core: improve alignment

Use the CPU alignment to align memory of allocated buffers. This makes
it more likely that all CPU optimizations can be done.

spa/plugins/audioconvert/audioadapter.c

@@ -23,6 +23,7 @@
  */
 
 #include <spa/support/log.h>
+#include <spa/support/cpu.h>
 
 #include <spa/node/node.h>
 #include <spa/node/io.h>
@@ -44,7 +45,9 @@ struct impl {
 	struct spa_node node;
 
 	struct spa_log *log;
+	struct spa_cpu *cpu;
 
+	uint32_t max_align;
 	enum spa_direction direction;
 
 	struct spa_node *target;
@@ -642,9 +645,9 @@ static int negotiate_buffers(struct impl *this)
 	struct spa_pod_builder b = SPA_POD_BUILDER_INIT(buffer, sizeof(buffer));
 	uint32_t state;
 	struct spa_pod *param;
-	int res, i;
+	int res;
 	bool slave_alloc, conv_alloc;
-	int32_t size, buffers, blocks, align, flags;
+	uint32_t i, size, buffers, blocks, align, flags;
 	uint32_t *aligns;
 	struct spa_data *datas;
 	uint32_t slave_flags, conv_flags;
@@ -702,6 +705,8 @@ static int negotiate_buffers(struct impl *this)
 	spa_log_debug(this->log, "%p: buffers %d, blocks %d, size %d, align %d %d:%d",
 			this, buffers, blocks, size, align, slave_alloc, conv_alloc);
 
+	align = SPA_MAX(align, this->max_align);
+
 	datas = alloca(sizeof(struct spa_data) * blocks);
 	memset(datas, 0, sizeof(struct spa_data) * blocks);
 	aligns = alloca(sizeof(uint32_t) * blocks);
@@ -962,6 +967,9 @@ impl_init(const struct spa_handle_factory *factory,
 		case SPA_TYPE_INTERFACE_Log:
 			this->log = support[i].data;
 			break;
+		case SPA_TYPE_INTERFACE_CPU:
+			this->cpu = support[i].data;
+			break;
 		}
 	}
 	if (info == NULL || (str = spa_dict_lookup(info, "audio.adapt.slave")) == NULL)
@@ -971,6 +979,9 @@ impl_init(const struct spa_handle_factory *factory,
 	if (this->slave == NULL)
 		return -EINVAL;
 
+	if (this->cpu)
+		this->max_align = spa_cpu_get_max_align(this->cpu);
+
 	spa_hook_list_init(&this->hooks);
 
 	this->node.iface = SPA_INTERFACE_INIT(

spa/plugins/audioconvert/audioconvert.c

@@ -27,6 +27,7 @@
 #include <stdio.h>
 
 #include <spa/support/log.h>
+#include <spa/support/cpu.h>
 #include <spa/utils/list.h>
 #include <spa/utils/names.h>
 #include <spa/node/node.h>
@@ -68,6 +69,9 @@ struct impl {
 	struct spa_node node;
 
 	struct spa_log *log;
+	struct spa_cpu *cpu;
+
+	uint32_t max_align;
 
 	struct spa_hook_list hooks;
 
@@ -345,6 +349,8 @@ static int negotiate_link_buffers(struct impl *this, struct link *link)
 	spa_log_debug(this->log, "%p: buffers %d, blocks %d, size %d, align %d %d:%d",
 			this, buffers, blocks, size, align, out_alloc, in_alloc);
 
+	align = SPA_MAX(align, this->max_align);
+
 	datas = alloca(sizeof(struct spa_data) * blocks);
 	memset(datas, 0, sizeof(struct spa_data) * blocks);
 	aligns = alloca(sizeof(uint32_t) * blocks);
@@ -1191,9 +1197,19 @@ impl_init(const struct spa_handle_factory *factory,
 	this = (struct impl *) handle;
 
 	for (i = 0; i < n_support; i++) {
-		if (support[i].type == SPA_TYPE_INTERFACE_Log)
+		switch (support[i].type) {
+		case SPA_TYPE_INTERFACE_Log:
 			this->log = support[i].data;
+			break;
+		case SPA_TYPE_INTERFACE_CPU:
+			this->cpu = support[i].data;
+			break;
 		}
+	}
+
+	if (this->cpu)
+		this->max_align = spa_cpu_get_max_align(this->cpu);
+
 	this->node.iface = SPA_INTERFACE_INIT(
 			SPA_TYPE_INTERFACE_Node,
 			SPA_VERSION_NODE,

spa/plugins/audioconvert/merger.c

@@ -73,7 +73,6 @@ struct port {
 	struct spa_param_info params[8];
 	char position[16];
 
