configure.ac | 28 +++++++++++ src/Makefile.am | 15 +++++ src/pulsecore/cpu-arm.c | 4 + src/pulsecore/cpu-arm.h | 8 +++ src/pulsecore/sconv_neon.c | 106 ++++++++++++++++++++++++++++++++++++++++++ src/tests/cpu-test.c | 113 ++++++++++++++++++++++++++++++++++++++++++--- 6 files changed, 266 insertions(+), 8 deletions(-)
New commits: commit a423ee0f3d36b98baff64cc31558632e738b402d Author: Tanu Kaskinen <[email protected]> Date: Sat Oct 6 14:43:07 2012 +0300 build: Add a2dp-codecs.h to libbluetooth-util sources. diff --git a/src/Makefile.am b/src/Makefile.am index 84598b2..88054cb 100644 --- a/src/Makefile.am +++ b/src/Makefile.am @@ -1960,7 +1960,10 @@ module_bluetooth_discover_la_LIBADD = $(MODULE_LIBADD) $(DBUS_LIBS) libbluetooth module_bluetooth_discover_la_CFLAGS = $(AM_CFLAGS) $(DBUS_CFLAGS) -libbluetooth_util_la_SOURCES = modules/bluetooth/bluetooth-util.c modules/bluetooth/bluetooth-util.h +libbluetooth_util_la_SOURCES = \ + modules/bluetooth/a2dp-codecs.h \ + modules/bluetooth/bluetooth-util.c \ + modules/bluetooth/bluetooth-util.h libbluetooth_util_la_LDFLAGS = -avoid-version libbluetooth_util_la_LIBADD = $(MODULE_LIBADD) $(DBUS_LIBS) libbluetooth_util_la_CFLAGS = $(AM_CFLAGS) $(DBUS_CFLAGS) commit 654444eda9e57cd8a9326f6ad0e6aeb354e9b83d Author: Arun Raghavan <[email protected]> Date: Mon Oct 29 14:40:00 2012 +0530 tests: Minor cpu-test fixes for non-NEON builds The s16->float test is only run on ARM with NEOn at the moment, so we don't define that code in other cases. diff --git a/src/tests/cpu-test.c b/src/tests/cpu-test.c index f3308d9..e38e632 100644 --- a/src/tests/cpu-test.c +++ b/src/tests/cpu-test.c @@ -255,6 +255,9 @@ static void run_conv_test_float_to_s16(pa_convert_func_t func, pa_convert_func_t } } +/* This test is currently only run under NEON */ +#if defined (__arm__) && defined (__linux__) +#ifdef HAVE_NEON static void run_conv_test_s16_to_float(pa_convert_func_t func, pa_convert_func_t orig_func, int align, pa_bool_t correct, pa_bool_t perf) { PA_DECLARE_ALIGNED(8, float, f[SAMPLES]) = { 0 }; @@ -298,6 +301,8 @@ static void run_conv_test_s16_to_float(pa_convert_func_t func, pa_convert_func_t } PA_CPU_TEST_RUN_STOP } } +#endif /* HAVE_NEON */ +#endif /* defined (__arm__) && defined (__linux__) */ #if defined (__i386__) || defined (__amd64__) START_TEST (sconv_sse_test) { @@ -329,6 +334,7 @@ END_TEST #endif /* defined (__i386__) || defined (__amd64__) */ #if defined (__arm__) && defined (__linux__) +#ifdef HAVE_NEON START_TEST (sconv_neon_test) { pa_cpu_arm_flag_t flags = 0; pa_convert_func_t orig_from_func, neon_from_func; @@ -368,6 +374,7 @@ START_TEST (sconv_neon_test) { run_conv_test_s16_to_float(neon_to_func, orig_to_func, 7, TRUE, TRUE); } END_TEST +#endif /* HAVE_NEON */ #endif /* defined (__arm__) && defined (__linux__) */ #undef SAMPLES commit 2ec891d892d995f09f119fb0a36b976d3a53e013 Author: Arun Raghavan <[email protected]> Date: Mon Oct 29 13:26:25 2012 +0530 tests: Print average outer-loop iteration time in cpu-test