SebastianBoblest commented on code in PR #12353:
URL: https://github.com/apache/tvm/pull/12353#discussion_r942434869
##########
python/tvm/relay/op/contrib/cmsisnn.py:
##########
@@ -122,6 +130,25 @@ def check_qnn_conv2d(pattern):
conv2d_input = conv2d.args[0]
conv2d_weight = conv2d.args[1]
+ # check if sum of paddings from pad() and conv2d() satisfies CMSIS-NN
constraints
+ can_pad_be_fused = True
+ if isinstance(conv2d_input, tvm.relay.expr.Call) and
str(conv2d_input.op.name) == "nn.pad":
+ data_layout = conv2d.attrs.data_layout
+ conv2d_padding = conv2d.attrs.padding # (top, left, bottom, right)
+ pad = conv2d_input
+ pad_width = pad.attrs.pad_width
+ pad_width_w = pad_width[data_layout.find("W")] # (left, right)
+ pad_width_h = pad_width[data_layout.find("H")] # (top, bottom)
+ # calculate effective padding post pad fusion
+ pad_top = pad_width_h[0] + conv2d_padding[0]
+ pad_bottom = pad_width_h[1] + conv2d_padding[2]
+ pad_left = pad_width_w[0] + conv2d_padding[1]
+ pad_right = pad_width_w[1] + conv2d_padding[3]
+ # check if difference in the side paddings is 1 along each
dimension
+ pad_w_diff = int(pad_right - pad_left)
Review Comment:
Why is this cast to int necessary? What are the data types of pad_right and
pad_left?
##########
tests/python/contrib/test_cmsisnn/test_conv2d.py:
##########
@@ -403,6 +401,109 @@ def test_conv2d_asymmetric_padding_int8(
)
[email protected]_cmsisnn
[email protected]("ifm_shape", [(1, 25, 25, 12), (1, 64, 100, 4)])
[email protected](
+ "pad_width",
+ [
+ ((0, 0), (0, 1), (1, 2), (0, 0)),
+ ((0, 0), (1, 1), (1, 1), (0, 0)),
+ ((0, 0), (2, 2), (3, 4), (0, 0)),
+ ],
+)
+def test_conv2d_separate_padding_int8(
+ ifm_shape,
+ pad_width,
+):
+ """Tests QNN Conv2D where the padding is asymmetric on different sides of
input"""
+ interface_api = "c"
+ use_unpacked_api = True
+ test_runner = AOT_USMP_CORSTONE300_RUNNER
+
+ ifm_shape = (1, 25, 25, 12)
+ kernel_size = (5, 5)
+ strides = (2, 2)
+ dilation = (1, 1)
+ padding = "SAME"
+ dtype = "int8"
+ enable_bias = True
+ relu_type = "NONE"
+ input_zero_point = 10
+ input_scale = 0.0128
+ kernel_scale = [0.11, 0.22]
+ out_channels = 2
+ groups = 1
+ weight_format = "HWIO"
+ kernel_h = kernel_size[0]
+ kernel_w = kernel_size[1]
+ kernel_shape = (kernel_h, kernel_w, ifm_shape[3] // groups, out_channels)
+ kernel_zero_point = 0
+ in_min, in_max = get_range_for_dtype_str(dtype)
+
+ output_scale, output_zero_point = get_conv2d_qnn_params(
+ kernel_shape,
+ input_scale,
+ input_zero_point,
+ kernel_scale,
+ kernel_zero_point,
+ dtype,
Review Comment:
I was not familiar with the signature of get_conv2d_qnn_params, so for me it
would be helpful to have
```suggestion
input_dtype=dtype,
```
##########
tests/python/contrib/test_cmsisnn/test_conv2d.py:
##########
@@ -403,6 +401,109 @@ def test_conv2d_asymmetric_padding_int8(
)
[email protected]_cmsisnn
[email protected]("ifm_shape", [(1, 25, 25, 12), (1, 64, 100, 4)])
[email protected](
+ "pad_width",
+ [
+ ((0, 0), (0, 1), (1, 2), (0, 0)),
+ ((0, 0), (1, 1), (1, 1), (0, 0)),
