[ARM][Performance] Improve ARM CPU depthwise convolution performance #2345
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Member
Author
|
Hi @merrymercy, could you spend some time reviewing? Thanks. |
merrymercy
reviewed
Jan 9, 2019
nnvm/src/top/nn/convolution.cc
Outdated
| << "output channels must divide group size"; | ||
|
|
||
| TShape wshape({param.channels / param.groups, | ||
| // Restore depthwise conv2d kernel layout |
Member
There was a problem hiding this comment.
I think the old code is strange and incorrect while your code is straightforward and correct. We can delete this comment.
merrymercy
reviewed
Jan 9, 2019
| # Currently, Mali schedule doesn't use it like conv2d. | ||
| if cfg.is_fallback: | ||
| ref_log = autotvm.tophub.load_reference_log('arm_cpu', 'rk3399', 'depthwise_conv2d_nchw', | ||
| 'direct') |
Member
There was a problem hiding this comment.
Suggested change
| 'direct') | |
| 'contrib_spatial_pack') |
merrymercy
requested changes
Jan 9, 2019
Member
There was a problem hiding this comment.
Overall looks good. But recently we updated the alter_op_layout to support relay (#2356, a new argument 'F' in introduced). Please resolve the conflict.
FrozenGene
force-pushed
the
arm_cpu_depthwise_convolution
branch
from
f26cdfd to
4cfc239
Compare
FrozenGene
requested review from
Huyuwei,
Laurawly,
nhynes and
phisiart
as code owners
FrozenGene
force-pushed
the
arm_cpu_depthwise_convolution
branch
from
4cfc239 to
1467f34
Compare
Member
Author
|
@merrymercy Have modified code as you suggest. Please review it again. |
FrozenGene
force-pushed
the
arm_cpu_depthwise_convolution
branch
3 times, most recently
from
6fccb6d to
ff30450
Compare
FrozenGene
force-pushed
the
arm_cpu_depthwise_convolution
branch
from
ff30450 to
0e86fe0
Compare
merrymercy
approved these changes
Jan 11, 2019
Closed
wweic
pushed a commit
to neo-ai/tvm
that referenced
this pull request
Feb 20, 2019
…pache#2345) * Add sptialpack schedule for arm cpu depthwise convolution * Supply comments.
wweic
pushed a commit
to neo-ai/tvm
that referenced
this pull request
Feb 20, 2019
…pache#2345) * Add sptialpack schedule for arm cpu depthwise convolution * Supply comments.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
original discussion about this PR:
#2028
To leverage existing spatial pack schedule and add tunable compute_at knob to re-implement ARM CPU's depthwise convolution. Currently, this PR named this schedule as
contrib_spatial_packas discussed in the original PR #2028.On my A53@2.0GHz ARM CPU (MTK6763), which can boost 1.6X performance compared with previous depthwise convolution in the Mobilenet V1 model (I have also checked the correctness of this schedule). However, @yzhliu has also proved that this PR can also boost the performance on x86 CPU.
The following is the Tensorflow Mobilenet V1 model auto tvm training GFLOPS log:
Currently:
[Task 2/20] Current/Best: 0.98/ 2.32 GFLOPS | Progress: (1427/2000) | 2679.82 s Done.
[Task 4/20] Current/Best: 0.56/ 1.15 GFLOPS | Progress: (1072/2000) | 2461.27 s Done.
[Task 6/20] Current/Best: 1.08/ 2.78 GFLOPS | Progress: (1084/2000) | 1987.91 s Done.
[Task 8/20] Current/Best: 0.39/ 1.19 GFLOPS | Progress: (1815/2000) | 2744.70 s Done.
[Task 10/20] Current/Best: 1.09/ 2.33 GFLOPS | Progress: (1222/2000) | 1866.02 s Done.
[Task 12/20] Current/Best: 0.42/ 0.90 GFLOPS | Progress: (1716/2000) | 2528.94 s Done.
[Task 14/20] Current/Best: 1.89/ 2.63 GFLOPS | Progress: (1284/2000) | 2288.55 s Done.
[Task 16/20] Current/Best: 0.47/ 0.96 GFLOPS | Progress: (1467/2000) | 2282.65 s Done.
[Task 18/20] Current/Best: 1.43/ 2.61 GFLOPS | Progress: (1007/2000) | 1525.76 s Done.
After this PR optimization:
[Task 2/20] Current/Best: 0.00/ 4.83 GFLOPS | Progress: (1682/2000) | 1470.40 s Done.
[Task 4/20] Current/Best: 1.35/ 3.17 GFLOPS | Progress: (1257/2000) | 1032.80 s Done.
[Task 6/20] Current/Best: 2.04/ 5.49 GFLOPS | Progress: (1904/2000) | 1623.10 s Done.
[Task 8/20] Current/Best: 0.75/ 3.15 GFLOPS | Progress: (1885/2000) | 1546.22 s Done.
[Task 10/20] Current/Best: 2.09/ 6.07 GFLOPS | Progress: (2000/2000) | 1640.41 s Done.
[Task 12/20] Current/Best: 2.99/ 3.80 GFLOPS | Progress: (1853/2000) | 1547.13 s Done.
[Task 14/20] Current/Best: 4.59/ 6.06 GFLOPS | Progress: (1355/2000) | 1091.93 s Done.
[Task 16/20] Current/Best: 1.96/ 4.01 GFLOPS | Progress: (2000/2000) | 1586.18 s Done.
[Task 18/20] Current/Best: 2.33/ 4.63 GFLOPS | Progress: (2000/2000) | 1599.89 s Done.
The depthwise convolution total execution time on single A53@2.0GHz time can be from 45.3839ms to 28.1945ms.
One thing you must notice to use this schedule: You MUST make the XGBTunner constructor’s feature type argument be feature_type= 'knob'. i.e. XGBTuner(tsk, loss_type='rank', feature_type='knob'). Otherwise your program maybe hang forever.
This schedule is not default schedule (i.e. direct) of arm cpu / x86 cpu depthwise convolution. I will update the auto tuning of ARM CPU tutorial to show how to use this
contrib_spatial_packschedule in the following PR.@merrymercy @yzhliu @tqchen pls review it.