audiomixer: improve sse and sse2 mixer

Add mixer benchmark.
Improve sse and sse2 mixer function by removin some read/write to the
temporary buffer at the expense of more scattered reads.

spa/plugins/audiomixer/benchmark-mix-ops.c

@@ -0,0 +1,217 @@
+/* Spa
+ *
+ * Copyright © 2019 Wim Taymans
+ *
+ * Permission is hereby granted, free of charge, to any person obtaining a
+ * copy of this software and associated documentation files (the "Software"),
+ * to deal in the Software without restriction, including without limitation
+ * the rights to use, copy, modify, merge, publish, distribute, sublicense,
+ * and/or sell copies of the Software, and to permit persons to whom the
+ * Software is furnished to do so, subject to the following conditions:
+ *
+ * The above copyright notice and this permission notice (including the next
+ * paragraph) shall be included in all copies or substantial portions of the
+ * Software.
+ *
+ * THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+ * IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+ * FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT.  IN NO EVENT SHALL
+ * THE AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+ * LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING
+ * FROM, OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER
+ * DEALINGS IN THE SOFTWARE.
+ */
+
+#include "config.h"
+
+#include <string.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <unistd.h>
+#include <errno.h>
+#include <time.h>
+
+#include "test-helper.h"
+#include "mix-ops.h"
+
+static uint32_t cpu_flags;
+
+typedef void (*mix_func_t) (struct mix_ops *ops, void * SPA_RESTRICT dst,
+		const void * SPA_RESTRICT src[], uint32_t n_src, uint32_t n_samples);
+struct stats {
+	uint32_t n_samples;
+	uint32_t n_src;
+	uint64_t perf;
+	const char *name;
+	const char *impl;
+};
+
+#define MAX_SAMPLES	4096
+#define MAX_SRC		11
+
+#define MAX_COUNT 100
+
+static uint8_t samp_in[MAX_SAMPLES * MAX_SRC * 8];
+static uint8_t samp_out[MAX_SAMPLES * 8];
+
+static const int sample_sizes[] = { 0, 1, 128, 513, 4096 };
+static const int src_counts[] = { 1, 2, 4, 6, 8, 11 };
+
+#define MAX_RESULTS	SPA_N_ELEMENTS(sample_sizes) * SPA_N_ELEMENTS(src_counts) * 70
+
+static uint32_t n_results = 0;
+static struct stats results[MAX_RESULTS];
+
+static void run_test1(const char *name, const char *impl, mix_func_t func, int n_src, int n_samples)
+{
+	int i, j;
+	const void *ip[n_src];
+	void *op;
+	struct timespec ts;
+	uint64_t count, t1, t2;
+	struct mix_ops mix;
+
+	mix.n_channels = 1;
+
+	for (j = 0; j < n_src; j++)
+		ip[j] = SPA_PTR_ALIGN(&samp_in[j * n_samples * 4], 16, void);
+	op = SPA_PTR_ALIGN(samp_out, 16, void);
+
+	clock_gettime(CLOCK_MONOTONIC, &ts);
+	t1 = SPA_TIMESPEC_TO_NSEC(&ts);
+
+	count = 0;
+	for (i = 0; i < MAX_COUNT; i++) {
+		func(&mix, op, ip, n_src, n_samples);
+		count++;
+	}
+	clock_gettime(CLOCK_MONOTONIC, &ts);
+	t2 = SPA_TIMESPEC_TO_NSEC(&ts);
+
+	spa_assert(n_results < MAX_RESULTS);
+
+	results[n_results++] = (struct stats) {
+		.n_samples = n_samples,
+		.n_src = n_src,
+		.perf = count * (uint64_t)SPA_NSEC_PER_SEC / (t2 - t1),
+		.name = name,
+		.impl = impl
+	};
+}
+
+static void run_test(const char *name, const char *impl, mix_func_t func)
+{
+	size_t i, j;
+
+	for (i = 0; i < SPA_N_ELEMENTS(sample_sizes); i++) {
+		for (j = 0; j < SPA_N_ELEMENTS(src_counts); j++) {
+			run_test1(name, impl, func, src_counts[j],
+				(sample_sizes[i] + (src_counts[j] -1)) / src_counts[j]);
+		}
+	}
+}
+
+static void test_s8(void)
+{
+	run_test("test_s8", "c", mix_s8_c);
+}
+static void test_u8(void)
+{
+	run_test("test_u8", "c", mix_u8_c);
+}
+
+static void test_s16(void)
+{
+	run_test("test_s16", "c", mix_s16_c);
+}
+static void test_u16(void)
+{
+	run_test("test_u8", "c", mix_u16_c);
+}
+
+static void test_s24(void)
+{
+	run_test("test_s24", "c", mix_s24_c);
+}
+static void test_u24(void)
+{
+	run_test("test_u24", "c", mix_u24_c);
+}
+static void test_s24_32(void)
+{
+	run_test("test_s24_32", "c", mix_s24_32_c);
+}
+static void test_u24_32(void)
+{
+	run_test("test_u24_32", "c", mix_u24_32_c);
+}
+
+static void test_s32(void)
+{
+	run_test("test_s32", "c", mix_s32_c);
+}
+static void test_u32(void)
+{
+	run_test("test_u32", "c", mix_u32_c);
+}
+
+static void test_f32(void)
+{
+	run_test("test_f32", "c", mix_f32_c);
+#if defined (HAVE_SSE)
+	if (cpu_flags & SPA_CPU_FLAG_SSE) {
+		run_test("test_f32", "sse", mix_f32_sse);
+	}
+#endif
+}
+
+static void test_f64(void)
+{
+	run_test("test_f64", "c", mix_f64_c);
+#if defined (HAVE_SSE2)
+	if (cpu_flags & SPA_CPU_FLAG_SSE2) {
+		run_test("test_f64", "sse2", mix_f64_sse2);
+	}
+#endif
+}
+
+static int compare_func(const void *_a, const void *_b)
+{
+	const struct stats *a = _a, *b = _b;
+	int diff;
+	if ((diff = strcmp(a->name, b->name)) != 0) return diff;
+	if ((diff = a->n_samples - b->n_samples) != 0) return diff;
+	if ((diff = a->n_src - b->n_src) != 0) return diff;
+	if ((diff = b->perf - a->perf) != 0) return diff;
+	return 0;
+}
+
+int main(int argc, char *argv[])
+{
+	uint32_t i;
+
+	cpu_flags = get_cpu_flags();
+	printf("got get CPU flags %d\n", cpu_flags);
+
+	test_s8();
+	test_u8();
+	test_s16();
+	test_u16();
+	test_s24();
+	test_u24();
+	test_s32();
+	test_u32();
+	test_s24_32();
+	test_u24_32();
+	test_f32();
+	test_f64();
+
+	qsort(results, n_results, sizeof(struct stats), compare_func);
+
+	for (i = 0; i < n_results; i++) {
+		struct stats *s = &results[i];
+		fprintf(stderr, "%-12."PRIu64" \t%-32.32s %s \t samples %d, src %d\n",
+				s->perf, s->name, s->impl, s->n_samples, s->n_src);
+	}
+	return 0;
+}

