[pulseaudio-discuss] [PATCH 4/6 v3] core: Add ARM NEON optimized mono-to-stereo/stereo-to-mono remapping code
Peter Meerwald
pmeerw at pmeerw.net
Tue Jul 24 01:20:34 PDT 2012
From: Peter Meerwald <p.meerwald at bct-electronic.com>
v3:
* fix test code: init float and int map_table
* different code path for Cortex-A8 and later (-A9, A15, unknown)
* convert from intrinsics to inline assembly
v2:
* add ARM NEON stereo-to-mono remapping code
* static __attribute__ ((noinline)) is necessary to prevent inlining and
work around gcc 4.6 ICE, see https://bugs.launchpad.net/bugs/936863
* call test code, the reference implementation is obtained using
pa_get_init_remap_func()
* remove check for NEON flags
v1:
* ARM NEON mono-to-stereo remapping code
compiled with 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] remap_neon.c: checking NEON remap_stereo_to_mono(float)
I: [pulseaudio] remap.c: Using stereo to mono remapping
I: [pulseaudio] remap_neon.c: NEON: 3082 usec.
I: [pulseaudio] remap_neon.c: ref: 24201 usec.
D: [pulseaudio] remap_neon.c: checking NEON remap_stereo_to_mono(s16)
I: [pulseaudio] remap.c: Using stereo to mono remapping
I: [pulseaudio] remap_neon.c: NEON: 1190 usec.
I: [pulseaudio] remap_neon.c: ref: 5615 usec.
D: [pulseaudio] remap_neon.c: checking NEON remap_mono_to_stereo(float)
I: [pulseaudio] remap.c: Using mono to stereo remapping
I: [pulseaudio] remap_neon.c: NEON/A8: 2350 usec.
I: [pulseaudio] remap_neon.c: NEON/A9: 4730 usec.
I: [pulseaudio] remap_neon.c: ref: 3601 usec.
D: [pulseaudio] remap_neon.c: checking NEON remap_mono_to_stereo(s16)
I: [pulseaudio] remap.c: Using mono to stereo remapping
I: [pulseaudio] remap_neon.c: NEON: 1403 usec.
I: [pulseaudio] remap_neon.c: ref: 3724 usec.
I: [pulseaudio] remap_neon.c: Initialising ARM NEON optimized remappers.
Signed-off-by: Peter Meerwald <p.meerwald at bct-electronic.com>
---
src/Makefile.am | 1 +
src/pulsecore/remap_neon.c | 528 ++++++++++++++++++++++++++++++++++++++++++++
2 files changed, 529 insertions(+), 0 deletions(-)
create mode 100644 src/pulsecore/remap_neon.c
diff --git a/src/Makefile.am b/src/Makefile.am
index 63ad837..df25efc 100644
--- a/src/Makefile.am
+++ b/src/Makefile.am
@@ -819,6 +819,7 @@ libpulsecore_ at PA_MAJORMINOR@_la_SOURCES = \
pulsecore/play-memchunk.c pulsecore/play-memchunk.h \
pulsecore/remap.c pulsecore/remap.h \
pulsecore/remap_mmx.c pulsecore/remap_sse.c \
+ pulsecore/remap_neon.c \
pulsecore/resampler.c pulsecore/resampler.h \
pulsecore/rtpoll.c pulsecore/rtpoll.h \
pulsecore/sample-util.c pulsecore/sample-util.h \
diff --git a/src/pulsecore/remap_neon.c b/src/pulsecore/remap_neon.c
new file mode 100644
index 0000000..0ecced1
--- /dev/null
+++ b/src/pulsecore/remap_neon.c
@@ -0,0 +1,528 @@
+/***
+ This file is part of PulseAudio.
