[ neon ] Apply openMP in HGEMM, HGEMV in neon for multithreading #2415
Conversation
skykongkong8
requested review from
myungjoo,
jijoongmoon,
again4you,
jaeyun-jung,
leemgs,
wooksong,
helloahn,
kparichay,
gichan-jang,
anyj0527,
zhoonit and
a team
as code owners
|
📝 TAOS-CI Version: 1.5.20200925. Thank you for submitting PR #2415. 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: 
There was a problem hiding this comment.
@skykongkong8, 💯 All CI checkers are successfully verified. Thanks.
skykongkong8
force-pushed
the
omp_neon
branch
from
71b54b5 to
e7e20a1
Compare
|
|
skykongkong8
force-pushed
the
omp_neon
branch
from
e7e20a1 to
902366d
Compare
|
|
skykongkong8
force-pushed
the
omp_neon
branch
2 times, most recently
from
3b40a41 to
ccfe651
Compare
skykongkong8
changed the title
skykongkong8
force-pushed
the
omp_neon
branch
from
ccfe651 to
2a6609f
Compare
There was a problem hiding this comment.
@skykongkong8, 💯 All CI checkers are successfully verified. Thanks.
skykongkong8
force-pushed
the
omp_neon
branch
2 times, most recently
from
b549840 to
83f65c5
Compare
There was a problem hiding this comment.
@skykongkong8, 💯 All CI checkers are successfully verified. Thanks.
skykongkong8
force-pushed
the
omp_neon
branch
from
83f65c5 to
d88a34b
Compare
skykongkong8
requested review from
songgot,
jihochu,
DonghakPark,
SeoHyungjun,
baek2sm and
EunjuYang
as code owners
There was a problem hiding this comment.
@skykongkong8, 💯 All CI checkers are successfully verified. Thanks.
skykongkong8
force-pushed
the
omp_neon
branch
from
d18cf56 to
32b896f
Compare
There was a problem hiding this comment.
@skykongkong8, 💯 All CI checkers are successfully verified. Thanks.
skykongkong8
force-pushed
the
omp_neon
branch
2 times, most recently
from
63d6e56 to
45a5fbc
Compare
There was a problem hiding this comment.
@skykongkong8, 💯 All CI checkers are successfully verified. Thanks.
test/include/nntrainer_test_util.h
Outdated
| for (int j = 0; j < channel; ++j) { \ | ||
| for (int k = 0; k < height_b; ++k) { \ | ||
| for (int l = 0; l < width_b; ++l) { \ | ||
| float val = equation_i_j_k_l; \ |
There was a problem hiding this comment.
| float val = equation_i_j_k_l; \ | |
| float val = (equation_i_j_k_l); \ |
There was a problem hiding this comment.
what difference would this make? just curious
will do it anyway
skykongkong8
force-pushed
the
omp_neon
branch
from
bba28d9 to
a476037
Compare
There was a problem hiding this comment.
@skykongkong8, 💯 All CI checkers are successfully verified. Thanks.
- get omp thread num option when build, and allocate thread accordingly. - getting system core num will be added in the near future - With precise indexing, we can integrate multithread function and general case function for GEMM and GEMV - For non-8-divisible case, an additional code block will be added, and it is expected to work as the same way for simd, but with value-by-value iteration to avoid segementation fault - For 8-divisible case, we need to save room for last 8-block because: 1. To apply openMP, we need for-loop only iterating with a single indexing variable 2. This will not deteriorate the function latency, but might make the code length longer than before. Need to discuss after. **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: skykongkong8 <[email protected]>
- Previously, there were only few half-precision GEMM, GEMV unittest cases with small dimensions. - In this commit, I would like to add more cases to evaluate current GEMM GEMV performance in practical use. - This commit might be helpful for: - Verification of custom-made GEMM, GEMV functions - Determining OMP_THRESHOLD - User can modify the distribution of Tensor values by modifying MOD and alpha for their own use - Additionally, GEN_TEST_INPUT_B is added. Previous TCs were praticing wrong input data generation when A (op) B = C. **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: skykongkong8 <[email protected]>
- With current version of HGEMV transpose, there is no need to modify how-much-batches-to partial sum **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: skykongkong8 <[email protected]>
- Redundant usage of int instead of unsigned int gives compiler warning. **Self evaluation:** 1. Build test: [X]Passed [ ]Failed [ ]Skipped 2. Run test: [X]Passed [ ]Failed [ ]Skipped Signed-off-by: skykongkong8 <[email protected]>
- For better intuistic code **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
omp_neon
branch
from
a476037 to
8a8b7a9
Compare
There was a problem hiding this comment.
@skykongkong8, 💯 All CI checkers are successfully verified. Thanks.
|
FYI) For GEMM, an updated version is on WIP |
Self evaluation: