You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@tvm.apache.org by GitBox <gi...@apache.org> on 2021/11/04 13:16:54 UTC
[GitHub] [tvm] NicolaLancellotti opened a new pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
NicolaLancellotti opened a new pull request #9442:
URL: https://github.com/apache/tvm/pull/9442
This commit adds support for the binary elementwise primitive operators for the Arm(R) Ethos(TM)-U NPU and includes a few minor rewording changes.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] mbaret commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
mbaret commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r743845777
##########
File path: python/tvm/relay/backend/contrib/ethosu/legalize.py
##########
@@ -423,11 +640,20 @@ class LegalizeEthosU:
def transform_module(
self, mod: tvm.ir.IRModule, ctx: tvm.ir.transform.PassContext
) -> tvm.ir.IRModule:
+ """This is the method that replace the operations with hardware/codegen supported
Review comment:
that replaces
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -343,5 +343,256 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_ethosu_binary_elementwise(
+ accel_type,
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ tf.config.run_functions_eagerly(True)
+
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, lhs, rhs):
+ if operator_type == "ADD":
+ op = tf.math.add(lhs, rhs)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(lhs, rhs)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(lhs, rhs)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(lhs, rhs)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(lhs, rhs)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
Review comment:
Not necessarily something to resolve in this patch, but I wonder if we can abstract some of the boilerplate code here into a utility function.
##########
File path: python/tvm/relay/op/contrib/ethosu.py
##########
@@ -458,6 +458,316 @@ def qnn_avgpool2d_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
return pattern
+class BinaryElementwiseParams:
+ """
+ This class will parse a call to a ethosu.binary_elementwise composite function
+ and extract the parameter information.
+ """
+
+ def __init__(self, func_body: Call, operator_type: str, has_quantization_parameters: bool):
+ clip = None
+ if str(func_body.op) == "clip":
+ clip = func_body
+ binary_op = clip.args[0]
+ else:
+ binary_op = func_body
+
+ layout = "NHWC"
+
+ if has_quantization_parameters:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm_zero_point.value],
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm2_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm2_zero_point.value],
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ofm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ofm_zero_point.value],
+ )
+ else:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ )
+ self.activation = clip
+ self.operator_type = operator_type
+
+ def brodcastable(x, y):
Review comment:
broadcastable (or 'can_broadcast' might be better)
##########
File path: python/tvm/relay/backend/contrib/ethosu/legalize.py
##########
@@ -423,11 +640,20 @@ class LegalizeEthosU:
def transform_module(
self, mod: tvm.ir.IRModule, ctx: tvm.ir.transform.PassContext
) -> tvm.ir.IRModule:
+ """This is the method that replace the operations with hardware/codegen supported
Review comment:
that replaces
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -343,5 +343,256 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_ethosu_binary_elementwise(
+ accel_type,
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ tf.config.run_functions_eagerly(True)
+
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, lhs, rhs):
+ if operator_type == "ADD":
+ op = tf.math.add(lhs, rhs)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(lhs, rhs)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(lhs, rhs)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(lhs, rhs)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(lhs, rhs)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
Review comment:
Not necessarily something to resolve in this patch, but I wonder if we can abstract some of the boilerplate code here into a utility function.
##########
File path: python/tvm/relay/op/contrib/ethosu.py
##########
@@ -458,6 +458,316 @@ def qnn_avgpool2d_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
return pattern
+class BinaryElementwiseParams:
+ """
+ This class will parse a call to a ethosu.binary_elementwise composite function
+ and extract the parameter information.
+ """
+
+ def __init__(self, func_body: Call, operator_type: str, has_quantization_parameters: bool):
+ clip = None
+ if str(func_body.op) == "clip":
+ clip = func_body
+ binary_op = clip.args[0]
+ else:
+ binary_op = func_body
+
+ layout = "NHWC"
+
+ if has_quantization_parameters:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm_zero_point.value],
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm2_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm2_zero_point.value],
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ofm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ofm_zero_point.value],
+ )
+ else:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ )
+ self.activation = clip
+ self.operator_type = operator_type
+
+ def brodcastable(x, y):
Review comment:
broadcastable (or 'can_broadcast' might be better)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] ekalda commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
ekalda commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r742898510
##########
File path: python/tvm/relay/backend/contrib/ethosu/te/binary_elementwise.py
##########
@@ -0,0 +1,169 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=invalid-name,unused-argument
+"""Tensor Expressions for binary_elementwise"""
+import operator
+from tvm import te
+from .dma import dma_ofm_compute, dma_ifm_compute
+
+
+def binary_elementwise_compute(
+ ifm: te.Tensor,
+ ifm2: te.Tensor,
+ lut: te.Tensor,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ activation: str,
+ clip_min: int,
+ clip_max: int,
+ ifm_layout: str,
+ ifm2_layout: str,
+ ofm_layout: str,
+) -> te.Tensor:
+ """A compute operator representing the capabilities of binary_elementwise for the NPU.
+
+ Parameters
+ ----------
+ ifm : te.Tensor
+ The Input Feature Map tensor (IFM).
+ ifm2 : te.Tensor
+ The Input Feature Map tensor 1 (IFM2).
+ lut : te.Tensor
+ The look-up table values to use if activation = "LUT".
+ operator_type: str
+ The type of the binary elementwise operator.
+ "ADD"
+ "SUB"
+ "MUL"
+ "MIN"
+ "MAX"
+ "SHR"
+ "SHL"
+ ifm_scale : float
+ The quantization scale for the Input Feature Map tensor.
+ ifm_zero_point : int
+ The quantization zero point for the Input Feature Map tensor.
+ ifm2_scale : float
+ The quantization scale for the Input Feature Map tensor 1.
Review comment:
Maybe call this "tensor 2" to comply with the parameter name and also with the documentation in C++ Relay operator implementation
##########
File path: tests/python/contrib/test_ethosu/test_legalize.py
##########
@@ -558,5 +558,193 @@ def verify(ext_func):
verify(mod["tvmgen_default_ethosu_main_0"])
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 4], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 4], True),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_tflite_binary_elemwise_legalize(
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ reversed_operands,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, x, y):
+ if operator_type == "ADD":
+ op = tf.math.add(x, y)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(x, y)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(x, y)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(x, y)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(x, y)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32), tf.TensorSpec(ifm2_shape, dtype=tf.float32)
+ )
+
+ # Convert the model
+ def representative_dataset():
+ for _ in range(100):
+ data = np.random.rand(*tuple(ifm_shape))
+ data2 = np.random.rand(*tuple(ifm2_shape)) * 2
+ yield [data.astype(np.float32), data2.astype(np.float32)]
+
+ converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
+ converter.optimizations = [tf.lite.Optimize.DEFAULT]
+ converter.representative_dataset = representative_dataset
+ converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+ converter.inference_input_type = tf.int8
+ converter.inference_output_type = tf.int8
+ tflite_model = converter.convert()
+ return tflite_model
+
+ def verify(ext_func):
+ out_shape = ifm2_shape if reversed_operands else ifm_shape
+ shapes = [ifm_shape, ifm2_shape]
+ ifm_index, ifm2_index = (1, 0) if reversed_operands else (0, 1)
+ op = ext_func.body
+ assert list(op.args[0].checked_type.shape) == shapes[ifm_index]
+ assert list(op.args[1].checked_type.shape) == shapes[ifm2_index]
+ assert op.args[0].checked_type.dtype == dtype
+ assert list(op.checked_type.shape) == out_shape
+ assert op.checked_type.dtype == dtype
+ assert op.attrs.operator_type == operator_type
+ assert op.attrs.reversed_operands == reversed_operands
+ if activation_function == "RELU":
+ assert str(op.attrs.activation) == "CLIP"
+
+ if operator_type == "ADD":
+ rewriter = legalize.AddRewriter()
+ pattern_table = [
+ (
+ ethosu.AddParams.composite_name,
+ ethosu.qnn_add_pattern(),
+ lambda pat: ethosu.AddParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "SUB":
+ rewriter = legalize.SubRewriter()
+ pattern_table = [
+ (
+ ethosu.SubParams.composite_name,
+ ethosu.qnn_subtract_pattern(),
+ lambda pat: ethosu.SubParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MUL":
+ rewriter = legalize.MulRewriter()
+ pattern_table = [
+ (
+ ethosu.MulParams.composite_name,
+ ethosu.qnn_mul_pattern(),
+ lambda pat: ethosu.MulParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MIN":
+ rewriter = legalize.MinRewriter()
+ pattern_table = [
+ (
+ ethosu.MinParams.composite_name,
+ ethosu.minimum_pattern(),
+ lambda pat: ethosu.MinParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MAX":
+ rewriter = legalize.MaxRewriter()
+ pattern_table = [
+ (
+ ethosu.MaxParams.composite_name,
+ ethosu.maximum_pattern(),
+ lambda pat: ethosu.MaxParams(pat).is_valid(),
+ ),
+ ]
+
+ tflite_graph = create_tflite_graph()
+ tflite_model = tflite.Model.Model.GetRootAsModel(tflite_graph, 0)
+
+ mod, _ = relay.frontend.from_tflite(
+ tflite_model,
+ shape_dict={"x": ifm_shape, "y": ifm2_shape},
+ dtype_dict={"x": dtype, "y": dtype},
+ )
+ mod = partition_ethosu_by_table(mod, pattern_table)
+
+ mod["tvmgen_default_ethosu_main_0"] = dataflow_pattern.rewrite(
+ rewriter, mod["tvmgen_default_ethosu_main_0"]
+ )
+ verify(mod["tvmgen_default_ethosu_main_0"])
+
+
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 4], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 4], True),
+ ],
+)
+def test_ethosu_left_shift_binary_elemwise_legalize(ifm_shape, ifm2_shape, reversed_operands):
+ dtype = "int32"
+ operator_type = "SHL"
+
+ def create_graph():
+ input1 = relay.var("x1", shape=ifm_shape, dtype=dtype)
+ input2 = relay.var("x2", shape=ifm2_shape, dtype=dtype)
+ c1 = relay.left_shift(input1, input2)
Review comment:
Since left shift legalization from Relay is tested, does it mean that the NPU left shift does the "actual" left shift, but right shift doesn't, so we shouldn't legalize the Relay right shift to Ethos-U right shift but should legalize the left shift? It seems a bit odd to me that one of the shifts has Relay legalization and the other doesn't, especially since I don't know what would be the use of Relay left shift for the NPU in the absence of corresponding TFLite operator. Any thoughts @manupa-arm @mbaret @lhutton1
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -343,5 +343,256 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_ethosu_binary_elementwise(
+ accel_type,
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ tf.config.run_functions_eagerly(True)
+
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, lhs, rhs):
+ if operator_type == "ADD":
+ op = tf.math.add(lhs, rhs)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(lhs, rhs)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(lhs, rhs)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(lhs, rhs)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(lhs, rhs)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32), tf.TensorSpec(ifm2_shape, dtype=tf.float32)
+ )
+
+ # Convert the model
+ def representative_dataset():
+ for _ in range(100):
+ data = np.random.rand(*tuple(ifm_shape))
+ data2 = np.random.rand(*tuple(ifm2_shape)) * 2
+ yield [data.astype(np.float32), data2.astype(np.float32)]
+
+ converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
+ converter.optimizations = [tf.lite.Optimize.DEFAULT]
+ converter.representative_dataset = representative_dataset
+ converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+ converter.inference_input_type = tf.int8
+ converter.inference_output_type = tf.int8
+ tflite_model = converter.convert()
+ return tflite_model
+
+ tflite_graph = create_tflite_graph()
+ tflite_model = tflite.Model.Model.GetRootAsModel(tflite_graph, 0)
+
+ mod, params = relay.frontend.from_tflite(
+ tflite_model,
+ shape_dict={"ifm": ifm_shape, "ifm2": ifm2_shape},
+ dtype_dict={"ifm": dtype, "ifm2": dtype},
+ )
+ mod = partition_for_ethosu(mod, params)
+
+ # Generate reference data
+ input_data, output_data = infra.generate_ref_data_tflite(tflite_graph)
+
+ compiled_models = infra.build_source(
+ mod,
+ input_data,
+ output_data,
+ accel_type,
+ output_tolerance=1 if operator_type == "MAX" else 0,
Review comment:
Why tolerance 1 for MAX?
##########
File path: python/tvm/relay/op/contrib/ethosu.py
##########
@@ -458,6 +458,348 @@ def qnn_avgpool2d_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
return pattern
+class BinaryElementwiseParams:
+ """
+ This class will parse a call to a ethosu.binary_elementwise composite function
+ and extract the parameter information.
+ """
+
+ def __init__(self, func_body: Call, operator_type: str, has_quantization_parameters: bool):
+ clip = None
+ if str(func_body.op) == "clip":
+ clip = func_body
+ binary_op = clip.args[0]
+ else:
+ binary_op = func_body
+
+ layout = "NHWC"
+
+ if has_quantization_parameters:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm_zero_point.value],
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm2_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm2_zero_point.value],
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ofm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ofm_zero_point.value],
+ )
+ else:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ )
+ self.activation = clip
+ self.operator_type = operator_type
+
+ def brodcastable(x, y):
+ for i in range(1, 4):
+ if x.shape[i] == y.shape[i] or y.shape[i] == 1:
+ continue
+ return False
+ return True
+
+ if brodcastable(self.ifm, self.ifm2):
+ self.reversed_operands = False
+ self.valid_broadcast = True
+ elif brodcastable(self.ifm2, self.ifm):
+ self.reversed_operands = True
+ self.ifm, self.ifm2 = self.ifm2, self.ifm
+ self.valid_broadcast = True
+ else:
+ self.valid_broadcast = False
+
+ def is_valid(self):
+ """
+ This function checks whether BinaryElementwise has compatible attributes with the NPU
+ """
+ if np.dtype(self.ofm) == np.int32 and self.activation is not None:
+ return False
+ if len(self.ifm.shape) != 4 or len(self.ifm2.shape) != 4:
+ return False
+ if self.ifm.shape[0] != 1 or self.ifm2.shape[0] != 1:
+ return False
+ if not self.valid_broadcast:
+ return False
+ return True
+
+
+class AddParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Add composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.add"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "ADD", True)
+
+ def is_valid(self):
+ """
+ This function checks whether Add has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8, np.int32]
+ ):
+ return False
+ return True
+
+
+def qnn_add_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for qnn.add with optional fused RELU activation.
+ """
+ pattern = is_op("qnn.add")(
+ wildcard(),
+ wildcard(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ )
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class SubParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Sub composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.sub"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "SUB", True)
+
+ def is_valid(self):
+ """
+ This function checks whether Sub has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8, np.int32]
+ ):
+ return False
+ return True
+
+
+def qnn_subtract_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for qnn.subtract with optional fused RELU activation.
+ """
+ pattern = is_op("qnn.subtract")(
+ wildcard(),
+ wildcard(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ )
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class MulParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Mul composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.mul"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "MUL", True)
+
+ def is_valid(self):
+ """
+ This function checks whether Mul has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8, np.int32]
+ ):
+ return False
+ return True
+
+
+def qnn_mul_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for qnn.mul with optional fused RELU activation.
+ """
+ pattern = is_op("qnn.mul")(
+ wildcard(),
+ wildcard(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ )
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class MinParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Min composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.min"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "MIN", False)
+
+ def is_valid(self):
+ """
+ This function checks whether Min has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if self.ifm.dtype != self.ifm2.dtype:
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8]
+ ):
+ return False
+ return True
+
+
+def minimum_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for minimum with optional fused RELU activation.
+ """
+ pattern = is_op("minimum")(wildcard(), wildcard())
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class MaxParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Max composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.max"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "MAX", False)
+
+ def is_valid(self):
+ """
+ This function checks whether Max has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if self.ifm.dtype != self.ifm2.dtype:
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8]
+ ):
+ return False
+ return True
+
+
+def maximum_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for maximum with optional fused RELU activation.
+ """
+ pattern = is_op("maximum")(wildcard(), wildcard())
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class ShrParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Shr composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.shr"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "SHR", False)
+
+ def is_valid(self):
+ """
+ This function checks whether Shr has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if not check_valid_dtypes([self.ifm, self.ifm2], supported_dtypes=[np.int32]):
+ return False
+ if not check_valid_dtypes([self.ofm], supported_dtypes=[np.uint8, np.int8, np.int32]):
+ return False
+ return True
+
+
+def shr_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for right_shift.
+ """
+ pattern = is_op("right_shift")(wildcard(), wildcard())
Review comment:
I think we don't need a pattern and params class for right shift if we are not going to legalize it for the NPU
##########
File path: python/tvm/relay/backend/contrib/ethosu/op/binary_elementwise.py
##########
@@ -0,0 +1,206 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=unused-argument
+"""Relay operators for binary elementwise operators for Arm(R) Ethos(TM)-U NPU"""
+import tvm
+from tvm.relay.op import _make
+from tvm.topi.generic import schedule_injective
+from tvm.relay.op.op import OpStrategy
+from tvm.relay.op import strategy as _strategy
+
+from ..te import binary_elementwise_compute
+
+
+def _extract_ethosu_binary_elementwise_params(attrs, args):
+ """Get the parameters necessary to construct a ethosu_binary_elementwise compute TE
+ from a ethosu_binary_elementwise Relay call."""
+ ifm = args[0]
+ ifm2 = args[1]
+ lut = args[2]
+ operator_type = attrs.operator_type
+ ifm_scale = attrs.ifm_scale
+ ifm_zero_point = attrs.ifm_zero_point
+ ifm2_scale = attrs.ifm2_scale
+ ifm2_zero_point = attrs.ifm2_zero_point
+ ofm_scale = attrs.ofm_scale
+ ofm_zero_point = attrs.ofm_zero_point
+ ofm_channels = attrs.ofm_channels
+ reversed_operands = attrs.reversed_operands
+ activation = attrs.activation
+ clip_min = attrs.clip_min
+ clip_max = attrs.clip_max
+ ifm_layout = attrs.ifm_layout
+ ifm2_layout = attrs.ifm2_layout
+ ofm_layout = attrs.ofm_layout
+
+ return (
+ ifm,
+ ifm2,
+ lut,
+ operator_type,
+ ifm_scale,
+ ifm_zero_point,
+ ifm2_scale,
+ ifm2_zero_point,
+ ofm_scale,
+ ofm_zero_point,
+ ofm_channels,
+ reversed_operands,
+ activation,
+ clip_min,
+ clip_max,
+ ifm_layout,
+ ifm2_layout,
+ ofm_layout,
+ )
+
+
+@tvm.ir.register_op_attr("contrib.ethosu.binary_elementwise", "FTVMCompute")
+def create_ethosu_binary_elementwise_compute(attrs, args, out_type):
+ """Create an ethosu_binary_elementwise compute op."""
