Skip to content

Register-tile sweep (#643)

Which micro-kernel register tile (mr × nr) the native GEMM uses on each ISA, and the sweeps that chose it. Each machine’s hardware and pinning policy are on the benchmark systems page.

Ryzen 9 8945HS: no tile clears the noise, so AVX-512 keeps 6×16

Section titled “Ryzen 9 8945HS: no tile clears the noise, so AVX-512 keeps 6×16”

bash benchmarks/machines/ryzen-9-8945hs-tile.sh under WSL, fp64, 5 rotated rounds, data in benchmarks/data/ryzen-9-8945hs-tile-double/. Only mr changes; nr = 16, kc = 128, mc = 128 and nc = 8192 are identical across the arms. Geometric-mean ratio of the median against default (mr = 6):

TileAccumulators (of 32)T=1T=8
8×16160.9951.046
10×16200.9961.036
12×16241.0321.021
14×16281.0020.994

The default arm’s round-to-round spread was 5–41%, as expected inside a VM (Windows places the virtual CPUs). 12×16 is nominally ahead at both thread counts (×1.032 and ×1.021), but neither gain, nor any other tile’s, clears that spread, so AVX-512 builds keep the AVX2 description. That is a measured reason to keep it, not a measured loss for a larger tile. A bare-metal Linux run would be the way to settle it.

Arduino Ventuno Q: 8×4 wins, and 12×4 loses at four threads

Section titled “Arduino Ventuno Q: 8×4 wins, and 12×4 loses at four threads”

benchmarks/machines/arduino-ventuno-q-tile.sh, schedutil, 5 rotated rounds, data in benchmarks/data/arduino-ventuno-q-schedutil-tile-{double,float}/. Every arm is the same source with only mr changed; nr, kc, mc and nc are identical across the arms (fp64 nr = 4, kc = 512, mc = 32; fp32 nr = 8, kc = 512, mc = 64), and the checksums agree across all arms. The ratios are median GFLOP/s against default (mr = 6 in both runs, which predate the change), as a geometric mean over the shapes:

Tile (fp64 / fp32)Accumulators (of 32)fp64 T=1fp64 T=4fp32 T=1fp32 T=4Worst shape (fp64 / fp32)
8×4 / 8×8161.1741.1021.1701.2191.060 / 1.156
10×4 / 10×8201.1561.0451.1711.1600.967 / 1.064
12×4 / 12×8241.0610.9301.0631.2060.840 / 1.048
14×4 / 14×8281.0490.8781.0301.0280.736 / 0.930
  • The peak is at half the register file, not at the 3/4 that #351 measured on AVX2. 12×4 is what the 3/4 rule would give on 32 registers, and it loses.
  • fp32 agrees on 8. 8×8 is best at T=4 on every shape and ties 10×8 at T=1. fp32 has no 12-row loss: its wide shapes are 128 rows (mc = 64), so each thread’s 32-row block takes 12-row panels at 36 rows of work, not 24 for 16.
  • The T=4 losses are a divisibility effect, not register pressure. They fall on the wide 64-row shapes, where each of the four threads gets a 16-row block. 8-row panels tile it exactly; 12-row panels cover it with 24 rows of work and 14-row panels with 28. It is the #408 family again: a register-tile change that moves how a block divides.
  • Shipped as a result: aarch64 builds now derive 8×4 fp64 and 8×8 fp32 (target_hw_traits, 32 registers with a measured budget of 16 accumulators). In these sessions that took one A78C from 8.3 to 9.7 GFLOP/s fp64, 57% of the 16.9 GFLOP/s peak at 2112 MHz, and from 17.1 to 19.9 GFLOP/s fp32, 59% of 33.8. Four cores went from 35.5 to 37.6 GFLOP/s fp64 and from 61.9 to 77.2 GFLOP/s fp32 at 2048³.
  • Noise. The clocks were unpinned, so the default arm spread 12–14% between rounds at T=1 (2–3% at T=4 on the squares). The 8×4 gain is consistent across every shape at both thread counts, which noise of that size would not produce. The absolute numbers are lower than the Ventuno’s cache-blocking A/B session (8.3 against 9.1 GFLOP/s for the same 6×4 tile, both under schedutil). This data does not say why; compare ratios within a session, not GFLOP/s across them.

Not measured, and covered by the same change: the Jetson’s A78AE and every other aarch64 core (Apple M-series, the Neoverse cores of the CI runners). The tile is a compile-time choice per ISA, so all of them moved together on one board’s fp64 data.