hexagon: new HMX-optimized GATED_DELTA_NET #29199
max-krasnyansky
merged
25 commits into
ggml-org:master
from
qualcomm:hexagon-gdn-hmx
hex-gdn: start putting together HMX support for GDN
69ad3710
hex-gdn: working hmx but not-pipelined and slow for now
09cef94c
hex-gdn: re-write vtcm layout handling and prep for pipelining
2b2c5c08
hex-gdn: starting to pipeline hmx and dmas
55f4b630
hex-gdn: add hvx threading for most pipeline stages
db7ed1e3
hex-gdb: add detailed trace events
cd5823f1
hex-gdn: vectorize expfs and use aligned hvx reads/writes
779b0e3d
hex-gnd: vectorize the rest of expf
b040435e
hex-gdn: optimize tail processing (pad partial chunks)
05c43eac
hex-gdb: avoid float up/down casts in hot loops
3eb6751c
hex-fa: remove float up/down casts from inner loops
26c8aabf
hex-gdn: do exp() in f16 to improve HVX utilization
268f3874
hex-gdn: optimize tiler
e9e9ef7d
hex-hmx: bump hmx-queue to 128 and dispatch all GDN gemms at once
9afd33ed
hex-gdn: further pipeline improvements
667073fe
hex-gdn: optimize gdn prep stage
ad56f1fa
hex-gdn: yet more tweaks to optimize GND_SOLVE task and pipeline
2f8412e5
hex-gdn: improve accuracy and optmize gdn-prep further
7d4b7c4d
hex-gdn: fix rebase conflict
44e9756f
hex-bufs: revert max_bufsize enforcement, it is enough to just enforc…
41b3393b
hex-scripts: improved inspect script to avoid false alarms in reg spi…
d85be32a
hex-fa: improve inline softmax with in-reg VKQ32 accum
7339763a
hex-fa: minor improvement for dma pipeline in hvx kernel
303ae1ca
hex-fa: reduce ddr reads by 20-30% during token gen
68160e57
hex-gdn: proper alignment for hvx vtcm spads
bc8bed78
CISC
approved these changes
on 2026-09-21
lhez
approved these changes
on 2026-09-21
Login to write a write a comment.
Login via GitHub