+ params = _extract_ethosu_binary_elementwise_params(attrs, args)
+ op = binary_elementwise_compute(*params)
+ return [op]
+
+
+@tvm.ir.register_op_attr("contrib.ethosu.binary_elementwise", "FTVMStrategy")
+def binary_elementwise_strategy_ethosu(attrs, inputs, out_type, target):
+ strategy = OpStrategy()
+ strategy.add_implementation(
+ create_ethosu_binary_elementwise_compute,
+ _strategy.wrap_topi_schedule(schedule_injective),
+ name="ethosu_binary_elementwise",
+ )
+ return strategy
+
+
+def ethosu_binary_elementwise(
+ ifm: tvm.relay.Expr,
+ ifm2: tvm.relay.Expr,
+ lut: tvm.relay.Expr,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ ofm_dtype: str,
+ activation: str = "NONE",
+ clip_min: int = 0,
+ clip_max: int = 0,
+ ifm_layout: str = "NHWC",
+ ifm2_layout: str = "NHWC",
+ ofm_layout: str = "NHWC",
+) -> tvm.relay.Call:
+ """This is a quantized binary elementwise operation as supported by
+ the NPU. It accepts either NHWC or NHCWB16 format
+ for the input data.
+
+ Parameters
+ ----------
+ ifm : tvm.relay.Expr
+ The Input Feature Map tensor (IFM).
+ ifm2 : tvm.relay.Expr
+ The Input Feature Map tensor 2 (IFM2).
+ lut : tvm.relay.Expr
+ The look-up table of values to use if activation = "LUT".
+ operator_type: str
+ The type of the binary elementwise operator.
+ "ADD"
+ "SUB"
+ "MUL"
+ "MIN"
+ "MAX"
+ "SHR"
+ "SHL"
+ ifm_scale : float
+ The quantization scale for the Input Feature Map tensor.
+ ifm_zero_point : int
+ The quantization zero point for the Input Feature Map tensor.
+ ifm2_scale : float
+ The quantization scale for the Input Feature Map tensor 2.
+ ifm2_zero_point : int
+ The quantization zero point for the Input Feature Map tensor 2.
+ ofm_scale : float
+ The quantization scale for the Output Feature Map tensor.
+ ofm_zero_point : int
+ The quantization zero point for the Output Feature Map tensor.
+ ofm_channels : int
+ The number of the Output Feature Map channels.
+ reversed_operands : bool
+ Specific if IFM2 is the first operand and IFM is the second operand.
Review comment:
Maybe "True" instead of "Specific"?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] lhutton1 commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
lhutton1 commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r742958078
##########
File path: tests/python/contrib/test_ethosu/test_legalize.py
##########
@@ -558,5 +558,193 @@ def verify(ext_func):
verify(mod["tvmgen_default_ethosu_main_0"])
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 4], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 4], True),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_tflite_binary_elemwise_legalize(
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ reversed_operands,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, x, y):
+ if operator_type == "ADD":
+ op = tf.math.add(x, y)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(x, y)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(x, y)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(x, y)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(x, y)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32), tf.TensorSpec(ifm2_shape, dtype=tf.float32)
+ )
+
+ # Convert the model
+ def representative_dataset():
+ for _ in range(100):
+ data = np.random.rand(*tuple(ifm_shape))
+ data2 = np.random.rand(*tuple(ifm2_shape)) * 2
+ yield [data.astype(np.float32), data2.astype(np.float32)]
+
+ converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
+ converter.optimizations = [tf.lite.Optimize.DEFAULT]
+ converter.representative_dataset = representative_dataset
+ converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+ converter.inference_input_type = tf.int8
+ converter.inference_output_type = tf.int8
+ tflite_model = converter.convert()
+ return tflite_model
+
+ def verify(ext_func):
+ out_shape = ifm2_shape if reversed_operands else ifm_shape
+ shapes = [ifm_shape, ifm2_shape]
+ ifm_index, ifm2_index = (1, 0) if reversed_operands else (0, 1)
+ op = ext_func.body
+ assert list(op.args[0].checked_type.shape) == shapes[ifm_index]
+ assert list(op.args[1].checked_type.shape) == shapes[ifm2_index]
+ assert op.args[0].checked_type.dtype == dtype
+ assert list(op.checked_type.shape) == out_shape
+ assert op.checked_type.dtype == dtype
+ assert op.attrs.operator_type == operator_type
+ assert op.attrs.reversed_operands == reversed_operands
+ if activation_function == "RELU":
+ assert str(op.attrs.activation) == "CLIP"
+
+ if operator_type == "ADD":
+ rewriter = legalize.AddRewriter()
+ pattern_table = [
+ (
+ ethosu.AddParams.composite_name,
+ ethosu.qnn_add_pattern(),
+ lambda pat: ethosu.AddParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "SUB":
+ rewriter = legalize.SubRewriter()
+ pattern_table = [
+ (
+ ethosu.SubParams.composite_name,
+ ethosu.qnn_subtract_pattern(),
+ lambda pat: ethosu.SubParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MUL":
+ rewriter = legalize.MulRewriter()
+ pattern_table = [
+ (
+ ethosu.MulParams.composite_name,
+ ethosu.qnn_mul_pattern(),
+ lambda pat: ethosu.MulParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MIN":
+ rewriter = legalize.MinRewriter()
+ pattern_table = [
+ (
+ ethosu.MinParams.composite_name,
+ ethosu.minimum_pattern(),
+ lambda pat: ethosu.MinParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MAX":
+ rewriter = legalize.MaxRewriter()
+ pattern_table = [
+ (
+ ethosu.MaxParams.composite_name,
+ ethosu.maximum_pattern(),
+ lambda pat: ethosu.MaxParams(pat).is_valid(),
+ ),
+ ]
+
+ tflite_graph = create_tflite_graph()
+ tflite_model = tflite.Model.Model.GetRootAsModel(tflite_graph, 0)
+
+ mod, _ = relay.frontend.from_tflite(
+ tflite_model,
+ shape_dict={"x": ifm_shape, "y": ifm2_shape},
+ dtype_dict={"x": dtype, "y": dtype},
+ )
+ mod = partition_ethosu_by_table(mod, pattern_table)
+
+ mod["tvmgen_default_ethosu_main_0"] = dataflow_pattern.rewrite(
+ rewriter, mod["tvmgen_default_ethosu_main_0"]
+ )
+ verify(mod["tvmgen_default_ethosu_main_0"])
+
+
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 4], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 4], True),
+ ],
+)
+def test_ethosu_left_shift_binary_elemwise_legalize(ifm_shape, ifm2_shape, reversed_operands):
+ dtype = "int32"
+ operator_type = "SHL"
+
+ def create_graph():
+ input1 = relay.var("x1", shape=ifm_shape, dtype=dtype)
+ input2 = relay.var("x2", shape=ifm2_shape, dtype=dtype)
+ c1 = relay.left_shift(input1, input2)
Review comment:
IIRC the difference is due to Relay using a `floor` while the NPU does something semantically equivalent to a `ceil`. If that's the case can we not add some trickery when legalizing a right shift operator directly? e.g. legalize right shift to: right shift followed by element-wise subtract of 1?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r745467423
##########
File path: python/tvm/relay/backend/contrib/ethosu/te/binary_elementwise.py
##########
@@ -0,0 +1,169 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=invalid-name,unused-argument
+"""Tensor Expressions for binary_elementwise"""
+import operator
+from tvm import te
+from .dma import dma_ofm_compute, dma_ifm_compute
+
+
+def binary_elementwise_compute(
+ ifm: te.Tensor,
+ ifm2: te.Tensor,
+ lut: te.Tensor,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ activation: str,
+ clip_min: int,
+ clip_max: int,
+ ifm_layout: str,
+ ifm2_layout: str,
+ ofm_layout: str,
+) -> te.Tensor:
+ """A compute operator representing the capabilities of binary_elementwise for the NPU.
+
+ Parameters
+ ----------
+ ifm : te.Tensor
+ The Input Feature Map tensor (IFM).
+ ifm2 : te.Tensor
+ The Input Feature Map tensor 1 (IFM2).
+ lut : te.Tensor
+ The look-up table values to use if activation = "LUT".
+ operator_type: str
+ The type of the binary elementwise operator.
+ "ADD"
+ "SUB"
+ "MUL"
+ "MIN"
+ "MAX"
+ "SHR"
+ "SHL"
+ ifm_scale : float
+ The quantization scale for the Input Feature Map tensor.
+ ifm_zero_point : int
+ The quantization zero point for the Input Feature Map tensor.
+ ifm2_scale : float
+ The quantization scale for the Input Feature Map tensor 2.
+ ifm2_zero_point : int
+ The quantization zero point for the Input Feature Map tensor 1.
+ ofm_scale : float
+ The quantization scale for the Output Feature Map tensor.
+ ofm_zero_point : int
+ The quantization zero point for the Output Feature Map tensor.
+ ofm_channels : int
+ The number of the Output Feature Map channels.
+ reversed_operands : bool
+ True if IFM2 is the first operand and IFM is the second operand.
+ activation : str
+ The activation function to use.
+ "NONE" - no activation function.
+ "CLIP" - clip the output between clip_min and clip_max.
+ "TANH" - tanh activation function.
+ "SIGMOID" - sigmoid activation function.
+ "LUT" - use a look-up table to perform the activation function.
+ Available activations for activation type:
+ {int8, uint8}: "NONE", "CLIP", "TANH", "SIGMOID", "LUT"
+ {int32}: "NONE"
+ clip_min : int
+ The minimum clipping value if activation = "CLIP".
+ clip_max : int
+ The maximum clipping value if activation = "CLIP".
+ ifm_layout : str, optional
+ The layout of the Input Feature Map tensor. Can be "NHWC" or "NHCWB16".
+ ifm2_layout : str, optional
+ The layout of the Input Feature Map tensor 1. Can be "NHWC" or "NHCWB16".
Review comment:
Done
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r745465992
##########
File path: python/tvm/relay/backend/contrib/ethosu/te/binary_elementwise.py
##########
@@ -0,0 +1,169 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=invalid-name,unused-argument
+"""Tensor Expressions for binary_elementwise"""
+import operator
+from tvm import te
+from .dma import dma_ofm_compute, dma_ifm_compute
+
+
+def binary_elementwise_compute(
+ ifm: te.Tensor,
+ ifm2: te.Tensor,
+ lut: te.Tensor,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ activation: str,
+ clip_min: int,
+ clip_max: int,
+ ifm_layout: str,
+ ifm2_layout: str,
+ ofm_layout: str,
+) -> te.Tensor:
+ """A compute operator representing the capabilities of binary_elementwise for the NPU.
+
+ Parameters
+ ----------
+ ifm : te.Tensor
+ The Input Feature Map tensor (IFM).
+ ifm2 : te.Tensor
+ The Input Feature Map tensor 1 (IFM2).
+ lut : te.Tensor
+ The look-up table values to use if activation = "LUT".
+ operator_type: str
+ The type of the binary elementwise operator.
+ "ADD"
+ "SUB"
+ "MUL"
+ "MIN"
+ "MAX"
+ "SHR"
+ "SHL"
+ ifm_scale : float
+ The quantization scale for the Input Feature Map tensor.
+ ifm_zero_point : int
+ The quantization zero point for the Input Feature Map tensor.
+ ifm2_scale : float
+ The quantization scale for the Input Feature Map tensor 2.
+ ifm2_zero_point : int
+ The quantization zero point for the Input Feature Map tensor 1.
+ ofm_scale : float
+ The quantization scale for the Output Feature Map tensor.
+ ofm_zero_point : int
+ The quantization zero point for the Output Feature Map tensor.
+ ofm_channels : int
+ The number of the Output Feature Map channels.
+ reversed_operands : bool
+ True if IFM2 is the first operand and IFM is the second operand.
+ activation : str
+ The activation function to use.
+ "NONE" - no activation function.
+ "CLIP" - clip the output between clip_min and clip_max.
+ "TANH" - tanh activation function.
+ "SIGMOID" - sigmoid activation function.
+ "LUT" - use a look-up table to perform the activation function.
+ Available activations for activation type:
+ {int8, uint8}: "NONE", "CLIP", "TANH", "SIGMOID", "LUT"
+ {int32}: "NONE"
+ clip_min : int
+ The minimum clipping value if activation = "CLIP".
+ clip_max : int
+ The maximum clipping value if activation = "CLIP".
+ ifm_layout : str, optional
+ The layout of the Input Feature Map tensor. Can be "NHWC" or "NHCWB16".
+ ifm2_layout : str, optional
+ The layout of the Input Feature Map tensor 1. Can be "NHWC" or "NHCWB16".
+ ofm_layout : str, optional
+ The layout of the Output Feature Map tensor. Can be "NHWC" or "NHCWB16".
+
+ Returns
+ -------
+ te.Tensor
+ The OFM tensor.
Review comment:
Done
##########
File path: src/relay/op/contrib/ethosu/binary_elementwise.cc
##########
@@ -0,0 +1,299 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file src/relay/op/contrib/ethosu/binary_elementwise.cc
+ * \brief Binary elementwise operators definitions for the Arm(R) Ethos(TM)-U NPU.
+ */
+#include <tvm/relay/op.h>
+
+#include "common.h"
+
+namespace tvm {
+namespace relay {
+namespace op {
+namespace contrib {
+namespace ethosu {
+
+/*! \brief Attributes used by the Ethos(TM)-U NPU binary elementwise operators */
+struct EthosuBinaryElementwiseAttrs : public tvm::AttrsNode<EthosuBinaryElementwiseAttrs> {
+ String operator_type;
+ double ifm_scale;
+ int ifm_zero_point;
+ double ifm2_scale;
+ int ifm2_zero_point;
+ double ofm_scale;
+ int ofm_zero_point;
+ IndexExpr ofm_channels;
+ bool reversed_operands;
+ String activation;
+ int clip_min;
+ int clip_max;
+ String ifm_layout;
+ String ifm2_layout;
+ String ofm_layout;
+ String ofm_dtype;
+
+ TVM_DECLARE_ATTRS(EthosuBinaryElementwiseAttrs, "relay.attrs.EthosuBinaryElementwiseAttrs") {
+ TVM_ATTR_FIELD(operator_type)
+ .describe(
+ "The type of the binary elementwise operator."
+ "'ADD'"
+ "'SUB'"
+ "'MUL'"
+ "'MIN'"
+ "'MAX'"
+ "'SHR'"
+ "'SHL'");
+ TVM_ATTR_FIELD(ifm_scale).describe("The quantization scale for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm2_scale)
+ .describe("The quantization scale for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ifm2_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ofm_scale).describe("The quantization scale for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_zero_point)
+ .describe("The quantization zero point for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_channels).describe("The number of the Output Feature Map channels.");
+ TVM_ATTR_FIELD(reversed_operands)
+ .describe("True if IFM2 is the first operand and IFM is the second operand.")
+ .set_default(false);
+ TVM_ATTR_FIELD(activation)
+ .describe(
+ "The activation function to use. "
+ "'NONE' - no activation function. "
+ "'CLIP' - clip the output between clip_min and clip_max. "
+ "'TANH' - tanh activation function. "
+ "'SIGMOID' - sigmoid activation function. "
+ "'LUT' - use a look-up table to perform the activation function."
+ "Available activations for activation type:"
+ "{int8, uint8}: 'NONE', 'CLIP', 'TANH', 'SIGMOID', 'LUT'"
+ "{int32}: 'NONE'")
+ .set_default("NONE");
+ TVM_ATTR_FIELD(clip_min)
+ .describe("The minimum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(clip_max)
+ .describe("The maximum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(ifm_layout)
+ .describe("The layout of the Input Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ifm2_layout)
+ .describe("The layout of the Input Feature Map tensor 2. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_layout)
+ .describe("The layout of the Output Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_dtype)
+ .describe(
+ "The Output Feature Map tensor type."
+ "MUL, ADD, SUB {IFM}->{OFM}:"
+ " {uint8, int8 int32} -> {uint8, int8, int32}, any pairing"
+ "MAX, MIN:"
+ " IFM and OFM must be of the same type, one of:"
+ " {int8, uint8}"
+ "SHR {IFM}->{OFM}:"
+ " {int32}->{int8, uint8, int32}, any pairing"
+ "SHL:"
+ " {int32}->{int32} only")
+ .set_default("NHWC");
+ }
+};
+
+TVM_REGISTER_NODE_TYPE(EthosuBinaryElementwiseAttrs);
+
+bool EthosuBinaryElementwiseRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
+ const TypeReporter& reporter) {
+ int ifm_index = 0;
+ int ifm2_index = 1;
+ int result_index = 3;
Review comment:
Done
##########
File path: src/relay/op/contrib/ethosu/binary_elementwise.cc
##########
@@ -0,0 +1,299 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file src/relay/op/contrib/ethosu/binary_elementwise.cc
+ * \brief Binary elementwise operators definitions for the Arm(R) Ethos(TM)-U NPU.
+ */
+#include <tvm/relay/op.h>
+
+#include "common.h"
+
+namespace tvm {
+namespace relay {
+namespace op {
+namespace contrib {
+namespace ethosu {
+
+/*! \brief Attributes used by the Ethos(TM)-U NPU binary elementwise operators */
+struct EthosuBinaryElementwiseAttrs : public tvm::AttrsNode<EthosuBinaryElementwiseAttrs> {
+ String operator_type;
+ double ifm_scale;
+ int ifm_zero_point;
+ double ifm2_scale;
+ int ifm2_zero_point;
+ double ofm_scale;
+ int ofm_zero_point;
+ IndexExpr ofm_channels;
+ bool reversed_operands;
+ String activation;
+ int clip_min;
+ int clip_max;
+ String ifm_layout;
+ String ifm2_layout;
+ String ofm_layout;
+ String ofm_dtype;
+
+ TVM_DECLARE_ATTRS(EthosuBinaryElementwiseAttrs, "relay.attrs.EthosuBinaryElementwiseAttrs") {
+ TVM_ATTR_FIELD(operator_type)
+ .describe(
+ "The type of the binary elementwise operator."
