Two buffers, inner and outer, are used in the transceiver
implementation. The outer buffer interfaces with the device receive
interface to guarantee timestamp aligned and contiguously allocated
sample buffers. The inner buffer absorbs vector size differences between
GSM bursts (156 or 157 samples) and the resampler interface (typically
fixed multiples of 65).
Reimplement the inner buffer with a ring buffer that allows fixed size
segments on the outer (resampler) portion and variable lengths (GSM
side) on the inner side. Compared to the previous stack-like version,
this implementation removes unnecessary copying of buffer contents.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
Slightly widen the search range to accommodate timing jitter
on certain classes of devices. The expanded range minimizes
the possibility of missing bursts that arrive too early or
too late due to timing error.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
EDGE 8-PSK soft slicer was receiving input from the output of the
downsampler. Equalization and derotation were missing causing the
soft symbol output to be invalid.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
Delay the EDGE downlink bursts by one symbol in order to match GMSK
pulse shaping group delay. The difference in group delay arises from
the dual pulse filter combination of the GMSK Laurent represenation
whereas 8-PSK uses a single pulse linear filter.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
Previously MAXDLY value was applied to Normal Bursts, which was nice
when working with sloppy test equipment like CMD57, but useless for
real world usage. At the same time documentation and de facto usage
of MAXDLY in OsmoBTS and OpenBTS assumed that it actually applies to
Access Bursts (RACH). So this patch changes osmo-rx behavior to apply
MAXDLY to RACH bursts and introduces a new command MAXDLYNB for the
old behavior.
Timing recovery and single tap channel compensation are identical
in both GMSK and EDGE receivers. This is the section ahead of and
including the optional 4-1 downsampler. GMSK and EDGE specific
sections operate at 1 SPS.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
Setup generators for empty, random, and dummy bursts. This moves error
prone burst length handling out of the Transceiver and into the signal
processing core.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
At 4 samples per symbol, we don't need to maintain the 156/157 sample
slot structure to account for the GSM 156.25 sample burst length.
Set the 4 SPS Laurent modulator to ignore the guard interval setting
and always output 625 sample sized bursts. The EDGE 8-PSK modulator
already has this behavior.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
Setup correlator and detection process similar to the GMSK
receiver chain. Require 4 SPS sampling on both Rx and Tx paths
as 1 SPS sampling adds too much distoration for 8-PSK recovery.
Core receiver operations still run at 1 SPS with the exception
of fractional delay filtering, which runs at the higher rate.
Perform linear equalization to handle the Gaussian pulse
induced ISI. The fixed impulse response used for equalizer tap
calculation consists of combined EDGE pulse shape filter and
effects of the downsampling filter. Note that the non-adaptive
equalizer corrects for modulation induced band limiting and
does not account for or compensate for fading channel effects.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
DFE equalizer is unused and has been experiencing code rot for
multiple years. The effect is a significant amount of baggage being
carried in the Transceiver and interfaces.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
Samples per symbol used by the transceiver is not configurable through
the socket interface once running, so stop pretending like it could be.
Initialize all tables and midambles at start.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
As a side change - get rid of passing toa and amp arguments as pointers and use
references instead.
The commit doesn't change behaviour, but makes the code cleaner.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
There are two primary changes in this commit:
1) Return values of detect functions changed form bool to int to actually pass
the return value from the inner function and notify higher levels about clipping.
Previously the information was lost due to conversion to bool.
2) Clipping level is not the final verdict now. We still try to demod a burst
and mark it as clipped only if the level is above the clipping level AND we can't
demod it. The reasoning for this is that in real life we want to do as much as
possible to demod the burst, because we want to get as much from our dynamic
range as possible. So a little bit of clipping is fine and is expected. We just
don't want too much of it to break our demod.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
R&S CMD57 complains about the start phase of bursts, particularly it shows
-15 to -30 deg of error for the bit 0.5 position (start tail bit). This patch
makes it happy. ETSI TS 145 004 section 2.2 describes this: "Before the first
bit of the bursts as defined in 3GPP TS 45.002 enters the modulator,
the modulator has an internal state as if a modulating bit stream consisting
of consecutive ones (di = 1) had entered the differential encoder."
