Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

[TOPHUB] use keys as a keyword for searching of existing statistics #13874

Merged
merged 5 commits into from
Feb 2, 2023
+26 −0
Merged

[TOPHUB] use keys as a keyword for searching of existing statistics #13874

Show file tree
Hide file tree
Changes from 4 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
e4ff001
[TOPHUB] use keys as a keyword for searching of existing statistics
elvin-n Jan 30, 2023
3fde8b3
Add test on target not having arm_cpu device
elvin-n Jan 31, 2023
8b84b43
minor fix, add comment
elvin-n Jan 31, 2023
4a7345c
Fix pylint
elvin-n Jan 31, 2023
812c36c
Fix comment
elvin-n Feb 2, 2023
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
4 changes: 4 additions & 0 deletions python/tvm/autotvm/tophub.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -110,6 +110,10 @@ def context(target, extra_files=None):
device = tgt.attrs.get("device", "")
if device != "":
possible_names.append(_alias(device))
# for cases when we do have explicitly defined -device in the target,
# we still might have information about it stored in keys container
# in other case we will load statistics for definitely irrelative stat
Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Any chance this comment can be made clearer? :) I read it several times and I don't get it.
might have information
What information?
about it
About what?
keys container
I guess that is Target.keys? Maybe make that more explicit
in other case
What case? When keys container is empty or when -device is not defined?
irrelative stat
What is irrelative stat?

Also, is it one or multiple sentences? Capital letters and punctuation would be helpful :)

Copy link
Contributor Author

@elvin-n elvin-n Jan 31, 2023 •
edited
Loading

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

:) tried to include logic of matching statistics without detailed description in this place.

  1. the only mandatory field in target is kind. it can be llvm/opencl/cuda/...
  2. it might be device specified, but not required. It might be "arm_cpu", or "adreno" for example
  3. it might be model specified, but not required. It might be "riscv32-unknown-linux-gnu" or "kirin970" for example
  4. the keys field is not specified in the Target string but is filled after analysis of the Target automatically by devices. It is done for hierarchical search in the topi implementations and do fallback to lower/parents layers. As a result it might be ["arm_cpu", "cpu"]

Before this change, dedicated tophub file was searched only by device and second by kind
As you remember, device is not mandatory in the Target. At the same time TVM understands that target belongs to arm_cpu, it treat it as arm_cpu, but before my change, the target like llvm -model=snapdragon835 -mtriple=arm64-linux-android -mattr=+neon was matched only to llvm stat file. I cannot even imagine on which device llvm stat was collected, but definitely not on ARM. I.e. statistics was irrelevant.

Adding keys to the priority before kind we increase the probability of searching proper statistics.

Do I need to comment line by line your answer?

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for the detailed explanation :) No need to respond line by line, I understand what it is doing and I think it is a really good change. However, would you mind updating that comment in your PR? Otherwise the next person reading it will be confused again. Reading that comment, I think it has the logic inverted, I suppose you meant "For cases when we do NOT have explicitly defined -device in the target"? In my opinion something along the lines "If we don't have explicitly defined -device in the target, check target.keys for device." would be enough. Or even better, explicitly list the order which we use to determine the device:

  1. target.device
  2. target.keys
  3. target.kind

In general, it is indeed pretty worrying that we can silently apply completely irrelevant parameters from the tophub. I suppose in the ideal world (out of scope for this PR) we'd have something like "Using tuning parameters determined by tuning for < specific core or architecture >" printed to the console during compilation, at least for tvmc, to warn the user about the possibly bad choice of params.

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

ok, will modify comment

Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

As for applying of the statistics, I am not even sure that taking statistics from another model of the same device is a good idea. Fortunately it more or less works for ARM CPU currently but if we take intel graphics, the previous models of Iris had different limitation of group worksize. Statistics which collected on newer graphics cannot be applied on previous one. I.e. we risk not to have bad performance but have just not working kernels.

I.e. I agree that we need to do something with tophub and search algorithm, but do not have idea how we need to do this.

And probably it is not so important assuming that Metaschedule should be the tuner #1. On the other hand MS is not a solution for default compilation without tuning.

Copy link
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yeah I agree that applying same tuning statistics across different implementations of the same architecture is quite sketchy, especially if it happens in this non-transparent way it currently does. I'm also not sure how often the database gets updated with new stats. I haven't seen compilation crashing because of unsuitable tuning records from tophub myself, but if it happens, that's really bad, since it makes tuning compulsory for being able to compile a network with TVM. Also, sounds like a really difficult thing to debug. But I suppose as long as AutoTVM is the default tuner (as it is for tvmc) and therefore the one new potential users will use to evaluate TVM, it should be kept in a state where its behaviour is representative of TVM's capabilities. In my opinion it would be good if we didn't have 3 tuners to maintain in parallel, but I suppose that's where the community is at the moment.

possible_names.extend(tgt.keys)
possible_names.append(tgt.kind.name)

all_packages = list(PACKAGE_VERSION.keys())
Expand Down
16 changes: 16 additions & 0 deletions tests/python/unittest/test_autotvm_dispatch_context.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -19,6 +19,7 @@
to the parameters of workload"""

from tvm import autotvm
import tvm


@autotvm.template("testing/dispatch_fallback")
Expand All @@ -31,5 +32,20 @@ def test_fallback():
simple_template(2, 3)


def test_tophub_kinds_match():
def verify_arm_cpu(target):
best_by_targetkey = autotvm.tophub.context(target).best_by_targetkey
assert len(best_by_targetkey)
found_arm_cpu = False
for a, _ in best_by_targetkey:
if "arm_cpu" in a:
found_arm_cpu = True
break
assert found_arm_cpu

verify_arm_cpu("llvm -device=arm_cpu -mtriple=aarch64-linux-gnu -mattr=+neon,+v8.2a,+dotprod")
verify_arm_cpu("llvm -model=snapdragon835 -mtriple=arm64-linux-android -mattr=+neon")


if __name__ == "__main__":
test_fallback()