hexagon: flash attention rework (optimizations, accuracy improvements, etc) - #25085

Merged

max-krasnyansky merged 48 commits into

ggml-org:masterfrom

qualcomm:hexagon-fa-rework

Jul 1, 2026

Commits on Jul 1, 2026

hex-mm: fold mm quant tasks into the main matmul threads
max-krasnyansky
committed
93879d1
hex-mm: minor formatting fixes
max-krasnyansky
committed
0353a74
hex-mm: cleanup is_quant checks in dma dispatch
max-krasnyansky
committed
56d55b2
hex-mm: fix dst-spad alignment
max-krasnyansky
committed
692951b
hex-mm: move fp kernels in the hvx-mm-kernels header
max-krasnyansky
committed
ba85f82
hex-mm: fuse with ADD
max-krasnyansky
committed
ab39f62
hex-fa: factor out ukernels into separate headers and unify the rest
max-krasnyansky
committed
919c5e9
hex-fa: move kernel-params compute into the host
max-krasnyansky
committed
4748d68
hex-fa: refactor vtcm alloc for consistency
max-krasnyansky
committed
0c5324d
hex-fa: add support for FA_SELECT
max-krasnyansky
committed
50d97ef
hex-fa: update tracing insrumentation to cover all functions
max-krasnyansky
committed
c5c33a8
hex-fa: update hvx fallback thresholds to recover t/g regressions
max-krasnyansky
committed
a2a8f1e
hex-fa: update tracing instrumentation
max-krasnyansky
committed
baaaf59
hex-fa: improved tracing with additional events
max-krasnyansky
committed
beaa2a0
hex-fa: optimize mask processing (fastdiv, etc)
max-krasnyansky
committed
950b44a
hex-fa: improve mask dma caching
max-krasnyansky
committed
39fc297
hmx-fa: change loop order to maximize mask cache hits
max-krasnyansky
committed
1a1bd4a
hex-fa: remove over instrumentation
max-krasnyansky
committed
022744f
hex-fa: breakdown QKV prep trace events
max-krasnyansky
committed
cf6060b
hmx-fa: further mask proc optimizations
max-krasnyansky
committed
998367c
hex-fa: mask broadcast is the common case, optimize for that
max-krasnyansky
committed
49ff8fd
hex-fa: use aligned loads where possible
max-krasnyansky
committed
f27a3b9
hex-fa: update loops to use uint32_t indices
max-krasnyansky
committed
ca0a185
hmx-fa: fold vtcm init into q prep task
max-krasnyansky
committed
b9529ef
hex-fa: update rest of the hmx funcs to use uint32_t
max-krasnyansky
committed
c8e1889
hmx-fa: fold build_d into the main softmax loop
max-krasnyansky
committed
3e38899
hmx-fa: start kv dmas earlier
max-krasnyansky
committed
ade6dd4
hmx-fa: start mask dma a bit earlier
max-krasnyansky
committed
c3497dd
hex-fa: precompute rows per task to avoid divs
max-krasnyansky
committed
a8c0be3
hmx-fa: specialize fa_o_store for f16 and f32
max-krasnyansky
committed
cb850e9
hmx-fa: prelim support for Sinks
max-krasnyansky
committed
12baed1
hmx-fa: keep softmax accumulators in fp32
max-krasnyansky
committed
7ef475c
hex-fa: add tanh_f16 and exp2_f16 and use that in FA
max-krasnyansky
committed
83dcc57
hex-fa: use fp16 math in the hvx kernel
max-krasnyansky
committed
a8c6152
hex-fa: avoid expensive float -> __fp16 cast for slopes and softcap
max-krasnyansky
committed
272637b
hex-fa: replace most vec_exp_f32 with vec_exp2_f16
max-krasnyansky
committed
4528f54
hmx-fa: vectorize sinks update
max-krasnyansky
committed
5251c93
hex-fa: minor formatting
max-krasnyansky
committed
b782567
hmx-fa: fold softcap loop into the tile load
max-krasnyansky
committed
32bf5a5
hmx-fa: use vectoralias to populate sinks
max-krasnyansky
committed
54dea95
hex-fa: remove redudant check
max-krasnyansky
committed
eb59d19
hex-fa: fix vtcm size compute to use fp32 for accumulators
max-krasnyansky
committed
adfdad4
hex-mm: fix trailing spaces
max-krasnyansky
committed
5120024
hmx-fa: dont use -inf to init mask to avoid conversion overflows
max-krasnyansky
committed
1a07912
hex-fa: no need to explicitly guard -inf in the f16->f32 converter now
max-krasnyansky
committed
61ce103
hmx-fa: cleanup fa sinks handling
max-krasnyansky
committed
bc6d133
hex-mm: fixed src2 stride handling when mm is fused with add
max-krasnyansky
committed
eb03f58
hex-fa: make lto happy
max-krasnyansky
committed
65a315b