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
Alert user of overdriven burst input indicated by a positive
threshold detector result. This indication serves as notification
that the receive RF gain level is too high for the configured
transceiver setup.
Signed-off-by: Tom Tsou <tom.tsou@ettus.com>
The call into table lookup will loop on values outside of the
table range. With continuously increasing phase, this leads
to an eventual permanent hard spin. Wrap the phase value to
prevent that from happening.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
Stack allocating the correlation output generates a call to the copy
constructor of an zero valued vector. We can avoid this extra copy
constructor with a pointer reference and dynamic allocation.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
Allow non-in-place use of the delay setting. Internally, the delay call
creates a new vector and copies the contents back into the original.
Instead, provide the option to return the computed output vector
directly and remove an an extra copy in the process.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
On Beagle Board the call into the sinc() function is generating a lot of
load on the peak interpolation. Simplify the sinc() function with a
dedicated table lookup. Eventually, this table may be removed in favour
of using a precomputed filterbank for fractional delay determination.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
Preallocate and compute a bank of fractional sample delay filters.
The number of filters to allocate is specified by the DELAYFILTS
preprocessor definition with a default value of 64. The filters
themselves are sinc pulse generated with 20 taps and Blackman-harris
windowed .
Signed-off-by: Thomas Tsou <tom@tsou.cc>
This prevents the use of a copy constructor in the downlink
modulator and prevents a secondary memory allocation during
the convolution. Avoid both cases by dynamically allocating
with preloaded head room. The latter provides enough memory
before the first sample in the burst to cover the length
of the filter taps.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
Complex-complex block multiples are used for phase rotation of
bursts. Optimization targeted from perf profiling.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
Similar to the existing Intel SSE cases, add support for NEON vector
floating point SIMD processing. In this case, use ARM assembly
directly as the NEON intrinsics do not generate preferential code
output.
Currently support NEON vectorized convolution and floating point
integer conversions.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
USRPping and sigProcLibTest are in an unmaintained state,
while the intended functionality remains unknown. Stored
filter taps are also unused and should also be removed.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
This patch applies oversampling, when selected with 4 sps,
to the downlink only, while running the receiver with
minimal sampling at 1 sps. These split sample rates allow
us to run a highly accurate downlink signal with minimal
distortion, while keeping receive path channel filtering
on the FPGA.
Without this patch, we oversample the receive path and
require a steep receive filter to get similar adjacent
channel suppression as the FPGA halfband / CIC filter
combination, which comes with a high computational cost.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
This requires an additional memcpy() on the signal vector
constructor, but allows the interpolation filter to use
SSE optimzationed convolution.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
Start the correlation search window at 4 symbols before
the expected correlation peak. End the search at 10
symbols and 4 + maximum expected delay for RACH and TSC
bursts respectively.
This change lowers receive side cpu utilization while
maintaining reasonable timing jitter and accuracy tolerance.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
Provides substantially improved transmit phase error
performance when enabled. Requires use of 4 samples
per symbol, and is enabled by default when set.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
Both RACH and normal bursts are detected with the same approach of
midamble correlation combined with peak-to-average ratio. The
difference is the midamble placements and lengths. Thus, there is
no reason to have independent implementations.
This patch creates a common call burstDetect(), while leaving the
correlation window indexing in the original calls.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
When 4 samples-per-symbol operation is selected, replace the
existing pulse approximation, which becomes inaccurate with
non-unit oversampling, with the primary pulse, C0, from the
Laurent linear pulse approximation.
Pierre Laurent, "Exact and Approximate Construction of Digital Phase
Modulations by Superposition of Amplitude Modulated Pulses", IEEE
Transactions of Communications, Vol. 34, No. 2, Feb 1986.
Octave pulse generation code for the first three pulses of the
linear approximation are included.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
The adaptive energy threshold gating suffers a near-far problem
at certain gain levels. This is due to exponential threshold
raising, but linear decreases. A large signal level followed by
a period low signal level causes (comparatively) weak signals to
go undetected. Additionally, the algorithm performs differently
at multiple RF gain levels.
