1. The encoder writes the output in sequence after the encoding tasks completed instead of writing within the tasks.
2. Simplify the output writing to make it more readable.
Co-authored-by: Laurent Thomas <laurent.thomas@open-cells.com>
There was a risk of a race condition between encoding tasks on the output array.
It could be that two task compete in writing a word at the border of their respective segment groups.
This commit offers to solve this issue by retaining the tail word in a struct that is returned after task completion so that segment group border can be fixed after completion of all the tasks.
Comes with additional cleanups of nrLDPC_coding_segment_encoder:
1. Move the section of code dedicated to writing to the encoding library output to a separate function.
2. Remove bit interleave debug in AVX512VBMI case of output writing for improving code readability.
1. nrLDPC_coding_segment_encoding: Clarifying encoder parameters
The way the encoder implementation params was used could be made clearer:
* `macro_num` was always used in order to calculate segment indexes.
It is clearer to pass directly the first segment index.
* The initialisation of a common implementation parameters structure can be made clearer.
2. nrLDPC_coding_segment_encoder: Remove multi segment dimension in encoder output
The output of `LDPCencoder` was a 2D unsigned char array.
But all the segments were actually packed in the first segment, 8 bits of different segments per byte of the first segment.
So the first dimension was useless. This commit removes this dimension.
3. nrLDPC_coding_segment_encoder: Rename `nrLDPC_prepare_TB_encoding` into `nrLDPC_launch_TB_encoding`
4. nrLDPC_coding_segment_encoder: Remove the useless second rate matching for E2, one rate matching with max(E, E2) does the job.
5. nr_interleaving_ldpc: cleanup
Rewrite code to use different SIMD sizes in a linear fashion.
This means cascading preprocessor directives are used to start from a high SIMD size an decrease it progressively.
6. ldpc_encoder_optim8segmulti: linear usage of SIMD
Use cascading preprocessor directives to include different SIMD sizes and types depending on which are available on the system.
The different loops use a common index that is progressively incremented.
Also add AVX512 version of bits distribution from 8 segments into one byte.
Co-authored-by: Laurent Thomas <laurent.thomas@open-cells.com>
Since this branch aims to bring all the encoded segments packed and contiguous,
it doesn't make anymore sense to have one output array per segment.
Therefore this commit moves the output array to the TB instead of the segment and adapts its size.
TODO: Adapt the size of f where it is relevant.
1. nrLDPC_coding_segment_encoder: pointer not aligned on 64 bits
An undefined behavior was happening because of a shift on a __m64 which
was actually a uint32_t which was not aligned on 64 bits.
This commit suggests a fix by properly loading into a __m64 with _mm_set_pi32
before performing the shift.
2. nr_rate_matching: Filler bits detection
Previously, if the bit selection was reaching Ncb and restarting selection from the beginning of encoded data, the filler bits were detected by testing the bytes again NR_NULL which was equal to 2.
But on this branch the bits of up to 8 segments are stored in one byte while doing rate matching so that one byte could be tested to be a filler while it is not.
So this commit changes the way to test filler bits.
Instead of relying on NR_NULL, bit selection now relies on the offset and size of filler bits.
3. nr_ulsch_coding: missing filling f with zeros
The array `f` of the UE UL HARQ process was not filled with zeros between transmissions.
This is a problem if the encoder does not properly overwrite its output.
In order to avoid such situation, this commit adds a memset to fill `f` with zeros before passing it to the encoder.
4. nrLDPC_coding_segment_encoder: missing init of f to zeros
In the encoder, `f` is not filled with zeros before being used.
`f` is instead somehow filled with arbitrary values.
When writing to `output` by 64 bit vectors, some arbitrary bits may be added after the segment.
Then the arbitrary added ones are not overwritten when writing the next segment because of writing with a logical or.
This commit properly fills `f` and `f2` with zeros.
5. nr_ulschsim & nr_dlschsim: Adapt modulation to new encoder output
The QPSK symbols that are feed to the channel model of nr_dlschsim were determined based on the bytes of the encoder output.
With the new output of the encoder, it had to be reworked to determine the QPSK symbols based on the bits of the encoder output.
1. 3GPP Rate-matching/interleaving on 8-segment blocks, deinterleaving at output of multi-segment processing.
2. Bit-packing done in nrLDPC_coding_segment_encoder.c now instead of nr_scrambling.
Interestingly, it improves performance significantly too.
Tested for AVX512 only for now.
Avoid the use of simde_mm_sign_epi8 which is inefficient on aarch64.
Decoding times are significantly improved on aarch64.
Co-authored-by: Romain Beurdouche <romain.beurdouche@eurecom.fr>
Add support for ARM build pipeline
Modify the existing python code to be able to build images where the
image tag can be prepended with a prefix, here arm_. This is preparatory
work to create the ARM build image pipeline, and reuse the existing
internal registry on porcepix to have x86 and ARM images coexist.
Fix various bugs in the build system to allow to build on a system with
as many cores as gracehopper.
Also, fix two programs (usim, nvram) to work correctly under ARM.
Various libraries in this commit use the log headers (that is the
direct dependency). The log headers in turn include the T headers
(transitive dependency). Building on gracehopper with its many cores
triggers the build of these libraries before T headers have been
generated.
To resolve this and force the generation of T headers before building
the libraries, correctly express their dependency on the log headers,
which in turn ensures that T headers have been generated.
