[microNPU][2b] Create CascaderGraphs from TE graphs #9471
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
mbaret
requested review from
anijain2305,
areusch,
comaniac,
icemelon,
jroesch,
junrushao,
jwfromm,
MarisaKirisame,
mbrookhart,
merrymercy,
slyubomirsky,
tqchen,
vinx13,
wweic,
yzhliu,
zhiics and
ZihengJiang
as code owners
12 tasks
mbaret
removed request for
areusch,
wweic,
yzhliu,
jroesch,
tqchen,
icemelon,
MarisaKirisame,
mbrookhart,
zhiics,
slyubomirsky,
vinx13 and
comaniac
mbaret
force-pushed
the
ethosu-cascader-3
branch
from
b92a2a2
to
6e1b3d4
Compare
mbaret
force-pushed
the
ethosu-cascader-3
branch
from
6e1b3d4
to
1279da0
Compare
mbaret
force-pushed
the
ethosu-cascader-3
branch
2 times, most recently
from
aba0a8e
to
950f280
Compare
mbaret
force-pushed
the
ethosu-cascader-3
branch
3 times, most recently
from
867ef3a
to
7a0cfb9
Compare
jacobbohlin
reviewed
Jan 5, 2022
The first step in the cascader is to create a CascaderGraph from a TE graph. To do this, every operator in the TE graph must get 'matched' by a Part matcher. This converts TE operations into cascader Parts by augmenting them with Propagators. In this initial commit, we include basic Part matchers for ethosu_conv2d and some transform operators so that the graph creation can be tested. Change-Id: I36bbbdbafd14f09e6d1a8abf32591d6755192915
Change-Id: I34e48e4ce4fd14e6c92601412f5029aa08efe8fc
Change-Id: Idcb92c26863837b06fa68f9122a1a79df79ee242
Change-Id: I0aad319b7722a56b5c7f116adcc791951223baa3
Change-Id: Ib1633341ed84e83c29c4220dd9bc4a7b6d17a5d0
Change-Id: I5223da462a8dadfda1798e1ea1c9b30b4e804217
Change-Id: Ib614c21e97723571aa002a2a4a3edda3afafb9d4
mbaret
force-pushed
the
ethosu-cascader-3
branch
from
7a0cfb9
to
a6f2857
Compare
|
cc @Mousius could you take a look? |
Mousius
requested changes
Jan 7, 2022
Comment on lines
+256
to
+273
| write = output_tensor | ||
| if write.op.name != "ethosu_write": | ||
| return None | ||
| convert_to_nhcwb16 = write.op.input_tensors[0] | ||
| if convert_to_nhcwb16.op.name != "ethosu_convert_to_nhcwb16": | ||
| return None | ||
| conv2d = convert_to_nhcwb16.op.input_tensors[0] | ||
| if conv2d.op.name != "ethosu_conv2d": | ||
| return None | ||
| pad = conv2d.op.input_tensors[0] | ||
| if pad.op.name != "ethosu_pad": | ||
| return None | ||
| convert_to_nhwc = pad.op.input_tensors[0] | ||
| if convert_to_nhwc.op.name != "ethosu_convert_to_nhwc": | ||
| return None | ||
| read = convert_to_nhwc.op.input_tensors[0] | ||
| if read.op.name != "ethosu_read": | ||
| return None |
There was a problem hiding this comment.
I can't see test cases for all of these clauses, so we're missing most of the logic of this function? Can you add some?
| part = match_ethosu_conv2d(out) | ||
|
|
||
| assert isinstance(part, cs.EthosuPart) | ||
| assert len(part.propagators) == 3 |
There was a problem hiding this comment.
Can we test the other properties of the part returned?
| part = match_ethosu_inline(out) | ||
|
|
||
| assert isinstance(part, cs.InlinePart) | ||
| assert len(part.propagators) == 1 |
There was a problem hiding this comment.
Can we test the other properties of the part ?
| ) | ||
| pytest.importorskip("ethosu.vela") | ||
|
|
||
| import tvm.contrib.ethosu.cascader as cs |
There was a problem hiding this comment.
As much as I appreciate using my initials here, it'd be really great to use something I could immediately associate with the cascader; potentially cascader rather than cs ?
Comment on lines
+44
to
+45
| assert part.propagators[0].transform == ifm_transform | ||
| assert part.propagators[0].offset == ifm_offset |
There was a problem hiding this comment.
It would be good to test at least a few cases just to ensure the function doesn't always return the same thing 😸
Comment on lines
+47
to
+48
| if output_tensor.op.name not in INLINE_OPS: | ||
| return None |
| input_tensors = part.subgraph.input_tensors | ||
| break | ||
|
|
||
| assert part is not None, f"The tensor {tensor} doesn't match any part." |
There was a problem hiding this comment.
Would be good to ensure this fires when we expect it to with pytest.raises.
AndrewZhaoLuo
pushed a commit
to AndrewZhaoLuo/tvm
that referenced
this pull request
Jan 7, 2022
The first step in the cascader is to create a CascaderGraph from a TE graph. To do this, every operator in the TE graph must get 'matched' by a Part matcher. This converts TE operations into cascader Parts by augmenting them with Propagators. In this initial commit, we include basic Part matchers for ethosu_conv2d and some transform operators so that the graph creation can be tested.
ylc
pushed a commit
to ylc/tvm
that referenced
this pull request
Feb 16, 2022
The first step in the cascader is to create a CascaderGraph from a TE graph. To do this, every operator in the TE graph must get 'matched' by a Part matcher. This converts TE operations into cascader Parts by augmenting them with Propagators. In this initial commit, we include basic Part matchers for ethosu_conv2d and some transform operators so that the graph creation can be tested.
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.
RFC: apache/tvm-rfcs#37
Issue: #9429
The first step in the cascader is to create a CascaderGraph from a TE graph. To do this, every operator in the TE graph must get 'matched' by a Part matcher. This converts TE operations into cascader Parts by augmenting them with Propagators.
In this initial commit, we include basic Part matchers for ethosu_conv2d and some transform operators so that the graph creation can be tested.