This patch switches solely to correlation based gating for burst
detection. The main computational load with this approach is
sub-sample width peak interpolation, which we disable for intial
detection and run after threshold passing.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
Replace the polyphase filter and resampler with a separate
implementation using SSE enabled convolution. The USRP2 (including
derived devices N200, N210) are the only supported devices that
require sample rate conversion, so set the default resampling
parameters for the 100 MHz FPGA clock. This changes the previous
resampling ratios.
270.833 kHz -> 400 kHz (65 / 96)
270.833 kHz -> 390.625 kHz (52 / 75)
The new resampling factor uses a USRP resampling factor of 256
instead of 250. On the device, this allows two halfband filters to
be used rather than one. The end result is reduced distortial and
aliasing effecits from CIC filter rolloff.
B100 and USRP1 will no be supported at 400 ksps with these changes.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
This large patch replaced the convolve() call with an SSE vector
enabled version. The lower C and SSE intrinsic based code operates
on fixed and aligned vectors for the filter taps. The storage format
of interleaved I/Q for both complex and real vectors is maintained.
SSE filter tap values must:
1. Start 16-byte aligned
2. Number with a multiple of 4 between 4 and 20 for real taps
3. Number with a multiple of 4 for complex taps
Non-compliant values will fall back to non-SSE usage. Fixed length
iterators mean that head and tail cases may require reallocation of
the input vector, which is automatically handled by the upper C++
interface.
Other calls are affected by these changes and adjusted or rewritten
accordingly. The underlying algorithms, however, are unchanged.
generateGSMPulse()
analyzeTrafficBurst()
detectRACHBurst()
Intel SSE configuration is automatically detected and configured at
build time with Autoconf macros.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
There is no temporal dependency on when the RACH sequence is generated,
so there is no need for transceiver to create it in response to a
command from GSM core. If we power on the transceiver, we will need
the RACH sequence, so just allocate it during initialization.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
The only logging outputs in the the signal processing library
are debug lines that generate copious amounts of output while
providing little useful information to the user. The relevant
information (time-of-arrival, channel gains, etc.) can and
should be logged from transceiver instance itself.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
Add destructor calls so we can avoid the nested vector deallocations.
Also remove the unnecessary pointer NULL checks prior to destruction.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
There is no reason expose the pulse shaping filter outside of the
signal processing calls. The main transceiver object makes no use
of the filter and there's no reason to pass it around.
Initialize the pulse shape with the signal processing library, and
maintain an internal static member like many of the other library
variables. Similarly destroy the object when the library is closed.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
Because repeatedly typing mSamplesPerSymbol is giving me
carpal tunnel syndrome. Replace with the much shorter,
easier to type, and just as clear name of 'sps'.
Signed-off-by: Thomas Tsou <tom@tsou.cc>
The correlation starting point for normal burst training sequence
calculation should be a scaled value of the same symbol regardless
of the samples-per-symbol used. Use of 2 samples-per-symbols double
the index values, but yields the following outputs, which results
in a late time-of-arrival value at the output of the correlation.
This patch modifies the parameter calculation accordingly.
1 sps parameters
maxTOA = 3
spanTOA = 5;
startIx = 61;
endIx = 87;
windowLen = 26;
corrLen = 7;
2 sps parameters (errant case)
maxTOA = 6;
spanTOA = 10;
startIx = 112;
endIx = 184;
windowLen = 72;
corrLen =13;
2 sps parameters (corrected)
maxTOA = 6;
spanTOA = 10;
startIx = 122;
endIx = 174;
windowLen = 52;
corrLen =13;
Signed-off-by: Thomas Tsou <tom@tsou.cc>
git-svn-id: http://wush.net/svn/range/software/public/openbts/trunk@5183 19bc5d8c-e614-43d4-8b26-e1612bc8e597
With transceiver integration, the resampling filter files were
dropped. This created a working resampling implementation for
the USRP2 / N200, but with spectrum irregulaties that likely
caused issues at longer range operation. Simply reinsert the
filter files and modify the filter initialization to use them.
Signed-off-by: Thomas Tsou <ttsou@vt.edu>
git-svn-id: http://wush.net/svn/range/software/public/openbts/trunk@2689 19bc5d8c-e614-43d4-8b26-e1612bc8e597
Tom Tsou