Depending on log_headers is preferrable because it is really the headers
that we depend on. Using instead e.g., utils, would trigger build of
source files, that we don't need for these libraries. This commit is
part of a larger cleanup that I have that is not merged yet. Follow-up
work will generalize this in the repo, and ensure that every(!) library
builds.
We have to add CONFIG_LIB, as the log module has a direct dependency
onto the configuration module (it uses it). I am not sure why this was
not relevant before.
The idea is to make LDPC iteration accounting more standard from a C coding perspective by making two changes:
1. Now the iteration counter starts at 0 for the first iteration and goes up to max_ldpc_iterations - 1 for the last iteration.
If it is equal to max_ldpc_iterations, it means that decoding failed.
2. A chain of useless returns in the UE DLSCH decoding call chain was removed.
The logs of `./nr_ulsim -n100 -m19 -s12 -S12 -z2 -L4` were showing that sometimes the decoding of one segment was failing while the CRC check succeeded.
This was happening because the iteration counter `numIter` was not properly managed so that when the last iteration was successful to decode the segment
still `numIter` at the exit was equal to `numMaxIter` plus 1 so that decoding was considered to have failed.
In other words the last iteration if it occured was always wasted.
This commit offers to solve this problem by simply moving the iteration counter increment at the end of the iteration loop
while remaining careful at fixing the tests relying on this counter.
fix(nrLDPC_coding): timers in DL encoding, fix(ldpctest): Confusion in sizes
Main improvements:
1. fix(ldpctest): Confusion in sizes
IMPORTANT: This fix makes that ldpctest now works with BG2 and K' != K!
(Worth double checking with more cases than I did check with though)
The variable block_length had an ambiguous role and was used as both K and
Kprime. This was leading to arrays with wrong sizes and functions with wrong
arguments. Now there is not anymore block_length but K and Kprime.
Kprime can be set with the command line argument -l.
There was also some problems with array allocation for segments with length
not divisible by 8. A proper sizing with ((size_in_bits + 7) & ~7) / 8 was
used.
A proper masking of the last byte containing payload bits was also added to
keep filler bits to 0 after random initialization.
2. fix(nrLDPC_coding): timers in DL encoding
While timers were available for rate matching and interleaving in the new
interface, these timers were not merged to rate matching and interleaving
timers in nr_dlsch_coding.
The variable block_length had an ambiguous role and was used as both K and Kprime.
This was leading to arrays with wrong sizes and functions with wrong arguments.
Now there is not anymore block_length but K and Kprime.
Kprime can be set with the command line argument `-l`.
There was also some problems with array allocation for segments with length not divisible by 8.
A proper sizing with `((size_in_bits + 7) & ~7) / 8` was used.
A proper masking of the last byte containing payload bits was also added to keep filler bits to 0 after random initialization.
The name ldpc8blocks_coding_segment was previously used to avoid any confusion with ldpc8blocks while both were coexisting.
They are not anymore coexisting so it can now be shortened.
Remove most of m64 type usage
MMX instructions were first SIMD version, the registers management is complex
(_mm_empty()) so, it is better to remove it now and use more recent instructions
anyway, gcc/clang should replace it automatically as long as we enable sse2,
that should always be the case but for code understanding and for ARM porting
for example, it is better to explicitly remove it
fix(nrLDPC_coding_t2): Miscellaneous fixes related to T2
1. HARQ combined offset:
The offset provided for each code block in the HARQ combined input and output
buffer shall not be above 512. A modulus was performed on the offset of the TB
in the buffer to avoid this situation. But the offset of the code block was
added afterward so that the total offset was sometimes going above 512.This
commit performs the modulus after adding the code block offset. Usually this
issue was triggering a segfault after less than 2 minutes of an iperf test on
some setup using T2. The same iperf test was successfully performed during 10
minutes on the same setup with this change.
2. Properly stop LDPC coding in gNB and UE:
LDPC coding was yet not properly stopped by using free_nrLDPC_coding_interface
in gNB and UE. This was not causing major issue until now. It was especially not
causing any issue with T2 LDPC coding using the PF. But using T2 VFs, if the
coding library is not properly stopped then the VF cannot be used anymore unless
the admin app is restarted. So this MR adds the missing free_nrLDPC_coding_interface.
LDPC coding was yet not properly stopped by using free_nrLDPC_coding_interface in gNB and UE.
This was not causing major issue until now. It was especially not causing any issue with T2 LDPC coding using the PF.
But using T2 VFs, if the coding library is not properly stopped then the VF cannot be used anymore unless the admin app is restarted.
So this MR adds the missing free_nrLDPC_coding_interface.
The offset provided for each code block in the HARQ combined input and output buffer shall not be above 512.
A modulus was performed on the offset of the TB in the buffer to avoid this situation.
But the offset of the code block was added afterward so that the total offset was sometimes going above 512.
This commit performs the modulus after adding the code block offset.
Remove spinlock from threadpool task return
The idea to use atomic aligned to cache line for counting finished jobs
is reused, but instead of polling a semaphore was added.
nrLDPC_coding: Fix naming, comments, superfluous variables and documentation typos in coding library interfaces
This MR is clarifying the interfaces to the LDPC coding libraries. It
aims to improve the readability of the interfaces and compliance of
naming with specification 38.212.
