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

[microNPU] Add relu6 relu_n1_to_1 test cases for Ethos-U #13645

Merged
merged 5 commits into from
Jan 12, 2023
Merged

[microNPU] Add relu6 relu_n1_to_1 test cases for Ethos-U #13645

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
5 commits
Select commit Hold shift + click to select a range
ee41a27
[microNPU] Add relu6 relu_n1_to_1 test cases for Ethos-U
Aleksei-grovety Dec 19, 2022
2da2b0e
add separate tests for relu6, relu_n1_to_1
Aleksei-grovety Dec 28, 2022
f692260
remain test max_relu_n1_to_1
Aleksei-grovety Dec 30, 2022
87b449d
add relu6, relu_n1_to_1 test cases for Ethos-U, remove changes for Bi…
Aleksei-grovety Jan 11, 2023
1332b95
remove pad operation from relu6, relu_n1_to_1 test cases
Aleksei-grovety Jan 12, 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
62 changes: 47 additions & 15 deletions python/tvm/relay/op/contrib/ethosu.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -688,15 +688,13 @@ def __init__(self, func_body: Call, operator_type: str, is_quantized_operation:
clip = None
requantize = None

if is_quantized_operation:
if str(current_call.op) == "clip":
clip = current_call
current_call = clip.args[0]
else:
if str(current_call.op) == "qnn.requantize":
requantize = current_call
clip = current_call.args[0]
current_call = clip.args[0]
if str(current_call.op) == "clip":
clip = current_call
current_call = clip.args[0]
elif str(current_call.op) == "qnn.requantize":
requantize = current_call
clip = current_call.args[0]
current_call = clip.args[0]
binary_op = current_call

layout = "NHWC"
Expand Down Expand Up @@ -929,6 +927,9 @@ def is_valid(self):
[self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8]
):
return False
# MIN with different scales is not supported on NPU.
if self.ifm.q_params.scale_f32 != self.ofm.q_params.scale_f32:
return False
return True


Expand All @@ -938,12 +939,21 @@ def minimum_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
"""
minimum = is_op("minimum")(wildcard(), wildcard())
optional_min_clip = is_op("clip")(minimum)
optional_min_clip = is_op("qnn.requantize")(
optional_min_clip, is_constant(), is_constant(), is_constant(), is_constant()
)
return minimum | optional_min_clip


def minimum_clip_requantize_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
"""
This function creates the pattern for minimum with fused RELU activation.
"""
pattern = is_op("minimum")(wildcard(), wildcard())
pattern = is_op("clip")(pattern)
pattern = is_op("qnn.requantize")(
pattern, is_constant(), is_constant(), is_constant(), is_constant()
)
return pattern
Copy link
Contributor

Choose a reason for hiding this comment

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

What's the motivation behind having two patterns instead of one with optional requantize if we use same params extractor (MinParams) with them?

Copy link
Contributor Author

Choose a reason for hiding this comment

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

There are two patterns to support two cases:
a) when we can offload minimum + clip + qnn.requantize to NPU with one operation ethosu_binary_elementwise if there are same scales
b) when we can offload minimum + clip + qnn.requantize to NPU with two operations ethosu_binary_elementwise + ethosu_identity if there are different scales

Since there is a hardware limitation, we cannot perform min or max operation fused with requantize (please look at NPU_SET_OFM_SCALE https://developer.arm.com/documentation/102420/0200/Programmers-model/Command-stream/cmd1-commands-) when we have different scales.

