*/
#include <stdio.h>
#include <cmath>
+#include <xmmintrin.h>
#include "pbd/compose.h"
-#include "pbd/cpus.h"
+#include "pbd/debug_rt_alloc.h"
#include "ardour/debug.h"
#include "ardour/graph.h"
using namespace PBD;
using namespace std;
+#ifdef DEBUG_RT_ALLOC
+static Graph* graph = 0;
-Graph::Graph (Session & session)
- : SessionHandleRef (session)
+extern "C" {
+
+int alloc_allowed ()
+{
+ return !graph->in_process_thread ();
+}
+
+}
+#endif
+
+Graph::Graph (Session & session)
+ : SessionHandleRef (session)
+ , _quit_threads (false)
+ , _execution_sem ("graph_execution", 0)
+ , _callback_start_sem ("graph_start", 0)
+ , _callback_done_sem ("graph_done", 0)
+ , _cleanup_sem ("graph_cleanup", 0)
{
pthread_mutex_init( &_trigger_mutex, NULL);
- sem_init( &_execution_sem, 0, 0 );
- sem_init( &_callback_start_sem, 0, 0 );
- sem_init( &_callback_done_sem, 0, 0 );
- sem_init( &_cleanup_sem, 0, 0 );
+ /* XXX: rather hacky `fix' to stop _trigger_queue.push_back() allocating
+ memory in the RT thread.
+ */
+ _trigger_queue.reserve (8192);
_execution_tokens = 0;
_quit_threads = false;
_graph_empty = true;
- int num_cpu = hardware_concurrency();
- int num_threads = num_cpu;
- int pu = Config->get_processor_usage ();
+ reset_thread_list ();
- if (pu < 0) {
- /* pu is negative: use "pu" less cores for DSP than appear to be available
- */
+#ifdef DEBUG_RT_ALLOC
+ graph = this;
+ pbd_alloc_allowed = &::alloc_allowed;
+#endif
+}
- if (-pu < num_threads) {
- num_threads += pu;
- } else {
- num_threads = 1;
- }
- } else {
- /* use "pu" cores, if available
- */
+/** Set up threads for running the graph */
+void
+Graph::reset_thread_list ()
+{
+ uint32_t num_threads = how_many_dsp_threads ();
- if (pu <= num_threads) {
- num_threads = pu;
- }
- }
+ /* For now, we shouldn't be using the graph code if we only have 1 DSP thread */
+ assert (num_threads > 1);
- info << string_compose (_("Using %2 threads on %1 CPUs"), num_cpu, num_threads) << endmsg;
+ /* don't bother doing anything here if we already have the right
+ number of threads.
+ */
+ if (_thread_list.size() == num_threads) {
+ return;
+ }
+
+ Glib::Mutex::Lock lm (_session.engine().process_lock());
pthread_t a_thread;
- if (AudioEngine::instance()->create_process_thread (boost::bind (&Graph::main_thread, this), &a_thread, 100000) == 0) {
- _thread_list.push_back (a_thread);
+ if (!_thread_list.empty()) {
+ drop_threads ();
+ }
+
+ if (AudioEngine::instance()->create_process_thread (boost::bind (&Graph::main_thread, this), &a_thread, 100000) != 0) {
+ throw failed_constructor ();
}
- for (int i = 1; i < num_threads; ++i) {
- if (AudioEngine::instance()->create_process_thread (boost::bind (&Graph::helper_thread, this), &a_thread, 100000) == 0) {
- _thread_list.push_back (a_thread);
+ _thread_list.push_back (a_thread);
+
+ for (uint32_t i = 1; i < num_threads; ++i) {
+ if (AudioEngine::instance()->create_process_thread (boost::bind (&Graph::helper_thread, this), &a_thread, 100000) != 0) {
+ throw failed_constructor ();
}
+
+ _thread_list.push_back (a_thread);
}
}
void
Graph::session_going_away()
+{
+ drop_threads ();
+
+ // now drop all references on the nodes.
