Feature #4585
closedARM NEON optimizied TS 05.03 encoding/decoding
100%
Description
We'd like to implement a NEON optimized version of our TS 05.03 convolutional encoding/decoding routines.
libosmocore already has generic C and SSE/AVX optimized versions today (see source:src/conv_acc_sse.c, source:src/conv/acc_sse_avx.c, source:src/conv_acc_sse_impl.h). Ideally the same infrastructure is re-used to introduce the ARM NEON optimized version.
(assigning to myself until we have figured out who and how).
Side note: There may also be other parts beyond the convolutional code that could benefit from NEON, such as the interleaving/deinterleaving, burst mapping. It may be worth looking at that afterwards.
Files
Updated by laforge almost 4 years ago
Updated by laforge almost 4 years ago
- Assignee changed from laforge to Hoernchen
Updated by Hoernchen over 3 years ago
- File neonlib_vs_neontest.log neonlib_vs_neontest.log added
- File standardlib_vs_neontest.log standardlib_vs_neontest.log added
- % Done changed from 0 to 100
Patch is in https://gerrit.osmocom.org/c/libosmocore/+/19372
Requires building libosmocore with --enable-neon configure flag
Benchmark code lives at https://github.com/Hoernchen/osmo-conv-test
Attached are two benchmark runs, original generic libosmosdr vs benchmark with neon, as well as modified libosmocore that used the neon code vs benchmark with the same neon code, to ensure performance actually matches the benchmark.
Speedup is 1.3-1.5 which is quite reasonable considering SSE only manages ~2-3, n2k5 needs a special case because generic is faster than NEON for very short RACH with len ~14 (other runs are len > 100).
Updated by fixeria over 3 years ago
- Status changed from New to Resolved
The patch has been merged today, and caused massive [master] build failures on rpi4-deb9build-ansible:
/usr/bin/ld: ../src/.libs/libosmocore.so: undefined reference to `osmo_conv_neon_metrics_k5_n2' /usr/bin/ld: ../src/.libs/libosmocore.so: undefined reference to `osmo_conv_neon_vdec_free' /usr/bin/ld: ../src/.libs/libosmocore.so: undefined reference to `osmo_conv_neon_metrics_k7_n3' /usr/bin/ld: ../src/.libs/libosmocore.so: undefined reference to `osmo_conv_neon_metrics_k5_n3' /usr/bin/ld: ../src/.libs/libosmocore.so: undefined reference to `osmo_conv_neon_metrics_k7_n2' /usr/bin/ld: ../src/.libs/libosmocore.so: undefined reference to `osmo_conv_neon_metrics_k5_n4' /usr/bin/ld: ../src/.libs/libosmocore.so: undefined reference to `osmo_conv_neon_metrics_k7_n4' /usr/bin/ld: ../src/.libs/libosmocore.so: undefined reference to `osmo_conv_neon_vdec_malloc' collect2: error: ld returned 1 exit status
I found and fixed the problem:
https://gerrit.osmocom.org/c/libosmocore/+/19545 configure.ac: fix: do not define HAVE_NEON unconditionally
and submitted a few more cosmetic / misc changes:
https://gerrit.osmocom.org/c/libosmocore/+/19543 src/Makefile.am: add conv_acc_neon_impl.h to EXTRA_DIST
https://gerrit.osmocom.org/c/libosmocore/+/19544 configure.ac: clarify description of --enable-neon
https://gerrit.osmocom.org/c/libosmocore/+/19546 configure.ac: print ARM NEON instructions support status
Tested on my ARMv6 based RPi. Closing this ticket.
