Using the new feature of pa_{source,sink}_get_latency_within_thread() leads to improved end to end latency estimation and to correct handling of port latency offsets. --- src/modules/module-loopback.c | 67 ++++++++++++++++++++++++++++++------------- 1 file changed, 47 insertions(+), 20 deletions(-) diff --git a/src/modules/module-loopback.c b/src/modules/module-loopback.c index f505295..aee8658 100644 --- a/src/modules/module-loopback.c +++ b/src/modules/module-loopback.c @@ -110,12 +110,12 @@ struct userdata { /* Used for sink input and source output snapshots */ struct { int64_t send_counter; - pa_usec_t source_latency; + int64_t source_latency; pa_usec_t source_timestamp; int64_t recv_counter; size_t loopback_memblockq_length; - pa_usec_t sink_latency; + int64_t sink_latency; pa_usec_t sink_timestamp; } latency_snapshot; @@ -128,6 +128,7 @@ struct userdata { /* Copied from main thread */ pa_usec_t effective_source_latency; + int64_t source_latency_offset; pa_usec_t minimum_latency; /* Various booleans */ @@ -171,7 +172,8 @@ enum { SINK_INPUT_MESSAGE_SINK_CHANGED, SINK_INPUT_MESSAGE_SOURCE_CHANGED, SINK_INPUT_MESSAGE_SET_EFFECTIVE_SOURCE_LATENCY, - SINK_INPUT_MESSAGE_UPDATE_MIN_LATENCY + SINK_INPUT_MESSAGE_UPDATE_MIN_LATENCY, + SINK_INPUT_MESSAGE_SOURCE_LATENCY_OFFSET_CHANGED }; enum { @@ -314,7 +316,8 @@ static void adjust_rates(struct userdata *u) { size_t buffer; uint32_t old_rate, base_rate, new_rate, run_hours; int32_t latency_difference; - pa_usec_t current_buffer_latency, snapshot_delay, current_source_sink_latency, current_latency, latency_at_optimum_rate; + pa_usec_t current_buffer_latency, snapshot_delay; + int64_t current_source_sink_latency, current_latency, latency_at_optimum_rate; pa_usec_t final_latency; pa_assert(u); @@ -361,7 +364,7 @@ static void adjust_rates(struct userdata *u) { latency_at_optimum_rate = current_source_sink_latency + current_buffer_latency * old_rate / base_rate; final_latency = PA_MAX(u->latency, u->minimum_latency); - latency_difference = (int32_t)((int64_t)latency_at_optimum_rate - final_latency); + latency_difference = (int32_t)(latency_at_optimum_rate - final_latency); pa_log_debug("Loopback overall latency is %0.2f ms + %0.2f ms + %0.2f ms = %0.2f ms", (double) u->latency_snapshot.sink_latency / PA_USEC_PER_MSEC, @@ -474,12 +477,23 @@ static void update_latency_boundaries(struct userdata *u, pa_source *source, pa_ /* Called from output context * Sets the memblockq to the configured latency corrected by latency_offset_usec */ -static void memblockq_adjust(struct userdata *u, pa_usec_t latency_offset_usec, bool allow_push) { +static void memblockq_adjust(struct userdata *u, int64_t latency_offset_usec, bool allow_push) { size_t current_memblockq_length, requested_memblockq_length, buffer_correction; - pa_usec_t requested_buffer_latency, final_latency; + int64_t requested_buffer_latency; + pa_usec_t final_latency, requested_sink_latency; final_latency = PA_MAX(u->latency, u->output_thread_info.minimum_latency); - requested_buffer_latency = PA_CLIP_SUB(final_latency, latency_offset_usec); + + /* If source or sink have some large negative latency offset, we might want to + * hold more than final_latency in the memblockq */ + requested_buffer_latency = (int64_t)final_latency - latency_offset_usec; + + /* Keep at least one sink latency in the queue to make sure that the sink + * never underruns initially */ + requested_sink_latency = pa_sink_get_requested_latency_within_thread(u->sink_input->sink); + if (requested_buffer_latency < (int64_t)requested_sink_latency) + requested_buffer_latency = requested_sink_latency; + requested_memblockq_length = pa_usec_to_bytes(requested_buffer_latency, &u->sink_input->sample_spec); current_memblockq_length = pa_memblockq_get_length(u->memblockq); @@ -500,7 +514,8 @@ static void memblockq_adjust(struct userdata *u, pa_usec_t latency_offset_usec, /* Called from input thread context */ static void source_output_push_cb(pa_source_output *o, const pa_memchunk *chunk) { struct userdata *u; - pa_usec_t push_time, current_source_latency; + pa_usec_t push_time; + int64_t current_source_latency; pa_source_output_assert_ref(o); pa_source_output_assert_io_context(o); @@ -508,9 +523,9 @@ static void source_output_push_cb(pa_source_output *o, const pa_memchunk *chunk) /* Send current source latency and timestamp with the message */ push_time = pa_rtclock_now(); - current_source_latency = pa_source_get_latency_within_thread(u->source_output->source, false); + current_source_latency = pa_source_get_latency_within_thread(u->source_output->source, true); - pa_asyncmsgq_post(u->asyncmsgq, PA_MSGOBJECT(u->sink_input), SINK_INPUT_MESSAGE_POST, PA_UINT_TO_PTR(current_source_latency), push_time, chunk, NULL); + pa_asyncmsgq_post(u->asyncmsgq, PA_MSGOBJECT(u->sink_input), SINK_INPUT_MESSAGE_POST, PA_INT_TO_PTR(current_source_latency), push_time, chunk, NULL); u->send_counter += (int64_t) chunk->length; } @@ -539,7 +554,7 @@ static int source_output_process_msg_cb(pa_msgobject *obj, int code, void *data, u->latency_snapshot.send_counter = u->send_counter; /* Add content of delay memblockq to the source latency */ - u->latency_snapshot.source_latency = pa_source_get_latency_within_thread(u->source_output->source, false) + + u->latency_snapshot.source_latency = pa_source_get_latency_within_thread(u->source_output->source, true) + pa_bytes_to_usec(length, &u->source_output->source->sample_spec); u->latency_snapshot.source_timestamp = pa_rtclock_now(); @@ -773,7 +788,7 @@ static int sink_input_pop_cb(pa_sink_input *i, size_t nbytes, pa_memchunk *chunk * when it starts pushing. Adjust the memblockq accordingly and ensure that there is * enough data in the queue to avoid underruns. */ if (!u->output_thread_info.push_called) - memblockq_adjust(u, u->output_thread_info.effective_source_latency, true); + memblockq_adjust(u, (int64_t)u->output_thread_info.effective_source_latency + u->output_thread_info.source_latency_offset, true); return 0; } @@ -815,19 +830,19 @@ static int sink_input_process_msg_cb(pa_msgobject *obj, int code, void *data, in * are enabled. Disable them on first push and correct the memblockq. Do the * same if the pop_cb() requested the adjustment */ if (!u->output_thread_info.push_called || u->output_thread_info.pop_adjust) { - pa_usec_t time_delta; + int64_t time_delta; - time_delta = PA_PTR_TO_UINT(data); + time_delta = PA_PTR_TO_INT(data); time_delta += pa_rtclock_now() - offset; - time_delta += pa_sink_get_latency_within_thread(u->sink_input->sink, false); + time_delta += pa_sink_get_latency_within_thread(u->sink_input->sink, true); /* If the source has overrun, assume that the maximum it should have pushed is * one full source latency. It may still be possible that the next push also * contains too much data, then the resulting latency will be wrong. */ if (pa_bytes_to_usec(chunk->length, &u->sink_input->sample_spec) > u->output_thread_info.effective_source_latency) - time_delta = PA_CLIP_SUB(time_delta, u->output_thread_info.effective_source_latency); + time_delta -= (int64_t)u->output_thread_info.effective_source_latency; else - time_delta = PA_CLIP_SUB(time_delta, pa_bytes_to_usec(chunk->length, &u->sink_input->sample_spec)); + time_delta -= (int64_t)pa_bytes_to_usec(chunk->length, &u->sink_input->sample_spec); memblockq_adjust(u, time_delta, true); @@ -838,7 +853,7 @@ static int sink_input_process_msg_cb(pa_msgobject *obj, int code, void *data, in /* If pop has not been called yet, make sure the latency does not grow too much. * Don't push any silence here, because we already have new data in the queue */ if (!u->output_thread_info.pop_called) - memblockq_adjust(u, pa_sink_get_latency_within_thread(u->sink_input->sink, false), false); + memblockq_adjust(u, pa_sink_get_latency_within_thread(u->sink_input->sink, true), false); /* Is this the end of an underrun? Then let's start things * right-away */ @@ -876,7 +891,7 @@ static int sink_input_process_msg_cb(pa_msgobject *obj, int code, void *data, in u->latency_snapshot.recv_counter = u->output_thread_info.recv_counter; u->latency_snapshot.loopback_memblockq_length = pa_memblockq_get_length(u->memblockq); /* Add content of render memblockq to sink latency */ - u->latency_snapshot.sink_latency = pa_sink_get_latency_within_thread(u->sink_input->sink, false) + + u->latency_snapshot.sink_latency = pa_sink_get_latency_within_thread(u->sink_input->sink, true) + pa_bytes_to_usec(length, &u->sink_input->sink->sample_spec); u->latency_snapshot.sink_timestamp = pa_rtclock_now(); @@ -907,6 +922,12 @@ static int sink_input_process_msg_cb(pa_msgobject *obj, int code, void *data, in u->output_thread_info.minimum_latency = (pa_usec_t)offset; return 0; + + case SINK_INPUT_MESSAGE_SOURCE_LATENCY_OFFSET_CHANGED: + + u->output_thread_info.source_latency_offset = offset; + + return 0; } return pa_sink_input_process_msg(obj, code, data, offset, chunk); @@ -1185,6 +1206,12 @@ static pa_hook_result_t source_port_latency_offset_changed_cb(pa_core *core, pa_ u->source_latency_offset = source->port_latency_offset; update_minimum_latency(u, u->sink_input->sink, true); + /* Update variable in output thread */ + if (u->sink_input->sink) + pa_asyncmsgq_send(u->sink_input->sink->asyncmsgq, PA_MSGOBJECT(u->sink_input), SINK_INPUT_MESSAGE_SOURCE_LATENCY_OFFSET_CHANGED, NULL, u->source_latency_offset, NULL); + else + u->output_thread_info.source_latency_offset = u->source_latency_offset; + return PA_HOOK_OK; } -- 2.10.1