+ _nodes_rt[0].clear();
+ _nodes_rt[1].clear();
+ _init_trigger_list[0].clear();
+ _init_trigger_list[1].clear();
+ _trigger_queue.clear();
+}
+
+void
+Graph::drop_threads ()
{
_quit_threads = true;
- for (unsigned int i=0; i<_thread_list.size(); i++) {
- sem_post( &_execution_sem);
+ for (unsigned int i=0; i< _thread_list.size(); i++) {
+ _execution_sem.signal ();
}
- sem_post( &_callback_start_sem);
+ _callback_start_sem.signal ();
- for (list<pthread_t>::iterator i = _thread_list.begin(); i != _thread_list.end(); i++) {
+ for (list<pthread_t>::iterator i = _thread_list.begin(); i != _thread_list.end(); ++i) {
void* status;
pthread_join (*i, &status);
}
- // now drop all references on the nodes.
- _nodes_rt[0].clear();
- _nodes_rt[1].clear();
- _init_trigger_list[0].clear();
- _init_trigger_list[1].clear();
- _trigger_queue.clear();
+ _thread_list.clear ();
+
+ _execution_tokens = 0;
+
+ _quit_threads = false;
}
void
/* setup chain == pending chain - we have
to wait till this is no longer true.
*/
- _cleanup_cond.wait (_swap_mutex);
+ _cleanup_cond.wait (_swap_mutex);
}
}
}
_finished_refcount = _init_finished_refcount[chain];
+ /* Trigger the initial nodes for processing, which are the ones at the `input' end */
+ pthread_mutex_lock (&_trigger_mutex);
for (i=_init_trigger_list[chain].begin(); i!=_init_trigger_list[chain].end(); i++) {
- this->trigger( i->get() );
+ /* don't use ::trigger here, as we have already locked the mutex */
+ _trigger_queue.push_back (i->get ());
}
+ pthread_mutex_unlock (&_trigger_mutex);
}
void
Graph::trigger (GraphNode* n)
{
- pthread_mutex_lock (&_trigger_mutex);
- _trigger_queue.push_back( n);
- pthread_mutex_unlock (&_trigger_mutex);
+ pthread_mutex_lock (&_trigger_mutex);
+ _trigger_queue.push_back (n);
+ pthread_mutex_unlock (&_trigger_mutex);
}
+/** Called when a node at the `output' end of the chain (ie one that has no-one to feed)
+ * is finished.
+ */
void
Graph::dec_ref()
{
if (g_atomic_int_dec_and_test (&_finished_refcount)) {
- // ok... this cycle is finished now.
- // we are the only thread alive.
-
- this->restart_cycle();
+ /* We have run all the nodes that are at the `output' end of
+ the graph, so there is nothing more to do this time around.
+ */
+
+ restart_cycle ();
}
}
void
Graph::restart_cycle()
{
- //printf( "cycle_done chain: %d\n", _current_chain);
-
// we are through. wakeup our caller.
+
again:
- sem_post( &_callback_done_sem);
+ _callback_done_sem.signal ();
+
+ /* Block until the a process callback triggers us */
+ _callback_start_sem.wait();
- // block until we are triggered.
- sem_wait( &_callback_start_sem);
- if (_quit_threads)
+ if (_quit_threads) {
return;
+ }
- //printf( "cycle_start\n" );
+ prep ();
- this->prep();
- if (_graph_empty)
+ if (_graph_empty) {
goto again;
- //printf( "cycle_start chain: %d\n", _current_chain);
+ }
// returning will restart the cycle.
- // starting with waking up the others.
-}
-
-static bool
-is_feedback (boost::shared_ptr<RouteList> routelist, Route* from, boost::shared_ptr<Route> to)
-{
- for (RouteList::iterator ri=routelist->begin(); ri!=routelist->end(); ri++) {
- if ((*ri).get() == from)
- return false;
- if ((*ri) == to)
- return true;
- }
- assert(0);
- return false;
+ // starting with waking up the others.
}
-static bool
-is_feedback (boost::shared_ptr<RouteList> routelist, boost::shared_ptr<Route> from, Route* to)
-{
- for (RouteList::iterator ri=routelist->begin(); ri!=routelist->end(); ri++) {
- if ((*ri).get() == to)
- return true;
- if ((*ri) == from)
- return false;
- }
- assert(0);
- return false;
-}
+/** Rechain our stuff using a list of routes (which can be in any order) and
+ * a directed graph of their interconnections, which is guaranteed to be
+ * acyclic.
