Fix and refactor channel average
- Fixing missing init average variable with zero and refactor the code #86b11ab7
Fixes just the missing avg1=0
- Refactor the code to a common SSE/AVX2 inline function using also horizontal add.
- We use int64_t result from now on.
- We decided to fix length to x12 instead of old x9 openair1/PHY/NR_UE_TRANSPORT/dci_nr.c
//compute average channel_level on each (TX,RX) antenna pair
void nr_pdcch_channel_level(int32_t rx_size,
c16_t dl_ch_estimates_ext[][rx_size],
NR_DL_FRAME_PARMS *frame_parms,
int32_t *avg,
int symbol,
int nb_rb)
{
for (int aarx = 0; aarx < frame_parms->nb_antennas_rx; aarx++) {
//clear average level
simde__m128i avg128P = simde_mm_setzero_si128();
simde__m128i *dl_ch128 = (simde__m128i *)&dl_ch_estimates_ext[aarx][symbol * nb_rb * 12];
for (int rb = 0; rb < (nb_rb * 12) >> 2; rb++) { <-------------- here
avg128P = simde_mm_add_epi32(avg128P, simde_mm_madd_epi16(dl_ch128[rb], dl_ch128[rb]));
}
DevAssert(nb_rb);
avg[aarx] = 0;
for (int i = 0; i < 4; i++)
avg[aarx] += ((int32_t *)&avg128P)[i] / (nb_rb * 12); <-------------- and here
LOG_DDD("Channel level : %d\n",avg[aarx]);
}
}
Optimize PHY_ofdm_mod CYCLIC_PREFIX in case of incidentally aligned pointers
It seems that the cyclic prefix is in most cases a multiple of 512 samples.
This means that in most cases the idft output pointer is already aligned and
there is no need to perform an extra memcpy. This saves the memcpy time in
most cyclic prefix insertion cases.
Alignment of temporary IDFT buffer was changed to 64 bytes. This is in preparation
for possible use of different IDFT implementation which could use AVX-512, requiring
64-byte alignment.
Add support for O1 in O-DU via telnet
This change set adds a simple telnet interface to nr-softmodem. Together
with oai/o1-adapter>, this implements a part of the O1 interface, for
the DU only. Currently supported uses cases:
- some stats are send to an SMO (such as bandwidth, PLMN, frequency, ...)
- re-configuration of bandwidth, frequency
- stop/start of the softmodem, while keeping the CU and L2 running
- the o1-adapter implements two alarms, load overload and base station
offline, using this interface.
It is possible to use the telnet interface by itself (i.e., "o1" without
the o1-adapter). To do this, compile with telnet support (./build_oai
--build-lib telnetsrv) and start the nr-softmodem with O1 module
(./nr-softmodem --telnetsrv --telnetsrv.shrmod o1). You can then do the
following:
- see some statistics as exposed to o1-adapter, in JSON: echo o1 stats | nc 127.0.0.1 9090 && echo
- stop the L1: echo o1 stop_modem | nc 127.0.0.1 9090 && echo
- reconfigure bandwidth: e.g., echo o1 bwconfig 20 | nc 127.0.0.1 9090 && echo (also supported: 40, 60, 100)
- start the L1: echo o1 start_modem | nc 127.0.0.1 9090 && echo
Note that the reconfiguration works with split-8 radios. Notably, it
does not work yet out-of-the-box with 7.2 radios, because they have to
be manually reconfigured on bandwith/frequency switch. The planned
M-plane should allow such use case, though.
To help with the above, this branch also includes fixes (for more
information, refer to the corresponding commits):
- fixes for stopping the L1, notably stop threads in right order, stop
thread pool of RU
- collect some general MAC metrics, fix some bugs in MAC
- fix problems in F1AP gNB-DU configuration update message, update test
- fix some problems in RRC
Need to keep track of number of RUs, as the RU initialization increments
the number of RUs. Therefore, subsequent initializations would fail if
we don't decrement first. There is no better place (or I did not find
it) to decrement the number of RUs than in nr_phy_free_RU(). For this
reason, we also have to manually set the number of RUs in nr_prachsim to
1, otherwise the assert will trigger.
