When interpolating with rate correction != 1.0, don't floor the
resulting input rate to the nearest smallest integer.
This allows rate corrections < 1/in_rate to have some effect, and
reduces jumps in the response. One of the jumps is inconveniently
between rate=1.0 and rate=1.0+eps and will cause rate corrections to
oscillate if target rate varies close to 1.0.
If phase is float, calculations in impl_native_in_len/out_len can
produce wrong results due to rounding error.
It's probably better to not be in the business of predicting
floating-point rounding, so replace this by fixed-point arithmetic.
Also make sure `offset+1` cannot overflow data->filter array in
do_resample_inter* due to float multiplication possibly rounding up.
If phase is float, calculations in impl_native_in_len/out_len don't
necessarily match with do_resample, because e.g.
float phase0 = 7999.99;
float phase = phase0;
int frac = 8000, out_rate = 8000, n = 64, count = 0;
for (int j = 0; j < n; ++j) {
phase += frac;
if (phase >= out_rate) {
phase -= out_rate;
count++;
}
}
printf("count = %d\n", count); /* count = 64 */
count = (int)(phase0 + n*frac) / out_rate;
printf("count = %d\n", count); /* count = 65 */
don't give the same result.
Also add test where floating point multiplication rounding up to nearest
in
float ph = phase * pm;
uint32_t offset = (uint32_t)floorf(ph);
computation results to offset+1 > data->n_phases, accessing filter array
beyond bounds. (The accessed value is still inside allocated memory
block, but contains unrelated values; the test passes silently.)
Using the parser for the spa_pod_sequence in the data buffers is
required in order to safely read the pods while there could be
concurrent writes.
See #4822
We don't actually need to calculate the GCD for each resampler rate
update. The GCD is only used to scale the in/out rates when using the
full resampler and this we can cache and reuse when we did the setup.
The interpolating resampler can work perfectly fine with a GCD of 1 and
so we can just assume that.
spa_alsa_read is called from the source process function when we are a
follower and no buffer is ready yet.
Part of the rate correction was performed by the ALSA driver when it
woke up but now, the resampler has updated the requested size and we
need to requery it before we can start reading samples.
Otherwise, we end up with requested samples from before the rate update
and we might not give enough samples to the resampler. In that case, the
adapter will call us again and we will again try to produce a buffer
worth of the requested samples, which will xrun.
Keep a list of active ports in the port_list. These are all ports added
with add_port and not yet removed. When a port is removed, move it to
the free_list and reuse the port later when needed.
Update a mix_list of ports when a valid io is set on a port. This then
makes it possible to more efficiently and safely iterate the ports in
the processing loop.
Use bits to capture the work that is needed. We clear the bit when
we added the stage, when all bits are cleared we have nothing more to
do. This avoids having to check multiple bookleans.
Make a helper function to calculate the destination buffer. When all
bits are cleared, we can use the output buffer.
If the timer was canceled, the discont flag needs to be set. But in the
next cycle, unless the timer was canceled again, that flag should not
remain set.
If the user alters the realtime clock (for example by using the "date"
command in the shell), and the node driver uses the realtime clock as
the timerfd clock, then the scheduled graph cycle invocation may not
take place, or may take place much later than planned, because the
timestamp that was passed to spa_system_timerfd_settime() is now invalid.
Configure the timer to automatically be canceled if the realtime clock
is modified so that the graph cycle can be rescheduled with an updated
timestamp that is actually usable with the altered realtime clock.
It uses the onnxruntime library to parse the onnx file and construct a
neural network. It uses the label field to setup the plugin and how to
map the various tensors of the model to input, output, control and
notify ports.
Add an example config for how to use the silero VAD ONNX model with the
noise gate.
There is an issue in the id allocation mechanism which can result
in the different devices having the same id. Specifically, consider
the scenario where there are only two cameras, which have just been
added. In this case `impl::devices` looks like this:
(0, camA) | (1, camB) | (?, nullptr) | ...
Now assume that `camA` is removed, after which the array appears
as follows:
(1, camB) | (1, nullptr) | (?, nullptr) | ...
Then assume that a new camera appears. When `get_free_id()` runs,
when `i == 1`, it will observe that `devices[i].camera == nullptr`,
so it selects `1` as the id. Leading to the following:
(1, camB) | (1, camC) | (?, nullptr) | ...
This is of course incorrect. The set of ids must be unique. When
wireplumber is faced with this situation it destroys the device
object for `camB` when `camC` is emitted.
Fix this by simply not moving elements in the `devices` array,
leaving everything where it is. In which case the array looks
like this:
(nullptr) | (camB) | (nullptr) | ... // after `camA` removal
(camC) | (camB) | (nullptr) | ... // after `camC` appearance
Note that `device::id` is removed, and the id is now derived from
the position in `impl::devices`.
If the libcamera `FrameMetadata` reports anything other than `FrameSuccess`,
then set `SPA_META_HEADER_FLAG_CORRUPTED`, notifying the application that
the frame may be unusable.
Use a union since only one member is active at a time, and use the
proper `libcamera::ControlType` enum to store the type instead of a
bare number. Also remove an unnecessary cast.
