llama.cpp
hexagon: tensor repacking rework and support for tiled and flat quantized and fp weights
#30635
Merged
Merged
hexagon: tensor repacking rework and support for tiled and flat quantized and fp weights #30635
max-krasnyansky
merged
35 commits into
ggml-org:master
from
qualcomm:hexagon-repack-rework
hex-repack: Remove need_repack hack and streamline the rules -- break…
aa92ea14
hex-get-rows: vectorize raw k-quant ukernels
6f70a90f
hex-binary: refactor dma-loop thread functions (no need for macros, etc)
49e65ed9
hex-quant: move/cleanup flat and tiled dequant ukernels
cd303615
hex-repack: add support for repacking/tiling F16/F32/BF16 2D weights
1a922c9c
hex-get-rows: remove redundant cast
b9f56936
hex-repack: proper filtering for BF16 that were leaking through durin…
44221add
hex-repack: streamling tiling rules (broken MM for now)
1400ad4f
hex-repack: fix mmid after tiling updates
5e5159b0
hex-repack: rename repacked to tiled in all flags/funcs/kernels/vars
f6545e01
hex-mm: fix valid_rows calc in various kernels
6e224332
hex-mm: remove unused hvx_mm_nx_2d_flat
1d3fa125
hex-mm: cleanup and optimize hvx-tiled vec_dot_f16
01e6ee0c
hex-mm: another pass at hvx-tiled f16 matmul
d38adec4
hex-get-rows: optimize small src1
a28c2166
hex-get-rows: one more pass at recovering get-rows perf regression
cac8a014
hex-repack: export GGML_HEXAGON_RT_REPACK via run.py
a731bd69
hex-mm-tiled: fix accuracy regression with hmx-tiled-f16 matmul
f44d502d
hex-mm: fold tiled_f16 checks into the general flow
1cdf7285
hex-get-rows: remove dead checks for buffer_is_hexagon and tiling dur…
987896dc
hex-quant: proper tail handling in new dequant ukernels
b214e380
hex-get-rows: remove unused code and cleanup support checks
667eb117
hex-mm: cleanup ne1 padding/tiling checks
3ce6c1a8
hex-mm: cleanup fusion checks for tiled tensors
d8d851d8
hex-repack: add a little can_repack helper to reduce duplication
e5f0ef89
hex-repack: never repack host buffers
44624d46
hex-repack: factor out repack functions into separate .h and .cpp
badf5e7e
hex-tensor: cleanup tensor ifaces for tiled/padded sizes
bd2c66db
hex-repack: update k-quant tiling comments and cleanup #defines
17228e3f
hex-dcache: factor out l2/dcache dirty tracking and flushing
22fc2e05
hex-tensor: update docs and remove #defs duplication
9e12ad10
hex-tensor: remove unused/incorrect get_data interface
f8fb610b
hex-get-rows: cleanup vtcm layout build
be1bb697
tests: disable huge MM tests
486059c2
lhez
approved these changes
on 2026-10-16
hex-dcache: add missing header
bffd6f8a
jhen0409
approved these changes
on 2026-10-16
Login to write a write a comment.
Login via GitHub