diff --git a/src/tests/cpu-test.c b/src/tests/cpu-test.c index b791e7f..f3308d9 100644 --- a/src/tests/cpu-test.c +++ b/src/tests/cpu-test.c @@ -55,8 +55,12 @@ _s1 += _stop - _start; \ _s2 += (_stop - _start) * (_stop - _start); \ } \ - pa_log_debug("%s: %llu usec (min = %llu, max = %llu, stddev = %g).", _label, (long long unsigned int)_s1, \ - (long long unsigned int)_min, (long long unsigned int)_max, sqrt(_times2 * _s2 - _s1 * _s1) / _times2); \ + pa_log_debug("%s: %llu usec (avg: %g, min = %llu, max = %llu, stddev = %g).", _label, \ + (long long unsigned int)_s1, \ + ((double)_s1 / _times2), \ + (long long unsigned int)_min, \ + (long long unsigned int)_max, \ + sqrt(_times2 * _s2 - _s1 * _s1) / _times2); \ } /* Common defines for svolume tests */ commit 17b5a3613022837525fe853de55df7fa6897ce50 Author: Arun Raghavan <[email protected]> Date: Fri Oct 26 16:45:32 2012 +0530 tests: Increase sconv cpu-test timeout This takes longer on ARM CPUs, especially older ones. diff --git a/src/tests/cpu-test.c b/src/tests/cpu-test.c index dd21607..b791e7f 100644 --- a/src/tests/cpu-test.c +++ b/src/tests/cpu-test.c @@ -403,6 +403,7 @@ int main(int argc, char *argv[]) { tcase_add_test(tc, sconv_neon_test); #endif #endif + tcase_set_timeout(tc, 120); suite_add_tcase(s, tc); sr = srunner_create(s); commit 32b4e848aeddaf1dc57a6fbb9327df216162a410 Author: Arun Raghavan <[email protected]> Date: Fri Oct 26 16:39:11 2012 +0530 tests: Allow off-by-one error in sconv test With some optimised sconv implementations (read NEON), rounding inaccuracy might lead to a difference of 1 with the reference implementation. The inaccuracy is worth the performance gain. Also increases floating-point accuracy while printing errors to make errors easier to analyse. diff --git a/src/tests/cpu-test.c b/src/tests/cpu-test.c index 3c7ad7d..dd21607 100644 --- a/src/tests/cpu-test.c +++ b/src/tests/cpu-test.c @@ -231,8 +231,8 @@ static void run_conv_test_float_to_s16(pa_convert_func_t func, pa_convert_func_t func(nsamples, floats, samples); for (i = 0; i < nsamples; i++) { - if (samples[i] != samples_ref[i]) { - pa_log_debug("%d: %04x != %04x (%f)\n", i, samples[i], samples_ref[i], floats[i]); + if (abs(samples[i] - samples_ref[i]) > 1) { + pa_log_debug("%d: %04x != %04x (%.24f)\n", i, samples[i], samples_ref[i], floats[i]); fail(); } } @@ -248,8 +248,6 @@ static void run_conv_test_float_to_s16(pa_convert_func_t func, pa_convert_func_t PA_CPU_TEST_RUN_START("orig", TIMES, TIMES2) { orig_func(nsamples, floats, samples_ref); } PA_CPU_TEST_RUN_STOP - - fail_unless(memcmp(samples_ref, samples, sizeof(nsamples)) == 0); } } @@ -277,8 +275,8 @@ static void run_conv_test_s16_to_float(pa_convert_func_t func, pa_convert_func_t func(nsamples, samples, floats); for (i = 0; i < nsamples; i++) { - if (floats[i] != floats_ref[i]) { - pa_log_debug("%d: %f != %f (%d)\n", i, floats[i], floats_ref[i], samples[i]); + if (abs(floats[i] - floats_ref[i]) > 1) { + pa_log_debug("%d: %.24f != %.24f (%d)\n", i, floats[i], floats_ref[i], samples[i]); fail(); } } @@ -294,8 +292,6 @@ static void run_conv_test_s16_to_float(pa_convert_func_t func, pa_convert_func_t PA_CPU_TEST_RUN_START("orig", TIMES, TIMES2) { orig_func(nsamples, samples, floats_ref); } PA_CPU_TEST_RUN_STOP - - fail_unless(memcmp(floats_ref, floats, nsamples * sizeof(float)) == 0); } } commit 1a8ec3c3e05f0f085c6831ee46205c1238fd2a08 Author: Arun Raghavan <[email protected]> Date: Fri Oct 26 16:36:49 2012 +0530 sconv: Fix NEON sconv rounding code Rounding with 0.5 causes us to always round up for any value of the form x.5. IEEE754 specifies round-to-nearest-even as the behaviour in this case. This might not always be possible with NEON code, but this change gets us much closer to it. diff --git a/src/pulsecore/sconv_neon.c b/src/pulsecore/sconv_neon.c index 81dd97e..6fd966d 100644 --- a/src/pulsecore/sconv_neon.c +++ b/src/pulsecore/sconv_neon.c @@ -40,18 +40,14 @@ static void pa_sconv_s16le_from_f32ne_neon(unsigned n, const float *src, int16_t "vneg.f32 q3, q2 \n\t" "vdup.f32 q4, %[scale] \n\t" "vdup.u32 q5, %[mask] \n\t" - "vdup.f32 q6, %[half] \n\t" "1: \n\t" "vld1.32 {q0}, [%[src]]! \n\t" "vmin.f32 q0, q0, q2 \n\t" /* clamp */ "vmax.f32 q0, q0, q3 \n\t" "vmul.f32 q0, q0, q4 \n\t" /* scale */ - "vand.u32 q1, q0, q5 \n\t" - "vorr.u32 q1, q1, q6 \n\t" /* round */ - "vadd.f32 q0, q0, q1 \n\t" - "vcvt.s32.f32 q0, q0 \n\t" /* narrow */ - "vmovn.i32 d0, q0 \n\t" + "vcvt.s32.f32 q0, q0, #16 \n\t" /* narrow */ + "vrshrn.s32 d0, q0, #16 \n\t" "subs %[n], %[n], #1 \n\t" "vst1.16 {d0}, [%[dst]]! \n\t" "bgt 1b \n\t" @@ -59,7 +55,7 @@ static void pa_sconv_s16le_from_f32ne_neon(unsigned n, const float *src, int16_t "2: \n\t" : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n) /* output operands (or input operands that get modified) */ - : [plusone] "r" (1.0f), [scale] "r" (32767.0f), [half] "r" (0.5f), [mask] "r" (0x80000000) /* input operands */ + : [plusone] "r" (1.0f), [scale] "r" (32767.0f), [mask] "r" (0x80000000) /* input operands */ : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5", "q6" /* clobber list */ ); commit f558327cca3188bd9ed04165d96d111efa078870 Author: Peter Meerwald <[email protected]> Date: Sat Oct 27 14:59:43 2012 +0200 tests: Implement test code for ARM NEON sconv s16_to_float Signed-off-by: Peter Meerwald <[email protected]> diff --git a/src/tests/cpu-test.c b/src/tests/cpu-test.c index 35775bd..3c7ad7d 100644 --- a/src/tests/cpu-test.c +++ b/src/tests/cpu-test.c @@ -253,6 +253,52 @@ static void run_conv_test_float_to_s16(pa_convert_func_t func, pa_convert_func_t } } +static void run_conv_test_s16_to_float(pa_convert_func_t func, pa_convert_func_t orig_func, int align, pa_bool_t correct, + pa_bool_t perf) { + PA_DECLARE_ALIGNED(8, float, f[SAMPLES]) = { 0 }; + PA_DECLARE_ALIGNED(8, float, f_ref[SAMPLES]) = { 0 }; + PA_DECLARE_ALIGNED(8, int16_t, s[SAMPLES]); + float *floats, *floats_ref; + int16_t *samples; + int i, nsamples; + + /* Force sample alignment as requested */ + floats = f + (8 - align); + floats_ref = f_ref + (8 - align); + samples = s + (8 - align); + nsamples = SAMPLES - (8 - align); + + pa_random(samples, nsamples * sizeof(int16_t)); + + if (correct) { + pa_log_debug("Testing sconv correctness with %d byte alignment", align); + + orig_func(nsamples, samples, floats_ref); + func(nsamples, samples, floats); + + for (i = 0; i < nsamples; i++) { + if (floats[i] != floats_ref[i]) { + pa_log_debug("%d: %f != %f (%d)\n", i, floats[i], floats_ref[i], samples[i]); + fail(); + } + } + } + + if (perf) { + pa_log_debug("Testing sconv performance with %d byte alignment", align); + + PA_CPU_TEST_RUN_START("func", TIMES, TIMES2) { + func(nsamples, samples, floats); + } PA_CPU_TEST_RUN_STOP + + PA_CPU_TEST_RUN_START("orig", TIMES, TIMES2) { + orig_func(nsamples, samples, floats_ref); + } PA_CPU_TEST_RUN_STOP + + fail_unless(memcmp(floats_ref, floats, nsamples * sizeof(float)) == 0); + } +} + #if defined (__i386__) || defined (__amd64__) START_TEST (sconv_sse_test) { pa_cpu_x86_flag_t flags = 0; @@ -285,7 +331,8 @@ END_TEST #if defined (__arm__) && defined (__linux__) START_TEST (sconv_neon_test) { pa_cpu_arm_flag_t flags = 0; - pa_convert_func_t orig_func, neon_func; + pa_convert_func_t orig_from_func, neon_from_func; + pa_convert_func_t orig_to_func, neon_to_func; pa_cpu_get_arm_flags(&flags); @@ -294,19 +341,31 @@ START_TEST (sconv_neon_test) { return; } - orig_func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE); + orig_from_func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE); + orig_to_func = pa_get_convert_to_float32ne_function(PA_SAMPLE_S16LE); pa_convert_func_init_neon(flags); - neon_func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE); + neon_from_func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE); + neon_to_func = pa_get_convert_to_float32ne_function(PA_SAMPLE_S16LE); pa_log_debug("Checking NEON sconv (float -> s16)"); - run_conv_test_float_to_s16(neon_func, orig_func, 0, TRUE, FALSE); - run_conv_test_float_to_s16(neon_func, orig_func, 1, TRUE, FALSE); - run_conv_test_float_to_s16(neon_func, orig_func, 2, TRUE, FALSE); - run_conv_test_float_to_s16(neon_func, orig_func, 3, TRUE, FALSE); - run_conv_test_float_to_s16(neon_func, orig_func, 4, TRUE, FALSE); - run_conv_test_float_to_s16(neon_func, orig_func, 5, TRUE, FALSE); - run_conv_test_float_to_s16(neon_func, orig_func, 6, TRUE, FALSE); - run_conv_test_float_to_s16(neon_func, orig_func, 7, TRUE, TRUE); + run_conv_test_float_to_s16(neon_from_func, orig_from_func, 0, TRUE, FALSE); + run_conv_test_float_to_s16(neon_from_func, orig_from_func, 1, TRUE, FALSE); + run_conv_test_float_to_s16(neon_from_func, orig_from_func, 2, TRUE, FALSE); + run_conv_test_float_to_s16(neon_from_func, orig_from_func, 3, TRUE, FALSE); + run_conv_test_float_to_s16(neon_from_func, orig_from_func, 4, TRUE, FALSE); + run_conv_test_float_to_s16(neon_from_func, orig_from_func, 5, TRUE, FALSE); + run_conv_test_float_to_s16(neon_from_func, orig_from_func, 