spa/plugins/audiomixer/meson.build

@@ -95,3 +95,32 @@ foreach a : test_apps
         )
   endif
 endforeach
+
+benchmark_apps = [
+  'benchmark-mix-ops',
+  ]
+
+foreach a : benchmark_apps
+  benchmark(a,
+    executable(a, a + '.c',
+      dependencies : [ spa_dep, dl_lib, pthread_lib, mathlib, audiomixer_dep ],
+      include_directories : [ configinc ],
+      c_args : [ simd_cargs ],
+      install_rpath : spa_plugindir / 'audiomixer',
+      install : installed_tests_enabled,
+      install_dir : installed_tests_execdir / 'audiomixer'),
+      env : [
+        'SPA_PLUGIN_DIR=@0@'.format(spa_dep.get_variable('plugindir')),
+        ])
+
+    if installed_tests_enabled
+      test_conf = configuration_data()
+      test_conf.set('exec', installed_tests_execdir / 'audiomixer' / a)
+      configure_file(
+        input: installed_tests_template,
+        output: a + '.test',
+        install_dir: installed_tests_metadir / 'audiomixer',
+        configuration: test_conf
+        )
+  endif
+endforeach

spa/plugins/audiomixer/mix-ops-c.c

@@ -39,7 +39,7 @@ void mix_ ##name## _c(struct mix_ops *ops,					\
 	type *d = dst;								\
 	const type **s = (const type **)src;					\
 	n_samples *= ops->n_channels;						\
-	if (n_src == 0  && zero)						\
+	if (n_src == 0 && zero)							\
 		memset(dst, 0, n_samples * sizeof(type));			\
 	else if (n_src == 1) {							\
 		if (dst != src[0])						\