-	bool have_format;
 	struct spa_audio_info format;
 	uint32_t blocks;
 	uint32_t stride;
@@ -82,6 +81,8 @@ struct port {
 	uint32_t n_buffers;
 
 	struct spa_list queue;
+
+	unsigned int have_format:1;
 };
 
 struct impl {

spa/plugins/audiomixer/mixer-dsp.c

@@ -44,6 +44,7 @@
 #define MAX_BUFFERS	64
 #define MAX_PORTS	128
 #define MAX_SAMPLES	1024
+#define MAX_ALIGN	64
 
 #define PORT_DEFAULT_VOLUME	1.0
 #define PORT_DEFAULT_MUTE	false
@@ -115,13 +116,14 @@ struct impl {
 	struct port in_ports[MAX_PORTS];
 	struct port out_ports[1];
 
-	bool have_format;
 	int n_formats;
 	struct spa_audio_info format;
 	uint32_t stride;
 
-	bool started;
+	unsigned int have_format:1;
-        float empty[MAX_SAMPLES + 15];
+	unsigned int started:1;
+
+	float empty[MAX_SAMPLES + MAX_ALIGN];
 };
 
 #define CHECK_FREE_IN_PORT(this,d,p) ((d) == SPA_DIRECTION_INPUT && (p) < MAX_PORTS && !this->in_ports[(p)].valid)
@@ -725,7 +727,7 @@ static int impl_node_process(void *object)
 	else {
 		outb->buffer->n_datas = 1;
 		outb->buffer->datas = outb->datas;
-		outb->datas[0].data = SPA_PTR_ALIGN(this->empty, 16, void);
+		outb->datas[0].data = SPA_PTR_ALIGN(this->empty, MAX_ALIGN, void);
 		outb->datas[0].chunk = outb->chunk;
 		outb->datas[0].chunk->offset = 0;
 		outb->datas[0].chunk->size = n_samples * sizeof(float);

src/pipewire/buffers.c

@@ -37,6 +37,7 @@
 
 #define NAME "buffers"
 
+#define MAX_ALIGN	32
 #define MAX_BUFFERS	64
 
 struct port {
@@ -246,8 +247,12 @@ int pw_buffers_negotiate(struct pw_core *core, uint32_t flags,
 	else
 		max_buffers = MAX_BUFFERS;
 
+	if ((str = pw_properties_get(core->properties, "cpu.max-align")) != NULL)
+		align = pw_properties_parse_int(str);
+	else
+		align = MAX_ALIGN;
+
 	minsize = stride = 0;
-	align = 8;
 	param = find_param(params, n_params, SPA_TYPE_OBJECT_ParamBuffers);
 	if (param) {
 		uint32_t qmax_buffers = max_buffers,

src/pipewire/core.c

@@ -29,6 +29,7 @@
 
 #include <pipewire/log.h>
 
+#include <spa/support/cpu.h>
 #include <spa/support/dbus.h>
 #include <spa/node/utils.h>
 #include <spa/utils/names.h>
@@ -479,6 +480,7 @@ struct pw_core *pw_core_new(struct pw_loop *main_loop,
 	void *dbus_iface = NULL;
 	uint32_t n_support;
 	struct pw_properties *pr;
+	struct spa_cpu *cpu;
 	int res = 0;
 
 	impl = calloc(1, sizeof(struct impl) + user_data_size);
@@ -526,6 +528,9 @@ struct pw_core *pw_core_new(struct pw_loop *main_loop,
 	this->support[n_support++] = SPA_SUPPORT_INIT(SPA_TYPE_INTERFACE_DataSystem, this->data_system);
 	this->support[n_support++] = SPA_SUPPORT_INIT(SPA_TYPE_INTERFACE_DataLoop, this->data_loop->loop);
 
+	if ((cpu = spa_support_find(this->support, n_support, SPA_TYPE_INTERFACE_CPU)) != NULL)
+		pw_properties_setf(properties, "cpu.max-align", "%u", spa_cpu_get_max_align(cpu));
+
 	lib = pw_properties_get(properties, PW_KEY_LIBRARY_NAME_DBUS);
 	if (lib == NULL)
 		lib = "support/libspa-dbus";