The existing code prevents double initialization if e.g., both gNB&UE
call these functions (like in a simulator). However, in the case of a
double-initialization such as stopping and re-starting the L1, the
memory will be freed, but not freed again. Basically, this works:
generate()
generate() # not initialized again
but this leads to problems:
generate()
free()
generate() # not initialized again, but should be!
Reimplement the check based on the data to be initialized, such that an
intermediate free() will be handled properly.
New LDPC coding library interface for decoding and encoding complete slots in one call to the library
Previous attempts to integrate LDPC decoding relying on offloading show
the limitation of the original code segment coding interface for the
library implementing LDPC coding. In order to perform the offload and
therefore the decoding in an optimal fashion, it is relevant to provide
the implementation with the broader amount of data possible per call,
which is a slot. A previous MR related to T2 offload managed to offer
an interface for decoding transport blocks by modifying only
nr_ulsch_decoding.c. But this may not be sufficient for performing an
efficient decoding in any situation. Slots could be shared between
multiple transport blocks in case multiple UEs are connected to the RAN,
which makes the transport block size shrink and therefore degrades the
efficiency of the offloading.
This MR offers a broader interface with a gNB decoding entry point in
phy_procedures_nr_gNB.c. The file nr_ulsch_decoding.c now lets the
decoding library implement deinterleaving, rate-unmatching and decoding
for the slot. In gNB encoding and UE decoding and encoding as well,
tasks are partitioned in a similar fashion.
As it allows to feed the implementation with the broadest amount of data
possible, it allows to fit any coding solution in a library. This will
avoid further modification to OAI source code and addition of options to
select an implementation when integrating new decoding solutions. An
implementation of this interface will therefore be offered for every
existing decoding solution in OAI and solutions added in the future will
be provided as independent module code without any modification to the
core of OAI code. The way to choose the decoding solution and to
configure it was also unified using configmodule. The library can be
chosen through the option --loader.ldpc.shlibversion. The slot coding
solutions shipped with this MR can be configured through configmodule
(e.g., T2 address and isolated cores).
This MR includes three coding implementations:
- A wrapper around segment decoding libraries for test purpose: libldpc.so
- A one shot slot coding and encoding with the T2 board: libldpc_t2.so
- A decoder on FPGA using the XDMA driver: libldpc_xdma.so
Improvements brought by this MR:
- Easy integration of a wider variety of LDPC coding solution at gNB and UE
- Easy selection of the LDPC coding solution (option --ldpc-offload-enable is kept for now but it can be ultimately removed)
- T2:
* More efficient when many UEs are connected
* Decoding and encoding in nr-uesoftmodem
SIB19 utilization on UE
Following SIB19 initialization and scheduling on gNB (!2899 - merged)
and Receive SIB19 on nrUE (!2920 - merged), this merge request aims to
introduce the actual use of the SIB19 provided parameters by gNB in UE.
Changes needed for utilization include: initialization of position
coordinates parameters for UE (which are needed to calculate propagation
delay between UE and SAT), propagation delay calculations and timing
advance adjusts. Moreover, this change removes the now unused arguments
for ntn_koffset, ntn_ta_common and ntn_ta_commondrift
It seems that the cyclic prefix is in most cases a multiple of 512 samples.
This means that in most cases the idft output pointer is already aligned and
there is no need to perform an extra memcpy. This saves the memcpy time in
most cyclic prefix insertion cases.
More complex implementations will be achieved using the new reworked interface
- t_nrLDPC_dec_params: remove F, Qm, rv and perCB
- encoder_implemparams_t: remove Qm, Tbslbrm, G, rv and perCB
- Remove all variables (like OPENAIR1), they serve no purpose and make
unnecessarily long lines
- Remove linking to T, it does not seem to be necessary
- Avoid intermediate variables to keep sources, and declare libraries
directly, making for shorter code
Fix several issues in ulsch_mmse_2layers
Fix incorrect array type of determ_fin which was set to int32_t while the
array holds the results of simde_mm_abs_epi32 which produces unsigned
values.
