llama.cpp
hexagon: Op Fusion refactoring, optimizations and new fusions like `BINARY_NX` (back-to-back ADDs, etc)
#30843
Merged
Merged
hexagon: Op Fusion refactoring, optimizations and new fusions like `BINARY_NX` (back-to-back ADDs, etc) #30843
max-krasnyansky
merged
35 commits into
ggml-org:master
from
qualcomm:hexagon-opfusion-v2
hex-fusion: support for GELU+MUL, REPEAT+GET_ROWS and BINARY_NX (back…
d76b2445
hex-fusion: factor out fusion logic into separate .h and .cpp and upd…
1b313736
hex-fusion: allow for RMS_NORM+MUL fusions for non-contig tensors
129ba38a
hex-fusion: streamline fusion eligibility checks and cover missing co…
9b54fd09
hex-fusion: streamline fusion flow
041ecb0a
hex-opbatch: make n_tensors limit more scalable
c8e0fcd0
hex-bufs: adjust the trace buff sizing
56055c52
hex-glu: cleanup glu-ops interfaces and add kernel_params
8d6c79bb
hex-kparams: clean up kparams handling in unary, ssm-conv, and other ops
4a719b6e
hex-kparams: update the rest of the ops PAD,DIAG,... to use kparams
d71f26cf
hex-unary: add support for collapsing dims
a5c90a80
hex-ops: cleanup redundant runtime checks
c0839381
hex-ops: consistent use of kernel_type in all kparams
f6ca0e8e
hex-fusion: remove redundant check in gdn+cpy
95e2081f
hex-fusion: further streamlining of fusion rules and code
316b708b
hex-fusion: cleanup naming
6a20c6e0
hex-opfmt: cleanup op formatting
5fe333d5
hex-fusion: bump fusion lookahead to 64
b8578e59
hex-fusion: fuse SSM_CONV+SILU
c5ae1e3c
hex-fusion: fix hvx register spills in fused ssm_conv+silu
d9458ada
hex-fusion: add more GLU fusions
0c247c5d
hex-fusion: remove duplicate checks and use ggml-common helpers
5bc4f50c
hex-kparams: refactor kparams into htp-kparams.h/cpp
8765b1ab
hex-fusion: fix binary-nx fusion regression
b69a6e54
hex-kparams: another pass at cleaning up kparams (removing unused, etc)
545e3982
hex-kparams: streamling nameing and move _check_ wrappers into .h
7b21c101
hex-kparams: templetize htp_kparams_check_<op> to reduce code bloat
fc23deb0
hex-ops: de-duplicate op dispatch switches
60af418c
hex-kparams: trim down exposed kparams calls
d76387cf
hex-kparams: simplify type casting of kparams structs
28e3ab98
hex-fence: de-dup fence tensor construction
7d962db0
hex-opbatch: simplify tensor shape init/checks
be7513f0
hex-mm: cleanup leftovers of the original QK defines
f8aee440
hex-fa: fix stack overflow on older npu firmware
326800fc
hex-ctx: move dma_caches into ops_ctx
05e91b46
lhez
approved these changes
on 2026-10-22
jhen0409
approved these changes
on 2026-10-23
Login to write a write a comment.
Login via GitHub