+ "'ADD'"
+ "'SUB'"
+ "'MUL'"
+ "'MIN'"
+ "'MAX'"
+ "'SHR'"
+ "'SHL'");
+ TVM_ATTR_FIELD(ifm_scale).describe("The quantization scale for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm2_scale)
+ .describe("The quantization scale for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ifm2_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ofm_scale).describe("The quantization scale for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_zero_point)
+ .describe("The quantization zero point for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_channels).describe("The number of the Output Feature Map channels.");
+ TVM_ATTR_FIELD(reversed_operands)
+ .describe("True if IFM2 is the first operand and IFM is the second operand.")
+ .set_default(false);
+ TVM_ATTR_FIELD(activation)
+ .describe(
+ "The activation function to use. "
+ "'NONE' - no activation function. "
+ "'CLIP' - clip the output between clip_min and clip_max. "
+ "'TANH' - tanh activation function. "
+ "'SIGMOID' - sigmoid activation function. "
+ "'LUT' - use a look-up table to perform the activation function."
+ "Available activations for activation type:"
+ "{int8, uint8}: 'NONE', 'CLIP', 'TANH', 'SIGMOID', 'LUT'"
+ "{int32}: 'NONE'")
+ .set_default("NONE");
+ TVM_ATTR_FIELD(clip_min)
+ .describe("The minimum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(clip_max)
+ .describe("The maximum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(ifm_layout)
+ .describe("The layout of the Input Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ifm2_layout)
+ .describe("The layout of the Input Feature Map tensor 2. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_layout)
+ .describe("The layout of the Output Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_dtype)
+ .describe(
+ "The Output Feature Map tensor type."
+ "MUL, ADD, SUB {IFM}->{OFM}:"
+ " {uint8, int8 int32} -> {uint8, int8, int32}, any pairing"
+ "MAX, MIN:"
+ " IFM and OFM must be of the same type, one of:"
+ " {int8, uint8}"
+ "SHR {IFM}->{OFM}:"
+ " {int32}->{int8, uint8, int32}, any pairing"
+ "SHL:"
+ " {int32}->{int32} only")
+ .set_default("NHWC");
+ }
+};
+
+TVM_REGISTER_NODE_TYPE(EthosuBinaryElementwiseAttrs);
+
+bool EthosuBinaryElementwiseRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
+ const TypeReporter& reporter) {
+ int ifm_index = 0;
+ int ifm2_index = 1;
+ int result_index = 3;
+ ICHECK_EQ(types.size(), result_index + 1);
+
+ const auto* ifm = types[ifm_index].as<TensorTypeNode>();
+ const auto* ifm2 = types[ifm2_index].as<TensorTypeNode>();
+ if (ifm == nullptr) return false;
+ if (ifm2 == nullptr) return false;
+
+ const auto* param = attrs.as<EthosuBinaryElementwiseAttrs>();
+ ICHECK(param != nullptr) << "EthosuBinaryElementwiseAttrs cannot be nullptr.";
+
+ String operator_type = param->operator_type;
+ auto ifm_dtype = ifm->dtype;
+ auto ifm2_dtype = ifm2->dtype;
+ DataType ofm_dtype;
+
+ if (param->ofm_dtype == "int8") {
+ ofm_dtype = DataType::Int(8);
+ } else if (param->ofm_dtype == "uint8") {
+ ofm_dtype = DataType::UInt(8);
+ } else if (param->ofm_dtype == "int32") {
+ ofm_dtype = DataType::Int(32);
+ }
+
+ if (ifm_dtype != ifm2_dtype) {
+ reporter->GetDiagCtx().EmitFatal(Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise "
+ << "type for ifm2 be the same of ifm but was " << ifm2_dtype
+ << " instead of " << ifm_dtype);
+ return false;
+ }
+
+ if (operator_type == "ADD" || operator_type == "SUB" || operator_type == "MUL") {
+ if (ifm_dtype != DataType::UInt(8) && ifm_dtype != DataType::Int(8) &&
+ ifm_dtype != DataType::Int(32)) {
+ reporter->GetDiagCtx().EmitFatal(
+ Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise " << operator_type
+ << " type(uint8) or type(int8) or type(int32) for ifm but was " << ifm_dtype);
+ return false;
+ }
+ if (ofm_dtype != DataType::UInt(8) && ofm_dtype != DataType::Int(8) &&
+ ofm_dtype != DataType::Int(32)) {
+ reporter->GetDiagCtx().EmitFatal(
+ Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise " << operator_type
+ << " type(uint8) or type(int8) or type(int32) for ofm but was " << ofm_dtype);
+ return false;
+ }
+ } else if (operator_type == "MIN" || operator_type == "MAX") {
+ if (ifm_dtype != DataType::UInt(8) && ifm_dtype != DataType::Int(8)) {
+ reporter->GetDiagCtx().EmitFatal(
+ Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise " << operator_type
+ << " type(uint8) or type(int8) for ifm but was " << ifm_dtype);
+ return false;
+ }
+ if (ifm_dtype != ofm_dtype) {
+ reporter->GetDiagCtx().EmitFatal(Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise "
+ << operator_type
+ << " type for ofm be the same of ifm but was " << ofm_dtype
+ << " instead of " << ifm_dtype);
+ return false;
+ }
+ } else if (operator_type == "SHR") {
+ if (ifm_dtype != DataType::Int(32)) {
+ reporter->GetDiagCtx().EmitFatal(Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise "
+ << operator_type << " type(int32) for ifm but was "
+ << ifm_dtype);
+ return false;
+ }
+ if (ofm_dtype != DataType::UInt(8) && ofm_dtype != DataType::Int(8) &&
+ ofm_dtype != DataType::Int(32)) {
+ reporter->GetDiagCtx().EmitFatal(
+ Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise " << operator_type
+ << " type(uint8) or type(int8) or type(int32) for ofm but was " << ofm_dtype);
+ return false;
+ }
+ } else if (operator_type == "SHL") {
+ if (ifm_dtype != DataType::Int(32)) {
+ reporter->GetDiagCtx().EmitFatal(Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise "
+ << operator_type << " type(int32) for ifm but was "
+ << ifm_dtype);
+
+ return false;
+ }
+ if (ofm_dtype != DataType::Int(32)) {
+ reporter->GetDiagCtx().EmitFatal(Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise "
+ << operator_type << " type(int32) for ofm but was "
+ << ofm_dtype);
+ return false;
+ }
+ } else {
+ reporter->GetDiagCtx().EmitFatal(
+ Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise 'ADD' or 'SUB' or 'MUL' or "
+ << "'MIN' or 'MAX' or 'SHR' or 'SHL' for operator_type but was " << param->operator_type);
+ return false;
+ }
+
+ // Assign ofm type
+ auto ofm_shape = EthosuInferBinaryElementwiseOutputShape(ifm->shape, param->ifm_layout,
+ param->ofm_layout, param->ofm_channels);
+ reporter->Assign(types[result_index], TensorType(ofm_shape, ofm_dtype));
+ return true;
+}
+
+Expr MakeEthosuBinaryElementwise(Expr ifm, Expr ifm2, Expr lut, String operator_type,
+ double ifm_scale, int ifm_zero_point, double ifm2_scale,
+ int ifm2_zero_point, double ofm_scale, int ofm_zero_point,
+ IndexExpr ofm_channels, bool reversed_operands, String activation,
+ int clip_min, int clip_max, String ifm_layout, String ifm2_layout,
+ String ofm_layout, String ofm_dtype) {
+ auto attrs = make_object<EthosuBinaryElementwiseAttrs>();
+
+ attrs->operator_type = std::move(operator_type);
+ attrs->ifm_scale = ifm_scale;
+ attrs->ifm_zero_point = ifm_zero_point;
+ attrs->ifm2_scale = ifm2_scale;
+ attrs->ifm2_zero_point = ifm2_zero_point;
+ attrs->ofm_scale = ofm_scale;
+ attrs->ofm_zero_point = ofm_zero_point;
+ attrs->ofm_channels = ofm_channels;
Review comment:
Done
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r745464026
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -343,5 +343,256 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_ethosu_binary_elementwise(
+ accel_type,
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ tf.config.run_functions_eagerly(True)
+
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, lhs, rhs):
+ if operator_type == "ADD":
+ op = tf.math.add(lhs, rhs)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(lhs, rhs)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(lhs, rhs)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(lhs, rhs)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(lhs, rhs)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32), tf.TensorSpec(ifm2_shape, dtype=tf.float32)
+ )
+
+ # Convert the model
+ def representative_dataset():
+ for _ in range(100):
+ data = np.random.rand(*tuple(ifm_shape))
+ data2 = np.random.rand(*tuple(ifm2_shape)) * 2
+ yield [data.astype(np.float32), data2.astype(np.float32)]
+
+ converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
+ converter.optimizations = [tf.lite.Optimize.DEFAULT]
+ converter.representative_dataset = representative_dataset
+ converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+ converter.inference_input_type = tf.int8
+ converter.inference_output_type = tf.int8
+ tflite_model = converter.convert()
+ return tflite_model
+
+ tflite_graph = create_tflite_graph()
+ tflite_model = tflite.Model.Model.GetRootAsModel(tflite_graph, 0)
+
+ mod, params = relay.frontend.from_tflite(
+ tflite_model,
+ shape_dict={"ifm": ifm_shape, "ifm2": ifm2_shape},
+ dtype_dict={"ifm": dtype, "ifm2": dtype},
+ )
+ mod = partition_for_ethosu(mod, params)
+
+ # Generate reference data
+ input_data, output_data = infra.generate_ref_data_tflite(tflite_graph)
+
+ compiled_models = infra.build_source(
+ mod,
+ input_data,
+ output_data,
+ accel_type,
+ output_tolerance=1 if operator_type == "MAX" else 0,
+ )
+
+ # Assumes only two runtime.Modules are created -- i.e. single offload module
+ imported_modules = compiled_models[0].executor_factory.lib.imported_modules
+ assert len(imported_modules) == 2
+ ethosu_module = imported_modules[0]
+
+ # Verify generated C source
+ get_cs = tvm._ffi.get_global_func("runtime.module.ethosu.getcs")
+ cmms = get_cs(ethosu_module)
+ cmms = bytes.fromhex(cmms)
+
+ infra.print_payload(cmms)
+ infra.verify_source(compiled_models, accel_type)
+
+
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+def test_ethosu_left_shift_binary_elemwise(
+ accel_type,
+ ifm_shape,
+ ifm2_shape,
+):
+ dtype = "int32"
+
+ def create_model():
+ ifm = relay.var("ifm", shape=ifm_shape, dtype=dtype)
+ ifm2 = relay.var("ifm2", shape=ifm2_shape, dtype=dtype)
+ c1 = relay.left_shift(ifm, ifm2)
+ f = relay.Function([ifm, ifm2], c1)
+ mod = tvm.IRModule()
+ mod["main"] = f
+ return mod
+
+ relay_mod = create_model()
+ mod = partition_for_ethosu(relay_mod)
+
+ # Generate reference data
+ in_min, in_max = util.get_range_for_dtype_str(dtype)
+ input_data = {
+ "ifm": np.random.randint(in_min, high=in_max, size=ifm_shape, dtype=dtype),
+ "ifm2": np.random.randint(0, high=32, size=ifm2_shape, dtype=dtype),
+ }
+ output_data = generate_ref_data(relay_mod, input_data)
+
+ compiled_models = infra.build_source(
+ mod,
+ input_data,
+ output_data,
+ accel_type,
+ )
+
+ # Assumes only two runtime.Modules are created -- i.e. single offload module
+ imported_modules = compiled_models[0].executor_factory.lib.imported_modules
+ assert len(imported_modules) == 2
+ ethosu_module = imported_modules[0]
+
+ # Verify generated C source
+ get_cs = tvm._ffi.get_global_func("runtime.module.ethosu.getcs")
+ cmms = get_cs(ethosu_module)
+ cmms = bytes.fromhex(cmms)
+
+ infra.print_payload(cmms)
+ infra.verify_source(compiled_models, accel_type)
+
+
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 1], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 1], True),
+ ],
+)
+def test_ethosu_right_shift_binary_elemwise(ifm_shape, ifm2_shape, reversed_operands, accel_type):
+ dtype = "int32"
Review comment:
Done
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -343,5 +343,256 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_ethosu_binary_elementwise(
+ accel_type,
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ tf.config.run_functions_eagerly(True)
+
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, lhs, rhs):
+ if operator_type == "ADD":
+ op = tf.math.add(lhs, rhs)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(lhs, rhs)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(lhs, rhs)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(lhs, rhs)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(lhs, rhs)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
Review comment:
I agree, let's do it in a follow-up pr.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] lhutton1 commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
lhutton1 commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r742958078
##########
File path: tests/python/contrib/test_ethosu/test_legalize.py
##########
@@ -558,5 +558,193 @@ def verify(ext_func):
verify(mod["tvmgen_default_ethosu_main_0"])
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 4], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 4], True),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_tflite_binary_elemwise_legalize(
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ reversed_operands,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, x, y):
+ if operator_type == "ADD":
+ op = tf.math.add(x, y)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(x, y)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(x, y)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(x, y)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(x, y)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32), tf.TensorSpec(ifm2_shape, dtype=tf.float32)
+ )
+
+ # Convert the model
+ def representative_dataset():
+ for _ in range(100):
+ data = np.random.rand(*tuple(ifm_shape))
+ data2 = np.random.rand(*tuple(ifm2_shape)) * 2
+ yield [data.astype(np.float32), data2.astype(np.float32)]
+
+ converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
+ converter.optimizations = [tf.lite.Optimize.DEFAULT]
+ converter.representative_dataset = representative_dataset
+ converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+ converter.inference_input_type = tf.int8
+ converter.inference_output_type = tf.int8
+ tflite_model = converter.convert()
+ return tflite_model
+
+ def verify(ext_func):
+ out_shape = ifm2_shape if reversed_operands else ifm_shape
+ shapes = [ifm_shape, ifm2_shape]
+ ifm_index, ifm2_index = (1, 0) if reversed_operands else (0, 1)
+ op = ext_func.body
+ assert list(op.args[0].checked_type.shape) == shapes[ifm_index]
+ assert list(op.args[1].checked_type.shape) == shapes[ifm2_index]
+ assert op.args[0].checked_type.dtype == dtype
+ assert list(op.checked_type.shape) == out_shape
+ assert op.checked_type.dtype == dtype
+ assert op.attrs.operator_type == operator_type
+ assert op.attrs.reversed_operands == reversed_operands
+ if activation_function == "RELU":
+ assert str(op.attrs.activation) == "CLIP"
+
+ if operator_type == "ADD":
+ rewriter = legalize.AddRewriter()
+ pattern_table = [
+ (
+ ethosu.AddParams.composite_name,
+ ethosu.qnn_add_pattern(),
+ lambda pat: ethosu.AddParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "SUB":
+ rewriter = legalize.SubRewriter()
+ pattern_table = [
+ (
+ ethosu.SubParams.composite_name,
+ ethosu.qnn_subtract_pattern(),
+ lambda pat: ethosu.SubParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MUL":
+ rewriter = legalize.MulRewriter()
+ pattern_table = [
+ (
+ ethosu.MulParams.composite_name,
+ ethosu.qnn_mul_pattern(),
+ lambda pat: ethosu.MulParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MIN":
+ rewriter = legalize.MinRewriter()
+ pattern_table = [
+ (
+ ethosu.MinParams.composite_name,
+ ethosu.minimum_pattern(),
+ lambda pat: ethosu.MinParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MAX":
+ rewriter = legalize.MaxRewriter()
+ pattern_table = [
+ (
+ ethosu.MaxParams.composite_name,
+ ethosu.maximum_pattern(),
+ lambda pat: ethosu.MaxParams(pat).is_valid(),
+ ),
+ ]
+
+ tflite_graph = create_tflite_graph()
+ tflite_model = tflite.Model.Model.GetRootAsModel(tflite_graph, 0)
+
+ mod, _ = relay.frontend.from_tflite(
+ tflite_model,
+ shape_dict={"x": ifm_shape, "y": ifm2_shape},
+ dtype_dict={"x": dtype, "y": dtype},
+ )
+ mod = partition_ethosu_by_table(mod, pattern_table)
+
+ mod["tvmgen_default_ethosu_main_0"] = dataflow_pattern.rewrite(
+ rewriter, mod["tvmgen_default_ethosu_main_0"]
+ )
+ verify(mod["tvmgen_default_ethosu_main_0"])
+
+
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 4], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 4], True),
+ ],
+)
+def test_ethosu_left_shift_binary_elemwise_legalize(ifm_shape, ifm2_shape, reversed_operands):
+ dtype = "int32"
+ operator_type = "SHL"
+
+ def create_graph():
+ input1 = relay.var("x1", shape=ifm_shape, dtype=dtype)
+ input2 = relay.var("x2", shape=ifm2_shape, dtype=dtype)
+ c1 = relay.left_shift(input1, input2)
Review comment:
Can we not legalize a right shift operator to right shift followed by element-wise subtract of 1?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] lhutton1 commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
lhutton1 commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r745672194
##########
File path: src/relay/op/contrib/ethosu/binary_elementwise.cc
##########
@@ -0,0 +1,299 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file src/relay/op/contrib/ethosu/binary_elementwise.cc
+ * \brief Binary elementwise operators definitions for the Arm(R) Ethos(TM)-U NPU.
+ */
+#include <tvm/relay/op.h>
+
+#include "common.h"
+
+namespace tvm {
+namespace relay {
+namespace op {
+namespace contrib {
+namespace ethosu {
+
+/*! \brief Attributes used by the Ethos(TM)-U NPU binary elementwise operators */
+struct EthosuBinaryElementwiseAttrs : public tvm::AttrsNode<EthosuBinaryElementwiseAttrs> {
+ String operator_type;
+ double ifm_scale;
+ int ifm_zero_point;
+ double ifm2_scale;
+ int ifm2_zero_point;
+ double ofm_scale;
+ int ofm_zero_point;
+ IndexExpr ofm_channels;
+ bool reversed_operands;
+ String activation;
+ int clip_min;
+ int clip_max;
+ String ifm_layout;
+ String ifm2_layout;
+ String ofm_layout;
+ String ofm_dtype;
+
+ TVM_DECLARE_ATTRS(EthosuBinaryElementwiseAttrs, "relay.attrs.EthosuBinaryElementwiseAttrs") {
+ TVM_ATTR_FIELD(operator_type)
+ .describe(
+ "The type of the binary elementwise operator."