+ */
void
-Graph::rechain (boost::shared_ptr<RouteList> routelist)
+Graph::rechain (boost::shared_ptr<RouteList> routelist, GraphEdges const & edges)
{
- node_list_t::iterator ni;
Glib::Mutex::Lock ls (_swap_mutex);
int chain = _setup_chain;
DEBUG_TRACE (DEBUG::Graph, string_compose ("============== setup %1\n", chain));
- // set all refcounts to 0;
+ /* This will become the number of nodes that do not feed any other node;
+ once we have processed this number of those nodes, we have finished.
+ */
_init_finished_refcount[chain] = 0;
+
+ /* This will become a list of nodes that are not fed by another node, ie
+ those at the `input' end.
+ */
_init_trigger_list[chain].clear();
_nodes_rt[chain].clear();
+ /* Clear things out, and make _nodes_rt[chain] a copy of routelist */
for (RouteList::iterator ri=routelist->begin(); ri!=routelist->end(); ri++) {
- node_ptr_t n = boost::dynamic_pointer_cast<GraphNode> (*ri);
-
- n->_init_refcount[chain] = 0;
- n->_activation_set[chain].clear();
- _nodes_rt[chain].push_back(n);
+ (*ri)->_init_refcount[chain] = 0;
+ (*ri)->_activation_set[chain].clear();
+ _nodes_rt[chain].push_back (*ri);
}
// now add refs for the connections.
- for (ni=_nodes_rt[chain].begin(); ni!=_nodes_rt[chain].end(); ni++) {
- bool has_input = false;
- bool has_output = false;
+ for (node_list_t::iterator ni = _nodes_rt[chain].begin(); ni != _nodes_rt[chain].end(); ni++) {
- boost::shared_ptr<Route> rp = boost::dynamic_pointer_cast<Route>( *ni);
+ boost::shared_ptr<Route> r = boost::dynamic_pointer_cast<Route> (*ni);
- for (RouteList::iterator ri=routelist->begin(); ri!=routelist->end(); ri++) {
- if (rp->direct_feeds (*ri)) {
- if (is_feedback (routelist, rp.get(), *ri)) {
- continue;
- }
+ /* The routes that are directly fed by r */
+ set<GraphVertex> fed_from_r = edges.from (r);
- has_output = true;
- (*ni)->_activation_set[chain].insert (boost::dynamic_pointer_cast<GraphNode> (*ri) );
- }
- }
+ /* Hence whether r has an output */
+ bool const has_output = !fed_from_r.empty ();
- for (Route::FedBy::iterator fi=rp->fed_by().begin(); fi!=rp->fed_by().end(); fi++) {
- if (boost::shared_ptr<Route> r = fi->r.lock()) {
- if (!is_feedback (routelist, r, rp.get())) {
- has_input = true;
- }
- }
- }
+ /* Set up r's activation set */
+ for (set<GraphVertex>::iterator i = fed_from_r.begin(); i != fed_from_r.end(); ++i) {
+ r->_activation_set[chain].insert (*i);
+ }
- for (node_set_t::iterator ai=(*ni)->_activation_set[chain].begin(); ai!=(*ni)->_activation_set[chain].end(); ai++) {
+ /* r has an input if there are some incoming edges to r in the graph */
+ bool const has_input = !edges.has_none_to (r);
+
+ /* Increment the refcount of any route that we directly feed */
+ for (node_set_t::iterator ai = r->_activation_set[chain].begin(); ai != r->_activation_set[chain].end(); ai++) {
(*ai)->_init_refcount[chain] += 1;
}
- if (!has_input)
+ if (!has_input) {
+ /* no input, so this node needs to be triggered initially to get things going */
_init_trigger_list[chain].push_back (*ni);
+ }
- if (!has_output)
+ if (!has_output) {
+ /* no output, so this is one of the nodes that we can count off to decide
+ if we've finished
+ */
_init_finished_refcount[chain] += 1;
- }
+ }
+ }
_pending_chain = chain;
dump(chain);
}
+/** Called by both the main thread and all helpers.
+ * @return true to quit, false to carry on.
+ */
bool
Graph::run_one()
{
to_run = 0;
}
+ /* the number of threads that are asleep */
int et = _execution_tokens;
+ /* the number of nodes that need to be run */
int ts = _trigger_queue.size();
+ /* hence how many threads to wake up */
int wakeup = min (et, ts);
+ /* update the number of threads that will still be sleeping */
_execution_tokens -= wakeup;
- for (int i=0; i<wakeup; i++ ) {
- sem_post (&_execution_sem);
+ DEBUG_TRACE(DEBUG::ProcessThreads, string_compose ("%1 signals %2\n", pthread_self(), wakeup));
+
+ for (int i = 0; i < wakeup; i++) {
+ _execution_sem.signal ();
}
while (to_run == 0) {
_execution_tokens += 1;
pthread_mutex_unlock (&_trigger_mutex);
DEBUG_TRACE (DEBUG::ProcessThreads, string_compose ("%1 goes to sleep\n", pthread_self()));
- sem_wait (&_execution_sem);
- if (_quit_threads)
+ _execution_sem.wait ();
+ if (_quit_threads) {
return true;
+ }
DEBUG_TRACE (DEBUG::ProcessThreads, string_compose ("%1 is awake\n", pthread_self()));
pthread_mutex_lock (&_trigger_mutex);
if (_trigger_queue.size()) {
to_run->process();
to_run->finish (_current_chain);
- return false;
-}
-
-static void get_rt()
-{
- if (!jack_is_realtime (AudioEngine::instance()->jack())) {
- return;
- }
-
- int priority = jack_client_real_time_priority (AudioEngine::instance()->jack());
+ DEBUG_TRACE(DEBUG::ProcessThreads, string_compose ("%1 has finished run_one()\n", pthread_self()));
- if (priority) {
- struct sched_param rtparam;
-
- memset (&rtparam, 0, sizeof (rtparam));
- rtparam.sched_priority = priority;
-
- pthread_setschedparam (pthread_self(), SCHED_FIFO, &rtparam);
- }
+ return false;
}
void
Graph::helper_thread()
{
- ProcessThread *pt = new ProcessThread;
+ suspend_rt_malloc_checks ();
+ ProcessThread* pt = new ProcessThread ();
+ resume_rt_malloc_checks ();
pt->get_buffers();
- get_rt();
while(1) {
if (run_one()) {
pt->drop_buffers();
}
+/** Here's the main graph thread */
void
Graph::main_thread()
{
- ProcessThread *pt = new ProcessThread;
+ suspend_rt_malloc_checks ();
+ ProcessThread* pt = new ProcessThread ();
+ resume_rt_malloc_checks ();
pt->get_buffers();
- get_rt();
again:
- sem_wait (&_callback_start_sem);
+ _callback_start_sem.wait ();
+
+ DEBUG_TRACE(DEBUG::ProcessThreads, "main thread is awake\n");
+
+ if (_quit_threads) {
+ return;
+ }
- this->prep();
+ prep ();
if (_graph_empty && !_quit_threads) {
- sem_post (&_callback_done_sem);
+ _callback_done_sem.signal ();
+ DEBUG_TRACE(DEBUG::ProcessThreads, "main thread sees graph done, goes back to sleep\n");
goto again;
}
+ /* This loop will run forever */
while (1) {
+ DEBUG_TRACE(DEBUG::ProcessThreads, "main thread runs one graph node\n");
if (run_one()) {
break;
}
}
int
-Graph::silent_process_routes (nframes_t nframes, framepos_t start_frame, framepos_t end_frame,
- bool can_record, bool rec_monitors_input, bool& need_butler)
+Graph::silent_process_routes (pframes_t nframes, framepos_t start_frame, framepos_t end_frame, bool& need_butler)
{
_process_nframes = nframes;
_process_start_frame = start_frame;
_process_end_frame = end_frame;
- _process_can_record = can_record;
- _process_rec_monitors_input = rec_monitors_input;
_process_silent = true;
_process_noroll = false;
_process_need_butler = false;
if (!_graph_empty) {
- sem_post (&_callback_start_sem);
- sem_wait (&_callback_done_sem);
+ DEBUG_TRACE(DEBUG::ProcessThreads, "wake graph for silent process\n");