6, TRUE, FALSE); + run_conv_test_float_to_s16(neon_from_func, orig_from_func, 7, TRUE, TRUE); + + pa_log_debug("Checking NEON sconv (s16 -> float)"); + run_conv_test_s16_to_float(neon_to_func, orig_to_func, 0, TRUE, FALSE); + run_conv_test_s16_to_float(neon_to_func, orig_to_func, 1, TRUE, FALSE); + run_conv_test_s16_to_float(neon_to_func, orig_to_func, 2, TRUE, FALSE); + run_conv_test_s16_to_float(neon_to_func, orig_to_func, 3, TRUE, FALSE); + run_conv_test_s16_to_float(neon_to_func, orig_to_func, 4, TRUE, FALSE); + run_conv_test_s16_to_float(neon_to_func, orig_to_func, 5, TRUE, FALSE); + run_conv_test_s16_to_float(neon_to_func, orig_to_func, 6, TRUE, FALSE); + run_conv_test_s16_to_float(neon_to_func, orig_to_func, 7, TRUE, TRUE); } END_TEST #endif /* defined (__arm__) && defined (__linux__) */ commit bd7229b38a4d2ddd08af65cccec8b1df905ea3a8 Author: Peter Meerwald <[email protected]> Date: Fri Oct 26 13:04:47 2012 +0200 tests: Fix test description in cpu-test Signed-off-by: Peter Meerwald <[email protected]> diff --git a/src/tests/cpu-test.c b/src/tests/cpu-test.c index 56b0cd7..35775bd 100644 --- a/src/tests/cpu-test.c +++ b/src/tests/cpu-test.c @@ -269,7 +269,7 @@ START_TEST (sconv_sse_test) { pa_convert_func_init_sse(flags); sse_func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE); - pa_log_debug("Checking SSE sconv (s16 -> float)"); + pa_log_debug("Checking SSE sconv (float -> s16)"); run_conv_test_float_to_s16(sse_func, orig_func, 0, TRUE, FALSE); run_conv_test_float_to_s16(sse_func, orig_func, 1, TRUE, FALSE); run_conv_test_float_to_s16(sse_func, orig_func, 2, TRUE, FALSE); @@ -298,7 +298,7 @@ START_TEST (sconv_neon_test) { pa_convert_func_init_neon(flags); neon_func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE); - pa_log_debug("Checking NEON sconv (s16 -> float)"); + pa_log_debug("Checking NEON sconv (float -> s16)"); run_conv_test_float_to_s16(neon_func, orig_func, 0, TRUE, FALSE); run_conv_test_float_to_s16(neon_func, orig_func, 1, TRUE, FALSE); run_conv_test_float_to_s16(neon_func, orig_func, 2, TRUE, FALSE); commit 1319c4533a7d3a2562907d043d238bfe3d51d335 Author: Peter Meerwald <[email protected]> Date: Tue Oct 23 17:54:57 2012 +0200 core: Add ARM NEON optimized sample conversion code final: * includes some minor style fixes and build-time changes to allow building a single binary for neon and non-neon systems v4: * fix for sample length < 4 v3: * convert from intrinsics to inline assembly v2: * load and store data with vld1/vld1q and vst1/vst1q, resp., to work around alignment issues of compiler-generated vldmia instruction * remove redundant check for NEON flags Ubuntu/Linaro gcc 4.6.3 arm-linux-gnueabi-gcc -O2 -mcpu=cortex-a8 -mfloat-abi=softfp -mfpu=neon runtime on beagle-xm: D: [pulseaudio] sconv_neon.c: checking NEON sconv_s16le_from_float I: [pulseaudio] sconv_neon.c: NEON: 3754 usec. I: [pulseaudio] sconv_neon.c: ref: 58594 usec. D: [pulseaudio] sconv_neon.c: checking NEON sconv_s16le_to_float I: [pulseaudio] sconv_neon.c: NEON: 