[TOPI] Add proper scheduling for dense on CUDA #3923
Merged
Conversation
This file contains 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
vinx13
approved these changes
Sep 9, 2019
|
@comaniac please look into ci error |
|
Could you also add fallback config for autotvm? |
|
@vinx13 Seems failed a case from NNVM. Will fix it soon. Thanks! |
|
The above commit:
|
wweic
reviewed
Sep 11, 2019
icemelon
reviewed
Sep 12, 2019
Huyuwei
reviewed
Sep 14, 2019
Huyuwei
requested changes
Sep 15, 2019
comaniac
force-pushed
the
dense_cuda_schedule
branch
from
28113b9 to
c5ddc78
Compare
comaniac
force-pushed
the
dense_cuda_schedule
branch
from
c5ddc78 to
a7b605f
Compare
comaniac
force-pushed
the
dense_cuda_schedule
branch
from
38cdeb6 to
f934a48
Compare
|
CI error doesn't relate to this PR and it passed locally. Re-run without change. |
Huyuwei
approved these changes
Sep 19, 2019
|
@comaniac This is merged, thanks! |
wweic
pushed a commit
to wweic/tvm
that referenced
this pull request
Sep 30, 2019
* add proper scheduling for dense on CUDA * add fallback config and fix unit test * fix corner cases * refactoring * fix bias and add testcase * let fusion happen
wweic
pushed a commit
to wweic/tvm
that referenced
this pull request
Sep 30, 2019
* add proper scheduling for dense on CUDA * add fallback config and fix unit test * fix corner cases * refactoring * fix bias and add testcase * let fusion happen
wweic
pushed a commit
to neo-ai/tvm
that referenced
this pull request
Oct 1, 2019
* add proper scheduling for dense on CUDA * add fallback config and fix unit test * fix corner cases * refactoring * fix bias and add testcase * let fusion happen
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
@icemelon9 please review this PR that adds scheduling for dense OP on CUDA since the original scheduling was too basic to achieve reasonable performance.
The added scheduling was modified from topi/recipe/gemm/cuda_gemm_square.py to achieve high performance (6TFlop/s for 2048x2048 dense matrix after AutoTVM tuning). For small batch (<64) dense, we are still based on the original scheduling but just added a parameter (tile_k) for AutoTVM to tune (~370 GFlop/s for batch size 1 dense computation).
One reason to have separate scheduling for different batch size is that I encountered invalid CUDA kernel errors when applying the high performance scheduling to small batch. I think that may be due to invalid splits, but I am not quite sure. You are welcome to comment and suggest improvements.