- Ensure packed struct through bitfield
- decoder tree buffer increase: i don't find evidence in the 3GPP
standard of the max size of polar coded PDU. even with this max size,
the way this tree is built is not trivial. so, it is hard to predict,
i just put 8 system pages (32KB) that looks enough [1]
- a lot of reformatting
[1] https://arxiv.org/pdf/1804.04389
according to this paper, the largest polar code in 5G is in UL, so
likely the biggest problem will be in gNB, not the UE. in DL, the max
output coded block is 512, in UL 1024. probably the tree we discuss is
related to this size.
"frozen bits" means only bits we don't encode/decode because polar code
size is only by power of 2.
We see this in OAI code: NR_POLAR_PUCCH_N_MAX is 10, so 2^10 is 1024,
that is the maximum size for 5G polar code length. in the code, we
assert if N > 512: "N = %d, not done yet\n",
polarParams->N". so, the today OAI code can't do PUCCH large blocks.
using current oai limitation of 2⁹ max code size, we can create up to
2^9+2⁸... nodes, so 1023 nodes in the tree. the problem then is that
each node have variable size, depending on the 'level' (the power of 2
we are currently). the algo provision large alpha and beta to be able
to store all lower levels in the tree (int leaf_sz = 1 << level)
therefore upper tree nodes are large (the top level for n=9 uses about
1.5KB) and the bottom nodes uses 64 bytes. if all nodes are populated,
the size will be very large: 1.5KB + 2 * 800 + .... but, we create
nodes in the tree only for the "information bits", so the payload, that
is limited to 128 bits (else we assert, the code is not "finished").
1.5KB * 128 is >> 8KB i allocated. but, we can't create more than one
mode at top level, so there is maximum 1.5KB + 2 * 800, + 4 * 400 + 8 *
200 + 16 * 100 + remaining of 128-32 nodes of 64 bytes. about 14KB.
i allocated 1024*sizeof(__m256i) bytes = 32KB that looks good.
- Reworked nrLDPC_decoder_offload.c to process all segments in a TB with a single call to LDPCdecoder() or LDPCencoder().
- Perform CRC checks on the T2.
- Modified nr_ulsch_decoding (gNB decoder) to:
- Count processed segments.
- Properly perform CRC checks.
- Created a constant NR_LDPC_MAX_NUM_CB to define the maximum number of codeblocks.
- Conditional library loading:
- On gNB: Load ldpc_t2 library if offload is enabled using the --ldpc-offload-enable flag.
- On UE: Load both ldpc_t2 and ldpc libraries since only encoder offload is supported.
- General cleanup of nrLDPC_decoder_offload.c for improved readability and maintainability.
- Modified the structure of LDPC encoder/decoder/offload parameters:
- Introduced a structure for code block (CB) related parameters.
- Removed parameter E from the encoder and offload parameter structures for clarity.
- Replaced E with perCB->E_cb in the UE encoder code.
Testing on Neoverse N1 (Ampere 3GHz). Tested only on a subset of phy-simulators.
- Changes:
- use SIMDE consistently
- adaptations of LDPC decoder generator for ARMv8 performance
- SIMDe modifications of Intel CRC to allow for aarch64 build.
optimizations for 128-bit to improve performance of LDPC encode/decode
on aarch64 (Neoverse 1)
- added BG2 files for 128-bit ldpc encoder (aarch64)
- testing on Xeon
- testing on x86
- minor changes to build/run on x86
- change in crc.h after returning to aarch64
- removed some warning in ldpc decoder generator for x86_64
- char
- Delete irrelevant constants
- Correctly declare variables
- Define stdbool for all architectures
- Remove definition of _MM_SHUFFLE and use SIMDE_MM_SHUFFLE
- Remove commented code
- Fix CMakeLists.txt
- Include SIMDE avx2 functions in tools_defs.h
Romain Beurdouche