
+        sse_func(SAMPLES, floats, samples);
+    }
+    stop = pa_rtclock_now();
+    pa_log_info("SSE: %llu usec.", (long long unsigned int)(stop - start));
+
+    start = pa_rtclock_now();
+    for (i = 0; i < TIMES; i++) {
+        orig_func(SAMPLES, floats, samples_ref);
+    }
+    stop = pa_rtclock_now();
+    pa_log_info("ref: %llu usec.", (long long unsigned int)(stop - start));
+
+#undef SAMPLES
+#undef TIMES
+}
+END_TEST
+
+int main(int argc, char *argv[]) {
+    int failed = 0;
+    Suite *s;
+    TCase *tc;
+    SRunner *sr;
+
+    if (!getenv("MAKE_CHECK"))
+        pa_log_set_level(PA_LOG_DEBUG);
+
+    s = suite_create("CPU");
+    tc = tcase_create("x86");
+    tcase_add_test(tc, svolume_mmx_test);
+    tcase_add_test(tc, svolume_sse_test);
+    tcase_add_test(tc, sconv_sse_test);
+    suite_add_tcase(s, tc);
+
+    sr = srunner_create(s);
+    srunner_run_all(sr, CK_NORMAL);
+    failed = srunner_ntests_failed(sr);
+    srunner_free(sr);
+
+    return (failed == 0) ? EXIT_SUCCESS : EXIT_FAILURE;
+}