Copy link
Contributor

Choose a reason for hiding this comment

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

Ok, I see how that works now :) I think so far the approach has been to match all the variations of the operator and then output the right combination of Ethos-U ops, e.g. Resize2d. If there is a reason it can't be done that way, it would be good to document it there since it is not immediately obvious from the code why we need two patterns.
I think it also needs some legalization tests to check that:
(1) min/max with matching scales -> ethosu_binary_elementwise
(2) min/max with different scales -> ethosu_binary_elementwise + ethosu_identity



class MaxParams(BinaryElementwiseParams):
"""
This class will parse a call to a ethosu.binary_elementwise Max composite function
Expand All @@ -967,6 +977,9 @@ def is_valid(self):
[self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8]
):
return False
# MAX with different scales is not supported on NPU.
if self.ifm.q_params.scale_f32 != self.ofm.q_params.scale_f32:
return False
return True


Expand All @@ -976,12 +989,21 @@ def maximum_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
"""
maximum = is_op("maximum")(wildcard(), wildcard())
optional_max_clip = is_op("clip")(maximum)
optional_max_clip = is_op("qnn.requantize")(
optional_max_clip, is_constant(), is_constant(), is_constant(), is_constant()
)
return maximum | optional_max_clip


def maximum_clip_requantize_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
"""
This function creates the pattern for maximum with fused RELU activation.
"""
pattern = is_op("maximum")(wildcard(), wildcard())
pattern = is_op("clip")(pattern)
pattern = is_op("qnn.requantize")(
pattern, is_constant(), is_constant(), is_constant(), is_constant()
)
return pattern


class ShlParams(BinaryElementwiseParams):
"""
This class will parse a call to a ethosu.binary_elementwise Shl composite function
Expand Down Expand Up @@ -1820,11 +1842,21 @@ def pattern_table() -> List[Tuple[str, tvm.relay.dataflow_pattern.DFPattern, Cal
qnn_mul_pattern(),
lambda pat: MulParams(pat).is_valid(),
),
(
MinParams.composite_name,
minimum_clip_requantize_pattern(),
lambda pat: MinParams(pat).is_valid(),
),
(
MinParams.composite_name,
minimum_pattern(),
lambda pat: MinParams(pat).is_valid(),
),
(
MaxParams.composite_name,
maximum_clip_requantize_pattern(),
lambda pat: MaxParams(pat).is_valid(),
),
(
MaxParams.composite_name,
maximum_pattern(),
Expand Down
20 changes: 20 additions & 0 deletions tests/python/contrib/test_ethosu/test_codegen.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1109,6 +1109,26 @@ def leaky_relu_func(x):
)


def test_tflite_relu_n1_to_1():
np.random.seed(0)
accel_type = "ethos-u55-128"
ifm_shape = (1, 12, 16, 8)

@tf.function
def max_relu_n1_to_1(lhs, rhs):
op = tf.math.maximum(lhs, rhs)
# The specific pattern will be replaced into RELU_N1_TO_1 by tflite.
return tf.math.maximum(-1.0, tf.math.minimum(op, 1.0))

infra.compare_tvm_with_tflite(
max_relu_n1_to_1,
[ifm_shape, ifm_shape],
accel_type,
enable_cascader=is_u55_accel_type(accel_type),
ranges=[(-1, 1), (0, 2)],
)


@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
@pytest.mark.parametrize("ifm_shape", [(1, 14), (1, 151)])
@pytest.mark.parametrize("ofm_channels", [32, 64])
Expand Down
10 changes: 10 additions & 0 deletions tests/python/contrib/test_ethosu/test_legalize.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -899,6 +899,11 @@ def verify(ext_func):
elif operator_type == "MIN":
rewriter = legalize.MinRewriter()
pattern_table = [
(
ethosu.MinParams.composite_name,
ethosu.minimum_clip_requantize_pattern(),
lambda pat: ethosu.MinParams(pat).is_valid(),
),
(
ethosu.MinParams.composite_name,
ethosu.minimum_pattern(),
Expand All @@ -908,6 +913,11 @@ def verify(ext_func):
elif operator_type == "MAX":
rewriter = legalize.MaxRewriter()
pattern_table = [
(
ethosu.MaxParams.composite_name,
ethosu.maximum_clip_requantize_pattern(),
lambda pat: ethosu.MaxParams(pat).is_valid(),
),
(
ethosu.MaxParams.composite_name,
ethosu.maximum_pattern(),
Expand Down