+
+ Copyright 2012 Peter Meerwald <p.meerwald at bct-electronic.com>
+
+ 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 <pulse/sample.h>
+#include <pulsecore/log.h>
+#include <pulsecore/macro.h>
+
+#include "cpu-arm.h"
+#include "remap.h"
+
+#if defined(__ARM_NEON__)
+
+#include <math.h>
+#include <arm_neon.h>
+
+#define RUN_TEST
+
+static inline void mono_to_stereo_float_neon_a8(float *dst, const float *src, unsigned n) {
+ int i = n & 3;
+
+ asm volatile (
+ "mov %[n], %[n], lsr #2\n\t"
+ "1:\n\t"
+ "vld1.32 {q0}, [%[src]]!\n\t"
+ "vmov q1, q0\n\t"
+ "subs %[n], %[n], #1\n\t"
+ "vst2.32 {q0,q1}, [%[dst]]!\n\t"
+ "bgt 1b\n\t"
+ /* output operands (or input operands that get modified) */
+ : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n)
+ : /* input operands */
+ : "memory", "cc", "q0", "q1" /* clobber list */
+ );
+
+ while (i--) {
+ dst[0] = dst[1] = src[0];
+ src++;
+ dst += 2;
+ }
+}
+
+static inline void mono_to_stereo_float_neon_a9(float *dst, const float *src, unsigned n) {
+ int i = n & 1;
+
+ asm volatile (
+ "mov %[n], %[n], lsr #1\n\t"
+ "1:\n\t"
+ "ldm %[src]!, {r4,r6}\n\t"
+ "mov r5, r4\n\t"
+ "mov r7, r6\n\t"
+ "subs %[n], %[n], #1\n\t"
+ "stm %[dst]!, {r4-r7}\n\t"
+ "bgt 1b\n\t"
+ /* output operands (or input operands that get modified) */
+ : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n)
+ : /* input operands */
+ : "memory", "cc", "r4", "r5", "r6", "r7" /* clobber list */
+ );
+
+ while (i--) {
+ dst[0] = dst[1] = src[0];
+ src++;
+ dst += 2;
+ }
+}
+
+static inline void mono_to_stereo_int16_neon(int16_t *dst, const int16_t *src, unsigned n) {
+ int i = n & 7;
+
+ asm volatile (
+ "mov %[n], %[n], lsr #3\n\t"
+ "1:\n\t"
+ "vld1.16 {q0}, [%[src]]!\n\t"
+ "vmov q1, q0\n\t"
+ "subs %[n], %[n], #1\n\t"
+ "vst2.16 {q0,q1}, [%[dst]]!\n\t"
+ "bgt 1b\n\t"
+ /* output operands (or input operands that get modified) */
+ : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n)
+ : /* input operands */
+ : "memory", "cc", "q0", "q1" /* clobber list */
+ );
+
+ while (i--) {
+ dst[0] = dst[1] = src[0];
+ src++;
+ dst += 2;
+ }
+}
+
+static void remap_mono_to_stereo_neon_a9(pa_remap_t *m, void *dst, const void *src, unsigned n) {
+ switch (*m->format) {
+ case PA_SAMPLE_FLOAT32NE:
+ mono_to_stereo_float_neon_a9(dst, src, n);
+ break;
+ case PA_SAMPLE_S16NE:
+ mono_to_stereo_int16_neon(dst, src, n);
+ break;
+ default:
+ pa_assert_not_reached();
+ }
+}
+
+static void remap_mono_to_stereo_neon_a8(pa_remap_t *m, void *dst, const void *src, unsigned n) {
+ switch (*m->format) {
+ case PA_SAMPLE_FLOAT32NE:
+ mono_to_stereo_float_neon_a8(dst, src, n);
+ break;
+ case PA_SAMPLE_S16NE:
+ mono_to_stereo_int16_neon(dst, src, n);
+ break;