+ _callback_start_sem.signal ();
+ _callback_done_sem.wait ();
}
need_butler = _process_need_butler;
}
int
-Graph::process_routes (nframes_t nframes, framepos_t start_frame, framepos_t end_frame, int declick,
- bool can_record, bool rec_monitors_input, bool& need_butler)
+Graph::process_routes (pframes_t nframes, framepos_t start_frame, framepos_t end_frame, int declick, bool& need_butler)
{
+ DEBUG_TRACE (DEBUG::ProcessThreads, string_compose ("graph execution from %1 to %2 = %3\n", start_frame, end_frame, nframes));
+
_process_nframes = nframes;
_process_start_frame = start_frame;
_process_end_frame = end_frame;
- _process_can_record = can_record;
- _process_rec_monitors_input = rec_monitors_input;
_process_declick = declick;
_process_silent = false;
_process_retval = 0;
_process_need_butler = false;
- sem_post (&_callback_start_sem);
- sem_wait (&_callback_done_sem);
+ DEBUG_TRACE(DEBUG::ProcessThreads, "wake graph for non-silent process\n");
+ _callback_start_sem.signal ();
+ _callback_done_sem.wait ();
+
+ DEBUG_TRACE (DEBUG::ProcessThreads, "graph execution complete\n");
need_butler = _process_need_butler;
}
int
-Graph::routes_no_roll (nframes_t nframes, framepos_t start_frame, framepos_t end_frame,
- bool non_rt_pending, bool can_record, int declick)
+Graph::routes_no_roll (pframes_t nframes, framepos_t start_frame, framepos_t end_frame,
+ bool non_rt_pending, int declick)
{
+ DEBUG_TRACE (DEBUG::ProcessThreads, string_compose ("no-roll graph execution from %1 to %2 = %3\n", start_frame, end_frame, nframes));
+
_process_nframes = nframes;
_process_start_frame = start_frame;
_process_end_frame = end_frame;
- _process_can_record = can_record;
_process_declick = declick;
_process_non_rt_pending = non_rt_pending;
_process_retval = 0;
_process_need_butler = false;
- sem_post (&_callback_start_sem);
- sem_wait (&_callback_done_sem);
+ DEBUG_TRACE(DEBUG::ProcessThreads, "wake graph for no-roll process\n");
+ _callback_start_sem.signal ();
+ _callback_done_sem.wait ();
return _process_retval;
}
DEBUG_TRACE (DEBUG::ProcessThreads, string_compose ("%1 runs route %2\n", pthread_self(), route->name()));
if (_process_silent) {
- retval = route->silent_roll (_process_nframes, _process_start_frame, _process_end_frame, _process_can_record, _process_rec_monitors_input, need_butler);
+ retval = route->silent_roll (_process_nframes, _process_start_frame, _process_end_frame, need_butler);
} else if (_process_noroll) {
route->set_pending_declick (_process_declick);
- retval = route->no_roll (_process_nframes, _process_start_frame, _process_end_frame, _process_non_rt_pending, _process_can_record, _process_declick);
+ retval = route->no_roll (_process_nframes, _process_start_frame, _process_end_frame, _process_non_rt_pending);
} else {
route->set_pending_declick (_process_declick);
- retval = route->roll (_process_nframes, _process_start_frame, _process_end_frame, _process_declick, _process_can_record, _process_rec_monitors_input, need_butler);
+ retval = route->roll (_process_nframes, _process_start_frame, _process_end_frame, _process_declick, need_butler);
}
if (retval) {
_process_retval = retval;
}
-
+
if (need_butler) {
_process_need_butler = true;
}
}
-
-
+bool
+Graph::in_process_thread () const
+{
+ for (list<pthread_t>::const_iterator i = _thread_list.begin (); i != _thread_list.end(); ++i) {
+ if (*i == pthread_self()) {
+ return true;
+ }
+ }
+ return false;
+}