+ ((0, 0), (2, 2), (3, 4), (0, 0)),
+ ],
+)
+def test_conv2d_separate_padding_int8(
+ ifm_shape,
+ pad_width,
+):
+ """Tests QNN Conv2D where the padding is asymmetric on different sides of
input"""
+ interface_api = "c"
+ use_unpacked_api = True
+ test_runner = AOT_USMP_CORSTONE300_RUNNER
+
+ ifm_shape = (1, 25, 25, 12)
+ kernel_size = (5, 5)
+ strides = (2, 2)
+ dilation = (1, 1)
+ padding = "SAME"
+ dtype = "int8"
+ enable_bias = True
+ relu_type = "NONE"
+ input_zero_point = 10
+ input_scale = 0.0128
+ kernel_scale = [0.11, 0.22]
+ out_channels = 2
+ groups = 1
+ weight_format = "HWIO"
+ kernel_h = kernel_size[0]
+ kernel_w = kernel_size[1]
+ kernel_shape = (kernel_h, kernel_w, ifm_shape[3] // groups, out_channels)
+ kernel_zero_point = 0
+ in_min, in_max = get_range_for_dtype_str(dtype)
+
+ output_scale, output_zero_point = get_conv2d_qnn_params(
+ kernel_shape,
+ input_scale,
+ input_zero_point,
+ kernel_scale,
+ kernel_zero_point,
+ dtype,
+ dtype,
Review Comment:
```suggestion
weights_dtype=dtype,
```
##########
tests/python/contrib/test_cmsisnn/test_fuse_pads.py:
##########
@@ -0,0 +1,285 @@
+# 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.
+
+"""CMSIS-NN integration tests: fuse_pads pass"""
+import numpy as np
+import pytest
+import tvm
+import tvm.testing
+from tvm import relay
+
+tvm._ffi._init_api("relay.ext.cmsisnn.transform", __name__)
+
+
+class CheckForPadsWithinCompositeFunc(tvm.relay.ExprVisitor):
+ """Provides methods to test number of constants present in a function"""
+
+ def __init__(self):
+ super().__init__()
+ self.num_pads_ = 0
+
+ def visit_call(self, call):
+ super().visit_call(call)
+ if (
+ isinstance(call, tvm.relay.Call)
+ and isinstance(call.op, tvm.ir.op.Op)
+ and call.op.name == "nn.pad"
+ ):
+ self.num_pads_ += 1
+
+ def check_num_pads(self):
+ assert self.num_pads_ == 0, "CMSIS-NN composite function should not
have pads"
+
+
+def set_external_func_attr(func, compiler, ext_symbol):
+ func = func.with_attr("Primitive", tvm.tir.IntImm("int32", 1))
+ func = func.with_attr("Compiler", compiler)
+ func = func.with_attr("global_symbol", ext_symbol)
+ return func
+
+
+def set_composite_func_attr(func, name):
+ func = func.with_attr("Composite", name)
+ return func
+
+
[email protected](
+ "ifm_shape, pad_width, conv2d_padding, ofm_shape",
+ [
+ [(1, 25, 25, 12), ((0, 0), (0, 2), (1, 2), (0, 0)), (1, 1, 1, 1), (1,
26, 28, 2)],
+ [(1, 64, 100, 4), ((0, 0), (1, 3), (1, 1), (0, 0)), (0, 0, 0, 0), (1,
64, 100, 2)],
+ [(1, 55, 55, 3), ((0, 0), (2, 1), (3, 5), (0, 0)), (0, 0, 1, 1), (1,
57, 59, 2)],
+ ],
+)
+def test_invalid_padding_for_fusion(ifm_shape, pad_width, conv2d_padding,
ofm_shape):
+ """Tests the pass ExternConstants when a composite function
+ is present within global function
+ """
+ dtype = "int8"
+ kernel_size = (3, 3)
+ ofm_channels = 2
+ local_input = relay.var("local_input", shape=ifm_shape, dtype=dtype)
+ pad = relay.nn.pad(
+ local_input,
+ pad_width=pad_width, # ((), (top, bottom), (left, right), ())
+ pad_value=10,
+ pad_mode="constant",
+ )
+ rng = np.random.default_rng(12321)
+ local_weight = tvm.nd.array(
+ rng.integers(
+ np.iinfo(dtype).min,
+ high=np.iinfo(dtype).max,
+ size=(ofm_channels, kernel_size[0], kernel_size[1], ifm_shape[3]),
+ dtype=dtype,
+ )
+ )
+ local_weight = relay.const(local_weight, dtype)
+ conv2d = relay.qnn.op.conv2d(
+ pad,
+ local_weight,
+ relay.const(1, "int32"),
+ relay.const(1, "int32"),
+ relay.const(1, "float32"),
+ relay.const(1, "float32"),
+ data_layout="NHWC",
+ kernel_layout="OHWI",
+ channels=ofm_channels,
+ kernel_size=(3, 3),
+ padding=conv2d_padding,
+ out_dtype="int32",
+ )
+ requantize = relay.qnn.op.requantize(
+ conv2d,
+ relay.const(1, "float32"),
+ relay.const(1, "int32"),
+ relay.const(1, "float32"),
+ relay.const(1, "int32"),
+ axis=0,
+ out_dtype=dtype,
+ )
+ local_func = relay.Function(relay.analysis.free_vars(requantize),
requantize)
+ local_func = set_composite_func_attr(local_func, "cmsis-nn.qnn_conv2d")
+
+ mod = tvm.IRModule()
+ ext_input = relay.var("ext_input", shape=ifm_shape, dtype=dtype)
+ call_local_func = relay.Call(local_func, [ext_input])
+ extern_func = relay.Function(relay.analysis.free_vars(call_local_func),
call_local_func)
+ extern_var = relay.GlobalVar("external_function")
+ extern_func = set_external_func_attr(extern_func, "cmsis-nn",
extern_var.name_hint)
+ mod[extern_var] = extern_func
+
+ main_input = relay.var("main_input", shape=ifm_shape, dtype=dtype)
+ call_extern_func = relay.Call(extern_var, [main_input])
+ main_func = relay.Function([main_input], call_extern_func,
relay.TensorType(ofm_shape, dtype))
+ main_var = relay.GlobalVar("main")
+ mod[main_var] = main_func
+
+ mod = relay.transform.InferType()(mod)
+
+ error_regex = r"Difference on each side of a dimension should be either 0
or 1"
+
+ with pytest.raises(tvm.TVMError, match=error_regex):
+ mod = CMSISNNFusePads()(mod)
+
+
[email protected](
+ "ifm_shape, pad_width, conv2d_padding, ofm_shape",
+ [
+ [(1, 25, 25, 12), ((0, 0), (0, 1), (1, 2), (0, 0)), (1, 1, 1, 1), (1,
26, 28, 2)],
+ [(1, 64, 100, 4), ((0, 0), (1, 1), (1, 1), (0, 0)), (0, 0, 0, 0), (1,
64, 100, 2)],
+ [(1, 55, 55, 3), ((0, 0), (2, 1), (3, 2), (0, 0)), (0, 0, 1, 1), (1,
57, 59, 2)],
+ ],
+)
+def test_pad_conv2d_fusion(ifm_shape, pad_width, conv2d_padding, ofm_shape):
+ """Tests the pass ExternConstants when a composite function
+ is present within global function
+ """
+ dtype = "int8"
+ kernel_size = (3, 3)
+ ofm_channels = 2
+ local_input = relay.var("local_input", shape=ifm_shape, dtype=dtype)
+ pad = relay.nn.pad(
+ local_input,
+ pad_width=pad_width, # ((), (top, bottom), (left, right), ())
+ pad_value=10,
+ pad_mode="constant",
+ )
+ rng = np.random.default_rng(12321)
+ local_weight = tvm.nd.array(
+ rng.integers(
+ np.iinfo(dtype).min,
+ high=np.iinfo(dtype).max,
+ size=(ofm_channels, kernel_size[0], kernel_size[1], ifm_shape[3]),
+ dtype=dtype,
+ )
+ )
+ local_weight = relay.const(local_weight, dtype)
+ conv2d = relay.qnn.op.conv2d(
+ pad,
+ local_weight,
+ relay.const(1, "int32"),
+ relay.const(1, "int32"),
+ relay.const(1, "float32"),
+ relay.const(1, "float32"),
+ data_layout="NHWC",
+ kernel_layout="OHWI",
+ channels=ofm_channels,
+ kernel_size=(3, 3),
+ padding=conv2d_padding,
+ out_dtype="int32",
+ )
+ requantize = relay.qnn.op.requantize(
+ conv2d,
+ relay.const(1, "float32"),
+ relay.const(1, "int32"),
+ relay.const(1, "float32"),
+ relay.const(1, "int32"),
+ axis=0,
+ out_dtype=dtype,
+ )
+ local_func = relay.Function(relay.analysis.free_vars(requantize),
requantize)
+ local_func = set_composite_func_attr(local_func, "cmsis-nn.qnn_conv2d")
+
+ mod = tvm.IRModule()
+ ext_input = relay.var("ext_input", shape=ifm_shape, dtype=dtype)
+ call_local_func = relay.Call(local_func, [ext_input])
+ extern_func = relay.Function(relay.analysis.free_vars(call_local_func),
call_local_func)
+ extern_var = relay.GlobalVar("external_function")
+ extern_func = set_external_func_attr(extern_func, "cmsis-nn",
extern_var.name_hint)
+ mod[extern_var] = extern_func
+
+ main_input = relay.var("main_input", shape=ifm_shape, dtype=dtype)
+ call_extern_func = relay.Call(extern_var, [main_input])
+ main_func = relay.Function([main_input], call_extern_func,
relay.TensorType(ofm_shape, dtype))
+ main_var = relay.GlobalVar("main")
+ mod[main_var] = main_func
+
+ mod = relay.transform.InferType()(mod)
+
+ mod = CMSISNNFusePads()(mod)
+ pad_verifier = CheckForPadsWithinCompositeFunc()
+ pad_verifier.visit_function(mod[extern_var])
+ pad_verifier.check_num_pads()
+
+
+def test_without_preceding_pad():
+ """Tests the pass ExternConstants when a composite function
Review Comment:
This test and the two preceding ones have identical doc strings. Could you
maybe add one sencentence to make the doc strings more informative?
##########
tests/python/contrib/test_cmsisnn/test_conv2d.py:
##########
@@ -403,6 +401,109 @@ def test_conv2d_asymmetric_padding_int8(
)
[email protected]_cmsisnn
[email protected]("ifm_shape", [(1, 25, 25, 12), (1, 64, 100, 4)])
[email protected](
+ "pad_width",
+ [
+ ((0, 0), (0, 1), (1, 2), (0, 0)),
+ ((0, 0), (1, 1), (1, 1), (0, 0)),
+ ((0, 0), (2, 2), (3, 4), (0, 0)),
+ ],
+)
+def test_conv2d_separate_padding_int8(
+ ifm_shape,
+ pad_width,
+):
+ """Tests QNN Conv2D where the padding is asymmetric on different sides of
input"""
+ interface_api = "c"
+ use_unpacked_api = True
+ test_runner = AOT_USMP_CORSTONE300_RUNNER
+
+ ifm_shape = (1, 25, 25, 12)
+ kernel_size = (5, 5)
+ strides = (2, 2)
+ dilation = (1, 1)
+ padding = "SAME"
+ dtype = "int8"
+ enable_bias = True
+ relu_type = "NONE"
+ input_zero_point = 10
+ input_scale = 0.0128
+ kernel_scale = [0.11, 0.22]
+ out_channels = 2
+ groups = 1
+ weight_format = "HWIO"
+ kernel_h = kernel_size[0]
+ kernel_w = kernel_size[1]
+ kernel_shape = (kernel_h, kernel_w, ifm_shape[3] // groups, out_channels)
+ kernel_zero_point = 0
+ in_min, in_max = get_range_for_dtype_str(dtype)
+
+ output_scale, output_zero_point = get_conv2d_qnn_params(
+ kernel_shape,
+ input_scale,
+ input_zero_point,
+ kernel_scale,
+ kernel_zero_point,
+ dtype,
+ dtype,
+ dtype,
+ )
+
+ invar = relay.var("input", shape=ifm_shape, dtype=dtype)
+ pad = relay.nn.pad(