+ default:
+ pa_assert_not_reached();
+ }
+}
+
+static inline void stereo_to_mono_float_neon(float *dst, const float *src, unsigned n) {
+ int i = n & 3;
+
+ asm volatile (
+ "mov %[n], %[n], lsr #2\n\t"
+ "1:\n\t"
+ "vld2.32 {q0,q1}, [%[src]]!\n\t"
+ "vadd.f32 q0, q0, q1\n\t"
+ "subs %[n], %[n], #1\n\t"
+ "vst1.32 {q0}, [%[dst]]!\n\t"
+ "bgt 1b\n\t"
+ /* output operands (or input operands that get modified) */
+ : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n)
+ : /* input operands */
+ : "memory", "cc" /* clobber list */
+ );
+
+ while (i--) {
+ dst[0] = src[0] + src[1];
+ src += 2;
+ dst++;
+ }
+}
+
+static inline void stereo_to_mono_int16_neon(int16_t *dst, const int16_t *src, unsigned n) {
+ int i = n & 7;
+
+ asm volatile (
+ "mov %[n], %[n], lsr #3\n\t"
+ "1:\n\t"
+ "vld2.16 {q0,q1}, [%[src]]!\n\t"
+ "vadd.s16 q0, q0, q1\n\t"
+ "subs %[n], %[n], #1\n\t"
+ "vst1.16 {q0}, [%[dst]]!\n\t"
+ "bgt 1b\n\t"
+ /* output operands (or input operands that get modified) */
+ : [dst] "+r" (dst), [src] "+r" (src), [n] "+r" (n)
+ : /* input operands */
+ : "memory", "cc" /* clobber list */
+ );
+
+ while (i--) {
+ dst[0] = src[0] + src[1];
+ src += 2;
+ dst++;
+ }
+}
+
+static void remap_stereo_to_mono_neon(pa_remap_t *m, void *dst, const void *src, unsigned n) {
+ switch (*m->format) {
+ case PA_SAMPLE_FLOAT32NE:
+ stereo_to_mono_float_neon(dst, src, n);
+ break;
+ case PA_SAMPLE_S16NE:
+ stereo_to_mono_int16_neon(dst, src, n);
+ break;
+ default:
+ pa_assert_not_reached();
+ }
+}
+
+#ifdef RUN_TEST
+#define SAMPLES 1019
+#define TIMES 1000
+
+static void run_test_float_mono_to_stereo(void) {
+ float stereo[2*SAMPLES];
+ float stereo_ref[2*SAMPLES];
+ float mono[SAMPLES];
+ int i;
+ pa_usec_t start, stop;
+ pa_sample_format_t sf;
+ pa_remap_t remap;
+ pa_sample_spec iss, oss;
+
+ pa_init_remap_func_t remap_init_func = pa_get_init_remap_func();
+
+ pa_log_debug("checking NEON remap_mono_to_stereo(float)");
+
+ memset(stereo_ref, 0, sizeof(stereo_ref));
+ memset(stereo, 0, sizeof(stereo));
+
+ for (i = 0; i < SAMPLES; i++) {
+ mono[i] = rand()/(float) RAND_MAX - 0.5f;
+ }
+
+ sf = PA_SAMPLE_FLOAT32NE;
+ remap.format = &sf;
+ iss.format = PA_SAMPLE_FLOAT32NE;
+ iss.channels = 1;
+ oss.format = PA_SAMPLE_FLOAT32NE;
+ oss.channels = 2;
+ remap.i_ss = &iss;
+ remap.o_ss = &oss;
+ remap.map_table_f[0][0] = 1.0;
+ remap.map_table_f[1][0] = 1.0;
+ remap.map_table_i[0][0] = 0x10000;
+ remap.map_table_i[1][0] = 0x10000;
+ remap_init_func(&remap);
+ if (!remap.do_remap) {
+ pa_log_debug("no reference remapping function, abort test");
+ return;
+ }
+
+ remap.do_remap(&remap, stereo_ref, mono, SAMPLES);
+ remap_mono_to_stereo_neon_a9(&remap, stereo, mono, SAMPLES);
