[ETHOSU][MicroNPU][Pass] Add a pass to replicate pads #14909
Merged
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
Collaborator
Contributor
Author
|
cc @neildhickey, @ekalda, @ilyag-grovety, @Aleksei-grovety @arina-grovety |
ekalda
reviewed
May 30, 2023
sergio-grovety
force-pushed
the
ethosu-branching-inside-composite
branch
from
ff1fd37 to
7836a42
Compare
sergio-grovety
changed the title
Contributor
|
@tvm-bot rerun |
Comment on lines
3689
to
3704
| if op_pairs[0] == "depthwise": | ||
| weight_shape = [kernel_shape[0], kernel_shape[1], ifm_shape[3], 1] | ||
| weight = tf.constant(np.random.uniform(size=weight_shape), dtype=tf.float32) | ||
| x1 = tf.nn.depthwise_conv2d( | ||
| x, weight, strides=tf_strides, padding=op_padding, dilations=dilation | ||
| ) | ||
| else: | ||
| weight_shape = [kernel_shape[0], kernel_shape[1], ifm_shape[3], 3] | ||
| weight = tf.constant(np.random.uniform(size=weight_shape), dtype=tf.float32) | ||
| x1 = tf.nn.conv2d( | ||
| x, | ||
| weight, | ||
| strides=tf_strides, | ||
| padding=op_padding, | ||
| dilations=dilation, | ||
| ) |
Contributor
There was a problem hiding this comment.
Here the same code as on lines 3706 - 3721, it can be put into a separate function.
| # KIND, either express or implied. See the License for the | ||
| # specific language governing permissions and limitations | ||
| # under the License. | ||
| "Adds pads so that each conv2d operator has only one consumer" |
Contributor
There was a problem hiding this comment.
Maybe "Adds pads to each conv2d operator so that each pad has only one consumer."?
| new_conv2d_args = [] | ||
| for i, arg in enumerate(call.args): | ||
| if i == 0: | ||
| new_conv2d_args.append(self.visit(new_pad)) |
Contributor
There was a problem hiding this comment.
new_pad is visited twice the first time on line 59.
Contributor
|
@tvm-bot rerun |
ekalda
requested changes
Jun 19, 2023
Contributor
ekalda
left a comment
ekalda
left a comment
There was a problem hiding this comment.
Thanks @sergio-grovety, great work! Bar @Aleksei-grovety suggestions, I only had one nit and two more general points:
- This pass is necessary due to how we do pattern matching for Ethos-U and is therefore coupled to Ethos-U integration, so I don't think it makes much sense to have it among generic Relay transformation passes. Maybe move it to somewhere into
ethosunamespace, e.g. into codegen. - Testing - while legalization test is the most important test for this kind of change, I think it would also be good to have a simple lightweight codegen test for this kind of graph topology. Since the pass creates a new conv2d, there is some risk of constants getting messed up, but a simple codegen test should eliminate that risk.
|
|
||
| def __init__(self): | ||
| ExprMutator.__init__(self) | ||
| self.hashes = set() |
Contributor
There was a problem hiding this comment.
Nit: Maybe add a comment what self.hashes represents
arina-grovety
force-pushed
the
ethosu-branching-inside-composite
branch
from
bf520bd to
4678cbf
Compare
Contributor
Author
|
@tvm-bot rerun |
ekalda
requested changes
Jul 4, 2023
Contributor
ekalda
left a comment
ekalda
left a comment
There was a problem hiding this comment.
Almost there, one suggestion regarding to testing!
Comment on lines
160
to
171
| @pytest.mark.parametrize("accel_type", ACCEL_TYPES) | ||
| @pytest.mark.parametrize("ifm_shape", [(1, 55, 32, 3)]) | ||
| @pytest.mark.parametrize( | ||
| "kernel_shape, activation_function", | ||
| [((3, 3), "RELU"), ((1, 2), "NONE")], | ||
| ) | ||
| @pytest.mark.parametrize("strides, dilation", [((3, 2), (1, 1))]) | ||
| @pytest.mark.parametrize("op_padding", ["SAME", "VALID"]) | ||
| @pytest.mark.parametrize("sep_padding", [(0, 0, 1, 1), (7, 5, 4, 5)]) | ||
| @pytest.mark.parametrize( | ||
| "op_pairs", [("conv2d", "conv2d"), ("depthwise", "depthwise"), ("conv2d", "depthwise")] | ||
| ) |
Contributor
There was a problem hiding this comment.
This is 128 codegen tests, which would noticeably increase CI time. I think the variations on kernel_shape, activation_function, op_padding and sep_padding are exercised by other tests as well, so these values could be kept constant. Also, if we are parametrizing over one value, it doesn't need to be expressed with parametrize and can be included into the test code.
ekalda
approved these changes
Jul 13, 2023
Contributor
ekalda
left a comment
ekalda
left a comment
There was a problem hiding this comment.
LGTM! Thanks @sergio-grovety, @Aleksei-grovety and @arina-grovety!
Contributor
|
Thanks all, this is merged now! |
junrushao
pushed a commit
to junrushao/tvm
that referenced
this pull request
Jul 24, 2023
Added a pass to to handle the situation when nn.pad operator has more than one qnn.conv2d consumer.
pad
/ \
Conv2D Conv2D
In this case, because of the peculiarities of pattern parsing, conv2d does not get into the composite for the NPU. Therefore, pads are added so that each has only one consumer.
---------
Co-authored-by: Sergey Smirnov <89378719+sergey-grovety@users.noreply.github.com>
Co-authored-by: Arina <117634809+arina-grovety@users.noreply.github.com>
Co-authored-by: arina.naumova <naumova@grovety.com>
junrushao
pushed a commit
to junrushao/tvm
that referenced
this pull request
Jul 27, 2023
Added a pass to to handle the situation when nn.pad operator has more than one qnn.conv2d consumer.
pad
/ \
Conv2D Conv2D
In this case, because of the peculiarities of pattern parsing, conv2d does not get into the composite for the NPU. Therefore, pads are added so that each has only one consumer.
---------
Co-authored-by: Sergey Smirnov <89378719+sergey-grovety@users.noreply.github.com>
Co-authored-by: Arina <117634809+arina-grovety@users.noreply.github.com>
Co-authored-by: arina.naumova <naumova@grovety.com>
junrushao
pushed a commit
to junrushao/tvm
that referenced
this pull request
Jul 30, 2023
Added a pass to to handle the situation when nn.pad operator has more than one qnn.conv2d consumer.
pad
/ \
Conv2D Conv2D
In this case, because of the peculiarities of pattern parsing, conv2d does not get into the composite for the NPU. Therefore, pads are added so that each has only one consumer.
---------
Co-authored-by: Sergey Smirnov <89378719+sergey-grovety@users.noreply.github.com>
Co-authored-by: Arina <117634809+arina-grovety@users.noreply.github.com>
Co-authored-by: arina.naumova <naumova@grovety.com>
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
5 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.
Added a pass to to handle the situation when nn.pad operator has more than one qnn.conv2d consumer.
In this case, because of the peculiarities of pattern parsing, conv2d does not get into the composite for the NPU. Therefore, pads are added so that each has only one consumer.