+ "'ADD'"
+ "'SUB'"
+ "'MUL'"
+ "'MIN'"
+ "'MAX'"
+ "'SHR'"
+ "'SHL'");
+ TVM_ATTR_FIELD(ifm_scale).describe("The quantization scale for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm2_scale)
+ .describe("The quantization scale for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ifm2_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ofm_scale).describe("The quantization scale for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_zero_point)
+ .describe("The quantization zero point for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_channels).describe("The number of the Output Feature Map channels.");
+ TVM_ATTR_FIELD(reversed_operands)
+ .describe("True if IFM2 is the first operand and IFM is the second operand.")
+ .set_default(false);
+ TVM_ATTR_FIELD(activation)
+ .describe(
+ "The activation function to use. "
+ "'NONE' - no activation function. "
+ "'CLIP' - clip the output between clip_min and clip_max. "
+ "'TANH' - tanh activation function. "
+ "'SIGMOID' - sigmoid activation function. "
+ "'LUT' - use a look-up table to perform the activation function."
+ "Available activations for activation type:"
+ "{int8, uint8}: 'NONE', 'CLIP', 'TANH', 'SIGMOID', 'LUT'"
+ "{int32}: 'NONE'")
+ .set_default("NONE");
+ TVM_ATTR_FIELD(clip_min)
+ .describe("The minimum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(clip_max)
+ .describe("The maximum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(ifm_layout)
+ .describe("The layout of the Input Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ifm2_layout)
+ .describe("The layout of the Input Feature Map tensor 2. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_layout)
+ .describe("The layout of the Output Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_dtype)
+ .describe(
+ "The Output Feature Map tensor type."
+ "MUL, ADD, SUB {IFM}->{OFM}:"
+ " {uint8, int8 int32} -> {uint8, int8, int32}, any pairing"
+ "MAX, MIN:"
+ " IFM and OFM must be of the same type, one of:"
+ " {int8, uint8}"
+ "SHR {IFM}->{OFM}:"
+ " {int32}->{int8, uint8, int32}, any pairing"
+ "SHL:"
+ " {int32}->{int32} only")
+ .set_default("NHWC");
Review comment:
Apologies missed this in my first pass,
```suggestion
.set_default("int8");
```
If `int8` is a suitable default value?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r743008661
##########
File path: python/tvm/relay/backend/contrib/ethosu/op/binary_elementwise.py
##########
@@ -0,0 +1,206 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=unused-argument
+"""Relay operators for binary elementwise operators for Arm(R) Ethos(TM)-U NPU"""
+import tvm
+from tvm.relay.op import _make
+from tvm.topi.generic import schedule_injective
+from tvm.relay.op.op import OpStrategy
+from tvm.relay.op import strategy as _strategy
+
+from ..te import binary_elementwise_compute
+
+
+def _extract_ethosu_binary_elementwise_params(attrs, args):
+ """Get the parameters necessary to construct a ethosu_binary_elementwise compute TE
+ from a ethosu_binary_elementwise Relay call."""
+ ifm = args[0]
+ ifm2 = args[1]
+ lut = args[2]
+ operator_type = attrs.operator_type
+ ifm_scale = attrs.ifm_scale
+ ifm_zero_point = attrs.ifm_zero_point
+ ifm2_scale = attrs.ifm2_scale
+ ifm2_zero_point = attrs.ifm2_zero_point
+ ofm_scale = attrs.ofm_scale
+ ofm_zero_point = attrs.ofm_zero_point
+ ofm_channels = attrs.ofm_channels
+ reversed_operands = attrs.reversed_operands
+ activation = attrs.activation
+ clip_min = attrs.clip_min
+ clip_max = attrs.clip_max
+ ifm_layout = attrs.ifm_layout
+ ifm2_layout = attrs.ifm2_layout
+ ofm_layout = attrs.ofm_layout
+
+ return (
+ ifm,
+ ifm2,
+ lut,
+ operator_type,
+ ifm_scale,
+ ifm_zero_point,
+ ifm2_scale,
+ ifm2_zero_point,
+ ofm_scale,
+ ofm_zero_point,
+ ofm_channels,
+ reversed_operands,
+ activation,
+ clip_min,
+ clip_max,
+ ifm_layout,
+ ifm2_layout,
+ ofm_layout,
+ )
+
+
+@tvm.ir.register_op_attr("contrib.ethosu.binary_elementwise", "FTVMCompute")
+def create_ethosu_binary_elementwise_compute(attrs, args, out_type):
+ """Create an ethosu_binary_elementwise compute op."""
+ params = _extract_ethosu_binary_elementwise_params(attrs, args)
+ op = binary_elementwise_compute(*params)
+ return [op]
+
+
+@tvm.ir.register_op_attr("contrib.ethosu.binary_elementwise", "FTVMStrategy")
+def binary_elementwise_strategy_ethosu(attrs, inputs, out_type, target):
+ strategy = OpStrategy()
+ strategy.add_implementation(
+ create_ethosu_binary_elementwise_compute,
+ _strategy.wrap_topi_schedule(schedule_injective),
+ name="ethosu_binary_elementwise",
+ )
+ return strategy
+
+
+def ethosu_binary_elementwise(
+ ifm: tvm.relay.Expr,
+ ifm2: tvm.relay.Expr,
+ lut: tvm.relay.Expr,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ ofm_dtype: str,
+ activation: str = "NONE",
+ clip_min: int = 0,
+ clip_max: int = 0,
+ ifm_layout: str = "NHWC",
+ ifm2_layout: str = "NHWC",
+ ofm_layout: str = "NHWC",
+) -> tvm.relay.Call:
+ """This is a quantized binary elementwise operation as supported by
+ the NPU. It accepts either NHWC or NHCWB16 format
+ for the input data.
+
+ Parameters
+ ----------
+ ifm : tvm.relay.Expr
+ The Input Feature Map tensor (IFM).
+ ifm2 : tvm.relay.Expr
+ The Input Feature Map tensor 2 (IFM2).
+ lut : tvm.relay.Expr
+ The look-up table of values to use if activation = "LUT".
+ operator_type: str
+ The type of the binary elementwise operator.
+ "ADD"
+ "SUB"
+ "MUL"
+ "MIN"
+ "MAX"
+ "SHR"
+ "SHL"
+ ifm_scale : float
+ The quantization scale for the Input Feature Map tensor.
+ ifm_zero_point : int
+ The quantization zero point for the Input Feature Map tensor.
+ ifm2_scale : float
+ The quantization scale for the Input Feature Map tensor 2.
+ ifm2_zero_point : int
+ The quantization zero point for the Input Feature Map tensor 2.
+ ofm_scale : float
+ The quantization scale for the Output Feature Map tensor.
+ ofm_zero_point : int
+ The quantization zero point for the Output Feature Map tensor.
+ ofm_channels : int
+ The number of the Output Feature Map channels.
+ reversed_operands : bool
+ Specific if IFM2 is the first operand and IFM is the second operand.
Review comment:
Done
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r745465259
##########
File path: python/tvm/relay/backend/contrib/ethosu/legalize.py
##########
@@ -423,11 +640,20 @@ class LegalizeEthosU:
def transform_module(
self, mod: tvm.ir.IRModule, ctx: tvm.ir.transform.PassContext
) -> tvm.ir.IRModule:
+ """This is the method that replace the operations with hardware/codegen supported
Review comment:
Done
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] mbaret merged pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
mbaret merged pull request #9442:
URL: https://github.com/apache/tvm/pull/9442
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] lhutton1 commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
lhutton1 commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r745672194
##########
File path: src/relay/op/contrib/ethosu/binary_elementwise.cc
##########
@@ -0,0 +1,299 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file src/relay/op/contrib/ethosu/binary_elementwise.cc
+ * \brief Binary elementwise operators definitions for the Arm(R) Ethos(TM)-U NPU.
+ */
+#include <tvm/relay/op.h>
+
+#include "common.h"
+
+namespace tvm {
+namespace relay {
+namespace op {
+namespace contrib {
+namespace ethosu {
+
+/*! \brief Attributes used by the Ethos(TM)-U NPU binary elementwise operators */
+struct EthosuBinaryElementwiseAttrs : public tvm::AttrsNode<EthosuBinaryElementwiseAttrs> {
+ String operator_type;
+ double ifm_scale;
+ int ifm_zero_point;
+ double ifm2_scale;
+ int ifm2_zero_point;
+ double ofm_scale;
+ int ofm_zero_point;
+ IndexExpr ofm_channels;
+ bool reversed_operands;
+ String activation;
+ int clip_min;
+ int clip_max;
+ String ifm_layout;
+ String ifm2_layout;
+ String ofm_layout;
+ String ofm_dtype;
+
+ TVM_DECLARE_ATTRS(EthosuBinaryElementwiseAttrs, "relay.attrs.EthosuBinaryElementwiseAttrs") {
+ TVM_ATTR_FIELD(operator_type)
+ .describe(
+ "The type of the binary elementwise operator."
+ "'ADD'"
+ "'SUB'"
+ "'MUL'"
+ "'MIN'"
+ "'MAX'"
+ "'SHR'"
+ "'SHL'");
+ TVM_ATTR_FIELD(ifm_scale).describe("The quantization scale for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm2_scale)
+ .describe("The quantization scale for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ifm2_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ofm_scale).describe("The quantization scale for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_zero_point)
+ .describe("The quantization zero point for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_channels).describe("The number of the Output Feature Map channels.");
+ TVM_ATTR_FIELD(reversed_operands)
+ .describe("True if IFM2 is the first operand and IFM is the second operand.")
+ .set_default(false);
+ TVM_ATTR_FIELD(activation)
+ .describe(
+ "The activation function to use. "
+ "'NONE' - no activation function. "
+ "'CLIP' - clip the output between clip_min and clip_max. "
+ "'TANH' - tanh activation function. "
+ "'SIGMOID' - sigmoid activation function. "
+ "'LUT' - use a look-up table to perform the activation function."
+ "Available activations for activation type:"
+ "{int8, uint8}: 'NONE', 'CLIP', 'TANH', 'SIGMOID', 'LUT'"
+ "{int32}: 'NONE'")
+ .set_default("NONE");
+ TVM_ATTR_FIELD(clip_min)
+ .describe("The minimum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(clip_max)
+ .describe("The maximum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(ifm_layout)
+ .describe("The layout of the Input Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ifm2_layout)
+ .describe("The layout of the Input Feature Map tensor 2. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_layout)
+ .describe("The layout of the Output Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_dtype)
+ .describe(
+ "The Output Feature Map tensor type."
+ "MUL, ADD, SUB {IFM}->{OFM}:"
+ " {uint8, int8 int32} -> {uint8, int8, int32}, any pairing"
+ "MAX, MIN:"
+ " IFM and OFM must be of the same type, one of:"
+ " {int8, uint8}"
+ "SHR {IFM}->{OFM}:"
+ " {int32}->{int8, uint8, int32}, any pairing"
+ "SHL:"
+ " {int32}->{int32} only")
+ .set_default("NHWC");
Review comment:
Apologies missed this in my first pass,
```suggestion
.set_default("int8");
```
If `int32` is a suitable default value?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r746713106
##########
File path: src/relay/op/contrib/ethosu/binary_elementwise.cc
##########
@@ -0,0 +1,299 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file src/relay/op/contrib/ethosu/binary_elementwise.cc
+ * \brief Binary elementwise operators definitions for the Arm(R) Ethos(TM)-U NPU.
+ */
+#include <tvm/relay/op.h>
+
+#include "common.h"
+
+namespace tvm {
+namespace relay {
+namespace op {
+namespace contrib {
+namespace ethosu {
+
+/*! \brief Attributes used by the Ethos(TM)-U NPU binary elementwise operators */
+struct EthosuBinaryElementwiseAttrs : public tvm::AttrsNode<EthosuBinaryElementwiseAttrs> {
+ String operator_type;
+ double ifm_scale;
+ int ifm_zero_point;
+ double ifm2_scale;
+ int ifm2_zero_point;
+ double ofm_scale;
+ int ofm_zero_point;
+ IndexExpr ofm_channels;
+ bool reversed_operands;
+ String activation;
+ int clip_min;
+ int clip_max;
+ String ifm_layout;
+ String ifm2_layout;
+ String ofm_layout;
+ String ofm_dtype;
+
+ TVM_DECLARE_ATTRS(EthosuBinaryElementwiseAttrs, "relay.attrs.EthosuBinaryElementwiseAttrs") {
+ TVM_ATTR_FIELD(operator_type)
+ .describe(
+ "The type of the binary elementwise operator."
+ "'ADD'"
+ "'SUB'"
+ "'MUL'"
+ "'MIN'"
+ "'MAX'"
+ "'SHR'"
+ "'SHL'");
+ TVM_ATTR_FIELD(ifm_scale).describe("The quantization scale for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm2_scale)
+ .describe("The quantization scale for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ifm2_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ofm_scale).describe("The quantization scale for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_zero_point)
+ .describe("The quantization zero point for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_channels).describe("The number of the Output Feature Map channels.");
+ TVM_ATTR_FIELD(reversed_operands)
+ .describe("True if IFM2 is the first operand and IFM is the second operand.")
+ .set_default(false);
+ TVM_ATTR_FIELD(activation)
+ .describe(
+ "The activation function to use. "
+ "'NONE' - no activation function. "
+ "'CLIP' - clip the output between clip_min and clip_max. "
+ "'TANH' - tanh activation function. "
+ "'SIGMOID' - sigmoid activation function. "
+ "'LUT' - use a look-up table to perform the activation function."
+ "Available activations for activation type:"
+ "{int8, uint8}: 'NONE', 'CLIP', 'TANH', 'SIGMOID', 'LUT'"
+ "{int32}: 'NONE'")
+ .set_default("NONE");
+ TVM_ATTR_FIELD(clip_min)
+ .describe("The minimum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(clip_max)
+ .describe("The maximum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(ifm_layout)
+ .describe("The layout of the Input Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ifm2_layout)
+ .describe("The layout of the Input Feature Map tensor 2. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_layout)
+ .describe("The layout of the Output Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_dtype)
+ .describe(
+ "The Output Feature Map tensor type."
+ "MUL, ADD, SUB {IFM}->{OFM}:"
+ " {uint8, int8 int32} -> {uint8, int8, int32}, any pairing"
+ "MAX, MIN:"
+ " IFM and OFM must be of the same type, one of:"
+ " {int8, uint8}"
+ "SHR {IFM}->{OFM}:"
+ " {int32}->{int8, uint8, int32}, any pairing"
+ "SHL:"
+ " {int32}->{int32} only")
+ .set_default("NHWC");
Review comment:
I think it is better to remove the default value.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r745466991
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -343,5 +343,256 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_ethosu_binary_elementwise(
+ accel_type,
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ tf.config.run_functions_eagerly(True)
Review comment:
Indeed, we don't need it, I removed it.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#issuecomment-960909495
@manupa-arm @mbaret @ekalda @lhutton1 @dchauhan-arm @Mousius
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] lhutton1 commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
lhutton1 commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r745672194
##########
File path: src/relay/op/contrib/ethosu/binary_elementwise.cc
##########
@@ -0,0 +1,299 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file src/relay/op/contrib/ethosu/binary_elementwise.cc
+ * \brief Binary elementwise operators definitions for the Arm(R) Ethos(TM)-U NPU.
+ */
+#include <tvm/relay/op.h>
+
+#include "common.h"
+
+namespace tvm {
+namespace relay {
+namespace op {
+namespace contrib {
+namespace ethosu {
+
+/*! \brief Attributes used by the Ethos(TM)-U NPU binary elementwise operators */
+struct EthosuBinaryElementwiseAttrs : public tvm::AttrsNode<EthosuBinaryElementwiseAttrs> {
+ String operator_type;
+ double ifm_scale;
+ int ifm_zero_point;
+ double ifm2_scale;
+ int ifm2_zero_point;
+ double ofm_scale;
+ int ofm_zero_point;
+ IndexExpr ofm_channels;
+ bool reversed_operands;
+ String activation;
+ int clip_min;
+ int clip_max;
+ String ifm_layout;
+ String ifm2_layout;
+ String ofm_layout;
+ String ofm_dtype;
+
+ TVM_DECLARE_ATTRS(EthosuBinaryElementwiseAttrs, "relay.attrs.EthosuBinaryElementwiseAttrs") {
+ TVM_ATTR_FIELD(operator_type)
+ .describe(
+ "The type of the binary elementwise operator."
+ "'ADD'"
+ "'SUB'"
+ "'MUL'"
+ "'MIN'"
+ "'MAX'"
+ "'SHR'"
+ "'SHL'");
+ TVM_ATTR_FIELD(ifm_scale).describe("The quantization scale for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm2_scale)
+ .describe("The quantization scale for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ifm2_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ofm_scale).describe("The quantization scale for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_zero_point)
+ .describe("The quantization zero point for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_channels).describe("The number of the Output Feature Map channels.");
+ TVM_ATTR_FIELD(reversed_operands)
+ .describe("True if IFM2 is the first operand and IFM is the second operand.")
+ .set_default(false);
+ TVM_ATTR_FIELD(activation)
+ .describe(
+ "The activation function to use. "
+ "'NONE' - no activation function. "
+ "'CLIP' - clip the output between clip_min and clip_max. "
+ "'TANH' - tanh activation function. "
+ "'SIGMOID' - sigmoid activation function. "
+ "'LUT' - use a look-up table to perform the activation function."
+ "Available activations for activation type:"
+ "{int8, uint8}: 'NONE', 'CLIP', 'TANH', 'SIGMOID', 'LUT'"
+ "{int32}: 'NONE'")
+ .set_default("NONE");
+ TVM_ATTR_FIELD(clip_min)
+ .describe("The minimum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(clip_max)
+ .describe("The maximum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(ifm_layout)
+ .describe("The layout of the Input Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ifm2_layout)
+ .describe("The layout of the Input Feature Map tensor 2. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_layout)
+ .describe("The layout of the Output Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_dtype)
+ .describe(
+ "The Output Feature Map tensor type."
+ "MUL, ADD, SUB {IFM}->{OFM}:"
+ " {uint8, int8 int32} -> {uint8, int8, int32}, any pairing"
+ "MAX, MIN:"
+ " IFM and OFM must be of the same type, one of:"
+ " {int8, uint8}"
+ "SHR {IFM}->{OFM}:"
+ " {int32}->{int8, uint8, int32}, any pairing"
+ "SHL:"
+ " {int32}->{int32} only")
+ .set_default("NHWC");
Review comment:
Apologies missed this in my first pass,
```suggestion
.set_default("int32");
```
If `int32` is a suitable default value?