+ for (i = 0; i < 2*SAMPLES; i++) {
+ if (fabsf(stereo[i] - stereo_ref[i]) > 0.00001) {
+ pa_log_debug("A9 %d: %.3f != %.3f (%.3f)", i, stereo[i], stereo_ref[i],
+ mono[i/2]);
+ break;
+ }
+ }
+
+ remap_mono_to_stereo_neon_a8(&remap, stereo, mono, SAMPLES);
+ for (i = 0; i < 2*SAMPLES; i++) {
+ if (fabsf(stereo[i] - stereo_ref[i]) > 0.00001) {
+ pa_log_debug("A8 %d: %.3f != %.3f (%.3f)", i, stereo[i], stereo_ref[i],
+ mono[i/2]);
+ break;
+ }
+ }
+
+ start = pa_rtclock_now();
+ for (i = 0; i < TIMES; i++) {
+ remap_mono_to_stereo_neon_a8(&remap, stereo, mono, SAMPLES);
+ }
+ stop = pa_rtclock_now();
+ pa_log_info("NEON/A8: %llu usec.", (long long unsigned int)(stop - start));
+
+ start = pa_rtclock_now();
+ for (i = 0; i < TIMES; i++) {
+ remap_mono_to_stereo_neon_a9(&remap, stereo, mono, SAMPLES);
+ }
+ stop = pa_rtclock_now();
+ pa_log_info("NEON/A9: %llu usec.", (long long unsigned int)(stop - start));
+
+ start = pa_rtclock_now();
+ for (i = 0; i < TIMES; i++) {
+ remap.do_remap(&remap, stereo_ref, mono, SAMPLES);
+ }
+ stop = pa_rtclock_now();
+ pa_log_info("ref: %llu usec.", (long long unsigned int)(stop - start));
+}
+
+static void run_test_s16_mono_to_stereo(void) {
+ int16_t stereo[2*SAMPLES];
+ int16_t stereo_ref[2*SAMPLES];
+ int16_t mono[SAMPLES];
+ int i;
+ pa_usec_t start, stop;
+ pa_sample_format_t sf;
+ pa_remap_t remap;
+ pa_sample_spec iss, oss;
+
+ pa_init_remap_func_t remap_init_func = pa_get_init_remap_func();
+
+ pa_log_debug("checking NEON remap_mono_to_stereo(s16)");
+
+ memset(stereo_ref, 0, sizeof(stereo_ref));
+ memset(stereo, 0, sizeof(stereo));
+
+ for (i = 0; i < SAMPLES; i++) {
+ mono[i] = rand() - RAND_MAX/2;
+ }
+
+ sf = PA_SAMPLE_S16NE;
+ remap.format = &sf;
+ iss.format = PA_SAMPLE_S16NE;
+ iss.channels = 1;
+ oss.format = PA_SAMPLE_S16NE;
+ oss.channels = 2;
+ remap.i_ss = &iss;
+ remap.o_ss = &oss;
+ remap.map_table_f[0][0] = 1.0;
+ remap.map_table_f[1][0] = 1.0;
+ remap.map_table_i[0][0] = 0x10000;
+ remap.map_table_i[1][0] = 0x10000;
+ remap_init_func(&remap);
+ if (!remap.do_remap) {
+ pa_log_debug("no reference remapping function, abort test");
+ return;
+ }
+
+ remap.do_remap(&remap, stereo_ref, mono, SAMPLES);
+
+ remap_mono_to_stereo_neon_a9(&remap, stereo, mono, SAMPLES);
+ for (i = 0; i < 2*SAMPLES; i++) {
+ if (abs(stereo[i] - stereo_ref[i]) > 0) {
+ pa_log_debug("A9 %d: %d != %d (%d)", i, stereo[i], stereo_ref[i],
+ mono[i/2]);
+ break;
+ }
+ }
+
+ remap_mono_to_stereo_neon_a8(&remap, stereo, mono, SAMPLES);
+ for (i = 0; i < 2*SAMPLES; i++) {
+ if (abs(stereo[i] - stereo_ref[i]) > 0) {
+ pa_log_debug("A8 %d: %d != %d (%d)", i, stereo[i], stereo_ref[i],
+ mono[i/2]);
+ break;
+ }
+ }
+
+ start = pa_rtclock_now();