1831 usec. I: [pulseaudio] sconv_neon.c: ref: 10528 usec. I: [pulseaudio] sconv_neon.c: Initialising ARM NEON optimized conversions. conversion may be off by one for some samples due to rounding issues diff --git a/configure.ac b/configure.ac index 2154429..c91c572 100644 --- a/configure.ac +++ b/configure.ac @@ -317,6 +317,34 @@ case $host in ;; esac +#### NEON optimisations #### +AC_ARG_ENABLE([neon-opt], + AS_HELP_STRING([--enable-neon-opt], [Enable NEON optimisations on ARM CPUs that support it])) + +AS_IF([test "x$enable_neon_opt" != "xno"], + [save_CFLAGS="$CFLAGS"; CFLAGS="$CFLAGS -mfpu=neon" + AC_COMPILE_IFELSE( + AC_LANG_PROGRAM([], []), + [ + HAVE_NEON=1 + NEON_CFLAGS="-mfpu=neon" + ], + [ + HAVE_NEON=0 + NEON_CFLAGS= + ]) + CFLAGS="$save_CFLAGS" + ], + [HAVE_NEON=0]) + +AS_IF([test "x$enable_neon_opt" = "xyes" && test "x$HAVE_NEON" = "x0"], + [AC_MSG_ERROR([*** Compiler does not support -mfpu=neon])]) + +AC_SUBST(HAVE_NEON) +AC_SUBST(NEON_CFLAGS) +AM_CONDITIONAL([HAVE_NEON], [test "x$HAVE_NEON" = x1]) +AS_IF([test "x$HAVE_NEON" = "x1"], AC_DEFINE([HAVE_NEON], 1, [Have NEON support?])) + #### libtool stuff #### diff --git a/src/Makefile.am b/src/Makefile.am index f80c448..84598b2 100644 --- a/src/Makefile.am +++ b/src/Makefile.am @@ -813,6 +813,7 @@ libpulsedsp_la_LDFLAGS = $(AM_LDFLAGS) -avoid-version -disable-static ################################### lib_LTLIBRARIES += libpulsecore-@[email protected] +noinst_LTLIBRARIES = # Pure core stuff libpulsecore_@PA_MAJORMINOR@_la_SOURCES = \ @@ -870,6 +871,13 @@ libpulsecore_@PA_MAJORMINOR@_la_CFLAGS = $(AM_CFLAGS) $(SERVER_CFLAGS) $(LIBSAMP libpulsecore_@PA_MAJORMINOR@_la_LDFLAGS = $(AM_LDFLAGS) -avoid-version libpulsecore_@PA_MAJORMINOR@_la_LIBADD = $(AM_LIBADD) $(LIBLTDL) $(LIBSAMPLERATE_LIBS) $(LIBSPEEX_LIBS) $(LIBSNDFILE_LIBS) $(WINSOCK_LIBS) $(LTLIBICONV) libpulsecommon-@[email protected] libpulse.la libpulsecore-foreign.la +if HAVE_NEON +noinst_LTLIBRARIES += libpulsecore_sconv_neon.la +libpulsecore_sconv_neon_la_SOURCES = pulsecore/sconv_neon.c +libpulsecore_sconv_neon_la_CFLAGS = $(AM_CFLAGS) $(NEON_CFLAGS) +libpulsecore_@PA_MAJORMINOR@_la_LIBADD += libpulsecore_sconv_neon.la +endif + if HAVE_ORC ORC_SOURCE += pulsecore/svolume libpulsecore_@PA_MAJORMINOR@_la_SOURCES += pulsecore/svolume_orc.c @@ -909,7 +917,7 @@ libpulsecore_@PA_MAJORMINOR@_la_SOURCES += pulsecore/database-simple.c endif # We split the foreign code off to not be annoyed by warnings we don't care about -noinst_LTLIBRARIES = libpulsecore-foreign.la +noinst_LTLIBRARIES += libpulsecore-foreign.la libpulsecore_foreign_la_SOURCES = \ pulsecore/ffmpeg/resample2.c pulsecore/ffmpeg/avcodec.h pulsecore/ffmpeg/dsputil.h diff --git a/src/pulsecore/cpu-arm.c b/src/pulsecore/cpu-arm.c index 934c7fe..05668f1 100644 --- a/src/pulsecore/cpu-arm.c +++ b/src/pulsecore/cpu-arm.c @@ -142,6 +142,10 @@ pa_bool_t pa_cpu_init_arm(pa_cpu_arm_flag_t *flags) { if (*flags & PA_CPU_ARM_V6) pa_volume_func_init_arm(*flags); +#ifdef HAVE_NEON + if (*flags & PA_CPU_ARM_NEON) + pa_convert_func_init_neon(*flags); +#endif return TRUE; diff --git a/src/pulsecore/cpu-arm.h b/src/pulsecore/cpu-arm.h index 91aa659..d2d3f5c 100644 --- a/src/pulsecore/cpu-arm.h +++ b/src/pulsecore/cpu-arm.h @@ -26,6 +26,10 @@ #include <stdint.h> #include <pulsecore/macro.h> +#ifndef PACKAGE +#error "Please include config.h before including this file!" +#endif + typedef enum pa_cpu_arm_flag { PA_CPU_ARM_V6 = (1 << 0), PA_CPU_ARM_V7 = (1 << 1), @@ -41,4 +45,8 @@ pa_bool_t pa_cpu_init_arm(pa_cpu_arm_flag_t *flags); /* some optimized functions */ void pa_volume_func_init_arm(pa_cpu_arm_flag_t flags); +#ifdef HAVE_NEON +void pa_convert_func_init_neon(pa_cpu_arm_flag_t flags); +#endif + #endif /* foocpuarmhfoo */ diff --git a/src/pulsecore/sconv_neon.c b/src/pulsecore/sconv_neon.c new file mode 100644 index 0000000..81dd97e --- /dev/null +++ b/src/pulsecore/sconv_neon.c @@ -0,0 +1,110 @@ +/*** + This file is part of PulseAudio. + + Copyright 2012 Peter Meerwald <[email protected]> + + PulseAudio is free software; you can redistribute it and/or modify + it under the terms of the GNU Lesser General Public License as published + by the Free Software Foundation; either version 2.1 of the License, + or (at your option) any later version. + + PulseAudio is distributed in the hope that it will be useful, but + WITHOUT ANY WARRANTY; without even the implied warranty of + MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the GNU + General Public License for more details. +***/ + +#ifdef HAVE_CONFIG_H +#include <config.h> +#endif + +#include <pulse/rtclock.h> + +#include <pulsecore/macro.h> +#include <pulsecore/endianmacros.h> + +#include "cpu-arm.h" +#include "sconv.h" + +#include <math.h> +#include <arm_neon.h> + +static void pa_sconv_s16le_from_f32ne_neon(unsigned n, const float *src, int16_t *dst) { + unsigned i = n & 3; + + __asm__ __volatile__ ( + "movs %[n], %[n], lsr #2 \n\t" + "beq 2f \n\t" + + "vdup.f32 q2, %[plusone] \n\t" + "vneg.f32 q3, q2 \n\t" + "vdup.f32 q4, %[scale] \n\t" + "vdup.u32 q5, %[mask] \n\t" + "vdup.f32 q6, %[half] \n\t" + + "1: \n\t" + "vld1.32 {q0}, [%[src]]! \n\t" + "vmin.f32 q0, q0, q2 \n\t" /* clamp */ + "vmax.f32 q0, q0, q3 \n\t" + "vmul.f32 q0, q0, q4 \n\t" /* scale */ + "vand.u32 q1, q0, q5 \n\t" + "vorr.u32 q1, q1, q6 \n\t" /* round */ + "vadd.f32 q0, q0, q1 \n\t" + "vcvt.s32.f32 q0, q0 \n\t" /* narrow */ + "vmovn.i32 d0, q0 \n\t" + "subs %[n], %[n], #1 \n\t" + "vst1.16 {d0}, [%[dst]]! \n\t" + "bgt 1b \n\t" + + "2: \n\t" + + : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n) /* output operands (or input operands that get modified) */ + : [plusone] "r" (1.0f), [scale] "r" (32767.0f), [half] "r" (0.5f), [mask] "r" (0x80000000) /* input operands */ + : "memory", "cc", "q0", "q1", "q2", "q3", "q4", "q5", "q6" /* clobber list */ + ); + + /* leftovers */ + while (i--) { + *dst++ = (int16_t) lrintf(PA_CLAMP_UNLIKELY(*src, -1.0f, 1.0f) * 0x7FFF); + src++; + } +} + +static void pa_sconv_s16le_to_f32ne_neon(unsigned