spa/plugins/audiomixer/mix-ops-sse.c

@@ -32,58 +32,57 @@
 
 #include <xmmintrin.h>
 
-static inline void mix_2(float * dst, const float * SPA_RESTRICT src, uint32_t n_samples)
-{
-	uint32_t n, unrolled;
-	__m128 in1[4], in2[4];
-
-	if (SPA_LIKELY(SPA_IS_ALIGNED(src, 16) &&
-	    SPA_IS_ALIGNED(dst, 16)))
-		unrolled = n_samples & ~15;
-	else
-		unrolled = 0;
-
-	for (n = 0; n < unrolled; n += 16) {
-		in1[0] = _mm_load_ps(&dst[n+ 0]);
-		in1[1] = _mm_load_ps(&dst[n+ 4]);
-		in1[2] = _mm_load_ps(&dst[n+ 8]);
-		in1[3] = _mm_load_ps(&dst[n+12]);
-
-		in2[0] = _mm_load_ps(&src[n+ 0]);
-		in2[1] = _mm_load_ps(&src[n+ 4]);
-		in2[2] = _mm_load_ps(&src[n+ 8]);
-		in2[3] = _mm_load_ps(&src[n+12]);
-
-		in1[0] = _mm_add_ps(in1[0], in2[0]);
-		in1[1] = _mm_add_ps(in1[1], in2[1]);
-		in1[2] = _mm_add_ps(in1[2], in2[2]);
-		in1[3] = _mm_add_ps(in1[3], in2[3]);
-
-		_mm_store_ps(&dst[n+ 0], in1[0]);
-		_mm_store_ps(&dst[n+ 4], in1[1]);
-		_mm_store_ps(&dst[n+ 8], in1[2]);
-		_mm_store_ps(&dst[n+12], in1[3]);
-	}
-	for (; n < n_samples; n++) {
-		in1[0] = _mm_load_ss(&dst[n]),
-		in2[0] = _mm_load_ss(&src[n]),
-		in1[0] = _mm_add_ss(in1[0], in2[0]);
-		_mm_store_ss(&dst[n], in1[0]);
-	}
-}
-
 void
 mix_f32_sse(struct mix_ops *ops, void * SPA_RESTRICT dst, const void * SPA_RESTRICT src[],
 		uint32_t n_src, uint32_t n_samples)
 {
-	uint32_t i;
+	n_samples *= ops->n_channels;
 
-	if (n_src == 0)
+	if (n_src == 0) {
-		memset(dst, 0, n_samples * ops->n_channels * sizeof(float));
+		memset(dst, 0, n_samples * sizeof(float));
-	else if (dst != src[0])
+	} else if (n_src == 1) {
-		spa_memcpy(dst, src[0], n_samples * ops->n_channels * sizeof(float));
+		if (dst != src[0])
+			spa_memcpy(dst, src[0], n_samples * sizeof(float));
+	} else {
+		uint32_t n, i, unrolled;
+		__m128 in[4];
+		const float **s = (const float **)src;
+		float *d = dst;
+		bool aligned = true;
 
-	for (i = 1; i < n_src; i++) {
+		if (SPA_UNLIKELY(!SPA_IS_ALIGNED(dst, 16)))
-		mix_2(dst, src[i], n_samples * ops->n_channels);
+			aligned = false;
+		else {
+			for (i = 0; i < n_src && aligned; i++) {
+				if (SPA_UNLIKELY(!SPA_IS_ALIGNED(src[i], 16)))
+					aligned = false;
+			}
+		}
+
+		unrolled = aligned ? n_samples & ~15 : 0;
+
+		for (n = 0; n < unrolled; n += 16) {
+			in[0] = _mm_load_ps(&s[0][n+ 0]);
+			in[1] = _mm_load_ps(&s[0][n+ 4]);
+			in[2] = _mm_load_ps(&s[0][n+ 8]);
+			in[3] = _mm_load_ps(&s[0][n+12]);
+
+			for (i = 1; i < n_src; i++) {
+				in[0] = _mm_add_ps(in[0], _mm_load_ps(&s[i][n+ 0]));
+				in[1] = _mm_add_ps(in[1], _mm_load_ps(&s[i][n+ 4]));
+				in[2] = _mm_add_ps(in[2], _mm_load_ps(&s[i][n+ 8]));
+				in[3] = _mm_add_ps(in[3], _mm_load_ps(&s[i][n+12]));
+			}
+			_mm_store_ps(&d[n+ 0], in[0]);
+			_mm_store_ps(&d[n+ 4], in[1]);
+			_mm_store_ps(&d[n+ 8], in[2]);
+			_mm_store_ps(&d[n+12], in[3]);
+		}
+		for (; n < n_samples; n++) {
+			in[0] = _mm_load_ss(&s[0][n]);
+			for (i = 1; i < n_src; i++)
+				in[0] = _mm_add_ss(in[0], _mm_load_ss(&s[i][n]));
+			_mm_store_ss(&d[n], in[0]);
+		}
 	}
 }

