llama.cpp
hexagon: further improved pipeline of the core bits (L2, DMA, MM, FA)
#26049
Merged

hexagon: further improved pipeline of the core bits (L2, DMA, MM, FA) #26049

max-krasnyansky
max-krasnyansky hex-l2: use dirty ranges for flushing
7c9e3c09
max-krasnyansky hex-l2: simplify range based flush logic
e72c4428
max-krasnyansky hex-l2: optimize dirty range scans
a8b4dea4
max-krasnyansky hex-hvx: support for reduce_max_i32
7cb851f3
max-krasnyansky hex-mm: optimize fused MUL_MAT+ADD to use vtcm for bias when it fits
c3875e43
max-krasnyansky hex-mmid: optimize mmid row-mapping generation
89d9991d
max-krasnyansky hex-mmid: optimize mmid row-mapping generation
fb1277db
max-krasnyansky hex-mmid: optimize mmid row-mapping generation (round2)
7e25ad31
max-krasnyansky hmx-mm: optimize output proc by tiling (col-chunking)
dfe66bdd
max-krasnyansky hex-fa: start the next q dmas a bit earlier
a969e8a8
max-krasnyansky hex-fa: prefetch Q even earlier
2b0d7a12
max-krasnyansky hvx-fa: optimize softmax to keep things in hvx registers
63816c2e
max-krasnyansky hex-fa: hoist const register init in softmax loop
6d488715
max-krasnyansky hmx-fa: kick off next-qkv DMAs before o-proc
5695533e
max-krasnyansky hmx-fa: hoist various checks out of the inner loop
5e634998
max-krasnyansky hmx-fa: adjust the cost model to better balance softmax work across h…
107fd636
max-krasnyansky hmx-fa: overlap diag rescale build with last HMX task
beb3c59a
max-krasnyansky hmx-fa: optimize idx update in output proc
39bee20e
max-krasnyansky hmx-fa: unroll the softmax loops for improved perf
e30f5147
max-krasnyansky hmx-fa: overlap qk-dot with softmax, double-buffer p and s tiles
d84fc3c7
max-krasnyansky hex-trace: double the default number of trace entries
6270178d
max-krasnyansky hex-trace: add trace events for opbatch and buffer mgmt
a99ba773
max-krasnyansky hex-trace: overhaul tracing to simplify runtime event handling and su…
a9416965
max-krasnyansky hex-trace: replace ascii timeline diagram with pipeline bubbles detector
1588334a
max-krasnyansky hex-trace: handle missing start/stop events
083755ab
max-krasnyansky hex-dma: always log stop/start trace events even for dummy dmas
9764cc7e
github-actions github-actions added ggml
github-actions github-actions added Hexagon
max-krasnyansky hex-scripts: fix flake warnings
caea4801
max-krasnyansky max-krasnyansky marked this pull request as ready for review 117 days ago
max-krasnyansky max-krasnyansky requested a review 117 days ago
max-krasnyansky
lhez
lhez approved these changes on 2026-07-23
max-krasnyansky
yomaytk
yomaytk approved these changes on 2026-07-24
max-krasnyansky max-krasnyansky merged 0a50d990 into master 117 days ago
max-krasnyansky max-krasnyansky deleted the hexagon-pipeline-v3 branch 117 days ago

Login to write a write a comment.

Login via GitHub

Reviewers
Assignees
No one assigned
Labels
Milestone