+ for (i = 0; i < TIMES; i++) {
+ remap_mono_to_stereo_neon_a8(&remap, stereo, mono, SAMPLES);
+ }
+ stop = pa_rtclock_now();
+ pa_log_info("NEON: %llu usec.", (long long unsigned int)(stop - start));
+
+ start = pa_rtclock_now();
+ for (i = 0; i < TIMES; i++) {
+ remap.do_remap(&remap, stereo_ref, mono, SAMPLES);
+ }
+ stop = pa_rtclock_now();
+ pa_log_info("ref: %llu usec.", (long long unsigned int)(stop - start));
+}
+
+static void run_test_float_stereo_to_mono(void) {
+ float stereo[2*SAMPLES];
+ float mono_ref[SAMPLES];
+ float mono[SAMPLES];
+ int i;
+ pa_usec_t start, stop;
+ pa_sample_format_t sf;
+ pa_remap_t remap;
+ pa_sample_spec iss, oss;
+
+ pa_init_remap_func_t remap_init_func = pa_get_init_remap_func();
+
+ pa_log_debug("checking NEON remap_stereo_to_mono(float)");
+
+ memset(mono_ref, 0, sizeof(mono_ref));
+ memset(mono, 0, sizeof(mono));
+
+ for (i = 0; i < 2*SAMPLES; i++) {
+ stereo[i] = rand()/(float) RAND_MAX - 0.5f;
+ }
+
+ sf = PA_SAMPLE_FLOAT32NE;
+ remap.format = &sf;
+ iss.format = PA_SAMPLE_FLOAT32NE;
+ iss.channels = 2;
+ oss.format = PA_SAMPLE_FLOAT32NE;
+ oss.channels = 1;
+ remap.i_ss = &iss;
+ remap.o_ss = &oss;
+ remap.map_table_f[0][0] = 1.0;
+ remap.map_table_f[0][1] = 1.0;
+ remap.map_table_i[0][0] = 0x10000;
+ remap.map_table_i[0][1] = 0x10000;
+ remap_init_func(&remap);
+ if (!remap.do_remap) {
+ pa_log_debug("no reference remapping function, abort test");
+ return;
+ }
+
+ remap.do_remap(&remap, mono_ref, stereo, SAMPLES);
+ remap_stereo_to_mono_neon(&remap, mono, stereo, SAMPLES);
+
+ for (i = 0; i < SAMPLES; i++) {
+ if (fabsf(mono[i] - mono_ref[i]) > 0.00001) {
+ pa_log_debug("%d: %.3f != %.3f (%.3f %.3f)", i, mono[i], mono_ref[i],
+ stereo[2*i+0], stereo[2*i+1]);
+ break;
+ }
+ }
+
+ start = pa_rtclock_now();
+ for (i = 0; i < TIMES; i++) {
+ remap_stereo_to_mono_neon(&remap, mono, stereo, SAMPLES);
+ }
+ stop = pa_rtclock_now();
+ pa_log_info("NEON: %llu usec.", (long long unsigned int)(stop - start));
+
+ start = pa_rtclock_now();
+ for (i = 0; i < TIMES; i++) {
+ remap.do_remap(&remap, mono_ref, stereo, SAMPLES);
+ }
+ stop = pa_rtclock_now();
+ pa_log_info("ref: %llu usec.", (long long unsigned int)(stop - start));
+}
+
+static void run_test_s16_stereo_to_mono(void) {
+ int16_t stereo[2*SAMPLES];
+ int16_t mono_ref[SAMPLES];
+ int16_t mono[SAMPLES];
+ int i;
+ pa_usec_t start, stop;
+ pa_sample_format_t sf;
+ pa_remap_t remap;
+ pa_sample_spec iss, oss;
+
+ pa_init_remap_func_t remap_init_func = pa_get_init_remap_func();
+
+ pa_log_debug("checking NEON remap_stereo_to_mono(s16)");
+
+ memset(mono_ref, 0, sizeof(mono_ref));
+ memset(mono, 0, sizeof(mono));
+
+ for (i = 0; i < 2*SAMPLES; i++) {
+ stereo[i] = rand() - RAND_MAX/2;