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] mbaret commented on pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
mbaret commented on pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#issuecomment-966485596
Thanks @NicolaLancellotti, @lhutton1, @ekalda, this is now merged.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r743008661
##########
File path: python/tvm/relay/backend/contrib/ethosu/op/binary_elementwise.py
##########
@@ -0,0 +1,206 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=unused-argument
+"""Relay operators for binary elementwise operators for Arm(R) Ethos(TM)-U NPU"""
+import tvm
+from tvm.relay.op import _make
+from tvm.topi.generic import schedule_injective
+from tvm.relay.op.op import OpStrategy
+from tvm.relay.op import strategy as _strategy
+
+from ..te import binary_elementwise_compute
+
+
+def _extract_ethosu_binary_elementwise_params(attrs, args):
+ """Get the parameters necessary to construct a ethosu_binary_elementwise compute TE
+ from a ethosu_binary_elementwise Relay call."""
+ ifm = args[0]
+ ifm2 = args[1]
+ lut = args[2]
+ operator_type = attrs.operator_type
+ ifm_scale = attrs.ifm_scale
+ ifm_zero_point = attrs.ifm_zero_point
+ ifm2_scale = attrs.ifm2_scale
+ ifm2_zero_point = attrs.ifm2_zero_point
+ ofm_scale = attrs.ofm_scale
+ ofm_zero_point = attrs.ofm_zero_point
+ ofm_channels = attrs.ofm_channels
+ reversed_operands = attrs.reversed_operands
+ activation = attrs.activation
+ clip_min = attrs.clip_min
+ clip_max = attrs.clip_max
+ ifm_layout = attrs.ifm_layout
+ ifm2_layout = attrs.ifm2_layout
+ ofm_layout = attrs.ofm_layout
+
+ return (
+ ifm,
+ ifm2,
+ lut,
+ operator_type,
+ ifm_scale,
+ ifm_zero_point,
+ ifm2_scale,
+ ifm2_zero_point,
+ ofm_scale,
+ ofm_zero_point,
+ ofm_channels,
+ reversed_operands,
+ activation,
+ clip_min,
+ clip_max,
+ ifm_layout,
+ ifm2_layout,
+ ofm_layout,
+ )
+
+
+@tvm.ir.register_op_attr("contrib.ethosu.binary_elementwise", "FTVMCompute")
+def create_ethosu_binary_elementwise_compute(attrs, args, out_type):
+ """Create an ethosu_binary_elementwise compute op."""
+ params = _extract_ethosu_binary_elementwise_params(attrs, args)
+ op = binary_elementwise_compute(*params)
+ return [op]
+
+
+@tvm.ir.register_op_attr("contrib.ethosu.binary_elementwise", "FTVMStrategy")
+def binary_elementwise_strategy_ethosu(attrs, inputs, out_type, target):
+ strategy = OpStrategy()
+ strategy.add_implementation(
+ create_ethosu_binary_elementwise_compute,
+ _strategy.wrap_topi_schedule(schedule_injective),
+ name="ethosu_binary_elementwise",
+ )
+ return strategy
+
+
+def ethosu_binary_elementwise(
+ ifm: tvm.relay.Expr,
+ ifm2: tvm.relay.Expr,
+ lut: tvm.relay.Expr,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ ofm_dtype: str,
+ activation: str = "NONE",
+ clip_min: int = 0,
+ clip_max: int = 0,
+ ifm_layout: str = "NHWC",
+ ifm2_layout: str = "NHWC",
+ ofm_layout: str = "NHWC",
+) -> tvm.relay.Call:
+ """This is a quantized binary elementwise operation as supported by
+ the NPU. It accepts either NHWC or NHCWB16 format
+ for the input data.
+
+ Parameters
+ ----------
+ ifm : tvm.relay.Expr
+ The Input Feature Map tensor (IFM).
+ ifm2 : tvm.relay.Expr
+ The Input Feature Map tensor 2 (IFM2).
+ lut : tvm.relay.Expr
+ The look-up table of values to use if activation = "LUT".
+ operator_type: str
+ The type of the binary elementwise operator.
+ "ADD"
+ "SUB"
+ "MUL"
+ "MIN"
+ "MAX"
+ "SHR"
+ "SHL"
+ ifm_scale : float
+ The quantization scale for the Input Feature Map tensor.
+ ifm_zero_point : int
+ The quantization zero point for the Input Feature Map tensor.
+ ifm2_scale : float
+ The quantization scale for the Input Feature Map tensor 2.
+ ifm2_zero_point : int
+ The quantization zero point for the Input Feature Map tensor 2.
+ ofm_scale : float
+ The quantization scale for the Output Feature Map tensor.
+ ofm_zero_point : int
+ The quantization zero point for the Output Feature Map tensor.
+ ofm_channels : int
+ The number of the Output Feature Map channels.
+ reversed_operands : bool
+ Specific if IFM2 is the first operand and IFM is the second operand.
Review comment:
Done
##########
File path: python/tvm/relay/backend/contrib/ethosu/te/binary_elementwise.py
##########
@@ -0,0 +1,169 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=invalid-name,unused-argument
+"""Tensor Expressions for binary_elementwise"""
+import operator
+from tvm import te
+from .dma import dma_ofm_compute, dma_ifm_compute
+
+
+def binary_elementwise_compute(
+ ifm: te.Tensor,
+ ifm2: te.Tensor,
+ lut: te.Tensor,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ activation: str,
+ clip_min: int,
+ clip_max: int,
+ ifm_layout: str,
+ ifm2_layout: str,
+ ofm_layout: str,
+) -> te.Tensor:
+ """A compute operator representing the capabilities of binary_elementwise for the NPU.
+
+ Parameters
+ ----------
+ ifm : te.Tensor
+ The Input Feature Map tensor (IFM).
+ ifm2 : te.Tensor
+ The Input Feature Map tensor 1 (IFM2).
+ lut : te.Tensor
+ The look-up table values to use if activation = "LUT".
+ operator_type: str
+ The type of the binary elementwise operator.
+ "ADD"
+ "SUB"
+ "MUL"
+ "MIN"
+ "MAX"
+ "SHR"
+ "SHL"
+ ifm_scale : float
+ The quantization scale for the Input Feature Map tensor.
+ ifm_zero_point : int
+ The quantization zero point for the Input Feature Map tensor.
+ ifm2_scale : float
+ The quantization scale for the Input Feature Map tensor 1.
Review comment:
Done
##########
File path: python/tvm/relay/op/contrib/ethosu.py
##########
@@ -458,6 +458,348 @@ def qnn_avgpool2d_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
return pattern
+class BinaryElementwiseParams:
+ """
+ This class will parse a call to a ethosu.binary_elementwise composite function
+ and extract the parameter information.
+ """
+
+ def __init__(self, func_body: Call, operator_type: str, has_quantization_parameters: bool):
+ clip = None
+ if str(func_body.op) == "clip":
+ clip = func_body
+ binary_op = clip.args[0]
+ else:
+ binary_op = func_body
+
+ layout = "NHWC"
+
+ if has_quantization_parameters:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm_zero_point.value],
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm2_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm2_zero_point.value],
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ofm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ofm_zero_point.value],
+ )
+ else:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ )
+ self.activation = clip
+ self.operator_type = operator_type
+
+ def brodcastable(x, y):
+ for i in range(1, 4):
+ if x.shape[i] == y.shape[i] or y.shape[i] == 1:
+ continue
+ return False
+ return True
+
+ if brodcastable(self.ifm, self.ifm2):
+ self.reversed_operands = False
+ self.valid_broadcast = True
+ elif brodcastable(self.ifm2, self.ifm):
+ self.reversed_operands = True
+ self.ifm, self.ifm2 = self.ifm2, self.ifm
+ self.valid_broadcast = True
+ else:
+ self.valid_broadcast = False
+
+ def is_valid(self):
+ """
+ This function checks whether BinaryElementwise has compatible attributes with the NPU
+ """
+ if np.dtype(self.ofm) == np.int32 and self.activation is not None:
+ return False
+ if len(self.ifm.shape) != 4 or len(self.ifm2.shape) != 4:
+ return False
+ if self.ifm.shape[0] != 1 or self.ifm2.shape[0] != 1:
+ return False
+ if not self.valid_broadcast:
+ return False
+ return True
+
+
+class AddParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Add composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.add"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "ADD", True)
+
+ def is_valid(self):
+ """
+ This function checks whether Add has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8, np.int32]
+ ):
+ return False
+ return True
+
+
+def qnn_add_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for qnn.add with optional fused RELU activation.
+ """
+ pattern = is_op("qnn.add")(
+ wildcard(),
+ wildcard(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ )
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class SubParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Sub composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.sub"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "SUB", True)
+
+ def is_valid(self):
+ """
+ This function checks whether Sub has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8, np.int32]
+ ):
+ return False
+ return True
+
+
+def qnn_subtract_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for qnn.subtract with optional fused RELU activation.
+ """
+ pattern = is_op("qnn.subtract")(
+ wildcard(),
+ wildcard(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ )
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class MulParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Mul composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.mul"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "MUL", True)
+
+ def is_valid(self):
+ """
+ This function checks whether Mul has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8, np.int32]
+ ):
+ return False
+ return True
+
+
+def qnn_mul_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for qnn.mul with optional fused RELU activation.
+ """
+ pattern = is_op("qnn.mul")(
+ wildcard(),
+ wildcard(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ )
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class MinParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Min composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.min"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "MIN", False)
+
+ def is_valid(self):
+ """
+ This function checks whether Min has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if self.ifm.dtype != self.ifm2.dtype:
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8]
+ ):
+ return False
+ return True
+
+
+def minimum_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for minimum with optional fused RELU activation.
+ """
+ pattern = is_op("minimum")(wildcard(), wildcard())
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class MaxParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Max composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.max"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "MAX", False)
+
+ def is_valid(self):
+ """
+ This function checks whether Max has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if self.ifm.dtype != self.ifm2.dtype:
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8]
+ ):
+ return False
+ return True
+
+
+def maximum_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for maximum with optional fused RELU activation.
+ """
+ pattern = is_op("maximum")(wildcard(), wildcard())
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class ShrParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Shr composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.shr"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "SHR", False)
+
+ def is_valid(self):
+ """
+ This function checks whether Shr has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if not check_valid_dtypes([self.ifm, self.ifm2], supported_dtypes=[np.int32]):
+ return False
+ if not check_valid_dtypes([self.ofm], supported_dtypes=[np.uint8, np.int8, np.int32]):
+ return False
+ return True
+
+
+def shr_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for right_shift.
+ """
+ pattern = is_op("right_shift")(wildcard(), wildcard())
Review comment:
Removed
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -343,5 +343,256 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_ethosu_binary_elementwise(
+ accel_type,
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ tf.config.run_functions_eagerly(True)
+
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, lhs, rhs):
+ if operator_type == "ADD":
+ op = tf.math.add(lhs, rhs)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(lhs, rhs)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(lhs, rhs)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(lhs, rhs)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(lhs, rhs)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32), tf.TensorSpec(ifm2_shape, dtype=tf.float32)
+ )
+
+ # Convert the model
+ def representative_dataset():
+ for _ in range(100):
+ data = np.random.rand(*tuple(ifm_shape))
+ data2 = np.random.rand(*tuple(ifm2_shape)) * 2
+ yield [data.astype(np.float32), data2.astype(np.float32)]
+
+ converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
+ converter.optimizations = [tf.lite.Optimize.DEFAULT]
+ converter.representative_dataset = representative_dataset
+ converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+ converter.inference_input_type = tf.int8
+ converter.inference_output_type = tf.int8
+ tflite_model = converter.convert()
+ return tflite_model
+
+ tflite_graph = create_tflite_graph()
+ tflite_model = tflite.Model.Model.GetRootAsModel(tflite_graph, 0)
+
+ mod, params = relay.frontend.from_tflite(
+ tflite_model,
+ shape_dict={"ifm": ifm_shape, "ifm2": ifm2_shape},
+ dtype_dict={"ifm": dtype, "ifm2": dtype},
+ )
+ mod = partition_for_ethosu(mod, params)
+
+ # Generate reference data
+ input_data, output_data = infra.generate_ref_data_tflite(tflite_graph)
+
+ compiled_models = infra.build_source(
+ mod,
+ input_data,
+ output_data,
+ accel_type,
+ output_tolerance=1 if operator_type == "MAX" else 0,
Review comment:
The Max operator with the RELU activation function is not bit accurate, it needs a tolerance of 1.
I'll investigate and address this problem in a follow-up pr.
##########
File path: tests/python/contrib/test_ethosu/test_legalize.py
##########
@@ -558,5 +558,193 @@ def verify(ext_func):
verify(mod["tvmgen_default_ethosu_main_0"])
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 4], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 4], True),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_tflite_binary_elemwise_legalize(
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ reversed_operands,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, x, y):
+ if operator_type == "ADD":
+ op = tf.math.add(x, y)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(x, y)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(x, y)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(x, y)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(x, y)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32), tf.TensorSpec(ifm2_shape, dtype=tf.float32)
+ )
+
+ # Convert the model
+ def representative_dataset():
+ for _ in range(100):
+ data = np.random.rand(*tuple(ifm_shape))
+ data2 = np.random.rand(*tuple(ifm2_shape)) * 2
+ yield [data.astype(np.float32), data2.astype(np.float32)]
+
+ converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
+ converter.optimizations = [tf.lite.Optimize.DEFAULT]
+ converter.representative_dataset = representative_dataset
+ converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+ converter.inference_input_type = tf.int8
+ converter.inference_output_type = tf.int8
+ tflite_model = converter.convert()
+ return tflite_model
+
+ def verify(ext_func):
+ out_shape = ifm2_shape if reversed_operands else ifm_shape
+ shapes = [ifm_shape, ifm2_shape]
+ ifm_index, ifm2_index = (1, 0) if reversed_operands else (0, 1)
+ op = ext_func.body
+ assert list(op.args[0].checked_type.shape) == shapes[ifm_index]
+ assert list(op.args[1].checked_type.shape) == shapes[ifm2_index]
+ assert op.args[0].checked_type.dtype == dtype
+ assert list(op.checked_type.shape) == out_shape
+ assert op.checked_type.dtype == dtype
+ assert op.attrs.operator_type == operator_type
+ assert op.attrs.reversed_operands == reversed_operands
+ if activation_function == "RELU":
+ assert str(op.attrs.activation) == "CLIP"
+
+ if operator_type == "ADD":
+ rewriter = legalize.AddRewriter()
+ pattern_table = [
+ (
+ ethosu.AddParams.composite_name,
+ ethosu.qnn_add_pattern(),
+ lambda pat: ethosu.AddParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "SUB":
+ rewriter = legalize.SubRewriter()
+ pattern_table = [
+ (
+ ethosu.SubParams.composite_name,
+ ethosu.qnn_subtract_pattern(),
+ lambda pat: ethosu.SubParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MUL":
+ rewriter = legalize.MulRewriter()
+ pattern_table = [
+ (
+ ethosu.MulParams.composite_name,
+ ethosu.qnn_mul_pattern(),
+ lambda pat: ethosu.MulParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MIN":
+ rewriter = legalize.MinRewriter()
+ pattern_table = [
+ (
+ ethosu.MinParams.composite_name,
+ ethosu.minimum_pattern(),
+ lambda pat: ethosu.MinParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MAX":
+ rewriter = legalize.MaxRewriter()
+ pattern_table = [
+ (
+ ethosu.MaxParams.composite_name,
+ ethosu.maximum_pattern(),
+ lambda pat: ethosu.MaxParams(pat).is_valid(),
+ ),
+ ]
+
+ tflite_graph = create_tflite_graph()
+ tflite_model = tflite.Model.Model.GetRootAsModel(tflite_graph, 0)
+
+ mod, _ = relay.frontend.from_tflite(
+ tflite_model,
+ shape_dict={"x": ifm_shape, "y": ifm2_shape},
+ dtype_dict={"x": dtype, "y": dtype},
+ )
+ mod = partition_ethosu_by_table(mod, pattern_table)
+
+ mod["tvmgen_default_ethosu_main_0"] = dataflow_pattern.rewrite(
+ rewriter, mod["tvmgen_default_ethosu_main_0"]
+ )
+ verify(mod["tvmgen_default_ethosu_main_0"])
+
+
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 4], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 4], True),
+ ],
+)
+def test_ethosu_left_shift_binary_elemwise_legalize(ifm_shape, ifm2_shape, reversed_operands):
+ dtype = "int32"
+ operator_type = "SHL"
+
+ def create_graph():
+ input1 = relay.var("x1", shape=ifm_shape, dtype=dtype)
+ input2 = relay.var("x2", shape=ifm2_shape, dtype=dtype)
+ c1 = relay.left_shift(input1, input2)
Review comment:
@ekalda Currently, all the NPU primitive operators are covered.
Because the NPU left shift operator does the "actual" left shift, we legalize the Relay left shift operator.
Conversely, the NPU right shift operator is actually a rounding right shift operator and so the Relay right shift operator cannot be legalized.
@lhutton1 The Relay right shift operator cannot be legalized to a rounding right shift followed by an element-wise subtract of 1.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r743013392
##########
File path: python/tvm/relay/op/contrib/ethosu.py
##########
@@ -458,6 +458,348 @@ def qnn_avgpool2d_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
return pattern
+class BinaryElementwiseParams:
+ """
+ This class will parse a call to a ethosu.binary_elementwise composite function
+ and extract the parameter information.
+ """
+
+ def __init__(self, func_body: Call, operator_type: str, has_quantization_parameters: bool):
+ clip = None
+ if str(func_body.op) == "clip":
+ clip = func_body
+ binary_op = clip.args[0]
+ else:
+ binary_op = func_body
+
+ layout = "NHWC"
+
+ if has_quantization_parameters:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm_zero_point.value],
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm2_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm2_zero_point.value],
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ofm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ofm_zero_point.value],
+ )
+ else:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ )
+ self.activation = clip
+ self.operator_type = operator_type
+
+ def brodcastable(x, y):
+ for i in range(1, 4):
+ if x.shape[i] == y.shape[i] or y.shape[i] == 1:
+ continue
+ return False
+ return True
+
+ if brodcastable(self.ifm, self.ifm2):
+ self.reversed_operands = False
+ self.valid_broadcast = True
+ elif brodcastable(self.ifm2, self.ifm):
+ self.reversed_operands = True
+ self.ifm, self.ifm2 = self.ifm2, self.ifm
+ self.valid_broadcast = True
+ else:
+ self.valid_broadcast = False
+
+ def is_valid(self):
+ """
+ This function checks whether BinaryElementwise has compatible attributes with the NPU
+ """
+ if np.dtype(self.ofm) == np.int32 and self.activation is not None:
+ return False
+ if len(self.ifm.shape) != 4 or len(self.ifm2.shape) != 4:
+ return False
+ if self.ifm.shape[0] != 1 or self.ifm2.shape[0] != 1:
+ return False
+ if not self.valid_broadcast:
+ return False
+ return True
+
+
+class AddParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Add composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.add"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "ADD", True)
+
+ def is_valid(self):
+ """
+ This function checks whether Add has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8, np.int32]
+ ):
+ return False
+ return True
+
+
+def qnn_add_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for qnn.add with optional fused RELU activation.
