onnxruntime
Optimize FastGelu with float2 and float4 vectorized kernels on ROCm
#11491
Merged
Merged
Optimize FastGelu with float2 and float4 vectorized kernels on ROCm #11491
zhangyaobit
merged
21 commits into
microsoft:master
from
ROCm:hubertlu/fastgelu
Using vectorized loads (float2) for fp16 to improve performance
664bb50e
Fix a few warnings from cpplint
68904e56
Fix a few warnings from cpplint
5dc6cb5c
Use __float2half2_rn and fix some cpplint warnings
64821fb1
Move some computaions to LaunchFastGeluKernel
4e998546
Fix some Lint C++ warning
e8c19264
Using vectorized loads (float4) for fp16 to improve performance
6d819560
Switch whether to optimize FastGelu with float4 vectorization
98686e6a
Switch to float4 memory access based on input_length in FastGelu
b87d0a9e
Merge branch 'hubertlu/fastgelu' of https://github.com/ROCmSoftwarePl…
c1e8b7c4
Comment how to set the threshold of float2 and float4 vectorized kernels
5264e642
hubertlu-tw
changed the title
Merge branch 'master' into hubertlu/fastgelu
4b6a1bb8
Add FastGelu fp16 unit tests for bias_length = 2 and 8
35e2caad
Make vectorized kernels generic with aligned_vector
96d548df
zhangyaobit
commented
on 2022-06-10
Unify the vectorized kernels with/without bias
81a571be
Refactor the code to suppress cpplint warnings
a78464dc
Solve formatting issues
07998575
Remove cudaDeviceProp from FastGeluKernel and LaunchFastGeluKernel
9269e11a
zhangyaobit
dismissed these changes
on 2022-06-20
Move fast_gelu_impl.h to rocm/bert
98d7c3ce
Merge remote-tracking branch 'upstream/master' into hubertlu/fastgelu
085ec9ae
tianleiwu
commented
on 2022-06-23
tianleiwu
commented
on 2022-06-23
tianleiwu
dismissed these changes
on 2022-06-23
Fix some Lint C++ warnings and code alignment
86dad9af
zhangyaobit
approved these changes
on 2022-06-24
tianleiwu
approved these changes
on 2022-06-24
Reviewers
Assignees
No one assigned
Labels
None yet
Milestone
No milestone
Login to write a write a comment.
Login via GitHub