+ }
+
+ sf = PA_SAMPLE_S16NE;
+ remap.format = &sf;
+ iss.format = PA_SAMPLE_S16NE;
+ iss.channels = 2;
+ oss.format = PA_SAMPLE_S16NE;
+ oss.channels = 1;
+ remap.i_ss = &iss;
+ remap.o_ss = &oss;
+ remap.map_table_f[0][0] = 1.0;
+ remap.map_table_f[0][1] = 1.0;
+ remap.map_table_i[0][0] = 0x10000;
+ remap.map_table_i[0][1] = 0x10000;
+ remap_init_func(&remap);
+ if (!remap.do_remap) {
+ pa_log_debug("no reference remapping function, abort test");
+ return;
+ }
+
+ remap.do_remap(&remap, mono_ref, stereo, SAMPLES);
+ remap_stereo_to_mono_neon(&remap, mono, stereo, SAMPLES);
+
+ for (i = 0; i < SAMPLES; i++) {
+ if (abs(mono[i] - mono_ref[i]) > 0) {
+ pa_log_debug("%d: %d != %d (%d %d)", i, mono[i], mono_ref[i],
+ stereo[2*i+0], stereo[2*i+1]);
+ break;
+ }
+ }
+
+ start = pa_rtclock_now();
+ for (i = 0; i < TIMES; i++) {
+ remap_stereo_to_mono_neon(&remap, mono, stereo, SAMPLES);
+ }
+ stop = pa_rtclock_now();
+ pa_log_info("NEON: %llu usec.", (long long unsigned int)(stop - start));
+
+ start = pa_rtclock_now();
+ for (i = 0; i < TIMES; i++) {
+ remap.do_remap(&remap, mono_ref, stereo, SAMPLES);
+ }
+ stop = pa_rtclock_now();
+ pa_log_info("ref: %llu usec.", (long long unsigned int)(stop - start));
+}
+#endif /* RUN_TEST */
+
+static pa_cpu_arm_flag_t arm_flags;
+
+static void init_remap_neon(pa_remap_t *m) {
+ unsigned n_oc, n_ic;
+
+ n_oc = m->o_ss->channels;
+ n_ic = m->i_ss->channels;
+
+ /* find some common channel remappings, fall back to full matrix operation. */
+ if (n_ic == 1 && n_oc == 2 &&
+ m->map_table_f[0][0] >= 1.0 && m->map_table_f[1][0] >= 1.0) {
+ if (arm_flags & PA_CPU_ARM_CORTEX_A8) {
+ m->do_remap = (pa_do_remap_func_t) remap_mono_to_stereo_neon_a8;
+ pa_log_info("Using ARM NEON/A8 mono to stereo remapping");
+ }
+ else {
+ m->do_remap = (pa_do_remap_func_t) remap_mono_to_stereo_neon_a9;
+ pa_log_info("Using ARM NEON mono to stereo remapping");
+ }
+ }
+ else if (n_ic == 2 && n_oc == 1 &&
+ m->map_table_f[0][0] >= 1.0 && m->map_table_f[0][1] >= 1.0) {
+ m->do_remap = (pa_do_remap_func_t) remap_stereo_to_mono_neon;
+ pa_log_info("Using ARM NEON stereo to mono remapping");
+ }
+}
+#endif /* defined (__ARM_NEON__) */
+
+void pa_remap_func_init_neon(pa_cpu_arm_flag_t flags) {
+#if defined (__ARM_NEON__)
+
+#ifdef RUN_TEST
+ run_test_float_stereo_to_mono();
+ run_test_s16_stereo_to_mono();
+ run_test_float_mono_to_stereo();
+ run_test_s16_mono_to_stereo();
+#endif
+
+ pa_log_info("Initialising ARM NEON optimized remappers.");
+ arm_flags = flags;
+ pa_set_init_remap_func((pa_init_remap_func_t) init_remap_neon);
+
+#endif /* defined (__ARM_NEON__) */
+}
--
1.7.5.4
More information about the pulseaudio-discuss
mailing list