Remove assert on right-shifting negative integers as it is impossible for
determ_fin to hold negative values.
Change type of sum_det to uint32_t to prevent possible overflow. Fix
the type cast (int -> uint32_t) to prevent casting before shift in
nr_ulsch_comp_muli_sum which might have resulted in right shifting
negative integers.
Closes: #753
Set the number of barriers to 512 and implement proper slot and frame
based selection. Now the dynamic barriers should not be overwritten in
case the rx-to-tx offset is high. The maximum offset supported is 512.
- Removed ntn_koffset, ntn_ta_common, ntn_ta_commondrift arguments in order to be used directly from SIB19
- Removed global NTN_UE_Koffset variable and it's use
- Introduced getter functions for ntn related parameters
Fix incorrect array type of determ_fin which was set to int32_t while the
array holds the results of simde_mm_abs_epi32 which produces unsigned
values.
Remove assert on right-shifting negative integers as it is impossible for
determ_fin to hold negative values.
Change type of sum_det to uint32_t to prevent possible overflow.
Fix the type cast (int -> uint32_t) to prevent casting before shift in
nr_ulsch_comp_muli_sum which might have resulted in right shifting negative
integers.
remove un-necessary global variables
Remove global variables that can be removed by a trivial way:
- simply dead globals
- or reference is a single source, so can be static
setting them static reduces the scope, that is a progress, but of course as
for all static variables, the code is still not re-entrant
Documentation: add multi-UE deployment with RFsimulator
Extended doc/NR_SA_Tutorial_OAI_nrUE.md with section about how to run multiple UEs
- multiple UEs with namespaces (provided new script)
- multiple UEs with Docker UEs
- updated rfsimulator and channel simulation tutorials
- updated README.md
T tracer: add macpdu traces for the NR UE
MIB, SIB1, random access and regular scheduling are traced.
At the beginning of a connection (rrc setup request and rrc setup), the
RNTI is reported as 0. Might be fixed, somehow, if absolutely needed (not
sure).
- removed some whitespace noise.
- cleaned up CMakeLists.txt
- reintegrated task_manager.c into thread-pool.c
- cleaned up some unnecessary code
Co-authored-by: Cedric Roux <cedric.roux@eurecom.fr>
Co-authored-by: Mikel Irazabal <mikel.irazabal@openairinterface.com>
MIB, SIB1, random access and regular scheduling are traced.
At the beginning of a connection (rrc setup request and rrc setup), the
RNTI is reported as 0. Might be fixed, somehow, if absolutely needed (not
sure).
FAPI P7 Refactoring and unitary test creation
This MR is a follow-up to the work done in !2714 this time separating
the P7 FAPI messages into its own library (nr_fapi_p7).
In the same way, utility functions to copy, free and compare have been
added for each message, as well as an addition function to get the
allocated size for a message, which is used for the messages that can be
too big to fit in a buffer that's allocated in the stack.
Unitary tests have been added for each of the messages, testing
pack/unpack and the utility functions added.
Imscope updates
- Using imgui docking branch - allows window docking.
- Moved scopes to different windows to allow docking.
- Disabled scatterplot for time domain samples - issues with amount of
vertices caused segfaults.
- Disabled IQ heatmaps for RX IQ samples on gNB and UE - this was
incorrect and currently I don't know how to fix it. This will be
reenabled once it is.
- added a simple menu and disabled demo windows by default.
Update NR UE threading model
This updates the NR UE threading model by preventing calling threadpool
from threadpool and fixes deadlock issues related to processSlotTX. By
making processSlotTX run to completion instead of starting and waiting
for other threads to finish it prevents locking thread pool cores,
therefore prevening deadlocks.
There is a slight difference here compared to the previous version where
I do not use the UL actor as I believe it is not necessary,
processSlotTX satisfies the run-to-completion requirement now. If in the
future any UL procedure would be parallelized using the thread pool it
should either be done in a run-to-completion model or the processSlotTX
function should be taken off the thread pool.
The documentation has been updated.
Laurent THOMAS