n, const int16_t *src, float *dst) { + unsigned i = n & 3; + + const float invscale = 1.0f / 0x7FFF; + + __asm__ __volatile__ ( + "movs %[n], %[n], lsr #2 \n\t" + "beq 2f \n\t" + + "vdup.f32 q1, %[invscale] \n\t" + + "1: \n\t" + "vld1.16 {d0}, [%[src]]! \n\t" + "vmovl.s16 q0, d0 \n\t" + "vcvt.f32.s32 q0, q0 \n\t" + "vmul.f32 q0, q0, q1 \n\t" + "subs %[n], %[n], #1 \n\t" + "vst1.32 {q0}, [%[dst]]! \n\t" + "bgt 1b \n\t" + + "2: \n\t" + + : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n) /* output operands (or input operands that get modified) */ + : [invscale] "r" (invscale) /* input operands */ + : "memory", "cc", "q0", "q1" /* clobber list */ + ); + + /* leftovers */ + while (i--) { + *dst++ = *src++ * invscale; + } +} + +void pa_convert_func_init_neon(pa_cpu_arm_flag_t flags) { + pa_log_info("Initialising ARM NEON optimized conversions."); + pa_set_convert_from_float32ne_function(PA_SAMPLE_S16LE, (pa_convert_func_t) pa_sconv_s16le_from_f32ne_neon); + pa_set_convert_to_float32ne_function(PA_SAMPLE_S16LE, (pa_convert_func_t) pa_sconv_s16le_to_f32ne_neon); +} diff --git a/src/tests/cpu-test.c b/src/tests/cpu-test.c index 5b83d84..56b0cd7 100644 --- a/src/tests/cpu-test.c +++ b/src/tests/cpu-test.c @@ -282,6 +282,35 @@ START_TEST (sconv_sse_test) { END_TEST #endif /* defined (__i386__) || defined (__amd64__) */ +#if defined (__arm__) && defined (__linux__) +START_TEST (sconv_neon_test) { + pa_cpu_arm_flag_t flags = 0; + pa_convert_func_t orig_func, neon_func; + + pa_cpu_get_arm_flags(&flags); + + if (!(flags & PA_CPU_ARM_NEON)) { + pa_log_info("NEON not supported. Skipping"); + return; + } + + orig_func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE); + pa_convert_func_init_neon(flags); + neon_func = pa_get_convert_from_float32ne_function(PA_SAMPLE_S16LE); + + pa_log_debug("Checking NEON sconv (s16 -> float)"); + run_conv_test_float_to_s16(neon_func, orig_func, 0, TRUE, FALSE); + run_conv_test_float_to_s16(neon_func, orig_func, 1, TRUE, FALSE); + run_conv_test_float_to_s16(neon_func, orig_func, 2, TRUE, FALSE); + run_conv_test_float_to_s16(neon_func, orig_func, 3, TRUE, FALSE); + run_conv_test_float_to_s16(neon_func, orig_func, 4, TRUE, FALSE); + run_conv_test_float_to_s16(neon_func, orig_func, 5, TRUE, FALSE); + run_conv_test_float_to_s16(neon_func, orig_func, 6, TRUE, FALSE); + run_conv_test_float_to_s16(neon_func, orig_func, 7, TRUE, TRUE); +} +END_TEST +#endif /* defined (__arm__) && defined (__linux__) */ + #undef SAMPLES #undef TIMES /* End conversion tests */ @@ -314,6 +343,11 @@ int main(int argc, char *argv[]) { #if defined (__i386__) || defined (__amd64__) tcase_add_test(tc, sconv_sse_test); #endif +#if defined (__arm__) && defined (__linux__) +#if HAVE_NEON + tcase_add_test(tc, sconv_neon_test); +#endif +#endif suite_add_tcase(s, tc); sr = srunner_create(s); _______________________________________________ pulseaudio-commits mailing list [email protected] http://lists.freedesktop.org/mailman/listinfo/pulseaudio-commits
