Skip to content

PDSCH precoding optimization

Quency Lin requested to merge pdsch-precoding-opti into develop

Based on !2395 (merged): PDSCH clean up.

This MR does:

  • Enable PDSCH precoding (pmi is no longer hardcoded to 0)
  • SIMD maltrix multiplication: Improves PDSCH layer precoding by 320%. We do 4 REs with 128 SIMDe within an iteration and 3 iterations for 1 RB with new function nr_layer_precoder_1RB(). Call the original non SIMD precoder nr_layer_precoder_cm() if the sub carrier offset is crossing ofdm symbol size.
  • Not yet in this MR:
    • 256 SIMD (maybe 2 RB in a row?)
    • Prefetch
    • Cache obilivious (we are doing Y = W*X, since W is not big, cache obilivious might not bring much improvments)

Measurements

On sithonia (x86, ubuntu)

sudo ./nr_dlsim -n200 -s35.8 -S40 -x2 -y4 -z4 -e18 -q1 -R106 -b106 -P -p1
2023.w40

!2395 (merged)

128 SIMD, 1 RB

0e24dc4c

256 SIMD, 2 RB

29dfdebc
Layer Precoding time 848.95 us 605.74 us 144.34 us 81.49 us
Edited by Quency Lin

Merge request reports