spa/plugins/audiomixer/mix-ops-sse2.c

@@ -32,58 +32,57 @@
 
 #include <emmintrin.h>
 
-static inline void mix_2(double * dst, const double * SPA_RESTRICT src, uint32_t n_samples)
-{
-	uint32_t n, unrolled;
-	__m128d in1[4], in2[4];
-
-	if (SPA_IS_ALIGNED(src, 16) &&
-	    SPA_IS_ALIGNED(dst, 16))
-		unrolled = n_samples & ~7;
-	else
-		unrolled = 0;
-
-	for (n = 0; n < unrolled; n += 8) {
-		in1[0] = _mm_load_pd(&dst[n+ 0]);
-		in1[1] = _mm_load_pd(&dst[n+ 2]);
-		in1[2] = _mm_load_pd(&dst[n+ 4]);
-		in1[3] = _mm_load_pd(&dst[n+ 6]);
-
-		in2[0] = _mm_load_pd(&src[n+ 0]);
-		in2[1] = _mm_load_pd(&src[n+ 2]);
-		in2[2] = _mm_load_pd(&src[n+ 4]);
-		in2[3] = _mm_load_pd(&src[n+ 6]);
-
-		in1[0] = _mm_add_pd(in1[0], in2[0]);
-		in1[1] = _mm_add_pd(in1[1], in2[1]);
-		in1[2] = _mm_add_pd(in1[2], in2[2]);
-		in1[3] = _mm_add_pd(in1[3], in2[3]);
-
-		_mm_store_pd(&dst[n+ 0], in1[0]);
-		_mm_store_pd(&dst[n+ 2], in1[1]);
-		_mm_store_pd(&dst[n+ 4], in1[2]);
-		_mm_store_pd(&dst[n+ 6], in1[3]);
-	}
-	for (; n < n_samples; n++) {
-		in1[0] = _mm_load_sd(&dst[n]),
-		in2[0] = _mm_load_sd(&src[n]),
-		in1[0] = _mm_add_sd(in1[0], in2[0]);
-		_mm_store_sd(&dst[n], in1[0]);
-	}
-}
-
 void
 mix_f64_sse2(struct mix_ops *ops, void * SPA_RESTRICT dst, const void * SPA_RESTRICT src[],
 		uint32_t n_src, uint32_t n_samples)
 {
-	uint32_t i;
+	n_samples *= ops->n_channels;
 
-	if (n_src == 0)
+	if (n_src == 0) {
-		memset(dst, 0, n_samples * ops->n_channels * sizeof(double));
+		memset(dst, 0, n_samples * sizeof(double));
-	else if (dst != src[0])
+	} else if (n_src == 1) {
-		spa_memcpy(dst, src[0], n_samples * ops->n_channels * sizeof(double));
+		if (dst != src[0])
+			spa_memcpy(dst, src[0], n_samples * sizeof(double));
+	} else {
+		uint32_t n, i, unrolled;
+		__m128d in[4];
+		const double **s = (const double **)src;
+		double *d = dst;
+		bool aligned = true;
 
-	for (i = 1; i < n_src; i++) {
+		if (SPA_UNLIKELY(!SPA_IS_ALIGNED(dst, 16)))
-		mix_2(dst, src[i], n_samples * ops->n_channels);
+			aligned = false;
+		else {
+			for (i = 0; i < n_src && aligned; i++) {
+				if (SPA_UNLIKELY(!SPA_IS_ALIGNED(src[i], 16)))
+					aligned = false;
+			}
+		}
+
+		unrolled = aligned ? n_samples & ~7 : 0;
+
+		for (n = 0; n < unrolled; n += 8) {
+			in[0] = _mm_load_pd(&s[0][n+0]);
+			in[1] = _mm_load_pd(&s[0][n+2]);
+			in[2] = _mm_load_pd(&s[0][n+4]);
+			in[3] = _mm_load_pd(&s[0][n+6]);
+
+			for (i = 1; i < n_src; i++) {
+				in[0] = _mm_add_pd(in[0], _mm_load_pd(&s[i][n+0]));
+				in[1] = _mm_add_pd(in[1], _mm_load_pd(&s[i][n+2]));
+				in[2] = _mm_add_pd(in[2], _mm_load_pd(&s[i][n+4]));
+				in[3] = _mm_add_pd(in[3], _mm_load_pd(&s[i][n+6]));
+			}
+			_mm_store_pd(&d[n+0], in[0]);
+			_mm_store_pd(&d[n+2], in[1]);
+			_mm_store_pd(&d[n+4], in[2]);
+			_mm_store_pd(&d[n+6], in[3]);
+		}
+		for (; n < n_samples; n++) {
+			in[0] = _mm_load_sd(&s[0][n]);
+			for (i = 1; i < n_src; i++)
+				in[0] = _mm_add_sd(in[0], _mm_load_sd(&s[i][n]));
+			_mm_store_sd(&d[n], in[0]);
+		}
 	}
 }