[ tensor ] Apply SIMD in matrix transpose #2603
Conversation
…matrix transpose - Previously, matrix transpose was relying on naive for-loop implementaion. - Using SIMD instructions, there is a room to be latency-optimized. - Note that current implementation only supports half-precision matrices. **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: skykongkong8 <[email protected]>
- Add new function "transpos_matrix" to use newly implemented matrix transpose code **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: skykongkong8 <[email protected]>
- If it is for height-width transpose, we can enjoy SIMD accelerated code. - Use SIMD version if possible, otherwise fallback. - Through this commit, followings are expected to be accelerated, or can be accelerated with ease in the near future: - "0:2:1" transpose - BiQHGEMM - HGEMM **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: skykongkong8 <[email protected]>
skykongkong8
requested review from
myungjoo,
jijoongmoon,
again4you,
jaeyun-jung,
leemgs,
wooksong,
helloahn,
kparichay,
gichan-jang,
anyj0527,
zhoonit,
lhs8928,
songgot,
jihochu,
DonghakPark,
SeoHyungjun,
baek2sm,
djeong20,
EunjuYang and
a team
as code owners
|
📝 TAOS-CI Version: 1.5.20200925. Thank you for submitting PR #2603. Please a submit 1commit/1PR (one commit per one PR) policy to get comments quickly from reviewers. Your PR must pass all verificiation processes of cibot before starting a review process from reviewers. If you are new member to join this project, please read manuals in documentation folder and wiki page. In order to monitor a progress status of your PR in more detail, visit http://ci.nnstreamer.ai/. |
|
|
cibot: 
- Previously, there was a code defect when transposing matrix with non-4-divisible col length. - Bugfix and refactor its using interface: move transpose fallback when NEON is supported. **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: skykongkong8 <[email protected]>
skykongkong8
force-pushed
the
pr/transpose/without_wait_for
branch
from
433ec98 to
4efa98b
Compare
|
|
skykongkong8
force-pushed
the
pr/transpose/without_wait_for
branch
2 times, most recently
from
41f6812 to
bc598c4
Compare
There was a problem hiding this comment.
@skykongkong8, 💯 All CI checkers are successfully verified. Thanks.
nntrainer/tensor/matrix_transpose_neon/matrix_transpose_kernels_neon.h
Outdated
Show resolved
Hide resolved
skykongkong8
force-pushed
the
pr/transpose/without_wait_for
branch
from
bc598c4 to
c42beca
Compare
There was a problem hiding this comment.
@skykongkong8, 💯 All CI checkers are successfully verified. Thanks.
skykongkong8
force-pushed
the
pr/transpose/without_wait_for
branch
from
c42beca to
c7daba7
Compare
- add doxygen tags to avoid CI fail - trivial formatting **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: skykongkong8 <[email protected]>
skykongkong8
force-pushed
the
pr/transpose/without_wait_for
branch
from
c7daba7 to
845c7d8
Compare
There was a problem hiding this comment.
@skykongkong8, 💯 All CI checkers are successfully verified. Thanks.
PR for issue raised in #2582
Matrix Transpose function in the latest NNTrainer (14.05.24) is implemented using for-loops.
Although current implementation is useful for general use in (b,c,h,w)-Tensor transpose, it would be a little bit naive implementation for the (h,w)-matrix transpose.
Latency measurement
transpose("0:2:1"))