mirror of
https://github.com/slsdetectorgroup/aare.git
synced 2026-11-03 00:53:37 +01:00
Ferjaoui
d3a5f842bf
throughput.py measures wall-clock FPS without nsys, for the copy-out (find_clusters_batched) and zero-copy (BatchView) paths, at 3x3 to 9x9 and 1 and 4 streams. On the A100, zero-copy with 4 streams reaches 94-102 % of the bound predicted by the nsys probes.
A100 probe results (Merlin7)
NVIDIA A100 on Merlin7 (gmerlin7, a100 partitions), CUDA 12.9.1, gcc 13.3.0,
nsys 2025.1.3. CUDA event trace is off (--cuda-event-trace=false) unless the
folder name says evtrace-on. Driver and git state of each run are in its
env.json.
| folder | precision | copy-out rounding | frames / batch | configs |
|---|---|---|---|---|
2026-10-28_f32_lround_evtrace-on |
f32 | lround |
20 000 / 2000 | 3x3, 9x9 at s1, s4 |
2026-10-28_f32_lround |
f32 | lround |
20 000 / 2000 | 3x3, 9x9 at s1, s4 |
2026-10-28_f64_lround |
f64 | lround |
20 000 / 2000 | 3x3, 9x9 at s1, s4 |
2026-10-29_f32_round_b2000_f20K |
f32 | round |
20 000 / 2000 | 3x3 to 9x9 at s1, s4, s8 |
2026-10-29_f32_round_b3000_f50K |
f32 | round |
50 000 / 3000 | 3x3 to 9x9 at s1, s4, s8 |
2026-10-29_f32_round_throughput |
f32 | round |
20 000 / 2000 | unprofiled wall clock (throughput.py), 3x3 to 9x9 at s1, s4 |
2026-10-29_f32_round_throughput_5x5 |
f32 | round |
20 000 / 2000 | 5x5 rerun, 5 passes (first run disturbed by other jobs on the node) |
lround(float) is emulated in software on the device and made the f32 9x9
kernel 1.8x slower than f64. It was replaced by round in 104e284. The
2026-10-28 runs used local, uncommitted edits to clusterfinder_kernel.cuh
(the precision aliases), so env.json shows git_dirty: true and does not
record the rounding.
Event trace on and off give the same kernel times; with it on, the idle gaps between steps are 10-16 % wider.