llama.cpp
ggml : use dynamic thread scheduling for matrix multiplication
#6915
Merged
Merged
ggml : use dynamic thread scheduling for matrix multiplication #6915
slaren
merged
27 commits into
ggml-org:master
from
kunnis:MMThreadingPerfChange
Just reordering some structs.
3024fd6b
Adding in the calls to mm_pause
5978b6eb
Passing around the state
e098171a
Renaming and moving a bunch of variables around.
a968553c
Extracting the logic to it's own function.
7b932e49
Moving some variable definitions into the chunk function.
4f95478e
Moving some variables around
086e5a82
moving src1_cont inside
209922f5
Moving row_size
bb1b1d00
adding the current_chunk
daa87b18
Reorg the code.
700c782d
Formatting to match the orig patch
891d5837
starting to setup the chunking variables
9acaec58
Starting the buildup of the loop
c0557fa2
The yield shouldn't be necessary.
4762d79d
adding the looping structure based on the chunk configuration.
fc7dc515
Add in the re-chunking code.
807c8252
Making it much more likely to rechunk.
974e43be
disable resizing if numa is enabled.
1c68ea8d
Updating comments with what we've learned.
bd80601e
Fix formatting
d9ba30a2
Couple more formatting fixes.
163dbfdd
More style fixes.
6b0c90fc
Fix Warnings
741a1981
Going with unused because there's conditional logic that needs it.
2dd9f017
slaren
changed the title
Update ggml.c
f2aabab4
Update ggml.c
14c104d1
ggerganov
approved these changes
on 2024-05-15
Login to write a write a comment.
Login via GitHub