+ invar,
+ pad_width=pad_width, # ((), (top, bottom), (left, right), ())
+ pad_value=input_zero_point,
+ pad_mode="constant",
+ )
+
+ model, params = make_model(
+ ifm_shape,
+ kernel_shape,
+ input_zero_point,
+ input_scale,
+ kernel_zero_point,
+ kernel_scale,
+ output_zero_point,
+ output_scale,
+ padding,
+ strides,
+ dilation,
+ groups,
+ dtype,
Review Comment:
Here, keyword argument syntax could make this a bit more readable
##########
tests/python/contrib/test_cmsisnn/test_conv2d.py:
##########
@@ -403,6 +401,109 @@ def test_conv2d_asymmetric_padding_int8(
)
[email protected]_cmsisnn
[email protected]("ifm_shape", [(1, 25, 25, 12), (1, 64, 100, 4)])
[email protected](
+ "pad_width",
+ [
+ ((0, 0), (0, 1), (1, 2), (0, 0)),
+ ((0, 0), (1, 1), (1, 1), (0, 0)),
+ ((0, 0), (2, 2), (3, 4), (0, 0)),
+ ],
+)
+def test_conv2d_separate_padding_int8(
+ ifm_shape,
+ pad_width,
+):
+ """Tests QNN Conv2D where the padding is asymmetric on different sides of
input"""
+ interface_api = "c"
+ use_unpacked_api = True
+ test_runner = AOT_USMP_CORSTONE300_RUNNER
+
+ ifm_shape = (1, 25, 25, 12)
+ kernel_size = (5, 5)
+ strides = (2, 2)
+ dilation = (1, 1)
+ padding = "SAME"
+ dtype = "int8"
+ enable_bias = True
+ relu_type = "NONE"
+ input_zero_point = 10
+ input_scale = 0.0128
+ kernel_scale = [0.11, 0.22]
+ out_channels = 2
+ groups = 1
+ weight_format = "HWIO"
+ kernel_h = kernel_size[0]
+ kernel_w = kernel_size[1]
+ kernel_shape = (kernel_h, kernel_w, ifm_shape[3] // groups, out_channels)
+ kernel_zero_point = 0
+ in_min, in_max = get_range_for_dtype_str(dtype)
+
+ output_scale, output_zero_point = get_conv2d_qnn_params(
+ kernel_shape,
+ input_scale,
+ input_zero_point,
+ kernel_scale,
+ kernel_zero_point,
+ dtype,
+ dtype,
+ dtype,
Review Comment:
```suggestion
output_dtype=dtype,
```
##########
tests/python/contrib/test_cmsisnn/utils.py:
##########
@@ -50,8 +50,19 @@ def visit_call(self, call):
return counter.count
-def assert_partitioned_function(orig_mod, cmsisnn_mod):
- """If kCompiler attribute is missing, this function raises assertion"""
+def assert_partitioned_function(orig_mod, cmsisnn_mod, is_num_calls_same=True):
Review Comment:
```suggestion
def assert_partitioned_function(orig_mod, cmsisnn_mod,
num_call_nodes_expected_unchanged=True):
```
##########
tests/python/contrib/test_cmsisnn/test_fuse_pads.py:
##########
@@ -0,0 +1,285 @@
+# 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.
+
+"""CMSIS-NN integration tests: fuse_pads pass"""
+import numpy as np
+import pytest
+import tvm
+import tvm.testing
+from tvm import relay
+
+tvm._ffi._init_api("relay.ext.cmsisnn.transform", __name__)
+
+
+class CheckForPadsWithinCompositeFunc(tvm.relay.ExprVisitor):
+ """Provides methods to test number of constants present in a function"""
Review Comment:
```suggestion
"""Provides methods to test how many constants are present in a
function"""
```
--
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: [email protected]
For queries about this service, please contact Infrastructure at:
[email protected]