+ """
+ pattern = is_op("qnn.add")(
+ wildcard(),
+ wildcard(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ )
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class SubParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Sub composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.sub"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "SUB", True)
+
+ def is_valid(self):
+ """
+ This function checks whether Sub has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8, np.int32]
+ ):
+ return False
+ return True
+
+
+def qnn_subtract_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for qnn.subtract with optional fused RELU activation.
+ """
+ pattern = is_op("qnn.subtract")(
+ wildcard(),
+ wildcard(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ )
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class MulParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Mul composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.mul"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "MUL", True)
+
+ def is_valid(self):
+ """
+ This function checks whether Mul has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8, np.int32]
+ ):
+ return False
+ return True
+
+
+def qnn_mul_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for qnn.mul with optional fused RELU activation.
+ """
+ pattern = is_op("qnn.mul")(
+ wildcard(),
+ wildcard(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ is_constant(),
+ )
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class MinParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Min composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.min"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "MIN", False)
+
+ def is_valid(self):
+ """
+ This function checks whether Min has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if self.ifm.dtype != self.ifm2.dtype:
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8]
+ ):
+ return False
+ return True
+
+
+def minimum_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for minimum with optional fused RELU activation.
+ """
+ pattern = is_op("minimum")(wildcard(), wildcard())
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class MaxParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Max composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.max"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "MAX", False)
+
+ def is_valid(self):
+ """
+ This function checks whether Max has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if self.ifm.dtype != self.ifm2.dtype:
+ return False
+ if not check_valid_dtypes(
+ [self.ifm, self.ifm2, self.ofm], supported_dtypes=[np.uint8, np.int8]
+ ):
+ return False
+ return True
+
+
+def maximum_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for maximum with optional fused RELU activation.
+ """
+ pattern = is_op("maximum")(wildcard(), wildcard())
+ pattern = pattern.optional(is_op("clip"))
+ return pattern
+
+
+class ShrParams(BinaryElementwiseParams):
+ """
+ This class will parse a call to a ethosu.binary_elementwise Shr composite function
+ and extract the parameter information.
+ """
+
+ composite_name = "ethosu.shr"
+
+ def __init__(self, func_body: Call):
+ BinaryElementwiseParams.__init__(self, func_body, "SHR", False)
+
+ def is_valid(self):
+ """
+ This function checks whether Shr has compatible attributes with the NPU
+ """
+ if not super().is_valid():
+ return False
+ if not check_valid_dtypes([self.ifm, self.ifm2], supported_dtypes=[np.int32]):
+ return False
+ if not check_valid_dtypes([self.ofm], supported_dtypes=[np.uint8, np.int8, np.int32]):
+ return False
+ return True
+
+
+def shr_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
+ """
+ This function creates the pattern for right_shift.
+ """
+ pattern = is_op("right_shift")(wildcard(), wildcard())
Review comment:
Removed
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r745465653
##########
File path: python/tvm/relay/op/contrib/ethosu.py
##########
@@ -458,6 +458,316 @@ def qnn_avgpool2d_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
return pattern
+class BinaryElementwiseParams:
+ """
+ This class will parse a call to a ethosu.binary_elementwise composite function
+ and extract the parameter information.
+ """
+
+ def __init__(self, func_body: Call, operator_type: str, has_quantization_parameters: bool):
+ clip = None
+ if str(func_body.op) == "clip":
+ clip = func_body
+ binary_op = clip.args[0]
+ else:
+ binary_op = func_body
+
+ layout = "NHWC"
+
+ if has_quantization_parameters:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm_zero_point.value],
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm2_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm2_zero_point.value],
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ofm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ofm_zero_point.value],
+ )
+ else:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ )
+ self.activation = clip
+ self.operator_type = operator_type
+
+ def brodcastable(x, y):
Review comment:
I replaced it with can_broadcast.
##########
File path: python/tvm/relay/backend/contrib/ethosu/op/binary_elementwise.py
##########
@@ -0,0 +1,206 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=unused-argument
+"""Relay operators for binary elementwise operators for Arm(R) Ethos(TM)-U NPU"""
+import tvm
+from tvm.relay.op import _make
+from tvm.topi.generic import schedule_injective
+from tvm.relay.op.op import OpStrategy
+from tvm.relay.op import strategy as _strategy
+
+from ..te import binary_elementwise_compute
+
+
+def _extract_ethosu_binary_elementwise_params(attrs, args):
+ """Get the parameters necessary to construct a ethosu_binary_elementwise compute TE
+ from a ethosu_binary_elementwise Relay call."""
+ ifm = args[0]
+ ifm2 = args[1]
+ lut = args[2]
+ operator_type = attrs.operator_type
+ ifm_scale = attrs.ifm_scale
+ ifm_zero_point = attrs.ifm_zero_point
+ ifm2_scale = attrs.ifm2_scale
+ ifm2_zero_point = attrs.ifm2_zero_point
+ ofm_scale = attrs.ofm_scale
+ ofm_zero_point = attrs.ofm_zero_point
+ ofm_channels = attrs.ofm_channels
+ reversed_operands = attrs.reversed_operands
+ activation = attrs.activation
+ clip_min = attrs.clip_min
+ clip_max = attrs.clip_max
+ ifm_layout = attrs.ifm_layout
+ ifm2_layout = attrs.ifm2_layout
+ ofm_layout = attrs.ofm_layout
+
+ return (
+ ifm,
+ ifm2,
+ lut,
+ operator_type,
+ ifm_scale,
+ ifm_zero_point,
+ ifm2_scale,
+ ifm2_zero_point,
+ ofm_scale,
+ ofm_zero_point,
+ ofm_channels,
+ reversed_operands,
+ activation,
+ clip_min,
+ clip_max,
+ ifm_layout,
+ ifm2_layout,
+ ofm_layout,
+ )
+
+
+@tvm.ir.register_op_attr("contrib.ethosu.binary_elementwise", "FTVMCompute")
+def create_ethosu_binary_elementwise_compute(attrs, args, out_type):
+ """Create an ethosu_binary_elementwise compute op."""
+ params = _extract_ethosu_binary_elementwise_params(attrs, args)
+ op = binary_elementwise_compute(*params)
+ return [op]
+
+
+@tvm.ir.register_op_attr("contrib.ethosu.binary_elementwise", "FTVMStrategy")
+def binary_elementwise_strategy_ethosu(attrs, inputs, out_type, target):
+ strategy = OpStrategy()
+ strategy.add_implementation(
+ create_ethosu_binary_elementwise_compute,
+ _strategy.wrap_topi_schedule(schedule_injective),
+ name="ethosu_binary_elementwise",
+ )
+ return strategy
+
+
+def ethosu_binary_elementwise(
+ ifm: tvm.relay.Expr,
+ ifm2: tvm.relay.Expr,
+ lut: tvm.relay.Expr,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ ofm_dtype: str,
+ activation: str = "NONE",
+ clip_min: int = 0,
+ clip_max: int = 0,
+ ifm_layout: str = "NHWC",
+ ifm2_layout: str = "NHWC",
+ ofm_layout: str = "NHWC",
Review comment:
Done
##########
File path: python/tvm/relay/backend/contrib/ethosu/te/binary_elementwise.py
##########
@@ -0,0 +1,169 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=invalid-name,unused-argument
+"""Tensor Expressions for binary_elementwise"""
+import operator
+from tvm import te
+from .dma import dma_ofm_compute, dma_ifm_compute
+
+
+def binary_elementwise_compute(
+ ifm: te.Tensor,
+ ifm2: te.Tensor,
+ lut: te.Tensor,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ activation: str,
+ clip_min: int,
+ clip_max: int,
+ ifm_layout: str,
+ ifm2_layout: str,
+ ofm_layout: str,
+) -> te.Tensor:
+ """A compute operator representing the capabilities of binary_elementwise for the NPU.
+
+ Parameters
+ ----------
+ ifm : te.Tensor
+ The Input Feature Map tensor (IFM).
+ ifm2 : te.Tensor
+ The Input Feature Map tensor 1 (IFM2).
Review comment:
Done
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] lhutton1 commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
lhutton1 commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r744649855
##########
File path: python/tvm/relay/backend/contrib/ethosu/te/binary_elementwise.py
##########
@@ -0,0 +1,169 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=invalid-name,unused-argument
+"""Tensor Expressions for binary_elementwise"""
+import operator
+from tvm import te
+from .dma import dma_ofm_compute, dma_ifm_compute
+
+
+def binary_elementwise_compute(
+ ifm: te.Tensor,
+ ifm2: te.Tensor,
+ lut: te.Tensor,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ activation: str,
+ clip_min: int,
+ clip_max: int,
+ ifm_layout: str,
+ ifm2_layout: str,
+ ofm_layout: str,
+) -> te.Tensor:
+ """A compute operator representing the capabilities of binary_elementwise for the NPU.
+
+ Parameters
+ ----------
+ ifm : te.Tensor
+ The Input Feature Map tensor (IFM).
+ ifm2 : te.Tensor
+ The Input Feature Map tensor 1 (IFM2).
Review comment:
```suggestion
The Input Feature Map tensor 2 (IFM2).
```
##########
File path: python/tvm/relay/backend/contrib/ethosu/te/binary_elementwise.py
##########
@@ -0,0 +1,169 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=invalid-name,unused-argument
+"""Tensor Expressions for binary_elementwise"""
+import operator
+from tvm import te
+from .dma import dma_ofm_compute, dma_ifm_compute
+
+
+def binary_elementwise_compute(
+ ifm: te.Tensor,
+ ifm2: te.Tensor,
+ lut: te.Tensor,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ activation: str,
+ clip_min: int,
+ clip_max: int,
+ ifm_layout: str,
+ ifm2_layout: str,
+ ofm_layout: str,
+) -> te.Tensor:
+ """A compute operator representing the capabilities of binary_elementwise for the NPU.
+
+ Parameters
+ ----------
+ ifm : te.Tensor
+ The Input Feature Map tensor (IFM).
+ ifm2 : te.Tensor
+ The Input Feature Map tensor 1 (IFM2).
+ lut : te.Tensor
+ The look-up table values to use if activation = "LUT".
+ operator_type: str
+ The type of the binary elementwise operator.
+ "ADD"
+ "SUB"
+ "MUL"
+ "MIN"
+ "MAX"
+ "SHR"
+ "SHL"
+ ifm_scale : float
+ The quantization scale for the Input Feature Map tensor.
+ ifm_zero_point : int
+ The quantization zero point for the Input Feature Map tensor.
+ ifm2_scale : float
+ The quantization scale for the Input Feature Map tensor 2.
+ ifm2_zero_point : int
+ The quantization zero point for the Input Feature Map tensor 1.
+ ofm_scale : float
+ The quantization scale for the Output Feature Map tensor.
+ ofm_zero_point : int
+ The quantization zero point for the Output Feature Map tensor.
+ ofm_channels : int
+ The number of the Output Feature Map channels.
+ reversed_operands : bool
+ True if IFM2 is the first operand and IFM is the second operand.
+ activation : str
+ The activation function to use.
+ "NONE" - no activation function.
+ "CLIP" - clip the output between clip_min and clip_max.
+ "TANH" - tanh activation function.
+ "SIGMOID" - sigmoid activation function.
+ "LUT" - use a look-up table to perform the activation function.
+ Available activations for activation type:
+ {int8, uint8}: "NONE", "CLIP", "TANH", "SIGMOID", "LUT"
+ {int32}: "NONE"
+ clip_min : int
+ The minimum clipping value if activation = "CLIP".
+ clip_max : int
+ The maximum clipping value if activation = "CLIP".
+ ifm_layout : str, optional
+ The layout of the Input Feature Map tensor. Can be "NHWC" or "NHCWB16".
+ ifm2_layout : str, optional
+ The layout of the Input Feature Map tensor 1. Can be "NHWC" or "NHCWB16".
+ ofm_layout : str, optional
+ The layout of the Output Feature Map tensor. Can be "NHWC" or "NHCWB16".
+
+ Returns
+ -------
+ te.Tensor
+ The OFM tensor.
Review comment:
Nit: Output Feature Map
##########
File path: python/tvm/relay/backend/contrib/ethosu/op/binary_elementwise.py
##########
@@ -0,0 +1,206 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=unused-argument
+"""Relay operators for binary elementwise operators for Arm(R) Ethos(TM)-U NPU"""
+import tvm
+from tvm.relay.op import _make
+from tvm.topi.generic import schedule_injective
+from tvm.relay.op.op import OpStrategy
+from tvm.relay.op import strategy as _strategy
+
+from ..te import binary_elementwise_compute
+
+
+def _extract_ethosu_binary_elementwise_params(attrs, args):
+ """Get the parameters necessary to construct a ethosu_binary_elementwise compute TE
+ from a ethosu_binary_elementwise Relay call."""
+ ifm = args[0]
+ ifm2 = args[1]
+ lut = args[2]
+ operator_type = attrs.operator_type
+ ifm_scale = attrs.ifm_scale
+ ifm_zero_point = attrs.ifm_zero_point
+ ifm2_scale = attrs.ifm2_scale
+ ifm2_zero_point = attrs.ifm2_zero_point
+ ofm_scale = attrs.ofm_scale
+ ofm_zero_point = attrs.ofm_zero_point
+ ofm_channels = attrs.ofm_channels
+ reversed_operands = attrs.reversed_operands
+ activation = attrs.activation
+ clip_min = attrs.clip_min
+ clip_max = attrs.clip_max
+ ifm_layout = attrs.ifm_layout
+ ifm2_layout = attrs.ifm2_layout
+ ofm_layout = attrs.ofm_layout
+
+ return (
+ ifm,
+ ifm2,
+ lut,
+ operator_type,
+ ifm_scale,
+ ifm_zero_point,
+ ifm2_scale,
+ ifm2_zero_point,
+ ofm_scale,
+ ofm_zero_point,
+ ofm_channels,
+ reversed_operands,
+ activation,
+ clip_min,
+ clip_max,
+ ifm_layout,
+ ifm2_layout,
+ ofm_layout,
+ )
+
+
+@tvm.ir.register_op_attr("contrib.ethosu.binary_elementwise", "FTVMCompute")
+def create_ethosu_binary_elementwise_compute(attrs, args, out_type):
+ """Create an ethosu_binary_elementwise compute op."""
+ params = _extract_ethosu_binary_elementwise_params(attrs, args)
+ op = binary_elementwise_compute(*params)
+ return [op]
+
+
+@tvm.ir.register_op_attr("contrib.ethosu.binary_elementwise", "FTVMStrategy")
+def binary_elementwise_strategy_ethosu(attrs, inputs, out_type, target):
+ strategy = OpStrategy()
+ strategy.add_implementation(
+ create_ethosu_binary_elementwise_compute,
+ _strategy.wrap_topi_schedule(schedule_injective),
+ name="ethosu_binary_elementwise",
+ )
+ return strategy
+
+
+def ethosu_binary_elementwise(
+ ifm: tvm.relay.Expr,
+ ifm2: tvm.relay.Expr,
+ lut: tvm.relay.Expr,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ ofm_dtype: str,
+ activation: str = "NONE",
+ clip_min: int = 0,
+ clip_max: int = 0,
+ ifm_layout: str = "NHWC",
+ ifm2_layout: str = "NHWC",
+ ofm_layout: str = "NHWC",
Review comment:
Nit: Optional[type]
##########
File path: python/tvm/relay/backend/contrib/ethosu/te/binary_elementwise.py
##########
@@ -0,0 +1,169 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=invalid-name,unused-argument
+"""Tensor Expressions for binary_elementwise"""
+import operator
+from tvm import te
+from .dma import dma_ofm_compute, dma_ifm_compute
+
+
+def binary_elementwise_compute(
+ ifm: te.Tensor,
+ ifm2: te.Tensor,
+ lut: te.Tensor,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ activation: str,
+ clip_min: int,
+ clip_max: int,
+ ifm_layout: str,
+ ifm2_layout: str,
+ ofm_layout: str,
+) -> te.Tensor:
+ """A compute operator representing the capabilities of binary_elementwise for the NPU.
+
+ Parameters
+ ----------
+ ifm : te.Tensor
+ The Input Feature Map tensor (IFM).
+ ifm2 : te.Tensor
+ The Input Feature Map tensor 1 (IFM2).
+ lut : te.Tensor
+ The look-up table values to use if activation = "LUT".
+ operator_type: str
+ The type of the binary elementwise operator.
+ "ADD"
+ "SUB"
+ "MUL"
+ "MIN"
+ "MAX"
+ "SHR"
+ "SHL"
+ ifm_scale : float
+ The quantization scale for the Input Feature Map tensor.
+ ifm_zero_point : int
+ The quantization zero point for the Input Feature Map tensor.
+ ifm2_scale : float
+ The quantization scale for the Input Feature Map tensor 2.
+ ifm2_zero_point : int
+ The quantization zero point for the Input Feature Map tensor 1.
+ ofm_scale : float
+ The quantization scale for the Output Feature Map tensor.
+ ofm_zero_point : int
+ The quantization zero point for the Output Feature Map tensor.
+ ofm_channels : int
+ The number of the Output Feature Map channels.
+ reversed_operands : bool
+ True if IFM2 is the first operand and IFM is the second operand.
+ activation : str
+ The activation function to use.
+ "NONE" - no activation function.
+ "CLIP" - clip the output between clip_min and clip_max.
+ "TANH" - tanh activation function.
+ "SIGMOID" - sigmoid activation function.
+ "LUT" - use a look-up table to perform the activation function.
+ Available activations for activation type:
+ {int8, uint8}: "NONE", "CLIP", "TANH", "SIGMOID", "LUT"
+ {int32}: "NONE"
+ clip_min : int
+ The minimum clipping value if activation = "CLIP".
+ clip_max : int
+ The maximum clipping value if activation = "CLIP".
+ ifm_layout : str, optional
+ The layout of the Input Feature Map tensor. Can be "NHWC" or "NHCWB16".
+ ifm2_layout : str, optional
+ The layout of the Input Feature Map tensor 1. Can be "NHWC" or "NHCWB16".
Review comment:
```suggestion
The layout of the Input Feature Map tensor 2. Can be "NHWC" or "NHCWB16".
```
##########
File path: src/relay/op/contrib/ethosu/binary_elementwise.cc
##########
@@ -0,0 +1,299 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file src/relay/op/contrib/ethosu/binary_elementwise.cc
+ * \brief Binary elementwise operators definitions for the Arm(R) Ethos(TM)-U NPU.
+ */
+#include <tvm/relay/op.h>
+
+#include "common.h"
+
+namespace tvm {
+namespace relay {
+namespace op {
+namespace contrib {
+namespace ethosu {
+
+/*! \brief Attributes used by the Ethos(TM)-U NPU binary elementwise operators */
+struct EthosuBinaryElementwiseAttrs : public tvm::AttrsNode<EthosuBinaryElementwiseAttrs> {
+ String operator_type;
+ double ifm_scale;
+ int ifm_zero_point;
+ double ifm2_scale;
+ int ifm2_zero_point;
+ double ofm_scale;
+ int ofm_zero_point;
+ IndexExpr ofm_channels;
+ bool reversed_operands;
+ String activation;
+ int clip_min;
+ int clip_max;
+ String ifm_layout;
+ String ifm2_layout;
+ String ofm_layout;
+ String ofm_dtype;
+
+ TVM_DECLARE_ATTRS(EthosuBinaryElementwiseAttrs, "relay.attrs.EthosuBinaryElementwiseAttrs") {
+ TVM_ATTR_FIELD(operator_type)
+ .describe(
+ "The type of the binary elementwise operator."
+ "'ADD'"
+ "'SUB'"
+ "'MUL'"
+ "'MIN'"
+ "'MAX'"
+ "'SHR'"
+ "'SHL'");
+ TVM_ATTR_FIELD(ifm_scale).describe("The quantization scale for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm2_scale)
+ .describe("The quantization scale for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ifm2_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ofm_scale).describe("The quantization scale for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_zero_point)
+ .describe("The quantization zero point for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_channels).describe("The number of the Output Feature Map channels.");
+ TVM_ATTR_FIELD(reversed_operands)
+ .describe("True if IFM2 is the first operand and IFM is the second operand.")
+ .set_default(false);
+ TVM_ATTR_FIELD(activation)
+ .describe(
+ "The activation function to use. "
+ "'NONE' - no activation function. "
+ "'CLIP' - clip the output between clip_min and clip_max. "
+ "'TANH' - tanh activation function. "
+ "'SIGMOID' - sigmoid activation function. "
+ "'LUT' - use a look-up table to perform the activation function."
+ "Available activations for activation type:"
+ "{int8, uint8}: 'NONE', 'CLIP', 'TANH', 'SIGMOID', 'LUT'"
+ "{int32}: 'NONE'")
+ .set_default("NONE");
+ TVM_ATTR_FIELD(clip_min)
+ .describe("The minimum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(clip_max)
+ .describe("The maximum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(ifm_layout)
+ .describe("The layout of the Input Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ifm2_layout)
+ .describe("The layout of the Input Feature Map tensor 2. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_layout)
+ .describe("The layout of the Output Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_dtype)
+ .describe(
+ "The Output Feature Map tensor type."
+ "MUL, ADD, SUB {IFM}->{OFM}:"
+ " {uint8, int8 int32} -> {uint8, int8, int32}, any pairing"
+ "MAX, MIN:"
+ " IFM and OFM must be of the same type, one of:"
+ " {int8, uint8}"
+ "SHR {IFM}->{OFM}:"
+ " {int32}->{int8, uint8, int32}, any pairing"
+ "SHL:"
+ " {int32}->{int32} only")
+ .set_default("NHWC");
+ }
+};
+
+TVM_REGISTER_NODE_TYPE(EthosuBinaryElementwiseAttrs);
+
+bool EthosuBinaryElementwiseRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
+ const TypeReporter& reporter) {
+ int ifm_index = 0;
+ int ifm2_index = 1;
+ int result_index = 3;
+ ICHECK_EQ(types.size(), result_index + 1);
+
+ const auto* ifm = types[ifm_index].as<TensorTypeNode>();
+ const auto* ifm2 = types[ifm2_index].as<TensorTypeNode>();
+ if (ifm == nullptr) return false;
+ if (ifm2 == nullptr) return false;
+
+ const auto* param = attrs.as<EthosuBinaryElementwiseAttrs>();
+ ICHECK(param != nullptr) << "EthosuBinaryElementwiseAttrs cannot be nullptr.";
+
+ String operator_type = param->operator_type;
+ auto ifm_dtype = ifm->dtype;
+ auto ifm2_dtype = ifm2->dtype;
+ DataType ofm_dtype;
+
+ if (param->ofm_dtype == "int8") {
+ ofm_dtype = DataType::Int(8);
+ } else if (param->ofm_dtype == "uint8") {
+ ofm_dtype = DataType::UInt(8);
+ } else if (param->ofm_dtype == "int32") {
+ ofm_dtype = DataType::Int(32);
+ }
+
+ if (ifm_dtype != ifm2_dtype) {
+ reporter->GetDiagCtx().EmitFatal(Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise "
+ << "type for ifm2 be the same of ifm but was " << ifm2_dtype
+ << " instead of " << ifm_dtype);
+ return false;
+ }
+
+ if (operator_type == "ADD" || operator_type == "SUB" || operator_type == "MUL") {
+ if (ifm_dtype != DataType::UInt(8) && ifm_dtype != DataType::Int(8) &&
+ ifm_dtype != DataType::Int(32)) {
+ reporter->GetDiagCtx().EmitFatal(
+ Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise " << operator_type
+ << " type(uint8) or type(int8) or type(int32) for ifm but was " << ifm_dtype);
+ return false;
+ }
+ if (ofm_dtype != DataType::UInt(8) && ofm_dtype != DataType::Int(8) &&
+ ofm_dtype != DataType::Int(32)) {
+ reporter->GetDiagCtx().EmitFatal(
+ Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise " << operator_type
+ << " type(uint8) or type(int8) or type(int32) for ofm but was " << ofm_dtype);
+ return false;
+ }
+ } else if (operator_type == "MIN" || operator_type == "MAX") {
+ if (ifm_dtype != DataType::UInt(8) && ifm_dtype != DataType::Int(8)) {
+ reporter->GetDiagCtx().EmitFatal(
+ Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise " << operator_type
+ << " type(uint8) or type(int8) for ifm but was " << ifm_dtype);
+ return false;
+ }
+ if (ifm_dtype != ofm_dtype) {
+ reporter->GetDiagCtx().EmitFatal(Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise "
+ << operator_type
+ << " type for ofm be the same of ifm but was " << ofm_dtype
+ << " instead of " << ifm_dtype);
+ return false;
+ }
+ } else if (operator_type == "SHR") {
+ if (ifm_dtype != DataType::Int(32)) {
+ reporter->GetDiagCtx().EmitFatal(Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise "
+ << operator_type << " type(int32) for ifm but was "
+ << ifm_dtype);
+ return false;
+ }
+ if (ofm_dtype != DataType::UInt(8) && ofm_dtype != DataType::Int(8) &&
+ ofm_dtype != DataType::Int(32)) {
+ reporter->GetDiagCtx().EmitFatal(
+ Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise " << operator_type
+ << " type(uint8) or type(int8) or type(int32) for ofm but was " << ofm_dtype);
+ return false;
+ }
+ } else if (operator_type == "SHL") {
+ if (ifm_dtype != DataType::Int(32)) {
+ reporter->GetDiagCtx().EmitFatal(Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise "
+ << operator_type << " type(int32) for ifm but was "
+ << ifm_dtype);
+
+ return false;
+ }
+ if (ofm_dtype != DataType::Int(32)) {
+ reporter->GetDiagCtx().EmitFatal(Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise "
+ << operator_type << " type(int32) for ofm but was "
+ << ofm_dtype);
+ return false;
+ }
+ } else {
+ reporter->GetDiagCtx().EmitFatal(
+ Diagnostic::Error(reporter->GetSpan())
+ << "Invalid operator: expected ethosu_binary_elementwise 'ADD' or 'SUB' or 'MUL' or "
+ << "'MIN' or 'MAX' or 'SHR' or 'SHL' for operator_type but was " << param->operator_type);
+ return false;
+ }
+
+ // Assign ofm type
+ auto ofm_shape = EthosuInferBinaryElementwiseOutputShape(ifm->shape, param->ifm_layout,
+ param->ofm_layout, param->ofm_channels);
+ reporter->Assign(types[result_index], TensorType(ofm_shape, ofm_dtype));
+ return true;
+}
+
+Expr MakeEthosuBinaryElementwise(Expr ifm, Expr ifm2, Expr lut, String operator_type,
+ double ifm_scale, int ifm_zero_point, double ifm2_scale,
+ int ifm2_zero_point, double ofm_scale, int ofm_zero_point,
+ IndexExpr ofm_channels, bool reversed_operands, String activation,
+ int clip_min, int clip_max, String ifm_layout, String ifm2_layout,
+ String ofm_layout, String ofm_dtype) {
+ auto attrs = make_object<EthosuBinaryElementwiseAttrs>();
+
+ attrs->operator_type = std::move(operator_type);
+ attrs->ifm_scale = ifm_scale;
+ attrs->ifm_zero_point = ifm_zero_point;
+ attrs->ifm2_scale = ifm2_scale;
+ attrs->ifm2_zero_point = ifm2_zero_point;
+ attrs->ofm_scale = ofm_scale;
+ attrs->ofm_zero_point = ofm_zero_point;
+ attrs->ofm_channels = ofm_channels;
Review comment:
```suggestion
attrs->ofm_channels = std::move(ofm_channels);
```
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -343,5 +343,256 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_ethosu_binary_elementwise(
+ accel_type,
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ tf.config.run_functions_eagerly(True)
Review comment:
Just curious, why do we need to enable eager execution?
##########
File path: src/relay/op/contrib/ethosu/binary_elementwise.cc
##########
@@ -0,0 +1,299 @@
+/*
+ * Licensed to the Apache Software Foundation (ASF) under one
+ * or more contributor license agreements. See the NOTICE file
+ * distributed with this work for additional information
+ * regarding copyright ownership. The ASF licenses this file
+ * to you under the Apache License, Version 2.0 (the
+ * "License"); you may not use this file except in compliance
+ * with the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing,
+ * software distributed under the License is distributed on an
+ * "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+ * KIND, either express or implied. See the License for the
+ * specific language governing permissions and limitations
+ * under the License.
+ */
+
+/*!
+ * \file src/relay/op/contrib/ethosu/binary_elementwise.cc
+ * \brief Binary elementwise operators definitions for the Arm(R) Ethos(TM)-U NPU.
+ */
+#include <tvm/relay/op.h>
+
+#include "common.h"
+
+namespace tvm {
+namespace relay {
+namespace op {
+namespace contrib {
+namespace ethosu {
+
+/*! \brief Attributes used by the Ethos(TM)-U NPU binary elementwise operators */
+struct EthosuBinaryElementwiseAttrs : public tvm::AttrsNode<EthosuBinaryElementwiseAttrs> {
+ String operator_type;
+ double ifm_scale;
+ int ifm_zero_point;
+ double ifm2_scale;
+ int ifm2_zero_point;
+ double ofm_scale;
+ int ofm_zero_point;
+ IndexExpr ofm_channels;
+ bool reversed_operands;
+ String activation;
+ int clip_min;
+ int clip_max;
+ String ifm_layout;
+ String ifm2_layout;
+ String ofm_layout;
+ String ofm_dtype;
+
+ TVM_DECLARE_ATTRS(EthosuBinaryElementwiseAttrs, "relay.attrs.EthosuBinaryElementwiseAttrs") {
+ TVM_ATTR_FIELD(operator_type)
+ .describe(
+ "The type of the binary elementwise operator."
+ "'ADD'"
+ "'SUB'"
+ "'MUL'"
+ "'MIN'"
+ "'MAX'"
+ "'SHR'"
+ "'SHL'");
+ TVM_ATTR_FIELD(ifm_scale).describe("The quantization scale for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor.");
+ TVM_ATTR_FIELD(ifm2_scale)
+ .describe("The quantization scale for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ifm2_zero_point)
+ .describe("The quantization zero point for the Input Feature Map tensor 2.");
+ TVM_ATTR_FIELD(ofm_scale).describe("The quantization scale for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_zero_point)
+ .describe("The quantization zero point for the Output Feature Map tensor.");
+ TVM_ATTR_FIELD(ofm_channels).describe("The number of the Output Feature Map channels.");
+ TVM_ATTR_FIELD(reversed_operands)
+ .describe("True if IFM2 is the first operand and IFM is the second operand.")
+ .set_default(false);
+ TVM_ATTR_FIELD(activation)
+ .describe(
+ "The activation function to use. "
+ "'NONE' - no activation function. "
+ "'CLIP' - clip the output between clip_min and clip_max. "
+ "'TANH' - tanh activation function. "
+ "'SIGMOID' - sigmoid activation function. "
+ "'LUT' - use a look-up table to perform the activation function."
+ "Available activations for activation type:"
+ "{int8, uint8}: 'NONE', 'CLIP', 'TANH', 'SIGMOID', 'LUT'"
+ "{int32}: 'NONE'")
+ .set_default("NONE");
+ TVM_ATTR_FIELD(clip_min)
+ .describe("The minimum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(clip_max)
+ .describe("The maximum clipping value if activation = 'CLIP'.")
+ .set_default(0);
+ TVM_ATTR_FIELD(ifm_layout)
+ .describe("The layout of the Input Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ifm2_layout)
+ .describe("The layout of the Input Feature Map tensor 2. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_layout)
+ .describe("The layout of the Output Feature Map tensor. Can be 'NHWC' or 'NHCWB16'.")
+ .set_default("NHWC");
+ TVM_ATTR_FIELD(ofm_dtype)
+ .describe(
+ "The Output Feature Map tensor type."
+ "MUL, ADD, SUB {IFM}->{OFM}:"
+ " {uint8, int8 int32} -> {uint8, int8, int32}, any pairing"
+ "MAX, MIN:"
+ " IFM and OFM must be of the same type, one of:"
+ " {int8, uint8}"
+ "SHR {IFM}->{OFM}:"
+ " {int32}->{int8, uint8, int32}, any pairing"
+ "SHL:"
+ " {int32}->{int32} only")
+ .set_default("NHWC");
+ }
+};
+
+TVM_REGISTER_NODE_TYPE(EthosuBinaryElementwiseAttrs);
+
+bool EthosuBinaryElementwiseRel(const Array<Type>& types, int num_inputs, const Attrs& attrs,
+ const TypeReporter& reporter) {
+ int ifm_index = 0;
+ int ifm2_index = 1;
+ int result_index = 3;
Review comment:
Can these be made `const`?
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -343,5 +343,256 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_ethosu_binary_elementwise(
+ accel_type,
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ tf.config.run_functions_eagerly(True)
+
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, lhs, rhs):
+ if operator_type == "ADD":
+ op = tf.math.add(lhs, rhs)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(lhs, rhs)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(lhs, rhs)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(lhs, rhs)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(lhs, rhs)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32), tf.TensorSpec(ifm2_shape, dtype=tf.float32)
+ )
+
+ # Convert the model
+ def representative_dataset():
+ for _ in range(100):
+ data = np.random.rand(*tuple(ifm_shape))
+ data2 = np.random.rand(*tuple(ifm2_shape)) * 2
+ yield [data.astype(np.float32), data2.astype(np.float32)]
+
+ converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
+ converter.optimizations = [tf.lite.Optimize.DEFAULT]
+ converter.representative_dataset = representative_dataset
+ converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+ converter.inference_input_type = tf.int8
+ converter.inference_output_type = tf.int8
+ tflite_model = converter.convert()
+ return tflite_model
+
+ tflite_graph = create_tflite_graph()
+ tflite_model = tflite.Model.Model.GetRootAsModel(tflite_graph, 0)
+
+ mod, params = relay.frontend.from_tflite(
+ tflite_model,
+ shape_dict={"ifm": ifm_shape, "ifm2": ifm2_shape},
+ dtype_dict={"ifm": dtype, "ifm2": dtype},
+ )
+ mod = partition_for_ethosu(mod, params)
+
+ # Generate reference data
+ input_data, output_data = infra.generate_ref_data_tflite(tflite_graph)
+
+ compiled_models = infra.build_source(
+ mod,
+ input_data,
+ output_data,
+ accel_type,
+ output_tolerance=1 if operator_type == "MAX" else 0,
+ )
+
+ # Assumes only two runtime.Modules are created -- i.e. single offload module
+ imported_modules = compiled_models[0].executor_factory.lib.imported_modules
+ assert len(imported_modules) == 2
+ ethosu_module = imported_modules[0]
+
+ # Verify generated C source
+ get_cs = tvm._ffi.get_global_func("runtime.module.ethosu.getcs")
+ cmms = get_cs(ethosu_module)
+ cmms = bytes.fromhex(cmms)
+
+ infra.print_payload(cmms)
+ infra.verify_source(compiled_models, accel_type)
+
+
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+def test_ethosu_left_shift_binary_elemwise(
+ accel_type,
+ ifm_shape,
+ ifm2_shape,
+):
+ dtype = "int32"
+
+ def create_model():
+ ifm = relay.var("ifm", shape=ifm_shape, dtype=dtype)
+ ifm2 = relay.var("ifm2", shape=ifm2_shape, dtype=dtype)
+ c1 = relay.left_shift(ifm, ifm2)
+ f = relay.Function([ifm, ifm2], c1)
+ mod = tvm.IRModule()
+ mod["main"] = f
+ return mod
+
+ relay_mod = create_model()
+ mod = partition_for_ethosu(relay_mod)
+
+ # Generate reference data
+ in_min, in_max = util.get_range_for_dtype_str(dtype)
+ input_data = {
+ "ifm": np.random.randint(in_min, high=in_max, size=ifm_shape, dtype=dtype),
+ "ifm2": np.random.randint(0, high=32, size=ifm2_shape, dtype=dtype),
+ }
+ output_data = generate_ref_data(relay_mod, input_data)
+
+ compiled_models = infra.build_source(
+ mod,
+ input_data,
+ output_data,
+ accel_type,
+ )
+
+ # Assumes only two runtime.Modules are created -- i.e. single offload module
+ imported_modules = compiled_models[0].executor_factory.lib.imported_modules
+ assert len(imported_modules) == 2
+ ethosu_module = imported_modules[0]
+
+ # Verify generated C source
+ get_cs = tvm._ffi.get_global_func("runtime.module.ethosu.getcs")
+ cmms = get_cs(ethosu_module)
+ cmms = bytes.fromhex(cmms)
+
+ infra.print_payload(cmms)
+ infra.verify_source(compiled_models, accel_type)
+
+
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 1], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 1], True),
+ ],
+)
+def test_ethosu_right_shift_binary_elemwise(ifm_shape, ifm2_shape, reversed_operands, accel_type):
+ dtype = "int32"
Review comment:
Is it possible to test the other datatype combinations? e.g. int32 -> int8
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r743017967
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -343,5 +343,256 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_ethosu_binary_elementwise(
+ accel_type,
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ tf.config.run_functions_eagerly(True)
+
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, lhs, rhs):
+ if operator_type == "ADD":
+ op = tf.math.add(lhs, rhs)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(lhs, rhs)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(lhs, rhs)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(lhs, rhs)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(lhs, rhs)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32), tf.TensorSpec(ifm2_shape, dtype=tf.float32)
+ )
+
+ # Convert the model
+ def representative_dataset():
+ for _ in range(100):
+ data = np.random.rand(*tuple(ifm_shape))
+ data2 = np.random.rand(*tuple(ifm2_shape)) * 2
+ yield [data.astype(np.float32), data2.astype(np.float32)]
+
+ converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
+ converter.optimizations = [tf.lite.Optimize.DEFAULT]
+ converter.representative_dataset = representative_dataset
+ converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+ converter.inference_input_type = tf.int8
+ converter.inference_output_type = tf.int8
+ tflite_model = converter.convert()
+ return tflite_model
+
+ tflite_graph = create_tflite_graph()
+ tflite_model = tflite.Model.Model.GetRootAsModel(tflite_graph, 0)
+
+ mod, params = relay.frontend.from_tflite(
+ tflite_model,
+ shape_dict={"ifm": ifm_shape, "ifm2": ifm2_shape},
+ dtype_dict={"ifm": dtype, "ifm2": dtype},
+ )
+ mod = partition_for_ethosu(mod, params)
+
+ # Generate reference data
+ input_data, output_data = infra.generate_ref_data_tflite(tflite_graph)
+
+ compiled_models = infra.build_source(
+ mod,
+ input_data,
+ output_data,
+ accel_type,
+ output_tolerance=1 if operator_type == "MAX" else 0,
Review comment:
The Max operator with the RELU activation function is not bit accurate, it needs a tolerance of 1.
I'll investigate and address this problem in a follow-up pr.
##########
File path: tests/python/contrib/test_ethosu/test_legalize.py
##########
@@ -558,5 +558,193 @@ def verify(ext_func):
verify(mod["tvmgen_default_ethosu_main_0"])
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 4], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 4], True),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_tflite_binary_elemwise_legalize(
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ reversed_operands,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, x, y):
+ if operator_type == "ADD":
+ op = tf.math.add(x, y)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(x, y)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(x, y)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(x, y)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(x, y)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32), tf.TensorSpec(ifm2_shape, dtype=tf.float32)
+ )
+
+ # Convert the model
+ def representative_dataset():
+ for _ in range(100):
+ data = np.random.rand(*tuple(ifm_shape))
+ data2 = np.random.rand(*tuple(ifm2_shape)) * 2
+ yield [data.astype(np.float32), data2.astype(np.float32)]
+
+ converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
+ converter.optimizations = [tf.lite.Optimize.DEFAULT]
+ converter.representative_dataset = representative_dataset
+ converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+ converter.inference_input_type = tf.int8
+ converter.inference_output_type = tf.int8
+ tflite_model = converter.convert()
+ return tflite_model
+
+ def verify(ext_func):
+ out_shape = ifm2_shape if reversed_operands else ifm_shape
+ shapes = [ifm_shape, ifm2_shape]
+ ifm_index, ifm2_index = (1, 0) if reversed_operands else (0, 1)
+ op = ext_func.body
+ assert list(op.args[0].checked_type.shape) == shapes[ifm_index]
+ assert list(op.args[1].checked_type.shape) == shapes[ifm2_index]
+ assert op.args[0].checked_type.dtype == dtype
+ assert list(op.checked_type.shape) == out_shape
+ assert op.checked_type.dtype == dtype
+ assert op.attrs.operator_type == operator_type
+ assert op.attrs.reversed_operands == reversed_operands
+ if activation_function == "RELU":
+ assert str(op.attrs.activation) == "CLIP"
+
+ if operator_type == "ADD":
+ rewriter = legalize.AddRewriter()
+ pattern_table = [
+ (
+ ethosu.AddParams.composite_name,
+ ethosu.qnn_add_pattern(),
+ lambda pat: ethosu.AddParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "SUB":
+ rewriter = legalize.SubRewriter()
+ pattern_table = [
+ (
+ ethosu.SubParams.composite_name,
+ ethosu.qnn_subtract_pattern(),
+ lambda pat: ethosu.SubParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MUL":
+ rewriter = legalize.MulRewriter()
+ pattern_table = [
+ (
+ ethosu.MulParams.composite_name,
+ ethosu.qnn_mul_pattern(),
+ lambda pat: ethosu.MulParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MIN":
+ rewriter = legalize.MinRewriter()
+ pattern_table = [
+ (
+ ethosu.MinParams.composite_name,
+ ethosu.minimum_pattern(),
+ lambda pat: ethosu.MinParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MAX":
+ rewriter = legalize.MaxRewriter()
+ pattern_table = [
+ (
+ ethosu.MaxParams.composite_name,
+ ethosu.maximum_pattern(),
+ lambda pat: ethosu.MaxParams(pat).is_valid(),
+ ),
+ ]
+
+ tflite_graph = create_tflite_graph()
+ tflite_model = tflite.Model.Model.GetRootAsModel(tflite_graph, 0)
+
+ mod, _ = relay.frontend.from_tflite(
+ tflite_model,
+ shape_dict={"x": ifm_shape, "y": ifm2_shape},
+ dtype_dict={"x": dtype, "y": dtype},
+ )
+ mod = partition_ethosu_by_table(mod, pattern_table)
+
+ mod["tvmgen_default_ethosu_main_0"] = dataflow_pattern.rewrite(
+ rewriter, mod["tvmgen_default_ethosu_main_0"]
+ )
+ verify(mod["tvmgen_default_ethosu_main_0"])
+
+
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 4], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 4], True),
+ ],
+)
+def test_ethosu_left_shift_binary_elemwise_legalize(ifm_shape, ifm2_shape, reversed_operands):
+ dtype = "int32"
+ operator_type = "SHL"
+
+ def create_graph():
+ input1 = relay.var("x1", shape=ifm_shape, dtype=dtype)
+ input2 = relay.var("x2", shape=ifm2_shape, dtype=dtype)
+ c1 = relay.left_shift(input1, input2)
Review comment:
@ekalda Currently, all the NPU primitive operators are covered.
Because the NPU left shift operator does the "actual" left shift, we legalize the Relay left shift operator.
Conversely, the NPU right shift operator is actually a rounding right shift operator and so the Relay right shift operator cannot be legalized.
@lhutton1 The Relay right shift operator cannot be legalized to a rounding right shift followed by an element-wise subtract of 1.
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] NicolaLancellotti commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
NicolaLancellotti commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r743012784
##########
File path: python/tvm/relay/backend/contrib/ethosu/te/binary_elementwise.py
##########
@@ -0,0 +1,169 @@
+# Licensed to the Apache Software Foundation (ASF) under one
+# or more contributor license agreements. See the NOTICE file
+# distributed with this work for additional information
+# regarding copyright ownership. The ASF licenses this file
+# to you under the Apache License, Version 2.0 (the
+# "License"); you may not use this file except in compliance
+# with the License. You may obtain a copy of the License at
+#
+# http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing,
+# software distributed under the License is distributed on an
+# "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY
+# KIND, either express or implied. See the License for the
+# specific language governing permissions and limitations
+# under the License.
+# pylint: disable=invalid-name,unused-argument
+"""Tensor Expressions for binary_elementwise"""
+import operator
+from tvm import te
+from .dma import dma_ofm_compute, dma_ifm_compute
+
+
+def binary_elementwise_compute(
+ ifm: te.Tensor,
+ ifm2: te.Tensor,
+ lut: te.Tensor,
+ operator_type: str,
+ ifm_scale: float,
+ ifm_zero_point: int,
+ ifm2_scale: float,
+ ifm2_zero_point: int,
+ ofm_scale: float,
+ ofm_zero_point: int,
+ ofm_channels: int,
+ reversed_operands: bool,
+ activation: str,
+ clip_min: int,
+ clip_max: int,
+ ifm_layout: str,
+ ifm2_layout: str,
+ ofm_layout: str,
+) -> te.Tensor:
+ """A compute operator representing the capabilities of binary_elementwise for the NPU.
+
+ Parameters
+ ----------
+ ifm : te.Tensor
+ The Input Feature Map tensor (IFM).
+ ifm2 : te.Tensor
+ The Input Feature Map tensor 1 (IFM2).
+ lut : te.Tensor
+ The look-up table values to use if activation = "LUT".
+ operator_type: str
+ The type of the binary elementwise operator.
+ "ADD"
+ "SUB"
+ "MUL"
+ "MIN"
+ "MAX"
+ "SHR"
+ "SHL"
+ ifm_scale : float
+ The quantization scale for the Input Feature Map tensor.
+ ifm_zero_point : int
+ The quantization zero point for the Input Feature Map tensor.
+ ifm2_scale : float
+ The quantization scale for the Input Feature Map tensor 1.
Review comment:
Done
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] mbaret commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
mbaret commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r743845777
##########
File path: python/tvm/relay/backend/contrib/ethosu/legalize.py
##########
@@ -423,11 +640,20 @@ class LegalizeEthosU:
def transform_module(
self, mod: tvm.ir.IRModule, ctx: tvm.ir.transform.PassContext
) -> tvm.ir.IRModule:
+ """This is the method that replace the operations with hardware/codegen supported
Review comment:
that replaces
##########
File path: tests/python/contrib/test_ethosu/test_codegen.py
##########
@@ -343,5 +343,256 @@ def representative_dataset():
infra.verify_source(compiled_models, accel_type)
+@pytest.mark.parametrize("accel_type", ACCEL_TYPES)
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4]),
+ ([1, 2, 3, 4], [1, 1, 3, 1]),
+ ([1, 1, 3, 1], [1, 2, 3, 4]),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_ethosu_binary_elementwise(
+ accel_type,
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ tf.config.run_functions_eagerly(True)
+
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, lhs, rhs):
+ if operator_type == "ADD":
+ op = tf.math.add(lhs, rhs)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(lhs, rhs)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(lhs, rhs)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(lhs, rhs)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(lhs, rhs)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
Review comment:
Not necessarily something to resolve in this patch, but I wonder if we can abstract some of the boilerplate code here into a utility function.
##########
File path: python/tvm/relay/op/contrib/ethosu.py
##########
@@ -458,6 +458,316 @@ def qnn_avgpool2d_pattern() -> tvm.relay.dataflow_pattern.DFPattern:
return pattern
+class BinaryElementwiseParams:
+ """
+ This class will parse a call to a ethosu.binary_elementwise composite function
+ and extract the parameter information.
+ """
+
+ def __init__(self, func_body: Call, operator_type: str, has_quantization_parameters: bool):
+ clip = None
+ if str(func_body.op) == "clip":
+ clip = func_body
+ binary_op = clip.args[0]
+ else:
+ binary_op = func_body
+
+ layout = "NHWC"
+
+ if has_quantization_parameters:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm_zero_point.value],
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ifm2_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ifm2_zero_point.value],
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ binary_op.args[BinaryElementwiseArgs.ofm_scale.value],
+ binary_op.args[BinaryElementwiseArgs.ofm_zero_point.value],
+ )
+ else:
+ self.ifm = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm.value],
+ layout,
+ )
+ self.ifm2 = TensorParams(
+ binary_op.args[BinaryElementwiseArgs.ifm2.value],
+ layout,
+ )
+ self.ofm = TensorParams(
+ binary_op,
+ layout,
+ )
+ self.activation = clip
+ self.operator_type = operator_type
+
+ def brodcastable(x, y):
Review comment:
broadcastable (or 'can_broadcast' might be better)
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org
[GitHub] [tvm] lhutton1 commented on a change in pull request #9442: Arm(R) Ethos(TM)-U NPU BinaryElementwise operators support
Posted by GitBox <gi...@apache.org>.
lhutton1 commented on a change in pull request #9442:
URL: https://github.com/apache/tvm/pull/9442#discussion_r747579851
##########
File path: tests/python/contrib/test_ethosu/test_legalize.py
##########
@@ -558,5 +558,193 @@ def verify(ext_func):
verify(mod["tvmgen_default_ethosu_main_0"])
+@pytest.mark.parametrize("operator_type", ["ADD", "SUB", "MUL", "MIN", "MAX"])
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 4], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 4], True),
+ ],
+)
+@pytest.mark.parametrize("activation_function", ["NONE", "RELU"])
+def test_tflite_binary_elemwise_legalize(
+ operator_type,
+ ifm_shape,
+ ifm2_shape,
+ reversed_operands,
+ activation_function,
+):
+ dtype = "int8"
+
+ def create_tflite_graph():
+ class Model(tf.Module):
+ @tf.function
+ def tf_function(self, x, y):
+ if operator_type == "ADD":
+ op = tf.math.add(x, y)
+ elif operator_type == "SUB":
+ op = tf.math.subtract(x, y)
+ elif operator_type == "MUL":
+ op = tf.math.multiply(x, y)
+ elif operator_type == "MIN":
+ op = tf.math.minimum(x, y)
+ elif operator_type == "MAX":
+ op = tf.math.maximum(x, y)
+ if activation_function == "RELU":
+ op = tf.nn.relu(op)
+ return op
+
+ model = Model()
+ concrete_func = model.tf_function.get_concrete_function(
+ tf.TensorSpec(ifm_shape, dtype=tf.float32), tf.TensorSpec(ifm2_shape, dtype=tf.float32)
+ )
+
+ # Convert the model
+ def representative_dataset():
+ for _ in range(100):
+ data = np.random.rand(*tuple(ifm_shape))
+ data2 = np.random.rand(*tuple(ifm2_shape)) * 2
+ yield [data.astype(np.float32), data2.astype(np.float32)]
+
+ converter = tf.lite.TFLiteConverter.from_concrete_functions([concrete_func])
+ converter.optimizations = [tf.lite.Optimize.DEFAULT]
+ converter.representative_dataset = representative_dataset
+ converter.target_spec.supported_ops = [tf.lite.OpsSet.TFLITE_BUILTINS_INT8]
+ converter.inference_input_type = tf.int8
+ converter.inference_output_type = tf.int8
+ tflite_model = converter.convert()
+ return tflite_model
+
+ def verify(ext_func):
+ out_shape = ifm2_shape if reversed_operands else ifm_shape
+ shapes = [ifm_shape, ifm2_shape]
+ ifm_index, ifm2_index = (1, 0) if reversed_operands else (0, 1)
+ op = ext_func.body
+ assert list(op.args[0].checked_type.shape) == shapes[ifm_index]
+ assert list(op.args[1].checked_type.shape) == shapes[ifm2_index]
+ assert op.args[0].checked_type.dtype == dtype
+ assert list(op.checked_type.shape) == out_shape
+ assert op.checked_type.dtype == dtype
+ assert op.attrs.operator_type == operator_type
+ assert op.attrs.reversed_operands == reversed_operands
+ if activation_function == "RELU":
+ assert str(op.attrs.activation) == "CLIP"
+
+ if operator_type == "ADD":
+ rewriter = legalize.AddRewriter()
+ pattern_table = [
+ (
+ ethosu.AddParams.composite_name,
+ ethosu.qnn_add_pattern(),
+ lambda pat: ethosu.AddParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "SUB":
+ rewriter = legalize.SubRewriter()
+ pattern_table = [
+ (
+ ethosu.SubParams.composite_name,
+ ethosu.qnn_subtract_pattern(),
+ lambda pat: ethosu.SubParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MUL":
+ rewriter = legalize.MulRewriter()
+ pattern_table = [
+ (
+ ethosu.MulParams.composite_name,
+ ethosu.qnn_mul_pattern(),
+ lambda pat: ethosu.MulParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MIN":
+ rewriter = legalize.MinRewriter()
+ pattern_table = [
+ (
+ ethosu.MinParams.composite_name,
+ ethosu.minimum_pattern(),
+ lambda pat: ethosu.MinParams(pat).is_valid(),
+ ),
+ ]
+ elif operator_type == "MAX":
+ rewriter = legalize.MaxRewriter()
+ pattern_table = [
+ (
+ ethosu.MaxParams.composite_name,
+ ethosu.maximum_pattern(),
+ lambda pat: ethosu.MaxParams(pat).is_valid(),
+ ),
+ ]
+
+ tflite_graph = create_tflite_graph()
+ tflite_model = tflite.Model.Model.GetRootAsModel(tflite_graph, 0)
+
+ mod, _ = relay.frontend.from_tflite(
+ tflite_model,
+ shape_dict={"x": ifm_shape, "y": ifm2_shape},
+ dtype_dict={"x": dtype, "y": dtype},
+ )
+ mod = partition_ethosu_by_table(mod, pattern_table)
+
+ mod["tvmgen_default_ethosu_main_0"] = dataflow_pattern.rewrite(
+ rewriter, mod["tvmgen_default_ethosu_main_0"]
+ )
+ verify(mod["tvmgen_default_ethosu_main_0"])
+
+
+@pytest.mark.parametrize(
+ "ifm_shape, ifm2_shape, reversed_operands",
+ [
+ ([1, 2, 3, 4], [1, 2, 3, 4], False),
+ ([1, 2, 3, 4], [1, 1, 3, 1], False),
+ ([1, 1, 3, 1], [1, 2, 3, 4], True),
+ ],
+)
+def test_ethosu_left_shift_binary_elemwise_legalize(ifm_shape, ifm2_shape, reversed_operands):
+ dtype = "int32"
+ operator_type = "SHL"
+
+ def create_graph():
+ input1 = relay.var("x1", shape=ifm_shape, dtype=dtype)
+ input2 = relay.var("x2", shape=ifm2_shape, dtype=dtype)
+ c1 = relay.left_shift(input1, input2)
Review comment:
I think we could follow-up on this in another PR
--
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.
To unsubscribe, e-mail: commits-unsubscribe@tvm.apache.org
For queries about this service, please contact Infrastructure at:
users@infra.apache.org